2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
24 #include "include/kernel_functions.c"
26 #include "common_nv.c"
27 #include "include/rp_gpu.h"
31 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
32 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
36 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
37 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
40 __device__ __constant__ u32 padding[8] =
52 __device__ static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
76 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
77 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
78 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
79 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
80 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
81 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
82 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
83 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
84 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
85 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
86 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
87 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
88 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
89 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
90 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
91 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
93 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
94 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
95 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
96 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
97 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
98 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
99 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
100 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
101 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
102 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
103 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
104 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
105 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
106 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
107 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
108 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
110 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
111 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
112 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
113 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
114 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
115 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
116 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
117 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
118 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
119 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
120 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
121 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
122 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
123 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
124 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
125 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
127 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
128 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
129 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
130 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
131 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
132 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
133 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
134 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
135 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
136 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
137 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
138 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
139 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
140 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
141 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
142 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
150 __device__ __constant__ gpu_rule_t c_rules[1024];
152 extern "C" __global__ void __launch_bounds__ (256, 1) m10420_m04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const pdf_t *pdf_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
158 const u32 lid = threadIdx.x;
164 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
166 if (gid >= gid_max) return;
170 pw_buf0[0] = pws[gid].i[ 0];
171 pw_buf0[1] = pws[gid].i[ 1];
172 pw_buf0[2] = pws[gid].i[ 2];
173 pw_buf0[3] = pws[gid].i[ 3];
177 pw_buf1[0] = pws[gid].i[ 4];
178 pw_buf1[1] = pws[gid].i[ 5];
179 pw_buf1[2] = pws[gid].i[ 6];
180 pw_buf1[3] = pws[gid].i[ 7];
182 const u32 pw_len = pws[gid].pw_len;
190 o_buf[0] = pdf_bufs[salt_pos].o_buf[0];
191 o_buf[1] = pdf_bufs[salt_pos].o_buf[1];
192 o_buf[2] = pdf_bufs[salt_pos].o_buf[2];
193 o_buf[3] = pdf_bufs[salt_pos].o_buf[3];
194 o_buf[4] = pdf_bufs[salt_pos].o_buf[4];
195 o_buf[5] = pdf_bufs[salt_pos].o_buf[5];
196 o_buf[6] = pdf_bufs[salt_pos].o_buf[6];
197 o_buf[7] = pdf_bufs[salt_pos].o_buf[7];
199 u32 P = pdf_bufs[salt_pos].P;
203 id_buf[0] = pdf_bufs[salt_pos].id_buf[0];
204 id_buf[1] = pdf_bufs[salt_pos].id_buf[1];
205 id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
206 id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
212 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
242 const u32 out_len = apply_rules (c_rules[il_pos].cmds, w0, w1, pw_len);
249 // max length supported by pdf11 is 32
251 w0_t[0] = padding[0];
252 w0_t[1] = padding[1];
253 w0_t[2] = padding[2];
254 w0_t[3] = padding[3];
255 w1_t[0] = padding[4];
256 w1_t[1] = padding[5];
257 w1_t[2] = padding[6];
258 w1_t[3] = padding[7];
268 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, pw_len);
271 // truncate at 32 is wanted, not a bug!
298 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
317 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
320 u32x b = digest[1] & 0xff;
327 #include VECT_COMPARE_M
331 extern "C" __global__ void __launch_bounds__ (256, 1) m10420_m08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const pdf_t *pdf_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
335 extern "C" __global__ void __launch_bounds__ (256, 1) m10420_m16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const pdf_t *pdf_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
339 extern "C" __global__ void __launch_bounds__ (256, 1) m10420_s04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const pdf_t *pdf_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
345 const u32 lid = threadIdx.x;
351 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
353 if (gid >= gid_max) return;
357 pw_buf0[0] = pws[gid].i[ 0];
358 pw_buf0[1] = pws[gid].i[ 1];
359 pw_buf0[2] = pws[gid].i[ 2];
360 pw_buf0[3] = pws[gid].i[ 3];
364 pw_buf1[0] = pws[gid].i[ 4];
365 pw_buf1[1] = pws[gid].i[ 5];
366 pw_buf1[2] = pws[gid].i[ 6];
367 pw_buf1[3] = pws[gid].i[ 7];
369 const u32 pw_len = pws[gid].pw_len;
375 const u32 search[4] =
377 digests_buf[digests_offset].digest_buf[DGST_R0],
378 digests_buf[digests_offset].digest_buf[DGST_R1],
379 digests_buf[digests_offset].digest_buf[DGST_R2],
380 digests_buf[digests_offset].digest_buf[DGST_R3]
389 o_buf[0] = pdf_bufs[salt_pos].o_buf[0];
390 o_buf[1] = pdf_bufs[salt_pos].o_buf[1];
391 o_buf[2] = pdf_bufs[salt_pos].o_buf[2];
392 o_buf[3] = pdf_bufs[salt_pos].o_buf[3];
393 o_buf[4] = pdf_bufs[salt_pos].o_buf[4];
394 o_buf[5] = pdf_bufs[salt_pos].o_buf[5];
395 o_buf[6] = pdf_bufs[salt_pos].o_buf[6];
396 o_buf[7] = pdf_bufs[salt_pos].o_buf[7];
398 u32 P = pdf_bufs[salt_pos].P;
402 id_buf[0] = pdf_bufs[salt_pos].id_buf[0];
403 id_buf[1] = pdf_bufs[salt_pos].id_buf[1];
404 id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
405 id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
411 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
441 const u32 out_len = apply_rules (c_rules[il_pos].cmds, w0, w1, pw_len);
448 // max length supported by pdf11 is 32
450 w0_t[0] = padding[0];
451 w0_t[1] = padding[1];
452 w0_t[2] = padding[2];
453 w0_t[3] = padding[3];
454 w1_t[0] = padding[4];
455 w1_t[1] = padding[5];
456 w1_t[2] = padding[6];
457 w1_t[3] = padding[7];
467 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, pw_len);
470 // truncate at 32 is wanted, not a bug!
497 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
516 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
519 u32x b = digest[1] & 0xff;
526 #include VECT_COMPARE_S
530 extern "C" __global__ void __launch_bounds__ (256, 1) m10420_s08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const pdf_t *pdf_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
534 extern "C" __global__ void __launch_bounds__ (256, 1) m10420_s16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const pdf_t *pdf_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)