From 8ae571454792327fc40641c72fe0b8de1e7d334f Mon Sep 17 00:00:00 2001 From: Jakub Sujak Date: Fri, 2 Dec 2022 16:09:06 +0000 Subject: Add Resize/Scale operator to Dynamic Fusion interface Resolves: COMPMID-5521 Change-Id: Id38a4ce18f9ea8805a151acb064e72795535d1a0 Signed-off-by: Jakub Sujak Signed-off-by: Gunes Bayir Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8859 Reviewed-by: Gian Marco Iodice Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- Android.bp | 4 + .../sketch/attributes/ResizeAttributes.h | 87 ++++ .../sketch/gpu/operators/GpuResize.h | 99 ++++ filelist.json | 4 + .../sketch/attributes/ResizeAttributes.cpp | 90 ++++ .../sketch/gpu/components/cl/ClComponentResize.cpp | 85 ++++ .../sketch/gpu/components/cl/ClComponentResize.h | 127 +++++ .../sketch/gpu/operators/GpuResize.cpp | 179 +++++++ .../template_writer/cl/ClTemplateDirectConv2d.cpp | 1 + .../gpu/template_writer/cl/ClTemplateResize.cpp | 310 ++++++++++++ .../gpu/template_writer/cl/ClTemplateResize.h | 120 +++++ tests/datasets/ScaleValidationDataset.h | 18 +- tests/validation/dynamic_fusion/gpu/cl/Resize.cpp | 529 +++++++++++++++++++++ .../dynamic_fusion/operators/ResizeFixture.h | 264 ++++++++++ utils/TypePrinter.h | 33 ++ 15 files changed, 1947 insertions(+), 3 deletions(-) create mode 100644 arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h create mode 100644 arm_compute/dynamic_fusion/sketch/gpu/operators/GpuResize.h create mode 100644 src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp create mode 100644 src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp create mode 100644 src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.h create mode 100644 src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp create mode 100644 src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp create mode 100644 src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.h create mode 100644 tests/validation/dynamic_fusion/gpu/cl/Resize.cpp create mode 100644 tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h diff --git a/Android.bp b/Android.bp index 963dd84f86..072bdb9fa5 100644 --- a/Android.bp +++ b/Android.bp @@ -591,6 +591,7 @@ cc_library_static { "src/dynamic_fusion/sketch/attributes/CastAttributes.cpp", "src/dynamic_fusion/sketch/attributes/ClampAttributes.cpp", "src/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.cpp", + "src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp", "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.cpp", "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp", "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.cpp", @@ -604,6 +605,7 @@ cc_library_static { "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp", "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp", "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentElementwiseBinary.cpp", + "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp", "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuAdd.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp", @@ -611,6 +613,7 @@ cc_library_static { "src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp", + "src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp", "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.cpp", @@ -618,6 +621,7 @@ cc_library_static { "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp", + "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp", "src/gpu/cl/ClContext.cpp", diff --git a/arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h b/arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h new file mode 100644 index 0000000000..8992693cd1 --- /dev/null +++ b/arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h @@ -0,0 +1,87 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#ifndef ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_ATTRIBUTES_RESIZEATTRIBUTES +#define ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_ATTRIBUTES_RESIZEATTRIBUTES + +#include "arm_compute/core/PixelValue.h" +#include "arm_compute/core/Types.h" +#include + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +/** Attributes are backend-agnostic parameters (in addition to the input/output tensors) of an operator. + */ + +/** Resize attributes */ +class ResizeAttributes +{ +public: + /** Set output width */ + ResizeAttributes &output_width(int32_t output_width); + + /** Get output width */ + int32_t output_width() const; + + /** Set output height */ + ResizeAttributes &output_height(int32_t output_height); + + /** Get output height */ + int32_t output_height() const; + + /** Set interpolation policy */ + ResizeAttributes &interpolation_policy(InterpolationPolicy interpolation_policy); + + /** Get interpolation policy */ + InterpolationPolicy interpolation_policy() const; + + /** Set sampling policy */ + ResizeAttributes &sampling_policy(SamplingPolicy sampling_policy); + + /** Get sampling policy */ + SamplingPolicy sampling_policy() const; + + /** Set align corners */ + ResizeAttributes &align_corners(bool align_corners); + + /** Get align corners */ + bool align_corners() const; + +private: + int32_t _output_width{}; + int32_t _output_height{}; + InterpolationPolicy _interpolation_policy{ InterpolationPolicy::BILINEAR }; + SamplingPolicy _sampling_policy{ SamplingPolicy::CENTER }; + bool _align_corners{ false }; +}; + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute + +#endif /* ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_ATTRIBUTES_RESIZEATTRIBUTES */ diff --git a/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuResize.h b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuResize.h new file mode 100644 index 0000000000..1387bf1cf0 --- /dev/null +++ b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuResize.h @@ -0,0 +1,99 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#ifndef ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPURESIZE +#define ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPURESIZE + +#include "arm_compute/core/ITensorInfo.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +/** Forward declaration */ +class GpuWorkloadContext; +class GpuWorkloadSketch; + +/** Operator interface. */ +class GpuResize final +{ +public: + /** Attributes are a set of backend-agnostic parameters that define what an operator does */ + using Attributes = ResizeAttributes; + + /** Create an operator and fuse it into the workload sketch. + * @note If @ref validate_op() fails, the creation also fails and may throw an error. + * @note If @ref validate_op() fails, @p sketch remains unchanged and valid. + * + * Valid data type configurations: + * |src |dst | + * |:--------------|:--------------| + * |QASYMM8 |QASYMM8 | + * |QASYMM8_SIGNED |QASYMM8_SIGNED | + * |F16 |F16 | + * |F32 |F32 | + * |U8 |U8 | + * |S16 |S16 | + * + * Valid data layouts: + * - NHWC + * + * @param[in,out] sketch Workload sketch into which the operator will be fused + * @param[in] src Left hand side tensor info. + * @param[out] dst Destination tensor info. + * If an uninitialized ITensorInfo is passed in, it will be auto-initialized + * @param[in] attributes Operator attributes + */ + static void create_op(GpuWorkloadSketch &sketch, + ITensorInfo *src, + ITensorInfo *dst, + const Attributes &attributes); + /** Check if the operator configuration is supported, irrespective of fusion + * + * @param[in] context Workload context within which the operator is running + * @param[in] src Left hand side tensor info. + * @param[out] dst Destination tensor info. + * If an uninitialized ITensorInfo is passed in, it will be auto-initialized + * @param[in] attributes Operator attributes + */ + static Status is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *src, + const ITensorInfo *dst, + const Attributes &attributes); + /** Validate the operator and check if the its configuration is supported and if it can be fused into the workload sketch. + * Similar to @ref GpuResize::create_op() + */ + static Status validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *src, + const ITensorInfo *dst, + const Attributes &attributes); +}; + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute +#endif /* ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPURESIZE */ diff --git a/filelist.json b/filelist.json index 146731b95e..33f6fc3178 100644 --- a/filelist.json +++ b/filelist.json @@ -2202,6 +2202,7 @@ "src/dynamic_fusion/sketch/attributes/CastAttributes.cpp", "src/dynamic_fusion/sketch/attributes/ClampAttributes.cpp", "src/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.cpp", + "src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp", "src/dynamic_fusion/sketch/OperatorAttributes.cpp", "src/dynamic_fusion/sketch/gpu/GpuKernelArgument.cpp", "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp", @@ -2216,12 +2217,14 @@ "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDepthwiseConv2d.cpp", "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.cpp", "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentElementwiseBinary.cpp", + "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp", "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuAdd.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuClamp.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp", + "src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp", "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.cpp", @@ -2229,6 +2232,7 @@ "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDepthwiseConv2d.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp", + "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateStore.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.cpp" diff --git a/src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp b/src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp new file mode 100644 index 0000000000..1919dbc72d --- /dev/null +++ b/src/dynamic_fusion/sketch/attributes/ResizeAttributes.cpp @@ -0,0 +1,90 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +ResizeAttributes &ResizeAttributes::output_width(int32_t output_width) +{ + _output_width = output_width; + return *this; +} + +int32_t ResizeAttributes::output_width() const +{ + return _output_width; +} + +ResizeAttributes &ResizeAttributes::output_height(int32_t output_height) +{ + _output_height = output_height; + return *this; +} + +int32_t ResizeAttributes::output_height() const +{ + return _output_height; +} + +ResizeAttributes &ResizeAttributes::interpolation_policy(InterpolationPolicy interpolation_policy) +{ + _interpolation_policy = interpolation_policy; + return *this; +} + +InterpolationPolicy ResizeAttributes::interpolation_policy() const +{ + return _interpolation_policy; +} + +ResizeAttributes &ResizeAttributes::sampling_policy(SamplingPolicy sampling_policy) +{ + _sampling_policy = sampling_policy; + return *this; +} + +SamplingPolicy ResizeAttributes::sampling_policy() const +{ + return _sampling_policy; +} + +ResizeAttributes &ResizeAttributes::align_corners(bool align_corners) +{ + _align_corners = align_corners; + return *this; +} + +bool ResizeAttributes::align_corners() const +{ + return _align_corners; +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp new file mode 100644 index 0000000000..895b854ae2 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.cpp @@ -0,0 +1,85 @@ +/* + * 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 "ClComponentResize.h" + +#include "arm_compute/core/Error.h" +#include "src/core/CL/CLValidate.h" +#include "src/core/utils/ScaleUtils.h" +#include "src/dynamic_fusion/sketch/ArgumentPack.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +Status ClComponentResize::validate(const IGpuKernelComponent::Properties &properties, + const ArgumentPack &tensors, + const ClComponentResize::Attributes &attributes) +{ + ARM_COMPUTE_UNUSED(properties); + + const ITensorInfo *src = tensors.get_const_tensor(TensorType::ACL_SRC_0); + const ITensorInfo *dst = tensors.get_const_tensor(TensorType::ACL_DST_0); + + // Mismatching data types and quantization info + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst); + + // Device requirements met + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); + + // Align corners and sampling policy conformance + ARM_COMPUTE_RETURN_ERROR_ON(attributes.align_corners() && !arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(attributes.sampling_policy())); + + // All tensor infos are initialized + ARM_COMPUTE_RETURN_ERROR_ON(src->tensor_shape().total_size() == 0); + ARM_COMPUTE_RETURN_ERROR_ON(dst->tensor_shape().total_size() == 0); + + return Status(); +} + +ClComponentResize::ClComponentResize(ComponentId id, + const IGpuKernelComponent::Properties &properties, + const ArgumentPack &tensors, + const ClComponentResize::Attributes &attributes) + : IGpuKernelComponent{ id, properties, tensors }, + _component_writer{ std::make_unique(id, tensors, attributes) } +{ +} + +ClComponentResize::~ClComponentResize() +{ +} + +const IGpuTemplateComponentWriter *ClComponentResize::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/ClComponentResize.h b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.h new file mode 100644 index 0000000000..87d5a61ce3 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.h @@ -0,0 +1,127 @@ +/* + * 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_CLCOMPONENTRESIZE +#define SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTRESIZE + +#include "arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h" +#include "src/dynamic_fusion/sketch/gpu/components/IGpuKernelComponent.h" + +namespace arm_compute +{ +/** Forward declaration */ +class ITensorInfo; +namespace experimental +{ +namespace dynamic_fusion +{ +/** Forward declaration */ +template +class ArgumentPack; + +/** Forward declaration */ +class ClTemplateResize; + +class ClComponentResize final : public IGpuKernelComponent +{ +public: + /** Attributes are a set of backend-agnostic parameters that define what a component does */ + using Attributes = ResizeAttributes; + + /** Validate the component + * + * @param[in] properties Component properties @ref Properties + * @param[in,out] tensors Tensor arguments to the component + * @param[in] attributes Component attributes @ref Attributes + * + * @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 | + * |:--------------|:--------------| + * |QASYMM8 |QASYMM8 | + * |QASYMM8_SIGNED |QASYMM8_SIGNED | + * |F16 |F16 | + * |F32 |F32 | + * |U8 |U8 | + * |S16 |S16 | + */ + static Status validate( + const Properties &properties, + const ArgumentPack &tensors, + const Attributes &attributes); + + /** Constructor + * + * Similar to @ref ClComponentResize::validate() + */ + ClComponentResize(ComponentId id, + const Properties &properties, + const ArgumentPack &tensors, + const Attributes &attributes); + + /** Destructor */ + ~ClComponentResize() override; + + /** Prevent instances of this class from being copy constructed */ + ClComponentResize(const ClComponentResize &component) = delete; + + /** Prevent instances of this class from being copied */ + ClComponentResize &operator=(const ClComponentResize &component) = delete; + + /** Allow instances of this class to be move constructed */ + ClComponentResize(ClComponentResize &&component) = default; + + /** Allow instances of this class to be moved */ + ClComponentResize &operator=(ClComponentResize &&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 _component_writer; +}; + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute + +#endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_COMPONENTS_CL_CLCOMPONENTRESIZE */ diff --git a/src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp b/src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp new file mode 100644 index 0000000000..aa45f4c1a5 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp @@ -0,0 +1,179 @@ +/* + * 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 "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuResize.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/dynamic_fusion/sketch/ArgumentPack.h" +#include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" +#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.h" + +#include "src/common/utils/Log.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +namespace +{ +void calculate_and_init_dst_if_empty(ITensorInfo *dst, const ITensorInfo *src, const ResizeAttributes &attributes) +{ + if(dst->total_size() == 0U) + { + TensorShape out_shape = src->tensor_shape(); + + out_shape.set(1, attributes.output_width()); + out_shape.set(2, attributes.output_height()); + + auto_init_if_empty(*dst, src->clone()->set_tensor_shape(out_shape)); + } +} + +constexpr GpuOperatorType operator_type = GpuOperatorType::Complex; +} +Status GpuResize::is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *src, + const ITensorInfo *dst, + const Attributes &attributes) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + + // Auto initialize dst tensor info + TensorInfo dst_info_to_validate = *dst; + calculate_and_init_dst_if_empty(&dst_info_to_validate, src, attributes); + + // Check support level + // Data type + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + // Data layout + ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC); + // Interpolation policy + ARM_COMPUTE_RETURN_ERROR_ON_MSG(attributes.interpolation_policy() != InterpolationPolicy::NEAREST_NEIGHBOR && attributes.interpolation_policy() != InterpolationPolicy::BILINEAR, + "Interpolation policy must be NEAREST_NEIGHBOR or BILINEAR"); + + // Check components + if(context.gpu_language() == GpuLanguage::OpenCL) + { + const auto cl_compile_ctx = context.cl_compile_context(); + ARM_COMPUTE_RETURN_ERROR_ON(cl_compile_ctx == nullptr); + + // Validate Activation Component + { + const KernelProperties properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); + + ArgumentPack arguments; + arguments.add_const_tensor(ACL_SRC_0, src); + arguments.add_const_tensor(ACL_DST_0, &dst_info_to_validate); + ARM_COMPUTE_RETURN_ON_ERROR(ClComponentResize::validate(properties, arguments, attributes)); + } + } + else + { + ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language"); + } + + return Status{}; +} + +Status GpuResize::validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *src, + const ITensorInfo *dst, + const GpuResize::Attributes &attributes) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id() || !dst->has_valid_id()); + + // Auto initialize dst tensor info if empty + TensorInfo dst_info_to_validate = *dst; + calculate_and_init_dst_if_empty(&dst_info_to_validate, src, attributes); + + // Perform fusion test + // Pack tensor infos + ArgumentPack tensors; + tensors.add_const_tensor(ACL_SRC_0, src); + tensors.add_const_tensor(ACL_DST_0, &dst_info_to_validate); + const Operator op = sketch.implementation().operator_group().new_operator(operator_type, tensors); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op), + "Operator fusion test failed. This operator cannot be fused into the workload"); + + // Check if configuration is supported + return is_supported_op(*sketch.gpu_context(), src, &dst_info_to_validate, attributes); +} + +void GpuResize::create_op(GpuWorkloadSketch &sketch, + ITensorInfo *src, + ITensorInfo *dst, + const GpuResize::Attributes &attributes) +{ + // Assert validation + ARM_COMPUTE_ERROR_THROW_ON(GpuResize::validate_op(sketch, src, dst, attributes)); + ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_LOG_PARAMS(src, dst, attributes); + + // Auto initialize dst tensor info if empty + calculate_and_init_dst_if_empty(dst, src, attributes); + + // Translate into components and add to component graph + GpuKernelComponentGraph &comp_graph = sketch.implementation().component_graph(); + const auto *sketch_ctx = sketch.implementation().context(); + + if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL) + { + ARM_COMPUTE_ERROR_ON_NULLPTR(sketch_ctx->cl_compile_context()); + + // Add Resize Component + { + const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); + + ArgumentPack arguments; + arguments.add_const_tensor(ACL_SRC_0, src); + arguments.add_const_tensor(ACL_DST_0, dst); + comp_graph.add_new_component(properties, arguments, attributes); + } + } + else + { + ARM_COMPUTE_ERROR("Unimplemented Gpu language"); + } + + // Set up fusion test by adding to the Operator Group + // Note this has to be performed after all the components have been successfully added to the component graph + + // Pack tensor infos + ArgumentPack tensors; + tensors.add_const_tensor(ACL_SRC_0, src); + tensors.add_const_tensor(ACL_DST_0, dst); + + const Operator op = sketch.implementation().operator_group().new_operator(operator_type, tensors); + sketch.implementation().operator_group().add_operator(op); +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp index 221addb7b5..26399c50a9 100644 --- a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateDirectConv2d.cpp @@ -330,6 +330,7 @@ CLBuildOptions ClTemplateDirectConv2d::get_build_options(const ComponentGroup &c // to disable -cl-finite-math-only, we only include -cl-unsafe-math-optimizations build_opts.add_option("-cl-unsafe-math-optimizations"); } + build_opts.add_option("-DIS_TILED"); build_opts.add_option("-DN0=" + support::cpp11::to_string(n0)); build_opts.add_option("-DM0=" + support::cpp11::to_string(m0)); diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp new file mode 100644 index 0000000000..7ee79e82af --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.cpp @@ -0,0 +1,310 @@ +/* + * 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 "ClTemplateResize.h" + +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/utils/ScaleUtils.h" +#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +ClTemplateResize::ClTemplateResize(ComponentId id, const ArgumentPack &tensors, const ClTemplateResize::Attributes &attributes) + : IGpuTemplateComponentWriter{ id, tensors }, _src{}, _dst{}, _attributes{ attributes } +{ + _src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0); + _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0); + + ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst); +} + +std::string ClTemplateResize::get_name() const +{ + return _attributes.interpolation_policy() == InterpolationPolicy::BILINEAR ? "resize_bilinear" : "resize_nearest"; +} + +std::string ClTemplateResize::get_component_code(const IGpuTemplateComponentWriter::ComponentGroup &comp_group) const +{ + ARM_COMPUTE_UNUSED(comp_group); + + std::string code = R"_( +//------------------ START KERNEL {{meta_kernel_id}} --------------------- +TILE({{DST_DATA_TYPE}}, 1, N0, {{dst}}); +TILE(uint, 1, 1, g_dst_indirect_y); +{ + const int yo = g_ind_2 % {{arg_dst}}_h; + const int bout = g_ind_2 / {{arg_dst}}_h; +)_"; + + if(_attributes.interpolation_policy() == InterpolationPolicy::NEAREST_NEIGHBOR) + { + if(_attributes.sampling_policy() == SamplingPolicy::TOP_LEFT) + { + code += R"_( + float xi_f = (g_ind_1 * SCALE_X); + float yi_f = (yo * SCALE_Y); +)_"; + } + else + { + code += R"_( + float xi_f = ((g_ind_1 + 0.5f) * SCALE_X); + float yi_f = ((yo + 0.5f) * SCALE_Y); +)_"; + } + + if(_attributes.align_corners()) + { + code += R"_( + xi_f = round(xi_f); + yi_f = round(yi_f); +)_"; + } + + code += R"_( + const int xi0 = clamp((int)xi_f, 0, (int){{src}}_w - 1); + const int yi0 = clamp((int)yi_f, 0, (int){{src}}_h - 1); + + T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi0, xi0, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, {{dst}}); +)_"; + } + else if(_attributes.interpolation_policy() == InterpolationPolicy::BILINEAR) + { + if(_attributes.sampling_policy() == SamplingPolicy::TOP_LEFT) + { + code += R"_( + float xi_f = (g_ind_1 * SCALE_X); + float yi_f = (yo * SCALE_Y); +)_"; + } + else + { + code += R"_( + float xi_f = ((g_ind_1 + 0.5f) * SCALE_X - 0.5f); + float yi_f = ((yo + 0.5f) * SCALE_Y - 0.5f); +)_"; + } + + code += R"_( + const int xi = (int)floor(xi_f); + const int yi = (int)floor(yi_f); + + TILE({{SRC_DATA_TYPE}}, 1, N0, in00); + TILE({{SRC_DATA_TYPE}}, 1, N0, in01); + TILE({{SRC_DATA_TYPE}}, 1, N0, in10); + TILE({{SRC_DATA_TYPE}}, 1, N0, in11); + + in00[0].v = {{CONSTANT_VALUE}}; + in01[0].v = {{CONSTANT_VALUE}}; + in10[0].v = {{CONSTANT_VALUE}}; + in11[0].v = {{CONSTANT_VALUE}}; + + const int xi0 = clamp(xi, 0, (int){{src}}_w - 1); + const int yi0 = clamp(yi, 0, (int){{src}}_h - 1); + const int xi1 = clamp(xi + 1, 0, (int){{src}}_w - 1); + const int yi1 = clamp(yi + 1, 0, (int){{src}}_h - 1); + + T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi0, xi0, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, in00); + T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi0, xi1, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, in01); + T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi1, xi0, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, in10); + T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi1, xi1, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, in11); +)_"; + + if(is_data_type_float(_src->data_type())) + { + code += R"_( + const {{SRC_DATA_TYPE}} a = ({{SRC_DATA_TYPE}})(xi_f - (float)xi); + const {{SRC_DATA_TYPE}} b = ({{SRC_DATA_TYPE}})(1.f - a); + const {{SRC_DATA_TYPE}} a1 = ({{SRC_DATA_TYPE}})(yi_f - (float)yi); + const {{SRC_DATA_TYPE}} b1 = ({{SRC_DATA_TYPE}})(1.f - a1); + + // Calculate the output + {{dst}}[0].v = ((in00[0].v * b * b1) + (in01[0].v * a * b1) + (in10[0].v * b * a1) + (in11[0].v * a * a1)); +)_"; + } + else + { + code += R"_( + TILE(float, 1, N0, out_f); + TILE(float, 1, N0, in00_f); + TILE(float, 1, N0, in01_f); + TILE(float, 1, N0, in10_f); + TILE(float, 1, N0, in11_f); + + const float a = (xi_f - (float)xi); + const float b = (1.f - a); + const float a1 = (yi_f - (float)yi); + const float b1 = (1.f - a1); +)_" + // Dequantize + R"_( + LOOP_UNROLLING(int, n0, 0, 1, N0, + { + in00_f[0].s[n0] = ((float)in00[0].s[n0] - (float){{OFFSET}}) * (float){{SCALE}}; + in01_f[0].s[n0] = ((float)in01[0].s[n0] - (float){{OFFSET}}) * (float){{SCALE}}; + in10_f[0].s[n0] = ((float)in10[0].s[n0] - (float){{OFFSET}}) * (float){{SCALE}}; + in11_f[0].s[n0] = ((float)in11[0].s[n0] - (float){{OFFSET}}) * (float){{SCALE}}; + }) +)_" + // Calculate the output in the floating-point domain + R"_( + out_f[0].v = ((in00_f[0].v * b * b1) + (in01_f[0].v * a * b1) + (in10_f[0].v * b * a1) + (in11_f[0].v * a * a1)); +)_" + // Quantize + R"_( + LOOP_UNROLLING(int, n0, 0, 1, N0, + { + {{dst}}[0].s[n0] = CONVERT_SAT(out_f[0].s[n0] / (float){{SCALE}} + (float){{OFFSET}}, {{DST_DATA_TYPE}}); + }) +)_"; + } + } + else + { + ARM_COMPUTE_ERROR("Unsupported interpolation policy"); + } + + code += R"_( + g_dst_indirect_y[0].v = g_ind_1 + (yo * (int)({{arg_dst}}_w)) + bout * (int)({{arg_dst}}_w * {{arg_dst}}_h); +} +//------------------ END KERNEL {{meta_kernel_id}} --------------------- +)_"; + + return code; +} + +void ClTemplateResize::declare_variables(GpuKernelVariableTable &vtable, const IGpuTemplateComponentWriter::ComponentGroup &comp_group) const +{ + vtable.declare_variable( + _src, + GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer), + comp_group.is_intermediate_tensor(_src), + "src"); + + vtable.declare_variable( + _dst, + GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer), + comp_group.is_intermediate_tensor(_dst), + "dst"); +} + +TagLUT ClTemplateResize::get_tag_lut(const GpuKernelVariableTable &vtable, const IGpuTemplateComponentWriter::ComponentGroup &comp_group) const +{ + TagLUT lut{}; + + // Arguments and global shared variables + lut["src"] = vtable.get_variable(_src); + lut["dst"] = vtable.get_variable(_dst); + + const auto dst_argument = vtable.get_variable(comp_group.get_any_dst_tensor()); + lut["arg_dst"] = dst_argument.uniq_name; + + // Local build options + lut["meta_kernel_id"] = id(); + lut["SRC_DATA_TYPE"] = get_cl_type_from_data_type(_src->data_type()); + lut["SRC_TENSOR_TYPE"] = "BUFFER"; + lut["DST_DATA_TYPE"] = get_cl_type_from_data_type(_dst->data_type()); + lut["CONSTANT_VALUE"] = string_from_pixel_value(0, _src->data_type()); + + const bool is_qasymm_bilinear = is_data_type_quantized_asymmetric(_src->data_type()) + && _attributes.interpolation_policy() == InterpolationPolicy::BILINEAR; + + if(is_qasymm_bilinear) + { + const UniformQuantizationInfo qinfo = _src->quantization_info().uniform(); + lut["SCALE"] = support::cpp11::to_string(qinfo.scale); + lut["OFFSET"] = support::cpp11::to_string(qinfo.offset); + } + else + { + lut["SCALE"] = support::cpp11::to_string(1); + lut["OFFSET"] = support::cpp11::to_string(0); + } + + return lut; +} + +CLBuildOptions ClTemplateResize::get_build_options(const IGpuTemplateComponentWriter::ComponentGroup &comp_group) const +{ + const Window root_window = comp_group.get_root_component()->template_writer()->get_window(); + const unsigned int n0 = root_window.x().step(); + const unsigned int m0 = root_window.y().step(); + const unsigned int partial_n0 = _dst->dimension(0) % n0; + + const float scale_x = scale_utils::calculate_resize_ratio(_src->dimension(1), _dst->dimension(1), _attributes.align_corners()); + const float scale_y = scale_utils::calculate_resize_ratio(_src->dimension(2), _dst->dimension(2), _attributes.align_corners()); + + CLBuildOptions build_opts; + + build_opts.add_option("-DN0=" + support::cpp11::to_string(n0)); + build_opts.add_option("-DM0=" + support::cpp11::to_string(m0)); + build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_n0)); + build_opts.add_option("-DSCALE_X=" + float_to_string_with_full_precision(scale_x)); + build_opts.add_option("-DSCALE_Y=" + float_to_string_with_full_precision(scale_y)); + + return build_opts; +} + +std::string ClTemplateResize::get_config_id() const +{ + std::string config_id{}; + + config_id += "resize_"; + config_id += (_attributes.interpolation_policy() == InterpolationPolicy::NEAREST_NEIGHBOR ? "NEAREST_NEIGHBOR" : ""); + config_id += (_attributes.interpolation_policy() == InterpolationPolicy::BILINEAR ? "BILINEAR" : ""); + config_id += "_"; + config_id += (_attributes.sampling_policy() == SamplingPolicy::CENTER ? "center" : "topleft"); + config_id += "_"; + config_id += support::cpp11::to_string(_dst->dimension(0)); + config_id += "_"; + config_id += support::cpp11::to_string(_dst->dimension(1)); + config_id += "_"; + config_id += support::cpp11::to_string(_dst->dimension(2)); + config_id += "_"; + config_id += support::cpp11::to_string(_dst->dimension(3)); + + return config_id; +} + +std::set ClTemplateResize::get_headers_list() const +{ + return std::set{ "helpers.h", "tile_helpers.h" }; +} + +Window ClTemplateResize::get_window() const +{ + ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized"); + + const unsigned int n0 = adjust_vec_size(16 / _src->element_size(), _src->dimension(0)); + Window win = calculate_max_window(*_dst, Steps(n0)); + return win.collapse(win, Window::DimZ); +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.h b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.h new file mode 100644 index 0000000000..4c69007185 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateResize.h @@ -0,0 +1,120 @@ +/* + * 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_TEMPLATE_WRITER_CL_CLTEMPLATERESIZE +#define SRC_DYNAMIC_FUSION_SKETCH_GPU_TEMPLATE_WRITER_CL_CLTEMPLATERESIZE + +#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentResize.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/GpuKernelVariableTable.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/IGpuTemplateComponentWriter.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +class ClTemplateResize final : public IGpuTemplateComponentWriter +{ +public: + using Attributes = ClComponentResize::Attributes; + + /** Constructor + * + * @param[in] id Component id + * @param[in] tensors Tensor arguments to the components + * @param[in] attributes Component attributes + */ + ClTemplateResize(ComponentId id, const ArgumentPack &tensors, const Attributes &attributes); + + /** Destructor */ + ~ClTemplateResize() override = default; + + /** Prevent instances of this class from being copy constructed */ + ClTemplateResize(const ClTemplateResize &resize) = delete; + + /** Prevent instances of this class from being copied */ + ClTemplateResize &operator=(const ClTemplateResize &resize) = delete; + + /** Allow instances of this class to be move constructed */ + ClTemplateResize(ClTemplateResize &&resize) = default; + + /** Allow instances of this class to be moved */ + ClTemplateResize &operator=(ClTemplateResize &&resize) = default; + + /** Generate kernel component name */ + std::string get_name() const override; + + /** Generate kernel component code template + * + * @param[in] comp_group Component group of which the component is a part of + * + * @return std::string Component code + */ + std::string get_component_code(const ComponentGroup &comp_group) const override; + + /** Declare all variables used by the component in the @p vtable + * + * @param[out] vtable Variable table + * @param[in] comp_group Component group of which the component is a part of + */ + void declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const override; + + /** Generate the tag look-up table used to instantiate the component code. + * + * @param[in] vtable Variable table + * @param[in] comp_group Component group of which the component is a part of + * + * @return TagLUT Tag lookup table + */ + TagLUT get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const override; + + /** Generate the build options used in the component + * + * @param[in] comp_group Component group of which the component is a part of + * + * @return CLBuildOptions Build options + */ + CLBuildOptions get_build_options(const ComponentGroup &comp_group) const override; + + /** Generate the component config id string used for tuning */ + std::string get_config_id() const override; + + /** Generate the header list used in the component */ + std::set get_headers_list() const override; + + /** Generate the execution window for the component */ + Window get_window() const override; + +private: + const ITensorInfo *_src; + const ITensorInfo *_dst; + Attributes _attributes; +}; + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute + +#endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_TEMPLATE_WRITER_CL_CLTEMPLATERESIZE */ diff --git a/tests/datasets/ScaleValidationDataset.h b/tests/datasets/ScaleValidationDataset.h index c6987c0908..8987c3a1c1 100644 --- a/tests/datasets/ScaleValidationDataset.h +++ b/tests/datasets/ScaleValidationDataset.h @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_SCALE_VALIDATION_DATASET -#define ARM_COMPUTE_TEST_SCALE_VALIDATION_DATASET +#ifndef TESTS_DATASETS_SCALEVALIDATIONDATASET +#define TESTS_DATASETS_SCALEVALIDATIONDATASET #include "arm_compute/core/Types.h" #include "tests/datasets/BorderModeDataset.h" @@ -173,6 +173,11 @@ framework::dataset::make("AlignCorners", { true })); datasets::BorderModes()), \ samping_policy_set) +#define ASSEMBLE_DATASET_DYNAMIC_FUSION(shape, samping_policy_set) \ + combine(combine(combine((shape), framework::dataset::make("DataLayout", { DataLayout::NHWC })), \ + ScaleInterpolationPolicySet), \ + samping_policy_set) + #define ASSEMBLE_S8_DATASET(shape, samping_policy_set) \ combine(combine(combine(combine((shape), framework::dataset::make("DataLayout", DataLayout::NHWC)), \ framework::dataset::make("InterpolationPolicy", { InterpolationPolicy::BILINEAR })), \ @@ -194,6 +199,13 @@ framework::dataset::make("AlignCorners", { true })); datasets::BorderModes()), \ sampling_policy_set) +#define ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(shape, sampling_policy_set, quantization_info_set) \ + combine(combine(combine(combine(shape, \ + quantization_info_set), \ + framework::dataset::make("DataLayout", { DataLayout::NHWC })), \ + ScaleInterpolationPolicySet), \ + sampling_policy_set) + /** Generating dataset for quantized data tyeps with the given shapes */ #define ASSEMBLE_DIFFERENTLY_QUANTIZED_DATASET(shape, sampling_policy_set, input_quant_info_set, output_quant_info_set) \ combine(combine(combine(combine(combine(combine(shape, \ @@ -207,4 +219,4 @@ framework::dataset::make("AlignCorners", { true })); } // namespace datasets } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_SCALE_VALIDATION_DATASET */ +#endif /* TESTS_DATASETS_SCALEVALIDATIONDATASET */ diff --git a/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp b/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp new file mode 100644 index 0000000000..17c341b803 --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp @@ -0,0 +1,529 @@ +/* +* 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 "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuResize.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ScaleValidationDataset.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h" + +using namespace arm_compute::experimental::dynamic_fusion; +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +using datasets::ScaleShapesBaseDataSet; +using datasets::ScaleInterpolationPolicySet; +using datasets::ScaleSamplingPolicySet; +using datasets::ScaleAlignCornersSamplingPolicySet; + +/** We consider vector size in byte 16 since the maximum size of + * a vector used by @ref CLScaleKernel is currently 16-byte (float4). + */ +constexpr uint32_t vector_byte = 16; + +template +constexpr uint32_t num_elements_per_vector() +{ + return vector_byte / sizeof(T); +} + +/** Quantization information data set */ +const auto QuantizationInfoSet = framework::dataset::make("QuantizationInfo", +{ + QuantizationInfo(0.5f, -1), +}); + +/** Tolerance */ +constexpr AbsoluteTolerance tolerance_q8(1); +constexpr AbsoluteTolerance tolerance_qs8(1); +constexpr AbsoluteTolerance tolerance_s16(1); +constexpr float tolerance_f32_absolute(0.001f); + +RelativeTolerance tolerance_f32(0.05); +constexpr float abs_tolerance_f16(0.1f); +RelativeTolerance tolerance_f16(half(0.1)); + +constexpr float tolerance_num_f32(0.01f); + +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(RESIZE) + +TEST_SUITE(Validate) + +const auto default_input_shape = TensorShape{ 2, 3, 3, 2 }; +const auto default_output_shape = TensorShape{ 4, 6, 3, 2 }; + +constexpr auto default_data_type = DataType::U8; +constexpr auto default_data_layout = DataLayout::NHWC; +constexpr auto default_interpolation_policy = InterpolationPolicy::NEAREST_NEIGHBOR; +constexpr bool default_use_padding = false; + +TEST_CASE(NullPtr, framework::DatasetMode::ALL) +{ + const TensorInfo input_info = TensorInfo{ default_input_shape, 1, default_data_type, default_data_layout }; + const TensorInfo output_info = TensorInfo{ default_output_shape, 1, default_data_type, default_data_layout }; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info); + const TensorInfo sketch_output_info = sketch.create_tensor_info(output_info); + + // nullptr is given as input + Status status = GpuResize::validate_op(sketch, nullptr, &sketch_output_info, ResizeAttributes()); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); + + // nullptr is given as output + status = GpuResize::validate_op(sketch, &sketch_input_info, nullptr, ResizeAttributes()); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_CASE(SupportDataType, framework::DatasetMode::ALL) +{ + const std::map supported_data_types = + { + { DataType::U8, true }, + { DataType::S8, false }, + { DataType::QSYMM8, false }, + { DataType::QASYMM8, true }, + { DataType::QASYMM8_SIGNED, true }, + { DataType::QSYMM8_PER_CHANNEL, false }, + { DataType::U16, false }, + { DataType::S16, true }, + { DataType::QSYMM16, false }, + { DataType::QASYMM16, false }, + { DataType::U32, false }, + { DataType::S32, false }, + { DataType::U64, false }, + { DataType::S64, false }, + { DataType::BFLOAT16, false }, + { DataType::F16, true }, + { DataType::F32, true }, + { DataType::F64, false }, + { DataType::SIZET, false }, + }; + + for(auto &kv : supported_data_types) + { + const TensorInfo input_info = TensorInfo{ default_input_shape, 1, kv.first, default_data_layout }; + const TensorInfo output_info = TensorInfo{ default_output_shape, 1, kv.first, default_data_layout }; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info); + const TensorInfo sketch_output_info = sketch.create_tensor_info(output_info); + + // nullptr is given as input + Status status = GpuResize::validate_op(sketch, &sketch_input_info, &sketch_output_info, ResizeAttributes()); + ARM_COMPUTE_EXPECT(bool(status) == kv.second, framework::LogLevel::ERRORS); + } +} + +TEST_CASE(MismatchingDataType, framework::DatasetMode::ALL) +{ + constexpr DataType non_default_data_type = DataType::F32; + + const TensorInfo input_info = TensorInfo{ default_input_shape, 1, default_data_type, default_data_layout }; + const TensorInfo output_info = TensorInfo{ default_output_shape, 1, non_default_data_type, default_data_layout }; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info); + const TensorInfo sketch_output_info = sketch.create_tensor_info(output_info); + + Status status = GpuResize::validate_op(sketch, &sketch_input_info, &sketch_output_info, ResizeAttributes()); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_CASE(AlignedCornerNotSupported, framework::DatasetMode::ALL) +{ + // Aligned corners require sampling policy to be TOP_LEFT. + constexpr InterpolationPolicy interpolation_policy = InterpolationPolicy::BILINEAR; + constexpr bool align_corners = true; + constexpr SamplingPolicy sampling_policy = SamplingPolicy::CENTER; + + const TensorInfo input_info = TensorInfo{ default_input_shape, 1, default_data_type, default_data_layout }; + const TensorInfo output_info = TensorInfo{ default_output_shape, 1, default_data_type, default_data_layout }; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info); + const TensorInfo sketch_output_info = sketch.create_tensor_info(output_info); + + ResizeAttributes attributes{}; + attributes.interpolation_policy(interpolation_policy) + .sampling_policy(sampling_policy) + .align_corners(align_corners); + + Status status = GpuResize::validate_op(sketch, &sketch_input_info, &sketch_output_info, attributes); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_CASE(UnsupportedInterpolationPolicy, framework::DatasetMode::ALL) +{ + const TensorInfo input_info = TensorInfo{ TensorShape(28U, 33U, 2U), 1, DataType::F32, default_data_layout }; + const TensorInfo output_info = TensorInfo{ TensorShape(26U, 21U, 2U), 1, DataType::F32, default_data_layout }; + constexpr auto interpolation_policy = InterpolationPolicy::AREA; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info); + const TensorInfo sketch_output_info = sketch.create_tensor_info(output_info); + + ResizeAttributes attributes{}; + attributes.interpolation_policy(interpolation_policy); + + Status status = GpuResize::validate_op(sketch, &sketch_input_info, &sketch_output_info, attributes); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_CASE(UnsupportedLayout, framework::DatasetMode::ALL) +{ + const TensorInfo input_info = TensorInfo{ default_input_shape, 1, default_data_type, DataLayout::NCHW }; + const TensorInfo output_info = TensorInfo{ default_output_shape, 1, default_data_type, DataLayout::NCHW }; + constexpr auto interpolation_policy = InterpolationPolicy::BILINEAR; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + const TensorInfo sketch_input_info = sketch.create_tensor_info(input_info); + const TensorInfo sketch_output_info = sketch.create_tensor_info(output_info); + + ResizeAttributes attributes{}; + attributes.interpolation_policy(interpolation_policy); + + Status status = GpuResize::validate_op(sketch, &sketch_input_info, &sketch_output_info, attributes); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_SUITE_END() // Validate + +template +using DynamicFusionResizeFixture = DynamicFusionResizeValidationFixture; + +TEST_SUITE(Float) +TEST_SUITE(FP32) + +const auto f32_shape = combine((SCALE_PRECOMMIT_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::F32)); + +FIXTURE_DATA_TEST_CASE(Run, DynamicFusionResizeFixture, framework::DatasetMode::ALL, ASSEMBLE_DATASET_DYNAMIC_FUSION(f32_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32, tolerance_f32_absolute); +} + +FIXTURE_DATA_TEST_CASE(RunAlignCorners, DynamicFusionResizeFixture, framework::DatasetMode::ALL, ASSEMBLE_DATASET_DYNAMIC_FUSION(f32_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32, tolerance_f32_absolute); +} +const auto f32_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::F32)); +FIXTURE_DATA_TEST_CASE(RunNightly, DynamicFusionResizeFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET_DYNAMIC_FUSION(f32_nightly_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32, tolerance_f32_absolute); +} +FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, DynamicFusionResizeFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET_DYNAMIC_FUSION(f32_nightly_shape, + ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32, tolerance_f32_absolute); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +const auto f16_shape = combine((SCALE_PRECOMMIT_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::F16)); +FIXTURE_DATA_TEST_CASE(Run, DynamicFusionResizeFixture, framework::DatasetMode::ALL, ASSEMBLE_DATASET_DYNAMIC_FUSION(f16_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunAlignCorners, DynamicFusionResizeFixture, framework::DatasetMode::ALL, ASSEMBLE_DATASET_DYNAMIC_FUSION(f16_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +const auto f16_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::F16)); +FIXTURE_DATA_TEST_CASE(RunNightly, DynamicFusionResizeFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET_DYNAMIC_FUSION(f16_nightly_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, DynamicFusionResizeFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET_DYNAMIC_FUSION(f16_nightly_shape, + ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float + +TEST_SUITE(Integer) +TEST_SUITE(U8) +const auto u8_shape = combine((SCALE_PRECOMMIT_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::U8)); +FIXTURE_DATA_TEST_CASE(Run, DynamicFusionResizeFixture, framework::DatasetMode::ALL, ASSEMBLE_DATASET_DYNAMIC_FUSION(u8_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_q8); +} +FIXTURE_DATA_TEST_CASE(RunAlignCorners, DynamicFusionResizeFixture, framework::DatasetMode::ALL, ASSEMBLE_DATASET_DYNAMIC_FUSION(u8_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_q8); +} +const auto u8_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::U8)); +FIXTURE_DATA_TEST_CASE(RunNightly, DynamicFusionResizeFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET_DYNAMIC_FUSION(u8_nightly_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_q8); +} +FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, DynamicFusionResizeFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET_DYNAMIC_FUSION(u8_nightly_shape, + ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_q8); +} +TEST_SUITE_END() // U8 + +TEST_SUITE(S16) +const auto s16_shape = combine((SCALE_PRECOMMIT_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::S16)); +FIXTURE_DATA_TEST_CASE(Run, DynamicFusionResizeFixture, framework::DatasetMode::ALL, ASSEMBLE_DATASET_DYNAMIC_FUSION(s16_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_s16); +} +FIXTURE_DATA_TEST_CASE(RunAlignCorners, DynamicFusionResizeFixture, framework::DatasetMode::ALL, ASSEMBLE_DATASET_DYNAMIC_FUSION(s16_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_s16); +} +const auto s16_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::S16)); +FIXTURE_DATA_TEST_CASE(RunNightly, DynamicFusionResizeFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET_DYNAMIC_FUSION(s16_nightly_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_s16); +} +FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, DynamicFusionResizeFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET_DYNAMIC_FUSION(s16_nightly_shape, + ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_s16); +} +TEST_SUITE_END() // S16 +TEST_SUITE_END() // Integer + +template +using DynamicFusionResizeQuantizedFixture = DynamicFusionResizeQuantizedValidationFixture; +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +const auto qasymm8_shape = combine((SCALE_PRECOMMIT_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::QASYMM8)); +FIXTURE_DATA_TEST_CASE(Run, DynamicFusionResizeQuantizedFixture, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(qasymm8_shape, ScaleSamplingPolicySet, + QuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_q8); +} +FIXTURE_DATA_TEST_CASE(RunAlignCorners, DynamicFusionResizeQuantizedFixture, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(qasymm8_shape, + ScaleAlignCornersSamplingPolicySet, + QuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_q8); +} +const auto qasymm8_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::QASYMM8)); +FIXTURE_DATA_TEST_CASE(RunNightly, DynamicFusionResizeQuantizedFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(qasymm8_nightly_shape, + ScaleSamplingPolicySet, + QuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_q8); +} +FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, DynamicFusionResizeQuantizedFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(qasymm8_nightly_shape, + ScaleAlignCornersSamplingPolicySet, + QuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_q8); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +const auto qasymm8_signed_shape = combine((SCALE_PRECOMMIT_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)); +FIXTURE_DATA_TEST_CASE(Run, DynamicFusionResizeQuantizedFixture, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(qasymm8_signed_shape, ScaleSamplingPolicySet, + QuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_qs8); +} +FIXTURE_DATA_TEST_CASE(RunAlignCorners, DynamicFusionResizeQuantizedFixture, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(qasymm8_signed_shape, + ScaleAlignCornersSamplingPolicySet, + QuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_qs8); +} +const auto qasymm8_signed_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector())), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)); +FIXTURE_DATA_TEST_CASE(RunNightly, DynamicFusionResizeQuantizedFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(qasymm8_signed_nightly_shape, + ScaleSamplingPolicySet, + QuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_qs8); +} +FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, DynamicFusionResizeQuantizedFixture, framework::DatasetMode::NIGHTLY, ASSEMBLE_QUANTIZED_DATASET_DYNAMIC_FUSION(qasymm8_signed_nightly_shape, + ScaleAlignCornersSamplingPolicySet, + QuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_qs8); +} +TEST_SUITE_END() // QASYMM8_SIGNED + +TEST_SUITE_END() // Quantized + +TEST_SUITE_END() // RESIZE +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h new file mode 100644 index 0000000000..5cdf52e62b --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h @@ -0,0 +1,264 @@ +/* +* 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 TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESIZEFIXTURE +#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESIZEFIXTURE + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/SimpleTensor.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/Validation.h" +#include "tests/validation/reference/Permute.h" +#include "tests/validation/reference/Scale.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template +class DynamicFusionResizeGenericValidationFixture : public framework::Fixture +{ +public: + template + void setup(TensorShape shape, DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout, + InterpolationPolicy interpolation_policy, SamplingPolicy sampling_policy, + bool align_corners, QuantizationInfo output_quantization_info) + { + _shape = shape; + _interpolation_policy = interpolation_policy; + _sampling_policy = sampling_policy; + _data_type = data_type; + _input_quantization_info = quantization_info; + _output_quantization_info = output_quantization_info; + _align_corners = align_corners; + _data_layout = data_layout; + + ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion resize supports only NHWC layout + + generate_scale(shape); + + std::mt19937 generator(library->seed()); + std::uniform_int_distribution distribution_u8(0, 255); + + _target = compute_target(shape); + _reference = compute_reference(shape); + } + +protected: + void generate_scale(const TensorShape &shape) + { + static constexpr float _min_scale{ 0.25f }; + static constexpr float _max_scale{ 3.f }; + + constexpr float max_width{ 8192.0f }; + constexpr float max_height{ 6384.0f }; + constexpr float min_width{ 1.f }; + constexpr float min_height{ 1.f }; + + std::mt19937 generator(library->seed()); + std::uniform_real_distribution distribution_float(_min_scale, _max_scale); + + auto generate = [&](size_t input_size, float min_output, float max_output) -> int + { + const float generated_scale = distribution_float(generator); + const int output_size = static_cast(utility::clamp(static_cast(input_size) * generated_scale, min_output, max_output)); + return output_size; + }; + + // Input shape is always given in NCHW layout. NHWC is dealt by permute in compute_target() + const int idx_width = get_data_layout_dimension_index(DataLayout::NCHW, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(DataLayout::NCHW, DataLayoutDimension::HEIGHT); + + _output_width = generate(shape[idx_width], min_width, max_width); + _output_height = generate(shape[idx_height], min_height, max_height); + } + + template + void fill(U &&tensor) + { + if(tensor.data_type() == DataType::F32) + { + std::uniform_real_distribution distribution(-5.0f, 5.0f); + library->fill(tensor, distribution, 0); + } + else if(tensor.data_type() == DataType::F16) + { + arm_compute::utils::uniform_real_distribution_16bit distribution{ -5.0f, 5.0f }; + library->fill(tensor, distribution, 0); + } + else if(is_data_type_quantized(tensor.data_type())) + { + std::uniform_int_distribution<> distribution(0, 100); + library->fill(tensor, distribution, 0); + } + else + { + library->fill_tensor_uniform(tensor, 0); + } + } + + TensorType compute_target(TensorShape shape) + { + // Our test shapes are assumed in NCHW data layout, thus the permutation + permute(shape, PermutationVector(2U, 0U, 1U)); + + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + // Create sketch tensors + TensorInfo src_info = sketch.create_tensor_info(TensorInfo(shape, 1, _data_type, _data_layout)); + src_info.set_quantization_info(_input_quantization_info); + TensorInfo dst_info = sketch.create_tensor_info(); + + ResizeAttributes attributes; + attributes.align_corners(_align_corners).sampling_policy(_sampling_policy).interpolation_policy(_interpolation_policy).output_width(_output_width).output_height(_output_height); + + TensorInfo scale_result_info = sketch.create_tensor_info(); + + FunctionType::create_op(sketch, &src_info, &scale_result_info, attributes); + GpuOutput::create_op(sketch, &scale_result_info, &dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + for(auto &data : runtime.get_auxiliary_tensors()) + { + auto tensor = data.first; + const auto aux_mem_req = data.second; + tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment); + tensor->allocator()->allocate(); + } + + // Construct user tensors + TensorType t_src{}; + TensorType t_dst{}; + + // Initialize user tensors + t_src.allocator()->init(src_info); + t_dst.allocator()->init(dst_info); + + // Allocate and fill user tensors + t_src.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_src)); + + // Run runtime + runtime.run({ &t_src, &t_dst }); + + return t_dst; + } + + SimpleTensor compute_reference(const TensorShape &shape) + { + // Create reference + SimpleTensor src{ shape, _data_type, 1, _input_quantization_info }; + + // Reference code is NCHW, so the input shapes are NCHW + const int idx_width = get_data_layout_dimension_index(DataLayout::NCHW, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(DataLayout::NCHW, DataLayoutDimension::HEIGHT); + + const float scale_x = static_cast(_output_width) / shape[idx_width]; + const float scale_y = static_cast(_output_height) / shape[idx_height]; + + // Fill reference + fill(src); + + return reference::scale(src, scale_x, scale_y, _interpolation_policy, + BorderMode::REPLICATE, static_cast(0), _sampling_policy, /* ceil_policy_scale */ false, + _align_corners, _output_quantization_info); + } + + TensorType _target{}; + SimpleTensor _reference{}; + TensorShape _shape{}; + InterpolationPolicy _interpolation_policy{}; + SamplingPolicy _sampling_policy{}; + DataType _data_type{}; + DataLayout _data_layout{}; + QuantizationInfo _input_quantization_info{}; + QuantizationInfo _output_quantization_info{}; + bool _align_corners{ false }; + int _output_width{ 0 }; + int _output_height{ 0 }; +}; + +template +class DynamicFusionResizeValidationFixture : public DynamicFusionResizeGenericValidationFixture +{ +public: + template + void setup(TensorShape shape, DataType data_type, DataLayout data_layout, InterpolationPolicy policy, SamplingPolicy sampling_policy, bool align_corners) + { + DynamicFusionResizeGenericValidationFixture::setup(shape, + data_type, + QuantizationInfo(), + data_layout, + policy, + sampling_policy, + align_corners, + QuantizationInfo()); + } +}; + +template +class DynamicFusionResizeQuantizedValidationFixture : public DynamicFusionResizeGenericValidationFixture +{ +public: + template + void setup(TensorShape shape, DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout, InterpolationPolicy policy, SamplingPolicy sampling_policy, + bool align_corners) + { + DynamicFusionResizeGenericValidationFixture::setup(shape, + data_type, + quantization_info, + data_layout, + policy, + sampling_policy, + align_corners, + quantization_info); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESIZEFIXTURE */ diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h index d4265bfdbd..d056bfbcf4 100644 --- a/utils/TypePrinter.h +++ b/utils/TypePrinter.h @@ -42,6 +42,7 @@ #include "arm_compute/dynamic_fusion/sketch/attributes/CastAttributes.h" #include "arm_compute/dynamic_fusion/sketch/attributes/ClampAttributes.h" #include "arm_compute/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h" #include "arm_compute/runtime/CL/CLTunerTypes.h" #include "arm_compute/runtime/CL/CLTypes.h" #include "arm_compute/runtime/FunctionDescriptors.h" @@ -3524,6 +3525,38 @@ inline std::string to_string(const experimental::dynamic_fusion::ClampAttributes return str.str(); } +/** Formatted output of the arm_compute::experimental::dynamic_fusion::ResizeAttributes type. + * + * @param[out] os Output stream. + * @param[in] resize_attr arm_compute::experimental::dynamic_fusion::ResizeAttributes type to output. + * + * @return Modified output stream. + */ +inline ::std::ostream &operator<<(::std::ostream &os, const experimental::dynamic_fusion::ResizeAttributes &resize_attr) +{ + os << "ResizeAttributes=" + << "[" + << "AlignCorners=" << resize_attr.align_corners() << ", " + << "InterpolationPolicy=" << resize_attr.interpolation_policy() << ", " + << "OutputHeight=" << resize_attr.output_height() << ", " + << "OutputWidth=" << resize_attr.output_width() << ", " + << "SamplingPolicy=" << resize_attr.sampling_policy() << "]"; + return os; +} + +/** Formatted output of the arm_compute::experimental::dynamic_fusion::ResizeAttributes type. + * + * @param[in] resize_attr arm_compute::experimental::dynamic_fusion::ResizeAttributes type to output. + * + * @return Formatted string. + */ +inline std::string to_string(const experimental::dynamic_fusion::ResizeAttributes &resize_attr) +{ + std::stringstream str; + str << resize_attr; + return str.str(); +} + } // namespace arm_compute #endif /* __ARM_COMPUTE_TYPE_PRINTER_H__ */ -- cgit v1.2.1