From 98aca0fda7f7c7c16bd2d1cf5386246ad796d9de Mon Sep 17 00:00:00 2001 From: Viet-Hoa Do Date: Thu, 2 Mar 2023 17:43:45 +0000 Subject: Add sigmoid and tanh for dynamic fusion * Add sigmoid and tanh activation functions for dynamic fusion. * Add corresponding tests, but both activation functions share the same fixture implementation. Resolves: COMPMID-5939 Signed-off-by: Viet-Hoa Do Change-Id: I0aae0eaa18b746ce89680d2773c66e09b0f854ce Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9257 Reviewed-by: Jakub Sujak Reviewed-by: Gunes Bayir Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- Android.bp | 2 + .../sketch/gpu/operators/GpuSigmoid.h | 89 +++++++++ .../dynamic_fusion/sketch/gpu/operators/GpuTanh.h | 89 +++++++++ filelist.json | 2 + .../sketch/gpu/operators/GpuSigmoid.cpp | 167 +++++++++++++++++ .../sketch/gpu/operators/GpuTanh.cpp | 167 +++++++++++++++++ tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp | 158 ++++++++++++++++ tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp | 158 ++++++++++++++++ .../dynamic_fusion/operators/ActivationFixture.h | 204 +++++++++++++++++++++ 9 files changed, 1036 insertions(+) create mode 100644 arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.h create mode 100644 arm_compute/dynamic_fusion/sketch/gpu/operators/GpuTanh.h create mode 100644 src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp create mode 100644 src/dynamic_fusion/sketch/gpu/operators/GpuTanh.cpp create mode 100644 tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp create mode 100644 tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp create mode 100644 tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h diff --git a/Android.bp b/Android.bp index 8d417e3d62..96a7b7a162 100644 --- a/Android.bp +++ b/Android.bp @@ -631,8 +631,10 @@ cc_library_static { "src/dynamic_fusion/sketch/gpu/operators/GpuPool2d.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuReshape.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp", + "src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuSub.cpp", + "src/dynamic_fusion/sketch/gpu/operators/GpuTanh.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", diff --git a/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.h b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.h new file mode 100644 index 0000000000..616a61e614 --- /dev/null +++ b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.h @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2023 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_GPUSIGMOID +#define ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPUSIGMOID + +#include "arm_compute/core/Types.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +/** Forward declaration */ +class GpuWorkloadContext; +class GpuWorkloadSketch; + +/** Operator interface. */ +class GpuSigmoid final +{ +public: + /** 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 | + * |:--------------|:--------------| + * |F16 |F16 | + * |F32 |F32 | + * + * Valid data layouts: + * - All + * + * @param[in, out] sketch Workload sketch into which the operator will be fused + * @param[in] src Source tensor info. Data types supported: F16/F32. + * + * @return Pointer for the destination tensor info + */ + static ITensorInfo *create_op(GpuWorkloadSketch &sketch, + ITensorInfo *src); + + /** Check if the operator configuration is supported, irrespective of fusion + * + * @param[in] context Workload context within which the operator is running + * @param[in] src Source tensor info. Data types supported: F16/F32. + * + * @return Status + */ + static Status is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *src); + + /** Validate the operator and check if it can be fused into the workload sketch. + * + * Parameters are similar to @ref GpuSigmoid::create_op() + * + * @return Status + */ + static Status validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *src); +}; + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute + +#endif //ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPUSIGMOID diff --git a/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuTanh.h b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuTanh.h new file mode 100644 index 0000000000..08b5032e93 --- /dev/null +++ b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuTanh.h @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2023 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_GPUTANH +#define ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPUTANH + +#include "arm_compute/core/Types.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +/** Forward declaration */ +class GpuWorkloadContext; +class GpuWorkloadSketch; + +/** Operator interface. */ +class GpuTanh final +{ +public: + /** 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 | + * |:--------------|:--------------| + * |F16 |F16 | + * |F32 |F32 | + * + * Valid data layouts: + * - All + * + * @param[in, out] sketch Workload sketch into which the operator will be fused + * @param[in] src Source tensor info. Data types supported: F16/F32. + * + * @return Pointer for the destination tensor info + */ + static ITensorInfo *create_op(GpuWorkloadSketch &sketch, + ITensorInfo *src); + + /** Check if the operator configuration is supported, irrespective of fusion + * + * @param[in] context Workload context within which the operator is running + * @param[in] src Source tensor info. Data types supported: F16/F32. + * + * @return Status + */ + static Status is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *src); + + /** Validate the operator and check if it can be fused into the workload sketch. + * + * Parameters are similar to @ref GpuTanh::create_op() + * + * @return Status + */ + static Status validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *src); +}; + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute + +#endif //ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPUTANH diff --git a/filelist.json b/filelist.json index cbdfb8443c..1b1db59379 100644 --- a/filelist.json +++ b/filelist.json @@ -2248,8 +2248,10 @@ "src/dynamic_fusion/sketch/gpu/operators/GpuPool2d.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuResize.cpp", + "src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.cpp", "src/dynamic_fusion/sketch/gpu/operators/GpuSub.cpp", + "src/dynamic_fusion/sketch/gpu/operators/GpuTanh.cpp", "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.cpp", "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateCast.cpp", diff --git a/src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp b/src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp new file mode 100644 index 0000000000..75b726bfda --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.cpp @@ -0,0 +1,167 @@ +/* + * Copyright (c) 2023 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/GpuSigmoid.h" +#include "arm_compute/core/experimental/Types.h" + +#include "src/dynamic_fusion/sketch/ArgumentPack.h" +#include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" +#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/common/utils/Log.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +namespace +{ +Status is_supported_op_helper(const GpuWorkloadContext &context, + const ITensorInfo *src, + const ITensorInfo *dst) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); + + TensorInfo dst_info_to_validate; + const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate; + + if(dst != nullptr) + { + dst_info_to_validate_ptr = dst; + } + + auto_init_if_empty(dst_info_to_validate, *src->clone()); + + const ClComponentActivation::Attributes act_info{ ActivationLayerInfo::ActivationFunction::LOGISTIC }; + + // Check components + if(context.gpu_language() == GpuLanguage::OpenCL) + { + // Validate Activation Component + const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); + + ArgumentPack arguments; + arguments.add_const_tensor(ACL_SRC, src); + arguments.add_const_tensor(ACL_DST, dst_info_to_validate_ptr); + ARM_COMPUTE_RETURN_ON_ERROR(ClComponentActivation::validate(properties, arguments, act_info)); + } + else + { + ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language"); + } + return Status{}; +} + +constexpr GpuOperatorType operator_type = GpuOperatorType::Simple; +} // namespace + +Status GpuSigmoid::is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *src) +{ + return is_supported_op_helper(context, src, nullptr); +} + +Status GpuSigmoid::validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *src) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); + + // Check if tensors have valid id, i.e. they are created from a sketch + ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id()); + + // Refer to GpuConv2d::validate_op() for id-validness of this TensorInfo object + TensorInfo dst_info_to_validate; + + // Auto initialize dst tensor info + auto_init_if_empty(dst_info_to_validate, *src->clone()); + + // Perform fusion test to check if the operator meets fusion constraints + ArgumentPack tensors; + tensors.add_const_tensor(ACL_SRC, src); + tensors.add_const_tensor(ACL_DST, &dst_info_to_validate); + const auto 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_helper(*sketch.gpu_context(), src, &dst_info_to_validate); +} + +ITensorInfo *GpuSigmoid::create_op(GpuWorkloadSketch &sketch, + ITensorInfo *src) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src); + ARM_COMPUTE_LOG_PARAMS(src); + ARM_COMPUTE_ERROR_THROW_ON(GpuSigmoid::validate_op(sketch, src)); + + ITensorInfo *dst = sketch.implementation().create_virtual_tensor(); + ARM_COMPUTE_ERROR_ON_NULLPTR(dst); + + // Auto initialize dst tensor + auto_init_if_empty(*dst, *src->clone()); + + // Translate into components and add to component graph + GpuKernelComponentGraph &comp_graph = sketch.implementation().component_graph(); + + const ClComponentActivation::Attributes act_info{ ActivationLayerInfo::ActivationFunction::LOGISTIC }; + + const auto *const sketch_ctx = sketch.implementation().context(); + + if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL) + { + // Add Activation Component + auto properties = IGpuKernelComponent::Properties(); + properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); + + ArgumentPack arguments; + arguments.add_const_tensor(ACL_SRC, src); + arguments.add_const_tensor(ACL_DST, dst); + comp_graph.add_new_component(properties, arguments, act_info); + } + 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, src); + tensors.add_const_tensor(ACL_DST, dst); + + const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors); + sketch.implementation().operator_group().add_operator(op); + + return dst; +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/operators/GpuTanh.cpp b/src/dynamic_fusion/sketch/gpu/operators/GpuTanh.cpp new file mode 100644 index 0000000000..c00716c76e --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/operators/GpuTanh.cpp @@ -0,0 +1,167 @@ +/* + * Copyright (c) 2023 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/GpuTanh.h" +#include "arm_compute/core/experimental/Types.h" + +#include "src/dynamic_fusion/sketch/ArgumentPack.h" +#include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" +#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.h" +#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/common/utils/Log.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +namespace +{ +Status is_supported_op_helper(const GpuWorkloadContext &context, + const ITensorInfo *src, + const ITensorInfo *dst) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); + + TensorInfo dst_info_to_validate; + const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate; + + if(dst != nullptr) + { + dst_info_to_validate_ptr = dst; + } + + auto_init_if_empty(dst_info_to_validate, *src->clone()); + + const ClComponentActivation::Attributes act_info{ ActivationLayerInfo::ActivationFunction::TANH }; + + // Check components + if(context.gpu_language() == GpuLanguage::OpenCL) + { + // Validate Activation Component + const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); + + ArgumentPack arguments; + arguments.add_const_tensor(ACL_SRC, src); + arguments.add_const_tensor(ACL_DST, dst_info_to_validate_ptr); + ARM_COMPUTE_RETURN_ON_ERROR(ClComponentActivation::validate(properties, arguments, act_info)); + } + else + { + ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language"); + } + return Status{}; +} + +constexpr GpuOperatorType operator_type = GpuOperatorType::Simple; +} // namespace + +Status GpuTanh::is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *src) +{ + return is_supported_op_helper(context, src, nullptr); +} + +Status GpuTanh::validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *src) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); + + // Check if tensors have valid id, i.e. they are created from a sketch + ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id()); + + // Refer to GpuConv2d::validate_op() for id-validness of this TensorInfo object + TensorInfo dst_info_to_validate; + + // Auto initialize dst tensor info + auto_init_if_empty(dst_info_to_validate, *src->clone()); + + // Perform fusion test to check if the operator meets fusion constraints + ArgumentPack tensors; + tensors.add_const_tensor(ACL_SRC, src); + tensors.add_const_tensor(ACL_DST, &dst_info_to_validate); + const auto 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_helper(*sketch.gpu_context(), src, &dst_info_to_validate); +} + +ITensorInfo *GpuTanh::create_op(GpuWorkloadSketch &sketch, + ITensorInfo *src) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(src); + ARM_COMPUTE_LOG_PARAMS(src); + ARM_COMPUTE_ERROR_THROW_ON(GpuTanh::validate_op(sketch, src)); + + ITensorInfo *dst = sketch.implementation().create_virtual_tensor(); + ARM_COMPUTE_ERROR_ON_NULLPTR(dst); + + // Auto initialize dst tensor + auto_init_if_empty(*dst, *src->clone()); + + // Translate into components and add to component graph + GpuKernelComponentGraph &comp_graph = sketch.implementation().component_graph(); + + const ClComponentActivation::Attributes act_info{ ActivationLayerInfo::ActivationFunction::TANH }; + + const auto *const sketch_ctx = sketch.implementation().context(); + + if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL) + { + // Add Activation Component + auto properties = IGpuKernelComponent::Properties(); + properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); + + ArgumentPack arguments; + arguments.add_const_tensor(ACL_SRC, src); + arguments.add_const_tensor(ACL_DST, dst); + comp_graph.add_new_component(properties, arguments, act_info); + } + 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, src); + tensors.add_const_tensor(ACL_DST, dst); + + const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors); + sketch.implementation().operator_group().add_operator(op); + + return dst; +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp b/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp new file mode 100644 index 0000000000..aace23eff4 --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp @@ -0,0 +1,158 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.h" +#include "arm_compute/runtime/CL/CLTensor.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +constexpr AbsoluteTolerance tolerance_f32(1e-6f); +constexpr AbsoluteTolerance tolerance_f16(0.001f); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(SIGMOID) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QASYMM8), // Unsupported data type + }), + framework::dataset::make("Expected", { true, true, false })), + input_info, expected) +{ + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + // Fuse sigmoid + const TensorInfo src_info = sketch.create_tensor_info(input_info); + + const bool res = static_cast(GpuSigmoid::validate_op(sketch, &src_info)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template +using DynamicFusionSigmoidOpFixture = DynamicFusionSigmoidValidationFixture; + +TEST_SUITE(Float) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionSigmoidOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), + framework::dataset::make("Fuse", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionSigmoidOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::Small5dShapes(), + framework::dataset::make("Fuse", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionSigmoidOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), + framework::dataset::make("Fuse", { true })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionSigmoidOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), + framework::dataset::make("Fuse", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionSigmoidOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::Small5dShapes(), + framework::dataset::make("Fuse", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionSigmoidOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), + framework::dataset::make("Fuse", { true })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // Float + +TEST_SUITE_END() // SIGMOID +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp b/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp new file mode 100644 index 0000000000..183cd079a3 --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp @@ -0,0 +1,158 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuTanh.h" +#include "arm_compute/runtime/CL/CLTensor.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +constexpr AbsoluteTolerance tolerance_f32(0.00001f); +constexpr AbsoluteTolerance tolerance_f16(0.001f); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(TANH) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QASYMM8), // Unsupported data type + }), + framework::dataset::make("Expected", { true, true, false })), + input_info, expected) +{ + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + // Fuse tanh + const TensorInfo src_info = sketch.create_tensor_info(input_info); + + const bool res = static_cast(GpuTanh::validate_op(sketch, &src_info)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template +using DynamicFusionTanhOpFixture = DynamicFusionTanhValidationFixture; + +TEST_SUITE(Float) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionTanhOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), + framework::dataset::make("Fuse", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionTanhOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::Small5dShapes(), + framework::dataset::make("Fuse", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionTanhOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), + framework::dataset::make("Fuse", { true })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionTanhOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), + framework::dataset::make("Fuse", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionTanhOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::Small5dShapes(), + framework::dataset::make("Fuse", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionTanhOpFixture, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), + framework::dataset::make("Fuse", { true })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // Float + +TEST_SUITE_END() // TANH +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/ActivationFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h new file mode 100644 index 0000000000..9656c497ea --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h @@ -0,0 +1,204 @@ +/* + * Copyright (c) 2023 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_ACTIVATIONFIXTURE +#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_ACTIVATIONFIXTURE + +#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/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/ActivationLayer.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template +class DynamicFusionActivationValidationFixture : public framework::Fixture +{ +public: + template + void setup(TensorShape shape, bool fuse, DataType data_type, ActivationLayerInfo act_info, TArgs... args) + { + _fuse = fuse; + _data_type = data_type; + _function = act_info.activation(); + _target = compute_target(shape, args...); + _reference = compute_reference(shape, act_info); + } + +protected: + std::vector get_boundary_values(T min, T max) + { + // This function will return a vector filled with the following values that can + // represent two partitions derived from equivalent partitioning. + // * Lower partition: min, min + delta, lower quarter (nominal), center - delta + // * Upper partition: center, center + delta, upper quarter (nominal), max - delta, max + const auto delta = is_data_type_float(_data_type) ? T(0.1f) : T(1); + const auto center_value = (min + max) / 2; + const auto lower_quarter = (min + center_value) / 2; + const auto upper_quarter = (center_value + max) / 2; + + std::vector boundary_values{}; + + // To ensure all the inserted values are within the given range after subtracing/adding delta + auto insert_values = [&boundary_values, &min, &max](const std::initializer_list &new_values) + { + for(auto &v : new_values) + { + if(v >= min && v <= max) + { + boundary_values.emplace_back(v); + } + } + }; + + insert_values({ min, static_cast(min + delta), static_cast(lower_quarter), static_cast(center_value - delta) }); // lower partition + insert_values({ static_cast(center_value), static_cast(center_value + delta), static_cast(upper_quarter), static_cast(max - delta), max }); // upper partition + + return boundary_values; + } + + template + void fill(U &&tensor) + { + float min_bound = 0; + float max_bound = 0; + std::tie(min_bound, max_bound) = get_activation_layer_test_bounds(_function, _data_type); + library->fill_static_values(tensor, get_boundary_values(static_cast(min_bound), static_cast(max_bound))); + } + + TensorType compute_target(const TensorShape &shape, TArgs... args) + { + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext gpu_ctx{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + // Create sketch tensors + TensorInfo src_info = sketch.create_tensor_info(TensorInfo(shape, 1, _data_type)); + TensorInfo dst_info = sketch.create_tensor_info(TensorInfo(shape, 1, _data_type)); + + ITensorInfo *ans_0_info = FunctionType::create_op(sketch, &src_info, args...); + if(_fuse) + { + ITensorInfo *ans_1_info = FunctionType::create_op(sketch, ans_0_info, args...); + GpuOutput::create_op(sketch, ans_1_info, &dst_info); + } + else + { + GpuOutput::create_op(sketch, ans_0_info, &dst_info); + } + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // 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, ActivationLayerInfo act_info) + { + // Create reference + SimpleTensor src{ shape, _data_type, 1 }; + + // Fill reference + fill(src); + + auto tmp = reference::activation_layer(src, act_info); + + if(_fuse) + { + auto dst = reference::activation_layer(tmp, act_info); + return dst; + } + else + { + return tmp; + } + } + +protected: + ActivationLayerInfo::ActivationFunction _function{}; + bool _fuse{ false }; + DataType _data_type{}; + TensorType _target{}; + SimpleTensor _reference{}; +}; + +template +class DynamicFusionSigmoidValidationFixture : public DynamicFusionActivationValidationFixture +{ +public: + template + void setup(TensorShape shape, bool fuse, DataType data_type) + { + ActivationLayerInfo act_info{ ActivationLayerInfo::ActivationFunction::LOGISTIC }; + DynamicFusionActivationValidationFixture::setup(shape, fuse, data_type, act_info); + } +}; + +template +class DynamicFusionTanhValidationFixture : public DynamicFusionActivationValidationFixture +{ +public: + template + void setup(TensorShape shape, bool fuse, DataType data_type) + { + ActivationLayerInfo act_info{ ActivationLayerInfo::ActivationFunction::TANH }; + DynamicFusionActivationValidationFixture::setup(shape, fuse, data_type, act_info); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_ACTIVATIONFIXTURE */ -- cgit v1.2.1