commit 52ba29e936b8e711e8acdfe819e36f884d4f3fe1
author Jenkins <bsgcomp@arm.com> Wed Aug 29 15:32:11 2018 +0000
committer Michele Di Giorgio <michele.digiorgio@arm.com> Thu Aug 30 17:15:09 2018 +0100
tree 23a8e15208ad7fcf07a37ed9119979cde01dcdc8
parent e2542c9f35ca427286822cd0c9296f49914f78b0

arm_compute v18.08

src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp

diff --git a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
index 49753ad..a70389a 100644
--- a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
@@ -33,17 +33,34 @@
 
 namespace
 {
-Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims)
+Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataLayout data_layout)
 {
     Size2D output_tile = Size2D{};
 
-    if(kernel_dims == Size2D(3U, 3U))
+    const unsigned int kernel_max_dim = std::max(kernel_dims.width, kernel_dims.height);
+
+    // Check if the input spatial dimensions are smaller than 4
+    const bool is_input_lt4_nchw = (input_dims.width <= 4 && input_dims.height <= 4) && (data_layout == DataLayout::NCHW);
+
+    if(kernel_max_dim == 3U)
     {
-        output_tile = (input_dims.width <= 4 && input_dims.height <= 4) ? Size2D(2U, 2U) : Size2D(4U, 4U);
+        if(kernel_dims == Size2D(3U, 3U))
+        {
+            output_tile = is_input_lt4_nchw ? Size2D(2U, 2U) : Size2D(4U, 4U);
+        }
+        else if(kernel_dims == Size2D(3U, 1U))
+        {
+            output_tile = is_input_lt4_nchw ? Size2D(2U, 1U) : Size2D(4U, 1U);
+        }
+        else
+        {
+            output_tile = is_input_lt4_nchw ? Size2D(1U, 2U) : Size2D(1U, 4U);
+        }
     }
-    else if(kernel_dims == Size2D(5U, 5U))
+    else if(kernel_max_dim == 5U)
     {
-        output_tile = Size2D(4U, 4U);
+        output_tile = Size2D(kernel_dims.width == 1 ? 1U : 4U,
+                             kernel_dims.height == 1 ? 1U : 4U);
     }
 
     return output_tile;
@@ -82,7 +99,7 @@
     // Input shape, kernel size and output tile
     const Size2D input_dims  = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]);
     const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]);
-    const Size2D output_tile = winograd_output_tile(input_dims, kernel_size);
+    const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->info()->data_layout());
 
     // Check if the Winograd configuration requires fast math
     if(!enable_fast_math)
@@ -139,7 +156,7 @@
     // Input shape, kernel size and output tile
     const Size2D input_dims  = Size2D(input->tensor_shape()[idx_width], input->tensor_shape()[idx_height]);
     const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]);
-    const Size2D output_tile = winograd_output_tile(input_dims, kernel_size);
+    const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->data_layout());
 
     // Check if the Winograd configuration requires fast math
     if(!enable_fast_math)