Prepare for a more dynamic #pragma unroll use

[hashcat.git] / OpenCL / m08600_a0.cl

diff --git a/OpenCL/m08600_a0.cl b/OpenCL/m08600_a0.cl
index 90ca64c..c4a043c 100644 (file)
--- a/OpenCL/m08600_a0.cl
+++ b/OpenCL/m08600_a0.cl
@@ -74,7 +74,7 @@ __constant u32 lotus_magic_table[256] =
 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
 #endif
 
-static void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
+void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
 {
   u32x p = 0;
 
@@ -82,7 +82,9 @@ static void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
   {
     u32 s = 48;
 
-    #pragma unroll 12
+    #ifdef _unroll
+    #pragma unroll
+    #endif
     for (int j = 0; j < 12; j++)
     {
       u32x tmp_in = in[j];
@@ -98,13 +100,15 @@ static void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
   }
 }
 
-static void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
+void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
 {
   u32x t = out[3] >> 24;
 
   u32x c;
 
-  #pragma unroll 4
+  #ifdef _unroll
+  #pragma unroll
+  #endif
   for (int i = 0; i < 4; i++)
   {
     t ^= (in[i] >>  0) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c <<  0; t = ((out[i] >>  0) & 0xff);
@@ -114,7 +118,7 @@ static void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lo
   }
 }
 
-static void pad (u32x w[4], const u32 len)
+void pad (u32x w[4], const u32 len)
 {
   const u32 val = 16 - len;
 
@@ -193,7 +197,7 @@ static void pad (u32x w[4], const u32 len)
   }
 }
 
-static void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
+void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
 {
   u32x x[12];
 
@@ -218,14 +222,14 @@ static void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_l
   state[3] = x[3];
 }
 
-static void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
+void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
 {
   mdtransform_norecalc (state, block, s_lotus_magic_table);
 
   lotus_transform_password (block, checksum, s_lotus_magic_table);
 }
 
-static void domino_big_md (const u32x saved_key[16], const u32x size, u32x state[4], __local u32 *s_lotus_magic_table)
+void domino_big_md (const u32x saved_key[16], const u32x size, u32x state[4], __local u32 *s_lotus_magic_table)
 {
   u32x checksum[4];
 
@@ -276,14 +280,12 @@ __kernel void m08600_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    */
 
   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];
-
-  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];
@@ -304,6 +306,10 @@ __kernel void m08600_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    /**
+     * domino
+     */
+
     u32x w[16];
 
     w[ 0] = w0[0];
@@ -395,14 +401,12 @@ __kernel void m08600_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
    */
 
   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];
-
-  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];
@@ -435,6 +439,10 @@ __kernel void m08600_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
 
     const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
 
+    /**
+     * domino
+     */
+
     u32x w[16];
 
     w[ 0] = w0[0];