Faster SpatialDownSampling

1 files changed, 4 insertions, 4 deletions

diff --git a/generic/SpatialDownSampling.c b/generic/SpatialDownSampling.c
index 9dc82bc..0feef5d 100644
--- a/generic/SpatialDownSampling.c
+++ b/generic/SpatialDownSampling.c
@@ -25,7 +25,7 @@ static int nn_(SpatialDownSampling_updateOutput)(lua_State *L) {
   real *output_data = THTensor_(data)(output);
 
   // resample each plane
-  real factor = 1.0f / (rH*rW), avg;
+  real avg;
   real *input_p = input_data, *output_p = output_data;
   int k, x, y, i, j;
   for (k = 0; k < ichannels; ++k, input_p += is[0], output_p += os[0])
@@ -35,9 +35,9 @@ static int nn_(SpatialDownSampling_updateOutput)(lua_State *L) {
 	for (i = y*rH; i < (y+1)*rH; ++i)
 	  for (j = x*rW; j < (x+1)*rW; ++j)
 	    avg += input_p[i*is[1]+j*is[2]];
-	avg *= factor;
         output_p[y*os[1] + x*os[2]] = avg;
       }
+  THTensor_(mul)(output, output, 1.0f/(rH*rW));
   return 1;
 }
 
@@ -64,7 +64,6 @@ static int nn_(SpatialDownSampling_updateGradInput)(lua_State *L) {
   real *gradOutput_data = THTensor_(data)(gradOutput);
 
   // compute gradients for each plane
-  real factor = 1.0f / (rH*rW);
   real *gradInput_p = gradInput_data, *gradOutput_p = gradOutput_data;
   int k, x, y, i, j;
   for (k = 0; k < ochannels; ++k, gradInput_p += gis[0], gradOutput_p += gos[0])
@@ -72,7 +71,8 @@ static int nn_(SpatialDownSampling_updateGradInput)(lua_State *L) {
       for (x = 0; x < owidth; ++x)
 	for (i = y*rH; i < (y+1)*rH; ++i)
 	  for (j = x*rW; j < (x+1)*rW; ++j)
-	    gradInput_p[i*gis[1]+j*gis[2]] += gradOutput_p[y*gos[1]+x*gos[2]] * factor;
+	    gradInput_p[i*gis[1]+j*gis[2]] += gradOutput_p[y*gos[1]+x*gos[2]];
+  THTensor_(mul)(gradInput, gradInput, 1.0f/(rH*rW));
   
   return 1;
 }