From cc2877368d5e15d9ea89d31c84ec651fc0fffd13 Mon Sep 17 00:00:00 2001 From: Gunes Bayir Date: Thu, 19 Jan 2023 15:56:00 +0000 Subject: Change dynamic fusion API to return destination tensor info The new dynamic fusion API is introduced in the following patch: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8906 For each operator (except Conv2D, which is migrated in the above patch), we - remove destination tensor from is_supported, validate and create calls - make create_op return ITensorInfo* to the intermediate destination object Affected operators: - DepthwiseConv2D - Cast - Elementwise Ops - Clamp - Reshape - Resize Resolves: COMPMID-5777 Change-Id: Ib60ec8a5f081752808455d7a7d790f2ed0627059 Signed-off-by: Gunes Bayir Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8991 Reviewed-by: Ramy Elgammal Reviewed-by: Jakub Sujak Dynamic-Fusion: Ramy Elgammal Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- .../internal/GpuElementwiseBinaryCommon.cpp | 119 ++++++++++----------- 1 file changed, 58 insertions(+), 61 deletions(-) (limited to 'src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp') diff --git a/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp b/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp index aec22e100c..7c087c9a7b 100644 --- a/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp +++ b/src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022 Arm Limited. + * Copyright (c) 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.h" +#include "src/common/utils/Log.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/dynamic_fusion/sketch/ArgumentPack.h" #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" @@ -43,30 +44,23 @@ void calculate_and_init_dst_if_empty(ITensorInfo *dst, const ITensorInfo *lhs, c 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) +Status is_supported_op_helper(const GpuWorkloadContext &context, + const ITensorInfo *lhs, + const ITensorInfo *rhs, + const ITensorInfo *dst, + const ElementwiseBinaryCommonAttributes &attributes) { - _operation = operation; - return *this; -} + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs); -ElementwiseBinaryCommonAttributes::ElementwiseOp ElementwiseBinaryCommonAttributes::operation() const -{ - return _operation; -} + TensorInfo dst_info_to_validate; + const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate; -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); + if(dst != nullptr) + { + dst_info_to_validate_ptr = 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 @@ -80,16 +74,8 @@ Status GpuElementwiseBinaryCommon::is_supported_op(const GpuWorkloadContext ArgumentPack arguments; arguments.add_const_tensor(ACL_SRC_0, lhs); arguments.add_const_tensor(ACL_SRC_1, rhs); + arguments.add_const_tensor(ACL_DST_0, dst_info_to_validate_ptr); - // 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)); } } @@ -101,18 +87,40 @@ Status GpuElementwiseBinaryCommon::is_supported_op(const GpuWorkloadContext return Status{}; } +GpuOperatorType operator_type = GpuOperatorType::Simple; +} // namespace + +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 ElementwiseBinaryCommonAttributes &attributes) +{ + return is_supported_op_helper(context, lhs, rhs, nullptr, attributes); +} + 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()); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs); + ARM_COMPUTE_RETURN_ERROR_ON(!lhs->has_valid_id() || !rhs->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 - TensorInfo dst_info_to_validate = *dst; calculate_and_init_dst_if_empty(&dst_info_to_validate, lhs, rhs); // Perform fusion test @@ -125,20 +133,21 @@ Status GpuElementwiseBinaryCommon::validate_op(const GpuWorkloadSketch 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); + // Check if configuration is supported + return is_supported_op_helper(*sketch.gpu_context(), lhs, rhs, &dst_info_to_validate, attributes); } -void GpuElementwiseBinaryCommon::create_op(GpuWorkloadSketch &sketch, - ITensorInfo *lhs, - ITensorInfo *rhs, - ITensorInfo *dst, - const ElementwiseBinaryCommonAttributes &attributes) +ITensorInfo *GpuElementwiseBinaryCommon::create_op(GpuWorkloadSketch &sketch, + ITensorInfo *lhs, + ITensorInfo *rhs, + 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{}; + ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs); + ARM_COMPUTE_LOG_PARAMS(lhs, rhs); + ARM_COMPUTE_ERROR_THROW_ON(GpuElementwiseBinaryCommon::validate_op(sketch, lhs, rhs, attributes)); + + ITensorInfo *dst = sketch.implementation().create_virtual_tensor(); + ARM_COMPUTE_ERROR_ON_NULLPTR(dst); // Auto initialize dst tensor calculate_and_init_dst_if_empty(dst, lhs, rhs); @@ -160,14 +169,7 @@ void GpuElementwiseBinaryCommon::create_op(GpuWorkloadSketch ArgumentPack 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); - } + arguments.add_const_tensor(ACL_DST_0, dst); comp_graph.add_new_component(properties, arguments, attributes); } } @@ -183,16 +185,11 @@ void GpuElementwiseBinaryCommon::create_op(GpuWorkloadSketch ArgumentPack 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); - } + 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); + + return dst; } } // namespace dynamic_fusion -- cgit v1.2.1