Merge branch 'master' of https://github.com/hashcat/oclHashcat

[hashcat.git] / OpenCL / m07500_a0.cl

diff --git a/OpenCL/m07500_a0.cl b/OpenCL/m07500_a0.cl
index a3a0298..61eee77 100644 (file)
--- a/OpenCL/m07500_a0.cl
+++ b/OpenCL/m07500_a0.cl
@@ -5,6 +5,9 @@
 
 #define _KRB5PA_
 
+//shared mem too small
+//#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -18,6 +21,7 @@
 #include "OpenCL/common.c"
 #include "include/rp_kernel.h"
 #include "OpenCL/rp.c"
+#include "OpenCL/simd.c"
 
 typedef struct
 {
@@ -550,30 +554,27 @@ __kernel void m07500_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * modifier
    */
 
-  __local RC4_KEY rc4_keys[64];
-
   const u32 lid = get_local_id (0);
-  const u32 gid = get_global_id (0);
-
-  if (gid >= gid_max) return;
 
   /**
    * base
    */
 
-  u32 pw_buf0[4];
+  const u32 gid = get_global_id (0);
 
-  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];
+  if (gid >= gid_max) return;
 
+  u32 pw_buf0[4];
   u32 pw_buf1[4];
 
-  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];
+  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;
 
@@ -600,40 +601,23 @@ __kernel void m07500_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
   timestamp_ct[7] = krb5pa_bufs[salt_pos].timestamp[7];
 
   /**
-   * loop
+   * shared
    */
 
-  for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
-  {
-    u32 w0[4];
-
-    w0[0] = pw_buf0[0];
-    w0[1] = pw_buf0[1];
-    w0[2] = pw_buf0[2];
-    w0[3] = pw_buf0[3];
-
-    u32 w1[4];
-
-    w1[0] = pw_buf1[0];
-    w1[1] = pw_buf1[1];
-    w1[2] = pw_buf1[2];
-    w1[3] = pw_buf1[3];
-
-    u32 w2[4];
-
-    w2[0] = 0;
-    w2[1] = 0;
-    w2[2] = 0;
-    w2[3] = 0;
+  __local RC4_KEY rc4_keys[64];
 
-    u32 w3[4];
+  /**
+   * loop
+   */
 
-    w3[0] = 0;
-    w3[1] = 0;
-    w3[2] = 0;
-    w3[3] = 0;
+  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 u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
+    const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
     /**
      * kerberos
@@ -673,30 +657,27 @@ __kernel void m07500_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * modifier
    */
 
-  __local RC4_KEY rc4_keys[64];
-
   const u32 lid = get_local_id (0);
-  const u32 gid = get_global_id (0);
-
-  if (gid >= gid_max) return;
 
   /**
    * base
    */
 
-  u32 pw_buf0[4];
+  const u32 gid = get_global_id (0);
 
-  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];
+  if (gid >= gid_max) return;
 
+  u32 pw_buf0[4];
   u32 pw_buf1[4];
 
-  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];
+  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;
 
@@ -723,40 +704,23 @@ __kernel void m07500_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
   timestamp_ct[7] = krb5pa_bufs[salt_pos].timestamp[7];
 
   /**
-   * loop
+   * shared
    */
 
-  for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
-  {
-    u32 w0[4];
-
-    w0[0] = pw_buf0[0];
-    w0[1] = pw_buf0[1];
-    w0[2] = pw_buf0[2];
-    w0[3] = pw_buf0[3];
-
-    u32 w1[4];
-
-    w1[0] = pw_buf1[0];
-    w1[1] = pw_buf1[1];
-    w1[2] = pw_buf1[2];
-    w1[3] = pw_buf1[3];
-
-    u32 w2[4];
-
-    w2[0] = 0;
-    w2[1] = 0;
-    w2[2] = 0;
-    w2[3] = 0;
+  __local RC4_KEY rc4_keys[64];
 
-    u32 w3[4];
+  /**
+   * loop
+   */
 
-    w3[0] = 0;
-    w3[1] = 0;
-    w3[2] = 0;
-    w3[3] = 0;
+  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 u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
+    const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
     /**
      * kerberos