Add two functions md5_update_global_utf16le_swap() and md5_update_global_swap() for later use

OpenCL/inc_hash_md5.cl

@@ -320,6 +320,92 @@ void md5_update_global (md5_ctx_t *ctx, const __global u32 *w, const int len)
   md5_update_64 (ctx, w0, w1, w2, w3, len - pos1);
 }
 
+void md5_update_global_swap (md5_ctx_t *ctx, const __global u32 *w, const int len)
+{
+  u32 w0[4];
+  u32 w1[4];
+  u32 w2[4];
+  u32 w3[4];
+
+  int pos1;
+  int pos4;
+
+  for (pos1 = 0, pos4 = 0; pos1 < len - 64; pos1 += 64, pos4 += 16)
+  {
+    w0[0] = w[pos4 +  0];
+    w0[1] = w[pos4 +  1];
+    w0[2] = w[pos4 +  2];
+    w0[3] = w[pos4 +  3];
+    w1[0] = w[pos4 +  4];
+    w1[1] = w[pos4 +  5];
+    w1[2] = w[pos4 +  6];
+    w1[3] = w[pos4 +  7];
+    w2[0] = w[pos4 +  8];
+    w2[1] = w[pos4 +  9];
+    w2[2] = w[pos4 + 10];
+    w2[3] = w[pos4 + 11];
+    w3[0] = w[pos4 + 12];
+    w3[1] = w[pos4 + 13];
+    w3[2] = w[pos4 + 14];
+    w3[3] = w[pos4 + 15];
+
+    w0[0] = swap32_S (w0[0]);
+    w0[1] = swap32_S (w0[1]);
+    w0[2] = swap32_S (w0[2]);
+    w0[3] = swap32_S (w0[3]);
+    w1[0] = swap32_S (w1[0]);
+    w1[1] = swap32_S (w1[1]);
+    w1[2] = swap32_S (w1[2]);
+    w1[3] = swap32_S (w1[3]);
+    w2[0] = swap32_S (w2[0]);
+    w2[1] = swap32_S (w2[1]);
+    w2[2] = swap32_S (w2[2]);
+    w2[3] = swap32_S (w2[3]);
+    w3[0] = swap32_S (w3[0]);
+    w3[1] = swap32_S (w3[1]);
+    w3[2] = swap32_S (w3[2]);
+    w3[3] = swap32_S (w3[3]);
+
+    md5_update_64 (ctx, w0, w1, w2, w3, 64);
+  }
+
+  w0[0] = w[pos4 +  0];
+  w0[1] = w[pos4 +  1];
+  w0[2] = w[pos4 +  2];
+  w0[3] = w[pos4 +  3];
+  w1[0] = w[pos4 +  4];
+  w1[1] = w[pos4 +  5];
+  w1[2] = w[pos4 +  6];
+  w1[3] = w[pos4 +  7];
+  w2[0] = w[pos4 +  8];
+  w2[1] = w[pos4 +  9];
+  w2[2] = w[pos4 + 10];
+  w2[3] = w[pos4 + 11];
+  w3[0] = w[pos4 + 12];
+  w3[1] = w[pos4 + 13];
+  w3[2] = w[pos4 + 14];
+  w3[3] = w[pos4 + 15];
+
+  w0[0] = swap32_S (w0[0]);
+  w0[1] = swap32_S (w0[1]);
+  w0[2] = swap32_S (w0[2]);
+  w0[3] = swap32_S (w0[3]);
+  w1[0] = swap32_S (w1[0]);
+  w1[1] = swap32_S (w1[1]);
+  w1[2] = swap32_S (w1[2]);
+  w1[3] = swap32_S (w1[3]);
+  w2[0] = swap32_S (w2[0]);
+  w2[1] = swap32_S (w2[1]);
+  w2[2] = swap32_S (w2[2]);
+  w2[3] = swap32_S (w2[3]);
+  w3[0] = swap32_S (w3[0]);
+  w3[1] = swap32_S (w3[1]);
+  w3[2] = swap32_S (w3[2]);
+  w3[3] = swap32_S (w3[3]);
+
+  md5_update_64 (ctx, w0, w1, w2, w3, len - pos1);
+}
+
 void md5_update_global_utf16le (md5_ctx_t *ctx, const __global u32 *w, const int len)
 {
   u32 w0[4];
@@ -362,6 +448,82 @@ void md5_update_global_utf16le (md5_ctx_t *ctx, const __global u32 *w, const int
   md5_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
 }
 
+void md5_update_global_utf16le_swap (md5_ctx_t *ctx, const __global u32 *w, const int len)
+{
+  u32 w0[4];
+  u32 w1[4];
+  u32 w2[4];
+  u32 w3[4];
+
+  int pos1;
+  int pos4;
+
+  for (pos1 = 0, pos4 = 0; pos1 < len - 32; pos1 += 32, pos4 += 8)
+  {
+    w0[0] = w[pos4 + 0];
+    w0[1] = w[pos4 + 1];
+    w0[2] = w[pos4 + 2];
+    w0[3] = w[pos4 + 3];
+    w1[0] = w[pos4 + 4];
+    w1[1] = w[pos4 + 5];
+    w1[2] = w[pos4 + 6];
+    w1[3] = w[pos4 + 7];
+
+    make_utf16le_S (w1, w2, w3);
+    make_utf16le_S (w0, w0, w1);
+
+    w0[0] = swap32_S (w0[0]);
+    w0[1] = swap32_S (w0[1]);
+    w0[2] = swap32_S (w0[2]);
+    w0[3] = swap32_S (w0[3]);
+    w1[0] = swap32_S (w1[0]);
+    w1[1] = swap32_S (w1[1]);
+    w1[2] = swap32_S (w1[2]);
+    w1[3] = swap32_S (w1[3]);
+    w2[0] = swap32_S (w2[0]);
+    w2[1] = swap32_S (w2[1]);
+    w2[2] = swap32_S (w2[2]);
+    w2[3] = swap32_S (w2[3]);
+    w3[0] = swap32_S (w3[0]);
+    w3[1] = swap32_S (w3[1]);
+    w3[2] = swap32_S (w3[2]);
+    w3[3] = swap32_S (w3[3]);
+
+    md5_update_64 (ctx, w0, w1, w2, w3, 32 * 2);
+  }
+
+  w0[0] = w[pos4 + 0];
+  w0[1] = w[pos4 + 1];
+  w0[2] = w[pos4 + 2];
+  w0[3] = w[pos4 + 3];
+  w1[0] = w[pos4 + 4];
+  w1[1] = w[pos4 + 5];
+  w1[2] = w[pos4 + 6];
+  w1[3] = w[pos4 + 7];
+
+  make_utf16le_S (w1, w2, w3);
+  make_utf16le_S (w0, w0, w1);
+
+  w0[0] = swap32_S (w0[0]);
+  w0[1] = swap32_S (w0[1]);
+  w0[2] = swap32_S (w0[2]);
+  w0[3] = swap32_S (w0[3]);
+  w1[0] = swap32_S (w1[0]);
+  w1[1] = swap32_S (w1[1]);
+  w1[2] = swap32_S (w1[2]);
+  w1[3] = swap32_S (w1[3]);
+  w2[0] = swap32_S (w2[0]);
+  w2[1] = swap32_S (w2[1]);
+  w2[2] = swap32_S (w2[2]);
+  w2[3] = swap32_S (w2[3]);
+  w3[0] = swap32_S (w3[0]);
+  w3[1] = swap32_S (w3[1]);
+  w3[2] = swap32_S (w3[2]);
+  w3[3] = swap32_S (w3[3]);
+
+  md5_update_64 (ctx, w0, w1, w2, w3, (len - pos1) * 2);
+}
+
 void md5_final (md5_ctx_t *ctx)
 {
   const int pos = ctx->len & 63;