arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-12-30 18:31:00 +00:00
Code Issues Releases Wiki Activity
17fdf8754d
hashcat/OpenCL/m04400_a1-optimized.cl

777 lines
32 KiB
Common Lisp
Raw Normal View History

Initial commit
2015-12-04 14:47:52 +00:00
/**
Added docs/credits.txt Added docs/team.txt
2016-09-11 20:20:15 +00:00
* Author......: See docs/credits.txt
Initial commit
2015-12-04 14:47:52 +00:00
* License.....: MIT
*/
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
#define NEW_SIMD_CODE
Move files from include/ to OpenCL/ if they are used within kernels Rename includes in OpenCL so that it's easier to recognize them as such
2016-05-25 21:04:26 +00:00
#include "inc_vendor.cl"
Move macros DGST_R0 - DGST_R3 to host, define dgst_size for opencl kernel from host; both at runtime
2016-06-26 21:39:42 +00:00
#include "inc_hash_constants.h"
Move files from include/ to OpenCL/ if they are used within kernels Rename includes in OpenCL so that it's easier to recognize them as such
2016-05-25 21:04:26 +00:00
#include "inc_hash_functions.cl"
#include "inc_types.cl"
#include "inc_common.cl"
#include "inc_simd.cl"
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
#if VECT_SIZE == 1
#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i)])
#elif VECT_SIZE == 2
#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1])
#elif VECT_SIZE == 4
#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3])
#elif VECT_SIZE == 8
#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7])
#elif VECT_SIZE == 16
#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
#endif
Initial commit
2015-12-04 14:47:52 +00:00
Make all kernel parameter macro helper a function
2018-11-16 13:17:01 +00:00
__kernel void m04400_m04 (KERN_ATTR_BASIC ())
Initial commit
2015-12-04 14:47:52 +00:00
{
/**
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
* modifier
Initial commit
2015-12-04 14:47:52 +00:00
*/
Switch the datatypes of the variables responsible for work-item count and work-item size from u32 to u64
2017-08-19 14:39:22 +00:00
const u64 gid = get_global_id (0);
const u64 lid = get_local_id (0);
const u64 lsz = get_local_size (0);
Initial commit
2015-12-04 14:47:52 +00:00
/**
Upgrade kernel to support dynamic local work sizes
2016-01-19 15:06:03 +00:00
* bin2asc table
Initial commit
2015-12-04 14:47:52 +00:00
*/
Upgrade kernel to support dynamic local work sizes
2016-01-19 15:06:03 +00:00
__local u32 l_bin2asc[256];
Workaround rocm OpenCL runtime bug when copy data from constant to local memory
2018-08-12 16:04:33 +00:00
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
Upgrade kernel to support dynamic local work sizes
2016-01-19 15:06:03 +00:00
{
const u32 i0 = (i >> 0) & 15;
const u32 i1 = (i >> 4) & 15;
l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 8
| ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 0;
}
barrier (CLK_LOCAL_MEM_FENCE);
if (gid >= gid_max) return;
/**
* base
*/
Initial commit
2015-12-04 14:47:52 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
u32 pw_buf0[4];
u32 pw_buf1[4];
Revert "Zero pws_buf before reuse" This reverts commit b409e5e9e162927b25af88e672326b246f17ec5f.
2016-02-22 20:32:38 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
pw_buf0[0] = pws[gid].i[0];
pw_buf0[1] = pws[gid].i[1];
pw_buf0[2] = pws[gid].i[2];
pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
Initial commit
2015-12-04 14:47:52 +00:00
Give the compiler a hint for automatic optimizations based on password length
2018-11-20 14:44:24 +00:00
const u32 pw_l_len = pws[gid].pw_len & 63;
Initial commit
2015-12-04 14:47:52 +00:00
/**
* loop
*/
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
Initial commit
2015-12-04 14:47:52 +00:00
{
Give the compiler a hint for automatic optimizations based on password length
2018-11-20 14:44:24 +00:00
const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos) & 63;
Revert "Zero pws_buf before reuse" This reverts commit b409e5e9e162927b25af88e672326b246f17ec5f.
2016-02-22 20:32:38 +00:00
Give the compiler a hint for automatic optimizations based on password length
2018-11-20 14:44:24 +00:00
const u32x pw_len = (pw_l_len + pw_r_len) & 63;
Initial commit
2015-12-04 14:47:52 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
/**
* concat password candidate
*/
Revert "Zero pws_buf before reuse" This reverts commit b409e5e9e162927b25af88e672326b246f17ec5f.
2016-02-22 20:32:38 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
u32x wordl3[4] = { 0 };
wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
Initial commit
2015-12-04 14:47:52 +00:00
if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
Initial commit
2015-12-04 14:47:52 +00:00
}
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
Initial commit
2015-12-04 14:47:52 +00:00
w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1];
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
w3[2] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
Initial commit
2015-12-04 14:47:52 +00:00
/**
* sha1
*/
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
u32x w0_t = swap32 (w0[0]);
u32x w1_t = swap32 (w0[1]);
u32x w2_t = swap32 (w0[2]);
u32x w3_t = swap32 (w0[3]);
u32x w4_t = swap32 (w1[0]);
u32x w5_t = swap32 (w1[1]);
u32x w6_t = swap32 (w1[2]);
u32x w7_t = swap32 (w1[3]);
u32x w8_t = swap32 (w2[0]);
u32x w9_t = swap32 (w2[1]);
u32x wa_t = swap32 (w2[2]);
u32x wb_t = swap32 (w2[3]);
u32x wc_t = swap32 (w3[0]);
u32x wd_t = swap32 (w3[1]);
u32x we_t = 0;
u32x wf_t = pw_len * 8;
u32x a = SHA1M_A;
u32x b = SHA1M_B;
u32x c = SHA1M_C;
u32x d = SHA1M_D;
u32x e = SHA1M_E;
Initial commit
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w2_t);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w4_t);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w5_t);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w6_t);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w7_t);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w8_t);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w9_t);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, wa_t);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, wb_t);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, wc_t);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, wd_t);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, we_t);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t);
#undef K
#define K SHA1C01
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w7_t);
#undef K
#define K SHA1C02
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wb_t);
#undef K
#define K SHA1C03
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);
a += SHA1M_A;
b += SHA1M_B;
c += SHA1M_C;
d += SHA1M_D;
e += SHA1M_E;
/**
* md5
*/
w0_t = uint_to_hex_lower8 ((a >> 24) & 255) << 0
| uint_to_hex_lower8 ((a >> 16) & 255) << 16;
w1_t = uint_to_hex_lower8 ((a >> 8) & 255) << 0
| uint_to_hex_lower8 ((a >> 0) & 255) << 16;
w2_t = uint_to_hex_lower8 ((b >> 24) & 255) << 0
| uint_to_hex_lower8 ((b >> 16) & 255) << 16;
w3_t = uint_to_hex_lower8 ((b >> 8) & 255) << 0
| uint_to_hex_lower8 ((b >> 0) & 255) << 16;
w4_t = uint_to_hex_lower8 ((c >> 24) & 255) << 0
| uint_to_hex_lower8 ((c >> 16) & 255) << 16;
w5_t = uint_to_hex_lower8 ((c >> 8) & 255) << 0
| uint_to_hex_lower8 ((c >> 0) & 255) << 16;
w6_t = uint_to_hex_lower8 ((d >> 24) & 255) << 0
| uint_to_hex_lower8 ((d >> 16) & 255) << 16;
w7_t = uint_to_hex_lower8 ((d >> 8) & 255) << 0
| uint_to_hex_lower8 ((d >> 0) & 255) << 16;
w8_t = uint_to_hex_lower8 ((e >> 24) & 255) << 0
| uint_to_hex_lower8 ((e >> 16) & 255) << 16;
w9_t = uint_to_hex_lower8 ((e >> 8) & 255) << 0
| uint_to_hex_lower8 ((e >> 0) & 255) << 16;
wa_t = 0x80;
wb_t = 0;
wc_t = 0;
wd_t = 0;
we_t = 40 * 8;
wf_t = 0;
a = MD5M_A;
b = MD5M_B;
c = MD5M_C;
d = MD5M_D;
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
OpenCL Kernels: Reactivate Dalibors XOR optimization on MD5_H on all MD5 based algorithms
2017-08-30 13:32:09 +00:00
u32x t;
MD5_STEP (MD5_H1, a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H2, d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H1, c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H2, b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H1, a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H2, d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H1, c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H2, b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H1, a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H2, d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H1, c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H2, b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H1, a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H2, d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H1, c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H2, b, c, d, a, w2_t, MD5C2f, MD5S23);
Initial commit
2015-12-04 14:47:52 +00:00
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
Revert "Zero pws_buf before reuse" This reverts commit b409e5e9e162927b25af88e672326b246f17ec5f.
2016-02-22 20:32:38 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
COMPARE_M_SIMD (a, d, c, b);
Initial commit
2015-12-04 14:47:52 +00:00
}
}
Make all kernel parameter macro helper a function
2018-11-16 13:17:01 +00:00
__kernel void m04400_m08 (KERN_ATTR_BASIC ())
Initial commit
2015-12-04 14:47:52 +00:00
{
}
Make all kernel parameter macro helper a function
2018-11-16 13:17:01 +00:00
__kernel void m04400_m16 (KERN_ATTR_BASIC ())
Initial commit
2015-12-04 14:47:52 +00:00
{
}
Make all kernel parameter macro helper a function
2018-11-16 13:17:01 +00:00
__kernel void m04400_s04 (KERN_ATTR_BASIC ())
Initial commit
2015-12-04 14:47:52 +00:00
{
/**
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
* modifier
Initial commit
2015-12-04 14:47:52 +00:00
*/
Switch the datatypes of the variables responsible for work-item count and work-item size from u32 to u64
2017-08-19 14:39:22 +00:00
const u64 gid = get_global_id (0);
const u64 lid = get_local_id (0);
const u64 lsz = get_local_size (0);
Initial commit
2015-12-04 14:47:52 +00:00
/**
Upgrade kernel to support dynamic local work sizes
2016-01-19 15:06:03 +00:00
* bin2asc table
Initial commit
2015-12-04 14:47:52 +00:00
*/
Upgrade kernel to support dynamic local work sizes
2016-01-19 15:06:03 +00:00
__local u32 l_bin2asc[256];
Workaround rocm OpenCL runtime bug when copy data from constant to local memory
2018-08-12 16:04:33 +00:00
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
Upgrade kernel to support dynamic local work sizes
2016-01-19 15:06:03 +00:00
{
const u32 i0 = (i >> 0) & 15;
const u32 i1 = (i >> 4) & 15;
l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 8
| ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 0;
}
barrier (CLK_LOCAL_MEM_FENCE);
if (gid >= gid_max) return;
/**
* base
*/
Initial commit
2015-12-04 14:47:52 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
u32 pw_buf0[4];
u32 pw_buf1[4];
Revert "Zero pws_buf before reuse" This reverts commit b409e5e9e162927b25af88e672326b246f17ec5f.
2016-02-22 20:32:38 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
pw_buf0[0] = pws[gid].i[0];
pw_buf0[1] = pws[gid].i[1];
pw_buf0[2] = pws[gid].i[2];
pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
Initial commit
2015-12-04 14:47:52 +00:00
Give the compiler a hint for automatic optimizations based on password length
2018-11-20 14:44:24 +00:00
const u32 pw_l_len = pws[gid].pw_len & 63;
Initial commit
2015-12-04 14:47:52 +00:00
/**
* digest
*/
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/**
* loop
*/
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
Initial commit
2015-12-04 14:47:52 +00:00
{
Give the compiler a hint for automatic optimizations based on password length
2018-11-20 14:44:24 +00:00
const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos) & 63;
Initial commit
2015-12-04 14:47:52 +00:00
Give the compiler a hint for automatic optimizations based on password length
2018-11-20 14:44:24 +00:00
const u32x pw_len = (pw_l_len + pw_r_len) & 63;
Initial commit
2015-12-04 14:47:52 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
/**
* concat password candidate
*/
Revert "Zero pws_buf before reuse" This reverts commit b409e5e9e162927b25af88e672326b246f17ec5f.
2016-02-22 20:32:38 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
u32x wordl3[4] = { 0 };
wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
Initial commit
2015-12-04 14:47:52 +00:00
if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
Initial commit
2015-12-04 14:47:52 +00:00
}
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
Initial commit
2015-12-04 14:47:52 +00:00
w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1];
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
w3[2] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
Initial commit
2015-12-04 14:47:52 +00:00
/**
* sha1
*/
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
u32x w0_t = swap32 (w0[0]);
u32x w1_t = swap32 (w0[1]);
u32x w2_t = swap32 (w0[2]);
u32x w3_t = swap32 (w0[3]);
u32x w4_t = swap32 (w1[0]);
u32x w5_t = swap32 (w1[1]);
u32x w6_t = swap32 (w1[2]);
u32x w7_t = swap32 (w1[3]);
u32x w8_t = swap32 (w2[0]);
u32x w9_t = swap32 (w2[1]);
u32x wa_t = swap32 (w2[2]);
u32x wb_t = swap32 (w2[3]);
u32x wc_t = swap32 (w3[0]);
u32x wd_t = swap32 (w3[1]);
u32x we_t = 0;
u32x wf_t = pw_len * 8;
u32x a = SHA1M_A;
u32x b = SHA1M_B;
u32x c = SHA1M_C;
u32x d = SHA1M_D;
u32x e = SHA1M_E;
Initial commit
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w2_t);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w4_t);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w5_t);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w6_t);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w7_t);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w8_t);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w9_t);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, wa_t);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, wb_t);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, wc_t);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, wd_t);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, we_t);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t);
#undef K
#define K SHA1C01
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w7_t);
#undef K
#define K SHA1C02
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wb_t);
#undef K
#define K SHA1C03
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);
a += SHA1M_A;
b += SHA1M_B;
c += SHA1M_C;
d += SHA1M_D;
e += SHA1M_E;
/**
* md5
*/
w0_t = uint_to_hex_lower8 ((a >> 24) & 255) << 0
| uint_to_hex_lower8 ((a >> 16) & 255) << 16;
w1_t = uint_to_hex_lower8 ((a >> 8) & 255) << 0
| uint_to_hex_lower8 ((a >> 0) & 255) << 16;
w2_t = uint_to_hex_lower8 ((b >> 24) & 255) << 0
| uint_to_hex_lower8 ((b >> 16) & 255) << 16;
w3_t = uint_to_hex_lower8 ((b >> 8) & 255) << 0
| uint_to_hex_lower8 ((b >> 0) & 255) << 16;
w4_t = uint_to_hex_lower8 ((c >> 24) & 255) << 0
| uint_to_hex_lower8 ((c >> 16) & 255) << 16;
w5_t = uint_to_hex_lower8 ((c >> 8) & 255) << 0
| uint_to_hex_lower8 ((c >> 0) & 255) << 16;
w6_t = uint_to_hex_lower8 ((d >> 24) & 255) << 0
| uint_to_hex_lower8 ((d >> 16) & 255) << 16;
w7_t = uint_to_hex_lower8 ((d >> 8) & 255) << 0
| uint_to_hex_lower8 ((d >> 0) & 255) << 16;
w8_t = uint_to_hex_lower8 ((e >> 24) & 255) << 0
| uint_to_hex_lower8 ((e >> 16) & 255) << 16;
w9_t = uint_to_hex_lower8 ((e >> 8) & 255) << 0
| uint_to_hex_lower8 ((e >> 0) & 255) << 16;
wa_t = 0x80;
wb_t = 0;
wc_t = 0;
wd_t = 0;
we_t = 40 * 8;
wf_t = 0;
a = MD5M_A;
b = MD5M_B;
c = MD5M_C;
d = MD5M_D;
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
OpenCL Kernels: Reactivate Dalibors XOR optimization on MD5_H on all MD5 based algorithms
2017-08-30 13:32:09 +00:00
u32x t;
MD5_STEP (MD5_H1, a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H2, d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H1, c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H2, b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H1, a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H2, d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H1, c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H2, b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H1, a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H2, d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H1, c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H2, b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H1, a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H2, d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H1, c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H2, b, c, d, a, w2_t, MD5C2f, MD5S23);
Initial commit
2015-12-04 14:47:52 +00:00
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
if (MATCHES_NONE_VS (a, search[0])) continue;
Initial commit
2015-12-04 14:47:52 +00:00
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
Revert "Zero pws_buf before reuse" This reverts commit b409e5e9e162927b25af88e672326b246f17ec5f.
2016-02-22 20:32:38 +00:00
Cleanup -m 44xx kernels to latest standard
2016-04-10 15:17:50 +00:00
COMPARE_S_SIMD (a, d, c, b);
Initial commit
2015-12-04 14:47:52 +00:00
}
}
Make all kernel parameter macro helper a function
2018-11-16 13:17:01 +00:00
__kernel void m04400_s08 (KERN_ATTR_BASIC ())
Initial commit
2015-12-04 14:47:52 +00:00
{
}
Make all kernel parameter macro helper a function
2018-11-16 13:17:01 +00:00
__kernel void m04400_s16 (KERN_ATTR_BASIC ())
Initial commit
2015-12-04 14:47:52 +00:00
{
}
Reference in New Issue Copy Permalink