arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-12-17 20:28:13 +00:00
Code Issues Releases Wiki Activity
811afddeda
hashcat/OpenCL/common.c

9234 lines
206 KiB
C
Raw Normal View History

- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
/**
* Author......: Jens Steube <jens.steube@gmail.com>
* License.....: MIT
*/
Fix append_* function calls
2015-12-15 12:42:37 +00:00
static int hash_comp (const u32 d1[4], __global u32 *d2)
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
{
if (d1[3] > d2[DGST_R3]) return ( 1);
if (d1[3] < d2[DGST_R3]) return (-1);
if (d1[2] > d2[DGST_R2]) return ( 1);
if (d1[2] < d2[DGST_R2]) return (-1);
if (d1[1] > d2[DGST_R1]) return ( 1);
if (d1[1] < d2[DGST_R1]) return (-1);
if (d1[0] > d2[DGST_R0]) return ( 1);
if (d1[0] < d2[DGST_R0]) return (-1);
return (0);
}
static int find_hash (const u32 digest[4], const u32 digests_cnt, __global digest_t *digests_buf)
{
for (u32 l = 0, r = digests_cnt; r; r >>= 1)
{
const u32 m = r >> 1;
const u32 c = l + m;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
const int cmp = hash_comp (digest, digests_buf[c].digest_buf);
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
if (cmp > 0)
{
l += m + 1;
r--;
}
if (cmp == 0) return (c);
}
return (-1);
}
static u32 check_bitmap (__global u32 *bitmap, const u32 bitmap_mask, const u32 bitmap_shift, const u32 digest)
{
return (bitmap[(digest >> bitmap_shift) & bitmap_mask] & (1 << (digest & 0x1f)));
}
static u32 check (const u32 digest[2], __global u32 *bitmap_s1_a, __global u32 *bitmap_s1_b, __global u32 *bitmap_s1_c, __global u32 *bitmap_s1_d, __global u32 *bitmap_s2_a, __global u32 *bitmap_s2_b, __global u32 *bitmap_s2_c, __global u32 *bitmap_s2_d, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2)
{
if (check_bitmap (bitmap_s1_a, bitmap_mask, bitmap_shift1, digest[0]) == 0) return (0);
if (check_bitmap (bitmap_s1_b, bitmap_mask, bitmap_shift1, digest[1]) == 0) return (0);
if (check_bitmap (bitmap_s1_c, bitmap_mask, bitmap_shift1, digest[2]) == 0) return (0);
if (check_bitmap (bitmap_s1_d, bitmap_mask, bitmap_shift1, digest[3]) == 0) return (0);
if (check_bitmap (bitmap_s2_a, bitmap_mask, bitmap_shift2, digest[0]) == 0) return (0);
if (check_bitmap (bitmap_s2_b, bitmap_mask, bitmap_shift2, digest[1]) == 0) return (0);
if (check_bitmap (bitmap_s2_c, bitmap_mask, bitmap_shift2, digest[2]) == 0) return (0);
if (check_bitmap (bitmap_s2_d, bitmap_mask, bitmap_shift2, digest[3]) == 0) return (0);
return (1);
}
static void mark_hash (__global plain_t *plains_buf, __global u32 *hashes_shown, const int hash_pos, const u32 gid, const u32 il_pos)
{
hashes_shown[hash_pos] = 1;
plains_buf[hash_pos].gidvid = (gid * 1) + 0;
plains_buf[hash_pos].il_pos = il_pos;
}
static void truncate_block (u32 w[4], const u32 len)
{
switch (len)
{
case 0: w[0] &= 0;
w[1] &= 0;
w[2] &= 0;
w[3] &= 0;
break;
case 1: w[0] &= 0x000000FF;
w[1] &= 0;
w[2] &= 0;
w[3] &= 0;
break;
case 2: w[0] &= 0x0000FFFF;
w[1] &= 0;
w[2] &= 0;
w[3] &= 0;
break;
case 3: w[0] &= 0x00FFFFFF;
w[1] &= 0;
w[2] &= 0;
w[3] &= 0;
break;
case 4: w[1] &= 0;
w[2] &= 0;
w[3] &= 0;
break;
case 5: w[1] &= 0x000000FF;
w[2] &= 0;
w[3] &= 0;
break;
case 6: w[1] &= 0x0000FFFF;
w[2] &= 0;
w[3] &= 0;
break;
case 7: w[1] &= 0x00FFFFFF;
w[2] &= 0;
w[3] &= 0;
break;
case 8: w[2] &= 0;
w[3] &= 0;
break;
case 9: w[2] &= 0x000000FF;
w[3] &= 0;
break;
case 10: w[2] &= 0x0000FFFF;
w[3] &= 0;
break;
case 11: w[2] &= 0x00FFFFFF;
w[3] &= 0;
break;
case 12: w[3] &= 0;
break;
case 13: w[3] &= 0x000000FF;
break;
case 14: w[3] &= 0x0000FFFF;
break;
case 15: w[3] &= 0x00FFFFFF;
break;
}
}
Prepare new SIMD code for kernel, -m 0, 10, 20, 1000 should work in -a 3 mode and other hopefully stay unaffected
2016-01-17 21:17:50 +00:00
static void make_unicode_S (const u32 in[4], u32 out1[4], u32 out2[4])
{
#ifdef IS_NV
out2[3] = __byte_perm_S (in[3], 0, 0x7372);
out2[2] = __byte_perm_S (in[3], 0, 0x7170);
out2[1] = __byte_perm_S (in[2], 0, 0x7372);
out2[0] = __byte_perm_S (in[2], 0, 0x7170);
out1[3] = __byte_perm_S (in[1], 0, 0x7372);
out1[2] = __byte_perm_S (in[1], 0, 0x7170);
out1[1] = __byte_perm_S (in[0], 0, 0x7372);
out1[0] = __byte_perm_S (in[0], 0, 0x7170);
#endif
#if defined IS_AMD || defined IS_GENERIC
out2[3] = ((in[3] >> 8) & 0x00FF0000) | ((in[3] >> 16) & 0x000000FF);
out2[2] = ((in[3] << 8) & 0x00FF0000) | ((in[3] >> 0) & 0x000000FF);
out2[1] = ((in[2] >> 8) & 0x00FF0000) | ((in[2] >> 16) & 0x000000FF);
out2[0] = ((in[2] << 8) & 0x00FF0000) | ((in[2] >> 0) & 0x000000FF);
out1[3] = ((in[1] >> 8) & 0x00FF0000) | ((in[1] >> 16) & 0x000000FF);
out1[2] = ((in[1] << 8) & 0x00FF0000) | ((in[1] >> 0) & 0x000000FF);
out1[1] = ((in[0] >> 8) & 0x00FF0000) | ((in[0] >> 16) & 0x000000FF);
out1[0] = ((in[0] << 8) & 0x00FF0000) | ((in[0] >> 0) & 0x000000FF);
#endif
}
static void make_unicode (const u32x in[4], u32x out1[4], u32x out2[4])
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
{
#ifdef IS_NV
out2[3] = __byte_perm (in[3], 0, 0x7372);
out2[2] = __byte_perm (in[3], 0, 0x7170);
out2[1] = __byte_perm (in[2], 0, 0x7372);
out2[0] = __byte_perm (in[2], 0, 0x7170);
out1[3] = __byte_perm (in[1], 0, 0x7372);
out1[2] = __byte_perm (in[1], 0, 0x7170);
out1[1] = __byte_perm (in[0], 0, 0x7372);
out1[0] = __byte_perm (in[0], 0, 0x7170);
#endif
Fix for -m 1500, -m 3000: Missing "defined" keyword Fix for -m 1000: MD4_H1 copy/paste error Fix for -m 8900, -m 9300: Invalid value for device_processor_cores for CPU devices Fix for -m 9100: Variable salt2 initializer needed some clauses for clearness Temporary limit gpu_accel for CPU devices to 1 for development phase Mark pocl as too bleeding edge for production use, recommend native drivers Remove workarounds for pocl Rename VENDOR_ID_UNKNOWN to VENDOR_ID_GENERIC in host code Rename IS_UNKNOWN to IS_GENERIC in kernel code
2016-01-07 19:14:34 +00:00
#if defined IS_AMD || defined IS_GENERIC
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
out2[3] = ((in[3] >> 8) & 0x00FF0000) | ((in[3] >> 16) & 0x000000FF);
out2[2] = ((in[3] << 8) & 0x00FF0000) | ((in[3] >> 0) & 0x000000FF);
out2[1] = ((in[2] >> 8) & 0x00FF0000) | ((in[2] >> 16) & 0x000000FF);
out2[0] = ((in[2] << 8) & 0x00FF0000) | ((in[2] >> 0) & 0x000000FF);
out1[3] = ((in[1] >> 8) & 0x00FF0000) | ((in[1] >> 16) & 0x000000FF);
out1[2] = ((in[1] << 8) & 0x00FF0000) | ((in[1] >> 0) & 0x000000FF);
out1[1] = ((in[0] >> 8) & 0x00FF0000) | ((in[0] >> 16) & 0x000000FF);
out1[0] = ((in[0] << 8) & 0x00FF0000) | ((in[0] >> 0) & 0x000000FF);
#endif
}
Prepare new SIMD code for kernel, -m 0, 10, 20, 1000 should work in -a 3 mode and other hopefully stay unaffected
2016-01-17 21:17:50 +00:00
static void undo_unicode_S (const u32 in1[4], const u32 in2[4], u32 out[4])
{
#ifdef IS_NV
out[0] = __byte_perm_S (in1[0], in1[1], 0x6420);
out[1] = __byte_perm_S (in1[2], in1[3], 0x6420);
out[2] = __byte_perm_S (in2[0], in2[1], 0x6420);
out[3] = __byte_perm_S (in2[2], in2[3], 0x6420);
#endif
#if defined IS_AMD || defined IS_GENERIC
out[0] = ((in1[0] & 0x000000ff) >> 0) | ((in1[0] & 0x00ff0000) >> 8)
| ((in1[1] & 0x000000ff) << 16) | ((in1[1] & 0x00ff0000) << 8);
out[1] = ((in1[2] & 0x000000ff) >> 0) | ((in1[2] & 0x00ff0000) >> 8)
| ((in1[3] & 0x000000ff) << 16) | ((in1[3] & 0x00ff0000) << 8);
out[2] = ((in2[0] & 0x000000ff) >> 0) | ((in2[0] & 0x00ff0000) >> 8)
| ((in2[1] & 0x000000ff) << 16) | ((in2[1] & 0x00ff0000) << 8);
out[3] = ((in2[2] & 0x000000ff) >> 0) | ((in2[2] & 0x00ff0000) >> 8)
| ((in2[3] & 0x000000ff) << 16) | ((in2[3] & 0x00ff0000) << 8);
#endif
}
static void undo_unicode (const u32x in1[4], const u32x in2[4], u32x out[4])
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
{
#ifdef IS_NV
out[0] = __byte_perm (in1[0], in1[1], 0x6420);
out[1] = __byte_perm (in1[2], in1[3], 0x6420);
out[2] = __byte_perm (in2[0], in2[1], 0x6420);
out[3] = __byte_perm (in2[2], in2[3], 0x6420);
#endif
Fix for -m 1500, -m 3000: Missing "defined" keyword Fix for -m 1000: MD4_H1 copy/paste error Fix for -m 8900, -m 9300: Invalid value for device_processor_cores for CPU devices Fix for -m 9100: Variable salt2 initializer needed some clauses for clearness Temporary limit gpu_accel for CPU devices to 1 for development phase Mark pocl as too bleeding edge for production use, recommend native drivers Remove workarounds for pocl Rename VENDOR_ID_UNKNOWN to VENDOR_ID_GENERIC in host code Rename IS_UNKNOWN to IS_GENERIC in kernel code
2016-01-07 19:14:34 +00:00
#if defined IS_AMD || defined IS_GENERIC
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
out[0] = ((in1[0] & 0x000000ff) >> 0) | ((in1[0] & 0x00ff0000) >> 8)
| ((in1[1] & 0x000000ff) << 16) | ((in1[1] & 0x00ff0000) << 8);
out[1] = ((in1[2] & 0x000000ff) >> 0) | ((in1[2] & 0x00ff0000) >> 8)
| ((in1[3] & 0x000000ff) << 16) | ((in1[3] & 0x00ff0000) << 8);
out[2] = ((in2[0] & 0x000000ff) >> 0) | ((in2[0] & 0x00ff0000) >> 8)
| ((in2[1] & 0x000000ff) << 16) | ((in2[1] & 0x00ff0000) << 8);
out[3] = ((in2[2] & 0x000000ff) >> 0) | ((in2[2] & 0x00ff0000) >> 8)
| ((in2[3] & 0x000000ff) << 16) | ((in2[3] & 0x00ff0000) << 8);
#endif
}
static void append_0x01_1x4 (u32 w0[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x01;
break;
case 1:
w0[0] = w0[0] | 0x0100;
break;
case 2:
w0[0] = w0[0] | 0x010000;
break;
case 3:
w0[0] = w0[0] | 0x01000000;
break;
case 4:
w0[1] = 0x01;
break;
case 5:
w0[1] = w0[1] | 0x0100;
break;
case 6:
w0[1] = w0[1] | 0x010000;
break;
case 7:
w0[1] = w0[1] | 0x01000000;
break;
case 8:
w0[2] = 0x01;
break;
case 9:
w0[2] = w0[2] | 0x0100;
break;
case 10:
w0[2] = w0[2] | 0x010000;
break;
case 11:
w0[2] = w0[2] | 0x01000000;
break;
case 12:
w0[3] = 0x01;
break;
case 13:
w0[3] = w0[3] | 0x0100;
break;
case 14:
w0[3] = w0[3] | 0x010000;
break;
case 15:
w0[3] = w0[3] | 0x01000000;
break;
}
}
static void append_0x01_2x4 (u32 w0[4], u32 w1[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x01;
break;
case 1:
w0[0] = w0[0] | 0x0100;
break;
case 2:
w0[0] = w0[0] | 0x010000;
break;
case 3:
w0[0] = w0[0] | 0x01000000;
break;
case 4:
w0[1] = 0x01;
break;
case 5:
w0[1] = w0[1] | 0x0100;
break;
case 6:
w0[1] = w0[1] | 0x010000;
break;
case 7:
w0[1] = w0[1] | 0x01000000;
break;
case 8:
w0[2] = 0x01;
break;
case 9:
w0[2] = w0[2] | 0x0100;
break;
case 10:
w0[2] = w0[2] | 0x010000;
break;
case 11:
w0[2] = w0[2] | 0x01000000;
break;
case 12:
w0[3] = 0x01;
break;
case 13:
w0[3] = w0[3] | 0x0100;
break;
case 14:
w0[3] = w0[3] | 0x010000;
break;
case 15:
w0[3] = w0[3] | 0x01000000;
break;
case 16:
w1[0] = 0x01;
break;
case 17:
w1[0] = w1[0] | 0x0100;
break;
case 18:
w1[0] = w1[0] | 0x010000;
break;
case 19:
w1[0] = w1[0] | 0x01000000;
break;
case 20:
w1[1] = 0x01;
break;
case 21:
w1[1] = w1[1] | 0x0100;
break;
case 22:
w1[1] = w1[1] | 0x010000;
break;
case 23:
w1[1] = w1[1] | 0x01000000;
break;
case 24:
w1[2] = 0x01;
break;
case 25:
w1[2] = w1[2] | 0x0100;
break;
case 26:
w1[2] = w1[2] | 0x010000;
break;
case 27:
w1[2] = w1[2] | 0x01000000;
break;
case 28:
w1[3] = 0x01;
break;
case 29:
w1[3] = w1[3] | 0x0100;
break;
case 30:
w1[3] = w1[3] | 0x010000;
break;
case 31:
w1[3] = w1[3] | 0x01000000;
break;
}
}
static void append_0x01_3x4 (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x01;
break;
case 1:
w0[0] = w0[0] | 0x0100;
break;
case 2:
w0[0] = w0[0] | 0x010000;
break;
case 3:
w0[0] = w0[0] | 0x01000000;
break;
case 4:
w0[1] = 0x01;
break;
case 5:
w0[1] = w0[1] | 0x0100;
break;
case 6:
w0[1] = w0[1] | 0x010000;
break;
case 7:
w0[1] = w0[1] | 0x01000000;
break;
case 8:
w0[2] = 0x01;
break;
case 9:
w0[2] = w0[2] | 0x0100;
break;
case 10:
w0[2] = w0[2] | 0x010000;
break;
case 11:
w0[2] = w0[2] | 0x01000000;
break;
case 12:
w0[3] = 0x01;
break;
case 13:
w0[3] = w0[3] | 0x0100;
break;
case 14:
w0[3] = w0[3] | 0x010000;
break;
case 15:
w0[3] = w0[3] | 0x01000000;
break;
case 16:
w1[0] = 0x01;
break;
case 17:
w1[0] = w1[0] | 0x0100;
break;
case 18:
w1[0] = w1[0] | 0x010000;
break;
case 19:
w1[0] = w1[0] | 0x01000000;
break;
case 20:
w1[1] = 0x01;
break;
case 21:
w1[1] = w1[1] | 0x0100;
break;
case 22:
w1[1] = w1[1] | 0x010000;
break;
case 23:
w1[1] = w1[1] | 0x01000000;
break;
case 24:
w1[2] = 0x01;
break;
case 25:
w1[2] = w1[2] | 0x0100;
break;
case 26:
w1[2] = w1[2] | 0x010000;
break;
case 27:
w1[2] = w1[2] | 0x01000000;
break;
case 28:
w1[3] = 0x01;
break;
case 29:
w1[3] = w1[3] | 0x0100;
break;
case 30:
w1[3] = w1[3] | 0x010000;
break;
case 31:
w1[3] = w1[3] | 0x01000000;
break;
case 32:
w2[0] = 0x01;
break;
case 33:
w2[0] = w2[0] | 0x0100;
break;
case 34:
w2[0] = w2[0] | 0x010000;
break;
case 35:
w2[0] = w2[0] | 0x01000000;
break;
case 36:
w2[1] = 0x01;
break;
case 37:
w2[1] = w2[1] | 0x0100;
break;
case 38:
w2[1] = w2[1] | 0x010000;
break;
case 39:
w2[1] = w2[1] | 0x01000000;
break;
case 40:
w2[2] = 0x01;
break;
case 41:
w2[2] = w2[2] | 0x0100;
break;
case 42:
w2[2] = w2[2] | 0x010000;
break;
case 43:
w2[2] = w2[2] | 0x01000000;
break;
case 44:
w2[3] = 0x01;
break;
case 45:
w2[3] = w2[3] | 0x0100;
break;
case 46:
w2[3] = w2[3] | 0x010000;
break;
case 47:
w2[3] = w2[3] | 0x01000000;
break;
}
}
static void append_0x01_4x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x01;
break;
case 1:
w0[0] = w0[0] | 0x0100;
break;
case 2:
w0[0] = w0[0] | 0x010000;
break;
case 3:
w0[0] = w0[0] | 0x01000000;
break;
case 4:
w0[1] = 0x01;
break;
case 5:
w0[1] = w0[1] | 0x0100;
break;
case 6:
w0[1] = w0[1] | 0x010000;
break;
case 7:
w0[1] = w0[1] | 0x01000000;
break;
case 8:
w0[2] = 0x01;
break;
case 9:
w0[2] = w0[2] | 0x0100;
break;
case 10:
w0[2] = w0[2] | 0x010000;
break;
case 11:
w0[2] = w0[2] | 0x01000000;
break;
case 12:
w0[3] = 0x01;
break;
case 13:
w0[3] = w0[3] | 0x0100;
break;
case 14:
w0[3] = w0[3] | 0x010000;
break;
case 15:
w0[3] = w0[3] | 0x01000000;
break;
case 16:
w1[0] = 0x01;
break;
case 17:
w1[0] = w1[0] | 0x0100;
break;
case 18:
w1[0] = w1[0] | 0x010000;
break;
case 19:
w1[0] = w1[0] | 0x01000000;
break;
case 20:
w1[1] = 0x01;
break;
case 21:
w1[1] = w1[1] | 0x0100;
break;
case 22:
w1[1] = w1[1] | 0x010000;
break;
case 23:
w1[1] = w1[1] | 0x01000000;
break;
case 24:
w1[2] = 0x01;
break;
case 25:
w1[2] = w1[2] | 0x0100;
break;
case 26:
w1[2] = w1[2] | 0x010000;
break;
case 27:
w1[2] = w1[2] | 0x01000000;
break;
case 28:
w1[3] = 0x01;
break;
case 29:
w1[3] = w1[3] | 0x0100;
break;
case 30:
w1[3] = w1[3] | 0x010000;
break;
case 31:
w1[3] = w1[3] | 0x01000000;
break;
case 32:
w2[0] = 0x01;
break;
case 33:
w2[0] = w2[0] | 0x0100;
break;
case 34:
w2[0] = w2[0] | 0x010000;
break;
case 35:
w2[0] = w2[0] | 0x01000000;
break;
case 36:
w2[1] = 0x01;
break;
case 37:
w2[1] = w2[1] | 0x0100;
break;
case 38:
w2[1] = w2[1] | 0x010000;
break;
case 39:
w2[1] = w2[1] | 0x01000000;
break;
case 40:
w2[2] = 0x01;
break;
case 41:
w2[2] = w2[2] | 0x0100;
break;
case 42:
w2[2] = w2[2] | 0x010000;
break;
case 43:
w2[2] = w2[2] | 0x01000000;
break;
case 44:
w2[3] = 0x01;
break;
case 45:
w2[3] = w2[3] | 0x0100;
break;
case 46:
w2[3] = w2[3] | 0x010000;
break;
case 47:
w2[3] = w2[3] | 0x01000000;
break;
case 48:
w3[0] = 0x01;
break;
case 49:
w3[0] = w3[0] | 0x0100;
break;
case 50:
w3[0] = w3[0] | 0x010000;
break;
case 51:
w3[0] = w3[0] | 0x01000000;
break;
case 52:
w3[1] = 0x01;
break;
case 53:
w3[1] = w3[1] | 0x0100;
break;
case 54:
w3[1] = w3[1] | 0x010000;
break;
case 55:
w3[1] = w3[1] | 0x01000000;
break;
case 56:
w3[2] = 0x01;
break;
case 57:
w3[2] = w3[2] | 0x0100;
break;
case 58:
w3[2] = w3[2] | 0x010000;
break;
case 59:
w3[2] = w3[2] | 0x01000000;
break;
case 60:
w3[3] = 0x01;
break;
case 61:
w3[3] = w3[3] | 0x0100;
break;
case 62:
w3[3] = w3[3] | 0x010000;
break;
case 63:
w3[3] = w3[3] | 0x01000000;
break;
}
}
static void append_0x01_8x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x01;
break;
case 1:
w0[0] = w0[0] | 0x0100;
break;
case 2:
w0[0] = w0[0] | 0x010000;
break;
case 3:
w0[0] = w0[0] | 0x01000000;
break;
case 4:
w0[1] = 0x01;
break;
case 5:
w0[1] = w0[1] | 0x0100;
break;
case 6:
w0[1] = w0[1] | 0x010000;
break;
case 7:
w0[1] = w0[1] | 0x01000000;
break;
case 8:
w0[2] = 0x01;
break;
case 9:
w0[2] = w0[2] | 0x0100;
break;
case 10:
w0[2] = w0[2] | 0x010000;
break;
case 11:
w0[2] = w0[2] | 0x01000000;
break;
case 12:
w0[3] = 0x01;
break;
case 13:
w0[3] = w0[3] | 0x0100;
break;
case 14:
w0[3] = w0[3] | 0x010000;
break;
case 15:
w0[3] = w0[3] | 0x01000000;
break;
case 16:
w1[0] = 0x01;
break;
case 17:
w1[0] = w1[0] | 0x0100;
break;
case 18:
w1[0] = w1[0] | 0x010000;
break;
case 19:
w1[0] = w1[0] | 0x01000000;
break;
case 20:
w1[1] = 0x01;
break;
case 21:
w1[1] = w1[1] | 0x0100;
break;
case 22:
w1[1] = w1[1] | 0x010000;
break;
case 23:
w1[1] = w1[1] | 0x01000000;
break;
case 24:
w1[2] = 0x01;
break;
case 25:
w1[2] = w1[2] | 0x0100;
break;
case 26:
w1[2] = w1[2] | 0x010000;
break;
case 27:
w1[2] = w1[2] | 0x01000000;
break;
case 28:
w1[3] = 0x01;
break;
case 29:
w1[3] = w1[3] | 0x0100;
break;
case 30:
w1[3] = w1[3] | 0x010000;
break;
case 31:
w1[3] = w1[3] | 0x01000000;
break;
case 32:
w2[0] = 0x01;
break;
case 33:
w2[0] = w2[0] | 0x0100;
break;
case 34:
w2[0] = w2[0] | 0x010000;
break;
case 35:
w2[0] = w2[0] | 0x01000000;
break;
case 36:
w2[1] = 0x01;
break;
case 37:
w2[1] = w2[1] | 0x0100;
break;
case 38:
w2[1] = w2[1] | 0x010000;
break;
case 39:
w2[1] = w2[1] | 0x01000000;
break;
case 40:
w2[2] = 0x01;
break;
case 41:
w2[2] = w2[2] | 0x0100;
break;
case 42:
w2[2] = w2[2] | 0x010000;
break;
case 43:
w2[2] = w2[2] | 0x01000000;
break;
case 44:
w2[3] = 0x01;
break;
case 45:
w2[3] = w2[3] | 0x0100;
break;
case 46:
w2[3] = w2[3] | 0x010000;
break;
case 47:
w2[3] = w2[3] | 0x01000000;
break;
case 48:
w3[0] = 0x01;
break;
case 49:
w3[0] = w3[0] | 0x0100;
break;
case 50:
w3[0] = w3[0] | 0x010000;
break;
case 51:
w3[0] = w3[0] | 0x01000000;
break;
case 52:
w3[1] = 0x01;
break;
case 53:
w3[1] = w3[1] | 0x0100;
break;
case 54:
w3[1] = w3[1] | 0x010000;
break;
case 55:
w3[1] = w3[1] | 0x01000000;
break;
case 56:
w3[2] = 0x01;
break;
case 57:
w3[2] = w3[2] | 0x0100;
break;
case 58:
w3[2] = w3[2] | 0x010000;
break;
case 59:
w3[2] = w3[2] | 0x01000000;
break;
case 60:
w3[3] = 0x01;
break;
case 61:
w3[3] = w3[3] | 0x0100;
break;
case 62:
w3[3] = w3[3] | 0x010000;
break;
case 63:
w3[3] = w3[3] | 0x01000000;
break;
case 64:
w4[0] = 0x01;
break;
case 65:
w4[0] = w4[0] | 0x0100;
break;
case 66:
w4[0] = w4[0] | 0x010000;
break;
case 67:
w4[0] = w4[0] | 0x01000000;
break;
case 68:
w4[1] = 0x01;
break;
case 69:
w4[1] = w4[1] | 0x0100;
break;
case 70:
w4[1] = w4[1] | 0x010000;
break;
case 71:
w4[1] = w4[1] | 0x01000000;
break;
case 72:
w4[2] = 0x01;
break;
case 73:
w4[2] = w4[2] | 0x0100;
break;
case 74:
w4[2] = w4[2] | 0x010000;
break;
case 75:
w4[2] = w4[2] | 0x01000000;
break;
case 76:
w4[3] = 0x01;
break;
case 77:
w4[3] = w4[3] | 0x0100;
break;
case 78:
w4[3] = w4[3] | 0x010000;
break;
case 79:
w4[3] = w4[3] | 0x01000000;
break;
case 80:
w5[0] = 0x01;
break;
case 81:
w5[0] = w5[0] | 0x0100;
break;
case 82:
w5[0] = w5[0] | 0x010000;
break;
case 83:
w5[0] = w5[0] | 0x01000000;
break;
case 84:
w5[1] = 0x01;
break;
case 85:
w5[1] = w5[1] | 0x0100;
break;
case 86:
w5[1] = w5[1] | 0x010000;
break;
case 87:
w5[1] = w5[1] | 0x01000000;
break;
case 88:
w5[2] = 0x01;
break;
case 89:
w5[2] = w5[2] | 0x0100;
break;
case 90:
w5[2] = w5[2] | 0x010000;
break;
case 91:
w5[2] = w5[2] | 0x01000000;
break;
case 92:
w5[3] = 0x01;
break;
case 93:
w5[3] = w5[3] | 0x0100;
break;
case 94:
w5[3] = w5[3] | 0x010000;
break;
case 95:
w5[3] = w5[3] | 0x01000000;
break;
case 96:
w6[0] = 0x01;
break;
case 97:
w6[0] = w6[0] | 0x0100;
break;
case 98:
w6[0] = w6[0] | 0x010000;
break;
case 99:
w6[0] = w6[0] | 0x01000000;
break;
case 100:
w6[1] = 0x01;
break;
case 101:
w6[1] = w6[1] | 0x0100;
break;
case 102:
w6[1] = w6[1] | 0x010000;
break;
case 103:
w6[1] = w6[1] | 0x01000000;
break;
case 104:
w6[2] = 0x01;
break;
case 105:
w6[2] = w6[2] | 0x0100;
break;
case 106:
w6[2] = w6[2] | 0x010000;
break;
case 107:
w6[2] = w6[2] | 0x01000000;
break;
case 108:
w6[3] = 0x01;
break;
case 109:
w6[3] = w6[3] | 0x0100;
break;
case 110:
w6[3] = w6[3] | 0x010000;
break;
case 111:
w6[3] = w6[3] | 0x01000000;
break;
case 112:
w7[0] = 0x01;
break;
case 113:
w7[0] = w7[0] | 0x0100;
break;
case 114:
w7[0] = w7[0] | 0x010000;
break;
case 115:
w7[0] = w7[0] | 0x01000000;
break;
case 116:
w7[1] = 0x01;
break;
case 117:
w7[1] = w7[1] | 0x0100;
break;
case 118:
w7[1] = w7[1] | 0x010000;
break;
case 119:
w7[1] = w7[1] | 0x01000000;
break;
case 120:
w7[2] = 0x01;
break;
case 121:
w7[2] = w7[2] | 0x0100;
break;
case 122:
w7[2] = w7[2] | 0x010000;
break;
case 123:
w7[2] = w7[2] | 0x01000000;
break;
case 124:
w7[3] = 0x01;
break;
case 125:
w7[3] = w7[3] | 0x0100;
break;
case 126:
w7[3] = w7[3] | 0x010000;
break;
case 127:
w7[3] = w7[3] | 0x01000000;
break;
}
}
static void append_0x02_1x4 (u32 w0[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x02;
break;
case 1:
w0[0] = w0[0] | 0x0200;
break;
case 2:
w0[0] = w0[0] | 0x020000;
break;
case 3:
w0[0] = w0[0] | 0x02000000;
break;
case 4:
w0[1] = 0x02;
break;
case 5:
w0[1] = w0[1] | 0x0200;
break;
case 6:
w0[1] = w0[1] | 0x020000;
break;
case 7:
w0[1] = w0[1] | 0x02000000;
break;
case 8:
w0[2] = 0x02;
break;
case 9:
w0[2] = w0[2] | 0x0200;
break;
case 10:
w0[2] = w0[2] | 0x020000;
break;
case 11:
w0[2] = w0[2] | 0x02000000;
break;
case 12:
w0[3] = 0x02;
break;
case 13:
w0[3] = w0[3] | 0x0200;
break;
case 14:
w0[3] = w0[3] | 0x020000;
break;
case 15:
w0[3] = w0[3] | 0x02000000;
break;
}
}
static void append_0x02_2x4 (u32 w0[4], u32 w1[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x02;
break;
case 1:
w0[0] = w0[0] | 0x0200;
break;
case 2:
w0[0] = w0[0] | 0x020000;
break;
case 3:
w0[0] = w0[0] | 0x02000000;
break;
case 4:
w0[1] = 0x02;
break;
case 5:
w0[1] = w0[1] | 0x0200;
break;
case 6:
w0[1] = w0[1] | 0x020000;
break;
case 7:
w0[1] = w0[1] | 0x02000000;
break;
case 8:
w0[2] = 0x02;
break;
case 9:
w0[2] = w0[2] | 0x0200;
break;
case 10:
w0[2] = w0[2] | 0x020000;
break;
case 11:
w0[2] = w0[2] | 0x02000000;
break;
case 12:
w0[3] = 0x02;
break;
case 13:
w0[3] = w0[3] | 0x0200;
break;
case 14:
w0[3] = w0[3] | 0x020000;
break;
case 15:
w0[3] = w0[3] | 0x02000000;
break;
case 16:
w1[0] = 0x02;
break;
case 17:
w1[0] = w1[0] | 0x0200;
break;
case 18:
w1[0] = w1[0] | 0x020000;
break;
case 19:
w1[0] = w1[0] | 0x02000000;
break;
case 20:
w1[1] = 0x02;
break;
case 21:
w1[1] = w1[1] | 0x0200;
break;
case 22:
w1[1] = w1[1] | 0x020000;
break;
case 23:
w1[1] = w1[1] | 0x02000000;
break;
case 24:
w1[2] = 0x02;
break;
case 25:
w1[2] = w1[2] | 0x0200;
break;
case 26:
w1[2] = w1[2] | 0x020000;
break;
case 27:
w1[2] = w1[2] | 0x02000000;
break;
case 28:
w1[3] = 0x02;
break;
case 29:
w1[3] = w1[3] | 0x0200;
break;
case 30:
w1[3] = w1[3] | 0x020000;
break;
case 31:
w1[3] = w1[3] | 0x02000000;
break;
}
}
static void append_0x02_3x4 (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x02;
break;
case 1:
w0[0] = w0[0] | 0x0200;
break;
case 2:
w0[0] = w0[0] | 0x020000;
break;
case 3:
w0[0] = w0[0] | 0x02000000;
break;
case 4:
w0[1] = 0x02;
break;
case 5:
w0[1] = w0[1] | 0x0200;
break;
case 6:
w0[1] = w0[1] | 0x020000;
break;
case 7:
w0[1] = w0[1] | 0x02000000;
break;
case 8:
w0[2] = 0x02;
break;
case 9:
w0[2] = w0[2] | 0x0200;
break;
case 10:
w0[2] = w0[2] | 0x020000;
break;
case 11:
w0[2] = w0[2] | 0x02000000;
break;
case 12:
w0[3] = 0x02;
break;
case 13:
w0[3] = w0[3] | 0x0200;
break;
case 14:
w0[3] = w0[3] | 0x020000;
break;
case 15:
w0[3] = w0[3] | 0x02000000;
break;
case 16:
w1[0] = 0x02;
break;
case 17:
w1[0] = w1[0] | 0x0200;
break;
case 18:
w1[0] = w1[0] | 0x020000;
break;
case 19:
w1[0] = w1[0] | 0x02000000;
break;
case 20:
w1[1] = 0x02;
break;
case 21:
w1[1] = w1[1] | 0x0200;
break;
case 22:
w1[1] = w1[1] | 0x020000;
break;
case 23:
w1[1] = w1[1] | 0x02000000;
break;
case 24:
w1[2] = 0x02;
break;
case 25:
w1[2] = w1[2] | 0x0200;
break;
case 26:
w1[2] = w1[2] | 0x020000;
break;
case 27:
w1[2] = w1[2] | 0x02000000;
break;
case 28:
w1[3] = 0x02;
break;
case 29:
w1[3] = w1[3] | 0x0200;
break;
case 30:
w1[3] = w1[3] | 0x020000;
break;
case 31:
w1[3] = w1[3] | 0x02000000;
break;
case 32:
w2[0] = 0x02;
break;
case 33:
w2[0] = w2[0] | 0x0200;
break;
case 34:
w2[0] = w2[0] | 0x020000;
break;
case 35:
w2[0] = w2[0] | 0x02000000;
break;
case 36:
w2[1] = 0x02;
break;
case 37:
w2[1] = w2[1] | 0x0200;
break;
case 38:
w2[1] = w2[1] | 0x020000;
break;
case 39:
w2[1] = w2[1] | 0x02000000;
break;
case 40:
w2[2] = 0x02;
break;
case 41:
w2[2] = w2[2] | 0x0200;
break;
case 42:
w2[2] = w2[2] | 0x020000;
break;
case 43:
w2[2] = w2[2] | 0x02000000;
break;
case 44:
w2[3] = 0x02;
break;
case 45:
w2[3] = w2[3] | 0x0200;
break;
case 46:
w2[3] = w2[3] | 0x020000;
break;
case 47:
w2[3] = w2[3] | 0x02000000;
break;
}
}
static void append_0x02_4x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x02;
break;
case 1:
w0[0] = w0[0] | 0x0200;
break;
case 2:
w0[0] = w0[0] | 0x020000;
break;
case 3:
w0[0] = w0[0] | 0x02000000;
break;
case 4:
w0[1] = 0x02;
break;
case 5:
w0[1] = w0[1] | 0x0200;
break;
case 6:
w0[1] = w0[1] | 0x020000;
break;
case 7:
w0[1] = w0[1] | 0x02000000;
break;
case 8:
w0[2] = 0x02;
break;
case 9:
w0[2] = w0[2] | 0x0200;
break;
case 10:
w0[2] = w0[2] | 0x020000;
break;
case 11:
w0[2] = w0[2] | 0x02000000;
break;
case 12:
w0[3] = 0x02;
break;
case 13:
w0[3] = w0[3] | 0x0200;
break;
case 14:
w0[3] = w0[3] | 0x020000;
break;
case 15:
w0[3] = w0[3] | 0x02000000;
break;
case 16:
w1[0] = 0x02;
break;
case 17:
w1[0] = w1[0] | 0x0200;
break;
case 18:
w1[0] = w1[0] | 0x020000;
break;
case 19:
w1[0] = w1[0] | 0x02000000;
break;
case 20:
w1[1] = 0x02;
break;
case 21:
w1[1] = w1[1] | 0x0200;
break;
case 22:
w1[1] = w1[1] | 0x020000;
break;
case 23:
w1[1] = w1[1] | 0x02000000;
break;
case 24:
w1[2] = 0x02;
break;
case 25:
w1[2] = w1[2] | 0x0200;
break;
case 26:
w1[2] = w1[2] | 0x020000;
break;
case 27:
w1[2] = w1[2] | 0x02000000;
break;
case 28:
w1[3] = 0x02;
break;
case 29:
w1[3] = w1[3] | 0x0200;
break;
case 30:
w1[3] = w1[3] | 0x020000;
break;
case 31:
w1[3] = w1[3] | 0x02000000;
break;
case 32:
w2[0] = 0x02;
break;
case 33:
w2[0] = w2[0] | 0x0200;
break;
case 34:
w2[0] = w2[0] | 0x020000;
break;
case 35:
w2[0] = w2[0] | 0x02000000;
break;
case 36:
w2[1] = 0x02;
break;
case 37:
w2[1] = w2[1] | 0x0200;
break;
case 38:
w2[1] = w2[1] | 0x020000;
break;
case 39:
w2[1] = w2[1] | 0x02000000;
break;
case 40:
w2[2] = 0x02;
break;
case 41:
w2[2] = w2[2] | 0x0200;
break;
case 42:
w2[2] = w2[2] | 0x020000;
break;
case 43:
w2[2] = w2[2] | 0x02000000;
break;
case 44:
w2[3] = 0x02;
break;
case 45:
w2[3] = w2[3] | 0x0200;
break;
case 46:
w2[3] = w2[3] | 0x020000;
break;
case 47:
w2[3] = w2[3] | 0x02000000;
break;
case 48:
w3[0] = 0x02;
break;
case 49:
w3[0] = w3[0] | 0x0200;
break;
case 50:
w3[0] = w3[0] | 0x020000;
break;
case 51:
w3[0] = w3[0] | 0x02000000;
break;
case 52:
w3[1] = 0x02;
break;
case 53:
w3[1] = w3[1] | 0x0200;
break;
case 54:
w3[1] = w3[1] | 0x020000;
break;
case 55:
w3[1] = w3[1] | 0x02000000;
break;
case 56:
w3[2] = 0x02;
break;
case 57:
w3[2] = w3[2] | 0x0200;
break;
case 58:
w3[2] = w3[2] | 0x020000;
break;
case 59:
w3[2] = w3[2] | 0x02000000;
break;
case 60:
w3[3] = 0x02;
break;
case 61:
w3[3] = w3[3] | 0x0200;
break;
case 62:
w3[3] = w3[3] | 0x020000;
break;
case 63:
w3[3] = w3[3] | 0x02000000;
break;
}
}
More cleanup
2015-12-15 12:23:32 +00:00
static void append_0x02_8x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
{
switch (offset)
{
case 0:
w0[0] = 0x02;
break;
case 1:
w0[0] = w0[0] | 0x0200;
break;
case 2:
w0[0] = w0[0] | 0x020000;
break;
case 3:
w0[0] = w0[0] | 0x02000000;
break;
case 4:
w0[1] = 0x02;
break;
case 5:
w0[1] = w0[1] | 0x0200;
break;
case 6:
w0[1] = w0[1] | 0x020000;
break;
case 7:
w0[1] = w0[1] | 0x02000000;
break;
case 8:
w0[2] = 0x02;
break;
case 9:
w0[2] = w0[2] | 0x0200;
break;
case 10:
w0[2] = w0[2] | 0x020000;
break;
case 11:
w0[2] = w0[2] | 0x02000000;
break;
case 12:
w0[3] = 0x02;
break;
case 13:
w0[3] = w0[3] | 0x0200;
break;
case 14:
w0[3] = w0[3] | 0x020000;
break;
case 15:
w0[3] = w0[3] | 0x02000000;
break;
case 16:
w1[0] = 0x02;
break;
case 17:
w1[0] = w1[0] | 0x0200;
break;
case 18:
w1[0] = w1[0] | 0x020000;
break;
case 19:
w1[0] = w1[0] | 0x02000000;
break;
case 20:
w1[1] = 0x02;
break;
case 21:
w1[1] = w1[1] | 0x0200;
break;
case 22:
w1[1] = w1[1] | 0x020000;
break;
case 23:
w1[1] = w1[1] | 0x02000000;
break;
case 24:
w1[2] = 0x02;
break;
case 25:
w1[2] = w1[2] | 0x0200;
break;
case 26:
w1[2] = w1[2] | 0x020000;
break;
case 27:
w1[2] = w1[2] | 0x02000000;
break;
case 28:
w1[3] = 0x02;
break;
case 29:
w1[3] = w1[3] | 0x0200;
break;
case 30:
w1[3] = w1[3] | 0x020000;
break;
case 31:
w1[3] = w1[3] | 0x02000000;
break;
case 32:
w2[0] = 0x02;
break;
case 33:
w2[0] = w2[0] | 0x0200;
break;
case 34:
w2[0] = w2[0] | 0x020000;
break;
case 35:
w2[0] = w2[0] | 0x02000000;
break;
case 36:
w2[1] = 0x02;
break;
case 37:
w2[1] = w2[1] | 0x0200;
break;
case 38:
w2[1] = w2[1] | 0x020000;
break;
case 39:
w2[1] = w2[1] | 0x02000000;
break;
case 40:
w2[2] = 0x02;
break;
case 41:
w2[2] = w2[2] | 0x0200;
break;
case 42:
w2[2] = w2[2] | 0x020000;
break;
case 43:
w2[2] = w2[2] | 0x02000000;
break;
case 44:
w2[3] = 0x02;
break;
case 45:
w2[3] = w2[3] | 0x0200;
break;
case 46:
w2[3] = w2[3] | 0x020000;
break;
case 47:
w2[3] = w2[3] | 0x02000000;
break;
case 48:
w3[0] = 0x02;
break;
case 49:
w3[0] = w3[0] | 0x0200;
break;
case 50:
w3[0] = w3[0] | 0x020000;
break;
case 51:
w3[0] = w3[0] | 0x02000000;
break;
case 52:
w3[1] = 0x02;
break;
case 53:
w3[1] = w3[1] | 0x0200;
break;
case 54:
w3[1] = w3[1] | 0x020000;
break;
case 55:
w3[1] = w3[1] | 0x02000000;
break;
case 56:
w3[2] = 0x02;
break;
case 57:
w3[2] = w3[2] | 0x0200;
break;
case 58:
w3[2] = w3[2] | 0x020000;
break;
case 59:
w3[2] = w3[2] | 0x02000000;
break;
case 60:
w3[3] = 0x02;
break;
case 61:
w3[3] = w3[3] | 0x0200;
break;
case 62:
w3[3] = w3[3] | 0x020000;
break;
case 63:
w3[3] = w3[3] | 0x02000000;
break;
case 64:
w4[0] = 0x02;
break;
case 65:
w4[0] = w4[0] | 0x0200;
break;
case 66:
w4[0] = w4[0] | 0x020000;
break;
case 67:
w4[0] = w4[0] | 0x02000000;
break;
case 68:
w4[1] = 0x02;
break;
case 69:
w4[1] = w4[1] | 0x0200;
break;
case 70:
w4[1] = w4[1] | 0x020000;
break;
case 71:
w4[1] = w4[1] | 0x02000000;
break;
case 72:
w4[2] = 0x02;
break;
case 73:
w4[2] = w4[2] | 0x0200;
break;
case 74:
w4[2] = w4[2] | 0x020000;
break;
case 75:
w4[2] = w4[2] | 0x02000000;
break;
case 76:
w4[3] = 0x02;
break;
case 77:
w4[3] = w4[3] | 0x0200;
break;
case 78:
w4[3] = w4[3] | 0x020000;
break;
case 79:
w4[3] = w4[3] | 0x02000000;
break;
case 80:
w5[0] = 0x02;
break;
case 81:
w5[0] = w5[0] | 0x0200;
break;
case 82:
w5[0] = w5[0] | 0x020000;
break;
case 83:
w5[0] = w5[0] | 0x02000000;
break;
case 84:
w5[1] = 0x02;
break;
case 85:
w5[1] = w5[1] | 0x0200;
break;
case 86:
w5[1] = w5[1] | 0x020000;
break;
case 87:
w5[1] = w5[1] | 0x02000000;
break;
case 88:
w5[2] = 0x02;
break;
case 89:
w5[2] = w5[2] | 0x0200;
break;
case 90:
w5[2] = w5[2] | 0x020000;
break;
case 91:
w5[2] = w5[2] | 0x02000000;
break;
case 92:
w5[3] = 0x02;
break;
case 93:
w5[3] = w5[3] | 0x0200;
break;
case 94:
w5[3] = w5[3] | 0x020000;
break;
case 95:
w5[3] = w5[3] | 0x02000000;
break;
case 96:
w6[0] = 0x02;
break;
case 97:
w6[0] = w6[0] | 0x0200;
break;
case 98:
w6[0] = w6[0] | 0x020000;
break;
case 99:
w6[0] = w6[0] | 0x02000000;
break;
case 100:
w6[1] = 0x02;
break;
case 101:
w6[1] = w6[1] | 0x0200;
break;
case 102:
w6[1] = w6[1] | 0x020000;
break;
case 103:
w6[1] = w6[1] | 0x02000000;
break;
case 104:
w6[2] = 0x02;
break;
case 105:
w6[2] = w6[2] | 0x0200;
break;
case 106:
w6[2] = w6[2] | 0x020000;
break;
case 107:
w6[2] = w6[2] | 0x02000000;
break;
case 108:
w6[3] = 0x02;
break;
case 109:
w6[3] = w6[3] | 0x0200;
break;
case 110:
w6[3] = w6[3] | 0x020000;
break;
case 111:
w6[3] = w6[3] | 0x02000000;
break;
case 112:
w7[0] = 0x02;
break;
case 113:
w7[0] = w7[0] | 0x0200;
break;
case 114:
w7[0] = w7[0] | 0x020000;
break;
case 115:
w7[0] = w7[0] | 0x02000000;
break;
case 116:
w7[1] = 0x02;
break;
case 117:
w7[1] = w7[1] | 0x0200;
break;
case 118:
w7[1] = w7[1] | 0x020000;
break;
case 119:
w7[1] = w7[1] | 0x02000000;
break;
case 120:
w7[2] = 0x02;
break;
case 121:
w7[2] = w7[2] | 0x0200;
break;
case 122:
w7[2] = w7[2] | 0x020000;
break;
case 123:
w7[2] = w7[2] | 0x02000000;
break;
case 124:
w7[3] = 0x02;
break;
case 125:
w7[3] = w7[3] | 0x0200;
break;
case 126:
w7[3] = w7[3] | 0x020000;
break;
case 127:
w7[3] = w7[3] | 0x02000000;
break;
}
}
static void append_0x80_1x4 (u32 w0[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x80;
break;
case 1:
w0[0] = w0[0] | 0x8000;
break;
case 2:
w0[0] = w0[0] | 0x800000;
break;
case 3:
w0[0] = w0[0] | 0x80000000;
break;
case 4:
w0[1] = 0x80;
break;
case 5:
w0[1] = w0[1] | 0x8000;
break;
case 6:
w0[1] = w0[1] | 0x800000;
break;
case 7:
w0[1] = w0[1] | 0x80000000;
break;
case 8:
w0[2] = 0x80;
break;
case 9:
w0[2] = w0[2] | 0x8000;
break;
case 10:
w0[2] = w0[2] | 0x800000;
break;
case 11:
w0[2] = w0[2] | 0x80000000;
break;
case 12:
w0[3] = 0x80;
break;
case 13:
w0[3] = w0[3] | 0x8000;
break;
case 14:
w0[3] = w0[3] | 0x800000;
break;
case 15:
w0[3] = w0[3] | 0x80000000;
break;
}
}
static void append_0x80_2x4 (u32 w0[4], u32 w1[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x80;
break;
case 1:
w0[0] = w0[0] | 0x8000;
break;
case 2:
w0[0] = w0[0] | 0x800000;
break;
case 3:
w0[0] = w0[0] | 0x80000000;
break;
case 4:
w0[1] = 0x80;
break;
case 5:
w0[1] = w0[1] | 0x8000;
break;
case 6:
w0[1] = w0[1] | 0x800000;
break;
case 7:
w0[1] = w0[1] | 0x80000000;
break;
case 8:
w0[2] = 0x80;
break;
case 9:
w0[2] = w0[2] | 0x8000;
break;
case 10:
w0[2] = w0[2] | 0x800000;
break;
case 11:
w0[2] = w0[2] | 0x80000000;
break;
case 12:
w0[3] = 0x80;
break;
case 13:
w0[3] = w0[3] | 0x8000;
break;
case 14:
w0[3] = w0[3] | 0x800000;
break;
case 15:
w0[3] = w0[3] | 0x80000000;
break;
case 16:
w1[0] = 0x80;
break;
case 17:
w1[0] = w1[0] | 0x8000;
break;
case 18:
w1[0] = w1[0] | 0x800000;
break;
case 19:
w1[0] = w1[0] | 0x80000000;
break;
case 20:
w1[1] = 0x80;
break;
case 21:
w1[1] = w1[1] | 0x8000;
break;
case 22:
w1[1] = w1[1] | 0x800000;
break;
case 23:
w1[1] = w1[1] | 0x80000000;
break;
case 24:
w1[2] = 0x80;
break;
case 25:
w1[2] = w1[2] | 0x8000;
break;
case 26:
w1[2] = w1[2] | 0x800000;
break;
case 27:
w1[2] = w1[2] | 0x80000000;
break;
case 28:
w1[3] = 0x80;
break;
case 29:
w1[3] = w1[3] | 0x8000;
break;
case 30:
w1[3] = w1[3] | 0x800000;
break;
case 31:
w1[3] = w1[3] | 0x80000000;
break;
}
}
static void append_0x80_3x4 (u32 w0[4], u32 w1[4], u32 w2[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x80;
break;
case 1:
w0[0] = w0[0] | 0x8000;
break;
case 2:
w0[0] = w0[0] | 0x800000;
break;
case 3:
w0[0] = w0[0] | 0x80000000;
break;
case 4:
w0[1] = 0x80;
break;
case 5:
w0[1] = w0[1] | 0x8000;
break;
case 6:
w0[1] = w0[1] | 0x800000;
break;
case 7:
w0[1] = w0[1] | 0x80000000;
break;
case 8:
w0[2] = 0x80;
break;
case 9:
w0[2] = w0[2] | 0x8000;
break;
case 10:
w0[2] = w0[2] | 0x800000;
break;
case 11:
w0[2] = w0[2] | 0x80000000;
break;
case 12:
w0[3] = 0x80;
break;
case 13:
w0[3] = w0[3] | 0x8000;
break;
case 14:
w0[3] = w0[3] | 0x800000;
break;
case 15:
w0[3] = w0[3] | 0x80000000;
break;
case 16:
w1[0] = 0x80;
break;
case 17:
w1[0] = w1[0] | 0x8000;
break;
case 18:
w1[0] = w1[0] | 0x800000;
break;
case 19:
w1[0] = w1[0] | 0x80000000;
break;
case 20:
w1[1] = 0x80;
break;
case 21:
w1[1] = w1[1] | 0x8000;
break;
case 22:
w1[1] = w1[1] | 0x800000;
break;
case 23:
w1[1] = w1[1] | 0x80000000;
break;
case 24:
w1[2] = 0x80;
break;
case 25:
w1[2] = w1[2] | 0x8000;
break;
case 26:
w1[2] = w1[2] | 0x800000;
break;
case 27:
w1[2] = w1[2] | 0x80000000;
break;
case 28:
w1[3] = 0x80;
break;
case 29:
w1[3] = w1[3] | 0x8000;
break;
case 30:
w1[3] = w1[3] | 0x800000;
break;
case 31:
w1[3] = w1[3] | 0x80000000;
break;
case 32:
w2[0] = 0x80;
break;
case 33:
w2[0] = w2[0] | 0x8000;
break;
case 34:
w2[0] = w2[0] | 0x800000;
break;
case 35:
w2[0] = w2[0] | 0x80000000;
break;
case 36:
w2[1] = 0x80;
break;
case 37:
w2[1] = w2[1] | 0x8000;
break;
case 38:
w2[1] = w2[1] | 0x800000;
break;
case 39:
w2[1] = w2[1] | 0x80000000;
break;
case 40:
w2[2] = 0x80;
break;
case 41:
w2[2] = w2[2] | 0x8000;
break;
case 42:
w2[2] = w2[2] | 0x800000;
break;
case 43:
w2[2] = w2[2] | 0x80000000;
break;
case 44:
w2[3] = 0x80;
break;
case 45:
w2[3] = w2[3] | 0x8000;
break;
case 46:
w2[3] = w2[3] | 0x800000;
break;
case 47:
w2[3] = w2[3] | 0x80000000;
break;
}
}
static void append_0x80_4x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x80;
break;
case 1:
w0[0] = w0[0] | 0x8000;
break;
case 2:
w0[0] = w0[0] | 0x800000;
break;
case 3:
w0[0] = w0[0] | 0x80000000;
break;
case 4:
w0[1] = 0x80;
break;
case 5:
w0[1] = w0[1] | 0x8000;
break;
case 6:
w0[1] = w0[1] | 0x800000;
break;
case 7:
w0[1] = w0[1] | 0x80000000;
break;
case 8:
w0[2] = 0x80;
break;
case 9:
w0[2] = w0[2] | 0x8000;
break;
case 10:
w0[2] = w0[2] | 0x800000;
break;
case 11:
w0[2] = w0[2] | 0x80000000;
break;
case 12:
w0[3] = 0x80;
break;
case 13:
w0[3] = w0[3] | 0x8000;
break;
case 14:
w0[3] = w0[3] | 0x800000;
break;
case 15:
w0[3] = w0[3] | 0x80000000;
break;
case 16:
w1[0] = 0x80;
break;
case 17:
w1[0] = w1[0] | 0x8000;
break;
case 18:
w1[0] = w1[0] | 0x800000;
break;
case 19:
w1[0] = w1[0] | 0x80000000;
break;
case 20:
w1[1] = 0x80;
break;
case 21:
w1[1] = w1[1] | 0x8000;
break;
case 22:
w1[1] = w1[1] | 0x800000;
break;
case 23:
w1[1] = w1[1] | 0x80000000;
break;
case 24:
w1[2] = 0x80;
break;
case 25:
w1[2] = w1[2] | 0x8000;
break;
case 26:
w1[2] = w1[2] | 0x800000;
break;
case 27:
w1[2] = w1[2] | 0x80000000;
break;
case 28:
w1[3] = 0x80;
break;
case 29:
w1[3] = w1[3] | 0x8000;
break;
case 30:
w1[3] = w1[3] | 0x800000;
break;
case 31:
w1[3] = w1[3] | 0x80000000;
break;
case 32:
w2[0] = 0x80;
break;
case 33:
w2[0] = w2[0] | 0x8000;
break;
case 34:
w2[0] = w2[0] | 0x800000;
break;
case 35:
w2[0] = w2[0] | 0x80000000;
break;
case 36:
w2[1] = 0x80;
break;
case 37:
w2[1] = w2[1] | 0x8000;
break;
case 38:
w2[1] = w2[1] | 0x800000;
break;
case 39:
w2[1] = w2[1] | 0x80000000;
break;
case 40:
w2[2] = 0x80;
break;
case 41:
w2[2] = w2[2] | 0x8000;
break;
case 42:
w2[2] = w2[2] | 0x800000;
break;
case 43:
w2[2] = w2[2] | 0x80000000;
break;
case 44:
w2[3] = 0x80;
break;
case 45:
w2[3] = w2[3] | 0x8000;
break;
case 46:
w2[3] = w2[3] | 0x800000;
break;
case 47:
w2[3] = w2[3] | 0x80000000;
break;
case 48:
w3[0] = 0x80;
break;
case 49:
w3[0] = w3[0] | 0x8000;
break;
case 50:
w3[0] = w3[0] | 0x800000;
break;
case 51:
w3[0] = w3[0] | 0x80000000;
break;
case 52:
w3[1] = 0x80;
break;
case 53:
w3[1] = w3[1] | 0x8000;
break;
case 54:
w3[1] = w3[1] | 0x800000;
break;
case 55:
w3[1] = w3[1] | 0x80000000;
break;
case 56:
w3[2] = 0x80;
break;
case 57:
w3[2] = w3[2] | 0x8000;
break;
case 58:
w3[2] = w3[2] | 0x800000;
break;
case 59:
w3[2] = w3[2] | 0x80000000;
break;
case 60:
w3[3] = 0x80;
break;
case 61:
w3[3] = w3[3] | 0x8000;
break;
case 62:
w3[3] = w3[3] | 0x800000;
break;
case 63:
w3[3] = w3[3] | 0x80000000;
break;
}
}
static void append_0x80_8x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 w4[4], u32 w5[4], u32 w6[4], u32 w7[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = 0x80;
break;
case 1:
w0[0] = w0[0] | 0x8000;
break;
case 2:
w0[0] = w0[0] | 0x800000;
break;
case 3:
w0[0] = w0[0] | 0x80000000;
break;
case 4:
w0[1] = 0x80;
break;
case 5:
w0[1] = w0[1] | 0x8000;
break;
case 6:
w0[1] = w0[1] | 0x800000;
break;
case 7:
w0[1] = w0[1] | 0x80000000;
break;
case 8:
w0[2] = 0x80;
break;
case 9:
w0[2] = w0[2] | 0x8000;
break;
case 10:
w0[2] = w0[2] | 0x800000;
break;
case 11:
w0[2] = w0[2] | 0x80000000;
break;
case 12:
w0[3] = 0x80;
break;
case 13:
w0[3] = w0[3] | 0x8000;
break;
case 14:
w0[3] = w0[3] | 0x800000;
break;
case 15:
w0[3] = w0[3] | 0x80000000;
break;
case 16:
w1[0] = 0x80;
break;
case 17:
w1[0] = w1[0] | 0x8000;
break;
case 18:
w1[0] = w1[0] | 0x800000;
break;
case 19:
w1[0] = w1[0] | 0x80000000;
break;
case 20:
w1[1] = 0x80;
break;
case 21:
w1[1] = w1[1] | 0x8000;
break;
case 22:
w1[1] = w1[1] | 0x800000;
break;
case 23:
w1[1] = w1[1] | 0x80000000;
break;
case 24:
w1[2] = 0x80;
break;
case 25:
w1[2] = w1[2] | 0x8000;
break;
case 26:
w1[2] = w1[2] | 0x800000;
break;
case 27:
w1[2] = w1[2] | 0x80000000;
break;
case 28:
w1[3] = 0x80;
break;
case 29:
w1[3] = w1[3] | 0x8000;
break;
case 30:
w1[3] = w1[3] | 0x800000;
break;
case 31:
w1[3] = w1[3] | 0x80000000;
break;
case 32:
w2[0] = 0x80;
break;
case 33:
w2[0] = w2[0] | 0x8000;
break;
case 34:
w2[0] = w2[0] | 0x800000;
break;
case 35:
w2[0] = w2[0] | 0x80000000;
break;
case 36:
w2[1] = 0x80;
break;
case 37:
w2[1] = w2[1] | 0x8000;
break;
case 38:
w2[1] = w2[1] | 0x800000;
break;
case 39:
w2[1] = w2[1] | 0x80000000;
break;
case 40:
w2[2] = 0x80;
break;
case 41:
w2[2] = w2[2] | 0x8000;
break;
case 42:
w2[2] = w2[2] | 0x800000;
break;
case 43:
w2[2] = w2[2] | 0x80000000;
break;
case 44:
w2[3] = 0x80;
break;
case 45:
w2[3] = w2[3] | 0x8000;
break;
case 46:
w2[3] = w2[3] | 0x800000;
break;
case 47:
w2[3] = w2[3] | 0x80000000;
break;
case 48:
w3[0] = 0x80;
break;
case 49:
w3[0] = w3[0] | 0x8000;
break;
case 50:
w3[0] = w3[0] | 0x800000;
break;
case 51:
w3[0] = w3[0] | 0x80000000;
break;
case 52:
w3[1] = 0x80;
break;
case 53:
w3[1] = w3[1] | 0x8000;
break;
case 54:
w3[1] = w3[1] | 0x800000;
break;
case 55:
w3[1] = w3[1] | 0x80000000;
break;
case 56:
w3[2] = 0x80;
break;
case 57:
w3[2] = w3[2] | 0x8000;
break;
case 58:
w3[2] = w3[2] | 0x800000;
break;
case 59:
w3[2] = w3[2] | 0x80000000;
break;
case 60:
w3[3] = 0x80;
break;
case 61:
w3[3] = w3[3] | 0x8000;
break;
case 62:
w3[3] = w3[3] | 0x800000;
break;
case 63:
w3[3] = w3[3] | 0x80000000;
break;
case 64:
w4[0] = 0x80;
break;
case 65:
w4[0] = w4[0] | 0x8000;
break;
case 66:
w4[0] = w4[0] | 0x800000;
break;
case 67:
w4[0] = w4[0] | 0x80000000;
break;
case 68:
w4[1] = 0x80;
break;
case 69:
w4[1] = w4[1] | 0x8000;
break;
case 70:
w4[1] = w4[1] | 0x800000;
break;
case 71:
w4[1] = w4[1] | 0x80000000;
break;
case 72:
w4[2] = 0x80;
break;
case 73:
w4[2] = w4[2] | 0x8000;
break;
case 74:
w4[2] = w4[2] | 0x800000;
break;
case 75:
w4[2] = w4[2] | 0x80000000;
break;
case 76:
w4[3] = 0x80;
break;
case 77:
w4[3] = w4[3] | 0x8000;
break;
case 78:
w4[3] = w4[3] | 0x800000;
break;
case 79:
w4[3] = w4[3] | 0x80000000;
break;
case 80:
w5[0] = 0x80;
break;
case 81:
w5[0] = w5[0] | 0x8000;
break;
case 82:
w5[0] = w5[0] | 0x800000;
break;
case 83:
w5[0] = w5[0] | 0x80000000;
break;
case 84:
w5[1] = 0x80;
break;
case 85:
w5[1] = w5[1] | 0x8000;
break;
case 86:
w5[1] = w5[1] | 0x800000;
break;
case 87:
w5[1] = w5[1] | 0x80000000;
break;
case 88:
w5[2] = 0x80;
break;
case 89:
w5[2] = w5[2] | 0x8000;
break;
case 90:
w5[2] = w5[2] | 0x800000;
break;
case 91:
w5[2] = w5[2] | 0x80000000;
break;
case 92:
w5[3] = 0x80;
break;
case 93:
w5[3] = w5[3] | 0x8000;
break;
case 94:
w5[3] = w5[3] | 0x800000;
break;
case 95:
w5[3] = w5[3] | 0x80000000;
break;
case 96:
w6[0] = 0x80;
break;
case 97:
w6[0] = w6[0] | 0x8000;
break;
case 98:
w6[0] = w6[0] | 0x800000;
break;
case 99:
w6[0] = w6[0] | 0x80000000;
break;
case 100:
w6[1] = 0x80;
break;
case 101:
w6[1] = w6[1] | 0x8000;
break;
case 102:
w6[1] = w6[1] | 0x800000;
break;
case 103:
w6[1] = w6[1] | 0x80000000;
break;
case 104:
w6[2] = 0x80;
break;
case 105:
w6[2] = w6[2] | 0x8000;
break;
case 106:
w6[2] = w6[2] | 0x800000;
break;
case 107:
w6[2] = w6[2] | 0x80000000;
break;
case 108:
w6[3] = 0x80;
break;
case 109:
w6[3] = w6[3] | 0x8000;
break;
case 110:
w6[3] = w6[3] | 0x800000;
break;
case 111:
w6[3] = w6[3] | 0x80000000;
break;
case 112:
w7[0] = 0x80;
break;
case 113:
w7[0] = w7[0] | 0x8000;
break;
case 114:
w7[0] = w7[0] | 0x800000;
break;
case 115:
w7[0] = w7[0] | 0x80000000;
break;
case 116:
w7[1] = 0x80;
break;
case 117:
w7[1] = w7[1] | 0x8000;
break;
case 118:
w7[1] = w7[1] | 0x800000;
break;
case 119:
w7[1] = w7[1] | 0x80000000;
break;
case 120:
w7[2] = 0x80;
break;
case 121:
w7[2] = w7[2] | 0x8000;
break;
case 122:
w7[2] = w7[2] | 0x800000;
break;
case 123:
w7[2] = w7[2] | 0x80000000;
break;
case 124:
w7[3] = 0x80;
break;
case 125:
w7[3] = w7[3] | 0x8000;
break;
case 126:
w7[3] = w7[3] | 0x800000;
break;
case 127:
w7[3] = w7[3] | 0x80000000;
break;
}
}
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
static void append_0x80_1x16 (u32 w[16], const u32 offset)
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
{
switch (offset)
{
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 0:
w[ 0] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 1:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 0] = w[ 0] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 2:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 0] = w[ 0] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 3:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 0] = w[ 0] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 4:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 1] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 5:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 1] = w[ 1] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 6:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 1] = w[ 1] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 7:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 1] = w[ 1] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 8:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 2] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 9:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 2] = w[ 2] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 10:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 2] = w[ 2] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 11:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 2] = w[ 2] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 12:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 3] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 13:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 3] = w[ 3] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 14:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 3] = w[ 3] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 15:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 3] = w[ 3] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 16:
w[ 4] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 17:
w[ 4] = w[ 4] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 18:
w[ 4] = w[ 4] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 19:
w[ 4] = w[ 4] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 20:
w[ 5] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 21:
w[ 5] = w[ 5] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 22:
w[ 5] = w[ 5] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 23:
w[ 5] = w[ 5] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 24:
w[ 6] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 25:
w[ 6] = w[ 6] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 26:
w[ 6] = w[ 6] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 27:
w[ 6] = w[ 6] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 28:
w[ 7] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 29:
w[ 7] = w[ 7] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 30:
w[ 7] = w[ 7] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 31:
w[ 7] = w[ 7] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 32:
w[ 8] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 33:
w[ 8] = w[ 8] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 34:
w[ 8] = w[ 8] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 35:
w[ 8] = w[ 8] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 36:
w[ 9] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 37:
w[ 9] = w[ 9] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 38:
w[ 9] = w[ 9] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 39:
w[ 9] = w[ 9] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 40:
w[10] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 41:
w[10] = w[10] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 42:
w[10] = w[10] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 43:
w[10] = w[10] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 44:
w[11] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 45:
w[11] = w[11] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 46:
w[11] = w[11] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 47:
w[11] = w[11] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 48:
w[12] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 49:
w[12] = w[12] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 50:
w[12] = w[12] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 51:
w[12] = w[12] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 52:
w[13] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 53:
w[13] = w[13] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 54:
w[13] = w[13] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 55:
w[13] = w[13] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 56:
w[14] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 57:
w[14] = w[14] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 58:
w[14] = w[14] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 59:
w[14] = w[14] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 60:
w[15] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 61:
w[15] = w[15] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 62:
w[15] = w[15] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 63:
w[15] = w[15] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
}
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Prepare new SIMD code for kernel, -m 0, 10, 20, 1000 should work in -a 3 mode and other hopefully stay unaffected
2016-01-17 21:17:50 +00:00
static void switch_buffer_by_offset_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
{
#if defined IS_AMD || defined IS_GENERIC
const int offset_mod_4 = offset & 3;
const int offset_minus_4 = 4 - offset;
switch (offset / 4)
{
case 0:
w3[2] = amd_bytealign_S ( 0, w3[1], offset_minus_4);
w3[1] = amd_bytealign_S (w3[1], w3[0], offset_minus_4);
w3[0] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
w2[3] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
w2[2] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
w2[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
w2[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
w1[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
w1[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
w1[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w1[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w0[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w0[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w0[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w0[0] = amd_bytealign_S (w0[0], 0, offset_minus_4);
if (offset_mod_4 == 0)
{
w0[0] = w0[1];
w0[1] = w0[2];
w0[2] = w0[3];
w0[3] = w1[0];
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 1:
w3[2] = amd_bytealign_S ( 0, w3[0], offset_minus_4);
w3[1] = amd_bytealign_S (w3[0], w2[3], offset_minus_4);
w3[0] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
w2[3] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
w2[2] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
w2[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
w2[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
w1[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
w1[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w1[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w1[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w0[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w0[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w0[1] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w0[0] = 0;
if (offset_mod_4 == 0)
{
w0[1] = w0[2];
w0[2] = w0[3];
w0[3] = w1[0];
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 2:
w3[2] = amd_bytealign_S ( 0, w2[3], offset_minus_4);
w3[1] = amd_bytealign_S (w2[3], w2[2], offset_minus_4);
w3[0] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
w2[3] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
w2[2] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
w2[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
w2[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
w1[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w1[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w1[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w1[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w0[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w0[2] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w0[2] = w0[3];
w0[3] = w1[0];
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 3:
w3[2] = amd_bytealign_S ( 0, w2[2], offset_minus_4);
w3[1] = amd_bytealign_S (w2[2], w2[1], offset_minus_4);
w3[0] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
w2[3] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
w2[2] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
w2[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
w2[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w1[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w1[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w1[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w1[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w0[3] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w0[3] = w1[0];
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 4:
w3[2] = amd_bytealign_S ( 0, w2[1], offset_minus_4);
w3[1] = amd_bytealign_S (w2[1], w2[0], offset_minus_4);
w3[0] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
w2[3] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
w2[2] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
w2[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w2[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w1[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w1[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w1[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w1[0] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 5:
w3[2] = amd_bytealign_S ( 0, w2[0], offset_minus_4);
w3[1] = amd_bytealign_S (w2[0], w1[3], offset_minus_4);
w3[0] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
w2[3] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
w2[2] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w2[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w2[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w1[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w1[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w1[1] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 6:
w3[2] = amd_bytealign_S ( 0, w1[3], offset_minus_4);
w3[1] = amd_bytealign_S (w1[3], w1[2], offset_minus_4);
w3[0] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
w2[3] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w2[2] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w2[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w2[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w1[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w1[2] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 7:
w3[2] = amd_bytealign_S ( 0, w1[2], offset_minus_4);
w3[1] = amd_bytealign_S (w1[2], w1[1], offset_minus_4);
w3[0] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w2[3] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w2[2] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w2[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w2[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w1[3] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 8:
w3[2] = amd_bytealign_S ( 0, w1[1], offset_minus_4);
w3[1] = amd_bytealign_S (w1[1], w1[0], offset_minus_4);
w3[0] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w2[3] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w2[2] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w2[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w2[0] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 9:
w3[2] = amd_bytealign_S ( 0, w1[0], offset_minus_4);
w3[1] = amd_bytealign_S (w1[0], w0[3], offset_minus_4);
w3[0] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w2[3] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w2[2] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w2[1] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 10:
w3[2] = amd_bytealign_S ( 0, w0[3], offset_minus_4);
w3[1] = amd_bytealign_S (w0[3], w0[2], offset_minus_4);
w3[0] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w2[3] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w2[2] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 11:
w3[2] = amd_bytealign_S ( 0, w0[2], offset_minus_4);
w3[1] = amd_bytealign_S (w0[2], w0[1], offset_minus_4);
w3[0] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w2[3] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 12:
w3[2] = amd_bytealign_S ( 0, w0[1], offset_minus_4);
w3[1] = amd_bytealign_S (w0[1], w0[0], offset_minus_4);
w3[0] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
break;
case 13:
w3[2] = amd_bytealign_S ( 0, w0[0], offset_minus_4);
w3[1] = amd_bytealign_S (w0[0], 0, offset_minus_4);
w3[0] = 0;
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w3[1] = w3[2];
w3[2] = 0;
}
break;
}
#endif
#ifdef IS_NV
const int offset_minus_4 = 4 - (offset % 4);
const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
switch (offset / 4)
{
case 0:
w3[1] = __byte_perm_S (w3[0], w3[1], selector);
w3[0] = __byte_perm_S (w2[3], w3[0], selector);
w2[3] = __byte_perm_S (w2[2], w2[3], selector);
w2[2] = __byte_perm_S (w2[1], w2[2], selector);
w2[1] = __byte_perm_S (w2[0], w2[1], selector);
w2[0] = __byte_perm_S (w1[3], w2[0], selector);
w1[3] = __byte_perm_S (w1[2], w1[3], selector);
w1[2] = __byte_perm_S (w1[1], w1[2], selector);
w1[1] = __byte_perm_S (w1[0], w1[1], selector);
w1[0] = __byte_perm_S (w0[3], w1[0], selector);
w0[3] = __byte_perm_S (w0[2], w0[3], selector);
w0[2] = __byte_perm_S (w0[1], w0[2], selector);
w0[1] = __byte_perm_S (w0[0], w0[1], selector);
w0[0] = __byte_perm_S ( 0, w0[0], selector);
break;
case 1:
w3[1] = __byte_perm_S (w2[3], w3[0], selector);
w3[0] = __byte_perm_S (w2[2], w2[3], selector);
w2[3] = __byte_perm_S (w2[1], w2[2], selector);
w2[2] = __byte_perm_S (w2[0], w2[1], selector);
w2[1] = __byte_perm_S (w1[3], w2[0], selector);
w2[0] = __byte_perm_S (w1[2], w1[3], selector);
w1[3] = __byte_perm_S (w1[1], w1[2], selector);
w1[2] = __byte_perm_S (w1[0], w1[1], selector);
w1[1] = __byte_perm_S (w0[3], w1[0], selector);
w1[0] = __byte_perm_S (w0[2], w0[3], selector);
w0[3] = __byte_perm_S (w0[1], w0[2], selector);
w0[2] = __byte_perm_S (w0[0], w0[1], selector);
w0[1] = __byte_perm_S ( 0, w0[0], selector);
w0[0] = 0;
break;
case 2:
w3[1] = __byte_perm_S (w2[2], w2[3], selector);
w3[0] = __byte_perm_S (w2[1], w2[2], selector);
w2[3] = __byte_perm_S (w2[0], w2[1], selector);
w2[2] = __byte_perm_S (w1[3], w2[0], selector);
w2[1] = __byte_perm_S (w1[2], w1[3], selector);
w2[0] = __byte_perm_S (w1[1], w1[2], selector);
w1[3] = __byte_perm_S (w1[0], w1[1], selector);
w1[2] = __byte_perm_S (w0[3], w1[0], selector);
w1[1] = __byte_perm_S (w0[2], w0[3], selector);
w1[0] = __byte_perm_S (w0[1], w0[2], selector);
w0[3] = __byte_perm_S (w0[0], w0[1], selector);
w0[2] = __byte_perm_S ( 0, w0[0], selector);
w0[1] = 0;
w0[0] = 0;
break;
case 3:
w3[1] = __byte_perm_S (w2[1], w2[2], selector);
w3[0] = __byte_perm_S (w2[0], w2[1], selector);
w2[3] = __byte_perm_S (w1[3], w2[0], selector);
w2[2] = __byte_perm_S (w1[2], w1[3], selector);
w2[1] = __byte_perm_S (w1[1], w1[2], selector);
w2[0] = __byte_perm_S (w1[0], w1[1], selector);
w1[3] = __byte_perm_S (w0[3], w1[0], selector);
w1[2] = __byte_perm_S (w0[2], w0[3], selector);
w1[1] = __byte_perm_S (w0[1], w0[2], selector);
w1[0] = __byte_perm_S (w0[0], w0[1], selector);
w0[3] = __byte_perm_S ( 0, w0[0], selector);
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 4:
w3[1] = __byte_perm_S (w2[0], w2[1], selector);
w3[0] = __byte_perm_S (w1[3], w2[0], selector);
w2[3] = __byte_perm_S (w1[2], w1[3], selector);
w2[2] = __byte_perm_S (w1[1], w1[2], selector);
w2[1] = __byte_perm_S (w1[0], w1[1], selector);
w2[0] = __byte_perm_S (w0[3], w1[0], selector);
w1[3] = __byte_perm_S (w0[2], w0[3], selector);
w1[2] = __byte_perm_S (w0[1], w0[2], selector);
w1[1] = __byte_perm_S (w0[0], w0[1], selector);
w1[0] = __byte_perm_S ( 0, w0[0], selector);
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 5:
w3[1] = __byte_perm_S (w1[3], w2[0], selector);
w3[0] = __byte_perm_S (w1[2], w1[3], selector);
w2[3] = __byte_perm_S (w1[1], w1[2], selector);
w2[2] = __byte_perm_S (w1[0], w1[1], selector);
w2[1] = __byte_perm_S (w0[3], w1[0], selector);
w2[0] = __byte_perm_S (w0[2], w0[3], selector);
w1[3] = __byte_perm_S (w0[1], w0[2], selector);
w1[2] = __byte_perm_S (w0[0], w0[1], selector);
w1[1] = __byte_perm_S ( 0, w0[0], selector);
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 6:
w3[1] = __byte_perm_S (w1[2], w1[3], selector);
w3[0] = __byte_perm_S (w1[1], w1[2], selector);
w2[3] = __byte_perm_S (w1[0], w1[1], selector);
w2[2] = __byte_perm_S (w0[3], w1[0], selector);
w2[1] = __byte_perm_S (w0[2], w0[3], selector);
w2[0] = __byte_perm_S (w0[1], w0[2], selector);
w1[3] = __byte_perm_S (w0[0], w0[1], selector);
w1[2] = __byte_perm_S ( 0, w0[0], selector);
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 7:
w3[1] = __byte_perm_S (w1[1], w1[2], selector);
w3[0] = __byte_perm_S (w1[0], w1[1], selector);
w2[3] = __byte_perm_S (w0[3], w1[0], selector);
w2[2] = __byte_perm_S (w0[2], w0[3], selector);
w2[1] = __byte_perm_S (w0[1], w0[2], selector);
w2[0] = __byte_perm_S (w0[0], w0[1], selector);
w1[3] = __byte_perm_S ( 0, w0[0], selector);
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 8:
w3[1] = __byte_perm_S (w1[0], w1[1], selector);
w3[0] = __byte_perm_S (w0[3], w1[0], selector);
w2[3] = __byte_perm_S (w0[2], w0[3], selector);
w2[2] = __byte_perm_S (w0[1], w0[2], selector);
w2[1] = __byte_perm_S (w0[0], w0[1], selector);
w2[0] = __byte_perm_S ( 0, w0[0], selector);
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 9:
w3[1] = __byte_perm_S (w0[3], w1[0], selector);
w3[0] = __byte_perm_S (w0[2], w0[3], selector);
w2[3] = __byte_perm_S (w0[1], w0[2], selector);
w2[2] = __byte_perm_S (w0[0], w0[1], selector);
w2[1] = __byte_perm_S ( 0, w0[0], selector);
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 10:
w3[1] = __byte_perm_S (w0[2], w0[3], selector);
w3[0] = __byte_perm_S (w0[1], w0[2], selector);
w2[3] = __byte_perm_S (w0[0], w0[1], selector);
w2[2] = __byte_perm_S ( 0, w0[0], selector);
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 11:
w3[1] = __byte_perm_S (w0[1], w0[2], selector);
w3[0] = __byte_perm_S (w0[0], w0[1], selector);
w2[3] = __byte_perm_S ( 0, w0[0], selector);
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 12:
w3[1] = __byte_perm_S (w0[0], w0[1], selector);
w3[0] = __byte_perm_S ( 0, w0[0], selector);
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 13:
w3[1] = __byte_perm_S ( 0, w0[0], selector);
w3[0] = 0;
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
}
#endif
}
static void switch_buffer_by_offset_be_S (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 offset)
{
#if defined IS_AMD || defined IS_GENERIC
switch (offset / 4)
{
case 0:
w3[2] = amd_bytealign_S (w3[1], 0, offset);
w3[1] = amd_bytealign_S (w3[0], w3[1], offset);
w3[0] = amd_bytealign_S (w2[3], w3[0], offset);
w2[3] = amd_bytealign_S (w2[2], w2[3], offset);
w2[2] = amd_bytealign_S (w2[1], w2[2], offset);
w2[1] = amd_bytealign_S (w2[0], w2[1], offset);
w2[0] = amd_bytealign_S (w1[3], w2[0], offset);
w1[3] = amd_bytealign_S (w1[2], w1[3], offset);
w1[2] = amd_bytealign_S (w1[1], w1[2], offset);
w1[1] = amd_bytealign_S (w1[0], w1[1], offset);
w1[0] = amd_bytealign_S (w0[3], w1[0], offset);
w0[3] = amd_bytealign_S (w0[2], w0[3], offset);
w0[2] = amd_bytealign_S (w0[1], w0[2], offset);
w0[1] = amd_bytealign_S (w0[0], w0[1], offset);
w0[0] = amd_bytealign_S ( 0, w0[0], offset);
break;
case 1:
w3[2] = amd_bytealign_S (w3[0], 0, offset);
w3[1] = amd_bytealign_S (w2[3], w3[0], offset);
w3[0] = amd_bytealign_S (w2[2], w2[3], offset);
w2[3] = amd_bytealign_S (w2[1], w2[2], offset);
w2[2] = amd_bytealign_S (w2[0], w2[1], offset);
w2[1] = amd_bytealign_S (w1[3], w2[0], offset);
w2[0] = amd_bytealign_S (w1[2], w1[3], offset);
w1[3] = amd_bytealign_S (w1[1], w1[2], offset);
w1[2] = amd_bytealign_S (w1[0], w1[1], offset);
w1[1] = amd_bytealign_S (w0[3], w1[0], offset);
w1[0] = amd_bytealign_S (w0[2], w0[3], offset);
w0[3] = amd_bytealign_S (w0[1], w0[2], offset);
w0[2] = amd_bytealign_S (w0[0], w0[1], offset);
w0[1] = amd_bytealign_S ( 0, w0[0], offset);
w0[0] = 0;
break;
case 2:
w3[2] = amd_bytealign_S (w2[3], 0, offset);
w3[1] = amd_bytealign_S (w2[2], w2[3], offset);
w3[0] = amd_bytealign_S (w2[1], w2[2], offset);
w2[3] = amd_bytealign_S (w2[0], w2[1], offset);
w2[2] = amd_bytealign_S (w1[3], w2[0], offset);
w2[1] = amd_bytealign_S (w1[2], w1[3], offset);
w2[0] = amd_bytealign_S (w1[1], w1[2], offset);
w1[3] = amd_bytealign_S (w1[0], w1[1], offset);
w1[2] = amd_bytealign_S (w0[3], w1[0], offset);
w1[1] = amd_bytealign_S (w0[2], w0[3], offset);
w1[0] = amd_bytealign_S (w0[1], w0[2], offset);
w0[3] = amd_bytealign_S (w0[0], w0[1], offset);
w0[2] = amd_bytealign_S ( 0, w0[0], offset);
w0[1] = 0;
w0[0] = 0;
break;
case 3:
w3[2] = amd_bytealign_S (w2[2], 0, offset);
w3[1] = amd_bytealign_S (w2[1], w2[2], offset);
w3[0] = amd_bytealign_S (w2[0], w2[1], offset);
w2[3] = amd_bytealign_S (w1[3], w2[0], offset);
w2[2] = amd_bytealign_S (w1[2], w1[3], offset);
w2[1] = amd_bytealign_S (w1[1], w1[2], offset);
w2[0] = amd_bytealign_S (w1[0], w1[1], offset);
w1[3] = amd_bytealign_S (w0[3], w1[0], offset);
w1[2] = amd_bytealign_S (w0[2], w0[3], offset);
w1[1] = amd_bytealign_S (w0[1], w0[2], offset);
w1[0] = amd_bytealign_S (w0[0], w0[1], offset);
w0[3] = amd_bytealign_S ( 0, w0[0], offset);
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 4:
w3[2] = amd_bytealign_S (w2[1], 0, offset);
w3[1] = amd_bytealign_S (w2[0], w2[1], offset);
w3[0] = amd_bytealign_S (w1[3], w2[0], offset);
w2[3] = amd_bytealign_S (w1[2], w1[3], offset);
w2[2] = amd_bytealign_S (w1[1], w1[2], offset);
w2[1] = amd_bytealign_S (w1[0], w1[1], offset);
w2[0] = amd_bytealign_S (w0[3], w1[0], offset);
w1[3] = amd_bytealign_S (w0[2], w0[3], offset);
w1[2] = amd_bytealign_S (w0[1], w0[2], offset);
w1[1] = amd_bytealign_S (w0[0], w0[1], offset);
w1[0] = amd_bytealign_S ( 0, w0[0], offset);
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 5:
w3[2] = amd_bytealign_S (w2[0], 0, offset);
w3[1] = amd_bytealign_S (w1[3], w2[0], offset);
w3[0] = amd_bytealign_S (w1[2], w1[3], offset);
w2[3] = amd_bytealign_S (w1[1], w1[2], offset);
w2[2] = amd_bytealign_S (w1[0], w1[1], offset);
w2[1] = amd_bytealign_S (w0[3], w1[0], offset);
w2[0] = amd_bytealign_S (w0[2], w0[3], offset);
w1[3] = amd_bytealign_S (w0[1], w0[2], offset);
w1[2] = amd_bytealign_S (w0[0], w0[1], offset);
w1[1] = amd_bytealign_S ( 0, w0[0], offset);
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 6:
w3[2] = amd_bytealign_S (w1[3], 0, offset);
w3[1] = amd_bytealign_S (w1[2], w1[3], offset);
w3[0] = amd_bytealign_S (w1[1], w1[2], offset);
w2[3] = amd_bytealign_S (w1[0], w1[1], offset);
w2[2] = amd_bytealign_S (w0[3], w1[0], offset);
w2[1] = amd_bytealign_S (w0[2], w0[3], offset);
w2[0] = amd_bytealign_S (w0[1], w0[2], offset);
w1[3] = amd_bytealign_S (w0[0], w0[1], offset);
w1[2] = amd_bytealign_S ( 0, w0[0], offset);
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 7:
w3[2] = amd_bytealign_S (w1[2], 0, offset);
w3[1] = amd_bytealign_S (w1[1], w1[2], offset);
w3[0] = amd_bytealign_S (w1[0], w1[1], offset);
w2[3] = amd_bytealign_S (w0[3], w1[0], offset);
w2[2] = amd_bytealign_S (w0[2], w0[3], offset);
w2[1] = amd_bytealign_S (w0[1], w0[2], offset);
w2[0] = amd_bytealign_S (w0[0], w0[1], offset);
w1[3] = amd_bytealign_S ( 0, w0[0], offset);
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 8:
w3[2] = amd_bytealign_S (w1[1], 0, offset);
w3[1] = amd_bytealign_S (w1[0], w1[1], offset);
w3[0] = amd_bytealign_S (w0[3], w1[0], offset);
w2[3] = amd_bytealign_S (w0[2], w0[3], offset);
w2[2] = amd_bytealign_S (w0[1], w0[2], offset);
w2[1] = amd_bytealign_S (w0[0], w0[1], offset);
w2[0] = amd_bytealign_S ( 0, w0[0], offset);
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 9:
w3[2] = amd_bytealign_S (w1[0], 0, offset);
w3[1] = amd_bytealign_S (w0[3], w1[0], offset);
w3[0] = amd_bytealign_S (w0[2], w0[3], offset);
w2[3] = amd_bytealign_S (w0[1], w0[2], offset);
w2[2] = amd_bytealign_S (w0[0], w0[1], offset);
w2[1] = amd_bytealign_S ( 0, w0[0], offset);
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 10:
w3[2] = amd_bytealign_S (w0[3], 0, offset);
w3[1] = amd_bytealign_S (w0[2], w0[3], offset);
w3[0] = amd_bytealign_S (w0[1], w0[2], offset);
w2[3] = amd_bytealign_S (w0[0], w0[1], offset);
w2[2] = amd_bytealign_S ( 0, w0[0], offset);
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 11:
w3[2] = amd_bytealign_S (w0[2], 0, offset);
w3[1] = amd_bytealign_S (w0[1], w0[2], offset);
w3[0] = amd_bytealign_S (w0[0], w0[1], offset);
w2[3] = amd_bytealign_S ( 0, w0[0], offset);
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 12:
w3[2] = amd_bytealign_S (w0[1], 0, offset);
w3[1] = amd_bytealign_S (w0[0], w0[1], offset);
w3[0] = amd_bytealign_S ( 0, w0[0], offset);
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 13:
w3[2] = amd_bytealign_S (w0[0], 0, offset);
w3[1] = amd_bytealign_S ( 0, w0[0], offset);
w3[0] = 0;
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
}
#endif
#ifdef IS_NV
const int selector = (0x76543210 >> ((offset & 3) * 4)) & 0xffff;
switch (offset / 4)
{
case 0:
w3[1] = __byte_perm_S (w3[1], w3[0], selector);
w3[0] = __byte_perm_S (w3[0], w2[3], selector);
w2[3] = __byte_perm_S (w2[3], w2[2], selector);
w2[2] = __byte_perm_S (w2[2], w2[1], selector);
w2[1] = __byte_perm_S (w2[1], w2[0], selector);
w2[0] = __byte_perm_S (w2[0], w1[3], selector);
w1[3] = __byte_perm_S (w1[3], w1[2], selector);
w1[2] = __byte_perm_S (w1[2], w1[1], selector);
w1[1] = __byte_perm_S (w1[1], w1[0], selector);
w1[0] = __byte_perm_S (w1[0], w0[3], selector);
w0[3] = __byte_perm_S (w0[3], w0[2], selector);
w0[2] = __byte_perm_S (w0[2], w0[1], selector);
w0[1] = __byte_perm_S (w0[1], w0[0], selector);
w0[0] = __byte_perm_S (w0[0], 0, selector);
break;
case 1:
w3[1] = __byte_perm_S (w3[0], w2[3], selector);
w3[0] = __byte_perm_S (w2[3], w2[2], selector);
w2[3] = __byte_perm_S (w2[2], w2[1], selector);
w2[2] = __byte_perm_S (w2[1], w2[0], selector);
w2[1] = __byte_perm_S (w2[0], w1[3], selector);
w2[0] = __byte_perm_S (w1[3], w1[2], selector);
w1[3] = __byte_perm_S (w1[2], w1[1], selector);
w1[2] = __byte_perm_S (w1[1], w1[0], selector);
w1[1] = __byte_perm_S (w1[0], w0[3], selector);
w1[0] = __byte_perm_S (w0[3], w0[2], selector);
w0[3] = __byte_perm_S (w0[2], w0[1], selector);
w0[2] = __byte_perm_S (w0[1], w0[0], selector);
w0[1] = __byte_perm_S (w0[0], 0, selector);
w0[0] = 0;
break;
case 2:
w3[1] = __byte_perm_S (w2[3], w2[2], selector);
w3[0] = __byte_perm_S (w2[2], w2[1], selector);
w2[3] = __byte_perm_S (w2[1], w2[0], selector);
w2[2] = __byte_perm_S (w2[0], w1[3], selector);
w2[1] = __byte_perm_S (w1[3], w1[2], selector);
w2[0] = __byte_perm_S (w1[2], w1[1], selector);
w1[3] = __byte_perm_S (w1[1], w1[0], selector);
w1[2] = __byte_perm_S (w1[0], w0[3], selector);
w1[1] = __byte_perm_S (w0[3], w0[2], selector);
w1[0] = __byte_perm_S (w0[2], w0[1], selector);
w0[3] = __byte_perm_S (w0[1], w0[0], selector);
w0[2] = __byte_perm_S (w0[0], 0, selector);
w0[1] = 0;
w0[0] = 0;
break;
case 3:
w3[1] = __byte_perm_S (w2[2], w2[1], selector);
w3[0] = __byte_perm_S (w2[1], w2[0], selector);
w2[3] = __byte_perm_S (w2[0], w1[3], selector);
w2[2] = __byte_perm_S (w1[3], w1[2], selector);
w2[1] = __byte_perm_S (w1[2], w1[1], selector);
w2[0] = __byte_perm_S (w1[1], w1[0], selector);
w1[3] = __byte_perm_S (w1[0], w0[3], selector);
w1[2] = __byte_perm_S (w0[3], w0[2], selector);
w1[1] = __byte_perm_S (w0[2], w0[1], selector);
w1[0] = __byte_perm_S (w0[1], w0[0], selector);
w0[3] = __byte_perm_S (w0[0], 0, selector);
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 4:
w3[1] = __byte_perm_S (w2[1], w2[0], selector);
w3[0] = __byte_perm_S (w2[0], w1[3], selector);
w2[3] = __byte_perm_S (w1[3], w1[2], selector);
w2[2] = __byte_perm_S (w1[2], w1[1], selector);
w2[1] = __byte_perm_S (w1[1], w1[0], selector);
w2[0] = __byte_perm_S (w1[0], w0[3], selector);
w1[3] = __byte_perm_S (w0[3], w0[2], selector);
w1[2] = __byte_perm_S (w0[2], w0[1], selector);
w1[1] = __byte_perm_S (w0[1], w0[0], selector);
w1[0] = __byte_perm_S (w0[0], 0, selector);
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 5:
w3[1] = __byte_perm_S (w2[0], w1[3], selector);
w3[0] = __byte_perm_S (w1[3], w1[2], selector);
w2[3] = __byte_perm_S (w1[2], w1[1], selector);
w2[2] = __byte_perm_S (w1[1], w1[0], selector);
w2[1] = __byte_perm_S (w1[0], w0[3], selector);
w2[0] = __byte_perm_S (w0[3], w0[2], selector);
w1[3] = __byte_perm_S (w0[2], w0[1], selector);
w1[2] = __byte_perm_S (w0[1], w0[0], selector);
w1[1] = __byte_perm_S (w0[0], 0, selector);
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 6:
w3[1] = __byte_perm_S (w1[3], w1[2], selector);
w3[0] = __byte_perm_S (w1[2], w1[1], selector);
w2[3] = __byte_perm_S (w1[1], w1[0], selector);
w2[2] = __byte_perm_S (w1[0], w0[3], selector);
w2[1] = __byte_perm_S (w0[3], w0[2], selector);
w2[0] = __byte_perm_S (w0[2], w0[1], selector);
w1[3] = __byte_perm_S (w0[1], w0[0], selector);
w1[2] = __byte_perm_S (w0[0], 0, selector);
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 7:
w3[1] = __byte_perm_S (w1[2], w1[1], selector);
w3[0] = __byte_perm_S (w1[1], w1[0], selector);
w2[3] = __byte_perm_S (w1[0], w0[3], selector);
w2[2] = __byte_perm_S (w0[3], w0[2], selector);
w2[1] = __byte_perm_S (w0[2], w0[1], selector);
w2[0] = __byte_perm_S (w0[1], w0[0], selector);
w1[3] = __byte_perm_S (w0[0], 0, selector);
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 8:
w3[1] = __byte_perm_S (w1[1], w1[0], selector);
w3[0] = __byte_perm_S (w1[0], w0[3], selector);
w2[3] = __byte_perm_S (w0[3], w0[2], selector);
w2[2] = __byte_perm_S (w0[2], w0[1], selector);
w2[1] = __byte_perm_S (w0[1], w0[0], selector);
w2[0] = __byte_perm_S (w0[0], 0, selector);
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 9:
w3[1] = __byte_perm_S (w1[0], w0[3], selector);
w3[0] = __byte_perm_S (w0[3], w0[2], selector);
w2[3] = __byte_perm_S (w0[2], w0[1], selector);
w2[2] = __byte_perm_S (w0[1], w0[0], selector);
w2[1] = __byte_perm_S (w0[0], 0, selector);
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 10:
w3[1] = __byte_perm_S (w0[3], w0[2], selector);
w3[0] = __byte_perm_S (w0[2], w0[1], selector);
w2[3] = __byte_perm_S (w0[1], w0[0], selector);
w2[2] = __byte_perm_S (w0[0], 0, selector);
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 11:
w3[1] = __byte_perm_S (w0[2], w0[1], selector);
w3[0] = __byte_perm_S (w0[1], w0[0], selector);
w2[3] = __byte_perm_S (w0[0], 0, selector);
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 12:
w3[1] = __byte_perm_S (w0[1], w0[0], selector);
w3[0] = __byte_perm_S (w0[0], 0, selector);
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 13:
w3[1] = __byte_perm_S (w0[0], 0, selector);
w3[0] = 0;
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
}
#endif
}
static void switch_buffer_by_offset (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
{
Fix for -m 1500, -m 3000: Missing "defined" keyword Fix for -m 1000: MD4_H1 copy/paste error Fix for -m 8900, -m 9300: Invalid value for device_processor_cores for CPU devices Fix for -m 9100: Variable salt2 initializer needed some clauses for clearness Temporary limit gpu_accel for CPU devices to 1 for development phase Mark pocl as too bleeding edge for production use, recommend native drivers Remove workarounds for pocl Rename VENDOR_ID_UNKNOWN to VENDOR_ID_GENERIC in host code Rename IS_UNKNOWN to IS_GENERIC in kernel code
2016-01-07 19:14:34 +00:00
#if defined IS_AMD || defined IS_GENERIC
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
const int offset_mod_4 = offset & 3;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
const int offset_minus_4 = 4 - offset;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
switch (offset / 4)
{
case 0:
w3[2] = amd_bytealign ( 0, w3[1], offset_minus_4);
w3[1] = amd_bytealign (w3[1], w3[0], offset_minus_4);
w3[0] = amd_bytealign (w3[0], w2[3], offset_minus_4);
w2[3] = amd_bytealign (w2[3], w2[2], offset_minus_4);
w2[2] = amd_bytealign (w2[2], w2[1], offset_minus_4);
w2[1] = amd_bytealign (w2[1], w2[0], offset_minus_4);
w2[0] = amd_bytealign (w2[0], w1[3], offset_minus_4);
w1[3] = amd_bytealign (w1[3], w1[2], offset_minus_4);
w1[2] = amd_bytealign (w1[2], w1[1], offset_minus_4);
w1[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w1[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w0[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w0[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w0[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w0[0] = amd_bytealign (w0[0], 0, offset_minus_4);
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w0[0] = w0[1];
w0[1] = w0[2];
w0[2] = w0[3];
w0[3] = w1[0];
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 1:
w3[2] = amd_bytealign ( 0, w3[0], offset_minus_4);
w3[1] = amd_bytealign (w3[0], w2[3], offset_minus_4);
w3[0] = amd_bytealign (w2[3], w2[2], offset_minus_4);
w2[3] = amd_bytealign (w2[2], w2[1], offset_minus_4);
w2[2] = amd_bytealign (w2[1], w2[0], offset_minus_4);
w2[1] = amd_bytealign (w2[0], w1[3], offset_minus_4);
w2[0] = amd_bytealign (w1[3], w1[2], offset_minus_4);
w1[3] = amd_bytealign (w1[2], w1[1], offset_minus_4);
w1[2] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w1[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w1[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w0[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w0[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w0[1] = amd_bytealign (w0[0], 0, offset_minus_4);
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w0[1] = w0[2];
w0[2] = w0[3];
w0[3] = w1[0];
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
w3[2] = amd_bytealign ( 0, w2[3], offset_minus_4);
w3[1] = amd_bytealign (w2[3], w2[2], offset_minus_4);
w3[0] = amd_bytealign (w2[2], w2[1], offset_minus_4);
w2[3] = amd_bytealign (w2[1], w2[0], offset_minus_4);
w2[2] = amd_bytealign (w2[0], w1[3], offset_minus_4);
w2[1] = amd_bytealign (w1[3], w1[2], offset_minus_4);
w2[0] = amd_bytealign (w1[2], w1[1], offset_minus_4);
w1[3] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w1[2] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w1[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w1[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w0[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w0[2] = amd_bytealign (w0[0], 0, offset_minus_4);
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w0[2] = w0[3];
w0[3] = w1[0];
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
w3[2] = amd_bytealign ( 0, w2[2], offset_minus_4);
w3[1] = amd_bytealign (w2[2], w2[1], offset_minus_4);
w3[0] = amd_bytealign (w2[1], w2[0], offset_minus_4);
w2[3] = amd_bytealign (w2[0], w1[3], offset_minus_4);
w2[2] = amd_bytealign (w1[3], w1[2], offset_minus_4);
w2[1] = amd_bytealign (w1[2], w1[1], offset_minus_4);
w2[0] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w1[3] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w1[2] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w1[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w1[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w0[3] = amd_bytealign (w0[0], 0, offset_minus_4);
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w0[3] = w1[0];
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
w3[2] = amd_bytealign ( 0, w2[1], offset_minus_4);
w3[1] = amd_bytealign (w2[1], w2[0], offset_minus_4);
w3[0] = amd_bytealign (w2[0], w1[3], offset_minus_4);
w2[3] = amd_bytealign (w1[3], w1[2], offset_minus_4);
w2[2] = amd_bytealign (w1[2], w1[1], offset_minus_4);
w2[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w2[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w1[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w1[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w1[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w1[0] = amd_bytealign (w0[0], 0, offset_minus_4);
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w1[0] = w1[1];
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
w3[2] = amd_bytealign ( 0, w2[0], offset_minus_4);
w3[1] = amd_bytealign (w2[0], w1[3], offset_minus_4);
w3[0] = amd_bytealign (w1[3], w1[2], offset_minus_4);
w2[3] = amd_bytealign (w1[2], w1[1], offset_minus_4);
w2[2] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w2[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w2[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w1[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w1[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w1[1] = amd_bytealign (w0[0], 0, offset_minus_4);
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w1[1] = w1[2];
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 6:
w3[2] = amd_bytealign ( 0, w1[3], offset_minus_4);
w3[1] = amd_bytealign (w1[3], w1[2], offset_minus_4);
w3[0] = amd_bytealign (w1[2], w1[1], offset_minus_4);
w2[3] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w2[2] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w2[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w2[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w1[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w1[2] = amd_bytealign (w0[0], 0, offset_minus_4);
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w1[2] = w1[3];
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 7:
w3[2] = amd_bytealign ( 0, w1[2], offset_minus_4);
w3[1] = amd_bytealign (w1[2], w1[1], offset_minus_4);
w3[0] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w2[3] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w2[2] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w2[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w2[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w1[3] = amd_bytealign (w0[0], 0, offset_minus_4);
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w1[3] = w2[0];
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 8:
w3[2] = amd_bytealign ( 0, w1[1], offset_minus_4);
w3[1] = amd_bytealign (w1[1], w1[0], offset_minus_4);
w3[0] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w2[3] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w2[2] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w2[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w2[0] = amd_bytealign (w0[0], 0, offset_minus_4);
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w2[0] = w2[1];
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 9:
w3[2] = amd_bytealign ( 0, w1[0], offset_minus_4);
w3[1] = amd_bytealign (w1[0], w0[3], offset_minus_4);
w3[0] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w2[3] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w2[2] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w2[1] = amd_bytealign (w0[0], 0, offset_minus_4);
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w2[1] = w2[2];
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 10:
w3[2] = amd_bytealign ( 0, w0[3], offset_minus_4);
w3[1] = amd_bytealign (w0[3], w0[2], offset_minus_4);
w3[0] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w2[3] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w2[2] = amd_bytealign (w0[0], 0, offset_minus_4);
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w2[2] = w2[3];
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 11:
w3[2] = amd_bytealign ( 0, w0[2], offset_minus_4);
w3[1] = amd_bytealign (w0[2], w0[1], offset_minus_4);
w3[0] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w2[3] = amd_bytealign (w0[0], 0, offset_minus_4);
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
if (offset_mod_4 == 0)
{
w2[3] = w3[0];
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 12:
w3[2] = amd_bytealign ( 0, w0[1], offset_minus_4);
w3[1] = amd_bytealign (w0[1], w0[0], offset_minus_4);
w3[0] = amd_bytealign (w0[0], 0, offset_minus_4);
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w3[0] = w3[1];
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 13:
w3[2] = amd_bytealign ( 0, w0[0], offset_minus_4);
w3[1] = amd_bytealign (w0[0], 0, offset_minus_4);
w3[0] = 0;
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
if (offset_mod_4 == 0)
{
w3[1] = w3[2];
w3[2] = 0;
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
}
#endif
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
#ifdef IS_NV
const int offset_minus_4 = 4 - (offset % 4);
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
switch (offset / 4)
{
case 0:
w3[1] = __byte_perm (w3[0], w3[1], selector);
w3[0] = __byte_perm (w2[3], w3[0], selector);
w2[3] = __byte_perm (w2[2], w2[3], selector);
w2[2] = __byte_perm (w2[1], w2[2], selector);
w2[1] = __byte_perm (w2[0], w2[1], selector);
w2[0] = __byte_perm (w1[3], w2[0], selector);
w1[3] = __byte_perm (w1[2], w1[3], selector);
w1[2] = __byte_perm (w1[1], w1[2], selector);
w1[1] = __byte_perm (w1[0], w1[1], selector);
w1[0] = __byte_perm (w0[3], w1[0], selector);
w0[3] = __byte_perm (w0[2], w0[3], selector);
w0[2] = __byte_perm (w0[1], w0[2], selector);
w0[1] = __byte_perm (w0[0], w0[1], selector);
w0[0] = __byte_perm ( 0, w0[0], selector);
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 1:
w3[1] = __byte_perm (w2[3], w3[0], selector);
w3[0] = __byte_perm (w2[2], w2[3], selector);
w2[3] = __byte_perm (w2[1], w2[2], selector);
w2[2] = __byte_perm (w2[0], w2[1], selector);
w2[1] = __byte_perm (w1[3], w2[0], selector);
w2[0] = __byte_perm (w1[2], w1[3], selector);
w1[3] = __byte_perm (w1[1], w1[2], selector);
w1[2] = __byte_perm (w1[0], w1[1], selector);
w1[1] = __byte_perm (w0[3], w1[0], selector);
w1[0] = __byte_perm (w0[2], w0[3], selector);
w0[3] = __byte_perm (w0[1], w0[2], selector);
w0[2] = __byte_perm (w0[0], w0[1], selector);
w0[1] = __byte_perm ( 0, w0[0], selector);
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
w3[1] = __byte_perm (w2[2], w2[3], selector);
w3[0] = __byte_perm (w2[1], w2[2], selector);
w2[3] = __byte_perm (w2[0], w2[1], selector);
w2[2] = __byte_perm (w1[3], w2[0], selector);
w2[1] = __byte_perm (w1[2], w1[3], selector);
w2[0] = __byte_perm (w1[1], w1[2], selector);
w1[3] = __byte_perm (w1[0], w1[1], selector);
w1[2] = __byte_perm (w0[3], w1[0], selector);
w1[1] = __byte_perm (w0[2], w0[3], selector);
w1[0] = __byte_perm (w0[1], w0[2], selector);
w0[3] = __byte_perm (w0[0], w0[1], selector);
w0[2] = __byte_perm ( 0, w0[0], selector);
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
w3[1] = __byte_perm (w2[1], w2[2], selector);
w3[0] = __byte_perm (w2[0], w2[1], selector);
w2[3] = __byte_perm (w1[3], w2[0], selector);
w2[2] = __byte_perm (w1[2], w1[3], selector);
w2[1] = __byte_perm (w1[1], w1[2], selector);
w2[0] = __byte_perm (w1[0], w1[1], selector);
w1[3] = __byte_perm (w0[3], w1[0], selector);
w1[2] = __byte_perm (w0[2], w0[3], selector);
w1[1] = __byte_perm (w0[1], w0[2], selector);
w1[0] = __byte_perm (w0[0], w0[1], selector);
w0[3] = __byte_perm ( 0, w0[0], selector);
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
w3[1] = __byte_perm (w2[0], w2[1], selector);
w3[0] = __byte_perm (w1[3], w2[0], selector);
w2[3] = __byte_perm (w1[2], w1[3], selector);
w2[2] = __byte_perm (w1[1], w1[2], selector);
w2[1] = __byte_perm (w1[0], w1[1], selector);
w2[0] = __byte_perm (w0[3], w1[0], selector);
w1[3] = __byte_perm (w0[2], w0[3], selector);
w1[2] = __byte_perm (w0[1], w0[2], selector);
w1[1] = __byte_perm (w0[0], w0[1], selector);
w1[0] = __byte_perm ( 0, w0[0], selector);
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
w3[1] = __byte_perm (w1[3], w2[0], selector);
w3[0] = __byte_perm (w1[2], w1[3], selector);
w2[3] = __byte_perm (w1[1], w1[2], selector);
w2[2] = __byte_perm (w1[0], w1[1], selector);
w2[1] = __byte_perm (w0[3], w1[0], selector);
w2[0] = __byte_perm (w0[2], w0[3], selector);
w1[3] = __byte_perm (w0[1], w0[2], selector);
w1[2] = __byte_perm (w0[0], w0[1], selector);
w1[1] = __byte_perm ( 0, w0[0], selector);
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 6:
w3[1] = __byte_perm (w1[2], w1[3], selector);
w3[0] = __byte_perm (w1[1], w1[2], selector);
w2[3] = __byte_perm (w1[0], w1[1], selector);
w2[2] = __byte_perm (w0[3], w1[0], selector);
w2[1] = __byte_perm (w0[2], w0[3], selector);
w2[0] = __byte_perm (w0[1], w0[2], selector);
w1[3] = __byte_perm (w0[0], w0[1], selector);
w1[2] = __byte_perm ( 0, w0[0], selector);
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 7:
w3[1] = __byte_perm (w1[1], w1[2], selector);
w3[0] = __byte_perm (w1[0], w1[1], selector);
w2[3] = __byte_perm (w0[3], w1[0], selector);
w2[2] = __byte_perm (w0[2], w0[3], selector);
w2[1] = __byte_perm (w0[1], w0[2], selector);
w2[0] = __byte_perm (w0[0], w0[1], selector);
w1[3] = __byte_perm ( 0, w0[0], selector);
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 8:
w3[1] = __byte_perm (w1[0], w1[1], selector);
w3[0] = __byte_perm (w0[3], w1[0], selector);
w2[3] = __byte_perm (w0[2], w0[3], selector);
w2[2] = __byte_perm (w0[1], w0[2], selector);
w2[1] = __byte_perm (w0[0], w0[1], selector);
w2[0] = __byte_perm ( 0, w0[0], selector);
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 9:
w3[1] = __byte_perm (w0[3], w1[0], selector);
w3[0] = __byte_perm (w0[2], w0[3], selector);
w2[3] = __byte_perm (w0[1], w0[2], selector);
w2[2] = __byte_perm (w0[0], w0[1], selector);
w2[1] = __byte_perm ( 0, w0[0], selector);
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 10:
w3[1] = __byte_perm (w0[2], w0[3], selector);
w3[0] = __byte_perm (w0[1], w0[2], selector);
w2[3] = __byte_perm (w0[0], w0[1], selector);
w2[2] = __byte_perm ( 0, w0[0], selector);
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 11:
w3[1] = __byte_perm (w0[1], w0[2], selector);
w3[0] = __byte_perm (w0[0], w0[1], selector);
w2[3] = __byte_perm ( 0, w0[0], selector);
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 12:
w3[1] = __byte_perm (w0[0], w0[1], selector);
w3[0] = __byte_perm ( 0, w0[0], selector);
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 13:
w3[1] = __byte_perm ( 0, w0[0], selector);
w3[0] = 0;
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
}
#endif
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Prepare new SIMD code for kernel, -m 0, 10, 20, 1000 should work in -a 3 mode and other hopefully stay unaffected
2016-01-17 21:17:50 +00:00
static void switch_buffer_by_offset_be (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const u32 offset)
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
{
Fix for -m 1500, -m 3000: Missing "defined" keyword Fix for -m 1000: MD4_H1 copy/paste error Fix for -m 8900, -m 9300: Invalid value for device_processor_cores for CPU devices Fix for -m 9100: Variable salt2 initializer needed some clauses for clearness Temporary limit gpu_accel for CPU devices to 1 for development phase Mark pocl as too bleeding edge for production use, recommend native drivers Remove workarounds for pocl Rename VENDOR_ID_UNKNOWN to VENDOR_ID_GENERIC in host code Rename IS_UNKNOWN to IS_GENERIC in kernel code
2016-01-07 19:14:34 +00:00
#if defined IS_AMD || defined IS_GENERIC
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
switch (offset / 4)
{
case 0:
w3[2] = amd_bytealign (w3[1], 0, offset);
w3[1] = amd_bytealign (w3[0], w3[1], offset);
w3[0] = amd_bytealign (w2[3], w3[0], offset);
w2[3] = amd_bytealign (w2[2], w2[3], offset);
w2[2] = amd_bytealign (w2[1], w2[2], offset);
w2[1] = amd_bytealign (w2[0], w2[1], offset);
w2[0] = amd_bytealign (w1[3], w2[0], offset);
w1[3] = amd_bytealign (w1[2], w1[3], offset);
w1[2] = amd_bytealign (w1[1], w1[2], offset);
w1[1] = amd_bytealign (w1[0], w1[1], offset);
w1[0] = amd_bytealign (w0[3], w1[0], offset);
w0[3] = amd_bytealign (w0[2], w0[3], offset);
w0[2] = amd_bytealign (w0[1], w0[2], offset);
w0[1] = amd_bytealign (w0[0], w0[1], offset);
w0[0] = amd_bytealign ( 0, w0[0], offset);
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 1:
w3[2] = amd_bytealign (w3[0], 0, offset);
w3[1] = amd_bytealign (w2[3], w3[0], offset);
w3[0] = amd_bytealign (w2[2], w2[3], offset);
w2[3] = amd_bytealign (w2[1], w2[2], offset);
w2[2] = amd_bytealign (w2[0], w2[1], offset);
w2[1] = amd_bytealign (w1[3], w2[0], offset);
w2[0] = amd_bytealign (w1[2], w1[3], offset);
w1[3] = amd_bytealign (w1[1], w1[2], offset);
w1[2] = amd_bytealign (w1[0], w1[1], offset);
w1[1] = amd_bytealign (w0[3], w1[0], offset);
w1[0] = amd_bytealign (w0[2], w0[3], offset);
w0[3] = amd_bytealign (w0[1], w0[2], offset);
w0[2] = amd_bytealign (w0[0], w0[1], offset);
w0[1] = amd_bytealign ( 0, w0[0], offset);
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
w3[2] = amd_bytealign (w2[3], 0, offset);
w3[1] = amd_bytealign (w2[2], w2[3], offset);
w3[0] = amd_bytealign (w2[1], w2[2], offset);
w2[3] = amd_bytealign (w2[0], w2[1], offset);
w2[2] = amd_bytealign (w1[3], w2[0], offset);
w2[1] = amd_bytealign (w1[2], w1[3], offset);
w2[0] = amd_bytealign (w1[1], w1[2], offset);
w1[3] = amd_bytealign (w1[0], w1[1], offset);
w1[2] = amd_bytealign (w0[3], w1[0], offset);
w1[1] = amd_bytealign (w0[2], w0[3], offset);
w1[0] = amd_bytealign (w0[1], w0[2], offset);
w0[3] = amd_bytealign (w0[0], w0[1], offset);
w0[2] = amd_bytealign ( 0, w0[0], offset);
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
w3[2] = amd_bytealign (w2[2], 0, offset);
w3[1] = amd_bytealign (w2[1], w2[2], offset);
w3[0] = amd_bytealign (w2[0], w2[1], offset);
w2[3] = amd_bytealign (w1[3], w2[0], offset);
w2[2] = amd_bytealign (w1[2], w1[3], offset);
w2[1] = amd_bytealign (w1[1], w1[2], offset);
w2[0] = amd_bytealign (w1[0], w1[1], offset);
w1[3] = amd_bytealign (w0[3], w1[0], offset);
w1[2] = amd_bytealign (w0[2], w0[3], offset);
w1[1] = amd_bytealign (w0[1], w0[2], offset);
w1[0] = amd_bytealign (w0[0], w0[1], offset);
w0[3] = amd_bytealign ( 0, w0[0], offset);
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
w3[2] = amd_bytealign (w2[1], 0, offset);
w3[1] = amd_bytealign (w2[0], w2[1], offset);
w3[0] = amd_bytealign (w1[3], w2[0], offset);
w2[3] = amd_bytealign (w1[2], w1[3], offset);
w2[2] = amd_bytealign (w1[1], w1[2], offset);
w2[1] = amd_bytealign (w1[0], w1[1], offset);
w2[0] = amd_bytealign (w0[3], w1[0], offset);
w1[3] = amd_bytealign (w0[2], w0[3], offset);
w1[2] = amd_bytealign (w0[1], w0[2], offset);
w1[1] = amd_bytealign (w0[0], w0[1], offset);
w1[0] = amd_bytealign ( 0, w0[0], offset);
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
w3[2] = amd_bytealign (w2[0], 0, offset);
w3[1] = amd_bytealign (w1[3], w2[0], offset);
w3[0] = amd_bytealign (w1[2], w1[3], offset);
w2[3] = amd_bytealign (w1[1], w1[2], offset);
w2[2] = amd_bytealign (w1[0], w1[1], offset);
w2[1] = amd_bytealign (w0[3], w1[0], offset);
w2[0] = amd_bytealign (w0[2], w0[3], offset);
w1[3] = amd_bytealign (w0[1], w0[2], offset);
w1[2] = amd_bytealign (w0[0], w0[1], offset);
w1[1] = amd_bytealign ( 0, w0[0], offset);
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 6:
w3[2] = amd_bytealign (w1[3], 0, offset);
w3[1] = amd_bytealign (w1[2], w1[3], offset);
w3[0] = amd_bytealign (w1[1], w1[2], offset);
w2[3] = amd_bytealign (w1[0], w1[1], offset);
w2[2] = amd_bytealign (w0[3], w1[0], offset);
w2[1] = amd_bytealign (w0[2], w0[3], offset);
w2[0] = amd_bytealign (w0[1], w0[2], offset);
w1[3] = amd_bytealign (w0[0], w0[1], offset);
w1[2] = amd_bytealign ( 0, w0[0], offset);
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 7:
w3[2] = amd_bytealign (w1[2], 0, offset);
w3[1] = amd_bytealign (w1[1], w1[2], offset);
w3[0] = amd_bytealign (w1[0], w1[1], offset);
w2[3] = amd_bytealign (w0[3], w1[0], offset);
w2[2] = amd_bytealign (w0[2], w0[3], offset);
w2[1] = amd_bytealign (w0[1], w0[2], offset);
w2[0] = amd_bytealign (w0[0], w0[1], offset);
w1[3] = amd_bytealign ( 0, w0[0], offset);
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 8:
w3[2] = amd_bytealign (w1[1], 0, offset);
w3[1] = amd_bytealign (w1[0], w1[1], offset);
w3[0] = amd_bytealign (w0[3], w1[0], offset);
w2[3] = amd_bytealign (w0[2], w0[3], offset);
w2[2] = amd_bytealign (w0[1], w0[2], offset);
w2[1] = amd_bytealign (w0[0], w0[1], offset);
w2[0] = amd_bytealign ( 0, w0[0], offset);
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 9:
w3[2] = amd_bytealign (w1[0], 0, offset);
w3[1] = amd_bytealign (w0[3], w1[0], offset);
w3[0] = amd_bytealign (w0[2], w0[3], offset);
w2[3] = amd_bytealign (w0[1], w0[2], offset);
w2[2] = amd_bytealign (w0[0], w0[1], offset);
w2[1] = amd_bytealign ( 0, w0[0], offset);
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 10:
w3[2] = amd_bytealign (w0[3], 0, offset);
w3[1] = amd_bytealign (w0[2], w0[3], offset);
w3[0] = amd_bytealign (w0[1], w0[2], offset);
w2[3] = amd_bytealign (w0[0], w0[1], offset);
w2[2] = amd_bytealign ( 0, w0[0], offset);
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 11:
w3[2] = amd_bytealign (w0[2], 0, offset);
w3[1] = amd_bytealign (w0[1], w0[2], offset);
w3[0] = amd_bytealign (w0[0], w0[1], offset);
w2[3] = amd_bytealign ( 0, w0[0], offset);
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 12:
w3[2] = amd_bytealign (w0[1], 0, offset);
w3[1] = amd_bytealign (w0[0], w0[1], offset);
w3[0] = amd_bytealign ( 0, w0[0], offset);
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 13:
w3[2] = amd_bytealign (w0[0], 0, offset);
w3[1] = amd_bytealign ( 0, w0[0], offset);
w3[0] = 0;
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
}
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
#endif
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
#ifdef IS_NV
const int selector = (0x76543210 >> ((offset & 3) * 4)) & 0xffff;
switch (offset / 4)
Fix append_* function calls
2015-12-15 12:42:37 +00:00
{
case 0:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w3[1], w3[0], selector);
w3[0] = __byte_perm (w3[0], w2[3], selector);
w2[3] = __byte_perm (w2[3], w2[2], selector);
w2[2] = __byte_perm (w2[2], w2[1], selector);
w2[1] = __byte_perm (w2[1], w2[0], selector);
w2[0] = __byte_perm (w2[0], w1[3], selector);
w1[3] = __byte_perm (w1[3], w1[2], selector);
w1[2] = __byte_perm (w1[2], w1[1], selector);
w1[1] = __byte_perm (w1[1], w1[0], selector);
w1[0] = __byte_perm (w1[0], w0[3], selector);
w0[3] = __byte_perm (w0[3], w0[2], selector);
w0[2] = __byte_perm (w0[2], w0[1], selector);
w0[1] = __byte_perm (w0[1], w0[0], selector);
w0[0] = __byte_perm (w0[0], 0, selector);
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 1:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w3[0], w2[3], selector);
w3[0] = __byte_perm (w2[3], w2[2], selector);
w2[3] = __byte_perm (w2[2], w2[1], selector);
w2[2] = __byte_perm (w2[1], w2[0], selector);
w2[1] = __byte_perm (w2[0], w1[3], selector);
w2[0] = __byte_perm (w1[3], w1[2], selector);
w1[3] = __byte_perm (w1[2], w1[1], selector);
w1[2] = __byte_perm (w1[1], w1[0], selector);
w1[1] = __byte_perm (w1[0], w0[3], selector);
w1[0] = __byte_perm (w0[3], w0[2], selector);
w0[3] = __byte_perm (w0[2], w0[1], selector);
w0[2] = __byte_perm (w0[1], w0[0], selector);
w0[1] = __byte_perm (w0[0], 0, selector);
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 2:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w2[3], w2[2], selector);
w3[0] = __byte_perm (w2[2], w2[1], selector);
w2[3] = __byte_perm (w2[1], w2[0], selector);
w2[2] = __byte_perm (w2[0], w1[3], selector);
w2[1] = __byte_perm (w1[3], w1[2], selector);
w2[0] = __byte_perm (w1[2], w1[1], selector);
w1[3] = __byte_perm (w1[1], w1[0], selector);
w1[2] = __byte_perm (w1[0], w0[3], selector);
w1[1] = __byte_perm (w0[3], w0[2], selector);
w1[0] = __byte_perm (w0[2], w0[1], selector);
w0[3] = __byte_perm (w0[1], w0[0], selector);
w0[2] = __byte_perm (w0[0], 0, selector);
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 3:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w2[2], w2[1], selector);
w3[0] = __byte_perm (w2[1], w2[0], selector);
w2[3] = __byte_perm (w2[0], w1[3], selector);
w2[2] = __byte_perm (w1[3], w1[2], selector);
w2[1] = __byte_perm (w1[2], w1[1], selector);
w2[0] = __byte_perm (w1[1], w1[0], selector);
w1[3] = __byte_perm (w1[0], w0[3], selector);
w1[2] = __byte_perm (w0[3], w0[2], selector);
w1[1] = __byte_perm (w0[2], w0[1], selector);
w1[0] = __byte_perm (w0[1], w0[0], selector);
w0[3] = __byte_perm (w0[0], 0, selector);
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 4:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w2[1], w2[0], selector);
w3[0] = __byte_perm (w2[0], w1[3], selector);
w2[3] = __byte_perm (w1[3], w1[2], selector);
w2[2] = __byte_perm (w1[2], w1[1], selector);
w2[1] = __byte_perm (w1[1], w1[0], selector);
w2[0] = __byte_perm (w1[0], w0[3], selector);
w1[3] = __byte_perm (w0[3], w0[2], selector);
w1[2] = __byte_perm (w0[2], w0[1], selector);
w1[1] = __byte_perm (w0[1], w0[0], selector);
w1[0] = __byte_perm (w0[0], 0, selector);
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 5:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w2[0], w1[3], selector);
w3[0] = __byte_perm (w1[3], w1[2], selector);
w2[3] = __byte_perm (w1[2], w1[1], selector);
w2[2] = __byte_perm (w1[1], w1[0], selector);
w2[1] = __byte_perm (w1[0], w0[3], selector);
w2[0] = __byte_perm (w0[3], w0[2], selector);
w1[3] = __byte_perm (w0[2], w0[1], selector);
w1[2] = __byte_perm (w0[1], w0[0], selector);
w1[1] = __byte_perm (w0[0], 0, selector);
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
break;
case 6:
w3[1] = __byte_perm (w1[3], w1[2], selector);
w3[0] = __byte_perm (w1[2], w1[1], selector);
w2[3] = __byte_perm (w1[1], w1[0], selector);
w2[2] = __byte_perm (w1[0], w0[3], selector);
w2[1] = __byte_perm (w0[3], w0[2], selector);
w2[0] = __byte_perm (w0[2], w0[1], selector);
w1[3] = __byte_perm (w0[1], w0[0], selector);
w1[2] = __byte_perm (w0[0], 0, selector);
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 7:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w1[2], w1[1], selector);
w3[0] = __byte_perm (w1[1], w1[0], selector);
w2[3] = __byte_perm (w1[0], w0[3], selector);
w2[2] = __byte_perm (w0[3], w0[2], selector);
w2[1] = __byte_perm (w0[2], w0[1], selector);
w2[0] = __byte_perm (w0[1], w0[0], selector);
w1[3] = __byte_perm (w0[0], 0, selector);
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 8:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w1[1], w1[0], selector);
w3[0] = __byte_perm (w1[0], w0[3], selector);
w2[3] = __byte_perm (w0[3], w0[2], selector);
w2[2] = __byte_perm (w0[2], w0[1], selector);
w2[1] = __byte_perm (w0[1], w0[0], selector);
w2[0] = __byte_perm (w0[0], 0, selector);
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 9:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w1[0], w0[3], selector);
w3[0] = __byte_perm (w0[3], w0[2], selector);
w2[3] = __byte_perm (w0[2], w0[1], selector);
w2[2] = __byte_perm (w0[1], w0[0], selector);
w2[1] = __byte_perm (w0[0], 0, selector);
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 10:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w0[3], w0[2], selector);
w3[0] = __byte_perm (w0[2], w0[1], selector);
w2[3] = __byte_perm (w0[1], w0[0], selector);
w2[2] = __byte_perm (w0[0], 0, selector);
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 11:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w0[2], w0[1], selector);
w3[0] = __byte_perm (w0[1], w0[0], selector);
w2[3] = __byte_perm (w0[0], 0, selector);
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 12:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w0[1], w0[0], selector);
w3[0] = __byte_perm (w0[0], 0, selector);
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 13:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w3[1] = __byte_perm (w0[0], 0, selector);
w3[0] = 0;
w2[3] = 0;
w2[2] = 0;
w2[1] = 0;
w2[0] = 0;
w1[3] = 0;
w1[2] = 0;
w1[1] = 0;
w1[0] = 0;
w0[3] = 0;
w0[2] = 0;
w0[1] = 0;
w0[0] = 0;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
}
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
#endif
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
}
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
/* not needed anymore?
// before: append_0x80_2_be
static void append_0x80_2x4_be (u32 w0[4], u32 w1[4], const u32 offset)
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
{
switch (offset)
{
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 0:
w0[0] |= 0x80000000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 1:
w0[0] |= 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
w0[0] |= 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
w0[0] |= 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
w0[1] |= 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
w0[1] |= 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 6:
w0[1] |= 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 7:
w0[1] |= 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 8:
w0[2] |= 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 9:
w0[2] |= 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 10:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w0[2] |= 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 11:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w0[2] |= 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 12:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w0[3] |= 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 13:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w0[3] |= 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 14:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w0[3] |= 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 15:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w0[3] |= 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 16:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[0] |= 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 17:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[0] |= 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 18:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[0] |= 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 19:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[0] |= 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 20:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[1] |= 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 21:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[1] |= 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 22:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[1] |= 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 23:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[1] |= 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 24:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[2] |= 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 25:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[2] |= 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 26:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[2] |= 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 27:
w1[2] |= 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 28:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[3] |= 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 29:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[3] |= 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 30:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[3] |= 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 31:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w1[3] |= 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
}
}
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
// before: append_0x80_8
static void append_0x80_1x32 (u32 w[32], const u32 offset)
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
{
switch (offset)
{
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 0:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 0] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 1:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 0] = w[ 0] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 2:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 0] = w[ 0] | 0x800000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 3:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 0] = w[ 0] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 4:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 1] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 5:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 1] = w[ 1] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 6:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 1] = w[ 1] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 7:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 1] = w[ 1] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 8:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 2] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 9:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 2] = w[ 2] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 10:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 2] = w[ 2] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 11:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 2] = w[ 2] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 12:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 3] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 13:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 3] = w[ 3] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 14:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 3] = w[ 3] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 15:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 3] = w[ 3] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 16:
w[ 4] = 0x80;
break;
case 17:
w[ 4] = w[ 4] | 0x8000;
break;
case 18:
w[ 4] = w[ 4] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 19:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 4] = w[ 4] | 0x80000000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 20:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 5] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 21:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 5] = w[ 5] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 22:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 5] = w[ 5] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 23:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 5] = w[ 5] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 24:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 6] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 25:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 6] = w[ 6] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 26:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 6] = w[ 6] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 27:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 6] = w[ 6] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 28:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 7] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 29:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 7] = w[ 7] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 30:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 7] = w[ 7] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 31:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 7] = w[ 7] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 32:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
w[ 8] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 33:
w[ 8] = w[ 8] | 0x8000;
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 34:
w[ 8] = w[ 8] | 0x800000;
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 35:
w[ 8] = w[ 8] | 0x80000000;
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 36:
w[ 9] = 0x80;
break;
case 37:
w[ 9] = w[ 9] | 0x8000;
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 38:
w[ 9] = w[ 9] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 39:
w[ 9] = w[ 9] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 40:
w[10] = 0x80;
break;
case 41:
w[10] = w[10] | 0x8000;
break;
case 42:
w[10] = w[10] | 0x800000;
break;
case 43:
w[10] = w[10] | 0x80000000;
break;
case 44:
w[11] = 0x80;
break;
case 45:
w[11] = w[11] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 46:
w[11] = w[11] | 0x800000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 47:
w[11] = w[11] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 48:
w[12] = 0x80;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 49:
w[12] = w[12] | 0x8000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 50:
w[12] = w[12] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 51:
w[12] = w[12] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 52:
w[13] = 0x80;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 53:
w[13] = w[13] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 54:
w[13] = w[13] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 55:
w[13] = w[13] | 0x80000000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 56:
w[14] = 0x80;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 57:
w[14] = w[14] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 58:
w[14] = w[14] | 0x800000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 59:
w[14] = w[14] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 60:
w[15] = 0x80;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 61:
w[15] = w[15] | 0x8000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 62:
w[15] = w[15] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 63:
w[15] = w[15] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 64:
w[16] = 0x80;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 65:
w[16] = w[16] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 66:
w[16] = w[16] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 67:
w[16] = w[16] | 0x80000000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 68:
w[17] = 0x80;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 69:
w[17] = w[17] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 70:
w[17] = w[17] | 0x800000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 71:
w[17] = w[17] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 72:
w[18] = 0x80;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 73:
w[18] = w[18] | 0x8000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 74:
w[18] = w[18] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 75:
w[18] = w[18] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 76:
w[19] = 0x80;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 77:
w[19] = w[19] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 78:
w[19] = w[19] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 79:
w[19] = w[19] | 0x80000000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 80:
w[20] = 0x80;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 81:
w[20] = w[20] | 0x8000;
break;
case 82:
w[20] = w[20] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 83:
w[20] = w[20] | 0x80000000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 84:
w[21] = 0x80;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 85:
w[21] = w[21] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 86:
w[21] = w[21] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 87:
w[21] = w[21] | 0x80000000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 88:
w[22] = 0x80;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 89:
w[22] = w[22] | 0x8000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 90:
w[22] = w[22] | 0x800000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 91:
w[22] = w[22] | 0x80000000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 92:
w[23] = 0x80;
break;
case 93:
w[23] = w[23] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 94:
w[23] = w[23] | 0x800000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 95:
w[23] = w[23] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 96:
w[24] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 97:
w[24] = w[24] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 98:
w[24] = w[24] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 99:
w[24] = w[24] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 100:
w[25] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 101:
w[25] = w[25] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 102:
w[25] = w[25] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 103:
w[25] = w[25] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 104:
w[26] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 105:
w[26] = w[26] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 106:
w[26] = w[26] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 107:
w[26] = w[26] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 108:
w[27] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 109:
w[27] = w[27] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 110:
w[27] = w[27] | 0x800000;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 111:
w[27] = w[27] | 0x80000000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 112:
w[28] = 0x80;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 113:
w[28] = w[28] | 0x8000;
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 114:
w[28] = w[28] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 115:
w[28] = w[28] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 116:
w[29] = 0x80;
break;
case 117:
w[29] = w[29] | 0x8000;
break;
case 118:
w[29] = w[29] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 119:
w[29] = w[29] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 120:
w[30] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 121:
w[30] = w[30] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 122:
w[30] = w[30] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 123:
w[30] = w[30] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 124:
w[31] = 0x80;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 125:
w[31] = w[31] | 0x8000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 126:
w[31] = w[31] | 0x800000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 127:
w[31] = w[31] | 0x80000000;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
}
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
// before: device_memcat2L
static void memcat_c7_d1x2_sl1x2_sr1x2 (const u32 offset, u32 dst0[2], u32 src_l0[2], u32 src_r0[2])
{
switch (offset)
{
case 1:
dst0[0] = src_l0[0] | src_r0[0] << 8;
dst0[1] = src_r0[0] >> 24 | src_r0[1] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
dst0[0] = src_l0[0] | src_r0[0] << 16;
dst0[1] = src_r0[0] >> 16 | src_r0[1] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
dst0[0] = src_l0[0] | src_r0[0] << 24;
dst0[1] = src_r0[0] >> 8 | src_r0[1] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
dst0[1] = src_r0[0];
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
dst0[1] = src_l0[1] | src_r0[0] << 8;
break;
case 6:
dst0[1] = src_l0[1] | src_r0[0] << 16;
break;
case 7:
dst0[1] = src_l0[1] | src_r0[0] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
}
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
// before: device_memcat4L
static void memcat_c15_d1x4_sl1x4_sr1x4 (const u32 offset, u32 dst0[4], u32 src_l0[4], u32 src_r0[4])
{
switch (offset)
{
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 1:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 8;
dst0[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst0[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst0[3] = src_r0[2] >> 24 | src_r0[3] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 2:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 16;
dst0[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst0[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst0[3] = src_r0[2] >> 16 | src_r0[3] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 3:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 24;
dst0[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst0[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst0[3] = src_r0[2] >> 8 | src_r0[3] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 4:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_r0[0];
dst0[2] = src_r0[1];
dst0[3] = src_r0[2];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 5:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 8;
dst0[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst0[3] = src_r0[1] >> 24 | src_r0[2] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 6:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 16;
dst0[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst0[3] = src_r0[1] >> 16 | src_r0[2] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 7:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 24;
dst0[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst0[3] = src_r0[1] >> 8 | src_r0[2] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 8:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_r0[0];
dst0[3] = src_r0[1];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 9:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 8;
dst0[3] = src_r0[0] >> 24 | src_r0[1] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 10:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 16;
dst0[3] = src_r0[0] >> 16 | src_r0[1] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 11:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 24;
dst0[3] = src_r0[0] >> 8 | src_r0[1] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 12:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_r0[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 13:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 8;
break;
case 14:
dst0[3] = src_l0[3] | src_r0[0] << 16;
break;
case 15:
dst0[3] = src_l0[3] | src_r0[0] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
}
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
// before: device_memcat8L
static void memcat_c31_d2x4_sl2x4_sr1x4 (const u32 offset, u32 dst0[4], u32 dst1[4], u32 src_l0[4], u32 src_l1[4], u32 src_r0[4])
Fix append_* function calls
2015-12-15 12:42:37 +00:00
{
switch (offset)
{
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 1:
dst0[0] = src_l0[0] | src_r0[0] << 8;
dst0[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst0[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst0[3] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[0] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
dst0[0] = src_l0[0] | src_r0[0] << 16;
dst0[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst0[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst0[3] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[0] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
dst0[0] = src_l0[0] | src_r0[0] << 24;
dst0[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst0[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst0[3] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[0] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
dst0[1] = src_r0[0];
dst0[2] = src_r0[1];
dst0[3] = src_r0[2];
dst1[0] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
dst0[1] = src_l0[1] | src_r0[0] << 8;
dst0[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst0[3] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[0] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[1] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 6:
dst0[1] = src_l0[1] | src_r0[0] << 16;
dst0[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst0[3] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[0] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[1] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 7:
dst0[1] = src_l0[1] | src_r0[0] << 24;
dst0[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst0[3] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[0] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[1] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 8:
dst0[2] = src_r0[0];
dst0[3] = src_r0[1];
dst1[0] = src_r0[2];
dst1[1] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 9:
dst0[2] = src_l0[2] | src_r0[0] << 8;
dst0[3] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[0] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[1] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[2] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 10:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 16;
dst0[3] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[0] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[1] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[2] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 11:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 24;
dst0[3] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[0] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[1] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[2] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 12:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_r0[0];
dst1[0] = src_r0[1];
dst1[1] = src_r0[2];
dst1[2] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 13:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 8;
dst1[0] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[1] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[2] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[3] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 14:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 16;
dst1[0] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[1] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[2] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[3] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 15:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 24;
dst1[0] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[1] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[2] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[3] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 16:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_r0[0];
dst1[1] = src_r0[1];
dst1[2] = src_r0[2];
dst1[3] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 17:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 8;
dst1[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[3] = src_r0[2] >> 24 | src_r0[3] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 18:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 16;
dst1[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[3] = src_r0[2] >> 16 | src_r0[3] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 19:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 24;
dst1[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[3] = src_r0[2] >> 8 | src_r0[3] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 20:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_r0[0];
dst1[2] = src_r0[1];
dst1[3] = src_r0[2];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 21:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 8;
dst1[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[3] = src_r0[1] >> 24 | src_r0[2] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 22:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 16;
dst1[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[3] = src_r0[1] >> 16 | src_r0[2] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 23:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 24;
dst1[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[3] = src_r0[1] >> 8 | src_r0[2] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 24:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_r0[0];
dst1[3] = src_r0[1];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 25:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 8;
dst1[3] = src_r0[0] >> 24 | src_r0[1] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 26:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 16;
dst1[3] = src_r0[0] >> 16 | src_r0[1] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 27:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 24;
dst1[3] = src_r0[0] >> 8 | src_r0[1] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 28:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_r0[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 29:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 30:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 31:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
}
}
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
// before: device_memcat12L
static void memcat_c47_d3x4_sl3x4_sr1x4 (const u32 offset, u32 dst0[4], u32 dst1[4], u32 dst2[4], u32 src_l0[4], u32 src_l1[4], u32 src_l2[4], u32 src_r0[4])
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
{
switch (offset)
{
case 1:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 8;
dst0[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst0[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst0[3] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[0] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 2:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 16;
dst0[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst0[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst0[3] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[0] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 3:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 24;
dst0[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst0[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst0[3] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[0] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 4:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_r0[0];
dst0[2] = src_r0[1];
dst0[3] = src_r0[2];
dst1[0] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 5:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 8;
dst0[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst0[3] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[0] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[1] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 6:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 16;
dst0[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst0[3] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[0] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[1] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 7:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 24;
dst0[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst0[3] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[0] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[1] = src_r0[3] >> 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 8:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_r0[0];
dst0[3] = src_r0[1];
dst1[0] = src_r0[2];
dst1[1] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 9:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 8;
dst0[3] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[0] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[1] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[2] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 10:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 16;
dst0[3] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[0] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[1] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[2] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 11:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 24;
dst0[3] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[0] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[1] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[2] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 12:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_r0[0];
dst1[0] = src_r0[1];
dst1[1] = src_r0[2];
dst1[2] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 13:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 8;
dst1[0] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[1] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[2] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[3] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 14:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 16;
dst1[0] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[1] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[2] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[3] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 15:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 24;
dst1[0] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[1] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[2] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[3] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 16:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_r0[0];
dst1[1] = src_r0[1];
dst1[2] = src_r0[2];
dst1[3] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 17:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 8;
dst1[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[3] = src_r0[2] >> 24 | src_r0[3] << 8;
dst2[0] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 18:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 16;
dst1[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[3] = src_r0[2] >> 16 | src_r0[3] << 16;
dst2[0] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 19:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 24;
dst1[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[3] = src_r0[2] >> 8 | src_r0[3] << 24;
dst2[0] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 20:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_r0[0];
dst1[2] = src_r0[1];
dst1[3] = src_r0[2];
dst2[0] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 21:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 8;
dst1[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[3] = src_r0[1] >> 24 | src_r0[2] << 8;
dst2[0] = src_r0[2] >> 24 | src_r0[3] << 8;
dst2[1] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 22:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 16;
dst1[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[3] = src_r0[1] >> 16 | src_r0[2] << 16;
dst2[0] = src_r0[2] >> 16 | src_r0[3] << 16;
dst2[1] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 23:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 24;
dst1[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[3] = src_r0[1] >> 8 | src_r0[2] << 24;
dst2[0] = src_r0[2] >> 8 | src_r0[3] << 24;
dst2[1] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 24:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_r0[0];
dst1[3] = src_r0[1];
dst2[0] = src_r0[2];
dst2[1] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 25:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 8;
dst1[3] = src_r0[0] >> 24 | src_r0[1] << 8;
dst2[0] = src_r0[1] >> 24 | src_r0[2] << 8;
dst2[1] = src_r0[2] >> 24 | src_r0[3] << 8;
dst2[2] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 26:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 16;
dst1[3] = src_r0[0] >> 16 | src_r0[1] << 16;
dst2[0] = src_r0[1] >> 16 | src_r0[2] << 16;
dst2[1] = src_r0[2] >> 16 | src_r0[3] << 16;
dst2[2] = src_r0[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 27:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 24;
dst1[3] = src_r0[0] >> 8 | src_r0[1] << 24;
dst2[0] = src_r0[1] >> 8 | src_r0[2] << 24;
dst2[1] = src_r0[2] >> 8 | src_r0[3] << 24;
dst2[2] = src_r0[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 28:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_r0[0];
dst2[0] = src_r0[1];
dst2[1] = src_r0[2];
dst2[2] = src_r0[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 29:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 8;
dst2[0] = src_r0[0] >> 24 | src_r0[1] << 8;
dst2[1] = src_r0[1] >> 24 | src_r0[2] << 8;
dst2[2] = src_r0[2] >> 24 | src_r0[3] << 8;
dst2[3] = src_r0[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
case 30:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 16;
dst2[0] = src_r0[0] >> 16 | src_r0[1] << 16;
dst2[1] = src_r0[1] >> 16 | src_r0[2] << 16;
dst2[2] = src_r0[2] >> 16 | src_r0[3] << 16;
dst2[3] = src_r0[3] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 31:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 24;
dst2[0] = src_r0[0] >> 8 | src_r0[1] << 24;
dst2[1] = src_r0[1] >> 8 | src_r0[2] << 24;
dst2[2] = src_r0[2] >> 8 | src_r0[3] << 24;
dst2[3] = src_r0[3] >> 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 32:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[0] = src_r0[0];
dst2[1] = src_r0[1];
dst2[2] = src_r0[2];
dst2[3] = src_r0[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 33:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[0] = src_l2[0] | src_r0[0] << 8;
dst2[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst2[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst2[3] = src_r0[2] >> 24 | src_r0[3] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 34:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[0] = src_l2[0] | src_r0[0] << 16;
dst2[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst2[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst2[3] = src_r0[2] >> 16 | src_r0[3] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 35:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[0] = src_l2[0] | src_r0[0] << 24;
dst2[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst2[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst2[3] = src_r0[2] >> 8 | src_r0[3] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 36:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[1] = src_r0[0];
dst2[2] = src_r0[1];
dst2[3] = src_r0[2];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 37:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[1] = src_l2[1] | src_r0[0] << 8;
dst2[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst2[3] = src_r0[1] >> 24 | src_r0[2] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 38:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[1] = src_l2[1] | src_r0[0] << 16;
dst2[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst2[3] = src_r0[1] >> 16 | src_r0[2] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 39:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[1] = src_l2[1] | src_r0[0] << 24;
dst2[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst2[3] = src_r0[1] >> 8 | src_r0[2] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 40:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[2] = src_r0[0];
dst2[3] = src_r0[1];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 41:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[2] = src_l2[2] | src_r0[0] << 8;
dst2[3] = src_r0[0] >> 24 | src_r0[1] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 42:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[2] = src_l2[2] | src_r0[0] << 16;
dst2[3] = src_r0[0] >> 16 | src_r0[1] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 43:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[2] = src_l2[2] | src_r0[0] << 24;
dst2[3] = src_r0[0] >> 8 | src_r0[1] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 44:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[3] = src_r0[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 45:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[3] = src_l2[3] | src_r0[0] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 46:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[3] = src_l2[3] | src_r0[0] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 47:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[3] = src_l2[3] | src_r0[0] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
}
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
// before: device_memcat12L
static void memcat_c47_d3x4_sl3x4_sr2x4 (const u32 offset, u32 dst0[4], u32 dst1[4], u32 dst2[4], u32 src_l0[4], u32 src_l1[4], u32 src_l2[4], u32 src_r0[4], u32 src_r1[4])
Fix append_* function calls
2015-12-15 12:42:37 +00:00
{
switch (offset)
{
case 0:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_r0[0];
dst0[1] = src_r0[1];
dst0[2] = src_r0[2];
dst0[3] = src_r0[3];
dst1[0] = src_r1[0];
dst1[1] = src_r1[1];
dst1[2] = src_r1[2];
dst1[3] = src_r1[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 1:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 8;
dst0[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst0[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst0[3] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[0] = src_r0[3] >> 24 | src_r1[0] << 8;
dst1[1] = src_r1[0] >> 24 | src_r1[1] << 8;
dst1[2] = src_r1[1] >> 24 | src_r1[2] << 8;
dst1[3] = src_r1[2] >> 24 | src_r1[3] << 8;
dst2[0] = src_r1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 2:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 16;
dst0[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst0[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst0[3] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[0] = src_r0[3] >> 16 | src_r1[0] << 16;
dst1[1] = src_r1[0] >> 16 | src_r1[1] << 16;
dst1[2] = src_r1[1] >> 16 | src_r1[2] << 16;
dst1[3] = src_r1[2] >> 16 | src_r1[3] << 16;
dst2[0] = src_r1[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 3:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[0] = src_l0[0] | src_r0[0] << 24;
dst0[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst0[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst0[3] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[0] = src_r0[3] >> 8 | src_r1[0] << 24;
dst1[1] = src_r1[0] >> 8 | src_r1[1] << 24;
dst1[2] = src_r1[1] >> 8 | src_r1[2] << 24;
dst1[3] = src_r1[2] >> 8 | src_r1[3] << 24;
dst2[0] = src_r1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 4:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_r0[0];
dst0[2] = src_r0[1];
dst0[3] = src_r0[2];
dst1[0] = src_r0[3];
dst1[1] = src_r1[0];
dst1[2] = src_r1[1];
dst1[3] = src_r1[2];
dst2[0] = src_r1[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 5:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 8;
dst0[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst0[3] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[0] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[1] = src_r0[3] >> 24 | src_r1[0] << 8;
dst1[2] = src_r1[0] >> 24 | src_r1[1] << 8;
dst1[3] = src_r1[1] >> 24 | src_r1[2] << 8;
dst2[0] = src_r1[2] >> 24 | src_r1[3] << 8;
dst2[1] = src_r1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 6:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 16;
dst0[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst0[3] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[0] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[1] = src_r0[3] >> 16 | src_r1[0] << 16;
dst1[2] = src_r1[0] >> 16 | src_r1[1] << 16;
dst1[3] = src_r1[1] >> 16 | src_r1[2] << 16;
dst2[0] = src_r1[2] >> 16 | src_r1[3] << 16;
dst2[1] = src_r1[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 7:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[1] = src_l0[1] | src_r0[0] << 24;
dst0[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst0[3] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[0] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[1] = src_r0[3] >> 8 | src_r1[0] << 24;
dst1[2] = src_r1[0] >> 8 | src_r1[1] << 24;
dst1[3] = src_r1[1] >> 8 | src_r1[2] << 24;
dst2[0] = src_r1[2] >> 8 | src_r1[3] << 24;
dst2[1] = src_r1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 8:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_r0[0];
dst0[3] = src_r0[1];
dst1[0] = src_r0[2];
dst1[1] = src_r0[3];
dst1[2] = src_r1[0];
dst1[3] = src_r1[1];
dst2[0] = src_r1[2];
dst2[1] = src_r1[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 9:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 8;
dst0[3] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[0] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[1] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[2] = src_r0[3] >> 24 | src_r1[0] << 8;
dst1[3] = src_r1[0] >> 24 | src_r1[1] << 8;
dst2[0] = src_r1[1] >> 24 | src_r1[2] << 8;
dst2[1] = src_r1[2] >> 24 | src_r1[3] << 8;
dst2[2] = src_r1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 10:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 16;
dst0[3] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[0] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[1] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[2] = src_r0[3] >> 16 | src_r1[0] << 16;
dst1[3] = src_r1[0] >> 16 | src_r1[1] << 16;
dst2[0] = src_r1[1] >> 16 | src_r1[2] << 16;
dst2[1] = src_r1[2] >> 16 | src_r1[3] << 16;
dst2[2] = src_r1[3] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 11:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[2] = src_l0[2] | src_r0[0] << 24;
dst0[3] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[0] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[1] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[2] = src_r0[3] >> 8 | src_r1[0] << 24;
dst1[3] = src_r1[0] >> 8 | src_r1[1] << 24;
dst2[0] = src_r1[1] >> 8 | src_r1[2] << 24;
dst2[1] = src_r1[2] >> 8 | src_r1[3] << 24;
dst2[2] = src_r1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 12:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_r0[0];
dst1[0] = src_r0[1];
dst1[1] = src_r0[2];
dst1[2] = src_r0[3];
dst1[3] = src_r1[0];
dst2[0] = src_r1[1];
dst2[1] = src_r1[2];
dst2[2] = src_r1[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 13:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 8;
dst1[0] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[1] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[2] = src_r0[2] >> 24 | src_r0[3] << 8;
dst1[3] = src_r0[3] >> 24 | src_r1[0] << 8;
dst2[0] = src_r1[0] >> 24 | src_r1[1] << 8;
dst2[1] = src_r1[1] >> 24 | src_r1[2] << 8;
dst2[2] = src_r1[2] >> 24 | src_r1[3] << 8;
dst2[3] = src_r1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 14:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 16;
dst1[0] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[1] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[2] = src_r0[2] >> 16 | src_r0[3] << 16;
dst1[3] = src_r0[3] >> 16 | src_r1[0] << 16;
dst2[0] = src_r1[0] >> 16 | src_r1[1] << 16;
dst2[1] = src_r1[1] >> 16 | src_r1[2] << 16;
dst2[2] = src_r1[2] >> 16 | src_r1[3] << 16;
dst2[3] = src_r1[3] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 15:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst0[3] = src_l0[3] | src_r0[0] << 24;
dst1[0] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[1] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[2] = src_r0[2] >> 8 | src_r0[3] << 24;
dst1[3] = src_r0[3] >> 8 | src_r1[0] << 24;
dst2[0] = src_r1[0] >> 8 | src_r1[1] << 24;
dst2[1] = src_r1[1] >> 8 | src_r1[2] << 24;
dst2[2] = src_r1[2] >> 8 | src_r1[3] << 24;
dst2[3] = src_r1[3] >> 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 16:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_r0[0];
dst1[1] = src_r0[1];
dst1[2] = src_r0[2];
dst1[3] = src_r0[3];
dst2[0] = src_r1[0];
dst2[1] = src_r1[1];
dst2[2] = src_r1[2];
dst2[3] = src_r1[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 17:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 8;
dst1[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst1[3] = src_r0[2] >> 24 | src_r0[3] << 8;
dst2[0] = src_r0[3] >> 24 | src_r1[0] << 8;
dst2[1] = src_r1[0] >> 24 | src_r1[1] << 8;
dst2[2] = src_r1[1] >> 24 | src_r1[2] << 8;
dst2[3] = src_r1[2] >> 24 | src_r1[3] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 18:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 16;
dst1[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst1[3] = src_r0[2] >> 16 | src_r0[3] << 16;
dst2[0] = src_r0[3] >> 16 | src_r1[0] << 16;
dst2[1] = src_r1[0] >> 16 | src_r1[1] << 16;
dst2[2] = src_r1[1] >> 16 | src_r1[2] << 16;
dst2[3] = src_r1[2] >> 16 | src_r1[3] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 19:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[0] = src_l1[0] | src_r0[0] << 24;
dst1[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst1[3] = src_r0[2] >> 8 | src_r0[3] << 24;
dst2[0] = src_r0[3] >> 8 | src_r1[0] << 24;
dst2[1] = src_r1[0] >> 8 | src_r1[1] << 24;
dst2[2] = src_r1[1] >> 8 | src_r1[2] << 24;
dst2[3] = src_r1[2] >> 8 | src_r1[3] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 20:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_r1[0];
dst1[2] = src_r0[1];
dst1[3] = src_r0[2];
dst2[0] = src_r0[3];
dst2[1] = src_r1[0];
dst2[2] = src_r1[1];
dst2[3] = src_r1[2];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 21:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 8;
dst1[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst1[3] = src_r0[1] >> 24 | src_r0[2] << 8;
dst2[0] = src_r0[2] >> 24 | src_r0[3] << 8;
dst2[1] = src_r0[3] >> 24 | src_r1[0] << 8;
dst2[2] = src_r1[0] >> 24 | src_r1[1] << 8;
dst2[3] = src_r1[1] >> 24 | src_r1[2] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 22:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 16;
dst1[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst1[3] = src_r0[1] >> 16 | src_r0[2] << 16;
dst2[0] = src_r0[2] >> 16 | src_r0[3] << 16;
dst2[1] = src_r0[3] >> 16 | src_r1[0] << 16;
dst2[2] = src_r1[0] >> 16 | src_r1[1] << 16;
dst2[3] = src_r1[1] >> 16 | src_r1[2] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 23:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[1] = src_l1[1] | src_r0[0] << 24;
dst1[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst1[3] = src_r0[1] >> 8 | src_r0[2] << 24;
dst2[0] = src_r0[2] >> 8 | src_r0[3] << 24;
dst2[1] = src_r0[3] >> 8 | src_r1[0] << 24;
dst2[2] = src_r1[0] >> 8 | src_r1[1] << 24;
dst2[3] = src_r1[1] >> 8 | src_r1[2] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 24:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_r1[0];
dst1[3] = src_r0[1];
dst2[0] = src_r0[2];
dst2[1] = src_r0[3];
dst2[2] = src_r1[0];
dst2[3] = src_r1[1];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 25:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 8;
dst1[3] = src_r0[0] >> 24 | src_r0[1] << 8;
dst2[0] = src_r0[1] >> 24 | src_r0[2] << 8;
dst2[1] = src_r0[2] >> 24 | src_r0[3] << 8;
dst2[2] = src_r0[3] >> 24 | src_r1[0] << 8;
dst2[3] = src_r1[0] >> 24 | src_r1[1] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 26:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 16;
dst1[3] = src_r0[0] >> 16 | src_r0[1] << 16;
dst2[0] = src_r0[1] >> 16 | src_r0[2] << 16;
dst2[1] = src_r0[2] >> 16 | src_r0[3] << 16;
dst2[2] = src_r0[3] >> 16 | src_r1[0] << 16;
dst2[3] = src_r1[0] >> 16 | src_r1[1] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 27:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[2] = src_l1[2] | src_r0[0] << 24;
dst1[3] = src_r0[0] >> 8 | src_r0[1] << 24;
dst2[0] = src_r0[1] >> 8 | src_r0[2] << 24;
dst2[1] = src_r0[2] >> 8 | src_r0[3] << 24;
dst2[2] = src_r0[3] >> 8 | src_r1[0] << 24;
dst2[3] = src_r1[0] >> 8 | src_r1[1] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 28:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_r1[0];
dst2[0] = src_r0[1];
dst2[1] = src_r0[2];
dst2[2] = src_r0[3];
dst2[3] = src_r1[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 29:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 8;
dst2[0] = src_r0[0] >> 24 | src_r0[1] << 8;
dst2[1] = src_r0[1] >> 24 | src_r0[2] << 8;
dst2[2] = src_r0[2] >> 24 | src_r0[3] << 8;
dst2[3] = src_r0[3] >> 24 | src_r1[0] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 30:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 16;
dst2[0] = src_r0[0] >> 16 | src_r0[1] << 16;
dst2[1] = src_r0[1] >> 16 | src_r0[2] << 16;
dst2[2] = src_r0[2] >> 16 | src_r0[3] << 16;
dst2[3] = src_r0[3] >> 16 | src_r1[0] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 31:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst1[3] = src_l1[3] | src_r0[0] << 24;
dst2[0] = src_r0[0] >> 8 | src_r0[1] << 24;
dst2[1] = src_r0[1] >> 8 | src_r0[2] << 24;
dst2[2] = src_r0[2] >> 8 | src_r0[3] << 24;
dst2[3] = src_r0[3] >> 8 | src_r1[0] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 32:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[0] = src_r0[0];
dst2[1] = src_r0[1];
dst2[2] = src_r0[2];
dst2[3] = src_r0[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 33:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[0] = src_l2[0] | src_r0[0] << 8;
dst2[1] = src_r0[0] >> 24 | src_r0[1] << 8;
dst2[2] = src_r0[1] >> 24 | src_r0[2] << 8;
dst2[3] = src_r0[2] >> 24 | src_r0[3] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 34:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[0] = src_l2[0] | src_r0[0] << 16;
dst2[1] = src_r0[0] >> 16 | src_r0[1] << 16;
dst2[2] = src_r0[1] >> 16 | src_r0[2] << 16;
dst2[3] = src_r0[2] >> 16 | src_r0[3] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 35:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[0] = src_l2[0] | src_r0[0] << 24;
dst2[1] = src_r0[0] >> 8 | src_r0[1] << 24;
dst2[2] = src_r0[1] >> 8 | src_r0[2] << 24;
dst2[3] = src_r0[2] >> 8 | src_r0[3] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 36:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[1] = src_r0[0];
dst2[2] = src_r0[1];
dst2[3] = src_r0[2];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 37:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[1] = src_l2[1] | src_r0[0] << 8;
dst2[2] = src_r0[0] >> 24 | src_r0[1] << 8;
dst2[3] = src_r0[1] >> 24 | src_r0[2] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 38:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[1] = src_l2[1] | src_r0[0] << 16;
dst2[2] = src_r0[0] >> 16 | src_r0[1] << 16;
dst2[3] = src_r0[1] >> 16 | src_r0[2] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 39:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[1] = src_l2[1] | src_r0[0] << 24;
dst2[2] = src_r0[0] >> 8 | src_r0[1] << 24;
dst2[3] = src_r0[1] >> 8 | src_r0[2] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 40:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[2] = src_r0[0];
dst2[3] = src_r0[1];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
case 41:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[2] = src_l2[2] | src_r0[0] << 8;
dst2[3] = src_r0[0] >> 24 | src_r0[1] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 42:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[2] = src_l2[2] | src_r0[0] << 16;
dst2[3] = src_r0[0] >> 16 | src_r0[1] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 43:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[2] = src_l2[2] | src_r0[0] << 24;
dst2[3] = src_r0[0] >> 8 | src_r0[1] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 44:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[3] = src_r0[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 45:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[3] = src_l2[3] | src_r0[0] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 46:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[3] = src_l2[3] | src_r0[0] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
case 47:
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
dst2[3] = src_l2[3] | src_r0[0] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
}
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
// before: memcat16_9
static void memcat_c15_w4x4_a3x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 append0[4], const u32 append1[4], const u32 append2[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = append0[0];
w0[1] = append0[1];
w0[2] = append0[2];
w0[3] = append0[3];
w1[0] = append1[0];
w1[1] = append1[1];
w1[2] = append1[2];
w1[3] = append1[3];
w2[0] = append2[0];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 1:
w0[0] = w0[0] | append0[0] << 8;
w0[1] = append0[0] >> 24 | append0[1] << 8;
w0[2] = append0[1] >> 24 | append0[2] << 8;
w0[3] = append0[2] >> 24 | append0[3] << 8;
w1[0] = append0[3] >> 24 | append1[0] << 8;
w1[1] = append1[0] >> 24 | append1[1] << 8;
w1[2] = append1[1] >> 24 | append1[2] << 8;
w1[3] = append1[2] >> 24 | append1[3] << 8;
w2[0] = append1[3] >> 24 | append2[0] << 8;
w2[1] = append2[0] >> 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
w0[0] = w0[0] | append0[0] << 16;
w0[1] = append0[0] >> 16 | append0[1] << 16;
w0[2] = append0[1] >> 16 | append0[2] << 16;
w0[3] = append0[2] >> 16 | append0[3] << 16;
w1[0] = append0[3] >> 16 | append1[0] << 16;
w1[1] = append1[0] >> 16 | append1[1] << 16;
w1[2] = append1[1] >> 16 | append1[2] << 16;
w1[3] = append1[2] >> 16 | append1[3] << 16;
w2[0] = append1[3] >> 16 | append2[0] << 16;
w2[1] = append2[0] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
w0[0] = w0[0] | append0[0] << 24;
w0[1] = append0[0] >> 8 | append0[1] << 24;
w0[2] = append0[1] >> 8 | append0[2] << 24;
w0[3] = append0[2] >> 8 | append0[3] << 24;
w1[0] = append0[3] >> 8 | append1[0] << 24;
w1[1] = append1[0] >> 8 | append1[1] << 24;
w1[2] = append1[1] >> 8 | append1[2] << 24;
w1[3] = append1[2] >> 8 | append1[3] << 24;
w2[0] = append1[3] >> 8 | append2[0] << 24;
w2[1] = append2[0] >> 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
w0[1] = append0[0];
w0[2] = append0[1];
w0[3] = append0[2];
w1[0] = append0[3];
w1[1] = append1[0];
w1[2] = append1[1];
w1[3] = append1[2];
w2[0] = append1[3];
w2[1] = append2[0];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
w0[1] = w0[1] | append0[0] << 8;
w0[2] = append0[0] >> 24 | append0[1] << 8;
w0[3] = append0[1] >> 24 | append0[2] << 8;
w1[0] = append0[2] >> 24 | append0[3] << 8;
w1[1] = append0[3] >> 24 | append1[0] << 8;
w1[2] = append1[0] >> 24 | append1[1] << 8;
w1[3] = append1[1] >> 24 | append1[2] << 8;
w2[0] = append1[2] >> 24 | append1[3] << 8;
w2[1] = append1[3] >> 24 | append2[0] << 8;
w2[2] = append2[0] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 6:
w0[1] = w0[1] | append0[0] << 16;
w0[2] = append0[0] >> 16 | append0[1] << 16;
w0[3] = append0[1] >> 16 | append0[2] << 16;
w1[0] = append0[2] >> 16 | append0[3] << 16;
w1[1] = append0[3] >> 16 | append1[0] << 16;
w1[2] = append1[0] >> 16 | append1[1] << 16;
w1[3] = append1[1] >> 16 | append1[2] << 16;
w2[0] = append1[2] >> 16 | append1[3] << 16;
w2[1] = append1[3] >> 16 | append2[0] << 16;
w2[2] = append2[0] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 7:
w0[1] = w0[1] | append0[0] << 24;
w0[2] = append0[0] >> 8 | append0[1] << 24;
w0[3] = append0[1] >> 8 | append0[2] << 24;
w1[0] = append0[2] >> 8 | append0[3] << 24;
w1[1] = append0[3] >> 8 | append1[0] << 24;
w1[2] = append1[0] >> 8 | append1[1] << 24;
w1[3] = append1[1] >> 8 | append1[2] << 24;
w2[0] = append1[2] >> 8 | append1[3] << 24;
w2[1] = append1[3] >> 8 | append2[0] << 24;
w2[2] = append2[0] >> 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 8:
w0[2] = append0[0];
w0[3] = append0[1];
w1[0] = append0[2];
w1[1] = append0[3];
w1[2] = append1[0];
w1[3] = append1[1];
w2[0] = append1[2];
w2[1] = append1[3];
w2[2] = append2[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 9:
w0[2] = w0[2] | append0[0] << 8;
w0[3] = append0[0] >> 24 | append0[1] << 8;
w1[0] = append0[1] >> 24 | append0[2] << 8;
w1[1] = append0[2] >> 24 | append0[3] << 8;
w1[2] = append0[3] >> 24 | append1[0] << 8;
w1[3] = append1[0] >> 24 | append1[1] << 8;
w2[0] = append1[1] >> 24 | append1[2] << 8;
w2[1] = append1[2] >> 24 | append1[3] << 8;
w2[2] = append1[3] >> 24 | append2[0] << 8;
w2[3] = append2[0] >> 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 10:
w0[2] = w0[2] | append0[0] << 16;
w0[3] = append0[0] >> 16 | append0[1] << 16;
w1[0] = append0[1] >> 16 | append0[2] << 16;
w1[1] = append0[2] >> 16 | append0[3] << 16;
w1[2] = append0[3] >> 16 | append1[0] << 16;
w1[3] = append1[0] >> 16 | append1[1] << 16;
w2[0] = append1[1] >> 16 | append1[2] << 16;
w2[1] = append1[2] >> 16 | append1[3] << 16;
w2[2] = append1[3] >> 16 | append2[0] << 16;
w2[3] = append2[0] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 11:
w0[2] = w0[2] | append0[0] << 24;
w0[3] = append0[0] >> 8 | append0[1] << 24;
w1[0] = append0[1] >> 8 | append0[2] << 24;
w1[1] = append0[2] >> 8 | append0[3] << 24;
w1[2] = append0[3] >> 8 | append1[0] << 24;
w1[3] = append1[0] >> 8 | append1[1] << 24;
w2[0] = append1[1] >> 8 | append1[2] << 24;
w2[1] = append1[2] >> 8 | append1[3] << 24;
w2[2] = append1[3] >> 8 | append2[0] << 24;
w2[3] = append2[0] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 12:
w0[3] = append0[0];
w1[0] = append0[1];
w1[1] = append0[2];
w1[2] = append0[3];
w1[3] = append1[0];
w2[0] = append1[1];
w2[1] = append1[2];
w2[2] = append1[3];
w2[3] = append2[0];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 13:
w0[3] = w0[3] | append0[0] << 8;
w1[0] = append0[0] >> 24 | append0[1] << 8;
w1[1] = append0[1] >> 24 | append0[2] << 8;
w1[2] = append0[2] >> 24 | append0[3] << 8;
w1[3] = append0[3] >> 24 | append1[0] << 8;
w2[0] = append1[0] >> 24 | append1[1] << 8;
w2[1] = append1[1] >> 24 | append1[2] << 8;
w2[2] = append1[2] >> 24 | append1[3] << 8;
w2[3] = append1[3] >> 24 | append2[0] << 8;
w3[0] = append2[0] >> 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 14:
w0[3] = w0[3] | append0[0] << 16;
w1[0] = append0[0] >> 16 | append0[1] << 16;
w1[1] = append0[1] >> 16 | append0[2] << 16;
w1[2] = append0[2] >> 16 | append0[3] << 16;
w1[3] = append0[3] >> 16 | append1[0] << 16;
w2[0] = append1[0] >> 16 | append1[1] << 16;
w2[1] = append1[1] >> 16 | append1[2] << 16;
w2[2] = append1[2] >> 16 | append1[3] << 16;
w2[3] = append1[3] >> 16 | append2[0] << 16;
w3[0] = append2[0] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 15:
w0[3] = w0[3] | append0[0] << 24;
w1[0] = append0[0] >> 8 | append0[1] << 24;
w1[1] = append0[1] >> 8 | append0[2] << 24;
w1[2] = append0[2] >> 8 | append0[3] << 24;
w1[3] = append0[3] >> 8 | append1[0] << 24;
w2[0] = append1[0] >> 8 | append1[1] << 24;
w2[1] = append1[1] >> 8 | append1[2] << 24;
w2[2] = append1[2] >> 8 | append1[3] << 24;
w2[3] = append1[3] >> 8 | append2[0] << 24;
w3[0] = append2[0] >> 8;
break;
}
}
// before: memcat32_8
static void memcat_c32_w4x4_a2x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 append0[4], const u32 append1[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = append0[0];
w0[1] = append0[1];
w0[2] = append0[2];
w0[3] = append0[3];
w1[0] = append1[0];
w1[1] = append1[1];
w1[2] = append1[2];
w1[3] = append1[3];
break;
case 1:
w0[0] = w0[0] | append0[0] << 8;
w0[1] = append0[0] >> 24 | append0[1] << 8;
w0[2] = append0[1] >> 24 | append0[2] << 8;
w0[3] = append0[2] >> 24 | append0[3] << 8;
w1[0] = append0[3] >> 24 | append1[0] << 8;
w1[1] = append1[0] >> 24 | append1[1] << 8;
w1[2] = append1[1] >> 24 | append1[2] << 8;
w1[3] = append1[2] >> 24 | append1[3] << 8;
w2[0] = append1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
w0[0] = w0[0] | append0[0] << 16;
w0[1] = append0[0] >> 16 | append0[1] << 16;
w0[2] = append0[1] >> 16 | append0[2] << 16;
w0[3] = append0[2] >> 16 | append0[3] << 16;
w1[0] = append0[3] >> 16 | append1[0] << 16;
w1[1] = append1[0] >> 16 | append1[1] << 16;
w1[2] = append1[1] >> 16 | append1[2] << 16;
w1[3] = append1[2] >> 16 | append1[3] << 16;
w2[0] = append1[3] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
w0[0] = w0[0] | append0[0] << 24;
w0[1] = append0[0] >> 8 | append0[1] << 24;
w0[2] = append0[1] >> 8 | append0[2] << 24;
w0[3] = append0[2] >> 8 | append0[3] << 24;
w1[0] = append0[3] >> 8 | append1[0] << 24;
w1[1] = append1[0] >> 8 | append1[1] << 24;
w1[2] = append1[1] >> 8 | append1[2] << 24;
w1[3] = append1[2] >> 8 | append1[3] << 24;
w2[0] = append1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
w0[1] = append0[0];
w0[2] = append0[1];
w0[3] = append0[2];
w1[0] = append0[3];
w1[1] = append1[0];
w1[2] = append1[1];
w1[3] = append1[2];
w2[0] = append1[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
w0[1] = w0[1] | append0[0] << 8;
w0[2] = append0[0] >> 24 | append0[1] << 8;
w0[3] = append0[1] >> 24 | append0[2] << 8;
w1[0] = append0[2] >> 24 | append0[3] << 8;
w1[1] = append0[3] >> 24 | append1[0] << 8;
w1[2] = append1[0] >> 24 | append1[1] << 8;
w1[3] = append1[1] >> 24 | append1[2] << 8;
w2[0] = append1[2] >> 24 | append1[3] << 8;
w2[1] = append1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 6:
w0[1] = w0[1] | append0[0] << 16;
w0[2] = append0[0] >> 16 | append0[1] << 16;
w0[3] = append0[1] >> 16 | append0[2] << 16;
w1[0] = append0[2] >> 16 | append0[3] << 16;
w1[1] = append0[3] >> 16 | append1[0] << 16;
w1[2] = append1[0] >> 16 | append1[1] << 16;
w1[3] = append1[1] >> 16 | append1[2] << 16;
w2[0] = append1[2] >> 16 | append1[3] << 16;
w2[1] = append1[3] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 7:
w0[1] = w0[1] | append0[0] << 24;
w0[2] = append0[0] >> 8 | append0[1] << 24;
w0[3] = append0[1] >> 8 | append0[2] << 24;
w1[0] = append0[2] >> 8 | append0[3] << 24;
w1[1] = append0[3] >> 8 | append1[0] << 24;
w1[2] = append1[0] >> 8 | append1[1] << 24;
w1[3] = append1[1] >> 8 | append1[2] << 24;
w2[0] = append1[2] >> 8 | append1[3] << 24;
w2[1] = append1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 8:
w0[2] = append0[0];
w0[3] = append0[1];
w1[0] = append0[2];
w1[1] = append0[3];
w1[2] = append1[0];
w1[3] = append1[1];
w2[0] = append1[2];
w2[1] = append1[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 9:
w0[2] = w0[2] | append0[0] << 8;
w0[3] = append0[0] >> 24 | append0[1] << 8;
w1[0] = append0[1] >> 24 | append0[2] << 8;
w1[1] = append0[2] >> 24 | append0[3] << 8;
w1[2] = append0[3] >> 24 | append1[0] << 8;
w1[3] = append1[0] >> 24 | append1[1] << 8;
w2[0] = append1[1] >> 24 | append1[2] << 8;
w2[1] = append1[2] >> 24 | append1[3] << 8;
w2[2] = append1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 10:
w0[2] = w0[2] | append0[0] << 16;
w0[3] = append0[0] >> 16 | append0[1] << 16;
w1[0] = append0[1] >> 16 | append0[2] << 16;
w1[1] = append0[2] >> 16 | append0[3] << 16;
w1[2] = append0[3] >> 16 | append1[0] << 16;
w1[3] = append1[0] >> 16 | append1[1] << 16;
w2[0] = append1[1] >> 16 | append1[2] << 16;
w2[1] = append1[2] >> 16 | append1[3] << 16;
w2[2] = append1[3] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 11:
w0[2] = w0[2] | append0[0] << 24;
w0[3] = append0[0] >> 8 | append0[1] << 24;
w1[0] = append0[1] >> 8 | append0[2] << 24;
w1[1] = append0[2] >> 8 | append0[3] << 24;
w1[2] = append0[3] >> 8 | append1[0] << 24;
w1[3] = append1[0] >> 8 | append1[1] << 24;
w2[0] = append1[1] >> 8 | append1[2] << 24;
w2[1] = append1[2] >> 8 | append1[3] << 24;
w2[2] = append1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 12:
w0[3] = append0[0];
w1[0] = append0[1];
w1[1] = append0[2];
w1[2] = append0[3];
w1[3] = append1[0];
w2[0] = append1[1];
w2[1] = append1[2];
w2[2] = append1[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 13:
w0[3] = w0[3] | append0[0] << 8;
w1[0] = append0[0] >> 24 | append0[1] << 8;
w1[1] = append0[1] >> 24 | append0[2] << 8;
w1[2] = append0[2] >> 24 | append0[3] << 8;
w1[3] = append0[3] >> 24 | append1[0] << 8;
w2[0] = append1[0] >> 24 | append1[1] << 8;
w2[1] = append1[1] >> 24 | append1[2] << 8;
w2[2] = append1[2] >> 24 | append1[3] << 8;
w2[3] = append1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 14:
w0[3] = w0[3] | append0[0] << 16;
w1[0] = append0[0] >> 16 | append0[1] << 16;
w1[1] = append0[1] >> 16 | append0[2] << 16;
w1[2] = append0[2] >> 16 | append0[3] << 16;
w1[3] = append0[3] >> 16 | append1[0] << 16;
w2[0] = append1[0] >> 16 | append1[1] << 16;
w2[1] = append1[1] >> 16 | append1[2] << 16;
w2[2] = append1[2] >> 16 | append1[3] << 16;
w2[3] = append1[3] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 15:
w0[3] = w0[3] | append0[0] << 24;
w1[0] = append0[0] >> 8 | append0[1] << 24;
w1[1] = append0[1] >> 8 | append0[2] << 24;
w1[2] = append0[2] >> 8 | append0[3] << 24;
w1[3] = append0[3] >> 8 | append1[0] << 24;
w2[0] = append1[0] >> 8 | append1[1] << 24;
w2[1] = append1[1] >> 8 | append1[2] << 24;
w2[2] = append1[2] >> 8 | append1[3] << 24;
w2[3] = append1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 16:
w1[0] = append0[0];
w1[1] = append0[1];
w1[2] = append0[2];
w1[3] = append0[3];
w2[0] = append1[0];
w2[1] = append1[1];
w2[2] = append1[2];
w2[3] = append1[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 17:
w1[0] = w1[0] | append0[0] << 8;
w1[1] = append0[0] >> 24 | append0[1] << 8;
w1[2] = append0[1] >> 24 | append0[2] << 8;
w1[3] = append0[2] >> 24 | append0[3] << 8;
w2[0] = append0[3] >> 24 | append1[0] << 8;
w2[1] = append1[0] >> 24 | append1[1] << 8;
w2[2] = append1[1] >> 24 | append1[2] << 8;
w2[3] = append1[2] >> 24 | append1[3] << 8;
w3[0] = append1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 18:
w1[0] = w1[0] | append0[0] << 16;
w1[1] = append0[0] >> 16 | append0[1] << 16;
w1[2] = append0[1] >> 16 | append0[2] << 16;
w1[3] = append0[2] >> 16 | append0[3] << 16;
w2[0] = append0[3] >> 16 | append1[0] << 16;
w2[1] = append1[0] >> 16 | append1[1] << 16;
w2[2] = append1[1] >> 16 | append1[2] << 16;
w2[3] = append1[2] >> 16 | append1[3] << 16;
w3[0] = append1[3] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 19:
w1[0] = w1[0] | append0[0] << 24;
w1[1] = append0[0] >> 8 | append0[1] << 24;
w1[2] = append0[1] >> 8 | append0[2] << 24;
w1[3] = append0[2] >> 8 | append0[3] << 24;
w2[0] = append0[3] >> 8 | append1[0] << 24;
w2[1] = append1[0] >> 8 | append1[1] << 24;
w2[2] = append1[1] >> 8 | append1[2] << 24;
w2[3] = append1[2] >> 8 | append1[3] << 24;
w3[0] = append1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 20:
w1[1] = append0[0];
w1[2] = append0[1];
w1[3] = append0[2];
w2[0] = append0[3];
w2[1] = append1[0];
w2[2] = append1[1];
w2[3] = append1[2];
w3[0] = append1[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 21:
w1[1] = w1[1] | append0[0] << 8;
w1[2] = append0[0] >> 24 | append0[1] << 8;
w1[3] = append0[1] >> 24 | append0[2] << 8;
w2[0] = append0[2] >> 24 | append0[3] << 8;
w2[1] = append0[3] >> 24 | append1[0] << 8;
w2[2] = append1[0] >> 24 | append1[1] << 8;
w2[3] = append1[1] >> 24 | append1[2] << 8;
w3[0] = append1[2] >> 24 | append1[3] << 8;
w3[1] = append1[3] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 22:
w1[1] = w1[1] | append0[0] << 16;
w1[2] = append0[0] >> 16 | append0[1] << 16;
w1[3] = append0[1] >> 16 | append0[2] << 16;
w2[0] = append0[2] >> 16 | append0[3] << 16;
w2[1] = append0[3] >> 16 | append1[0] << 16;
w2[2] = append1[0] >> 16 | append1[1] << 16;
w2[3] = append1[1] >> 16 | append1[2] << 16;
w3[0] = append1[2] >> 16 | append1[3] << 16;
w3[1] = append1[3] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 23:
w1[1] = w1[1] | append0[0] << 24;
w1[2] = append0[0] >> 8 | append0[1] << 24;
w1[3] = append0[1] >> 8 | append0[2] << 24;
w2[0] = append0[2] >> 8 | append0[3] << 24;
w2[1] = append0[3] >> 8 | append1[0] << 24;
w2[2] = append1[0] >> 8 | append1[1] << 24;
w2[3] = append1[1] >> 8 | append1[2] << 24;
w3[0] = append1[2] >> 8 | append1[3] << 24;
w3[1] = append1[3] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 24:
w1[2] = append0[0];
w1[3] = append0[1];
w2[0] = append0[2];
w2[1] = append0[3];
w2[2] = append1[0];
w2[3] = append1[1];
w3[0] = append1[2];
w3[1] = append1[3];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 25:
w1[2] = w1[2] | append0[0] << 8;
w1[3] = append0[0] >> 24 | append0[1] << 8;
w2[0] = append0[1] >> 24 | append0[2] << 8;
w2[1] = append0[2] >> 24 | append0[3] << 8;
w2[2] = append0[3] >> 24 | append1[0] << 8;
w2[3] = append1[0] >> 24 | append1[1] << 8;
w3[0] = append1[1] >> 24 | append1[2] << 8;
w3[1] = append1[2] >> 24 | append1[3] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 26:
w1[2] = w1[2] | append0[0] << 16;
w1[3] = append0[0] >> 16 | append0[1] << 16;
w2[0] = append0[1] >> 16 | append0[2] << 16;
w2[1] = append0[2] >> 16 | append0[3] << 16;
w2[2] = append0[3] >> 16 | append1[0] << 16;
w2[3] = append1[0] >> 16 | append1[1] << 16;
w3[0] = append1[1] >> 16 | append1[2] << 16;
w3[1] = append1[2] >> 16 | append1[3] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 27:
w1[2] = w1[2] | append0[0] << 24;
w1[3] = append0[0] >> 8 | append0[1] << 24;
w2[0] = append0[1] >> 8 | append0[2] << 24;
w2[1] = append0[2] >> 8 | append0[3] << 24;
w2[2] = append0[3] >> 8 | append1[0] << 24;
w2[3] = append1[0] >> 8 | append1[1] << 24;
w3[0] = append1[1] >> 8 | append1[2] << 24;
w3[1] = append1[2] >> 8 | append1[3] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 28:
w1[3] = append0[0];
w2[0] = append0[1];
w2[1] = append0[2];
w2[2] = append0[3];
w2[3] = append1[0];
w3[0] = append1[1];
w3[1] = append1[2];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 29:
w1[3] = w1[3] | append0[0] << 8;
w2[0] = append0[0] >> 24 | append0[1] << 8;
w2[1] = append0[1] >> 24 | append0[2] << 8;
w2[2] = append0[2] >> 24 | append0[3] << 8;
w2[3] = append0[3] >> 24 | append1[0] << 8;
w3[0] = append1[0] >> 24 | append1[1] << 8;
w3[1] = append1[1] >> 24 | append1[2] << 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 30:
w1[3] = w1[3] | append0[0] << 16;
w2[0] = append0[0] >> 16 | append0[1] << 16;
w2[1] = append0[1] >> 16 | append0[2] << 16;
w2[2] = append0[2] >> 16 | append0[3] << 16;
w2[3] = append0[3] >> 16 | append1[0] << 16;
w3[0] = append1[0] >> 16 | append1[1] << 16;
w3[1] = append1[1] >> 16 | append1[2] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 31:
w1[3] = w1[3] | append0[0] << 24;
w2[0] = append0[0] >> 8 | append0[1] << 24;
w2[1] = append0[1] >> 8 | append0[2] << 24;
w2[2] = append0[2] >> 8 | append0[3] << 24;
w2[3] = append0[3] >> 8 | append1[0] << 24;
w3[0] = append1[0] >> 8 | append1[1] << 24;
w3[1] = append1[1] >> 8 | append1[2] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 32:
w2[0] = append0[0];
w2[1] = append0[1];
w2[2] = append0[2];
w2[3] = append0[3];
w3[0] = append1[0];
w3[1] = append1[1];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
}
}
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
// before: memcat32_9
static void memcat_c32_w4x4_a3x4 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 append0[4], const u32 append1[4], const u32 append2[4], const u32 offset)
{
switch (offset)
{
case 0:
w0[0] = append0[0];
w0[1] = append0[1];
w0[2] = append0[2];
w0[3] = append0[3];
w1[0] = append1[0];
w1[1] = append1[1];
w1[2] = append1[2];
w1[3] = append1[3];
w2[0] = append2[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 1:
w0[0] = w0[0] | append0[0] << 8;
w0[1] = append0[0] >> 24 | append0[1] << 8;
w0[2] = append0[1] >> 24 | append0[2] << 8;
w0[3] = append0[2] >> 24 | append0[3] << 8;
w1[0] = append0[3] >> 24 | append1[0] << 8;
w1[1] = append1[0] >> 24 | append1[1] << 8;
w1[2] = append1[1] >> 24 | append1[2] << 8;
w1[3] = append1[2] >> 24 | append1[3] << 8;
w2[0] = append1[3] >> 24 | append2[0] << 8;
w2[1] = append2[0] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 2:
w0[0] = w0[0] | append0[0] << 16;
w0[1] = append0[0] >> 16 | append0[1] << 16;
w0[2] = append0[1] >> 16 | append0[2] << 16;
w0[3] = append0[2] >> 16 | append0[3] << 16;
w1[0] = append0[3] >> 16 | append1[0] << 16;
w1[1] = append1[0] >> 16 | append1[1] << 16;
w1[2] = append1[1] >> 16 | append1[2] << 16;
w1[3] = append1[2] >> 16 | append1[3] << 16;
w2[0] = append1[3] >> 16 | append2[0] << 16;
w2[1] = append2[0] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 3:
w0[0] = w0[0] | append0[0] << 24;
w0[1] = append0[0] >> 8 | append0[1] << 24;
w0[2] = append0[1] >> 8 | append0[2] << 24;
w0[3] = append0[2] >> 8 | append0[3] << 24;
w1[0] = append0[3] >> 8 | append1[0] << 24;
w1[1] = append1[0] >> 8 | append1[1] << 24;
w1[2] = append1[1] >> 8 | append1[2] << 24;
w1[3] = append1[2] >> 8 | append1[3] << 24;
w2[0] = append1[3] >> 8 | append2[0] << 24;
w2[1] = append2[0] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 4:
w0[1] = append0[0];
w0[2] = append0[1];
w0[3] = append0[2];
w1[0] = append0[3];
w1[1] = append1[0];
w1[2] = append1[1];
w1[3] = append1[2];
w2[0] = append1[3];
w2[1] = append2[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 5:
w0[1] = w0[1] | append0[0] << 8;
w0[2] = append0[0] >> 24 | append0[1] << 8;
w0[3] = append0[1] >> 24 | append0[2] << 8;
w1[0] = append0[2] >> 24 | append0[3] << 8;
w1[1] = append0[3] >> 24 | append1[0] << 8;
w1[2] = append1[0] >> 24 | append1[1] << 8;
w1[3] = append1[1] >> 24 | append1[2] << 8;
w2[0] = append1[2] >> 24 | append1[3] << 8;
w2[1] = append1[3] >> 24 | append2[0] << 8;
w2[2] = append2[0] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 6:
w0[1] = w0[1] | append0[0] << 16;
w0[2] = append0[0] >> 16 | append0[1] << 16;
w0[3] = append0[1] >> 16 | append0[2] << 16;
w1[0] = append0[2] >> 16 | append0[3] << 16;
w1[1] = append0[3] >> 16 | append1[0] << 16;
w1[2] = append1[0] >> 16 | append1[1] << 16;
w1[3] = append1[1] >> 16 | append1[2] << 16;
w2[0] = append1[2] >> 16 | append1[3] << 16;
w2[1] = append1[3] >> 16 | append2[0] << 16;
w2[2] = append2[0] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 7:
w0[1] = w0[1] | append0[0] << 24;
w0[2] = append0[0] >> 8 | append0[1] << 24;
w0[3] = append0[1] >> 8 | append0[2] << 24;
w1[0] = append0[2] >> 8 | append0[3] << 24;
w1[1] = append0[3] >> 8 | append1[0] << 24;
w1[2] = append1[0] >> 8 | append1[1] << 24;
w1[3] = append1[1] >> 8 | append1[2] << 24;
w2[0] = append1[2] >> 8 | append1[3] << 24;
w2[1] = append1[3] >> 8 | append2[0] << 24;
w2[2] = append2[0] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 8:
w0[2] = append0[0];
w0[3] = append0[1];
w1[0] = append0[2];
w1[1] = append0[3];
w1[2] = append1[0];
w1[3] = append1[1];
w2[0] = append1[2];
w2[1] = append1[3];
w2[2] = append2[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 9:
w0[2] = w0[2] | append0[0] << 8;
w0[3] = append0[0] >> 24 | append0[1] << 8;
w1[0] = append0[1] >> 24 | append0[2] << 8;
w1[1] = append0[2] >> 24 | append0[3] << 8;
w1[2] = append0[3] >> 24 | append1[0] << 8;
w1[3] = append1[0] >> 24 | append1[1] << 8;
w2[0] = append1[1] >> 24 | append1[2] << 8;
w2[1] = append1[2] >> 24 | append1[3] << 8;
w2[2] = append1[3] >> 24 | append2[0] << 8;
w2[3] = append2[0] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 10:
w0[2] = w0[2] | append0[0] << 16;
w0[3] = append0[0] >> 16 | append0[1] << 16;
w1[0] = append0[1] >> 16 | append0[2] << 16;
w1[1] = append0[2] >> 16 | append0[3] << 16;
w1[2] = append0[3] >> 16 | append1[0] << 16;
w1[3] = append1[0] >> 16 | append1[1] << 16;
w2[0] = append1[1] >> 16 | append1[2] << 16;
w2[1] = append1[2] >> 16 | append1[3] << 16;
w2[2] = append1[3] >> 16 | append2[0] << 16;
w2[3] = append2[0] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 11:
w0[2] = w0[2] | append0[0] << 24;
w0[3] = append0[0] >> 8 | append0[1] << 24;
w1[0] = append0[1] >> 8 | append0[2] << 24;
w1[1] = append0[2] >> 8 | append0[3] << 24;
w1[2] = append0[3] >> 8 | append1[0] << 24;
w1[3] = append1[0] >> 8 | append1[1] << 24;
w2[0] = append1[1] >> 8 | append1[2] << 24;
w2[1] = append1[2] >> 8 | append1[3] << 24;
w2[2] = append1[3] >> 8 | append2[0] << 24;
w2[3] = append2[0] >> 8;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 12:
w0[3] = append0[0];
w1[0] = append0[1];
w1[1] = append0[2];
w1[2] = append0[3];
w1[3] = append1[0];
w2[0] = append1[1];
w2[1] = append1[2];
w2[2] = append1[3];
w2[3] = append2[0];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 13:
w0[3] = w0[3] | append0[0] << 8;
w1[0] = append0[0] >> 24 | append0[1] << 8;
w1[1] = append0[1] >> 24 | append0[2] << 8;
w1[2] = append0[2] >> 24 | append0[3] << 8;
w1[3] = append0[3] >> 24 | append1[0] << 8;
w2[0] = append1[0] >> 24 | append1[1] << 8;
w2[1] = append1[1] >> 24 | append1[2] << 8;
w2[2] = append1[2] >> 24 | append1[3] << 8;
w2[3] = append1[3] >> 24 | append2[0] << 8;
w3[0] = append2[0] >> 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 14:
w0[3] = w0[3] | append0[0] << 16;
w1[0] = append0[0] >> 16 | append0[1] << 16;
w1[1] = append0[1] >> 16 | append0[2] << 16;
w1[2] = append0[2] >> 16 | append0[3] << 16;
w1[3] = append0[3] >> 16 | append1[0] << 16;
w2[0] = append1[0] >> 16 | append1[1] << 16;
w2[1] = append1[1] >> 16 | append1[2] << 16;
w2[2] = append1[2] >> 16 | append1[3] << 16;
w2[3] = append1[3] >> 16 | append2[0] << 16;
w3[0] = append2[0] >> 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 15:
w0[3] = w0[3] | append0[0] << 24;
w1[0] = append0[0] >> 8 | append0[1] << 24;
w1[1] = append0[1] >> 8 | append0[2] << 24;
w1[2] = append0[2] >> 8 | append0[3] << 24;
w1[3] = append0[3] >> 8 | append1[0] << 24;
w2[0] = append1[0] >> 8 | append1[1] << 24;
w2[1] = append1[1] >> 8 | append1[2] << 24;
w2[2] = append1[2] >> 8 | append1[3] << 24;
w2[3] = append1[3] >> 8 | append2[0] << 24;
w3[0] = append2[0] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 16:
w1[0] = append0[0];
w1[1] = append0[1];
w1[2] = append0[2];
w1[3] = append0[3];
w2[0] = append1[0];
w2[1] = append1[1];
w2[2] = append1[2];
w2[3] = append1[3];
w3[0] = append2[0];
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 17:
w1[0] = w1[0] | append0[0] << 8;
w1[1] = append0[0] >> 24 | append0[1] << 8;
w1[2] = append0[1] >> 24 | append0[2] << 8;
w1[3] = append0[2] >> 24 | append0[3] << 8;
w2[0] = append0[3] >> 24 | append1[0] << 8;
w2[1] = append1[0] >> 24 | append1[1] << 8;
w2[2] = append1[1] >> 24 | append1[2] << 8;
w2[3] = append1[2] >> 24 | append1[3] << 8;
w3[0] = append1[3] >> 24 | append2[0] << 8;
w3[1] = append2[0] >> 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 18:
w1[0] = w1[0] | append0[0] << 16;
w1[1] = append0[0] >> 16 | append0[1] << 16;
w1[2] = append0[1] >> 16 | append0[2] << 16;
w1[3] = append0[2] >> 16 | append0[3] << 16;
w2[0] = append0[3] >> 16 | append1[0] << 16;
w2[1] = append1[0] >> 16 | append1[1] << 16;
w2[2] = append1[1] >> 16 | append1[2] << 16;
w2[3] = append1[2] >> 16 | append1[3] << 16;
w3[0] = append1[3] >> 16 | append2[0] << 16;
w3[1] = append2[0] >> 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 19:
w1[0] = w1[0] | append0[0] << 24;
w1[1] = append0[0] >> 8 | append0[1] << 24;
w1[2] = append0[1] >> 8 | append0[2] << 24;
w1[3] = append0[2] >> 8 | append0[3] << 24;
w2[0] = append0[3] >> 8 | append1[0] << 24;
w2[1] = append1[0] >> 8 | append1[1] << 24;
w2[2] = append1[1] >> 8 | append1[2] << 24;
w2[3] = append1[2] >> 8 | append1[3] << 24;
w3[0] = append1[3] >> 8 | append2[0] << 24;
w3[1] = append2[0] >> 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 20:
w1[1] = append0[0];
w1[2] = append0[1];
w1[3] = append0[2];
w2[0] = append0[3];
w2[1] = append1[0];
w2[2] = append1[1];
w2[3] = append1[2];
w3[0] = append1[3];
w3[1] = append2[0];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 21:
w1[1] = w1[1] | append0[0] << 8;
w1[2] = append0[0] >> 24 | append0[1] << 8;
w1[3] = append0[1] >> 24 | append0[2] << 8;
w2[0] = append0[2] >> 24 | append0[3] << 8;
w2[1] = append0[3] >> 24 | append1[0] << 8;
w2[2] = append1[0] >> 24 | append1[1] << 8;
w2[3] = append1[1] >> 24 | append1[2] << 8;
w3[0] = append1[2] >> 24 | append1[3] << 8;
w3[1] = append1[3] >> 24 | append2[0] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 22:
w1[1] = w1[1] | append0[0] << 16;
w1[2] = append0[0] >> 16 | append0[1] << 16;
w1[3] = append0[1] >> 16 | append0[2] << 16;
w2[0] = append0[2] >> 16 | append0[3] << 16;
w2[1] = append0[3] >> 16 | append1[0] << 16;
w2[2] = append1[0] >> 16 | append1[1] << 16;
w2[3] = append1[1] >> 16 | append1[2] << 16;
w3[0] = append1[2] >> 16 | append1[3] << 16;
w3[1] = append1[3] >> 16 | append2[0] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 23:
w1[1] = w1[1] | append0[0] << 24;
w1[2] = append0[0] >> 8 | append0[1] << 24;
w1[3] = append0[1] >> 8 | append0[2] << 24;
w2[0] = append0[2] >> 8 | append0[3] << 24;
w2[1] = append0[3] >> 8 | append1[0] << 24;
w2[2] = append1[0] >> 8 | append1[1] << 24;
w2[3] = append1[1] >> 8 | append1[2] << 24;
w3[0] = append1[2] >> 8 | append1[3] << 24;
w3[1] = append1[3] >> 8 | append2[0] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 24:
w1[2] = append0[0];
w1[3] = append0[1];
w2[0] = append0[2];
w2[1] = append0[3];
w2[2] = append1[0];
w2[3] = append1[1];
w3[0] = append1[2];
w3[1] = append1[3];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 25:
w1[2] = w1[2] | append0[0] << 8;
w1[3] = append0[0] >> 24 | append0[1] << 8;
w2[0] = append0[1] >> 24 | append0[2] << 8;
w2[1] = append0[2] >> 24 | append0[3] << 8;
w2[2] = append0[3] >> 24 | append1[0] << 8;
w2[3] = append1[0] >> 24 | append1[1] << 8;
w3[0] = append1[1] >> 24 | append1[2] << 8;
w3[1] = append1[2] >> 24 | append1[3] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 26:
w1[2] = w1[2] | append0[0] << 16;
w1[3] = append0[0] >> 16 | append0[1] << 16;
w2[0] = append0[1] >> 16 | append0[2] << 16;
w2[1] = append0[2] >> 16 | append0[3] << 16;
w2[2] = append0[3] >> 16 | append1[0] << 16;
w2[3] = append1[0] >> 16 | append1[1] << 16;
w3[0] = append1[1] >> 16 | append1[2] << 16;
w3[1] = append1[2] >> 16 | append1[3] << 16;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 27:
w1[2] = w1[2] | append0[0] << 24;
w1[3] = append0[0] >> 8 | append0[1] << 24;
w2[0] = append0[1] >> 8 | append0[2] << 24;
w2[1] = append0[2] >> 8 | append0[3] << 24;
w2[2] = append0[3] >> 8 | append1[0] << 24;
w2[3] = append1[0] >> 8 | append1[1] << 24;
w3[0] = append1[1] >> 8 | append1[2] << 24;
w3[1] = append1[2] >> 8 | append1[3] << 24;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 28:
w1[3] = append0[0];
w2[0] = append0[1];
w2[1] = append0[2];
w2[2] = append0[3];
w2[3] = append1[0];
w3[0] = append1[1];
w3[1] = append1[2];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 29:
w1[3] = w1[3] | append0[0] << 8;
w2[0] = append0[0] >> 24 | append0[1] << 8;
w2[1] = append0[1] >> 24 | append0[2] << 8;
w2[2] = append0[2] >> 24 | append0[3] << 8;
w2[3] = append0[3] >> 24 | append1[0] << 8;
w3[0] = append1[0] >> 24 | append1[1] << 8;
w3[1] = append1[1] >> 24 | append1[2] << 8;
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 30:
w1[3] = w1[3] | append0[0] << 16;
w2[0] = append0[0] >> 16 | append0[1] << 16;
w2[1] = append0[1] >> 16 | append0[2] << 16;
w2[2] = append0[2] >> 16 | append0[3] << 16;
w2[3] = append0[3] >> 16 | append1[0] << 16;
w3[0] = append1[0] >> 16 | append1[1] << 16;
w3[1] = append1[1] >> 16 | append1[2] << 16;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 31:
w1[3] = w1[3] | append0[0] << 24;
w2[0] = append0[0] >> 8 | append0[1] << 24;
w2[1] = append0[1] >> 8 | append0[2] << 24;
w2[2] = append0[2] >> 8 | append0[3] << 24;
w2[3] = append0[3] >> 8 | append1[0] << 24;
w3[0] = append1[0] >> 8 | append1[1] << 24;
w3[1] = append1[1] >> 8 | append1[2] << 24;
Fix append_* function calls
2015-12-15 12:42:37 +00:00
break;
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
case 32:
w2[0] = append0[0];
w2[1] = append0[1];
w2[2] = append0[2];
w2[3] = append0[3];
w3[0] = append1[0];
w3[1] = append1[1];
- Dropped all vector code since new GPU's are all scalar, makes the code much easier - Some performance on low-end GPU may drop because of that, but only for a few hash-modes - Dropped scalar code (aka warp) since we do not have any vector datatypes anymore - Renamed C++ overloading functions memcat32_9 -> memcat_c32_w4x4_a3x4 - Still need to fix kernels to new function names, needs to be done manually - Temperature Management needs to be rewritten partially because of conflicting datatypes names - Added code to create different codepaths for NV on AMD in runtime in host (see data.vendor_id) - Added code to create different codepaths for NV on AMD in runtime in kernels (see IS_NV and IS_AMD) - First tests working for -m 0, for example - Great performance increases in general for NV so far - Tested amp_* and markov_* kernel - Migrated special NV optimizations for rule processor
2015-12-15 11:04:22 +00:00
break;
}
}
Fix more append_* functions in kernels
2015-12-15 15:50:21 +00:00
Fix append_* function calls
2015-12-15 12:42:37 +00:00
*/
Reference in New Issue Copy Permalink