Fall back to old rotr64 optimization for AMD

OpenCL/inc_types.cl

@@ -198,9 +198,17 @@ u32 rotl32_S (const u32 a, const u32 n)
   return rotate (a, n);
 }
 
-u64 rotr64_S (const u64 a, const u32 n)
+inline u64 rotr64_S (const u64 a, const u32 n)
 {
-  return rotate (a, (u64) (64 - n));
+  const u32 a0 = h32_from_64_S (a);
+  const u32 a1 = l32_from_64_S (a);
+
+  const u32 t0 = (n >= 32) ? amd_bitalign (a0, a1, n - 32) : amd_bitalign (a1, a0, n);
+  const u32 t1 = (n >= 32) ? amd_bitalign (a1, a0, n - 32) : amd_bitalign (a0, a1, n);
+
+  const u64 r = hl32_to_64_S (t0, t1);
+
+  return r;
 }
 
 u64 rotl64_S (const u64 a, const u32 n)
@@ -232,9 +240,17 @@ u32x rotl32 (const u32x a, const u32 n)
   return rotate (a, n);
 }
 
-u64x rotr64 (const u64x a, const u32 n)
+inline u64x rotr64 (const u64x a, const u32 n)
 {
-  return rotate (a, (u64x) (64 - n));
+  const u32x a0 = h32_from_64 (a);
+  const u32x a1 = l32_from_64 (a);
+
+  const u32x t0 = (n >= 32) ? amd_bitalign (a0, a1, n - 32) : amd_bitalign (a1, a0, n);
+  const u32x t1 = (n >= 32) ? amd_bitalign (a1, a0, n - 32) : amd_bitalign (a0, a1, n);
+
+  const u64x r = hl32_to_64 (t0, t1);
+
+  return r;
 }
 
 u64x rotl64 (const u64x a, const u32 n)