Cleanup -m 131xx kernels to latest standard

diff --git a/OpenCL/m13100_a0.cl b/OpenCL/m13100_a0.cl
index 9d55fbc..28d1e7d 100644 (file)
--- a/OpenCL/m13100_a0.cl
+++ b/OpenCL/m13100_a0.cl
@@ -8,6 +8,9 @@
 
 #define _KRB5TGS_
 
+//too much register pressure
+//#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -21,6 +24,7 @@
 #include "OpenCL/common.c"
 #include "include/rp_kernel.h"
 #include "OpenCL/rp.c"
+#include "OpenCL/simd.c"
 
 typedef struct
 {
@@ -784,8 +788,6 @@ __kernel void m13100_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);
 
@@ -811,6 +813,12 @@ __kernel void m13100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   const u32 pw_len = pws[gid].pw_len;
 
+  /**
+   * shared
+   */
+
+  __local RC4_KEY rc4_keys[64];
+
   /**
    * salt
    */
@@ -826,37 +834,14 @@ __kernel void m13100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    u32 w0[4];
+    u32x w0[4] = { 0 };
+    u32x w1[4] = { 0 };
+    u32x w2[4] = { 0 };
+    u32x w3[4] = { 0 };
 
-    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;
-
-    u32 w3[4];
-
-    w3[0] = 0;
-    w3[1] = 0;
-    w3[2] = 0;
-    w3[3] = 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
@@ -866,7 +851,7 @@ __kernel void m13100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     u32 K2[4];
 
-    kerb_prepare (w0, w1, pw_len, checksum, digest, K2);
+    kerb_prepare (w0, w1, out_len, checksum, digest, K2);
 
     u32 tmp[4];
 
@@ -898,8 +883,6 @@ __kernel void m13100_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);
 
@@ -925,6 +908,12 @@ __kernel void m13100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
   const u32 pw_len = pws[gid].pw_len;
 
+  /**
+   * shared
+   */
+
+  __local RC4_KEY rc4_keys[64];
+
   /**
    * salt
    */
@@ -940,37 +929,14 @@ __kernel void m13100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    u32 w0[4];
-
-    w0[0] = pw_buf0[0];
-    w0[1] = pw_buf0[1];
-    w0[2] = pw_buf0[2];
-    w0[3] = pw_buf0[3];
+    u32x w0[4] = { 0 };
+    u32x w1[4] = { 0 };
+    u32x w2[4] = { 0 };
+    u32x w3[4] = { 0 };
 
-    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;
-
-    u32 w3[4];
-
-    w3[0] = 0;
-    w3[1] = 0;
-    w3[2] = 0;
-    w3[3] = 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
@@ -980,7 +946,7 @@ __kernel void m13100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     u32 K2[4];
 
-    kerb_prepare (w0, w1, pw_len, checksum, digest, K2);
+    kerb_prepare (w0, w1, out_len, checksum, digest, K2);
 
     u32 tmp[4];
 

diff --git a/OpenCL/m13100_a1.cl b/OpenCL/m13100_a1.cl
index 93e7757..65b8517 100644 (file)
--- a/OpenCL/m13100_a1.cl
+++ b/OpenCL/m13100_a1.cl
@@ -8,6 +8,9 @@
 
 #define _KRB5TGS_
 
+//too much register pressure
+//#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -19,6 +22,7 @@
 #include "include/kernel_functions.c"
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
+#include "OpenCL/simd.c"
 
 typedef struct
 {
@@ -728,51 +732,35 @@ __kernel void m13100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * modifier
    */
 
-  __local RC4_KEY rc4_keys[64];
-
-  const u32 gid = get_global_id (0);
   const u32 lid = get_local_id (0);
 
-  if (gid >= gid_max) return;
-
   /**
    * 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];
+  const u32 gid = get_global_id (0);
 
-  wordl2[0] = 0;
-  wordl2[1] = 0;
-  wordl2[2] = 0;
-  wordl2[3] = 0;
+  if (gid >= gid_max) return;
 
-  u32 wordl3[4];
+  u32 pw_buf0[4];
+  u32 pw_buf1[4];
 
-  wordl3[0] = 0;
-  wordl3[1] = 0;
-  wordl3[2] = 0;
-  wordl3[3] = 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];
+  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_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);
-  }
+  /**
+   * shared
+   */
+
+  __local RC4_KEY rc4_keys[64];
 
   /**
    * salt
@@ -789,73 +777,65 @@ __kernel void m13100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
   {
-    const u32 pw_r_len = combs_buf[il_pos].pw_len;
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    const u32 pw_len = pw_l_len + pw_r_len;
+    const u32x 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];
-
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-
-    u32 wordr3[4];
+    /**
+     * concat password candidate
+     */
 
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    wordl0[0] = pw_buf0[0];
+    wordl0[1] = pw_buf0[1];
+    wordl0[2] = pw_buf0[2];
+    wordl0[3] = pw_buf0[3];
+    wordl1[0] = pw_buf1[0];
+    wordl1[1] = pw_buf1[1];
+    wordl1[2] = pw_buf1[2];
+    wordl1[3] = pw_buf1[3];
+
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 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);
+      switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+    }
+    else
+    {
+      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
     }
 
-    u32 w0[4];
+    u32x w0[4];
+    u32x w1[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
     w0[2] = wordl0[2] | wordr0[2];
     w0[3] = wordl0[3] | wordr0[3];
-
-    u32 w1[4];
-
     w1[0] = wordl1[0] | wordr1[0];
     w1[1] = wordl1[1] | wordr1[1];
     w1[2] = wordl1[2] | wordr1[2];
     w1[3] = wordl1[3] | wordr1[3];
 
-    u32 w2[4];
-
-    w2[0] = wordl2[0] | wordr2[0];
-    w2[1] = wordl2[1] | wordr2[1];
-    w2[2] = wordl2[2] | wordr2[2];
-    w2[3] = wordl2[3] | wordr2[3];
-
-    u32 w3[4];
-
-    w3[0] = wordl3[0] | wordr3[0];
-    w3[1] = wordl3[1] | wordr3[1];
-    w3[2] = 0;
-    w3[3] = 0;
-
     /**
      * kerberos
      */
@@ -896,51 +876,35 @@ __kernel void m13100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * modifier
    */
 
-  __local RC4_KEY rc4_keys[64];
-
-  const u32 gid = get_global_id (0);
   const u32 lid = get_local_id (0);
 
-  if (gid >= gid_max) return;
-
   /**
    * 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];
+  const u32 gid = get_global_id (0);
 
-  wordl2[0] = 0;
-  wordl2[1] = 0;
-  wordl2[2] = 0;
-  wordl2[3] = 0;
+  if (gid >= gid_max) return;
 
-  u32 wordl3[4];
+  u32 pw_buf0[4];
+  u32 pw_buf1[4];
 
-  wordl3[0] = 0;
-  wordl3[1] = 0;
-  wordl3[2] = 0;
-  wordl3[3] = 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];
+  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_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);
-  }
+  /**
+   * shared
+   */
+
+  __local RC4_KEY rc4_keys[64];
 
   /**
    * salt
@@ -957,73 +921,65 @@ __kernel void m13100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < il_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;
+    const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
 
-    u32 wordr0[4];
+    const u32x pw_len = pw_l_len + pw_r_len;
 
-    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];
-
-    wordr2[0] = 0;
-    wordr2[1] = 0;
-    wordr2[2] = 0;
-    wordr2[3] = 0;
-
-    u32 wordr3[4];
+    /**
+     * concat password candidate
+     */
 
-    wordr3[0] = 0;
-    wordr3[1] = 0;
-    wordr3[2] = 0;
-    wordr3[3] = 0;
+    u32x wordl0[4] = { 0 };
+    u32x wordl1[4] = { 0 };
+    u32x wordl2[4] = { 0 };
+    u32x wordl3[4] = { 0 };
+
+    wordl0[0] = pw_buf0[0];
+    wordl0[1] = pw_buf0[1];
+    wordl0[2] = pw_buf0[2];
+    wordl0[3] = pw_buf0[3];
+    wordl1[0] = pw_buf1[0];
+    wordl1[1] = pw_buf1[1];
+    wordl1[2] = pw_buf1[2];
+    wordl1[3] = pw_buf1[3];
+
+    u32x wordr0[4] = { 0 };
+    u32x wordr1[4] = { 0 };
+    u32x wordr2[4] = { 0 };
+    u32x wordr3[4] = { 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);
+      switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+    }
+    else
+    {
+      switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
     }
 
-    u32 w0[4];
+    u32x w0[4];
+    u32x w1[4];
 
     w0[0] = wordl0[0] | wordr0[0];
     w0[1] = wordl0[1] | wordr0[1];
     w0[2] = wordl0[2] | wordr0[2];
     w0[3] = wordl0[3] | wordr0[3];
-
-    u32 w1[4];
-
     w1[0] = wordl1[0] | wordr1[0];
     w1[1] = wordl1[1] | wordr1[1];
     w1[2] = wordl1[2] | wordr1[2];
     w1[3] = wordl1[3] | wordr1[3];
 
-    u32 w2[4];
-
-    w2[0] = wordl2[0] | wordr2[0];
-    w2[1] = wordl2[1] | wordr2[1];
-    w2[2] = wordl2[2] | wordr2[2];
-    w2[3] = wordl2[3] | wordr2[3];
-
-    u32 w3[4];
-
-    w3[0] = wordl3[0] | wordr3[0];
-    w3[1] = wordl3[1] | wordr3[1];
-    w3[2] = 0;
-    w3[3] = 0;
-
     /**
      * kerberos
      */

diff --git a/OpenCL/m13100_a3.cl b/OpenCL/m13100_a3.cl
index 1b4a6bf..7f7099c 100644 (file)
--- a/OpenCL/m13100_a3.cl
+++ b/OpenCL/m13100_a3.cl
@@ -8,6 +8,9 @@
 
 #define _KRB5TGS_
 
+//too much register pressure
+//#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -19,6 +22,7 @@
 #include "include/kernel_functions.c"
 #include "OpenCL/types_ocl.c"
 #include "OpenCL/common.c"
+#include "OpenCL/simd.c"
 
 typedef struct
 {
@@ -754,6 +758,10 @@ static void m13100 (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
 
     w0[0] = w0l | w0r;
 
+    /**
+     * kerberos
+     */
+
     u32 digest[4];
 
     u32 K2[4];