Decompose complex lines.

Create a bunch of const variables to better profile.

.gitignore

@@ -3,6 +3,7 @@
 *.swp
 *.o
 
+rotation
 rotated*.ppm
 rotated*.pnm
 *.png

Makefile

@@ -1,9 +1,8 @@
 include Makefile.rules
 
 CXXFLAGS  = -std=c++11 -W -Wall -O3 -ffast-math -g $(CXXFLAGS_PLAFORM)
-LFLAGS    = #-flto
 DEFINES   = #-DSIMD
-BUILD_DIR = /tmp
+BUILD_DIR = .
 SRC       = rotation.cpp \
 	    image.cpp \
             pnm.cpp
@@ -19,7 +18,7 @@ all: $(OBJS)
 	$(CXX) $(CXXFLAGS) $(DEFINES) $< -c -o $@
 
 clean:
-	@rm -f *~ *.o .*.swp *.ppm *.pgm *.pnm cachegrind.out.*
+	@rm -f rotation *~ *.o .*.swp *.ppm *.pgm *.pnm cachegrind.out.* callgrind.out.*
 
 run: all
 	$(BUILD_DIR)/rotation $(IMG)

configure

@@ -4,12 +4,14 @@ case `uname -s` in
 
   Darwin)
     echo 'CXX = clang++' > Makefile.rules
-    echo 'CXXFLAGS_PLATFORM = ' > Makefile.rules
+    echo 'CXXFLAGS_PLATFORM = ' >> Makefile.rules
+    echo 'LFLAGS = -flto' >> Makefile.rules
     ;;
 
   Linux)
     echo 'CXX = g++' > Makefile.rules
-    echo 'CXXFLAGS_PLATFORM = ' > Makefile.rules
+    echo 'CXXFLAGS_PLATFORM = ' >> Makefile.rules
+    echo 'LFLAGS = ' >> Makefile.rules
     ;;
 
 esac

rotation.cpp

@@ -11,6 +11,7 @@
 
 #include <xmmintrin.h>
 #include <emmintrin.h>
+#include <tmmintrin.h>
 
 #include "image.h"
 
@@ -177,6 +178,7 @@ void rotate_pixel(Image const& src,
 
   // Bilinear interpolation
   unsigned int const src_index_1 = (src_y * src.width + src_x) * src.pixel_size;
+  unsigned int const src_index_2 = src_index_1 + src.pixel_size;
   unsigned int const src_index_3 = src_index_1 + src.pixel_size * src.width;
   unsigned int const src_index_4 = src_index_3 + src.pixel_size;
 
@@ -186,6 +188,11 @@ void rotate_pixel(Image const& src,
     return;
   }
 
+  pvalue_t const src_tl = src.buffer[src_index_1];
+  pvalue_t const src_tr = src.buffer[src_index_2];
+  pvalue_t const src_bl = src.buffer[src_index_3];
+  pvalue_t const src_br = src.buffer[src_index_4];
+
   unsigned int const x_delta = (src_rotated_point.x >> (q_pow - q_inter_pow)) & mask;
   unsigned int const y_delta = (src_rotated_point.y >> (q_pow - q_inter_pow)) & mask;
   unsigned int const inv_x = q_inter - x_delta;
@@ -193,10 +200,9 @@ void rotate_pixel(Image const& src,
 
 #ifndef SIMD
 
-  unsigned int src_index_2 = src_index_1 + src.pixel_size;
+  pvalue_t interpolated = ((src_tl * inv_x + src_tr * x_delta) * inv_y
-
+                         + (src_bl * inv_x + src_br * x_delta) * y_delta) >> (q_inter_pow << 1);
-  rotate_buffer[rot_index] = ((src.buffer[src_index_1] * inv_x + src.buffer[src_index_2] * x_delta) * inv_y
+  rotate_buffer[rot_index] = interpolated;
-                            + (src.buffer[src_index_3] * inv_x + src.buffer[src_index_4] * x_delta) * y_delta) >> (q_inter_pow << 1);
   // rotate_buffer[rot_index + 1] = ((src.buffer[src_index_1 + 1] * inv_x + src.buffer[src_index_2 + 1] * x_delta) * inv_y
   //                               + (src.buffer[src_index_3 + 1] * inv_x + src.buffer[src_index_4 + 1] * x_delta) * y_delta) >> 6;
   // rotate_buffer[rot_index + 2] = ((src.buffer[src_index_1 + 2] * inv_x + src.buffer[src_index_2 + 2] * x_delta) * inv_y
@@ -260,13 +266,16 @@ Image* rotate(Image const& src, double angle)
   int64_t const src_qheight = (int64_t) src.height * q_pos;
 
   unsigned int const src_limit = src.width * src.height * src.pixel_size;
+  size_t const buffer_step = rotated->pixel_size;
+  int const width = rotated->width;
+  int const height = rotated->height;
 
-  for (int y = 0; y < (int) rotated->height; ++y)
+  for (int y = 0; y < height; ++y)
   {
     Point src_rotated_point((rot_origin_in_src.x * q_pos) + y * qdy.x,
                             (rot_origin_in_src.y * q_pos) + y * qdy.y);
 
-    for (unsigned int x = 0; x < rotated->width; ++x)
+    for (int x = 0; x < width; ++x)
     {
       if (src_rotated_point.x >= 0 && src_rotated_point.x < src_qwidth
           && src_rotated_point.y >= 0 && src_rotated_point.y < src_qheight)
@@ -278,8 +287,7 @@ Image* rotate(Image const& src, double angle)
       }
 
       src_rotated_point += qdx;
-
+      buffer_index += buffer_step;
-      buffer_index += rotated->pixel_size;
     }
   }