arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-11-26 18:08:20 +00:00
Code Issues Releases Wiki Activity

Mix in pure kernel functions in various optimized kernels

Browse Source
This commit is contained in:
jsteube 2017-08-10 12:38:42 +02:00
parent a228e296da
commit 66f7590883
22 changed files with 168 additions and 6878 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

111
OpenCL/m05300_a0-optimized.cl
View File

@ -13,104 +13,7 @@
#include "inc_rp.h" #include "inc_rp.h"
#include "inc_rp.cl" #include "inc_rp.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md5.cl"
void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4]) void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
{ {
@ -136,7 +39,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
ipad[2] = MD5M_C; ipad[2] = MD5M_C;
ipad[3] = MD5M_D; ipad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -160,7 +63,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
opad[2] = MD5M_C; opad[2] = MD5M_C;
opad[3] = MD5M_D; opad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, opad); md5_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4]) void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
@ -170,7 +73,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = ipad[2]; digest[2] = ipad[2];
digest[3] = ipad[3]; digest[3] = ipad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -194,7 +97,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = opad[2]; digest[2] = opad[2];
digest[3] = opad[3]; digest[3] = opad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
} }
__kernel void m05300_m04 (__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 ikepsk_t *ikepsk_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 u32 gid_max) __kernel void m05300_m04 (__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 ikepsk_t *ikepsk_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 u32 gid_max)
@ -340,7 +243,7 @@ __kernel void m05300_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
@ -529,7 +432,7 @@ __kernel void m05300_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];

111
OpenCL/m05300_a1-optimized.cl
View File

@ -11,104 +11,7 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md5.cl"
void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4]) void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
{ {
@ -134,7 +37,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
ipad[2] = MD5M_C; ipad[2] = MD5M_C;
ipad[3] = MD5M_D; ipad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -158,7 +61,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
opad[2] = MD5M_C; opad[2] = MD5M_C;
opad[3] = MD5M_D; opad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, opad); md5_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4]) void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
@ -168,7 +71,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = ipad[2]; digest[2] = ipad[2];
digest[3] = ipad[3]; digest[3] = ipad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -192,7 +95,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = opad[2]; digest[2] = opad[2];
digest[3] = opad[3]; digest[3] = opad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
} }
__kernel void m05300_m04 (__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 ikepsk_t *ikepsk_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 u32 gid_max) __kernel void m05300_m04 (__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 ikepsk_t *ikepsk_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 u32 gid_max)
@ -398,7 +301,7 @@ __kernel void m05300_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
@ -647,7 +550,7 @@ __kernel void m05300_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];

111
OpenCL/m05300_a3-optimized.cl
View File

@ -11,104 +11,7 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md5.cl"
void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4]) void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
{ {
@ -134,7 +37,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
ipad[2] = MD5M_C; ipad[2] = MD5M_C;
ipad[3] = MD5M_D; ipad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -158,7 +61,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
opad[2] = MD5M_C; opad[2] = MD5M_C;
opad[3] = MD5M_D; opad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, opad); md5_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4]) void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
@ -168,7 +71,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = ipad[2]; digest[2] = ipad[2];
digest[3] = ipad[3]; digest[3] = ipad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -192,7 +95,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = opad[2]; digest[2] = opad[2];
digest[3] = opad[3]; digest[3] = opad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
} }
void m05300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 ikepsk_t *ikepsk_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, __local u32 *s_msg_buf) void m05300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 ikepsk_t *ikepsk_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, __local u32 *s_msg_buf)
@ -316,7 +219,7 @@ void m05300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = s_msg_buf[off + 14]; w3_t[2] = s_msg_buf[off + 14];
w3_t[3] = s_msg_buf[off + 15]; w3_t[3] = s_msg_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_msg_buf[off + 0]; w0_t[0] = s_msg_buf[off + 0];
@ -475,7 +378,7 @@ void m05300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = s_msg_buf[off + 14]; w3_t[2] = s_msg_buf[off + 14];
w3_t[3] = s_msg_buf[off + 15]; w3_t[3] = s_msg_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_msg_buf[off + 0]; w0_t[0] = s_msg_buf[off + 0];

141
OpenCL/m05400_a0-optimized.cl
View File

@ -13,134 +13,7 @@
#include "inc_rp.h" #include "inc_rp.h"
#include "inc_rp.cl" #include "inc_rp.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_sha1.cl"
void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
{
u32x A = digest[0];
u32x B = digest[1];
u32x C = digest[2];
u32x D = digest[3];
u32x E = digest[4];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w0_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w1_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w2_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w3_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w4_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w5_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w6_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w7_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w8_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w9_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wa_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, wb_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, wc_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, wd_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, we_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, E, A, B, C, D, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, D, E, A, B, C, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, C, D, E, A, B, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, B, C, D, E, A, w3_t);
#undef K
#define K SHA1C01
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w7_t);
#undef K
#define K SHA1C02
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wb_t);
#undef K
#define K SHA1C03
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wf_t);
digest[0] += A;
digest[1] += B;
digest[2] += C;
digest[3] += D;
digest[4] += E;
}
void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5]) void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
{ {
@ -167,7 +40,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
ipad[3] = SHA1M_D; ipad[3] = SHA1M_D;
ipad[4] = SHA1M_E; ipad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -192,7 +65,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
opad[3] = SHA1M_D; opad[3] = SHA1M_D;
opad[4] = SHA1M_E; opad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, opad); sha1_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5]) void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
@ -203,7 +76,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = ipad[3]; digest[3] = ipad[3];
digest[4] = ipad[4]; digest[4] = ipad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -228,7 +101,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = opad[3]; digest[3] = opad[3];
digest[4] = opad[4]; digest[4] = opad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
} }
__kernel void m05400_m04 (__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 ikepsk_t *ikepsk_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 u32 gid_max) __kernel void m05400_m04 (__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 ikepsk_t *ikepsk_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 u32 gid_max)
@ -383,7 +256,7 @@ __kernel void m05400_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
@ -581,7 +454,7 @@ __kernel void m05400_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];

141
OpenCL/m05400_a1-optimized.cl
View File

@ -11,134 +11,7 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_sha1.cl"
void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
{
u32x A = digest[0];
u32x B = digest[1];
u32x C = digest[2];
u32x D = digest[3];
u32x E = digest[4];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w0_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w1_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w2_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w3_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w4_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w5_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w6_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w7_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w8_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w9_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wa_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, wb_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, wc_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, wd_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, we_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, E, A, B, C, D, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, D, E, A, B, C, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, C, D, E, A, B, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, B, C, D, E, A, w3_t);
#undef K
#define K SHA1C01
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w7_t);
#undef K
#define K SHA1C02
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wb_t);
#undef K
#define K SHA1C03
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wf_t);
digest[0] += A;
digest[1] += B;
digest[2] += C;
digest[3] += D;
digest[4] += E;
}
void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5]) void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
{ {
@ -165,7 +38,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
ipad[3] = SHA1M_D; ipad[3] = SHA1M_D;
ipad[4] = SHA1M_E; ipad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -190,7 +63,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
opad[3] = SHA1M_D; opad[3] = SHA1M_D;
opad[4] = SHA1M_E; opad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, opad); sha1_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5]) void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
@ -201,7 +74,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = ipad[3]; digest[3] = ipad[3];
digest[4] = ipad[4]; digest[4] = ipad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -226,7 +99,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = opad[3]; digest[3] = opad[3];
digest[4] = opad[4]; digest[4] = opad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
} }
__kernel void m05400_m04 (__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 ikepsk_t *ikepsk_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 u32 gid_max) __kernel void m05400_m04 (__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 ikepsk_t *ikepsk_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 u32 gid_max)
@ -449,7 +322,7 @@ __kernel void m05400_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];
@ -715,7 +588,7 @@ __kernel void m05400_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = s_msg_buf[off + 14]; w3[2] = s_msg_buf[off + 14];
w3[3] = s_msg_buf[off + 15]; w3[3] = s_msg_buf[off + 15];
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = s_msg_buf[off + 0]; w0[0] = s_msg_buf[off + 0];

141
OpenCL/m05400_a3-optimized.cl
View File

@ -11,134 +11,7 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_sha1.cl"
void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
{
u32x A = digest[0];
u32x B = digest[1];
u32x C = digest[2];
u32x D = digest[3];
u32x E = digest[4];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w0_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w1_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w2_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w3_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w4_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w5_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w6_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w7_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w8_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w9_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wa_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, wb_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, wc_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, wd_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, we_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, E, A, B, C, D, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, D, E, A, B, C, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, C, D, E, A, B, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, B, C, D, E, A, w3_t);
#undef K
#define K SHA1C01
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w7_t);
#undef K
#define K SHA1C02
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wb_t);
#undef K
#define K SHA1C03
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wf_t);
digest[0] += A;
digest[1] += B;
digest[2] += C;
digest[3] += D;
digest[4] += E;
}
void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5]) void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
{ {
@ -165,7 +38,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
ipad[3] = SHA1M_D; ipad[3] = SHA1M_D;
ipad[4] = SHA1M_E; ipad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -190,7 +63,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
opad[3] = SHA1M_D; opad[3] = SHA1M_D;
opad[4] = SHA1M_E; opad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, opad); sha1_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5]) void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
@ -201,7 +74,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = ipad[3]; digest[3] = ipad[3];
digest[4] = ipad[4]; digest[4] = ipad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -226,7 +99,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = opad[3]; digest[3] = opad[3];
digest[4] = opad[4]; digest[4] = opad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
} }
void m05400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 ikepsk_t *ikepsk_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, __local u32 *s_msg_buf) void m05400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 ikepsk_t *ikepsk_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, __local u32 *s_msg_buf)
@ -350,7 +223,7 @@ void m05400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = s_msg_buf[off + 14]; w3_t[2] = s_msg_buf[off + 14];
w3_t[3] = s_msg_buf[off + 15]; w3_t[3] = s_msg_buf[off + 15];
sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha1_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_msg_buf[off + 0]; w0_t[0] = s_msg_buf[off + 0];
@ -509,7 +382,7 @@ void m05400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = s_msg_buf[off + 14]; w3_t[2] = s_msg_buf[off + 14];
w3_t[3] = s_msg_buf[off + 15]; w3_t[3] = s_msg_buf[off + 15];
sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha1_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_msg_buf[off + 0]; w0_t[0] = s_msg_buf[off + 0];

201
OpenCL/m05600_a0-optimized.cl
View File

@ -13,185 +13,8 @@
#include "inc_rp.h" #include "inc_rp.h"
#include "inc_rp.cl" #include "inc_rp.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md4.cl"
void md4_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4]) #include "inc_hash_md5.cl"
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4]) void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
{ {
@ -217,7 +40,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
ipad[2] = MD5M_C; ipad[2] = MD5M_C;
ipad[3] = MD5M_D; ipad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -241,7 +64,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
opad[2] = MD5M_C; opad[2] = MD5M_C;
opad[3] = MD5M_D; opad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, opad); md5_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4]) void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
@ -251,7 +74,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = ipad[2]; digest[2] = ipad[2];
digest[3] = ipad[3]; digest[3] = ipad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -275,7 +98,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = opad[2]; digest[2] = opad[2];
digest[3] = opad[3]; digest[3] = opad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
} }
__kernel void m05600_m04 (__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 netntlm_t *netntlm_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 u32 gid_max) __kernel void m05600_m04 (__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 netntlm_t *netntlm_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 u32 gid_max)
@ -367,7 +190,7 @@ __kernel void m05600_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
digest[2] = MD4M_C; digest[2] = MD4M_C;
digest[3] = MD4M_D; digest[3] = MD4M_D;
md4_transform (w0_t, w1_t, w2_t, w3_t, digest); md4_transform_vector (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = digest[0]; w0_t[0] = digest[0];
w0_t[1] = digest[1]; w0_t[1] = digest[1];
@ -418,7 +241,7 @@ __kernel void m05600_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3_t[2] = s_userdomain_buf[off + 14]; w3_t[2] = s_userdomain_buf[off + 14];
w3_t[3] = s_userdomain_buf[off + 15]; w3_t[3] = s_userdomain_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_userdomain_buf[off + 0]; w0_t[0] = s_userdomain_buf[off + 0];
@ -483,7 +306,7 @@ __kernel void m05600_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3_t[2] = s_chall_buf[off + 14]; w3_t[2] = s_chall_buf[off + 14];
w3_t[3] = s_chall_buf[off + 15]; w3_t[3] = s_chall_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_chall_buf[off + 0]; w0_t[0] = s_chall_buf[off + 0];
@ -618,7 +441,7 @@ __kernel void m05600_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
digest[2] = MD4M_C; digest[2] = MD4M_C;
digest[3] = MD4M_D; digest[3] = MD4M_D;
md4_transform (w0_t, w1_t, w2_t, w3_t, digest); md4_transform_vector (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = digest[0]; w0_t[0] = digest[0];
w0_t[1] = digest[1]; w0_t[1] = digest[1];
@ -669,7 +492,7 @@ __kernel void m05600_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3_t[2] = s_userdomain_buf[off + 14]; w3_t[2] = s_userdomain_buf[off + 14];
w3_t[3] = s_userdomain_buf[off + 15]; w3_t[3] = s_userdomain_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_userdomain_buf[off + 0]; w0_t[0] = s_userdomain_buf[off + 0];
@ -734,7 +557,7 @@ __kernel void m05600_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3_t[2] = s_chall_buf[off + 14]; w3_t[2] = s_chall_buf[off + 14];
w3_t[3] = s_chall_buf[off + 15]; w3_t[3] = s_chall_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_chall_buf[off + 0]; w0_t[0] = s_chall_buf[off + 0];

201
OpenCL/m05600_a1-optimized.cl
View File

@ -11,185 +11,8 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md4.cl"
void md4_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4]) #include "inc_hash_md5.cl"
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4]) void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
{ {
@ -215,7 +38,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
ipad[2] = MD5M_C; ipad[2] = MD5M_C;
ipad[3] = MD5M_D; ipad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -239,7 +62,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
opad[2] = MD5M_C; opad[2] = MD5M_C;
opad[3] = MD5M_D; opad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, opad); md5_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4]) void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
@ -249,7 +72,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = ipad[2]; digest[2] = ipad[2];
digest[3] = ipad[3]; digest[3] = ipad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -273,7 +96,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = opad[2]; digest[2] = opad[2];
digest[3] = opad[3]; digest[3] = opad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
} }
__kernel void m05600_m04 (__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 netntlm_t *netntlm_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 u32 gid_max) __kernel void m05600_m04 (__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 netntlm_t *netntlm_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 u32 gid_max)
@ -423,7 +246,7 @@ __kernel void m05600_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
digest[2] = MD4M_C; digest[2] = MD4M_C;
digest[3] = MD4M_D; digest[3] = MD4M_D;
md4_transform (w0_t, w1_t, w2_t, w3_t, digest); md4_transform_vector (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = digest[0]; w0_t[0] = digest[0];
w0_t[1] = digest[1]; w0_t[1] = digest[1];
@ -474,7 +297,7 @@ __kernel void m05600_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3_t[2] = s_userdomain_buf[off + 14]; w3_t[2] = s_userdomain_buf[off + 14];
w3_t[3] = s_userdomain_buf[off + 15]; w3_t[3] = s_userdomain_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_userdomain_buf[off + 0]; w0_t[0] = s_userdomain_buf[off + 0];
@ -539,7 +362,7 @@ __kernel void m05600_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3_t[2] = s_chall_buf[off + 14]; w3_t[2] = s_chall_buf[off + 14];
w3_t[3] = s_chall_buf[off + 15]; w3_t[3] = s_chall_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_chall_buf[off + 0]; w0_t[0] = s_chall_buf[off + 0];
@ -732,7 +555,7 @@ __kernel void m05600_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
digest[2] = MD4M_C; digest[2] = MD4M_C;
digest[3] = MD4M_D; digest[3] = MD4M_D;
md4_transform (w0_t, w1_t, w2_t, w3_t, digest); md4_transform_vector (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = digest[0]; w0_t[0] = digest[0];
w0_t[1] = digest[1]; w0_t[1] = digest[1];
@ -783,7 +606,7 @@ __kernel void m05600_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3_t[2] = s_userdomain_buf[off + 14]; w3_t[2] = s_userdomain_buf[off + 14];
w3_t[3] = s_userdomain_buf[off + 15]; w3_t[3] = s_userdomain_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_userdomain_buf[off + 0]; w0_t[0] = s_userdomain_buf[off + 0];
@ -848,7 +671,7 @@ __kernel void m05600_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3_t[2] = s_chall_buf[off + 14]; w3_t[2] = s_chall_buf[off + 14];
w3_t[3] = s_chall_buf[off + 15]; w3_t[3] = s_chall_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_chall_buf[off + 0]; w0_t[0] = s_chall_buf[off + 0];

201
OpenCL/m05600_a3-optimized.cl
View File

@ -11,185 +11,8 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md4.cl"
void md4_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4]) #include "inc_hash_md5.cl"
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{
u32x a = digest[0];
u32x b = digest[1];
u32x c = digest[2];
u32x d = digest[3];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4]) void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
{ {
@ -215,7 +38,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
ipad[2] = MD5M_C; ipad[2] = MD5M_C;
ipad[3] = MD5M_D; ipad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, ipad); md5_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -239,7 +62,7 @@ void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
opad[2] = MD5M_C; opad[2] = MD5M_C;
opad[3] = MD5M_D; opad[3] = MD5M_D;
md5_transform (w0, w1, w2, w3, opad); md5_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4]) void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
@ -249,7 +72,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = ipad[2]; digest[2] = ipad[2];
digest[3] = ipad[3]; digest[3] = ipad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -273,7 +96,7 @@ void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4],
digest[2] = opad[2]; digest[2] = opad[2];
digest[3] = opad[3]; digest[3] = opad[3];
md5_transform (w0, w1, w2, w3, digest); md5_transform_vector (w0, w1, w2, w3, digest);
} }
void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 netntlm_t *netntlm_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, __local u32 *s_userdomain_buf, __local u32 *s_chall_buf) void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 netntlm_t *netntlm_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, __local u32 *s_userdomain_buf, __local u32 *s_chall_buf)
@ -340,7 +163,7 @@ void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
digest[2] = MD4M_C; digest[2] = MD4M_C;
digest[3] = MD4M_D; digest[3] = MD4M_D;
md4_transform (w0_t, w1_t, w2_t, w3_t, digest); md4_transform_vector (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = digest[0]; w0_t[0] = digest[0];
w0_t[1] = digest[1]; w0_t[1] = digest[1];
@ -391,7 +214,7 @@ void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = s_userdomain_buf[off + 14]; w3_t[2] = s_userdomain_buf[off + 14];
w3_t[3] = s_userdomain_buf[off + 15]; w3_t[3] = s_userdomain_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_userdomain_buf[off + 0]; w0_t[0] = s_userdomain_buf[off + 0];
@ -456,7 +279,7 @@ void m05600m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = s_chall_buf[off + 14]; w3_t[2] = s_chall_buf[off + 14];
w3_t[3] = s_chall_buf[off + 15]; w3_t[3] = s_chall_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_chall_buf[off + 0]; w0_t[0] = s_chall_buf[off + 0];
@ -558,7 +381,7 @@ void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
digest[2] = MD4M_C; digest[2] = MD4M_C;
digest[3] = MD4M_D; digest[3] = MD4M_D;
md4_transform (w0_t, w1_t, w2_t, w3_t, digest); md4_transform_vector (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = digest[0]; w0_t[0] = digest[0];
w0_t[1] = digest[1]; w0_t[1] = digest[1];
@ -609,7 +432,7 @@ void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = s_userdomain_buf[off + 14]; w3_t[2] = s_userdomain_buf[off + 14];
w3_t[3] = s_userdomain_buf[off + 15]; w3_t[3] = s_userdomain_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_userdomain_buf[off + 0]; w0_t[0] = s_userdomain_buf[off + 0];
@ -674,7 +497,7 @@ void m05600s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = s_chall_buf[off + 14]; w3_t[2] = s_chall_buf[off + 14];
w3_t[3] = s_chall_buf[off + 15]; w3_t[3] = s_chall_buf[off + 15];
md5_transform (w0_t, w1_t, w2_t, w3_t, ipad); md5_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = s_chall_buf[off + 0]; w0_t[0] = s_chall_buf[off + 0];

201
OpenCL/m06000_a0-optimized.cl
View File

@ -13,202 +13,11 @@
#include "inc_rp.h" #include "inc_rp.h"
#include "inc_rp.cl" #include "inc_rp.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_ripemd160.cl"
void ripemd160_transform (const u32x w[16], u32x dgst[5]) void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
{ {
u32x a1 = dgst[0]; ripemd160_transform_vector (w + 0, w + 4, w + 8, w + 12, dgst);
u32x b1 = dgst[1];
u32x c1 = dgst[2];
u32x d1 = dgst[3];
u32x e1 = dgst[4];
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[ 0], RIPEMD160C00, RIPEMD160S00);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[ 1], RIPEMD160C00, RIPEMD160S01);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[ 2], RIPEMD160C00, RIPEMD160S02);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[ 3], RIPEMD160C00, RIPEMD160S03);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[ 4], RIPEMD160C00, RIPEMD160S04);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[ 5], RIPEMD160C00, RIPEMD160S05);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[ 6], RIPEMD160C00, RIPEMD160S06);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[ 7], RIPEMD160C00, RIPEMD160S07);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[ 8], RIPEMD160C00, RIPEMD160S08);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[ 9], RIPEMD160C00, RIPEMD160S09);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[10], RIPEMD160C00, RIPEMD160S0A);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[11], RIPEMD160C00, RIPEMD160S0B);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[12], RIPEMD160C00, RIPEMD160S0C);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[13], RIPEMD160C00, RIPEMD160S0D);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[14], RIPEMD160C00, RIPEMD160S0E);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[15], RIPEMD160C00, RIPEMD160S0F);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 7], RIPEMD160C10, RIPEMD160S10);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[ 4], RIPEMD160C10, RIPEMD160S11);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[13], RIPEMD160C10, RIPEMD160S12);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[ 1], RIPEMD160C10, RIPEMD160S13);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[10], RIPEMD160C10, RIPEMD160S14);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 6], RIPEMD160C10, RIPEMD160S15);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[15], RIPEMD160C10, RIPEMD160S16);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[ 3], RIPEMD160C10, RIPEMD160S17);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[12], RIPEMD160C10, RIPEMD160S18);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[ 0], RIPEMD160C10, RIPEMD160S19);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 9], RIPEMD160C10, RIPEMD160S1A);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[ 5], RIPEMD160C10, RIPEMD160S1B);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[ 2], RIPEMD160C10, RIPEMD160S1C);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[14], RIPEMD160C10, RIPEMD160S1D);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[11], RIPEMD160C10, RIPEMD160S1E);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 8], RIPEMD160C10, RIPEMD160S1F);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[ 3], RIPEMD160C20, RIPEMD160S20);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[10], RIPEMD160C20, RIPEMD160S21);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[14], RIPEMD160C20, RIPEMD160S22);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[ 4], RIPEMD160C20, RIPEMD160S23);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 9], RIPEMD160C20, RIPEMD160S24);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[15], RIPEMD160C20, RIPEMD160S25);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[ 8], RIPEMD160C20, RIPEMD160S26);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[ 1], RIPEMD160C20, RIPEMD160S27);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[ 2], RIPEMD160C20, RIPEMD160S28);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 7], RIPEMD160C20, RIPEMD160S29);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[ 0], RIPEMD160C20, RIPEMD160S2A);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[ 6], RIPEMD160C20, RIPEMD160S2B);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[13], RIPEMD160C20, RIPEMD160S2C);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[11], RIPEMD160C20, RIPEMD160S2D);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 5], RIPEMD160C20, RIPEMD160S2E);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[12], RIPEMD160C20, RIPEMD160S2F);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 1], RIPEMD160C30, RIPEMD160S30);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[ 9], RIPEMD160C30, RIPEMD160S31);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[11], RIPEMD160C30, RIPEMD160S32);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[10], RIPEMD160C30, RIPEMD160S33);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 0], RIPEMD160C30, RIPEMD160S34);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 8], RIPEMD160C30, RIPEMD160S35);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[12], RIPEMD160C30, RIPEMD160S36);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[ 4], RIPEMD160C30, RIPEMD160S37);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[13], RIPEMD160C30, RIPEMD160S38);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 3], RIPEMD160C30, RIPEMD160S39);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 7], RIPEMD160C30, RIPEMD160S3A);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[15], RIPEMD160C30, RIPEMD160S3B);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[14], RIPEMD160C30, RIPEMD160S3C);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[ 5], RIPEMD160C30, RIPEMD160S3D);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 6], RIPEMD160C30, RIPEMD160S3E);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 2], RIPEMD160C30, RIPEMD160S3F);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[ 4], RIPEMD160C40, RIPEMD160S40);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 0], RIPEMD160C40, RIPEMD160S41);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[ 5], RIPEMD160C40, RIPEMD160S42);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[ 9], RIPEMD160C40, RIPEMD160S43);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[ 7], RIPEMD160C40, RIPEMD160S44);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[12], RIPEMD160C40, RIPEMD160S45);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 2], RIPEMD160C40, RIPEMD160S46);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[10], RIPEMD160C40, RIPEMD160S47);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[14], RIPEMD160C40, RIPEMD160S48);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[ 1], RIPEMD160C40, RIPEMD160S49);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[ 3], RIPEMD160C40, RIPEMD160S4A);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 8], RIPEMD160C40, RIPEMD160S4B);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[11], RIPEMD160C40, RIPEMD160S4C);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[ 6], RIPEMD160C40, RIPEMD160S4D);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[15], RIPEMD160C40, RIPEMD160S4E);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[13], RIPEMD160C40, RIPEMD160S4F);
u32x a2 = dgst[0];
u32x b2 = dgst[1];
u32x c2 = dgst[2];
u32x d2 = dgst[3];
u32x e2 = dgst[4];
RIPEMD160_STEP_WORKAROUND_BUG (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 5], RIPEMD160C50, RIPEMD160S50);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[14], RIPEMD160C50, RIPEMD160S51);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 7], RIPEMD160C50, RIPEMD160S52);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[ 0], RIPEMD160C50, RIPEMD160S53);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 9], RIPEMD160C50, RIPEMD160S54);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 2], RIPEMD160C50, RIPEMD160S55);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[11], RIPEMD160C50, RIPEMD160S56);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 4], RIPEMD160C50, RIPEMD160S57);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[13], RIPEMD160C50, RIPEMD160S58);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 6], RIPEMD160C50, RIPEMD160S59);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[15], RIPEMD160C50, RIPEMD160S5A);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[ 8], RIPEMD160C50, RIPEMD160S5B);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 1], RIPEMD160C50, RIPEMD160S5C);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[10], RIPEMD160C50, RIPEMD160S5D);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 3], RIPEMD160C50, RIPEMD160S5E);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[12], RIPEMD160C50, RIPEMD160S5F);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 6], RIPEMD160C60, RIPEMD160S60);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[11], RIPEMD160C60, RIPEMD160S61);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[ 3], RIPEMD160C60, RIPEMD160S62);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[ 7], RIPEMD160C60, RIPEMD160S63);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[ 0], RIPEMD160C60, RIPEMD160S64);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[13], RIPEMD160C60, RIPEMD160S65);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[ 5], RIPEMD160C60, RIPEMD160S66);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[10], RIPEMD160C60, RIPEMD160S67);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[14], RIPEMD160C60, RIPEMD160S68);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[15], RIPEMD160C60, RIPEMD160S69);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 8], RIPEMD160C60, RIPEMD160S6A);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[12], RIPEMD160C60, RIPEMD160S6B);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[ 4], RIPEMD160C60, RIPEMD160S6C);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[ 9], RIPEMD160C60, RIPEMD160S6D);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[ 1], RIPEMD160C60, RIPEMD160S6E);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 2], RIPEMD160C60, RIPEMD160S6F);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[15], RIPEMD160C70, RIPEMD160S70);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 5], RIPEMD160C70, RIPEMD160S71);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[ 1], RIPEMD160C70, RIPEMD160S72);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[ 3], RIPEMD160C70, RIPEMD160S73);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 7], RIPEMD160C70, RIPEMD160S74);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[14], RIPEMD160C70, RIPEMD160S75);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 6], RIPEMD160C70, RIPEMD160S76);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[ 9], RIPEMD160C70, RIPEMD160S77);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[11], RIPEMD160C70, RIPEMD160S78);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 8], RIPEMD160C70, RIPEMD160S79);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[12], RIPEMD160C70, RIPEMD160S7A);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 2], RIPEMD160C70, RIPEMD160S7B);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[10], RIPEMD160C70, RIPEMD160S7C);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[ 0], RIPEMD160C70, RIPEMD160S7D);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 4], RIPEMD160C70, RIPEMD160S7E);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[13], RIPEMD160C70, RIPEMD160S7F);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[ 8], RIPEMD160C80, RIPEMD160S80);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[ 6], RIPEMD160C80, RIPEMD160S81);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 4], RIPEMD160C80, RIPEMD160S82);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 1], RIPEMD160C80, RIPEMD160S83);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[ 3], RIPEMD160C80, RIPEMD160S84);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[11], RIPEMD160C80, RIPEMD160S85);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[15], RIPEMD160C80, RIPEMD160S86);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 0], RIPEMD160C80, RIPEMD160S87);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 5], RIPEMD160C80, RIPEMD160S88);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[12], RIPEMD160C80, RIPEMD160S89);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[ 2], RIPEMD160C80, RIPEMD160S8A);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[13], RIPEMD160C80, RIPEMD160S8B);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 9], RIPEMD160C80, RIPEMD160S8C);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 7], RIPEMD160C80, RIPEMD160S8D);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[10], RIPEMD160C80, RIPEMD160S8E);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[14], RIPEMD160C80, RIPEMD160S8F);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[12], RIPEMD160C90, RIPEMD160S90);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[15], RIPEMD160C90, RIPEMD160S91);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[10], RIPEMD160C90, RIPEMD160S92);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 4], RIPEMD160C90, RIPEMD160S93);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 1], RIPEMD160C90, RIPEMD160S94);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[ 5], RIPEMD160C90, RIPEMD160S95);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[ 8], RIPEMD160C90, RIPEMD160S96);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[ 7], RIPEMD160C90, RIPEMD160S97);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 6], RIPEMD160C90, RIPEMD160S98);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 2], RIPEMD160C90, RIPEMD160S99);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[13], RIPEMD160C90, RIPEMD160S9A);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[14], RIPEMD160C90, RIPEMD160S9B);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[ 0], RIPEMD160C90, RIPEMD160S9C);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 3], RIPEMD160C90, RIPEMD160S9D);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 9], RIPEMD160C90, RIPEMD160S9E);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[11], RIPEMD160C90, RIPEMD160S9F);
const u32x a = dgst[1] + c1 + d2;
const u32x b = dgst[2] + d1 + e2;
const u32x c = dgst[3] + e1 + a2;
const u32x d = dgst[4] + a1 + b2;
const u32x e = dgst[0] + b1 + c2;
dgst[0] = a;
dgst[1] = b;
dgst[2] = c;
dgst[3] = d;
dgst[4] = e;
} }
__kernel void m06000_m04 (__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 netntlm_t *netntlm_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 u32 gid_max) __kernel void m06000_m04 (__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 netntlm_t *netntlm_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 u32 gid_max)
@ -287,7 +96,7 @@ __kernel void m06000_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
dgst[3] = RIPEMD160M_D; dgst[3] = RIPEMD160M_D;
dgst[4] = RIPEMD160M_E; dgst[4] = RIPEMD160M_E;
ripemd160_transform (w, dgst); ripemd160_transform_transport_vector (w, dgst);
COMPARE_M_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]); COMPARE_M_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]);
} }
@ -389,7 +198,7 @@ __kernel void m06000_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
dgst[3] = RIPEMD160M_D; dgst[3] = RIPEMD160M_D;
dgst[4] = RIPEMD160M_E; dgst[4] = RIPEMD160M_E;
ripemd160_transform (w, dgst); ripemd160_transform_transport_vector (w, dgst);
COMPARE_S_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]); COMPARE_S_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]);
} }

201
OpenCL/m06000_a1-optimized.cl
View File

@ -11,202 +11,11 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_ripemd160.cl"
void ripemd160_transform (const u32x w[16], u32x dgst[5]) void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
{ {
u32x a1 = dgst[0]; ripemd160_transform_vector (w + 0, w + 4, w + 8, w + 12, dgst);
u32x b1 = dgst[1];
u32x c1 = dgst[2];
u32x d1 = dgst[3];
u32x e1 = dgst[4];
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[ 0], RIPEMD160C00, RIPEMD160S00);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[ 1], RIPEMD160C00, RIPEMD160S01);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[ 2], RIPEMD160C00, RIPEMD160S02);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[ 3], RIPEMD160C00, RIPEMD160S03);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[ 4], RIPEMD160C00, RIPEMD160S04);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[ 5], RIPEMD160C00, RIPEMD160S05);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[ 6], RIPEMD160C00, RIPEMD160S06);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[ 7], RIPEMD160C00, RIPEMD160S07);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[ 8], RIPEMD160C00, RIPEMD160S08);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[ 9], RIPEMD160C00, RIPEMD160S09);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[10], RIPEMD160C00, RIPEMD160S0A);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[11], RIPEMD160C00, RIPEMD160S0B);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[12], RIPEMD160C00, RIPEMD160S0C);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[13], RIPEMD160C00, RIPEMD160S0D);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[14], RIPEMD160C00, RIPEMD160S0E);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[15], RIPEMD160C00, RIPEMD160S0F);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 7], RIPEMD160C10, RIPEMD160S10);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[ 4], RIPEMD160C10, RIPEMD160S11);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[13], RIPEMD160C10, RIPEMD160S12);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[ 1], RIPEMD160C10, RIPEMD160S13);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[10], RIPEMD160C10, RIPEMD160S14);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 6], RIPEMD160C10, RIPEMD160S15);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[15], RIPEMD160C10, RIPEMD160S16);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[ 3], RIPEMD160C10, RIPEMD160S17);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[12], RIPEMD160C10, RIPEMD160S18);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[ 0], RIPEMD160C10, RIPEMD160S19);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 9], RIPEMD160C10, RIPEMD160S1A);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[ 5], RIPEMD160C10, RIPEMD160S1B);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[ 2], RIPEMD160C10, RIPEMD160S1C);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[14], RIPEMD160C10, RIPEMD160S1D);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[11], RIPEMD160C10, RIPEMD160S1E);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 8], RIPEMD160C10, RIPEMD160S1F);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[ 3], RIPEMD160C20, RIPEMD160S20);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[10], RIPEMD160C20, RIPEMD160S21);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[14], RIPEMD160C20, RIPEMD160S22);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[ 4], RIPEMD160C20, RIPEMD160S23);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 9], RIPEMD160C20, RIPEMD160S24);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[15], RIPEMD160C20, RIPEMD160S25);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[ 8], RIPEMD160C20, RIPEMD160S26);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[ 1], RIPEMD160C20, RIPEMD160S27);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[ 2], RIPEMD160C20, RIPEMD160S28);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 7], RIPEMD160C20, RIPEMD160S29);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[ 0], RIPEMD160C20, RIPEMD160S2A);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[ 6], RIPEMD160C20, RIPEMD160S2B);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[13], RIPEMD160C20, RIPEMD160S2C);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[11], RIPEMD160C20, RIPEMD160S2D);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 5], RIPEMD160C20, RIPEMD160S2E);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[12], RIPEMD160C20, RIPEMD160S2F);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 1], RIPEMD160C30, RIPEMD160S30);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[ 9], RIPEMD160C30, RIPEMD160S31);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[11], RIPEMD160C30, RIPEMD160S32);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[10], RIPEMD160C30, RIPEMD160S33);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 0], RIPEMD160C30, RIPEMD160S34);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 8], RIPEMD160C30, RIPEMD160S35);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[12], RIPEMD160C30, RIPEMD160S36);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[ 4], RIPEMD160C30, RIPEMD160S37);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[13], RIPEMD160C30, RIPEMD160S38);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 3], RIPEMD160C30, RIPEMD160S39);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 7], RIPEMD160C30, RIPEMD160S3A);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[15], RIPEMD160C30, RIPEMD160S3B);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[14], RIPEMD160C30, RIPEMD160S3C);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[ 5], RIPEMD160C30, RIPEMD160S3D);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 6], RIPEMD160C30, RIPEMD160S3E);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 2], RIPEMD160C30, RIPEMD160S3F);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[ 4], RIPEMD160C40, RIPEMD160S40);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 0], RIPEMD160C40, RIPEMD160S41);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[ 5], RIPEMD160C40, RIPEMD160S42);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[ 9], RIPEMD160C40, RIPEMD160S43);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[ 7], RIPEMD160C40, RIPEMD160S44);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[12], RIPEMD160C40, RIPEMD160S45);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 2], RIPEMD160C40, RIPEMD160S46);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[10], RIPEMD160C40, RIPEMD160S47);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[14], RIPEMD160C40, RIPEMD160S48);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[ 1], RIPEMD160C40, RIPEMD160S49);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[ 3], RIPEMD160C40, RIPEMD160S4A);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 8], RIPEMD160C40, RIPEMD160S4B);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[11], RIPEMD160C40, RIPEMD160S4C);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[ 6], RIPEMD160C40, RIPEMD160S4D);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[15], RIPEMD160C40, RIPEMD160S4E);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[13], RIPEMD160C40, RIPEMD160S4F);
u32x a2 = dgst[0];
u32x b2 = dgst[1];
u32x c2 = dgst[2];
u32x d2 = dgst[3];
u32x e2 = dgst[4];
RIPEMD160_STEP_WORKAROUND_BUG (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 5], RIPEMD160C50, RIPEMD160S50);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[14], RIPEMD160C50, RIPEMD160S51);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 7], RIPEMD160C50, RIPEMD160S52);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[ 0], RIPEMD160C50, RIPEMD160S53);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 9], RIPEMD160C50, RIPEMD160S54);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 2], RIPEMD160C50, RIPEMD160S55);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[11], RIPEMD160C50, RIPEMD160S56);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 4], RIPEMD160C50, RIPEMD160S57);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[13], RIPEMD160C50, RIPEMD160S58);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 6], RIPEMD160C50, RIPEMD160S59);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[15], RIPEMD160C50, RIPEMD160S5A);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[ 8], RIPEMD160C50, RIPEMD160S5B);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 1], RIPEMD160C50, RIPEMD160S5C);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[10], RIPEMD160C50, RIPEMD160S5D);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 3], RIPEMD160C50, RIPEMD160S5E);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[12], RIPEMD160C50, RIPEMD160S5F);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 6], RIPEMD160C60, RIPEMD160S60);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[11], RIPEMD160C60, RIPEMD160S61);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[ 3], RIPEMD160C60, RIPEMD160S62);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[ 7], RIPEMD160C60, RIPEMD160S63);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[ 0], RIPEMD160C60, RIPEMD160S64);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[13], RIPEMD160C60, RIPEMD160S65);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[ 5], RIPEMD160C60, RIPEMD160S66);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[10], RIPEMD160C60, RIPEMD160S67);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[14], RIPEMD160C60, RIPEMD160S68);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[15], RIPEMD160C60, RIPEMD160S69);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 8], RIPEMD160C60, RIPEMD160S6A);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[12], RIPEMD160C60, RIPEMD160S6B);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[ 4], RIPEMD160C60, RIPEMD160S6C);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[ 9], RIPEMD160C60, RIPEMD160S6D);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[ 1], RIPEMD160C60, RIPEMD160S6E);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 2], RIPEMD160C60, RIPEMD160S6F);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[15], RIPEMD160C70, RIPEMD160S70);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 5], RIPEMD160C70, RIPEMD160S71);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[ 1], RIPEMD160C70, RIPEMD160S72);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[ 3], RIPEMD160C70, RIPEMD160S73);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 7], RIPEMD160C70, RIPEMD160S74);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[14], RIPEMD160C70, RIPEMD160S75);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 6], RIPEMD160C70, RIPEMD160S76);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[ 9], RIPEMD160C70, RIPEMD160S77);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[11], RIPEMD160C70, RIPEMD160S78);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 8], RIPEMD160C70, RIPEMD160S79);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[12], RIPEMD160C70, RIPEMD160S7A);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 2], RIPEMD160C70, RIPEMD160S7B);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[10], RIPEMD160C70, RIPEMD160S7C);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[ 0], RIPEMD160C70, RIPEMD160S7D);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 4], RIPEMD160C70, RIPEMD160S7E);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[13], RIPEMD160C70, RIPEMD160S7F);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[ 8], RIPEMD160C80, RIPEMD160S80);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[ 6], RIPEMD160C80, RIPEMD160S81);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 4], RIPEMD160C80, RIPEMD160S82);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 1], RIPEMD160C80, RIPEMD160S83);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[ 3], RIPEMD160C80, RIPEMD160S84);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[11], RIPEMD160C80, RIPEMD160S85);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[15], RIPEMD160C80, RIPEMD160S86);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 0], RIPEMD160C80, RIPEMD160S87);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 5], RIPEMD160C80, RIPEMD160S88);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[12], RIPEMD160C80, RIPEMD160S89);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[ 2], RIPEMD160C80, RIPEMD160S8A);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[13], RIPEMD160C80, RIPEMD160S8B);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 9], RIPEMD160C80, RIPEMD160S8C);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 7], RIPEMD160C80, RIPEMD160S8D);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[10], RIPEMD160C80, RIPEMD160S8E);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[14], RIPEMD160C80, RIPEMD160S8F);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[12], RIPEMD160C90, RIPEMD160S90);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[15], RIPEMD160C90, RIPEMD160S91);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[10], RIPEMD160C90, RIPEMD160S92);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 4], RIPEMD160C90, RIPEMD160S93);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 1], RIPEMD160C90, RIPEMD160S94);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[ 5], RIPEMD160C90, RIPEMD160S95);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[ 8], RIPEMD160C90, RIPEMD160S96);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[ 7], RIPEMD160C90, RIPEMD160S97);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 6], RIPEMD160C90, RIPEMD160S98);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 2], RIPEMD160C90, RIPEMD160S99);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[13], RIPEMD160C90, RIPEMD160S9A);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[14], RIPEMD160C90, RIPEMD160S9B);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[ 0], RIPEMD160C90, RIPEMD160S9C);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 3], RIPEMD160C90, RIPEMD160S9D);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 9], RIPEMD160C90, RIPEMD160S9E);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[11], RIPEMD160C90, RIPEMD160S9F);
const u32x a = dgst[1] + c1 + d2;
const u32x b = dgst[2] + d1 + e2;
const u32x c = dgst[3] + e1 + a2;
const u32x d = dgst[4] + a1 + b2;
const u32x e = dgst[0] + b1 + c2;
dgst[0] = a;
dgst[1] = b;
dgst[2] = c;
dgst[3] = d;
dgst[4] = e;
} }
__kernel void m06000_m04 (__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 netntlm_t *netntlm_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 u32 gid_max) __kernel void m06000_m04 (__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 netntlm_t *netntlm_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 u32 gid_max)
@ -343,7 +152,7 @@ __kernel void m06000_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
dgst[3] = RIPEMD160M_D; dgst[3] = RIPEMD160M_D;
dgst[4] = RIPEMD160M_E; dgst[4] = RIPEMD160M_E;
ripemd160_transform (w, dgst); ripemd160_transform_transport_vector (w, dgst);
COMPARE_M_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]); COMPARE_M_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]);
} }
@ -503,7 +312,7 @@ __kernel void m06000_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
dgst[3] = RIPEMD160M_D; dgst[3] = RIPEMD160M_D;
dgst[4] = RIPEMD160M_E; dgst[4] = RIPEMD160M_E;
ripemd160_transform (w, dgst); ripemd160_transform_transport_vector (w, dgst);
COMPARE_S_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]); COMPARE_S_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]);
} }

201
OpenCL/m06000_a3-optimized.cl
View File

@ -11,202 +11,11 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_ripemd160.cl"
void ripemd160_transform (const u32x w[16], u32x dgst[5]) void ripemd160_transform_transport_vector (const u32x w[16], u32x dgst[5])
{ {
u32x a1 = dgst[0]; ripemd160_transform_vector (w + 0, w + 4, w + 8, w + 12, dgst);
u32x b1 = dgst[1];
u32x c1 = dgst[2];
u32x d1 = dgst[3];
u32x e1 = dgst[4];
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[ 0], RIPEMD160C00, RIPEMD160S00);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[ 1], RIPEMD160C00, RIPEMD160S01);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[ 2], RIPEMD160C00, RIPEMD160S02);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[ 3], RIPEMD160C00, RIPEMD160S03);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[ 4], RIPEMD160C00, RIPEMD160S04);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[ 5], RIPEMD160C00, RIPEMD160S05);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[ 6], RIPEMD160C00, RIPEMD160S06);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[ 7], RIPEMD160C00, RIPEMD160S07);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[ 8], RIPEMD160C00, RIPEMD160S08);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[ 9], RIPEMD160C00, RIPEMD160S09);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[10], RIPEMD160C00, RIPEMD160S0A);
RIPEMD160_STEP (RIPEMD160_F , e1, a1, b1, c1, d1, w[11], RIPEMD160C00, RIPEMD160S0B);
RIPEMD160_STEP (RIPEMD160_F , d1, e1, a1, b1, c1, w[12], RIPEMD160C00, RIPEMD160S0C);
RIPEMD160_STEP (RIPEMD160_F , c1, d1, e1, a1, b1, w[13], RIPEMD160C00, RIPEMD160S0D);
RIPEMD160_STEP (RIPEMD160_F , b1, c1, d1, e1, a1, w[14], RIPEMD160C00, RIPEMD160S0E);
RIPEMD160_STEP (RIPEMD160_F , a1, b1, c1, d1, e1, w[15], RIPEMD160C00, RIPEMD160S0F);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 7], RIPEMD160C10, RIPEMD160S10);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[ 4], RIPEMD160C10, RIPEMD160S11);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[13], RIPEMD160C10, RIPEMD160S12);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[ 1], RIPEMD160C10, RIPEMD160S13);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[10], RIPEMD160C10, RIPEMD160S14);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 6], RIPEMD160C10, RIPEMD160S15);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[15], RIPEMD160C10, RIPEMD160S16);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[ 3], RIPEMD160C10, RIPEMD160S17);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[12], RIPEMD160C10, RIPEMD160S18);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[ 0], RIPEMD160C10, RIPEMD160S19);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 9], RIPEMD160C10, RIPEMD160S1A);
RIPEMD160_STEP (RIPEMD160_Go, d1, e1, a1, b1, c1, w[ 5], RIPEMD160C10, RIPEMD160S1B);
RIPEMD160_STEP (RIPEMD160_Go, c1, d1, e1, a1, b1, w[ 2], RIPEMD160C10, RIPEMD160S1C);
RIPEMD160_STEP (RIPEMD160_Go, b1, c1, d1, e1, a1, w[14], RIPEMD160C10, RIPEMD160S1D);
RIPEMD160_STEP (RIPEMD160_Go, a1, b1, c1, d1, e1, w[11], RIPEMD160C10, RIPEMD160S1E);
RIPEMD160_STEP (RIPEMD160_Go, e1, a1, b1, c1, d1, w[ 8], RIPEMD160C10, RIPEMD160S1F);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[ 3], RIPEMD160C20, RIPEMD160S20);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[10], RIPEMD160C20, RIPEMD160S21);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[14], RIPEMD160C20, RIPEMD160S22);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[ 4], RIPEMD160C20, RIPEMD160S23);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 9], RIPEMD160C20, RIPEMD160S24);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[15], RIPEMD160C20, RIPEMD160S25);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[ 8], RIPEMD160C20, RIPEMD160S26);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[ 1], RIPEMD160C20, RIPEMD160S27);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[ 2], RIPEMD160C20, RIPEMD160S28);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 7], RIPEMD160C20, RIPEMD160S29);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[ 0], RIPEMD160C20, RIPEMD160S2A);
RIPEMD160_STEP (RIPEMD160_H , c1, d1, e1, a1, b1, w[ 6], RIPEMD160C20, RIPEMD160S2B);
RIPEMD160_STEP (RIPEMD160_H , b1, c1, d1, e1, a1, w[13], RIPEMD160C20, RIPEMD160S2C);
RIPEMD160_STEP (RIPEMD160_H , a1, b1, c1, d1, e1, w[11], RIPEMD160C20, RIPEMD160S2D);
RIPEMD160_STEP (RIPEMD160_H , e1, a1, b1, c1, d1, w[ 5], RIPEMD160C20, RIPEMD160S2E);
RIPEMD160_STEP (RIPEMD160_H , d1, e1, a1, b1, c1, w[12], RIPEMD160C20, RIPEMD160S2F);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 1], RIPEMD160C30, RIPEMD160S30);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[ 9], RIPEMD160C30, RIPEMD160S31);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[11], RIPEMD160C30, RIPEMD160S32);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[10], RIPEMD160C30, RIPEMD160S33);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 0], RIPEMD160C30, RIPEMD160S34);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 8], RIPEMD160C30, RIPEMD160S35);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[12], RIPEMD160C30, RIPEMD160S36);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[ 4], RIPEMD160C30, RIPEMD160S37);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[13], RIPEMD160C30, RIPEMD160S38);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 3], RIPEMD160C30, RIPEMD160S39);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 7], RIPEMD160C30, RIPEMD160S3A);
RIPEMD160_STEP (RIPEMD160_Io, b1, c1, d1, e1, a1, w[15], RIPEMD160C30, RIPEMD160S3B);
RIPEMD160_STEP (RIPEMD160_Io, a1, b1, c1, d1, e1, w[14], RIPEMD160C30, RIPEMD160S3C);
RIPEMD160_STEP (RIPEMD160_Io, e1, a1, b1, c1, d1, w[ 5], RIPEMD160C30, RIPEMD160S3D);
RIPEMD160_STEP (RIPEMD160_Io, d1, e1, a1, b1, c1, w[ 6], RIPEMD160C30, RIPEMD160S3E);
RIPEMD160_STEP (RIPEMD160_Io, c1, d1, e1, a1, b1, w[ 2], RIPEMD160C30, RIPEMD160S3F);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[ 4], RIPEMD160C40, RIPEMD160S40);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 0], RIPEMD160C40, RIPEMD160S41);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[ 5], RIPEMD160C40, RIPEMD160S42);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[ 9], RIPEMD160C40, RIPEMD160S43);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[ 7], RIPEMD160C40, RIPEMD160S44);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[12], RIPEMD160C40, RIPEMD160S45);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 2], RIPEMD160C40, RIPEMD160S46);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[10], RIPEMD160C40, RIPEMD160S47);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[14], RIPEMD160C40, RIPEMD160S48);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[ 1], RIPEMD160C40, RIPEMD160S49);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[ 3], RIPEMD160C40, RIPEMD160S4A);
RIPEMD160_STEP (RIPEMD160_J , a1, b1, c1, d1, e1, w[ 8], RIPEMD160C40, RIPEMD160S4B);
RIPEMD160_STEP (RIPEMD160_J , e1, a1, b1, c1, d1, w[11], RIPEMD160C40, RIPEMD160S4C);
RIPEMD160_STEP (RIPEMD160_J , d1, e1, a1, b1, c1, w[ 6], RIPEMD160C40, RIPEMD160S4D);
RIPEMD160_STEP (RIPEMD160_J , c1, d1, e1, a1, b1, w[15], RIPEMD160C40, RIPEMD160S4E);
RIPEMD160_STEP (RIPEMD160_J , b1, c1, d1, e1, a1, w[13], RIPEMD160C40, RIPEMD160S4F);
u32x a2 = dgst[0];
u32x b2 = dgst[1];
u32x c2 = dgst[2];
u32x d2 = dgst[3];
u32x e2 = dgst[4];
RIPEMD160_STEP_WORKAROUND_BUG (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 5], RIPEMD160C50, RIPEMD160S50);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[14], RIPEMD160C50, RIPEMD160S51);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 7], RIPEMD160C50, RIPEMD160S52);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[ 0], RIPEMD160C50, RIPEMD160S53);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 9], RIPEMD160C50, RIPEMD160S54);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[ 2], RIPEMD160C50, RIPEMD160S55);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[11], RIPEMD160C50, RIPEMD160S56);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 4], RIPEMD160C50, RIPEMD160S57);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[13], RIPEMD160C50, RIPEMD160S58);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 6], RIPEMD160C50, RIPEMD160S59);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[15], RIPEMD160C50, RIPEMD160S5A);
RIPEMD160_STEP (RIPEMD160_J , e2, a2, b2, c2, d2, w[ 8], RIPEMD160C50, RIPEMD160S5B);
RIPEMD160_STEP (RIPEMD160_J , d2, e2, a2, b2, c2, w[ 1], RIPEMD160C50, RIPEMD160S5C);
RIPEMD160_STEP (RIPEMD160_J , c2, d2, e2, a2, b2, w[10], RIPEMD160C50, RIPEMD160S5D);
RIPEMD160_STEP (RIPEMD160_J , b2, c2, d2, e2, a2, w[ 3], RIPEMD160C50, RIPEMD160S5E);
RIPEMD160_STEP (RIPEMD160_J , a2, b2, c2, d2, e2, w[12], RIPEMD160C50, RIPEMD160S5F);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 6], RIPEMD160C60, RIPEMD160S60);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[11], RIPEMD160C60, RIPEMD160S61);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[ 3], RIPEMD160C60, RIPEMD160S62);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[ 7], RIPEMD160C60, RIPEMD160S63);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[ 0], RIPEMD160C60, RIPEMD160S64);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[13], RIPEMD160C60, RIPEMD160S65);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[ 5], RIPEMD160C60, RIPEMD160S66);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[10], RIPEMD160C60, RIPEMD160S67);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[14], RIPEMD160C60, RIPEMD160S68);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[15], RIPEMD160C60, RIPEMD160S69);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 8], RIPEMD160C60, RIPEMD160S6A);
RIPEMD160_STEP (RIPEMD160_Io, d2, e2, a2, b2, c2, w[12], RIPEMD160C60, RIPEMD160S6B);
RIPEMD160_STEP (RIPEMD160_Io, c2, d2, e2, a2, b2, w[ 4], RIPEMD160C60, RIPEMD160S6C);
RIPEMD160_STEP (RIPEMD160_Io, b2, c2, d2, e2, a2, w[ 9], RIPEMD160C60, RIPEMD160S6D);
RIPEMD160_STEP (RIPEMD160_Io, a2, b2, c2, d2, e2, w[ 1], RIPEMD160C60, RIPEMD160S6E);
RIPEMD160_STEP (RIPEMD160_Io, e2, a2, b2, c2, d2, w[ 2], RIPEMD160C60, RIPEMD160S6F);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[15], RIPEMD160C70, RIPEMD160S70);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 5], RIPEMD160C70, RIPEMD160S71);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[ 1], RIPEMD160C70, RIPEMD160S72);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[ 3], RIPEMD160C70, RIPEMD160S73);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 7], RIPEMD160C70, RIPEMD160S74);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[14], RIPEMD160C70, RIPEMD160S75);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 6], RIPEMD160C70, RIPEMD160S76);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[ 9], RIPEMD160C70, RIPEMD160S77);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[11], RIPEMD160C70, RIPEMD160S78);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 8], RIPEMD160C70, RIPEMD160S79);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[12], RIPEMD160C70, RIPEMD160S7A);
RIPEMD160_STEP (RIPEMD160_H , c2, d2, e2, a2, b2, w[ 2], RIPEMD160C70, RIPEMD160S7B);
RIPEMD160_STEP (RIPEMD160_H , b2, c2, d2, e2, a2, w[10], RIPEMD160C70, RIPEMD160S7C);
RIPEMD160_STEP (RIPEMD160_H , a2, b2, c2, d2, e2, w[ 0], RIPEMD160C70, RIPEMD160S7D);
RIPEMD160_STEP (RIPEMD160_H , e2, a2, b2, c2, d2, w[ 4], RIPEMD160C70, RIPEMD160S7E);
RIPEMD160_STEP (RIPEMD160_H , d2, e2, a2, b2, c2, w[13], RIPEMD160C70, RIPEMD160S7F);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[ 8], RIPEMD160C80, RIPEMD160S80);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[ 6], RIPEMD160C80, RIPEMD160S81);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 4], RIPEMD160C80, RIPEMD160S82);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 1], RIPEMD160C80, RIPEMD160S83);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[ 3], RIPEMD160C80, RIPEMD160S84);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[11], RIPEMD160C80, RIPEMD160S85);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[15], RIPEMD160C80, RIPEMD160S86);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 0], RIPEMD160C80, RIPEMD160S87);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 5], RIPEMD160C80, RIPEMD160S88);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[12], RIPEMD160C80, RIPEMD160S89);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[ 2], RIPEMD160C80, RIPEMD160S8A);
RIPEMD160_STEP (RIPEMD160_Go, b2, c2, d2, e2, a2, w[13], RIPEMD160C80, RIPEMD160S8B);
RIPEMD160_STEP (RIPEMD160_Go, a2, b2, c2, d2, e2, w[ 9], RIPEMD160C80, RIPEMD160S8C);
RIPEMD160_STEP (RIPEMD160_Go, e2, a2, b2, c2, d2, w[ 7], RIPEMD160C80, RIPEMD160S8D);
RIPEMD160_STEP (RIPEMD160_Go, d2, e2, a2, b2, c2, w[10], RIPEMD160C80, RIPEMD160S8E);
RIPEMD160_STEP (RIPEMD160_Go, c2, d2, e2, a2, b2, w[14], RIPEMD160C80, RIPEMD160S8F);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[12], RIPEMD160C90, RIPEMD160S90);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[15], RIPEMD160C90, RIPEMD160S91);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[10], RIPEMD160C90, RIPEMD160S92);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 4], RIPEMD160C90, RIPEMD160S93);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 1], RIPEMD160C90, RIPEMD160S94);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[ 5], RIPEMD160C90, RIPEMD160S95);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[ 8], RIPEMD160C90, RIPEMD160S96);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[ 7], RIPEMD160C90, RIPEMD160S97);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 6], RIPEMD160C90, RIPEMD160S98);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 2], RIPEMD160C90, RIPEMD160S99);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[13], RIPEMD160C90, RIPEMD160S9A);
RIPEMD160_STEP (RIPEMD160_F , a2, b2, c2, d2, e2, w[14], RIPEMD160C90, RIPEMD160S9B);
RIPEMD160_STEP (RIPEMD160_F , e2, a2, b2, c2, d2, w[ 0], RIPEMD160C90, RIPEMD160S9C);
RIPEMD160_STEP (RIPEMD160_F , d2, e2, a2, b2, c2, w[ 3], RIPEMD160C90, RIPEMD160S9D);
RIPEMD160_STEP (RIPEMD160_F , c2, d2, e2, a2, b2, w[ 9], RIPEMD160C90, RIPEMD160S9E);
RIPEMD160_STEP (RIPEMD160_F , b2, c2, d2, e2, a2, w[11], RIPEMD160C90, RIPEMD160S9F);
const u32x a = dgst[1] + c1 + d2;
const u32x b = dgst[2] + d1 + e2;
const u32x c = dgst[3] + e1 + a2;
const u32x d = dgst[4] + a1 + b2;
const u32x e = dgst[0] + b1 + c2;
dgst[0] = a;
dgst[1] = b;
dgst[2] = c;
dgst[3] = d;
dgst[4] = e;
} }
void m06000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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) void m06000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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)
@ -261,7 +70,7 @@ void m06000m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
dgst[3] = RIPEMD160M_D; dgst[3] = RIPEMD160M_D;
dgst[4] = RIPEMD160M_E; dgst[4] = RIPEMD160M_E;
ripemd160_transform (w, dgst); ripemd160_transform_transport_vector (w, dgst);
COMPARE_M_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]); COMPARE_M_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]);
} }
@ -331,7 +140,7 @@ void m06000s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
dgst[3] = RIPEMD160M_D; dgst[3] = RIPEMD160M_D;
dgst[4] = RIPEMD160M_E; dgst[4] = RIPEMD160M_E;
ripemd160_transform (w, dgst); ripemd160_transform_transport_vector (w, dgst);
COMPARE_S_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]); COMPARE_S_SIMD (dgst[0], dgst[1], dgst[2], dgst[3]);
} }

1333
OpenCL/m06100_a0-optimized.cl
View File

File diff suppressed because it is too large Load Diff

1333
OpenCL/m06100_a1-optimized.cl
View File

File diff suppressed because it is too large Load Diff

1333
OpenCL/m06100_a3-optimized.cl
View File

File diff suppressed because it is too large Load Diff

141
OpenCL/m07300_a0-optimized.cl
View File

@ -13,134 +13,7 @@
#include "inc_rp.h" #include "inc_rp.h"
#include "inc_rp.cl" #include "inc_rp.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_sha1.cl"
void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
{
u32x A = digest[0];
u32x B = digest[1];
u32x C = digest[2];
u32x D = digest[3];
u32x E = digest[4];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w0_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w1_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w2_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w3_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w4_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w5_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w6_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w7_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w8_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w9_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wa_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, wb_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, wc_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, wd_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, we_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, E, A, B, C, D, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, D, E, A, B, C, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, C, D, E, A, B, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, B, C, D, E, A, w3_t);
#undef K
#define K SHA1C01
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w7_t);
#undef K
#define K SHA1C02
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wb_t);
#undef K
#define K SHA1C03
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wf_t);
digest[0] += A;
digest[1] += B;
digest[2] += C;
digest[3] += D;
digest[4] += E;
}
void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5]) void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
{ {
@ -167,7 +40,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
ipad[3] = SHA1M_D; ipad[3] = SHA1M_D;
ipad[4] = SHA1M_E; ipad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -192,7 +65,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
opad[3] = SHA1M_D; opad[3] = SHA1M_D;
opad[4] = SHA1M_E; opad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, opad); sha1_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5]) void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
@ -203,7 +76,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = ipad[3]; digest[3] = ipad[3];
digest[4] = ipad[4]; digest[4] = ipad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -228,7 +101,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = opad[3]; digest[3] = opad[3];
digest[4] = opad[4]; digest[4] = opad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
} }
__kernel void m07300_m04 (__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 rakp_t *rakp_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 u32 gid_max) __kernel void m07300_m04 (__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 rakp_t *rakp_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 u32 gid_max)
@ -322,7 +195,7 @@ __kernel void m07300_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14]; w3[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14];
w3[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15]; w3[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15];
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0]; w0[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0];
@ -461,7 +334,7 @@ __kernel void m07300_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14]; w3[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14];
w3[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15]; w3[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15];
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0]; w0[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0];

141
OpenCL/m07300_a1-optimized.cl
View File

@ -11,134 +11,7 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_sha1.cl"
void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
{
u32x A = digest[0];
u32x B = digest[1];
u32x C = digest[2];
u32x D = digest[3];
u32x E = digest[4];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w0_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w1_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w2_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w3_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w4_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w5_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w6_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w7_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w8_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w9_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wa_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, wb_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, wc_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, wd_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, we_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, E, A, B, C, D, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, D, E, A, B, C, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, C, D, E, A, B, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, B, C, D, E, A, w3_t);
#undef K
#define K SHA1C01
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w7_t);
#undef K
#define K SHA1C02
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wb_t);
#undef K
#define K SHA1C03
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wf_t);
digest[0] += A;
digest[1] += B;
digest[2] += C;
digest[3] += D;
digest[4] += E;
}
void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5]) void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
{ {
@ -165,7 +38,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
ipad[3] = SHA1M_D; ipad[3] = SHA1M_D;
ipad[4] = SHA1M_E; ipad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -190,7 +63,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
opad[3] = SHA1M_D; opad[3] = SHA1M_D;
opad[4] = SHA1M_E; opad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, opad); sha1_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5]) void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
@ -201,7 +74,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = ipad[3]; digest[3] = ipad[3];
digest[4] = ipad[4]; digest[4] = ipad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -226,7 +99,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = opad[3]; digest[3] = opad[3];
digest[4] = opad[4]; digest[4] = opad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
} }
__kernel void m07300_m04 (__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 rakp_t *rakp_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 u32 gid_max) __kernel void m07300_m04 (__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 rakp_t *rakp_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 u32 gid_max)
@ -388,7 +261,7 @@ __kernel void m07300_m04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14]; w3[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14];
w3[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15]; w3[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15];
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0]; w0[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0];
@ -595,7 +468,7 @@ __kernel void m07300_s04 (__global pw_t *pws, __global const kernel_rule_t *rule
w3[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14]; w3[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14];
w3[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15]; w3[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15];
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
} }
w0[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0]; w0[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0];

141
OpenCL/m07300_a3-optimized.cl
View File

@ -11,134 +11,7 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_sha1.cl"
void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
{
u32x A = digest[0];
u32x B = digest[1];
u32x C = digest[2];
u32x D = digest[3];
u32x E = digest[4];
u32x w0_t = w0[0];
u32x w1_t = w0[1];
u32x w2_t = w0[2];
u32x w3_t = w0[3];
u32x w4_t = w1[0];
u32x w5_t = w1[1];
u32x w6_t = w1[2];
u32x w7_t = w1[3];
u32x w8_t = w2[0];
u32x w9_t = w2[1];
u32x wa_t = w2[2];
u32x wb_t = w2[3];
u32x wc_t = w3[0];
u32x wd_t = w3[1];
u32x we_t = w3[2];
u32x wf_t = w3[3];
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w0_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w1_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w2_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w3_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w4_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, w5_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, w6_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, w7_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, w8_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, w9_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wa_t);
SHA1_STEP (SHA1_F0o, E, A, B, C, D, wb_t);
SHA1_STEP (SHA1_F0o, D, E, A, B, C, wc_t);
SHA1_STEP (SHA1_F0o, C, D, E, A, B, wd_t);
SHA1_STEP (SHA1_F0o, B, C, D, E, A, we_t);
SHA1_STEP (SHA1_F0o, A, B, C, D, E, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, E, A, B, C, D, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, D, E, A, B, C, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, C, D, E, A, B, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, B, C, D, E, A, w3_t);
#undef K
#define K SHA1C01
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w7_t);
#undef K
#define K SHA1C02
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, A, B, C, D, E, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, E, A, B, C, D, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, D, E, A, B, C, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, C, D, E, A, B, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, B, C, D, E, A, wb_t);
#undef K
#define K SHA1C03
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, wf_t);
w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w0_t);
w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w1_t);
w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w2_t);
w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w3_t);
w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w4_t);
w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, w5_t);
w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, w6_t);
w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, w7_t);
w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, w8_t);
w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, w9_t);
wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wa_t);
wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, A, B, C, D, E, wb_t);
wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, E, A, B, C, D, wc_t);
wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, D, E, A, B, C, wd_t);
we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, C, D, E, A, B, we_t);
wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, B, C, D, E, A, wf_t);
digest[0] += A;
digest[1] += B;
digest[2] += C;
digest[3] += D;
digest[4] += E;
}
void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5]) void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5])
{ {
@ -165,7 +38,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
ipad[3] = SHA1M_D; ipad[3] = SHA1M_D;
ipad[4] = SHA1M_E; ipad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, ipad); sha1_transform_vector (w0, w1, w2, w3, ipad);
w0[0] = w0[0] ^ 0x6a6a6a6a; w0[0] = w0[0] ^ 0x6a6a6a6a;
w0[1] = w0[1] ^ 0x6a6a6a6a; w0[1] = w0[1] ^ 0x6a6a6a6a;
@ -190,7 +63,7 @@ void hmac_sha1_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
opad[3] = SHA1M_D; opad[3] = SHA1M_D;
opad[4] = SHA1M_E; opad[4] = SHA1M_E;
sha1_transform (w0, w1, w2, w3, opad); sha1_transform_vector (w0, w1, w2, w3, opad);
} }
void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5]) void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5], u32x opad[5], u32x digest[5])
@ -201,7 +74,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = ipad[3]; digest[3] = ipad[3];
digest[4] = ipad[4]; digest[4] = ipad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
w0[0] = digest[0]; w0[0] = digest[0];
w0[1] = digest[1]; w0[1] = digest[1];
@ -226,7 +99,7 @@ void hmac_sha1_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[5]
digest[3] = opad[3]; digest[3] = opad[3];
digest[4] = opad[4]; digest[4] = opad[4];
sha1_transform (w0, w1, w2, w3, digest); sha1_transform_vector (w0, w1, w2, w3, digest);
} }
void m07300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 rakp_t *rakp_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) void m07300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 rakp_t *rakp_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)
@ -311,7 +184,7 @@ void m07300m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14]; w3_t[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14];
w3_t[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15]; w3_t[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15];
sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha1_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0]; w0_t[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0];
@ -433,7 +306,7 @@ void m07300s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl
w3_t[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14]; w3_t[2] = rakp_bufs[digests_offset].salt_buf[esalt_off + 14];
w3_t[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15]; w3_t[3] = rakp_bufs[digests_offset].salt_buf[esalt_off + 15];
sha1_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha1_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad);
} }
w0_t[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0]; w0_t[0] = rakp_bufs[digests_offset].salt_buf[esalt_off + 0];

170
OpenCL/m07400-optimized.cl
View File

@ -8,140 +8,36 @@
#include "inc_hash_functions.cl" #include "inc_hash_functions.cl"
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_hash_sha256.cl"
#define COMPARE_S "inc_comp_single.cl" #define COMPARE_S "inc_comp_single.cl"
#define COMPARE_M "inc_comp_multi.cl" #define COMPARE_M "inc_comp_multi.cl"
__constant u32a k_sha256[64] = void sha256_transform_transport (const u32 w[16], u32 digest[8])
{ {
SHA256C00, SHA256C01, SHA256C02, SHA256C03, u32 w0[4];
SHA256C04, SHA256C05, SHA256C06, SHA256C07, u32 w1[4];
SHA256C08, SHA256C09, SHA256C0a, SHA256C0b, u32 w2[4];
SHA256C0c, SHA256C0d, SHA256C0e, SHA256C0f, u32 w3[4];
SHA256C10, SHA256C11, SHA256C12, SHA256C13,
SHA256C14, SHA256C15, SHA256C16, SHA256C17,
SHA256C18, SHA256C19, SHA256C1a, SHA256C1b,
SHA256C1c, SHA256C1d, SHA256C1e, SHA256C1f,
SHA256C20, SHA256C21, SHA256C22, SHA256C23,
SHA256C24, SHA256C25, SHA256C26, SHA256C27,
SHA256C28, SHA256C29, SHA256C2a, SHA256C2b,
SHA256C2c, SHA256C2d, SHA256C2e, SHA256C2f,
SHA256C30, SHA256C31, SHA256C32, SHA256C33,
SHA256C34, SHA256C35, SHA256C36, SHA256C37,
SHA256C38, SHA256C39, SHA256C3a, SHA256C3b,
SHA256C3c, SHA256C3d, SHA256C3e, SHA256C3f,
};
void sha256_transform (const u32 w[16], u32 digest[8]) w0[0] = swap32_S (w[ 0]);
{ w0[1] = swap32_S (w[ 1]);
u32 a = digest[0]; w0[2] = swap32_S (w[ 2]);
u32 b = digest[1]; w0[3] = swap32_S (w[ 3]);
u32 c = digest[2]; w1[0] = swap32_S (w[ 4]);
u32 d = digest[3]; w1[1] = swap32_S (w[ 5]);
u32 e = digest[4]; w1[2] = swap32_S (w[ 6]);
u32 f = digest[5]; w1[3] = swap32_S (w[ 7]);
u32 g = digest[6]; w2[0] = swap32_S (w[ 8]);
u32 h = digest[7]; w2[1] = swap32_S (w[ 9]);
w2[2] = swap32_S (w[10]);
w2[3] = swap32_S (w[11]);
w3[0] = swap32_S (w[12]);
w3[1] = swap32_S (w[13]);
w3[2] = swap32_S (w[14]);
w3[3] = swap32_S (w[15]);
u32 w0_t = swap32 (w[ 0]); sha256_transform (w0, w1, w2, w3, digest);
u32 w1_t = swap32 (w[ 1]);
u32 w2_t = swap32 (w[ 2]);
u32 w3_t = swap32 (w[ 3]);
u32 w4_t = swap32 (w[ 4]);
u32 w5_t = swap32 (w[ 5]);
u32 w6_t = swap32 (w[ 6]);
u32 w7_t = swap32 (w[ 7]);
u32 w8_t = swap32 (w[ 8]);
u32 w9_t = swap32 (w[ 9]);
u32 wa_t = swap32 (w[10]);
u32 wb_t = swap32 (w[11]);
u32 wc_t = swap32 (w[12]);
u32 wd_t = swap32 (w[13]);
u32 we_t = swap32 (w[14]);
u32 wf_t = swap32 (w[15]);
#define ROUND_EXPAND() \
{ \
w0_t = SHA256_EXPAND (we_t, w9_t, w1_t, w0_t); \
w1_t = SHA256_EXPAND (wf_t, wa_t, w2_t, w1_t); \
w2_t = SHA256_EXPAND (w0_t, wb_t, w3_t, w2_t); \
w3_t = SHA256_EXPAND (w1_t, wc_t, w4_t, w3_t); \
w4_t = SHA256_EXPAND (w2_t, wd_t, w5_t, w4_t); \
w5_t = SHA256_EXPAND (w3_t, we_t, w6_t, w5_t); \
w6_t = SHA256_EXPAND (w4_t, wf_t, w7_t, w6_t); \
w7_t = SHA256_EXPAND (w5_t, w0_t, w8_t, w7_t); \
w8_t = SHA256_EXPAND (w6_t, w1_t, w9_t, w8_t); \
w9_t = SHA256_EXPAND (w7_t, w2_t, wa_t, w9_t); \
wa_t = SHA256_EXPAND (w8_t, w3_t, wb_t, wa_t); \
wb_t = SHA256_EXPAND (w9_t, w4_t, wc_t, wb_t); \
wc_t = SHA256_EXPAND (wa_t, w5_t, wd_t, wc_t); \
wd_t = SHA256_EXPAND (wb_t, w6_t, we_t, wd_t); \
we_t = SHA256_EXPAND (wc_t, w7_t, wf_t, we_t); \
wf_t = SHA256_EXPAND (wd_t, w8_t, w0_t, wf_t); \
}
#define ROUND_STEP(i) \
{ \
SHA256_STEP (SHA256_F0o, SHA256_F1o, a, b, c, d, e, f, g, h, w0_t, k_sha256[i + 0]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, h, a, b, c, d, e, f, g, w1_t, k_sha256[i + 1]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, g, h, a, b, c, d, e, f, w2_t, k_sha256[i + 2]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, f, g, h, a, b, c, d, e, w3_t, k_sha256[i + 3]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, e, f, g, h, a, b, c, d, w4_t, k_sha256[i + 4]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, d, e, f, g, h, a, b, c, w5_t, k_sha256[i + 5]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, c, d, e, f, g, h, a, b, w6_t, k_sha256[i + 6]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, b, c, d, e, f, g, h, a, w7_t, k_sha256[i + 7]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, a, b, c, d, e, f, g, h, w8_t, k_sha256[i + 8]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, h, a, b, c, d, e, f, g, w9_t, k_sha256[i + 9]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, g, h, a, b, c, d, e, f, wa_t, k_sha256[i + 10]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, f, g, h, a, b, c, d, e, wb_t, k_sha256[i + 11]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, e, f, g, h, a, b, c, d, wc_t, k_sha256[i + 12]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, d, e, f, g, h, a, b, c, wd_t, k_sha256[i + 13]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, c, d, e, f, g, h, a, b, we_t, k_sha256[i + 14]); \
SHA256_STEP (SHA256_F0o, SHA256_F1o, b, c, d, e, f, g, h, a, wf_t, k_sha256[i + 15]); \
}
ROUND_STEP (0);
#ifdef _unroll
#pragma unroll
#endif
for (int i = 16; i < 64; i += 16)
{
ROUND_EXPAND (); ROUND_STEP (i);
}
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
digest[4] += e;
digest[5] += f;
digest[6] += g;
digest[7] += h;
}
void sha256_transform_no14 (const u32 w[16], u32 digest[8])
{
u32 w_t[16];
w_t[ 0] = w[ 0];
w_t[ 1] = w[ 1];
w_t[ 2] = w[ 2];
w_t[ 3] = w[ 3];
w_t[ 4] = w[ 4];
w_t[ 5] = w[ 5];
w_t[ 6] = w[ 6];
w_t[ 7] = w[ 7];
w_t[ 8] = w[ 8];
w_t[ 9] = w[ 9];
w_t[10] = w[10];
w_t[11] = w[11];
w_t[12] = w[12];
w_t[13] = w[13];
w_t[14] = 0;
w_t[15] = w[15];
sha256_transform (w_t, digest);
} }
void init_ctx (u32 digest[8]) void init_ctx (u32 digest[8])
@ -483,7 +379,7 @@ u32 memcat16c (u32 block[16], const u32 offset, const u32 append[4], const u32 a
{ {
new_len -= 64; new_len -= 64;
sha256_transform (block, digest); sha256_transform_transport (block, digest);
bzero16 (block); bzero16 (block);
@ -865,7 +761,7 @@ __kernel void m07400_init (__global pw_t *pws, __global const kernel_rule_t *rul
init_ctx (alt_result); init_ctx (alt_result);
sha256_transform (block, alt_result); sha256_transform_transport (block, alt_result);
bswap8 (alt_result); bswap8 (alt_result);
@ -938,14 +834,14 @@ __kernel void m07400_init (__global pw_t *pws, __global const kernel_rule_t *rul
if (block_len >= 56) if (block_len >= 56)
{ {
sha256_transform (block, alt_result); sha256_transform_transport (block, alt_result);
bzero16 (block); bzero16 (block);
} }
block[15] = swap32 (transform_len * 8); block[15] = swap32 (transform_len * 8);
sha256_transform (block, alt_result); sha256_transform_transport (block, alt_result);
bswap8 (alt_result); bswap8 (alt_result);
@ -983,14 +879,14 @@ __kernel void m07400_init (__global pw_t *pws, __global const kernel_rule_t *rul
if (block_len >= 56) if (block_len >= 56)
{ {
sha256_transform (block, p_bytes); sha256_transform_transport (block, p_bytes);
bzero16 (block); bzero16 (block);
} }
block[15] = swap32 (transform_len * 8); block[15] = swap32 (transform_len * 8);
sha256_transform (block, p_bytes); sha256_transform_transport (block, p_bytes);
bswap8 (p_bytes); bswap8 (p_bytes);
@ -1026,14 +922,14 @@ __kernel void m07400_init (__global pw_t *pws, __global const kernel_rule_t *rul
if (block_len >= 56) if (block_len >= 56)
{ {
sha256_transform (block, s_bytes); sha256_transform_transport (block, s_bytes);
bzero16 (block); bzero16 (block);
} }
block[15] = swap32 (transform_len * 8); block[15] = swap32 (transform_len * 8);
sha256_transform (block, s_bytes); sha256_transform_transport (block, s_bytes);
bswap8 (s_bytes); bswap8 (s_bytes);
@ -1162,7 +1058,7 @@ __kernel void m07400_loop (__global pw_t *pws, __global const kernel_rule_t *rul
if (block_len >= 56) if (block_len >= 56)
{ {
sha256_transform (block, tmp); sha256_transform_transport (block, tmp);
block[ 0] = block[16]; block[ 0] = block[16];
block[ 1] = block[17]; block[ 1] = block[17];
@ -1184,7 +1080,7 @@ __kernel void m07400_loop (__global pw_t *pws, __global const kernel_rule_t *rul
block[15] = swap32 (block_len * 8); block[15] = swap32 (block_len * 8);
sha256_transform_no14 (block, tmp); sha256_transform_transport (block, tmp);
bswap8 (tmp); bswap8 (tmp);

164
OpenCL/m07500_a0-optimized.cl
View File

@ -14,6 +14,8 @@
#include "inc_rp.h" #include "inc_rp.h"
#include "inc_rp.cl" #include "inc_rp.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md4.cl"
#include "inc_hash_md5.cl"
typedef struct typedef struct
{ {
@ -168,168 +170,6 @@ int decrypt_and_check (__local RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8
return 1; return 1;
} }
void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 c = digest[2];
u32 d = digest[3];
MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 c = digest[2];
u32 d = digest[3];
u32 w0_t = w0[0];
u32 w1_t = w0[1];
u32 w2_t = w0[2];
u32 w3_t = w0[3];
u32 w4_t = w1[0];
u32 w5_t = w1[1];
u32 w6_t = w1[2];
u32 w7_t = w1[3];
u32 w8_t = w2[0];
u32 w9_t = w2[1];
u32 wa_t = w2[2];
u32 wb_t = w2[3];
u32 wc_t = w3[0];
u32 wd_t = w3[1];
u32 we_t = w3[2];
u32 wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4]) void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
{ {
w0[0] = w0[0] ^ 0x36363636; w0[0] = w0[0] ^ 0x36363636;

164
OpenCL/m07500_a1-optimized.cl
View File

@ -12,6 +12,8 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md4.cl"
#include "inc_hash_md5.cl"
typedef struct typedef struct
{ {
@ -166,168 +168,6 @@ int decrypt_and_check (__local RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8
return 1; return 1;
} }
void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 c = digest[2];
u32 d = digest[3];
MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 c = digest[2];
u32 d = digest[3];
u32 w0_t = w0[0];
u32 w1_t = w0[1];
u32 w2_t = w0[2];
u32 w3_t = w0[3];
u32 w4_t = w1[0];
u32 w5_t = w1[1];
u32 w6_t = w1[2];
u32 w7_t = w1[3];
u32 w8_t = w2[0];
u32 w9_t = w2[1];
u32 wa_t = w2[2];
u32 wb_t = w2[3];
u32 wc_t = w3[0];
u32 wd_t = w3[1];
u32 we_t = w3[2];
u32 wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4]) void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
{ {
w0[0] = w0[0] ^ 0x36363636; w0[0] = w0[0] ^ 0x36363636;

164
OpenCL/m07500_a3-optimized.cl
View File

@ -12,6 +12,8 @@
#include "inc_types.cl" #include "inc_types.cl"
#include "inc_common.cl" #include "inc_common.cl"
#include "inc_simd.cl" #include "inc_simd.cl"
#include "inc_hash_md4.cl"
#include "inc_hash_md5.cl"
typedef struct typedef struct
{ {
@ -166,168 +168,6 @@ int decrypt_and_check (__local RC4_KEY *rc4_key, u32 data[4], u32 timestamp_ct[8
return 1; return 1;
} }
void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 c = digest[2];
u32 d = digest[3];
MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
{
u32 a = digest[0];
u32 b = digest[1];
u32 c = digest[2];
u32 d = digest[3];
u32 w0_t = w0[0];
u32 w1_t = w0[1];
u32 w2_t = w0[2];
u32 w3_t = w0[3];
u32 w4_t = w1[0];
u32 w5_t = w1[1];
u32 w6_t = w1[2];
u32 w7_t = w1[3];
u32 w8_t = w2[0];
u32 w9_t = w2[1];
u32 wa_t = w2[2];
u32 wb_t = w2[3];
u32 wc_t = w3[0];
u32 wd_t = w3[1];
u32 we_t = w3[2];
u32 wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
digest[0] += a;
digest[1] += b;
digest[2] += c;
digest[3] += d;
}
void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4]) void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
{ {
w0[0] = w0[0] ^ 0x36363636; w0[0] = w0[0] ^ 0x36363636;