Zero pws_buf before reuse

[hashcat.git] / OpenCL / m08700_a1.cl

diff --git a/OpenCL/m08700_a1.cl b/OpenCL/m08700_a1.cl
index 7ee90f4..4298d6d 100644 (file)
--- a/OpenCL/m08700_a1.cl
+++ b/OpenCL/m08700_a1.cl
@@ -7,6 +7,8 @@
 
 #define _LOTUS6_
 
+#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -18,9 +20,7 @@
 #include "include/kernel_functions.c"
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
-
-#define COMPARE_S "OpenCL/check_single_comp4.c"
-#define COMPARE_M "OpenCL/check_multi_comp4.c"
+#include "OpenCL/simd.c"
 
 __constant u32 lotus_magic_table[256] =
 {
@@ -292,41 +292,20 @@ __kernel void m08700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * base
    */
 
-  u32 wordl0[4];
-
-  wordl0[0] = pws[gid].i[ 0];
-  wordl0[1] = pws[gid].i[ 1];
-  wordl0[2] = pws[gid].i[ 2];
-  wordl0[3] = pws[gid].i[ 3];
-
-  u32 wordl1[4];
-
-  wordl1[0] = pws[gid].i[ 4];
-  wordl1[1] = pws[gid].i[ 5];
-  wordl1[2] = pws[gid].i[ 6];
-  wordl1[3] = pws[gid].i[ 7];
-
-  u32 wordl2[4];
-
-  wordl2[0] = 0;
-  wordl2[1] = 0;
-  wordl2[2] = 0;
-  wordl2[3] = 0;
-
-  u32 wordl3[4];
+  u32 pws0[4] = { 0 };
+  u32 pws1[4] = { 0 };
 
-  wordl3[0] = 0;
-  wordl3[1] = 0;
-  wordl3[2] = 0;
-  wordl3[3] = 0;
+  pws0[0] = pws[gid].i[0];
+  pws0[1] = pws[gid].i[1];
+  pws0[2] = pws[gid].i[2];
+  pws0[3] = pws[gid].i[3];
+  pws1[0] = pws[gid].i[4];
+  pws1[1] = pws[gid].i[5];
+  pws1[2] = pws[gid].i[6];
+  pws1[3] = pws[gid].i[7];
 
   const u32 pw_l_len = pws[gid].pw_len;
 
-  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
-  {
-    switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
@@ -338,46 +317,32 @@ __kernel void m08700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < combs_cnt; il_pos += VECT_SIZE)
   {
-    const u32 pw_r_len = combs_buf[il_pos].pw_len;
-
-    const u32 pw_len = pw_l_len + pw_r_len;
-
-    u32 wordr0[4];
-
-    wordr0[0] = combs_buf[il_pos].i[0];
-    wordr0[1] = combs_buf[il_pos].i[1];
-    wordr0[2] = combs_buf[il_pos].i[2];
-    wordr0[3] = combs_buf[il_pos].i[3];
-
-    u32 wordr1[4];
-
-    wordr1[0] = combs_buf[il_pos].i[4];
-    wordr1[1] = combs_buf[il_pos].i[5];
-    wordr1[2] = combs_buf[il_pos].i[6];
-    wordr1[3] = combs_buf[il_pos].i[7];
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    u32 wordr2[4];
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 0 };
 
-    u32 wordr3[4];
-
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
+    wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
+    wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
+    wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
+    wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
+    wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
+    wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
+    wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
 
     if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
     {
       switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
     }
 
-    u32 w[16];
+    u32x w[16];
 
     w[ 0] = wordl0[0] | wordr0[0];
     w[ 1] = wordl0[1] | wordr0[1];
@@ -396,7 +361,7 @@ __kernel void m08700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w[14] = wordl3[2] | wordr3[2];
     w[15] = wordl3[3] | wordr3[3];
 
-    u32 state[4];
+    u32x state[4];
 
     state[0] = 0;
     state[1] = 0;
@@ -469,10 +434,10 @@ __kernel void m08700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     domino_big_md (w, 34, state, s_lotus_magic_table);
 
-    u32 a = state[0] & 0xffffffff;
-    u32 b = state[1] & 0xffffffff;
-    u32 c = state[2] & 0x000000ff;
-    u32 d = state[3] & 0x00000000;
+    u32x a = state[0] & 0xffffffff;
+    u32x b = state[1] & 0xffffffff;
+    u32x c = state[2] & 0x000000ff;
+    u32x d = state[3] & 0x00000000;
 
     const u32 r0 = a;
     const u32 r1 = b;
@@ -531,41 +496,20 @@ __kernel void m08700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * base
    */
 
-  u32 wordl0[4];
-
-  wordl0[0] = pws[gid].i[ 0];
-  wordl0[1] = pws[gid].i[ 1];
-  wordl0[2] = pws[gid].i[ 2];
-  wordl0[3] = pws[gid].i[ 3];
-
-  u32 wordl1[4];
-
-  wordl1[0] = pws[gid].i[ 4];
-  wordl1[1] = pws[gid].i[ 5];
-  wordl1[2] = pws[gid].i[ 6];
-  wordl1[3] = pws[gid].i[ 7];
+  u32 pws0[4] = { 0 };
+  u32 pws1[4] = { 0 };
 
-  u32 wordl2[4];
-
-  wordl2[0] = 0;
-  wordl2[1] = 0;
-  wordl2[2] = 0;
-  wordl2[3] = 0;
-
-  u32 wordl3[4];
-
-  wordl3[0] = 0;
-  wordl3[1] = 0;
-  wordl3[2] = 0;
-  wordl3[3] = 0;
+  pws0[0] = pws[gid].i[0];
+  pws0[1] = pws[gid].i[1];
+  pws0[2] = pws[gid].i[2];
+  pws0[3] = pws[gid].i[3];
+  pws1[0] = pws[gid].i[4];
+  pws1[1] = pws[gid].i[5];
+  pws1[2] = pws[gid].i[6];
+  pws1[3] = pws[gid].i[7];
 
   const u32 pw_l_len = pws[gid].pw_len;
 
-  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
-  {
-    switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
-  }
-
   /**
    * salt
    */
@@ -589,46 +533,32 @@ __kernel void m08700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < combs_cnt; il_pos += VECT_SIZE)
   {
-    const u32 pw_r_len = combs_buf[il_pos].pw_len;
-
-    const u32 pw_len = pw_l_len + pw_r_len;
-
-    u32 wordr0[4];
-
-    wordr0[0] = combs_buf[il_pos].i[0];
-    wordr0[1] = combs_buf[il_pos].i[1];
-    wordr0[2] = combs_buf[il_pos].i[2];
-    wordr0[3] = combs_buf[il_pos].i[3];
-
-    u32 wordr1[4];
-
-    wordr1[0] = combs_buf[il_pos].i[4];
-    wordr1[1] = combs_buf[il_pos].i[5];
-    wordr1[2] = combs_buf[il_pos].i[6];
-    wordr1[3] = combs_buf[il_pos].i[7];
-
-    u32 wordr2[4];
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    u32 wordr3[4];
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 0 };
 
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
+    wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
+    wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
+    wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
+    wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
+    wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
+    wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
+    wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
 
     if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
     {
       switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
     }
 
-    u32 w[16];
+    u32x w[16];
 
     w[ 0] = wordl0[0] | wordr0[0];
     w[ 1] = wordl0[1] | wordr0[1];
@@ -647,7 +577,7 @@ __kernel void m08700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     w[14] = wordl3[2] | wordr3[2];
     w[15] = wordl3[3] | wordr3[3];
 
-    u32 state[4];
+    u32x state[4];
 
     state[0] = 0;
     state[1] = 0;
@@ -720,10 +650,10 @@ __kernel void m08700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     domino_big_md (w, 34, state, s_lotus_magic_table);
 
-    u32 a = state[0] & 0xffffffff;
-    u32 b = state[1] & 0xffffffff;
-    u32 c = state[2] & 0x000000ff;
-    u32 d = state[3] & 0x00000000;
+    u32x a = state[0] & 0xffffffff;
+    u32x b = state[1] & 0xffffffff;
+    u32x c = state[2] & 0x000000ff;
+    u32x d = state[3] & 0x00000000;
 
     const u32 r0 = a;
     const u32 r1 = b;