Fix more append_* functions in kernels

OpenCL/gpu_aes256_amd.c

@@ -784,14 +784,14 @@ static void aes256_set_encrypt_key (u32 *ks, const u32 *ukey)
 {
   u32 ukey_s[8];
 
-  ukey_s[0] = swap_workaround (ukey[0]);
-  ukey_s[1] = swap_workaround (ukey[1]);
-  ukey_s[2] = swap_workaround (ukey[2]);
-  ukey_s[3] = swap_workaround (ukey[3]);
-  ukey_s[4] = swap_workaround (ukey[4]);
-  ukey_s[5] = swap_workaround (ukey[5]);
-  ukey_s[6] = swap_workaround (ukey[6]);
-  ukey_s[7] = swap_workaround (ukey[7]);
+  ukey_s[0] = swap32 (ukey[0]);
+  ukey_s[1] = swap32 (ukey[1]);
+  ukey_s[2] = swap32 (ukey[2]);
+  ukey_s[3] = swap32 (ukey[3]);
+  ukey_s[4] = swap32 (ukey[4]);
+  ukey_s[5] = swap32 (ukey[5]);
+  ukey_s[6] = swap32 (ukey[6]);
+  ukey_s[7] = swap32 (ukey[7]);
 
   aes256_ExpandKey (ks, ukey_s);
 }
@@ -800,14 +800,14 @@ static void aes256_set_decrypt_key (u32 *ks, const u32 *ukey)
 {
   u32 ukey_s[8];
 
-  ukey_s[0] = swap_workaround (ukey[0]);
-  ukey_s[1] = swap_workaround (ukey[1]);
-  ukey_s[2] = swap_workaround (ukey[2]);
-  ukey_s[3] = swap_workaround (ukey[3]);
-  ukey_s[4] = swap_workaround (ukey[4]);
-  ukey_s[5] = swap_workaround (ukey[5]);
-  ukey_s[6] = swap_workaround (ukey[6]);
-  ukey_s[7] = swap_workaround (ukey[7]);
+  ukey_s[0] = swap32 (ukey[0]);
+  ukey_s[1] = swap32 (ukey[1]);
+  ukey_s[2] = swap32 (ukey[2]);
+  ukey_s[3] = swap32 (ukey[3]);
+  ukey_s[4] = swap32 (ukey[4]);
+  ukey_s[5] = swap32 (ukey[5]);
+  ukey_s[6] = swap32 (ukey[6]);
+  ukey_s[7] = swap32 (ukey[7]);
 
   aes256_ExpandKey (ks, ukey_s);
 
@@ -818,10 +818,10 @@ static void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out)
 {
   u32 in_s[4];
 
-  in_s[0] = swap_workaround (in[0]);
-  in_s[1] = swap_workaround (in[1]);
-  in_s[2] = swap_workaround (in[2]);
-  in_s[3] = swap_workaround (in[3]);
+  in_s[0] = swap32 (in[0]);
+  in_s[1] = swap32 (in[1]);
+  in_s[2] = swap32 (in[2]);
+  in_s[3] = swap32 (in[3]);
 
   u32 s0 = in_s[0] ^ ks[0];
   u32 s1 = in_s[1] ^ ks[1];
@@ -910,20 +910,20 @@ static void aes256_decrypt (const u32 *ks, const u32 *in, u32 *out)
          ^ (td4[(t0 >>  0) & 0xff] & 0x000000ff)
          ^ ks[59];
 
-  out[0] = swap_workaround (out[0]);
-  out[1] = swap_workaround (out[1]);
-  out[2] = swap_workaround (out[2]);
-  out[3] = swap_workaround (out[3]);
+  out[0] = swap32 (out[0]);
+  out[1] = swap32 (out[1]);
+  out[2] = swap32 (out[2]);
+  out[3] = swap32 (out[3]);
 }
 
 static void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out)
 {
   u32 in_s[4];
 
-  in_s[0] = swap_workaround (in[0]);
-  in_s[1] = swap_workaround (in[1]);
-  in_s[2] = swap_workaround (in[2]);
-  in_s[3] = swap_workaround (in[3]);
+  in_s[0] = swap32 (in[0]);
+  in_s[1] = swap32 (in[1]);
+  in_s[2] = swap32 (in[2]);
+  in_s[3] = swap32 (in[3]);
 
   u32 s0 = in_s[0] ^ ks[0];
   u32 s1 = in_s[1] ^ ks[1];
@@ -1012,10 +1012,10 @@ static void aes256_encrypt (const u32 *ks, const u32 *in, u32 *out)
          ^ (te4[(t2 >>  0) & 0xff] & 0x000000ff)
          ^ ks[59];
 
-  out[0] = swap_workaround (out[0]);
-  out[1] = swap_workaround (out[1]);
-  out[2] = swap_workaround (out[2]);
-  out[3] = swap_workaround (out[3]);
+  out[0] = swap32 (out[0]);
+  out[1] = swap32 (out[1]);
+  out[2] = swap32 (out[2]);
+  out[3] = swap32 (out[3]);
 }
 
 static void aes256_decrypt_xts (const u32 *ukey1, const u32 *ukey2, const u32 *in, u32 *out)

OpenCL/m00100_a0.cl

@@ -96,20 +96,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00100_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 
@@ -325,20 +325,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00100_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 

OpenCL/m00100_a1.cl

@@ -150,20 +150,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00100_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 
@@ -435,20 +435,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00100_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 

OpenCL/m00110_a0.cl

@@ -172,20 +172,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00110_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 
@@ -477,20 +477,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00110_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 

OpenCL/m00110_a1.cl

@@ -204,20 +204,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00110_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 
@@ -543,20 +543,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00110_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 

OpenCL/m00110_a3.cl

@@ -63,22 +63,22 @@ static void m00110m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global g
 
   switch_buffer_by_offset (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
-  w[ 0] |= swap_workaround (salt_buf0[0]);
-  w[ 1] |= swap_workaround (salt_buf0[1]);
-  w[ 2] |= swap_workaround (salt_buf0[2]);
-  w[ 3] |= swap_workaround (salt_buf0[3]);
-  w[ 4] |= swap_workaround (salt_buf1[0]);
-  w[ 5] |= swap_workaround (salt_buf1[1]);
-  w[ 6] |= swap_workaround (salt_buf1[2]);
-  w[ 7] |= swap_workaround (salt_buf1[3]);
-  w[ 8] |= swap_workaround (salt_buf2[0]);
-  w[ 9] |= swap_workaround (salt_buf2[1]);
-  w[10] |= swap_workaround (salt_buf2[2]);
-  w[11] |= swap_workaround (salt_buf2[3]);
-  w[12] |= swap_workaround (salt_buf3[0]);
-  w[13] |= swap_workaround (salt_buf3[1]);
-  w[14] |= swap_workaround (salt_buf3[2]);
-  w[15] |= swap_workaround (salt_buf3[3]);
+  w[ 0] |= swap32 (salt_buf0[0]);
+  w[ 1] |= swap32 (salt_buf0[1]);
+  w[ 2] |= swap32 (salt_buf0[2]);
+  w[ 3] |= swap32 (salt_buf0[3]);
+  w[ 4] |= swap32 (salt_buf1[0]);
+  w[ 5] |= swap32 (salt_buf1[1]);
+  w[ 6] |= swap32 (salt_buf1[2]);
+  w[ 7] |= swap32 (salt_buf1[3]);
+  w[ 8] |= swap32 (salt_buf2[0]);
+  w[ 9] |= swap32 (salt_buf2[1]);
+  w[10] |= swap32 (salt_buf2[2]);
+  w[11] |= swap32 (salt_buf2[3]);
+  w[12] |= swap32 (salt_buf3[0]);
+  w[13] |= swap32 (salt_buf3[1]);
+  w[14] |= swap32 (salt_buf3[2]);
+  w[15] |= swap32 (salt_buf3[3]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 

OpenCL/m00120_a0.cl

@@ -151,22 +151,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00120_m04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;
@@ -435,22 +435,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00120_s04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;

OpenCL/m00120_a1.cl

@@ -197,22 +197,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00120_m04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;
@@ -529,22 +529,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00120_s04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;

OpenCL/m00120_a3.cl

@@ -169,22 +169,22 @@ static void m00120m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   u32 w2_t[4];
   u32 w3_t[4];
 
-  w0_t[0] = swap_workaround (w0[0]);
-  w0_t[1] = swap_workaround (w0[1]);
-  w0_t[2] = swap_workaround (w0[2]);
-  w0_t[3] = swap_workaround (w0[3]);
-  w1_t[0] = swap_workaround (w1[0]);
-  w1_t[1] = swap_workaround (w1[1]);
-  w1_t[2] = swap_workaround (w1[2]);
-  w1_t[3] = swap_workaround (w1[3]);
-  w2_t[0] = swap_workaround (w2[0]);
-  w2_t[1] = swap_workaround (w2[1]);
-  w2_t[2] = swap_workaround (w2[2]);
-  w2_t[3] = swap_workaround (w2[3]);
-  w3_t[0] = swap_workaround (w3[0]);
-  w3_t[1] = swap_workaround (w3[1]);
-  w3_t[2] = swap_workaround (w3[2]);
-  w3_t[3] = swap_workaround (w3[3]);
+  w0_t[0] = swap32 (w0[0]);
+  w0_t[1] = swap32 (w0[1]);
+  w0_t[2] = swap32 (w0[2]);
+  w0_t[3] = swap32 (w0[3]);
+  w1_t[0] = swap32 (w1[0]);
+  w1_t[1] = swap32 (w1[1]);
+  w1_t[2] = swap32 (w1[2]);
+  w1_t[3] = swap32 (w1[3]);
+  w2_t[0] = swap32 (w2[0]);
+  w2_t[1] = swap32 (w2[1]);
+  w2_t[2] = swap32 (w2[2]);
+  w2_t[3] = swap32 (w2[3]);
+  w3_t[0] = swap32 (w3[0]);
+  w3_t[1] = swap32 (w3[1]);
+  w3_t[2] = swap32 (w3[2]);
+  w3_t[3] = swap32 (w3[3]);
 
   switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -205,22 +205,22 @@ static void m00120m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   w3_t[2] |= salt_buf3[2];
   w3_t[3] |= salt_buf3[3];
 
-  w0_t[0] = swap_workaround (w0_t[0]);
-  w0_t[1] = swap_workaround (w0_t[1]);
-  w0_t[2] = swap_workaround (w0_t[2]);
-  w0_t[3] = swap_workaround (w0_t[3]);
-  w1_t[0] = swap_workaround (w1_t[0]);
-  w1_t[1] = swap_workaround (w1_t[1]);
-  w1_t[2] = swap_workaround (w1_t[2]);
-  w1_t[3] = swap_workaround (w1_t[3]);
-  w2_t[0] = swap_workaround (w2_t[0]);
-  w2_t[1] = swap_workaround (w2_t[1]);
-  w2_t[2] = swap_workaround (w2_t[2]);
-  w2_t[3] = swap_workaround (w2_t[3]);
-  w3_t[0] = swap_workaround (w3_t[0]);
-  w3_t[1] = swap_workaround (w3_t[1]);
-  w3_t[2] = swap_workaround (w3_t[2]);
-  w3_t[3] = swap_workaround (w3_t[3]);
+  w0_t[0] = swap32 (w0_t[0]);
+  w0_t[1] = swap32 (w0_t[1]);
+  w0_t[2] = swap32 (w0_t[2]);
+  w0_t[3] = swap32 (w0_t[3]);
+  w1_t[0] = swap32 (w1_t[0]);
+  w1_t[1] = swap32 (w1_t[1]);
+  w1_t[2] = swap32 (w1_t[2]);
+  w1_t[3] = swap32 (w1_t[3]);
+  w2_t[0] = swap32 (w2_t[0]);
+  w2_t[1] = swap32 (w2_t[1]);
+  w2_t[2] = swap32 (w2_t[2]);
+  w2_t[3] = swap32 (w2_t[3]);
+  w3_t[0] = swap32 (w3_t[0]);
+  w3_t[1] = swap32 (w3_t[1]);
+  w3_t[2] = swap32 (w3_t[2]);
+  w3_t[3] = swap32 (w3_t[3]);
 
   /**
    * loop
@@ -464,22 +464,22 @@ static void m00120s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   u32 w2_t[4];
   u32 w3_t[4];
 
-  w0_t[0] = swap_workaround (w0[0]);
-  w0_t[1] = swap_workaround (w0[1]);
-  w0_t[2] = swap_workaround (w0[2]);
-  w0_t[3] = swap_workaround (w0[3]);
-  w1_t[0] = swap_workaround (w1[0]);
-  w1_t[1] = swap_workaround (w1[1]);
-  w1_t[2] = swap_workaround (w1[2]);
-  w1_t[3] = swap_workaround (w1[3]);
-  w2_t[0] = swap_workaround (w2[0]);
-  w2_t[1] = swap_workaround (w2[1]);
-  w2_t[2] = swap_workaround (w2[2]);
-  w2_t[3] = swap_workaround (w2[3]);
-  w3_t[0] = swap_workaround (w3[0]);
-  w3_t[1] = swap_workaround (w3[1]);
-  w3_t[2] = swap_workaround (w3[2]);
-  w3_t[3] = swap_workaround (w3[3]);
+  w0_t[0] = swap32 (w0[0]);
+  w0_t[1] = swap32 (w0[1]);
+  w0_t[2] = swap32 (w0[2]);
+  w0_t[3] = swap32 (w0[3]);
+  w1_t[0] = swap32 (w1[0]);
+  w1_t[1] = swap32 (w1[1]);
+  w1_t[2] = swap32 (w1[2]);
+  w1_t[3] = swap32 (w1[3]);
+  w2_t[0] = swap32 (w2[0]);
+  w2_t[1] = swap32 (w2[1]);
+  w2_t[2] = swap32 (w2[2]);
+  w2_t[3] = swap32 (w2[3]);
+  w3_t[0] = swap32 (w3[0]);
+  w3_t[1] = swap32 (w3[1]);
+  w3_t[2] = swap32 (w3[2]);
+  w3_t[3] = swap32 (w3[3]);
 
   switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -500,22 +500,22 @@ static void m00120s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   w3_t[2] |= salt_buf3[2];
   w3_t[3] |= salt_buf3[3];
 
-  w0_t[0] = swap_workaround (w0_t[0]);
-  w0_t[1] = swap_workaround (w0_t[1]);
-  w0_t[2] = swap_workaround (w0_t[2]);
-  w0_t[3] = swap_workaround (w0_t[3]);
-  w1_t[0] = swap_workaround (w1_t[0]);
-  w1_t[1] = swap_workaround (w1_t[1]);
-  w1_t[2] = swap_workaround (w1_t[2]);
-  w1_t[3] = swap_workaround (w1_t[3]);
-  w2_t[0] = swap_workaround (w2_t[0]);
-  w2_t[1] = swap_workaround (w2_t[1]);
-  w2_t[2] = swap_workaround (w2_t[2]);
-  w2_t[3] = swap_workaround (w2_t[3]);
-  w3_t[0] = swap_workaround (w3_t[0]);
-  w3_t[1] = swap_workaround (w3_t[1]);
-  w3_t[2] = swap_workaround (w3_t[2]);
-  w3_t[3] = swap_workaround (w3_t[3]);
+  w0_t[0] = swap32 (w0_t[0]);
+  w0_t[1] = swap32 (w0_t[1]);
+  w0_t[2] = swap32 (w0_t[2]);
+  w0_t[3] = swap32 (w0_t[3]);
+  w1_t[0] = swap32 (w1_t[0]);
+  w1_t[1] = swap32 (w1_t[1]);
+  w1_t[2] = swap32 (w1_t[2]);
+  w1_t[3] = swap32 (w1_t[3]);
+  w2_t[0] = swap32 (w2_t[0]);
+  w2_t[1] = swap32 (w2_t[1]);
+  w2_t[2] = swap32 (w2_t[2]);
+  w2_t[3] = swap32 (w2_t[3]);
+  w3_t[0] = swap32 (w3_t[0]);
+  w3_t[1] = swap32 (w3_t[1]);
+  w3_t[2] = swap32 (w3_t[2]);
+  w3_t[3] = swap32 (w3_t[3]);
 
   /**
    * loop

OpenCL/m00130_a0.cl

@@ -181,22 +181,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00130_m04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;
@@ -493,22 +493,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00130_s04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;

OpenCL/m00130_a1.cl

@@ -227,22 +227,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00130_m04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;
@@ -587,22 +587,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00130_s04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;

OpenCL/m00130_a3.cl

@@ -63,22 +63,22 @@ static void m00130m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global g
 
   switch_buffer_by_offset (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
-  w[ 0] |= swap_workaround (salt_buf0[0]);
-  w[ 1] |= swap_workaround (salt_buf0[1]);
-  w[ 2] |= swap_workaround (salt_buf0[2]);
-  w[ 3] |= swap_workaround (salt_buf0[3]);
-  w[ 4] |= swap_workaround (salt_buf1[0]);
-  w[ 5] |= swap_workaround (salt_buf1[1]);
-  w[ 6] |= swap_workaround (salt_buf1[2]);
-  w[ 7] |= swap_workaround (salt_buf1[3]);
-  w[ 8] |= swap_workaround (salt_buf2[0]);
-  w[ 9] |= swap_workaround (salt_buf2[1]);
-  w[10] |= swap_workaround (salt_buf2[2]);
-  w[11] |= swap_workaround (salt_buf2[3]);
-  w[12] |= swap_workaround (salt_buf3[0]);
-  w[13] |= swap_workaround (salt_buf3[1]);
-  w[14] |= swap_workaround (salt_buf3[2]);
-  w[15] |= swap_workaround (salt_buf3[3]);
+  w[ 0] |= swap32 (salt_buf0[0]);
+  w[ 1] |= swap32 (salt_buf0[1]);
+  w[ 2] |= swap32 (salt_buf0[2]);
+  w[ 3] |= swap32 (salt_buf0[3]);
+  w[ 4] |= swap32 (salt_buf1[0]);
+  w[ 5] |= swap32 (salt_buf1[1]);
+  w[ 6] |= swap32 (salt_buf1[2]);
+  w[ 7] |= swap32 (salt_buf1[3]);
+  w[ 8] |= swap32 (salt_buf2[0]);
+  w[ 9] |= swap32 (salt_buf2[1]);
+  w[10] |= swap32 (salt_buf2[2]);
+  w[11] |= swap32 (salt_buf2[3]);
+  w[12] |= swap32 (salt_buf3[0]);
+  w[13] |= swap32 (salt_buf3[1]);
+  w[14] |= swap32 (salt_buf3[2]);
+  w[15] |= swap32 (salt_buf3[3]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 

OpenCL/m00140_a0.cl

@@ -145,22 +145,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00140_m04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;
@@ -421,22 +421,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00140_s04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;

OpenCL/m00140_a1.cl

@@ -191,22 +191,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00140_m04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;
@@ -515,22 +515,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00140_s04 (__glo
      * sha1
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u32 a = SHA1M_A;
     u32 b = SHA1M_B;

OpenCL/m00140_a3.cl

@@ -169,22 +169,22 @@ static void m00140m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   u32 w2_t[4];
   u32 w3_t[4];
 
-  w0_t[0] = swap_workaround (w0[0]);
-  w0_t[1] = swap_workaround (w0[1]);
-  w0_t[2] = swap_workaround (w0[2]);
-  w0_t[3] = swap_workaround (w0[3]);
-  w1_t[0] = swap_workaround (w1[0]);
-  w1_t[1] = swap_workaround (w1[1]);
-  w1_t[2] = swap_workaround (w1[2]);
-  w1_t[3] = swap_workaround (w1[3]);
-  w2_t[0] = swap_workaround (w2[0]);
-  w2_t[1] = swap_workaround (w2[1]);
-  w2_t[2] = swap_workaround (w2[2]);
-  w2_t[3] = swap_workaround (w2[3]);
-  w3_t[0] = swap_workaround (w3[0]);
-  w3_t[1] = swap_workaround (w3[1]);
-  w3_t[2] = swap_workaround (w3[2]);
-  w3_t[3] = swap_workaround (w3[3]);
+  w0_t[0] = swap32 (w0[0]);
+  w0_t[1] = swap32 (w0[1]);
+  w0_t[2] = swap32 (w0[2]);
+  w0_t[3] = swap32 (w0[3]);
+  w1_t[0] = swap32 (w1[0]);
+  w1_t[1] = swap32 (w1[1]);
+  w1_t[2] = swap32 (w1[2]);
+  w1_t[3] = swap32 (w1[3]);
+  w2_t[0] = swap32 (w2[0]);
+  w2_t[1] = swap32 (w2[1]);
+  w2_t[2] = swap32 (w2[2]);
+  w2_t[3] = swap32 (w2[3]);
+  w3_t[0] = swap32 (w3[0]);
+  w3_t[1] = swap32 (w3[1]);
+  w3_t[2] = swap32 (w3[2]);
+  w3_t[3] = swap32 (w3[3]);
 
   switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -205,22 +205,22 @@ static void m00140m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   w3_t[2] |= salt_buf3[2];
   w3_t[3] |= salt_buf3[3];
 
-  w0_t[0] = swap_workaround (w0_t[0]);
-  w0_t[1] = swap_workaround (w0_t[1]);
-  w0_t[2] = swap_workaround (w0_t[2]);
-  w0_t[3] = swap_workaround (w0_t[3]);
-  w1_t[0] = swap_workaround (w1_t[0]);
-  w1_t[1] = swap_workaround (w1_t[1]);
-  w1_t[2] = swap_workaround (w1_t[2]);
-  w1_t[3] = swap_workaround (w1_t[3]);
-  w2_t[0] = swap_workaround (w2_t[0]);
-  w2_t[1] = swap_workaround (w2_t[1]);
-  w2_t[2] = swap_workaround (w2_t[2]);
-  w2_t[3] = swap_workaround (w2_t[3]);
-  w3_t[0] = swap_workaround (w3_t[0]);
-  w3_t[1] = swap_workaround (w3_t[1]);
-  w3_t[2] = swap_workaround (w3_t[2]);
-  w3_t[3] = swap_workaround (w3_t[3]);
+  w0_t[0] = swap32 (w0_t[0]);
+  w0_t[1] = swap32 (w0_t[1]);
+  w0_t[2] = swap32 (w0_t[2]);
+  w0_t[3] = swap32 (w0_t[3]);
+  w1_t[0] = swap32 (w1_t[0]);
+  w1_t[1] = swap32 (w1_t[1]);
+  w1_t[2] = swap32 (w1_t[2]);
+  w1_t[3] = swap32 (w1_t[3]);
+  w2_t[0] = swap32 (w2_t[0]);
+  w2_t[1] = swap32 (w2_t[1]);
+  w2_t[2] = swap32 (w2_t[2]);
+  w2_t[3] = swap32 (w2_t[3]);
+  w3_t[0] = swap32 (w3_t[0]);
+  w3_t[1] = swap32 (w3_t[1]);
+  w3_t[2] = swap32 (w3_t[2]);
+  w3_t[3] = swap32 (w3_t[3]);
 
   /**
    * loop
@@ -464,22 +464,22 @@ static void m00140s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   u32 w2_t[4];
   u32 w3_t[4];
 
-  w0_t[0] = swap_workaround (w0[0]);
-  w0_t[1] = swap_workaround (w0[1]);
-  w0_t[2] = swap_workaround (w0[2]);
-  w0_t[3] = swap_workaround (w0[3]);
-  w1_t[0] = swap_workaround (w1[0]);
-  w1_t[1] = swap_workaround (w1[1]);
-  w1_t[2] = swap_workaround (w1[2]);
-  w1_t[3] = swap_workaround (w1[3]);
-  w2_t[0] = swap_workaround (w2[0]);
-  w2_t[1] = swap_workaround (w2[1]);
-  w2_t[2] = swap_workaround (w2[2]);
-  w2_t[3] = swap_workaround (w2[3]);
-  w3_t[0] = swap_workaround (w3[0]);
-  w3_t[1] = swap_workaround (w3[1]);
-  w3_t[2] = swap_workaround (w3[2]);
-  w3_t[3] = swap_workaround (w3[3]);
+  w0_t[0] = swap32 (w0[0]);
+  w0_t[1] = swap32 (w0[1]);
+  w0_t[2] = swap32 (w0[2]);
+  w0_t[3] = swap32 (w0[3]);
+  w1_t[0] = swap32 (w1[0]);
+  w1_t[1] = swap32 (w1[1]);
+  w1_t[2] = swap32 (w1[2]);
+  w1_t[3] = swap32 (w1[3]);
+  w2_t[0] = swap32 (w2[0]);
+  w2_t[1] = swap32 (w2[1]);
+  w2_t[2] = swap32 (w2[2]);
+  w2_t[3] = swap32 (w2[3]);
+  w3_t[0] = swap32 (w3[0]);
+  w3_t[1] = swap32 (w3[1]);
+  w3_t[2] = swap32 (w3[2]);
+  w3_t[3] = swap32 (w3[3]);
 
   switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -500,22 +500,22 @@ static void m00140s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
   w3_t[2] |= salt_buf3[2];
   w3_t[3] |= salt_buf3[3];
 
-  w0_t[0] = swap_workaround (w0_t[0]);
-  w0_t[1] = swap_workaround (w0_t[1]);
-  w0_t[2] = swap_workaround (w0_t[2]);
-  w0_t[3] = swap_workaround (w0_t[3]);
-  w1_t[0] = swap_workaround (w1_t[0]);
-  w1_t[1] = swap_workaround (w1_t[1]);
-  w1_t[2] = swap_workaround (w1_t[2]);
-  w1_t[3] = swap_workaround (w1_t[3]);
-  w2_t[0] = swap_workaround (w2_t[0]);
-  w2_t[1] = swap_workaround (w2_t[1]);
-  w2_t[2] = swap_workaround (w2_t[2]);
-  w2_t[3] = swap_workaround (w2_t[3]);
-  w3_t[0] = swap_workaround (w3_t[0]);
-  w3_t[1] = swap_workaround (w3_t[1]);
-  w3_t[2] = swap_workaround (w3_t[2]);
-  w3_t[3] = swap_workaround (w3_t[3]);
+  w0_t[0] = swap32 (w0_t[0]);
+  w0_t[1] = swap32 (w0_t[1]);
+  w0_t[2] = swap32 (w0_t[2]);
+  w0_t[3] = swap32 (w0_t[3]);
+  w1_t[0] = swap32 (w1_t[0]);
+  w1_t[1] = swap32 (w1_t[1]);
+  w1_t[2] = swap32 (w1_t[2]);
+  w1_t[3] = swap32 (w1_t[3]);
+  w2_t[0] = swap32 (w2_t[0]);
+  w2_t[1] = swap32 (w2_t[1]);
+  w2_t[2] = swap32 (w2_t[2]);
+  w2_t[3] = swap32 (w2_t[3]);
+  w3_t[0] = swap32 (w3_t[0]);
+  w3_t[1] = swap32 (w3_t[1]);
+  w3_t[2] = swap32 (w3_t[2]);
+  w3_t[3] = swap32 (w3_t[3]);
 
   /**
    * loop

OpenCL/m00150_a0.cl

@@ -333,17 +333,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00150_m04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -364,14 +364,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00150_m04 (__glo
 
     hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -508,17 +508,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00150_s04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -539,14 +539,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00150_s04 (__glo
 
     hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m00150_a1.cl

@@ -385,17 +385,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00150_m04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -416,14 +416,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00150_m04 (__glo
 
     hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -614,17 +614,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00150_s04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -645,14 +645,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00150_s04 (__glo
 
     hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m00150_a3.cl

@@ -315,14 +315,14 @@ static void m00150m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -435,14 +435,14 @@ static void m00150s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     hmac_sha1_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m00160_a0.cl

@@ -295,17 +295,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00160_m04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -364,14 +364,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00160_m04 (__glo
 
     append_0x80_2x4 (w0, w1, out_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -458,17 +458,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00160_s04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -539,14 +539,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00160_s04 (__glo
 
     append_0x80_2x4 (w0, w1, out_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m00160_a1.cl

@@ -312,17 +312,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00160_m04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -416,20 +416,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00160_m04 (__glo
 
     append_0x80_4x4 (w0, w1, w2, w3, pw_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = (64 + pw_len) * 8;
 
@@ -529,17 +529,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00160_s04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -645,20 +645,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00160_s04 (__glo
 
     append_0x80_4x4 (w0, w1, w2, w3, pw_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = (64 + pw_len) * 8;
 

OpenCL/m00160_a3.cl

@@ -270,17 +270,17 @@ static void m00160m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -376,17 +376,17 @@ static void m00160s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 

OpenCL/m00190_a0.cl

@@ -96,20 +96,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00190_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 
@@ -336,20 +336,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00190_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 

OpenCL/m00190_a1.cl

@@ -150,20 +150,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00190_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 
@@ -446,20 +446,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00190_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 

OpenCL/m00300_a0.cl

@@ -96,20 +96,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00300_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 
@@ -450,20 +450,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00300_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 

OpenCL/m00300_a1.cl

@@ -150,20 +150,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00300_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 
@@ -560,20 +560,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00300_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 

OpenCL/m01400_a0.cl

@@ -96,14 +96,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01400_m04 (__glo
      * SHA256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
     u32 w8_t = 0;
     u32 w9_t = 0;
     u32 wa_t = 0;
@@ -294,14 +294,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01400_s04 (__glo
      * SHA256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
     u32 w8_t = 0;
     u32 w9_t = 0;
     u32 wa_t = 0;

OpenCL/m01400_a1.cl

@@ -144,20 +144,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01400_m04 (__glo
      * SHA256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 
@@ -392,20 +392,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01400_s04 (__glo
      * SHA256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 

OpenCL/m01410_a0.cl

@@ -172,20 +172,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01410_m04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_salt_len * 8;
 
@@ -446,20 +446,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01410_s04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_salt_len * 8;
 

OpenCL/m01410_a1.cl

@@ -198,20 +198,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01410_m04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 
@@ -500,20 +500,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01410_s04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 

OpenCL/m01410_a3.cl

@@ -63,22 +63,22 @@ static void m01410m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global g
 
   switch_buffer_by_offset (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
-  w[ 0] |= swap_workaround (salt_buf0[0]);
-  w[ 1] |= swap_workaround (salt_buf0[1]);
-  w[ 2] |= swap_workaround (salt_buf0[2]);
-  w[ 3] |= swap_workaround (salt_buf0[3]);
-  w[ 4] |= swap_workaround (salt_buf1[0]);
-  w[ 5] |= swap_workaround (salt_buf1[1]);
-  w[ 6] |= swap_workaround (salt_buf1[2]);
-  w[ 7] |= swap_workaround (salt_buf1[3]);
-  w[ 8] |= swap_workaround (salt_buf2[0]);
-  w[ 9] |= swap_workaround (salt_buf2[1]);
-  w[10] |= swap_workaround (salt_buf2[2]);
-  w[11] |= swap_workaround (salt_buf2[3]);
-  w[12] |= swap_workaround (salt_buf3[0]);
-  w[13] |= swap_workaround (salt_buf3[1]);
-  w[14] |= swap_workaround (salt_buf3[2]);
-  w[15] |= swap_workaround (salt_buf3[3]);
+  w[ 0] |= swap32 (salt_buf0[0]);
+  w[ 1] |= swap32 (salt_buf0[1]);
+  w[ 2] |= swap32 (salt_buf0[2]);
+  w[ 3] |= swap32 (salt_buf0[3]);
+  w[ 4] |= swap32 (salt_buf1[0]);
+  w[ 5] |= swap32 (salt_buf1[1]);
+  w[ 6] |= swap32 (salt_buf1[2]);
+  w[ 7] |= swap32 (salt_buf1[3]);
+  w[ 8] |= swap32 (salt_buf2[0]);
+  w[ 9] |= swap32 (salt_buf2[1]);
+  w[10] |= swap32 (salt_buf2[2]);
+  w[11] |= swap32 (salt_buf2[3]);
+  w[12] |= swap32 (salt_buf3[0]);
+  w[13] |= swap32 (salt_buf3[1]);
+  w[14] |= swap32 (salt_buf3[2]);
+  w[15] |= swap32 (salt_buf3[3]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 

OpenCL/m01420_a0.cl

@@ -133,20 +133,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01420_m04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_salt_len * 8;
 
@@ -368,20 +368,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01420_s04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_salt_len * 8;
 

OpenCL/m01420_a1.cl

@@ -173,20 +173,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01420_m04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 
@@ -450,20 +450,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01420_s04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 

OpenCL/m01420_a3.cl

@@ -86,22 +86,22 @@ static void m01420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32 w2_t2[4];
     u32 w3_t2[4];
 
-    w0_t2[0] = swap_workaround (w0[0]);
-    w0_t2[1] = swap_workaround (w0[1]);
-    w0_t2[2] = swap_workaround (w0[2]);
-    w0_t2[3] = swap_workaround (w0[3]);
-    w1_t2[0] = swap_workaround (w1[0]);
-    w1_t2[1] = swap_workaround (w1[1]);
-    w1_t2[2] = swap_workaround (w1[2]);
-    w1_t2[3] = swap_workaround (w1[3]);
-    w2_t2[0] = swap_workaround (w2[0]);
-    w2_t2[1] = swap_workaround (w2[1]);
-    w2_t2[2] = swap_workaround (w2[2]);
-    w2_t2[3] = swap_workaround (w2[3]);
-    w3_t2[0] = swap_workaround (w3[0]);
-    w3_t2[1] = swap_workaround (w3[1]);
-    w3_t2[2] = swap_workaround (w3[2]);
-    w3_t2[3] = swap_workaround (w3[3]);
+    w0_t2[0] = swap32 (w0[0]);
+    w0_t2[1] = swap32 (w0[1]);
+    w0_t2[2] = swap32 (w0[2]);
+    w0_t2[3] = swap32 (w0[3]);
+    w1_t2[0] = swap32 (w1[0]);
+    w1_t2[1] = swap32 (w1[1]);
+    w1_t2[2] = swap32 (w1[2]);
+    w1_t2[3] = swap32 (w1[3]);
+    w2_t2[0] = swap32 (w2[0]);
+    w2_t2[1] = swap32 (w2[1]);
+    w2_t2[2] = swap32 (w2[2]);
+    w2_t2[3] = swap32 (w2[3]);
+    w3_t2[0] = swap32 (w3[0]);
+    w3_t2[1] = swap32 (w3[1]);
+    w3_t2[2] = swap32 (w3[2]);
+    w3_t2[3] = swap32 (w3[3]);
 
     switch_buffer_by_offset (w0_t2, w1_t2, w2_t2, w3_t2, salt_len);
 
@@ -126,20 +126,20 @@ static void m01420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 
@@ -308,22 +308,22 @@ static void m01420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32 w2_t2[4];
     u32 w3_t2[4];
 
-    w0_t2[0] = swap_workaround (w0[0]);
-    w0_t2[1] = swap_workaround (w0[1]);
-    w0_t2[2] = swap_workaround (w0[2]);
-    w0_t2[3] = swap_workaround (w0[3]);
-    w1_t2[0] = swap_workaround (w1[0]);
-    w1_t2[1] = swap_workaround (w1[1]);
-    w1_t2[2] = swap_workaround (w1[2]);
-    w1_t2[3] = swap_workaround (w1[3]);
-    w2_t2[0] = swap_workaround (w2[0]);
-    w2_t2[1] = swap_workaround (w2[1]);
-    w2_t2[2] = swap_workaround (w2[2]);
-    w2_t2[3] = swap_workaround (w2[3]);
-    w3_t2[0] = swap_workaround (w3[0]);
-    w3_t2[1] = swap_workaround (w3[1]);
-    w3_t2[2] = swap_workaround (w3[2]);
-    w3_t2[3] = swap_workaround (w3[3]);
+    w0_t2[0] = swap32 (w0[0]);
+    w0_t2[1] = swap32 (w0[1]);
+    w0_t2[2] = swap32 (w0[2]);
+    w0_t2[3] = swap32 (w0[3]);
+    w1_t2[0] = swap32 (w1[0]);
+    w1_t2[1] = swap32 (w1[1]);
+    w1_t2[2] = swap32 (w1[2]);
+    w1_t2[3] = swap32 (w1[3]);
+    w2_t2[0] = swap32 (w2[0]);
+    w2_t2[1] = swap32 (w2[1]);
+    w2_t2[2] = swap32 (w2[2]);
+    w2_t2[3] = swap32 (w2[3]);
+    w3_t2[0] = swap32 (w3[0]);
+    w3_t2[1] = swap32 (w3[1]);
+    w3_t2[2] = swap32 (w3[2]);
+    w3_t2[3] = swap32 (w3[3]);
 
     switch_buffer_by_offset (w0_t2, w1_t2, w2_t2, w3_t2, salt_len);
 
@@ -348,20 +348,20 @@ static void m01420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 

OpenCL/m01430_a0.cl

@@ -177,20 +177,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01430_m04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = out_salt_len * 8;
 
@@ -456,20 +456,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01430_s04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = out_salt_len * 8;
 

OpenCL/m01430_a1.cl

@@ -217,20 +217,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01430_m04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 
@@ -538,20 +538,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01430_s04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 

OpenCL/m01430_a3.cl

@@ -63,22 +63,22 @@ static void m01430m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global g
 
   switch_buffer_by_offset (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
-  w[ 0] |= swap_workaround (salt_buf0[0]);
-  w[ 1] |= swap_workaround (salt_buf0[1]);
-  w[ 2] |= swap_workaround (salt_buf0[2]);
-  w[ 3] |= swap_workaround (salt_buf0[3]);
-  w[ 4] |= swap_workaround (salt_buf1[0]);
-  w[ 5] |= swap_workaround (salt_buf1[1]);
-  w[ 6] |= swap_workaround (salt_buf1[2]);
-  w[ 7] |= swap_workaround (salt_buf1[3]);
-  w[ 8] |= swap_workaround (salt_buf2[0]);
-  w[ 9] |= swap_workaround (salt_buf2[1]);
-  w[10] |= swap_workaround (salt_buf2[2]);
-  w[11] |= swap_workaround (salt_buf2[3]);
-  w[12] |= swap_workaround (salt_buf3[0]);
-  w[13] |= swap_workaround (salt_buf3[1]);
-  w[14] |= swap_workaround (salt_buf3[2]);
-  w[15] |= swap_workaround (salt_buf3[3]);
+  w[ 0] |= swap32 (salt_buf0[0]);
+  w[ 1] |= swap32 (salt_buf0[1]);
+  w[ 2] |= swap32 (salt_buf0[2]);
+  w[ 3] |= swap32 (salt_buf0[3]);
+  w[ 4] |= swap32 (salt_buf1[0]);
+  w[ 5] |= swap32 (salt_buf1[1]);
+  w[ 6] |= swap32 (salt_buf1[2]);
+  w[ 7] |= swap32 (salt_buf1[3]);
+  w[ 8] |= swap32 (salt_buf2[0]);
+  w[ 9] |= swap32 (salt_buf2[1]);
+  w[10] |= swap32 (salt_buf2[2]);
+  w[11] |= swap32 (salt_buf2[3]);
+  w[12] |= swap32 (salt_buf3[0]);
+  w[13] |= swap32 (salt_buf3[1]);
+  w[14] |= swap32 (salt_buf3[2]);
+  w[15] |= swap32 (salt_buf3[3]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 

OpenCL/m01440_a0.cl

@@ -135,20 +135,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01440_m04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = out_salt_len * 8;
 
@@ -372,20 +372,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01440_s04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = out_salt_len * 8;
 

OpenCL/m01440_a1.cl

@@ -181,20 +181,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01440_m04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 
@@ -466,20 +466,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01440_s04 (__glo
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 

OpenCL/m01440_a3.cl

@@ -86,22 +86,22 @@ static void m01440m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32 w2_t2[4];
     u32 w3_t2[4];
 
-    w0_t2[0] = swap_workaround (w0[0]);
-    w0_t2[1] = swap_workaround (w0[1]);
-    w0_t2[2] = swap_workaround (w0[2]);
-    w0_t2[3] = swap_workaround (w0[3]);
-    w1_t2[0] = swap_workaround (w1[0]);
-    w1_t2[1] = swap_workaround (w1[1]);
-    w1_t2[2] = swap_workaround (w1[2]);
-    w1_t2[3] = swap_workaround (w1[3]);
-    w2_t2[0] = swap_workaround (w2[0]);
-    w2_t2[1] = swap_workaround (w2[1]);
-    w2_t2[2] = swap_workaround (w2[2]);
-    w2_t2[3] = swap_workaround (w2[3]);
-    w3_t2[0] = swap_workaround (w3[0]);
-    w3_t2[1] = swap_workaround (w3[1]);
-    w3_t2[2] = swap_workaround (w3[2]);
-    w3_t2[3] = swap_workaround (w3[3]);
+    w0_t2[0] = swap32 (w0[0]);
+    w0_t2[1] = swap32 (w0[1]);
+    w0_t2[2] = swap32 (w0[2]);
+    w0_t2[3] = swap32 (w0[3]);
+    w1_t2[0] = swap32 (w1[0]);
+    w1_t2[1] = swap32 (w1[1]);
+    w1_t2[2] = swap32 (w1[2]);
+    w1_t2[3] = swap32 (w1[3]);
+    w2_t2[0] = swap32 (w2[0]);
+    w2_t2[1] = swap32 (w2[1]);
+    w2_t2[2] = swap32 (w2[2]);
+    w2_t2[3] = swap32 (w2[3]);
+    w3_t2[0] = swap32 (w3[0]);
+    w3_t2[1] = swap32 (w3[1]);
+    w3_t2[2] = swap32 (w3[2]);
+    w3_t2[3] = swap32 (w3[3]);
 
     switch_buffer_by_offset (w0_t2, w1_t2, w2_t2, w3_t2, salt_len);
 
@@ -126,20 +126,20 @@ static void m01440m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 
@@ -308,22 +308,22 @@ static void m01440s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32 w2_t2[4];
     u32 w3_t2[4];
 
-    w0_t2[0] = swap_workaround (w0[0]);
-    w0_t2[1] = swap_workaround (w0[1]);
-    w0_t2[2] = swap_workaround (w0[2]);
-    w0_t2[3] = swap_workaround (w0[3]);
-    w1_t2[0] = swap_workaround (w1[0]);
-    w1_t2[1] = swap_workaround (w1[1]);
-    w1_t2[2] = swap_workaround (w1[2]);
-    w1_t2[3] = swap_workaround (w1[3]);
-    w2_t2[0] = swap_workaround (w2[0]);
-    w2_t2[1] = swap_workaround (w2[1]);
-    w2_t2[2] = swap_workaround (w2[2]);
-    w2_t2[3] = swap_workaround (w2[3]);
-    w3_t2[0] = swap_workaround (w3[0]);
-    w3_t2[1] = swap_workaround (w3[1]);
-    w3_t2[2] = swap_workaround (w3[2]);
-    w3_t2[3] = swap_workaround (w3[3]);
+    w0_t2[0] = swap32 (w0[0]);
+    w0_t2[1] = swap32 (w0[1]);
+    w0_t2[2] = swap32 (w0[2]);
+    w0_t2[3] = swap32 (w0[3]);
+    w1_t2[0] = swap32 (w1[0]);
+    w1_t2[1] = swap32 (w1[1]);
+    w1_t2[2] = swap32 (w1[2]);
+    w1_t2[3] = swap32 (w1[3]);
+    w2_t2[0] = swap32 (w2[0]);
+    w2_t2[1] = swap32 (w2[1]);
+    w2_t2[2] = swap32 (w2[2]);
+    w2_t2[3] = swap32 (w2[3]);
+    w3_t2[0] = swap32 (w3[0]);
+    w3_t2[1] = swap32 (w3[1]);
+    w3_t2[2] = swap32 (w3[2]);
+    w3_t2[3] = swap32 (w3[3]);
 
     switch_buffer_by_offset (w0_t2, w1_t2, w2_t2, w3_t2, salt_len);
 
@@ -348,20 +348,20 @@ static void m01440s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha256
      */
 
-    u32 w0_t = swap_workaround (w0_t2[0]);
-    u32 w1_t = swap_workaround (w0_t2[1]);
-    u32 w2_t = swap_workaround (w0_t2[2]);
-    u32 w3_t = swap_workaround (w0_t2[3]);
-    u32 w4_t = swap_workaround (w1_t2[0]);
-    u32 w5_t = swap_workaround (w1_t2[1]);
-    u32 w6_t = swap_workaround (w1_t2[2]);
-    u32 w7_t = swap_workaround (w1_t2[3]);
-    u32 w8_t = swap_workaround (w2_t2[0]);
-    u32 w9_t = swap_workaround (w2_t2[1]);
-    u32 wa_t = swap_workaround (w2_t2[2]);
-    u32 wb_t = swap_workaround (w2_t2[3]);
-    u32 wc_t = swap_workaround (w3_t2[0]);
-    u32 wd_t = swap_workaround (w3_t2[1]);
+    u32 w0_t = swap32 (w0_t2[0]);
+    u32 w1_t = swap32 (w0_t2[1]);
+    u32 w2_t = swap32 (w0_t2[2]);
+    u32 w3_t = swap32 (w0_t2[3]);
+    u32 w4_t = swap32 (w1_t2[0]);
+    u32 w5_t = swap32 (w1_t2[1]);
+    u32 w6_t = swap32 (w1_t2[2]);
+    u32 w7_t = swap32 (w1_t2[3]);
+    u32 w8_t = swap32 (w2_t2[0]);
+    u32 w9_t = swap32 (w2_t2[1]);
+    u32 wa_t = swap32 (w2_t2[2]);
+    u32 wb_t = swap32 (w2_t2[3]);
+    u32 wc_t = swap32 (w3_t2[0]);
+    u32 wd_t = swap32 (w3_t2[1]);
     u32 we_t = 0;
     u32 wf_t = pw_salt_len * 8;
 

OpenCL/m01450_a0.cl

@@ -323,17 +323,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01450_m04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -354,14 +354,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01450_m04 (__glo
 
     hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -498,17 +498,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01450_s04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -529,14 +529,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01450_s04 (__glo
 
     hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m01450_a1.cl

@@ -375,17 +375,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01450_m04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -406,14 +406,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01450_m04 (__glo
 
     hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -604,17 +604,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01450_s04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -635,14 +635,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01450_s04 (__glo
 
     hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m01450_a3.cl

@@ -305,14 +305,14 @@ static void m01450m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -425,14 +425,14 @@ static void m01450s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     hmac_sha256_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m01460_a0.cl

@@ -285,17 +285,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01460_m04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -354,14 +354,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01460_m04 (__glo
 
     append_0x80_2x4 (w0, w1, out_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -448,17 +448,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01460_s04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -529,14 +529,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01460_s04 (__glo
 
     append_0x80_2x4 (w0, w1, out_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m01460_a1.cl

@@ -302,17 +302,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01460_m04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -406,20 +406,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01460_m04 (__glo
 
     append_0x80_4x4 (w0, w1, w2, w3, pw_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = (64 + pw_len) * 8;
 
@@ -519,17 +519,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01460_s04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -635,20 +635,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01460_s04 (__glo
 
     append_0x80_4x4 (w0, w1, w2, w3, pw_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = (64 + pw_len) * 8;
 

OpenCL/m01460_a3.cl

@@ -260,17 +260,17 @@ static void m01460m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -366,17 +366,17 @@ static void m01460s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 

OpenCL/m01600.cl

@@ -17,13 +17,8 @@
 #include "types_ocl.c"
 #include "common.c"
 
-#ifdef  VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
+#define COMPARE_S "check_single_comp4.c"
+#define COMPARE_M "check_multi_comp4.c"
 
 #define md5apr1_magic0 0x72706124
 #define md5apr1_magic1 0x00002431

OpenCL/m01700_a0.cl

@@ -222,20 +222,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01700_m04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = out_len * 8;
 
@@ -361,20 +361,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01700_s04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = out_len * 8;
 

OpenCL/m01700_a1.cl

@@ -270,20 +270,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01700_m04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_len * 8;
 
@@ -459,20 +459,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01700_s04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_len * 8;
 

OpenCL/m01710_a0.cl

@@ -298,20 +298,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01710_m04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 
@@ -513,20 +513,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01710_s04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 

OpenCL/m01710_a1.cl

@@ -324,20 +324,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01710_m04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 
@@ -567,20 +567,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01710_s04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 

OpenCL/m01710_a3.cl

@@ -184,22 +184,22 @@ static void m01710m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global g
 
   switch_buffer_by_offset (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
-  w[ 0] |= swap_workaround (salt_buf0[0]);
-  w[ 1] |= swap_workaround (salt_buf0[1]);
-  w[ 2] |= swap_workaround (salt_buf0[2]);
-  w[ 3] |= swap_workaround (salt_buf0[3]);
-  w[ 4] |= swap_workaround (salt_buf1[0]);
-  w[ 5] |= swap_workaround (salt_buf1[1]);
-  w[ 6] |= swap_workaround (salt_buf1[2]);
-  w[ 7] |= swap_workaround (salt_buf1[3]);
-  w[ 8] |= swap_workaround (salt_buf2[0]);
-  w[ 9] |= swap_workaround (salt_buf2[1]);
-  w[10] |= swap_workaround (salt_buf2[2]);
-  w[11] |= swap_workaround (salt_buf2[3]);
-  w[12] |= swap_workaround (salt_buf3[0]);
-  w[13] |= swap_workaround (salt_buf3[1]);
-  w[14] |= swap_workaround (salt_buf3[2]);
-  w[15] |= swap_workaround (salt_buf3[3]);
+  w[ 0] |= swap32 (salt_buf0[0]);
+  w[ 1] |= swap32 (salt_buf0[1]);
+  w[ 2] |= swap32 (salt_buf0[2]);
+  w[ 3] |= swap32 (salt_buf0[3]);
+  w[ 4] |= swap32 (salt_buf1[0]);
+  w[ 5] |= swap32 (salt_buf1[1]);
+  w[ 6] |= swap32 (salt_buf1[2]);
+  w[ 7] |= swap32 (salt_buf1[3]);
+  w[ 8] |= swap32 (salt_buf2[0]);
+  w[ 9] |= swap32 (salt_buf2[1]);
+  w[10] |= swap32 (salt_buf2[2]);
+  w[11] |= swap32 (salt_buf2[3]);
+  w[12] |= swap32 (salt_buf3[0]);
+  w[13] |= swap32 (salt_buf3[1]);
+  w[14] |= swap32 (salt_buf3[2]);
+  w[15] |= swap32 (salt_buf3[3]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 

OpenCL/m01720_a0.cl

@@ -259,20 +259,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_m04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 
@@ -435,20 +435,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_s04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 

OpenCL/m01720_a1.cl

@@ -299,20 +299,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_m04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 
@@ -517,20 +517,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01720_s04 (__glo
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 

OpenCL/m01720_a3.cl

@@ -207,22 +207,22 @@ static void m01720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
-    w3_t[2] = swap_workaround (w3[2]);
-    w3_t[3] = swap_workaround (w3[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
+    w3_t[2] = swap32 (w3[2]);
+    w3_t[3] = swap32 (w3[3]);
 
     switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -247,22 +247,22 @@ static void m01720m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u64 digest[8];
 
@@ -365,22 +365,22 @@ static void m01720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
-    w3_t[2] = swap_workaround (w3[2]);
-    w3_t[3] = swap_workaround (w3[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
+    w3_t[2] = swap32 (w3[2]);
+    w3_t[3] = swap32 (w3[3]);
 
     switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -405,22 +405,22 @@ static void m01720s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u64 digest[8];
 

OpenCL/m01730_a0.cl

@@ -298,20 +298,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01730_m04 (__glo
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 
@@ -513,20 +513,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01730_s04 (__glo
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 

OpenCL/m01730_a1.cl

@@ -338,20 +338,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01730_m04 (__glo
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 
@@ -595,20 +595,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01730_s04 (__glo
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 

OpenCL/m01730_a3.cl

@@ -184,22 +184,22 @@ static void m01730m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global g
 
   switch_buffer_by_offset (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
 
-  w[ 0] |= swap_workaround (salt_buf0[0]);
-  w[ 1] |= swap_workaround (salt_buf0[1]);
-  w[ 2] |= swap_workaround (salt_buf0[2]);
-  w[ 3] |= swap_workaround (salt_buf0[3]);
-  w[ 4] |= swap_workaround (salt_buf1[0]);
-  w[ 5] |= swap_workaround (salt_buf1[1]);
-  w[ 6] |= swap_workaround (salt_buf1[2]);
-  w[ 7] |= swap_workaround (salt_buf1[3]);
-  w[ 8] |= swap_workaround (salt_buf2[0]);
-  w[ 9] |= swap_workaround (salt_buf2[1]);
-  w[10] |= swap_workaround (salt_buf2[2]);
-  w[11] |= swap_workaround (salt_buf2[3]);
-  w[12] |= swap_workaround (salt_buf3[0]);
-  w[13] |= swap_workaround (salt_buf3[1]);
-  w[14] |= swap_workaround (salt_buf3[2]);
-  w[15] |= swap_workaround (salt_buf3[3]);
+  w[ 0] |= swap32 (salt_buf0[0]);
+  w[ 1] |= swap32 (salt_buf0[1]);
+  w[ 2] |= swap32 (salt_buf0[2]);
+  w[ 3] |= swap32 (salt_buf0[3]);
+  w[ 4] |= swap32 (salt_buf1[0]);
+  w[ 5] |= swap32 (salt_buf1[1]);
+  w[ 6] |= swap32 (salt_buf1[2]);
+  w[ 7] |= swap32 (salt_buf1[3]);
+  w[ 8] |= swap32 (salt_buf2[0]);
+  w[ 9] |= swap32 (salt_buf2[1]);
+  w[10] |= swap32 (salt_buf2[2]);
+  w[11] |= swap32 (salt_buf2[3]);
+  w[12] |= swap32 (salt_buf3[0]);
+  w[13] |= swap32 (salt_buf3[1]);
+  w[14] |= swap32 (salt_buf3[2]);
+  w[15] |= swap32 (salt_buf3[3]);
 
   const u32 salt_len = salt_bufs[salt_pos].salt_len;
 

OpenCL/m01740_a0.cl

@@ -256,20 +256,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01740_m04 (__glo
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 
@@ -429,20 +429,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01740_s04 (__glo
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
     w3_t[2] = 0;
     w3_t[3] = out_salt_len * 8;
 

OpenCL/m01740_a1.cl

@@ -302,20 +302,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01740_m04 (__glo
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 
@@ -523,20 +523,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01740_s04 (__glo
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
     w3_t[2] = 0;
     w3_t[3] = pw_salt_len * 8;
 

OpenCL/m01740_a3.cl

@@ -207,22 +207,22 @@ static void m01740m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
-    w3_t[2] = swap_workaround (w3[2]);
-    w3_t[3] = swap_workaround (w3[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
+    w3_t[2] = swap32 (w3[2]);
+    w3_t[3] = swap32 (w3[3]);
 
     switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -247,22 +247,22 @@ static void m01740m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u64 digest[8];
 
@@ -365,22 +365,22 @@ static void m01740s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     u32 w2_t[4];
     u32 w3_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
-    w3_t[2] = swap_workaround (w3[2]);
-    w3_t[3] = swap_workaround (w3[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
+    w3_t[2] = swap32 (w3[2]);
+    w3_t[3] = swap32 (w3[3]);
 
     switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
 
@@ -405,22 +405,22 @@ static void m01740s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
      * sha512
      */
 
-    w0_t[0] = swap_workaround (w0_t[0]);
-    w0_t[1] = swap_workaround (w0_t[1]);
-    w0_t[2] = swap_workaround (w0_t[2]);
-    w0_t[3] = swap_workaround (w0_t[3]);
-    w1_t[0] = swap_workaround (w1_t[0]);
-    w1_t[1] = swap_workaround (w1_t[1]);
-    w1_t[2] = swap_workaround (w1_t[2]);
-    w1_t[3] = swap_workaround (w1_t[3]);
-    w2_t[0] = swap_workaround (w2_t[0]);
-    w2_t[1] = swap_workaround (w2_t[1]);
-    w2_t[2] = swap_workaround (w2_t[2]);
-    w2_t[3] = swap_workaround (w2_t[3]);
-    w3_t[0] = swap_workaround (w3_t[0]);
-    w3_t[1] = swap_workaround (w3_t[1]);
-    //w3_t[2] = swap_workaround (w3_t[2]);
-    //w3_t[3] = swap_workaround (w3_t[3]);
+    w0_t[0] = swap32 (w0_t[0]);
+    w0_t[1] = swap32 (w0_t[1]);
+    w0_t[2] = swap32 (w0_t[2]);
+    w0_t[3] = swap32 (w0_t[3]);
+    w1_t[0] = swap32 (w1_t[0]);
+    w1_t[1] = swap32 (w1_t[1]);
+    w1_t[2] = swap32 (w1_t[2]);
+    w1_t[3] = swap32 (w1_t[3]);
+    w2_t[0] = swap32 (w2_t[0]);
+    w2_t[1] = swap32 (w2_t[1]);
+    w2_t[2] = swap32 (w2_t[2]);
+    w2_t[3] = swap32 (w2_t[3]);
+    w3_t[0] = swap32 (w3_t[0]);
+    w3_t[1] = swap32 (w3_t[1]);
+    //w3_t[2] = swap32 (w3_t[2]);
+    //w3_t[3] = swap32 (w3_t[3]);
 
     u64 digest[8];
 

OpenCL/m01750_a0.cl

@@ -354,17 +354,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01750_m04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -385,14 +385,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01750_m04 (__glo
 
     hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -530,17 +530,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01750_s04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -561,14 +561,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01750_s04 (__glo
 
     hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m01750_a1.cl

@@ -406,17 +406,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01750_m04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -437,14 +437,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01750_m04 (__glo
 
     hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -636,17 +636,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01750_s04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -667,14 +667,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01750_s04 (__glo
 
     hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m01750_a3.cl

@@ -336,14 +336,14 @@ static void m01750m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -457,14 +457,14 @@ static void m01750s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
     hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
 
-    w0_t[0] = swap_workaround (salt_buf0[0]);
-    w0_t[1] = swap_workaround (salt_buf0[1]);
-    w0_t[2] = swap_workaround (salt_buf0[2]);
-    w0_t[3] = swap_workaround (salt_buf0[3]);
-    w1_t[0] = swap_workaround (salt_buf1[0]);
-    w1_t[1] = swap_workaround (salt_buf1[1]);
-    w1_t[2] = swap_workaround (salt_buf1[2]);
-    w1_t[3] = swap_workaround (salt_buf1[3]);
+    w0_t[0] = swap32 (salt_buf0[0]);
+    w0_t[1] = swap32 (salt_buf0[1]);
+    w0_t[2] = swap32 (salt_buf0[2]);
+    w0_t[3] = swap32 (salt_buf0[3]);
+    w1_t[0] = swap32 (salt_buf1[0]);
+    w1_t[1] = swap32 (salt_buf1[1]);
+    w1_t[2] = swap32 (salt_buf1[2]);
+    w1_t[3] = swap32 (salt_buf1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m01760_a0.cl

@@ -316,17 +316,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01760_m04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -385,14 +385,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01760_m04 (__glo
 
     append_0x80_2x4 (w0, w1, out_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;
@@ -480,17 +480,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01760_s04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -561,14 +561,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01760_s04 (__glo
 
     append_0x80_2x4 (w0, w1, out_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
     w2_t[0] = 0;
     w2_t[1] = 0;
     w2_t[2] = 0;

OpenCL/m01760_a1.cl

@@ -333,17 +333,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01760_m04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -437,20 +437,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01760_m04 (__glo
 
     append_0x80_4x4 (w0, w1, w2, w3, pw_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = (128 + pw_len) * 8;
 
@@ -551,17 +551,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01760_s04 (__glo
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -667,20 +667,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01760_s04 (__glo
 
     append_0x80_4x4 (w0, w1, w2, w3, pw_len);
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
-    w2_t[0] = swap_workaround (w2[0]);
-    w2_t[1] = swap_workaround (w2[1]);
-    w2_t[2] = swap_workaround (w2[2]);
-    w2_t[3] = swap_workaround (w2[3]);
-    w3_t[0] = swap_workaround (w3[0]);
-    w3_t[1] = swap_workaround (w3[1]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
+    w2_t[0] = swap32 (w2[0]);
+    w2_t[1] = swap32 (w2[1]);
+    w2_t[2] = swap32 (w2[2]);
+    w2_t[3] = swap32 (w2[3]);
+    w3_t[0] = swap32 (w3[0]);
+    w3_t[1] = swap32 (w3[1]);
     w3_t[2] = 0;
     w3_t[3] = (128 + pw_len) * 8;
 

OpenCL/m01760_a3.cl

@@ -291,17 +291,17 @@ static void m01760m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 
@@ -398,17 +398,17 @@ static void m01760s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 w0_t[4];
 
-  w0_t[0] = swap_workaround (salt_buf0[0]);
-  w0_t[1] = swap_workaround (salt_buf0[1]);
-  w0_t[2] = swap_workaround (salt_buf0[2]);
-  w0_t[3] = swap_workaround (salt_buf0[3]);
+  w0_t[0] = swap32 (salt_buf0[0]);
+  w0_t[1] = swap32 (salt_buf0[1]);
+  w0_t[2] = swap32 (salt_buf0[2]);
+  w0_t[3] = swap32 (salt_buf0[3]);
 
   u32 w1_t[4];
 
-  w1_t[0] = swap_workaround (salt_buf1[0]);
-  w1_t[1] = swap_workaround (salt_buf1[1]);
-  w1_t[2] = swap_workaround (salt_buf1[2]);
-  w1_t[3] = swap_workaround (salt_buf1[3]);
+  w1_t[0] = swap32 (salt_buf1[0]);
+  w1_t[1] = swap32 (salt_buf1[1]);
+  w1_t[2] = swap32 (salt_buf1[2]);
+  w1_t[3] = swap32 (salt_buf1[3]);
 
   u32 w2_t[4];
 

OpenCL/m01800.cl

@@ -330,13 +330,13 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01800_init (__gl
 
   u64 pw[2];
 
-  pw[0] = swap_workaround (hl32_to_64 (w0[1], w0[0]));
-  pw[1] = swap_workaround (hl32_to_64 (w0[3], w0[2]));
+  pw[0] = swap32 (hl32_to_64 (w0[1], w0[0]));
+  pw[1] = swap32 (hl32_to_64 (w0[3], w0[2]));
 
   u64 salt[2];
 
-  salt[0] = swap_workaround (hl32_to_64 (salt_buf[1], salt_buf[0]));
-  salt[1] = swap_workaround (hl32_to_64 (salt_buf[3], salt_buf[2]));
+  salt[0] = swap32 (hl32_to_64 (salt_buf[1], salt_buf[0]));
+  salt[1] = swap32 (hl32_to_64 (salt_buf[3], salt_buf[2]));
 
   /**
    * begin
@@ -609,8 +609,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m01800_comp (__gl
 
   const u32 lid = get_local_id (0);
 
-  const u64 a = swap_workaround (tmps[gid].l_alt_result[0]);
-  const u64 b = swap_workaround (tmps[gid].l_alt_result[1]);
+  const u64 a = swap32 (tmps[gid].l_alt_result[0]);
+  const u64 b = swap32 (tmps[gid].l_alt_result[1]);
 
   const u32 r0 = l32_from_64 (a);
   const u32 r1 = h32_from_64 (a);

OpenCL/m02100.cl

@@ -418,10 +418,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02100_init (__gl
    * pads
    */
 
-  w0[0] = swap_workaround (digest_md4[0]);
-  w0[1] = swap_workaround (digest_md4[1]);
-  w0[2] = swap_workaround (digest_md4[2]);
-  w0[3] = swap_workaround (digest_md4[3]);
+  w0[0] = swap32 (digest_md4[0]);
+  w0[1] = swap32 (digest_md4[1]);
+  w0[2] = swap32 (digest_md4[2]);
+  w0[3] = swap32 (digest_md4[3]);
   w1[0] = 0;
   w1[1] = 0;
   w1[2] = 0;
@@ -476,20 +476,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02100_init (__gl
   append_0x01_4x4 (w0, w1, w2, w3, salt_len + 3);
   append_0x80_4x4 (w0, w1, w2, w3, salt_len + 4);
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
 
   u32 digest[5];
 

OpenCL/m02500.cl

@@ -17,17 +17,8 @@
 #include "types_ocl.c"
 #include "common.c"
 
-#ifdef  VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE2
-#define COMPARE_M "check_multi_vect2_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
+#define COMPARE_S "check_single_comp4.c"
+#define COMPARE_M "check_multi_comp4.c"
 
 static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
 {
@@ -491,22 +482,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02500_init (__gl
    * pads
    */
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
-  w3[2] = swap_workaround (w3[2]);
-  w3[3] = swap_workaround (w3[3]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
+  w3[2] = swap32 (w3[2]);
+  w3[3] = swap32 (w3[3]);
 
   u32 ipad[5];
   u32 opad[5];
@@ -551,16 +542,16 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02500_init (__gl
 
     append_0x80_3x4 (w0, w1, w2, salt_len + 4);
 
-    w0[0] = swap_workaround (w0[0]);
-    w0[1] = swap_workaround (w0[1]);
-    w0[2] = swap_workaround (w0[2]);
-    w0[3] = swap_workaround (w0[3]);
-    w1[0] = swap_workaround (w1[0]);
-    w1[1] = swap_workaround (w1[1]);
-    w1[2] = swap_workaround (w1[2]);
-    w1[3] = swap_workaround (w1[3]);
-    w2[0] = swap_workaround (w2[0]);
-    w2[1] = swap_workaround (w2[1]);
+    w0[0] = swap32 (w0[0]);
+    w0[1] = swap32 (w0[1]);
+    w0[2] = swap32 (w0[2]);
+    w0[3] = swap32 (w0[3]);
+    w1[0] = swap32 (w1[0]);
+    w1[1] = swap32 (w1[1]);
+    w1[2] = swap32 (w1[2]);
+    w1[3] = swap32 (w1[3]);
+    w2[0] = swap32 (w2[0]);
+    w2[1] = swap32 (w2[1]);
     w2[2] = 0;
     w2[3] = 0;
     w3[0] = 0;
@@ -747,10 +738,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02500_comp (__gl
   hmac_sha1_run (w0, w1, w2, w3, ipad, opad, digest);
 
   {
-    w0[0] = swap_workaround (digest[0]);
-    w0[1] = swap_workaround (digest[1]);
-    w0[2] = swap_workaround (digest[2]);
-    w0[3] = swap_workaround (digest[3]);
+    w0[0] = swap32 (digest[0]);
+    w0[1] = swap32 (digest[1]);
+    w0[2] = swap32 (digest[2]);
+    w0[3] = swap32 (digest[3]);
     w1[0] = 0;
     w1[1] = 0;
     w1[2] = 0;

OpenCL/m03200.cl

@@ -404,24 +404,24 @@ __kernel void __attribute__((reqd_work_group_size (8, 1, 1))) m03200_init (__glo
 
   expand_key (E, w, pw_len);
 
-  E[ 0] = swap_workaround (E[ 0]);
-  E[ 1] = swap_workaround (E[ 1]);
-  E[ 2] = swap_workaround (E[ 2]);
-  E[ 3] = swap_workaround (E[ 3]);
-  E[ 4] = swap_workaround (E[ 4]);
-  E[ 5] = swap_workaround (E[ 5]);
-  E[ 6] = swap_workaround (E[ 6]);
-  E[ 7] = swap_workaround (E[ 7]);
-  E[ 8] = swap_workaround (E[ 8]);
-  E[ 9] = swap_workaround (E[ 9]);
-  E[10] = swap_workaround (E[10]);
-  E[11] = swap_workaround (E[11]);
-  E[12] = swap_workaround (E[12]);
-  E[13] = swap_workaround (E[13]);
-  E[14] = swap_workaround (E[14]);
-  E[15] = swap_workaround (E[15]);
-  E[16] = swap_workaround (E[16]);
-  E[17] = swap_workaround (E[17]);
+  E[ 0] = swap32 (E[ 0]);
+  E[ 1] = swap32 (E[ 1]);
+  E[ 2] = swap32 (E[ 2]);
+  E[ 3] = swap32 (E[ 3]);
+  E[ 4] = swap32 (E[ 4]);
+  E[ 5] = swap32 (E[ 5]);
+  E[ 6] = swap32 (E[ 6]);
+  E[ 7] = swap32 (E[ 7]);
+  E[ 8] = swap32 (E[ 8]);
+  E[ 9] = swap32 (E[ 9]);
+  E[10] = swap32 (E[10]);
+  E[11] = swap32 (E[11]);
+  E[12] = swap32 (E[12]);
+  E[13] = swap32 (E[13]);
+  E[14] = swap32 (E[14]);
+  E[15] = swap32 (E[15]);
+  E[16] = swap32 (E[16]);
+  E[17] = swap32 (E[17]);
 
   /**
    * salt
@@ -615,24 +615,24 @@ __kernel void __attribute__((reqd_work_group_size (8, 1, 1))) m03200_loop (__glo
 
   expand_key (E, w, pw_len);
 
-  E[ 0] = swap_workaround (E[ 0]);
-  E[ 1] = swap_workaround (E[ 1]);
-  E[ 2] = swap_workaround (E[ 2]);
-  E[ 3] = swap_workaround (E[ 3]);
-  E[ 4] = swap_workaround (E[ 4]);
-  E[ 5] = swap_workaround (E[ 5]);
-  E[ 6] = swap_workaround (E[ 6]);
-  E[ 7] = swap_workaround (E[ 7]);
-  E[ 8] = swap_workaround (E[ 8]);
-  E[ 9] = swap_workaround (E[ 9]);
-  E[10] = swap_workaround (E[10]);
-  E[11] = swap_workaround (E[11]);
-  E[12] = swap_workaround (E[12]);
-  E[13] = swap_workaround (E[13]);
-  E[14] = swap_workaround (E[14]);
-  E[15] = swap_workaround (E[15]);
-  E[16] = swap_workaround (E[16]);
-  E[17] = swap_workaround (E[17]);
+  E[ 0] = swap32 (E[ 0]);
+  E[ 1] = swap32 (E[ 1]);
+  E[ 2] = swap32 (E[ 2]);
+  E[ 3] = swap32 (E[ 3]);
+  E[ 4] = swap32 (E[ 4]);
+  E[ 5] = swap32 (E[ 5]);
+  E[ 6] = swap32 (E[ 6]);
+  E[ 7] = swap32 (E[ 7]);
+  E[ 8] = swap32 (E[ 8]);
+  E[ 9] = swap32 (E[ 9]);
+  E[10] = swap32 (E[10]);
+  E[11] = swap32 (E[11]);
+  E[12] = swap32 (E[12]);
+  E[13] = swap32 (E[13]);
+  E[14] = swap32 (E[14]);
+  E[15] = swap32 (E[15]);
+  E[16] = swap32 (E[16]);
+  E[17] = swap32 (E[17]);
 
   // load
 

OpenCL/m04400_a0.cl

@@ -141,20 +141,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04400_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 
@@ -507,20 +507,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04400_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 

OpenCL/m04400_a1.cl

@@ -195,20 +195,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04400_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 
@@ -617,20 +617,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04400_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 

OpenCL/m04500_a0.cl

@@ -141,20 +141,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04500_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 
@@ -543,20 +543,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04500_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = out_len * 8;
 

OpenCL/m04500_a1.cl

@@ -195,20 +195,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04500_m04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 
@@ -652,20 +652,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04500_s04 (__glo
      * sha1
      */
 
-    u32 w0_t = swap_workaround (w0[0]);
-    u32 w1_t = swap_workaround (w0[1]);
-    u32 w2_t = swap_workaround (w0[2]);
-    u32 w3_t = swap_workaround (w0[3]);
-    u32 w4_t = swap_workaround (w1[0]);
-    u32 w5_t = swap_workaround (w1[1]);
-    u32 w6_t = swap_workaround (w1[2]);
-    u32 w7_t = swap_workaround (w1[3]);
-    u32 w8_t = swap_workaround (w2[0]);
-    u32 w9_t = swap_workaround (w2[1]);
-    u32 wa_t = swap_workaround (w2[2]);
-    u32 wb_t = swap_workaround (w2[3]);
-    u32 wc_t = swap_workaround (w3[0]);
-    u32 wd_t = swap_workaround (w3[1]);
+    u32 w0_t = swap32 (w0[0]);
+    u32 w1_t = swap32 (w0[1]);
+    u32 w2_t = swap32 (w0[2]);
+    u32 w3_t = swap32 (w0[3]);
+    u32 w4_t = swap32 (w1[0]);
+    u32 w5_t = swap32 (w1[1]);
+    u32 w6_t = swap32 (w1[2]);
+    u32 w7_t = swap32 (w1[3]);
+    u32 w8_t = swap32 (w2[0]);
+    u32 w9_t = swap32 (w2[1]);
+    u32 wa_t = swap32 (w2[2]);
+    u32 wb_t = swap32 (w2[3]);
+    u32 wc_t = swap32 (w3[0]);
+    u32 wd_t = swap32 (w3[1]);
     u32 we_t = 0;
     u32 wf_t = pw_len * 8;
 

OpenCL/m04900_a0.cl

@@ -202,20 +202,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04900_m04 (__glo
 
     append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
 
-    u32 w0 = swap_workaround (w0_t[0]);
-    u32 w1 = swap_workaround (w0_t[1]);
-    u32 w2 = swap_workaround (w0_t[2]);
-    u32 w3 = swap_workaround (w0_t[3]);
-    u32 w4 = swap_workaround (w1_t[0]);
-    u32 w5 = swap_workaround (w1_t[1]);
-    u32 w6 = swap_workaround (w1_t[2]);
-    u32 w7 = swap_workaround (w1_t[3]);
-    u32 w8 = swap_workaround (w2_t[0]);
-    u32 w9 = swap_workaround (w2_t[1]);
-    u32 wa = swap_workaround (w2_t[2]);
-    u32 wb = swap_workaround (w2_t[3]);
-    u32 wc = swap_workaround (w3_t[0]);
-    u32 wd = swap_workaround (w3_t[1]);
+    u32 w0 = swap32 (w0_t[0]);
+    u32 w1 = swap32 (w0_t[1]);
+    u32 w2 = swap32 (w0_t[2]);
+    u32 w3 = swap32 (w0_t[3]);
+    u32 w4 = swap32 (w1_t[0]);
+    u32 w5 = swap32 (w1_t[1]);
+    u32 w6 = swap32 (w1_t[2]);
+    u32 w7 = swap32 (w1_t[3]);
+    u32 w8 = swap32 (w2_t[0]);
+    u32 w9 = swap32 (w2_t[1]);
+    u32 wa = swap32 (w2_t[2]);
+    u32 wb = swap32 (w2_t[3]);
+    u32 wc = swap32 (w3_t[0]);
+    u32 wd = swap32 (w3_t[1]);
     u32 we = 0;
     u32 wf = pw_salt_len * 8;
 
@@ -540,20 +540,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04900_s04 (__glo
 
     append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
 
-    u32 w0 = swap_workaround (w0_t[0]);
-    u32 w1 = swap_workaround (w0_t[1]);
-    u32 w2 = swap_workaround (w0_t[2]);
-    u32 w3 = swap_workaround (w0_t[3]);
-    u32 w4 = swap_workaround (w1_t[0]);
-    u32 w5 = swap_workaround (w1_t[1]);
-    u32 w6 = swap_workaround (w1_t[2]);
-    u32 w7 = swap_workaround (w1_t[3]);
-    u32 w8 = swap_workaround (w2_t[0]);
-    u32 w9 = swap_workaround (w2_t[1]);
-    u32 wa = swap_workaround (w2_t[2]);
-    u32 wb = swap_workaround (w2_t[3]);
-    u32 wc = swap_workaround (w3_t[0]);
-    u32 wd = swap_workaround (w3_t[1]);
+    u32 w0 = swap32 (w0_t[0]);
+    u32 w1 = swap32 (w0_t[1]);
+    u32 w2 = swap32 (w0_t[2]);
+    u32 w3 = swap32 (w0_t[3]);
+    u32 w4 = swap32 (w1_t[0]);
+    u32 w5 = swap32 (w1_t[1]);
+    u32 w6 = swap32 (w1_t[2]);
+    u32 w7 = swap32 (w1_t[3]);
+    u32 w8 = swap32 (w2_t[0]);
+    u32 w9 = swap32 (w2_t[1]);
+    u32 wa = swap32 (w2_t[2]);
+    u32 wb = swap32 (w2_t[3]);
+    u32 wc = swap32 (w3_t[0]);
+    u32 wd = swap32 (w3_t[1]);
     u32 we = 0;
     u32 wf = pw_salt_len * 8;
 

OpenCL/m04900_a1.cl

@@ -254,20 +254,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04900_m04 (__glo
 
     append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
 
-    u32 w0 = swap_workaround (w0_t[0]);
-    u32 w1 = swap_workaround (w0_t[1]);
-    u32 w2 = swap_workaround (w0_t[2]);
-    u32 w3 = swap_workaround (w0_t[3]);
-    u32 w4 = swap_workaround (w1_t[0]);
-    u32 w5 = swap_workaround (w1_t[1]);
-    u32 w6 = swap_workaround (w1_t[2]);
-    u32 w7 = swap_workaround (w1_t[3]);
-    u32 w8 = swap_workaround (w2_t[0]);
-    u32 w9 = swap_workaround (w2_t[1]);
-    u32 wa = swap_workaround (w2_t[2]);
-    u32 wb = swap_workaround (w2_t[3]);
-    u32 wc = swap_workaround (w3_t[0]);
-    u32 wd = swap_workaround (w3_t[1]);
+    u32 w0 = swap32 (w0_t[0]);
+    u32 w1 = swap32 (w0_t[1]);
+    u32 w2 = swap32 (w0_t[2]);
+    u32 w3 = swap32 (w0_t[3]);
+    u32 w4 = swap32 (w1_t[0]);
+    u32 w5 = swap32 (w1_t[1]);
+    u32 w6 = swap32 (w1_t[2]);
+    u32 w7 = swap32 (w1_t[3]);
+    u32 w8 = swap32 (w2_t[0]);
+    u32 w9 = swap32 (w2_t[1]);
+    u32 wa = swap32 (w2_t[2]);
+    u32 wb = swap32 (w2_t[3]);
+    u32 wc = swap32 (w3_t[0]);
+    u32 wd = swap32 (w3_t[1]);
     u32 we = 0;
     u32 wf = pw_salt_len * 8;
 
@@ -646,20 +646,20 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04900_s04 (__glo
 
     append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
 
-    u32 w0 = swap_workaround (w0_t[0]);
-    u32 w1 = swap_workaround (w0_t[1]);
-    u32 w2 = swap_workaround (w0_t[2]);
-    u32 w3 = swap_workaround (w0_t[3]);
-    u32 w4 = swap_workaround (w1_t[0]);
-    u32 w5 = swap_workaround (w1_t[1]);
-    u32 w6 = swap_workaround (w1_t[2]);
-    u32 w7 = swap_workaround (w1_t[3]);
-    u32 w8 = swap_workaround (w2_t[0]);
-    u32 w9 = swap_workaround (w2_t[1]);
-    u32 wa = swap_workaround (w2_t[2]);
-    u32 wb = swap_workaround (w2_t[3]);
-    u32 wc = swap_workaround (w3_t[0]);
-    u32 wd = swap_workaround (w3_t[1]);
+    u32 w0 = swap32 (w0_t[0]);
+    u32 w1 = swap32 (w0_t[1]);
+    u32 w2 = swap32 (w0_t[2]);
+    u32 w3 = swap32 (w0_t[3]);
+    u32 w4 = swap32 (w1_t[0]);
+    u32 w5 = swap32 (w1_t[1]);
+    u32 w6 = swap32 (w1_t[2]);
+    u32 w7 = swap32 (w1_t[3]);
+    u32 w8 = swap32 (w2_t[0]);
+    u32 w9 = swap32 (w2_t[1]);
+    u32 wa = swap32 (w2_t[2]);
+    u32 wb = swap32 (w2_t[3]);
+    u32 wc = swap32 (w3_t[0]);
+    u32 wd = swap32 (w3_t[1]);
     u32 we = 0;
     u32 wf = pw_salt_len * 8;
 

OpenCL/m04900_a3.cl

@@ -183,21 +183,21 @@ static void m04900m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     w3_t[1] |= salt_buf3[1];
     w3_t[2] |= salt_buf3[2];
 
-    u32 w0 = swap_workaround (w0_t[0]);
-    u32 w1 = swap_workaround (w0_t[1]);
-    u32 w2 = swap_workaround (w0_t[2]);
-    u32 w3 = swap_workaround (w0_t[3]);
-    u32 w4 = swap_workaround (w1_t[0]);
-    u32 w5 = swap_workaround (w1_t[1]);
-    u32 w6 = swap_workaround (w1_t[2]);
-    u32 w7 = swap_workaround (w1_t[3]);
-    u32 w8 = swap_workaround (w2_t[0]);
-    u32 w9 = swap_workaround (w2_t[1]);
-    u32 wa = swap_workaround (w2_t[2]);
-    u32 wb = swap_workaround (w2_t[3]);
-    u32 wc = swap_workaround (w3_t[0]);
-    u32 wd = swap_workaround (w3_t[1]);
-    u32 we = swap_workaround (w3_t[2]);
+    u32 w0 = swap32 (w0_t[0]);
+    u32 w1 = swap32 (w0_t[1]);
+    u32 w2 = swap32 (w0_t[2]);
+    u32 w3 = swap32 (w0_t[3]);
+    u32 w4 = swap32 (w1_t[0]);
+    u32 w5 = swap32 (w1_t[1]);
+    u32 w6 = swap32 (w1_t[2]);
+    u32 w7 = swap32 (w1_t[3]);
+    u32 w8 = swap32 (w2_t[0]);
+    u32 w9 = swap32 (w2_t[1]);
+    u32 wa = swap32 (w2_t[2]);
+    u32 wb = swap32 (w2_t[3]);
+    u32 wc = swap32 (w3_t[0]);
+    u32 wd = swap32 (w3_t[1]);
+    u32 we = swap32 (w3_t[2]);
     u32 wf = pw_salt_len * 8;
 
     /**
@@ -496,21 +496,21 @@ static void m04900s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
     w3_t[1] |= salt_buf3[1];
     w3_t[2] |= salt_buf3[2];
 
-    u32 w0 = swap_workaround (w0_t[0]);
-    u32 w1 = swap_workaround (w0_t[1]);
-    u32 w2 = swap_workaround (w0_t[2]);
-    u32 w3 = swap_workaround (w0_t[3]);
-    u32 w4 = swap_workaround (w1_t[0]);
-    u32 w5 = swap_workaround (w1_t[1]);
-    u32 w6 = swap_workaround (w1_t[2]);
-    u32 w7 = swap_workaround (w1_t[3]);
-    u32 w8 = swap_workaround (w2_t[0]);
-    u32 w9 = swap_workaround (w2_t[1]);
-    u32 wa = swap_workaround (w2_t[2]);
-    u32 wb = swap_workaround (w2_t[3]);
-    u32 wc = swap_workaround (w3_t[0]);
-    u32 wd = swap_workaround (w3_t[1]);
-    u32 we = swap_workaround (w3_t[2]);
+    u32 w0 = swap32 (w0_t[0]);
+    u32 w1 = swap32 (w0_t[1]);
+    u32 w2 = swap32 (w0_t[2]);
+    u32 w3 = swap32 (w0_t[3]);
+    u32 w4 = swap32 (w1_t[0]);
+    u32 w5 = swap32 (w1_t[1]);
+    u32 w6 = swap32 (w1_t[2]);
+    u32 w7 = swap32 (w1_t[3]);
+    u32 w8 = swap32 (w2_t[0]);
+    u32 w9 = swap32 (w2_t[1]);
+    u32 wa = swap32 (w2_t[2]);
+    u32 wb = swap32 (w2_t[3]);
+    u32 wc = swap32 (w3_t[0]);
+    u32 wd = swap32 (w3_t[1]);
+    u32 we = swap32 (w3_t[2]);
     u32 wf = pw_salt_len * 8;
 
     /**

OpenCL/m05200.cl

@@ -17,17 +17,8 @@
 #include "types_ocl.c"
 #include "common.c"
 
-#ifdef  VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE2
-#define COMPARE_M "check_multi_vect2_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
+#define COMPARE_S "check_single_comp4.c"
+#define COMPARE_M "check_multi_comp4.c"
 
 __constant u32 k_sha256[64] =
 {
@@ -239,23 +230,23 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05200_init (__gl
    * init
    */
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
 
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
 
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
 
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
   w3[2] = 0;
   w3[3] = block_len * 8;
 

OpenCL/m05400_a0.cl

@@ -278,29 +278,29 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m04 (__glo
 
   u32 salt_buf0[4];
 
-  salt_buf0[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 0]);
-  salt_buf0[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 1]);
-  salt_buf0[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 2]);
-  salt_buf0[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 3]);
+  salt_buf0[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 0]);
+  salt_buf0[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 1]);
+  salt_buf0[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 2]);
+  salt_buf0[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 3]);
 
   u32 salt_buf1[4];
 
-  salt_buf1[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 4]);
-  salt_buf1[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 5]);
-  salt_buf1[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 6]);
-  salt_buf1[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 7]);
+  salt_buf1[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 4]);
+  salt_buf1[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 5]);
+  salt_buf1[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 6]);
+  salt_buf1[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 7]);
 
   u32 salt_buf2[4];
 
-  salt_buf2[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 8]);
-  salt_buf2[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 9]);
-  salt_buf2[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[10]);
-  salt_buf2[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[11]);
+  salt_buf2[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 8]);
+  salt_buf2[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 9]);
+  salt_buf2[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[10]);
+  salt_buf2[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[11]);
 
   u32 salt_buf3[4];
 
-  salt_buf3[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[12]);
-  salt_buf3[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[13]);
+  salt_buf3[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[12]);
+  salt_buf3[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[13]);
   salt_buf3[2] = 0;
   salt_buf3[3] = 0;
 
@@ -308,8 +308,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m04 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -357,17 +357,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -528,29 +528,29 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s04 (__glo
 
   u32 salt_buf0[4];
 
-  salt_buf0[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 0]);
-  salt_buf0[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 1]);
-  salt_buf0[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 2]);
-  salt_buf0[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 3]);
+  salt_buf0[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 0]);
+  salt_buf0[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 1]);
+  salt_buf0[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 2]);
+  salt_buf0[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 3]);
 
   u32 salt_buf1[4];
 
-  salt_buf1[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 4]);
-  salt_buf1[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 5]);
-  salt_buf1[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 6]);
-  salt_buf1[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 7]);
+  salt_buf1[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 4]);
+  salt_buf1[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 5]);
+  salt_buf1[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 6]);
+  salt_buf1[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 7]);
 
   u32 salt_buf2[4];
 
-  salt_buf2[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 8]);
-  salt_buf2[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 9]);
-  salt_buf2[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[10]);
-  salt_buf2[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[11]);
+  salt_buf2[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 8]);
+  salt_buf2[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 9]);
+  salt_buf2[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[10]);
+  salt_buf2[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[11]);
 
   u32 salt_buf3[4];
 
-  salt_buf3[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[12]);
-  salt_buf3[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[13]);
+  salt_buf3[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[12]);
+  salt_buf3[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[13]);
   salt_buf3[2] = 0;
   salt_buf3[3] = 0;
 
@@ -558,8 +558,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s04 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -619,17 +619,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 

OpenCL/m05400_a1.cl

@@ -295,29 +295,29 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m04 (__glo
 
   u32 salt_buf0[4];
 
-  salt_buf0[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 0]);
-  salt_buf0[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 1]);
-  salt_buf0[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 2]);
-  salt_buf0[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 3]);
+  salt_buf0[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 0]);
+  salt_buf0[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 1]);
+  salt_buf0[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 2]);
+  salt_buf0[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 3]);
 
   u32 salt_buf1[4];
 
-  salt_buf1[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 4]);
-  salt_buf1[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 5]);
-  salt_buf1[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 6]);
-  salt_buf1[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 7]);
+  salt_buf1[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 4]);
+  salt_buf1[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 5]);
+  salt_buf1[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 6]);
+  salt_buf1[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 7]);
 
   u32 salt_buf2[4];
 
-  salt_buf2[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 8]);
-  salt_buf2[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 9]);
-  salt_buf2[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[10]);
-  salt_buf2[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[11]);
+  salt_buf2[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 8]);
+  salt_buf2[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 9]);
+  salt_buf2[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[10]);
+  salt_buf2[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[11]);
 
   u32 salt_buf3[4];
 
-  salt_buf3[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[12]);
-  salt_buf3[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[13]);
+  salt_buf3[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[12]);
+  salt_buf3[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[13]);
   salt_buf3[2] = 0;
   salt_buf3[3] = 0;
 
@@ -325,8 +325,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m04 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -409,17 +409,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 
@@ -599,29 +599,29 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s04 (__glo
 
   u32 salt_buf0[4];
 
-  salt_buf0[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 0]);
-  salt_buf0[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 1]);
-  salt_buf0[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 2]);
-  salt_buf0[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 3]);
+  salt_buf0[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 0]);
+  salt_buf0[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 1]);
+  salt_buf0[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 2]);
+  salt_buf0[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 3]);
 
   u32 salt_buf1[4];
 
-  salt_buf1[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 4]);
-  salt_buf1[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 5]);
-  salt_buf1[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 6]);
-  salt_buf1[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 7]);
+  salt_buf1[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 4]);
+  salt_buf1[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 5]);
+  salt_buf1[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 6]);
+  salt_buf1[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 7]);
 
   u32 salt_buf2[4];
 
-  salt_buf2[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 8]);
-  salt_buf2[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 9]);
-  salt_buf2[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[10]);
-  salt_buf2[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[11]);
+  salt_buf2[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 8]);
+  salt_buf2[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 9]);
+  salt_buf2[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[10]);
+  salt_buf2[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[11]);
 
   u32 salt_buf3[4];
 
-  salt_buf3[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[12]);
-  salt_buf3[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[13]);
+  salt_buf3[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[12]);
+  salt_buf3[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[13]);
   salt_buf3[2] = 0;
   salt_buf3[3] = 0;
 
@@ -629,8 +629,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s04 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -725,17 +725,17 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s04 (__glo
 
     u32 w0_t[4];
 
-    w0_t[0] = swap_workaround (w0[0]);
-    w0_t[1] = swap_workaround (w0[1]);
-    w0_t[2] = swap_workaround (w0[2]);
-    w0_t[3] = swap_workaround (w0[3]);
+    w0_t[0] = swap32 (w0[0]);
+    w0_t[1] = swap32 (w0[1]);
+    w0_t[2] = swap32 (w0[2]);
+    w0_t[3] = swap32 (w0[3]);
 
     u32 w1_t[4];
 
-    w1_t[0] = swap_workaround (w1[0]);
-    w1_t[1] = swap_workaround (w1[1]);
-    w1_t[2] = swap_workaround (w1[2]);
-    w1_t[3] = swap_workaround (w1[3]);
+    w1_t[0] = swap32 (w1[0]);
+    w1_t[1] = swap32 (w1[1]);
+    w1_t[2] = swap32 (w1[2]);
+    w1_t[3] = swap32 (w1[3]);
 
     u32 w2_t[4];
 

OpenCL/m05400_a3.cl

@@ -255,29 +255,29 @@ static void m05400m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 salt_buf0[4];
 
-  salt_buf0[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 0]);
-  salt_buf0[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 1]);
-  salt_buf0[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 2]);
-  salt_buf0[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 3]);
+  salt_buf0[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 0]);
+  salt_buf0[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 1]);
+  salt_buf0[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 2]);
+  salt_buf0[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 3]);
 
   u32 salt_buf1[4];
 
-  salt_buf1[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 4]);
-  salt_buf1[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 5]);
-  salt_buf1[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 6]);
-  salt_buf1[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 7]);
+  salt_buf1[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 4]);
+  salt_buf1[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 5]);
+  salt_buf1[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 6]);
+  salt_buf1[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 7]);
 
   u32 salt_buf2[4];
 
-  salt_buf2[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 8]);
-  salt_buf2[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 9]);
-  salt_buf2[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[10]);
-  salt_buf2[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[11]);
+  salt_buf2[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 8]);
+  salt_buf2[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 9]);
+  salt_buf2[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[10]);
+  salt_buf2[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[11]);
 
   u32 salt_buf3[4];
 
-  salt_buf3[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[12]);
-  salt_buf3[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[13]);
+  salt_buf3[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[12]);
+  salt_buf3[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[13]);
   salt_buf3[2] = 0;
   salt_buf3[3] = 0;
 
@@ -441,29 +441,29 @@ static void m05400s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
 
   u32 salt_buf0[4];
 
-  salt_buf0[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 0]);
-  salt_buf0[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 1]);
-  salt_buf0[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 2]);
-  salt_buf0[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 3]);
+  salt_buf0[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 0]);
+  salt_buf0[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 1]);
+  salt_buf0[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 2]);
+  salt_buf0[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 3]);
 
   u32 salt_buf1[4];
 
-  salt_buf1[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 4]);
-  salt_buf1[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 5]);
-  salt_buf1[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 6]);
-  salt_buf1[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 7]);
+  salt_buf1[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 4]);
+  salt_buf1[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 5]);
+  salt_buf1[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 6]);
+  salt_buf1[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 7]);
 
   u32 salt_buf2[4];
 
-  salt_buf2[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 8]);
-  salt_buf2[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[ 9]);
-  salt_buf2[2] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[10]);
-  salt_buf2[3] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[11]);
+  salt_buf2[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 8]);
+  salt_buf2[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[ 9]);
+  salt_buf2[2] = swap32 (ikepsk_bufs[salt_pos].nr_buf[10]);
+  salt_buf2[3] = swap32 (ikepsk_bufs[salt_pos].nr_buf[11]);
 
   u32 salt_buf3[4];
 
-  salt_buf3[0] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[12]);
-  salt_buf3[1] = swap_workaround (ikepsk_bufs[salt_pos].nr_buf[13]);
+  salt_buf3[0] = swap32 (ikepsk_bufs[salt_pos].nr_buf[12]);
+  salt_buf3[1] = swap32 (ikepsk_bufs[salt_pos].nr_buf[13]);
   salt_buf3[2] = 0;
   salt_buf3[3] = 0;
 
@@ -673,8 +673,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m04 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -739,8 +739,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m08 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -805,8 +805,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_m16 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -871,8 +871,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s04 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -937,8 +937,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s08 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 
@@ -1003,8 +1003,8 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05400_s16 (__glo
 
   const u32 lid2 = lid * 2;
 
-  s_msg_buf[lid2 + 0] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
-  s_msg_buf[lid2 + 1] = swap_workaround (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
+  s_msg_buf[lid2 + 0] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 0]);
+  s_msg_buf[lid2 + 1] = swap32 (ikepsk_bufs[salt_pos].msg_buf[lid2 + 1]);
 
   barrier (CLK_LOCAL_MEM_FENCE);
 

OpenCL/m05800.cl

@@ -17,13 +17,8 @@
 #include "types_ocl.c"
 #include "common.c"
 
-#ifdef  VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
+#define COMPARE_S "check_single_comp4.c"
+#define COMPARE_M "check_multi_comp4.c"
 
 typedef struct
 {
@@ -545,16 +540,16 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05800_init (__gl
   u32 w2[4];
   u32 w3[4];
 
-  w0[0] = swap_workaround (data0[0]);
-  w0[1] = swap_workaround (data0[1]);
-  w0[2] = swap_workaround (data0[2]);
-  w0[3] = swap_workaround (data0[3]);
-  w1[0] = swap_workaround (data1[0]);
-  w1[1] = swap_workaround (data1[1]);
-  w1[2] = swap_workaround (data1[2]);
-  w1[3] = swap_workaround (data1[3]);
-  w2[0] = swap_workaround (data2[0]);
-  w2[1] = swap_workaround (data2[1]);
+  w0[0] = swap32 (data0[0]);
+  w0[1] = swap32 (data0[1]);
+  w0[2] = swap32 (data0[2]);
+  w0[3] = swap32 (data0[3]);
+  w1[0] = swap32 (data1[0]);
+  w1[1] = swap32 (data1[1]);
+  w1[2] = swap32 (data1[2]);
+  w1[3] = swap32 (data1[3]);
+  w2[0] = swap32 (data2[0]);
+  w2[1] = swap32 (data2[1]);
   w2[2] = 0;
   w2[3] = 0;
   w3[0] = 0;
@@ -678,15 +673,15 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05800_loop (__gl
     w0[2] = digest[2];
     w0[3] = digest[3];
     w1[0] = digest[4];
-    w1[1] = swap_workaround (data0[0]);
-    w1[2] = swap_workaround (data0[1]);
-    w1[3] = swap_workaround (data0[2]);
-    w2[0] = swap_workaround (data0[3]);
-    w2[1] = swap_workaround (data1[0]);
-    w2[2] = swap_workaround (data1[1]);
-    w2[3] = swap_workaround (data1[2]);
-    w3[0] = swap_workaround (data1[3]);
-    w3[1] = swap_workaround (data2[0]);
+    w1[1] = swap32 (data0[0]);
+    w1[2] = swap32 (data0[1]);
+    w1[3] = swap32 (data0[2]);
+    w2[0] = swap32 (data0[3]);
+    w2[1] = swap32 (data1[0]);
+    w2[2] = swap32 (data1[1]);
+    w2[3] = swap32 (data1[2]);
+    w3[0] = swap32 (data1[3]);
+    w3[1] = swap32 (data2[0]);
     w3[2] = 0;
     w3[3] = (20 + pc_len + pw_len + salt_len) * 8;
 

OpenCL/m06100_a0.cl

@@ -1436,14 +1436,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06100_m04 (__glo
 
     u32 wl[16];
 
-    wl[ 0] = swap_workaround (w0[0]);
-    wl[ 1] = swap_workaround (w0[1]);
-    wl[ 2] = swap_workaround (w0[2]);
-    wl[ 3] = swap_workaround (w0[3]);
-    wl[ 4] = swap_workaround (w1[0]);
-    wl[ 5] = swap_workaround (w1[1]);
-    wl[ 6] = swap_workaround (w1[2]);
-    wl[ 7] = swap_workaround (w1[3]);
+    wl[ 0] = swap32 (w0[0]);
+    wl[ 1] = swap32 (w0[1]);
+    wl[ 2] = swap32 (w0[2]);
+    wl[ 3] = swap32 (w0[3]);
+    wl[ 4] = swap32 (w1[0]);
+    wl[ 5] = swap32 (w1[1]);
+    wl[ 6] = swap32 (w1[2]);
+    wl[ 7] = swap32 (w1[3]);
     wl[ 8] = 0;
     wl[ 9] = 0;
     wl[10] = 0;
@@ -1598,14 +1598,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06100_s04 (__glo
 
     u32 wl[16];
 
-    wl[ 0] = swap_workaround (w0[0]);
-    wl[ 1] = swap_workaround (w0[1]);
-    wl[ 2] = swap_workaround (w0[2]);
-    wl[ 3] = swap_workaround (w0[3]);
-    wl[ 4] = swap_workaround (w1[0]);
-    wl[ 5] = swap_workaround (w1[1]);
-    wl[ 6] = swap_workaround (w1[2]);
-    wl[ 7] = swap_workaround (w1[3]);
+    wl[ 0] = swap32 (w0[0]);
+    wl[ 1] = swap32 (w0[1]);
+    wl[ 2] = swap32 (w0[2]);
+    wl[ 3] = swap32 (w0[3]);
+    wl[ 4] = swap32 (w1[0]);
+    wl[ 5] = swap32 (w1[1]);
+    wl[ 6] = swap32 (w1[2]);
+    wl[ 7] = swap32 (w1[3]);
     wl[ 8] = 0;
     wl[ 9] = 0;
     wl[10] = 0;

OpenCL/m06100_a1.cl

@@ -1488,14 +1488,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06100_m04 (__glo
 
     u32 wl[16];
 
-    wl[ 0] = swap_workaround (w0[0]);
-    wl[ 1] = swap_workaround (w0[1]);
-    wl[ 2] = swap_workaround (w0[2]);
-    wl[ 3] = swap_workaround (w0[3]);
-    wl[ 4] = swap_workaround (w1[0]);
-    wl[ 5] = swap_workaround (w1[1]);
-    wl[ 6] = swap_workaround (w1[2]);
-    wl[ 7] = swap_workaround (w1[3]);
+    wl[ 0] = swap32 (w0[0]);
+    wl[ 1] = swap32 (w0[1]);
+    wl[ 2] = swap32 (w0[2]);
+    wl[ 3] = swap32 (w0[3]);
+    wl[ 4] = swap32 (w1[0]);
+    wl[ 5] = swap32 (w1[1]);
+    wl[ 6] = swap32 (w1[2]);
+    wl[ 7] = swap32 (w1[3]);
     wl[ 8] = 0;
     wl[ 9] = 0;
     wl[10] = 0;
@@ -1704,14 +1704,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06100_s04 (__glo
 
     u32 wl[16];
 
-    wl[ 0] = swap_workaround (w0[0]);
-    wl[ 1] = swap_workaround (w0[1]);
-    wl[ 2] = swap_workaround (w0[2]);
-    wl[ 3] = swap_workaround (w0[3]);
-    wl[ 4] = swap_workaround (w1[0]);
-    wl[ 5] = swap_workaround (w1[1]);
-    wl[ 6] = swap_workaround (w1[2]);
-    wl[ 7] = swap_workaround (w1[3]);
+    wl[ 0] = swap32 (w0[0]);
+    wl[ 1] = swap32 (w0[1]);
+    wl[ 2] = swap32 (w0[2]);
+    wl[ 3] = swap32 (w0[3]);
+    wl[ 4] = swap32 (w1[0]);
+    wl[ 5] = swap32 (w1[1]);
+    wl[ 6] = swap32 (w1[2]);
+    wl[ 7] = swap32 (w1[3]);
     wl[ 8] = 0;
     wl[ 9] = 0;
     wl[10] = 0;

OpenCL/m06211.cl

@@ -512,7 +512,7 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06211_init (__gl
 
   for (u32 i = 0, j = 1; i < (truecrypt_mdlen / 8 / 4); i += 5, j += 1)
   {
-    salt_buf2[0] = swap_workaround (j);
+    salt_buf2[0] = swap32 (j);
 
     u32 dgst[5];
 

OpenCL/m06212.cl

@@ -512,7 +512,7 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06212_init (__gl
 
   for (u32 i = 0, j = 1; i < (truecrypt_mdlen / 8 / 4); i += 5, j += 1)
   {
-    salt_buf2[0] = swap_workaround (j);
+    salt_buf2[0] = swap32 (j);
 
     u32 dgst[5];
 

OpenCL/m06213.cl

@@ -512,7 +512,7 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06213_init (__gl
 
   for (u32 i = 0, j = 1; i < (truecrypt_mdlen / 8 / 4); i += 5, j += 1)
   {
-    salt_buf2[0] = swap_workaround (j);
+    salt_buf2[0] = swap32 (j);
 
     u32 dgst[5];
 

OpenCL/m06221.cl

@@ -326,14 +326,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06221_init (__gl
 
   // swap fehlt
 
-  salt_buf[ 0] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 0])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 1]);
-  salt_buf[ 1] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 2])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 3]);
-  salt_buf[ 2] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 4])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 5]);
-  salt_buf[ 3] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 6])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 7]);
-  salt_buf[ 4] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 8])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 9]);
-  salt_buf[ 5] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[10])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[11]);
-  salt_buf[ 6] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[12])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[13]);
-  salt_buf[ 7] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[14])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[15]);
+  salt_buf[ 0] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 0])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf[ 1] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 2])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf[ 2] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 4])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf[ 3] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 6])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf[ 4] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 8])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf[ 5] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[10])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[11]);
+  salt_buf[ 6] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[12])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[13]);
+  salt_buf[ 7] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[14])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[15]);
   salt_buf[ 8] = 0;
   salt_buf[ 9] = 0;
   salt_buf[10] = 0;
@@ -347,14 +347,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06221_init (__gl
 
   u64 w[16];
 
-  w[ 0] = ((u64) swap_workaround (w0[0])) << 32 | (u64) swap_workaround (w0[1]);
-  w[ 1] = ((u64) swap_workaround (w0[2])) << 32 | (u64) swap_workaround (w0[3]);
-  w[ 2] = ((u64) swap_workaround (w1[0])) << 32 | (u64) swap_workaround (w1[1]);
-  w[ 3] = ((u64) swap_workaround (w1[2])) << 32 | (u64) swap_workaround (w1[3]);
-  w[ 4] = ((u64) swap_workaround (w2[0])) << 32 | (u64) swap_workaround (w2[1]);
-  w[ 5] = ((u64) swap_workaround (w2[2])) << 32 | (u64) swap_workaround (w2[3]);
-  w[ 6] = ((u64) swap_workaround (w3[0])) << 32 | (u64) swap_workaround (w3[1]);
-  w[ 7] = ((u64) swap_workaround (w3[2])) << 32 | (u64) swap_workaround (w3[3]);
+  w[ 0] = ((u64) swap32 (w0[0])) << 32 | (u64) swap32 (w0[1]);
+  w[ 1] = ((u64) swap32 (w0[2])) << 32 | (u64) swap32 (w0[3]);
+  w[ 2] = ((u64) swap32 (w1[0])) << 32 | (u64) swap32 (w1[1]);
+  w[ 3] = ((u64) swap32 (w1[2])) << 32 | (u64) swap32 (w1[3]);
+  w[ 4] = ((u64) swap32 (w2[0])) << 32 | (u64) swap32 (w2[1]);
+  w[ 5] = ((u64) swap32 (w2[2])) << 32 | (u64) swap32 (w2[3]);
+  w[ 6] = ((u64) swap32 (w3[0])) << 32 | (u64) swap32 (w3[1]);
+  w[ 7] = ((u64) swap32 (w3[2])) << 32 | (u64) swap32 (w3[3]);
   w[ 8] = 0;
   w[ 9] = 0;
   w[10] = 0;
@@ -535,25 +535,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06221_comp (__gl
 
   u32 ukey1[8];
 
-  ukey1[0] = swap_workaround (h32_from_64 (tmps[gid].out[ 0]));
-  ukey1[1] = swap_workaround (l32_from_64 (tmps[gid].out[ 0]));
-  ukey1[2] = swap_workaround (h32_from_64 (tmps[gid].out[ 1]));
-  ukey1[3] = swap_workaround (l32_from_64 (tmps[gid].out[ 1]));
-  ukey1[4] = swap_workaround (h32_from_64 (tmps[gid].out[ 2]));
-  ukey1[5] = swap_workaround (l32_from_64 (tmps[gid].out[ 2]));
-  ukey1[6] = swap_workaround (h32_from_64 (tmps[gid].out[ 3]));
-  ukey1[7] = swap_workaround (l32_from_64 (tmps[gid].out[ 3]));
+  ukey1[0] = swap32 (h32_from_64 (tmps[gid].out[ 0]));
+  ukey1[1] = swap32 (l32_from_64 (tmps[gid].out[ 0]));
+  ukey1[2] = swap32 (h32_from_64 (tmps[gid].out[ 1]));
+  ukey1[3] = swap32 (l32_from_64 (tmps[gid].out[ 1]));
+  ukey1[4] = swap32 (h32_from_64 (tmps[gid].out[ 2]));
+  ukey1[5] = swap32 (l32_from_64 (tmps[gid].out[ 2]));
+  ukey1[6] = swap32 (h32_from_64 (tmps[gid].out[ 3]));
+  ukey1[7] = swap32 (l32_from_64 (tmps[gid].out[ 3]));
 
   u32 ukey2[8];
 
-  ukey2[0] = swap_workaround (h32_from_64 (tmps[gid].out[ 4]));
-  ukey2[1] = swap_workaround (l32_from_64 (tmps[gid].out[ 4]));
-  ukey2[2] = swap_workaround (h32_from_64 (tmps[gid].out[ 5]));
-  ukey2[3] = swap_workaround (l32_from_64 (tmps[gid].out[ 5]));
-  ukey2[4] = swap_workaround (h32_from_64 (tmps[gid].out[ 6]));
-  ukey2[5] = swap_workaround (l32_from_64 (tmps[gid].out[ 6]));
-  ukey2[6] = swap_workaround (h32_from_64 (tmps[gid].out[ 7]));
-  ukey2[7] = swap_workaround (l32_from_64 (tmps[gid].out[ 7]));
+  ukey2[0] = swap32 (h32_from_64 (tmps[gid].out[ 4]));
+  ukey2[1] = swap32 (l32_from_64 (tmps[gid].out[ 4]));
+  ukey2[2] = swap32 (h32_from_64 (tmps[gid].out[ 5]));
+  ukey2[3] = swap32 (l32_from_64 (tmps[gid].out[ 5]));
+  ukey2[4] = swap32 (h32_from_64 (tmps[gid].out[ 6]));
+  ukey2[5] = swap32 (l32_from_64 (tmps[gid].out[ 6]));
+  ukey2[6] = swap32 (h32_from_64 (tmps[gid].out[ 7]));
+  ukey2[7] = swap32 (l32_from_64 (tmps[gid].out[ 7]));
 
   u32 data[4];
 

OpenCL/m06222.cl

@@ -326,14 +326,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06222_init (__gl
 
   // swap fehlt
 
-  salt_buf[ 0] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 0])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 1]);
-  salt_buf[ 1] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 2])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 3]);
-  salt_buf[ 2] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 4])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 5]);
-  salt_buf[ 3] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 6])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 7]);
-  salt_buf[ 4] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 8])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 9]);
-  salt_buf[ 5] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[10])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[11]);
-  salt_buf[ 6] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[12])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[13]);
-  salt_buf[ 7] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[14])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[15]);
+  salt_buf[ 0] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 0])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf[ 1] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 2])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf[ 2] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 4])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf[ 3] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 6])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf[ 4] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 8])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf[ 5] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[10])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[11]);
+  salt_buf[ 6] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[12])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[13]);
+  salt_buf[ 7] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[14])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[15]);
   salt_buf[ 8] = 0;
   salt_buf[ 9] = 0;
   salt_buf[10] = 0;
@@ -347,14 +347,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06222_init (__gl
 
   u64 w[16];
 
-  w[ 0] = ((u64) swap_workaround (w0[0])) << 32 | (u64) swap_workaround (w0[1]);
-  w[ 1] = ((u64) swap_workaround (w0[2])) << 32 | (u64) swap_workaround (w0[3]);
-  w[ 2] = ((u64) swap_workaround (w1[0])) << 32 | (u64) swap_workaround (w1[1]);
-  w[ 3] = ((u64) swap_workaround (w1[2])) << 32 | (u64) swap_workaround (w1[3]);
-  w[ 4] = ((u64) swap_workaround (w2[0])) << 32 | (u64) swap_workaround (w2[1]);
-  w[ 5] = ((u64) swap_workaround (w2[2])) << 32 | (u64) swap_workaround (w2[3]);
-  w[ 6] = ((u64) swap_workaround (w3[0])) << 32 | (u64) swap_workaround (w3[1]);
-  w[ 7] = ((u64) swap_workaround (w3[2])) << 32 | (u64) swap_workaround (w3[3]);
+  w[ 0] = ((u64) swap32 (w0[0])) << 32 | (u64) swap32 (w0[1]);
+  w[ 1] = ((u64) swap32 (w0[2])) << 32 | (u64) swap32 (w0[3]);
+  w[ 2] = ((u64) swap32 (w1[0])) << 32 | (u64) swap32 (w1[1]);
+  w[ 3] = ((u64) swap32 (w1[2])) << 32 | (u64) swap32 (w1[3]);
+  w[ 4] = ((u64) swap32 (w2[0])) << 32 | (u64) swap32 (w2[1]);
+  w[ 5] = ((u64) swap32 (w2[2])) << 32 | (u64) swap32 (w2[3]);
+  w[ 6] = ((u64) swap32 (w3[0])) << 32 | (u64) swap32 (w3[1]);
+  w[ 7] = ((u64) swap32 (w3[2])) << 32 | (u64) swap32 (w3[3]);
   w[ 8] = 0;
   w[ 9] = 0;
   w[10] = 0;
@@ -535,25 +535,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06222_comp (__gl
 
   u32 ukey1[8];
 
-  ukey1[0] = swap_workaround (h32_from_64 (tmps[gid].out[ 0]));
-  ukey1[1] = swap_workaround (l32_from_64 (tmps[gid].out[ 0]));
-  ukey1[2] = swap_workaround (h32_from_64 (tmps[gid].out[ 1]));
-  ukey1[3] = swap_workaround (l32_from_64 (tmps[gid].out[ 1]));
-  ukey1[4] = swap_workaround (h32_from_64 (tmps[gid].out[ 2]));
-  ukey1[5] = swap_workaround (l32_from_64 (tmps[gid].out[ 2]));
-  ukey1[6] = swap_workaround (h32_from_64 (tmps[gid].out[ 3]));
-  ukey1[7] = swap_workaround (l32_from_64 (tmps[gid].out[ 3]));
+  ukey1[0] = swap32 (h32_from_64 (tmps[gid].out[ 0]));
+  ukey1[1] = swap32 (l32_from_64 (tmps[gid].out[ 0]));
+  ukey1[2] = swap32 (h32_from_64 (tmps[gid].out[ 1]));
+  ukey1[3] = swap32 (l32_from_64 (tmps[gid].out[ 1]));
+  ukey1[4] = swap32 (h32_from_64 (tmps[gid].out[ 2]));
+  ukey1[5] = swap32 (l32_from_64 (tmps[gid].out[ 2]));
+  ukey1[6] = swap32 (h32_from_64 (tmps[gid].out[ 3]));
+  ukey1[7] = swap32 (l32_from_64 (tmps[gid].out[ 3]));
 
   u32 ukey2[8];
 
-  ukey2[0] = swap_workaround (h32_from_64 (tmps[gid].out[ 4]));
-  ukey2[1] = swap_workaround (l32_from_64 (tmps[gid].out[ 4]));
-  ukey2[2] = swap_workaround (h32_from_64 (tmps[gid].out[ 5]));
-  ukey2[3] = swap_workaround (l32_from_64 (tmps[gid].out[ 5]));
-  ukey2[4] = swap_workaround (h32_from_64 (tmps[gid].out[ 6]));
-  ukey2[5] = swap_workaround (l32_from_64 (tmps[gid].out[ 6]));
-  ukey2[6] = swap_workaround (h32_from_64 (tmps[gid].out[ 7]));
-  ukey2[7] = swap_workaround (l32_from_64 (tmps[gid].out[ 7]));
+  ukey2[0] = swap32 (h32_from_64 (tmps[gid].out[ 4]));
+  ukey2[1] = swap32 (l32_from_64 (tmps[gid].out[ 4]));
+  ukey2[2] = swap32 (h32_from_64 (tmps[gid].out[ 5]));
+  ukey2[3] = swap32 (l32_from_64 (tmps[gid].out[ 5]));
+  ukey2[4] = swap32 (h32_from_64 (tmps[gid].out[ 6]));
+  ukey2[5] = swap32 (l32_from_64 (tmps[gid].out[ 6]));
+  ukey2[6] = swap32 (h32_from_64 (tmps[gid].out[ 7]));
+  ukey2[7] = swap32 (l32_from_64 (tmps[gid].out[ 7]));
 
   u32 data[4];
 
@@ -614,25 +614,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06222_comp (__gl
 
   u32 ukey3[8];
 
-  ukey3[0] = swap_workaround (h32_from_64 (tmps[gid].out[ 8]));
-  ukey3[1] = swap_workaround (l32_from_64 (tmps[gid].out[ 8]));
-  ukey3[2] = swap_workaround (h32_from_64 (tmps[gid].out[ 9]));
-  ukey3[3] = swap_workaround (l32_from_64 (tmps[gid].out[ 9]));
-  ukey3[4] = swap_workaround (h32_from_64 (tmps[gid].out[10]));
-  ukey3[5] = swap_workaround (l32_from_64 (tmps[gid].out[10]));
-  ukey3[6] = swap_workaround (h32_from_64 (tmps[gid].out[11]));
-  ukey3[7] = swap_workaround (l32_from_64 (tmps[gid].out[11]));
+  ukey3[0] = swap32 (h32_from_64 (tmps[gid].out[ 8]));
+  ukey3[1] = swap32 (l32_from_64 (tmps[gid].out[ 8]));
+  ukey3[2] = swap32 (h32_from_64 (tmps[gid].out[ 9]));
+  ukey3[3] = swap32 (l32_from_64 (tmps[gid].out[ 9]));
+  ukey3[4] = swap32 (h32_from_64 (tmps[gid].out[10]));
+  ukey3[5] = swap32 (l32_from_64 (tmps[gid].out[10]));
+  ukey3[6] = swap32 (h32_from_64 (tmps[gid].out[11]));
+  ukey3[7] = swap32 (l32_from_64 (tmps[gid].out[11]));
 
   u32 ukey4[8];
 
-  ukey4[0] = swap_workaround (h32_from_64 (tmps[gid].out[12]));
-  ukey4[1] = swap_workaround (l32_from_64 (tmps[gid].out[12]));
-  ukey4[2] = swap_workaround (h32_from_64 (tmps[gid].out[13]));
-  ukey4[3] = swap_workaround (l32_from_64 (tmps[gid].out[13]));
-  ukey4[4] = swap_workaround (h32_from_64 (tmps[gid].out[14]));
-  ukey4[5] = swap_workaround (l32_from_64 (tmps[gid].out[14]));
-  ukey4[6] = swap_workaround (h32_from_64 (tmps[gid].out[15]));
-  ukey4[7] = swap_workaround (l32_from_64 (tmps[gid].out[15]));
+  ukey4[0] = swap32 (h32_from_64 (tmps[gid].out[12]));
+  ukey4[1] = swap32 (l32_from_64 (tmps[gid].out[12]));
+  ukey4[2] = swap32 (h32_from_64 (tmps[gid].out[13]));
+  ukey4[3] = swap32 (l32_from_64 (tmps[gid].out[13]));
+  ukey4[4] = swap32 (h32_from_64 (tmps[gid].out[14]));
+  ukey4[5] = swap32 (l32_from_64 (tmps[gid].out[14]));
+  ukey4[6] = swap32 (h32_from_64 (tmps[gid].out[15]));
+  ukey4[7] = swap32 (l32_from_64 (tmps[gid].out[15]));
 
   {
     tmp[0] = data[0];

OpenCL/m06223.cl

@@ -326,14 +326,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06223_init (__gl
 
   // swap fehlt
 
-  salt_buf[ 0] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 0])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 1]);
-  salt_buf[ 1] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 2])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 3]);
-  salt_buf[ 2] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 4])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 5]);
-  salt_buf[ 3] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 6])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 7]);
-  salt_buf[ 4] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 8])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[ 9]);
-  salt_buf[ 5] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[10])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[11]);
-  salt_buf[ 6] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[12])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[13]);
-  salt_buf[ 7] = ((u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[14])) << 32 | (u64) swap_workaround (esalt_bufs[salt_pos].salt_buf[15]);
+  salt_buf[ 0] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 0])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf[ 1] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 2])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf[ 2] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 4])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf[ 3] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 6])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf[ 4] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 8])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf[ 5] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[10])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[11]);
+  salt_buf[ 6] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[12])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[13]);
+  salt_buf[ 7] = ((u64) swap32 (esalt_bufs[salt_pos].salt_buf[14])) << 32 | (u64) swap32 (esalt_bufs[salt_pos].salt_buf[15]);
   salt_buf[ 8] = 0;
   salt_buf[ 9] = 0;
   salt_buf[10] = 0;
@@ -347,14 +347,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06223_init (__gl
 
   u64 w[16];
 
-  w[ 0] = ((u64) swap_workaround (w0[0])) << 32 | (u64) swap_workaround (w0[1]);
-  w[ 1] = ((u64) swap_workaround (w0[2])) << 32 | (u64) swap_workaround (w0[3]);
-  w[ 2] = ((u64) swap_workaround (w1[0])) << 32 | (u64) swap_workaround (w1[1]);
-  w[ 3] = ((u64) swap_workaround (w1[2])) << 32 | (u64) swap_workaround (w1[3]);
-  w[ 4] = ((u64) swap_workaround (w2[0])) << 32 | (u64) swap_workaround (w2[1]);
-  w[ 5] = ((u64) swap_workaround (w2[2])) << 32 | (u64) swap_workaround (w2[3]);
-  w[ 6] = ((u64) swap_workaround (w3[0])) << 32 | (u64) swap_workaround (w3[1]);
-  w[ 7] = ((u64) swap_workaround (w3[2])) << 32 | (u64) swap_workaround (w3[3]);
+  w[ 0] = ((u64) swap32 (w0[0])) << 32 | (u64) swap32 (w0[1]);
+  w[ 1] = ((u64) swap32 (w0[2])) << 32 | (u64) swap32 (w0[3]);
+  w[ 2] = ((u64) swap32 (w1[0])) << 32 | (u64) swap32 (w1[1]);
+  w[ 3] = ((u64) swap32 (w1[2])) << 32 | (u64) swap32 (w1[3]);
+  w[ 4] = ((u64) swap32 (w2[0])) << 32 | (u64) swap32 (w2[1]);
+  w[ 5] = ((u64) swap32 (w2[2])) << 32 | (u64) swap32 (w2[3]);
+  w[ 6] = ((u64) swap32 (w3[0])) << 32 | (u64) swap32 (w3[1]);
+  w[ 7] = ((u64) swap32 (w3[2])) << 32 | (u64) swap32 (w3[3]);
   w[ 8] = 0;
   w[ 9] = 0;
   w[10] = 0;
@@ -535,25 +535,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06223_comp (__gl
 
   u32 ukey1[8];
 
-  ukey1[0] = swap_workaround (h32_from_64 (tmps[gid].out[ 0]));
-  ukey1[1] = swap_workaround (l32_from_64 (tmps[gid].out[ 0]));
-  ukey1[2] = swap_workaround (h32_from_64 (tmps[gid].out[ 1]));
-  ukey1[3] = swap_workaround (l32_from_64 (tmps[gid].out[ 1]));
-  ukey1[4] = swap_workaround (h32_from_64 (tmps[gid].out[ 2]));
-  ukey1[5] = swap_workaround (l32_from_64 (tmps[gid].out[ 2]));
-  ukey1[6] = swap_workaround (h32_from_64 (tmps[gid].out[ 3]));
-  ukey1[7] = swap_workaround (l32_from_64 (tmps[gid].out[ 3]));
+  ukey1[0] = swap32 (h32_from_64 (tmps[gid].out[ 0]));
+  ukey1[1] = swap32 (l32_from_64 (tmps[gid].out[ 0]));
+  ukey1[2] = swap32 (h32_from_64 (tmps[gid].out[ 1]));
+  ukey1[3] = swap32 (l32_from_64 (tmps[gid].out[ 1]));
+  ukey1[4] = swap32 (h32_from_64 (tmps[gid].out[ 2]));
+  ukey1[5] = swap32 (l32_from_64 (tmps[gid].out[ 2]));
+  ukey1[6] = swap32 (h32_from_64 (tmps[gid].out[ 3]));
+  ukey1[7] = swap32 (l32_from_64 (tmps[gid].out[ 3]));
 
   u32 ukey2[8];
 
-  ukey2[0] = swap_workaround (h32_from_64 (tmps[gid].out[ 4]));
-  ukey2[1] = swap_workaround (l32_from_64 (tmps[gid].out[ 4]));
-  ukey2[2] = swap_workaround (h32_from_64 (tmps[gid].out[ 5]));
-  ukey2[3] = swap_workaround (l32_from_64 (tmps[gid].out[ 5]));
-  ukey2[4] = swap_workaround (h32_from_64 (tmps[gid].out[ 6]));
-  ukey2[5] = swap_workaround (l32_from_64 (tmps[gid].out[ 6]));
-  ukey2[6] = swap_workaround (h32_from_64 (tmps[gid].out[ 7]));
-  ukey2[7] = swap_workaround (l32_from_64 (tmps[gid].out[ 7]));
+  ukey2[0] = swap32 (h32_from_64 (tmps[gid].out[ 4]));
+  ukey2[1] = swap32 (l32_from_64 (tmps[gid].out[ 4]));
+  ukey2[2] = swap32 (h32_from_64 (tmps[gid].out[ 5]));
+  ukey2[3] = swap32 (l32_from_64 (tmps[gid].out[ 5]));
+  ukey2[4] = swap32 (h32_from_64 (tmps[gid].out[ 6]));
+  ukey2[5] = swap32 (l32_from_64 (tmps[gid].out[ 6]));
+  ukey2[6] = swap32 (h32_from_64 (tmps[gid].out[ 7]));
+  ukey2[7] = swap32 (l32_from_64 (tmps[gid].out[ 7]));
 
   u32 data[4];
 
@@ -614,25 +614,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06223_comp (__gl
 
   u32 ukey3[8];
 
-  ukey3[0] = swap_workaround (h32_from_64 (tmps[gid].out[ 8]));
-  ukey3[1] = swap_workaround (l32_from_64 (tmps[gid].out[ 8]));
-  ukey3[2] = swap_workaround (h32_from_64 (tmps[gid].out[ 9]));
-  ukey3[3] = swap_workaround (l32_from_64 (tmps[gid].out[ 9]));
-  ukey3[4] = swap_workaround (h32_from_64 (tmps[gid].out[10]));
-  ukey3[5] = swap_workaround (l32_from_64 (tmps[gid].out[10]));
-  ukey3[6] = swap_workaround (h32_from_64 (tmps[gid].out[11]));
-  ukey3[7] = swap_workaround (l32_from_64 (tmps[gid].out[11]));
+  ukey3[0] = swap32 (h32_from_64 (tmps[gid].out[ 8]));
+  ukey3[1] = swap32 (l32_from_64 (tmps[gid].out[ 8]));
+  ukey3[2] = swap32 (h32_from_64 (tmps[gid].out[ 9]));
+  ukey3[3] = swap32 (l32_from_64 (tmps[gid].out[ 9]));
+  ukey3[4] = swap32 (h32_from_64 (tmps[gid].out[10]));
+  ukey3[5] = swap32 (l32_from_64 (tmps[gid].out[10]));
+  ukey3[6] = swap32 (h32_from_64 (tmps[gid].out[11]));
+  ukey3[7] = swap32 (l32_from_64 (tmps[gid].out[11]));
 
   u32 ukey4[8];
 
-  ukey4[0] = swap_workaround (h32_from_64 (tmps[gid].out[12]));
-  ukey4[1] = swap_workaround (l32_from_64 (tmps[gid].out[12]));
-  ukey4[2] = swap_workaround (h32_from_64 (tmps[gid].out[13]));
-  ukey4[3] = swap_workaround (l32_from_64 (tmps[gid].out[13]));
-  ukey4[4] = swap_workaround (h32_from_64 (tmps[gid].out[14]));
-  ukey4[5] = swap_workaround (l32_from_64 (tmps[gid].out[14]));
-  ukey4[6] = swap_workaround (h32_from_64 (tmps[gid].out[15]));
-  ukey4[7] = swap_workaround (l32_from_64 (tmps[gid].out[15]));
+  ukey4[0] = swap32 (h32_from_64 (tmps[gid].out[12]));
+  ukey4[1] = swap32 (l32_from_64 (tmps[gid].out[12]));
+  ukey4[2] = swap32 (h32_from_64 (tmps[gid].out[13]));
+  ukey4[3] = swap32 (l32_from_64 (tmps[gid].out[13]));
+  ukey4[4] = swap32 (h32_from_64 (tmps[gid].out[14]));
+  ukey4[5] = swap32 (l32_from_64 (tmps[gid].out[14]));
+  ukey4[6] = swap32 (h32_from_64 (tmps[gid].out[15]));
+  ukey4[7] = swap32 (l32_from_64 (tmps[gid].out[15]));
 
   {
     tmp[0] = data[0];
@@ -687,25 +687,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06223_comp (__gl
 
   u32 ukey5[8];
 
-  ukey5[0] = swap_workaround (h32_from_64 (tmps[gid].out[16]));
-  ukey5[1] = swap_workaround (l32_from_64 (tmps[gid].out[16]));
-  ukey5[2] = swap_workaround (h32_from_64 (tmps[gid].out[17]));
-  ukey5[3] = swap_workaround (l32_from_64 (tmps[gid].out[17]));
-  ukey5[4] = swap_workaround (h32_from_64 (tmps[gid].out[18]));
-  ukey5[5] = swap_workaround (l32_from_64 (tmps[gid].out[18]));
-  ukey5[6] = swap_workaround (h32_from_64 (tmps[gid].out[19]));
-  ukey5[7] = swap_workaround (l32_from_64 (tmps[gid].out[19]));
+  ukey5[0] = swap32 (h32_from_64 (tmps[gid].out[16]));
+  ukey5[1] = swap32 (l32_from_64 (tmps[gid].out[16]));
+  ukey5[2] = swap32 (h32_from_64 (tmps[gid].out[17]));
+  ukey5[3] = swap32 (l32_from_64 (tmps[gid].out[17]));
+  ukey5[4] = swap32 (h32_from_64 (tmps[gid].out[18]));
+  ukey5[5] = swap32 (l32_from_64 (tmps[gid].out[18]));
+  ukey5[6] = swap32 (h32_from_64 (tmps[gid].out[19]));
+  ukey5[7] = swap32 (l32_from_64 (tmps[gid].out[19]));
 
   u32 ukey6[8];
 
-  ukey6[0] = swap_workaround (h32_from_64 (tmps[gid].out[20]));
-  ukey6[1] = swap_workaround (l32_from_64 (tmps[gid].out[20]));
-  ukey6[2] = swap_workaround (h32_from_64 (tmps[gid].out[21]));
-  ukey6[3] = swap_workaround (l32_from_64 (tmps[gid].out[21]));
-  ukey6[4] = swap_workaround (h32_from_64 (tmps[gid].out[22]));
-  ukey6[5] = swap_workaround (l32_from_64 (tmps[gid].out[22]));
-  ukey6[6] = swap_workaround (h32_from_64 (tmps[gid].out[23]));
-  ukey6[7] = swap_workaround (l32_from_64 (tmps[gid].out[23]));
+  ukey6[0] = swap32 (h32_from_64 (tmps[gid].out[20]));
+  ukey6[1] = swap32 (l32_from_64 (tmps[gid].out[20]));
+  ukey6[2] = swap32 (h32_from_64 (tmps[gid].out[21]));
+  ukey6[3] = swap32 (l32_from_64 (tmps[gid].out[21]));
+  ukey6[4] = swap32 (h32_from_64 (tmps[gid].out[22]));
+  ukey6[5] = swap32 (l32_from_64 (tmps[gid].out[22]));
+  ukey6[6] = swap32 (h32_from_64 (tmps[gid].out[23]));
+  ukey6[7] = swap32 (l32_from_64 (tmps[gid].out[23]));
 
   {
     tmp[0] = data[0];

OpenCL/m06231.cl

@@ -1559,22 +1559,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06231_init (__gl
 
   u32 salt_buf1[16];
 
-  salt_buf1[ 0] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 0]);
-  salt_buf1[ 1] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 1]);
-  salt_buf1[ 2] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 2]);
-  salt_buf1[ 3] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 3]);
-  salt_buf1[ 4] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 4]);
-  salt_buf1[ 5] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 5]);
-  salt_buf1[ 6] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 6]);
-  salt_buf1[ 7] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 7]);
-  salt_buf1[ 8] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 8]);
-  salt_buf1[ 9] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 9]);
-  salt_buf1[10] = swap_workaround (esalt_bufs[salt_pos].salt_buf[10]);
-  salt_buf1[11] = swap_workaround (esalt_bufs[salt_pos].salt_buf[11]);
-  salt_buf1[12] = swap_workaround (esalt_bufs[salt_pos].salt_buf[12]);
-  salt_buf1[13] = swap_workaround (esalt_bufs[salt_pos].salt_buf[13]);
-  salt_buf1[14] = swap_workaround (esalt_bufs[salt_pos].salt_buf[14]);
-  salt_buf1[15] = swap_workaround (esalt_bufs[salt_pos].salt_buf[15]);
+  salt_buf1[ 0] = swap32 (esalt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf1[ 1] = swap32 (esalt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf1[ 2] = swap32 (esalt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf1[ 3] = swap32 (esalt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[ 4] = swap32 (esalt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[ 5] = swap32 (esalt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[ 6] = swap32 (esalt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[ 7] = swap32 (esalt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf1[ 8] = swap32 (esalt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf1[ 9] = swap32 (esalt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf1[10] = swap32 (esalt_bufs[salt_pos].salt_buf[10]);
+  salt_buf1[11] = swap32 (esalt_bufs[salt_pos].salt_buf[11]);
+  salt_buf1[12] = swap32 (esalt_bufs[salt_pos].salt_buf[12]);
+  salt_buf1[13] = swap32 (esalt_bufs[salt_pos].salt_buf[13]);
+  salt_buf1[14] = swap32 (esalt_bufs[salt_pos].salt_buf[14]);
+  salt_buf1[15] = swap32 (esalt_bufs[salt_pos].salt_buf[15]);
 
   u32 salt_buf2[16];
 
@@ -1599,22 +1599,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06231_init (__gl
 
   u32 w[16];
 
-  w[ 0] = swap_workaround (w0[0]);
-  w[ 1] = swap_workaround (w0[1]);
-  w[ 2] = swap_workaround (w0[2]);
-  w[ 3] = swap_workaround (w0[3]);
-  w[ 4] = swap_workaround (w1[0]);
-  w[ 5] = swap_workaround (w1[1]);
-  w[ 6] = swap_workaround (w1[2]);
-  w[ 7] = swap_workaround (w1[3]);
-  w[ 8] = swap_workaround (w2[0]);
-  w[ 9] = swap_workaround (w2[1]);
-  w[10] = swap_workaround (w2[2]);
-  w[11] = swap_workaround (w2[3]);
-  w[12] = swap_workaround (w3[0]);
-  w[13] = swap_workaround (w3[1]);
-  w[14] = swap_workaround (w3[2]);
-  w[15] = swap_workaround (w3[3]);
+  w[ 0] = swap32 (w0[0]);
+  w[ 1] = swap32 (w0[1]);
+  w[ 2] = swap32 (w0[2]);
+  w[ 3] = swap32 (w0[3]);
+  w[ 4] = swap32 (w1[0]);
+  w[ 5] = swap32 (w1[1]);
+  w[ 6] = swap32 (w1[2]);
+  w[ 7] = swap32 (w1[3]);
+  w[ 8] = swap32 (w2[0]);
+  w[ 9] = swap32 (w2[1]);
+  w[10] = swap32 (w2[2]);
+  w[11] = swap32 (w2[3]);
+  w[12] = swap32 (w3[0]);
+  w[13] = swap32 (w3[1]);
+  w[14] = swap32 (w3[2]);
+  w[15] = swap32 (w3[3]);
 
   u32 ipad[16];
   u32 opad[16];
@@ -1914,25 +1914,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06231_comp (__gl
 
   u32 ukey1[8];
 
-  ukey1[0] = swap_workaround (tmps[gid].out[ 0]);
-  ukey1[1] = swap_workaround (tmps[gid].out[ 1]);
-  ukey1[2] = swap_workaround (tmps[gid].out[ 2]);
-  ukey1[3] = swap_workaround (tmps[gid].out[ 3]);
-  ukey1[4] = swap_workaround (tmps[gid].out[ 4]);
-  ukey1[5] = swap_workaround (tmps[gid].out[ 5]);
-  ukey1[6] = swap_workaround (tmps[gid].out[ 6]);
-  ukey1[7] = swap_workaround (tmps[gid].out[ 7]);
+  ukey1[0] = swap32 (tmps[gid].out[ 0]);
+  ukey1[1] = swap32 (tmps[gid].out[ 1]);
+  ukey1[2] = swap32 (tmps[gid].out[ 2]);
+  ukey1[3] = swap32 (tmps[gid].out[ 3]);
+  ukey1[4] = swap32 (tmps[gid].out[ 4]);
+  ukey1[5] = swap32 (tmps[gid].out[ 5]);
+  ukey1[6] = swap32 (tmps[gid].out[ 6]);
+  ukey1[7] = swap32 (tmps[gid].out[ 7]);
 
   u32 ukey2[8];
 
-  ukey2[0] = swap_workaround (tmps[gid].out[ 8]);
-  ukey2[1] = swap_workaround (tmps[gid].out[ 9]);
-  ukey2[2] = swap_workaround (tmps[gid].out[10]);
-  ukey2[3] = swap_workaround (tmps[gid].out[11]);
-  ukey2[4] = swap_workaround (tmps[gid].out[12]);
-  ukey2[5] = swap_workaround (tmps[gid].out[13]);
-  ukey2[6] = swap_workaround (tmps[gid].out[14]);
-  ukey2[7] = swap_workaround (tmps[gid].out[15]);
+  ukey2[0] = swap32 (tmps[gid].out[ 8]);
+  ukey2[1] = swap32 (tmps[gid].out[ 9]);
+  ukey2[2] = swap32 (tmps[gid].out[10]);
+  ukey2[3] = swap32 (tmps[gid].out[11]);
+  ukey2[4] = swap32 (tmps[gid].out[12]);
+  ukey2[5] = swap32 (tmps[gid].out[13]);
+  ukey2[6] = swap32 (tmps[gid].out[14]);
+  ukey2[7] = swap32 (tmps[gid].out[15]);
 
   u32 data[4];
 

OpenCL/m06232.cl

@@ -1559,22 +1559,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06232_init (__gl
 
   u32 salt_buf1[16];
 
-  salt_buf1[ 0] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 0]);
-  salt_buf1[ 1] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 1]);
-  salt_buf1[ 2] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 2]);
-  salt_buf1[ 3] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 3]);
-  salt_buf1[ 4] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 4]);
-  salt_buf1[ 5] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 5]);
-  salt_buf1[ 6] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 6]);
-  salt_buf1[ 7] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 7]);
-  salt_buf1[ 8] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 8]);
-  salt_buf1[ 9] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 9]);
-  salt_buf1[10] = swap_workaround (esalt_bufs[salt_pos].salt_buf[10]);
-  salt_buf1[11] = swap_workaround (esalt_bufs[salt_pos].salt_buf[11]);
-  salt_buf1[12] = swap_workaround (esalt_bufs[salt_pos].salt_buf[12]);
-  salt_buf1[13] = swap_workaround (esalt_bufs[salt_pos].salt_buf[13]);
-  salt_buf1[14] = swap_workaround (esalt_bufs[salt_pos].salt_buf[14]);
-  salt_buf1[15] = swap_workaround (esalt_bufs[salt_pos].salt_buf[15]);
+  salt_buf1[ 0] = swap32 (esalt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf1[ 1] = swap32 (esalt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf1[ 2] = swap32 (esalt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf1[ 3] = swap32 (esalt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[ 4] = swap32 (esalt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[ 5] = swap32 (esalt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[ 6] = swap32 (esalt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[ 7] = swap32 (esalt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf1[ 8] = swap32 (esalt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf1[ 9] = swap32 (esalt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf1[10] = swap32 (esalt_bufs[salt_pos].salt_buf[10]);
+  salt_buf1[11] = swap32 (esalt_bufs[salt_pos].salt_buf[11]);
+  salt_buf1[12] = swap32 (esalt_bufs[salt_pos].salt_buf[12]);
+  salt_buf1[13] = swap32 (esalt_bufs[salt_pos].salt_buf[13]);
+  salt_buf1[14] = swap32 (esalt_bufs[salt_pos].salt_buf[14]);
+  salt_buf1[15] = swap32 (esalt_bufs[salt_pos].salt_buf[15]);
 
   u32 salt_buf2[16];
 
@@ -1599,22 +1599,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06232_init (__gl
 
   u32 w[16];
 
-  w[ 0] = swap_workaround (w0[0]);
-  w[ 1] = swap_workaround (w0[1]);
-  w[ 2] = swap_workaround (w0[2]);
-  w[ 3] = swap_workaround (w0[3]);
-  w[ 4] = swap_workaround (w1[0]);
-  w[ 5] = swap_workaround (w1[1]);
-  w[ 6] = swap_workaround (w1[2]);
-  w[ 7] = swap_workaround (w1[3]);
-  w[ 8] = swap_workaround (w2[0]);
-  w[ 9] = swap_workaround (w2[1]);
-  w[10] = swap_workaround (w2[2]);
-  w[11] = swap_workaround (w2[3]);
-  w[12] = swap_workaround (w3[0]);
-  w[13] = swap_workaround (w3[1]);
-  w[14] = swap_workaround (w3[2]);
-  w[15] = swap_workaround (w3[3]);
+  w[ 0] = swap32 (w0[0]);
+  w[ 1] = swap32 (w0[1]);
+  w[ 2] = swap32 (w0[2]);
+  w[ 3] = swap32 (w0[3]);
+  w[ 4] = swap32 (w1[0]);
+  w[ 5] = swap32 (w1[1]);
+  w[ 6] = swap32 (w1[2]);
+  w[ 7] = swap32 (w1[3]);
+  w[ 8] = swap32 (w2[0]);
+  w[ 9] = swap32 (w2[1]);
+  w[10] = swap32 (w2[2]);
+  w[11] = swap32 (w2[3]);
+  w[12] = swap32 (w3[0]);
+  w[13] = swap32 (w3[1]);
+  w[14] = swap32 (w3[2]);
+  w[15] = swap32 (w3[3]);
 
   u32 ipad[16];
   u32 opad[16];
@@ -1914,25 +1914,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06232_comp (__gl
 
   u32 ukey1[8];
 
-  ukey1[0] = swap_workaround (tmps[gid].out[ 0]);
-  ukey1[1] = swap_workaround (tmps[gid].out[ 1]);
-  ukey1[2] = swap_workaround (tmps[gid].out[ 2]);
-  ukey1[3] = swap_workaround (tmps[gid].out[ 3]);
-  ukey1[4] = swap_workaround (tmps[gid].out[ 4]);
-  ukey1[5] = swap_workaround (tmps[gid].out[ 5]);
-  ukey1[6] = swap_workaround (tmps[gid].out[ 6]);
-  ukey1[7] = swap_workaround (tmps[gid].out[ 7]);
+  ukey1[0] = swap32 (tmps[gid].out[ 0]);
+  ukey1[1] = swap32 (tmps[gid].out[ 1]);
+  ukey1[2] = swap32 (tmps[gid].out[ 2]);
+  ukey1[3] = swap32 (tmps[gid].out[ 3]);
+  ukey1[4] = swap32 (tmps[gid].out[ 4]);
+  ukey1[5] = swap32 (tmps[gid].out[ 5]);
+  ukey1[6] = swap32 (tmps[gid].out[ 6]);
+  ukey1[7] = swap32 (tmps[gid].out[ 7]);
 
   u32 ukey2[8];
 
-  ukey2[0] = swap_workaround (tmps[gid].out[ 8]);
-  ukey2[1] = swap_workaround (tmps[gid].out[ 9]);
-  ukey2[2] = swap_workaround (tmps[gid].out[10]);
-  ukey2[3] = swap_workaround (tmps[gid].out[11]);
-  ukey2[4] = swap_workaround (tmps[gid].out[12]);
-  ukey2[5] = swap_workaround (tmps[gid].out[13]);
-  ukey2[6] = swap_workaround (tmps[gid].out[14]);
-  ukey2[7] = swap_workaround (tmps[gid].out[15]);
+  ukey2[0] = swap32 (tmps[gid].out[ 8]);
+  ukey2[1] = swap32 (tmps[gid].out[ 9]);
+  ukey2[2] = swap32 (tmps[gid].out[10]);
+  ukey2[3] = swap32 (tmps[gid].out[11]);
+  ukey2[4] = swap32 (tmps[gid].out[12]);
+  ukey2[5] = swap32 (tmps[gid].out[13]);
+  ukey2[6] = swap32 (tmps[gid].out[14]);
+  ukey2[7] = swap32 (tmps[gid].out[15]);
 
   u32 data[4];
 
@@ -1993,25 +1993,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06232_comp (__gl
 
   u32 ukey3[8];
 
-  ukey3[0] = swap_workaround (tmps[gid].out[16]);
-  ukey3[1] = swap_workaround (tmps[gid].out[17]);
-  ukey3[2] = swap_workaround (tmps[gid].out[18]);
-  ukey3[3] = swap_workaround (tmps[gid].out[19]);
-  ukey3[4] = swap_workaround (tmps[gid].out[20]);
-  ukey3[5] = swap_workaround (tmps[gid].out[21]);
-  ukey3[6] = swap_workaround (tmps[gid].out[22]);
-  ukey3[7] = swap_workaround (tmps[gid].out[23]);
+  ukey3[0] = swap32 (tmps[gid].out[16]);
+  ukey3[1] = swap32 (tmps[gid].out[17]);
+  ukey3[2] = swap32 (tmps[gid].out[18]);
+  ukey3[3] = swap32 (tmps[gid].out[19]);
+  ukey3[4] = swap32 (tmps[gid].out[20]);
+  ukey3[5] = swap32 (tmps[gid].out[21]);
+  ukey3[6] = swap32 (tmps[gid].out[22]);
+  ukey3[7] = swap32 (tmps[gid].out[23]);
 
   u32 ukey4[8];
 
-  ukey4[0] = swap_workaround (tmps[gid].out[24]);
-  ukey4[1] = swap_workaround (tmps[gid].out[25]);
-  ukey4[2] = swap_workaround (tmps[gid].out[26]);
-  ukey4[3] = swap_workaround (tmps[gid].out[27]);
-  ukey4[4] = swap_workaround (tmps[gid].out[28]);
-  ukey4[5] = swap_workaround (tmps[gid].out[29]);
-  ukey4[6] = swap_workaround (tmps[gid].out[30]);
-  ukey4[7] = swap_workaround (tmps[gid].out[31]);
+  ukey4[0] = swap32 (tmps[gid].out[24]);
+  ukey4[1] = swap32 (tmps[gid].out[25]);
+  ukey4[2] = swap32 (tmps[gid].out[26]);
+  ukey4[3] = swap32 (tmps[gid].out[27]);
+  ukey4[4] = swap32 (tmps[gid].out[28]);
+  ukey4[5] = swap32 (tmps[gid].out[29]);
+  ukey4[6] = swap32 (tmps[gid].out[30]);
+  ukey4[7] = swap32 (tmps[gid].out[31]);
 
   {
     tmp[0] = data[0];

OpenCL/m06233.cl

@@ -1559,22 +1559,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06233_init (__gl
 
   u32 salt_buf1[16];
 
-  salt_buf1[ 0] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 0]);
-  salt_buf1[ 1] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 1]);
-  salt_buf1[ 2] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 2]);
-  salt_buf1[ 3] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 3]);
-  salt_buf1[ 4] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 4]);
-  salt_buf1[ 5] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 5]);
-  salt_buf1[ 6] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 6]);
-  salt_buf1[ 7] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 7]);
-  salt_buf1[ 8] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 8]);
-  salt_buf1[ 9] = swap_workaround (esalt_bufs[salt_pos].salt_buf[ 9]);
-  salt_buf1[10] = swap_workaround (esalt_bufs[salt_pos].salt_buf[10]);
-  salt_buf1[11] = swap_workaround (esalt_bufs[salt_pos].salt_buf[11]);
-  salt_buf1[12] = swap_workaround (esalt_bufs[salt_pos].salt_buf[12]);
-  salt_buf1[13] = swap_workaround (esalt_bufs[salt_pos].salt_buf[13]);
-  salt_buf1[14] = swap_workaround (esalt_bufs[salt_pos].salt_buf[14]);
-  salt_buf1[15] = swap_workaround (esalt_bufs[salt_pos].salt_buf[15]);
+  salt_buf1[ 0] = swap32 (esalt_bufs[salt_pos].salt_buf[ 0]);
+  salt_buf1[ 1] = swap32 (esalt_bufs[salt_pos].salt_buf[ 1]);
+  salt_buf1[ 2] = swap32 (esalt_bufs[salt_pos].salt_buf[ 2]);
+  salt_buf1[ 3] = swap32 (esalt_bufs[salt_pos].salt_buf[ 3]);
+  salt_buf1[ 4] = swap32 (esalt_bufs[salt_pos].salt_buf[ 4]);
+  salt_buf1[ 5] = swap32 (esalt_bufs[salt_pos].salt_buf[ 5]);
+  salt_buf1[ 6] = swap32 (esalt_bufs[salt_pos].salt_buf[ 6]);
+  salt_buf1[ 7] = swap32 (esalt_bufs[salt_pos].salt_buf[ 7]);
+  salt_buf1[ 8] = swap32 (esalt_bufs[salt_pos].salt_buf[ 8]);
+  salt_buf1[ 9] = swap32 (esalt_bufs[salt_pos].salt_buf[ 9]);
+  salt_buf1[10] = swap32 (esalt_bufs[salt_pos].salt_buf[10]);
+  salt_buf1[11] = swap32 (esalt_bufs[salt_pos].salt_buf[11]);
+  salt_buf1[12] = swap32 (esalt_bufs[salt_pos].salt_buf[12]);
+  salt_buf1[13] = swap32 (esalt_bufs[salt_pos].salt_buf[13]);
+  salt_buf1[14] = swap32 (esalt_bufs[salt_pos].salt_buf[14]);
+  salt_buf1[15] = swap32 (esalt_bufs[salt_pos].salt_buf[15]);
 
   u32 salt_buf2[16];
 
@@ -1599,22 +1599,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06233_init (__gl
 
   u32 w[16];
 
-  w[ 0] = swap_workaround (w0[0]);
-  w[ 1] = swap_workaround (w0[1]);
-  w[ 2] = swap_workaround (w0[2]);
-  w[ 3] = swap_workaround (w0[3]);
-  w[ 4] = swap_workaround (w1[0]);
-  w[ 5] = swap_workaround (w1[1]);
-  w[ 6] = swap_workaround (w1[2]);
-  w[ 7] = swap_workaround (w1[3]);
-  w[ 8] = swap_workaround (w2[0]);
-  w[ 9] = swap_workaround (w2[1]);
-  w[10] = swap_workaround (w2[2]);
-  w[11] = swap_workaround (w2[3]);
-  w[12] = swap_workaround (w3[0]);
-  w[13] = swap_workaround (w3[1]);
-  w[14] = swap_workaround (w3[2]);
-  w[15] = swap_workaround (w3[3]);
+  w[ 0] = swap32 (w0[0]);
+  w[ 1] = swap32 (w0[1]);
+  w[ 2] = swap32 (w0[2]);
+  w[ 3] = swap32 (w0[3]);
+  w[ 4] = swap32 (w1[0]);
+  w[ 5] = swap32 (w1[1]);
+  w[ 6] = swap32 (w1[2]);
+  w[ 7] = swap32 (w1[3]);
+  w[ 8] = swap32 (w2[0]);
+  w[ 9] = swap32 (w2[1]);
+  w[10] = swap32 (w2[2]);
+  w[11] = swap32 (w2[3]);
+  w[12] = swap32 (w3[0]);
+  w[13] = swap32 (w3[1]);
+  w[14] = swap32 (w3[2]);
+  w[15] = swap32 (w3[3]);
 
   u32 ipad[16];
   u32 opad[16];
@@ -1914,25 +1914,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06233_comp (__gl
 
   u32 ukey1[8];
 
-  ukey1[0] = swap_workaround (tmps[gid].out[ 0]);
-  ukey1[1] = swap_workaround (tmps[gid].out[ 1]);
-  ukey1[2] = swap_workaround (tmps[gid].out[ 2]);
-  ukey1[3] = swap_workaround (tmps[gid].out[ 3]);
-  ukey1[4] = swap_workaround (tmps[gid].out[ 4]);
-  ukey1[5] = swap_workaround (tmps[gid].out[ 5]);
-  ukey1[6] = swap_workaround (tmps[gid].out[ 6]);
-  ukey1[7] = swap_workaround (tmps[gid].out[ 7]);
+  ukey1[0] = swap32 (tmps[gid].out[ 0]);
+  ukey1[1] = swap32 (tmps[gid].out[ 1]);
+  ukey1[2] = swap32 (tmps[gid].out[ 2]);
+  ukey1[3] = swap32 (tmps[gid].out[ 3]);
+  ukey1[4] = swap32 (tmps[gid].out[ 4]);
+  ukey1[5] = swap32 (tmps[gid].out[ 5]);
+  ukey1[6] = swap32 (tmps[gid].out[ 6]);
+  ukey1[7] = swap32 (tmps[gid].out[ 7]);
 
   u32 ukey2[8];
 
-  ukey2[0] = swap_workaround (tmps[gid].out[ 8]);
-  ukey2[1] = swap_workaround (tmps[gid].out[ 9]);
-  ukey2[2] = swap_workaround (tmps[gid].out[10]);
-  ukey2[3] = swap_workaround (tmps[gid].out[11]);
-  ukey2[4] = swap_workaround (tmps[gid].out[12]);
-  ukey2[5] = swap_workaround (tmps[gid].out[13]);
-  ukey2[6] = swap_workaround (tmps[gid].out[14]);
-  ukey2[7] = swap_workaround (tmps[gid].out[15]);
+  ukey2[0] = swap32 (tmps[gid].out[ 8]);
+  ukey2[1] = swap32 (tmps[gid].out[ 9]);
+  ukey2[2] = swap32 (tmps[gid].out[10]);
+  ukey2[3] = swap32 (tmps[gid].out[11]);
+  ukey2[4] = swap32 (tmps[gid].out[12]);
+  ukey2[5] = swap32 (tmps[gid].out[13]);
+  ukey2[6] = swap32 (tmps[gid].out[14]);
+  ukey2[7] = swap32 (tmps[gid].out[15]);
 
   u32 data[4];
 
@@ -1993,25 +1993,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06233_comp (__gl
 
   u32 ukey3[8];
 
-  ukey3[0] = swap_workaround (tmps[gid].out[16]);
-  ukey3[1] = swap_workaround (tmps[gid].out[17]);
-  ukey3[2] = swap_workaround (tmps[gid].out[18]);
-  ukey3[3] = swap_workaround (tmps[gid].out[19]);
-  ukey3[4] = swap_workaround (tmps[gid].out[20]);
-  ukey3[5] = swap_workaround (tmps[gid].out[21]);
-  ukey3[6] = swap_workaround (tmps[gid].out[22]);
-  ukey3[7] = swap_workaround (tmps[gid].out[23]);
+  ukey3[0] = swap32 (tmps[gid].out[16]);
+  ukey3[1] = swap32 (tmps[gid].out[17]);
+  ukey3[2] = swap32 (tmps[gid].out[18]);
+  ukey3[3] = swap32 (tmps[gid].out[19]);
+  ukey3[4] = swap32 (tmps[gid].out[20]);
+  ukey3[5] = swap32 (tmps[gid].out[21]);
+  ukey3[6] = swap32 (tmps[gid].out[22]);
+  ukey3[7] = swap32 (tmps[gid].out[23]);
 
   u32 ukey4[8];
 
-  ukey4[0] = swap_workaround (tmps[gid].out[24]);
-  ukey4[1] = swap_workaround (tmps[gid].out[25]);
-  ukey4[2] = swap_workaround (tmps[gid].out[26]);
-  ukey4[3] = swap_workaround (tmps[gid].out[27]);
-  ukey4[4] = swap_workaround (tmps[gid].out[28]);
-  ukey4[5] = swap_workaround (tmps[gid].out[29]);
-  ukey4[6] = swap_workaround (tmps[gid].out[30]);
-  ukey4[7] = swap_workaround (tmps[gid].out[31]);
+  ukey4[0] = swap32 (tmps[gid].out[24]);
+  ukey4[1] = swap32 (tmps[gid].out[25]);
+  ukey4[2] = swap32 (tmps[gid].out[26]);
+  ukey4[3] = swap32 (tmps[gid].out[27]);
+  ukey4[4] = swap32 (tmps[gid].out[28]);
+  ukey4[5] = swap32 (tmps[gid].out[29]);
+  ukey4[6] = swap32 (tmps[gid].out[30]);
+  ukey4[7] = swap32 (tmps[gid].out[31]);
 
   {
     tmp[0] = data[0];
@@ -2066,25 +2066,25 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06233_comp (__gl
 
   u32 ukey5[8];
 
-  ukey5[0] = swap_workaround (tmps[gid].out[32]);
-  ukey5[1] = swap_workaround (tmps[gid].out[33]);
-  ukey5[2] = swap_workaround (tmps[gid].out[34]);
-  ukey5[3] = swap_workaround (tmps[gid].out[35]);
-  ukey5[4] = swap_workaround (tmps[gid].out[36]);
-  ukey5[5] = swap_workaround (tmps[gid].out[37]);
-  ukey5[6] = swap_workaround (tmps[gid].out[38]);
-  ukey5[7] = swap_workaround (tmps[gid].out[39]);
+  ukey5[0] = swap32 (tmps[gid].out[32]);
+  ukey5[1] = swap32 (tmps[gid].out[33]);
+  ukey5[2] = swap32 (tmps[gid].out[34]);
+  ukey5[3] = swap32 (tmps[gid].out[35]);
+  ukey5[4] = swap32 (tmps[gid].out[36]);
+  ukey5[5] = swap32 (tmps[gid].out[37]);
+  ukey5[6] = swap32 (tmps[gid].out[38]);
+  ukey5[7] = swap32 (tmps[gid].out[39]);
 
   u32 ukey6[8];
 
-  ukey6[0] = swap_workaround (tmps[gid].out[40]);
-  ukey6[1] = swap_workaround (tmps[gid].out[41]);
-  ukey6[2] = swap_workaround (tmps[gid].out[42]);
-  ukey6[3] = swap_workaround (tmps[gid].out[43]);
-  ukey6[4] = swap_workaround (tmps[gid].out[44]);
-  ukey6[5] = swap_workaround (tmps[gid].out[45]);
-  ukey6[6] = swap_workaround (tmps[gid].out[46]);
-  ukey6[7] = swap_workaround (tmps[gid].out[47]);
+  ukey6[0] = swap32 (tmps[gid].out[40]);
+  ukey6[1] = swap32 (tmps[gid].out[41]);
+  ukey6[2] = swap32 (tmps[gid].out[42]);
+  ukey6[3] = swap32 (tmps[gid].out[43]);
+  ukey6[4] = swap32 (tmps[gid].out[44]);
+  ukey6[5] = swap32 (tmps[gid].out[45]);
+  ukey6[6] = swap32 (tmps[gid].out[46]);
+  ukey6[7] = swap32 (tmps[gid].out[47]);
 
   {
     tmp[0] = data[0];

OpenCL/m06300.cl

@@ -17,13 +17,8 @@
 #include "types_ocl.c"
 #include "common.c"
 
-#ifdef  VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
+#define COMPARE_S "check_single_comp4.c"
+#define COMPARE_M "check_multi_comp4.c"
 
 static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
 {

OpenCL/m06400.cl

@@ -17,17 +17,8 @@
 #include "types_ocl.c"
 #include "common.c"
 
-#ifdef  VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE2
-#define COMPARE_M "check_multi_vect2_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
+#define COMPARE_S "check_single_comp4.c"
+#define COMPARE_M "check_multi_comp4.c"
 
 __constant u32 k_sha256[64] =
 {
@@ -316,22 +307,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06400_init (__gl
    * pads
    */
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
-  w3[2] = swap_workaround (w3[2]);
-  w3[3] = swap_workaround (w3[3]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
+  w3[2] = swap32 (w3[2]);
+  w3[3] = swap32 (w3[3]);
 
   u32 ipad[8];
   u32 opad[8];
@@ -373,21 +364,21 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06400_init (__gl
   w3[2] = salt_buf3[2];
   // w3[3] = salt_buf3[3];
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
-  w3[2] = swap_workaround (w3[2]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
+  w3[2] = swap32 (w3[2]);
   w3[3] = (64 + salt_len + 4) * 8;
 
   u32 dgst[8];

OpenCL/m06500.cl

@@ -17,17 +17,8 @@
 #include "types_ocl.c"
 #include "common.c"
 
-#ifdef  VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE2
-#define COMPARE_M "check_multi_vect2_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
+#define COMPARE_S "check_single_comp4.c"
+#define COMPARE_M "check_multi_comp4.c"
 
 __constant u64 k_sha512[80] =
 {
@@ -320,22 +311,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06500_init (__gl
    * pads
    */
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
-  w3[2] = swap_workaround (w3[2]);
-  w3[3] = swap_workaround (w3[3]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
+  w3[2] = swap32 (w3[2]);
+  w3[3] = swap32 (w3[3]);
 
   u64 w0l[4];
   u64 w1l[4];
@@ -399,14 +390,14 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06500_init (__gl
   w3l[2] = 0;
   w3l[3] = 0;
 
-  w0l[0] = swap_workaround (w0l[0]);
-  w0l[1] = swap_workaround (w0l[1]);
-  w0l[2] = swap_workaround (w0l[2]);
-  w0l[3] = swap_workaround (w0l[3]);
-  w1l[0] = swap_workaround (w1l[0]);
-  w1l[1] = swap_workaround (w1l[1]);
-  w1l[2] = swap_workaround (w1l[2]);
-  w1l[3] = swap_workaround (w1l[3]);
+  w0l[0] = swap32 (w0l[0]);
+  w0l[1] = swap32 (w0l[1]);
+  w0l[2] = swap32 (w0l[2]);
+  w0l[3] = swap32 (w0l[3]);
+  w1l[0] = swap32 (w1l[0]);
+  w1l[1] = swap32 (w1l[1]);
+  w1l[2] = swap32 (w1l[2]);
+  w1l[3] = swap32 (w1l[3]);
   w2l[0] = 0;
   w2l[1] = 0;
   w2l[2] = 0;

OpenCL/m06600.cl

@@ -1123,22 +1123,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06600_init (__gl
    * pads
    */
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
-  w3[2] = swap_workaround (w3[2]);
-  w3[3] = swap_workaround (w3[3]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
+  w3[2] = swap32 (w3[2]);
+  w3[3] = swap32 (w3[3]);
 
   u32 ipad[5];
   u32 opad[5];
@@ -1177,10 +1177,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06600_init (__gl
   append_0x01_1x4 (w0, salt_len + 3);
   append_0x80_1x4 (w0, salt_len + 4);
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
   w1[0] = 0;
   w1[1] = 0;
   w1[2] = 0;

OpenCL/m06700.cl

@@ -17,17 +17,8 @@
 #include "types_ocl.c"
 #include "common.c"
 
-#ifdef  VECT_SIZE1
-#define COMPARE_M "check_multi_vect1_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE2
-#define COMPARE_M "check_multi_vect2_comp4.c"
-#endif
-
-#ifdef  VECT_SIZE4
-#define COMPARE_M "check_multi_vect4_comp4.c"
-#endif
+#define COMPARE_S "check_single_comp4.c"
+#define COMPARE_M "check_multi_comp4.c"
 
 static void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
 {
@@ -326,22 +317,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06700_init (__gl
    * pads
    */
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
-  w3[2] = swap_workaround (w3[2]);
-  w3[3] = swap_workaround (w3[3]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
+  w3[2] = swap32 (w3[2]);
+  w3[3] = swap32 (w3[3]);
 
   u32 ipad[5];
   u32 opad[5];
@@ -377,21 +368,21 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06700_init (__gl
   w3[2] = salt_buf3[2];
   //w3[3] = salt_buf3[3];
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
-  w3[2] = swap_workaround (w3[2]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
+  w3[2] = swap32 (w3[2]);
   w3[3] = (64 + salt_len + 4) * 8;
 
   u32 dgst[5];

OpenCL/m06800.cl

@@ -1255,22 +1255,22 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06800_init (__gl
    * pads
    */
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
-  w2[2] = swap_workaround (w2[2]);
-  w2[3] = swap_workaround (w2[3]);
-  w3[0] = swap_workaround (w3[0]);
-  w3[1] = swap_workaround (w3[1]);
-  w3[2] = swap_workaround (w3[2]);
-  w3[3] = swap_workaround (w3[3]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
+  w2[2] = swap32 (w2[2]);
+  w2[3] = swap32 (w2[3]);
+  w3[0] = swap32 (w3[0]);
+  w3[1] = swap32 (w3[1]);
+  w3[2] = swap32 (w3[2]);
+  w3[3] = swap32 (w3[3]);
 
   u32 ipad[8];
   u32 opad[8];
@@ -1315,16 +1315,16 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06800_init (__gl
   append_0x01_3x4 (w0, w1, w2, salt_len + 3);
   append_0x80_3x4 (w0, w1, w2, salt_len + 4);
 
-  w0[0] = swap_workaround (w0[0]);
-  w0[1] = swap_workaround (w0[1]);
-  w0[2] = swap_workaround (w0[2]);
-  w0[3] = swap_workaround (w0[3]);
-  w1[0] = swap_workaround (w1[0]);
-  w1[1] = swap_workaround (w1[1]);
-  w1[2] = swap_workaround (w1[2]);
-  w1[3] = swap_workaround (w1[3]);
-  w2[0] = swap_workaround (w2[0]);
-  w2[1] = swap_workaround (w2[1]);
+  w0[0] = swap32 (w0[0]);
+  w0[1] = swap32 (w0[1]);
+  w0[2] = swap32 (w0[2]);
+  w0[3] = swap32 (w0[3]);
+  w1[0] = swap32 (w1[0]);
+  w1[1] = swap32 (w1[1]);
+  w1[2] = swap32 (w1[2]);
+  w1[3] = swap32 (w1[3]);
+  w2[0] = swap32 (w2[0]);
+  w2[1] = swap32 (w2[1]);
   w2[2] = 0;
   w2[3] = 0;
   w3[0] = 0;
@@ -1596,10 +1596,10 @@ __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m06800_comp (__gl
   salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
   salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
 
-  out[0] = swap_workaround (out[0]);
-  out[1] = swap_workaround (out[1]);
-  out[2] = swap_workaround (out[2]);
-  out[3] = swap_workaround (out[3]);
+  out[0] = swap32 (out[0]);
+  out[1] = swap32 (out[1]);
+  out[2] = swap32 (out[2]);
+  out[3] = swap32 (out[3]);
 
   truncate_block (out, salt_len);