diff --git a/.vscode/settings.json b/.vscode/settings.json new file mode 100644 index 000000000..9ddf6b280 --- /dev/null +++ b/.vscode/settings.json @@ -0,0 +1,3 @@ +{ + "cmake.ignoreCMakeListsMissing": true +} \ No newline at end of file diff --git a/OpenCL/m33000_a0-optimized.cl b/OpenCL/m33000_a0-optimized.cl deleted file mode 100644 index 64e38e67f..000000000 --- a/OpenCL/m33000_a0-optimized.cl +++ /dev/null @@ -1,450 +0,0 @@ -/** - * Author......: See docs/credits.txt - * License.....: MIT - */ - -#define NEW_SIMD_CODE - -#ifdef KERNEL_STATIC -#include M2S(INCLUDE_PATH/inc_vendor.h) -#include M2S(INCLUDE_PATH/inc_types.h) -#include M2S(INCLUDE_PATH/inc_platform.cl) -#include M2S(INCLUDE_PATH/inc_common.cl) -#include M2S(INCLUDE_PATH/inc_rp_optimized.h) -#include M2S(INCLUDE_PATH/inc_rp_optimized.cl) -#include M2S(INCLUDE_PATH/inc_simd.cl) -#include M2S(INCLUDE_PATH/inc_hash_md5.cl) -#endif - -KERNEL_FQ void m33000_m04 (KERN_ATTR_RULES ()) -{ - /** - * modifier - */ - - const u64 lid = get_local_id (0); - - /** - * base - */ - - const u64 gid = get_global_id (0); - - if (gid >= GID_CNT) return; - - u32 pw_buf0[4]; - u32 pw_buf1[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]; - 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 & 63; - - /** - * salt - */ - - u32 salt_buf0[4]; - u32 salt_buf1[4]; - u32 salt_buf2[4]; - u32 salt_buf3[4]; - - salt_buf0[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 0]; - salt_buf0[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 1]; - salt_buf0[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 2]; - salt_buf0[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 3]; - salt_buf1[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 4]; - salt_buf1[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 5]; - salt_buf1[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 6]; - salt_buf1[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 7]; - salt_buf2[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 8]; - salt_buf2[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 9]; - salt_buf2[2] = salt_bufs[SALT_POS_HOST].salt_buf[10]; - salt_buf2[3] = salt_bufs[SALT_POS_HOST].salt_buf[11]; - salt_buf3[0] = salt_bufs[SALT_POS_HOST].salt_buf[12]; - salt_buf3[1] = salt_bufs[SALT_POS_HOST].salt_buf[13]; - salt_buf3[2] = salt_bufs[SALT_POS_HOST].salt_buf[14]; - salt_buf3[3] = salt_bufs[SALT_POS_HOST].salt_buf[15]; - - const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len; - - /** - * loop - */ - - for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE) - { - u32x w0[4] = { 0 }; - u32x w1[4] = { 0 }; - u32x w2[4] = { 0 }; - u32x w3[4] = { 0 }; - - const u32x out_len = apply_rules_vect_optimized (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1); - - /** - * append salt - */ - - u32x s0[4]; - u32x s1[4]; - u32x s2[4]; - u32x s3[4]; - - s0[0] = salt_buf0[0]; - s0[1] = salt_buf0[1]; - s0[2] = salt_buf0[2]; - s0[3] = salt_buf0[3]; - s1[0] = salt_buf1[0]; - s1[1] = salt_buf1[1]; - s1[2] = salt_buf1[2]; - s1[3] = salt_buf1[3]; - s2[0] = salt_buf2[0]; - s2[1] = salt_buf2[1]; - s2[2] = salt_buf2[2]; - s2[3] = salt_buf2[3]; - s3[0] = salt_buf3[0]; - s3[1] = salt_buf3[1]; - s3[2] = salt_buf3[2]; - s3[3] = salt_buf3[3]; - - switch_buffer_by_offset_le_VV (s0, s1, s2, s3, out_len); - - const u32x pw_salt_len = out_len + salt_len; - - w0[0] |= s0[0]; - w0[1] |= s0[1]; - w0[2] |= s0[2]; - w0[3] |= s0[3]; - w1[0] |= s1[0]; - w1[1] |= s1[1]; - w1[2] |= s1[2]; - w1[3] |= s1[3]; - w2[0] |= s2[0]; - w2[1] |= s2[1]; - w2[2] |= s2[2]; - w2[3] |= s2[3]; - w3[0] |= s3[0]; - w3[1] |= s3[1]; - w3[2] = pw_salt_len * 8; - w3[3] = 0; - - /** - * md5 - */ - - u32x a = MD5M_A; - u32x b = MD5M_B; - u32x c = MD5M_C; - u32x d = MD5M_D; - - MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00); - MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01); - MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02); - MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03); - MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00); - MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01); - MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02); - MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03); - MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00); - MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01); - MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02); - MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03); - MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00); - MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01); - MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02); - MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03); - - MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10); - MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11); - MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12); - MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13); - MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10); - MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11); - MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12); - MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13); - MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10); - MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11); - MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12); - MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13); - MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10); - MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11); - MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12); - MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13); - - u32x t; - - MD5_STEP (MD5_H1, a, b, c, d, w1[1], MD5C20, MD5S20); - MD5_STEP (MD5_H2, d, a, b, c, w2[0], MD5C21, MD5S21); - MD5_STEP (MD5_H1, c, d, a, b, w2[3], MD5C22, MD5S22); - MD5_STEP (MD5_H2, b, c, d, a, w3[2], MD5C23, MD5S23); - MD5_STEP (MD5_H1, a, b, c, d, w0[1], MD5C24, MD5S20); - MD5_STEP (MD5_H2, d, a, b, c, w1[0], MD5C25, MD5S21); - MD5_STEP (MD5_H1, c, d, a, b, w1[3], MD5C26, MD5S22); - MD5_STEP (MD5_H2, b, c, d, a, w2[2], MD5C27, MD5S23); - MD5_STEP (MD5_H1, a, b, c, d, w3[1], MD5C28, MD5S20); - MD5_STEP (MD5_H2, d, a, b, c, w0[0], MD5C29, MD5S21); - MD5_STEP (MD5_H1, c, d, a, b, w0[3], MD5C2a, MD5S22); - MD5_STEP (MD5_H2, b, c, d, a, w1[2], MD5C2b, MD5S23); - MD5_STEP (MD5_H1, a, b, c, d, w2[1], MD5C2c, MD5S20); - MD5_STEP (MD5_H2, d, a, b, c, w3[0], MD5C2d, MD5S21); - MD5_STEP (MD5_H1, c, d, a, b, w3[3], MD5C2e, MD5S22); - MD5_STEP (MD5_H2, b, c, d, a, w0[2], MD5C2f, MD5S23); - - MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30); - MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31); - MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32); - MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33); - MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30); - MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31); - MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32); - MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33); - MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30); - MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31); - MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32); - MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33); - MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30); - MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31); - MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32); - MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33); - - COMPARE_M_SIMD (a, d, c, b); - } -} - -KERNEL_FQ void m33000_m08 (KERN_ATTR_RULES ()) -{ -} - -KERNEL_FQ void m33000_m16 (KERN_ATTR_RULES ()) -{ -} - -KERNEL_FQ void m33000_s04 (KERN_ATTR_RULES ()) -{ - /** - * modifier - */ - - const u64 lid = get_local_id (0); - - /** - * base - */ - - const u64 gid = get_global_id (0); - - if (gid >= GID_CNT) return; - - u32 pw_buf0[4]; - u32 pw_buf1[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]; - 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 & 63; - - /** - * salt - */ - - u32 salt_buf0[4]; - u32 salt_buf1[4]; - u32 salt_buf2[4]; - u32 salt_buf3[4]; - - salt_buf0[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 0]; - salt_buf0[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 1]; - salt_buf0[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 2]; - salt_buf0[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 3]; - salt_buf1[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 4]; - salt_buf1[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 5]; - salt_buf1[2] = salt_bufs[SALT_POS_HOST].salt_buf[ 6]; - salt_buf1[3] = salt_bufs[SALT_POS_HOST].salt_buf[ 7]; - salt_buf2[0] = salt_bufs[SALT_POS_HOST].salt_buf[ 8]; - salt_buf2[1] = salt_bufs[SALT_POS_HOST].salt_buf[ 9]; - salt_buf2[2] = salt_bufs[SALT_POS_HOST].salt_buf[10]; - salt_buf2[3] = salt_bufs[SALT_POS_HOST].salt_buf[11]; - salt_buf3[0] = salt_bufs[SALT_POS_HOST].salt_buf[12]; - salt_buf3[1] = salt_bufs[SALT_POS_HOST].salt_buf[13]; - salt_buf3[2] = salt_bufs[SALT_POS_HOST].salt_buf[14]; - salt_buf3[3] = salt_bufs[SALT_POS_HOST].salt_buf[15]; - - const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len; - - /** - * digest - */ - - const u32 search[4] = - { - digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R0], - digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R1], - digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R2], - digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R3] - }; - - /** - * loop - */ - - for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE) - { - u32x w0[4] = { 0 }; - u32x w1[4] = { 0 }; - u32x w2[4] = { 0 }; - u32x w3[4] = { 0 }; - - const u32x out_len = apply_rules_vect_optimized (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1); - - /** - * append salt - */ - - u32x s0[4]; - u32x s1[4]; - u32x s2[4]; - u32x s3[4]; - - s0[0] = salt_buf0[0]; - s0[1] = salt_buf0[1]; - s0[2] = salt_buf0[2]; - s0[3] = salt_buf0[3]; - s1[0] = salt_buf1[0]; - s1[1] = salt_buf1[1]; - s1[2] = salt_buf1[2]; - s1[3] = salt_buf1[3]; - s2[0] = salt_buf2[0]; - s2[1] = salt_buf2[1]; - s2[2] = salt_buf2[2]; - s2[3] = salt_buf2[3]; - s3[0] = salt_buf3[0]; - s3[1] = salt_buf3[1]; - s3[2] = salt_buf3[2]; - s3[3] = salt_buf3[3]; - - switch_buffer_by_offset_le_VV (s0, s1, s2, s3, out_len); - - const u32x pw_salt_len = out_len + salt_len; - - w0[0] |= s0[0]; - w0[1] |= s0[1]; - w0[2] |= s0[2]; - w0[3] |= s0[3]; - w1[0] |= s1[0]; - w1[1] |= s1[1]; - w1[2] |= s1[2]; - w1[3] |= s1[3]; - w2[0] |= s2[0]; - w2[1] |= s2[1]; - w2[2] |= s2[2]; - w2[3] |= s2[3]; - w3[0] |= s3[0]; - w3[1] |= s3[1]; - w3[2] = pw_salt_len * 8; - w3[3] = 0; - - /** - * md5 - */ - - u32x a = MD5M_A; - u32x b = MD5M_B; - u32x c = MD5M_C; - u32x d = MD5M_D; - - MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00); - MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01); - MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02); - MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03); - MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00); - MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01); - MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02); - MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03); - MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00); - MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01); - MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02); - MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03); - MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00); - MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01); - MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02); - MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03); - - MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10); - MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11); - MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12); - MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13); - MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10); - MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11); - MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12); - MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13); - MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10); - MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11); - MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12); - MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13); - MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10); - MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11); - MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12); - MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13); - - u32x t; - - MD5_STEP (MD5_H1, a, b, c, d, w1[1], MD5C20, MD5S20); - MD5_STEP (MD5_H2, d, a, b, c, w2[0], MD5C21, MD5S21); - MD5_STEP (MD5_H1, c, d, a, b, w2[3], MD5C22, MD5S22); - MD5_STEP (MD5_H2, b, c, d, a, w3[2], MD5C23, MD5S23); - MD5_STEP (MD5_H1, a, b, c, d, w0[1], MD5C24, MD5S20); - MD5_STEP (MD5_H2, d, a, b, c, w1[0], MD5C25, MD5S21); - MD5_STEP (MD5_H1, c, d, a, b, w1[3], MD5C26, MD5S22); - MD5_STEP (MD5_H2, b, c, d, a, w2[2], MD5C27, MD5S23); - MD5_STEP (MD5_H1, a, b, c, d, w3[1], MD5C28, MD5S20); - MD5_STEP (MD5_H2, d, a, b, c, w0[0], MD5C29, MD5S21); - MD5_STEP (MD5_H1, c, d, a, b, w0[3], MD5C2a, MD5S22); - MD5_STEP (MD5_H2, b, c, d, a, w1[2], MD5C2b, MD5S23); - MD5_STEP (MD5_H1, a, b, c, d, w2[1], MD5C2c, MD5S20); - MD5_STEP (MD5_H2, d, a, b, c, w3[0], MD5C2d, MD5S21); - MD5_STEP (MD5_H1, c, d, a, b, w3[3], MD5C2e, MD5S22); - MD5_STEP (MD5_H2, b, c, d, a, w0[2], MD5C2f, MD5S23); - - MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30); - MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31); - MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32); - MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33); - MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30); - MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31); - MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32); - MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33); - MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30); - MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31); - MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32); - MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33); - MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30); - - if (MATCHES_NONE_VS (a, search[0])) continue; - - MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31); - MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32); - MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33); - - COMPARE_S_SIMD (a, d, c, b); - } -} - -KERNEL_FQ void m33000_s08 (KERN_ATTR_RULES ()) -{ -} - -KERNEL_FQ void m33000_s16 (KERN_ATTR_RULES ()) -{ -} diff --git a/OpenCL/m33000_a3-pure.cl b/OpenCL/m33000_a3-pure.cl new file mode 100644 index 000000000..b188becc8 --- /dev/null +++ b/OpenCL/m33000_a3-pure.cl @@ -0,0 +1,216 @@ +/** + * Author......: See docs/credits.txt + * License.....: MIT + */ + +#define NEW_SIMD_CODE + +#ifdef KERNEL_STATIC +#include M2S(INCLUDE_PATH/inc_vendor.h) +#include M2S(INCLUDE_PATH/inc_types.h) +#include M2S(INCLUDE_PATH/inc_platform.cl) +#include M2S(INCLUDE_PATH/inc_common.cl) +#include M2S(INCLUDE_PATH/inc_simd.cl) +#include M2S(INCLUDE_PATH/inc_hash_md5.cl) +#endif + +/** + * 破解逻辑: + * 1. 输入的密文格式为 {enc8}BASE64STRING + * 2. 去掉{enc8}前缀后进行base64解码 + * 3. 解码后的数据: + * - 前16字节为MD5哈希值 + * - 后4字节为salt值 + * 4. 使用密码和salt组合: $pass.$salt + * 5. 对组合后的字符串进行MD5哈希 + * 6. 比较计算结果与目标哈希值 + */ + +KERNEL_FQ void m33000_mxx (KERN_ATTR_VECTOR ()) +{ + /** + * 修饰符 + */ + + const u64 lid = get_local_id (0); + const u64 gid = get_global_id (0); + + if (gid >= GID_CNT) return; + + /** + * 基础 + */ + + const u32 pw_len = pws[gid].pw_len; + + // 存储密码 + u32x w[64] = { 0 }; + + for (u32 i = 0, idx = 0; i < pw_len; i += 4, idx += 1) + { + w[idx] = pws[gid].i[idx]; + } + + // 获取salt + const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len; + + u32x s[64] = { 0 }; + + for (u32 i = 0, idx = 0; i < salt_len; i += 4, idx += 1) + { + s[idx] = salt_bufs[SALT_POS_HOST].salt_buf[idx]; + } + + // 打印第一个密码尝试的信息 + if (gid == 0 && lid == 0) + { + printf("[DEBUG-GPU] Salt length: %u\n", salt_len); + printf("[DEBUG-GPU] Salt value: %08x\n", s[0]); + printf("[DEBUG-GPU] Password length: %u\n", pw_len); + printf("[DEBUG-GPU] First word: %08x\n", w[0]); + } + + /** + * 循环 + */ + + u32x w0l = w[0]; + + for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE) + { + const u32x w0r = words_buf_r[il_pos / VECT_SIZE]; + + const u32x w0 = w0l | w0r; + + w[0] = w0; + + // 初始化MD5上下文 + md5_ctx_vector_t ctx; + + md5_init_vector (&ctx); + + // 先更新密码 + md5_update_vector (&ctx, w, pw_len); + + // 再更新salt + md5_update_vector (&ctx, s, salt_len); + + // 计算最终哈希值 + md5_final_vector (&ctx); + + const u32x r0 = ctx.h[DGST_R0]; + const u32x r1 = ctx.h[DGST_R1]; + const u32x r2 = ctx.h[DGST_R2]; + const u32x r3 = ctx.h[DGST_R3]; + + // 打印第一个哈希计算的结果 + if (gid == 0 && il_pos == 0) + { + printf("[DEBUG-GPU] Password length: %u\n", pw_len); + printf("[DEBUG-GPU] Salt length: %u\n", salt_len); + printf("[DEBUG-GPU] First password bytes: %02x %02x %02x %02x\n", + w[0] & 0xff, (w[0] >> 8) & 0xff, (w[0] >> 16) & 0xff, (w[0] >> 24) & 0xff); + printf("[DEBUG-GPU] Salt bytes: %02x %02x %02x %02x\n", + s[0] & 0xff, (s[0] >> 8) & 0xff, (s[0] >> 16) & 0xff, (s[0] >> 24) & 0xff); + printf("[DEBUG-GPU] Computed hash: %08x %08x %08x %08x\n", r0, r1, r2, r3); + } + + // 打印计算得到的哈希值(仅第一个线程) + if (gid == 0 && il_pos == 0) + { + printf("[DEBUG-GPU] Computed hash: %08x %08x %08x %08x\n", r0, r1, r2, r3); + } + + COMPARE_M_SIMD (r0, r1, r2, r3); + } +} + +KERNEL_FQ void m33000_sxx (KERN_ATTR_VECTOR ()) +{ + /** + * 修饰符 + */ + + const u64 lid = get_local_id (0); + const u64 gid = get_global_id (0); + + if (gid >= GID_CNT) return; + + /** + * 摘要 + */ + + const u32 search[4] = + { + digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R0], + digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R1], + digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R2], + digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R3] + }; + + /** + * 基础 + */ + + const u32 pw_len = pws[gid].pw_len; + + // 存储密码 + u32x w[64] = { 0 }; + + for (u32 i = 0, idx = 0; i < pw_len; i += 4, idx += 1) + { + w[idx] = pws[gid].i[idx]; + } + + // 获取salt + const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len; + + u32x s[64] = { 0 }; + + for (u32 i = 0, idx = 0; i < salt_len; i += 4, idx += 1) + { + s[idx] = salt_bufs[SALT_POS_HOST].salt_buf[idx]; + } + + /** + * 循环 + */ + + u32x w0l = w[0]; + + for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE) + { + const u32x w0r = words_buf_r[il_pos / VECT_SIZE]; + + const u32x w0 = w0l | w0r; + + w[0] = w0; + + // 初始化MD5上下文 + md5_ctx_vector_t ctx; + + md5_init_vector (&ctx); + + // 先更新密码 + md5_update_vector (&ctx, w, pw_len); + + // 再更新salt + md5_update_vector (&ctx, s, salt_len); + + // 计算最终哈希值 + md5_final_vector (&ctx); + + const u32x r0 = ctx.h[DGST_R0]; + const u32x r1 = ctx.h[DGST_R1]; + const u32x r2 = ctx.h[DGST_R2]; + const u32x r3 = ctx.h[DGST_R3]; + + // 打印计算得到的哈希值(仅第一个线程) + if (gid == 0 && il_pos == 0) + { + printf("[DEBUG-GPU] Computed hash: %08x %08x %08x %08x\n", r0, r1, r2, r3); + } + + COMPARE_S_SIMD (r0, r1, r2, r3); + } +} \ No newline at end of file diff --git a/src/modules/module_33000.c b/src/modules/module_33000.c index bb708be75..a2d1e25f3 100644 --- a/src/modules/module_33000.c +++ b/src/modules/module_33000.c @@ -12,6 +12,10 @@ #include "emu_inc_hash_md5.h" #include "memory.h" +// 函数声明 +void module_hash_decode_test(void); + +// 定义攻击类型为内核内部执行 static const u32 ATTACK_EXEC = ATTACK_EXEC_INSIDE_KERNEL; static const u32 DGST_POS0 = 0; static const u32 DGST_POS1 = 3; @@ -34,7 +38,7 @@ static const u64 OPTS_TYPE = OPTS_TYPE_PT_GENERATE_LE | OPTS_TYPE_ST_ADDBITS14; static const u32 SALT_TYPE = SALT_TYPE_GENERIC; static const char *ST_PASS = "123456"; -static const char *ST_HASH = "{enc8}EUxNIpbzGlnJbM4KKjYl+za4fmA="; +static const char *ST_HASH = "{enc8}EUxNIpbzGlnJbM4KKjYl+za4fmA="; // 示例哈希值 u32 module_attack_exec (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const user_options_t *user_options, MAYBE_UNUSED const user_options_extra_t *user_options_extra); u32 module_dgst_pos0 (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const user_options_t *user_options, MAYBE_UNUSED const user_options_extra_t *user_options_extra); @@ -148,14 +152,17 @@ int module_hash_decode (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSE memset (&token, 0, sizeof (hc_token_t)); + // 设置令牌解析规则 token.token_cnt = 2; token.signatures_cnt = 1; - token.signatures_buf[0] = "{enc8}"; + token.signatures_buf[0] = "{enc8}"; // 验证前缀 + // 验证前缀长度 token.len[0] = 6; token.attr[0] = TOKEN_ATTR_FIXED_LENGTH | TOKEN_ATTR_VERIFY_SIGNATURE; + // 验证base64编码部分长度 token.len_min[1] = 28; token.len_max[1] = 28; token.attr[1] = TOKEN_ATTR_VERIFY_LENGTH @@ -163,38 +170,97 @@ int module_hash_decode (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSE const int rc_tokenizer = input_tokenizer ((const u8 *) line_buf, line_len, &token); - if (rc_tokenizer != PARSER_OK) return (rc_tokenizer); + if (rc_tokenizer != PARSER_OK) { + return (rc_tokenizer); + } + // 获取base64编码部分 const u8 *hash_pos = token.buf[1]; const int hash_len = token.len[1]; + // base64解码 u8 tmp_buf[100] = { 0 }; int tmp_len = base64_decode (base64_to_int, hash_pos, hash_len, tmp_buf); - if (tmp_len != 20) return (PARSER_HASH_LENGTH); + if (tmp_len != 20) { + return (PARSER_HASH_LENGTH); + } + // 提取MD5哈希值 memcpy (digest, tmp_buf, 16); - digest[0] = byte_swap_32 (digest[0]); - digest[1] = byte_swap_32 (digest[1]); - digest[2] = byte_swap_32 (digest[2]); - digest[3] = byte_swap_32 (digest[3]); + + // 不进行字节交换,保持原始字节序 + // digest[0] = byte_swap_32(digest[0]); + // digest[1] = byte_swap_32(digest[1]); + // digest[2] = byte_swap_32(digest[2]); + // digest[3] = byte_swap_32(digest[3]); + + // 提取4字节salt + salt->salt_len = 4; + memcpy (salt->salt_buf, tmp_buf + 16, 4); + + // 不需要转换字节序，保持原样 + + return (PARSER_OK); +} + +int module_hash_encode (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const void *digest_buf, MAYBE_UNUSED const salt_t *salt, MAYBE_UNUSED const void *esalt_buf, MAYBE_UNUSED const void *hook_salt_buf, MAYBE_UNUSED const hashinfo_t *hash_info, char *line_buf, MAYBE_UNUSED const int line_size) +{ + const u32 *digest = (const u32 *) digest_buf; + + // 准备MD5哈希值 + u32 tmp[4]; + + tmp[0] = digest[0]; + tmp[1] = digest[1]; + tmp[2] = digest[2]; + tmp[3] = digest[3]; if (hashconfig->opti_type & OPTI_TYPE_OPTIMIZED_KERNEL) { - digest[0] -= MD5M_A; - digest[1] -= MD5M_B; - digest[2] -= MD5M_C; - digest[3] -= MD5M_D; + tmp[0] += MD5M_A; + tmp[1] += MD5M_B; + tmp[2] += MD5M_C; + tmp[3] += MD5M_D; } - u8 salt_hex[9] = { 0 }; - u32 salt_val = byte_swap_32 (*(u32 *)(tmp_buf + 16)); - snprintf ((char *)salt_hex, sizeof(salt_hex), "%08x", salt_val); + // 准备20字节缓冲区(16字节MD5 + 4字节salt) + u8 tmp_buf[20] = { 0 }; + u8 *ptr = tmp_buf; - const bool parse_rc = generic_salt_decode (hashconfig, salt_hex, 8, (u8 *) salt->salt_buf, (int *) &salt->salt_len); - if (parse_rc == false) return (PARSER_SALT_LENGTH); + // 转换MD5哈希值为大端序 + ptr[0] = (tmp[0] >> 24) & 0xff; + ptr[1] = (tmp[0] >> 16) & 0xff; + ptr[2] = (tmp[0] >> 8) & 0xff; + ptr[3] = (tmp[0] >> 0) & 0xff; + ptr[4] = (tmp[1] >> 24) & 0xff; + ptr[5] = (tmp[1] >> 16) & 0xff; + ptr[6] = (tmp[1] >> 8) & 0xff; + ptr[7] = (tmp[1] >> 0) & 0xff; + ptr[8] = (tmp[2] >> 24) & 0xff; + ptr[9] = (tmp[2] >> 16) & 0xff; + ptr[10] = (tmp[2] >> 8) & 0xff; + ptr[11] = (tmp[2] >> 0) & 0xff; + ptr[12] = (tmp[3] >> 24) & 0xff; + ptr[13] = (tmp[3] >> 16) & 0xff; + ptr[14] = (tmp[3] >> 8) & 0xff; + ptr[15] = (tmp[3] >> 0) & 0xff; - return (PARSER_OK); + // 添加salt + memcpy(tmp_buf + 16, salt->salt_buf, 4); + + // 生成最终输出 + char *out_buf = line_buf; + int out_len = 0; + + // 添加{enc8}前缀 + memcpy (out_buf, "{enc8}", 6); + out_len += 6; + + // base64编码 + out_len += base64_encode (int_to_base64, tmp_buf, 20, (u8 *) out_buf + out_len); + + return out_len; } char *module_jit_build_options (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const user_options_t *user_options, MAYBE_UNUSED const user_options_extra_t *user_options_extra, MAYBE_UNUSED const hashes_t *hashes, MAYBE_UNUSED const hc_device_param_t *device_param) @@ -244,60 +310,11 @@ char *module_jit_build_options (MAYBE_UNUSED const hashconfig_t *hashconfig, MAY return jit_build_options; } -int module_hash_encode (MAYBE_UNUSED const hashconfig_t *hashconfig, MAYBE_UNUSED const void *digest_buf, MAYBE_UNUSED const salt_t *salt, MAYBE_UNUSED const void *esalt_buf, MAYBE_UNUSED const void *hook_salt_buf, MAYBE_UNUSED const hashinfo_t *hash_info, char *line_buf, MAYBE_UNUSED const int line_size) -{ - const u32 *digest = (const u32 *) digest_buf; - - u32 tmp[4]; - - tmp[0] = digest[0]; - tmp[1] = digest[1]; - tmp[2] = digest[2]; - tmp[3] = digest[3]; - - if (hashconfig->opti_type & OPTI_TYPE_OPTIMIZED_KERNEL) - { - tmp[0] += MD5M_A; - tmp[1] += MD5M_B; - tmp[2] += MD5M_C; - tmp[3] += MD5M_D; - } - - u8 tmp_buf[20] = { 0 }; - u8 *ptr = tmp_buf; - - ptr[0] = (tmp[0] >> 24) & 0xff; - ptr[1] = (tmp[0] >> 16) & 0xff; - ptr[2] = (tmp[0] >> 8) & 0xff; - ptr[3] = (tmp[0] >> 0) & 0xff; - ptr[4] = (tmp[1] >> 24) & 0xff; - ptr[5] = (tmp[1] >> 16) & 0xff; - ptr[6] = (tmp[1] >> 8) & 0xff; - ptr[7] = (tmp[1] >> 0) & 0xff; - ptr[8] = (tmp[2] >> 24) & 0xff; - ptr[9] = (tmp[2] >> 16) & 0xff; - ptr[10] = (tmp[2] >> 8) & 0xff; - ptr[11] = (tmp[2] >> 0) & 0xff; - ptr[12] = (tmp[3] >> 24) & 0xff; - ptr[13] = (tmp[3] >> 16) & 0xff; - ptr[14] = (tmp[3] >> 8) & 0xff; - ptr[15] = (tmp[3] >> 0) & 0xff; - - memcpy (tmp_buf + 16, salt->salt_buf, 4); - - char *out_buf = line_buf; - int out_len = 0; - - memcpy (out_buf, "{enc8}", 6); - out_len += 6; - - out_len += base64_encode (int_to_base64, tmp_buf, 20, (u8 *) out_buf + out_len); - - return out_len; -} - void module_init (module_ctx_t *module_ctx) { + // 运行测试用例 + module_hash_decode_test(); + module_ctx->module_context_size = MODULE_CONTEXT_SIZE_CURRENT; module_ctx->module_interface_version = MODULE_INTERFACE_VERSION_CURRENT; module_ctx->module_hash_decode_zero_hash = module_hash_decode_zero_hash; @@ -375,3 +392,25 @@ void module_init (module_ctx_t *module_ctx) module_ctx->module_unstable_warning = MODULE_DEFAULT; module_ctx->module_warmup_disable = MODULE_DEFAULT; } + +void module_hash_decode_test() +{ + printf("\n[TEST] Running enc8 hash decode test\n"); + + // 测试用例1: 使用示例哈希 + const char *test_hash = "{enc8}EUxNIpbzGlnJbM4KKjYl+za4fmA="; + u32 test_digest[4] = {0}; + salt_t test_salt; + memset(&test_salt, 0, sizeof(salt_t)); + + hashconfig_t test_config; + memset(&test_config, 0, sizeof(hashconfig_t)); + test_config.opti_type = OPTI_TYPE_ZERO_BYTE | OPTI_TYPE_OPTIMIZED_KERNEL; + + printf("\n[TEST] Testing hash: %s\n", test_hash); + + int rc = module_hash_decode(&test_config, test_digest, &test_salt, NULL, NULL, NULL, test_hash, strlen(test_hash)); + + printf("[TEST] Decode result: %s (code: %d)\n", + rc == PARSER_OK ? "SUCCESS" : "FAILED", rc); +} diff --git a/test_decode.py b/test_decode.py deleted file mode 100644 index 044072f4f..000000000 --- a/test_decode.py +++ /dev/null @@ -1,25 +0,0 @@ -import base64 - -def analyze_enc8_hash(enc8_hash): - # Remove {enc8} prefix - if enc8_hash.startswith('{enc8}'): - enc8_hash = enc8_hash[6:] - - # Decode base64 - decoded = base64.b64decode(enc8_hash) - - # Split into hash and salt - hash_portion = decoded[:-4] - salt_portion = decoded[-4:] - - print(f'Analysis of enc8 hash:') - print(f'Total length: {len(decoded)} bytes') - print(f'Hash portion (hex): {hash_portion.hex()}') - print(f'Salt portion (hex): {salt_portion.hex()}') - print(f'Salt (little-endian): 0x{salt_portion[::-1].hex()}') - - return hash_portion, salt_portion - -# Test vector -test_hash = '{enc8}EUxNIpbzGlnJbM4KKjYl+za4fmA=' -analyze_enc8_hash(test_hash)