diff --git a/OpenCL/m07300_a0-optimized.cl b/OpenCL/m07300_a0-optimized.cl index 35f4a3639..2deb27bac 100644 --- a/OpenCL/m07300_a0-optimized.cl +++ b/OpenCL/m07300_a0-optimized.cl @@ -118,13 +118,22 @@ KERNEL_FQ void m07300_m04 (KERN_ATTR_RULES_ESALT (rakp_t)) * modifier */ + const u64 gid = get_global_id (0); const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); /** - * base + * s_msg */ - const u64 gid = get_global_id (0); + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; @@ -186,40 +195,40 @@ KERNEL_FQ void m07300_m04 (KERN_ATTR_RULES_ESALT (rakp_t)) for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16) { - w0[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; - w3[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 14]; - w3[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 15]; + w0[0] = s_esalt_buf[esalt_off + 0]; + w0[1] = s_esalt_buf[esalt_off + 1]; + w0[2] = s_esalt_buf[esalt_off + 2]; + w0[3] = s_esalt_buf[esalt_off + 3]; + w1[0] = s_esalt_buf[esalt_off + 4]; + w1[1] = s_esalt_buf[esalt_off + 5]; + w1[2] = s_esalt_buf[esalt_off + 6]; + w1[3] = s_esalt_buf[esalt_off + 7]; + w2[0] = s_esalt_buf[esalt_off + 8]; + w2[1] = s_esalt_buf[esalt_off + 9]; + w2[2] = s_esalt_buf[esalt_off + 10]; + w2[3] = s_esalt_buf[esalt_off + 11]; + w3[0] = s_esalt_buf[esalt_off + 12]; + w3[1] = s_esalt_buf[esalt_off + 13]; + w3[2] = s_esalt_buf[esalt_off + 14]; + w3[3] = s_esalt_buf[esalt_off + 15]; sha1_transform_vector (w0, w1, w2, w3, ipad); } - w0[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; + w0[0] = s_esalt_buf[esalt_off + 0]; + w0[1] = s_esalt_buf[esalt_off + 1]; + w0[2] = s_esalt_buf[esalt_off + 2]; + w0[3] = s_esalt_buf[esalt_off + 3]; + w1[0] = s_esalt_buf[esalt_off + 4]; + w1[1] = s_esalt_buf[esalt_off + 5]; + w1[2] = s_esalt_buf[esalt_off + 6]; + w1[3] = s_esalt_buf[esalt_off + 7]; + w2[0] = s_esalt_buf[esalt_off + 8]; + w2[1] = s_esalt_buf[esalt_off + 9]; + w2[2] = s_esalt_buf[esalt_off + 10]; + w2[3] = s_esalt_buf[esalt_off + 11]; + w3[0] = s_esalt_buf[esalt_off + 12]; + w3[1] = s_esalt_buf[esalt_off + 13]; w3[2] = 0; w3[3] = (64 + esalt_size) * 8; @@ -245,13 +254,22 @@ KERNEL_FQ void m07300_s04 (KERN_ATTR_RULES_ESALT (rakp_t)) * modifier */ + const u64 gid = get_global_id (0); const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); /** - * base + * s_msg */ - const u64 gid = get_global_id (0); + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; @@ -325,40 +343,40 @@ KERNEL_FQ void m07300_s04 (KERN_ATTR_RULES_ESALT (rakp_t)) for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16) { - w0[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; - w3[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 14]; - w3[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 15]; + w0[0] = s_esalt_buf[esalt_off + 0]; + w0[1] = s_esalt_buf[esalt_off + 1]; + w0[2] = s_esalt_buf[esalt_off + 2]; + w0[3] = s_esalt_buf[esalt_off + 3]; + w1[0] = s_esalt_buf[esalt_off + 4]; + w1[1] = s_esalt_buf[esalt_off + 5]; + w1[2] = s_esalt_buf[esalt_off + 6]; + w1[3] = s_esalt_buf[esalt_off + 7]; + w2[0] = s_esalt_buf[esalt_off + 8]; + w2[1] = s_esalt_buf[esalt_off + 9]; + w2[2] = s_esalt_buf[esalt_off + 10]; + w2[3] = s_esalt_buf[esalt_off + 11]; + w3[0] = s_esalt_buf[esalt_off + 12]; + w3[1] = s_esalt_buf[esalt_off + 13]; + w3[2] = s_esalt_buf[esalt_off + 14]; + w3[3] = s_esalt_buf[esalt_off + 15]; sha1_transform_vector (w0, w1, w2, w3, ipad); } - w0[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; + w0[0] = s_esalt_buf[esalt_off + 0]; + w0[1] = s_esalt_buf[esalt_off + 1]; + w0[2] = s_esalt_buf[esalt_off + 2]; + w0[3] = s_esalt_buf[esalt_off + 3]; + w1[0] = s_esalt_buf[esalt_off + 4]; + w1[1] = s_esalt_buf[esalt_off + 5]; + w1[2] = s_esalt_buf[esalt_off + 6]; + w1[3] = s_esalt_buf[esalt_off + 7]; + w2[0] = s_esalt_buf[esalt_off + 8]; + w2[1] = s_esalt_buf[esalt_off + 9]; + w2[2] = s_esalt_buf[esalt_off + 10]; + w2[3] = s_esalt_buf[esalt_off + 11]; + w3[0] = s_esalt_buf[esalt_off + 12]; + w3[1] = s_esalt_buf[esalt_off + 13]; w3[2] = 0; w3[3] = (64 + esalt_size) * 8; diff --git a/OpenCL/m07300_a1-optimized.cl b/OpenCL/m07300_a1-optimized.cl index 83fce7cf9..21e1e0d2f 100644 --- a/OpenCL/m07300_a1-optimized.cl +++ b/OpenCL/m07300_a1-optimized.cl @@ -116,13 +116,22 @@ KERNEL_FQ void m07300_m04 (KERN_ATTR_ESALT (rakp_t)) * modifier */ + const u64 gid = get_global_id (0); const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); /** - * base + * s_msg */ - const u64 gid = get_global_id (0); + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; @@ -252,40 +261,40 @@ KERNEL_FQ void m07300_m04 (KERN_ATTR_ESALT (rakp_t)) for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16) { - w0[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; - w3[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 14]; - w3[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 15]; + w0[0] = s_esalt_buf[esalt_off + 0]; + w0[1] = s_esalt_buf[esalt_off + 1]; + w0[2] = s_esalt_buf[esalt_off + 2]; + w0[3] = s_esalt_buf[esalt_off + 3]; + w1[0] = s_esalt_buf[esalt_off + 4]; + w1[1] = s_esalt_buf[esalt_off + 5]; + w1[2] = s_esalt_buf[esalt_off + 6]; + w1[3] = s_esalt_buf[esalt_off + 7]; + w2[0] = s_esalt_buf[esalt_off + 8]; + w2[1] = s_esalt_buf[esalt_off + 9]; + w2[2] = s_esalt_buf[esalt_off + 10]; + w2[3] = s_esalt_buf[esalt_off + 11]; + w3[0] = s_esalt_buf[esalt_off + 12]; + w3[1] = s_esalt_buf[esalt_off + 13]; + w3[2] = s_esalt_buf[esalt_off + 14]; + w3[3] = s_esalt_buf[esalt_off + 15]; sha1_transform_vector (w0, w1, w2, w3, ipad); } - w0[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; + w0[0] = s_esalt_buf[esalt_off + 0]; + w0[1] = s_esalt_buf[esalt_off + 1]; + w0[2] = s_esalt_buf[esalt_off + 2]; + w0[3] = s_esalt_buf[esalt_off + 3]; + w1[0] = s_esalt_buf[esalt_off + 4]; + w1[1] = s_esalt_buf[esalt_off + 5]; + w1[2] = s_esalt_buf[esalt_off + 6]; + w1[3] = s_esalt_buf[esalt_off + 7]; + w2[0] = s_esalt_buf[esalt_off + 8]; + w2[1] = s_esalt_buf[esalt_off + 9]; + w2[2] = s_esalt_buf[esalt_off + 10]; + w2[3] = s_esalt_buf[esalt_off + 11]; + w3[0] = s_esalt_buf[esalt_off + 12]; + w3[1] = s_esalt_buf[esalt_off + 13]; w3[2] = 0; w3[3] = (64 + esalt_size) * 8; @@ -311,13 +320,22 @@ KERNEL_FQ void m07300_s04 (KERN_ATTR_ESALT (rakp_t)) * modifier */ + const u64 gid = get_global_id (0); const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); /** - * base + * s_msg */ - const u64 gid = get_global_id (0); + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; @@ -459,40 +477,40 @@ KERNEL_FQ void m07300_s04 (KERN_ATTR_ESALT (rakp_t)) for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16) { - w0[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; - w3[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 14]; - w3[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 15]; + w0[0] = s_esalt_buf[esalt_off + 0]; + w0[1] = s_esalt_buf[esalt_off + 1]; + w0[2] = s_esalt_buf[esalt_off + 2]; + w0[3] = s_esalt_buf[esalt_off + 3]; + w1[0] = s_esalt_buf[esalt_off + 4]; + w1[1] = s_esalt_buf[esalt_off + 5]; + w1[2] = s_esalt_buf[esalt_off + 6]; + w1[3] = s_esalt_buf[esalt_off + 7]; + w2[0] = s_esalt_buf[esalt_off + 8]; + w2[1] = s_esalt_buf[esalt_off + 9]; + w2[2] = s_esalt_buf[esalt_off + 10]; + w2[3] = s_esalt_buf[esalt_off + 11]; + w3[0] = s_esalt_buf[esalt_off + 12]; + w3[1] = s_esalt_buf[esalt_off + 13]; + w3[2] = s_esalt_buf[esalt_off + 14]; + w3[3] = s_esalt_buf[esalt_off + 15]; sha1_transform_vector (w0, w1, w2, w3, ipad); } - w0[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; + w0[0] = s_esalt_buf[esalt_off + 0]; + w0[1] = s_esalt_buf[esalt_off + 1]; + w0[2] = s_esalt_buf[esalt_off + 2]; + w0[3] = s_esalt_buf[esalt_off + 3]; + w1[0] = s_esalt_buf[esalt_off + 4]; + w1[1] = s_esalt_buf[esalt_off + 5]; + w1[2] = s_esalt_buf[esalt_off + 6]; + w1[3] = s_esalt_buf[esalt_off + 7]; + w2[0] = s_esalt_buf[esalt_off + 8]; + w2[1] = s_esalt_buf[esalt_off + 9]; + w2[2] = s_esalt_buf[esalt_off + 10]; + w2[3] = s_esalt_buf[esalt_off + 11]; + w3[0] = s_esalt_buf[esalt_off + 12]; + w3[1] = s_esalt_buf[esalt_off + 13]; w3[2] = 0; w3[3] = (64 + esalt_size) * 8; diff --git a/OpenCL/m07300_a3-optimized.cl b/OpenCL/m07300_a3-optimized.cl index b3b74c963..a8a2904a7 100644 --- a/OpenCL/m07300_a3-optimized.cl +++ b/OpenCL/m07300_a3-optimized.cl @@ -110,7 +110,7 @@ DECLSPEC void hmac_sha1_run (u32x *w0, u32x *w1, u32x *w2, u32x *w3, u32x *ipad, sha1_transform_vector (w0, w1, w2, w3, digest); } -DECLSPEC void m07300m (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KERN_ATTR_ESALT (rakp_t)) +DECLSPEC void m07300m (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KERN_ATTR_ESALT (rakp_t), LOCAL_AS u32 *s_esalt_buf) { /** * modifier @@ -175,40 +175,40 @@ DECLSPEC void m07300m (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KER for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16) { - w0_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; - w3_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 14]; - w3_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 15]; + w0_t[0] = s_esalt_buf[esalt_off + 0]; + w0_t[1] = s_esalt_buf[esalt_off + 1]; + w0_t[2] = s_esalt_buf[esalt_off + 2]; + w0_t[3] = s_esalt_buf[esalt_off + 3]; + w1_t[0] = s_esalt_buf[esalt_off + 4]; + w1_t[1] = s_esalt_buf[esalt_off + 5]; + w1_t[2] = s_esalt_buf[esalt_off + 6]; + w1_t[3] = s_esalt_buf[esalt_off + 7]; + w2_t[0] = s_esalt_buf[esalt_off + 8]; + w2_t[1] = s_esalt_buf[esalt_off + 9]; + w2_t[2] = s_esalt_buf[esalt_off + 10]; + w2_t[3] = s_esalt_buf[esalt_off + 11]; + w3_t[0] = s_esalt_buf[esalt_off + 12]; + w3_t[1] = s_esalt_buf[esalt_off + 13]; + w3_t[2] = s_esalt_buf[esalt_off + 14]; + w3_t[3] = s_esalt_buf[esalt_off + 15]; sha1_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad); } - w0_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; + w0_t[0] = s_esalt_buf[esalt_off + 0]; + w0_t[1] = s_esalt_buf[esalt_off + 1]; + w0_t[2] = s_esalt_buf[esalt_off + 2]; + w0_t[3] = s_esalt_buf[esalt_off + 3]; + w1_t[0] = s_esalt_buf[esalt_off + 4]; + w1_t[1] = s_esalt_buf[esalt_off + 5]; + w1_t[2] = s_esalt_buf[esalt_off + 6]; + w1_t[3] = s_esalt_buf[esalt_off + 7]; + w2_t[0] = s_esalt_buf[esalt_off + 8]; + w2_t[1] = s_esalt_buf[esalt_off + 9]; + w2_t[2] = s_esalt_buf[esalt_off + 10]; + w2_t[3] = s_esalt_buf[esalt_off + 11]; + w3_t[0] = s_esalt_buf[esalt_off + 12]; + w3_t[1] = s_esalt_buf[esalt_off + 13]; w3_t[2] = 0; w3_t[3] = (64 + esalt_size) * 8; @@ -220,7 +220,7 @@ DECLSPEC void m07300m (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KER } } -DECLSPEC void m07300s (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KERN_ATTR_ESALT (rakp_t)) +DECLSPEC void m07300s (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KERN_ATTR_ESALT (rakp_t), LOCAL_AS u32 *s_esalt_buf) { /** * modifier @@ -297,40 +297,40 @@ DECLSPEC void m07300s (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KER for (esalt_left = esalt_size, esalt_off = 0; esalt_left >= 56; esalt_left -= 64, esalt_off += 16) { - w0_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; - w3_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 14]; - w3_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 15]; + w0_t[0] = s_esalt_buf[esalt_off + 0]; + w0_t[1] = s_esalt_buf[esalt_off + 1]; + w0_t[2] = s_esalt_buf[esalt_off + 2]; + w0_t[3] = s_esalt_buf[esalt_off + 3]; + w1_t[0] = s_esalt_buf[esalt_off + 4]; + w1_t[1] = s_esalt_buf[esalt_off + 5]; + w1_t[2] = s_esalt_buf[esalt_off + 6]; + w1_t[3] = s_esalt_buf[esalt_off + 7]; + w2_t[0] = s_esalt_buf[esalt_off + 8]; + w2_t[1] = s_esalt_buf[esalt_off + 9]; + w2_t[2] = s_esalt_buf[esalt_off + 10]; + w2_t[3] = s_esalt_buf[esalt_off + 11]; + w3_t[0] = s_esalt_buf[esalt_off + 12]; + w3_t[1] = s_esalt_buf[esalt_off + 13]; + w3_t[2] = s_esalt_buf[esalt_off + 14]; + w3_t[3] = s_esalt_buf[esalt_off + 15]; sha1_transform_vector (w0_t, w1_t, w2_t, w3_t, ipad); } - w0_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 0]; - w0_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 1]; - w0_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 2]; - w0_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 3]; - w1_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 4]; - w1_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 5]; - w1_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 6]; - w1_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 7]; - w2_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 8]; - w2_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 9]; - w2_t[2] = esalt_bufs[digests_offset].salt_buf[esalt_off + 10]; - w2_t[3] = esalt_bufs[digests_offset].salt_buf[esalt_off + 11]; - w3_t[0] = esalt_bufs[digests_offset].salt_buf[esalt_off + 12]; - w3_t[1] = esalt_bufs[digests_offset].salt_buf[esalt_off + 13]; + w0_t[0] = s_esalt_buf[esalt_off + 0]; + w0_t[1] = s_esalt_buf[esalt_off + 1]; + w0_t[2] = s_esalt_buf[esalt_off + 2]; + w0_t[3] = s_esalt_buf[esalt_off + 3]; + w1_t[0] = s_esalt_buf[esalt_off + 4]; + w1_t[1] = s_esalt_buf[esalt_off + 5]; + w1_t[2] = s_esalt_buf[esalt_off + 6]; + w1_t[3] = s_esalt_buf[esalt_off + 7]; + w2_t[0] = s_esalt_buf[esalt_off + 8]; + w2_t[1] = s_esalt_buf[esalt_off + 9]; + w2_t[2] = s_esalt_buf[esalt_off + 10]; + w2_t[3] = s_esalt_buf[esalt_off + 11]; + w3_t[0] = s_esalt_buf[esalt_off + 12]; + w3_t[1] = s_esalt_buf[esalt_off + 13]; w3_t[2] = 0; w3_t[3] = (64 + esalt_size) * 8; @@ -345,13 +345,32 @@ DECLSPEC void m07300s (u32 *w0, u32 *w1, u32 *w2, u32 *w3, const u32 pw_len, KER KERNEL_FQ void m07300_m04 (KERN_ATTR_ESALT (rakp_t)) { /** - * base + * modifier */ const u64 gid = get_global_id (0); + const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); + + /** + * s_msg + */ + + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; + /** + * base + */ + u32 w0[4]; w0[0] = pws[gid].i[ 0]; @@ -386,19 +405,38 @@ KERNEL_FQ void m07300_m04 (KERN_ATTR_ESALT (rakp_t)) * main */ - m07300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max); + m07300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max, s_esalt_buf); } KERNEL_FQ void m07300_m08 (KERN_ATTR_ESALT (rakp_t)) { /** - * base + * modifier */ const u64 gid = get_global_id (0); + const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); + + /** + * s_msg + */ + + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; + /** + * base + */ + u32 w0[4]; w0[0] = pws[gid].i[ 0]; @@ -433,19 +471,38 @@ KERNEL_FQ void m07300_m08 (KERN_ATTR_ESALT (rakp_t)) * main */ - m07300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max); + m07300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max, s_esalt_buf); } KERNEL_FQ void m07300_m16 (KERN_ATTR_ESALT (rakp_t)) { /** - * base + * modifier */ const u64 gid = get_global_id (0); + const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); + + /** + * s_msg + */ + + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; + /** + * base + */ + u32 w0[4]; w0[0] = pws[gid].i[ 0]; @@ -480,19 +537,38 @@ KERNEL_FQ void m07300_m16 (KERN_ATTR_ESALT (rakp_t)) * main */ - m07300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max); + m07300m (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max, s_esalt_buf); } KERNEL_FQ void m07300_s04 (KERN_ATTR_ESALT (rakp_t)) { /** - * base + * modifier */ const u64 gid = get_global_id (0); + const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); + + /** + * s_msg + */ + + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; + /** + * base + */ + u32 w0[4]; w0[0] = pws[gid].i[ 0]; @@ -527,19 +603,38 @@ KERNEL_FQ void m07300_s04 (KERN_ATTR_ESALT (rakp_t)) * main */ - m07300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max); + m07300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max, s_esalt_buf); } KERNEL_FQ void m07300_s08 (KERN_ATTR_ESALT (rakp_t)) { /** - * base + * modifier */ const u64 gid = get_global_id (0); + const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); + + /** + * s_msg + */ + + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; + /** + * base + */ + u32 w0[4]; w0[0] = pws[gid].i[ 0]; @@ -574,19 +669,38 @@ KERNEL_FQ void m07300_s08 (KERN_ATTR_ESALT (rakp_t)) * main */ - m07300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max); + m07300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max, s_esalt_buf); } KERNEL_FQ void m07300_s16 (KERN_ATTR_ESALT (rakp_t)) { /** - * base + * modifier */ const u64 gid = get_global_id (0); + const u64 lid = get_local_id (0); + const u64 lsz = get_local_size (0); + + /** + * s_msg + */ + + LOCAL_VK u32 s_esalt_buf[128]; + + for (u32 i = lid; i < 128; i += lsz) + { + s_esalt_buf[i] = esalt_bufs[digests_offset].salt_buf[i]; + } + + SYNC_THREADS (); if (gid >= gid_max) return; + /** + * base + */ + u32 w0[4]; w0[0] = pws[gid].i[ 0]; @@ -621,5 +735,5 @@ KERNEL_FQ void m07300_s16 (KERN_ATTR_ESALT (rakp_t)) * main */ - m07300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max); + m07300s (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_extra0_buf, d_extra1_buf, d_extra2_buf, d_extra3_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, combs_mode, gid_max, s_esalt_buf); }