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

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

Change-Id: Iac35a54389266701b7d8f5434a7a37df85b7b187
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/133315
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

8 files changed, 467 insertions, 47 deletions

diff --git a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
index b92ff2f60c..58e8291161 100644
--- a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
+++ b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
@@ -49,6 +49,7 @@ public:
      * @note Winograd input transform supports the following configurations:
      *       F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
      *       Strides: only unit strides
+     *       Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3)
      *
      * @param[in] input         The input tensor to transform. Data types supported: F32
      * @param[in] output        The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_input_transform_shape. Data types supported: Same as @p input
@@ -60,6 +61,7 @@ public:
      * @note Winograd input transform supports the following configurations:
      *       F(output tile, kernel size):F(2x2, 3x3), F(4x4, 3x3), F(4x4, 5x5)
      *       Strides: only unit strides
+     *       Data Layout: NCHW for all configurations, NHWC for F(4x4, 3x3)
      *
      * @param[in] input         The input tensor to transform. Data types supported: F32
      * @param[in] output        The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_input_transform_shape. Data types supported: Same as @p input
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 9666702749..f64cf9d6ae 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -250,11 +250,15 @@ inline TensorShape compute_winograd_input_transform_shape(const ITensorInfo &inp
     const Size2D        output_tile_size = winograd_info.output_tile_size;
     const Size2D        input_tile_size  = Size2D(output_tile_size.width + kernel_size.width - 1, output_tile_size.height + kernel_size.height - 1);
 
+    const size_t idx_w = get_data_layout_dimension_index(input.data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(input.data_layout(), DataLayoutDimension::HEIGHT);
+    const size_t idx_c = get_data_layout_dimension_index(input.data_layout(), DataLayoutDimension::CHANNEL);
+
     // Compute height
-    const unsigned int num_tiles_x = std::ceil((input.tensor_shape().x() - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right()) / static_cast<float>(output_tile_size.width));
-    const unsigned int num_tiles_y = std::ceil((input.tensor_shape().y() - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / static_cast<float>(output_tile_size.height));
+    const unsigned int num_tiles_x = std::ceil((input.tensor_shape()[idx_w] - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right()) / static_cast<float>(output_tile_size.width));
+    const unsigned int num_tiles_y = std::ceil((input.tensor_shape()[idx_h] - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / static_cast<float>(output_tile_size.height));
 
-    const unsigned int width  = input.tensor_shape()[get_data_layout_dimension_index(input.data_layout(), DataLayoutDimension::CHANNEL)];
+    const unsigned int width  = input.tensor_shape()[idx_c];
     const unsigned int height = num_tiles_x * num_tiles_y;
     const unsigned int depth  = input_tile_size.area();
 
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 0b2f414c71..75d4feb637 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -377,6 +377,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "winograd_input_transform_2x2_3x3_stepz1_nchw", "winograd.cl" },
     { "winograd_input_transform_2x2_3x3_stepz2_nchw", "winograd.cl" },
     { "winograd_input_transform_4x4_3x3_stepz1_nchw", "winograd.cl" },
+    { "winograd_input_transform_4x4_3x3_stepz1_nhwc", "winograd.cl" },
     { "winograd_output_transform_2x2_3x3_nchw", "winograd.cl" },
     { "winograd_output_transform_4x4_3x3_nchw", "winograd.cl" },
     { "winograd_output_transform_4x4_5x5_nchw", "winograd.cl" },
diff --git a/src/core/CL/cl_kernels/winograd.cl b/src/core/CL/cl_kernels/winograd.cl
index c7ca8f6752..383a3a718a 100644
--- a/src/core/CL/cl_kernels/winograd.cl
+++ b/src/core/CL/cl_kernels/winograd.cl
@@ -25,7 +25,7 @@
 
 #if defined(SRC_DIM_Z)
 
-/** This OpenCL kernel performs Winograd filter transform 3x3 when the data format is NCHW and the output tile is 2x2
+/** This OpenCL kernel performs Winograd filter transform 3x3 when the data layout is NCHW and the output tile is 2x2
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  *
@@ -117,7 +117,7 @@ __kernel void winograd_filter_transform_2x2_3x3_nchw(
     *(__global float *)(dst_addr + 15 * dst_stride_z) = out3.s3;
 }
 
-/** This OpenCL kernel performs Winograd filter transform 3x3 when the data format is NCHW and the output tile is 4x4
+/** This OpenCL kernel performs Winograd filter transform 3x3 when the data layout is NCHW and the output tile is 4x4
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  *
@@ -255,7 +255,7 @@ __kernel void winograd_filter_transform_4x4_3x3_nchw(
     *(__global float *)(dst_addr + 35 * dst_stride_z) = out5.s5;
 }
 
-/** This OpenCL kernel performs Winograd filter transform 3x3 when the data format is NHWC and the output tile is 4x4
+/** This OpenCL kernel performs Winograd filter transform 3x3 when the data layout is NHWC and the output tile is 4x4
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  *
@@ -404,7 +404,7 @@ __kernel void winograd_filter_transform_4x4_3x3_nhwc(
     *(__global float *)(dst_addr + 34 * dst_stride_z) = out54;
     *(__global float *)(dst_addr + 35 * dst_stride_z) = out55;
 }
-/** This OpenCL kernel performs Winograd filter transform 5x5 when the data format is NCHW and the output tile is 4x4
+/** This OpenCL kernel performs Winograd filter transform 5x5 when the data layout is NCHW and the output tile is 4x4
  *
  * @note In order to correctly split the input tensor in batches, its dimension across the Z axis (channels for NCHW, height for NHWC) must be passed at compile time using -DSRC_DIM_Z: e.g. -DSRC_DIM_Z=64
  *
@@ -857,7 +857,7 @@ __kernel void winograd_input_transform_2x2_3x3_stepz2_nchw(
     vstore2(out33, 0, (__global float *)(dst_addr + 15 * dst_stride_z));
 }
 
-/** This OpenCL kernel computes the input transform when the output tile is 4x4, the filter size 3x3 and the data format is NCHW
+/** This OpenCL kernel computes the input transform when the output tile is 4x4, the filter size 3x3 and the data layout is NCHW
  *
  * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5).
  * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0).
@@ -1116,6 +1116,348 @@ __kernel void winograd_input_transform_4x4_3x3_stepz1_nchw(
     dst_addr += dst_plane_stride;
 }
 
+#if defined(SRC_DIM_1) && defined(SRC_DIM_2)
+/** This OpenCL kernel computes the input transform when the output tile is 4x4, the filter size 3x3 and the data layout is NHWC
+ *
+ * @note The number of tiles in the x axis must be passed at compile time using -DNUM_TILES_X (i.e.-DNUM_TILES_X=5).
+ * @note The pad left and pad top must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (i.e.-DPAD_LEFT=1 and -DPAD_TOP=0).
+ *
+ * @param[in] src_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in] src_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in] src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] src_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in] src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in] src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z                        src_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_ptr                           Pointer to the destination tensor. Supported data types: as @p src_ptr
+ * @param[in] dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z                        dst_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void winograd_input_transform_4x4_3x3_stepz1_nhwc(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    int x = get_global_id(0);
+    int y = get_global_id(1);
+    int z = get_global_id(2);
+
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * src_stride_x;
+
+    // Clamp coordinates. This clamp is valid for all rows
+    int4 y_coord0 = (int4)(y * 4) + (int4)(0, 1, 2, 3) - (int4)PAD_LEFT;
+    int2 y_coord1 = (int2)(y * 4) + (int2)(4, 5) - (int2)PAD_LEFT;
+    y_coord0      = clamp(y_coord0, -1, SRC_DIM_1);
+    y_coord1      = clamp(y_coord1, -1, SRC_DIM_1);
+
+    // Row4
+    int z_coord = (z * 4) - PAD_TOP + 4;
+
+    // If z < 0, set y to -1
+    int4 valid_y0 = select(y_coord0, -1, (int4)z_coord < 0);
+    int2 valid_y1 = select(y_coord1, -1, (int2)z_coord < 0);
+    // If z >= SRC_DIM_2, set y to SRC_DIM_2
+    valid_y0 = select(valid_y0, SRC_DIM_1, (int4)z_coord >= SRC_DIM_2);
+    valid_y1 = select(valid_y1, SRC_DIM_1, (int2)z_coord >= SRC_DIM_2);
+
+    // Clamp z coordinate
+    z_coord = clamp(z_coord, 0, SRC_DIM_2 - 1);
+
+    float d40 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d41 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    float d42 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    float d43 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    float d44 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d45 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+
+    float k0 = d44;
+    float k1 = d44;
+    float k2 = d44;
+    float k3 = d44;
+    float k4 = d44;
+    float k5 = (float)0.0f;
+
+    k0 += 4.0f * d40 - 5.0f * d42;
+    k1 += -4.0f * d41 - 4.0f * d42 + d43;
+    k2 += 4.0f * d41 - 4.0f * d42 - d43;
+    k3 += -2.0f * d41 + 2.0f * d43 - d42;
+    k4 += 2.0f * d41 - 2.0f * d43 - d42;
+    k5 += 4.0f * d41 - 5.0f * d43 + d45;
+
+    // Row0
+    z_coord = (z * 4) - PAD_TOP + 0;
+
+#if PAD_TOP != 0
+    valid_y0 = select(y_coord0, -1, (int4)z_coord < 0);
+    valid_y1 = select(y_coord1, -1, (int2)z_coord < 0);
+    valid_y0 = select(valid_y0, SRC_DIM_1, (int4)z_coord >= SRC_DIM_2);
+    valid_y1 = select(valid_y1, SRC_DIM_1, (int2)z_coord >= SRC_DIM_2);
+    z_coord  = clamp(z_coord, 0, SRC_DIM_2 - 1);
+#else  // PAD_TOP != 0
+    valid_y0 = y_coord0;
+    valid_y1 = y_coord1;
+#endif // if PAD_TOP == 0, we cannot read out of bound
+
+    float d00 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d01 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    float d02 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    float d03 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    float d04 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d05 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+
+    // Row2
+    z_coord  = (z * 4) - PAD_TOP + 2;
+    valid_y0 = select(y_coord0, -1, (int4)z_coord < 0);
+    valid_y1 = select(y_coord1, -1, (int2)z_coord < 0);
+    valid_y0 = select(valid_y0, SRC_DIM_1, (int4)z_coord >= SRC_DIM_2);
+    valid_y1 = select(valid_y1, SRC_DIM_1, (int2)z_coord >= SRC_DIM_2);
+    z_coord  = clamp(z_coord, 0, SRC_DIM_2 - 1);
+
+    float d20 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d21 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    float d22 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    float d23 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    float d24 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d25 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+
+    // Compute destination address
+    __global float *dst_addr = (__global float *)(dst_ptr + dst_offset_first_element_in_bytes + x * dst_stride_x + (y + z * (int)NUM_TILES_X) * dst_stride_y);
+
+    uint dst_plane_stride = dst_stride_z / sizeof(float);
+
+    float out0  = k0;
+    float out1  = k1;
+    float out2  = k2;
+    float out3  = k3;
+    float out4  = k4;
+    float out5  = k5;
+    float out6  = k0;
+    float out7  = k1;
+    float out8  = k2;
+    float out9  = k3;
+    float out10 = k4;
+    float out11 = k5;
+    float out12 = k0;
+    float out13 = k1;
+    float out14 = k2;
+    float out15 = k3;
+    float out16 = k4;
+    float out17 = k5;
+    float out18 = k0;
+    float out19 = k1;
+    float out20 = k2;
+    float out21 = k3;
+    float out22 = k4;
+    float out23 = k5;
+    float out24 = k0;
+    float out25 = k1;
+    float out26 = k2;
+    float out27 = k3;
+    float out28 = k4;
+    float out29 = k5;
+
+    // Channels [0, 5]: [out00, out01, out02, out03, out04, out05]
+    out0 += 16.0f * d00 - 20.0f * d02 - 20.0f * d20 + 25.0f * d22 + 4.0f * d04 - 5.0f * d24;
+    out1 += -16.0f * d01 - 16.0f * d02 + 4.0f * d03 + 20.0f * d21 + 20.0f * d22 - 5.0f * d23 + 4.0f * d04 - 5.0f * d24;
+    out2 += 16.0f * d01 - 16.0f * d02 - 4.0f * d03 - 20.0f * d21 + 20.0f * d22 + 5.0f * d23 + 4.0f * d04 - 5.0f * d24;
+    out3 += -8.0f * d01 - 4.0f * d02 + 8.0f * d03 + 10.0f * d21 + 5.0f * d22 - 10.0f * d23 + 4.0f * d04 - 5.0f * d24;
+    out4 += 8.0f * d01 - 4.0f * d02 - 8.0f * d03 - 10.0f * d21 + 5.0f * d22 + 10.0f * d23 + 4.0f * d04 - 5.0f * d24;
+    out5 += 16.0f * d01 - 20.0f * d03 - 20.0f * d21 + 4.0f * d05 + 25.0f * d23 - 5.0f * d25;
+
+    *((__global float *)dst_addr) = out0;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out1;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out2;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out3;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out4;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out5;
+    dst_addr += dst_plane_stride;
+
+    // Row1
+    z_coord = (z * 4) - PAD_TOP + 1;
+    // Row1 can never be out of bounds
+    valid_y0 = y_coord0;
+    valid_y1 = y_coord1;
+
+    float d10 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d11 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    float d12 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    float d13 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    float d14 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d15 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+
+    // Row3
+    z_coord  = (z * 4) - PAD_TOP + 3;
+    valid_y0 = select(y_coord0, -1, (int4)z_coord < 0);
+    valid_y1 = select(y_coord1, -1, (int2)z_coord < 0);
+    valid_y0 = select(valid_y0, SRC_DIM_1, (int4)z_coord >= SRC_DIM_2);
+    valid_y1 = select(valid_y1, SRC_DIM_1, (int2)z_coord >= SRC_DIM_2);
+    z_coord  = clamp(z_coord, 0, SRC_DIM_2 - 1);
+    z_coord  = clamp(z_coord, 0, SRC_DIM_2 - 1);
+
+    float d30 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d31 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    float d32 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    float d33 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    float d34 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d35 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+
+    // Compute common parts for the channels between [6, 29]
+    // Channels [6, 11]:  [out10, out11, out12, out13, out14, out15]
+    // Channels [12, 17]: [out20, out21, out22, out23, out24, out25]
+    float part0  = -16.0f * d20 + 20.0f * d22 - 4.0f * d24;
+    float part1  = 16.0f * d10 - 20.0f * d12 + 4.0f * d14 - 4.0f * d30 + 5.0f * d32 - d34;
+    float part2  = 16.0f * d22 - 4.0f * d24;
+    float part3  = 16.0f * d21 - 4.0f * d23;
+    float part4  = 16.0f * d12 - 4.0f * d14 - 4.0f * d32 + d34;
+    float part5  = 16.0f * d11 - 4.0f * d13 - 4.0f * d31 + d33;
+    float part6  = 4.0f * d22 - 4.0f * d24;
+    float part7  = 8.0f * d11 - 8.0f * d13 - 2.0f * d31 + 2.0f * d33;
+    float part8  = 4.0f * d12 - 4.0f * d14 - d32 + d34;
+    float part9  = 8.0f * d21 - 8.0f * d23;
+    float part10 = -16.0f * d21 + 20.0f * d23 - 4.0f * d25;
+    float part11 = -16.0f * d11 + 20.0f * d13 - 4.0f * d15 + 4.0f * d31 - 5.0f * d33 + d35;
+
+    // Channels [18, 23]: [out30, out31, out32, out33, out34, out35]
+    // Channels [24, 29]: [out40, out41, out42, out43, out44, out45]
+    float part12 = 8.0f * d10 - 10.0f * d12 + 2.0f * d14 - 8.0f * d30 + 10.0f * d32 - 2.0f * d34;
+    float part13 = part0 * 0.25f; // -4.0f * d20 + 5.0f * d22 - d24
+    float part14 = part2 * 0.25f; // 4.0f * d22 - d24
+    float part15 = 8.0f * d11 - 2.0f * d13 - 8.0f * d31 + 2.0f * d33;
+    float part16 = 8.0f * d12 - 2.0f * d14 - 8.0f * d32 + 2.0f * d34;
+    float part17 = part3 * 0.25f; // 4.0f * d21 - d23
+    float part18 = part6 * 0.25f; // d22 - d24
+    float part19 = 4.0f * d11 - 4.0f * d13 - 4.0f * d31 + 4.0f * d33;
+    float part20 = 2.0f * d12 - 2.0f * d14 - 2.0f * d32 + 2.0f * d34;
+    float part21 = part9 * 0.25f;                                        // 2.0f * (d21 - d23)
+    float part22 = part10 * 0.25f;                                       // - 4.0f * d21 + 5.0f * d23 - d25
+    float part23 = part11 * 0.5f + 6.0f * d31 - 7.5f * d33 + 1.5f * d35; // - 8.0f * d11 + 10.0f * d13 - 2.0f * d15 + 8.0f * d31 - 10.0f * d33 + 2.0f * d35;
+
+    out6 += part0 - part1;
+    out12 += part0 + part1;
+    out7 += part2 + part3 + part4 + part5;
+    out8 += part2 - part3 + part4 - part5;
+    out13 += part2 + part3 - part4 - part5;
+    out14 += part2 - part3 - part4 + part5;
+    out9 += part6 + part7 + part8 + part9;
+    out10 += part6 - part7 + part8 - part9;
+    out15 += part6 - part7 - part8 + part9;
+    out16 += part6 + part7 - part8 - part9;
+    out11 += part10 + part11;
+    out17 += part10 - part11;
+
+    out18 += part13 - part12;
+    out24 += part13 + part12;
+    out19 += part14 + part15 + part16 + part17;
+    out20 += part14 - part15 + part16 - part17;
+    out25 += part14 - part15 - part16 + part17;
+    out26 += part14 + part15 - part16 - part17;
+    out21 += part18 + part19 + part20 + part21;
+    out22 += part18 - part19 + part20 - part21;
+    out27 += part18 - part19 - part20 + part21;
+    out28 += part18 + part19 - part20 - part21;
+    out23 += part22 + part23;
+    out29 += part22 - part23;
+
+    *((__global float *)dst_addr) = out6;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out7;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out8;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out9;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out10;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out11;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out12;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out13;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out14;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out15;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out16;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out17;
+    dst_addr += dst_plane_stride;
+
+    *((__global float *)dst_addr) = out18;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out19;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out20;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out21;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out22;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out23;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out24;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out25;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out26;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out27;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out28;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out29;
+    dst_addr += dst_plane_stride;
+
+    // Row5
+    z_coord  = (z * 4) - PAD_TOP + 5;
+    valid_y0 = select(y_coord0, -1, (int4)z_coord < 0);
+    valid_y1 = select(y_coord1, -1, (int2)z_coord < 0);
+    valid_y0 = select(valid_y0, SRC_DIM_1, (int4)z_coord >= SRC_DIM_2);
+    valid_y1 = select(valid_y1, SRC_DIM_1, (int2)z_coord >= SRC_DIM_2);
+    z_coord  = clamp(z_coord, 0, SRC_DIM_2 - 1);
+    z_coord  = clamp(z_coord, 0, SRC_DIM_2 - 1);
+
+    float d50 = *(__global float *)(src_addr + valid_y0.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d51 = *(__global float *)(src_addr + valid_y0.s1 * (int)src_stride_y + z_coord * src_stride_z);
+    float d52 = *(__global float *)(src_addr + valid_y0.s2 * (int)src_stride_y + z_coord * src_stride_z);
+    float d53 = *(__global float *)(src_addr + valid_y0.s3 * (int)src_stride_y + z_coord * src_stride_z);
+    float d54 = *(__global float *)(src_addr + valid_y1.s0 * (int)src_stride_y + z_coord * src_stride_z);
+    float d55 = *(__global float *)(src_addr + valid_y1.s1 * (int)src_stride_y + z_coord * src_stride_z);
+
+    // Channels [30, 35]
+    out0 = 16.0f * d10 - 20.0f * d12 - 20.0f * d30 + 25.0f * d32 + 4.0f * d50 - 5.0f * d52 + d54 + 4.0f * d14 - 5.0f * d34;
+    out1 = -16.0f * d11 - 16.0f * d12 + 4.0f * d13 + 20.0f * d31 + 20.0f * d32 - 5.0f * d33 - 4.0f * d51 - 4.0f * d52 + d53 + d54 + 4.0f * d14 - 5.0f * d34;
+    out2 = 16.0f * d11 - 16.0f * d12 - 4.0f * d13 - 20.0f * d31 + 20.0f * d32 + 5.0f * d33 + 4.0f * d51 - 4.0f * d52 - d53 + d54 + 4.0f * d14 - 5.0f * d34;
+    out3 = -8.0f * d11 - 4.0f * d12 + 8.0f * d13 + 10.0f * d31 - 10.0f * d33 + 5.0f * d32 - 2.0f * d51 + 2.0f * d53 - d52 + d54 + 4.0f * d14 - 5.0f * d34;
+    out4 = 8.0f * d11 - 4.0f * d12 - 8.0f * d13 - 10.0f * d31 + 5.0f * d32 + 10.0f * d33 + 2.0f * d51 - 2.0f * d53 - d52 + d54 + 4.0f * d14 - 5.0f * d34;
+    out5 = 16.0f * d11 - 20.0f * d13 + 4.0f * d15 - 20.0f * d31 + 25.0f * d33 - 5.0f * d35 + 4.0f * d51 - 5.0f * d53 + d55;
+
+    *((__global float *)dst_addr) = out0;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out1;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out2;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out3;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out4;
+    dst_addr += dst_plane_stride;
+    *((__global float *)dst_addr) = out5;
+    dst_addr += dst_plane_stride;
+}
+
+#endif /* defined(SRC_DIM_1) && defined(SRC_DIM_2) */
+
 #define OUTPUT_ROW_4x4_5x5(out, tmp, comm_fact)                     \
     ({                                                              \
         comm_fact.s0 = tmp.s2 - 4.25f * tmp.s4 + tmp.s6;            \
@@ -1287,7 +1629,7 @@ __kernel void winograd_input_transform_4x4_5x5_stepz1_nchw(
 #endif // defined(NUM_TILES_X) && defined(PAD_LEFT) && defined(PAD_TOP)
 
 #if defined(NUM_TILES_X)
-/** This OpenCL kernel performs Winograd output transform when the output tile is 2x2, the filter size 3x3 and the data format is NCHW
+/** This OpenCL kernel performs Winograd output transform when the output tile is 2x2, the filter size 3x3 and the data layout is NCHW
  *
  * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16
  *
@@ -1389,7 +1731,7 @@ __kernel void winograd_output_transform_2x2_3x3_nchw(
     vstore2((float2)(out10, out11), 0, (__global float *)(dst_addr + 1 * dst_stride_y));
 }
 
-/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, the filter size 3x3 and the data format is NCHW
+/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, the filter size 3x3 and the data layout is NCHW
  *
  * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16
  *
@@ -1566,7 +1908,7 @@ __kernel void winograd_output_transform_4x4_3x3_nchw(
     vstore4((float4)(out30, out31, out32, out33), 0, (__global float *)(dst_addr + 3 * dst_stride_y));
 }
 
-/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, the filter size 3x3 and the data format is NHWC
+/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, the filter size 3x3 and the data layout is NHWC
  *
  * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16
  *
@@ -1774,7 +2116,7 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
         col.s3 = comm_fact.s0 + 8.f * comm_fact.s1 + comm_fact.s2 + d7;  \
     })
 
-/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, the filter size 5x5 and the data format is NCHW
+/** This OpenCL kernel performs Winograd output transform when the output tile is 4x4, the filter size 5x5 and the data layout is NCHW
  *
  * @note The number of tiles along the X direction must be passed at compile time using -DNUM_TILES_X: e.g. -DNUM_TILES_X=16
  *
diff --git a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
index febd22b04e..e73ac7df76 100644
--- a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
+++ b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
@@ -40,13 +40,13 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() != DataLayout::NCHW);
 
     const PadStrideInfo conv_info        = winograd_info.convolution_info;
     const Size2D        output_tile_size = winograd_info.output_tile_size;
     const Size2D        kernel_size      = winograd_info.kernel_size;
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd input transform only supports unit strides");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size != Size2D(3U, 3U) && kernel_size != Size2D(5U, 5U), "Winograd input transform only supports 3x3 and 5x5 kernels");
+    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::NHWC && (output_tile_size != Size2D(4U, 4U) || kernel_size != Size2D(3U, 3U)));
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(3U, 3U) && output_tile_size != Size2D(2U, 2U)
                                     && output_tile_size != Size2D(4U, 4U),
                                     "Winograd input transform only supports 2x2 or 4x4 output tile for 3x3 kernels");
@@ -75,12 +75,28 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     const Size2D        output_tile_size = winograd_info.output_tile_size;
     const Size2D        kernel_size      = winograd_info.kernel_size;
 
-    const unsigned int num_elems_read_per_iteration_x = output_tile_size.width + kernel_size.width - 1;
-    const unsigned int num_elems_read_per_iteration_y = output_tile_size.height + kernel_size.height - 1;
+    unsigned int num_elems_read_per_iteration_x = 0;
+    unsigned int num_elems_read_per_iteration_y = 0;
+    unsigned int pad_left                       = 0;
+    unsigned int pad_top                        = 0;
+
+    if(input->data_layout() == DataLayout::NCHW)
+    {
+        num_elems_read_per_iteration_x = output_tile_size.width + kernel_size.width - 1;
+        num_elems_read_per_iteration_y = output_tile_size.height + kernel_size.height - 1;
+        pad_left                       = conv_info.pad_left();
+        pad_top                        = conv_info.pad_top();
+    }
+    else
+    {
+        num_elems_read_per_iteration_x = 1;
+        num_elems_read_per_iteration_y = output_tile_size.width + kernel_size.width - 1;
+        pad_top                        = 1;
+    }
 
     Window win = calculate_max_window(*input, Steps(1, 1));
 
-    AccessWindowRectangle input_access(input, -conv_info.pad_left(), -conv_info.pad_top(), num_elems_read_per_iteration_x, num_elems_read_per_iteration_y);
+    AccessWindowRectangle input_access(input, -pad_left, -pad_top, num_elems_read_per_iteration_x, num_elems_read_per_iteration_y);
 
     bool window_changed = update_window_and_padding(win, input_access);
 
@@ -108,17 +124,27 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
     const Size2D        output_tile_size = winograd_info.output_tile_size;
     const Size2D        kernel_size      = winograd_info.kernel_size;
 
+    const size_t idx_w = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
+
     // Compute number of elements to process in the X and Y direction
-    const int num_elements_x = input->info()->dimension(0) - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right();
-    const int num_elements_y = input->info()->dimension(1) - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom();
+    const int num_elements_x = input->info()->dimension(idx_w) - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right();
+    const int num_elements_y = input->info()->dimension(idx_h) - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom();
 
-    // Check if we need to extend the right or bottom border
-    const unsigned int extra_border_right  = ((num_elements_x % output_tile_size.width) == 0) ? 0u : static_cast<unsigned int>(output_tile_size.width - 1);
-    const unsigned int extra_border_bottom = ((num_elements_y % output_tile_size.height) == 0) ? 0u : static_cast<unsigned int>(output_tile_size.height - 1);
+    _input  = input;
+    _output = output;
+    if(input->info()->data_layout() == DataLayout::NCHW)
+    {
+        // Check if we need to extend the right or bottom border
+        const unsigned int extra_border_right  = ((num_elements_x % output_tile_size.width) == 0) ? 0u : static_cast<unsigned int>(output_tile_size.width - 1);
+        const unsigned int extra_border_bottom = ((num_elements_y % output_tile_size.height) == 0) ? 0u : static_cast<unsigned int>(output_tile_size.height - 1);
 
-    _input       = input;
-    _output      = output;
-    _border_size = BorderSize(conv_info.pad_top(), conv_info.pad_right() + extra_border_right, conv_info.pad_bottom() + extra_border_bottom, conv_info.pad_left());
+        _border_size = BorderSize(conv_info.pad_top(), conv_info.pad_right() + extra_border_right, conv_info.pad_bottom() + extra_border_bottom, conv_info.pad_left());
+    }
+    else
+    {
+        _border_size = BorderSize(1U, 0U, 1U, 0);
+    }
     _num_tiles_x = std::ceil(num_elements_x / static_cast<float>(output_tile_size.width));
     _num_tiles_y = std::ceil(num_elements_y / static_cast<float>(output_tile_size.height));
 
@@ -134,6 +160,12 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
     build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
     build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
 
+    if(input->info()->data_layout() == DataLayout::NHWC)
+    {
+        build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
+        build_opts.add_option("-DSRC_DIM_2=" + support::cpp11::to_string(_input->info()->dimension(2)));
+    }
+
     // Create kernel
     std::string kernel_name = "winograd_input_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string();
 
@@ -148,7 +180,7 @@ void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor
     // Append stepz and data layout
     kernel_name += "_stepz";
     kernel_name += support::cpp11::to_string(_step_z);
-    kernel_name += "_nchw";
+    kernel_name += "_" + lower_string(string_from_data_layout(input->info()->data_layout()));
 
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
@@ -183,12 +215,16 @@ void CLWinogradInputTransformKernel::run(const Window &window, cl::CommandQueue
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 
+    const size_t idx_w = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_h = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::HEIGHT);
+    const size_t idx_c = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::CHANNEL);
+
     Window slice = window.first_slice_window_3D();
-    slice.set(Window::DimX, Window::Dimension(0, _num_tiles_x, 1));
-    slice.set(Window::DimY, Window::Dimension(0, _num_tiles_y, 1));
+    slice.set(idx_w, Window::Dimension(0, _num_tiles_x, 1));
+    slice.set(idx_h, Window::Dimension(0, _num_tiles_y, 1));
 
-    ARM_COMPUTE_ERROR_ON(((slice.z().end() - slice.z().start()) % _step_z) != 0);
-    slice.set(Window::DimZ, Window::Dimension(slice.z().start(), slice.z().end(), _step_z));
+    ARM_COMPUTE_ERROR_ON(((slice[idx_c].end() - slice[idx_c].start()) % _step_z) != 0);
+    slice.set(idx_c, Window::Dimension(slice[idx_c].start(), slice[idx_c].end(), _step_z));
 
     do
     {
diff --git a/tests/datasets/WinogradInputTransformDataset.h b/tests/datasets/WinogradInputTransformDataset.h
index 59fb2adbd6..e365f9657f 100644
--- a/tests/datasets/WinogradInputTransformDataset.h
+++ b/tests/datasets/WinogradInputTransformDataset.h
@@ -97,12 +97,11 @@ private:
     std::vector<WinogradInfo> _infos{};
 };
 
-class SmallWinogradInputTransformDataset final : public WinogradInputTransformDataset
+class SmallWinogradInputTransformDataset2x2_3x3 final : public WinogradInputTransformDataset
 {
 public:
-    SmallWinogradInputTransformDataset()
+    SmallWinogradInputTransformDataset2x2_3x3()
     {
-        // (2x2, 3x3)
         add_config(TensorShape(9U, 9U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(9U, 9U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(27U, 13U, 2U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(27U, 13U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(128U, 64U, 1U, 3U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(128U, 64U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
@@ -110,8 +109,14 @@ public:
         add_config(TensorShape(27U, 13U, 2U, 4U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(27U, 13U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(9U, 9U, 3U, 5U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(9U, 9U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(14U, 14U, 512U, 2U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
+    }
+};
 
-        // (4x4, 3x3)
+class SmallWinogradInputTransformDataset4x4_3x3 final : public WinogradInputTransformDataset
+{
+public:
+    SmallWinogradInputTransformDataset4x4_3x3()
+    {
         add_config(TensorShape(9U, 9U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(9U, 9U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(27U, 13U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(27U, 13U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(128U, 64U, 1U, 3U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(128U, 64U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
@@ -119,8 +124,14 @@ public:
         add_config(TensorShape(27U, 13U, 2U, 4U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(27U, 13U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(9U, 9U, 3U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(9U, 9U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(14U, 14U, 512U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(14U, 14U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
+    }
+};
 
-        // (4x4, 5x5)
+class SmallWinogradInputTransformDataset4x4_5x5 final : public WinogradInputTransformDataset
+{
+public:
+    SmallWinogradInputTransformDataset4x4_5x5()
+    {
         add_config(TensorShape(9U, 9U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(9U, 9U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(27U, 13U, 2U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(27U, 13U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(128U, 64U, 1U, 3U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(128U, 64U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
@@ -131,24 +142,35 @@ public:
     }
 };
 
-class LargeWinogradInputTransformDataset final : public WinogradInputTransformDataset
+class LargeWinogradInputTransformDataset2x2_3x3 final : public WinogradInputTransformDataset
 {
 public:
-    LargeWinogradInputTransformDataset()
+    LargeWinogradInputTransformDataset2x2_3x3()
     {
-        // (2x2, 3x3)
         add_config(TensorShape(42U, 37U, 8U, 15U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(42U, 37U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(57U, 60U, 13U, 8U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(57U, 60U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(128U, 64U, 21U, 13U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(128U, 64U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(83U, 72U, 14U, 5U), WinogradInfo(Size2D(2U, 2U), Size2D(3U, 3U), Size2D(83U, 72U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+    }
+};
 
-        // (4x4, 3x3)
+class LargeWinogradInputTransformDataset4x4_3x3 final : public WinogradInputTransformDataset
+{
+public:
+    LargeWinogradInputTransformDataset4x4_3x3()
+    {
         add_config(TensorShape(42U, 37U, 8U, 15U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(42U, 37U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(57U, 60U, 13U, 8U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(57U, 60U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(128U, 64U, 21U, 13U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(128U, 64U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
         add_config(TensorShape(83U, 72U, 14U, 5U), WinogradInfo(Size2D(4U, 4U), Size2D(3U, 3U), Size2D(83U, 72U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
+    }
+};
 
-        // (4x4, 5x5)
+class LargeWinogradInputTransformDataset4x4_5x5 final : public WinogradInputTransformDataset
+{
+public:
+    LargeWinogradInputTransformDataset4x4_5x5()
+    {
         add_config(TensorShape(42U, 37U, 8U, 15U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(42U, 37U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(57U, 60U, 13U, 8U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(57U, 60U), PadStrideInfo(1, 1, 1, 1), DataLayout::NCHW));
         add_config(TensorShape(128U, 64U, 21U, 13U), WinogradInfo(Size2D(4U, 4U), Size2D(5U, 5U), Size2D(128U, 64U), PadStrideInfo(1, 1, 0, 0), DataLayout::NCHW));
diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp
index 4565c2d59d..59fe25db73 100644
--- a/tests/validation/CL/Winograd.cpp
+++ b/tests/validation/CL/Winograd.cpp
@@ -53,6 +53,10 @@ namespace
 {
 constexpr AbsoluteTolerance<float> tolerance_f32(0.001f);
 constexpr AbsoluteTolerance<float> tolerance_convolution_layer_f32(0.1f);
+const auto                         SmallWinogradInputTransformDataset = framework::dataset::concat(datasets::SmallWinogradInputTransformDataset2x2_3x3(),
+                                                                                                   framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x4_3x3(), datasets::SmallWinogradInputTransformDataset4x4_5x5()));
+const auto LargeWinogradInputTransformDataset = framework::dataset::concat(datasets::LargeWinogradInputTransformDataset2x2_3x3(),
+                                                                           framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x4_3x3(), datasets::LargeWinogradInputTransformDataset4x4_5x5()));
 } // namespace
 
 using namespace arm_compute::misc::shape_calculator;
@@ -102,7 +106,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
 
 using CLWinogradInputTransformFixture = WinogradInputTransformValidationFixture<CLTensor, CLAccessor, CLWinogradInputTransform, float>;
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallWinogradInputTransformDataset(), datasets::LargeWinogradInputTransformDataset()),
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(SmallWinogradInputTransformDataset, LargeWinogradInputTransformDataset),
                                                                            framework::dataset::make("DataLayout", { DataLayout::NCHW })),
                                                                    framework::dataset::make("DataType", { DataType::F32 })),
                shape_in, winograd_info, data_layout, data_type)
@@ -123,15 +127,19 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame
     winograd_input_transform.configure(&in, &out, winograd_info);
 }
 
-FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradInputTransformFixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallWinogradInputTransformDataset(),
-                                                                                                                     framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradInputTransformFixture, framework::DatasetMode::PRECOMMIT, combine(framework::dataset::concat(combine(SmallWinogradInputTransformDataset,
+                                                                                                             framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+                                                                                                             combine(datasets::SmallWinogradInputTransformDataset4x4_3x3(),
+                                                                                                                     framework::dataset::make("DataLayout", { DataLayout::NHWC }))),
                                                                                                              framework::dataset::make("DataType", { DataType::F32 })))
 {
     validate(CLAccessor(_target), _reference, tolerance_f32);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradInputTransformFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeWinogradInputTransformDataset(),
-                                                                                                                   framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradInputTransformFixture, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(combine(LargeWinogradInputTransformDataset,
+                                                                                                           framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+                                                                                                           combine(datasets::LargeWinogradInputTransformDataset4x4_3x3(),
+                                                                                                                   framework::dataset::make("DataLayout", { DataLayout::NHWC }))),
                                                                                                            framework::dataset::make("DataType", { DataType::F32 })))
 {
     validate(CLAccessor(_target), _reference, tolerance_f32);
diff --git a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
index f40f3d2e43..07795c2361 100644
--- a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
@@ -314,10 +314,15 @@ protected:
         }
     }
 
-    TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, const WinogradInfo &winograd_info, DataLayout data_layout, DataType data_type)
+    TensorType compute_target(TensorShape input_shape, const TensorShape &output_shape, const WinogradInfo &winograd_info, DataLayout data_layout, DataType data_type)
     {
+        if(data_layout == DataLayout::NHWC)
+        {
+            permute(input_shape, PermutationVector(2U, 0U, 1U));
+        }
+
         TensorType src = create_tensor<TensorType>(input_shape, data_type, 1, 0, QuantizationInfo(), data_layout);
-        TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, 0, QuantizationInfo(), data_layout);
+        TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, 0, QuantizationInfo());
 
         // Create and configure function
         FunctionType transf;
@@ -345,7 +350,7 @@ protected:
     SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, const WinogradInfo &winograd_info, DataLayout data_layout, DataType data_type)
     {
         // Create reference
-        SimpleTensor<T> src{ input_shape, data_type, 1, 0, QuantizationInfo(), data_layout };
+        SimpleTensor<T> src{ input_shape, data_type, 1, 0, QuantizationInfo() };
 
         // Fill reference
         fill(src, 0, -1.f, 1.f);