From 32741725ac6e6c9658c51ed8585d314a1707ee8c Mon Sep 17 00:00:00 2001 From: Jakub Sujak Date: Fri, 25 Nov 2022 16:43:18 +0000 Subject: Add CLAMP operator to Dynamic Fusion interface Add the CLAMP activation function for GPU backend with generic activation Component and TemplateWriter modules. CLAMP is internally implemented as LU_BOUNDED_RELU activation function with the alpha and beta variables swapped. We do NOT consider in-place computation cases in this patch. * CLAMP operator for GPU backend. * Activation Component and TemplateWriter for CL backend. * TemplateWriter generates tiled kernel code. * Supported data types: F16, F32. * Validation tests for CLAMP operation. Resolves: COMPMID-5519 Change-Id: Ieb097d6b1e6a7ed2b882518e88314454efb402f6 Signed-off-by: Jakub Sujak Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8762 Comments-Addressed: Arm Jenkins Reviewed-by: Gunes Bayir Reviewed-by: SiCong Li Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- .../dynamic_fusion/operators/ClampFixture.h | 189 +++++++++++++++++++++ 1 file changed, 189 insertions(+) create mode 100644 tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h (limited to 'tests/validation/fixtures/dynamic_fusion/operators') diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h new file mode 100644 index 0000000000..dbac29fd22 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h @@ -0,0 +1,189 @@ +/* + * 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_CLAMPFIXTURE +#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CLAMPFIXTURE + +#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 "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 DynamicFusionClampValidationFixture : public framework::Fixture +{ +public: + template + void setup(TensorShape shape, ClampAttributes attributes, bool fuse, DataType data_type) + { + // CLAMP is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped. + ActivationLayerInfo act_info{ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val() }; + + _fuse = fuse; + _attributes = attributes; + _data_type = data_type; + _target = compute_target(shape, attributes); + _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(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, _data_type); + library->fill_static_values(tensor, get_boundary_values(static_cast(min_bound), static_cast(max_bound))); + } + + TensorType compute_target(const TensorShape &shape, ClampAttributes attributes) + { + // 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_0_info = sketch.create_tensor_info(TensorInfo(shape, 1, _data_type)); + TensorInfo dst_1_info; + + FunctionType::create_op(sketch, &src_info, &dst_0_info, attributes); + if(_fuse) + { + dst_1_info = sketch.create_tensor_info(shape, 1, _data_type); + FunctionType::create_op(sketch, &dst_0_info, &dst_1_info, attributes); + } + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // Construct user tensors + TensorType t_src{}; + TensorType t_dst_0{}; + TensorType t_dst_1{}; + + // Initialize user tensors + t_src.allocator()->init(src_info); + t_dst_0.allocator()->init(dst_0_info); + if(_fuse) + { + t_dst_1.allocator()->init(dst_1_info); + } + + // Allocate and fill user tensors + t_src.allocator()->allocate(); + t_dst_0.allocator()->allocate(); + if(_fuse) + { + t_dst_1.allocator()->allocate(); + } + + fill(AccessorType(t_src)); + + // Run runtime + if(_fuse) + { + runtime.run({ &t_src, &t_dst_1 }); + } + else + { + runtime.run({ &t_src, &t_dst_0 }); + } + + if(_fuse) + { + return t_dst_1; + } + + return t_dst_0; + } + + SimpleTensor compute_reference(const TensorShape &shape, ActivationLayerInfo act_info) + { + // Create reference + SimpleTensor src{ shape, _data_type, 1, _quantization_info }; + + // Fill reference + fill(src); + + auto dst = reference::activation_layer(src, act_info, _quantization_info); + return dst; + } + +protected: + QuantizationInfo _quantization_info{}; + ClampAttributes _attributes{}; + bool _fuse{ false }; + DataType _data_type{}; + TensorType _target{}; + SimpleTensor _reference{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CLAMPFIXTURE */ -- cgit v1.2.1