stdin

stdin

diff --git a/libavutil/imgutils.c b/libavutil/imgutils.c
index 37808e5..96e94bf 100644
--- a/libavutil/imgutils.c
+++ b/libavutil/imgutils.c
@@ -31,6 +31,8 @@
 #include "pixdesc.h"
 #include "rational.h"
 
+#include <smmintrin.h>
+
 void av_image_fill_max_pixsteps(int max_pixsteps[4], int max_pixstep_comps[4],
                                 const AVPixFmtDescriptor *pixdesc)
 {
@@ -284,6 +286,58 @@ int av_image_check_sar(unsigned int w, unsigned int h, AVRational sar)
     return AVERROR(EINVAL);
 }
 
+static void copy_from_uswc(uint8_t *pDest, uint32_t dstPitch,
+                           const uint8_t *pSrc, uint32_t srcPitch,
+                           uint32_t width, uint32_t height)
+{
+    register __m128i x0 __asm ("xmm1");
+    register __m128i x1 __asm ("xmm2");
+    register __m128i x2 __asm ("xmm3");
+    register __m128i x3 __asm ("xmm4");
+
+    uint32_t y;
+
+    _mm_mfence();
+
+    for (y = 0; y < height; ++y)
+    {
+        const uint32_t unaligned = ((uintptr_t)pSrc) & 0x0f;
+        uint32_t x = 0;
+
+        for (; x < unaligned; ++x)
+            pDest[x] = pSrc[x];
+
+        for (; x + 63 < width; x += 64)
+        {
+            x0 = _mm_stream_load_si128((__m128i*)(&pSrc[x] + 0));
+            x1 = _mm_stream_load_si128((__m128i*)(&pSrc[x] + 16));
+            x2 = _mm_stream_load_si128((__m128i*)(&pSrc[x] + 32));
+            x3 = _mm_stream_load_si128((__m128i*)(&pSrc[x] + 48));
+
+            if (!unaligned)
+            {
+                _mm_store_si128((__m128i*)(&pDest[x] + 0), x0);
+                _mm_store_si128((__m128i*)(&pDest[x] + 16), x1);
+                _mm_store_si128((__m128i*)(&pDest[x] + 32), x2);
+                _mm_store_si128((__m128i*)(&pDest[x] + 48), x3);
+            }
+            else
+            {
+                _mm_storeu_si128((__m128i*)(&pDest[x] + 0), x0);
+                _mm_storeu_si128((__m128i*)(&pDest[x] + 16), x1);
+                _mm_storeu_si128((__m128i*)(&pDest[x] + 32), x2);
+                _mm_storeu_si128((__m128i*)(&pDest[x] + 48), x3);
+            }
+        }
+
+        for (; x < width; ++x)
+            pDest[x] = pSrc[x];
+
+        pDest += dstPitch;
+        pSrc += srcPitch;
+    }
+}
+
 void av_image_copy_plane(uint8_t       *dst, int dst_linesize,
                          const uint8_t *src, int src_linesize,
                          int bytewidth, int height)
@@ -292,11 +346,15 @@ void av_image_copy_plane(uint8_t       *dst, int dst_linesize,
         return;
     av_assert0(abs(src_linesize) >= bytewidth);
     av_assert0(abs(dst_linesize) >= bytewidth);
+#if 1
+    copy_from_uswc(dst, dst_linesize, src, src_linesize, bytewidth, height);
+#else
     for (;height > 0; height--) {
         memcpy(dst, src, bytewidth);
         dst += dst_linesize;
         src += src_linesize;
     }
+#endif
 }
 
 void av_image_copy(uint8_t *dst_data[4], int dst_linesizes[4],