COMPMID-935 Implementing Convolution with Winograd on OpenCL (part 3)

Change-Id: I51f92f30602fb0a02314f344fa67061f448694bf
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122793
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

16 files changed, 1276 insertions, 2 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index e7fa74d00e..ca2cb0411a 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -109,5 +109,6 @@
 #include "arm_compute/core/CL/kernels/CLWarpAffineKernel.h"
 #include "arm_compute/core/CL/kernels/CLWarpPerspectiveKernel.h"
 #include "arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h"
+#include "arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h"
 
 #endif /* __ARM_COMPUTE_CLKERNELS_H__ */
diff --git a/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
new file mode 100644
index 0000000000..15cd6e2649
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_CLWINOGRADINPUTTRANSFORMKERNEL_H__
+#define __ARM_COMPUTE_CLWINOGRADINPUTTRANSFORMKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** OpenCL kernel to perform Winograd input transform.*/
+class CLWinogradInputTransformKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLWinogradInputTransformKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLWinogradInputTransformKernel(const CLWinogradInputTransformKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLWinogradInputTransformKernel &operator=(const CLWinogradInputTransformKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLWinogradInputTransformKernel(CLWinogradInputTransformKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLWinogradInputTransformKernel &operator=(CLWinogradInputTransformKernel &&) = default;
+    /** Set the input and output of the kernel.
+     *
+     * @param[in] input       The input tensor to permute. Data types supported: F32
+     * @param[in] output      The output tensor. Data types supported: Same as @p input
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 kernels are supported
+     */
+    void configure(const ICLTensor *input, ICLTensor *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradInputTransformKernel
+     *
+     * @param[in] input       First tensor input info. Data types supported: F32.
+     * @param[in] output      Output tensor info. Data types supported: same as @p input.
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 kernels are supported
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+    BorderSize border_size() const override;
+
+private:
+    BorderSize       _border_size;
+    const ICLTensor *_input;
+    ICLTensor       *_output;
+    int              _num_tiles_x;
+    int              _num_tiles_y;
+    unsigned int     _step_z;
+};
+} // arm_compute
+#endif /*__ARM_COMPUTE_CLWINOGRADINPUTTRANSFORMKERNEL_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index e21e5cd0d6..2919625511 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -194,6 +194,24 @@ inline TensorShape compute_fully_connected_reshaped_weights_shape(const ITensorI
 
     return output_shape;
 }
+
+inline TensorShape compute_winograd_input_transform_shape(const ITensorInfo &input, const PadStrideInfo &conv_info, const Size2D &kernel_size)
+{
+    // 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()) / 2.f);
+    const unsigned int num_tiles_y = std::ceil((input.tensor_shape().y() - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / 2.f);
+
+    const unsigned int width  = input.tensor_shape()[get_data_layout_dimension_index(input.data_layout(), DataLayoutDimension::CHANNEL)];
+    const unsigned int height = num_tiles_x * num_tiles_y;
+    const unsigned int depth  = 16; // COMPMID-990
+
+    TensorShape output_shape{ input.tensor_shape() };
+    output_shape.set(0, width);
+    output_shape.set(1, height);
+    output_shape.set(2, depth);
+
+    return output_shape;
+}
 } // namespace shape_calculator
 } // namespace misc
 } // namespace arm_compute
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index a5bbc41a17..7c2377aa09 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -107,5 +107,6 @@
 #include "arm_compute/runtime/CL/functions/CLTranspose.h"
 #include "arm_compute/runtime/CL/functions/CLWarpAffine.h"
 #include "arm_compute/runtime/CL/functions/CLWarpPerspective.h"
+#include "arm_compute/runtime/CL/functions/CLWinogradInputTransform.h"
 
 #endif /* __ARM_COMPUTE_CLFUNCTIONS_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h b/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h
new file mode 100644
index 0000000000..54b8bdecba
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLWinogradInputTransform.h
@@ -0,0 +1,60 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_CLWINOGRADINPUTTRANSFORM_H__
+#define __ARM_COMPUTE_CLWINOGRADINPUTTRANSFORM_H__
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+
+#include <cstdint>
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Basic function to execute a @ref CLWinogradInputTransformKernel. */
+class CLWinogradInputTransform : public ICLSimpleFunction
+{
+public:
+    /** Set the input and output tensors.
+     *
+     * @param[in] input       The input tensor to transform. Data types supported: F32
+     * @param[in] output      The output tensor. Data types supported: Same as @p input
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 kernels are supported
+     */
+    void configure(ICLTensor *input, ICLTensor *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+    /**  Static function to check if given info will lead to a valid configuration of @ref CLWinogradInputTransform.
+     *
+     * @param[in] input       First tensor input info. Data types supported: F32.
+     * @param[in] output      Output tensor info. Data types supported: same as @p input.
+     * @param[in] conv_info   Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
+     * @param[in] kernel_dims Kernel dimensions. Currently only 3x3 kernels are supported
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+};
+}
+#endif /*__ARM_COMPUTE_CLWINOGRADINPUTTRANSFORM_H__ */
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index aba20448e7..40aceb702a 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -351,6 +351,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "warp_affine_bilinear", "warp_affine.cl" },
     { "warp_perspective_nearest_neighbour", "warp_perspective.cl" },
     { "warp_perspective_bilinear", "warp_perspective.cl" },
+    { "winograd_input_transform_2x2_3x3_stepz1_nchw", "winograd.cl" },
+    { "winograd_input_transform_2x2_3x3_stepz2_nchw", "winograd.cl" },
     { "YUYV422_to_IYUV_bt709", "color_convert.cl" },
     { "YUYV422_to_NV12_bt709", "color_convert.cl" },
     { "YUYV422_to_RGB888_bt709", "color_convert.cl" },
@@ -676,6 +678,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
         "warp_perspective.cl",
 #include "./cl_kernels/warp_perspective.clembed"
     },
+    {
+        "winograd.cl",
+#include "./cl_kernels/winograd.clembed"
+    },
 #endif /* EMBEDDED_KERNELS */
 };
 
diff --git a/src/core/CL/cl_kernels/winograd.cl b/src/core/CL/cl_kernels/winograd.cl
new file mode 100644
index 0000000000..fa06601c50
--- /dev/null
+++ b/src/core/CL/cl_kernels/winograd.cl
@@ -0,0 +1,208 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers.h"
+
+#if defined(NUM_TILES_X)
+
+/** This OpenCL kernel computes the input transform when the kernel size is 3x3 and the output tile is 2x2
+ *
+ * @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_2x2_3x3_stepz1_nchw(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    int x = get_global_id(0);
+    int y = get_global_id(1);
+    int z = get_global_id(2);
+
+    // Compute input address
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * 2 * src_stride_x + y * 2 * src_stride_y + z * src_stride_z;
+
+    src_addr = src_addr - ((int)PAD_LEFT * src_stride_x) - ((int)PAD_TOP * src_stride_y);
+
+    float4 in_row0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float4 in_row1 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
+    float4 in_row2 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
+    float4 in_row3 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
+
+    float4 tmp0 = in_row0 - in_row2;
+    float4 tmp1 = in_row1 + in_row2;
+    float4 tmp2 = in_row2 - in_row1;
+    float4 tmp3 = in_row1 - in_row3;
+
+    float out00 = tmp0.s0 - tmp0.s2;
+    float out01 = tmp0.s1 + tmp0.s2;
+    float out02 = tmp0.s2 - tmp0.s1;
+    float out03 = tmp0.s1 - tmp0.s3;
+
+    float out10 = tmp1.s0 - tmp1.s2;
+    float out11 = tmp1.s1 + tmp1.s2;
+    float out12 = tmp1.s2 - tmp1.s1;
+    float out13 = tmp1.s1 - tmp1.s3;
+
+    float out20 = tmp2.s0 - tmp2.s2;
+    float out21 = tmp2.s1 + tmp2.s2;
+    float out22 = tmp2.s2 - tmp2.s1;
+    float out23 = tmp2.s1 - tmp2.s3;
+
+    float out30 = tmp3.s0 - tmp3.s2;
+    float out31 = tmp3.s1 + tmp3.s2;
+    float out32 = tmp3.s2 - tmp3.s1;
+    float out33 = tmp3.s1 - tmp3.s3;
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * dst_stride_x + (x + y * (int)NUM_TILES_X) * dst_stride_y;
+
+    *((__global float *)(dst_addr + 0 * dst_stride_z))  = out00;
+    *((__global float *)(dst_addr + 1 * dst_stride_z))  = out01;
+    *((__global float *)(dst_addr + 2 * dst_stride_z))  = out02;
+    *((__global float *)(dst_addr + 3 * dst_stride_z))  = out03;
+    *((__global float *)(dst_addr + 4 * dst_stride_z))  = out10;
+    *((__global float *)(dst_addr + 5 * dst_stride_z))  = out11;
+    *((__global float *)(dst_addr + 6 * dst_stride_z))  = out12;
+    *((__global float *)(dst_addr + 7 * dst_stride_z))  = out13;
+    *((__global float *)(dst_addr + 8 * dst_stride_z))  = out20;
+    *((__global float *)(dst_addr + 9 * dst_stride_z))  = out21;
+    *((__global float *)(dst_addr + 10 * dst_stride_z)) = out22;
+    *((__global float *)(dst_addr + 11 * dst_stride_z)) = out23;
+    *((__global float *)(dst_addr + 12 * dst_stride_z)) = out30;
+    *((__global float *)(dst_addr + 13 * dst_stride_z)) = out31;
+    *((__global float *)(dst_addr + 14 * dst_stride_z)) = out32;
+    *((__global float *)(dst_addr + 15 * dst_stride_z)) = out33;
+}
+
+/** This OpenCL kernel computes the input transform when the kernel size is 3x3, the output tile is 2x2 and the number of channels is multiple of 2
+ *
+ * @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_2x2_3x3_stepz2_nchw(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    int x = get_global_id(0);
+    int y = get_global_id(1);
+    int z = get_global_id(2) * 2;
+
+    // Compute input address
+    __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * 2 * src_stride_x + y * 2 * src_stride_y + z * src_stride_z;
+
+    src_addr = src_addr - ((int)PAD_LEFT * src_stride_x) - ((int)PAD_TOP * src_stride_y);
+
+    float4 in_row0 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float4 in_row1 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
+    float4 in_row2 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
+    float4 in_row3 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
+
+    src_addr += src_stride_z;
+    float4 in_row4 = vload4(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float4 in_row5 = vload4(0, (__global float *)(src_addr + 1 * src_stride_y));
+    float4 in_row6 = vload4(0, (__global float *)(src_addr + 2 * src_stride_y));
+    float4 in_row7 = vload4(0, (__global float *)(src_addr + 3 * src_stride_y));
+
+    float4 tmp0 = in_row0 - in_row2;
+    float4 tmp1 = in_row1 + in_row2;
+    float4 tmp2 = in_row2 - in_row1;
+    float4 tmp3 = in_row1 - in_row3;
+
+    float4 tmp4 = in_row4 - in_row6;
+    float4 tmp5 = in_row5 + in_row6;
+    float4 tmp6 = in_row6 - in_row5;
+    float4 tmp7 = in_row5 - in_row7;
+
+    float2 out00 = (float2)(tmp0.s0 - tmp0.s2, tmp4.s0 - tmp4.s2);
+    float2 out01 = (float2)(tmp0.s1 + tmp0.s2, tmp4.s1 + tmp4.s2);
+    float2 out02 = (float2)(tmp0.s2 - tmp0.s1, tmp4.s2 - tmp4.s1);
+    float2 out03 = (float2)(tmp0.s1 - tmp0.s3, tmp4.s1 - tmp4.s3);
+
+    float2 out10 = (float2)(tmp1.s0 - tmp1.s2, tmp5.s0 - tmp5.s2);
+    float2 out11 = (float2)(tmp1.s1 + tmp1.s2, tmp5.s1 + tmp5.s2);
+    float2 out12 = (float2)(tmp1.s2 - tmp1.s1, tmp5.s2 - tmp5.s1);
+    float2 out13 = (float2)(tmp1.s1 - tmp1.s3, tmp5.s1 - tmp5.s3);
+
+    float2 out20 = (float2)(tmp2.s0 - tmp2.s2, tmp6.s0 - tmp6.s2);
+    float2 out21 = (float2)(tmp2.s1 + tmp2.s2, tmp6.s1 + tmp6.s2);
+    float2 out22 = (float2)(tmp2.s2 - tmp2.s1, tmp6.s2 - tmp6.s1);
+    float2 out23 = (float2)(tmp2.s1 - tmp2.s3, tmp6.s1 - tmp6.s3);
+
+    float2 out30 = (float2)(tmp3.s0 - tmp3.s2, tmp7.s0 - tmp7.s2);
+    float2 out31 = (float2)(tmp3.s1 + tmp3.s2, tmp7.s1 + tmp7.s2);
+    float2 out32 = (float2)(tmp3.s2 - tmp3.s1, tmp7.s2 - tmp7.s1);
+    float2 out33 = (float2)(tmp3.s1 - tmp3.s3, tmp7.s1 - tmp7.s3);
+
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + z * dst_stride_x + (x + y * (int)NUM_TILES_X) * dst_stride_y;
+
+    vstore2(out00, 0, (__global float *)(dst_addr + 0 * dst_stride_z));
+    vstore2(out01, 0, (__global float *)(dst_addr + 1 * dst_stride_z));
+    vstore2(out02, 0, (__global float *)(dst_addr + 2 * dst_stride_z));
+    vstore2(out03, 0, (__global float *)(dst_addr + 3 * dst_stride_z));
+    vstore2(out10, 0, (__global float *)(dst_addr + 4 * dst_stride_z));
+    vstore2(out11, 0, (__global float *)(dst_addr + 5 * dst_stride_z));
+    vstore2(out12, 0, (__global float *)(dst_addr + 6 * dst_stride_z));
+    vstore2(out13, 0, (__global float *)(dst_addr + 7 * dst_stride_z));
+    vstore2(out20, 0, (__global float *)(dst_addr + 8 * dst_stride_z));
+    vstore2(out21, 0, (__global float *)(dst_addr + 9 * dst_stride_z));
+    vstore2(out22, 0, (__global float *)(dst_addr + 10 * dst_stride_z));
+    vstore2(out23, 0, (__global float *)(dst_addr + 11 * dst_stride_z));
+    vstore2(out30, 0, (__global float *)(dst_addr + 12 * dst_stride_z));
+    vstore2(out31, 0, (__global float *)(dst_addr + 13 * dst_stride_z));
+    vstore2(out32, 0, (__global float *)(dst_addr + 14 * dst_stride_z));
+    vstore2(out33, 0, (__global float *)(dst_addr + 15 * dst_stride_z));
+}
+#endif //defined(NUM_TILES_X)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
new file mode 100644
index 0000000000..72adb5f358
--- /dev/null
+++ b/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp
@@ -0,0 +1,180 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    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_dims.width != 3 || kernel_dims.height != 3, "Winograd input transform only supports 3x3 kernels");
+    ARM_COMPUTE_UNUSED(kernel_dims);
+
+    const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, conv_info, Size2D(3U, 3U));
+
+    // Validate configured output
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_UNUSED(output);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_ON(kernel_dims.width != 3 || kernel_dims.height != 3);
+    ARM_COMPUTE_UNUSED(kernel_dims);
+
+    constexpr unsigned int num_elems_read_per_iteration_x = 4u;
+    constexpr unsigned int num_elems_read_per_iteration_y = 4u;
+
+    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);
+
+    bool window_changed = update_window_and_padding(win, input_access);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+CLWinogradInputTransformKernel::CLWinogradInputTransformKernel()
+    : _border_size(0), _input(nullptr), _output(nullptr), _num_tiles_x(0), _num_tiles_y(0), _step_z(1)
+{
+}
+
+BorderSize CLWinogradInputTransformKernel::border_size() const
+{
+    return _border_size;
+}
+
+void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), conv_info, kernel_dims));
+
+    // Compute number of elements to process in the X and Y direction
+    const int num_elements_x = input->info()->dimension(0) - 2 + conv_info.pad_left() + conv_info.pad_right();
+    const int num_elements_y = input->info()->dimension(1) - 2 + 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 % 2 == 0) ? 0u : 1u;
+    const unsigned int extra_border_bottom = (num_elements_y % 2 == 0) ? 0u : 1u;
+
+    _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());
+    _num_tiles_x = std::ceil(num_elements_x / 2.0f);
+    _num_tiles_y = std::ceil(num_elements_y / 2.0f);
+
+    const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input->info(), conv_info, Size2D(3U, 3U));
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+
+    ARM_COMPUTE_ERROR_ON(_num_tiles_x * _num_tiles_y != static_cast<int>(output->info()->dimension(1)));
+
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(_num_tiles_x));
+    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()));
+
+    // Create kernel
+    if((_input->info()->dimension(2) % 2) != 0)
+    {
+        _step_z = 1;
+        _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("winograd_input_transform_2x2_3x3_stepz1_nchw", build_opts.options()));
+    }
+    else
+    {
+        _step_z   = 2;
+        _kernel   = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("winograd_input_transform_2x2_3x3_stepz2_nchw", build_opts.options()));
+        _lws_hint = cl::NDRange(1, 1, 8);
+    }
+
+    // Create window and update padding
+    auto win_config = validate_and_configure_window(input->info(), output->info(), conv_info, kernel_dims);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure(win_config.second);
+
+    _config_id = "winograd_transform_input_2x2_3x3_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(2));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(conv_info.pad_left());
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(conv_info.pad_top());
+}
+
+Status CLWinogradInputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(validate_arguments(input, output, conv_info, kernel_dims));
+
+    return Status{};
+}
+
+void CLWinogradInputTransformKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    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));
+
+    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));
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
+
+        enqueue(queue, *this, slice, _lws_hint);
+    }
+    while(window.slide_window_slice_3D(slice));
+}
diff --git a/src/runtime/CL/functions/CLWinogradInputTransform.cpp b/src/runtime/CL/functions/CLWinogradInputTransform.cpp
new file mode 100644
index 0000000000..652f31ac74
--- /dev/null
+++ b/src/runtime/CL/functions/CLWinogradInputTransform.cpp
@@ -0,0 +1,45 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/CL/functions/CLWinogradInputTransform.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h"
+#include "arm_compute/core/Error.h"
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+
+void CLWinogradInputTransform::configure(ICLTensor *input, ICLTensor *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLWinogradInputTransformKernel>();
+    k->configure(input, output, conv_info, kernel_dims);
+    _kernel = std::move(k);
+    _border_handler.configure(input, _kernel->border_size(), BorderMode::CONSTANT, PixelValue(0));
+}
+
+Status CLWinogradInputTransform::validate(const ITensorInfo *input, const ITensorInfo *output, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON(CLWinogradInputTransformKernel::validate(input, output, conv_info, kernel_dims));
+    return Status{};
+}
diff --git a/tests/datasets/WinogradInputTransformDataset.h b/tests/datasets/WinogradInputTransformDataset.h
new file mode 100644
index 0000000000..07e41ebf7a
--- /dev/null
+++ b/tests/datasets/WinogradInputTransformDataset.h
@@ -0,0 +1,140 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_TEST_WINOGRAD_INPUT_TRANSFORM_DATASET
+#define ARM_COMPUTE_TEST_WINOGRAD_INPUT_TRANSFORM_DATASET
+
+#include "utils/TypePrinter.h"
+
+#include "arm_compute/core/Types.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+class WinogradInputTransformDataset
+{
+public:
+    using type = std::tuple<TensorShape, PadStrideInfo, Size2D, bool>;
+
+    struct iterator
+    {
+        iterator(std::vector<TensorShape>::const_iterator in_it, std::vector<PadStrideInfo>::const_iterator info_it, std::vector<Size2D>::const_iterator kernel_dims_it,
+                 std::vector<bool>::const_iterator format_it)
+            : _in_it{ std::move(in_it) }, _info_it{ std::move(info_it) }, _kernel_dims_it{ std::move(kernel_dims_it) }, _format_it{ std::move(format_it) }
+        {
+        }
+
+        std::string description() const
+        {
+            std::stringstream description;
+            description << "In=" << *_in_it << ":";
+            description << "Info=" << *_info_it;
+            description << "KernelDims=" << *_kernel_dims_it;
+            description << "IsNCHW=" << *_format_it;
+            return description.str();
+        }
+
+        WinogradInputTransformDataset::type operator*() const
+        {
+            return std::make_tuple(*_in_it, *_info_it, *_kernel_dims_it, *_format_it);
+        }
+
+        iterator &operator++()
+        {
+            ++_in_it;
+            ++_info_it;
+            ++_kernel_dims_it;
+            ++_format_it;
+
+            return *this;
+        }
+
+    private:
+        std::vector<TensorShape>::const_iterator   _in_it;
+        std::vector<PadStrideInfo>::const_iterator _info_it;
+        std::vector<Size2D>::const_iterator        _kernel_dims_it;
+        std::vector<bool>::const_iterator          _format_it;
+    };
+
+    iterator begin() const
+    {
+        return iterator(_in_shapes.begin(), _infos.begin(), _kernel_dims.begin(), _format.begin());
+    }
+
+    int size() const
+    {
+        return std::min(_in_shapes.size(), std::min(_infos.size(), std::min(_kernel_dims.size(), _format.size())));
+    }
+
+    void add_config(TensorShape in, PadStrideInfo info, Size2D kernel_dims, bool format)
+    {
+        _in_shapes.emplace_back(std::move(in));
+        _infos.emplace_back(std::move(info));
+        _kernel_dims.emplace_back(std::move(kernel_dims));
+        _format.emplace_back(std::move(format));
+    }
+
+protected:
+    WinogradInputTransformDataset()                                 = default;
+    WinogradInputTransformDataset(WinogradInputTransformDataset &&) = default;
+
+private:
+    std::vector<TensorShape>   _in_shapes{};
+    std::vector<PadStrideInfo> _infos{};
+    std::vector<Size2D>        _kernel_dims{};
+    std::vector<bool>          _format{};
+};
+
+class SmallWinogradInputTransformDataset final : public WinogradInputTransformDataset
+{
+public:
+    SmallWinogradInputTransformDataset()
+    {
+        add_config(TensorShape(9U, 9U), PadStrideInfo(1, 1, 1, 1), Size2D(3U, 3U), true);
+        add_config(TensorShape(27U, 13U, 2U), PadStrideInfo(1, 1, 0, 0), Size2D(3U, 3U), true);
+        add_config(TensorShape(128U, 64U, 1U, 3U), PadStrideInfo(1, 1, 1, 1), Size2D(3U, 3U), true);
+        add_config(TensorShape(9U, 9U, 3U, 4U), PadStrideInfo(1, 1, 0, 0), Size2D(3U, 3U), true);
+        add_config(TensorShape(27U, 13U, 2U, 4U), PadStrideInfo(1, 1, 1, 1), Size2D(3U, 3U), true);
+        add_config(TensorShape(9U, 9U, 3U, 5U), PadStrideInfo(1, 1, 0, 0), Size2D(3U, 3U), true);
+        add_config(TensorShape(14U, 14U, 512U, 2U), PadStrideInfo(1, 1, 1, 1), Size2D(3U, 3U), true);
+    }
+};
+
+class LargeWinogradInputTransformDataset final : public WinogradInputTransformDataset
+{
+public:
+    LargeWinogradInputTransformDataset()
+    {
+        add_config(TensorShape(42U, 37U, 8U, 15U), PadStrideInfo(1, 1, 1, 1), Size2D(3U, 3U), true);
+        add_config(TensorShape(57U, 60U, 13U, 8U), PadStrideInfo(1, 1, 1, 1), Size2D(3U, 3U), true);
+        add_config(TensorShape(128U, 64U, 21U, 13U), PadStrideInfo(1, 1, 0, 0), Size2D(3U, 3U), true);
+        add_config(TensorShape(83U, 72U, 14U, 5U), PadStrideInfo(1, 1, 0, 0), Size2D(3U, 3U), true);
+    }
+};
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_WINOGRAD_INPUT_TRANSFORM_DATASET */
diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp
new file mode 100644
index 0000000000..664b3f4ef8
--- /dev/null
+++ b/tests/validation/CL/Winograd.cpp
@@ -0,0 +1,135 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONCLCTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "arm_compute/runtime/CL/functions/CLWinogradInputTransform.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/WinogradInputTransformDataset.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/WinogradLayerFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+TEST_SUITE(CL)
+TEST_SUITE(Winograd)
+
+TEST_SUITE(InputTransform)
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
+                                                framework::dataset::make("InputInfo",{
+                                                                                        TensorInfo(TensorShape(53U, 21U, 5U, 3U), 1, DataType::F16),     // F16 not supported
+                                                                                        TensorInfo(TensorShape(53U, 21U, 5U, 3U), 1, DataType::QASYMM8), // QASYMM8 not supported
+                                                                                        TensorInfo(TensorShape(53U, 21U, 5U, 3U), 1, DataType::F32),     // Kernel size not supported
+                                                                                        TensorInfo(TensorShape(53U, 21U, 5U, 3U), 1, DataType::F32),     // Strides not supported
+                                                                                        TensorInfo(TensorShape(53U, 33U, 4U), 1, DataType::F32),         // valid
+                                                                                        TensorInfo(TensorShape(34U, 42U, 7U, 3U), 1, DataType::F32),     // valid
+                                                                                        TensorInfo(TensorShape(31U, 37U, 37U), 1, DataType::F32)         // valid
+                                                                                    }),
+                                                framework::dataset::make("OutputInfo", {
+                                                                                        TensorInfo(TensorShape(5U, 5U, 16U, 3U), 1, DataType::F16),
+                                                                                        TensorInfo(TensorShape(5U, 5U, 16U, 3U), 1, DataType::QASYMM8),
+                                                                                        TensorInfo(TensorShape(5U, 5U, 16U, 3U), 1, DataType::F32),
+                                                                                        TensorInfo(TensorShape(5U, 1U, 16U, 3U), 1, DataType::F32),
+                                                                                        TensorInfo(TensorShape(4U, 442U, 16U), 1, DataType::F32),
+                                                                                        TensorInfo(TensorShape(7U, 320U, 16U, 3U), 1, DataType::F32),
+                                                                                        TensorInfo(TensorShape(37U, 304U, 16U), 1, DataType::F32)
+                                                                                    })),
+                                                framework::dataset::make("PadStrideInfo", {
+                                                                                        PadStrideInfo(1, 1, 1, 0),
+                                                                                        PadStrideInfo(1, 1, 0, 0),
+                                                                                        PadStrideInfo(1, 1, 1, 1),
+                                                                                        PadStrideInfo(2, 1, 1, 1),
+                                                                                        PadStrideInfo(1, 1, 0, 1),
+                                                                                        PadStrideInfo(1, 1, 0, 0),
+                                                                                        PadStrideInfo(1, 1, 1, 1)
+                                                                                    })),
+                                                framework::dataset::make("KernelDims", {
+                                                                                        Size2D(3U, 3U),
+                                                                                        Size2D(3U, 3U),
+                                                                                        Size2D(5U, 5U),
+                                                                                        Size2D(3U, 3U),
+                                                                                        Size2D(3U, 3U),
+                                                                                        Size2D(3U, 3U),
+                                                                                        Size2D(3U, 3U)
+                                                                                    })),
+                                                framework::dataset::make("Expected", { false, false, false, false, true, true, true })),
+                                            input_info, output_info, conv_info, kernel_dims, expected)
+{
+    ARM_COMPUTE_EXPECT(bool(CLWinogradInputTransform::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, kernel_dims)) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+using CLWinogradInputTransformFixture = WinogradInputTransformValidationFixture<CLTensor, CLAccessor, CLWinogradInputTransform, float>;
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallWinogradInputTransformDataset(), datasets::LargeWinogradInputTransformDataset()),
+                                                                   framework::dataset::make("DataType", { DataType::F32 })),
+               shape_in, conv_info, kernel_dims, is_nchw_format, data_type)
+{
+    ARM_COMPUTE_UNUSED(is_nchw_format);
+
+    TensorShape shape_out = compute_winograd_input_transform_shape(TensorInfo(shape_in, 1, data_type), conv_info, kernel_dims);
+
+    // Create tensors
+    CLTensor in  = create_tensor<CLTensor>(shape_in, data_type);
+    CLTensor out = create_tensor<CLTensor>(shape_out, data_type);
+
+    ARM_COMPUTE_EXPECT(in.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(out.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLWinogradInputTransform winograd_input_transform;
+
+    // Configure the function
+    winograd_input_transform.configure(&in, &out, conv_info, kernel_dims);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradInputTransformFixture, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallWinogradInputTransformDataset(), framework::dataset::make("DataType", { DataType::F32 })))
+{
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradInputTransformFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeWinogradInputTransformDataset(), framework::dataset::make("DataType", { DataType::F32 })))
+{
+    validate(CLAccessor(_target), _reference);
+}
+
+TEST_SUITE_END()
+
+TEST_SUITE_END()
+TEST_SUITE_END()
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/Helpers.cpp b/tests/validation/Helpers.cpp
index 313b059a8c..3d554f0d25 100644
--- a/tests/validation/Helpers.cpp
+++ b/tests/validation/Helpers.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,6 +23,9 @@
  */
 #include "tests/validation/Helpers.h"
 
+#include <algorithm>
+#include <cmath>
+
 namespace arm_compute
 {
 namespace test
@@ -133,6 +136,111 @@ SimpleTensor<uint8_t> convert_to_asymmetric(const SimpleTensor<float> &src, cons
     }
     return dst;
 }
+
+void matrix_multiply(const SimpleTensor<float> &a, const SimpleTensor<float> &b, SimpleTensor<float> &out)
+{
+    ARM_COMPUTE_ERROR_ON(a.shape()[0] != b.shape()[1]);
+    ARM_COMPUTE_ERROR_ON(a.shape()[1] != out.shape()[1]);
+    ARM_COMPUTE_ERROR_ON(b.shape()[0] != out.shape()[0]);
+
+    const int M = a.shape()[1]; // Rows
+    const int N = b.shape()[0]; // Cols
+    const int K = b.shape()[1];
+
+    for(int y = 0; y < M; ++y)
+    {
+        for(int x = 0; x < N; ++x)
+        {
+            float acc = 0.0f;
+            for(int k = 0; k < K; ++k)
+            {
+                acc += a[y * K + k] * b[x + k * N];
+            }
+
+            out[x + y * N] = acc;
+        }
+    }
+}
+
+void transpose_matrix(const SimpleTensor<float> &in, SimpleTensor<float> &out)
+{
+    ARM_COMPUTE_ERROR_ON((in.shape()[0] != out.shape()[1]) || (in.shape()[1] != out.shape()[0]));
+
+    const int width  = in.shape()[0];
+    const int height = in.shape()[1];
+
+    for(int y = 0; y < height; ++y)
+    {
+        for(int x = 0; x < width; ++x)
+        {
+            const float val = in[x + y * width];
+
+            out[x * height + y] = val;
+        }
+    }
+}
+
+template <typename T>
+void get_tile(const SimpleTensor<T> &in, SimpleTensor<T> &tile, const Coordinates &coord)
+{
+    ARM_COMPUTE_ERROR_ON(tile.shape().num_dimensions() != 2);
+
+    const int w_tile = tile.shape()[0];
+    const int h_tile = tile.shape()[1];
+
+    // Fill the tile with zeros
+    std::fill(tile.data() + 0, (tile.data() + (w_tile * h_tile)), static_cast<T>(0));
+
+    // Check if with the dimensions greater than 2 we could have out-of-bound reads
+    for(size_t d = 2; d < Coordinates::num_max_dimensions; ++d)
+    {
+        if(coord[d] < 0 || coord[d] >= static_cast<int>(in.shape()[d]))
+        {
+            ARM_COMPUTE_ERROR("coord[d] < 0 || coord[d] >= in.shape()[d] with d >= 2");
+        }
+    }
+
+    // Since we could have out-of-bound reads along the X and Y dimensions,
+    // we start calculating the input address with x = 0 and y = 0
+    Coordinates start_coord = coord;
+    start_coord[0]          = 0;
+    start_coord[1]          = 0;
+
+    // Get input and roi pointers
+    auto in_ptr  = static_cast<const T *>(in(start_coord));
+    auto roi_ptr = static_cast<T *>(tile.data());
+
+    const int x_in_start = std::max(0, coord[0]);
+    const int y_in_start = std::max(0, coord[1]);
+    const int x_in_end   = std::min(static_cast<int>(in.shape()[0]), coord[0] + w_tile);
+    const int y_in_end   = std::min(static_cast<int>(in.shape()[1]), coord[1] + h_tile);
+
+    // Number of elements to copy per row
+    const int n = x_in_end - x_in_start;
+
+    // Starting coordinates for the ROI
+    const int x_tile_start = coord[0] > 0 ? 0 : std::abs(coord[0]);
+    const int y_tile_start = coord[1] > 0 ? 0 : std::abs(coord[1]);
+
+    // Update input pointer
+    in_ptr += x_in_start;
+    in_ptr += (y_in_start * in.shape()[0]);
+
+    // Update ROI pointer
+    roi_ptr += x_tile_start;
+    roi_ptr += (y_tile_start * tile.shape()[0]);
+
+    for(int y = y_in_start; y < y_in_end; ++y)
+    {
+        // Copy per row
+        std::copy(in_ptr, in_ptr + n, roi_ptr);
+
+        in_ptr += in.shape()[0];
+        roi_ptr += tile.shape()[0];
+    }
+}
+
+template void get_tile(const SimpleTensor<float> &in, SimpleTensor<float> &roi, const Coordinates &coord);
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/Helpers.h b/tests/validation/Helpers.h
index ba45968392..b192f317b4 100755
--- a/tests/validation/Helpers.h
+++ b/tests/validation/Helpers.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -232,6 +232,34 @@ SimpleTensor<float> convert_from_asymmetric(const SimpleTensor<uint8_t> &src);
  * @return Quantized tensor.
  */
 SimpleTensor<uint8_t> convert_to_asymmetric(const SimpleTensor<float> &src, const QuantizationInfo &quantization_info);
+
+/** Matrix multiply between 2 float simple tensors
+ *
+ * @param[in]  a   Input tensor A
+ * @param[in]  b   Input tensor B
+ * @param[out] out Output tensor
+ *
+ */
+void matrix_multiply(const SimpleTensor<float> &a, const SimpleTensor<float> &b, SimpleTensor<float> &out);
+
+/** Transpose matrix
+ *
+ * @param[in]  in  Input tensor
+ * @param[out] out Output tensor
+ *
+ */
+void transpose_matrix(const SimpleTensor<float> &in, SimpleTensor<float> &out);
+
+/** Get a 2D tile from a tensor
+ *
+ * @note In case of out-of-bound reads, the tile will be filled with zeros
+ *
+ * @param[in]  in    Input tensor
+ * @param[out] tile  Tile
+ * @param[in]  coord Coordinates
+ */
+template <typename T>
+void get_tile(const SimpleTensor<T> &in, SimpleTensor<T> &tile, const Coordinates &coord);
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/fixtures/WinogradLayerFixture.h b/tests/validation/fixtures/WinogradLayerFixture.h
index d7f0cbfdf5..95e331560d 100644
--- a/tests/validation/fixtures/WinogradLayerFixture.h
+++ b/tests/validation/fixtures/WinogradLayerFixture.h
@@ -26,6 +26,7 @@
 
 #include "arm_compute/core/TensorShape.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 #include "tests/AssetsLibrary.h"
 #include "tests/Globals.h"
@@ -35,6 +36,7 @@
 #include "tests/validation/Helpers.h"
 #include "tests/validation/reference/ConvolutionLayer.h"
 #include "tests/validation/reference/Utils.h"
+#include "tests/validation/reference/Winograd.h"
 
 #include <random>
 
@@ -46,6 +48,8 @@ namespace test
 {
 namespace validation
 {
+using namespace arm_compute::misc::shape_calculator;
+
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
 class WinogradLayerValidationFixture : public framework::Fixture
 {
@@ -139,6 +143,87 @@ protected:
     DataType        _data_type{};
 };
 
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class WinogradInputTransformValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, PadStrideInfo conv_info, Size2D kernel_dims, bool is_nchw_format, DataType data_type)
+    {
+        TensorShape output_shape = compute_winograd_input_transform_shape(TensorInfo(input_shape, 1, data_type), conv_info, kernel_dims);
+
+        _target    = compute_target(input_shape, output_shape, conv_info, kernel_dims, is_nchw_format, data_type);
+        _reference = compute_reference(input_shape, output_shape, conv_info, kernel_dims, is_nchw_format, data_type);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i, float min, float max)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::F32:
+            {
+                std::uniform_real_distribution<> distribution(min, max);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            default:
+            {
+                ARM_COMPUTE_ERROR("Not supported");
+                library->fill_tensor_uniform(tensor, i);
+                break;
+            }
+        }
+    }
+
+    TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims, bool is_nchw_format, DataType data_type)
+    {
+        ARM_COMPUTE_UNUSED(is_nchw_format);
+
+        // Create tensors
+        TensorType src = create_tensor<TensorType>(input_shape, data_type);
+        TensorType dst = create_tensor<TensorType>(output_shape, data_type);
+
+        // Create and configure function
+        FunctionType transf;
+        transf.configure(&src, &dst, conv_info, kernel_dims);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src), 0, -1.f, 1.f);
+
+        // Compute CLWinogradInputTransform function
+        transf.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims, bool is_nchw_format, DataType data_type)
+    {
+        ARM_COMPUTE_UNUSED(is_nchw_format);
+
+        // Create reference
+        SimpleTensor<T> src{ input_shape, data_type };
+
+        // Fill reference
+        fill(src, 0, -1.f, 1.f);
+
+        return reference::winograd_input_transform<T>(src, output_shape, conv_info, kernel_dims);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/reference/Winograd.cpp b/tests/validation/reference/Winograd.cpp
new file mode 100644
index 0000000000..371bb6348e
--- /dev/null
+++ b/tests/validation/reference/Winograd.cpp
@@ -0,0 +1,137 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "Winograd.h"
+
+#include "tests/validation/Helpers.h"
+#include "tests/validation/reference/Utils.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+namespace
+{
+template <typename T>
+void winograd_input_transform3x3(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const PadStrideInfo &conv_info)
+{
+    TensorShape shape4x4(4u, 4u);
+
+    // Simple tensor for the 4x4 input tile
+    SimpleTensor<T> src_tile{ shape4x4, src.data_type() };
+
+    // Simple tensor for the 4x4 temporary tile
+    SimpleTensor<T> tmp_tile{ shape4x4, src.data_type() };
+
+    // Simple tensor for the 4x4 output tile
+    SimpleTensor<T> dst_tile{ shape4x4, src.data_type() };
+
+    // Simple tensor for the transformation matrix
+    SimpleTensor<T> matrix{ shape4x4, src.data_type() };
+
+    // Simple tensor for the transformation matrix transposed
+    SimpleTensor<T> matrix_transposed{ shape4x4, src.data_type() };
+
+    const float matrix_values[] = { 1.f, 0.f, -1.f, 0.f,
+                                    0.f, 1.f, 1.f, 0.f,
+                                    0.f, -1.f, 1.f, 0.f,
+                                    0.f, 1.f, 0.f, -1.f
+                                  };
+
+    for(int i = 0; i < matrix.num_elements(); ++i)
+    {
+        matrix[i] = matrix_values[i];
+    }
+
+    transpose_matrix(matrix, matrix_transposed);
+
+    const int in_w        = src.shape().x();
+    const int in_h        = src.shape().y();
+    const int in_d        = src.shape().z();
+    const int num_batches = src.shape().total_size() / (in_w * in_h * in_d);
+    const int num_tiles_x = std::ceil((in_w - 2 + conv_info.pad_left() + conv_info.pad_right()) / 2.0f);
+    const int num_tiles_y = std::ceil((in_h - 2 + conv_info.pad_top() + conv_info.pad_bottom()) / 2.0f);
+
+    ARM_COMPUTE_ERROR_ON((num_tiles_x * num_tiles_y) != static_cast<int>(dst.shape().y()));
+
+    for(int b = 0; b < num_batches; ++b)
+    {
+        for(int z = 0; z < in_d; ++z)
+        {
+            for(int y = 0; y < num_tiles_y; ++y)
+            {
+                for(int x = 0; x < num_tiles_x; ++x)
+                {
+                    int xi = x * 2 - conv_info.pad_left();
+                    int yi = y * 2 - conv_info.pad_top();
+
+                    // Get the 4x4 tile from the input tensor
+                    get_tile(src, src_tile, Coordinates(xi, yi, z, b));
+
+                    // Compute the transformation
+                    matrix_multiply(matrix, src_tile, tmp_tile);
+                    matrix_multiply(tmp_tile, matrix_transposed, dst_tile);
+
+                    // Store the 4x4 output tile across the 16 channels
+                    for(int i = 0; i < 16; ++i)
+                    {
+                        int xo = z;
+                        int yo = x + y * num_tiles_x;
+                        dst[coords2index(dst.shape(), Coordinates(xo, yo, i, b))] = dst_tile[i];
+                    }
+                }
+            }
+        }
+    }
+}
+} // namespace
+
+template <typename T>
+SimpleTensor<T> winograd_input_transform(const SimpleTensor<T> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims)
+{
+    ARM_COMPUTE_ERROR_ON(kernel_dims.width != kernel_dims.height);
+    ARM_COMPUTE_ERROR_ON(src.data_layout() != DataLayout::NCHW);
+
+    SimpleTensor<T> dst{ dst_shape, src.data_type() };
+
+    switch(kernel_dims.width)
+    {
+        case 3:
+            winograd_input_transform3x3(src, dst, conv_info);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Only 3x3 kernels are supported");
+    }
+
+    return dst;
+}
+
+template SimpleTensor<float> winograd_input_transform(const SimpleTensor<float> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/reference/Winograd.h b/tests/validation/reference/Winograd.h
new file mode 100644
index 0000000000..ed95239db3
--- /dev/null
+++ b/tests/validation/reference/Winograd.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_TEST_WINOGRAD_H__
+#define __ARM_COMPUTE_TEST_WINOGRAD_H__
+
+#include "tests/SimpleTensor.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace reference
+{
+template <typename T>
+SimpleTensor<T> winograd_input_transform(const SimpleTensor<T> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+} // namespace reference
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_TEST_WINOGRAD_H__ */