Extract one 64-bit value for SSE.

Instead of extracting 4 16-bit value, we directly extract a 64-bit
value.

configure

@@ -4,7 +4,7 @@ case `uname -s` in
 
   Darwin)
     echo 'CXX = clang++' > Makefile.rules
-    echo 'CXXFLAGS_PLATFORM = ' >> Makefile.rules
+    echo 'CXXFLAGS_PLATFORM = -march=native -msse4.2' >> Makefile.rules
     echo 'LFLAGS = -flto' >> Makefile.rules
     ;;
 

rotation.cpp

@@ -13,6 +13,7 @@
 #include <xmmintrin.h>
 #include <emmintrin.h>
 #include <tmmintrin.h>
+#include <smmintrin.h>
 
 #include "image.h"
 
@@ -363,7 +364,7 @@ void fill_row(Image const& src,
 inline
 void interpolate_row(uint8_t* row_buffer, uint8_t* row_coefs,
                      unsigned int row_index,
-                     pvalue_t* rotate_buffer, unsigned int rot_index)
+                     uint64_t* rotate_buffer, unsigned int rot_index)
 {
   __m128i pixels = _mm_loadu_si128((__m128i*) &row_buffer[row_index]);
   __m128i coefs = _mm_loadu_si128((__m128i*) &row_coefs[row_index]);
@@ -374,10 +375,7 @@ void interpolate_row(uint8_t* row_buffer, uint8_t* row_coefs,
   pixels = _mm_hadd_epi16(pixels, zero); // 1 bin per pixel, 4 pixels
   pixels = _mm_srli_epi16(pixels, 6);
 
-  rotate_buffer[rot_index] = _mm_extract_epi16(pixels, 0);
-  rotate_buffer[rot_index + 1] = _mm_extract_epi16(pixels, 1);
-  rotate_buffer[rot_index + 2] = _mm_extract_epi16(pixels, 2);
-  rotate_buffer[rot_index + 3] = _mm_extract_epi16(pixels, 3);
+  rotate_buffer[rot_index / 4] = _mm_extract_epi64(pixels, 0);
 }
 
 inline
@@ -443,6 +441,7 @@ Image* rotate(Image const& src, double angle)
 
   unsigned int buffer_index = 0;
   pvalue_t* buffer = rotated->buffer;
+  uint64_t* buffer64 = (uint64_t*) rotated->buffer;
 
   int const width = rotated->width;
   int const height = rotated->height;
@@ -508,7 +507,7 @@ Image* rotate(Image const& src, double angle)
     // We process 4 pixels at a time
     for (int x = 0; x < core_pixels / 4; ++x)
     {
-      interpolate_row(row_buffer.get(), row_coefs.get(), x * 16, buffer, buffer_index);
+      interpolate_row(row_buffer.get(), row_coefs.get(), x * 16, buffer64, buffer_index);
       buffer_index += 4;
     }
     buffer_index += core_pixels % 4;