Fix incorrect indirection size computation for DWCONV

Slightly reduce memory footprint for depthwise convolutions PiperOrigin-RevId: 283990702
google · Dec 5, 2019 · 03ff294 · 03ff294
1 parent ad74a7b
commit 03ff294
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/bench/f32-dwconv.cc b/bench/f32-dwconv.cc
@@ -66,7 +66,7 @@ static void DWConvBenchmark(benchmark::State& state,
   const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
   const size_t output_size = output_height * output_width;
   const size_t step_width = dilation == 1 ? subsampling : kernel_width;
-  const size_t step_height = kernel_size + (output_width * step_width - 1) * kernel_height;
+  const size_t step_height = kernel_size + (output_width - 1) * step_width * kernel_height;
 
   const size_t c_stride = benchmark::utils::RoundUp<size_t>(channels, cr);
 

diff --git a/src/convolution-nhwc.c b/src/convolution-nhwc.c
@@ -967,7 +967,7 @@ static enum xnn_status setup_convolution2d_nhwc(
       const size_t output_height = convolution_op->output_height;
       const size_t output_width = convolution_op->output_width;
       const size_t step_width = convolution_op->dilation_width == 1 ? convolution_op->stride_width : kernel_width;
-      const size_t step_height = kernel_size + (output_width * step_width - 1) * kernel_height;
+      const size_t step_height = kernel_size + (output_width - 1) * step_width * kernel_height;
       const size_t indirection_buffer_size = sizeof(void*) * batch_size * output_height * step_height;
 
       const void** indirection_buffer =