diff options
author | Manuel Bottini <manuel.bottini@arm.com> | 2021-05-18 18:41:56 +0100 |
---|---|---|
committer | Manuel Bottini <manuel.bottini@arm.com> | 2021-06-15 16:33:52 +0000 |
commit | c6f4ec377027b21a67061efd21b65609079f98f9 (patch) | |
tree | d864f2092fff63790944fea7c8de5be46293bb43 /src | |
parent | 94f799e8f6f605333d40472860fb472e8ba6d83d (diff) | |
download | ComputeLibrary-c6f4ec377027b21a67061efd21b65609079f98f9.tar.gz |
Port CLWinogradConvolutionLayer with ClWinogradConv2d
Port CLWinogradInputTransformKernel
Port CLWinogradFilterTransformKernel
Port CLWinogradOutputTransformKernel
Resolves: COMPMID-4504
Change-Id: I3177dda0b9c2f56b36cb317027e94abe8d47229e
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5680
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src')
14 files changed, 847 insertions, 729 deletions
diff --git a/src/core/CL/CLKernels.h b/src/core/CL/CLKernels.h index c59eebacbb..5dc95dae27 100644 --- a/src/core/CL/CLKernels.h +++ b/src/core/CL/CLKernels.h @@ -80,8 +80,5 @@ #include "src/core/CL/kernels/CLStridedSliceKernel.h" #include "src/core/CL/kernels/CLTileKernel.h" #include "src/core/CL/kernels/CLWeightsReshapeKernel.h" -#include "src/core/CL/kernels/CLWinogradFilterTransformKernel.h" -#include "src/core/CL/kernels/CLWinogradInputTransformKernel.h" -#include "src/core/CL/kernels/CLWinogradOutputTransformKernel.h" #endif /* ARM_COMPUTE_CLKERNELS_H */ diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.h b/src/core/CL/kernels/CLWinogradFilterTransformKernel.h deleted file mode 100644 index d22fedebcd..0000000000 --- a/src/core/CL/kernels/CLWinogradFilterTransformKernel.h +++ /dev/null @@ -1,115 +0,0 @@ -/* - * Copyright (c) 2018-2020 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 "src/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. - * - * @note Winograd filter transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd filter transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] input Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout) or [IFM, kernel_x, kernel_y, OFM] (NHWC data layout). Data types supported: F16/F32. - * @param[out] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_filter_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - */ - void configure(const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info); - /** Set the input and output tensor. - * - * @note Winograd filter transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd filter transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] compile_context The compile context to be used. - * @param[in] input Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout) or [IFM, kernel_x, kernel_y, OFM] (NHWC data layout). Data types supported: F16/F32. - * @param[out] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_filter_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info); - /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradFilterTransformKernel - * - * @note Winograd filter transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd filter transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] input Source tensor. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout) or [IFM, kernel_x, kernel_y, OFM] (NHWC data layout). Data types supported: F16/F32. - * @param[out] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_filter_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info); - - // 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/src/core/CL/kernels/CLWinogradInputTransformKernel.h b/src/core/CL/kernels/CLWinogradInputTransformKernel.h deleted file mode 100644 index 25301877e6..0000000000 --- a/src/core/CL/kernels/CLWinogradInputTransformKernel.h +++ /dev/null @@ -1,121 +0,0 @@ -/* - * Copyright (c) 2018-2020 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 "src/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. - * - * @note Winograd input transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd input transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] input The input tensor to transform. Data types supported: F16/F32 - * @param[in] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_input_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo. - */ - void configure(const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info); - /** Set the input and output of the kernel. - * - * @note Winograd input transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd input transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] compile_context The compile context to be used. - * @param[in] input The input tensor to transform. Data types supported: F16/F32 - * @param[in] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_input_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo. - */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info); - /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradInputTransformKernel - * - * @note Winograd input transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd input transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] input The input tensor to transform. Data types supported: F16/F32 - * @param[in] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_input_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo. - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info); - - // Inherited methods overridden: - void run(const Window &window, cl::CommandQueue &queue) override; - BorderSize border_size() const override; - -private: - using WinogradKey = std::pair<std::pair<int, int>, std::pair<int, int>>; - - BorderSize _border_size; - const ICLTensor *_input; - ICLTensor *_output; - DataLayout _data_layout; - int _num_tiles_x; - int _num_tiles_y; - unsigned int _step_z; -}; -} // arm_compute -#endif /*ARM_COMPUTE_CLWINOGRADINPUTTRANSFORMKERNEL_H */ diff --git a/src/core/CL/kernels/CLWinogradOutputTransformKernel.h b/src/core/CL/kernels/CLWinogradOutputTransformKernel.h deleted file mode 100644 index 632a5629d9..0000000000 --- a/src/core/CL/kernels/CLWinogradOutputTransformKernel.h +++ /dev/null @@ -1,127 +0,0 @@ -/* - * Copyright (c) 2018-2020 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_CLWINOGRADOUTPUTTRANSFORMKERNEL_H -#define ARM_COMPUTE_CLWINOGRADOUTPUTTRANSFORMKERNEL_H - -#include "src/core/CL/ICLKernel.h" - -namespace arm_compute -{ -class ICLTensor; - -/** Interface for the Winograd output transform kernel. */ -class CLWinogradOutputTransformKernel : public ICLKernel -{ -public: - /** Default constructor */ - CLWinogradOutputTransformKernel(); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - CLWinogradOutputTransformKernel(const CLWinogradOutputTransformKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - CLWinogradOutputTransformKernel &operator=(const CLWinogradOutputTransformKernel &) = delete; - /** Allow instances of this class to be moved */ - CLWinogradOutputTransformKernel(CLWinogradOutputTransformKernel &&) = default; - /** Allow instances of this class to be moved */ - CLWinogradOutputTransformKernel &operator=(CLWinogradOutputTransformKernel &&) = default; - /** Default destructor */ - ~CLWinogradOutputTransformKernel() = default; - /** Set the input and output tensor. - * - * @note Winograd output transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd output transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F16/F32. - * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input - * @param[out] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. - */ - void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info = ActivationLayerInfo()); - /** Set the input and output tensor. - * - * @note Winograd output transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd output transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] compile_context The compile context to be used. - * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F16/F32. - * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input - * @param[out] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - * @param[in] act_info (Optional) Activation layer information in case of a fused activation. - */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info, - const ActivationLayerInfo &act_info = ActivationLayerInfo()); - - /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradOutputTransformKernel - * - * @note Winograd output transform supports the following configurations for NCWH data layout - * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), - * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * @note Winograd output transform supports the following configurations for NHWC data layout - * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), - * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) - * - * Strides: only unit strides - * - * @param[in] input Source tensor with shape [C, N, K, batches]. Data types supported: F16/F32. - * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input - * @param[out] output The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - * @param[in] act_info (Optional) Activation layer information in case of a fused activation @ref ActivationLayerInfo. Only RELU, BOUNDED_RELU, LU_BOUNDED_RELU, LEAKY_RELU and SOFT_RELU supported. - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info = ActivationLayerInfo()); - - // Inherited methods overridden: - void run(const Window &window, cl::CommandQueue &queue) override; - -private: - using WinogradKey = std::pair<std::pair<int, int>, std::pair<int, int>>; - - const ICLTensor *_input; - const ICLTensor *_bias; - ICLTensor *_output; - bool _is_nhwc; -}; -} // namespace arm_compute -#endif /*ARM_COMPUTE_CLWINOGRADOUTPUTTRANSFORMKERNEL_H */ diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp index 138f4cf947..381b4bcae9 100644 --- a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp +++ b/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.cpp @@ -21,7 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "src/core/CL/kernels/CLWinogradFilterTransformKernel.h" +#include "src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h" #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" @@ -36,13 +36,17 @@ #include "src/core/CL/CLValidate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" - +#include "support/Cast.h" #include "support/StringSupport.h" using namespace arm_compute::misc::shape_calculator; namespace arm_compute { +namespace opencl +{ +namespace kernels +{ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) @@ -87,69 +91,61 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen } } // namespace -CLWinogradFilterTransformKernel::CLWinogradFilterTransformKernel() - : _input(nullptr), _output(nullptr) -{ -} - -void CLWinogradFilterTransformKernel::configure(const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info) +void ClWinogradFilterTransformKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info) { - configure(CLKernelLibrary::get().get_compile_context(), input, output, winograd_info); -} - -void CLWinogradFilterTransformKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); // Output auto initialization if not yet initialized - auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*input->info(), winograd_info))); + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*src, winograd_info))); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), winograd_info)); - auto padding_info = get_padding_info({ input, output }); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, winograd_info)); + auto padding_info = get_padding_info({ src, dst }); // Set build options CLBuildOptions build_opts; - build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(input->info()->dimension(2))); - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DSRC_DIM_Z=" + support::cpp11::to_string(src->dimension(2))); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_FILTER_TRANSFORM_HORIZONTAL"); build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_FILTER_TRANSFORM_VERTICAL"); const Size2D kernel_size = winograd_info.kernel_size; const Size2D output_tile_size = winograd_info.output_tile_size; // Create kernel - std::string kernel_name = "winograd_filter_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string() + "_" + lower_string(string_from_data_layout(input->info()->data_layout())); + std::string kernel_name = "winograd_filter_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string() + "_" + lower_string(string_from_data_layout(src->data_layout())); _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - _input = input; - _output = output; - // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), output->info()); + auto win_config = validate_and_configure_window(src, dst); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - ICLKernel::configure_internal(win_config.second); + IClKernel::configure_internal(win_config.second); ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } -Status CLWinogradFilterTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) +Status ClWinogradFilterTransformKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_info) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, winograd_info)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, winograd_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get()).first); return Status{}; } -void CLWinogradFilterTransformKernel::run(const Window &window, cl::CommandQueue &queue) +void ClWinogradFilterTransformKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IClKernel::window(), window); + + auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); + auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); // Setup output window Window window_out; - window_out.use_tensor_dimensions(_output->info()->tensor_shape(), 0); + window_out.use_tensor_dimensions(dst->info()->tensor_shape(), 0); unsigned int idx = 0; - add_4D_tensor_argument(idx, _input, window); - add_3D_tensor_argument(idx, _output, window_out); + add_4D_tensor_argument(idx, src, window); + add_3D_tensor_argument(idx, dst, window_out); enqueue(queue, *this, window, lws_hint()); } +} // namespace kernels +} // namespace opencl } // namespace arm_compute
\ No newline at end of file diff --git a/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h b/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h new file mode 100644 index 0000000000..2bc2ceb36e --- /dev/null +++ b/src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h @@ -0,0 +1,78 @@ +/* + * Copyright (c) 2018-2021 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_WINOGRAD_FILTER_TRANSFORM_KERNEL_H +#define ARM_COMPUTE_CL_WINOGRAD_FILTER_TRANSFORM_KERNEL_H + +#include "arm_compute/core/KernelDescriptors.h" +#include "src/core/common/Macros.h" +#include "src/core/gpu/cl/ClCompileContext.h" +#include "src/core/gpu/cl/IClKernel.h" + +namespace arm_compute +{ +namespace opencl +{ +namespace kernels +{ +/** Interface for the Winograd filter transform kernel. */ +class ClWinogradFilterTransformKernel : public IClKernel +{ +public: + /** Default constructor */ + ClWinogradFilterTransformKernel() = default; + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWinogradFilterTransformKernel); + /** Set the input and output tensor. + * + * @note Winograd filter transform supports the following configurations for NCWH data layout + * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), + * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * + * @note Winograd filter transform supports the following configurations for NHWC data layout + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * + * Strides: only unit strides + * + * @param[in] compile_context The compile context to be used. + * @param[in] src Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout) or [IFM, kernel_x, kernel_y, OFM] (NHWC data layout). Data types supported: F16/F32. + * @param[out] dst The output tensor info. The shape for this tensor can be calculated using the utility function @p compute_winograd_filter_transform_shape. Data types supported: Same as @p input + * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo + */ + void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info); + /** Static function to check if given info will lead to a valid configuration + * + * Similar to ClWinogradFilterTransformKernel::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_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_WINOGRAD_FILTER_TRANSFORM_KERNEL_H */ diff --git a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp b/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp index 3399f47d5f..17f0eb9e2c 100644 --- a/src/core/CL/kernels/CLWinogradInputTransformKernel.cpp +++ b/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp @@ -21,7 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "src/core/CL/kernels/CLWinogradInputTransformKernel.h" +#include "src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h" #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" @@ -36,10 +36,15 @@ #include "src/core/CL/CLValidate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" +#include "support/Cast.h" #include "support/StringSupport.h" -using namespace arm_compute; - +namespace arm_compute +{ +namespace opencl +{ +namespace kernels +{ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) @@ -95,69 +100,62 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen } } // namespace -CLWinogradInputTransformKernel::CLWinogradInputTransformKernel() - : _border_size(0), _input(nullptr), _output(nullptr), _data_layout(DataLayout::UNKNOWN), _num_tiles_x(0), _num_tiles_y(0), _step_z(1) +ClWinogradInputTransformKernel::ClWinogradInputTransformKernel() + : _border_size(0), _data_layout(DataLayout::UNKNOWN), _num_tiles_x(0), _num_tiles_y(0), _step_z(1) { } -BorderSize CLWinogradInputTransformKernel::border_size() const +BorderSize ClWinogradInputTransformKernel::border_size() const { return _border_size; } -void CLWinogradInputTransformKernel::configure(const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info) -{ - configure(CLKernelLibrary::get().get_compile_context(), input, output, winograd_info); -} - -void CLWinogradInputTransformKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info) +void ClWinogradInputTransformKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), winograd_info)); + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, winograd_info)); - auto padding_info = get_padding_info({ input, output }); + auto padding_info = get_padding_info({ src, dst }); const PadStrideInfo conv_info = winograd_info.convolution_info; const Size2D output_tile_size = winograd_info.output_tile_size; const Size2D kernel_size = winograd_info.kernel_size; - _data_layout = input->info()->data_layout(); + _data_layout = src->data_layout(); const size_t idx_w = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); const size_t idx_h = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); // Compute the number of output tiles along the x and y direction of size "output_tile_size" - const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(input->info()->dimension(idx_w), input->info()->dimension(idx_h)), + const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(src->dimension(idx_w), src->dimension(idx_h)), kernel_size, output_tile_size, conv_info); - _input = input; - _output = output; _num_tiles_x = num_tiles.width; _num_tiles_y = num_tiles.height; - const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input->info(), winograd_info); + const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*src, winograd_info); // Output auto initialization if not yet initialized - auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape)); + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(output_shape)); - ARM_COMPUTE_ERROR_ON(_num_tiles_x * _num_tiles_y != static_cast<int>(output->info()->dimension(1))); - const size_t total_batches = input->info()->tensor_shape().total_size_upper(3); + ARM_COMPUTE_ERROR_ON(_num_tiles_x * _num_tiles_y != static_cast<int>(dst->dimension(1))); + const size_t total_batches = src->tensor_shape().total_size_upper(3); CLBuildOptions build_opts; if(_data_layout == DataLayout::NHWC) { build_opts.add_option("-DNHWC"); - build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_w))); - build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(idx_h))); + build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(idx_w))); + build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_h))); build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(_num_tiles_x)); build_opts.add_option("-DNUM_TILES_Y=" + support::cpp11::to_string(_num_tiles_y)); 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())); build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width)); build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height)); - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL"); build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_INPUT_TRANSFORM_VERTICAL"); } @@ -168,10 +166,10 @@ void CLWinogradInputTransformKernel::configure(const CLCompileContext &compile_c build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top())); build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width)); build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height)); - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL"); build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_INPUT_TRANSFORM_VERTICAL"); - build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(_input->info()->dimension(2))); + build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(2))); } // Create kernel @@ -183,7 +181,7 @@ void CLWinogradInputTransformKernel::configure(const CLCompileContext &compile_c // Check optimized kernel if output_dims == 2x2 if((tile_max_dim == 2) && (_data_layout == DataLayout::NCHW)) { - _step_z = (_input->info()->dimension(2) % 2) != 0 ? 1 : 2; + _step_z = (src->dimension(2) % 2) != 0 ? 1 : 2; } // Append stepz and data layout @@ -194,20 +192,20 @@ void CLWinogradInputTransformKernel::configure(const CLCompileContext &compile_c _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); // Create window and update padding - auto win_config = validate_and_configure_window(input->info(), output->info(), winograd_info); + auto win_config = validate_and_configure_window(src, dst, winograd_info); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - ICLKernel::configure_internal(win_config.second, cl::NDRange(1, 1, 8)); + IClKernel::configure_internal(win_config.second, cl::NDRange(1, 1, 8)); - _border_size = BorderSize(_input->info()->padding()); + _border_size = BorderSize(src->padding()); - ARM_COMPUTE_ERROR_ON((input->info()->data_layout() == DataLayout::NHWC) && has_padding_changed(padding_info)); + ARM_COMPUTE_ERROR_ON((src->data_layout() == DataLayout::NHWC) && has_padding_changed(padding_info)); _config_id = kernel_name; - _config_id += support::cpp11::to_string(input->info()->dimension(0)); + _config_id += support::cpp11::to_string(src->dimension(0)); _config_id += "_"; - _config_id += support::cpp11::to_string(input->info()->dimension(1)); + _config_id += support::cpp11::to_string(src->dimension(1)); _config_id += "_"; - _config_id += support::cpp11::to_string(input->info()->dimension(2)); + _config_id += support::cpp11::to_string(src->dimension(2)); _config_id += "_"; _config_id += support::cpp11::to_string(conv_info.pad_left()); _config_id += "_"; @@ -216,27 +214,29 @@ void CLWinogradInputTransformKernel::configure(const CLCompileContext &compile_c _config_id += lower_string(string_from_data_layout(_data_layout)); } -Status CLWinogradInputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) +Status ClWinogradInputTransformKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_info) { - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, winograd_info)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), winograd_info).first); - + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, winograd_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), winograd_info).first); return Status{}; } -void CLWinogradInputTransformKernel::run(const Window &window, cl::CommandQueue &queue) +void ClWinogradInputTransformKernel::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); + auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); + auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); + const size_t idx_w = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); const size_t idx_h = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); const size_t idx_c = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); const size_t total_batches = window.shape().total_size_upper(3); // Collapse window - Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); + Window window_collapsed = window.collapse_if_possible(IClKernel::window(), Window::DimZ); if(_data_layout == DataLayout::NHWC) { @@ -245,8 +245,8 @@ void CLWinogradInputTransformKernel::run(const Window &window, cl::CommandQueue slice.set(2, Window::Dimension(0, total_batches, 1)); unsigned int idx = 0; - add_4D_tensor_argument(idx, _input, slice); - add_4D_tensor_argument(idx, _output, slice); + add_4D_tensor_argument(idx, src, slice); + add_4D_tensor_argument(idx, dst, slice); enqueue(queue, *this, slice, lws_hint()); } else @@ -259,17 +259,20 @@ void CLWinogradInputTransformKernel::run(const Window &window, cl::CommandQueue slice.set(idx_c, Window::Dimension(slice[idx_c].start(), slice[idx_c].end(), _step_z)); unsigned int idx = 2 * num_arguments_per_3D_tensor(); - _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input->info()->strides_in_bytes()[3])); - _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[3])); + _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src->info()->strides_in_bytes()[3])); + _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[3])); do { unsigned int idx = 0; - add_3D_tensor_argument(idx, _input, slice); - add_3D_tensor_argument(idx, _output, slice); + add_3D_tensor_argument(idx, src, slice); + add_3D_tensor_argument(idx, dst, slice); enqueue(queue, *this, slice, lws_hint()); } while(window_collapsed.slide_window_slice_3D(slice)); } } +} // namespace kernels +} // namespace opencl +} // namespace arm_compute
\ No newline at end of file diff --git a/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h b/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h new file mode 100644 index 0000000000..76b45279a4 --- /dev/null +++ b/src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h @@ -0,0 +1,88 @@ +/* + * Copyright (c) 2018-2021 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_WINOGRAD_INPUT_TRANSFORM_KERNEL_H +#define ARM_COMPUTE_CL_WINOGRAD_INPUT_TRANSFORM_KERNEL_H + +#include "arm_compute/core/KernelDescriptors.h" +#include "src/core/common/Macros.h" +#include "src/core/gpu/cl/ClCompileContext.h" +#include "src/core/gpu/cl/IClKernel.h" + +namespace arm_compute +{ +namespace opencl +{ +namespace kernels +{ +/** OpenCL kernel to perform Winograd input transform.*/ +class ClWinogradInputTransformKernel : public IClKernel +{ +public: + /** Default constructor */ + ClWinogradInputTransformKernel(); + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWinogradInputTransformKernel); + /** Set the input and output of the kernel. + * + * @note Winograd input transform supports the following configurations for NCWH data layout + * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), + * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * + * @note Winograd input transform supports the following configurations for NHWC data layout + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * + * Strides: only unit strides + * + * @param[in] compile_context The compile context to be used. + * @param[in] src The input tensor info to transform. Data types supported: F16/F32 + * @param[in] dst The output tensor info. The shape for this tensor can be calculated using the utility function @p compute_winograd_input_transform_shape. Data types supported: Same as @p input + * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo. + */ + void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info); + /** Static function to check if given info will lead to a valid configuration + * + * Similar to ClWinogradInputTransformKernel::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_info); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; + BorderSize border_size() const override; + +private: + using WinogradKey = std::pair<std::pair<int, int>, std::pair<int, int>>; + + BorderSize _border_size; + DataLayout _data_layout; + int _num_tiles_x; + int _num_tiles_y; + unsigned int _step_z; +}; +} // namespace kernels +} // namespace opencl +} // namespace arm_compute +#endif /*ARM_COMPUTE_CL_WINOGRAD_INPUT_TRANSFORM_KERNEL_H */ diff --git a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp b/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp index 965bf9df77..a6c05420ed 100644 --- a/src/core/CL/kernels/CLWinogradOutputTransformKernel.cpp +++ b/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.cpp @@ -21,7 +21,7 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "src/core/CL/kernels/CLWinogradOutputTransformKernel.h" +#include "src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h" #include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" @@ -38,15 +38,19 @@ #include "src/core/CL/CLValidate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" - +#include "support/Cast.h" #include "support/StringSupport.h" #include <cmath> -namespace arm_compute -{ using namespace arm_compute::misc::shape_calculator; +namespace arm_compute +{ +namespace opencl +{ +namespace kernels +{ namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) @@ -118,36 +122,23 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen } } // namespace -CLWinogradOutputTransformKernel::CLWinogradOutputTransformKernel() - : _input(nullptr), _bias(nullptr), _output(nullptr), _is_nhwc(false) -{ -} - -void CLWinogradOutputTransformKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - configure(CLKernelLibrary::get().get_compile_context(), input, bias, output, winograd_info, act_info); -} - -void CLWinogradOutputTransformKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info, +void ClWinogradOutputTransformKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input->info(), winograd_info))); + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*src, winograd_info))); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info(), winograd_info, act_info)); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, winograd_info, act_info)); // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info(), winograd_info.output_tile_size); + auto win_config = validate_and_configure_window(src, bias, dst, winograd_info.output_tile_size); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - ICLKernel::configure_internal(win_config.second); + IClKernel::configure_internal(win_config.second); - auto padding_info = get_padding_info({ input, bias, output }); + auto padding_info = get_padding_info({ src, bias, dst }); - _input = input; - _bias = bias; - _output = output; _is_nhwc = winograd_info.output_data_layout == DataLayout::NHWC; // Compute num_tiles_x @@ -163,7 +154,7 @@ void CLWinogradOutputTransformKernel::configure(const CLCompileContext &compile_ kernel_size, output_tile_size, conv_info); - const size_t total_batches = output->info()->tensor_shape().total_size_upper(3); + const size_t total_batches = dst->tensor_shape().total_size_upper(3); // Set build options CLBuildOptions build_opts; @@ -180,17 +171,17 @@ void CLWinogradOutputTransformKernel::configure(const CLCompileContext &compile_ build_opts.add_option("-DVEC_SIZE=4"); } - build_opts.add_option_if(_bias != nullptr, std::string("-DHAS_BIAS")); + build_opts.add_option_if(bias != nullptr, std::string("-DHAS_BIAS")); build_opts.add_option("-cl-fast-relaxed-math"); build_opts.add_option("-DN0=" + support::cpp11::to_string(win_config.second.x().step())); build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width)); build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width)); build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height)); - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); - build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(_input->info()->dimension(1))); - build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(idx_width))); - build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(_output->info()->dimension(idx_height))); - build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(_input->info()->dimension(2))); + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); + build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(1))); + build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst->dimension(idx_width))); + build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(idx_height))); + build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(2))); build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL"); build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_OUTPUT_TRANSFORM_VERTICAL"); @@ -201,36 +192,39 @@ void CLWinogradOutputTransformKernel::configure(const CLCompileContext &compile_ // Set config_id for enabling LWS tuning _config_id = kernel_name; _config_id += "_"; - _config_id += lower_string(string_from_data_type(input->info()->data_type())); + _config_id += lower_string(string_from_data_type(src->data_type())); _config_id += "_"; - _config_id += support::cpp11::to_string(input->info()->dimension(0)); + _config_id += support::cpp11::to_string(src->dimension(0)); _config_id += "_"; - _config_id += support::cpp11::to_string(input->info()->dimension(1)); + _config_id += support::cpp11::to_string(src->dimension(1)); _config_id += "_"; - _config_id += support::cpp11::to_string(output->info()->dimension(0)); + _config_id += support::cpp11::to_string(dst->dimension(0)); _config_id += "_"; - _config_id += support::cpp11::to_string(output->info()->dimension(1)); + _config_id += support::cpp11::to_string(dst->dimension(1)); _config_id += "_"; _config_id += lower_string(string_from_data_layout(winograd_info.output_data_layout)); ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info) && _is_nhwc); } -Status CLWinogradOutputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +Status ClWinogradOutputTransformKernel::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, (bias != nullptr ? bias->clone().get() : nullptr), output, winograd_info, act_info)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), output->clone().get(), winograd_info.output_tile_size).first); - + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, (bias != nullptr ? bias->clone().get() : nullptr), dst, winograd_info, act_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), dst->clone().get(), winograd_info.output_tile_size).first); return Status{}; } -void CLWinogradOutputTransformKernel::run(const Window &window, cl::CommandQueue &queue) +void ClWinogradOutputTransformKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) { ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IClKernel::window(), window); + + auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); + auto bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); + auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); // Collapse window - Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); + Window window_collapsed = window.collapse_if_possible(IClKernel::window(), Window::DimZ); // Get initial windows Window slice = window_collapsed.first_slice_window_4D(); @@ -241,27 +235,29 @@ void CLWinogradOutputTransformKernel::run(const Window &window, cl::CommandQueue slice_out.set(Window::DimX, Window::Dimension(0, 0, 0)); slice_out.set(Window::DimY, Window::Dimension(0, 0, 0)); - if(_bias != nullptr) + if(bias != nullptr) { unsigned int idx1 = 2 * num_arguments_per_4D_tensor(); Window slice_biases; - slice_biases.use_tensor_dimensions(_bias->info()->tensor_shape()); - add_1D_tensor_argument(idx1, _bias, slice_biases); + slice_biases.use_tensor_dimensions(bias->info()->tensor_shape()); + add_1D_tensor_argument(idx1, bias, slice_biases); } if(_is_nhwc) { - unsigned int idx2 = 2 * num_arguments_per_4D_tensor() + ((_bias != nullptr) ? num_arguments_per_1D_tensor() : 0); - _kernel.setArg(idx2, static_cast<int>(_output->info()->total_size() - _output->info()->strides_in_bytes().y())); + unsigned int idx2 = 2 * num_arguments_per_4D_tensor() + ((bias != nullptr) ? num_arguments_per_1D_tensor() : 0); + _kernel.setArg(idx2, static_cast<int>(dst->info()->total_size() - dst->info()->strides_in_bytes().y())); } do { unsigned int idx = 0; - add_4D_tensor_argument(idx, _input, slice); - add_4D_tensor_argument(idx, _output, slice_out); + add_4D_tensor_argument(idx, src, slice); + add_4D_tensor_argument(idx, dst, slice_out); enqueue(queue, *this, slice, lws_hint()); } while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_out)); } +} // namespace kernels +} // namespace opencl } // namespace arm_compute diff --git a/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h b/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h new file mode 100644 index 0000000000..48b27e658c --- /dev/null +++ b/src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h @@ -0,0 +1,87 @@ +/* + * Copyright (c) 2018-2021 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_WINOGRAD_OUTPUT_TRANSFORM_KERNEL_H +#define ARM_COMPUTE_CL_WINOGRAD_OUTPUT_TRANSFORM_KERNEL_H + +#include "arm_compute/core/KernelDescriptors.h" +#include "src/core/common/Macros.h" +#include "src/core/gpu/cl/ClCompileContext.h" +#include "src/core/gpu/cl/IClKernel.h" + +namespace arm_compute +{ +namespace opencl +{ +namespace kernels +{ +/** Interface for the Winograd output transform kernel. */ +class ClWinogradOutputTransformKernel : public IClKernel +{ +public: + /** Default constructor */ + ClWinogradOutputTransformKernel() = default; + ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClWinogradOutputTransformKernel); + /** Set the input and output tensor. + * + * @note Winograd output transform supports the following configurations for NCWH data layout + * F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3), + * F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * + * @note Winograd output transform supports the following configurations for NHWC data layout + * F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3), + * F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5) + * + * Strides: only unit strides + * + * @param[in] compile_context The compile context to be used. + * @param[in] src Source tensor info with shape [C, N, K, batches]. Data types supported: F16/F32. + * @param[in] bias Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p src + * @param[out] dst The output tensor info. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p src + * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + */ + void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, const WinogradInfo &winograd_info, + const ActivationLayerInfo &act_info = ActivationLayerInfo()); + + /** Static function to check if given info will lead to a valid configuration + * + * Similar to ClWinogradOutputTransformKernel::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override; + +private: + using WinogradKey = std::pair<std::pair<int, int>, std::pair<int, int>>; + + bool _is_nhwc{ false }; +}; +} // namespace kernels +} // namespace opencl +} // namespace arm_compute +#endif /*ARM_COMPUTE_CL_WINOGRAD_OUTPUT_TRANSFORM_KERNEL_H */ diff --git a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp index 6b8b00414a..f758c3d0b3 100644 --- a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp @@ -23,79 +23,34 @@ */ #include "arm_compute/runtime/CL/functions/CLWinogradConvolutionLayer.h" +#include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/runtime/CL/CLScheduler.h" -#include "src/core/CL/kernels/CLFillBorderKernel.h" -#include "src/core/CL/kernels/CLWinogradFilterTransformKernel.h" -#include "src/core/CL/kernels/CLWinogradOutputTransformKernel.h" +#include "arm_compute/core/KernelDescriptors.h" +#include "src/core/CL/ICLKernel.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/runtime/gpu/cl/operators/ClWinogradConv2d.h" +#include "support/Cast.h" -using namespace arm_compute; - -namespace +namespace arm_compute { -Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataLayout data_layout) +struct CLWinogradConvolutionLayer::Impl { - Size2D output_tile = Size2D{}; - - const unsigned int kernel_max_dim = std::max(kernel_dims.width, kernel_dims.height); - - // Check if the input spatial dimensions are smaller than 4 - const bool is_input_lt4_nchw = (input_dims.width <= 4 && input_dims.height <= 4) && (data_layout == DataLayout::NCHW); - - if(kernel_max_dim == 3U) - { - if(kernel_dims == Size2D(3U, 3U)) - { - output_tile = is_input_lt4_nchw ? Size2D(2U, 2U) : Size2D(4U, 4U); - } - else if(kernel_dims == Size2D(3U, 1U)) - { - output_tile = is_input_lt4_nchw ? Size2D(2U, 1U) : Size2D(4U, 1U); - } - else - { - output_tile = is_input_lt4_nchw ? Size2D(1U, 2U) : Size2D(1U, 4U); - } - } - else if(kernel_max_dim == 5U) - { - output_tile = Size2D(kernel_dims.width == 1 ? 1U : 4U, - kernel_dims.height == 1 ? 1U : 4U); - } - else if(kernel_max_dim == 7U) - { - output_tile = Size2D(kernel_dims.width == 1 ? 1U : 2U, - kernel_dims.height == 1 ? 1U : 2U); - } - - return output_tile; -} - -bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size) -{ - // Check if we want to configure a Winograd configuration which requires fast math - using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>; - - std::vector<WinogradConfiguration> fast_math_winograd = - { - WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)), - WinogradConfiguration(std::pair<int, int>(2, 2), std::pair<int, int>(7, 7)) - }; - - auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height), - std::pair<int, int>(kernel_size.width, kernel_size.height)); - - return std::find(fast_math_winograd.begin(), fast_math_winograd.end(), p) != fast_math_winograd.end(); -} -} // namespace + const ICLTensor *src{ nullptr }; + const ICLTensor *weights{ nullptr }; + const ICLTensor *biases{ nullptr }; + ICLTensor *dst{ nullptr }; + std::unique_ptr<opencl::ClWinogradConv2d> op{ nullptr }; + ITensorPack run_pack{}; + ITensorPack prep_pack{}; + MemoryGroup memory_group{}; + WorkspaceData<CLTensor> workspace_tensors{}; + bool is_prepared{ false }; +}; CLWinogradConvolutionLayer::CLWinogradConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(memory_manager), _batched_mm(memory_manager), _input_transform(), _filter_transform(std::make_unique<CLWinogradFilterTransformKernel>()), - _output_transform(std::make_unique<CLWinogradOutputTransformKernel>()), _input0(), _input1(), _batched_mm_output(), _original_weights(nullptr), _is_prepared(false) + : _impl(std::make_unique<Impl>()) { + _impl->memory_group = MemoryGroup(memory_manager); } CLWinogradConvolutionLayer::~CLWinogradConvolutionLayer() = default; @@ -110,139 +65,45 @@ void CLWinogradConvolutionLayer::configure(const CLCompileContext &compile_conte const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) { - // Get indices for the width and height - const size_t idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH); - const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT); + _impl->src = input; + _impl->weights = weights; + _impl->biases = biases; + _impl->dst = output; - // Input shape, kernel size and output tile - const Size2D input_dims = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]); - const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]); - const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->info()->data_layout()); + _impl->op = std::make_unique<opencl::ClWinogradConv2d>(); + _impl->op->configure(compile_context, input->info(), weights->info(), (biases != nullptr ? biases->info() : nullptr), output->info(), conv_info, act_info, enable_fast_math); - // Check if the Winograd configuration requires fast math - if(!enable_fast_math) + _impl->run_pack = { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); //disable winograd for fp16 if fast math is false. - ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); - } - const WinogradInfo winograd_info = WinogradInfo(output_tile, - kernel_size, - input_dims, - conv_info, - input->info()->data_layout()); - - _is_prepared = false; - _original_weights = weights; - - // Manage intermediate tensors - _memory_group.manage(&_input0); - _memory_group.manage(&_batched_mm_output); - - // Do not manage _input1 as it contains the weights - - // Configure input transform - _input_transform.configure(compile_context, input, &_input0, winograd_info); - - // Configure filter transform - _filter_transform->configure(compile_context, weights, &_input1, winograd_info); - - // Configure batched matrix multiply - _batched_mm.configure(compile_context, &_input0, &_input1, nullptr, &_batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, 0, false, false, - GEMMLowpOutputStageInfo(), - (input->info()->data_type() == DataType::F16))); - - // Configure output transform - _output_transform->configure(compile_context, &_batched_mm_output, biases, output, winograd_info, act_info); + { TensorType::ACL_SRC_0, _impl->src }, + { TensorType::ACL_SRC_1, _impl->weights }, + { TensorType::ACL_SRC_2, _impl->biases }, + { TensorType::ACL_DST, _impl->dst } + }; - // Allocate temporary tensors - _input0.allocator()->allocate(); - _batched_mm_output.allocator()->allocate(); + _impl->prep_pack = { { TensorType::ACL_SRC_1, _impl->weights } }; + _impl->workspace_tensors = manage_workspace<CLTensor>(_impl->op->workspace(), _impl->memory_group, _impl->run_pack, _impl->prep_pack); } Status CLWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) { - // Get indeces for the width and height - const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); - const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); - - // Input shape, kernel size and output tile - const Size2D input_dims = Size2D(input->tensor_shape()[idx_width], input->tensor_shape()[idx_height]); - const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]); - const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->data_layout()); - - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((conv_info.pad_left() > (kernel_size.x() / 2u)) || (conv_info.pad_right() > (kernel_size.x() / 2u))), "Winograd only supports padding up to half kernel size"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((conv_info.pad_top() > (kernel_size.y() / 2u)) || (conv_info.pad_bottom() > (kernel_size.y() / 2u))), "Winograd only supports padding up to half kernel size"); - - // Check if the Winograd configuration requires fast math - if(!enable_fast_math) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); //disable winograd for fp16 if fast math is false. - ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); - } - - const WinogradInfo winograd_info = WinogradInfo(output_tile, - kernel_size, - input_dims, - conv_info, - input->data_layout()); - - // Validate input transform - const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); - const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape); - ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradInputTransform::validate(input, &input0, winograd_info)); - - // Validate filter transform - const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); - const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); - ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradFilterTransformKernel::validate(weights, &input1, winograd_info)); - - // Validate batched matrix multiply - TensorShape batched_mm_output_shape = input0.tensor_shape(); - batched_mm_output_shape[0] = input1.tensor_shape()[0]; - const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); - ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, 0, false, false, - GEMMLowpOutputStageInfo(), (input->data_type() == DataType::F16)))); - - // Configure output transform - ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradOutputTransformKernel::validate(&batched_mm_output, biases, output, winograd_info, act_info)); - - return Status{}; + return opencl::ClWinogradConv2d::validate(input, weights, biases, output, conv_info, act_info, enable_fast_math); } void CLWinogradConvolutionLayer::run() { + MemoryGroupResourceScope scope_mg(_impl->memory_group); prepare(); - - MemoryGroupResourceScope scope_mg(_memory_group); - - // Run input transform - _input_transform.run(); - - // Run batched matrix multiplication - _batched_mm.run(); - - // Run output transform - CLScheduler::get().enqueue(*_output_transform); + _impl->op->run(_impl->run_pack); } void CLWinogradConvolutionLayer::prepare() { - if(!_is_prepared) + if(!_impl->is_prepared) { - // Run filter transform and mark original weights as unused - _input1.allocator()->allocate(); - CLScheduler::get().enqueue(*_filter_transform, false); - _original_weights->mark_as_unused(); - - // Prepare GEMM and release reshaped weights if marked unused by CLGEMM - _batched_mm.prepare(); - if(!_input1.is_used()) - { - _input1.allocator()->free(); - } - - CLScheduler::get().queue().finish(); - _is_prepared = true; + _impl->op->prepare(_impl->prep_pack); + _impl->is_prepared = true; } } +} // namespace arm_compute
\ No newline at end of file diff --git a/src/runtime/CL/functions/CLWinogradInputTransform.cpp b/src/runtime/CL/functions/CLWinogradInputTransform.cpp deleted file mode 100644 index 6d5a692bc3..0000000000 --- a/src/runtime/CL/functions/CLWinogradInputTransform.cpp +++ /dev/null @@ -1,50 +0,0 @@ -/* - * Copyright (c) 2018-2020 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/Error.h" -#include "src/core/CL/kernels/CLFillBorderKernel.h" -#include "src/core/CL/kernels/CLWinogradInputTransformKernel.h" - -using namespace arm_compute; - -void CLWinogradInputTransform::configure(ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info) -{ - configure(CLKernelLibrary::get().get_compile_context(), input, output, winograd_info); -} - -void CLWinogradInputTransform::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info) -{ - auto k = std::make_unique<CLWinogradInputTransformKernel>(); - k->configure(compile_context, input, output, winograd_info); - _kernel = std::move(k); - _border_handler->configure(compile_context, input, _kernel->border_size(), BorderMode::CONSTANT, PixelValue()); -} - -Status CLWinogradInputTransform::validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradInputTransformKernel::validate(input, output, winograd_info)); - return Status{}; -} diff --git a/src/runtime/gpu/cl/operators/ClWinogradConv2d.cpp b/src/runtime/gpu/cl/operators/ClWinogradConv2d.cpp new file mode 100644 index 0000000000..c8db697778 --- /dev/null +++ b/src/runtime/gpu/cl/operators/ClWinogradConv2d.cpp @@ -0,0 +1,299 @@ +/* + * Copyright (c) 2018-2021 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/runtime/gpu/cl/operators/ClWinogradConv2d.h" + +#include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/experimental/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/runtime/CL/CLScheduler.h" +#include "src/core/CL/kernels/CLFillBorderKernel.h" +#include "src/core/CL/kernels/CLFillBorderKernel.h" +#include "src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h" +#include "src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h" +#include "src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/runtime/gpu/cl/utils/ClAuxTensorHandler.h" +#include "support/Cast.h" + +using namespace arm_compute::experimental; + +namespace arm_compute +{ +namespace opencl +{ +namespace +{ +Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataLayout data_layout) +{ + Size2D output_tile = Size2D{}; + + const unsigned int kernel_max_dim = std::max(kernel_dims.width, kernel_dims.height); + + // Check if the input spatial dimensions are smaller than 4 + const bool is_input_lt4_nchw = (input_dims.width <= 4 && input_dims.height <= 4) && (data_layout == DataLayout::NCHW); + + if(kernel_max_dim == 3U) + { + if(kernel_dims == Size2D(3U, 3U)) + { + output_tile = is_input_lt4_nchw ? Size2D(2U, 2U) : Size2D(4U, 4U); + } + else if(kernel_dims == Size2D(3U, 1U)) + { + output_tile = is_input_lt4_nchw ? Size2D(2U, 1U) : Size2D(4U, 1U); + } + else + { + output_tile = is_input_lt4_nchw ? Size2D(1U, 2U) : Size2D(1U, 4U); + } + } + else if(kernel_max_dim == 5U) + { + output_tile = Size2D(kernel_dims.width == 1 ? 1U : 4U, + kernel_dims.height == 1 ? 1U : 4U); + } + else if(kernel_max_dim == 7U) + { + output_tile = Size2D(kernel_dims.width == 1 ? 1U : 2U, + kernel_dims.height == 1 ? 1U : 2U); + } + + return output_tile; +} + +bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size) +{ + // Check if we want to configure a Winograd configuration which requires fast math + using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>; + + std::vector<WinogradConfiguration> fast_math_winograd = + { + WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)), + WinogradConfiguration(std::pair<int, int>(2, 2), std::pair<int, int>(7, 7)) + }; + + auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height), + std::pair<int, int>(kernel_size.width, kernel_size.height)); + + return std::find(fast_math_winograd.begin(), fast_math_winograd.end(), p) != fast_math_winograd.end(); +} + +Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, + const ActivationLayerInfo &act_info, bool enable_fast_math) +{ + // Get indeces for the width and height + const size_t idx_width = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::WIDTH); + const size_t idx_height = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::HEIGHT); + + // Input shape, kernel size and output tile + const Size2D input_dims = Size2D(src->tensor_shape()[idx_width], src->tensor_shape()[idx_height]); + const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]); + const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, src->data_layout()); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((conv_info.pad_left() > (kernel_size.x() / 2u)) || (conv_info.pad_right() > (kernel_size.x() / 2u))), "Winograd only supports padding up to half kernel size"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((conv_info.pad_top() > (kernel_size.y() / 2u)) || (conv_info.pad_bottom() > (kernel_size.y() / 2u))), "Winograd only supports padding up to half kernel size"); + + // Check if the Winograd configuration requires fast math + if(!enable_fast_math) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); //disable winograd for fp16 if fast math is false. + ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); + } + + const WinogradInfo winograd_info = WinogradInfo(output_tile, + kernel_size, + input_dims, + conv_info, + src->data_layout()); + + // Validate input transform + const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*src, winograd_info); + const TensorInfo input0 = src->clone()->set_tensor_shape(input0_shape); + ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWinogradInputTransformKernel::validate(src, &input0, winograd_info)); + + // Validate filter transform + const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); + const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); + ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWinogradFilterTransformKernel::validate(weights, &input1, winograd_info)); + + // Validate batched matrix multiply + TensorShape batched_mm_output_shape = input0.tensor_shape(); + batched_mm_output_shape[0] = input1.tensor_shape()[0]; + const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); + ARM_COMPUTE_RETURN_ON_ERROR(ClGemm::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, 0, false, false, + GEMMLowpOutputStageInfo(), (src->data_type() == DataType::F16)))); + + // Configure output transform + ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWinogradOutputTransformKernel::validate(&batched_mm_output, biases, dst, winograd_info, act_info)); + return Status{}; +} + +} // namespace + +ClWinogradConv2d::ClWinogradConv2d() + : _batched_mm(), + _input_transform(std::make_unique<kernels::ClWinogradInputTransformKernel>()), + _filter_transform(std::make_unique<kernels::ClWinogradFilterTransformKernel>()), + _output_transform(std::make_unique<kernels::ClWinogradOutputTransformKernel>()), + _border_handler(), + _input0(), + _input1(), + _batched_mm_output(), + _is_prepared(false), + _aux_mem() +{ +} + +ClWinogradConv2d::~ClWinogradConv2d() = default; + +void ClWinogradConv2d::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, + const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) +{ + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, weights, biases, dst, conv_info, act_info, enable_fast_math)); + // Get indices for the width and height + const size_t idx_width = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::WIDTH); + const size_t idx_height = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::HEIGHT); + + // Input shape, kernel size and output tile + const Size2D input_dims = Size2D(src->tensor_shape()[idx_width], src->tensor_shape()[idx_height]); + const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]); + const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, src->data_layout()); + + // Check if the Winograd configuration requires fast math + if(!enable_fast_math) + { + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); //disable winograd for fp16 if fast math is false. + ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); + } + const WinogradInfo winograd_info = WinogradInfo(output_tile, + kernel_size, + input_dims, + conv_info, + src->data_layout()); + + _is_prepared = false; + + // Configure input transform + _input_transform->configure(compile_context, src, &_input0, winograd_info); + _border_handler.configure(compile_context, src, _input_transform->border_size(), BorderMode::CONSTANT, PixelValue()); + + // Configure filter transform + _filter_transform->configure(compile_context, weights, &_input1, winograd_info); + + // Configure batched matrix multiply + _batched_mm.configure(compile_context, &_input0, &_input1, nullptr, &_batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, 0, + false, false, + GEMMLowpOutputStageInfo(), + (src->data_type() == DataType::F16))); + + // Configure output transform + _output_transform->configure(compile_context, &_batched_mm_output, biases, dst, winograd_info, act_info); + + _aux_mem = _batched_mm.workspace(); + _aux_mem.push_back(MemoryInfo(offset_int_vec(2), MemoryLifetime::Temporary, _input0.total_size())); + _aux_mem.push_back(MemoryInfo(offset_int_vec(3), MemoryLifetime::Persistent, _input1.total_size())); + _aux_mem.push_back(MemoryInfo(offset_int_vec(4), MemoryLifetime::Temporary, _batched_mm_output.total_size())); +} + +Status ClWinogradConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, + const ActivationLayerInfo &act_info, bool enable_fast_math) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, conv_info, act_info, enable_fast_math)); + return Status{}; +} + +void ClWinogradConv2d::run(ITensorPack &tensors) +{ + prepare(tensors); + + // Run input transform + auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); + 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)); + + CLAuxTensorHandler input0(offset_int_vec(2), _input0, tensors, true); + CLAuxTensorHandler input1(offset_int_vec(3), _input1, tensors, true); + CLAuxTensorHandler batched_mm_output(offset_int_vec(4), _batched_mm_output, tensors, true); + + ITensorPack pack_it + { + { TensorType::ACL_SRC, src }, + { TensorType::ACL_DST, input0.get() }, + }; + CLScheduler::get().enqueue_op(_border_handler, pack_it); + CLScheduler::get().enqueue_op(*_input_transform, pack_it); + + // Run batched matrix multiplication + ITensorPack pack_mm + { + { TensorType::ACL_SRC_0, input0.get() }, + { TensorType::ACL_SRC_1, input1.get() }, + { TensorType::ACL_DST, batched_mm_output.get() }, + }; + _batched_mm.run(pack_mm); + + // Run output transform + ITensorPack pack_ot + { + { TensorType::ACL_SRC_0, batched_mm_output.get() }, + { TensorType::ACL_SRC_1, biases }, + { TensorType::ACL_DST, dst }, + }; + CLScheduler::get().enqueue_op(*_output_transform, pack_ot); +} + +void ClWinogradConv2d::prepare(ITensorPack &tensors) +{ + if(!_is_prepared) + { + auto weights = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); + ICLTensor *in1_aux = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(offset_int_vec(3))); + + CLAuxTensorHandler input1(_input1, *in1_aux); + ITensorPack pack_ft + { + { TensorType::ACL_SRC, weights }, + { TensorType::ACL_DST, input1.get() }, + }; + // Run filter transform and mark original weights as unused + CLScheduler::get().enqueue_op(*_filter_transform, pack_ft, false); + weights->mark_as_unused(); + + tensors.add_tensor(ACL_SRC_1, input1.get()); + // Prepare GEMM and release reshaped weights if marked unused by ClGemm + _batched_mm.prepare(tensors); + + CLScheduler::get().queue().finish(); + _is_prepared = true; + } +} + +experimental::MemoryRequirements ClWinogradConv2d::workspace() const +{ + return _aux_mem; +} +} // namespace opencl +} // namespace arm_compute
\ No newline at end of file diff --git a/src/runtime/gpu/cl/operators/ClWinogradConv2d.h b/src/runtime/gpu/cl/operators/ClWinogradConv2d.h new file mode 100644 index 0000000000..83b31f1c99 --- /dev/null +++ b/src/runtime/gpu/cl/operators/ClWinogradConv2d.h @@ -0,0 +1,126 @@ +/* + * Copyright (c) 2018-2021 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_WINOGRADCONV2D_H +#define ARM_COMPUTE_CL_WINOGRADCONV2D_H + +#include "arm_compute/runtime/CL/CLTensor.h" +#include "src/core/CL/kernels/CLFillBorderKernel.h" +#include "src/core/gpu/cl/ClCompileContext.h" +#include "src/runtime/gpu/cl/IClOperator.h" +#include "src/runtime/gpu/cl/operators/ClGemm.h" + +namespace arm_compute +{ +class CLCompileContext; +class ITensorInfo; +namespace opencl +{ +namespace kernels +{ +class ClWinogradInputTransformKernel; +class ClWinogradFilterTransformKernel; +class ClWinogradOutputTransformKernel; +} // kernels +/** Basic function to execute Winograd-based convolution on OpenCL. This function calls the following OpenCL functions/kernels: + * + * -# @ref kernels::ClWinogradInputTransformKernel + * -# @ref kernels::ClWinogradFilterTransformKernel (only once) + * -# @ref ClGemm + * -# @ref kernels::ClWinogradOutputTransformKernel + * + */ +class ClWinogradConv2d : public IClOperator +{ +public: + /** Default constructor */ + ClWinogradConv2d(); + /** Default destructor */ + ~ClWinogradConv2d(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + ClWinogradConv2d(const ClWinogradConv2d &) = delete; + /** Default move constructor */ + ClWinogradConv2d(ClWinogradConv2d &&) = default; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + ClWinogradConv2d &operator=(const ClWinogradConv2d &) = delete; + /** Default move assignment operator */ + ClWinogradConv2d &operator=(ClWinogradConv2d &&) = default; + /** Set the input and output tensors. + * + * Valid data layouts: + * - NHWC + * - NCHW + * + * Valid data type configurations: + * |src0 |src1 |src2 |dst | + * |:--------------|:--------------|:------|:--------------| + * |F16 |F16 |F16 |F16 | + * |F32 |F32 |F32 |F32 | + * + * @note: This function only works with 3x3,3x1,1x3,5x5,5x1,1x5,7x1 and 1x7 kernels along with unit strides for both NCHW and NHWC data layout + * @note Some Winograd configurations (i.e. F(4x4, 5x5)) are supported only with enable_fast_math = true + * + * @param[in] compile_context The compile context to be used. + * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: F16/F32. + * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p src. + * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].Data type supported: Same as @p src + * @param[out] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p src. + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation + * available which may introduce a drop of accuracy as well. Default is false + */ + void configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, const PadStrideInfo &conv_info, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); + /** Static function to check if given info will lead to a valid configuration + * + * Similar to ClWinogradConv2d::configure() + * + * @return a status + */ + static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); + + // Inherited method overridden + void run(ITensorPack &tensors) override; + void prepare(ITensorPack &tensors) override; + experimental::MemoryRequirements workspace() const override; + +private: + ClGemm _batched_mm; + std::unique_ptr<kernels::ClWinogradInputTransformKernel> _input_transform; + std::unique_ptr<kernels::ClWinogradFilterTransformKernel> _filter_transform; + std::unique_ptr<kernels::ClWinogradOutputTransformKernel> _output_transform; + CLFillBorderKernel _border_handler; + TensorInfo _input0; + TensorInfo _input1; + TensorInfo _batched_mm_output; + bool _is_prepared; + experimental::MemoryRequirements _aux_mem{}; +}; +} // namespace opencl +} // namespace arm_compute +#endif /* ARM_COMPUTE_CL_WINOGRADCONV2D_H */ |