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>

3 files changed, 394 insertions, 0 deletions

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));
+}