diff --git a/OpenCL/inc_hash_md4.cl b/OpenCL/inc_hash_md4.cl new file mode 100644 index 000000000..a376ead1b --- /dev/null +++ b/OpenCL/inc_hash_md4.cl @@ -0,0 +1,639 @@ + +// important notes on this: +// input buf unused bytes needs to be set to zero +// input buf need to be in algorithm native byte order (md5 = LE, sha1 = BE, etc) +// input buf need to be 64 byte aligned when usin md5_update() + +typedef struct md4_ctx_scalar +{ + u32 h[4]; + + u32 w0[4]; + u32 w1[4]; + u32 w2[4]; + u32 w3[4]; + + int len; + +} md4_ctx_scalar_t; + +void md4_transform_scalar (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]; + + MD4_STEP_S (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00); + MD4_STEP_S (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01); + MD4_STEP_S (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02); + MD4_STEP_S (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03); + MD4_STEP_S (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00); + MD4_STEP_S (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01); + MD4_STEP_S (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02); + MD4_STEP_S (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03); + MD4_STEP_S (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00); + MD4_STEP_S (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01); + MD4_STEP_S (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02); + MD4_STEP_S (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03); + MD4_STEP_S (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00); + MD4_STEP_S (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01); + MD4_STEP_S (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02); + MD4_STEP_S (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03); + + MD4_STEP_S (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10); + MD4_STEP_S (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11); + MD4_STEP_S (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12); + MD4_STEP_S (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13); + MD4_STEP_S (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10); + MD4_STEP_S (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11); + MD4_STEP_S (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12); + MD4_STEP_S (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13); + MD4_STEP_S (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10); + MD4_STEP_S (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11); + MD4_STEP_S (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12); + MD4_STEP_S (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13); + MD4_STEP_S (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10); + MD4_STEP_S (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11); + MD4_STEP_S (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12); + MD4_STEP_S (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13); + + MD4_STEP_S (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20); + MD4_STEP_S (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21); + MD4_STEP_S (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22); + MD4_STEP_S (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23); + MD4_STEP_S (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20); + MD4_STEP_S (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21); + MD4_STEP_S (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22); + MD4_STEP_S (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23); + MD4_STEP_S (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20); + MD4_STEP_S (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21); + MD4_STEP_S (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22); + MD4_STEP_S (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23); + MD4_STEP_S (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20); + MD4_STEP_S (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21); + MD4_STEP_S (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22); + MD4_STEP_S (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23); + + digest[0] += a; + digest[1] += b; + digest[2] += c; + digest[3] += d; +} + +void md4_init_scalar (md4_ctx_scalar_t *md4_ctx) +{ + md4_ctx->h[0] = MD4M_A; + md4_ctx->h[1] = MD4M_B; + md4_ctx->h[2] = MD4M_C; + md4_ctx->h[3] = MD4M_D; + + md4_ctx->w0[0] = 0; + md4_ctx->w0[1] = 0; + md4_ctx->w0[2] = 0; + md4_ctx->w0[3] = 0; + md4_ctx->w1[0] = 0; + md4_ctx->w1[1] = 0; + md4_ctx->w1[2] = 0; + md4_ctx->w1[3] = 0; + md4_ctx->w2[0] = 0; + md4_ctx->w2[1] = 0; + md4_ctx->w2[2] = 0; + md4_ctx->w2[3] = 0; + md4_ctx->w3[0] = 0; + md4_ctx->w3[1] = 0; + md4_ctx->w3[2] = 0; + md4_ctx->w3[3] = 0; + + md4_ctx->len = 0; +} + +void md4_update_scalar_64 (md4_ctx_scalar_t *md4_ctx, u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const int len) +{ + const int pos = md4_ctx->len & 0x3f; + + md4_ctx->len += len; + + if ((pos + len) < 64) + { + switch_buffer_by_offset_le_S (w0, w1, w2, w3, pos); + + md4_ctx->w0[0] |= w0[0]; + md4_ctx->w0[1] |= w0[1]; + md4_ctx->w0[2] |= w0[2]; + md4_ctx->w0[3] |= w0[3]; + md4_ctx->w1[0] |= w1[0]; + md4_ctx->w1[1] |= w1[1]; + md4_ctx->w1[2] |= w1[2]; + md4_ctx->w1[3] |= w1[3]; + md4_ctx->w2[0] |= w2[0]; + md4_ctx->w2[1] |= w2[1]; + md4_ctx->w2[2] |= w2[2]; + md4_ctx->w2[3] |= w2[3]; + md4_ctx->w3[0] |= w3[0]; + md4_ctx->w3[1] |= w3[1]; + md4_ctx->w3[2] |= w3[2]; + md4_ctx->w3[3] |= w3[3]; + } + else + { + u32 c0[4] = { 0 }; + u32 c1[4] = { 0 }; + u32 c2[4] = { 0 }; + u32 c3[4] = { 0 }; + + switch_buffer_by_offset_carry_le_S (w0, w1, w2, w3, c0, c1, c2, c3, pos); + + md4_ctx->w0[0] |= w0[0]; + md4_ctx->w0[1] |= w0[1]; + md4_ctx->w0[2] |= w0[2]; + md4_ctx->w0[3] |= w0[3]; + md4_ctx->w1[0] |= w1[0]; + md4_ctx->w1[1] |= w1[1]; + md4_ctx->w1[2] |= w1[2]; + md4_ctx->w1[3] |= w1[3]; + md4_ctx->w2[0] |= w2[0]; + md4_ctx->w2[1] |= w2[1]; + md4_ctx->w2[2] |= w2[2]; + md4_ctx->w2[3] |= w2[3]; + md4_ctx->w3[0] |= w3[0]; + md4_ctx->w3[1] |= w3[1]; + md4_ctx->w3[2] |= w3[2]; + md4_ctx->w3[3] |= w3[3]; + + md4_transform_scalar (md4_ctx->w0, md4_ctx->w1, md4_ctx->w2, md4_ctx->w3, md4_ctx->h); + + md4_ctx->w0[0] = c0[0]; + md4_ctx->w0[1] = c0[1]; + md4_ctx->w0[2] = c0[2]; + md4_ctx->w0[3] = c0[3]; + md4_ctx->w1[0] = c1[0]; + md4_ctx->w1[1] = c1[1]; + md4_ctx->w1[2] = c1[2]; + md4_ctx->w1[3] = c1[3]; + md4_ctx->w2[0] = c2[0]; + md4_ctx->w2[1] = c2[1]; + md4_ctx->w2[2] = c2[2]; + md4_ctx->w2[3] = c2[3]; + md4_ctx->w3[0] = c3[0]; + md4_ctx->w3[1] = c3[1]; + md4_ctx->w3[2] = c3[2]; + md4_ctx->w3[3] = c3[3]; + } +} + +void md4_update_scalar (md4_ctx_scalar_t *md4_ctx, const u32 *w, const int len) +{ + u32 w0[4]; + u32 w1[4]; + u32 w2[4]; + u32 w3[4]; + + int i; + int j; + + for (i = 0, j = 0; i < len - 64; i += 64, j += 16) + { + w0[0] = w[i + 0]; + w0[1] = w[i + 1]; + w0[2] = w[i + 2]; + w0[3] = w[i + 3]; + w1[0] = w[i + 4]; + w1[1] = w[i + 5]; + w1[2] = w[i + 6]; + w1[3] = w[i + 7]; + w2[0] = w[i + 8]; + w2[1] = w[i + 9]; + w2[2] = w[i + 10]; + w2[3] = w[i + 11]; + w3[0] = w[i + 12]; + w3[1] = w[i + 13]; + w3[2] = w[i + 14]; + w3[3] = w[i + 15]; + + md4_update_scalar_64 (md4_ctx, w0, w1, w2, w3, 64); + } + + w0[0] = w[i + 0]; + w0[1] = w[i + 1]; + w0[2] = w[i + 2]; + w0[3] = w[i + 3]; + w1[0] = w[i + 4]; + w1[1] = w[i + 5]; + w1[2] = w[i + 6]; + w1[3] = w[i + 7]; + w2[0] = w[i + 8]; + w2[1] = w[i + 9]; + w2[2] = w[i + 10]; + w2[3] = w[i + 11]; + w3[0] = w[i + 12]; + w3[1] = w[i + 13]; + w3[2] = w[i + 14]; + w3[3] = w[i + 15]; + + md4_update_scalar_64 (md4_ctx, w0, w1, w2, w3, len & 0x3f); +} + +void md4_update_scalar_global (md4_ctx_scalar_t *md4_ctx, const __global u32 *w, const int len) +{ + u32 w0[4]; + u32 w1[4]; + u32 w2[4]; + u32 w3[4]; + + int i; + int j; + + for (i = 0, j = 0; i < len - 64; i += 64, j += 16) + { + w0[0] = w[i + 0]; + w0[1] = w[i + 1]; + w0[2] = w[i + 2]; + w0[3] = w[i + 3]; + w1[0] = w[i + 4]; + w1[1] = w[i + 5]; + w1[2] = w[i + 6]; + w1[3] = w[i + 7]; + w2[0] = w[i + 8]; + w2[1] = w[i + 9]; + w2[2] = w[i + 10]; + w2[3] = w[i + 11]; + w3[0] = w[i + 12]; + w3[1] = w[i + 13]; + w3[2] = w[i + 14]; + w3[3] = w[i + 15]; + + md4_update_scalar_64 (md4_ctx, w0, w1, w2, w3, 64); + } + + w0[0] = w[i + 0]; + w0[1] = w[i + 1]; + w0[2] = w[i + 2]; + w0[3] = w[i + 3]; + w1[0] = w[i + 4]; + w1[1] = w[i + 5]; + w1[2] = w[i + 6]; + w1[3] = w[i + 7]; + w2[0] = w[i + 8]; + w2[1] = w[i + 9]; + w2[2] = w[i + 10]; + w2[3] = w[i + 11]; + w3[0] = w[i + 12]; + w3[1] = w[i + 13]; + w3[2] = w[i + 14]; + w3[3] = w[i + 15]; + + md4_update_scalar_64 (md4_ctx, w0, w1, w2, w3, len & 0x3f); +} + +void md4_update_scalar_global_utf16le (md4_ctx_scalar_t *md4_ctx, const __global u32 *w, const int len) +{ + u32 w0[4]; + u32 w1[4]; + u32 w2[4]; + u32 w3[4]; + + int i; + int j; + + for (i = 0, j = 0; i < len - 32; i += 32, j += 8) + { + w0[0] = w[i + 0]; + w0[1] = w[i + 1]; + w0[2] = w[i + 2]; + w0[3] = w[i + 3]; + w1[0] = w[i + 4]; + w1[1] = w[i + 5]; + w1[2] = w[i + 6]; + w1[3] = w[i + 7]; + + make_utf16le_S (w1, w2, w3); + make_utf16le_S (w0, w0, w1); + + md4_update_scalar_64 (md4_ctx, w0, w1, w2, w3, 64); + } + + w0[0] = w[i + 0]; + w0[1] = w[i + 1]; + w0[2] = w[i + 2]; + w0[3] = w[i + 3]; + w1[0] = w[i + 4]; + w1[1] = w[i + 5]; + w1[2] = w[i + 6]; + w1[3] = w[i + 7]; + + make_utf16le_S (w1, w2, w3); + make_utf16le_S (w0, w0, w1); + + md4_update_scalar_64 (md4_ctx, w0, w1, w2, w3, (len * 2) & 0x3f); +} + +void md4_final_scalar (md4_ctx_scalar_t *md4_ctx) +{ + const int pos = md4_ctx->len & 0x3f; + + append_0x80_4x4_S (md4_ctx->w0, md4_ctx->w1, md4_ctx->w2, md4_ctx->w3, pos); + + if (pos >= 56) + { + md4_transform_scalar (md4_ctx->w0, md4_ctx->w1, md4_ctx->w2, md4_ctx->w3, md4_ctx->h); + + md4_ctx->w0[0] = 0; + md4_ctx->w0[1] = 0; + md4_ctx->w0[2] = 0; + md4_ctx->w0[3] = 0; + md4_ctx->w1[0] = 0; + md4_ctx->w1[1] = 0; + md4_ctx->w1[2] = 0; + md4_ctx->w1[3] = 0; + md4_ctx->w2[0] = 0; + md4_ctx->w2[1] = 0; + md4_ctx->w2[2] = 0; + md4_ctx->w2[3] = 0; + md4_ctx->w3[0] = 0; + md4_ctx->w3[1] = 0; + md4_ctx->w3[2] = 0; + md4_ctx->w3[3] = 0; + } + + md4_ctx->w3[2] = md4_ctx->len * 8; + + md4_transform_scalar (md4_ctx->w0, md4_ctx->w1, md4_ctx->w2, md4_ctx->w3, md4_ctx->h); +} + +void md4_optimize_max_length_scalar (md4_ctx_scalar_t *md4_ctx, const int bits) +{ + md4_ctx->len &= (1 << bits) - 1; +} + +// while input buf can be a vector datatype, the length of the different elements can not + +typedef struct md4_ctx_vector +{ + u32x h[4]; + + u32x w0[4]; + u32x w1[4]; + u32x w2[4]; + u32x w3[4]; + + int len; + +} md4_ctx_vector_t; + +void md4_transform_vector (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4]) +{ + u32x a = digest[0]; + u32x b = digest[1]; + u32x c = digest[2]; + u32x d = digest[3]; + + 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[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[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); + + digest[0] += a; + digest[1] += b; + digest[2] += c; + digest[3] += d; +} + +void md4_init_vector (md4_ctx_vector_t *md4_ctx) +{ + md4_ctx->h[0] = MD4M_A; + md4_ctx->h[1] = MD4M_B; + md4_ctx->h[2] = MD4M_C; + md4_ctx->h[3] = MD4M_D; + + md4_ctx->w0[0] = 0; + md4_ctx->w0[1] = 0; + md4_ctx->w0[2] = 0; + md4_ctx->w0[3] = 0; + md4_ctx->w1[0] = 0; + md4_ctx->w1[1] = 0; + md4_ctx->w1[2] = 0; + md4_ctx->w1[3] = 0; + md4_ctx->w2[0] = 0; + md4_ctx->w2[1] = 0; + md4_ctx->w2[2] = 0; + md4_ctx->w2[3] = 0; + md4_ctx->w3[0] = 0; + md4_ctx->w3[1] = 0; + md4_ctx->w3[2] = 0; + md4_ctx->w3[3] = 0; + + md4_ctx->len = 0; +} + +void md4_update_vector_64 (md4_ctx_vector_t *md4_ctx, u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], const int len) +{ + const int pos = md4_ctx->len & 0x3f; + + md4_ctx->len += len; + + if ((pos + len) < 64) + { + switch_buffer_by_offset_le (w0, w1, w2, w3, pos); + + md4_ctx->w0[0] |= w0[0]; + md4_ctx->w0[1] |= w0[1]; + md4_ctx->w0[2] |= w0[2]; + md4_ctx->w0[3] |= w0[3]; + md4_ctx->w1[0] |= w1[0]; + md4_ctx->w1[1] |= w1[1]; + md4_ctx->w1[2] |= w1[2]; + md4_ctx->w1[3] |= w1[3]; + md4_ctx->w2[0] |= w2[0]; + md4_ctx->w2[1] |= w2[1]; + md4_ctx->w2[2] |= w2[2]; + md4_ctx->w2[3] |= w2[3]; + md4_ctx->w3[0] |= w3[0]; + md4_ctx->w3[1] |= w3[1]; + md4_ctx->w3[2] |= w3[2]; + md4_ctx->w3[3] |= w3[3]; + } + else + { + u32x c0[4] = { 0 }; + u32x c1[4] = { 0 }; + u32x c2[4] = { 0 }; + u32x c3[4] = { 0 }; + + switch_buffer_by_offset_carry_le (w0, w1, w2, w3, c0, c1, c2, c3, pos); + + md4_ctx->w0[0] |= w0[0]; + md4_ctx->w0[1] |= w0[1]; + md4_ctx->w0[2] |= w0[2]; + md4_ctx->w0[3] |= w0[3]; + md4_ctx->w1[0] |= w1[0]; + md4_ctx->w1[1] |= w1[1]; + md4_ctx->w1[2] |= w1[2]; + md4_ctx->w1[3] |= w1[3]; + md4_ctx->w2[0] |= w2[0]; + md4_ctx->w2[1] |= w2[1]; + md4_ctx->w2[2] |= w2[2]; + md4_ctx->w2[3] |= w2[3]; + md4_ctx->w3[0] |= w3[0]; + md4_ctx->w3[1] |= w3[1]; + md4_ctx->w3[2] |= w3[2]; + md4_ctx->w3[3] |= w3[3]; + + md4_transform_vector (md4_ctx->w0, md4_ctx->w1, md4_ctx->w2, md4_ctx->w3, md4_ctx->h); + + md4_ctx->w0[0] = c0[0]; + md4_ctx->w0[1] = c0[1]; + md4_ctx->w0[2] = c0[2]; + md4_ctx->w0[3] = c0[3]; + md4_ctx->w1[0] = c1[0]; + md4_ctx->w1[1] = c1[1]; + md4_ctx->w1[2] = c1[2]; + md4_ctx->w1[3] = c1[3]; + md4_ctx->w2[0] = c2[0]; + md4_ctx->w2[1] = c2[1]; + md4_ctx->w2[2] = c2[2]; + md4_ctx->w2[3] = c2[3]; + md4_ctx->w3[0] = c3[0]; + md4_ctx->w3[1] = c3[1]; + md4_ctx->w3[2] = c3[2]; + md4_ctx->w3[3] = c3[3]; + } +} + +void md4_update_vector (md4_ctx_vector_t *md4_ctx, const u32x *w, const int len) +{ + u32x w0[4]; + u32x w1[4]; + u32x w2[4]; + u32x w3[4]; + + int i; + int j; + + for (i = 0, j = 0; i < len - 64; i += 64, j += 16) + { + w0[0] = w[i + 0]; + w0[1] = w[i + 1]; + w0[2] = w[i + 2]; + w0[3] = w[i + 3]; + w1[0] = w[i + 4]; + w1[1] = w[i + 5]; + w1[2] = w[i + 6]; + w1[3] = w[i + 7]; + w2[0] = w[i + 8]; + w2[1] = w[i + 9]; + w2[2] = w[i + 10]; + w2[3] = w[i + 11]; + w3[0] = w[i + 12]; + w3[1] = w[i + 13]; + w3[2] = w[i + 14]; + w3[3] = w[i + 15]; + + md4_update_vector_64 (md4_ctx, w0, w1, w2, w3, 64); + } + + w0[0] = w[i + 0]; + w0[1] = w[i + 1]; + w0[2] = w[i + 2]; + w0[3] = w[i + 3]; + w1[0] = w[i + 4]; + w1[1] = w[i + 5]; + w1[2] = w[i + 6]; + w1[3] = w[i + 7]; + w2[0] = w[i + 8]; + w2[1] = w[i + 9]; + w2[2] = w[i + 10]; + w2[3] = w[i + 11]; + w3[0] = w[i + 12]; + w3[1] = w[i + 13]; + w3[2] = w[i + 14]; + w3[3] = w[i + 15]; + + md4_update_vector_64 (md4_ctx, w0, w1, w2, w3, len & 0x3f); +} + +void md4_final_vector (md4_ctx_vector_t *md4_ctx) +{ + const int pos = md4_ctx->len & 0x3f; + + append_0x80_4x4 (md4_ctx->w0, md4_ctx->w1, md4_ctx->w2, md4_ctx->w3, pos); + + if (pos >= 56) + { + md4_transform_vector (md4_ctx->w0, md4_ctx->w1, md4_ctx->w2, md4_ctx->w3, md4_ctx->h); + + md4_ctx->w0[0] = 0; + md4_ctx->w0[1] = 0; + md4_ctx->w0[2] = 0; + md4_ctx->w0[3] = 0; + md4_ctx->w1[0] = 0; + md4_ctx->w1[1] = 0; + md4_ctx->w1[2] = 0; + md4_ctx->w1[3] = 0; + md4_ctx->w2[0] = 0; + md4_ctx->w2[1] = 0; + md4_ctx->w2[2] = 0; + md4_ctx->w2[3] = 0; + md4_ctx->w3[0] = 0; + md4_ctx->w3[1] = 0; + md4_ctx->w3[2] = 0; + md4_ctx->w3[3] = 0; + } + + md4_ctx->w3[2] = md4_ctx->len * 8; + + md4_transform_vector (md4_ctx->w0, md4_ctx->w1, md4_ctx->w2, md4_ctx->w3, md4_ctx->h); +} + +void md4_optimize_max_length_vector (md4_ctx_vector_t *md4_ctx, const int bits) +{ + md4_ctx->len &= (1 << bits) - 1; +} + + diff --git a/OpenCL/m02100.cl b/OpenCL/m02100.cl index 0249709f8..df717f27e 100644 --- a/OpenCL/m02100.cl +++ b/OpenCL/m02100.cl @@ -11,6 +11,7 @@ #include "inc_types.cl" #include "inc_common.cl" #include "inc_simd.cl" +#include "inc_hash_md4.cl" #define COMPARE_S "inc_comp_single.cl" #define COMPARE_M "inc_comp_multi.cl" @@ -470,113 +471,51 @@ __kernel void m02100_init (__global pw_t *pws, __global const kernel_rule_t *rul if (gid >= gid_max) return; - u32 w0[4]; + md4_ctx_scalar_t md4_ctx1; - w0[0] = pws[gid].i[ 0]; - w0[1] = pws[gid].i[ 1]; - w0[2] = pws[gid].i[ 2]; - w0[3] = pws[gid].i[ 3]; + md4_init_scalar (&md4_ctx1); - u32 w1[4]; + md4_update_scalar_global_utf16le (&md4_ctx1, pws[gid].i, pws[gid].pw_len); - w1[0] = pws[gid].i[ 4]; - w1[1] = pws[gid].i[ 5]; - w1[2] = pws[gid].i[ 6]; - w1[3] = pws[gid].i[ 7]; - - u32 w2[4]; - - 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[3] = 0; - - const u32 pw_len = pws[gid].pw_len; + md4_final_scalar (&md4_ctx1); /** * salt */ - u32 salt_len = salt_bufs[salt_pos].salt_len; + md4_ctx_scalar_t md4_ctx2; - u32 salt_buf0[4]; - u32 salt_buf1[4]; - u32 salt_buf2[4]; + md4_init_scalar (&md4_ctx2); - salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0]; - salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1]; - salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2]; - salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3]; - salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4]; - salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5]; - salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6]; - salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7]; - salt_buf2[0] = salt_bufs[salt_pos].salt_buf[8]; - salt_buf2[1] = salt_bufs[salt_pos].salt_buf[9]; - salt_buf2[2] = 0; - salt_buf2[3] = 0; + md4_ctx2.w0[0] = md4_ctx1.h[0]; + md4_ctx2.w0[1] = md4_ctx1.h[1]; + md4_ctx2.w0[2] = md4_ctx1.h[2]; + md4_ctx2.w0[3] = md4_ctx1.h[3]; - /** - * generate dcc - */ + md4_ctx2.len = 16; - append_0x80_4x4_S (w0, w1, w2, w3, pw_len); + md4_update_scalar_global (&md4_ctx2, salt_bufs[salt_pos].salt_buf, salt_bufs[salt_pos].salt_len); - make_utf16le_S (w1, w2, w3); - make_utf16le_S (w0, w0, w1); + md4_final_scalar (&md4_ctx2); - w3[2] = pw_len * 2 * 8; - - u32 digest_md4[4]; - - digest_md4[0] = MD4M_A; - digest_md4[1] = MD4M_B; - digest_md4[2] = MD4M_C; - digest_md4[3] = MD4M_D; - - md4_transform_S (w0, w1, w2, w3, digest_md4); - - w0[0] = digest_md4[0]; - w0[1] = digest_md4[1]; - w0[2] = digest_md4[2]; - w0[3] = digest_md4[3]; - w1[0] = salt_buf0[0]; - w1[1] = salt_buf0[1]; - w1[2] = salt_buf0[2]; - w1[3] = salt_buf0[3]; - w2[0] = salt_buf1[0]; - w2[1] = salt_buf1[1]; - w2[2] = salt_buf1[2]; - w2[3] = salt_buf1[3]; - w3[0] = salt_buf2[0]; - w3[1] = salt_buf2[1]; - w3[2] = (16 + salt_len) * 8; - w3[3] = 0; - - append_0x80_4x4_S (w0, w1, w2, w3, 16 + salt_len); - - digest_md4[0] = MD4M_A; - digest_md4[1] = MD4M_B; - digest_md4[2] = MD4M_C; - digest_md4[3] = MD4M_D; - - md4_transform_S (w0, w1, w2, w3, digest_md4); + md4_ctx2.h[0] = swap32_S (md4_ctx2.h[0]); + md4_ctx2.h[1] = swap32_S (md4_ctx2.h[1]); + md4_ctx2.h[2] = swap32_S (md4_ctx2.h[2]); + md4_ctx2.h[3] = swap32_S (md4_ctx2.h[3]); /** * pads */ - w0[0] = swap32_S (digest_md4[0]); - w0[1] = swap32_S (digest_md4[1]); - w0[2] = swap32_S (digest_md4[2]); - w0[3] = swap32_S (digest_md4[3]); + u32 w0[4]; + u32 w1[4]; + u32 w2[4]; + u32 w3[4]; + + w0[0] = md4_ctx2.h[0]; + w0[1] = md4_ctx2.h[1]; + w0[2] = md4_ctx2.h[2]; + w0[3] = md4_ctx2.h[3]; w1[0] = 0; w1[1] = 0; w1[2] = 0; @@ -611,19 +550,21 @@ __kernel void m02100_init (__global pw_t *pws, __global const kernel_rule_t *rul * hmac1 */ - w0[0] = salt_buf0[0]; - w0[1] = salt_buf0[1]; - w0[2] = salt_buf0[2]; - w0[3] = salt_buf0[3]; - w1[0] = salt_buf1[0]; - w1[1] = salt_buf1[1]; - w1[2] = salt_buf1[2]; - w1[3] = salt_buf1[3]; - w2[0] = salt_buf2[0]; - w2[1] = salt_buf2[1]; - w2[2] = 0; - w2[3] = 0; - w3[0] = 0; + const u32 salt_len = salt_bufs[salt_pos].salt_len; + + w0[0] = salt_bufs[salt_pos].salt_buf[ 0]; + w0[1] = salt_bufs[salt_pos].salt_buf[ 1]; + w0[2] = salt_bufs[salt_pos].salt_buf[ 2]; + w0[3] = salt_bufs[salt_pos].salt_buf[ 3]; + w1[0] = salt_bufs[salt_pos].salt_buf[ 4]; + w1[1] = salt_bufs[salt_pos].salt_buf[ 5]; + w1[2] = salt_bufs[salt_pos].salt_buf[ 6]; + w1[3] = salt_bufs[salt_pos].salt_buf[ 7]; + w2[0] = salt_bufs[salt_pos].salt_buf[ 8]; + w2[1] = salt_bufs[salt_pos].salt_buf[ 9]; + w2[2] = salt_bufs[salt_pos].salt_buf[10]; + w2[3] = salt_bufs[salt_pos].salt_buf[11]; + w3[0] = salt_bufs[salt_pos].salt_buf[12]; w3[1] = 0; w3[2] = 0; w3[3] = (64 + salt_len + 4) * 8; diff --git a/include/types.h b/include/types.h index c67712ee7..610b91550 100644 --- a/include/types.h +++ b/include/types.h @@ -326,25 +326,27 @@ typedef enum salt_type typedef enum opti_type { - OPTI_TYPE_ZERO_BYTE = (1 << 1), - OPTI_TYPE_PRECOMPUTE_INIT = (1 << 2), - OPTI_TYPE_PRECOMPUTE_MERKLE = (1 << 3), - OPTI_TYPE_PRECOMPUTE_PERMUT = (1 << 4), - OPTI_TYPE_MEET_IN_MIDDLE = (1 << 5), - OPTI_TYPE_EARLY_SKIP = (1 << 6), - OPTI_TYPE_NOT_SALTED = (1 << 7), - OPTI_TYPE_NOT_ITERATED = (1 << 8), - OPTI_TYPE_PREPENDED_SALT = (1 << 9), - OPTI_TYPE_APPENDED_SALT = (1 << 10), - OPTI_TYPE_SINGLE_HASH = (1 << 11), - OPTI_TYPE_SINGLE_SALT = (1 << 12), - OPTI_TYPE_BRUTE_FORCE = (1 << 13), - OPTI_TYPE_RAW_HASH = (1 << 14), - OPTI_TYPE_SLOW_HASH_SIMD = (1 << 15), - OPTI_TYPE_USES_BITS_8 = (1 << 16), - OPTI_TYPE_USES_BITS_16 = (1 << 17), - OPTI_TYPE_USES_BITS_32 = (1 << 18), - OPTI_TYPE_USES_BITS_64 = (1 << 19) + OPTI_TYPE_ZERO_BYTE = (1 << 1), + OPTI_TYPE_PRECOMPUTE_INIT = (1 << 2), + OPTI_TYPE_PRECOMPUTE_MERKLE = (1 << 3), + OPTI_TYPE_PRECOMPUTE_PERMUT = (1 << 4), + OPTI_TYPE_MEET_IN_MIDDLE = (1 << 5), + OPTI_TYPE_EARLY_SKIP = (1 << 6), + OPTI_TYPE_NOT_SALTED = (1 << 7), + OPTI_TYPE_NOT_ITERATED = (1 << 8), + OPTI_TYPE_PREPENDED_SALT = (1 << 9), + OPTI_TYPE_APPENDED_SALT = (1 << 10), + OPTI_TYPE_SINGLE_HASH = (1 << 11), + OPTI_TYPE_SINGLE_SALT = (1 << 12), + OPTI_TYPE_BRUTE_FORCE = (1 << 13), + OPTI_TYPE_RAW_HASH = (1 << 14), + OPTI_TYPE_SLOW_HASH_SIMD_INIT = (1 << 15), + OPTI_TYPE_SLOW_HASH_SIMD_LOOP = (1 << 16), + OPTI_TYPE_SLOW_HASH_SIMD_COMP = (1 << 17), + OPTI_TYPE_USES_BITS_8 = (1 << 18), + OPTI_TYPE_USES_BITS_16 = (1 << 19), + OPTI_TYPE_USES_BITS_32 = (1 << 20), + OPTI_TYPE_USES_BITS_64 = (1 << 21) } opti_type_t; diff --git a/src/interface.c b/src/interface.c index ee2744092..c63b20d90 100644 --- a/src/interface.c +++ b/src/interface.c @@ -262,25 +262,27 @@ static char ST_HASH_15600[] = "$ethereum$p*1024*38353131353831333338313138363430 static char ST_HASH_15700[] = "$ethereum$s*1024*1*1*3033363133373132373638333437323331383637383437333631373038323434*69eaf081695cf971ef7ee5a49997c1a3922e7efef59068109e83853755ee31c3*64a1adec1750ee4416b22b81111dd2a3c2fede820d6da8bf788dca2641d5b181"; static char ST_HASH_99999[] = "hashcat"; -static const char OPTI_STR_ZERO_BYTE[] = "Zero-Byte"; -static const char OPTI_STR_PRECOMPUTE_INIT[] = "Precompute-Init"; -static const char OPTI_STR_PRECOMPUTE_MERKLE[] = "Precompute-Merkle-Demgard"; -static const char OPTI_STR_PRECOMPUTE_PERMUT[] = "Precompute-Final-Permutation"; -static const char OPTI_STR_MEET_IN_MIDDLE[] = "Meet-In-The-Middle"; -static const char OPTI_STR_EARLY_SKIP[] = "Early-Skip"; -static const char OPTI_STR_NOT_SALTED[] = "Not-Salted"; -static const char OPTI_STR_NOT_ITERATED[] = "Not-Iterated"; -static const char OPTI_STR_PREPENDED_SALT[] = "Prepended-Salt"; -static const char OPTI_STR_APPENDED_SALT[] = "Appended-Salt"; -static const char OPTI_STR_SINGLE_HASH[] = "Single-Hash"; -static const char OPTI_STR_SINGLE_SALT[] = "Single-Salt"; -static const char OPTI_STR_BRUTE_FORCE[] = "Brute-Force"; -static const char OPTI_STR_RAW_HASH[] = "Raw-Hash"; -static const char OPTI_STR_SLOW_HASH_SIMD[] = "Slow-Hash-SIMD"; -static const char OPTI_STR_USES_BITS_8[] = "Uses-8-Bit"; -static const char OPTI_STR_USES_BITS_16[] = "Uses-16-Bit"; -static const char OPTI_STR_USES_BITS_32[] = "Uses-32-Bit"; -static const char OPTI_STR_USES_BITS_64[] = "Uses-64-Bit"; +static const char OPTI_STR_ZERO_BYTE[] = "Zero-Byte"; +static const char OPTI_STR_PRECOMPUTE_INIT[] = "Precompute-Init"; +static const char OPTI_STR_PRECOMPUTE_MERKLE[] = "Precompute-Merkle-Demgard"; +static const char OPTI_STR_PRECOMPUTE_PERMUT[] = "Precompute-Final-Permutation"; +static const char OPTI_STR_MEET_IN_MIDDLE[] = "Meet-In-The-Middle"; +static const char OPTI_STR_EARLY_SKIP[] = "Early-Skip"; +static const char OPTI_STR_NOT_SALTED[] = "Not-Salted"; +static const char OPTI_STR_NOT_ITERATED[] = "Not-Iterated"; +static const char OPTI_STR_PREPENDED_SALT[] = "Prepended-Salt"; +static const char OPTI_STR_APPENDED_SALT[] = "Appended-Salt"; +static const char OPTI_STR_SINGLE_HASH[] = "Single-Hash"; +static const char OPTI_STR_SINGLE_SALT[] = "Single-Salt"; +static const char OPTI_STR_BRUTE_FORCE[] = "Brute-Force"; +static const char OPTI_STR_RAW_HASH[] = "Raw-Hash"; +static const char OPTI_STR_SLOW_HASH_SIMD_INIT[] = "Slow-Hash-SIMD-INIT"; +static const char OPTI_STR_SLOW_HASH_SIMD_LOOP[] = "Slow-Hash-SIMD-LOOP"; +static const char OPTI_STR_SLOW_HASH_SIMD_COMP[] = "Slow-Hash-SIMD-COMP"; +static const char OPTI_STR_USES_BITS_8[] = "Uses-8-Bit"; +static const char OPTI_STR_USES_BITS_16[] = "Uses-16-Bit"; +static const char OPTI_STR_USES_BITS_32[] = "Uses-32-Bit"; +static const char OPTI_STR_USES_BITS_64[] = "Uses-64-Bit"; static const char PA_000[] = "OK"; static const char PA_001[] = "Ignored due to comment"; @@ -15671,25 +15673,27 @@ char *stroptitype (const u32 opti_type) { switch (opti_type) { - case OPTI_TYPE_ZERO_BYTE: return ((char *) OPTI_STR_ZERO_BYTE); - case OPTI_TYPE_PRECOMPUTE_INIT: return ((char *) OPTI_STR_PRECOMPUTE_INIT); - case OPTI_TYPE_PRECOMPUTE_MERKLE: return ((char *) OPTI_STR_PRECOMPUTE_MERKLE); - case OPTI_TYPE_PRECOMPUTE_PERMUT: return ((char *) OPTI_STR_PRECOMPUTE_PERMUT); - case OPTI_TYPE_MEET_IN_MIDDLE: return ((char *) OPTI_STR_MEET_IN_MIDDLE); - case OPTI_TYPE_EARLY_SKIP: return ((char *) OPTI_STR_EARLY_SKIP); - case OPTI_TYPE_NOT_SALTED: return ((char *) OPTI_STR_NOT_SALTED); - case OPTI_TYPE_NOT_ITERATED: return ((char *) OPTI_STR_NOT_ITERATED); - case OPTI_TYPE_PREPENDED_SALT: return ((char *) OPTI_STR_PREPENDED_SALT); - case OPTI_TYPE_APPENDED_SALT: return ((char *) OPTI_STR_APPENDED_SALT); - case OPTI_TYPE_SINGLE_HASH: return ((char *) OPTI_STR_SINGLE_HASH); - case OPTI_TYPE_SINGLE_SALT: return ((char *) OPTI_STR_SINGLE_SALT); - case OPTI_TYPE_BRUTE_FORCE: return ((char *) OPTI_STR_BRUTE_FORCE); - case OPTI_TYPE_RAW_HASH: return ((char *) OPTI_STR_RAW_HASH); - case OPTI_TYPE_SLOW_HASH_SIMD: return ((char *) OPTI_STR_SLOW_HASH_SIMD); - case OPTI_TYPE_USES_BITS_8: return ((char *) OPTI_STR_USES_BITS_8); - case OPTI_TYPE_USES_BITS_16: return ((char *) OPTI_STR_USES_BITS_16); - case OPTI_TYPE_USES_BITS_32: return ((char *) OPTI_STR_USES_BITS_32); - case OPTI_TYPE_USES_BITS_64: return ((char *) OPTI_STR_USES_BITS_64); + case OPTI_TYPE_ZERO_BYTE: return ((char *) OPTI_STR_ZERO_BYTE); + case OPTI_TYPE_PRECOMPUTE_INIT: return ((char *) OPTI_STR_PRECOMPUTE_INIT); + case OPTI_TYPE_PRECOMPUTE_MERKLE: return ((char *) OPTI_STR_PRECOMPUTE_MERKLE); + case OPTI_TYPE_PRECOMPUTE_PERMUT: return ((char *) OPTI_STR_PRECOMPUTE_PERMUT); + case OPTI_TYPE_MEET_IN_MIDDLE: return ((char *) OPTI_STR_MEET_IN_MIDDLE); + case OPTI_TYPE_EARLY_SKIP: return ((char *) OPTI_STR_EARLY_SKIP); + case OPTI_TYPE_NOT_SALTED: return ((char *) OPTI_STR_NOT_SALTED); + case OPTI_TYPE_NOT_ITERATED: return ((char *) OPTI_STR_NOT_ITERATED); + case OPTI_TYPE_PREPENDED_SALT: return ((char *) OPTI_STR_PREPENDED_SALT); + case OPTI_TYPE_APPENDED_SALT: return ((char *) OPTI_STR_APPENDED_SALT); + case OPTI_TYPE_SINGLE_HASH: return ((char *) OPTI_STR_SINGLE_HASH); + case OPTI_TYPE_SINGLE_SALT: return ((char *) OPTI_STR_SINGLE_SALT); + case OPTI_TYPE_BRUTE_FORCE: return ((char *) OPTI_STR_BRUTE_FORCE); + case OPTI_TYPE_RAW_HASH: return ((char *) OPTI_STR_RAW_HASH); + case OPTI_TYPE_SLOW_HASH_SIMD_INIT: return ((char *) OPTI_STR_SLOW_HASH_SIMD_INIT); + case OPTI_TYPE_SLOW_HASH_SIMD_LOOP: return ((char *) OPTI_STR_SLOW_HASH_SIMD_LOOP); + case OPTI_TYPE_SLOW_HASH_SIMD_COMP: return ((char *) OPTI_STR_SLOW_HASH_SIMD_COMP); + case OPTI_TYPE_USES_BITS_8: return ((char *) OPTI_STR_USES_BITS_8); + case OPTI_TYPE_USES_BITS_16: return ((char *) OPTI_STR_USES_BITS_16); + case OPTI_TYPE_USES_BITS_32: return ((char *) OPTI_STR_USES_BITS_32); + case OPTI_TYPE_USES_BITS_64: return ((char *) OPTI_STR_USES_BITS_64); } return (NULL); @@ -20332,7 +20336,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_4; hashconfig->parse_func = phpass_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -21047,7 +21051,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_4; hashconfig->parse_func = dcc2_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -21106,7 +21110,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_4; hashconfig->parse_func = wpa_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -22156,7 +22160,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->parse_func = sha512osx_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE | OPTI_TYPE_USES_BITS_64 - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -22174,7 +22178,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->parse_func = sha512grub_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE | OPTI_TYPE_USES_BITS_64 - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -22520,7 +22524,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_32; hashconfig->parse_func = cisco8_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -22740,7 +22744,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_32; hashconfig->parse_func = djangopbkdf2_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -22948,7 +22952,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_32; hashconfig->parse_func = pbkdf2_sha256_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -23131,7 +23135,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_32; hashconfig->parse_func = pbkdf2_md5_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -23150,7 +23154,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_32; hashconfig->parse_func = pbkdf2_sha1_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -23168,7 +23172,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_32; hashconfig->parse_func = atlassian_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -23188,7 +23192,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->parse_func = pbkdf2_sha512_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE | OPTI_TYPE_USES_BITS_64 - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -23892,7 +23896,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_4; // we actually do not have a digest hashconfig->parse_func = itunes_backup_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -23913,7 +23917,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_4; // we actually do not have a digest hashconfig->parse_func = itunes_backup_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; @@ -24060,7 +24064,7 @@ int hashconfig_init (hashcat_ctx_t *hashcat_ctx) hashconfig->dgst_size = DGST_SIZE_4_8; hashconfig->parse_func = ethereum_pbkdf2_parse_hash; hashconfig->opti_type = OPTI_TYPE_ZERO_BYTE - | OPTI_TYPE_SLOW_HASH_SIMD; + | OPTI_TYPE_SLOW_HASH_SIMD_LOOP; hashconfig->dgst_pos0 = 0; hashconfig->dgst_pos1 = 1; hashconfig->dgst_pos2 = 2; diff --git a/src/opencl.c b/src/opencl.c index ac417fe54..8ef90359f 100644 --- a/src/opencl.c +++ b/src/opencl.c @@ -1402,9 +1402,23 @@ int run_kernel (hashcat_ctx_t *hashcat_ctx, hc_device_param_t *device_param, con } else { - if (kern_run == KERN_RUN_2) + if (kern_run == KERN_RUN_1) { - if (hashconfig->opti_type & OPTI_TYPE_SLOW_HASH_SIMD) + if (hashconfig->opti_type & OPTI_TYPE_SLOW_HASH_SIMD_INIT) + { + num_elements = CEILDIV (num_elements, device_param->vector_width); + } + } + else if (kern_run == KERN_RUN_2) + { + if (hashconfig->opti_type & OPTI_TYPE_SLOW_HASH_SIMD_LOOP) + { + num_elements = CEILDIV (num_elements, device_param->vector_width); + } + } + else if (kern_run == KERN_RUN_3) + { + if (hashconfig->opti_type & OPTI_TYPE_SLOW_HASH_SIMD_COMP) { num_elements = CEILDIV (num_elements, device_param->vector_width); }