[host] Added SMID BGRAToRGB support and implemented it in DXGI

host/Util.h

@@ -18,6 +18,9 @@ Place, Suite 330, Boston, MA 02111-1307 USA
 
 #pragma once
 #include <string>
+#include <assert.h>
+#include <inttypes.h>
+#include <tmmintrin.h>
 
 #include "common\debug.h"
 
@@ -56,4 +59,45 @@ public:
 #endif
     return defaultPath;
   }
+
+  static void BGRAtoRGB(uint8_t * orig, size_t imagesize, uint8_t * dest)
+  {
+    assert((uintptr_t)orig % 16 == 0);
+    assert((uintptr_t)dest % 16 == 0);
+    assert(imagesize % 16 == 0);
+
+    __m128i mask_right = _mm_set_epi8
+    (
+        12,   13,   14,    8,
+         9,   10,    4,    5, 
+         6,    0,    1,    2,
+      -128, -128, -128, -128
+    );
+
+    __m128i mask_left = _mm_set_epi8
+    (
+      -128, -128, -128, -128,
+        12,   13,   14,    8,
+         9,   10,    4,    5, 
+         6,    0,    1,    2
+    );
+
+
+    uint8_t *end = orig + imagesize * 4;
+    for (; orig != end; orig += 64, dest += 48)
+    {
+      __m128i v0 = _mm_shuffle_epi8(_mm_load_si128((__m128i *)&orig[0 ]), mask_right);
+      __m128i v1 = _mm_shuffle_epi8(_mm_load_si128((__m128i *)&orig[16]), mask_left );
+      __m128i v2 = _mm_shuffle_epi8(_mm_load_si128((__m128i *)&orig[32]), mask_left );
+      __m128i v3 = _mm_shuffle_epi8(_mm_load_si128((__m128i *)&orig[48]), mask_left );
+
+      v0 = _mm_alignr_epi8(v1, v0, 4);
+      v1 = _mm_alignr_epi8(v2, _mm_slli_si128(v1, 4),  8);
+      v2 = _mm_alignr_epi8(v3, _mm_slli_si128(v2, 4), 12);
+      
+      _mm_stream_si128((__m128i *)&dest[0 ], v0);
+      _mm_stream_si128((__m128i *)&dest[16], v1);
+      _mm_stream_si128((__m128i *)&dest[32], v2);
+    }
+  }
 };