From 116d7620c56dfbecb44d4bd34d5a51c0b9c0a5de Mon Sep 17 00:00:00 2001 From: Jens Steube Date: Fri, 4 Mar 2016 16:54:22 +0100 Subject: [PATCH] Cleanup -m 1000 kernels to latest standard --- OpenCL/m01000_a0.cl | 252 ++++++++++----------- OpenCL/m01000_a1.cl | 521 ++++++++++++++++++++------------------------ 2 files changed, 356 insertions(+), 417 deletions(-) diff --git a/OpenCL/m01000_a0.cl b/OpenCL/m01000_a0.cl index 3c5d799f1..70fc825d0 100644 --- a/OpenCL/m01000_a0.cl +++ b/OpenCL/m01000_a0.cl @@ -39,18 +39,16 @@ __kernel void m01000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, if (gid >= gid_max) return; 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]; - 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]; + 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]; + 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; @@ -69,71 +67,67 @@ __kernel void m01000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, append_0x80_2x4_VV (w0, w1, out_len); - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + make_unicode (w1, w2, w3); + make_unicode (w0, w0, w1); - make_unicode (w0, w0_t, w1_t); - make_unicode (w1, w2_t, w3_t); - - w3_t[2] = out_len * 8 * 2; + w3[2] = out_len * 8 * 2; + w3[3] = 0; u32x a = MD4M_A; u32x b = MD4M_B; u32x c = MD4M_C; u32x d = MD4M_D; - MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); + MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00); + MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01); + MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02); + MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03); + MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00); + MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01); + MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02); + MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03); + MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00); + MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01); + MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02); + MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03); + MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00); + MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01); + MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02); + MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03); - MD4_STEP (MD4_Go, a, b, c, d, w0_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_Go, a, b, c, d, w0[0], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13); - MD4_STEP (MD4_H , a, b, c, d, w0_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); + MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23); COMPARE_M_SIMD (a, d, c, b); } @@ -164,18 +158,16 @@ __kernel void m01000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, if (gid >= gid_max) return; 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]; - 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]; + 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]; + 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; @@ -206,74 +198,70 @@ __kernel void m01000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, append_0x80_2x4_VV (w0, w1, out_len); - u32x w0_t[4]; - u32x w1_t[4]; - u32x w2_t[4]; - u32x w3_t[4]; + make_unicode (w1, w2, w3); + make_unicode (w0, w0, w1); - make_unicode (w0, w0_t, w1_t); - make_unicode (w1, w2_t, w3_t); - - w3_t[2] = out_len * 8 * 2; + w3[2] = out_len * 8 * 2; + w3[3] = 0; u32x a = MD4M_A; u32x b = MD4M_B; u32x c = MD4M_C; u32x d = MD4M_D; - MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); + MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00); + MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01); + MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02); + MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03); + MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00); + MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01); + MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02); + MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03); + MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00); + MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01); + MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02); + MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03); + MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00); + MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01); + MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02); + MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03); - MD4_STEP (MD4_Go, a, b, c, d, w0_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_Go, a, b, c, d, w0[0], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13); - MD4_STEP (MD4_H , a, b, c, d, w0_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 , a, b, c, d, w0[0], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20); 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 , c, d, a, b, w1_t[3], MD4C02, MD4S22); - MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23); + 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); COMPARE_S_SIMD (a, d, c, b); } diff --git a/OpenCL/m01000_a1.cl b/OpenCL/m01000_a1.cl index 89c93761f..d5455aa08 100644 --- a/OpenCL/m01000_a1.cl +++ b/OpenCL/m01000_a1.cl @@ -5,6 +5,8 @@ #define _MD4_ +#define NEW_SIMD_CODE + #include "include/constants.h" #include "include/kernel_vendor.h" @@ -16,9 +18,7 @@ #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" __kernel void m01000_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) { @@ -36,186 +36,160 @@ __kernel void m01000_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, if (gid >= gid_max) return; - u32 wordl0[4]; + u32 pw_buf0[4]; + u32 pw_buf1[4]; - wordl0[0] = pws[gid].i[ 0]; - wordl0[1] = pws[gid].i[ 1]; - wordl0[2] = pws[gid].i[ 2]; - wordl0[3] = pws[gid].i[ 3]; - - u32 wordl1[4]; - - wordl1[0] = pws[gid].i[ 4]; - 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; + 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]; + 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_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); - } - /** * 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]; - wordr0[1] = combs_buf[il_pos].i[1]; - wordr0[2] = combs_buf[il_pos].i[2]; - wordr0[3] = combs_buf[il_pos].i[3]; + u32x wordl0[4] = { 0 }; + u32x wordl1[4] = { 0 }; + u32x wordl2[4] = { 0 }; + 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]; - wordr1[1] = combs_buf[il_pos].i[5]; - wordr1[2] = combs_buf[il_pos].i[6]; - wordr1[3] = combs_buf[il_pos].i[7]; + u32x wordr0[4] = { 0 }; + u32x wordr1[4] = { 0 }; + u32x wordr2[4] = { 0 }; + u32x wordr3[4] = { 0 }; - u32 wordr2[4]; - - wordr2[0] = 0; - wordr2[1] = 0; - wordr2[2] = 0; - wordr2[3] = 0; - - u32 wordr3[4]; - - wordr3[0] = 0; - wordr3[1] = 0; - wordr3[2] = 0; - wordr3[3] = 0; + wordr0[0] = ix_create_combt (combs_buf, il_pos, 0); + wordr0[1] = ix_create_combt (combs_buf, il_pos, 1); + wordr0[2] = ix_create_combt (combs_buf, il_pos, 2); + wordr0[3] = ix_create_combt (combs_buf, il_pos, 3); + wordr1[0] = ix_create_combt (combs_buf, il_pos, 4); + wordr1[1] = ix_create_combt (combs_buf, il_pos, 5); + wordr1[2] = ix_create_combt (combs_buf, il_pos, 6); + wordr1[3] = ix_create_combt (combs_buf, il_pos, 7); 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[1] = wordl0[1] | wordr0[1]; w0[2] = wordl0[2] | wordr0[2]; w0[3] = wordl0[3] | wordr0[3]; - - u32 w1[4]; - w1[0] = wordl1[0] | wordr1[0]; w1[1] = wordl1[1] | wordr1[1]; w1[2] = wordl1[2] | wordr1[2]; 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; - w2[1] = 0; - w2[2] = 0; - w2[3] = 0; - - u32 w3[4]; - - w3[0] = 0; - w3[1] = 0; - w3[2] = 0; + w3[2] = pw_len * 8 * 2; w3[3] = 0; - u32 w0_t[4]; - u32 w1_t[4]; - u32 w2_t[4]; - u32 w3_t[4]; + /** + * md4 + */ - make_unicode (w0, w0_t, w1_t); - make_unicode (w1, w2_t, w3_t); + u32x a = MD4M_A; + 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; - u32 b = MD4M_B; - u32 c = MD4M_C; - u32 d = MD4M_D; + MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13); - MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); + MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23); - 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); - - const u32 r0 = a; - const u32 r1 = d; - const u32 r2 = c; - const u32 r3 = b; - - #include COMPARE_M + COMPARE_M_SIMD (a, d, c, b); } } @@ -243,43 +217,20 @@ __kernel void m01000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, if (gid >= gid_max) return; - u32 wordl0[4]; + u32 pw_buf0[4]; + u32 pw_buf1[4]; - wordl0[0] = pws[gid].i[ 0]; - wordl0[1] = pws[gid].i[ 1]; - wordl0[2] = pws[gid].i[ 2]; - wordl0[3] = pws[gid].i[ 3]; - - u32 wordl1[4]; - - wordl1[0] = pws[gid].i[ 4]; - 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; + 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]; + 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_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); - } - /** * digest */ @@ -296,145 +247,145 @@ __kernel void m01000_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, * 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]; - wordr0[1] = combs_buf[il_pos].i[1]; - wordr0[2] = combs_buf[il_pos].i[2]; - wordr0[3] = combs_buf[il_pos].i[3]; + u32x wordl0[4] = { 0 }; + u32x wordl1[4] = { 0 }; + u32x wordl2[4] = { 0 }; + 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]; - wordr1[1] = combs_buf[il_pos].i[5]; - wordr1[2] = combs_buf[il_pos].i[6]; - wordr1[3] = combs_buf[il_pos].i[7]; + u32x wordr0[4] = { 0 }; + u32x wordr1[4] = { 0 }; + u32x wordr2[4] = { 0 }; + u32x wordr3[4] = { 0 }; - u32 wordr2[4]; - - wordr2[0] = 0; - wordr2[1] = 0; - wordr2[2] = 0; - wordr2[3] = 0; - - u32 wordr3[4]; - - wordr3[0] = 0; - wordr3[1] = 0; - wordr3[2] = 0; - wordr3[3] = 0; + wordr0[0] = ix_create_combt (combs_buf, il_pos, 0); + wordr0[1] = ix_create_combt (combs_buf, il_pos, 1); + wordr0[2] = ix_create_combt (combs_buf, il_pos, 2); + wordr0[3] = ix_create_combt (combs_buf, il_pos, 3); + wordr1[0] = ix_create_combt (combs_buf, il_pos, 4); + wordr1[1] = ix_create_combt (combs_buf, il_pos, 5); + wordr1[2] = ix_create_combt (combs_buf, il_pos, 6); + wordr1[3] = ix_create_combt (combs_buf, il_pos, 7); 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[1] = wordl0[1] | wordr0[1]; w0[2] = wordl0[2] | wordr0[2]; w0[3] = wordl0[3] | wordr0[3]; - - u32 w1[4]; - w1[0] = wordl1[0] | wordr1[0]; w1[1] = wordl1[1] | wordr1[1]; w1[2] = wordl1[2] | wordr1[2]; 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; - w2[1] = 0; - w2[2] = 0; - w2[3] = 0; - - u32 w3[4]; - - w3[0] = 0; - w3[1] = 0; - w3[2] = 0; + w3[2] = pw_len * 8 * 2; w3[3] = 0; - u32 w0_t[4]; - u32 w1_t[4]; - u32 w2_t[4]; - u32 w3_t[4]; + /** + * md4 + */ - make_unicode (w0, w0_t, w1_t); - make_unicode (w1, w2_t, w3_t); + u32x a = MD4M_A; + 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; - u32 b = MD4M_B; - u32 c = MD4M_C; - u32 d = MD4M_D; + MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13); + MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10); + MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11); + MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12); + MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13); - MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03); - MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00); - MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01); - MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02); - MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03); + MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20); + MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21); + MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22); + MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23); + MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20); - MD4_STEP (MD4_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); + if (MATCHES_NONE_VS (a, search[0])) continue; - 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); + 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); - const u32 r0 = a; - const u32 r1 = d; - const u32 r2 = c; - const u32 r3 = b; - - #include COMPARE_S + COMPARE_S_SIMD (a, d, c, b); } }