COMPMID-1162: Enable NHWC data layout support for NEWinogradConvolutionLayer - part1

In this first part we reworked the configuration of the kernels as before we
passed the raw pointer to the buffer within the configuration of the function

Change-Id: I83d3cb64c562303093c7f0ae52395ecd080a5d52
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/133560
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

4 files changed, 257 insertions, 164 deletions

diff --git a/arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h b/arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h
index 6b8866cb2e..68c133ee37 100644
--- a/arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h
@@ -42,15 +42,15 @@ public:
     /** Determine how much memory (in units of TIn) to allocate for the
      * transformed input.
      *
-     * @param[in] n_batches    Number of batches in the input tensor.
-     * @param[in] n_channels   Number of feature maps in the input tensor.
-     * @param[in] n_rows       Number of rows in each feature map.
-     * @param[in] n_cols       Number of columns in each feature map.
+     * @param[in] num_batches  Number of batches in the input tensor.
+     * @param[in] num_channels Number of feature maps in the input tensor.
+     * @param[in] num_rows     Number of rows in each feature map.
+     * @param[in] num_cols     Number of columns in each feature map.
      * @param[in] same_padding Use "SAME" padding, otherwise use "VALID".
      *
      * @return Storage size (in units of TIn) required.
      */
-    virtual unsigned int get_input_storage_size(int n_batches, int n_channels, int n_rows, int n_cols, bool same_padding) const = 0;
+    virtual unsigned int get_input_storage_size(int num_batches, int num_channels, int num_rows, int num_cols, bool same_padding) const = 0;
 
     /** Gets the stride between matrices in the input worspace
      *
@@ -64,16 +64,17 @@ public:
 
     /** Configure the output transform kernel.
      *
-     * @param[in]  input         Input tensor data
-     * @param[in]  n_batches     Number of batches in input tensor.
-     * @param[in]  n_rows        Number of rows in input tensor.
-     * @param[in]  n_cols        Number of columns in input tensor.
-     * @param[in]  n_channels    Number of channels in input tensor.
+     * @param[in]  input_nhwc    Input tensor in NHWC data layout format.
+     * @param[in]  num_batches   Number of batches in input tensor.
+     * @param[in]  num_rows      Number of rows in input tensor.
+     * @param[in]  num_cols      Number of columns in input tensor.
+     * @param[in]  num_channels  Number of channels in input tensor.
      * @param[in]  padding       Padding type.
      * @param[out] output        Base of output matrices.
      * @param[in]  matrix_stride Stride between output matrices.
      */
-    virtual void configure(const T *const input, const int n_batches, const int n_rows, const int n_cols, const int n_channels, const PaddingType padding, T *const output, const int matrix_stride) = 0;
+    virtual void configure(const ITensor *input_nhwc, const int num_batches, const int num_rows, const int num_cols, const int num_channels,
+                           const PaddingType padding, T *const output, const int matrix_stride) = 0;
 
     /** Destructor */
     virtual ~INEWinogradLayerTransformInputKernel()
@@ -86,22 +87,33 @@ template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, in
 class NEWinogradLayerTransformInputKernel : public INEWinogradLayerTransformInputKernel<T>
 {
 public:
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEWinogradLayerTransformInputKernel(const NEWinogradLayerTransformInputKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEWinogradLayerTransformInputKernel &operator=(const NEWinogradLayerTransformInputKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEWinogradLayerTransformInputKernel(NEWinogradLayerTransformInputKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEWinogradLayerTransformInputKernel &operator=(NEWinogradLayerTransformInputKernel &&) = default;
+    /** Default destructor */
+    ~NEWinogradLayerTransformInputKernel() = default;
+
     /** Determine how much memory (in units of TIn) to allocate for the
      * transformed input.
      *
-     * @param[in] n_batches    Number of batches in the input tensor.
-     * @param[in] n_channels   Number of feature maps in the input tensor.
-     * @param[in] n_rows       Number of rows in each feature map.
-     * @param[in] n_cols       Number of columns in each feature map.
+     * @param[in] num_batches  Number of batches in the input tensor.
+     * @param[in] num_channels Number of feature maps in the input tensor.
+     * @param[in] num_rows     Number of rows in each feature map.
+     * @param[in] num_cols     Number of columns in each feature map.
      * @param[in] same_padding Use "SAME" padding, otherwise use "VALID".
      *
      * @return Storage size (in units of TIn) required.
      */
     unsigned int get_input_storage_size(
-        int  n_batches,
-        int  n_channels,
-        int  n_rows,
-        int  n_cols,
+        int  num_batches,
+        int  num_channels,
+        int  num_rows,
+        int  num_cols,
         bool same_padding) const override;
 
     /** Gets the stride between matrices in the input worspace
@@ -124,21 +136,21 @@ public:
 
     /** Configure the output transform kernel.
      *
-     * @param[in]  input         Input tensor data. Data types supported: F32.
-     * @param[in]  n_batches     Number of batches in input tensor.
-     * @param[in]  n_rows        Number of rows in input tensor.
-     * @param[in]  n_cols        Number of columns in input tensor.
-     * @param[in]  n_channels    Number of channels in input tensor.
+     * @param[in]  input_nhwc    Input tensor.  Data types supported: F32. Layout supported NHWC.
+     * @param[in]  num_batches   Number of batches in input tensor.
+     * @param[in]  num_rows      Number of rows in input tensor.
+     * @param[in]  num_cols      Number of columns in input tensor.
+     * @param[in]  num_channels  Number of channels in input tensor.
      * @param[in]  padding       Padding type.
      * @param[out] output        Base of output matrices.
      * @param[in]  matrix_stride Stride between output matrices.
      */
     void configure(
-        const T *const    input,
-        const int         n_batches,
-        const int         n_rows,
-        const int         n_cols,
-        const int         n_channels,
+        const ITensor    *input_nhwc,
+        const int         num_batches,
+        const int         num_rows,
+        const int         num_cols,
+        const int         num_channels,
         const PaddingType padding,
         T *const          output,
         const int         matrix_stride) override;
@@ -163,7 +175,14 @@ public:
 
 private:
     using InputTransform = typename WinogradBase::template InputTransform<T>;
-    std::unique_ptr<InputTransform> _transform;
+    const ITensor *_input_nhwc;
+    int            _num_batches;   /**< Number of batches in input tensor. */
+    int            _num_rows;      /**< Number of rows in input tensor. */
+    int            _num_cols;      /**< Number of columns in input tensor. */
+    int            _num_channels;  /**< Number of channels in input tensor. */
+    PaddingType    _padding;       /**< Padding type. */
+    T             *_output;        /**< Base of output matrices. */
+    int            _matrix_stride; /**< Stride between output matrices. */
 };
 
 /** Interface for the NEON kernel to perform Winograd output transform. */
@@ -174,15 +193,15 @@ public:
     /** Determine how much memory (in units of TOut) to allocate for the
      * (Winograd domain) output.
      *
-     * @param[in] n_batches         Number of batches in the output tensor.
-     * @param[in] n_rows            Number of rows in each feature map of the input tensor.
-     * @param[in] n_cols            Number of columns in each feature map of the input tensor.
-     * @param[in] n_output_channels Number of feature maps in the output tensor.
-     * @param[in] same_padding      Use "SAME" padding, otherwise use "VALID".
+     * @param[in] num_batches         Number of batches in the output tensor.
+     * @param[in] num_rows            Number of rows in each feature map of the input tensor.
+     * @param[in] num_cols            Number of columns in each feature map of the input tensor.
+     * @param[in] num_output_channels Number of feature maps in the output tensor.
+     * @param[in] same_padding        Use "SAME" padding, otherwise use "VALID".
      *
      * @return Storage size (in units of TOut) required.
      */
-    virtual unsigned int get_output_storage_size(int n_batches, int n_rows, int n_cols, int n_output_channels, bool same_padding) const = 0;
+    virtual unsigned int get_output_storage_size(int num_batches, int num_rows, int num_cols, int num_output_channels, bool same_padding) const = 0;
 
     /** Gets the stride between matrices in the output worspace
      *
@@ -209,21 +228,21 @@ public:
      * @param[in]  biases              Pointer to the biases tensor.
      * @param[in]  output_workingspace Pointer to working space for the output tensor in the Winograd domain.
      * @param[in]  matrix_stride       Output matrix stride, can be computed with winograd::WinogradGEMM<2, 2, 3, 3>::Convolution<float, float>::get_output_matrix_stride()
-     * @param[out] output              Pointer to NHWC ordered output tensor, in the spatial domain.
-     * @param[in]  n_batches           Number of batches in the input tensor.
-     * @param[in]  n_rows              Number of rows in output tensor.
-     * @param[in]  n_cols              Number of columns in output tensor.
-     * @param[in]  n_channels          Number of feature maps in the output tensor.
+     * @param[out] output_nhwc         Pointer to a tensor in NHWC data layout ordered output tensor, in the spatial domain.
+     * @param[in]  num_batches         Number of batches in the input tensor.
+     * @param[in]  num_rows            Number of rows in output tensor.
+     * @param[in]  num_cols            Number of columns in output tensor.
+     * @param[in]  num_channels        Number of feature maps in the output tensor.
      */
     virtual void configure(
         const ITensor *biases,
         const T *const output_workingspace,
         const int      matrix_stride,
-        T *const       output,
-        const int      n_batches,
-        const int      n_rows,
-        const int      n_cols,
-        const int      n_channels) = 0;
+        ITensor *const output_nhwc,
+        const int      num_batches,
+        const int      num_rows,
+        const int      num_cols,
+        const int      num_channels) = 0;
 
     virtual ~INEWinogradLayerTransformOutputKernel()
     {
@@ -257,15 +276,15 @@ public:
     /** Determine how much memory (in units of TOut) to allocate for the
      * (Winograd domain) output.
      *
-     * @param[in] n_batches         Number of batches in the output tensor.
-     * @param[in] n_rows            Number of rows in each feature map of the input tensor.
-     * @param[in] n_cols            Number of columns in each feature map of the input tensor.
-     * @param[in] n_output_channels Number of feature maps in the output tensor.
-     * @param[in] same_padding      Use "SAME" padding, otherwise use "VALID".
+     * @param[in] num_batches         Number of batches in the output tensor.
+     * @param[in] num_rows            Number of rows in each feature map of the input tensor.
+     * @param[in] num_cols            Number of columns in each feature map of the input tensor.
+     * @param[in] num_output_channels Number of feature maps in the output tensor.
+     * @param[in] same_padding        Use "SAME" padding, otherwise use "VALID".
      *
      * @return Storage size (in units of TOut) required.
      */
-    unsigned int get_output_storage_size(int n_batches, int n_rows, int n_cols, int n_output_channels, bool same_padding) const override;
+    unsigned int get_output_storage_size(int num_batches, int num_rows, int num_cols, int num_output_channels, bool same_padding) const override;
 
     /** Gets the stride between matrices in the output worspace
      *
@@ -291,21 +310,21 @@ public:
      * @param[in]  biases              Pointer to the biases tensor.
      * @param[in]  output_workingspace Pointer to working space for the output tensor in the Winograd domain.
      * @param[in]  matrix_stride       Output matrix stride, can be computed with winograd::WinogradGEMM<2, 2, 3, 3>::Convolution<float, float>::get_output_matrix_stride()
-     * @param[out] output              Pointer to NHWC ordered output tensor, in the spatial domain.
-     * @param[in]  n_batches           Number of batches in the input tensor.
-     * @param[in]  n_rows              Number of rows in output tensor.
-     * @param[in]  n_cols              Number of columns in output tensor.
-     * @param[in]  n_channels          Number of feature maps in the output tensor.
+     * @param[out] output_nhwc         Pointer to a tensor with NHWC data layout, in the spatial domain.
+     * @param[in]  num_batches         Number of batches in the input tensor.
+     * @param[in]  num_rows            Number of rows in output tensor.
+     * @param[in]  num_cols            Number of columns in output tensor.
+     * @param[in]  num_channels        Number of feature maps in the output tensor.
      */
     void configure(
         const ITensor *biases,
         const T *const output_workingspace,
         const int      matrix_stride,
-        T *const       output,
-        const int      n_batches,
-        const int      n_rows,
-        const int      n_cols,
-        const int      n_channels) override;
+        ITensor *const output_nhwc,
+        const int      num_batches,
+        const int      num_rows,
+        const int      num_cols,
+        const int      num_channels) override;
 
     void run(const Window &window, const ThreadInfo &info) override;
 
@@ -329,11 +348,11 @@ private:
     const T       *_output_workspace;
     int            _matrix_stride;
     int            _matrix_row_stride;
-    T             *_output;
-    int            _n_batches;
-    int            _n_rows;
-    int            _n_cols;
-    int            _n_channels;
+    ITensor       *_output_nhwc;
+    int            _num_batches;
+    int            _num_rows;
+    int            _num_cols;
+    int            _num_channels;
 };
 
 /** Interface for the NEON kernel to perform Winograd weights transform. */
@@ -344,12 +363,12 @@ public:
     /** Determine how much memory (in units of T) to allocate for the
      * transformed weights.
      *
-     * @param[in] n_output_channels Number of output feature maps.
-     * @param[in] n_input_channels  Number of input feature maps.
+     * @param[in] num_output_channels Number of output feature maps.
+     * @param[in] num_input_channels  Number of input feature maps.
      *
      * @return Storage size (in units of T) required.
      */
-    virtual unsigned int get_weight_storage_size(int n_output_channels, int n_input_channels) const = 0;
+    virtual unsigned int get_weight_storage_size(int num_output_channels, int num_input_channels) const = 0;
     /** Gets the stride between matrices in the kernel worspace
      *
      * @param[in] kernel_shape The shape of the weights tensor.
@@ -360,13 +379,14 @@ public:
 
     /** Configure the weights transform kernel.
      *
-     * @param[in] weights_hwio      Pointer to the weights tensor
-     * @param[in] output            Pointer to working space for the output tensor in the Winograd domain.
-     * @param[in] matrix_stride     Stride across matrices in the output workspace.
-     * @param[in] n_output_channels Number of filters.
-     * @param[in] n_input_channels  Number of channels in each filter.
+     * @param[in] weights_hwio        Pointer to the weights tensor
+     * @param[in] output              Pointer to working space for the output tensor in the Winograd domain.
+     * @param[in] matrix_stride       Stride across matrices in the output workspace.
+     * @param[in] num_output_channels Number of filters.
+     * @param[in] num_input_channels  Number of channels in each filter.
      */
-    virtual void configure(const ITensor *weights_hwio, T *const output, const int matrix_stride, const int n_output_channels, const int n_input_channels) = 0;
+
+    virtual void configure(const ITensor *weights_hwio, T *const output, const int matrix_stride, const int num_output_channels, const int num_input_channels) = 0;
 
     virtual ~INEWinogradLayerTransformWeightsKernel()
     {
@@ -378,6 +398,17 @@ template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, in
 class NEWinogradLayerTransformWeightsKernel final : public INEWinogradLayerTransformWeightsKernel<T>
 {
 public:
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEWinogradLayerTransformWeightsKernel(const NEWinogradLayerTransformWeightsKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    NEWinogradLayerTransformWeightsKernel &operator=(const NEWinogradLayerTransformWeightsKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    NEWinogradLayerTransformWeightsKernel(NEWinogradLayerTransformWeightsKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    NEWinogradLayerTransformWeightsKernel &operator=(NEWinogradLayerTransformWeightsKernel &&) = default;
+    /** Default destructor */
+    ~NEWinogradLayerTransformWeightsKernel() = default;
+
     /** Default constructor. */
     NEWinogradLayerTransformWeightsKernel();
     const char *name() const override
@@ -397,8 +428,8 @@ public:
     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info);
 
     // Inherited methods overridden:
-    void configure(const ITensor *weights_hwio, T *const output, const int matrix_stride, const int n_output_channels, const int n_input_channels) override;
-    unsigned int get_weight_storage_size(int n_output_channels, int n_input_channels) const override;
+    void configure(const ITensor *weights_hwio, T *const output, const int matrix_stride, const int num_output_channels, const int num_input_channels) override;
+    unsigned int get_weight_storage_size(int num_output_channels, int num_input_channels) const override;
     int get_matrix_stride(const KernelShape &kernel_shape) const override;
     void run(const Window &window, const ThreadInfo &info) override;
     bool is_parallelisable() const override;
@@ -407,7 +438,12 @@ private:
     using WinogradBase     = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols>;
     using WinogradConv     = typename WinogradBase::template Convolution<T, T>;
     using WeightsTransform = typename WinogradBase::template WeightsTransform<T>;
-    std::unique_ptr<WeightsTransform> _transform;
+
+    const ITensor *_weights_hwio;
+    T             *_output;
+    int            _matrix_stride;
+    int            _num_output_channels;
+    int            _num_input_channels;
 };
 
 /** Interface for the NEON kernel to perform Winograd. */
@@ -421,7 +457,7 @@ public:
     /** Initialise the kernel
      *
      * @param[in]  n_gemms         Number of GEMMs to compute.
-     * @param[in]  M               in_shape.n_batches * tile_rows * tile_cols.
+     * @param[in]  M               in_shape.num_batches * tile_rows * tile_cols.
      * @param[in]  K               Number of channels in the input tensor.
      * @param[in]  N               Number of channels in the output tensor.
      * @param[in]  a_matrix_stride Stride between input matrices.
@@ -498,7 +534,7 @@ public:
     /** Initialise the kernel
      *
      * @param[in]  n_gemms         Number of GEMMs to compute.
-     * @param[in]  M               in_shape.n_batches * tile_rows * tile_cols.
+     * @param[in]  M               in_shape.num_batches * tile_rows * tile_cols.
      * @param[in]  K               Number of channels in the input tensor.
      * @param[in]  N               Number of channels in the output tensor.
      * @param[in]  a_matrix_stride Stride between input matrices.
diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
index 672684d14f..cfd53d7082 100644
--- a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
@@ -309,9 +309,9 @@ template class NEWinogradLayerBatchedGEMMKernel<float, float, 2, 2, 5, 5>;
 // Weights transform
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
-unsigned int NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_weight_storage_size(int n_output_channels, int n_input_channels) const
+unsigned int NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_weight_storage_size(int num_output_channels, int num_input_channels) const
 {
-    const KernelShape shape(n_output_channels, KernelRows, KernelCols, n_input_channels);
+    const KernelShape shape(num_output_channels, KernelRows, KernelCols, num_input_channels);
     return static_cast<unsigned int>(
                // WinogradConv returns the size in bytes, we divide by `sizeof(T)` to express that in units of T
                WinogradConv::get_kernel_storage_size(shape) / sizeof(T));
@@ -319,7 +319,8 @@ unsigned int NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTile
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
 NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformWeightsKernel()
-    : _transform()
+    : _weights_hwio(nullptr), _output(nullptr), _matrix_stride(0), _num_output_channels(0), _num_input_channels(0)
+
 {
 }
 
@@ -333,15 +334,20 @@ template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, in
 void NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure(
     const ITensor *weights_hwio,
     T *const       output,
-    const int      matrix_stride,     /** Stride across matrices in the output. */
-    const int      n_output_channels, /** Number of filters. */
-    const int      n_input_channels)  /** Number of channels in each filter. */
-{
-    const int matrix_row_stride = roundup(n_output_channels, WinogradConv::N_BLOCK);
-    _transform                  = support::cpp14::make_unique<WeightsTransform>(reinterpret_cast<T *>(weights_hwio->buffer()), output, matrix_stride, matrix_row_stride, n_output_channels,
-                                                                                n_input_channels);
-    Window win;
-    auto   win_last = _transform->get_window();
+    const int      matrix_stride,       /** Stride across matrices in the output. */
+    const int      num_output_channels, /** Number of filters. */
+    const int      num_input_channels)  /** Number of channels in each filter. */
+{
+    _weights_hwio        = weights_hwio;
+    _output              = output;
+    _matrix_stride       = matrix_stride;
+    _num_output_channels = num_output_channels;
+    _num_input_channels  = num_input_channels;
+
+    const int        matrix_row_stride = roundup(num_output_channels, WinogradConv::N_BLOCK);
+    WeightsTransform transform(nullptr, output, matrix_stride, matrix_row_stride, num_output_channels, num_input_channels);
+    Window           win;
+    auto             win_last = transform.get_window();
     win.set(Window::DimX, Window::Dimension(0, win_last, 1));
     INEKernel::configure(win);
 }
@@ -351,9 +357,12 @@ void NEWinogradLayerTransformWeightsKernel<T, OutputTileRows, OutputTileCols, Ke
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    const size_t fst = window.x().start();
-    const size_t lst = window.x().end();
-    _transform->run(fst, lst);
+
+    const int        matrix_row_stride = roundup(_num_output_channels, WinogradConv::N_BLOCK);
+    WeightsTransform transform(reinterpret_cast<T *>(_weights_hwio->buffer()), _output, _matrix_stride, matrix_row_stride, _num_output_channels, _num_input_channels);
+    const size_t     fst = window.x().start();
+    const size_t     lst = window.x().end();
+    transform.run(fst, lst);
 }
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
@@ -379,16 +388,16 @@ template class NEWinogradLayerTransformWeightsKernel<float, 2, 2, 5, 5>;
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
 unsigned int NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_input_storage_size(
-    int  n_batches,   /** Number of batches in the input tensor. */
-    int  n_channels,  /** Number of feature maps in the input tensor. */
-    int  n_rows,      /** Number of rows in each feature map. */
-    int  n_cols,      /** Number of columns in each feature map. */
-    bool same_padding /** Use "SAME" padding, otherwise use "VALID". */
+    int  num_batches,  /* Number of batches in the input tensor. */
+    int  num_channels, /* Number of feature maps in the input tensor. */
+    int  num_rows,     /* Number of rows in each feature map. */
+    int  num_cols,     /* Number of columns in each feature map. */
+    bool same_padding  /* Use "SAME" padding, otherwise use "VALID". */
 ) const
 {
     // Construct shapes for the input and kernel tensors.
-    const Tensor4DShape input_shape(n_batches, n_rows, n_cols, n_channels);
-    const KernelShape   kern_shape(1, KernelRows, KernelCols, n_channels);
+    const Tensor4DShape input_shape(num_batches, num_rows, num_cols, num_channels);
+    const KernelShape   kern_shape(1, KernelRows, KernelCols, num_channels);
     const PaddingType   padding = (same_padding) ? PADDING_SAME : PADDING_VALID;
     // Return the size, converted into units of TIn
     return static_cast<unsigned int>(WinogradConv::get_input_storage_size(kern_shape, input_shape, padding) / sizeof(T));
@@ -403,25 +412,32 @@ int NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, Kerne
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
 NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformInputKernel()
-    : _transform()
+    : _input_nhwc(), _num_batches(0), _num_rows(0), _num_cols(0), _num_channels(0), _padding(), _output(nullptr), _matrix_stride(0)
 {
 }
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
 void NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::configure(
-    const T *const    input,         /** Input tensor data */
-    const int         n_batches,     /** Number of batches in input tensor. */
-    const int         n_rows,        /** Number of rows in input tensor. */
-    const int         n_cols,        /** Number of columns in input tensor. */
-    const int         n_channels,    /** Number of channels in input tensor. */
-    const PaddingType padding,       /** Padding type. */
-    T *const          output,        /** Base of output matrices. */
-    const int         matrix_stride) /** Stride between output matrices. */
-{
-    //  _input_matrix_row_stride(n_input_channels),
-    _transform = support::cpp14::make_unique<InputTransform>(input, n_batches, n_rows, n_cols, n_channels, padding, output, matrix_stride, n_channels);
-    Window win;
-    auto   win_last = _transform->get_window();
+    const ITensor    *input_nhwc,
+    const int         num_batches,   /* Number of batches in input tensor. */
+    const int         num_rows,      /* Number of rows in input tensor. */
+    const int         num_cols,      /* Number of columns in input tensor. */
+    const int         num_channels,  /* Number of channels in input tensor. */
+    const PaddingType padding,       /* Padding type. */
+    T *const          output,        /* Base of output matrices. */
+    const int         matrix_stride) /* Stride between output matrices. */
+{
+    _input_nhwc    = input_nhwc;
+    _num_batches   = num_batches;
+    _num_rows      = num_rows;
+    _num_cols      = num_cols;
+    _num_channels  = num_channels;
+    _padding       = padding;
+    _output        = output;
+    _matrix_stride = matrix_stride;
+    InputTransform transform(nullptr, num_batches, num_rows, num_cols, num_channels, padding, output, matrix_stride, num_channels);
+    Window         win;
+    auto           win_last = transform.get_window();
     win.set(Window::DimX, Window::Dimension(0, win_last, 1));
     INEKernel::configure(win);
 }
@@ -431,9 +447,13 @@ void NEWinogradLayerTransformInputKernel<T, OutputTileRows, OutputTileCols, Kern
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+
+    InputTransform input_transform(reinterpret_cast<const T *>(_input_nhwc->buffer()), _num_batches, _num_rows, _num_cols, _num_channels, _padding, _output, _matrix_stride, _num_channels);
+
+    // The code below cannot be moved to configure because biases hasn't been allocated at that point
     const size_t fst = window.x().start();
     const size_t lst = window.x().end();
-    _transform->run(fst, lst);
+    input_transform.run(fst, lst);
 }
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
@@ -453,16 +473,16 @@ template class NEWinogradLayerTransformInputKernel<float, 2, 2, 5, 5>;
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
 unsigned int NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_output_storage_size(
-    int  n_batches,         /** Number of batches in the output tensor. */
-    int  n_rows,            /** Number of rows in each feature map of the input tensor. */
-    int  n_cols,            /** Number of columns in each feature map of the input tensor. */
-    int  n_output_channels, /** Number of feature maps in the output tensor. */
-    bool same_padding       /** Use "SAME" padding, otherwise use "VALID". */
+    int  num_batches,         /* Number of batches in the output tensor. */
+    int  num_rows,            /* Number of rows in each feature map of the input tensor. */
+    int  num_cols,            /* Number of columns in each feature map of the input tensor. */
+    int  num_output_channels, /* Number of feature maps in the output tensor. */
+    bool same_padding         /* Use "SAME" padding, otherwise use "VALID". */
 ) const
 {
     // Construct shapes for the input and kernel tensors.
-    const Tensor4DShape input_shape(n_batches, n_rows, n_cols, 1);
-    const KernelShape   kern_shape(n_output_channels, KernelRows, KernelCols, 1);
+    const Tensor4DShape input_shape(num_batches, num_rows, num_cols, 1);
+    const KernelShape   kern_shape(num_output_channels, KernelRows, KernelCols, 1);
     const PaddingType   padding = (same_padding) ? PADDING_SAME : PADDING_VALID;
 
     // Return the size, converted into units of TOut
@@ -472,7 +492,7 @@ unsigned int NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileC
 
 template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols>
 NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::NEWinogradLayerTransformOutputKernel()
-    : _biases(nullptr), _output_workspace(nullptr), _matrix_stride(0), _matrix_row_stride(0), _output(nullptr), _n_batches(0), _n_rows(0), _n_cols(0), _n_channels(0)
+    : _biases(nullptr), _output_workspace(nullptr), _matrix_stride(0), _matrix_row_stride(0), _output_nhwc(nullptr), _num_batches(0), _num_rows(0), _num_cols(0), _num_channels(0)
 {
 }
 
@@ -494,24 +514,24 @@ void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Ker
     const ITensor *biases,
     const T *const output_workingspace,
     const int      matrix_stride,
-    T *const       output,
-    const int      n_batches,
-    const int      n_rows,
-    const int      n_cols,
-    const int      n_channels)
+    ITensor *const output_nhwc,
+    const int      num_batches,
+    const int      num_rows,
+    const int      num_cols,
+    const int      num_channels)
 {
     _biases            = biases;
     _output_workspace  = output_workingspace;
     _matrix_stride     = matrix_stride;
-    _matrix_row_stride = roundup(n_channels, WinogradConv::N_BLOCK);
-    _output            = output;
-    _n_batches         = n_batches;
-    _n_rows            = n_rows;
-    _n_cols            = n_cols;
-    _n_channels        = n_channels;
+    _matrix_row_stride = roundup(num_channels, WinogradConv::N_BLOCK);
+    _output_nhwc       = output_nhwc;
+    _num_batches       = num_batches;
+    _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, _output, _n_batches, _n_rows, _n_cols, _n_channels);
+    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();
     win.set(Window::DimX, Window::Dimension(0, win_last, 1));
@@ -524,11 +544,11 @@ void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Ker
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_NULLPTR(_output_workspace);
-    ARM_COMPUTE_ERROR_ON_NULLPTR(_output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(_output_nhwc);
 
     OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride,
-                                     (_biases ? reinterpret_cast<T *>(_biases->buffer()) : nullptr), _output,
-                                     _n_batches, _n_rows, _n_cols, _n_channels);
+                                     (_biases ? reinterpret_cast<T *>(_biases->buffer()) : nullptr), reinterpret_cast<T *>(_output_nhwc->buffer()),
+                                     _num_batches, _num_rows, _num_cols, _num_channels);
 
     // 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 1a9c72965b..d6bc5cfd9a 100644
--- a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
@@ -60,8 +60,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
     ARM_COMPUTE_UNUSED(output);
     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(data_layout != DataLayout::NCHW); // COMPMID-1162
     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.");
@@ -107,6 +107,7 @@ bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_siz
 
     return std::find(fast_math_winograd.begin(), fast_math_winograd.end(), p) != fast_math_winograd.end();
 }
+
 } //namespace
 
 NEWinogradConvolutionLayer::NEWinogradConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
@@ -218,33 +219,60 @@ void NEWinogradConvolutionLayer::configure(const ITensor *input, const ITensor *
     _output_nhwc.allocator()->init(info);
     _output_nhwc.allocator()->allocate();
 
-    // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map]
-    _permute_weights.configure(weights, &_weights_hwio, PermutationVector(3U, 2U, 0U, 1U));
-    _weights_hwio.allocator()->allocate();
-
-    // configure the kernel to transform the input tensor from NCHW -> NHWC
-    _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U));
-    _input_nhwc.allocator()->allocate();
-
     const KernelShape kernel_shape({ out_channels, static_cast<int>(kernel_size.height), static_cast<int>(kernel_size.width), in_channels });
 
     // Configure the InputTransform
     const int input_matrix_stride = transform_input_kernel->get_matrix_stride(kernel_shape, in_shape, use_padding_type);
-    transform_input_kernel->configure(reinterpret_cast<float *>(_input_nhwc.buffer()), in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type,
-                                      reinterpret_cast<float *>(_input_workspace.buffer()), input_matrix_stride);
+
+    if(data_layout == DataLayout::NCHW)
+    {
+        // configure the kernel to transform the input tensor from NCHW -> NHWC
+        _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U));
+        _input_nhwc.allocator()->allocate();
+        transform_input_kernel->configure(&_input_nhwc, in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type,
+                                          reinterpret_cast<float *>(_input_workspace.buffer()), input_matrix_stride);
+    }
+    else
+    {
+        transform_input_kernel->configure(_input, in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type,
+                                          reinterpret_cast<float *>(_input_workspace.buffer()), input_matrix_stride);
+    }
 
     // Configure WeightsTransform
     const int kernel_matrix_stride = transform_weights_kernel->get_matrix_stride(kernel_shape);
-    transform_weights_kernel->configure(&_weights_hwio, reinterpret_cast<float *>(_kernel_storage.buffer()), kernel_matrix_stride, out_channels, in_channels);
+    if(data_layout == DataLayout::NCHW)
+    {
+        // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map]
+        _permute_weights.configure(weights, &_weights_hwio, PermutationVector(3U, 2U, 0U, 1U));
+
+        transform_weights_kernel->configure(&_weights_hwio, reinterpret_cast<float *>(_kernel_storage.buffer()), kernel_matrix_stride, out_channels, in_channels);
+    }
+    else
+    {
+        // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map]
+        _permute_weights.configure(weights, &_weights_hwio, PermutationVector(3U, 0U, 1U, 2U));
+
+        transform_weights_kernel->configure(&_weights_hwio, reinterpret_cast<float *>(_kernel_storage.buffer()), kernel_matrix_stride, out_channels, in_channels);
+    }
+    _weights_hwio.allocator()->allocate();
 
     // Configure OutputTransform
     //The biases tensor has not been allocated at this point in time, the output transform will add the biases to the final result in the run() method
     const int  output_matrix_stride = transform_output_kernel->get_matrix_stride(kernel_shape, in_shape, use_padding_type);
     const auto output_shape(transform_output_kernel->get_output_shape(kernel_shape, in_shape, use_padding_type));
 
-    transform_output_kernel->configure(biases, reinterpret_cast<float *>(_output_workspace.buffer()),
-                                       output_matrix_stride, reinterpret_cast<float *>(_output_nhwc.buffer()),
-                                       in_shape.n_batches, output_shape.n_rows, output_shape.n_cols, out_channels);
+    if(data_layout == DataLayout::NCHW)
+    {
+        transform_output_kernel->configure(biases, reinterpret_cast<float *>(_output_workspace.buffer()),
+                                           output_matrix_stride, &_output_nhwc,
+                                           in_shape.n_batches, output_shape.n_rows, output_shape.n_cols, out_channels);
+    }
+    else
+    {
+        transform_output_kernel->configure(biases, reinterpret_cast<float *>(_output_workspace.buffer()),
+                                           output_matrix_stride, _output,
+                                           in_shape.n_batches, output_shape.n_rows, output_shape.n_cols, out_channels);
+    }
 
     // Configure GEMM
     const int    tile_rows                = iceildiv(output_shape.n_rows, output_tile.height);
@@ -293,14 +321,16 @@ void NEWinogradConvolutionLayer::configure(const ITensor *input, const ITensor *
 
     //Configure Activation Layer
     _is_activationlayer_enabled = act_info.enabled();
-    if(_is_activationlayer_enabled)
+    if(data_layout == DataLayout::NCHW && _is_activationlayer_enabled)
     {
-        _activationlayer_function.configure(output, nullptr, act_info);
+        _activationlayer_function.configure(_output, nullptr, act_info);
     }
 }
 
 void NEWinogradConvolutionLayer::run()
 {
+    const DataLayout data_layout = _input->info()->data_layout();
+
     _memory_group.acquire();
     if(!_reshaped_kernel)
     {
@@ -308,9 +338,12 @@ void NEWinogradConvolutionLayer::run()
         _permute_weights.run();
         NEScheduler::get().schedule(_transform_weights_kernel.get(), Window::DimX);
     }
-    //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC
-    _permute_input.run();
 
+    if(data_layout == DataLayout::NCHW)
+    {
+        //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC
+        _permute_input.run();
+    }
     // Transform input tensor to the winograd domain
     NEScheduler::get().schedule(_transform_input_kernel.get(), Window::DimX);
 
@@ -320,8 +353,11 @@ void NEWinogradConvolutionLayer::run()
     // Transform output tensor to the spatial domain
     NEScheduler::get().schedule(_transform_output_kernel.get(), Window::DimX);
 
-    // Reorder the convoluted output to ACL's ordering NCHW
-    _permute_output.run();
+    if(data_layout == DataLayout::NCHW)
+    {
+        // Reorder the convoluted output to ACL's ordering NCHW
+        _permute_output.run();
+    }
 
     if(_is_activationlayer_enabled)
     {
diff --git a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
index 6381b99131..f40f3d2e43 100644
--- a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
@@ -55,7 +55,8 @@ class WinogradConvolutionLayerValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, ActivationLayerInfo act_info)
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation,
+               DataType data_type, ActivationLayerInfo act_info)
     {
         ARM_COMPUTE_UNUSED(dilation);
 
@@ -84,7 +85,7 @@ protected:
         }
     }
 
-    TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info,
+    TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, const PadStrideInfo &info,
                               DataType data_type, ActivationLayerInfo act_info)
     {
         // Create tensors