2017-08-06 11:54:02 +00:00
|
|
|
/**
|
|
|
|
* Author......: See docs/credits.txt
|
|
|
|
* License.....: MIT
|
|
|
|
*/
|
|
|
|
|
|
|
|
#define NEW_SIMD_CODE
|
|
|
|
|
2019-03-22 14:16:25 +00:00
|
|
|
#ifdef KERNEL_STATIC
|
2019-03-21 22:00:38 +00:00
|
|
|
#include "inc_vendor.h"
|
|
|
|
#include "inc_types.h"
|
2019-04-26 11:59:43 +00:00
|
|
|
#include "inc_platform.cl"
|
2017-08-06 11:54:02 +00:00
|
|
|
#include "inc_common.cl"
|
|
|
|
#include "inc_simd.cl"
|
|
|
|
#include "inc_hash_md5.cl"
|
2019-03-22 14:16:25 +00:00
|
|
|
#endif
|
2017-08-06 11:54:02 +00:00
|
|
|
|
|
|
|
#if VECT_SIZE == 1
|
2019-05-08 13:21:22 +00:00
|
|
|
#define uint_to_hex_lower8(i) make_u32x (l_bin2asc[(i)])
|
2017-08-06 11:54:02 +00:00
|
|
|
#elif VECT_SIZE == 2
|
2019-05-08 13:21:22 +00:00
|
|
|
#define uint_to_hex_lower8(i) make_u32x (l_bin2asc[(i).s0], l_bin2asc[(i).s1])
|
2017-08-06 11:54:02 +00:00
|
|
|
#elif VECT_SIZE == 4
|
2019-05-08 13:21:22 +00:00
|
|
|
#define uint_to_hex_lower8(i) make_u32x (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3])
|
2017-08-06 11:54:02 +00:00
|
|
|
#elif VECT_SIZE == 8
|
2019-05-08 13:21:22 +00:00
|
|
|
#define uint_to_hex_lower8(i) make_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])
|
2017-08-06 11:54:02 +00:00
|
|
|
#elif VECT_SIZE == 16
|
2019-05-08 13:21:22 +00:00
|
|
|
#define uint_to_hex_lower8(i) make_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])
|
2017-08-06 11:54:02 +00:00
|
|
|
#endif
|
|
|
|
|
2019-03-22 21:27:58 +00:00
|
|
|
KERNEL_FQ void m11100_mxx (KERN_ATTR_VECTOR ())
|
2017-08-06 11:54:02 +00:00
|
|
|
{
|
|
|
|
/**
|
|
|
|
* modifier
|
|
|
|
*/
|
|
|
|
|
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);
|
2017-08-06 11:54:02 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* bin2asc table
|
|
|
|
*/
|
|
|
|
|
2019-05-07 07:01:32 +00:00
|
|
|
LOCAL_VK u32 l_bin2asc[256];
|
2017-08-06 11:54:02 +00:00
|
|
|
|
2019-03-03 18:18:56 +00:00
|
|
|
for (u32 i = lid; i < 256; i += lsz)
|
2017-08-06 11:54:02 +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;
|
|
|
|
}
|
|
|
|
|
2019-04-26 11:34:07 +00:00
|
|
|
SYNC_THREADS ();
|
2017-08-06 11:54:02 +00:00
|
|
|
|
|
|
|
if (gid >= gid_max) return;
|
|
|
|
|
|
|
|
/**
|
|
|
|
* salt
|
|
|
|
*/
|
|
|
|
|
|
|
|
u32 challenge;
|
|
|
|
|
|
|
|
challenge = salt_bufs[salt_pos].salt_buf[0];
|
|
|
|
|
|
|
|
u32 salt_buf0[4];
|
|
|
|
u32 salt_buf1[4];
|
|
|
|
|
|
|
|
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[1]; // not a bug, see challenge
|
|
|
|
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[2];
|
|
|
|
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[3];
|
|
|
|
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[4];
|
|
|
|
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[5];
|
|
|
|
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[6];
|
|
|
|
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[7];
|
|
|
|
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[8];
|
|
|
|
|
|
|
|
const u32 salt_len = salt_bufs[salt_pos].salt_len - 4;
|
|
|
|
|
|
|
|
/**
|
|
|
|
* base
|
|
|
|
*/
|
|
|
|
|
2019-02-26 20:20:07 +00:00
|
|
|
const u32 pw_len = pws[gid].pw_len;
|
2017-08-06 11:54:02 +00:00
|
|
|
|
|
|
|
u32x w[64] = { 0 };
|
|
|
|
|
2019-04-10 10:23:39 +00:00
|
|
|
for (u32 i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
2017-08-06 11:54:02 +00:00
|
|
|
{
|
|
|
|
w[idx] = pws[gid].i[idx];
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* loop
|
|
|
|
*/
|
|
|
|
|
|
|
|
u32x w0l = w[0];
|
|
|
|
|
|
|
|
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
|
|
|
|
{
|
|
|
|
const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
|
|
|
|
|
|
|
|
const u32x w0lr = w0l | w0r;
|
|
|
|
|
|
|
|
w[0] = w0lr;
|
|
|
|
|
|
|
|
md5_ctx_vector_t ctx1;
|
|
|
|
|
|
|
|
md5_init_vector (&ctx1);
|
|
|
|
|
|
|
|
md5_update_vector (&ctx1, w, pw_len);
|
|
|
|
|
|
|
|
u32x s0[4];
|
|
|
|
u32x s1[4];
|
|
|
|
u32x s2[4];
|
|
|
|
u32x s3[4];
|
|
|
|
|
|
|
|
s0[0] = salt_buf0[0];
|
|
|
|
s0[1] = salt_buf0[1];
|
|
|
|
s0[2] = salt_buf0[2];
|
|
|
|
s0[3] = salt_buf0[3];
|
|
|
|
s1[0] = salt_buf1[0];
|
|
|
|
s1[1] = salt_buf1[1];
|
|
|
|
s1[2] = salt_buf1[2];
|
|
|
|
s1[3] = salt_buf1[3];
|
|
|
|
s2[0] = 0;
|
|
|
|
s2[1] = 0;
|
|
|
|
s2[2] = 0;
|
|
|
|
s2[3] = 0;
|
|
|
|
s3[0] = 0;
|
|
|
|
s3[1] = 0;
|
|
|
|
s3[2] = 0;
|
|
|
|
s3[3] = 0;
|
|
|
|
|
|
|
|
md5_update_vector_64 (&ctx1, s0, s1, s2, s3, salt_len);
|
|
|
|
|
|
|
|
md5_final_vector (&ctx1);
|
|
|
|
|
|
|
|
const u32x a = ctx1.h[0];
|
|
|
|
const u32x b = ctx1.h[1];
|
|
|
|
const u32x c = ctx1.h[2];
|
|
|
|
const u32x d = ctx1.h[3];
|
|
|
|
|
|
|
|
u32x w0[4];
|
|
|
|
u32x w1[4];
|
|
|
|
u32x w2[4];
|
|
|
|
u32x w3[4];
|
|
|
|
|
|
|
|
w0[0] = uint_to_hex_lower8 ((a >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((a >> 8) & 255) << 16;
|
|
|
|
w0[1] = uint_to_hex_lower8 ((a >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((a >> 24) & 255) << 16;
|
|
|
|
w0[2] = uint_to_hex_lower8 ((b >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((b >> 8) & 255) << 16;
|
|
|
|
w0[3] = uint_to_hex_lower8 ((b >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((b >> 24) & 255) << 16;
|
|
|
|
w1[0] = uint_to_hex_lower8 ((c >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((c >> 8) & 255) << 16;
|
|
|
|
w1[1] = uint_to_hex_lower8 ((c >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((c >> 24) & 255) << 16;
|
|
|
|
w1[2] = uint_to_hex_lower8 ((d >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
|
|
|
|
w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
|
|
|
|
w2[0] = challenge;
|
|
|
|
w2[1] = 0;
|
|
|
|
w2[2] = 0;
|
|
|
|
w2[3] = 0;
|
|
|
|
w3[0] = 0;
|
|
|
|
w3[1] = 0;
|
|
|
|
w3[2] = 0;
|
|
|
|
w3[3] = 0;
|
|
|
|
|
|
|
|
md5_ctx_vector_t ctx;
|
|
|
|
|
|
|
|
md5_init_vector (&ctx);
|
|
|
|
|
|
|
|
ctx.w0[0] = uint_to_hex_lower8 ((a >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((a >> 8) & 255) << 16;
|
|
|
|
ctx.w0[1] = uint_to_hex_lower8 ((a >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((a >> 24) & 255) << 16;
|
|
|
|
ctx.w0[2] = uint_to_hex_lower8 ((b >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((b >> 8) & 255) << 16;
|
|
|
|
ctx.w0[3] = uint_to_hex_lower8 ((b >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((b >> 24) & 255) << 16;
|
|
|
|
ctx.w1[0] = uint_to_hex_lower8 ((c >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((c >> 8) & 255) << 16;
|
|
|
|
ctx.w1[1] = uint_to_hex_lower8 ((c >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((c >> 24) & 255) << 16;
|
|
|
|
ctx.w1[2] = uint_to_hex_lower8 ((d >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
|
|
|
|
ctx.w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
|
|
|
|
ctx.w2[0] = challenge;
|
|
|
|
ctx.w2[1] = 0;
|
|
|
|
ctx.w2[2] = 0;
|
|
|
|
ctx.w2[3] = 0;
|
|
|
|
ctx.w3[0] = 0;
|
|
|
|
ctx.w3[1] = 0;
|
|
|
|
ctx.w3[2] = 0;
|
|
|
|
ctx.w3[3] = 0;
|
|
|
|
|
|
|
|
ctx.len = 32 + 4;
|
|
|
|
|
|
|
|
md5_final_vector (&ctx);
|
|
|
|
|
|
|
|
const u32x r0 = ctx.h[DGST_R0];
|
|
|
|
const u32x r1 = ctx.h[DGST_R1];
|
|
|
|
const u32x r2 = ctx.h[DGST_R2];
|
|
|
|
const u32x r3 = ctx.h[DGST_R3];
|
|
|
|
|
|
|
|
COMPARE_M_SIMD (r0, r1, r2, r3);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-03-22 21:27:58 +00:00
|
|
|
KERNEL_FQ void m11100_sxx (KERN_ATTR_VECTOR ())
|
2017-08-06 11:54:02 +00:00
|
|
|
{
|
|
|
|
/**
|
|
|
|
* modifier
|
|
|
|
*/
|
|
|
|
|
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);
|
2017-08-06 11:54:02 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* bin2asc table
|
|
|
|
*/
|
|
|
|
|
2019-05-07 07:01:32 +00:00
|
|
|
LOCAL_VK u32 l_bin2asc[256];
|
2017-08-06 11:54:02 +00:00
|
|
|
|
2019-03-03 18:18:56 +00:00
|
|
|
for (u32 i = lid; i < 256; i += lsz)
|
2017-08-06 11:54:02 +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;
|
|
|
|
}
|
|
|
|
|
2019-04-26 11:34:07 +00:00
|
|
|
SYNC_THREADS ();
|
2017-08-06 11:54:02 +00:00
|
|
|
|
|
|
|
if (gid >= gid_max) return;
|
|
|
|
|
|
|
|
/**
|
|
|
|
* 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]
|
|
|
|
};
|
|
|
|
|
|
|
|
/**
|
|
|
|
* salt
|
|
|
|
*/
|
|
|
|
|
|
|
|
u32 challenge;
|
|
|
|
|
|
|
|
challenge = salt_bufs[salt_pos].salt_buf[0];
|
|
|
|
|
|
|
|
u32 salt_buf0[4];
|
|
|
|
u32 salt_buf1[4];
|
|
|
|
|
|
|
|
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[1]; // not a bug, see challenge
|
|
|
|
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[2];
|
|
|
|
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[3];
|
|
|
|
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[4];
|
|
|
|
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[5];
|
|
|
|
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[6];
|
|
|
|
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[7];
|
|
|
|
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[8];
|
|
|
|
|
|
|
|
const u32 salt_len = salt_bufs[salt_pos].salt_len - 4;
|
|
|
|
|
|
|
|
/**
|
|
|
|
* base
|
|
|
|
*/
|
|
|
|
|
2019-02-26 20:20:07 +00:00
|
|
|
const u32 pw_len = pws[gid].pw_len;
|
2017-08-06 11:54:02 +00:00
|
|
|
|
|
|
|
u32x w[64] = { 0 };
|
|
|
|
|
2019-04-10 10:23:39 +00:00
|
|
|
for (u32 i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
2017-08-06 11:54:02 +00:00
|
|
|
{
|
|
|
|
w[idx] = pws[gid].i[idx];
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* loop
|
|
|
|
*/
|
|
|
|
|
|
|
|
u32x w0l = w[0];
|
|
|
|
|
|
|
|
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
|
|
|
|
{
|
|
|
|
const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
|
|
|
|
|
|
|
|
const u32x w0lr = w0l | w0r;
|
|
|
|
|
|
|
|
w[0] = w0lr;
|
|
|
|
|
|
|
|
md5_ctx_vector_t ctx1;
|
|
|
|
|
|
|
|
md5_init_vector (&ctx1);
|
|
|
|
|
|
|
|
md5_update_vector (&ctx1, w, pw_len);
|
|
|
|
|
|
|
|
u32x s0[4];
|
|
|
|
u32x s1[4];
|
|
|
|
u32x s2[4];
|
|
|
|
u32x s3[4];
|
|
|
|
|
|
|
|
s0[0] = salt_buf0[0];
|
|
|
|
s0[1] = salt_buf0[1];
|
|
|
|
s0[2] = salt_buf0[2];
|
|
|
|
s0[3] = salt_buf0[3];
|
|
|
|
s1[0] = salt_buf1[0];
|
|
|
|
s1[1] = salt_buf1[1];
|
|
|
|
s1[2] = salt_buf1[2];
|
|
|
|
s1[3] = salt_buf1[3];
|
|
|
|
s2[0] = 0;
|
|
|
|
s2[1] = 0;
|
|
|
|
s2[2] = 0;
|
|
|
|
s2[3] = 0;
|
|
|
|
s3[0] = 0;
|
|
|
|
s3[1] = 0;
|
|
|
|
s3[2] = 0;
|
|
|
|
s3[3] = 0;
|
|
|
|
|
|
|
|
md5_update_vector_64 (&ctx1, s0, s1, s2, s3, salt_len);
|
|
|
|
|
|
|
|
md5_final_vector (&ctx1);
|
|
|
|
|
|
|
|
const u32x a = ctx1.h[0];
|
|
|
|
const u32x b = ctx1.h[1];
|
|
|
|
const u32x c = ctx1.h[2];
|
|
|
|
const u32x d = ctx1.h[3];
|
|
|
|
|
|
|
|
md5_ctx_vector_t ctx;
|
|
|
|
|
|
|
|
md5_init_vector (&ctx);
|
|
|
|
|
|
|
|
ctx.w0[0] = uint_to_hex_lower8 ((a >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((a >> 8) & 255) << 16;
|
|
|
|
ctx.w0[1] = uint_to_hex_lower8 ((a >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((a >> 24) & 255) << 16;
|
|
|
|
ctx.w0[2] = uint_to_hex_lower8 ((b >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((b >> 8) & 255) << 16;
|
|
|
|
ctx.w0[3] = uint_to_hex_lower8 ((b >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((b >> 24) & 255) << 16;
|
|
|
|
ctx.w1[0] = uint_to_hex_lower8 ((c >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((c >> 8) & 255) << 16;
|
|
|
|
ctx.w1[1] = uint_to_hex_lower8 ((c >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((c >> 24) & 255) << 16;
|
|
|
|
ctx.w1[2] = uint_to_hex_lower8 ((d >> 0) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
|
|
|
|
ctx.w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
|
|
|
|
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
|
|
|
|
ctx.w2[0] = challenge;
|
|
|
|
ctx.w2[1] = 0;
|
|
|
|
ctx.w2[2] = 0;
|
|
|
|
ctx.w2[3] = 0;
|
|
|
|
ctx.w3[0] = 0;
|
|
|
|
ctx.w3[1] = 0;
|
|
|
|
ctx.w3[2] = 0;
|
|
|
|
ctx.w3[3] = 0;
|
|
|
|
|
|
|
|
ctx.len = 32 + 4;
|
|
|
|
|
|
|
|
md5_final_vector (&ctx);
|
|
|
|
|
|
|
|
const u32x r0 = ctx.h[DGST_R0];
|
|
|
|
const u32x r1 = ctx.h[DGST_R1];
|
|
|
|
const u32x r2 = ctx.h[DGST_R2];
|
|
|
|
const u32x r3 = ctx.h[DGST_R3];
|
|
|
|
|
|
|
|
COMPARE_S_SIMD (r0, r1, r2, r3);
|
|
|
|
}
|
|
|
|
}
|