From 1009e87127e08a58561373c3f83f22b5f0922641 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Mon, 27 Jul 2020 12:48:34 +0100 Subject: COMPMID-3385: Async support to CLArithmetic* kernels/functions Pt.2 Signed-off-by: Michalis Spyrou Change-Id: Idc5ac2dd2ba5295c00c88b44a783645327a27e15 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3617 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas Comments-Addressed: Arm Jenkins --- .../CL/functions/CLPixelWiseMultiplication.cpp | 152 +++++++++++++++++++-- 1 file changed, 139 insertions(+), 13 deletions(-) (limited to 'src/runtime/CL/functions/CLPixelWiseMultiplication.cpp') diff --git a/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp b/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp index 139209c4e0..34e06a3d03 100644 --- a/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp +++ b/src/runtime/CL/functions/CLPixelWiseMultiplication.cpp @@ -25,30 +25,50 @@ #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h" +#include "arm_compute/runtime/CL/CLScheduler.h" #include "support/MemorySupport.h" #include namespace arm_compute { -void CLPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale, - ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) +namespace { - configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, scale, overflow_policy, rounding_policy, act_info); +void select_border_input(InputTensorMap &tensor_map, InputTensorMap &inputs, OutputTensorMap &outputs) +{ + if(outputs.at(TensorType::ACL_DST)->info()->dimension(0) > 1) + { + if(inputs.at(TensorType::ACL_SRC_1)->info()->dimension(0) == 1) + { + tensor_map[TensorType::ACL_SRC] = inputs.at(TensorType::ACL_SRC_1); + } + else + { + tensor_map[TensorType::ACL_SRC] = inputs.at(TensorType::ACL_SRC_0); + } + } } +} // namespace -void CLPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale, +namespace experimental +{ +CLPixelWiseMultiplication::CLPixelWiseMultiplication() + : _border_handler() +{ +} + +void CLPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) { auto k = arm_compute::support::cpp14::make_unique(); k->configure(compile_context, input1, input2, output, scale, overflow_policy, rounding_policy, act_info); _kernel = std::move(k); - if(output->info()->dimension(0) > 1) + if(output->dimension(0) > 1) { - ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2; + ITensorInfo *broadcasted_info = (input1->dimension(0) == 1) ? input1 : input2; - if(broadcasted_info->info()->dimension(0) == 1) + if(broadcasted_info->dimension(0) == 1) { _border_handler.configure(compile_context, broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE); } @@ -61,22 +81,30 @@ Status CLPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITen return CLPixelWiseMultiplicationKernel::validate(input1, input2, output, scale, overflow_policy, rounding_policy, act_info); } -void CLComplexPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +void CLPixelWiseMultiplication::run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) { - configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, act_info); + InputTensorMap src; + select_border_input(src, inputs, outputs); + CLScheduler::get().enqueue_op(_border_handler, src, {}); + ICLOperator::run(inputs, outputs, workspace); } -void CLComplexPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +CLComplexPixelWiseMultiplication::CLComplexPixelWiseMultiplication() + : _border_handler() +{ +} + +void CLComplexPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info) { auto k = arm_compute::support::cpp14::make_unique(); k->configure(compile_context, input1, input2, output, act_info); _kernel = std::move(k); - if(output->info()->dimension(0) > 1) + if(output->dimension(0) > 1) { - ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2; + ITensorInfo *broadcasted_info = (input1->dimension(0) == 1) ? input1 : input2; - if(broadcasted_info->info()->dimension(0) == 1) + if(broadcasted_info->dimension(0) == 1) { _border_handler.configure(compile_context, broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE); } @@ -87,4 +115,102 @@ Status CLComplexPixelWiseMultiplication::validate(const ITensorInfo *input1, con { return CLComplexPixelWiseMultiplicationKernel::validate(input1, input2, output, act_info); } + +void CLComplexPixelWiseMultiplication::run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) +{ + InputTensorMap src; + select_border_input(src, inputs, outputs); + CLScheduler::get().enqueue_op(_border_handler, src, {}); + ICLOperator::run(inputs, outputs, workspace); +} +} // namespace experimental + +struct CLPixelWiseMultiplication::Impl +{ + const ICLTensor *src_0{ nullptr }; + const ICLTensor *src_1{ nullptr }; + ICLTensor *dst{ nullptr }; + std::unique_ptr op{ nullptr }; +}; + +CLPixelWiseMultiplication::CLPixelWiseMultiplication() + : _impl(support::cpp14::make_unique()) +{ +} +CLPixelWiseMultiplication::CLPixelWiseMultiplication(CLPixelWiseMultiplication &&) = default; +CLPixelWiseMultiplication &CLPixelWiseMultiplication::operator=(CLPixelWiseMultiplication &&) = default; +CLPixelWiseMultiplication::~CLPixelWiseMultiplication() = default; + +void CLPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale, + ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) +{ + configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, scale, overflow_policy, rounding_policy, act_info); +} + +void CLPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale, + ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) +{ + _impl->src_0 = input1; + _impl->src_1 = input2; + _impl->dst = output; + _impl->op = arm_compute::support::cpp14::make_unique(); + _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), scale, overflow_policy, rounding_policy, act_info); +} + +Status CLPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale, + ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info) +{ + return experimental::CLPixelWiseMultiplication::validate(input1, input2, output, scale, overflow_policy, rounding_policy, act_info); +} + +void CLPixelWiseMultiplication::run() +{ + const InputTensorMap src{ { TensorType::ACL_SRC_0, _impl->src_0 }, { TensorType::ACL_SRC_1, _impl->src_1 } }; + const OutputTensorMap dst{ { TensorType::ACL_DST, _impl->dst } }; + + _impl->op->run(src, dst, {}); +} + +struct CLComplexPixelWiseMultiplication::Impl +{ + const ICLTensor *src_0{ nullptr }; + const ICLTensor *src_1{ nullptr }; + ICLTensor *dst{ nullptr }; + std::unique_ptr op{ nullptr }; +}; + +CLComplexPixelWiseMultiplication::CLComplexPixelWiseMultiplication() + : _impl(support::cpp14::make_unique()) +{ +} +CLComplexPixelWiseMultiplication::CLComplexPixelWiseMultiplication(CLComplexPixelWiseMultiplication &&) = default; +CLComplexPixelWiseMultiplication &CLComplexPixelWiseMultiplication::operator=(CLComplexPixelWiseMultiplication &&) = default; +CLComplexPixelWiseMultiplication::~CLComplexPixelWiseMultiplication() = default; + +void CLComplexPixelWiseMultiplication::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +{ + configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, act_info); +} + +void CLComplexPixelWiseMultiplication::configure(const CLCompileContext &compile_context, ICLTensor *input1, ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info) +{ + _impl->src_0 = input1; + _impl->src_1 = input2; + _impl->dst = output; + _impl->op = arm_compute::support::cpp14::make_unique(); + _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), act_info); +} + +Status CLComplexPixelWiseMultiplication::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info) +{ + return experimental::CLComplexPixelWiseMultiplication::validate(input1, input2, output, act_info); +} + +void CLComplexPixelWiseMultiplication::run() +{ + const InputTensorMap src{ { TensorType::ACL_SRC_0, _impl->src_0 }, { TensorType::ACL_SRC_1, _impl->src_1 } }; + const OutputTensorMap dst{ { TensorType::ACL_DST, _impl->dst } }; + + _impl->op->run(src, dst, {}); +} } // namespace arm_compute -- cgit v1.2.1