arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-11-22 16:18:09 +00:00
Code Issues Releases Wiki Activity

Cleanup -m 1100 kernels to latest standard

Browse Source
This commit is contained in:
Jens Steube 2016-03-04 16:54:28 +01:00
parent 116d7620c5
commit 555e339674
3 changed files with 653 additions and 743 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

532
OpenCL/m01100_a0.cl
View File

@ -39,18 +39,16 @@ __kernel void m01100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 pw_buf0[4]; u32 pw_buf0[4];
pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4]; u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4]; pw_buf0[0] = pws[gid].i[0];
pw_buf1[1] = pws[gid].i[ 5]; pw_buf0[1] = pws[gid].i[1];
pw_buf1[2] = pws[gid].i[ 6]; pw_buf0[2] = pws[gid].i[2];
pw_buf1[3] = pws[gid].i[ 7]; pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
const u32 pw_len = pws[gid].pw_len; const u32 pw_len = pws[gid].pw_len;
@ -59,25 +57,21 @@ __kernel void m01100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
*/ */
u32 salt_buf0[4]; u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0]; salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1]; salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2]; salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3]; salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
u32 salt_buf1[4];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4]; salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5]; salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6]; salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7]; salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
u32 salt_buf2[4];
salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8]; salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9]; salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
salt_buf2[2] = 0; salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
salt_buf2[3] = 0; salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
const u32 salt_len = salt_bufs[salt_pos].salt_len; const u32 salt_len = salt_bufs[salt_pos].salt_len;
@ -96,149 +90,145 @@ __kernel void m01100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
append_0x80_2x4_VV (w0, w1, out_len); append_0x80_2x4_VV (w0, w1, out_len);
u32x w0_t[4]; make_unicode (w1, w2, w3);
u32x w1_t[4]; make_unicode (w0, w0, w1);
u32x w2_t[4];
u32x w3_t[4];
make_unicode (w0, w0_t, w1_t); w3[2] = out_len * 2 * 8;
make_unicode (w1, w2_t, w3_t); w3[3] = 0;
w3_t[2] = out_len * 8 * 2;
u32x a = MD4M_A; u32x a = MD4M_A;
u32x b = MD4M_B; u32x b = MD4M_B;
u32x c = MD4M_C; u32x c = MD4M_C;
u32x d = MD4M_D; u32x d = MD4M_D;
MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
a += MD4M_A; a += MD4M_A;
b += MD4M_B; b += MD4M_B;
c += MD4M_C; c += MD4M_C;
d += MD4M_D; d += MD4M_D;
w0_t[0] = a; w0[0] = a;
w0_t[1] = b; w0[1] = b;
w0_t[2] = c; w0[2] = c;
w0_t[3] = d; w0[3] = d;
w1_t[0] = salt_buf0[0]; w1[0] = salt_buf0[0];
w1_t[1] = salt_buf0[1]; w1[1] = salt_buf0[1];
w1_t[2] = salt_buf0[2]; w1[2] = salt_buf0[2];
w1_t[3] = salt_buf0[3]; w1[3] = salt_buf0[3];
w2_t[0] = salt_buf1[0]; w2[0] = salt_buf1[0];
w2_t[1] = salt_buf1[1]; w2[1] = salt_buf1[1];
w2_t[2] = salt_buf1[2]; w2[2] = salt_buf1[2];
w2_t[3] = salt_buf1[3]; w2[3] = salt_buf1[3];
w3_t[0] = salt_buf2[0]; w3[0] = salt_buf2[0];
w3_t[1] = salt_buf2[1]; w3[1] = salt_buf2[1];
w3_t[2] = (16 + salt_len) * 8; w3[2] = (16 + salt_len) * 8;
w3_t[3] = 0; w3[3] = 0;
a = MD4M_A; a = MD4M_A;
b = MD4M_B; b = MD4M_B;
c = MD4M_C; c = MD4M_C;
d = MD4M_D; d = MD4M_D;
MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
COMPARE_M_SIMD (a, d, c, b); COMPARE_M_SIMD (a, d, c, b);
} }
@ -269,18 +259,16 @@ __kernel void m01100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 pw_buf0[4]; u32 pw_buf0[4];
pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4]; u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4]; pw_buf0[0] = pws[gid].i[0];
pw_buf1[1] = pws[gid].i[ 5]; pw_buf0[1] = pws[gid].i[1];
pw_buf1[2] = pws[gid].i[ 6]; pw_buf0[2] = pws[gid].i[2];
pw_buf1[3] = pws[gid].i[ 7]; pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
const u32 pw_len = pws[gid].pw_len; const u32 pw_len = pws[gid].pw_len;
@ -301,25 +289,21 @@ __kernel void m01100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
*/ */
u32 salt_buf0[4]; u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0]; salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1]; salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2]; salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3]; salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
u32 salt_buf1[4];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4]; salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5]; salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6]; salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7]; salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
u32 salt_buf2[4];
salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8]; salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9]; salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
salt_buf2[2] = 0; salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
salt_buf2[3] = 0; salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
const u32 salt_len = salt_bufs[salt_pos].salt_len; const u32 salt_len = salt_bufs[salt_pos].salt_len;
@ -338,152 +322,148 @@ __kernel void m01100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
append_0x80_2x4_VV (w0, w1, out_len); append_0x80_2x4_VV (w0, w1, out_len);
u32x w0_t[4]; make_unicode (w1, w2, w3);
u32x w1_t[4]; make_unicode (w0, w0, w1);
u32x w2_t[4];
u32x w3_t[4];
make_unicode (w0, w0_t, w1_t); w3[2] = out_len * 2 * 8;
make_unicode (w1, w2_t, w3_t); w3[3] = 0;
w3_t[2] = out_len * 8 * 2;
u32x a = MD4M_A; u32x a = MD4M_A;
u32x b = MD4M_B; u32x b = MD4M_B;
u32x c = MD4M_C; u32x c = MD4M_C;
u32x d = MD4M_D; u32x d = MD4M_D;
MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
a += MD4M_A; a += MD4M_A;
b += MD4M_B; b += MD4M_B;
c += MD4M_C; c += MD4M_C;
d += MD4M_D; d += MD4M_D;
w0_t[0] = a; w0[0] = a;
w0_t[1] = b; w0[1] = b;
w0_t[2] = c; w0[2] = c;
w0_t[3] = d; w0[3] = d;
w1_t[0] = salt_buf0[0]; w1[0] = salt_buf0[0];
w1_t[1] = salt_buf0[1]; w1[1] = salt_buf0[1];
w1_t[2] = salt_buf0[2]; w1[2] = salt_buf0[2];
w1_t[3] = salt_buf0[3]; w1[3] = salt_buf0[3];
w2_t[0] = salt_buf1[0]; w2[0] = salt_buf1[0];
w2_t[1] = salt_buf1[1]; w2[1] = salt_buf1[1];
w2_t[2] = salt_buf1[2]; w2[2] = salt_buf1[2];
w2_t[3] = salt_buf1[3]; w2[3] = salt_buf1[3];
w3_t[0] = salt_buf2[0]; w3[0] = salt_buf2[0];
w3_t[1] = salt_buf2[1]; w3[1] = salt_buf2[1];
w3_t[2] = (16 + salt_len) * 8; w3[2] = (16 + salt_len) * 8;
w3_t[3] = 0; w3[3] = 0;
a = MD4M_A; a = MD4M_A;
b = MD4M_B; b = MD4M_B;
c = MD4M_C; c = MD4M_C;
d = MD4M_D; d = MD4M_D;
MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
if (MATCHES_NONE_VS (a, search[0])) continue; if (MATCHES_NONE_VS (a, search[0])) continue;
MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
COMPARE_S_SIMD (a, d, c, b); COMPARE_S_SIMD (a, d, c, b);
} }

840
OpenCL/m01100_a1.cl
View File

@ -5,6 +5,8 @@
#define _MD4_ #define _MD4_
#define NEW_SIMD_CODE
#include "include/constants.h" #include "include/constants.h"
#include "include/kernel_vendor.h" #include "include/kernel_vendor.h"
@ -16,9 +18,7 @@
#include "include/kernel_functions.c" #include "include/kernel_functions.c"
#include "OpenCL/types_ocl.c" #include "OpenCL/types_ocl.c"
#include "OpenCL/common.c" #include "OpenCL/common.c"
#include "OpenCL/simd.c"
#define COMPARE_S "OpenCL/check_single_comp4.c"
#define COMPARE_M "OpenCL/check_multi_comp4.c"
__kernel void m01100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_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 m01100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_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)
{ {
@ -36,67 +36,40 @@ __kernel void m01100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 wordl0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
wordl0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[0];
wordl0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[1];
wordl0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[2];
wordl0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
u32 wordl1[4]; pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
wordl1[0] = pws[gid].i[ 4]; pw_buf1[3] = pws[gid].i[7];
wordl1[1] = pws[gid].i[ 5];
wordl1[2] = pws[gid].i[ 6];
wordl1[3] = pws[gid].i[ 7];
u32 wordl2[4];
wordl2[0] = 0;
wordl2[1] = 0;
wordl2[2] = 0;
wordl2[3] = 0;
u32 wordl3[4];
wordl3[0] = 0;
wordl3[1] = 0;
wordl3[2] = 0;
wordl3[3] = 0;
const u32 pw_l_len = pws[gid].pw_len; const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
append_0x80_2x4 (wordl0, wordl1, pw_l_len);
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/** /**
* salt * salt
*/ */
u32 salt_buf0[4]; u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0]; salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1]; salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2]; salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3]; salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
u32 salt_buf1[4];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4]; salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5]; salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6]; salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7]; salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
u32 salt_buf2[4];
salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8]; salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9]; salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
salt_buf2[2] = 0; salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
salt_buf2[3] = 0; salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
const u32 salt_len = salt_bufs[salt_pos].salt_len; const u32 salt_len = salt_bufs[salt_pos].salt_len;
@ -104,223 +77,220 @@ __kernel void m01100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* loop * loop
*/ */
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{ {
const u32 pw_r_len = combs_buf[il_pos].pw_len; const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32 pw_len = pw_l_len + pw_r_len; const u32x pw_len = pw_l_len + pw_r_len;
u32 wordr0[4]; /**
* concat password candidate
*/
wordr0[0] = combs_buf[il_pos].i[0]; u32x wordl0[4] = { 0 };
wordr0[1] = combs_buf[il_pos].i[1]; u32x wordl1[4] = { 0 };
wordr0[2] = combs_buf[il_pos].i[2]; u32x wordl2[4] = { 0 };
wordr0[3] = combs_buf[il_pos].i[3]; u32x wordl3[4] = { 0 };
u32 wordr1[4]; wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
wordr1[0] = combs_buf[il_pos].i[4]; u32x wordr0[4] = { 0 };
wordr1[1] = combs_buf[il_pos].i[5]; u32x wordr1[4] = { 0 };
wordr1[2] = combs_buf[il_pos].i[6]; u32x wordr2[4] = { 0 };
wordr1[3] = combs_buf[il_pos].i[7]; u32x wordr3[4] = { 0 };
u32 wordr2[4]; wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr2[0] = 0; wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr2[1] = 0; wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr2[2] = 0; wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr2[3] = 0; wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
u32 wordr3[4]; wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
if (combs_mode == COMBINATOR_MODE_BASE_LEFT) if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{ {
switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len); switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
} }
u32 w0[4]; u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = wordl0[0] | wordr0[0]; w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1]; w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2]; w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3]; w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0]; w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1]; w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2]; w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3]; w1[3] = wordl1[3] | wordr1[3];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1];
w3[2] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
u32 w2[4]; make_unicode (w1, w2, w3);
make_unicode (w0, w0, w1);
w2[0] = 0; w3[2] = pw_len * 2 * 8;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
u32 w3[4];
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0; w3[3] = 0;
u32 w0_t[4]; /**
u32 w1_t[4]; * md4
u32 w2_t[4]; */
u32 w3_t[4];
make_unicode (w0, w0_t, w1_t); u32x a = MD4M_A;
make_unicode (w1, w2_t, w3_t); u32x b = MD4M_B;
u32x c = MD4M_C;
u32x d = MD4M_D;
w3_t[2] = pw_len * 8 * 2; 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);
u32 a = MD4M_A; MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
u32 b = MD4M_B; MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
u32 c = MD4M_C; MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
u32 d = MD4M_D; 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_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
a += MD4M_A; a += MD4M_A;
b += MD4M_B; b += MD4M_B;
c += MD4M_C; c += MD4M_C;
d += MD4M_D; d += MD4M_D;
w0_t[0] = a; w0[0] = a;
w0_t[1] = b; w0[1] = b;
w0_t[2] = c; w0[2] = c;
w0_t[3] = d; w0[3] = d;
w1_t[0] = salt_buf0[0]; w1[0] = salt_buf0[0];
w1_t[1] = salt_buf0[1]; w1[1] = salt_buf0[1];
w1_t[2] = salt_buf0[2]; w1[2] = salt_buf0[2];
w1_t[3] = salt_buf0[3]; w1[3] = salt_buf0[3];
w2_t[0] = salt_buf1[0]; w2[0] = salt_buf1[0];
w2_t[1] = salt_buf1[1]; w2[1] = salt_buf1[1];
w2_t[2] = salt_buf1[2]; w2[2] = salt_buf1[2];
w2_t[3] = salt_buf1[3]; w2[3] = salt_buf1[3];
w3_t[0] = salt_buf2[0]; w3[0] = salt_buf2[0];
w3_t[1] = salt_buf2[1]; w3[1] = salt_buf2[1];
w3_t[2] = (16 + salt_len) * 8; w3[2] = (16 + salt_len) * 8;
w3_t[3] = 0; w3[3] = 0;
a = MD4M_A; a = MD4M_A;
b = MD4M_B; b = MD4M_B;
c = MD4M_C; c = MD4M_C;
d = MD4M_D; d = MD4M_D;
MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
const u32 r0 = a; COMPARE_M_SIMD (a, d, c, b);
const u32 r1 = d;
const u32 r2 = c;
const u32 r3 = b;
#include COMPARE_M
} }
} }
@ -348,42 +318,42 @@ __kernel void m01100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 wordl0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
wordl0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[0];
wordl0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[1];
wordl0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[2];
wordl0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
u32 wordl1[4]; pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
wordl1[0] = pws[gid].i[ 4]; pw_buf1[3] = pws[gid].i[7];
wordl1[1] = pws[gid].i[ 5];
wordl1[2] = pws[gid].i[ 6];
wordl1[3] = pws[gid].i[ 7];
u32 wordl2[4];
wordl2[0] = 0;
wordl2[1] = 0;
wordl2[2] = 0;
wordl2[3] = 0;
u32 wordl3[4];
wordl3[0] = 0;
wordl3[1] = 0;
wordl3[2] = 0;
wordl3[3] = 0;
const u32 pw_l_len = pws[gid].pw_len; const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT) /**
{ * salt
append_0x80_2x4 (wordl0, wordl1, pw_l_len); */
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len); u32 salt_buf0[4];
} u32 salt_buf1[4];
u32 salt_buf2[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
const u32 salt_len = salt_bufs[salt_pos].salt_len;
/** /**
* digest * digest
@ -397,259 +367,227 @@ __kernel void m01100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digests_buf[digests_offset].digest_buf[DGST_R3] digests_buf[digests_offset].digest_buf[DGST_R3]
}; };
/**
* salt
*/
u32 salt_buf0[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
u32 salt_buf1[4];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
u32 salt_buf2[4];
salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
salt_buf2[2] = 0;
salt_buf2[3] = 0;
const u32 salt_len = salt_bufs[salt_pos].salt_len;
/** /**
* loop * loop
*/ */
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{ {
const u32 pw_r_len = combs_buf[il_pos].pw_len; const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32 pw_len = pw_l_len + pw_r_len; const u32x pw_len = pw_l_len + pw_r_len;
u32 wordr0[4]; /**
* concat password candidate
*/
wordr0[0] = combs_buf[il_pos].i[0]; u32x wordl0[4] = { 0 };
wordr0[1] = combs_buf[il_pos].i[1]; u32x wordl1[4] = { 0 };
wordr0[2] = combs_buf[il_pos].i[2]; u32x wordl2[4] = { 0 };
wordr0[3] = combs_buf[il_pos].i[3]; u32x wordl3[4] = { 0 };
u32 wordr1[4]; wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
wordr1[0] = combs_buf[il_pos].i[4]; u32x wordr0[4] = { 0 };
wordr1[1] = combs_buf[il_pos].i[5]; u32x wordr1[4] = { 0 };
wordr1[2] = combs_buf[il_pos].i[6]; u32x wordr2[4] = { 0 };
wordr1[3] = combs_buf[il_pos].i[7]; u32x wordr3[4] = { 0 };
u32 wordr2[4]; wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr2[0] = 0; wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr2[1] = 0; wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr2[2] = 0; wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr2[3] = 0; wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
u32 wordr3[4]; wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
if (combs_mode == COMBINATOR_MODE_BASE_LEFT) if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{ {
switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len); switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
} }
u32 w0[4]; u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = wordl0[0] | wordr0[0]; w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1]; w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2]; w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3]; w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0]; w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1]; w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2]; w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3]; w1[3] = wordl1[3] | wordr1[3];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1];
w3[2] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
u32 w2[4]; make_unicode (w1, w2, w3);
make_unicode (w0, w0, w1);
w2[0] = 0; w3[2] = pw_len * 2 * 8;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
u32 w3[4];
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0; w3[3] = 0;
u32 w0_t[4]; /**
u32 w1_t[4]; * md4
u32 w2_t[4]; */
u32 w3_t[4];
make_unicode (w0, w0_t, w1_t); u32x a = MD4M_A;
make_unicode (w1, w2_t, w3_t); u32x b = MD4M_B;
u32x c = MD4M_C;
u32x d = MD4M_D;
w3_t[2] = pw_len * 8 * 2; 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);
u32 a = MD4M_A; MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
u32 b = MD4M_B; MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
u32 c = MD4M_C; MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
u32 d = MD4M_D; 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_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
a += MD4M_A; a += MD4M_A;
b += MD4M_B; b += MD4M_B;
c += MD4M_C; c += MD4M_C;
d += MD4M_D; d += MD4M_D;
w0_t[0] = a; w0[0] = a;
w0_t[1] = b; w0[1] = b;
w0_t[2] = c; w0[2] = c;
w0_t[3] = d; w0[3] = d;
w1_t[0] = salt_buf0[0]; w1[0] = salt_buf0[0];
w1_t[1] = salt_buf0[1]; w1[1] = salt_buf0[1];
w1_t[2] = salt_buf0[2]; w1[2] = salt_buf0[2];
w1_t[3] = salt_buf0[3]; w1[3] = salt_buf0[3];
w2_t[0] = salt_buf1[0]; w2[0] = salt_buf1[0];
w2_t[1] = salt_buf1[1]; w2[1] = salt_buf1[1];
w2_t[2] = salt_buf1[2]; w2[2] = salt_buf1[2];
w2_t[3] = salt_buf1[3]; w2[3] = salt_buf1[3];
w3_t[0] = salt_buf2[0]; w3[0] = salt_buf2[0];
w3_t[1] = salt_buf2[1]; w3[1] = salt_buf2[1];
w3_t[2] = (16 + salt_len) * 8; w3[2] = (16 + salt_len) * 8;
w3_t[3] = 0; w3[3] = 0;
a = MD4M_A; a = MD4M_A;
b = MD4M_B; b = MD4M_B;
c = MD4M_C; c = MD4M_C;
d = MD4M_D; d = MD4M_D;
MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10); MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11); MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12); MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13); MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21); MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22); MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23); MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20); MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
bool q_cond = allx (search[0] != a); if (MATCHES_NONE_VS (a, search[0])) continue;
if (q_cond) continue; 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);
MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21); COMPARE_S_SIMD (a, d, c, b);
MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
const u32 r0 = a;
const u32 r1 = d;
const u32 r2 = c;
const u32 r3 = b;
#include COMPARE_S
} }
} }

24
OpenCL/m01100_a3.cl
View File

@ -34,25 +34,21 @@ static void m01100m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
*/ */
u32 salt_buf0[4]; u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0]; salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1]; salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2]; salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3]; salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
u32 salt_buf1[4];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4]; salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5]; salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6]; salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7]; salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
u32 salt_buf2[4];
salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8]; salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9]; salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
salt_buf2[2] = 0; salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
salt_buf2[3] = 0; salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
const u32 salt_len = salt_bufs[salt_pos].salt_len; const u32 salt_len = salt_bufs[salt_pos].salt_len;
@ -280,25 +276,21 @@ static void m01100s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global k
*/ */
u32 salt_buf0[4]; u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0]; salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1]; salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2]; salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3]; salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
u32 salt_buf1[4];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4]; salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5]; salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6]; salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7]; salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
u32 salt_buf2[4];
salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8]; salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9]; salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
salt_buf2[2] = 0; salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
salt_buf2[3] = 0; salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
const u32 salt_len = salt_bufs[salt_pos].salt_len; const u32 salt_len = salt_bufs[salt_pos].salt_len;