Mohammed Suhail Munshi | a18d85c | 2023-01-03 10:16:16 +0000 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2023 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 | |
| 25 | #include "arm_compute/core/CL/CLCompileContext.h" |
| 26 | #include "arm_compute/core/Validate.h" |
| 27 | #include "arm_compute/core/experimental/Types.h" |
| 28 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 29 | #include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h" |
| 30 | |
| 31 | #include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" |
| 32 | #include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuPool2d.h" |
| 33 | #include "src/common/utils/Log.h" |
| 34 | #include "src/core/helpers/AutoConfiguration.h" |
| 35 | #include "src/dynamic_fusion/sketch/ArgumentPack.h" |
| 36 | #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" |
| 37 | #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSourceCode.h" |
| 38 | #include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentPool2d.h" |
| 39 | #include "src/dynamic_fusion/utils/Utils.h" |
| 40 | |
| 41 | namespace arm_compute |
| 42 | { |
| 43 | namespace experimental |
| 44 | { |
| 45 | namespace dynamic_fusion |
| 46 | { |
| 47 | namespace |
| 48 | { |
| 49 | constexpr GpuOperatorType operator_type = GpuOperatorType::Unfusable; |
| 50 | } // namespace |
| 51 | |
| 52 | GpuPool2dSettings &GpuPool2dSettings::mixed_precision(bool mixed_precision) |
| 53 | { |
| 54 | _mixed_precision = mixed_precision; |
| 55 | return *this; |
| 56 | } |
| 57 | |
| 58 | bool GpuPool2dSettings::mixed_precision() const |
| 59 | { |
| 60 | return _mixed_precision; |
| 61 | } |
| 62 | |
| 63 | Status GpuPool2d::validate_op(const GpuWorkloadSketch &sketch, |
| 64 | const ITensorInfo *src, |
| 65 | const ITensorInfo *dst, |
| 66 | const Pool2dAttributes &attributes, |
| 67 | const GpuPool2dSettings &settings) |
| 68 | { |
| 69 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); |
| 70 | ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id() || !dst->has_valid_id()); |
| 71 | |
| 72 | // Auto initialize dst tensor info |
| 73 | TensorInfo dst_info_to_validate = *dst; |
| 74 | { |
| 75 | auto shape = misc::shape_calculator::compute_pool_shape(*src, convert_pool_attr_to_pool_info(attributes, settings.mixed_precision())); |
| 76 | auto_init_if_empty(dst_info_to_validate, src->clone()->set_tensor_shape(shape)); |
| 77 | } |
| 78 | |
| 79 | // Perform fusion test |
| 80 | // Pack tensor infos |
| 81 | ArgumentPack<ITensorInfo> tensors; |
| 82 | tensors.add_const_tensor(ACL_SRC_0, src); |
| 83 | tensors.add_const_tensor(ACL_DST_0, &dst_info_to_validate); |
| 84 | |
| 85 | const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors); |
| 86 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op), |
| 87 | "Operator fusion test failed. This operator cannot be fused into the workload"); |
| 88 | |
| 89 | // Check if configuration is supported |
| 90 | return is_supported_op(*sketch.gpu_context(), src, &dst_info_to_validate, attributes, settings); |
| 91 | } |
| 92 | |
| 93 | Status GpuPool2d::is_supported_op(const GpuWorkloadContext &context, |
| 94 | const ITensorInfo *src, |
| 95 | const ITensorInfo *dst, |
| 96 | const Pool2dAttributes &attributes, |
| 97 | const GpuPool2dSettings &settings) |
| 98 | { |
| 99 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); |
| 100 | // Data type |
| 101 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); |
| 102 | // Data layout |
| 103 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC); |
| 104 | // Check exclude padding is not false |
| 105 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(!attributes.exclude_padding(), "Exclude padding must be set to true in Attributes!"); |
| 106 | |
| 107 | // Auto initialize dst tensor info |
| 108 | TensorInfo dst_info_to_validate = *dst; |
| 109 | { |
| 110 | auto shape = misc::shape_calculator::compute_pool_shape(*src, convert_pool_attr_to_pool_info(attributes, settings.mixed_precision())); |
| 111 | auto_init_if_empty(dst_info_to_validate, src->clone()->set_tensor_shape(shape)); |
| 112 | } |
| 113 | |
| 114 | // Check components |
| 115 | if(context.gpu_language() == GpuLanguage::OpenCL) |
| 116 | { |
| 117 | const auto cl_compile_ctx = context.cl_compile_context(); |
| 118 | ARM_COMPUTE_RETURN_ERROR_ON(cl_compile_ctx == nullptr); |
| 119 | |
| 120 | // Validate Component |
| 121 | { |
| 122 | const KernelProperties properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); |
| 123 | |
| 124 | ArgumentPack<ITensorInfo> arguments; |
| 125 | arguments.add_const_tensor(ACL_SRC_0, src); |
| 126 | arguments.add_const_tensor(ACL_DST_0, &dst_info_to_validate); |
| 127 | ARM_COMPUTE_RETURN_ON_ERROR(ClComponentPool2d::validate(properties, arguments, attributes, settings)); |
| 128 | } |
| 129 | } |
| 130 | else |
| 131 | { |
| 132 | ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language"); |
| 133 | } |
| 134 | return Status{}; |
| 135 | } |
| 136 | |
| 137 | void GpuPool2d::create_op(GpuWorkloadSketch &sketch, |
| 138 | ITensorInfo *src, |
| 139 | ITensorInfo *dst, |
| 140 | const Pool2dAttributes &attributes, |
| 141 | const GpuPool2dSettings &settings) |
| 142 | { |
| 143 | // Assert validation |
| 144 | ARM_COMPUTE_ERROR_THROW_ON(GpuPool2d::validate_op(sketch, src, dst, attributes, settings)); |
| 145 | ARM_COMPUTE_LOG_PARAMS(src, dst, attributes, settings); |
| 146 | |
| 147 | // Auto initialize dst tensor |
| 148 | { |
| 149 | auto shape = misc::shape_calculator::compute_pool_shape(*src, convert_pool_attr_to_pool_info(attributes, settings.mixed_precision())); // use the default DimensionRoundingType |
| 150 | auto_init_if_empty(*dst, src->clone()->set_tensor_shape(shape)); |
| 151 | } |
| 152 | |
| 153 | // Translate into components and add to component graph |
| 154 | auto &comp_graph = sketch.implementation().component_graph(); |
| 155 | |
| 156 | const auto sketch_ctx = sketch.implementation().context(); |
| 157 | |
| 158 | if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL) |
| 159 | { |
| 160 | const auto cl_compile_ctx = sketch_ctx->cl_compile_context(); |
| 161 | ARM_COMPUTE_ERROR_ON(cl_compile_ctx == nullptr); |
| 162 | |
| 163 | // Add Component |
| 164 | { |
| 165 | auto properties = IGpuKernelComponent::Properties(); |
| 166 | properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); |
| 167 | |
| 168 | ArgumentPack<ITensorInfo> arguments; |
| 169 | arguments.add_const_tensor(ACL_SRC_0, src); |
| 170 | arguments.add_const_tensor(ACL_DST_0, dst); |
| 171 | comp_graph.add_new_component<ClComponentPool2d>(properties, arguments, attributes, settings); |
| 172 | } |
| 173 | } |
| 174 | else |
| 175 | { |
| 176 | ARM_COMPUTE_ERROR("Unimplemented Gpu language"); |
| 177 | } |
| 178 | |
| 179 | // Set up fusion test by adding to the Operator Group |
| 180 | // Note this has to be performed after all the components have been successfully added to the component graph |
| 181 | |
| 182 | // Pack tensor infos |
| 183 | ArgumentPack<ITensorInfo> tensors; |
| 184 | tensors.add_const_tensor(ACL_SRC_0, src); |
| 185 | tensors.add_tensor(ACL_DST_0, dst); |
| 186 | |
| 187 | const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors); |
| 188 | sketch.implementation().operator_group().add_operator(op); |
| 189 | } |
| 190 | |
| 191 | } // namespace dynamic_fusion |
| 192 | } // namespace experimental |
| 193 | } // namespace arm_compute |