diff options
author | Ramy Elgammal <ramy.elgammal@arm.com> | 2022-11-08 02:14:46 +0000 |
---|---|---|
committer | Ramy Elgammal <ramy.elgammal@arm.com> | 2022-11-29 10:24:59 +0000 |
commit | 404462af4ca002ece819161a03a4bdb19a87abf2 (patch) | |
tree | 09cf812530afcbe3fc524ce7eded5f06129e0889 /src/dynamic_fusion/sketch/gpu/operators/internal | |
parent | 03b2971ac69a86f10a1566938d1a25afee15746c (diff) | |
download | ComputeLibrary-404462af4ca002ece819161a03a4bdb19a87abf2.tar.gz |
Adding GpuAdd to dynamic fusion operators
- Provide support for Add operator
- Auto initialize the destination tensor before testing fusion in conv2d
and elementwise binary ops.
Resolves: COMPMID-5518
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Ibd815020f02b57f88eea7c2921bdcf98605d99c5
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8617
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/dynamic_fusion/sketch/gpu/operators/internal')
-rw-r--r-- | src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp | 200 | ||||
-rw-r--r-- | src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.h | 113 |
2 files changed, 313 insertions, 0 deletions
diff --git a/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp b/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp new file mode 100644 index 0000000000..073924947c --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp @@ -0,0 +1,200 @@ +/* + * 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 "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.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/ClComponentElementwiseBinary.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +namespace +{ +void calculate_and_init_dst_if_empty(ITensorInfo *dst, const ITensorInfo *lhs, const ITensorInfo *rhs) +{ + if(dst->total_size() == 0U) + { + const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*lhs, *rhs); + auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(broadcast_pair.first)); + } +} +GpuOperatorType operator_type = GpuOperatorType::Simple; +} + +ElementwiseBinaryCommonAttributes &ElementwiseBinaryCommonAttributes::operation(const ElementwiseBinaryCommonAttributes::ElementwiseOp &operation) +{ + _operation = operation; + return *this; +} + +ElementwiseBinaryCommonAttributes::ElementwiseOp ElementwiseBinaryCommonAttributes::operation() const +{ + return _operation; +} + +Status GpuElementwiseBinaryCommon::is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *lhs, + const ITensorInfo *rhs, + const ITensorInfo *dst, + const ElementwiseBinaryCommonAttributes &attributes) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs, dst); + + // Auto initialize dst tensor info + TensorInfo dst_info_to_validate = *dst; + calculate_and_init_dst_if_empty(&dst_info_to_validate, lhs, rhs); + + // 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 ElementwiseBinary Component + { + ArgumentPack<ITensorInfo> arguments; + arguments.add_const_tensor(ACL_SRC_0, lhs); + arguments.add_const_tensor(ACL_SRC_1, rhs); + + // We needed to pass the original dst pointer for in-place detection, in case its shape is not empty + if(dst->tensor_shape().total_size() == 0) + { + arguments.add_const_tensor(ACL_DST_0, &dst_info_to_validate); + } + else + { + arguments.add_const_tensor(ACL_DST_0, dst); + } + ARM_COMPUTE_RETURN_ON_ERROR(ClComponentElementwiseBinary::validate(arguments, attributes)); + } + } + else + { + ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language"); + } + + return Status{}; +} + +Status GpuElementwiseBinaryCommon::validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *lhs, + const ITensorInfo *rhs, + const ITensorInfo *dst, + const ElementwiseBinaryCommonAttributes &attributes) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs, dst); + ARM_COMPUTE_RETURN_ERROR_ON( + !lhs->has_valid_id() || !rhs->has_valid_id() || !dst->has_valid_id()); + + // Auto initialize dst tensor info + TensorInfo dst_info_to_validate = *dst; + calculate_and_init_dst_if_empty(&dst_info_to_validate, lhs, rhs); + + // Perform fusion test + // Pack tensor infos + ArgumentPack<ITensorInfo> tensors; + tensors.add_const_tensor(ACL_SRC_0, lhs); + tensors.add_const_tensor(ACL_SRC_1, rhs); + tensors.add_const_tensor(ACL_DST_0, &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, and passing the original dst for in-place detection + return is_supported_op(*sketch.gpu_context(), lhs, rhs, dst, attributes); +} + +void GpuElementwiseBinaryCommon::create_op(GpuWorkloadSketch &sketch, + ITensorInfo *lhs, + ITensorInfo *rhs, + ITensorInfo *dst, + const ElementwiseBinaryCommonAttributes &attributes) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst); + const bool in_place = (lhs == dst) || (rhs == dst); + static TensorInfo in_place_dst; + in_place_dst = in_place ? sketch.create_tensor_info(*lhs) : TensorInfo{}; + + // Auto initialize dst tensor + calculate_and_init_dst_if_empty(dst, lhs, rhs); + + // Translate into components and add to component graph + auto &comp_graph = sketch.implementation().component_graph(); + + const auto sketch_ctx = sketch.implementation().context(); + + if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL) + { + const auto cl_compile_ctx = sketch_ctx->cl_compile_context(); + ARM_COMPUTE_ERROR_ON(cl_compile_ctx == nullptr); + + // Add ElementwiseBinary Component + { + auto properties = IGpuKernelComponent::Properties(); + properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); + + ArgumentPack<ITensorInfo> arguments; + arguments.add_const_tensor(ACL_SRC_0, lhs); + arguments.add_const_tensor(ACL_SRC_1, rhs); + if(in_place) + { + arguments.add_const_tensor(ACL_DST_0, &in_place_dst); + } + else + { + arguments.add_const_tensor(ACL_DST_0, dst); + } + comp_graph.add_new_component<ClComponentElementwiseBinary>(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<ITensorInfo> tensors; + tensors.add_const_tensor(ACL_SRC_0, lhs); + tensors.add_const_tensor(ACL_SRC_1, rhs); + if(in_place) + { + tensors.add_const_tensor(ACL_DST_0, &in_place_dst); + } + else + { + tensors.add_tensor(ACL_DST_0, dst); + } + const auto 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/operators/internal/GpuElementwiseBinaryCommon.h b/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.h new file mode 100644 index 0000000000..ffae801e47 --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.h @@ -0,0 +1,113 @@ +/* + * 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_OPERATORS_INTERNAL_GPUELEMENTWISEBINARYCOMMON +#define SRC_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_INTERNAL_GPUELEMENTWISEBINARYCOMMON + +#include "arm_compute/core/Error.h" +#include "arm_compute/dynamic_fusion/sketch/OperatorAttributes.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +class ElementwiseBinaryCommonAttributes +{ +public: + enum class ElementwiseOp + { + ADD, /**< (x + y) */ + SUB, /**< (x - y) */ + DIV, /**< (x / y) */ + MIN, /**< Min(x, y) */ + MAX, /**< Max(x, y) */ + SQUARED_DIFF, /**< (x - y)^2 */ + POWER, /**< x ^ y */ + PRELU, /**< y*x if x < 0, x otherwise */ + }; + /** Set operation*/ + ElementwiseBinaryCommonAttributes &operation(const ElementwiseBinaryCommonAttributes::ElementwiseOp &operation); + /** Get operation*/ + ElementwiseOp operation() const; + +private: + ElementwiseOp _operation; /**< Elementwise operation */ +}; + +/** Forward declaration */ +class GpuWorkloadContext; +class GpuWorkloadSketch; + +/** Operator interface. */ +class GpuElementwiseBinaryCommon 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 are checked at the operator level i.e. GpuAdd::validate_op(), GpuSub::validate_op(), ... etc. + * + * Valid data layouts: + * - Any + * + * @param[in,out] sketch Workload sketch into which the operator will be fused + * @param[in] lhs Left hand side tensor info. Data types supported: U8/S16/S32/F16/F32. + * @param[in] rhs Right hand side tensor info. Data types supported: U8/S16/S32/F16/F32. + * @param[out] dst Destination tensor info. Data types supported: U8/S16/S32/F16/F32. If an uninitialized ITensorInfo is passed in, it will be auto-initialized + * @param[in] attributes ElementwiseBinaryCommonAttributes containing the operator type: ADD, SUB, DIV, ... etc. + */ + static void create_op(GpuWorkloadSketch &sketch, + ITensorInfo *lhs, + ITensorInfo *rhs, + ITensorInfo *dst, + const ElementwiseBinaryCommonAttributes &attributes); + /** Check if the operator configuration is supported, irrespective of fusion + * Similar to @ref GpuElementwiseBinaryCommon::create_op() + * + * @param[in] context Workload context within which the operator is running + * @param[in] lhs Left hand side tensor info. Data types supported: U8/S16/S32/F16/F32. + * @param[in] rhs Right hand side tensor info. Data types supported: U8/S16/S32/F16/F32. + * @param[in] dst Destination tensor info. Data types supported: U8/S16/S32/F16/F32. If an uninitialized ITensorInfo is passed in, it will be auto-initialized + * @param[in] attributes ElementwiseBinaryCommonAttributes containing the operator type: ADD, SUB, DIV, ... etc. + */ + static Status is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *lhs, + const ITensorInfo *rhs, + const ITensorInfo *dst, + const ElementwiseBinaryCommonAttributes &attributes); + /** Validate the operator and check if it can be fused into the workload sketch. + * Similar to @ref GpuElementwiseBinaryCommon::create_op() + */ + static Status validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *rhs, + const ITensorInfo *lhs, + const ITensorInfo *dst, + const ElementwiseBinaryCommonAttributes &attributes); +}; +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute +#endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_INTERNAL_GPUELEMENTWISEBINARYCOMMON */ |