X-Git-Url: https://www.flypig.org.uk/git/?a=blobdiff_plain;f=OpenCL%2Fm06900_a3.cl;h=21f8e46a461c245ec77ee10f0daee8f97b7de07a;hb=51e1d11b2d6dff4ae331cff6098169455733c412;hp=fdb442be75076eca350c4a3fd1f73fd08856dba7;hpb=e4465d7dcb88426532f6b0488cf9bc1cd3b06cc6;p=hashcat.git diff --git a/OpenCL/m06900_a3.cl b/OpenCL/m06900_a3.cl index fdb442b..21f8e46 100644 --- a/OpenCL/m06900_a3.cl +++ b/OpenCL/m06900_a3.cl @@ -1,24 +1,21 @@ /** - * Author......: Jens Steube + * Authors.....: Jens Steube + * Gabriele Gristina + * magnum + * * License.....: MIT */ #define _GOST_ -#include "include/constants.h" -#include "include/kernel_vendor.h" +#define NEW_SIMD_CODE -#define DGST_R0 0 -#define DGST_R1 1 -#define DGST_R2 2 -#define DGST_R3 3 - -#include "include/kernel_functions.c" -#include "types_ocl.c" -#include "common.c" - -#define COMPARE_S "check_single_comp4.c" -#define COMPARE_M "check_multi_comp4.c" +#include "inc_vendor.cl" +#include "inc_hash_constants.h" +#include "inc_hash_functions.cl" +#include "inc_types.cl" +#include "inc_common.cl" +#include "inc_simd.cl" __constant u32 c_tables[4][256] = { @@ -288,11 +285,21 @@ __constant u32 c_tables[4][256] = } }; +#if VECT_SIZE == 1 #define BOX(i,n,S) (S)[(n)][(i)] - -#define round(k1,k2,tbl) \ +#elif VECT_SIZE == 2 +#define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1]) +#elif VECT_SIZE == 4 +#define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3]) +#elif VECT_SIZE == 8 +#define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7]) +#elif VECT_SIZE == 16 +#define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf]) +#endif + +#define _round(k1,k2,tbl) \ { \ - u32 t; \ + u32x t; \ t = (k1) + r; \ l ^= BOX (((t >> 0) & 0xff), 0, tbl) ^ \ BOX (((t >> 8) & 0xff), 1, tbl) ^ \ @@ -307,26 +314,26 @@ __constant u32 c_tables[4][256] = #define R(k,h,s,i,t) \ { \ - u32 r; \ - u32 l; \ + u32x r; \ + u32x l; \ r = h[i + 0]; \ l = h[i + 1]; \ - round (k[0], k[1], t); \ - round (k[2], k[3], t); \ - round (k[4], k[5], t); \ - round (k[6], k[7], t); \ - round (k[0], k[1], t); \ - round (k[2], k[3], t); \ - round (k[4], k[5], t); \ - round (k[6], k[7], t); \ - round (k[0], k[1], t); \ - round (k[2], k[3], t); \ - round (k[4], k[5], t); \ - round (k[6], k[7], t); \ - round (k[7], k[6], t); \ - round (k[5], k[4], t); \ - round (k[3], k[2], t); \ - round (k[1], k[0], t); \ + _round (k[0], k[1], t); \ + _round (k[2], k[3], t); \ + _round (k[4], k[5], t); \ + _round (k[6], k[7], t); \ + _round (k[0], k[1], t); \ + _round (k[2], k[3], t); \ + _round (k[4], k[5], t); \ + _round (k[6], k[7], t); \ + _round (k[0], k[1], t); \ + _round (k[2], k[3], t); \ + _round (k[4], k[5], t); \ + _round (k[6], k[7], t); \ + _round (k[7], k[6], t); \ + _round (k[5], k[4], t); \ + _round (k[3], k[2], t); \ + _round (k[1], k[0], t); \ s[i + 0] = l; \ s[i + 1] = r; \ } @@ -377,8 +384,8 @@ __constant u32 c_tables[4][256] = #define A(x) \ { \ - u32 l; \ - u32 r; \ + u32x l; \ + u32x r; \ l = x[0] ^ x[2]; \ r = x[1] ^ x[3]; \ x[0] = x[2]; \ @@ -393,8 +400,8 @@ __constant u32 c_tables[4][256] = #define AA(x) \ { \ - u32 l; \ - u32 r; \ + u32x l; \ + u32x r; \ l = x[0]; \ r = x[2]; \ x[0] = x[4]; \ @@ -652,8 +659,8 @@ __constant u32 c_tables[4][256] = #define PASS0(h,s,u,v,t) \ { \ - u32 k[8]; \ - u32 w[8]; \ + u32x k[8]; \ + u32x w[8]; \ X (w, u, v); \ P (k, w); \ R (k, h, s, 0, t); \ @@ -663,8 +670,8 @@ __constant u32 c_tables[4][256] = #define PASS2(h,s,u,v,t) \ { \ - u32 k[8]; \ - u32 w[8]; \ + u32x k[8]; \ + u32x w[8]; \ X (w, u, v); \ P (k, w); \ R (k, h, s, 2, t); \ @@ -675,8 +682,8 @@ __constant u32 c_tables[4][256] = #define PASS4(h,s,u,v,t) \ { \ - u32 k[8]; \ - u32 w[8]; \ + u32x k[8]; \ + u32x w[8]; \ X (w, u, v); \ P (k, w); \ R (k, h, s, 4, t); \ @@ -686,14 +693,14 @@ __constant u32 c_tables[4][256] = #define PASS6(h,s,u,v,t) \ { \ - u32 k[8]; \ - u32 w[8]; \ + u32x k[8]; \ + u32x w[8]; \ X (w, u, v); \ P (k, w); \ R (k, h, s, 6, t); \ } -static void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 s_tables[4][256]) +void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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_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_tables)[256]) { /** * modifier @@ -702,27 +709,25 @@ static void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le const u32 gid = get_global_id (0); const u32 lid = get_local_id (0); - /** - * base - */ - - const u32 w14 = pw_len * 8; - /** * loop */ u32 w0l = w0[0]; - for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++) + for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE) { - const u32 w0r = bfs_buf[il_pos].i; + const u32x w0r = ix_create_bft (bfs_buf, il_pos); + + const u32x w0lr = w0l | w0r; - w0[0] = w0l | w0r; + /** + * GOST + */ - u32 data[8]; + u32x data[8]; - data[0] = w0[0]; + data[0] = w0lr; data[1] = w0[1]; data[2] = w0[2]; data[3] = w0[3]; @@ -731,7 +736,7 @@ static void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le data[6] = w1[2]; data[7] = w1[3]; - u32 state[16]; + u32x state[16]; state[ 0] = 0; state[ 1] = 0; @@ -750,8 +755,8 @@ static void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le state[14] = data[6]; state[15] = data[7]; - u32 state_m[8]; - u32 data_m[8]; + u32x state_m[8]; + u32x data_m[8]; /* gost1 */ @@ -773,9 +778,9 @@ static void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le data_m[6] = data[6]; data_m[7] = data[7]; - u32 tmp[8]; + u32x tmp[8]; - if (pw_len > 0) + //if (pw_len > 0) // not really SIMD compatible { PASS0 (state, tmp, state_m, data_m, s_tables); PASS2 (state, tmp, state_m, data_m, s_tables); @@ -787,7 +792,7 @@ static void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le SHIFT61 (state, data_m); } - data[0] = w14; + data[0] = pw_len * 8; data[1] = 0; data[2] = 0; data[3] = 0; @@ -865,15 +870,11 @@ static void m06900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le /* store */ - const u32 r0 = state[0]; - const u32 r1 = state[1]; - const u32 r2 = state[2]; - const u32 r3 = state[3]; - - #include COMPARE_M + COMPARE_M_SIMD (state[0], state[1], state[2], state[3]); } } -static void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 s_tables[4][256]) + +void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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_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_tables)[256]) { /** * modifier @@ -882,12 +883,6 @@ static void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le const u32 gid = get_global_id (0); const u32 lid = get_local_id (0); - /** - * base - */ - - const u32 w14 = pw_len * 8; - /** * digest */ @@ -906,15 +901,19 @@ static void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le u32 w0l = w0[0]; - for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++) + for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE) { - const u32 w0r = bfs_buf[il_pos].i; + const u32x w0r = ix_create_bft (bfs_buf, il_pos); - w0[0] = w0l | w0r; + const u32x w0lr = w0l | w0r; - u32 data[8]; + /** + * GOST + */ - data[0] = w0[0]; + u32x data[8]; + + data[0] = w0lr; data[1] = w0[1]; data[2] = w0[2]; data[3] = w0[3]; @@ -923,7 +922,7 @@ static void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le data[6] = w1[2]; data[7] = w1[3]; - u32 state[16]; + u32x state[16]; state[ 0] = 0; state[ 1] = 0; @@ -942,8 +941,8 @@ static void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le state[14] = data[6]; state[15] = data[7]; - u32 state_m[8]; - u32 data_m[8]; + u32x state_m[8]; + u32x data_m[8]; /* gost1 */ @@ -965,9 +964,9 @@ static void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le data_m[6] = data[6]; data_m[7] = data[7]; - u32 tmp[8]; + u32x tmp[8]; - if (pw_len > 0) + //if (pw_len > 0) // not really SIMD compatible { PASS0 (state, tmp, state_m, data_m, s_tables); PASS2 (state, tmp, state_m, data_m, s_tables); @@ -979,7 +978,7 @@ static void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le SHIFT61 (state, data_m); } - data[0] = w14; + data[0] = pw_len * 8; data[1] = 0; data[2] = 0; data[3] = 0; @@ -1057,28 +1056,41 @@ static void m06900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le /* store */ - const u32 r0 = state[0]; - const u32 r1 = state[1]; - const u32 r2 = state[2]; - const u32 r3 = state[3]; - - #include COMPARE_S + COMPARE_M_SIMD (state[0], state[1], state[2], state[3]); } } -__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_m04 (__global pw_t *pws, __global gpu_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max) +__kernel void m06900_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_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) { /** * base */ const u32 gid = get_global_id (0); + const u32 lid = get_local_id (0); + const u32 lsz = get_local_size (0); /** - * modifier + * sbox */ - const u32 lid = get_local_id (0); + __local u32 s_tables[4][256]; + + for (u32 i = lid; i < 256; i += lsz) + { + s_tables[0][i] = c_tables[0][i]; + s_tables[1][i] = c_tables[1][i]; + s_tables[2][i] = c_tables[2][i]; + s_tables[3][i] = c_tables[3][i]; + } + + barrier (CLK_LOCAL_MEM_FENCE); + + if (gid >= gid_max) return; + + /** + * modifier + */ u32 w0[4]; @@ -1110,58 +1122,44 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_m04 (__glo const u32 pw_len = pws[gid].pw_len; - /** - * sbox - */ - - __local u32 s_tables[4][256]; - - const u32 lid4 = lid * 4; - - s_tables[0][lid4 + 0] = c_tables[0][lid4 + 0]; - s_tables[0][lid4 + 1] = c_tables[0][lid4 + 1]; - s_tables[0][lid4 + 2] = c_tables[0][lid4 + 2]; - s_tables[0][lid4 + 3] = c_tables[0][lid4 + 3]; - - s_tables[1][lid4 + 0] = c_tables[1][lid4 + 0]; - s_tables[1][lid4 + 1] = c_tables[1][lid4 + 1]; - s_tables[1][lid4 + 2] = c_tables[1][lid4 + 2]; - s_tables[1][lid4 + 3] = c_tables[1][lid4 + 3]; - - s_tables[2][lid4 + 0] = c_tables[2][lid4 + 0]; - s_tables[2][lid4 + 1] = c_tables[2][lid4 + 1]; - s_tables[2][lid4 + 2] = c_tables[2][lid4 + 2]; - s_tables[2][lid4 + 3] = c_tables[2][lid4 + 3]; - - s_tables[3][lid4 + 0] = c_tables[3][lid4 + 0]; - s_tables[3][lid4 + 1] = c_tables[3][lid4 + 1]; - s_tables[3][lid4 + 2] = c_tables[3][lid4 + 2]; - s_tables[3][lid4 + 3] = c_tables[3][lid4 + 3]; - - barrier (CLK_LOCAL_MEM_FENCE); - - if (gid >= gid_max) return; - /** * main */ - m06900m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_tables); + m06900m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, s_tables); } -__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_m08 (__global pw_t *pws, __global gpu_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max) +__kernel void m06900_m08 (__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_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) { /** * base */ const u32 gid = get_global_id (0); + const u32 lid = get_local_id (0); + const u32 lsz = get_local_size (0); /** - * modifier + * sbox */ - const u32 lid = get_local_id (0); + __local u32 s_tables[4][256]; + + for (u32 i = lid; i < 256; i += lsz) + { + s_tables[0][i] = c_tables[0][i]; + s_tables[1][i] = c_tables[1][i]; + s_tables[2][i] = c_tables[2][i]; + s_tables[3][i] = c_tables[3][i]; + } + + barrier (CLK_LOCAL_MEM_FENCE); + + if (gid >= gid_max) return; + + /** + * modifier + */ u32 w0[4]; @@ -1193,62 +1191,48 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_m08 (__glo const u32 pw_len = pws[gid].pw_len; - /** - * sbox - */ - - __local u32 s_tables[4][256]; - - const u32 lid4 = lid * 4; - - s_tables[0][lid4 + 0] = c_tables[0][lid4 + 0]; - s_tables[0][lid4 + 1] = c_tables[0][lid4 + 1]; - s_tables[0][lid4 + 2] = c_tables[0][lid4 + 2]; - s_tables[0][lid4 + 3] = c_tables[0][lid4 + 3]; - - s_tables[1][lid4 + 0] = c_tables[1][lid4 + 0]; - s_tables[1][lid4 + 1] = c_tables[1][lid4 + 1]; - s_tables[1][lid4 + 2] = c_tables[1][lid4 + 2]; - s_tables[1][lid4 + 3] = c_tables[1][lid4 + 3]; - - s_tables[2][lid4 + 0] = c_tables[2][lid4 + 0]; - s_tables[2][lid4 + 1] = c_tables[2][lid4 + 1]; - s_tables[2][lid4 + 2] = c_tables[2][lid4 + 2]; - s_tables[2][lid4 + 3] = c_tables[2][lid4 + 3]; - - s_tables[3][lid4 + 0] = c_tables[3][lid4 + 0]; - s_tables[3][lid4 + 1] = c_tables[3][lid4 + 1]; - s_tables[3][lid4 + 2] = c_tables[3][lid4 + 2]; - s_tables[3][lid4 + 3] = c_tables[3][lid4 + 3]; - - barrier (CLK_LOCAL_MEM_FENCE); - - if (gid >= gid_max) return; - /** * main */ - m06900m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_tables); + m06900m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, s_tables); } -__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_m16 (__global pw_t *pws, __global gpu_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max) +__kernel void m06900_m16 (__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_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 __attribute__((reqd_work_group_size (64, 1, 1))) m06900_s04 (__global pw_t *pws, __global gpu_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max) +__kernel void m06900_s04 (__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_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) { /** * base */ const u32 gid = get_global_id (0); + const u32 lid = get_local_id (0); + const u32 lsz = get_local_size (0); /** - * modifier + * sbox */ - const u32 lid = get_local_id (0); + __local u32 s_tables[4][256]; + + for (u32 i = lid; i < 256; i += lsz) + { + s_tables[0][i] = c_tables[0][i]; + s_tables[1][i] = c_tables[1][i]; + s_tables[2][i] = c_tables[2][i]; + s_tables[3][i] = c_tables[3][i]; + } + + barrier (CLK_LOCAL_MEM_FENCE); + + if (gid >= gid_max) return; + + /** + * modifier + */ u32 w0[4]; @@ -1280,58 +1264,44 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_s04 (__glo const u32 pw_len = pws[gid].pw_len; - /** - * sbox - */ - - __local u32 s_tables[4][256]; - - const u32 lid4 = lid * 4; - - s_tables[0][lid4 + 0] = c_tables[0][lid4 + 0]; - s_tables[0][lid4 + 1] = c_tables[0][lid4 + 1]; - s_tables[0][lid4 + 2] = c_tables[0][lid4 + 2]; - s_tables[0][lid4 + 3] = c_tables[0][lid4 + 3]; - - s_tables[1][lid4 + 0] = c_tables[1][lid4 + 0]; - s_tables[1][lid4 + 1] = c_tables[1][lid4 + 1]; - s_tables[1][lid4 + 2] = c_tables[1][lid4 + 2]; - s_tables[1][lid4 + 3] = c_tables[1][lid4 + 3]; - - s_tables[2][lid4 + 0] = c_tables[2][lid4 + 0]; - s_tables[2][lid4 + 1] = c_tables[2][lid4 + 1]; - s_tables[2][lid4 + 2] = c_tables[2][lid4 + 2]; - s_tables[2][lid4 + 3] = c_tables[2][lid4 + 3]; - - s_tables[3][lid4 + 0] = c_tables[3][lid4 + 0]; - s_tables[3][lid4 + 1] = c_tables[3][lid4 + 1]; - s_tables[3][lid4 + 2] = c_tables[3][lid4 + 2]; - s_tables[3][lid4 + 3] = c_tables[3][lid4 + 3]; - - barrier (CLK_LOCAL_MEM_FENCE); - - if (gid >= gid_max) return; - /** * main */ - m06900s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_tables); + m06900s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, s_tables); } -__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_s08 (__global pw_t *pws, __global gpu_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max) +__kernel void m06900_s08 (__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_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) { /** * base */ const u32 gid = get_global_id (0); + const u32 lid = get_local_id (0); + const u32 lsz = get_local_size (0); /** - * modifier + * sbox */ - const u32 lid = get_local_id (0); + __local u32 s_tables[4][256]; + + for (u32 i = lid; i < 256; i += lsz) + { + s_tables[0][i] = c_tables[0][i]; + s_tables[1][i] = c_tables[1][i]; + s_tables[2][i] = c_tables[2][i]; + s_tables[3][i] = c_tables[3][i]; + } + + barrier (CLK_LOCAL_MEM_FENCE); + + if (gid >= gid_max) return; + + /** + * modifier + */ u32 w0[4]; @@ -1363,45 +1333,13 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_s08 (__glo const u32 pw_len = pws[gid].pw_len; - /** - * sbox - */ - - __local u32 s_tables[4][256]; - - const u32 lid4 = lid * 4; - - s_tables[0][lid4 + 0] = c_tables[0][lid4 + 0]; - s_tables[0][lid4 + 1] = c_tables[0][lid4 + 1]; - s_tables[0][lid4 + 2] = c_tables[0][lid4 + 2]; - s_tables[0][lid4 + 3] = c_tables[0][lid4 + 3]; - - s_tables[1][lid4 + 0] = c_tables[1][lid4 + 0]; - s_tables[1][lid4 + 1] = c_tables[1][lid4 + 1]; - s_tables[1][lid4 + 2] = c_tables[1][lid4 + 2]; - s_tables[1][lid4 + 3] = c_tables[1][lid4 + 3]; - - s_tables[2][lid4 + 0] = c_tables[2][lid4 + 0]; - s_tables[2][lid4 + 1] = c_tables[2][lid4 + 1]; - s_tables[2][lid4 + 2] = c_tables[2][lid4 + 2]; - s_tables[2][lid4 + 3] = c_tables[2][lid4 + 3]; - - s_tables[3][lid4 + 0] = c_tables[3][lid4 + 0]; - s_tables[3][lid4 + 1] = c_tables[3][lid4 + 1]; - s_tables[3][lid4 + 2] = c_tables[3][lid4 + 2]; - s_tables[3][lid4 + 3] = c_tables[3][lid4 + 3]; - - barrier (CLK_LOCAL_MEM_FENCE); - - if (gid >= gid_max) return; - /** * main */ - m06900s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_tables); + m06900s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV0_buf, d_scryptV1_buf, d_scryptV2_buf, d_scryptV3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, s_tables); } -__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06900_s16 (__global pw_t *pws, __global gpu_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max) +__kernel void m06900_s16 (__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_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) { }