diff options
Diffstat (limited to 'src/dynamic_fusion/sketch/gpu/components/cl')
4 files changed, 458 insertions, 0 deletions
diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp new file mode 100644 index 0000000000..e94cfd1581 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp @@ -0,0 +1,152 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "ClComponentDirectConv2d.h" + +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/dynamic_fusion/sketch/OperatorAttributes.h" +#include "src/core/CL/CLValidate.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +using Settings = ClComponentDirectConv2dSettings; + +Settings &Settings::export_to_cl_image(bool cl_image) +{ + _export_to_cl_image = cl_image; + return *this; +} + +bool Settings::export_to_cl_image() const +{ + return _export_to_cl_image; +} + +Settings &Settings::fast_relaxed_math(bool fast_relaxed_math) +{ + _fast_relaxed_math = fast_relaxed_math; + return *this; +} + +bool Settings::fast_relaxed_math() const +{ + return _fast_relaxed_math; +} + +Status ClComponentDirectConv2d::validate( + const Properties &properties, + const ArgumentPack<ITensorInfo> &tensors, + const Attributes &attributes, + const Settings &settings) +{ + ARM_COMPUTE_UNUSED(properties, settings); + const auto src = tensors.get_const_tensor(TensorType::ACL_SRC_0); + const auto wei = tensors.get_const_tensor(TensorType::ACL_SRC_1); + const auto bia = tensors.get_const_tensor(TensorType::ACL_SRC_2); + const auto dst = tensors.get_const_tensor(TensorType::ACL_DST_0); + + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, wei, dst); + + // 1. Check validity + // Matching data type + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, wei); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); + if(bia != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, bia); + } + + // Matching data layout + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, wei); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, dst); + if(bia != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, bia); + } + + // All tensor infos are initialized + ARM_COMPUTE_RETURN_ERROR_ON(src->tensor_shape().total_size() == 0); + ARM_COMPUTE_RETURN_ERROR_ON(wei->tensor_shape().total_size() == 0); + ARM_COMPUTE_RETURN_ERROR_ON(dst->tensor_shape().total_size() == 0); + if(bia != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON(bia->tensor_shape().total_size() == 0); + } + // Device requirements are met + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); + // wei shape is correct + const DataLayout data_layout = src->data_layout(); + const int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(wei->dimension(channel_idx) != src->dimension(channel_idx), "Weights feature map dimension should match the respective src's one"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(wei->num_dimensions() > 4, "Weights can be at most 4 dimensional"); + + // dst shape is correct + PadStrideInfo legacy_pad_stride(attributes.stride().x(), attributes.stride().y(), attributes.pad().left, attributes.pad().right, attributes.pad().top, + attributes.pad().bottom, DimensionRoundingType{}); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), + misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, legacy_pad_stride)); + + // bia shape is correct + if(bia != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MSG(bia->dimension(0) != wei->dimension(3), + "Biases size and number of dst feature maps should match"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(bia->num_dimensions() > 1, + "Biases should be one dimensional"); + } + + // 2. Check support level + // Data type + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); + // Data layout + ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC); + + return Status{}; +} + +ClComponentDirectConv2d::ClComponentDirectConv2d( + ComponentId id, + const Properties &properties, + const ArgumentPack<ITensorInfo> &tensors, + const Attributes &attributes, + const Settings &settings) + : IGpuKernelComponent{ id, properties, tensors }, + _component_writer{ std::make_unique<ClTemplateDirectConv2d>(id, tensors, attributes, settings) } +{ +} +ClComponentDirectConv2d::~ClComponentDirectConv2d() +{ +} +const IGpuTemplateComponentWriter *ClComponentDirectConv2d::template_writer() const +{ + return _component_writer.get(); +} +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h new file mode 100644 index 0000000000..fec22b84a5 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h @@ -0,0 +1,147 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTDIRECTCONV2D +#define SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTDIRECTCONV2D + +#include "arm_compute/core/Error.h" +#include "src/dynamic_fusion/sketch/gpu/components/IGpuKernelComponent.h" +#include <memory> + +namespace arm_compute +{ +/** Forward declaration */ +class ITensorInfo; +namespace experimental +{ +namespace dynamic_fusion +{ +/** Forward declaration */ +template <typename T> +class ArgumentPack; +class Conv2dAttributes; + +/** Component specific settings + */ +class ClComponentDirectConv2dSettings +{ +public: + /** Set export_to_cl_image flag */ + ClComponentDirectConv2dSettings &export_to_cl_image(bool cl_image); + /** Get export_to_cl_image flag */ + bool export_to_cl_image() const; + + /** Set fast_relaxed_math flag */ + ClComponentDirectConv2dSettings &fast_relaxed_math(bool fast_relaxed_math); + /** Get fast_relaxed_math flag */ + bool fast_relaxed_math() const; + +private: + bool _export_to_cl_image{ false }; + bool _fast_relaxed_math{ true }; +}; + +/** Forward declaration */ +class ClTemplateDirectConv2d; + +class ClComponentDirectConv2d final : public IGpuKernelComponent +{ +public: + /** Attributes are a set of backend-agnostic parameters that define what a component does */ + using Attributes = Conv2dAttributes; + /** Settings are a set of backend-specific parameters that influence the implementation of a component */ + using Settings = ClComponentDirectConv2dSettings; + +public: + /** Validate the component + * + * @param[in] properties Component properties + * @param[in,out] tensors Tensor arguments to the component + * @param[in] attributes Component attributes + * @param[in] settings Component settings + * + * @return Status Validation results + * + * Tensor argument names: + * - ACL_SRC_0: Input + * - ACL_SRC_1: Weight + * - ACL_SRC_2: Bias (Optional) + * - ACL_DST_0: Output + * + * Tensor argument constness: + * - ACL_SRC_0: Const + * - ACL_SRC_1: Const + * - ACL_SRC_2: Const + * - ACL_DST_0: Const + * + * Valid data layouts: + * - NHWC + * + * Valid data type configurations: + * |ACL_SRC_0 |ACL_SRC_1 |ACL_SRC_2 |ACL_DST_0 | + * |:--------------|:--------------|:--------------|:--------------| + * |F16 |F16 |F16 |F16 | + * |F32 |F32 |F32 |F32 | + */ + static Status validate( + const Properties &properties, + const ArgumentPack<ITensorInfo> &tensors, + const Attributes &attributes, + const Settings &settings); + + /** Constructor + * + * Similar to @ref ClComponentDirectConv2d::validate() + */ + ClComponentDirectConv2d( + ComponentId id, + const Properties &properties, + const ArgumentPack<ITensorInfo> &tensors, + const Attributes &attributes, + const Settings &settings); + + /** Destructor */ + ~ClComponentDirectConv2d() override; + /** Prevent instances of this class from being copy constructed */ + ClComponentDirectConv2d(const ClComponentDirectConv2d &component) = delete; + /** Prevent instances of this class from being copied */ + ClComponentDirectConv2d &operator=(const ClComponentDirectConv2d &component) = delete; + /** Allow instances of this class to be move constructed */ + ClComponentDirectConv2d(ClComponentDirectConv2d &&component) = default; + /** Allow instances of this class to be moved */ + ClComponentDirectConv2d &operator=(ClComponentDirectConv2d &&component) = default; + /** Get template writer for the component */ + const IGpuTemplateComponentWriter *template_writer() const override; + /** Get component type */ + GpuComponentType type() const override + { + return GpuComponentType::Complex; + } + +private: + std::unique_ptr<ClTemplateDirectConv2d> _component_writer; +}; +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute +#endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTDIRECTCONV2D */ diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp new file mode 100644 index 0000000000..f49f397ec1 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp @@ -0,0 +1,57 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "ClComponentStore.h" + +#include "src/dynamic_fusion/sketch/ArgumentPack.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.h" + +#include <memory> + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +Status ClComponentStore::validate( + const Properties &properties, + const ArgumentPack<ITensorInfo> &tensors) +{ + ARM_COMPUTE_UNUSED(properties, tensors); + return Status{}; +} +ClComponentStore::ClComponentStore(ComponentId id, const Properties &properties, const ArgumentPack<ITensorInfo> &tensors) + : IGpuKernelComponent{ id, properties, tensors }, _component_writer{ std::make_unique<ClTemplateStore>(id, tensors) } +{ +} +ClComponentStore::~ClComponentStore() +{ +} +const IGpuTemplateComponentWriter *ClComponentStore::template_writer() const +{ + return _component_writer.get(); +} +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.h b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.h new file mode 100644 index 0000000000..bf8c9f031e --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.h @@ -0,0 +1,102 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTSTORE +#define SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTSTORE + +#include "arm_compute/core/Error.h" +#include "src/dynamic_fusion/sketch/gpu/components/IGpuKernelComponent.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.h" +#include <memory> + +namespace arm_compute +{ +/** Forward declaration */ +class ITensorInfo; +namespace experimental +{ +namespace dynamic_fusion +{ +/** Forward declaration */ +template <typename T> +class ArgumentPack; + +class ClComponentStore final : public IGpuKernelComponent +{ +public: + /** Validate the component + * + * @param[in] properties Component properties + * @param[in] tensors Tensor arguments to the components + * + * @return Status Validation results + * + * Tensor argument names: + * - ACL_SRC_0: Input + * - ACL_DST_0: Output + * + * Tensor argument constness: + * - ACL_SRC_0: Const + * - ACL_DST_0: Const + * + * Valid data layouts: + * - NHWC + * + * Valid data type configurations: + * |ACL_SRC_0 |ACL_DST_0 | + * |:--------------|:--------------| + * |All |All | + */ + static Status validate( + const Properties &properties, + const ArgumentPack<ITensorInfo> &tensors); + /** Constructor + * + * Similar to @ref ClComponentStore::validate() + */ + ClComponentStore(ComponentId id, const Properties &properties, const ArgumentPack<ITensorInfo> &tensors); + /** Destructor */ + ~ClComponentStore() override; + /** Prevent instances of this class from being copy constructed */ + ClComponentStore(const ClComponentStore &component) = delete; + /** Prevent instances of this class from being copied */ + ClComponentStore &operator=(const ClComponentStore &component) = delete; + /** Allow instances of this class to be move constructed */ + ClComponentStore(ClComponentStore &&component) = default; + /** Allow instances of this class to be moved */ + ClComponentStore &operator=(ClComponentStore &&component) = default; + /** Get template writer for the component */ + const IGpuTemplateComponentWriter *template_writer() const override; + /** Get component type */ + GpuComponentType type() const override + { + return GpuComponentType::Output; + } + +private: + std::unique_ptr<ClTemplateStore> _component_writer; +}; +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute +#endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTSTORE */ |