Converted to new SIMD: -m 3710 -a 0

diff --git a/OpenCL/m03710_a0.cl b/OpenCL/m03710_a0.cl
index b614f71..0dbf39f 100644 (file)
--- a/OpenCL/m03710_a0.cl
+++ b/OpenCL/m03710_a0.cl
@@ -5,6 +5,8 @@
 
 #define _MD5_
 
+#define NEW_SIMD_CODE
+
 #include "include/constants.h"
 #include "include/kernel_vendor.h"
 
@@ -18,9 +20,7 @@
 #include "OpenCL/common.c"
 #include "include/rp_kernel.h"
 #include "OpenCL/rp.c"
-
-#define COMPARE_S "OpenCL/check_single_comp4.c"
-#define COMPARE_M "OpenCL/check_multi_comp4.c"
+#include "OpenCL/simd.c"
 
 #if   VECT_SIZE == 1
 #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i)])
@@ -121,46 +121,23 @@ __kernel void m03710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < rules_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];
-
-    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;
+    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 u32 out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
     append_0x80_2x4 (w0, w1, out_len);
 
     w3[2] = out_len * 8;
 
-    u32 a = MD5M_A;
-    u32 b = MD5M_B;
-    u32 c = MD5M_C;
-    u32 d = MD5M_D;
+    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);
@@ -235,10 +212,10 @@ __kernel void m03710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     c += MD5M_C;
     d += MD5M_D;
 
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
+    u32x w0_t[4];
+    u32x w1_t[4];
+    u32x w2_t[4];
+    u32x w3_t[4];
 
     w0_t[0] = uint_to_hex_lower8 ((a >>  0) & 255) <<  0
             | uint_to_hex_lower8 ((a >>  8) & 255) << 16;
@@ -369,12 +346,7 @@ __kernel void m03710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
     MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
 
-    const u32 r0 = a;
-    const u32 r1 = d;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_M
+    COMPARE_M_SIMD (a, d, c, b);
   }
 }
 
@@ -487,46 +459,23 @@ __kernel void m03710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    * loop
    */
 
-  for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
+  for (u32 il_pos = 0; il_pos < rules_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];
-
-    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];
+    u32x w0[4] = { 0 };
+    u32x w1[4] = { 0 };
+    u32x w2[4] = { 0 };
+    u32x w3[4] = { 0 };
 
-    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 u32 out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
     append_0x80_2x4 (w0, w1, out_len);
 
     w3[2] = out_len * 8;
 
-    u32 a = MD5M_A;
-    u32 b = MD5M_B;
-    u32 c = MD5M_C;
-    u32 d = MD5M_D;
+    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);
@@ -601,10 +550,10 @@ __kernel void m03710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     c += MD5M_C;
     d += MD5M_D;
 
-    u32 w0_t[4];
-    u32 w1_t[4];
-    u32 w2_t[4];
-    u32 w3_t[4];
+    u32x w0_t[4];
+    u32x w1_t[4];
+    u32x w2_t[4];
+    u32x w3_t[4];
 
     w0_t[0] = uint_to_hex_lower8 ((a >>  0) & 255) <<  0
             | uint_to_hex_lower8 ((a >>  8) & 255) << 16;
@@ -735,12 +684,7 @@ __kernel void m03710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
     MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
     MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
 
-    const u32 r0 = a;
-    const u32 r1 = d;
-    const u32 r2 = c;
-    const u32 r3 = b;
-
-    #include COMPARE_S
+    COMPARE_S_SIMD (a, d, c, b);
   }
 }