Optimize Transposed Convolution for CL backend (FP32/16)

This patch optimizes transposed convolution for CL backend by rewriting it in a single kernel instead of three (flip_kernel + upsample + conv). The new kernel skips the upsampling step which reduces the input space of convolution by stride_x * stride_y, resulting in significant performance improvement. It also skips the kernel flipping by traversing the weights accordingly, thus reduces the memory footprint.

Resolves: COMPMID-5676
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: I8a333212dc7c5f7f0597aa58b0d56d44814baa14
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8588
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 450 insertions, 0 deletions

diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp
index 0f08f5d044..4e036399db 100644
--- a/src/gpu/cl/ClKernelLibrary.cpp
+++ b/src/gpu/cl/ClKernelLibrary.cpp
@@ -448,6 +448,7 @@ const std::map<std::string, std::string> ClKernelLibrary::_kernel_program_map =
     { "space_to_batch_nhwc", "nhwc/space_to_batch.cl" },
     { "space_to_batch_static_nhwc", "nhwc/space_to_batch.cl" },
     { "space_to_depth_nhwc", "nhwc/space_to_depth.cl" },
+    { "transposed_convolution_nhwc", "nhwc/transposed_convolution.cl" },
     { "upsample_layer_nhwc", "nhwc/upsample_layer.cl" },
     { "winograd_filter_transform_4x1_3x1_nhwc", "nhwc/winograd_filter_transform.cl" },
     { "winograd_filter_transform_1x4_1x3_nhwc", "nhwc/winograd_filter_transform.cl" },
@@ -943,6 +944,10 @@ const std::map<std::string, std::string> ClKernelLibrary::_program_source_map =
 #include "./cl_kernels/nhwc/space_to_depth.clembed"
     },
     {
+        "nhwc/transposed_convolution.cl",
+#include "./cl_kernels/nhwc/transposed_convolution.clembed"
+    },
+    {
         "nhwc/winograd_filter_transform.cl",
 #include "./cl_kernels/nhwc/winograd_filter_transform.clembed"
     },
diff --git a/src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp b/src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp
new file mode 100644
index 0000000000..16c6ad9a9b
--- /dev/null
+++ b/src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp
@@ -0,0 +1,231 @@
+/*
+ * Copyright (c) 2022 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 "src/gpu/cl/kernels/ClTransposedConvolutionKernel.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "src/core/CL/CLValidate.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/WindowHelpers.h"
+#include "support/Cast.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
+                          const PadStrideInfo &deconv_info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(weights, DataLayout::NHWC);
+
+    constexpr unsigned int channel_idx = 0;
+    constexpr unsigned int width_idx   = 1;
+    constexpr unsigned int height_idx  = 2;
+    constexpr unsigned int batch_idx   = 3;
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(channel_idx) != input->dimension(channel_idx), "Weights feature map dimension should match the respective src's one");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional");
+
+    if(biases != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->dimension(channel_idx) != weights->dimension(batch_idx),
+                                        "Biases size and number of dst feature maps should match");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->num_dimensions() > 1, "Biases should be one dimensional");
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
+    }
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        const size_t input_width    = input->dimension(width_idx);
+        const size_t input_height   = input->dimension(height_idx);
+        const size_t weights_width  = weights->dimension(width_idx);
+        const size_t weights_height = weights->dimension(height_idx);
+
+        auto        out_dims     = deconvolution_output_dimensions(input_width, input_height, weights_width, weights_height, deconv_info);
+        TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input, *weights);
+
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(output, DataLayout::NHWC);
+    }
+
+    return Status{};
+}
+} // namespace
+
+void ClTransposedConvolutionKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *input, const ITensorInfo *weights,
+                                              const ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &deconv_info)
+{
+    ARM_COMPUTE_UNUSED(biases, deconv_info);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+
+    // Perform validation
+    ARM_COMPUTE_ERROR_THROW_ON(validate(input, weights, biases, output, deconv_info));
+
+    constexpr unsigned int channel_idx = 0;
+    constexpr unsigned int width_idx   = 1;
+    constexpr unsigned int height_idx  = 2;
+
+    const size_t input_channels  = input->dimension(channel_idx); // same as weight channels
+    const size_t input_width     = input->dimension(width_idx);
+    const size_t input_height    = input->dimension(height_idx);
+    const size_t weights_width   = weights->dimension(width_idx);
+    const size_t weights_height  = weights->dimension(height_idx);
+    const size_t output_width    = output->dimension(width_idx);
+    const size_t output_height   = output->dimension(height_idx);
+    const size_t output_channels = output->dimension(channel_idx);
+
+    // Calculate output shape
+    auto        out_dims     = deconvolution_output_dimensions(input_width, input_height, weights_width, weights_height, deconv_info);
+    TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input, *weights);
+    auto_init_if_empty(*output, output_shape, 1, input->data_type(), input->quantization_info());
+
+    // Calculate updated paddings
+    // p' = k - p - 1 (k: kernel dimensions)
+    const uint32_t pad_left = weights_width - deconv_info.pad_left() - 1;
+    const uint32_t pad_top  = weights_height - deconv_info.pad_top() - 1;
+
+    // Configure kernel window
+    Window win;
+    output_shape.collapse(2U, 1U); // Collapse width and height into single dimension
+
+    // Create window and update padding
+    win = calculate_max_window(output_shape, Steps(1, 1));
+    ICLKernel::configure_internal(win);
+
+    const std::string kernel_name = "transposed_convolution_nhwc";
+    CLBuildOptions    build_options;
+
+    const DataType input_data_type = input->data_type();        // Fp32 or Fp16 only
+    const auto     strides         = deconv_info.stride();
+
+    const unsigned int n0               = 1;
+    const unsigned int m0               = 1;
+    const unsigned int k0               = adjust_vec_size(input_data_type == DataType::F32 ? 4 : 8, input_channels);
+    const unsigned int partial_store_n0 = output_channels % n0;
+
+    if(biases != nullptr)
+    {
+        build_options.add_option(std::string("-DHAS_BIAS"));
+        build_options.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->data_type())));
+    }
+
+    const auto output_data_type = output->data_type();
+
+    build_options.add_option("-cl-fast-relaxed-math");
+    build_options.add_option("-DSRC_TENSOR_TYPE=BUFFER");
+    build_options.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(input_data_type));
+    build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(input_channels));
+    build_options.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input_width));
+    build_options.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input_height));
+    build_options.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(output_channels));
+    build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(output_width));
+    build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(output_height));
+    build_options.add_option("-DDST_TENSOR_TYPE=BUFFER");
+    build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(output_data_type));
+    build_options.add_option("-DWEI_TENSOR_TYPE=BUFFER");
+    build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights_width));
+    build_options.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights_height));
+    build_options.add_option("-DWEI_DATA_TYPE=" + get_cl_type_from_data_type(weights->data_type()));
+    build_options.add_option("-DSTRIDE_X=" + support::cpp11::to_string(strides.first));
+    build_options.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(strides.second));
+    build_options.add_option("-DPAD_LEFT=" + support::cpp11::to_string(pad_left));
+    build_options.add_option("-DPAD_TOP=" + support::cpp11::to_string(pad_top));
+    build_options.add_option("-DN0=" + support::cpp11::to_string(n0));
+    build_options.add_option("-DM0=" + support::cpp11::to_string(m0));
+    build_options.add_option("-DK0=" + support::cpp11::to_string(k0));
+    build_options.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
+    build_options.add_option_if((input_channels % k0) != 0, "-DLEFTOVER_LOOP");
+    build_options.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(input_data_type));
+
+    if(compile_context.get_ddk_version() >= 30)
+    {
+        build_options.add_option("-fregister-allocation=64");
+    }
+
+    _kernel = create_kernel(compile_context, kernel_name, build_options.options());
+
+    // Set config_id for enabling LWS tuning
+    _config_id = kernel_name;
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_type(input_data_type));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(weights_width);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(strides.first);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(strides.second);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(output_width);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(m0);
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(n0);
+}
+
+Status ClTransposedConvolutionKernel::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases,
+                                               const ITensorInfo *dst, const PadStrideInfo &deconv_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, deconv_info));
+    return Status{};
+}
+
+void ClTransposedConvolutionKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    // Get initial windows
+    Window slice = window.first_slice_window_3D();
+
+    const auto src     = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
+    const auto weights = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
+    const auto biases  = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
+    auto       dst     = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+
+    unsigned int idx = 0;
+    add_4d_tensor_nhwc_argument(idx, src);
+    add_4d_tensor_nhwc_argument(idx, dst);
+
+    add_4d_tensor_nhwc_argument(idx, weights);
+    if(biases != nullptr)
+    {
+        add_1D_tensor_argument(idx, biases, slice);
+    }
+
+    enqueue(queue, *this, slice, lws_hint());
+}
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/gpu/cl/kernels/ClTransposedConvolutionKernel.h b/src/gpu/cl/kernels/ClTransposedConvolutionKernel.h
new file mode 100644
index 0000000000..d4350dda50
--- /dev/null
+++ b/src/gpu/cl/kernels/ClTransposedConvolutionKernel.h
@@ -0,0 +1,66 @@
+/*
+ * Copyright (c) 2022 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_CL_TRANSPOSED_CONVOLUTION_KERNEL_H
+#define ARM_COMPUTE_CL_TRANSPOSED_CONVOLUTION_KERNEL_H
+
+#include "src/core/common/Macros.h"
+#include "src/gpu/cl/ClCompileContext.h"
+#include "src/gpu/cl/IClKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+namespace kernels
+{
+/** OpenCL kernel for transposed convolution. */
+class ClTransposedConvolutionKernel : public IClKernel
+{
+public:
+    ClTransposedConvolutionKernel() = default;
+    ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClTransposedConvolutionKernel);
+    /** Set the input, weights, biases and output tensors.
+     *
+     * Similar to @ref ClTransposedConvolution::configure()
+     *
+     */
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *input, const ITensorInfo *weights,
+                   const ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &deconv_info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to @ref ClTransposedConvolution::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases,
+                           const ITensorInfo *output, const PadStrideInfo &deconv_info);
+
+    // Inherited methods overridden:
+    void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
+};
+} // namespace kernels
+} // namespace opencl
+} // namespace arm_compute
+
+#endif /* ARM_COMPUTE_CL_TRANSPOSED_CONVOLUTION_KERNEL_H */
\ No newline at end of file
diff --git a/src/gpu/cl/operators/ClTransposedConvolution.cpp b/src/gpu/cl/operators/ClTransposedConvolution.cpp
new file mode 100644
index 0000000000..90dbe7f291
--- /dev/null
+++ b/src/gpu/cl/operators/ClTransposedConvolution.cpp
@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2022 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 "src/gpu/cl/operators/ClTransposedConvolution.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+#include "src/common/utils/Log.h"
+#include "src/gpu/cl/kernels/ClTransposedConvolutionKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+void ClTransposedConvolution::configure(const CLCompileContext &compile_context, const ITensorInfo *input, const ITensorInfo *weights,
+                                        const ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &deconv_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_LOG_PARAMS(input, weights, biases, output, deconv_info);
+    auto kernel_object = std::make_unique<kernels::ClTransposedConvolutionKernel>();
+    kernel_object->set_target(CLScheduler::get().target());
+    kernel_object->configure(compile_context, input, weights, biases, output, deconv_info);
+    _transposed_conv_kernel = std::move(kernel_object);
+}
+
+Status ClTransposedConvolution::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases,
+                                         const ITensorInfo *output, const PadStrideInfo &deconv_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClTransposedConvolutionKernel::validate(input, weights, biases, output, deconv_info));
+    return Status{};
+}
+
+void ClTransposedConvolution::run(ITensorPack &tensors)
+{
+    CLScheduler::get().enqueue_op(*_transposed_conv_kernel.get(), tensors, false);
+}
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/gpu/cl/operators/ClTransposedConvolution.h b/src/gpu/cl/operators/ClTransposedConvolution.h
new file mode 100644
index 0000000000..bc04387df5
--- /dev/null
+++ b/src/gpu/cl/operators/ClTransposedConvolution.h
@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) 2022 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_CL_TRANSPOSED_CONVOLUTION_H
+#define ARM_COMPUTE_CL_TRANSPOSED_CONVOLUTION_H
+
+#include "src/gpu/cl/ClCompileContext.h"
+#include "src/gpu/cl/IClKernel.h"
+#include "src/gpu/cl/IClOperator.h"
+namespace arm_compute
+{
+namespace opencl
+{
+/** Basic function to simulate a directly convolution layer. This function calls the following OpenCL kernels:
+ *
+ * -# @ref opencl::ClTransposedConvolution
+ */
+class ClTransposedConvolution : public IClOperator
+{
+public:
+    /** Default constructor */
+    ClTransposedConvolution() = default;
+    /** Default Destructor */
+    ~ClTransposedConvolution() = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    ClTransposedConvolution(const ClTransposedConvolution &) = delete;
+    /** Default move constructor */
+    ClTransposedConvolution(ClTransposedConvolution &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    ClTransposedConvolution &operator=(const ClTransposedConvolution &) = delete;
+    /** Default move assignment operator */
+    ClTransposedConvolution &operator=(ClTransposedConvolution &&) = default;
+
+    /** Set the input, weights, biases and output tensors.
+     *
+     * @note: Only NHWC data layout is supported
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  input           Input tensor info with dimensions [IFM, width, height, batch]
+     *                             Data types supported: F16/F32.
+     * @param[in]  weights         Weight tensor info with dimensions [IFM, width, height, OFM].
+     *                             Data type supported: Same as @p input
+     * @param[in]  biases          (Optional) Biases tensor info. Biases are 1D tensor with dimension [OFM].
+     *                             Data type supported: Should match @p input data type
+     * @param[out] output          Output tensor info with dimensions [OFM, width, height, batch]
+     *                             The 1st dimension must be equal to the 4th dimension of the @p weights tensor.
+     *                             Data types supported: Same as @p input.
+     * @param[in]  deconv_info     Contains padding and stride information described in @ref PadStrideInfo.
+     *
+     */
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *input, const ITensorInfo *weights,
+                   const ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &deconv_info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClTransposedConvolution::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases,
+                           const ITensorInfo *output, const PadStrideInfo &deconv_info);
+
+    // Inherited method overridden
+    void run(ITensorPack &tensors) override;
+
+private:
+    std::unique_ptr<IClKernel> _transposed_conv_kernel{ nullptr };
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_TRANSPOSED_CONVOLUTION_H */
\ No newline at end of file