diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index ca2cb04..ef629c2 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -109,6 +109,7 @@
 #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/CLWinogradFilterTransformKernel.h"
 #include "arm_compute/core/CL/kernels/CLWinogradInputTransformKernel.h"
 
 #endif /* __ARM_COMPUTE_CLKERNELS_H__ */
diff --git a/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
new file mode 100644
index 0000000..ec5e514
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h
@@ -0,0 +1,74 @@
+/*
+ * 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_CLWINOGRADFILTERTRANSFORMKERNEL_H__
+#define __ARM_COMPUTE_CLWINOGRADFILTERTRANSFORMKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Interface for the Winograd filter transform kernel. */
+class CLWinogradFilterTransformKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLWinogradFilterTransformKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLWinogradFilterTransformKernel(const CLWinogradFilterTransformKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLWinogradFilterTransformKernel &operator=(const CLWinogradFilterTransformKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLWinogradFilterTransformKernel(CLWinogradFilterTransformKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLWinogradFilterTransformKernel &operator=(CLWinogradFilterTransformKernel &&) = default;
+    /** Default destructor */
+    ~CLWinogradFilterTransformKernel() = default;
+    /** Set the input and output tensor.
+     *
+     * @param[in]  input  Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
+     *                    kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
+     * @param[out] output Destination tensor. The output is a 3D tensor with dimensions [OFM, IFM, 16]. Data type supported: same as @p input
+     */
+    void configure(const ICLTensor *input, ICLTensor *output);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradFilterTransformKernel
+     *
+     * @param[in] input  Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout).
+     *                   kernel_x must be 3 and equal to kernel_y. Data types supported: F32.
+     * @param[in] output Destination tensor info. The output is a 3D tensor with dimensions [OFM, IFM, 16]. Data type supported: same as @p input
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    const ICLTensor *_input;
+    ICLTensor       *_output;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLWINOGRADFILTERTRANSFORMKERNEL_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index 354f60d..9cb8023 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -184,7 +184,7 @@
         output_shape = compute_transposed_shape(*input);
     }
 
-    // If the we run multiple batches we need 1xW transpose, too.
+    // If we run multiple batches we need 1xW transpose, too.
     if(is_batched_fc_layer)
     {
         output_shape = compute_transposed_shape(input->clone()->set_tensor_shape(output_shape));
@@ -193,6 +193,29 @@
 
     return output_shape;
 }
+
+inline TensorShape compute_winograd_filter_transform_shape(const ITensorInfo &input)
+{
+    // COMPMID-984 (giaiod01)
+    TensorShape tensor_shape{ input.tensor_shape() };
+
+    if(input.data_layout() == DataLayout::NCHW)
+    {
+        tensor_shape.remove_dimension(0);
+        tensor_shape.set(Window::DimX, input.dimension(3));
+        tensor_shape.set(Window::DimY, input.dimension(2));
+        tensor_shape.set(Window::DimZ, 16);
+    }
+    else
+    {
+        tensor_shape.remove_dimension(1);
+        tensor_shape.set(Window::DimY, input.dimension(2));
+        tensor_shape.set(Window::DimZ, 16);
+    }
+
+    return tensor_shape;
+}
+
 inline TensorShape compute_winograd_input_transform_shape(const ITensorInfo &input, const PadStrideInfo &conv_info, const Size2D &kernel_size)
 {
     // Compute height
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 40aceb7..4b7fa8a 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -351,6 +351,7 @@
     { "warp_affine_bilinear", "warp_affine.cl" },
     { "warp_perspective_nearest_neighbour", "warp_perspective.cl" },
     { "warp_perspective_bilinear", "warp_perspective.cl" },
+    { "winograd_filter_transform_2x2_3x3_nchw", "winograd.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" },
diff --git a/src/core/CL/cl_kernels/winograd.cl b/src/core/CL/cl_kernels/winograd.cl
index fa06601..238e21a 100644
--- a/src/core/CL/cl_kernels/winograd.cl
+++ b/src/core/CL/cl_kernels/winograd.cl
@@ -205,4 +205,99 @@
     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
+#endif //defined(NUM_TILES_X)
+
+#if defined(NUM_CHANNELS)
+
+/** This OpenCL kernel performs Winograd filter transform 3x3 when the data format is NCHW and the output tile is 2x2
+ *
+ * @note The number of channels must be passed at compile time using -DNUM_CHANNELS: e.g. -DNUM_CHANNELS=64
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor 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 tensor 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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  src_step_w                        src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same 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]  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 Z 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_filter_transform_2x2_3x3_nchw(
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, NUM_CHANNELS);
+
+    const __global uchar *src_addr = tensor4D_offset(&src, 0, 0, 0, 0);
+
+    // Load the values from the input tensor
+    float3 w0 = vload3(0, (__global float *)(src_addr + 0 * src_stride_y));
+    float3 w1 = vload3(0, (__global float *)(src_addr + 1 * src_stride_y));
+    float3 w2 = vload3(0, (__global float *)(src_addr + 2 * src_stride_y));
+
+    // Transform the 3x3 tile in a 4x4 tile
+    float4 out0 = 0.0f;
+    float4 out1 = 0.0f;
+    float4 out2 = 0.0f;
+    float4 out3 = 0.0f;
+
+    // Row 0
+    out0.s0 = (w0.s0);
+    out0.s1 = (w0.s0 + w0.s1 + w0.s2) * 0.5f;
+    out0.s2 = (w0.s0 + w0.s2 - w0.s1) * 0.5f;
+    out0.s3 = (w0.s2);
+
+    // Row 1
+    out1.s0 = (w0.s0 + w1.s0 + w2.s0) * 0.5f;
+    out1.s1 = (w0.s0 + w1.s0 + w2.s0 + w0.s1 + w1.s1 + w2.s1 + w0.s2 + w1.s2 + w2.s2) * 0.25f;
+    out1.s2 = (w0.s0 + w1.s0 + w2.s0 + w0.s2 + w1.s2 + w2.s2 - w0.s1 - w1.s1 - w2.s1) * 0.25f;
+    out1.s3 = (w0.s2 + w1.s2 + w2.s2) * 0.5f;
+
+    // Row 2
+    out2.s0 = (w0.s0 + w2.s0 - w1.s0) * 0.5f;
+    out2.s1 = (w0.s0 + w2.s0 + w0.s1 + w2.s1 + w0.s2 + w2.s2 - w1.s0 - w1.s1 - w1.s2) * 0.25f;
+    out2.s2 = (w0.s0 + w2.s0 + w1.s1 + w0.s2 + w2.s2 - w1.s0 - w0.s1 - w2.s1 - w1.s2) * 0.25f;
+    out2.s3 = (w0.s2 + w2.s2 - w1.s2) * 0.5f;
+
+    // Row 3
+    out3.s0 = (w2.s0);
+    out3.s1 = (w2.s0 + w2.s1 + w2.s2) * 0.5f;
+    out3.s2 = (w2.s0 + w2.s2 - w2.s1) * 0.5f;
+    out3.s3 = (w2.s2);
+
+    int z  = get_global_id(2);
+    int x0 = z / NUM_CHANNELS; // idx filter
+    int y0 = z % NUM_CHANNELS; // idx channel
+
+    // Get output address
+    __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x0 * dst_stride_x + y0 * dst_stride_y;
+
+    // Store the 16 values across the 16 channels
+    *(__global float *)(dst_addr + 0 * dst_stride_z)  = out0.s0;
+    *(__global float *)(dst_addr + 1 * dst_stride_z)  = out0.s1;
+    *(__global float *)(dst_addr + 2 * dst_stride_z)  = out0.s2;
+    *(__global float *)(dst_addr + 3 * dst_stride_z)  = out0.s3;
+    *(__global float *)(dst_addr + 4 * dst_stride_z)  = out1.s0;
+    *(__global float *)(dst_addr + 5 * dst_stride_z)  = out1.s1;
+    *(__global float *)(dst_addr + 6 * dst_stride_z)  = out1.s2;
+    *(__global float *)(dst_addr + 7 * dst_stride_z)  = out1.s3;
+    *(__global float *)(dst_addr + 8 * dst_stride_z)  = out2.s0;
+    *(__global float *)(dst_addr + 9 * dst_stride_z)  = out2.s1;
+    *(__global float *)(dst_addr + 10 * dst_stride_z) = out2.s2;
+    *(__global float *)(dst_addr + 11 * dst_stride_z) = out2.s3;
+    *(__global float *)(dst_addr + 12 * dst_stride_z) = out3.s0;
+    *(__global float *)(dst_addr + 13 * dst_stride_z) = out3.s1;
+    *(__global float *)(dst_addr + 14 * dst_stride_z) = out3.s2;
+    *(__global float *)(dst_addr + 15 * dst_stride_z) = out3.s3;
+}
+#endif // defined(NUM_CHANNELS)
diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
new file mode 100644
index 0000000..3dbbe15
--- /dev/null
+++ b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp
@@ -0,0 +1,139 @@
+/*
+ * 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/CLWinogradFilterTransformKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.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/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "support/ToolchainSupport.h"
+
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != 3);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != input->dimension(1));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*input));
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    constexpr unsigned int num_elems_processed_per_iteration_x = 3;
+    constexpr unsigned int num_elems_processed_per_iteration_y = 3;
+
+    Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+    bool   window_changed = false;
+
+    AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+    AccessWindowStatic    output_access(output, 0, 0, output->dimension(0), output->dimension(1));
+    window_changed = update_window_and_padding(win, input_access, output_access);
+    output_access.set_valid_region(win, input->valid_region());
+
+    Window win_collapsed = win.collapse(win, Window::DimZ);
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win_collapsed);
+}
+} // namespace
+
+CLWinogradFilterTransformKernel::CLWinogradFilterTransformKernel()
+    : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLWinogradFilterTransformKernel::configure(const ICLTensor *input, ICLTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Output tensor auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*input->info())));
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DNUM_CHANNELS=" + support::cpp11::to_string(input->info()->dimension(2)));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("winograd_filter_transform_2x2_3x3_nchw", build_opts.options()));
+
+    _input  = input;
+    _output = output;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure(win_config.second);
+}
+
+Status CLWinogradFilterTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
+
+    return Status{};
+}
+
+void CLWinogradFilterTransformKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    // Setup output window
+    Window window_out;
+    window_out.use_tensor_dimensions(_output->info()->tensor_shape(), 0);
+
+    unsigned int idx = 0;
+    add_4D_tensor_argument(idx, _input, window);
+    add_3D_tensor_argument(idx, _output, window_out);
+    enqueue(queue, *this, window);
+}
diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h
index 4b56370..e939a6f 100644
--- a/tests/datasets/ShapeDatasets.h
+++ b/tests/datasets/ShapeDatasets.h
@@ -238,6 +238,38 @@
     }
 };
 
+/** Data set containing medium 3D tensor shapes. */
+class Medium3DShapes final : public ShapeDataset
+{
+public:
+    Medium3DShapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 42U, 37U, 8U },
+                     TensorShape{ 57U, 60U, 13U },
+                     TensorShape{ 128U, 64U, 21U },
+                     TensorShape{ 83U, 72U, 14U }
+    })
+    {
+    }
+};
+
+/** Data set containing medium 4D tensor shapes. */
+class Medium4DShapes final : public ShapeDataset
+{
+public:
+    Medium4DShapes()
+        : ShapeDataset("Shape",
+    {
+        TensorShape{ 42U, 37U, 8U, 15U },
+                     TensorShape{ 57U, 60U, 13U, 8U },
+                     TensorShape{ 128U, 64U, 21U, 13U },
+                     TensorShape{ 83U, 72U, 14U, 5U }
+    })
+    {
+    }
+};
+
 /** Data set containing large tensor shapes. */
 class LargeShapes final : public ShapeDataset
 {
diff --git a/tests/datasets/WinogradFilterTransformDataset.h b/tests/datasets/WinogradFilterTransformDataset.h
new file mode 100644
index 0000000..07d0283
--- /dev/null
+++ b/tests/datasets/WinogradFilterTransformDataset.h
@@ -0,0 +1,128 @@
+/*
+ * 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_FILTER_TRANSFORM_DATASET
+#define ARM_COMPUTE_TEST_WINOGRAD_FILTER_TRANSFORM_DATASET
+
+#include "utils/TypePrinter.h"
+
+#include "arm_compute/core/TensorShape.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+class WinogradFilterTransformDataset
+{
+public:
+    using type = std::tuple<TensorShape, bool>;
+
+    struct iterator
+    {
+        iterator(std::vector<TensorShape>::const_iterator a_it,
+                 std::vector<bool>::const_iterator        is_nchw_it)
+            : _a_it{ std::move(a_it) },
+              _is_nchw_it{ std::move(is_nchw_it) }
+        {
+        }
+
+        std::string description() const
+        {
+            std::stringstream description;
+            description << "Input=" << *_a_it << ":";
+            description << "IsNCHW=" << *_is_nchw_it << ":";
+            return description.str();
+        }
+
+        WinogradFilterTransformDataset::type operator*() const
+        {
+            return std::make_tuple(*_a_it, *_is_nchw_it);
+        }
+
+        iterator &operator++()
+        {
+            ++_a_it;
+            ++_is_nchw_it;
+
+            return *this;
+        }
+
+    private:
+        std::vector<TensorShape>::const_iterator _a_it;
+        std::vector<bool>::const_iterator        _is_nchw_it;
+    };
+
+    iterator begin() const
+    {
+        return iterator(_a_shapes.begin(), _is_nchw.begin());
+    }
+
+    int size() const
+    {
+        return std::min(_a_shapes.size(), _is_nchw.size());
+    }
+
+    void add_config(TensorShape a, bool is_nchw)
+    {
+        _a_shapes.emplace_back(std::move(a));
+        _is_nchw.emplace_back(std::move(is_nchw));
+    }
+
+protected:
+    WinogradFilterTransformDataset()                                  = default;
+    WinogradFilterTransformDataset(WinogradFilterTransformDataset &&) = default;
+
+private:
+    std::vector<TensorShape> _a_shapes{};
+    std::vector<bool>        _is_nchw{};
+};
+
+class SmallWinogradFilterTransformDataset final : public WinogradFilterTransformDataset
+{
+public:
+    SmallWinogradFilterTransformDataset()
+    {
+        add_config(TensorShape(3U, 3U, 7U, 4U), true);
+        add_config(TensorShape(3U, 3U, 4U, 13U), true);
+        add_config(TensorShape(3U, 3U, 9U, 2U), true);
+        add_config(TensorShape(3U, 3U, 3U, 5U), true);
+    }
+};
+
+class LargeWinogradFilterTransformDataset final : public WinogradFilterTransformDataset
+{
+public:
+    LargeWinogradFilterTransformDataset()
+    {
+        add_config(TensorShape(3U, 3U, 32U, 64U), true);
+        add_config(TensorShape(3U, 3U, 51U, 13U), true);
+        add_config(TensorShape(3U, 3U, 53U, 47U), true);
+        add_config(TensorShape(3U, 3U, 128U, 384U), true);
+    }
+};
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_WINOGRAD_FILTER_TRANSFORM_DATASET */
diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp
index 664b3f4..0b21ed2 100644
--- a/tests/validation/CL/Winograd.cpp
+++ b/tests/validation/CL/Winograd.cpp
@@ -18,15 +18,20 @@
  * 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
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
+#include "arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h"
 #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/CL/Helper.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/datasets/WinogradFilterTransformDataset.h"
 #include "tests/datasets/WinogradInputTransformDataset.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Macros.h"
@@ -40,6 +45,13 @@
 {
 namespace validation
 {
+namespace
+{
+constexpr AbsoluteTolerance<float> tolerance_f32(0.0001f);
+} // namespace
+
+using namespace arm_compute::misc::shape_calculator;
+
 TEST_SUITE(CL)
 TEST_SUITE(Winograd)
 
@@ -125,11 +137,76 @@
 {
     validate(CLAccessor(_target), _reference);
 }
+TEST_SUITE_END() // InputTransform
 
-TEST_SUITE_END()
+TEST_SUITE(FilterTransform)
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(
+                                                framework::dataset::make("InputInfo",{
+                                                                                        TensorInfo(TensorShape(3U, 3U, 5U, 3U), 1, DataType::F16),     // F16 not supported
+                                                                                        TensorInfo(TensorShape(3U, 3U, 5U, 3U), 1, DataType::QASYMM8), // QASYMM8 not supported
+                                                                                        TensorInfo(TensorShape(5U, 5U, 5U, 3U), 1, DataType::F32),     // Kernel size not supported
+                                                                                        TensorInfo(TensorShape(3U, 3U), 1, DataType::F32),             // valid
+                                                                                        TensorInfo(TensorShape(3U, 3U, 5U, 3U), 1, DataType::F32),     // valid
+                                                                                        TensorInfo(TensorShape(3U, 3U, 37U, 2U), 1, DataType::F32),    // valid
+                                                                                        TensorInfo(TensorShape(3U, 3U, 37U, 22U), 1, DataType::F32)    // valid
+                                                                                    }),
+                                                framework::dataset::make("OutputInfo", {
+                                                                                        TensorInfo(TensorShape(3U, 5U, 16U), 1, DataType::F16),
+                                                                                        TensorInfo(TensorShape(3U, 5U, 16U), 1, DataType::QASYMM8),
+                                                                                        TensorInfo(TensorShape(3U, 5U, 16U), 1, DataType::F32),
+                                                                                        TensorInfo(TensorShape(1U, 1U, 16U), 1, DataType::F32),
+                                                                                        TensorInfo(TensorShape(3U, 5U, 16U), 1, DataType::F32),
+                                                                                        TensorInfo(TensorShape(2U, 37U, 16U), 1, DataType::F32),
+                                                                                        TensorInfo(TensorShape(22U, 37U, 16U), 1, DataType::F32)
+                                                                                    })),
+                                                framework::dataset::make("Expected", { false, false, false, true, true, true, true })),
+                                            input_info, output_info, expected)
+{
+    ARM_COMPUTE_EXPECT(bool(CLWinogradFilterTransformKernel::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+using CLWinogradFilterTransform        = CLSynthetizeFunctionWithZeroConstantBorder<CLWinogradFilterTransformKernel, 0>;
+using CLWinogradFilterTransformFixture = WinogradFilterTransformValidationFixture<CLTensor, CLAccessor, CLWinogradFilterTransform, float>;
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallWinogradFilterTransformDataset(), datasets::LargeWinogradFilterTransformDataset()),
+                                                                   framework::dataset::make("DataType", { DataType::F32 })),
+               shape_a, is_nchw_format, data_type)
+{
+    ARM_COMPUTE_UNUSED(is_nchw_format);
+
+    TensorShape shape_b = compute_winograd_filter_transform_shape(TensorInfo(shape_a, 1, data_type));
+
+    // Create tensors
+    CLTensor a = create_tensor<CLTensor>(shape_a, data_type);
+    CLTensor b = create_tensor<CLTensor>(shape_b, data_type);
+
+    ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLWinogradFilterTransform winograd_filter_transform;
+    winograd_filter_transform.configure(&a, &b);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixture, framework::DatasetMode::ALL, combine(datasets::SmallWinogradFilterTransformDataset(), framework::dataset::make("DataType", { DataType::F32 })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeWinogradFilterTransformDataset(), framework::dataset::make("DataType", { DataType::F32 })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // FilterTransform
+
+TEST_SUITE_END() // Winograd
+TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/Helpers.h b/tests/validation/Helpers.h
old mode 100755
new mode 100644
diff --git a/tests/validation/fixtures/WinogradLayerFixture.h b/tests/validation/fixtures/WinogradLayerFixture.h
index 95e3315..bfe1efc 100644
--- a/tests/validation/fixtures/WinogradLayerFixture.h
+++ b/tests/validation/fixtures/WinogradLayerFixture.h
@@ -27,7 +27,6 @@
 #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"
 #include "tests/IAccessor.h"
@@ -42,8 +41,6 @@
 
 namespace arm_compute
 {
-class NEWinogradLayer;
-
 namespace test
 {
 namespace validation
@@ -224,6 +221,87 @@
     TensorType      _target{};
     SimpleTensor<T> _reference{};
 };
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class WinogradFilterTransformValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, bool is_nchw_format, DataType data_type)
+    {
+        TensorShape output_shape = compute_winograd_filter_transform_shape(TensorInfo(input_shape, 1, data_type));
+
+        _target    = compute_target(input_shape, output_shape, is_nchw_format, data_type);
+        _reference = compute_reference(input_shape, output_shape, 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, 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 filter_transform;
+        filter_transform.configure(&src, &dst);
+
+        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);
+
+        filter_transform.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, bool is_nchw_format, DataType data_type)
+    {
+        ARM_COMPUTE_ERROR_ON(!is_nchw_format);
+
+        // Create reference
+        SimpleTensor<T> src{ input_shape, data_type, 1 };
+
+        // Fill reference
+        fill(src, 0, -1.f, 1.f);
+
+        return reference::winograd_filter_transform<T>(src, output_shape);
+    }
+
+    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
index 371bb63..3ed55fb 100644
--- a/tests/validation/reference/Winograd.cpp
+++ b/tests/validation/reference/Winograd.cpp
@@ -26,6 +26,8 @@
 #include "tests/validation/Helpers.h"
 #include "tests/validation/reference/Utils.h"
 
+#include "arm_compute/core/Types.h"
+
 namespace arm_compute
 {
 namespace test
@@ -108,6 +110,87 @@
         }
     }
 }
+
+template <typename T>
+void winograd_filter_transform3x3(const SimpleTensor<T> &in, SimpleTensor<T> &out)
+{
+    // Simple tensor for the 3x3 input tile
+    SimpleTensor<T> input_tile{ TensorShape(3u, 3u), in.data_type(), 1 };
+
+    // Simple tensor for the transformation matrix
+    SimpleTensor<T> trans_matrix{ TensorShape(3u, 4u), in.data_type(), 1 };
+
+    // Simple tensor for the transformation matrix transpose
+    SimpleTensor<T> trans_matrix_transposed{ TensorShape(4u, 3u), in.data_type(), 1 };
+
+    // Simple tensor for the 4x3 temporary tile
+    SimpleTensor<T> tmp_tile{ TensorShape(3u, 4u), in.data_type(), 1 };
+
+    // Simple tensor for the 4x4 output tile
+    SimpleTensor<T> output_tile{ TensorShape(4u, 4u), in.data_type(), 1 };
+
+    // Initialize transformation matrix
+    // 1   | 0   | 0
+    // 0.5 | 0.5 | 0.5
+    // 0.5 |-0.5 | 0.5
+    // 0   | 0   | 1
+    trans_matrix[0 + 0 * 3] = 1.0f;
+    trans_matrix[1 + 0 * 3] = 0.0f;
+    trans_matrix[2 + 0 * 3] = 0.0f;
+    trans_matrix[0 + 1 * 3] = 0.5f;
+    trans_matrix[1 + 1 * 3] = 0.5f;
+    trans_matrix[2 + 1 * 3] = 0.5f;
+    trans_matrix[0 + 2 * 3] = 0.5f;
+    trans_matrix[1 + 2 * 3] = -0.5f;
+    trans_matrix[2 + 2 * 3] = 0.5f;
+    trans_matrix[0 + 3 * 3] = 0.0f;
+    trans_matrix[1 + 3 * 3] = 0.0f;
+    trans_matrix[2 + 3 * 3] = 1.0f;
+
+    // Transpose the transformation matrix
+    transpose_matrix(trans_matrix, trans_matrix_transposed);
+
+    const int num_channels = in.shape()[2];
+    const int num_filters  = in.shape()[3];
+    const int num_batches  = in.shape().total_size() / (9 * num_channels * num_filters);
+
+    for(int n = 0; n < num_batches; ++n)
+    {
+        for(int w = 0; w < num_filters; ++w)
+        {
+            for(int z = 0; z < num_channels; ++z)
+            {
+                // Load the 3x3 tile from the input tensor
+                get_tile(in, input_tile, Coordinates(0, 0, z, w, n));
+
+                // First transformation
+                matrix_multiply(trans_matrix, input_tile, tmp_tile);
+
+                // Second transformation
+                matrix_multiply(tmp_tile, trans_matrix_transposed, output_tile);
+
+                // Store the 4x4 output tile across the 16 channels
+                const int output_offset                              = w + z * num_filters;
+                out[output_offset + 0 * num_filters * num_channels]  = output_tile[0 + 0 * 4];
+                out[output_offset + 1 * num_filters * num_channels]  = output_tile[1 + 0 * 4];
+                out[output_offset + 2 * num_filters * num_channels]  = output_tile[2 + 0 * 4];
+                out[output_offset + 3 * num_filters * num_channels]  = output_tile[3 + 0 * 4];
+                out[output_offset + 4 * num_filters * num_channels]  = output_tile[0 + 1 * 4];
+                out[output_offset + 5 * num_filters * num_channels]  = output_tile[1 + 1 * 4];
+                out[output_offset + 6 * num_filters * num_channels]  = output_tile[2 + 1 * 4];
+                out[output_offset + 7 * num_filters * num_channels]  = output_tile[3 + 1 * 4];
+                out[output_offset + 8 * num_filters * num_channels]  = output_tile[0 + 2 * 4];
+                out[output_offset + 9 * num_filters * num_channels]  = output_tile[1 + 2 * 4];
+                out[output_offset + 10 * num_filters * num_channels] = output_tile[2 + 2 * 4];
+                out[output_offset + 11 * num_filters * num_channels] = output_tile[3 + 2 * 4];
+                out[output_offset + 12 * num_filters * num_channels] = output_tile[0 + 3 * 4];
+                out[output_offset + 13 * num_filters * num_channels] = output_tile[1 + 3 * 4];
+                out[output_offset + 14 * num_filters * num_channels] = output_tile[2 + 3 * 4];
+                out[output_offset + 15 * num_filters * num_channels] = output_tile[3 + 3 * 4];
+            }
+        }
+    }
+}
 } // namespace
 
 template <typename T>
@@ -130,7 +213,29 @@
     return dst;
 }
 
+template <typename T>
+SimpleTensor<T> winograd_filter_transform(const SimpleTensor<T> &in, const TensorShape &output_shape)
+{
+    ARM_COMPUTE_ERROR_ON_MSG(in.data_layout() != DataLayout::NCHW, "Only supported NCHW data format");
+
+    // Create reference
+    SimpleTensor<T> out{ output_shape, in.data_type(), 1 };
+
+    switch(in.shape()[0])
+    {
+        case 3:
+            winograd_filter_transform3x3(in, out);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Only supported 3x3 kernel");
+            break;
+    }
+
+    return out;
+}
+
 template SimpleTensor<float> winograd_input_transform(const SimpleTensor<float> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+template SimpleTensor<float> winograd_filter_transform(const SimpleTensor<float> &in, const TensorShape &output_shape);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/Winograd.h b/tests/validation/reference/Winograd.h
index ed95239..ba8e5c1 100644
--- a/tests/validation/reference/Winograd.h
+++ b/tests/validation/reference/Winograd.h
@@ -24,6 +24,8 @@
 #ifndef __ARM_COMPUTE_TEST_WINOGRAD_H__
 #define __ARM_COMPUTE_TEST_WINOGRAD_H__
 
+#include "arm_compute/core/TensorShape.h"
+
 #include "tests/SimpleTensor.h"
 
 namespace arm_compute
@@ -36,6 +38,9 @@
 {
 template <typename T>
 SimpleTensor<T> winograd_input_transform(const SimpleTensor<T> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims);
+
+template <typename T>
+SimpleTensor<T> winograd_filter_transform(const SimpleTensor<T> &in, const TensorShape &output_shape);
 } // namespace reference
 } // namespace validation
 } // namespace test