Jakub Sujak | 3274172 | 2022-11-25 16:43:18 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2022 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuClamp.h" |
| 25 | |
| 26 | #include "arm_compute/core/experimental/Types.h" |
| 27 | |
| 28 | #include "src/core/helpers/AutoConfiguration.h" |
| 29 | #include "src/dynamic_fusion/sketch/ArgumentPack.h" |
| 30 | #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" |
| 31 | #include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.h" |
| 32 | #include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.h" |
| 33 | |
| 34 | #include "src/common/utils/Log.h" |
| 35 | |
| 36 | namespace arm_compute |
| 37 | { |
| 38 | namespace experimental |
| 39 | { |
| 40 | namespace dynamic_fusion |
| 41 | { |
| 42 | namespace |
| 43 | { |
| 44 | constexpr GpuOperatorType operator_type = GpuOperatorType::Simple; |
| 45 | } // namespace |
| 46 | |
| 47 | Status GpuClamp::is_supported_op(const GpuWorkloadContext &context, |
| 48 | const ITensorInfo *src, |
| 49 | const ITensorInfo *dst, |
| 50 | const ClampAttributes &attributes) |
| 51 | { |
| 52 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); |
| 53 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); |
| 54 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(attributes.max_val() < attributes.min_val(), "Maximum clamp value cannot be lower than minimum value"); |
| 55 | |
| 56 | // Auto initialize dst tensor info |
| 57 | TensorInfo dst_info_to_validate = *dst; |
| 58 | auto_init_if_empty(dst_info_to_validate, *src->clone()); |
| 59 | |
| 60 | // CLAMP operator is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped |
| 61 | const ClComponentActivation::Attributes act_info{ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val() }; |
| 62 | |
| 63 | // Check components |
| 64 | if(context.gpu_language() == GpuLanguage::OpenCL) |
| 65 | { |
| 66 | // Validate Activation Component |
| 67 | const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); |
| 68 | |
| 69 | ArgumentPack<ITensorInfo> arguments; |
| 70 | arguments.add_const_tensor(ACL_SRC, src); |
| 71 | arguments.add_const_tensor(ACL_DST, &dst_info_to_validate); |
| 72 | ARM_COMPUTE_RETURN_ON_ERROR(ClComponentActivation::validate(properties, arguments, act_info)); |
| 73 | } |
| 74 | else |
| 75 | { |
| 76 | ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language"); |
| 77 | } |
| 78 | return Status{}; |
| 79 | } |
| 80 | |
| 81 | Status GpuClamp::validate_op(const GpuWorkloadSketch &sketch, |
| 82 | const ITensorInfo *src, |
| 83 | const ITensorInfo *dst, |
| 84 | const ClampAttributes &attributes) |
| 85 | { |
| 86 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); |
| 87 | |
| 88 | // Check if tensors have valid id, i.e. they are created from a sketch |
| 89 | ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id() || !dst->has_valid_id()); |
| 90 | |
| 91 | // Auto initialize dst tensor info |
| 92 | TensorInfo dst_info_to_validate = *dst; |
| 93 | auto_init_if_empty(dst_info_to_validate, *src->clone()); |
| 94 | |
| 95 | // Perform fusion test to check if the operator meets fusion constraints |
| 96 | ArgumentPack<ITensorInfo> tensors; |
| 97 | tensors.add_const_tensor(ACL_SRC, src); |
| 98 | tensors.add_const_tensor(ACL_DST, &dst_info_to_validate); |
| 99 | const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors); |
| 100 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op), |
| 101 | "Operator fusion test failed. This operator cannot be fused into the workload"); |
| 102 | |
| 103 | // Check if configuration is supported |
| 104 | return is_supported_op(*sketch.gpu_context(), src, &dst_info_to_validate, attributes); |
| 105 | } |
| 106 | |
| 107 | void GpuClamp::create_op(GpuWorkloadSketch &sketch, |
| 108 | ITensorInfo *src, |
| 109 | ITensorInfo *dst, |
| 110 | const ClampAttributes &attributes) |
| 111 | { |
| 112 | // Assert validation |
| 113 | ARM_COMPUTE_ERROR_THROW_ON(GpuClamp::validate_op(sketch, src, dst, attributes)); |
| 114 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); |
| 115 | ARM_COMPUTE_LOG_PARAMS(src, dst, attributes); |
| 116 | |
| 117 | // Auto initialize dst tensor |
| 118 | auto_init_if_empty(*dst, *src->clone()); |
| 119 | |
| 120 | // Translate into components and add to component graph |
| 121 | GpuKernelComponentGraph &comp_graph = sketch.implementation().component_graph(); |
| 122 | |
| 123 | // CLAMP operator is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped |
| 124 | const ClComponentActivation::Attributes act_info{ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val() }; |
| 125 | |
| 126 | const auto *const sketch_ctx = sketch.implementation().context(); |
| 127 | |
| 128 | if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL) |
| 129 | { |
| 130 | // Add Activation Component |
| 131 | auto properties = IGpuKernelComponent::Properties(); |
| 132 | properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); |
| 133 | |
| 134 | ArgumentPack<ITensorInfo> arguments; |
| 135 | arguments.add_const_tensor(ACL_SRC, src); |
| 136 | arguments.add_const_tensor(ACL_DST, dst); |
| 137 | comp_graph.add_new_component<ClComponentActivation>(properties, arguments, act_info); |
| 138 | } |
| 139 | else |
| 140 | { |
| 141 | ARM_COMPUTE_ERROR("Unimplemented Gpu language"); |
| 142 | } |
| 143 | |
| 144 | // Set up fusion test by adding to the Operator Group |
| 145 | // Note this has to be performed after all the components have been successfully added to the component graph |
| 146 | |
| 147 | // Pack tensor infos |
| 148 | ArgumentPack<ITensorInfo> tensors; |
| 149 | tensors.add_const_tensor(ACL_SRC, src); |
| 150 | tensors.add_const_tensor(ACL_DST, dst); |
| 151 | |
| 152 | const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors); |
| 153 | sketch.implementation().operator_group().add_operator(op); |
| 154 | } |
| 155 | |
| 156 | } // namespace dynamic_fusion |
| 157 | } // namespace experimental |
| 158 | } // namespace arm_compute |