mirror of https://github.com/hashcat/hashcat.git
VeraCrypt added the possibility to use Streebog-512 as hashing algorithm for the key derivation. This commit adds the necessary VeraCrypt kernels as well as additional HMAC-Streebog kernels. - Add hash-mode 13771: VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 512 bit - Add hash-mode 13772: VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 1024 bit - Add hash-mode 13773: VeraCrypt PBKDF2-HMAC-Streebog-512 + XTS 1536 bit - Add hash-mode 11750: HMAC-Streebog-256 (key = $pass), big-endian - Add hash-mode 11760: HMAC-Streebog-256 (key = $salt), big-endian - Add hash-mode 11860: HMAC-Streebog-512 (key = $salt), big-endian - Add test suite for hash-modes 11750, 11760 and 11860 - Improve pure Streebog kernelspull/1771/head
R. Yushaev
6 years ago
parent
e1011c51a3
commit
47bd838e25
@ -0,0 +1,192 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_rp.h"
|
||||
#include "inc_rp.cl"
|
||||
#include "inc_scalar.cl"
|
||||
#include "inc_hash_streebog256.cl"
|
||||
|
||||
__kernel void m11750_mxx (__global pw_t *pws, __constant const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
COPY_PW (pws[gid]);
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
pw_t tmp = PASTE_PW;
|
||||
|
||||
tmp.pw_len = apply_rules (rules_buf[il_pos].cmds, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog256_hmac_ctx_t ctx;
|
||||
|
||||
streebog256_hmac_init_swap (&ctx, tmp.i, tmp.pw_len, s_sbob_sl64);
|
||||
|
||||
streebog256_hmac_update (&ctx, s, salt_len);
|
||||
|
||||
streebog256_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11750_sxx (__global pw_t *pws, __constant const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
COPY_PW (pws[gid]);
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
pw_t tmp = PASTE_PW;
|
||||
|
||||
tmp.pw_len = apply_rules (rules_buf[il_pos].cmds, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog256_hmac_ctx_t ctx;
|
||||
|
||||
streebog256_hmac_init_swap (&ctx, tmp.i, tmp.pw_len, s_sbob_sl64);
|
||||
|
||||
streebog256_hmac_update (&ctx, s, salt_len);
|
||||
|
||||
streebog256_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,240 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_scalar.cl"
|
||||
#include "inc_hash_streebog256.cl"
|
||||
|
||||
__kernel void m11750_mxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32 w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = swap32_S (pws[gid].i[idx]);
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
const u32 comb_len = combs_buf[il_pos].pw_len;
|
||||
|
||||
u32 c[64];
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int idx = 0; idx < 64; idx++)
|
||||
{
|
||||
c[idx] = swap32_S (combs_buf[il_pos].i[idx]);
|
||||
}
|
||||
|
||||
switch_buffer_by_offset_1x64_be_S (c, pw_len);
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int i = 0; i < 64; i++)
|
||||
{
|
||||
c[i] |= w[i];
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_t ctx;
|
||||
|
||||
streebog256_hmac_init (&ctx, c, pw_len + comb_len, s_sbob_sl64);
|
||||
|
||||
streebog256_hmac_update (&ctx, s, salt_len);
|
||||
|
||||
streebog256_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11750_sxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32 w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = swap32_S (pws[gid].i[idx]);
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
const u32 comb_len = combs_buf[il_pos].pw_len;
|
||||
|
||||
u32 c[64];
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int idx = 0; idx < 64; idx++)
|
||||
{
|
||||
c[idx] = swap32_S (combs_buf[il_pos].i[idx]);
|
||||
}
|
||||
|
||||
switch_buffer_by_offset_1x64_be_S (c, pw_len);
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int i = 0; i < 64; i++)
|
||||
{
|
||||
c[i] |= w[i];
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_t ctx;
|
||||
|
||||
streebog256_hmac_init (&ctx, c, pw_len + comb_len, s_sbob_sl64);
|
||||
|
||||
streebog256_hmac_update (&ctx, s, salt_len);
|
||||
|
||||
streebog256_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,212 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_simd.cl"
|
||||
#include "inc_hash_streebog256.cl"
|
||||
|
||||
__kernel void m11750_mxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32x w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = pws[gid].i[idx];
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32x s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[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 w0 = w0l | w0r;
|
||||
|
||||
w[0] = w0;
|
||||
|
||||
streebog256_hmac_ctx_vector_t ctx;
|
||||
|
||||
streebog256_hmac_init_vector_swap (&ctx, w, pw_len, s_sbob_sl64);
|
||||
|
||||
streebog256_hmac_update_vector (&ctx, s, salt_len);
|
||||
|
||||
streebog256_hmac_final_vector (&ctx);
|
||||
|
||||
const u32x r0 = l32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r1 = h32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r2 = l32_from_64 (ctx.opad.h[1]);
|
||||
const u32x r3 = h32_from_64 (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SIMD (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11750_sxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32x w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = pws[gid].i[idx];
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32x s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[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 w0 = w0l | w0r;
|
||||
|
||||
w[0] = w0;
|
||||
|
||||
streebog256_hmac_ctx_vector_t ctx;
|
||||
|
||||
streebog256_hmac_init_vector_swap (&ctx, w, pw_len, s_sbob_sl64);
|
||||
|
||||
streebog256_hmac_update_vector (&ctx, s, salt_len);
|
||||
|
||||
streebog256_hmac_final_vector (&ctx);
|
||||
|
||||
const u32x r0 = l32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r1 = h32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r2 = l32_from_64 (ctx.opad.h[1]);
|
||||
const u32x r3 = h32_from_64 (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SIMD (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,196 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_rp.h"
|
||||
#include "inc_rp.cl"
|
||||
#include "inc_scalar.cl"
|
||||
#include "inc_hash_streebog256.cl"
|
||||
|
||||
__kernel void m11760_mxx (__global pw_t *pws, __constant const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
COPY_PW (pws[gid]);
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_t ctx0;
|
||||
|
||||
streebog256_hmac_init (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
pw_t tmp = PASTE_PW;
|
||||
|
||||
tmp.pw_len = apply_rules (rules_buf[il_pos].cmds, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog256_hmac_ctx_t ctx = ctx0;
|
||||
|
||||
streebog256_hmac_update_swap (&ctx, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog256_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11760_sxx (__global pw_t *pws, __constant const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
COPY_PW (pws[gid]);
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_t ctx0;
|
||||
|
||||
streebog256_hmac_init (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
pw_t tmp = PASTE_PW;
|
||||
|
||||
tmp.pw_len = apply_rules (rules_buf[il_pos].cmds, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog256_hmac_ctx_t ctx = ctx0;
|
||||
|
||||
streebog256_hmac_update_swap (&ctx, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog256_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,244 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_scalar.cl"
|
||||
#include "inc_hash_streebog256.cl"
|
||||
|
||||
__kernel void m11760_mxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32 w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = swap32_S (pws[gid].i[idx]);
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_t ctx0;
|
||||
|
||||
streebog256_hmac_init (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
const u32 comb_len = combs_buf[il_pos].pw_len;
|
||||
|
||||
u32 c[64];
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int idx = 0; idx < 64; idx++)
|
||||
{
|
||||
c[idx] = swap32_S (combs_buf[il_pos].i[idx]);
|
||||
}
|
||||
|
||||
switch_buffer_by_offset_1x64_be_S (c, pw_len);
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int i = 0; i < 64; i++)
|
||||
{
|
||||
c[i] |= w[i];
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_t ctx = ctx0;
|
||||
|
||||
streebog256_hmac_update (&ctx, c, pw_len + comb_len);
|
||||
|
||||
streebog256_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11760_sxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32 w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = swap32_S (pws[gid].i[idx]);
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_t ctx0;
|
||||
|
||||
streebog256_hmac_init (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
const u32 comb_len = combs_buf[il_pos].pw_len;
|
||||
|
||||
u32 c[64];
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int idx = 0; idx < 64; idx++)
|
||||
{
|
||||
c[idx] = swap32_S (combs_buf[il_pos].i[idx]);
|
||||
}
|
||||
|
||||
switch_buffer_by_offset_1x64_be_S (c, pw_len);
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int i = 0; i < 64; i++)
|
||||
{
|
||||
c[i] |= w[i];
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_t ctx = ctx0;
|
||||
|
||||
streebog256_hmac_update (&ctx, c, pw_len + comb_len);
|
||||
|
||||
streebog256_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,216 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_simd.cl"
|
||||
#include "inc_hash_streebog256.cl"
|
||||
|
||||
__kernel void m11760_mxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32x w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = pws[gid].i[idx];
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32x s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_vector_t ctx0;
|
||||
|
||||
streebog256_hmac_init_vector (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* 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 w0 = w0l | w0r;
|
||||
|
||||
w[0] = w0;
|
||||
|
||||
streebog256_hmac_ctx_vector_t ctx = ctx0;
|
||||
|
||||
streebog256_hmac_update_vector_swap (&ctx, w, pw_len);
|
||||
|
||||
streebog256_hmac_final_vector (&ctx);
|
||||
|
||||
const u32x r0 = l32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r1 = h32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r2 = l32_from_64 (ctx.opad.h[1]);
|
||||
const u32x r3 = h32_from_64 (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SIMD (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11760_sxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32x w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = pws[gid].i[idx];
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32x s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog256_hmac_ctx_vector_t ctx0;
|
||||
|
||||
streebog256_hmac_init_vector (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* 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 w0 = w0l | w0r;
|
||||
|
||||
w[0] = w0;
|
||||
|
||||
streebog256_hmac_ctx_vector_t ctx = ctx0;
|
||||
|
||||
streebog256_hmac_update_vector_swap (&ctx, w, pw_len);
|
||||
|
||||
streebog256_hmac_final_vector (&ctx);
|
||||
|
||||
const u32x r0 = l32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r1 = h32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r2 = l32_from_64 (ctx.opad.h[1]);
|
||||
const u32x r3 = h32_from_64 (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SIMD (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,196 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_rp.h"
|
||||
#include "inc_rp.cl"
|
||||
#include "inc_scalar.cl"
|
||||
#include "inc_hash_streebog512.cl"
|
||||
|
||||
__kernel void m11860_mxx (__global pw_t *pws, __constant const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
COPY_PW (pws[gid]);
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog512_hmac_ctx_t ctx0;
|
||||
|
||||
streebog512_hmac_init (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
pw_t tmp = PASTE_PW;
|
||||
|
||||
tmp.pw_len = apply_rules (rules_buf[il_pos].cmds, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog512_hmac_ctx_t ctx = ctx0;
|
||||
|
||||
streebog512_hmac_update_swap (&ctx, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog512_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11860_sxx (__global pw_t *pws, __constant const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
COPY_PW (pws[gid]);
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog512_hmac_ctx_t ctx0;
|
||||
|
||||
streebog512_hmac_init (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
pw_t tmp = PASTE_PW;
|
||||
|
||||
tmp.pw_len = apply_rules (rules_buf[il_pos].cmds, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog512_hmac_ctx_t ctx = ctx0;
|
||||
|
||||
streebog512_hmac_update_swap (&ctx, tmp.i, tmp.pw_len);
|
||||
|
||||
streebog512_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,244 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_scalar.cl"
|
||||
#include "inc_hash_streebog512.cl"
|
||||
|
||||
__kernel void m11860_mxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32 w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = swap32_S (pws[gid].i[idx]);
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog512_hmac_ctx_t ctx0;
|
||||
|
||||
streebog512_hmac_init (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
const u32 comb_len = combs_buf[il_pos].pw_len;
|
||||
|
||||
u32 c[64];
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int idx = 0; idx < 64; idx++)
|
||||
{
|
||||
c[idx] = swap32_S (combs_buf[il_pos].i[idx]);
|
||||
}
|
||||
|
||||
switch_buffer_by_offset_1x64_be_S (c, pw_len);
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int i = 0; i < 64; i++)
|
||||
{
|
||||
c[i] |= w[i];
|
||||
}
|
||||
|
||||
streebog512_hmac_ctx_t ctx = ctx0;
|
||||
|
||||
streebog512_hmac_update (&ctx, c, pw_len + comb_len);
|
||||
|
||||
streebog512_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11860_sxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32 w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = swap32_S (pws[gid].i[idx]);
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32 s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog512_hmac_ctx_t ctx0;
|
||||
|
||||
streebog512_hmac_init (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* loop
|
||||
*/
|
||||
|
||||
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
|
||||
{
|
||||
const u32 comb_len = combs_buf[il_pos].pw_len;
|
||||
|
||||
u32 c[64];
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int idx = 0; idx < 64; idx++)
|
||||
{
|
||||
c[idx] = swap32_S (combs_buf[il_pos].i[idx]);
|
||||
}
|
||||
|
||||
switch_buffer_by_offset_1x64_be_S (c, pw_len);
|
||||
|
||||
#ifdef _unroll
|
||||
#pragma unroll
|
||||
#endif
|
||||
for (int i = 0; i < 64; i++)
|
||||
{
|
||||
c[i] |= w[i];
|
||||
}
|
||||
|
||||
streebog512_hmac_ctx_t ctx = ctx0;
|
||||
|
||||
streebog512_hmac_update (&ctx, c, pw_len + comb_len);
|
||||
|
||||
streebog512_hmac_final (&ctx);
|
||||
|
||||
const u32 r0 = l32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r1 = h32_from_64_S (ctx.opad.h[0]);
|
||||
const u32 r2 = l32_from_64_S (ctx.opad.h[1]);
|
||||
const u32 r3 = h32_from_64_S (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SCALAR (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,216 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_simd.cl"
|
||||
#include "inc_hash_streebog512.cl"
|
||||
|
||||
__kernel void m11860_mxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32x w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = pws[gid].i[idx];
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32x s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog512_hmac_ctx_vector_t ctx0;
|
||||
|
||||
streebog512_hmac_init_vector (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* 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 w0 = w0l | w0r;
|
||||
|
||||
w[0] = w0;
|
||||
|
||||
streebog512_hmac_ctx_vector_t ctx = ctx0;
|
||||
|
||||
streebog512_hmac_update_vector_swap (&ctx, w, pw_len);
|
||||
|
||||
streebog512_hmac_final_vector (&ctx);
|
||||
|
||||
const u32x r0 = l32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r1 = h32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r2 = l32_from_64 (ctx.opad.h[1]);
|
||||
const u32x r3 = h32_from_64 (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_M_SIMD (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m11860_sxx (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __constant const u32x *words_buf_r, __global void *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* modifier
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
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]
|
||||
};
|
||||
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u32 pw_len = pws[gid].pw_len;
|
||||
|
||||
u32x w[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < pw_len; i += 4, idx += 1)
|
||||
{
|
||||
w[idx] = pws[gid].i[idx];
|
||||
}
|
||||
|
||||
const u32 salt_len = salt_bufs[salt_pos].salt_len;
|
||||
|
||||
u32x s[64] = { 0 };
|
||||
|
||||
for (int i = 0, idx = 0; i < salt_len; i += 4, idx += 1)
|
||||
{
|
||||
s[idx] = swap32_S (salt_bufs[salt_pos].salt_buf[idx]);
|
||||
}
|
||||
|
||||
streebog512_hmac_ctx_vector_t ctx0;
|
||||
|
||||
streebog512_hmac_init_vector (&ctx0, s, salt_len, s_sbob_sl64);
|
||||
|
||||
/**
|
||||
* 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 w0 = w0l | w0r;
|
||||
|
||||
w[0] = w0;
|
||||
|
||||
streebog512_hmac_ctx_vector_t ctx = ctx0;
|
||||
|
||||
streebog512_hmac_update_vector_swap (&ctx, w, pw_len);
|
||||
|
||||
streebog512_hmac_final_vector (&ctx);
|
||||
|
||||
const u32x r0 = l32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r1 = h32_from_64 (ctx.opad.h[0]);
|
||||
const u32x r2 = l32_from_64 (ctx.opad.h[1]);
|
||||
const u32x r3 = h32_from_64 (ctx.opad.h[1]);
|
||||
|
||||
COMPARE_S_SIMD (r0, r1, r2, r3);
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,576 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_simd.cl"
|
||||
#include "inc_hash_streebog512.cl"
|
||||
|
||||
#include "inc_cipher_aes.cl"
|
||||
#include "inc_cipher_twofish.cl"
|
||||
#include "inc_cipher_serpent.cl"
|
||||
|
||||
#include "inc_truecrypt_crc32.cl"
|
||||
#include "inc_truecrypt_xts.cl"
|
||||
|
||||
DECLSPEC u32 u8add (const u32 a, const u32 b)
|
||||
{
|
||||
const u32 a1 = (a >> 0) & 0xff;
|
||||
const u32 a2 = (a >> 8) & 0xff;
|
||||
const u32 a3 = (a >> 16) & 0xff;
|
||||
const u32 a4 = (a >> 24) & 0xff;
|
||||
|
||||
const u32 b1 = (b >> 0) & 0xff;
|
||||
const u32 b2 = (b >> 8) & 0xff;
|
||||
const u32 b3 = (b >> 16) & 0xff;
|
||||
const u32 b4 = (b >> 24) & 0xff;
|
||||
|
||||
const u32 r1 = (a1 + b1) & 0xff;
|
||||
const u32 r2 = (a2 + b2) & 0xff;
|
||||
const u32 r3 = (a3 + b3) & 0xff;
|
||||
const u32 r4 = (a4 + b4) & 0xff;
|
||||
|
||||
const u32 r = r1 << 0
|
||||
| r2 << 8
|
||||
| r3 << 16
|
||||
| r4 << 24;
|
||||
|
||||
return r;
|
||||
}
|
||||
|
||||
DECLSPEC void hmac_streebog512_run_V (u32x *w0, u32x *w1, u32x *w2, u32x *w3, u64x *ipad_hash, u64x *opad_hash, u64x *ipad_raw, u64x *opad_raw, u64x *digest, SHM_TYPE u64a (*s_sbob_sl64)[256])
|
||||
{
|
||||
const u64x nullbuf[8] = { 0 };
|
||||
u64x counterbuf[8] = { 0 };
|
||||
u64x padding[8] = { 0 };
|
||||
u64x message[8];
|
||||
|
||||
padding[7] = swap64 ((u64x) 0x01);
|
||||
|
||||
//inner HMAC: ipad + message
|
||||
|
||||
//first transform: precalculated ipad hash
|
||||
counterbuf[7] = swap64 ((u64x) 0x200);
|
||||
|
||||
//second transform: message = previous HMAC digest
|
||||
message[7] = hl32_to_64 (w3[2], w3[3]);
|
||||
message[6] = hl32_to_64 (w3[0], w3[1]);
|
||||
message[5] = hl32_to_64 (w2[2], w2[3]);
|
||||
message[4] = hl32_to_64 (w2[0], w2[1]);
|
||||
message[3] = hl32_to_64 (w1[2], w1[3]);
|
||||
message[2] = hl32_to_64 (w1[0], w1[1]);
|
||||
message[1] = hl32_to_64 (w0[2], w0[3]);
|
||||
message[0] = hl32_to_64 (w0[0], w0[1]);
|
||||
|
||||
digest[0] = ipad_hash[0];
|
||||
digest[1] = ipad_hash[1];
|
||||
digest[2] = ipad_hash[2];
|
||||
digest[3] = ipad_hash[3];
|
||||
digest[4] = ipad_hash[4];
|
||||
digest[5] = ipad_hash[5];
|
||||
digest[6] = ipad_hash[6];
|
||||
digest[7] = ipad_hash[7];
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, message, s_sbob_sl64);
|
||||
|
||||
counterbuf[7] = swap64 ((u64x) 0x400);
|
||||
|
||||
//final: padding byte
|
||||
streebog512_g_vector (digest, counterbuf, padding, s_sbob_sl64);
|
||||
|
||||
streebog512_add_vector (message, ipad_raw);
|
||||
streebog512_add_vector (message, padding);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, counterbuf, s_sbob_sl64);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, message, s_sbob_sl64);
|
||||
|
||||
//outer HMAC: opad + digest
|
||||
|
||||
//first transform: precalculated opad hash
|
||||
counterbuf[7] = swap64 ((u64x) 0x200);
|
||||
|
||||
//second transform: message = inner HMAC digest
|
||||
message[0] = digest[0];
|
||||
message[1] = digest[1];
|
||||
message[2] = digest[2];
|
||||
message[3] = digest[3];
|
||||
message[4] = digest[4];
|
||||
message[5] = digest[5];
|
||||
message[6] = digest[6];
|
||||
message[7] = digest[7];
|
||||
|
||||
digest[0] = opad_hash[0];
|
||||
digest[1] = opad_hash[1];
|
||||
digest[2] = opad_hash[2];
|
||||
digest[3] = opad_hash[3];
|
||||
digest[4] = opad_hash[4];
|
||||
digest[5] = opad_hash[5];
|
||||
digest[6] = opad_hash[6];
|
||||
digest[7] = opad_hash[7];
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, message, s_sbob_sl64);
|
||||
|
||||
counterbuf[7] = swap64 ((u64x) 0x400);
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, padding, s_sbob_sl64);
|
||||
|
||||
streebog512_add_vector (message, opad_raw);
|
||||
streebog512_add_vector (message, padding);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, counterbuf, s_sbob_sl64);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, message, s_sbob_sl64);
|
||||
}
|
||||
|
||||
__kernel void m13771_init (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
u32 w0[4];
|
||||
u32 w1[4];
|
||||
u32 w2[4];
|
||||
u32 w3[4];
|
||||
|
||||
w0[0] = pws[gid].i[ 0];
|
||||
w0[1] = pws[gid].i[ 1];
|
||||
w0[2] = pws[gid].i[ 2];
|
||||
w0[3] = pws[gid].i[ 3];
|
||||
w1[0] = pws[gid].i[ 4];
|
||||
w1[1] = pws[gid].i[ 5];
|
||||
w1[2] = pws[gid].i[ 6];
|
||||
w1[3] = pws[gid].i[ 7];
|
||||
w2[0] = pws[gid].i[ 8];
|
||||
w2[1] = pws[gid].i[ 9];
|
||||
w2[2] = pws[gid].i[10];
|
||||
w2[3] = pws[gid].i[11];
|
||||
w3[0] = pws[gid].i[12];
|
||||
w3[1] = pws[gid].i[13];
|
||||
w3[2] = pws[gid].i[14];
|
||||
w3[3] = pws[gid].i[15];
|
||||
|
||||
w0[0] = u8add (w0[0], esalt_bufs[digests_offset].keyfile_buf[ 0]);
|
||||
w0[1] = u8add (w0[1], esalt_bufs[digests_offset].keyfile_buf[ 1]);
|
||||
w0[2] = u8add (w0[2], esalt_bufs[digests_offset].keyfile_buf[ 2]);
|
||||
w0[3] = u8add (w0[3], esalt_bufs[digests_offset].keyfile_buf[ 3]);
|
||||
w1[0] = u8add (w1[0], esalt_bufs[digests_offset].keyfile_buf[ 4]);
|
||||
w1[1] = u8add (w1[1], esalt_bufs[digests_offset].keyfile_buf[ 5]);
|
||||
w1[2] = u8add (w1[2], esalt_bufs[digests_offset].keyfile_buf[ 6]);
|
||||
w1[3] = u8add (w1[3], esalt_bufs[digests_offset].keyfile_buf[ 7]);
|
||||
w2[0] = u8add (w2[0], esalt_bufs[digests_offset].keyfile_buf[ 8]);
|
||||
w2[1] = u8add (w2[1], esalt_bufs[digests_offset].keyfile_buf[ 9]);
|
||||
w2[2] = u8add (w2[2], esalt_bufs[digests_offset].keyfile_buf[10]);
|
||||
w2[3] = u8add (w2[3], esalt_bufs[digests_offset].keyfile_buf[11]);
|
||||
w3[0] = u8add (w3[0], esalt_bufs[digests_offset].keyfile_buf[12]);
|
||||
w3[1] = u8add (w3[1], esalt_bufs[digests_offset].keyfile_buf[13]);
|
||||
w3[2] = u8add (w3[2], esalt_bufs[digests_offset].keyfile_buf[14]);
|
||||
w3[3] = u8add (w3[3], esalt_bufs[digests_offset].keyfile_buf[15]);
|
||||
|
||||
w0[0] = swap32_S (w0[0]);
|
||||
w0[1] = swap32_S (w0[1]);
|
||||
w0[2] = swap32_S (w0[2]);
|
||||
w0[3] = swap32_S (w0[3]);
|
||||
w1[0] = swap32_S (w1[0]);
|
||||
w1[1] = swap32_S (w1[1]);
|
||||
w1[2] = swap32_S (w1[2]);
|
||||
w1[3] = swap32_S (w1[3]);
|
||||
w2[0] = swap32_S (w2[0]);
|
||||
w2[1] = swap32_S (w2[1]);
|
||||
w2[2] = swap32_S (w2[2]);
|
||||
w2[3] = swap32_S (w2[3]);
|
||||
w3[0] = swap32_S (w3[0]);
|
||||
w3[1] = swap32_S (w3[1]);
|
||||
w3[2] = swap32_S (w3[2]);
|
||||
w3[3] = swap32_S (w3[3]);
|
||||
|
||||
streebog512_hmac_ctx_t streebog512_hmac_ctx;
|
||||
|
||||
streebog512_hmac_init_64 (&streebog512_hmac_ctx, w0, w1, w2, w3, s_sbob_sl64);
|
||||
|
||||
tmps[gid].ipad_hash[0] = streebog512_hmac_ctx.ipad.h[0];
|
||||
tmps[gid].ipad_hash[1] = streebog512_hmac_ctx.ipad.h[1];
|
||||
tmps[gid].ipad_hash[2] = streebog512_hmac_ctx.ipad.h[2];
|
||||
tmps[gid].ipad_hash[3] = streebog512_hmac_ctx.ipad.h[3];
|
||||
tmps[gid].ipad_hash[4] = streebog512_hmac_ctx.ipad.h[4];
|
||||
tmps[gid].ipad_hash[5] = streebog512_hmac_ctx.ipad.h[5];
|
||||
tmps[gid].ipad_hash[6] = streebog512_hmac_ctx.ipad.h[6];
|
||||
tmps[gid].ipad_hash[7] = streebog512_hmac_ctx.ipad.h[7];
|
||||
|
||||
tmps[gid].opad_hash[0] = streebog512_hmac_ctx.opad.h[0];
|
||||
tmps[gid].opad_hash[1] = streebog512_hmac_ctx.opad.h[1];
|
||||
tmps[gid].opad_hash[2] = streebog512_hmac_ctx.opad.h[2];
|
||||
tmps[gid].opad_hash[3] = streebog512_hmac_ctx.opad.h[3];
|
||||
tmps[gid].opad_hash[4] = streebog512_hmac_ctx.opad.h[4];
|
||||
tmps[gid].opad_hash[5] = streebog512_hmac_ctx.opad.h[5];
|
||||
tmps[gid].opad_hash[6] = streebog512_hmac_ctx.opad.h[6];
|
||||
tmps[gid].opad_hash[7] = streebog512_hmac_ctx.opad.h[7];
|
||||
|
||||
tmps[gid].ipad_raw[0] = streebog512_hmac_ctx.ipad.s[0];
|
||||
tmps[gid].ipad_raw[1] = streebog512_hmac_ctx.ipad.s[1];
|
||||
tmps[gid].ipad_raw[2] = streebog512_hmac_ctx.ipad.s[2];
|
||||
tmps[gid].ipad_raw[3] = streebog512_hmac_ctx.ipad.s[3];
|
||||
tmps[gid].ipad_raw[4] = streebog512_hmac_ctx.ipad.s[4];
|
||||
tmps[gid].ipad_raw[5] = streebog512_hmac_ctx.ipad.s[5];
|
||||
tmps[gid].ipad_raw[6] = streebog512_hmac_ctx.ipad.s[6];
|
||||
tmps[gid].ipad_raw[7] = streebog512_hmac_ctx.ipad.s[7];
|
||||
|
||||
tmps[gid].opad_raw[0] = streebog512_hmac_ctx.opad.s[0];
|
||||
tmps[gid].opad_raw[1] = streebog512_hmac_ctx.opad.s[1];
|
||||
tmps[gid].opad_raw[2] = streebog512_hmac_ctx.opad.s[2];
|
||||
tmps[gid].opad_raw[3] = streebog512_hmac_ctx.opad.s[3];
|
||||
tmps[gid].opad_raw[4] = streebog512_hmac_ctx.opad.s[4];
|
||||
tmps[gid].opad_raw[5] = streebog512_hmac_ctx.opad.s[5];
|
||||
tmps[gid].opad_raw[6] = streebog512_hmac_ctx.opad.s[6];
|
||||
tmps[gid].opad_raw[7] = streebog512_hmac_ctx.opad.s[7];
|
||||
|
||||
streebog512_hmac_update_global_swap (&streebog512_hmac_ctx, esalt_bufs[digests_offset].salt_buf, 64);
|
||||
|
||||
for (u32 i = 0, j = 1; i < 8; i += 8, j += 1)
|
||||
{
|
||||
streebog512_hmac_ctx_t streebog512_hmac_ctx2 = streebog512_hmac_ctx;
|
||||
|
||||
w0[0] = j;
|
||||
w0[1] = 0;
|
||||
w0[2] = 0;
|
||||
w0[3] = 0;
|
||||
w1[0] = 0;
|
||||
w1[1] = 0;
|
||||
w1[2] = 0;
|
||||
w1[3] = 0;
|
||||
w2[0] = 0;
|
||||
w2[1] = 0;
|
||||
w2[2] = 0;
|
||||
w2[3] = 0;
|
||||
w3[0] = 0;
|
||||
w3[1] = 0;
|
||||
w3[2] = 0;
|
||||
w3[3] = 0;
|
||||
|
||||
streebog512_hmac_update_64 (&streebog512_hmac_ctx2, w0, w1, w2, w3, 4);
|
||||
|
||||
streebog512_hmac_final (&streebog512_hmac_ctx2);
|
||||
|
||||
tmps[gid].dgst[i + 0] = streebog512_hmac_ctx2.opad.h[0];
|
||||
tmps[gid].dgst[i + 1] = streebog512_hmac_ctx2.opad.h[1];
|
||||
tmps[gid].dgst[i + 2] = streebog512_hmac_ctx2.opad.h[2];
|
||||
tmps[gid].dgst[i + 3] = streebog512_hmac_ctx2.opad.h[3];
|
||||
tmps[gid].dgst[i + 4] = streebog512_hmac_ctx2.opad.h[4];
|
||||
tmps[gid].dgst[i + 5] = streebog512_hmac_ctx2.opad.h[5];
|
||||
tmps[gid].dgst[i + 6] = streebog512_hmac_ctx2.opad.h[6];
|
||||
tmps[gid].dgst[i + 7] = streebog512_hmac_ctx2.opad.h[7];
|
||||
|
||||
tmps[gid].out[i + 0] = tmps[gid].dgst[i + 0];
|
||||
tmps[gid].out[i + 1] = tmps[gid].dgst[i + 1];
|
||||
tmps[gid].out[i + 2] = tmps[gid].dgst[i + 2];
|
||||
tmps[gid].out[i + 3] = tmps[gid].dgst[i + 3];
|
||||
tmps[gid].out[i + 4] = tmps[gid].dgst[i + 4];
|
||||
tmps[gid].out[i + 5] = tmps[gid].dgst[i + 5];
|
||||
tmps[gid].out[i + 6] = tmps[gid].dgst[i + 6];
|
||||
tmps[gid].out[i + 7] = tmps[gid].dgst[i + 7];
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m13771_loop (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if ((gid * VECT_SIZE) >= gid_max) return;
|
||||
|
||||
u64x ipad_hash[8];
|
||||
u64x opad_hash[8];
|
||||
|
||||
ipad_hash[0] = pack64v (tmps, ipad_hash, gid, 0);
|
||||
ipad_hash[1] = pack64v (tmps, ipad_hash, gid, 1);
|
||||
ipad_hash[2] = pack64v (tmps, ipad_hash, gid, 2);
|
||||
ipad_hash[3] = pack64v (tmps, ipad_hash, gid, 3);
|
||||
ipad_hash[4] = pack64v (tmps, ipad_hash, gid, 4);
|
||||
ipad_hash[5] = pack64v (tmps, ipad_hash, gid, 5);
|
||||
ipad_hash[6] = pack64v (tmps, ipad_hash, gid, 6);
|
||||
ipad_hash[7] = pack64v (tmps, ipad_hash, gid, 7);
|
||||
|
||||
opad_hash[0] = pack64v (tmps, opad_hash, gid, 0);
|
||||
opad_hash[1] = pack64v (tmps, opad_hash, gid, 1);
|
||||
opad_hash[2] = pack64v (tmps, opad_hash, gid, 2);
|
||||
opad_hash[3] = pack64v (tmps, opad_hash, gid, 3);
|
||||
opad_hash[4] = pack64v (tmps, opad_hash, gid, 4);
|
||||
opad_hash[5] = pack64v (tmps, opad_hash, gid, 5);
|
||||
opad_hash[6] = pack64v (tmps, opad_hash, gid, 6);
|
||||
opad_hash[7] = pack64v (tmps, opad_hash, gid, 7);
|
||||
|
||||
u64x ipad_raw[8];
|
||||
u64x opad_raw[8];
|
||||
|
||||
ipad_raw[0] = pack64v (tmps, ipad_raw, gid, 0);
|
||||
ipad_raw[1] = pack64v (tmps, ipad_raw, gid, 1);
|
||||
ipad_raw[2] = pack64v (tmps, ipad_raw, gid, 2);
|
||||
ipad_raw[3] = pack64v (tmps, ipad_raw, gid, 3);
|
||||
ipad_raw[4] = pack64v (tmps, ipad_raw, gid, 4);
|
||||
ipad_raw[5] = pack64v (tmps, ipad_raw, gid, 5);
|
||||
ipad_raw[6] = pack64v (tmps, ipad_raw, gid, 6);
|
||||
ipad_raw[7] = pack64v (tmps, ipad_raw, gid, 7);
|
||||
|
||||
opad_raw[0] = pack64v (tmps, opad_raw, gid, 0);
|
||||
opad_raw[1] = pack64v (tmps, opad_raw, gid, 1);
|
||||
opad_raw[2] = pack64v (tmps, opad_raw, gid, 2);
|
||||
opad_raw[3] = pack64v (tmps, opad_raw, gid, 3);
|
||||
opad_raw[4] = pack64v (tmps, opad_raw, gid, 4);
|
||||
opad_raw[5] = pack64v (tmps, opad_raw, gid, 5);
|
||||
opad_raw[6] = pack64v (tmps, opad_raw, gid, 6);
|
||||
opad_raw[7] = pack64v (tmps, opad_raw, gid, 7);
|
||||
|
||||
for (u32 i = 0; i < 8; i += 8)
|
||||
{
|
||||
u64x dgst[8];
|
||||
u64x out[8];
|
||||
|
||||
dgst[0] = pack64v (tmps, dgst, gid, i + 0);
|
||||
dgst[1] = pack64v (tmps, dgst, gid, i + 1);
|
||||
dgst[2] = pack64v (tmps, dgst, gid, i + 2);
|
||||
dgst[3] = pack64v (tmps, dgst, gid, i + 3);
|
||||
dgst[4] = pack64v (tmps, dgst, gid, i + 4);
|
||||
dgst[5] = pack64v (tmps, dgst, gid, i + 5);
|
||||
dgst[6] = pack64v (tmps, dgst, gid, i + 6);
|
||||
dgst[7] = pack64v (tmps, dgst, gid, i + 7);
|
||||
|
||||
out[0] = pack64v (tmps, out, gid, i + 0);
|
||||
out[1] = pack64v (tmps, out, gid, i + 1);
|
||||
out[2] = pack64v (tmps, out, gid, i + 2);
|
||||
out[3] = pack64v (tmps, out, gid, i + 3);
|
||||
out[4] = pack64v (tmps, out, gid, i + 4);
|
||||
out[5] = pack64v (tmps, out, gid, i + 5);
|
||||
out[6] = pack64v (tmps, out, gid, i + 6);
|
||||
out[7] = pack64v (tmps, out, gid, i + 7);
|
||||
|
||||
for (u32 j = 0; j < loop_cnt; j++)
|
||||
{
|
||||
u32x w0[4];
|
||||
u32x w1[4];
|
||||
u32x w2[4];
|
||||
u32x w3[4];
|
||||
|
||||
w0[0] = h32_from_64 (dgst[0]);
|
||||
w0[1] = l32_from_64 (dgst[0]);
|
||||
w0[2] = h32_from_64 (dgst[1]);
|
||||
w0[3] = l32_from_64 (dgst[1]);
|
||||
w1[0] = h32_from_64 (dgst[2]);
|
||||
w1[1] = l32_from_64 (dgst[2]);
|
||||
w1[2] = h32_from_64 (dgst[3]);
|
||||
w1[3] = l32_from_64 (dgst[3]);
|
||||
w2[0] = h32_from_64 (dgst[4]);
|
||||
w2[1] = l32_from_64 (dgst[4]);
|
||||
w2[2] = h32_from_64 (dgst[5]);
|
||||
w2[3] = l32_from_64 (dgst[5]);
|
||||
w3[0] = h32_from_64 (dgst[6]);
|
||||
w3[1] = l32_from_64 (dgst[6]);
|
||||
w3[2] = h32_from_64 (dgst[7]);
|
||||
w3[3] = l32_from_64 (dgst[7]);
|
||||
|
||||
hmac_streebog512_run_V (w0, w1, w2, w3, ipad_hash, opad_hash, ipad_raw, opad_raw, dgst, s_sbob_sl64);
|
||||
|
||||
out[0] ^= dgst[0];
|
||||
out[1] ^= dgst[1];
|
||||
out[2] ^= dgst[2];
|
||||
out[3] ^= dgst[3];
|
||||
out[4] ^= dgst[4];
|
||||
out[5] ^= dgst[5];
|
||||
out[6] ^= dgst[6];
|
||||
out[7] ^= dgst[7];
|
||||
}
|
||||
|
||||
unpack64v (tmps, dgst, gid, i + 0, dgst[0]);
|
||||
unpack64v (tmps, dgst, gid, i + 1, dgst[1]);
|
||||
unpack64v (tmps, dgst, gid, i + 2, dgst[2]);
|
||||
unpack64v (tmps, dgst, gid, i + 3, dgst[3]);
|
||||
unpack64v (tmps, dgst, gid, i + 4, dgst[4]);
|
||||
unpack64v (tmps, dgst, gid, i + 5, dgst[5]);
|
||||
unpack64v (tmps, dgst, gid, i + 6, dgst[6]);
|
||||
unpack64v (tmps, dgst, gid, i + 7, dgst[7]);
|
||||
|
||||
unpack64v (tmps, out, gid, i + 0, out[0]);
|
||||
unpack64v (tmps, out, gid, i + 1, out[1]);
|
||||
unpack64v (tmps, out, gid, i + 2, out[2]);
|
||||
unpack64v (tmps, out, gid, i + 3, out[3]);
|
||||
unpack64v (tmps, out, gid, i + 4, out[4]);
|
||||
unpack64v (tmps, out, gid, i + 5, out[5]);
|
||||
unpack64v (tmps, out, gid, i + 6, out[6]);
|
||||
unpack64v (tmps, out, gid, i + 7, out[7]);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m13771_comp (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* aes shared
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u32 s_td0[256];
|
||||
__local u32 s_td1[256];
|
||||
__local u32 s_td2[256];
|
||||
__local u32 s_td3[256];
|
||||
__local u32 s_td4[256];
|
||||
|
||||
__local u32 s_te0[256];
|
||||
__local u32 s_te1[256];
|
||||
__local u32 s_te2[256];
|
||||
__local u32 s_te3[256];
|
||||
__local u32 s_te4[256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_td0[i] = td0[i];
|
||||
s_td1[i] = td1[i];
|
||||
s_td2[i] = td2[i];
|
||||
s_td3[i] = td3[i];
|
||||
s_td4[i] = td4[i];
|
||||
|
||||
s_te0[i] = te0[i];
|
||||
s_te1[i] = te1[i];
|
||||
s_te2[i] = te2[i];
|
||||
s_te3[i] = te3[i];
|
||||
s_te4[i] = te4[i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u32a *s_td0 = td0;
|
||||
__constant u32a *s_td1 = td1;
|
||||
__constant u32a *s_td2 = td2;
|
||||
__constant u32a *s_td3 = td3;
|
||||
__constant u32a *s_td4 = td4;
|
||||
|
||||
__constant u32a *s_te0 = te0;
|
||||
__constant u32a *s_te1 = te1;
|
||||
__constant u32a *s_te2 = te2;
|
||||
__constant u32a *s_te3 = te3;
|
||||
__constant u32a *s_te4 = te4;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
u32 ukey1[8];
|
||||
u32 ukey2[8];
|
||||
|
||||
ukey1[0] = swap32_S (h32_from_64_S (tmps[gid].out[7]));
|
||||
ukey1[1] = swap32_S (l32_from_64_S (tmps[gid].out[7]));
|
||||
ukey1[2] = swap32_S (h32_from_64_S (tmps[gid].out[6]));
|
||||
ukey1[3] = swap32_S (l32_from_64_S (tmps[gid].out[6]));
|
||||
ukey1[4] = swap32_S (h32_from_64_S (tmps[gid].out[5]));
|
||||
ukey1[5] = swap32_S (l32_from_64_S (tmps[gid].out[5]));
|
||||
ukey1[6] = swap32_S (h32_from_64_S (tmps[gid].out[4]));
|
||||
ukey1[7] = swap32_S (l32_from_64_S (tmps[gid].out[4]));
|
||||
|
||||
ukey2[0] = swap32_S (h32_from_64_S (tmps[gid].out[3]));
|
||||
ukey2[1] = swap32_S (l32_from_64_S (tmps[gid].out[3]));
|
||||
ukey2[2] = swap32_S (h32_from_64_S (tmps[gid].out[2]));
|
||||
ukey2[3] = swap32_S (l32_from_64_S (tmps[gid].out[2]));
|
||||
ukey2[4] = swap32_S (h32_from_64_S (tmps[gid].out[1]));
|
||||
ukey2[5] = swap32_S (l32_from_64_S (tmps[gid].out[1]));
|
||||
ukey2[6] = swap32_S (h32_from_64_S (tmps[gid].out[0]));
|
||||
ukey2[7] = swap32_S (l32_from_64_S (tmps[gid].out[0]));
|
||||
|
||||
if (verify_header_aes (esalt_bufs, ukey1, ukey2, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_serpent (esalt_bufs, ukey1, ukey2) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_twofish (esalt_bufs, ukey1, ukey2) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,623 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_simd.cl"
|
||||
#include "inc_hash_streebog512.cl"
|
||||
|
||||
#include "inc_cipher_aes.cl"
|
||||
#include "inc_cipher_twofish.cl"
|
||||
#include "inc_cipher_serpent.cl"
|
||||
|
||||
#include "inc_truecrypt_crc32.cl"
|
||||
#include "inc_truecrypt_xts.cl"
|
||||
|
||||
DECLSPEC u32 u8add (const u32 a, const u32 b)
|
||||
{
|
||||
const u32 a1 = (a >> 0) & 0xff;
|
||||
const u32 a2 = (a >> 8) & 0xff;
|
||||
const u32 a3 = (a >> 16) & 0xff;
|
||||
const u32 a4 = (a >> 24) & 0xff;
|
||||
|
||||
const u32 b1 = (b >> 0) & 0xff;
|
||||
const u32 b2 = (b >> 8) & 0xff;
|
||||
const u32 b3 = (b >> 16) & 0xff;
|
||||
const u32 b4 = (b >> 24) & 0xff;
|
||||
|
||||
const u32 r1 = (a1 + b1) & 0xff;
|
||||
const u32 r2 = (a2 + b2) & 0xff;
|
||||
const u32 r3 = (a3 + b3) & 0xff;
|
||||
const u32 r4 = (a4 + b4) & 0xff;
|
||||
|
||||
const u32 r = r1 << 0
|
||||
| r2 << 8
|
||||
| r3 << 16
|
||||
| r4 << 24;
|
||||
|
||||
return r;
|
||||
}
|
||||
|
||||
DECLSPEC void hmac_streebog512_run_V (u32x *w0, u32x *w1, u32x *w2, u32x *w3, u64x *ipad_hash, u64x *opad_hash, u64x *ipad_raw, u64x *opad_raw, u64x *digest, SHM_TYPE u64a (*s_sbob_sl64)[256])
|
||||
{
|
||||
const u64x nullbuf[8] = { 0 };
|
||||
u64x counterbuf[8] = { 0 };
|
||||
u64x padding[8] = { 0 };
|
||||
u64x message[8];
|
||||
|
||||
padding[7] = swap64 ((u64x) 0x01);
|
||||
|
||||
//inner HMAC: ipad + message
|
||||
|
||||
//first transform: precalculated ipad hash
|
||||
counterbuf[7] = swap64 ((u64x) 0x200);
|
||||
|
||||
//second transform: message = previous HMAC digest
|
||||
message[7] = hl32_to_64 (w3[2], w3[3]);
|
||||
message[6] = hl32_to_64 (w3[0], w3[1]);
|
||||
message[5] = hl32_to_64 (w2[2], w2[3]);
|
||||
message[4] = hl32_to_64 (w2[0], w2[1]);
|
||||
message[3] = hl32_to_64 (w1[2], w1[3]);
|
||||
message[2] = hl32_to_64 (w1[0], w1[1]);
|
||||
message[1] = hl32_to_64 (w0[2], w0[3]);
|
||||
message[0] = hl32_to_64 (w0[0], w0[1]);
|
||||
|
||||
digest[0] = ipad_hash[0];
|
||||
digest[1] = ipad_hash[1];
|
||||
digest[2] = ipad_hash[2];
|
||||
digest[3] = ipad_hash[3];
|
||||
digest[4] = ipad_hash[4];
|
||||
digest[5] = ipad_hash[5];
|
||||
digest[6] = ipad_hash[6];
|
||||
digest[7] = ipad_hash[7];
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, message, s_sbob_sl64);
|
||||
|
||||
counterbuf[7] = swap64 ((u64x) 0x400);
|
||||
|
||||
//final: padding byte
|
||||
streebog512_g_vector (digest, counterbuf, padding, s_sbob_sl64);
|
||||
|
||||
streebog512_add_vector (message, ipad_raw);
|
||||
streebog512_add_vector (message, padding);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, counterbuf, s_sbob_sl64);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, message, s_sbob_sl64);
|
||||
|
||||
//outer HMAC: opad + digest
|
||||
|
||||
//first transform: precalculated opad hash
|
||||
counterbuf[7] = swap64 ((u64x) 0x200);
|
||||
|
||||
//second transform: message = inner HMAC digest
|
||||
message[0] = digest[0];
|
||||
message[1] = digest[1];
|
||||
message[2] = digest[2];
|
||||
message[3] = digest[3];
|
||||
message[4] = digest[4];
|
||||
message[5] = digest[5];
|
||||
message[6] = digest[6];
|
||||
message[7] = digest[7];
|
||||
|
||||
digest[0] = opad_hash[0];
|
||||
digest[1] = opad_hash[1];
|
||||
digest[2] = opad_hash[2];
|
||||
digest[3] = opad_hash[3];
|
||||
digest[4] = opad_hash[4];
|
||||
digest[5] = opad_hash[5];
|
||||
digest[6] = opad_hash[6];
|
||||
digest[7] = opad_hash[7];
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, message, s_sbob_sl64);
|
||||
|
||||
counterbuf[7] = swap64 ((u64x) 0x400);
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, padding, s_sbob_sl64);
|
||||
|
||||
streebog512_add_vector (message, opad_raw);
|
||||
streebog512_add_vector (message, padding);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, counterbuf, s_sbob_sl64);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, message, s_sbob_sl64);
|
||||
}
|
||||
|
||||
__kernel void m13772_init (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
u32 w0[4];
|
||||
u32 w1[4];
|
||||
u32 w2[4];
|
||||
u32 w3[4];
|
||||
|
||||
w0[0] = pws[gid].i[ 0];
|
||||
w0[1] = pws[gid].i[ 1];
|
||||
w0[2] = pws[gid].i[ 2];
|
||||
w0[3] = pws[gid].i[ 3];
|
||||
w1[0] = pws[gid].i[ 4];
|
||||
w1[1] = pws[gid].i[ 5];
|
||||
w1[2] = pws[gid].i[ 6];
|
||||
w1[3] = pws[gid].i[ 7];
|
||||
w2[0] = pws[gid].i[ 8];
|
||||
w2[1] = pws[gid].i[ 9];
|
||||
w2[2] = pws[gid].i[10];
|
||||
w2[3] = pws[gid].i[11];
|
||||
w3[0] = pws[gid].i[12];
|
||||
w3[1] = pws[gid].i[13];
|
||||
w3[2] = pws[gid].i[14];
|
||||
w3[3] = pws[gid].i[15];
|
||||
|
||||
w0[0] = u8add (w0[0], esalt_bufs[digests_offset].keyfile_buf[ 0]);
|
||||
w0[1] = u8add (w0[1], esalt_bufs[digests_offset].keyfile_buf[ 1]);
|
||||
w0[2] = u8add (w0[2], esalt_bufs[digests_offset].keyfile_buf[ 2]);
|
||||
w0[3] = u8add (w0[3], esalt_bufs[digests_offset].keyfile_buf[ 3]);
|
||||
w1[0] = u8add (w1[0], esalt_bufs[digests_offset].keyfile_buf[ 4]);
|
||||
w1[1] = u8add (w1[1], esalt_bufs[digests_offset].keyfile_buf[ 5]);
|
||||
w1[2] = u8add (w1[2], esalt_bufs[digests_offset].keyfile_buf[ 6]);
|
||||
w1[3] = u8add (w1[3], esalt_bufs[digests_offset].keyfile_buf[ 7]);
|
||||
w2[0] = u8add (w2[0], esalt_bufs[digests_offset].keyfile_buf[ 8]);
|
||||
w2[1] = u8add (w2[1], esalt_bufs[digests_offset].keyfile_buf[ 9]);
|
||||
w2[2] = u8add (w2[2], esalt_bufs[digests_offset].keyfile_buf[10]);
|
||||
w2[3] = u8add (w2[3], esalt_bufs[digests_offset].keyfile_buf[11]);
|
||||
w3[0] = u8add (w3[0], esalt_bufs[digests_offset].keyfile_buf[12]);
|
||||
w3[1] = u8add (w3[1], esalt_bufs[digests_offset].keyfile_buf[13]);
|
||||
w3[2] = u8add (w3[2], esalt_bufs[digests_offset].keyfile_buf[14]);
|
||||
w3[3] = u8add (w3[3], esalt_bufs[digests_offset].keyfile_buf[15]);
|
||||
|
||||
w0[0] = swap32_S (w0[0]);
|
||||
w0[1] = swap32_S (w0[1]);
|
||||
w0[2] = swap32_S (w0[2]);
|
||||
w0[3] = swap32_S (w0[3]);
|
||||
w1[0] = swap32_S (w1[0]);
|
||||
w1[1] = swap32_S (w1[1]);
|
||||
w1[2] = swap32_S (w1[2]);
|
||||
w1[3] = swap32_S (w1[3]);
|
||||
w2[0] = swap32_S (w2[0]);
|
||||
w2[1] = swap32_S (w2[1]);
|
||||
w2[2] = swap32_S (w2[2]);
|
||||
w2[3] = swap32_S (w2[3]);
|
||||
w3[0] = swap32_S (w3[0]);
|
||||
w3[1] = swap32_S (w3[1]);
|
||||
w3[2] = swap32_S (w3[2]);
|
||||
w3[3] = swap32_S (w3[3]);
|
||||
|
||||
streebog512_hmac_ctx_t streebog512_hmac_ctx;
|
||||
|
||||
streebog512_hmac_init_64 (&streebog512_hmac_ctx, w0, w1, w2, w3, s_sbob_sl64);
|
||||
|
||||
tmps[gid].ipad_hash[0] = streebog512_hmac_ctx.ipad.h[0];
|
||||
tmps[gid].ipad_hash[1] = streebog512_hmac_ctx.ipad.h[1];
|
||||
tmps[gid].ipad_hash[2] = streebog512_hmac_ctx.ipad.h[2];
|
||||
tmps[gid].ipad_hash[3] = streebog512_hmac_ctx.ipad.h[3];
|
||||
tmps[gid].ipad_hash[4] = streebog512_hmac_ctx.ipad.h[4];
|
||||
tmps[gid].ipad_hash[5] = streebog512_hmac_ctx.ipad.h[5];
|
||||
tmps[gid].ipad_hash[6] = streebog512_hmac_ctx.ipad.h[6];
|
||||
tmps[gid].ipad_hash[7] = streebog512_hmac_ctx.ipad.h[7];
|
||||
|
||||
tmps[gid].opad_hash[0] = streebog512_hmac_ctx.opad.h[0];
|
||||
tmps[gid].opad_hash[1] = streebog512_hmac_ctx.opad.h[1];
|
||||
tmps[gid].opad_hash[2] = streebog512_hmac_ctx.opad.h[2];
|
||||
tmps[gid].opad_hash[3] = streebog512_hmac_ctx.opad.h[3];
|
||||
tmps[gid].opad_hash[4] = streebog512_hmac_ctx.opad.h[4];
|
||||
tmps[gid].opad_hash[5] = streebog512_hmac_ctx.opad.h[5];
|
||||
tmps[gid].opad_hash[6] = streebog512_hmac_ctx.opad.h[6];
|
||||
tmps[gid].opad_hash[7] = streebog512_hmac_ctx.opad.h[7];
|
||||
|
||||
tmps[gid].ipad_raw[0] = streebog512_hmac_ctx.ipad.s[0];
|
||||
tmps[gid].ipad_raw[1] = streebog512_hmac_ctx.ipad.s[1];
|
||||
tmps[gid].ipad_raw[2] = streebog512_hmac_ctx.ipad.s[2];
|
||||
tmps[gid].ipad_raw[3] = streebog512_hmac_ctx.ipad.s[3];
|
||||
tmps[gid].ipad_raw[4] = streebog512_hmac_ctx.ipad.s[4];
|
||||
tmps[gid].ipad_raw[5] = streebog512_hmac_ctx.ipad.s[5];
|
||||
tmps[gid].ipad_raw[6] = streebog512_hmac_ctx.ipad.s[6];
|
||||
tmps[gid].ipad_raw[7] = streebog512_hmac_ctx.ipad.s[7];
|
||||
|
||||
tmps[gid].opad_raw[0] = streebog512_hmac_ctx.opad.s[0];
|
||||
tmps[gid].opad_raw[1] = streebog512_hmac_ctx.opad.s[1];
|
||||
tmps[gid].opad_raw[2] = streebog512_hmac_ctx.opad.s[2];
|
||||
tmps[gid].opad_raw[3] = streebog512_hmac_ctx.opad.s[3];
|
||||
tmps[gid].opad_raw[4] = streebog512_hmac_ctx.opad.s[4];
|
||||
tmps[gid].opad_raw[5] = streebog512_hmac_ctx.opad.s[5];
|
||||
tmps[gid].opad_raw[6] = streebog512_hmac_ctx.opad.s[6];
|
||||
tmps[gid].opad_raw[7] = streebog512_hmac_ctx.opad.s[7];
|
||||
|
||||
streebog512_hmac_update_global_swap (&streebog512_hmac_ctx, esalt_bufs[digests_offset].salt_buf, 64);
|
||||
|
||||
for (u32 i = 0, j = 1; i < 16; i += 8, j += 1)
|
||||
{
|
||||
streebog512_hmac_ctx_t streebog512_hmac_ctx2 = streebog512_hmac_ctx;
|
||||
|
||||
w0[0] = j;
|
||||
w0[1] = 0;
|
||||
w0[2] = 0;
|
||||
w0[3] = 0;
|
||||
w1[0] = 0;
|
||||
w1[1] = 0;
|
||||
w1[2] = 0;
|
||||
w1[3] = 0;
|
||||
w2[0] = 0;
|
||||
w2[1] = 0;
|
||||
w2[2] = 0;
|
||||
w2[3] = 0;
|
||||
w3[0] = 0;
|
||||
w3[1] = 0;
|
||||
w3[2] = 0;
|
||||
w3[3] = 0;
|
||||
|
||||
streebog512_hmac_update_64 (&streebog512_hmac_ctx2, w0, w1, w2, w3, 4);
|
||||
|
||||
streebog512_hmac_final (&streebog512_hmac_ctx2);
|
||||
|
||||
tmps[gid].dgst[i + 0] = streebog512_hmac_ctx2.opad.h[0];
|
||||
tmps[gid].dgst[i + 1] = streebog512_hmac_ctx2.opad.h[1];
|
||||
tmps[gid].dgst[i + 2] = streebog512_hmac_ctx2.opad.h[2];
|
||||
tmps[gid].dgst[i + 3] = streebog512_hmac_ctx2.opad.h[3];
|
||||
tmps[gid].dgst[i + 4] = streebog512_hmac_ctx2.opad.h[4];
|
||||
tmps[gid].dgst[i + 5] = streebog512_hmac_ctx2.opad.h[5];
|
||||
tmps[gid].dgst[i + 6] = streebog512_hmac_ctx2.opad.h[6];
|
||||
tmps[gid].dgst[i + 7] = streebog512_hmac_ctx2.opad.h[7];
|
||||
|
||||
tmps[gid].out[i + 0] = tmps[gid].dgst[i + 0];
|
||||
tmps[gid].out[i + 1] = tmps[gid].dgst[i + 1];
|
||||
tmps[gid].out[i + 2] = tmps[gid].dgst[i + 2];
|
||||
tmps[gid].out[i + 3] = tmps[gid].dgst[i + 3];
|
||||
tmps[gid].out[i + 4] = tmps[gid].dgst[i + 4];
|
||||
tmps[gid].out[i + 5] = tmps[gid].dgst[i + 5];
|
||||
tmps[gid].out[i + 6] = tmps[gid].dgst[i + 6];
|
||||
tmps[gid].out[i + 7] = tmps[gid].dgst[i + 7];
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m13772_loop (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if ((gid * VECT_SIZE) >= gid_max) return;
|
||||
|
||||
u64x ipad_hash[8];
|
||||
u64x opad_hash[8];
|
||||
|
||||
ipad_hash[0] = pack64v (tmps, ipad_hash, gid, 0);
|
||||
ipad_hash[1] = pack64v (tmps, ipad_hash, gid, 1);
|
||||
ipad_hash[2] = pack64v (tmps, ipad_hash, gid, 2);
|
||||
ipad_hash[3] = pack64v (tmps, ipad_hash, gid, 3);
|
||||
ipad_hash[4] = pack64v (tmps, ipad_hash, gid, 4);
|
||||
ipad_hash[5] = pack64v (tmps, ipad_hash, gid, 5);
|
||||
ipad_hash[6] = pack64v (tmps, ipad_hash, gid, 6);
|
||||
ipad_hash[7] = pack64v (tmps, ipad_hash, gid, 7);
|
||||
|
||||
opad_hash[0] = pack64v (tmps, opad_hash, gid, 0);
|
||||
opad_hash[1] = pack64v (tmps, opad_hash, gid, 1);
|
||||
opad_hash[2] = pack64v (tmps, opad_hash, gid, 2);
|
||||
opad_hash[3] = pack64v (tmps, opad_hash, gid, 3);
|
||||
opad_hash[4] = pack64v (tmps, opad_hash, gid, 4);
|
||||
opad_hash[5] = pack64v (tmps, opad_hash, gid, 5);
|
||||
opad_hash[6] = pack64v (tmps, opad_hash, gid, 6);
|
||||
opad_hash[7] = pack64v (tmps, opad_hash, gid, 7);
|
||||
|
||||
u64x ipad_raw[8];
|
||||
u64x opad_raw[8];
|
||||
|
||||
ipad_raw[0] = pack64v (tmps, ipad_raw, gid, 0);
|
||||
ipad_raw[1] = pack64v (tmps, ipad_raw, gid, 1);
|
||||
ipad_raw[2] = pack64v (tmps, ipad_raw, gid, 2);
|
||||
ipad_raw[3] = pack64v (tmps, ipad_raw, gid, 3);
|
||||
ipad_raw[4] = pack64v (tmps, ipad_raw, gid, 4);
|
||||
ipad_raw[5] = pack64v (tmps, ipad_raw, gid, 5);
|
||||
ipad_raw[6] = pack64v (tmps, ipad_raw, gid, 6);
|
||||
ipad_raw[7] = pack64v (tmps, ipad_raw, gid, 7);
|
||||
|
||||
opad_raw[0] = pack64v (tmps, opad_raw, gid, 0);
|
||||
opad_raw[1] = pack64v (tmps, opad_raw, gid, 1);
|
||||
opad_raw[2] = pack64v (tmps, opad_raw, gid, 2);
|
||||
opad_raw[3] = pack64v (tmps, opad_raw, gid, 3);
|
||||
opad_raw[4] = pack64v (tmps, opad_raw, gid, 4);
|
||||
opad_raw[5] = pack64v (tmps, opad_raw, gid, 5);
|
||||
opad_raw[6] = pack64v (tmps, opad_raw, gid, 6);
|
||||
opad_raw[7] = pack64v (tmps, opad_raw, gid, 7);
|
||||
|
||||
for (u32 i = 0; i < 16; i += 8)
|
||||
{
|
||||
u64x dgst[8];
|
||||
u64x out[8];
|
||||
|
||||
dgst[0] = pack64v (tmps, dgst, gid, i + 0);
|
||||
dgst[1] = pack64v (tmps, dgst, gid, i + 1);
|
||||
dgst[2] = pack64v (tmps, dgst, gid, i + 2);
|
||||
dgst[3] = pack64v (tmps, dgst, gid, i + 3);
|
||||
dgst[4] = pack64v (tmps, dgst, gid, i + 4);
|
||||
dgst[5] = pack64v (tmps, dgst, gid, i + 5);
|
||||
dgst[6] = pack64v (tmps, dgst, gid, i + 6);
|
||||
dgst[7] = pack64v (tmps, dgst, gid, i + 7);
|
||||
|
||||
out[0] = pack64v (tmps, out, gid, i + 0);
|
||||
out[1] = pack64v (tmps, out, gid, i + 1);
|
||||
out[2] = pack64v (tmps, out, gid, i + 2);
|
||||
out[3] = pack64v (tmps, out, gid, i + 3);
|
||||
out[4] = pack64v (tmps, out, gid, i + 4);
|
||||
out[5] = pack64v (tmps, out, gid, i + 5);
|
||||
out[6] = pack64v (tmps, out, gid, i + 6);
|
||||
out[7] = pack64v (tmps, out, gid, i + 7);
|
||||
|
||||
for (u32 j = 0; j < loop_cnt; j++)
|
||||
{
|
||||
u32x w0[4];
|
||||
u32x w1[4];
|
||||
u32x w2[4];
|
||||
u32x w3[4];
|
||||
|
||||
w0[0] = h32_from_64 (dgst[0]);
|
||||
w0[1] = l32_from_64 (dgst[0]);
|
||||
w0[2] = h32_from_64 (dgst[1]);
|
||||
w0[3] = l32_from_64 (dgst[1]);
|
||||
w1[0] = h32_from_64 (dgst[2]);
|
||||
w1[1] = l32_from_64 (dgst[2]);
|
||||
w1[2] = h32_from_64 (dgst[3]);
|
||||
w1[3] = l32_from_64 (dgst[3]);
|
||||
w2[0] = h32_from_64 (dgst[4]);
|
||||
w2[1] = l32_from_64 (dgst[4]);
|
||||
w2[2] = h32_from_64 (dgst[5]);
|
||||
w2[3] = l32_from_64 (dgst[5]);
|
||||
w3[0] = h32_from_64 (dgst[6]);
|
||||
w3[1] = l32_from_64 (dgst[6]);
|
||||
w3[2] = h32_from_64 (dgst[7]);
|
||||
w3[3] = l32_from_64 (dgst[7]);
|
||||
|
||||
hmac_streebog512_run_V (w0, w1, w2, w3, ipad_hash, opad_hash, ipad_raw, opad_raw, dgst, s_sbob_sl64);
|
||||
|
||||
out[0] ^= dgst[0];
|
||||
out[1] ^= dgst[1];
|
||||
out[2] ^= dgst[2];
|
||||
out[3] ^= dgst[3];
|
||||
out[4] ^= dgst[4];
|
||||
out[5] ^= dgst[5];
|
||||
out[6] ^= dgst[6];
|
||||
out[7] ^= dgst[7];
|
||||
}
|
||||
|
||||
unpack64v (tmps, dgst, gid, i + 0, dgst[0]);
|
||||
unpack64v (tmps, dgst, gid, i + 1, dgst[1]);
|
||||
unpack64v (tmps, dgst, gid, i + 2, dgst[2]);
|
||||
unpack64v (tmps, dgst, gid, i + 3, dgst[3]);
|
||||
unpack64v (tmps, dgst, gid, i + 4, dgst[4]);
|
||||
unpack64v (tmps, dgst, gid, i + 5, dgst[5]);
|
||||
unpack64v (tmps, dgst, gid, i + 6, dgst[6]);
|
||||
unpack64v (tmps, dgst, gid, i + 7, dgst[7]);
|
||||
|
||||
unpack64v (tmps, out, gid, i + 0, out[0]);
|
||||
unpack64v (tmps, out, gid, i + 1, out[1]);
|
||||
unpack64v (tmps, out, gid, i + 2, out[2]);
|
||||
unpack64v (tmps, out, gid, i + 3, out[3]);
|
||||
unpack64v (tmps, out, gid, i + 4, out[4]);
|
||||
unpack64v (tmps, out, gid, i + 5, out[5]);
|
||||
unpack64v (tmps, out, gid, i + 6, out[6]);
|
||||
unpack64v (tmps, out, gid, i + 7, out[7]);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m13772_comp (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* aes shared
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u32 s_td0[256];
|
||||
__local u32 s_td1[256];
|
||||
__local u32 s_td2[256];
|
||||
__local u32 s_td3[256];
|
||||
__local u32 s_td4[256];
|
||||
|
||||
__local u32 s_te0[256];
|
||||
__local u32 s_te1[256];
|
||||
__local u32 s_te2[256];
|
||||
__local u32 s_te3[256];
|
||||
__local u32 s_te4[256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_td0[i] = td0[i];
|
||||
s_td1[i] = td1[i];
|
||||
s_td2[i] = td2[i];
|
||||
s_td3[i] = td3[i];
|
||||
s_td4[i] = td4[i];
|
||||
|
||||
s_te0[i] = te0[i];
|
||||
s_te1[i] = te1[i];
|
||||
s_te2[i] = te2[i];
|
||||
s_te3[i] = te3[i];
|
||||
s_te4[i] = te4[i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u32a *s_td0 = td0;
|
||||
__constant u32a *s_td1 = td1;
|
||||
__constant u32a *s_td2 = td2;
|
||||
__constant u32a *s_td3 = td3;
|
||||
__constant u32a *s_td4 = td4;
|
||||
|
||||
__constant u32a *s_te0 = te0;
|
||||
__constant u32a *s_te1 = te1;
|
||||
__constant u32a *s_te2 = te2;
|
||||
__constant u32a *s_te3 = te3;
|
||||
__constant u32a *s_te4 = te4;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
u32 ukey1[8];
|
||||
|
||||
ukey1[0] = swap32_S (h32_from_64_S (tmps[gid].out[7]));
|
||||
ukey1[1] = swap32_S (l32_from_64_S (tmps[gid].out[7]));
|
||||
ukey1[2] = swap32_S (h32_from_64_S (tmps[gid].out[6]));
|
||||
ukey1[3] = swap32_S (l32_from_64_S (tmps[gid].out[6]));
|
||||
ukey1[4] = swap32_S (h32_from_64_S (tmps[gid].out[5]));
|
||||
ukey1[5] = swap32_S (l32_from_64_S (tmps[gid].out[5]));
|
||||
ukey1[6] = swap32_S (h32_from_64_S (tmps[gid].out[4]));
|
||||
ukey1[7] = swap32_S (l32_from_64_S (tmps[gid].out[4]));
|
||||
|
||||
u32 ukey2[8];
|
||||
|
||||
ukey2[0] = swap32_S (h32_from_64_S (tmps[gid].out[3]));
|
||||
ukey2[1] = swap32_S (l32_from_64_S (tmps[gid].out[3]));
|
||||
ukey2[2] = swap32_S (h32_from_64_S (tmps[gid].out[2]));
|
||||
ukey2[3] = swap32_S (l32_from_64_S (tmps[gid].out[2]));
|
||||
ukey2[4] = swap32_S (h32_from_64_S (tmps[gid].out[1]));
|
||||
ukey2[5] = swap32_S (l32_from_64_S (tmps[gid].out[1]));
|
||||
ukey2[6] = swap32_S (h32_from_64_S (tmps[gid].out[0]));
|
||||
ukey2[7] = swap32_S (l32_from_64_S (tmps[gid].out[0]));
|
||||
|
||||
if (verify_header_aes (esalt_bufs, ukey1, ukey2, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_serpent (esalt_bufs, ukey1, ukey2) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_twofish (esalt_bufs, ukey1, ukey2) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
u32 ukey3[8];
|
||||
|
||||
ukey3[0] = swap32_S (h32_from_64_S (tmps[gid].out[15]));
|
||||
ukey3[1] = swap32_S (l32_from_64_S (tmps[gid].out[15]));
|
||||
ukey3[2] = swap32_S (h32_from_64_S (tmps[gid].out[14]));
|
||||
ukey3[3] = swap32_S (l32_from_64_S (tmps[gid].out[14]));
|
||||
ukey3[4] = swap32_S (h32_from_64_S (tmps[gid].out[13]));
|
||||
ukey3[5] = swap32_S (l32_from_64_S (tmps[gid].out[13]));
|
||||
ukey3[6] = swap32_S (h32_from_64_S (tmps[gid].out[12]));
|
||||
ukey3[7] = swap32_S (l32_from_64_S (tmps[gid].out[12]));
|
||||
|
||||
u32 ukey4[8];
|
||||
|
||||
ukey4[0] = swap32_S (h32_from_64_S (tmps[gid].out[11]));
|
||||
ukey4[1] = swap32_S (l32_from_64_S (tmps[gid].out[11]));
|
||||
ukey4[2] = swap32_S (h32_from_64_S (tmps[gid].out[10]));
|
||||
ukey4[3] = swap32_S (l32_from_64_S (tmps[gid].out[10]));
|
||||
ukey4[4] = swap32_S (h32_from_64_S (tmps[gid].out[ 9]));
|
||||
ukey4[5] = swap32_S (l32_from_64_S (tmps[gid].out[ 9]));
|
||||
ukey4[6] = swap32_S (h32_from_64_S (tmps[gid].out[ 8]));
|
||||
ukey4[7] = swap32_S (l32_from_64_S (tmps[gid].out[ 8]));
|
||||
|
||||
if (verify_header_aes_twofish (esalt_bufs, ukey1, ukey2, ukey3, ukey4, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_serpent_aes (esalt_bufs, ukey1, ukey2, ukey3, ukey4, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_twofish_serpent (esalt_bufs, ukey1, ukey2, ukey3, ukey4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
}
|
||||
@ -0,0 +1,661 @@
|
||||
/**
|
||||
* Author......: See docs/credits.txt
|
||||
* License.....: MIT
|
||||
*/
|
||||
|
||||
//#define NEW_SIMD_CODE
|
||||
|
||||
#include "inc_vendor.cl"
|
||||
#include "inc_hash_constants.h"
|
||||
#include "inc_hash_functions.cl"
|
||||
#include "inc_types.cl"
|
||||
#include "inc_common.cl"
|
||||
#include "inc_simd.cl"
|
||||
#include "inc_hash_streebog512.cl"
|
||||
|
||||
#include "inc_cipher_aes.cl"
|
||||
#include "inc_cipher_twofish.cl"
|
||||
#include "inc_cipher_serpent.cl"
|
||||
|
||||
#include "inc_truecrypt_crc32.cl"
|
||||
#include "inc_truecrypt_xts.cl"
|
||||
|
||||
DECLSPEC u32 u8add (const u32 a, const u32 b)
|
||||
{
|
||||
const u32 a1 = (a >> 0) & 0xff;
|
||||
const u32 a2 = (a >> 8) & 0xff;
|
||||
const u32 a3 = (a >> 16) & 0xff;
|
||||
const u32 a4 = (a >> 24) & 0xff;
|
||||
|
||||
const u32 b1 = (b >> 0) & 0xff;
|
||||
const u32 b2 = (b >> 8) & 0xff;
|
||||
const u32 b3 = (b >> 16) & 0xff;
|
||||
const u32 b4 = (b >> 24) & 0xff;
|
||||
|
||||
const u32 r1 = (a1 + b1) & 0xff;
|
||||
const u32 r2 = (a2 + b2) & 0xff;
|
||||
const u32 r3 = (a3 + b3) & 0xff;
|
||||
const u32 r4 = (a4 + b4) & 0xff;
|
||||
|
||||
const u32 r = r1 << 0
|
||||
| r2 << 8
|
||||
| r3 << 16
|
||||
| r4 << 24;
|
||||
|
||||
return r;
|
||||
}
|
||||
|
||||
DECLSPEC void hmac_streebog512_run_V (u32x *w0, u32x *w1, u32x *w2, u32x *w3, u64x *ipad_hash, u64x *opad_hash, u64x *ipad_raw, u64x *opad_raw, u64x *digest, SHM_TYPE u64a (*s_sbob_sl64)[256])
|
||||
{
|
||||
const u64x nullbuf[8] = { 0 };
|
||||
u64x counterbuf[8] = { 0 };
|
||||
u64x padding[8] = { 0 };
|
||||
u64x message[8];
|
||||
|
||||
padding[7] = swap64 ((u64x) 0x01);
|
||||
|
||||
//inner HMAC: ipad + message
|
||||
|
||||
//first transform: precalculated ipad hash
|
||||
counterbuf[7] = swap64 ((u64x) 0x200);
|
||||
|
||||
//second transform: message = previous HMAC digest
|
||||
message[7] = hl32_to_64 (w3[2], w3[3]);
|
||||
message[6] = hl32_to_64 (w3[0], w3[1]);
|
||||
message[5] = hl32_to_64 (w2[2], w2[3]);
|
||||
message[4] = hl32_to_64 (w2[0], w2[1]);
|
||||
message[3] = hl32_to_64 (w1[2], w1[3]);
|
||||
message[2] = hl32_to_64 (w1[0], w1[1]);
|
||||
message[1] = hl32_to_64 (w0[2], w0[3]);
|
||||
message[0] = hl32_to_64 (w0[0], w0[1]);
|
||||
|
||||
digest[0] = ipad_hash[0];
|
||||
digest[1] = ipad_hash[1];
|
||||
digest[2] = ipad_hash[2];
|
||||
digest[3] = ipad_hash[3];
|
||||
digest[4] = ipad_hash[4];
|
||||
digest[5] = ipad_hash[5];
|
||||
digest[6] = ipad_hash[6];
|
||||
digest[7] = ipad_hash[7];
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, message, s_sbob_sl64);
|
||||
|
||||
counterbuf[7] = swap64 ((u64x) 0x400);
|
||||
|
||||
//final: padding byte
|
||||
streebog512_g_vector (digest, counterbuf, padding, s_sbob_sl64);
|
||||
|
||||
streebog512_add_vector (message, ipad_raw);
|
||||
streebog512_add_vector (message, padding);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, counterbuf, s_sbob_sl64);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, message, s_sbob_sl64);
|
||||
|
||||
//outer HMAC: opad + digest
|
||||
|
||||
//first transform: precalculated opad hash
|
||||
counterbuf[7] = swap64 ((u64x) 0x200);
|
||||
|
||||
//second transform: message = inner HMAC digest
|
||||
message[0] = digest[0];
|
||||
message[1] = digest[1];
|
||||
message[2] = digest[2];
|
||||
message[3] = digest[3];
|
||||
message[4] = digest[4];
|
||||
message[5] = digest[5];
|
||||
message[6] = digest[6];
|
||||
message[7] = digest[7];
|
||||
|
||||
digest[0] = opad_hash[0];
|
||||
digest[1] = opad_hash[1];
|
||||
digest[2] = opad_hash[2];
|
||||
digest[3] = opad_hash[3];
|
||||
digest[4] = opad_hash[4];
|
||||
digest[5] = opad_hash[5];
|
||||
digest[6] = opad_hash[6];
|
||||
digest[7] = opad_hash[7];
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, message, s_sbob_sl64);
|
||||
|
||||
counterbuf[7] = swap64 ((u64x) 0x400);
|
||||
|
||||
streebog512_g_vector (digest, counterbuf, padding, s_sbob_sl64);
|
||||
|
||||
streebog512_add_vector (message, opad_raw);
|
||||
streebog512_add_vector (message, padding);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, counterbuf, s_sbob_sl64);
|
||||
|
||||
streebog512_g_vector (digest, nullbuf, message, s_sbob_sl64);
|
||||
}
|
||||
|
||||
__kernel void m13773_init (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
/**
|
||||
* base
|
||||
*/
|
||||
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
u32 w0[4];
|
||||
u32 w1[4];
|
||||
u32 w2[4];
|
||||
u32 w3[4];
|
||||
|
||||
w0[0] = pws[gid].i[ 0];
|
||||
w0[1] = pws[gid].i[ 1];
|
||||
w0[2] = pws[gid].i[ 2];
|
||||
w0[3] = pws[gid].i[ 3];
|
||||
w1[0] = pws[gid].i[ 4];
|
||||
w1[1] = pws[gid].i[ 5];
|
||||
w1[2] = pws[gid].i[ 6];
|
||||
w1[3] = pws[gid].i[ 7];
|
||||
w2[0] = pws[gid].i[ 8];
|
||||
w2[1] = pws[gid].i[ 9];
|
||||
w2[2] = pws[gid].i[10];
|
||||
w2[3] = pws[gid].i[11];
|
||||
w3[0] = pws[gid].i[12];
|
||||
w3[1] = pws[gid].i[13];
|
||||
w3[2] = pws[gid].i[14];
|
||||
w3[3] = pws[gid].i[15];
|
||||
|
||||
w0[0] = u8add (w0[0], esalt_bufs[digests_offset].keyfile_buf[ 0]);
|
||||
w0[1] = u8add (w0[1], esalt_bufs[digests_offset].keyfile_buf[ 1]);
|
||||
w0[2] = u8add (w0[2], esalt_bufs[digests_offset].keyfile_buf[ 2]);
|
||||
w0[3] = u8add (w0[3], esalt_bufs[digests_offset].keyfile_buf[ 3]);
|
||||
w1[0] = u8add (w1[0], esalt_bufs[digests_offset].keyfile_buf[ 4]);
|
||||
w1[1] = u8add (w1[1], esalt_bufs[digests_offset].keyfile_buf[ 5]);
|
||||
w1[2] = u8add (w1[2], esalt_bufs[digests_offset].keyfile_buf[ 6]);
|
||||
w1[3] = u8add (w1[3], esalt_bufs[digests_offset].keyfile_buf[ 7]);
|
||||
w2[0] = u8add (w2[0], esalt_bufs[digests_offset].keyfile_buf[ 8]);
|
||||
w2[1] = u8add (w2[1], esalt_bufs[digests_offset].keyfile_buf[ 9]);
|
||||
w2[2] = u8add (w2[2], esalt_bufs[digests_offset].keyfile_buf[10]);
|
||||
w2[3] = u8add (w2[3], esalt_bufs[digests_offset].keyfile_buf[11]);
|
||||
w3[0] = u8add (w3[0], esalt_bufs[digests_offset].keyfile_buf[12]);
|
||||
w3[1] = u8add (w3[1], esalt_bufs[digests_offset].keyfile_buf[13]);
|
||||
w3[2] = u8add (w3[2], esalt_bufs[digests_offset].keyfile_buf[14]);
|
||||
w3[3] = u8add (w3[3], esalt_bufs[digests_offset].keyfile_buf[15]);
|
||||
|
||||
w0[0] = swap32_S (w0[0]);
|
||||
w0[1] = swap32_S (w0[1]);
|
||||
w0[2] = swap32_S (w0[2]);
|
||||
w0[3] = swap32_S (w0[3]);
|
||||
w1[0] = swap32_S (w1[0]);
|
||||
w1[1] = swap32_S (w1[1]);
|
||||
w1[2] = swap32_S (w1[2]);
|
||||
w1[3] = swap32_S (w1[3]);
|
||||
w2[0] = swap32_S (w2[0]);
|
||||
w2[1] = swap32_S (w2[1]);
|
||||
w2[2] = swap32_S (w2[2]);
|
||||
w2[3] = swap32_S (w2[3]);
|
||||
w3[0] = swap32_S (w3[0]);
|
||||
w3[1] = swap32_S (w3[1]);
|
||||
w3[2] = swap32_S (w3[2]);
|
||||
w3[3] = swap32_S (w3[3]);
|
||||
|
||||
streebog512_hmac_ctx_t streebog512_hmac_ctx;
|
||||
|
||||
streebog512_hmac_init_64 (&streebog512_hmac_ctx, w0, w1, w2, w3, s_sbob_sl64);
|
||||
|
||||
tmps[gid].ipad_hash[0] = streebog512_hmac_ctx.ipad.h[0];
|
||||
tmps[gid].ipad_hash[1] = streebog512_hmac_ctx.ipad.h[1];
|
||||
tmps[gid].ipad_hash[2] = streebog512_hmac_ctx.ipad.h[2];
|
||||
tmps[gid].ipad_hash[3] = streebog512_hmac_ctx.ipad.h[3];
|
||||
tmps[gid].ipad_hash[4] = streebog512_hmac_ctx.ipad.h[4];
|
||||
tmps[gid].ipad_hash[5] = streebog512_hmac_ctx.ipad.h[5];
|
||||
tmps[gid].ipad_hash[6] = streebog512_hmac_ctx.ipad.h[6];
|
||||
tmps[gid].ipad_hash[7] = streebog512_hmac_ctx.ipad.h[7];
|
||||
|
||||
tmps[gid].opad_hash[0] = streebog512_hmac_ctx.opad.h[0];
|
||||
tmps[gid].opad_hash[1] = streebog512_hmac_ctx.opad.h[1];
|
||||
tmps[gid].opad_hash[2] = streebog512_hmac_ctx.opad.h[2];
|
||||
tmps[gid].opad_hash[3] = streebog512_hmac_ctx.opad.h[3];
|
||||
tmps[gid].opad_hash[4] = streebog512_hmac_ctx.opad.h[4];
|
||||
tmps[gid].opad_hash[5] = streebog512_hmac_ctx.opad.h[5];
|
||||
tmps[gid].opad_hash[6] = streebog512_hmac_ctx.opad.h[6];
|
||||
tmps[gid].opad_hash[7] = streebog512_hmac_ctx.opad.h[7];
|
||||
|
||||
tmps[gid].ipad_raw[0] = streebog512_hmac_ctx.ipad.s[0];
|
||||
tmps[gid].ipad_raw[1] = streebog512_hmac_ctx.ipad.s[1];
|
||||
tmps[gid].ipad_raw[2] = streebog512_hmac_ctx.ipad.s[2];
|
||||
tmps[gid].ipad_raw[3] = streebog512_hmac_ctx.ipad.s[3];
|
||||
tmps[gid].ipad_raw[4] = streebog512_hmac_ctx.ipad.s[4];
|
||||
tmps[gid].ipad_raw[5] = streebog512_hmac_ctx.ipad.s[5];
|
||||
tmps[gid].ipad_raw[6] = streebog512_hmac_ctx.ipad.s[6];
|
||||
tmps[gid].ipad_raw[7] = streebog512_hmac_ctx.ipad.s[7];
|
||||
|
||||
tmps[gid].opad_raw[0] = streebog512_hmac_ctx.opad.s[0];
|
||||
tmps[gid].opad_raw[1] = streebog512_hmac_ctx.opad.s[1];
|
||||
tmps[gid].opad_raw[2] = streebog512_hmac_ctx.opad.s[2];
|
||||
tmps[gid].opad_raw[3] = streebog512_hmac_ctx.opad.s[3];
|
||||
tmps[gid].opad_raw[4] = streebog512_hmac_ctx.opad.s[4];
|
||||
tmps[gid].opad_raw[5] = streebog512_hmac_ctx.opad.s[5];
|
||||
tmps[gid].opad_raw[6] = streebog512_hmac_ctx.opad.s[6];
|
||||
tmps[gid].opad_raw[7] = streebog512_hmac_ctx.opad.s[7];
|
||||
|
||||
streebog512_hmac_update_global_swap (&streebog512_hmac_ctx, esalt_bufs[digests_offset].salt_buf, 64);
|
||||
|
||||
for (u32 i = 0, j = 1; i < 24; i += 8, j += 1)
|
||||
{
|
||||
streebog512_hmac_ctx_t streebog512_hmac_ctx2 = streebog512_hmac_ctx;
|
||||
|
||||
w0[0] = j;
|
||||
w0[1] = 0;
|
||||
w0[2] = 0;
|
||||
w0[3] = 0;
|
||||
w1[0] = 0;
|
||||
w1[1] = 0;
|
||||
w1[2] = 0;
|
||||
w1[3] = 0;
|
||||
w2[0] = 0;
|
||||
w2[1] = 0;
|
||||
w2[2] = 0;
|
||||
w2[3] = 0;
|
||||
w3[0] = 0;
|
||||
w3[1] = 0;
|
||||
w3[2] = 0;
|
||||
w3[3] = 0;
|
||||
|
||||
streebog512_hmac_update_64 (&streebog512_hmac_ctx2, w0, w1, w2, w3, 4);
|
||||
|
||||
streebog512_hmac_final (&streebog512_hmac_ctx2);
|
||||
|
||||
tmps[gid].dgst[i + 0] = streebog512_hmac_ctx2.opad.h[0];
|
||||
tmps[gid].dgst[i + 1] = streebog512_hmac_ctx2.opad.h[1];
|
||||
tmps[gid].dgst[i + 2] = streebog512_hmac_ctx2.opad.h[2];
|
||||
tmps[gid].dgst[i + 3] = streebog512_hmac_ctx2.opad.h[3];
|
||||
tmps[gid].dgst[i + 4] = streebog512_hmac_ctx2.opad.h[4];
|
||||
tmps[gid].dgst[i + 5] = streebog512_hmac_ctx2.opad.h[5];
|
||||
tmps[gid].dgst[i + 6] = streebog512_hmac_ctx2.opad.h[6];
|
||||
tmps[gid].dgst[i + 7] = streebog512_hmac_ctx2.opad.h[7];
|
||||
|
||||
tmps[gid].out[i + 0] = tmps[gid].dgst[i + 0];
|
||||
tmps[gid].out[i + 1] = tmps[gid].dgst[i + 1];
|
||||
tmps[gid].out[i + 2] = tmps[gid].dgst[i + 2];
|
||||
tmps[gid].out[i + 3] = tmps[gid].dgst[i + 3];
|
||||
tmps[gid].out[i + 4] = tmps[gid].dgst[i + 4];
|
||||
tmps[gid].out[i + 5] = tmps[gid].dgst[i + 5];
|
||||
tmps[gid].out[i + 6] = tmps[gid].dgst[i + 6];
|
||||
tmps[gid].out[i + 7] = tmps[gid].dgst[i + 7];
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m13773_loop (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* shared lookup table
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u64a s_sbob_sl64[8][256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_sbob_sl64[0][i] = sbob_sl64[0][i];
|
||||
s_sbob_sl64[1][i] = sbob_sl64[1][i];
|
||||
s_sbob_sl64[2][i] = sbob_sl64[2][i];
|
||||
s_sbob_sl64[3][i] = sbob_sl64[3][i];
|
||||
s_sbob_sl64[4][i] = sbob_sl64[4][i];
|
||||
s_sbob_sl64[5][i] = sbob_sl64[5][i];
|
||||
s_sbob_sl64[6][i] = sbob_sl64[6][i];
|
||||
s_sbob_sl64[7][i] = sbob_sl64[7][i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u64a (*s_sbob_sl64)[256] = sbob_sl64;
|
||||
|
||||
#endif
|
||||
|
||||
if ((gid * VECT_SIZE) >= gid_max) return;
|
||||
|
||||
u64x ipad_hash[8];
|
||||
u64x opad_hash[8];
|
||||
|
||||
ipad_hash[0] = pack64v (tmps, ipad_hash, gid, 0);
|
||||
ipad_hash[1] = pack64v (tmps, ipad_hash, gid, 1);
|
||||
ipad_hash[2] = pack64v (tmps, ipad_hash, gid, 2);
|
||||
ipad_hash[3] = pack64v (tmps, ipad_hash, gid, 3);
|
||||
ipad_hash[4] = pack64v (tmps, ipad_hash, gid, 4);
|
||||
ipad_hash[5] = pack64v (tmps, ipad_hash, gid, 5);
|
||||
ipad_hash[6] = pack64v (tmps, ipad_hash, gid, 6);
|
||||
ipad_hash[7] = pack64v (tmps, ipad_hash, gid, 7);
|
||||
|
||||
opad_hash[0] = pack64v (tmps, opad_hash, gid, 0);
|
||||
opad_hash[1] = pack64v (tmps, opad_hash, gid, 1);
|
||||
opad_hash[2] = pack64v (tmps, opad_hash, gid, 2);
|
||||
opad_hash[3] = pack64v (tmps, opad_hash, gid, 3);
|
||||
opad_hash[4] = pack64v (tmps, opad_hash, gid, 4);
|
||||
opad_hash[5] = pack64v (tmps, opad_hash, gid, 5);
|
||||
opad_hash[6] = pack64v (tmps, opad_hash, gid, 6);
|
||||
opad_hash[7] = pack64v (tmps, opad_hash, gid, 7);
|
||||
|
||||
u64x ipad_raw[8];
|
||||
u64x opad_raw[8];
|
||||
|
||||
ipad_raw[0] = pack64v (tmps, ipad_raw, gid, 0);
|
||||
ipad_raw[1] = pack64v (tmps, ipad_raw, gid, 1);
|
||||
ipad_raw[2] = pack64v (tmps, ipad_raw, gid, 2);
|
||||
ipad_raw[3] = pack64v (tmps, ipad_raw, gid, 3);
|
||||
ipad_raw[4] = pack64v (tmps, ipad_raw, gid, 4);
|
||||
ipad_raw[5] = pack64v (tmps, ipad_raw, gid, 5);
|
||||
ipad_raw[6] = pack64v (tmps, ipad_raw, gid, 6);
|
||||
ipad_raw[7] = pack64v (tmps, ipad_raw, gid, 7);
|
||||
|
||||
opad_raw[0] = pack64v (tmps, opad_raw, gid, 0);
|
||||
opad_raw[1] = pack64v (tmps, opad_raw, gid, 1);
|
||||
opad_raw[2] = pack64v (tmps, opad_raw, gid, 2);
|
||||
opad_raw[3] = pack64v (tmps, opad_raw, gid, 3);
|
||||
opad_raw[4] = pack64v (tmps, opad_raw, gid, 4);
|
||||
opad_raw[5] = pack64v (tmps, opad_raw, gid, 5);
|
||||
opad_raw[6] = pack64v (tmps, opad_raw, gid, 6);
|
||||
opad_raw[7] = pack64v (tmps, opad_raw, gid, 7);
|
||||
|
||||
for (u32 i = 0; i < 24; i += 8)
|
||||
{
|
||||
u64x dgst[8];
|
||||
u64x out[8];
|
||||
|
||||
dgst[0] = pack64v (tmps, dgst, gid, i + 0);
|
||||
dgst[1] = pack64v (tmps, dgst, gid, i + 1);
|
||||
dgst[2] = pack64v (tmps, dgst, gid, i + 2);
|
||||
dgst[3] = pack64v (tmps, dgst, gid, i + 3);
|
||||
dgst[4] = pack64v (tmps, dgst, gid, i + 4);
|
||||
dgst[5] = pack64v (tmps, dgst, gid, i + 5);
|
||||
dgst[6] = pack64v (tmps, dgst, gid, i + 6);
|
||||
dgst[7] = pack64v (tmps, dgst, gid, i + 7);
|
||||
|
||||
out[0] = pack64v (tmps, out, gid, i + 0);
|
||||
out[1] = pack64v (tmps, out, gid, i + 1);
|
||||
out[2] = pack64v (tmps, out, gid, i + 2);
|
||||
out[3] = pack64v (tmps, out, gid, i + 3);
|
||||
out[4] = pack64v (tmps, out, gid, i + 4);
|
||||
out[5] = pack64v (tmps, out, gid, i + 5);
|
||||
out[6] = pack64v (tmps, out, gid, i + 6);
|
||||
out[7] = pack64v (tmps, out, gid, i + 7);
|
||||
|
||||
for (u32 j = 0; j < loop_cnt; j++)
|
||||
{
|
||||
u32x w0[4];
|
||||
u32x w1[4];
|
||||
u32x w2[4];
|
||||
u32x w3[4];
|
||||
|
||||
w0[0] = h32_from_64 (dgst[0]);
|
||||
w0[1] = l32_from_64 (dgst[0]);
|
||||
w0[2] = h32_from_64 (dgst[1]);
|
||||
w0[3] = l32_from_64 (dgst[1]);
|
||||
w1[0] = h32_from_64 (dgst[2]);
|
||||
w1[1] = l32_from_64 (dgst[2]);
|
||||
w1[2] = h32_from_64 (dgst[3]);
|
||||
w1[3] = l32_from_64 (dgst[3]);
|
||||
w2[0] = h32_from_64 (dgst[4]);
|
||||
w2[1] = l32_from_64 (dgst[4]);
|
||||
w2[2] = h32_from_64 (dgst[5]);
|
||||
w2[3] = l32_from_64 (dgst[5]);
|
||||
w3[0] = h32_from_64 (dgst[6]);
|
||||
w3[1] = l32_from_64 (dgst[6]);
|
||||
w3[2] = h32_from_64 (dgst[7]);
|
||||
w3[3] = l32_from_64 (dgst[7]);
|
||||
|
||||
hmac_streebog512_run_V (w0, w1, w2, w3, ipad_hash, opad_hash, ipad_raw, opad_raw, dgst, s_sbob_sl64);
|
||||
|
||||
out[0] ^= dgst[0];
|
||||
out[1] ^= dgst[1];
|
||||
out[2] ^= dgst[2];
|
||||
out[3] ^= dgst[3];
|
||||
out[4] ^= dgst[4];
|
||||
out[5] ^= dgst[5];
|
||||
out[6] ^= dgst[6];
|
||||
out[7] ^= dgst[7];
|
||||
}
|
||||
|
||||
unpack64v (tmps, dgst, gid, i + 0, dgst[0]);
|
||||
unpack64v (tmps, dgst, gid, i + 1, dgst[1]);
|
||||
unpack64v (tmps, dgst, gid, i + 2, dgst[2]);
|
||||
unpack64v (tmps, dgst, gid, i + 3, dgst[3]);
|
||||
unpack64v (tmps, dgst, gid, i + 4, dgst[4]);
|
||||
unpack64v (tmps, dgst, gid, i + 5, dgst[5]);
|
||||
unpack64v (tmps, dgst, gid, i + 6, dgst[6]);
|
||||
unpack64v (tmps, dgst, gid, i + 7, dgst[7]);
|
||||
|
||||
unpack64v (tmps, out, gid, i + 0, out[0]);
|
||||
unpack64v (tmps, out, gid, i + 1, out[1]);
|
||||
unpack64v (tmps, out, gid, i + 2, out[2]);
|
||||
unpack64v (tmps, out, gid, i + 3, out[3]);
|
||||
unpack64v (tmps, out, gid, i + 4, out[4]);
|
||||
unpack64v (tmps, out, gid, i + 5, out[5]);
|
||||
unpack64v (tmps, out, gid, i + 6, out[6]);
|
||||
unpack64v (tmps, out, gid, i + 7, out[7]);
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void m13773_comp (__global pw_t *pws, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, __global vc64_sbog_tmp_t *tmps, __global void *hooks, __global const u32 *bitmaps_buf_s1_a, __global const u32 *bitmaps_buf_s1_b, __global const u32 *bitmaps_buf_s1_c, __global const u32 *bitmaps_buf_s1_d, __global const u32 *bitmaps_buf_s2_a, __global const u32 *bitmaps_buf_s2_b, __global const u32 *bitmaps_buf_s2_c, __global const u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global const digest_t *digests_buf, __global u32 *hashes_shown, __global const salt_t *salt_bufs, __global const tc_t *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u64 gid_max)
|
||||
{
|
||||
const u64 gid = get_global_id (0);
|
||||
const u64 lid = get_local_id (0);
|
||||
const u64 lsz = get_local_size (0);
|
||||
|
||||
/**
|
||||
* aes shared
|
||||
*/
|
||||
|
||||
#ifdef REAL_SHM
|
||||
|
||||
__local u32 s_td0[256];
|
||||
__local u32 s_td1[256];
|
||||
__local u32 s_td2[256];
|
||||
__local u32 s_td3[256];
|
||||
__local u32 s_td4[256];
|
||||
|
||||
__local u32 s_te0[256];
|
||||
__local u32 s_te1[256];
|
||||
__local u32 s_te2[256];
|
||||
__local u32 s_te3[256];
|
||||
__local u32 s_te4[256];
|
||||
|
||||
for (MAYBE_VOLATILE u32 i = lid; i < 256; i += lsz)
|
||||
{
|
||||
s_td0[i] = td0[i];
|
||||
s_td1[i] = td1[i];
|
||||
s_td2[i] = td2[i];
|
||||
s_td3[i] = td3[i];
|
||||
s_td4[i] = td4[i];
|
||||
|
||||
s_te0[i] = te0[i];
|
||||
s_te1[i] = te1[i];
|
||||
s_te2[i] = te2[i];
|
||||
s_te3[i] = te3[i];
|
||||
s_te4[i] = te4[i];
|
||||
}
|
||||
|
||||
barrier (CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
#else
|
||||
|
||||
__constant u32a *s_td0 = td0;
|
||||
__constant u32a *s_td1 = td1;
|
||||
__constant u32a *s_td2 = td2;
|
||||
__constant u32a *s_td3 = td3;
|
||||
__constant u32a *s_td4 = td4;
|
||||
|
||||
__constant u32a *s_te0 = te0;
|
||||
__constant u32a *s_te1 = te1;
|
||||
__constant u32a *s_te2 = te2;
|
||||
__constant u32a *s_te3 = te3;
|
||||
__constant u32a *s_te4 = te4;
|
||||
|
||||
#endif
|
||||
|
||||
if (gid >= gid_max) return;
|
||||
|
||||
u32 ukey1[8];
|
||||
|
||||
ukey1[0] = swap32_S (h32_from_64_S (tmps[gid].out[7]));
|
||||
ukey1[1] = swap32_S (l32_from_64_S (tmps[gid].out[7]));
|
||||
ukey1[2] = swap32_S (h32_from_64_S (tmps[gid].out[6]));
|
||||
ukey1[3] = swap32_S (l32_from_64_S (tmps[gid].out[6]));
|
||||
ukey1[4] = swap32_S (h32_from_64_S (tmps[gid].out[5]));
|
||||
ukey1[5] = swap32_S (l32_from_64_S (tmps[gid].out[5]));
|
||||
ukey1[6] = swap32_S (h32_from_64_S (tmps[gid].out[4]));
|
||||
ukey1[7] = swap32_S (l32_from_64_S (tmps[gid].out[4]));
|
||||
|
||||
u32 ukey2[8];
|
||||
|
||||
ukey2[0] = swap32_S (h32_from_64_S (tmps[gid].out[3]));
|
||||
ukey2[1] = swap32_S (l32_from_64_S (tmps[gid].out[3]));
|
||||
ukey2[2] = swap32_S (h32_from_64_S (tmps[gid].out[2]));
|
||||
ukey2[3] = swap32_S (l32_from_64_S (tmps[gid].out[2]));
|
||||
ukey2[4] = swap32_S (h32_from_64_S (tmps[gid].out[1]));
|
||||
ukey2[5] = swap32_S (l32_from_64_S (tmps[gid].out[1]));
|
||||
ukey2[6] = swap32_S (h32_from_64_S (tmps[gid].out[0]));
|
||||
ukey2[7] = swap32_S (l32_from_64_S (tmps[gid].out[0]));
|
||||
|
||||
if (verify_header_aes (esalt_bufs, ukey1, ukey2, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_serpent (esalt_bufs, ukey1, ukey2) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_twofish (esalt_bufs, ukey1, ukey2) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
u32 ukey3[8];
|
||||
|
||||
ukey3[0] = swap32_S (h32_from_64_S (tmps[gid].out[15]));
|
||||
ukey3[1] = swap32_S (l32_from_64_S (tmps[gid].out[15]));
|
||||
ukey3[2] = swap32_S (h32_from_64_S (tmps[gid].out[14]));
|
||||
ukey3[3] = swap32_S (l32_from_64_S (tmps[gid].out[14]));
|
||||
ukey3[4] = swap32_S (h32_from_64_S (tmps[gid].out[13]));
|
||||
ukey3[5] = swap32_S (l32_from_64_S (tmps[gid].out[13]));
|
||||
ukey3[6] = swap32_S (h32_from_64_S (tmps[gid].out[12]));
|
||||
ukey3[7] = swap32_S (l32_from_64_S (tmps[gid].out[12]));
|
||||
|
||||
u32 ukey4[8];
|
||||
|
||||
ukey4[0] = swap32_S (h32_from_64_S (tmps[gid].out[11]));
|
||||
ukey4[1] = swap32_S (l32_from_64_S (tmps[gid].out[11]));
|
||||
ukey4[2] = swap32_S (h32_from_64_S (tmps[gid].out[10]));
|
||||
ukey4[3] = swap32_S (l32_from_64_S (tmps[gid].out[10]));
|
||||
ukey4[4] = swap32_S (h32_from_64_S (tmps[gid].out[ 9]));
|
||||
ukey4[5] = swap32_S (l32_from_64_S (tmps[gid].out[ 9]));
|
||||
ukey4[6] = swap32_S (h32_from_64_S (tmps[gid].out[ 8]));
|
||||
ukey4[7] = swap32_S (l32_from_64_S (tmps[gid].out[ 8]));
|
||||
|
||||
if (verify_header_aes_twofish (esalt_bufs, ukey1, ukey2, ukey3, ukey4, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_serpent_aes (esalt_bufs, ukey1, ukey2, ukey3, ukey4, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_twofish_serpent (esalt_bufs, ukey1, ukey2, ukey3, ukey4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
u32 ukey5[8];
|
||||
|
||||
ukey5[0] = swap32_S (h32_from_64_S (tmps[gid].out[23]));
|
||||
ukey5[1] = swap32_S (l32_from_64_S (tmps[gid].out[23]));
|
||||
ukey5[2] = swap32_S (h32_from_64_S (tmps[gid].out[22]));
|
||||
ukey5[3] = swap32_S (l32_from_64_S (tmps[gid].out[22]));
|
||||
ukey5[4] = swap32_S (h32_from_64_S (tmps[gid].out[21]));
|
||||
ukey5[5] = swap32_S (l32_from_64_S (tmps[gid].out[21]));
|
||||
ukey5[6] = swap32_S (h32_from_64_S (tmps[gid].out[20]));
|
||||
ukey5[7] = swap32_S (l32_from_64_S (tmps[gid].out[20]));
|
||||
|
||||
u32 ukey6[8];
|
||||
|
||||
ukey6[0] = swap32_S (h32_from_64_S (tmps[gid].out[19]));
|
||||
ukey6[1] = swap32_S (l32_from_64_S (tmps[gid].out[19]));
|
||||
ukey6[2] = swap32_S (h32_from_64_S (tmps[gid].out[18]));
|
||||
ukey6[3] = swap32_S (l32_from_64_S (tmps[gid].out[18]));
|
||||
ukey6[4] = swap32_S (h32_from_64_S (tmps[gid].out[17]));
|
||||
ukey6[5] = swap32_S (l32_from_64_S (tmps[gid].out[17]));
|
||||
ukey6[6] = swap32_S (h32_from_64_S (tmps[gid].out[16]));
|
||||
ukey6[7] = swap32_S (l32_from_64_S (tmps[gid].out[16]));
|
||||
|
||||
if (verify_header_aes_twofish_serpent (esalt_bufs, ukey1, ukey2, ukey3, ukey4, ukey5, ukey6, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
|
||||
if (verify_header_serpent_twofish_aes (esalt_bufs, ukey1, ukey2, ukey3, ukey4, ukey5, ukey6, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4) == 1)
|
||||
{
|
||||
if (atomic_inc (&hashes_shown[0]) == 0)
|
||||
{
|
||||
mark_hash (plains_buf, d_return_buf, salt_pos, digests_cnt, 0, 0, gid, 0);
|
||||
}
|
||||
}
|
||||
}
|
||||
Loading…
Reference in new issue