Merge pull request #321 from raven02/patch-2

gl4: Swap high and low 16 bits within a 32 bit word

src/xenia/base/memory.h

@@ -46,6 +46,8 @@ void copy_and_swap_64_aligned(uint64_t* dest, const uint64_t* src,
                               size_t count);
 void copy_and_swap_64_unaligned(uint64_t* dest, const uint64_t* src,
                                 size_t count);
+void copy_and_swap_16_in_32_aligned(uint32_t* dest, const uint32_t* src,
+                                    size_t count);
 
 template <typename T>
 void copy_and_swap(T* dest, const T* src, size_t count) {

src/xenia/base/memory_generic.cc

@@ -132,4 +132,18 @@ void copy_and_swap_64_unaligned(uint64_t* dest, const uint64_t* src,
   }
 }
 
+void copy_and_swap_16_in_32_aligned(uint32_t* dest, const uint32_t* src,
+                                    size_t count) {
+  size_t i;
+  __m128i input, output;
+  for (i = 0; i + 4 <= count; i += 4) {
+    input = _mm_loadu_si128((__m128i*)&src[i]);
+    output = _mm_or_si128(_mm_slli_epi32(input, 16), _mm_srli_epi32(input, 16));
+    _mm_storeu_si128((__m128i*)&dest[i], output);
+  }
+  for (; i < count; ++i) {  // handle residual elements
+    dest[i] = (src[i] >> 16) | (src[i] << 16);
+  }
+}
+
 }  // namespace xe

src/xenia/gpu/gl4/texture_cache.cc

@@ -687,13 +687,10 @@ void TextureSwap(Endian endianness, void* dest, const void* src,
                                    reinterpret_cast<const uint32_t*>(src),
                                    length / 4);
       break;
-    case Endian::k16in32:
+    case Endian::k16in32:  // Swap high and low 16 bits within a 32 bit word
-      // TODO(benvanik): make more efficient.
+      xe::copy_and_swap_16_in_32_aligned(reinterpret_cast<uint32_t*>(dest),
-      /*for (uint32_t i = 0; i < length; i += 4, src += 4, dest += 4) {
+                                         reinterpret_cast<const uint32_t*>(src),
-        uint32_t value = *(uint32_t*)src;
+                                         length);
-        *(uint32_t*)dest = ((value >> 16) & 0xFFFF) | (value << 16);
-      }*/
-      assert_always("16in32 not supported");
       break;
     default:
     case Endian::kUnspecified: