From aa55adf6522e958c3a54bf1788f0ab45b49c00b6 Mon Sep 17 00:00:00 2001 From: Jens Steube Date: Sat, 23 Jan 2016 22:27:31 +0100 Subject: [PATCH] Revert some RC4 based kernels back to scalar processing --- OpenCL/m09700_a3.cl | 1082 +++++++++++++++++++++---------------------- OpenCL/m09710_a3.cl | 88 ++-- OpenCL/m09800_a3.cl | 94 ++-- OpenCL/m09810_a3.cl | 94 ++-- OpenCL/m10400_a3.cl | 22 +- OpenCL/m10410_a3.cl | 38 +- 6 files changed, 698 insertions(+), 720 deletions(-) diff --git a/OpenCL/m09700_a3.cl b/OpenCL/m09700_a3.cl index c84cfe02e..72f04f03c 100644 --- a/OpenCL/m09700_a3.cl +++ b/OpenCL/m09700_a3.cl @@ -5,8 +5,6 @@ #define _OLDOFFICE01_ -#define NEW_SIMD_CODE - #include "include/constants.h" #include "include/kernel_vendor.h" @@ -37,10 +35,10 @@ static void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j) rc4_key->S[j] = tmp; } -static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) +static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4]) { - u32x v = 0x03020100; - u32x a = 0x04040404; + u32 v = 0x03020100; + u32 a = 0x04040404; __local u32 *ptr = (__local u32 *) rc4_key->S; @@ -88,7 +86,7 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) } } -static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32x in[4], u32x out[4]) +static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4]) { #pragma unroll for (u32 k = 0; k < 4; k++) @@ -139,29 +137,29 @@ static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32x in[4], u return j; } -static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4]) +static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4]) { - u32x a = digest[0]; - u32x b = digest[1]; - u32x c = digest[2]; - u32x d = digest[3]; + u32 a = digest[0]; + u32 b = digest[1]; + u32 c = digest[2]; + u32 d = digest[3]; - u32x w0_t = w0[0]; - u32x w1_t = w0[1]; - u32x w2_t = w0[2]; - u32x w3_t = w0[3]; - u32x w4_t = w1[0]; - u32x w5_t = w1[1]; - u32x w6_t = w1[2]; - u32x w7_t = w1[3]; - u32x w8_t = w2[0]; - u32x w9_t = w2[1]; - u32x wa_t = w2[2]; - u32x wb_t = w2[3]; - u32x wc_t = w3[0]; - u32x wd_t = w3[1]; - u32x we_t = w3[2]; - u32x wf_t = w3[3]; + u32 w0_t = w0[0]; + u32 w1_t = w0[1]; + u32 w2_t = w0[2]; + u32 w3_t = w0[3]; + u32 w4_t = w1[0]; + u32 w5_t = w1[1]; + u32 w6_t = w1[2]; + u32 w7_t = w1[3]; + u32 w8_t = w2[0]; + u32 w9_t = w2[1]; + u32 wa_t = w2[2]; + u32 wb_t = w2[3]; + u32 wc_t = w3[0]; + u32 wd_t = w3[1]; + u32 we_t = w3[2]; + u32 wf_t = w3[3]; MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00); MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01); @@ -303,14 +301,14 @@ static void m09700m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + u32 w0_t[4]; + u32 w1_t[4]; + u32 w2_t[4]; + u32 w3_t[4]; w0_t[0] = w0lr; w0_t[1] = w0[1]; @@ -329,513 +327,10 @@ static void m09700m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ w3_t[2] = pw_len * 8; w3_t[3] = 0; - u32x digest_t0[4]; - u32x digest_t1[2]; // need only first 5 byte - u32x digest_t2[2]; - u32x digest_t3[2]; - - digest_t0[0] = MD5M_A; - digest_t0[1] = MD5M_B; - digest_t0[2] = MD5M_C; - digest_t0[3] = MD5M_D; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest_t0); - - // prepare 16 * 21 buffer stuff - - u32x digest[4]; - - digest[0] = MD5M_A; - digest[1] = MD5M_B; - digest[2] = MD5M_C; - digest[3] = MD5M_D; - - // offsets - - digest_t0[0] &= 0xffffffff; - digest_t0[1] &= 0x000000ff; - digest_t0[2] &= 0x00000000; - digest_t0[3] &= 0x00000000; - - digest_t1[0] = digest_t0[0] << 8; - digest_t1[1] = digest_t0[0] >> 24 | digest_t0[1] << 8; - - digest_t2[0] = digest_t0[0] << 16; - digest_t2[1] = digest_t0[0] >> 16 | digest_t0[1] << 16; - - digest_t3[0] = digest_t0[0] << 24; - digest_t3[1] = digest_t0[0] >> 8 | digest_t0[1] << 24; - - // generate the 16 * 21 buffer - - w0_t[0] = 0; - w0_t[1] = 0; - w0_t[2] = 0; - w0_t[3] = 0; - w1_t[0] = 0; - w1_t[1] = 0; - w1_t[2] = 0; - w1_t[3] = 0; - w2_t[0] = 0; - w2_t[1] = 0; - w2_t[2] = 0; - w2_t[3] = 0; - w3_t[0] = 0; - w3_t[1] = 0; - w3_t[2] = 0; - w3_t[3] = 0; - - // 0..5 - w0_t[0] = digest_t0[0]; - w0_t[1] = digest_t0[1]; - - // 5..21 - w0_t[1] |= salt_buf_t1[0]; - w0_t[2] = salt_buf_t1[1]; - w0_t[3] = salt_buf_t1[2]; - w1_t[0] = salt_buf_t1[3]; - w1_t[1] = salt_buf_t1[4]; - - // 21..26 - w1_t[1] |= digest_t1[0]; - w1_t[2] = digest_t1[1]; - - // 26..42 - w1_t[2] |= salt_buf_t2[0]; - w1_t[3] = salt_buf_t2[1]; - w2_t[0] = salt_buf_t2[2]; - w2_t[1] = salt_buf_t2[3]; - w2_t[2] = salt_buf_t2[4]; - - // 42..47 - w2_t[2] |= digest_t2[0]; - w2_t[3] = digest_t2[1]; - - // 47..63 - w2_t[3] |= salt_buf_t3[0]; - w3_t[0] = salt_buf_t3[1]; - w3_t[1] = salt_buf_t3[2]; - w3_t[2] = salt_buf_t3[3]; - w3_t[3] = salt_buf_t3[4]; - - // 63.. - - w3_t[3] |= digest_t3[0]; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); - - w0_t[0] = 0; - w0_t[1] = 0; - w0_t[2] = 0; - w0_t[3] = 0; - w1_t[0] = 0; - w1_t[1] = 0; - w1_t[2] = 0; - w1_t[3] = 0; - w2_t[0] = 0; - w2_t[1] = 0; - w2_t[2] = 0; - w2_t[3] = 0; - w3_t[0] = 0; - w3_t[1] = 0; - w3_t[2] = 0; - w3_t[3] = 0; - - // 0..4 - w0_t[0] = digest_t3[1]; - - // 4..20 - w0_t[1] = salt_buf_t0[0]; - w0_t[2] = salt_buf_t0[1]; - w0_t[3] = salt_buf_t0[2]; - w1_t[0] = salt_buf_t0[3]; - - // 20..25 - w1_t[1] = digest_t0[0]; - w1_t[2] = digest_t0[1]; - - // 25..41 - w1_t[2] |= salt_buf_t1[0]; - w1_t[3] = salt_buf_t1[1]; - w2_t[0] = salt_buf_t1[2]; - w2_t[1] = salt_buf_t1[3]; - w2_t[2] = salt_buf_t1[4]; - - // 41..46 - w2_t[2] |= digest_t1[0]; - w2_t[3] = digest_t1[1]; - - // 46..62 - w2_t[3] |= salt_buf_t2[0]; - w3_t[0] = salt_buf_t2[1]; - w3_t[1] = salt_buf_t2[2]; - w3_t[2] = salt_buf_t2[3]; - w3_t[3] = salt_buf_t2[4]; - - // 62.. - w3_t[3] |= digest_t2[0]; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); - - w0_t[0] = 0; - w0_t[1] = 0; - w0_t[2] = 0; - w0_t[3] = 0; - w1_t[0] = 0; - w1_t[1] = 0; - w1_t[2] = 0; - w1_t[3] = 0; - w2_t[0] = 0; - w2_t[1] = 0; - w2_t[2] = 0; - w2_t[3] = 0; - w3_t[0] = 0; - w3_t[1] = 0; - w3_t[2] = 0; - w3_t[3] = 0; - - // 0..3 - w0_t[0] = digest_t2[1]; - - // 3..19 - w0_t[0] |= salt_buf_t3[0]; - w0_t[1] = salt_buf_t3[1]; - w0_t[2] = salt_buf_t3[2]; - w0_t[3] = salt_buf_t3[3]; - w1_t[0] = salt_buf_t3[4]; - - // 19..24 - w1_t[0] |= digest_t3[0]; - w1_t[1] = digest_t3[1]; - - // 24..40 - w1_t[2] = salt_buf_t0[0]; - w1_t[3] = salt_buf_t0[1]; - w2_t[0] = salt_buf_t0[2]; - w2_t[1] = salt_buf_t0[3]; - - // 40..45 - w2_t[2] = digest_t0[0]; - w2_t[3] = digest_t0[1]; - - // 45..61 - w2_t[3] |= salt_buf_t1[0]; - w3_t[0] = salt_buf_t1[1]; - w3_t[1] = salt_buf_t1[2]; - w3_t[2] = salt_buf_t1[3]; - w3_t[3] = salt_buf_t1[4]; - - // 61.. - w3_t[3] |= digest_t1[0]; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); - - w0_t[0] = 0; - w0_t[1] = 0; - w0_t[2] = 0; - w0_t[3] = 0; - w1_t[0] = 0; - w1_t[1] = 0; - w1_t[2] = 0; - w1_t[3] = 0; - w2_t[0] = 0; - w2_t[1] = 0; - w2_t[2] = 0; - w2_t[3] = 0; - w3_t[0] = 0; - w3_t[1] = 0; - w3_t[2] = 0; - w3_t[3] = 0; - - // 0..2 - w0_t[0] = digest_t1[1]; - - // 2..18 - w0_t[0] |= salt_buf_t2[0]; - w0_t[1] = salt_buf_t2[1]; - w0_t[2] = salt_buf_t2[2]; - w0_t[3] = salt_buf_t2[3]; - w1_t[0] = salt_buf_t2[4]; - - // 18..23 - w1_t[0] |= digest_t2[0]; - w1_t[1] = digest_t2[1]; - - // 23..39 - w1_t[1] |= salt_buf_t3[0]; - w1_t[2] = salt_buf_t3[1]; - w1_t[3] = salt_buf_t3[2]; - w2_t[0] = salt_buf_t3[3]; - w2_t[1] = salt_buf_t3[4]; - - // 39..44 - w2_t[1] |= digest_t3[0]; - w2_t[2] = digest_t3[1]; - - // 44..60 - w2_t[3] = salt_buf_t0[0]; - w3_t[0] = salt_buf_t0[1]; - w3_t[1] = salt_buf_t0[2]; - w3_t[2] = salt_buf_t0[3]; - - // 60.. - w3_t[3] = digest_t0[0]; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); - - w0_t[0] = 0; - w0_t[1] = 0; - w0_t[2] = 0; - w0_t[3] = 0; - w1_t[0] = 0; - w1_t[1] = 0; - w1_t[2] = 0; - w1_t[3] = 0; - w2_t[0] = 0; - w2_t[1] = 0; - w2_t[2] = 0; - w2_t[3] = 0; - w3_t[0] = 0; - w3_t[1] = 0; - w3_t[2] = 0; - w3_t[3] = 0; - - // 0..1 - w0_t[0] = digest_t0[1]; - - // 1..17 - w0_t[0] |= salt_buf_t1[0]; - w0_t[1] = salt_buf_t1[1]; - w0_t[2] = salt_buf_t1[2]; - w0_t[3] = salt_buf_t1[3]; - w1_t[0] = salt_buf_t1[4]; - - // 17..22 - w1_t[0] |= digest_t1[0]; - w1_t[1] = digest_t1[1]; - - // 22..38 - w1_t[1] |= salt_buf_t2[0]; - w1_t[2] = salt_buf_t2[1]; - w1_t[3] = salt_buf_t2[2]; - w2_t[0] = salt_buf_t2[3]; - w2_t[1] = salt_buf_t2[4]; - - // 38..43 - w2_t[1] |= digest_t2[0]; - w2_t[2] = digest_t2[1]; - - // 43..59 - w2_t[2] |= salt_buf_t3[0]; - w2_t[3] = salt_buf_t3[1]; - w3_t[0] = salt_buf_t3[2]; - w3_t[1] = salt_buf_t3[3]; - w3_t[2] = salt_buf_t3[4]; - - // 59.. - w3_t[2] |= digest_t3[0]; - w3_t[3] = digest_t3[1]; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); - - w0_t[0] = salt_buf_t0[0]; - w0_t[1] = salt_buf_t0[1]; - w0_t[2] = salt_buf_t0[2]; - w0_t[3] = salt_buf_t0[3]; - w1_t[0] = 0x80; - w1_t[1] = 0; - w1_t[2] = 0; - w1_t[3] = 0; - w2_t[0] = 0; - w2_t[1] = 0; - w2_t[2] = 0; - w2_t[3] = 0; - w3_t[0] = 0; - w3_t[1] = 0; - w3_t[2] = 21 * 16 * 8; - w3_t[3] = 0; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); - - // now the 40 bit input for the MD5 which then will generate the RC4 key, so it's precomputable! - - w0_t[0] = digest[0]; - w0_t[1] = digest[1] & 0xff; - w0_t[2] = 0x8000; - w0_t[3] = 0; - w1_t[0] = 0; - w1_t[1] = 0; - w1_t[2] = 0; - w1_t[3] = 0; - w2_t[0] = 0; - w2_t[1] = 0; - w2_t[2] = 0; - w2_t[3] = 0; - w3_t[0] = 0; - w3_t[1] = 0; - w3_t[2] = 9 * 8; - w3_t[3] = 0; - - digest[0] = MD5M_A; - digest[1] = MD5M_B; - digest[2] = MD5M_C; - digest[3] = MD5M_D; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); - - // now the RC4 part - - u32x key[4]; - - key[0] = digest[0]; - key[1] = digest[1]; - key[2] = digest[2]; - key[3] = digest[3]; - - rc4_init_16 (rc4_key, key); - - u32x out[4]; - - u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); - - w0_t[0] = out[0]; - w0_t[1] = out[1]; - w0_t[2] = out[2]; - w0_t[3] = out[3]; - w1_t[0] = 0x80; - w1_t[1] = 0; - w1_t[2] = 0; - w1_t[3] = 0; - w2_t[0] = 0; - w2_t[1] = 0; - w2_t[2] = 0; - w2_t[3] = 0; - w3_t[0] = 0; - w3_t[1] = 0; - w3_t[2] = 16 * 8; - w3_t[3] = 0; - - digest[0] = MD5M_A; - digest[1] = MD5M_B; - digest[2] = MD5M_C; - digest[3] = MD5M_D; - - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); - - rc4_next_16 (rc4_key, 16, j, digest, out); - - COMPARE_M_SIMD (out[0], out[1], out[2], out[3]); - } -} - -static void m09700s (__local RC4_KEY rc4_keys[64], 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 oldoffice01_t *oldoffice01_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) -{ - /** - * modifier - */ - - const u32 gid = get_global_id (0); - const u32 lid = get_local_id (0); - - __local RC4_KEY *rc4_key = &rc4_keys[lid]; - - /** - * digest - */ - - const u32 search[4] = - { - digests_buf[digests_offset].digest_buf[DGST_R0], - digests_buf[digests_offset].digest_buf[DGST_R1], - digests_buf[digests_offset].digest_buf[DGST_R2], - digests_buf[digests_offset].digest_buf[DGST_R3] - }; - - /** - * salt - */ - - u32 salt_buf_t0[4]; - u32 salt_buf_t1[5]; - u32 salt_buf_t2[5]; - u32 salt_buf_t3[5]; - - salt_buf_t0[0] = salt_bufs[salt_pos].salt_buf[0]; - salt_buf_t0[1] = salt_bufs[salt_pos].salt_buf[1]; - salt_buf_t0[2] = salt_bufs[salt_pos].salt_buf[2]; - salt_buf_t0[3] = salt_bufs[salt_pos].salt_buf[3]; - - salt_buf_t1[0] = salt_buf_t0[0] << 8; - salt_buf_t1[1] = salt_buf_t0[0] >> 24 | salt_buf_t0[1] << 8; - salt_buf_t1[2] = salt_buf_t0[1] >> 24 | salt_buf_t0[2] << 8; - salt_buf_t1[3] = salt_buf_t0[2] >> 24 | salt_buf_t0[3] << 8; - salt_buf_t1[4] = salt_buf_t0[3] >> 24; - - salt_buf_t2[0] = salt_buf_t0[0] << 16; - salt_buf_t2[1] = salt_buf_t0[0] >> 16 | salt_buf_t0[1] << 16; - salt_buf_t2[2] = salt_buf_t0[1] >> 16 | salt_buf_t0[2] << 16; - salt_buf_t2[3] = salt_buf_t0[2] >> 16 | salt_buf_t0[3] << 16; - salt_buf_t2[4] = salt_buf_t0[3] >> 16; - - salt_buf_t3[0] = salt_buf_t0[0] << 24; - salt_buf_t3[1] = salt_buf_t0[0] >> 8 | salt_buf_t0[1] << 24; - salt_buf_t3[2] = salt_buf_t0[1] >> 8 | salt_buf_t0[2] << 24; - salt_buf_t3[3] = salt_buf_t0[2] >> 8 | salt_buf_t0[3] << 24; - salt_buf_t3[4] = salt_buf_t0[3] >> 8; - - const u32 salt_len = 16; - - /** - * esalt - */ - - const u32 version = oldoffice01_bufs[salt_pos].version; - - u32 encryptedVerifier[4]; - - encryptedVerifier[0] = oldoffice01_bufs[salt_pos].encryptedVerifier[0]; - encryptedVerifier[1] = oldoffice01_bufs[salt_pos].encryptedVerifier[1]; - encryptedVerifier[2] = oldoffice01_bufs[salt_pos].encryptedVerifier[2]; - encryptedVerifier[3] = oldoffice01_bufs[salt_pos].encryptedVerifier[3]; - - /** - * loop - */ - - u32 w0l = w0[0]; - - for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) - { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); - - const u32x w0lr = w0l | w0r; - - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; - - w0_t[0] = w0lr; - w0_t[1] = w0[1]; - w0_t[2] = w0[2]; - w0_t[3] = w0[3]; - w1_t[0] = w1[0]; - w1_t[1] = w1[1]; - w1_t[2] = w1[2]; - w1_t[3] = w1[3]; - w2_t[0] = w2[0]; - w2_t[1] = w2[1]; - w2_t[2] = w2[2]; - w2_t[3] = w2[3]; - w3_t[0] = w3[0]; - w3_t[1] = w3[1]; - w3_t[2] = pw_len * 8; - w3_t[3] = 0; - - u32x digest_t0[4]; - u32x digest_t1[2]; // need only first 5 byte - u32x digest_t2[2]; - u32x digest_t3[2]; + u32 digest_t0[4]; + u32 digest_t1[2]; // need only first 5 byte + u32 digest_t2[2]; + u32 digest_t3[2]; digest_t0[0] = MD5M_A; digest_t0[1] = MD5M_B; @@ -1188,7 +683,7 @@ static void m09700s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ // now the RC4 part - u32x key[4]; + u32 key[4]; key[0] = digest[0]; key[1] = digest[1]; @@ -1197,7 +692,510 @@ static void m09700s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; + + u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); + + w0_t[0] = out[0]; + w0_t[1] = out[1]; + w0_t[2] = out[2]; + w0_t[3] = out[3]; + w1_t[0] = 0x80; + w1_t[1] = 0; + w1_t[2] = 0; + w1_t[3] = 0; + w2_t[0] = 0; + w2_t[1] = 0; + w2_t[2] = 0; + w2_t[3] = 0; + w3_t[0] = 0; + w3_t[1] = 0; + w3_t[2] = 16 * 8; + w3_t[3] = 0; + + digest[0] = MD5M_A; + digest[1] = MD5M_B; + digest[2] = MD5M_C; + digest[3] = MD5M_D; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + + rc4_next_16 (rc4_key, 16, j, digest, out); + + COMPARE_M_SIMD (out[0], out[1], out[2], out[3]); + } +} + +static void m09700s (__local RC4_KEY rc4_keys[64], 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 oldoffice01_t *oldoffice01_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) +{ + /** + * modifier + */ + + const u32 gid = get_global_id (0); + const u32 lid = get_local_id (0); + + __local RC4_KEY *rc4_key = &rc4_keys[lid]; + + /** + * digest + */ + + const u32 search[4] = + { + digests_buf[digests_offset].digest_buf[DGST_R0], + digests_buf[digests_offset].digest_buf[DGST_R1], + digests_buf[digests_offset].digest_buf[DGST_R2], + digests_buf[digests_offset].digest_buf[DGST_R3] + }; + + /** + * salt + */ + + u32 salt_buf_t0[4]; + u32 salt_buf_t1[5]; + u32 salt_buf_t2[5]; + u32 salt_buf_t3[5]; + + salt_buf_t0[0] = salt_bufs[salt_pos].salt_buf[0]; + salt_buf_t0[1] = salt_bufs[salt_pos].salt_buf[1]; + salt_buf_t0[2] = salt_bufs[salt_pos].salt_buf[2]; + salt_buf_t0[3] = salt_bufs[salt_pos].salt_buf[3]; + + salt_buf_t1[0] = salt_buf_t0[0] << 8; + salt_buf_t1[1] = salt_buf_t0[0] >> 24 | salt_buf_t0[1] << 8; + salt_buf_t1[2] = salt_buf_t0[1] >> 24 | salt_buf_t0[2] << 8; + salt_buf_t1[3] = salt_buf_t0[2] >> 24 | salt_buf_t0[3] << 8; + salt_buf_t1[4] = salt_buf_t0[3] >> 24; + + salt_buf_t2[0] = salt_buf_t0[0] << 16; + salt_buf_t2[1] = salt_buf_t0[0] >> 16 | salt_buf_t0[1] << 16; + salt_buf_t2[2] = salt_buf_t0[1] >> 16 | salt_buf_t0[2] << 16; + salt_buf_t2[3] = salt_buf_t0[2] >> 16 | salt_buf_t0[3] << 16; + salt_buf_t2[4] = salt_buf_t0[3] >> 16; + + salt_buf_t3[0] = salt_buf_t0[0] << 24; + salt_buf_t3[1] = salt_buf_t0[0] >> 8 | salt_buf_t0[1] << 24; + salt_buf_t3[2] = salt_buf_t0[1] >> 8 | salt_buf_t0[2] << 24; + salt_buf_t3[3] = salt_buf_t0[2] >> 8 | salt_buf_t0[3] << 24; + salt_buf_t3[4] = salt_buf_t0[3] >> 8; + + const u32 salt_len = 16; + + /** + * esalt + */ + + const u32 version = oldoffice01_bufs[salt_pos].version; + + u32 encryptedVerifier[4]; + + encryptedVerifier[0] = oldoffice01_bufs[salt_pos].encryptedVerifier[0]; + encryptedVerifier[1] = oldoffice01_bufs[salt_pos].encryptedVerifier[1]; + encryptedVerifier[2] = oldoffice01_bufs[salt_pos].encryptedVerifier[2]; + encryptedVerifier[3] = oldoffice01_bufs[salt_pos].encryptedVerifier[3]; + + /** + * loop + */ + + u32 w0l = w0[0]; + + for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) + { + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); + + const u32 w0lr = w0l | w0r; + + u32 w0_t[4]; + u32 w1_t[4]; + u32 w2_t[4]; + u32 w3_t[4]; + + w0_t[0] = w0lr; + w0_t[1] = w0[1]; + w0_t[2] = w0[2]; + w0_t[3] = w0[3]; + w1_t[0] = w1[0]; + w1_t[1] = w1[1]; + w1_t[2] = w1[2]; + w1_t[3] = w1[3]; + w2_t[0] = w2[0]; + w2_t[1] = w2[1]; + w2_t[2] = w2[2]; + w2_t[3] = w2[3]; + w3_t[0] = w3[0]; + w3_t[1] = w3[1]; + w3_t[2] = pw_len * 8; + w3_t[3] = 0; + + u32 digest_t0[4]; + u32 digest_t1[2]; // need only first 5 byte + u32 digest_t2[2]; + u32 digest_t3[2]; + + digest_t0[0] = MD5M_A; + digest_t0[1] = MD5M_B; + digest_t0[2] = MD5M_C; + digest_t0[3] = MD5M_D; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest_t0); + + // prepare 16 * 21 buffer stuff + + u32 digest[4]; + + digest[0] = MD5M_A; + digest[1] = MD5M_B; + digest[2] = MD5M_C; + digest[3] = MD5M_D; + + // offsets + + digest_t0[0] &= 0xffffffff; + digest_t0[1] &= 0x000000ff; + digest_t0[2] &= 0x00000000; + digest_t0[3] &= 0x00000000; + + digest_t1[0] = digest_t0[0] << 8; + digest_t1[1] = digest_t0[0] >> 24 | digest_t0[1] << 8; + + digest_t2[0] = digest_t0[0] << 16; + digest_t2[1] = digest_t0[0] >> 16 | digest_t0[1] << 16; + + digest_t3[0] = digest_t0[0] << 24; + digest_t3[1] = digest_t0[0] >> 8 | digest_t0[1] << 24; + + // generate the 16 * 21 buffer + + w0_t[0] = 0; + w0_t[1] = 0; + w0_t[2] = 0; + w0_t[3] = 0; + w1_t[0] = 0; + w1_t[1] = 0; + w1_t[2] = 0; + w1_t[3] = 0; + w2_t[0] = 0; + w2_t[1] = 0; + w2_t[2] = 0; + w2_t[3] = 0; + w3_t[0] = 0; + w3_t[1] = 0; + w3_t[2] = 0; + w3_t[3] = 0; + + // 0..5 + w0_t[0] = digest_t0[0]; + w0_t[1] = digest_t0[1]; + + // 5..21 + w0_t[1] |= salt_buf_t1[0]; + w0_t[2] = salt_buf_t1[1]; + w0_t[3] = salt_buf_t1[2]; + w1_t[0] = salt_buf_t1[3]; + w1_t[1] = salt_buf_t1[4]; + + // 21..26 + w1_t[1] |= digest_t1[0]; + w1_t[2] = digest_t1[1]; + + // 26..42 + w1_t[2] |= salt_buf_t2[0]; + w1_t[3] = salt_buf_t2[1]; + w2_t[0] = salt_buf_t2[2]; + w2_t[1] = salt_buf_t2[3]; + w2_t[2] = salt_buf_t2[4]; + + // 42..47 + w2_t[2] |= digest_t2[0]; + w2_t[3] = digest_t2[1]; + + // 47..63 + w2_t[3] |= salt_buf_t3[0]; + w3_t[0] = salt_buf_t3[1]; + w3_t[1] = salt_buf_t3[2]; + w3_t[2] = salt_buf_t3[3]; + w3_t[3] = salt_buf_t3[4]; + + // 63.. + + w3_t[3] |= digest_t3[0]; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + + w0_t[0] = 0; + w0_t[1] = 0; + w0_t[2] = 0; + w0_t[3] = 0; + w1_t[0] = 0; + w1_t[1] = 0; + w1_t[2] = 0; + w1_t[3] = 0; + w2_t[0] = 0; + w2_t[1] = 0; + w2_t[2] = 0; + w2_t[3] = 0; + w3_t[0] = 0; + w3_t[1] = 0; + w3_t[2] = 0; + w3_t[3] = 0; + + // 0..4 + w0_t[0] = digest_t3[1]; + + // 4..20 + w0_t[1] = salt_buf_t0[0]; + w0_t[2] = salt_buf_t0[1]; + w0_t[3] = salt_buf_t0[2]; + w1_t[0] = salt_buf_t0[3]; + + // 20..25 + w1_t[1] = digest_t0[0]; + w1_t[2] = digest_t0[1]; + + // 25..41 + w1_t[2] |= salt_buf_t1[0]; + w1_t[3] = salt_buf_t1[1]; + w2_t[0] = salt_buf_t1[2]; + w2_t[1] = salt_buf_t1[3]; + w2_t[2] = salt_buf_t1[4]; + + // 41..46 + w2_t[2] |= digest_t1[0]; + w2_t[3] = digest_t1[1]; + + // 46..62 + w2_t[3] |= salt_buf_t2[0]; + w3_t[0] = salt_buf_t2[1]; + w3_t[1] = salt_buf_t2[2]; + w3_t[2] = salt_buf_t2[3]; + w3_t[3] = salt_buf_t2[4]; + + // 62.. + w3_t[3] |= digest_t2[0]; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + + w0_t[0] = 0; + w0_t[1] = 0; + w0_t[2] = 0; + w0_t[3] = 0; + w1_t[0] = 0; + w1_t[1] = 0; + w1_t[2] = 0; + w1_t[3] = 0; + w2_t[0] = 0; + w2_t[1] = 0; + w2_t[2] = 0; + w2_t[3] = 0; + w3_t[0] = 0; + w3_t[1] = 0; + w3_t[2] = 0; + w3_t[3] = 0; + + // 0..3 + w0_t[0] = digest_t2[1]; + + // 3..19 + w0_t[0] |= salt_buf_t3[0]; + w0_t[1] = salt_buf_t3[1]; + w0_t[2] = salt_buf_t3[2]; + w0_t[3] = salt_buf_t3[3]; + w1_t[0] = salt_buf_t3[4]; + + // 19..24 + w1_t[0] |= digest_t3[0]; + w1_t[1] = digest_t3[1]; + + // 24..40 + w1_t[2] = salt_buf_t0[0]; + w1_t[3] = salt_buf_t0[1]; + w2_t[0] = salt_buf_t0[2]; + w2_t[1] = salt_buf_t0[3]; + + // 40..45 + w2_t[2] = digest_t0[0]; + w2_t[3] = digest_t0[1]; + + // 45..61 + w2_t[3] |= salt_buf_t1[0]; + w3_t[0] = salt_buf_t1[1]; + w3_t[1] = salt_buf_t1[2]; + w3_t[2] = salt_buf_t1[3]; + w3_t[3] = salt_buf_t1[4]; + + // 61.. + w3_t[3] |= digest_t1[0]; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + + w0_t[0] = 0; + w0_t[1] = 0; + w0_t[2] = 0; + w0_t[3] = 0; + w1_t[0] = 0; + w1_t[1] = 0; + w1_t[2] = 0; + w1_t[3] = 0; + w2_t[0] = 0; + w2_t[1] = 0; + w2_t[2] = 0; + w2_t[3] = 0; + w3_t[0] = 0; + w3_t[1] = 0; + w3_t[2] = 0; + w3_t[3] = 0; + + // 0..2 + w0_t[0] = digest_t1[1]; + + // 2..18 + w0_t[0] |= salt_buf_t2[0]; + w0_t[1] = salt_buf_t2[1]; + w0_t[2] = salt_buf_t2[2]; + w0_t[3] = salt_buf_t2[3]; + w1_t[0] = salt_buf_t2[4]; + + // 18..23 + w1_t[0] |= digest_t2[0]; + w1_t[1] = digest_t2[1]; + + // 23..39 + w1_t[1] |= salt_buf_t3[0]; + w1_t[2] = salt_buf_t3[1]; + w1_t[3] = salt_buf_t3[2]; + w2_t[0] = salt_buf_t3[3]; + w2_t[1] = salt_buf_t3[4]; + + // 39..44 + w2_t[1] |= digest_t3[0]; + w2_t[2] = digest_t3[1]; + + // 44..60 + w2_t[3] = salt_buf_t0[0]; + w3_t[0] = salt_buf_t0[1]; + w3_t[1] = salt_buf_t0[2]; + w3_t[2] = salt_buf_t0[3]; + + // 60.. + w3_t[3] = digest_t0[0]; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + + w0_t[0] = 0; + w0_t[1] = 0; + w0_t[2] = 0; + w0_t[3] = 0; + w1_t[0] = 0; + w1_t[1] = 0; + w1_t[2] = 0; + w1_t[3] = 0; + w2_t[0] = 0; + w2_t[1] = 0; + w2_t[2] = 0; + w2_t[3] = 0; + w3_t[0] = 0; + w3_t[1] = 0; + w3_t[2] = 0; + w3_t[3] = 0; + + // 0..1 + w0_t[0] = digest_t0[1]; + + // 1..17 + w0_t[0] |= salt_buf_t1[0]; + w0_t[1] = salt_buf_t1[1]; + w0_t[2] = salt_buf_t1[2]; + w0_t[3] = salt_buf_t1[3]; + w1_t[0] = salt_buf_t1[4]; + + // 17..22 + w1_t[0] |= digest_t1[0]; + w1_t[1] = digest_t1[1]; + + // 22..38 + w1_t[1] |= salt_buf_t2[0]; + w1_t[2] = salt_buf_t2[1]; + w1_t[3] = salt_buf_t2[2]; + w2_t[0] = salt_buf_t2[3]; + w2_t[1] = salt_buf_t2[4]; + + // 38..43 + w2_t[1] |= digest_t2[0]; + w2_t[2] = digest_t2[1]; + + // 43..59 + w2_t[2] |= salt_buf_t3[0]; + w2_t[3] = salt_buf_t3[1]; + w3_t[0] = salt_buf_t3[2]; + w3_t[1] = salt_buf_t3[3]; + w3_t[2] = salt_buf_t3[4]; + + // 59.. + w3_t[2] |= digest_t3[0]; + w3_t[3] = digest_t3[1]; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + + w0_t[0] = salt_buf_t0[0]; + w0_t[1] = salt_buf_t0[1]; + w0_t[2] = salt_buf_t0[2]; + w0_t[3] = salt_buf_t0[3]; + w1_t[0] = 0x80; + w1_t[1] = 0; + w1_t[2] = 0; + w1_t[3] = 0; + w2_t[0] = 0; + w2_t[1] = 0; + w2_t[2] = 0; + w2_t[3] = 0; + w3_t[0] = 0; + w3_t[1] = 0; + w3_t[2] = 21 * 16 * 8; + w3_t[3] = 0; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + + // now the 40 bit input for the MD5 which then will generate the RC4 key, so it's precomputable! + + w0_t[0] = digest[0]; + w0_t[1] = digest[1] & 0xff; + w0_t[2] = 0x8000; + w0_t[3] = 0; + w1_t[0] = 0; + w1_t[1] = 0; + w1_t[2] = 0; + w1_t[3] = 0; + w2_t[0] = 0; + w2_t[1] = 0; + w2_t[2] = 0; + w2_t[3] = 0; + w3_t[0] = 0; + w3_t[1] = 0; + w3_t[2] = 9 * 8; + w3_t[3] = 0; + + digest[0] = MD5M_A; + digest[1] = MD5M_B; + digest[2] = MD5M_C; + digest[3] = MD5M_D; + + md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + + // now the RC4 part + + u32 key[4]; + + key[0] = digest[0]; + key[1] = digest[1]; + key[2] = digest[2]; + key[3] = digest[3]; + + rc4_init_16 (rc4_key, key); + + u32 out[4]; u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); diff --git a/OpenCL/m09710_a3.cl b/OpenCL/m09710_a3.cl index 76cfef13d..229e83653 100644 --- a/OpenCL/m09710_a3.cl +++ b/OpenCL/m09710_a3.cl @@ -5,8 +5,6 @@ #define _OLDOFFICE01_ -#define NEW_SIMD_CODE - #include "include/constants.h" #include "include/kernel_vendor.h" @@ -37,10 +35,10 @@ static void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j) rc4_key->S[j] = tmp; } -static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) +static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4]) { - u32x v = 0x03020100; - u32x a = 0x04040404; + u32 v = 0x03020100; + u32 a = 0x04040404; __local u32 *ptr = (__local u32 *) rc4_key->S; @@ -88,7 +86,7 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) } } -static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32x in[4], u32x out[4]) +static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4]) { #pragma unroll for (u32 k = 0; k < 4; k++) @@ -139,29 +137,29 @@ static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32x in[4], u return j; } -static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4]) +static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4]) { - u32x a = digest[0]; - u32x b = digest[1]; - u32x c = digest[2]; - u32x d = digest[3]; + u32 a = digest[0]; + u32 b = digest[1]; + u32 c = digest[2]; + u32 d = digest[3]; - u32x w0_t = w0[0]; - u32x w1_t = w0[1]; - u32x w2_t = w0[2]; - u32x w3_t = w0[3]; - u32x w4_t = w1[0]; - u32x w5_t = w1[1]; - u32x w6_t = w1[2]; - u32x w7_t = w1[3]; - u32x w8_t = w2[0]; - u32x w9_t = w2[1]; - u32x wa_t = w2[2]; - u32x wb_t = w2[3]; - u32x wc_t = w3[0]; - u32x wd_t = w3[1]; - u32x we_t = w3[2]; - u32x wf_t = w3[3]; + u32 w0_t = w0[0]; + u32 w1_t = w0[1]; + u32 w2_t = w0[2]; + u32 w3_t = w0[3]; + u32 w4_t = w1[0]; + u32 w5_t = w1[1]; + u32 w6_t = w1[2]; + u32 w7_t = w1[3]; + u32 w8_t = w2[0]; + u32 w9_t = w2[1]; + u32 wa_t = w2[2]; + u32 wb_t = w2[3]; + u32 wc_t = w3[0]; + u32 wd_t = w3[1]; + u32 we_t = w3[2]; + u32 wf_t = w3[3]; MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00); MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01); @@ -269,16 +267,16 @@ static void m09710m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; // first md5 to generate RC4 128 bit key - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + u32 w0_t[4]; + u32 w1_t[4]; + u32 w2_t[4]; + u32 w3_t[4]; w0_t[0] = w0lr; w0_t[1] = w0[1] & 0xff; @@ -297,7 +295,7 @@ static void m09710m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ w3_t[2] = 9 * 8; w3_t[3] = 0; - u32x digest[4]; + u32 digest[4]; digest[0] = MD5M_A; digest[1] = MD5M_B; @@ -308,7 +306,7 @@ static void m09710m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ // now the RC4 part - u32x key[4]; + u32 key[4]; key[0] = digest[0]; key[1] = digest[1]; @@ -317,7 +315,7 @@ static void m09710m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); @@ -395,16 +393,16 @@ static void m09710s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; // first md5 to generate RC4 128 bit key - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + u32 w0_t[4]; + u32 w1_t[4]; + u32 w2_t[4]; + u32 w3_t[4]; w0_t[0] = w0lr; w0_t[1] = w0[1] & 0xff; @@ -423,7 +421,7 @@ static void m09710s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ w3_t[2] = 9 * 8; w3_t[3] = 0; - u32x digest[4]; + u32 digest[4]; digest[0] = MD5M_A; digest[1] = MD5M_B; @@ -434,7 +432,7 @@ static void m09710s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ // now the RC4 part - u32x key[4]; + u32 key[4]; key[0] = digest[0]; key[1] = digest[1]; @@ -443,7 +441,7 @@ static void m09710s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); diff --git a/OpenCL/m09800_a3.cl b/OpenCL/m09800_a3.cl index 652756e41..31f2ea6bb 100644 --- a/OpenCL/m09800_a3.cl +++ b/OpenCL/m09800_a3.cl @@ -5,8 +5,6 @@ #define _OLDOFFICE34_ -#define NEW_SIMD_CODE - #include "include/constants.h" #include "include/kernel_vendor.h" @@ -37,10 +35,10 @@ static void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j) rc4_key->S[j] = tmp; } -static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) +static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4]) { - u32x v = 0x03020100; - u32x a = 0x04040404; + u32 v = 0x03020100; + u32 a = 0x04040404; __local u32 *ptr = (__local u32 *) rc4_key->S; @@ -56,7 +54,7 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) { u32 idx = i * 16; - u32x v; + u32 v; v = data[0]; @@ -88,12 +86,12 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) } } -static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32x in[4], u32x out[4]) +static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4]) { #pragma unroll for (u32 k = 0; k < 4; k++) { - u32x xor4 = 0; + u32 xor4 = 0; u8 idx; @@ -139,30 +137,30 @@ static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32x in[4], u return j; } -static void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5]) +static void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5]) { - u32x A = digest[0]; - u32x B = digest[1]; - u32x C = digest[2]; - u32x D = digest[3]; - u32x E = digest[4]; + u32 A = digest[0]; + u32 B = digest[1]; + u32 C = digest[2]; + u32 D = digest[3]; + u32 E = digest[4]; - u32x w0_t = w0[0]; - u32x w1_t = w0[1]; - u32x w2_t = w0[2]; - u32x w3_t = w0[3]; - u32x w4_t = w1[0]; - u32x w5_t = w1[1]; - u32x w6_t = w1[2]; - u32x w7_t = w1[3]; - u32x w8_t = w2[0]; - u32x w9_t = w2[1]; - u32x wa_t = w2[2]; - u32x wb_t = w2[3]; - u32x wc_t = w3[0]; - u32x wd_t = w3[1]; - u32x we_t = w3[2]; - u32x wf_t = w3[3]; + u32 w0_t = w0[0]; + u32 w1_t = w0[1]; + u32 w2_t = w0[2]; + u32 w3_t = w0[3]; + u32 w4_t = w1[0]; + u32 w5_t = w1[1]; + u32 w6_t = w1[2]; + u32 w7_t = w1[3]; + u32 w8_t = w2[0]; + u32 w9_t = w2[1]; + u32 wa_t = w2[2]; + u32 wb_t = w2[3]; + u32 wc_t = w3[0]; + u32 wd_t = w3[1]; + u32 we_t = w3[2]; + u32 wf_t = w3[3]; #undef K #define K SHA1C00 @@ -314,14 +312,14 @@ static void m09800m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + u32 w0_t[4]; + u32 w1_t[4]; + u32 w2_t[4]; + u32 w3_t[4]; w0_t[0] = salt_buf[0]; w0_t[1] = salt_buf[1]; @@ -340,7 +338,7 @@ static void m09800m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ w3_t[2] = 0; w3_t[3] = pw_salt_len * 8; - u32x digest[5]; + u32 digest[5]; digest[0] = SHA1M_A; digest[1] = SHA1M_B; @@ -375,7 +373,7 @@ static void m09800m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); - u32x key[4]; + u32 key[4]; key[0] = swap32 (digest[0]); key[1] = swap32 (digest[1]); @@ -391,7 +389,7 @@ static void m09800m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); @@ -490,14 +488,14 @@ static void m09800s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + u32 w0_t[4]; + u32 w1_t[4]; + u32 w2_t[4]; + u32 w3_t[4]; w0_t[0] = salt_buf[0]; w0_t[1] = salt_buf[1]; @@ -516,7 +514,7 @@ static void m09800s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ w3_t[2] = 0; w3_t[3] = pw_salt_len * 8; - u32x digest[5]; + u32 digest[5]; digest[0] = SHA1M_A; digest[1] = SHA1M_B; @@ -551,7 +549,7 @@ static void m09800s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ sha1_transform (w0_t, w1_t, w2_t, w3_t, digest); - u32x key[4]; + u32 key[4]; key[0] = swap32 (digest[0]); key[1] = swap32 (digest[1]); @@ -567,7 +565,7 @@ static void m09800s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); diff --git a/OpenCL/m09810_a3.cl b/OpenCL/m09810_a3.cl index a6fd83f19..270e83a4a 100644 --- a/OpenCL/m09810_a3.cl +++ b/OpenCL/m09810_a3.cl @@ -5,8 +5,6 @@ #define _OLDOFFICE34_ -#define NEW_SIMD_CODE - #include "include/constants.h" #include "include/kernel_vendor.h" @@ -37,10 +35,10 @@ static void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j) rc4_key->S[j] = tmp; } -static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) +static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4]) { - u32x v = 0x03020100; - u32x a = 0x04040404; + u32 v = 0x03020100; + u32 a = 0x04040404; __local u32 *ptr = (__local u32 *) rc4_key->S; @@ -56,7 +54,7 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) { u32 idx = i * 16; - u32x v; + u32 v; v = data[0]; @@ -88,12 +86,12 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) } } -static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32x in[4], u32x out[4]) +static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32 in[4], u32 out[4]) { #pragma unroll for (u32 k = 0; k < 4; k++) { - u32x xor4 = 0; + u32 xor4 = 0; u8 idx; @@ -139,30 +137,30 @@ static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, const u32x in[4], u return j; } -static void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5]) +static void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5]) { - u32x A = digest[0]; - u32x B = digest[1]; - u32x C = digest[2]; - u32x D = digest[3]; - u32x E = digest[4]; + u32 A = digest[0]; + u32 B = digest[1]; + u32 C = digest[2]; + u32 D = digest[3]; + u32 E = digest[4]; - u32x w0_t = w0[0]; - u32x w1_t = w0[1]; - u32x w2_t = w0[2]; - u32x w3_t = w0[3]; - u32x w4_t = w1[0]; - u32x w5_t = w1[1]; - u32x w6_t = w1[2]; - u32x w7_t = w1[3]; - u32x w8_t = w2[0]; - u32x w9_t = w2[1]; - u32x wa_t = w2[2]; - u32x wb_t = w2[3]; - u32x wc_t = w3[0]; - u32x wd_t = w3[1]; - u32x we_t = w3[2]; - u32x wf_t = w3[3]; + u32 w0_t = w0[0]; + u32 w1_t = w0[1]; + u32 w2_t = w0[2]; + u32 w3_t = w0[3]; + u32 w4_t = w1[0]; + u32 w5_t = w1[1]; + u32 w6_t = w1[2]; + u32 w7_t = w1[3]; + u32 w8_t = w2[0]; + u32 w9_t = w2[1]; + u32 wa_t = w2[2]; + u32 wb_t = w2[3]; + u32 wc_t = w3[0]; + u32 wd_t = w3[1]; + u32 we_t = w3[2]; + u32 wf_t = w3[3]; #undef K #define K SHA1C00 @@ -299,11 +297,11 @@ static void m09810m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; - u32x key[4]; + u32 key[4]; key[0] = w0lr; key[1] = w0[1] & 0xff; @@ -312,14 +310,14 @@ static void m09810m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + u32 w0_t[4]; + u32 w1_t[4]; + u32 w2_t[4]; + u32 w3_t[4]; w0_t[0] = swap32 (out[0]); w0_t[1] = swap32 (out[1]); @@ -338,7 +336,7 @@ static void m09810m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ w3_t[2] = 0; w3_t[3] = 16 * 8; - u32x digest[5]; + u32 digest[5]; digest[0] = SHA1M_A; digest[1] = SHA1M_B; @@ -403,11 +401,11 @@ static void m09810s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; - u32x key[4]; + u32 key[4]; key[0] = w0lr; key[1] = w0[1] & 0xff; @@ -416,14 +414,14 @@ static void m09810s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; u8 j = rc4_next_16 (rc4_key, 0, 0, encryptedVerifier, out); - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + u32 w0_t[4]; + u32 w1_t[4]; + u32 w2_t[4]; + u32 w3_t[4]; w0_t[0] = swap32 (out[0]); w0_t[1] = swap32 (out[1]); @@ -442,7 +440,7 @@ static void m09810s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ w3_t[2] = 0; w3_t[3] = 16 * 8; - u32x digest[5]; + u32 digest[5]; digest[0] = SHA1M_A; digest[1] = SHA1M_B; diff --git a/OpenCL/m10400_a3.cl b/OpenCL/m10400_a3.cl index 5730bb926..c626ef853 100644 --- a/OpenCL/m10400_a3.cl +++ b/OpenCL/m10400_a3.cl @@ -16,9 +16,7 @@ #include "include/kernel_functions.c" #include "OpenCL/types_ocl.c" #include "OpenCL/common.c" - -#define COMPARE_S "OpenCL/check_single_comp4.c" -#define COMPARE_M "OpenCL/check_multi_comp4.c" +#include "OpenCL/simd.c" __constant u32 padding[8] = { @@ -275,7 +273,7 @@ static void m10400m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++) { - const u32 w0r = bfs_buf[il_pos].i; + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); w0[0] = w0l | w0r; @@ -369,12 +367,7 @@ static void m10400m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_next_16 (rc4_key, 0, 0, padding, out); - const u32 r0 = out[0]; - const u32 r1 = out[1]; - const u32 r2 = out[2]; - const u32 r3 = out[3]; - - #include COMPARE_M + COMPARE_M_SIMD (out[0], out[1], out[2], out[3]); } } @@ -433,7 +426,7 @@ static void m10400s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++) { - const u32 w0r = bfs_buf[il_pos].i; + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); w0[0] = w0l | w0r; @@ -527,12 +520,7 @@ static void m10400s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_next_16 (rc4_key, 0, 0, padding, out); - const u32 r0 = out[0]; - const u32 r1 = out[1]; - const u32 r2 = out[2]; - const u32 r3 = out[3]; - - #include COMPARE_S + COMPARE_S_SIMD (out[0], out[1], out[2], out[3]); } } diff --git a/OpenCL/m10410_a3.cl b/OpenCL/m10410_a3.cl index 53f9c0713..497c2063d 100644 --- a/OpenCL/m10410_a3.cl +++ b/OpenCL/m10410_a3.cl @@ -5,8 +5,6 @@ #define _MD5_ -#define NEW_SIMD_CODE - #include "include/constants.h" #include "include/kernel_vendor.h" @@ -49,10 +47,10 @@ static void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j) rc4_key->S[j] = tmp; } -static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) +static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4]) { - u32x v = 0x03020100; - u32x a = 0x04040404; + u32 v = 0x03020100; + u32 a = 0x04040404; __local u32 *ptr = (__local u32 *) rc4_key->S; @@ -62,11 +60,11 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) ptr[i] = v; v += a; } - const u32x d0 = data[0] >> 0; - const u32x d1 = data[0] >> 8; - const u32x d2 = data[0] >> 16; - const u32x d3 = data[0] >> 24; - const u32x d4 = data[1] >> 0; + const u32 d0 = data[0] >> 0; + const u32 d1 = data[0] >> 8; + const u32 d2 = data[0] >> 16; + const u32 d3 = data[0] >> 24; + const u32 d4 = data[1] >> 0; u32 j = 0; @@ -83,11 +81,11 @@ static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32x data[4]) j += rc4_key->S[255] + d0; swap (rc4_key, 255, j); } -static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32x in[4], u32x out[4]) +static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 in[4], u32 out[4]) { for (u32 k = 0; k < 4; k++) { - u32x xor4 = 0; + u32 xor4 = 0; u8 idx; @@ -152,13 +150,13 @@ static void m10410m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; // now the RC4 part - u32x key[4]; + u32 key[4]; key[0] = w0lr; key[1] = w0[1]; @@ -167,7 +165,7 @@ static void m10410m (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; rc4_next_16 (rc4_key, 0, 0, padding, out); @@ -206,13 +204,13 @@ static void m10410s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE) { - const u32x w0r = w0r_create_bft (bfs_buf, il_pos); + const u32 w0r = w0r_create_bft (bfs_buf, il_pos); - const u32x w0lr = w0l | w0r; + const u32 w0lr = w0l | w0r; // now the RC4 part - u32x key[4]; + u32 key[4]; key[0] = w0lr; key[1] = w0[1]; @@ -221,7 +219,7 @@ static void m10410s (__local RC4_KEY rc4_keys[64], u32 w0[4], u32 w1[4], u32 w2[ rc4_init_16 (rc4_key, key); - u32x out[4]; + u32 out[4]; rc4_next_16 (rc4_key, 0, 0, padding, out);