diff --git a/OpenCL/m10420_a3.cl b/OpenCL/m10420_a3.cl index f6c11b152..4457c1ae4 100644 --- a/OpenCL/m10420_a3.cl +++ b/OpenCL/m10420_a3.cl @@ -26,6 +26,104 @@ __constant u32 padding[8] = 0x7a695364 }; +void md5_transform_S (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4]) +{ + u32 a = digest[0]; + u32 b = digest[1]; + u32 c = digest[2]; + u32 d = digest[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_S (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00); + MD5_STEP_S (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01); + MD5_STEP_S (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02); + MD5_STEP_S (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03); + MD5_STEP_S (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00); + MD5_STEP_S (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01); + MD5_STEP_S (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02); + MD5_STEP_S (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03); + MD5_STEP_S (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00); + MD5_STEP_S (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01); + MD5_STEP_S (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02); + MD5_STEP_S (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03); + MD5_STEP_S (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00); + MD5_STEP_S (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01); + MD5_STEP_S (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02); + MD5_STEP_S (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03); + + MD5_STEP_S (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10); + MD5_STEP_S (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11); + MD5_STEP_S (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12); + MD5_STEP_S (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13); + MD5_STEP_S (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10); + MD5_STEP_S (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11); + MD5_STEP_S (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12); + MD5_STEP_S (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13); + MD5_STEP_S (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10); + MD5_STEP_S (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11); + MD5_STEP_S (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12); + MD5_STEP_S (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13); + MD5_STEP_S (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10); + MD5_STEP_S (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11); + MD5_STEP_S (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12); + MD5_STEP_S (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13); + + MD5_STEP_S (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20); + MD5_STEP_S (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21); + MD5_STEP_S (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22); + MD5_STEP_S (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23); + MD5_STEP_S (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20); + MD5_STEP_S (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21); + MD5_STEP_S (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22); + MD5_STEP_S (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23); + MD5_STEP_S (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20); + MD5_STEP_S (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21); + MD5_STEP_S (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22); + MD5_STEP_S (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23); + MD5_STEP_S (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20); + MD5_STEP_S (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21); + MD5_STEP_S (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22); + MD5_STEP_S (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23); + + MD5_STEP_S (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30); + MD5_STEP_S (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31); + MD5_STEP_S (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32); + MD5_STEP_S (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33); + MD5_STEP_S (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30); + MD5_STEP_S (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31); + MD5_STEP_S (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32); + MD5_STEP_S (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33); + MD5_STEP_S (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30); + MD5_STEP_S (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31); + MD5_STEP_S (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32); + MD5_STEP_S (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33); + MD5_STEP_S (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30); + MD5_STEP_S (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31); + MD5_STEP_S (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32); + MD5_STEP_S (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33); + + digest[0] += a; + digest[1] += b; + digest[2] += c; + digest[3] += d; +} + void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4]) { u32x a = digest[0]; @@ -179,7 +277,7 @@ void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl p3[2] = 0; p3[3] = 0; - switch_buffer_by_offset_le (p0, p1, p2, p3, pw_len); + switch_buffer_by_offset_le_S (p0, p1, p2, p3, pw_len); w0[0] |= p0[0]; w0[1] |= p0[1]; @@ -206,7 +304,7 @@ void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) { - const u32 w0r = ix_create_bft (bfs_buf, il_pos); + const u32 w0r = (u32) bfs_buf[il_pos + 0].i; w0[0] = w0l | w0r; @@ -247,7 +345,7 @@ void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl digest[2] = MD5M_C; digest[3] = MD5M_D; - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + md5_transform_S (w0_t, w1_t, w2_t, w3_t, digest); w0_t[0] = P; w0_t[1] = id_buf[0]; @@ -266,7 +364,7 @@ void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl w3_t[2] = 84 * 8; w3_t[3] = 0; - md5_transform (w0_t, w1_t, w2_t, w3_t, digest); + md5_transform_S (w0_t, w1_t, w2_t, w3_t, digest); u32x a = digest[0]; u32x b = digest[1] & 0xff; @@ -332,7 +430,7 @@ void m10420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl p3[2] = 0; p3[3] = 0; - switch_buffer_by_offset_le (p0, p1, p2, p3, pw_len); + switch_buffer_by_offset_le_S (p0, p1, p2, p3, pw_len); w0[0] |= p0[0]; w0[1] |= p0[1]; @@ -371,7 +469,7 @@ void m10420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __gl for (u32 il_pos = 0; il_pos < il_cnt; il_pos++) { - const u32 w0r = ix_create_bft (bfs_buf, il_pos); + const u32 w0r = (u32) bfs_buf[il_pos].i; w0[0] = w0l | w0r;