COMPMID-1048 Add NHWC data format support to Winograd output transform 4x4_3x3

https://confluence.arm.com/display/MLENG/Winograd+Output+Transform%3A+NCHW+vs+NHWC+on+OpenCL

Change-Id: I6995f5cef759ba70ebd96d545b952041b6f1f36e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/128729
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

3 files changed, 24 insertions, 9 deletions

diff --git a/tests/datasets/WinogradOutputTransformDataset.h b/tests/datasets/WinogradOutputTransformDataset.h
index b28df390d1..bf6f5cd2a9 100644
--- a/tests/datasets/WinogradOutputTransformDataset.h
+++ b/tests/datasets/WinogradOutputTransformDataset.h
@@ -118,6 +118,12 @@ public:
         add_config(TensorShape(7U, 4U, 36U, 3U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(24U, 16U, 36U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(7U, 12U, 16U, 5U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+        // NHWC
+        add_config(TensorShape(13U, 4U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(10U, 9U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC));
+        add_config(TensorShape(13U, 6U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(10U, 11U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC));
+        add_config(TensorShape(7U, 117U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(53U, 33U), PadStrideInfo(1, 1, 0, 1), DataLayout::NHWC));
+        add_config(TensorShape(7U, 4U, 36U, 3U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(8U, 10U), PadStrideInfo(1, 1, 0, 0), DataLayout::NHWC));
+        add_config(TensorShape(24U, 16U, 36U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 1), DataLayout::NHWC));
 
         // (4x4, 5x5)
         add_config(TensorShape(13U, 1U, 64U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(7U, 6U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
@@ -149,6 +155,13 @@ public:
         add_config(TensorShape(64U, 3136U, 36U, 3U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(224U, 224U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(32U, 784U, 36U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(112U, 112U), PadStrideInfo(1, 1, 1, 0), DataLayout::NCHW));
         add_config(TensorShape(13U, 196U, 36U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NCHW));
+        // NHWC
+        add_config(TensorShape(64U, 3136U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(224U, 224U), PadStrideInfo(1, 1, 1, 1), DataLayout::NHWC));
+        add_config(TensorShape(32U, 784U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(112U, 112U), PadStrideInfo(1, 1, 1, 0), DataLayout::NHWC));
+        add_config(TensorShape(13U, 196U, 36U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NHWC));
+        add_config(TensorShape(64U, 3136U, 36U, 3U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(224U, 224U), PadStrideInfo(1, 1, 1, 1), DataLayout::NHWC));
+        add_config(TensorShape(32U, 784U, 36U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(112U, 112U), PadStrideInfo(1, 1, 1, 0), DataLayout::NHWC));
+        add_config(TensorShape(13U, 196U, 36U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(56U, 56U), PadStrideInfo(1, 1, 0, 1), DataLayout::NHWC));
     }
 };
 } // namespace datasets
diff --git a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
index ef596e0bae..e23368add6 100644
--- a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
@@ -36,6 +36,7 @@
 #include "tests/validation/reference/ActivationLayer.h"
 #include "tests/validation/reference/ConvolutionLayer.h"
 #include "tests/validation/reference/GEMM.h"
+#include "tests/validation/reference/Permute.h"
 #include "tests/validation/reference/Utils.h"
 #include "tests/validation/reference/Winograd.h"
 
@@ -440,10 +441,8 @@ public:
     template <typename...>
     void setup(TensorShape input_shape, WinogradInfo winograd_info, DataType data_type)
     {
-        TensorShape output_shape = compute_winograd_output_transform_shape(TensorInfo(input_shape, 1, data_type), winograd_info);
-
-        _target    = compute_target(input_shape, output_shape, winograd_info, data_type);
-        _reference = compute_reference(input_shape, output_shape, winograd_info, data_type);
+        _target    = compute_target(input_shape, winograd_info, data_type);
+        _reference = compute_reference(input_shape, winograd_info, data_type);
     }
 
 protected:
@@ -467,8 +466,10 @@ protected:
         }
     }
 
-    TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, const WinogradInfo &winograd_info, DataType data_type)
+    TensorType compute_target(const TensorShape &input_shape, const WinogradInfo &winograd_info, DataType data_type)
     {
+        TensorShape output_shape = compute_winograd_output_transform_shape(TensorInfo(input_shape, 1, data_type), winograd_info);
+
         // Create tensors
         TensorType src = create_tensor<TensorType>(input_shape, data_type);
         TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, 0, QuantizationInfo(), winograd_info.output_data_layout);
@@ -495,8 +496,11 @@ protected:
         return dst;
     }
 
-    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, const WinogradInfo &winograd_info, DataType data_type)
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, WinogradInfo winograd_info, DataType data_type)
     {
+        winograd_info.output_data_layout = DataLayout::NCHW;
+        TensorShape output_shape         = compute_winograd_output_transform_shape(TensorInfo(input_shape, 1, data_type), winograd_info);
+
         // Create reference
         SimpleTensor<T> src{ input_shape, data_type };
         SimpleTensor<T> bias{ TensorShape(input_shape[0]), data_type };
diff --git a/tests/validation/reference/Winograd.cpp b/tests/validation/reference/Winograd.cpp
index 194a78e95f..197d218129 100644
--- a/tests/validation/reference/Winograd.cpp
+++ b/tests/validation/reference/Winograd.cpp
@@ -333,8 +333,6 @@ SimpleTensor<T> winograd_filter_transform(const SimpleTensor<T> &in, const Tenso
 template <typename T>
 SimpleTensor<T> winograd_output_transform(const SimpleTensor<T> &in, const SimpleTensor<T> &b, const TensorShape &output_shape, const WinogradInfo &winograd_info)
 {
-    ARM_COMPUTE_ERROR_ON_MSG(winograd_info.output_data_layout != DataLayout::NCHW, "Only supported NCHW data format");
-
     const PadStrideInfo conv_info        = winograd_info.convolution_info;
     const Size2D        input_dimensions = winograd_info.input_dimensions;
     const Size2D        output_tile_size = winograd_info.output_tile_size;
@@ -350,7 +348,7 @@ SimpleTensor<T> winograd_output_transform(const SimpleTensor<T> &in, const Simpl
     const unsigned int out_tile_h = output_tile_size.height;
 
     ARM_COMPUTE_ERROR_ON(in.shape()[2] != (in_tile_w * in_tile_h));
-    ARM_COMPUTE_ERROR_ON(in.shape()[0] != out.shape()[2]);
+    ARM_COMPUTE_ERROR_ON(in.shape()[0] != out.shape()[get_data_layout_dimension_index(winograd_info.output_data_layout, DataLayoutDimension::CHANNEL)]);
 
     // Compute tile dimensions
     // Input tile dimensions