Fix -m 1500 multihash and benchmark

[hashcat.git] / OpenCL / m03100_a0.cl

diff --git a/OpenCL/m03100_a0.cl b/OpenCL/m03100_a0.cl
index 9b539df..9220eb7 100644 (file)
--- a/OpenCL/m03100_a0.cl
+++ b/OpenCL/m03100_a0.cl
@@ -372,7 +372,7 @@ __constant u32 c_skb[8][64] =
 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3], (S)[(n)][(i).s4], (S)[(n)][(i).s5], (S)[(n)][(i).s6], (S)[(n)][(i).s7], (S)[(n)][(i).s8], (S)[(n)][(i).s9], (S)[(n)][(i).sa], (S)[(n)][(i).sb], (S)[(n)][(i).sc], (S)[(n)][(i).sd], (S)[(n)][(i).se], (S)[(n)][(i).sf])
 #endif
 
-static void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
+void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 (*s_SPtrans)[64])
 {
   u32x tt;
 
@@ -424,7 +424,7 @@ static void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[1
   iv[1] = r;
 }
 
-static void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
+void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 (*s_skb)[64])
 {
   u32x tt;
 
@@ -541,18 +541,16 @@ __kernel void m03100_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   u32 pw_buf0[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];
-  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;
 
@@ -561,14 +559,12 @@ __kernel void m03100_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   u32 salt_buf0[4];
+  u32 salt_buf1[4];
 
   salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
   salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
   salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
   salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-
-  u32 salt_buf1[4];
-
   salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
   salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
   salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
@@ -577,7 +573,7 @@ __kernel void m03100_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 
   /**
-   * loop
+   * main
    */
 
   for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
@@ -595,63 +591,26 @@ __kernel void m03100_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
      * prepend salt
      */
 
-    u32x w0_t[4];
-
-    w0_t[0] = w0[0];
-    w0_t[1] = w0[1];
-    w0_t[2] = w0[2];
-    w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
-    w1_t[0] = w1[0];
-    w1_t[1] = w1[1];
-    w1_t[2] = w1[2];
-    w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
-    w2_t[0] = w2[0];
-    w2_t[1] = w2[1];
-    w2_t[2] = w2[2];
-    w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
-    w3_t[0] = w3[0];
-    w3_t[1] = w3[1];
-    w3_t[2] = w3[2];
-    w3_t[3] = w3[3];
-
-    switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
-    w0_t[0] |= salt_buf0[0];
-    w0_t[1] |= salt_buf0[1];
-    w0_t[2] |= salt_buf0[2];
-    w0_t[3] |= salt_buf0[3];
-    w1_t[0] |= salt_buf1[0];
-    w1_t[1] |= salt_buf1[1];
-    w1_t[2] |= salt_buf1[2];
-    w1_t[3] |= salt_buf1[3];
+    switch_buffer_by_offset_le (w0, w1, w2, w3, salt_len);
 
     u32x dst[16];
 
-    dst[ 0] = w0_t[0];
-    dst[ 1] = w0_t[1];
-    dst[ 2] = w0_t[2];
-    dst[ 3] = w0_t[3];
-    dst[ 4] = w1_t[0];
-    dst[ 5] = w1_t[1];
-    dst[ 6] = w1_t[2];
-    dst[ 7] = w1_t[3];
-    dst[ 8] = w2_t[0];
-    dst[ 9] = w2_t[1];
-    dst[10] = w2_t[2];
-    dst[11] = w2_t[3];
-    dst[12] = 0;
-    dst[13] = 0;
-    dst[14] = 0;
-    dst[15] = 0;
+    dst[ 0] = w0[0] | salt_buf0[0];
+    dst[ 1] = w0[1] | salt_buf0[1];
+    dst[ 2] = w0[2] | salt_buf0[2];
+    dst[ 3] = w0[3] | salt_buf0[3];
+    dst[ 4] = w1[0] | salt_buf1[0];
+    dst[ 5] = w1[1] | salt_buf1[1];
+    dst[ 6] = w1[2] | salt_buf1[2];
+    dst[ 7] = w1[3] | salt_buf1[3];
+    dst[ 8] = w2[0];
+    dst[ 9] = w2[1];
+    dst[10] = w2[2];
+    dst[11] = w2[3];
+    dst[12] = w3[0];
+    dst[13] = w3[1];
+    dst[14] = w3[2];
+    dst[15] = w3[3];
 
     /**
      * precompute key1 since key is static: 0x0123456789abcdef
@@ -744,12 +703,9 @@ __kernel void m03100_m04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
      * cmp
      */
 
-    const u32x r0 = iv[0];
-    const u32x r1 = iv[1];
-    const u32x r2 = 0;
-    const u32x r3 = 0;
+    u32x z = 0;
 
-    COMPARE_M_SIMD (r0, r1, r2, r3);
+    COMPARE_M_SIMD (iv[0], iv[1], z, z);
   }
 }
 
@@ -808,18 +764,16 @@ __kernel void m03100_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   u32 pw_buf0[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];
-  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;
 
@@ -828,14 +782,12 @@ __kernel void m03100_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
    */
 
   u32 salt_buf0[4];
+  u32 salt_buf1[4];
 
   salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
   salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
   salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
   salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
-
-  u32 salt_buf1[4];
-
   salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
   salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
   salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
@@ -851,12 +803,12 @@ __kernel void m03100_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
   {
     digests_buf[digests_offset].digest_buf[DGST_R0],
     digests_buf[digests_offset].digest_buf[DGST_R1],
-    digests_buf[digests_offset].digest_buf[DGST_R2],
-    digests_buf[digests_offset].digest_buf[DGST_R3]
+    0,
+    0
   };
 
   /**
-   * loop
+   * main
    */
 
   for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
@@ -874,63 +826,26 @@ __kernel void m03100_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
      * prepend salt
      */
 
-    u32x w0_t[4];
-
-    w0_t[0] = w0[0];
-    w0_t[1] = w0[1];
-    w0_t[2] = w0[2];
-    w0_t[3] = w0[3];
-
-    u32x w1_t[4];
-
-    w1_t[0] = w1[0];
-    w1_t[1] = w1[1];
-    w1_t[2] = w1[2];
-    w1_t[3] = w1[3];
-
-    u32x w2_t[4];
-
-    w2_t[0] = w2[0];
-    w2_t[1] = w2[1];
-    w2_t[2] = w2[2];
-    w2_t[3] = w2[3];
-
-    u32x w3_t[4];
-
-    w3_t[0] = w3[0];
-    w3_t[1] = w3[1];
-    w3_t[2] = w3[2];
-    w3_t[3] = w3[3];
-
-    switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
-    w0_t[0] |= salt_buf0[0];
-    w0_t[1] |= salt_buf0[1];
-    w0_t[2] |= salt_buf0[2];
-    w0_t[3] |= salt_buf0[3];
-    w1_t[0] |= salt_buf1[0];
-    w1_t[1] |= salt_buf1[1];
-    w1_t[2] |= salt_buf1[2];
-    w1_t[3] |= salt_buf1[3];
+    switch_buffer_by_offset_le (w0, w1, w2, w3, salt_len);
 
     u32x dst[16];
 
-    dst[ 0] = w0_t[0];
-    dst[ 1] = w0_t[1];
-    dst[ 2] = w0_t[2];
-    dst[ 3] = w0_t[3];
-    dst[ 4] = w1_t[0];
-    dst[ 5] = w1_t[1];
-    dst[ 6] = w1_t[2];
-    dst[ 7] = w1_t[3];
-    dst[ 8] = w2_t[0];
-    dst[ 9] = w2_t[1];
-    dst[10] = w2_t[2];
-    dst[11] = w2_t[3];
-    dst[12] = 0;
-    dst[13] = 0;
-    dst[14] = 0;
-    dst[15] = 0;
+    dst[ 0] = w0[0] | salt_buf0[0];
+    dst[ 1] = w0[1] | salt_buf0[1];
+    dst[ 2] = w0[2] | salt_buf0[2];
+    dst[ 3] = w0[3] | salt_buf0[3];
+    dst[ 4] = w1[0] | salt_buf1[0];
+    dst[ 5] = w1[1] | salt_buf1[1];
+    dst[ 6] = w1[2] | salt_buf1[2];
+    dst[ 7] = w1[3] | salt_buf1[3];
+    dst[ 8] = w2[0];
+    dst[ 9] = w2[1];
+    dst[10] = w2[2];
+    dst[11] = w2[3];
+    dst[12] = w3[0];
+    dst[13] = w3[1];
+    dst[14] = w3[2];
+    dst[15] = w3[3];
 
     /**
      * precompute key1 since key is static: 0x0123456789abcdef
@@ -1023,12 +938,9 @@ __kernel void m03100_s04 (__global pw_t *pws, __global kernel_rule_t *  rules_bu
      * cmp
      */
 
-    const u32x r0 = iv[0];
-    const u32x r1 = iv[1];
-    const u32x r2 = 0;
-    const u32x r3 = 0;
+    u32x z = 0;
 
-    COMPARE_M_SIMD (r0, r1, r2, r3);
+    COMPARE_S_SIMD (iv[0], iv[1], z, z);
   }
 }