COMPMID-1287: Extending NEWinogradLayer test suite

Added NHWC to the dataset to the validation tests
Fixed a problem in the output transform which made the Activation to fail
because way/ordering the output transform wrote the data to the output tensor.

Change-Id: I9609f86605dbfef70b47a0fb043287bf0e5d675b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/136015
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

2 files changed, 10 insertions, 8 deletions

diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
index cfd53d7082..50e69a8adf 100644
--- a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
@@ -182,7 +182,6 @@ Status validate_arguments_winograd_output_trans(const ITensorInfo *input, const
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON(winograd_info.output_data_layout != DataLayout::NCHW);
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area());
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd output transform only supports 3x3 and 5x5 kernels");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd output transform only supports 3x3 and 5x5 kernels");
@@ -529,12 +528,15 @@ void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Ker
     _num_rows          = num_rows;
     _num_cols          = num_cols;
     _num_channels      = num_channels;
-
     // We don't have the biases buffer at this stage as it hasn't been allocated, we pass in nullptr OutputTransform is only used here to compute the window
     OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride, nullptr, nullptr, _num_batches, _num_rows, _num_cols, _num_channels);
-    Window          win;
-    auto            win_last = output_transform.get_window();
+
+    Window win;
+    auto   win_last = output_transform.get_window();
     win.set(Window::DimX, Window::Dimension(0, win_last, 1));
+
+    _output_nhwc->info()->set_valid_region(ValidRegion(Coordinates(), _output_nhwc->info()->tensor_shape()));
+
     INEKernel::configure(win);
 }
 
@@ -548,7 +550,7 @@ void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Ker
 
     OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride,
                                      (_biases ? reinterpret_cast<T *>(_biases->buffer()) : nullptr), reinterpret_cast<T *>(_output_nhwc->buffer()),
-                                     _num_batches, _num_rows, _num_cols, _num_channels);
+                                     _num_batches, _num_rows, _num_cols, _num_channels, 0, _output_nhwc->info()->strides_in_bytes()[2] / sizeof(T), _output_nhwc->info()->strides_in_bytes()[1] / sizeof(T));
 
     // The code below cannot be moved to configure because biases hasn't been allocated at that point
     const size_t fst = window.x().start();
diff --git a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
index d9f6c0e0f8..39175c26c6 100644
--- a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
@@ -61,7 +61,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != 3 && weights->dimension(height_idx) != 5, "Only 3 and 5 kernels are supported");
-    ARM_COMPUTE_RETURN_ERROR_ON(data_layout != DataLayout::NCHW); // COMPMID-1287
     ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
 
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd layer only supports unit strides.");
@@ -325,7 +324,7 @@ void NEWinogradConvolutionLayer::configure(const ITensor *input, const ITensor *
 
     //Configure Activation Layer
     _is_activationlayer_enabled = act_info.enabled();
-    if(data_layout == DataLayout::NCHW && _is_activationlayer_enabled)
+    if(_is_activationlayer_enabled)
     {
         _activationlayer_function.configure(_output, nullptr, act_info);
     }
@@ -363,6 +362,7 @@ void NEWinogradConvolutionLayer::run()
     {
         _activationlayer_function.run();
     }
+
     _memory_group.release();
 }
 
@@ -396,6 +396,7 @@ Status NEWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITen
     // Validate input transform
     const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info);
     const TensorInfo  input0       = input->clone()->set_tensor_shape(input0_shape);
+
     switch(weights->dimension(idx_width))
     {
         case 3:
@@ -482,7 +483,6 @@ Status NEWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITen
             break;
         }
     }
-
     // Validate Activation Layer
     if(act_info.enabled())
     {