From 4667dddc0ed403c636348294cd7f70261e5540cf Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 13 Jul 2020 21:21:33 +0100 Subject: COMPMID-3374: Remove memory state from NEConcatenateLayer kernels * Allow the following kernels to accept backing memory at run-time: * NEBatchConcatenateLayerKernel * NEDepthConcatenateLayerKernel * NEHeightConcatenateLayerKernel * NEWidthConcatenateLayerKernel * Allow the following functions to accept backing memory at run-time: * NEConcatenateLayer Signed-off-by: Georgios Pinitas Change-Id: Ib0b6714cff7f06a52dc74d294bc3e0d72a1c2419 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3569 Tested-by: Arm Jenkins Reviewed-by: Michalis Spyrou Comments-Addressed: Arm Jenkins --- src/runtime/NEON/functions/NEConcatenateLayer.cpp | 131 +++++++++++++++------- src/runtime/NEON/functions/NELSTMLayer.cpp | 4 +- src/runtime/NEON/functions/NEPadLayer.cpp | 2 +- 3 files changed, 93 insertions(+), 44 deletions(-) (limited to 'src/runtime/NEON') diff --git a/src/runtime/NEON/functions/NEConcatenateLayer.cpp b/src/runtime/NEON/functions/NEConcatenateLayer.cpp index 9c480a0d50..37cdd15529 100644 --- a/src/runtime/NEON/functions/NEConcatenateLayer.cpp +++ b/src/runtime/NEON/functions/NEConcatenateLayer.cpp @@ -39,58 +39,31 @@ namespace arm_compute { -NEConcatenateLayer::NEConcatenateLayer() - : _concat_kernels(), - _num_inputs(0), - _axis(Window::DimX) -{ -} - -void NEConcatenateLayer::configure(std::vector inputs_vector, ITensor *output, size_t axis) -{ - configure_internal(std::move(inputs_vector), output, axis); -} - -void NEConcatenateLayer::configure(std::vector inputs_vector, ITensor *output, size_t axis) +namespace experimental { - configure_internal(std::move(inputs_vector), output, axis); -} - -Status NEConcatenateLayer::validate(const std::vector &inputs_vector, const ITensorInfo *output, size_t axis) -{ - return validate_internal(inputs_vector, output, axis); -} - -Status NEConcatenateLayer::validate(const std::vector &inputs_vector, const ITensorInfo *output, size_t axis) +NEConcatenateLayer::NEConcatenateLayer() + : _concat_kernels(), _num_inputs(0), _axis(0) { - return validate_internal(inputs_vector, output, axis); } -template -void NEConcatenateLayer::configure_internal(std::vector &&inputs_vector, ITensor *output, size_t axis) +void NEConcatenateLayer::configure(const std::vector &inputs_vector, ITensorInfo *output, size_t axis) { ARM_COMPUTE_ERROR_ON(output == nullptr); + _axis = axis; _num_inputs = inputs_vector.size(); - std::vector inputs_vector_info; - inputs_vector_info.reserve(_num_inputs); - for(unsigned int i = 0; i < _num_inputs; ++i) - { - ARM_COMPUTE_ERROR_ON_NULLPTR(inputs_vector.at(i)); - inputs_vector_info.emplace_back(inputs_vector.at(i)->info()); - } - TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, _axis); + TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, axis); // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type()); - ARM_COMPUTE_ERROR_THROW_ON(NEConcatenateLayer::validate(inputs_vector_info, output->info(), axis)); + auto_init_if_empty(*output, output_shape, 1, inputs_vector[0]->data_type()); + ARM_COMPUTE_ERROR_THROW_ON(NEConcatenateLayer::validate(inputs_vector, output, axis)); unsigned int offset = 0; for(unsigned int i = 0; i < _num_inputs; ++i) { - switch(_axis) + switch(axis) { case Window::DimX: { @@ -123,12 +96,11 @@ void NEConcatenateLayer::configure_internal(std::vector &&inputs_v default: ARM_COMPUTE_ERROR("Axis not supported"); } - offset += inputs_vector.at(i)->info()->dimension(_axis); + offset += inputs_vector.at(i)->dimension(axis); } } -template -Status NEConcatenateLayer::validate_internal(const std::vector &inputs_vector, const ITensorInfo *output, size_t axis) +Status NEConcatenateLayer::validate(const std::vector &inputs_vector, const ITensorInfo *output, size_t axis) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); ARM_COMPUTE_RETURN_ERROR_ON(inputs_vector.size() < 2); @@ -174,11 +146,88 @@ Status NEConcatenateLayer::validate_internal(const std::vector return Status{}; } +MemoryRequirements NEConcatenateLayer::workspace() const +{ + return MemoryRequirements{}; +} + +void NEConcatenateLayer::run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) +{ + ARM_COMPUTE_UNUSED(workspace); + + if(inputs.empty() || outputs.empty()) + { + ARM_COMPUTE_ERROR("No inputs provided"); + } + + if(inputs.size() != _num_inputs) + { + ARM_COMPUTE_ERROR("Configured with different number of inputs"); + } + + int i = 0; + for(auto &k : _concat_kernels) + { + const InputTensorMap input = { { TensorType::ACL_SRC, inputs.at(ACL_SRC_VEC + i) } }; + NEScheduler::get().schedule_op(k.get(), Window::DimY, input, outputs); + ++i; + } +} +} // namespace experimental + +struct NEConcatenateLayer::Impl +{ + std::vector srcs{}; + ITensor *dst{ nullptr }; + unsigned int num_inputs{ 0 }; + unsigned int axis{ 0 }; + std::unique_ptr op{ nullptr }; +}; + +NEConcatenateLayer::NEConcatenateLayer() + : _impl(support::cpp14::make_unique()) +{ +} + +NEConcatenateLayer::NEConcatenateLayer(NEConcatenateLayer &&) = default; + +NEConcatenateLayer &NEConcatenateLayer::operator=(NEConcatenateLayer &&) = default; + +NEConcatenateLayer::~NEConcatenateLayer() = default; + +void NEConcatenateLayer::configure(std::vector inputs_vector, ITensor *output, size_t axis) +{ + ARM_COMPUTE_ERROR_ON(output == nullptr); + + _impl->srcs = inputs_vector; + _impl->dst = output; + _impl->axis = axis; + _impl->num_inputs = inputs_vector.size(); + _impl->op = arm_compute::support::cpp14::make_unique(); + + std::vector inputs_vector_info; + for(unsigned int i = 0; i < inputs_vector.size(); ++i) + { + ARM_COMPUTE_ERROR_ON_NULLPTR(inputs_vector.at(i)); + inputs_vector_info.emplace_back(inputs_vector.at(i)->info()); + } + _impl->op->configure(inputs_vector_info, _impl->dst->info(), axis); +} + +Status NEConcatenateLayer::validate(const std::vector &inputs_vector, const ITensorInfo *output, size_t axis) +{ + return experimental::NEConcatenateLayer::validate(inputs_vector, output, axis); +} + void NEConcatenateLayer::run() { - for(auto &kernel : _concat_kernels) + InputTensorMap srcs; + for(unsigned i = 0; i < _impl->num_inputs; ++i) { - NEScheduler::get().schedule(kernel.get(), Window::DimY); + srcs.insert(std::make_pair(TensorType::ACL_SRC_VEC + i, _impl->srcs.at(i))); } + const OutputTensorMap dst{ { TensorType::ACL_DST, _impl->dst } }; + + _impl->op->run(srcs, dst, {}); } } // namespace arm_compute diff --git a/src/runtime/NEON/functions/NELSTMLayer.cpp b/src/runtime/NEON/functions/NELSTMLayer.cpp index f89b3e999c..dca274acd2 100644 --- a/src/runtime/NEON/functions/NELSTMLayer.cpp +++ b/src/runtime/NEON/functions/NELSTMLayer.cpp @@ -347,7 +347,7 @@ void NELSTMLayer::configure(const ITensor *input, _copy_output.configure(output_state_out, output); // Vector for holding the tensors to store in scratch buffer - std::vector scratch_inputs; + std::vector scratch_inputs; if(!lstm_params.has_cifg_opt()) { scratch_inputs.emplace_back(input_gate_out); @@ -579,7 +579,7 @@ Status NELSTMLayer::validate(const ITensorInfo *input, ARM_COMPUTE_RETURN_ON_ERROR(NECopyKernel::validate(output_state_out, output)); // Validate scratch concatenation - std::vector inputs_vector_info_raw; + std::vector inputs_vector_info_raw; if(!lstm_params.has_cifg_opt()) { inputs_vector_info_raw.push_back(&input_gate); diff --git a/src/runtime/NEON/functions/NEPadLayer.cpp b/src/runtime/NEON/functions/NEPadLayer.cpp index da9a425d9b..21c349ba95 100644 --- a/src/runtime/NEON/functions/NEPadLayer.cpp +++ b/src/runtime/NEON/functions/NEPadLayer.cpp @@ -117,7 +117,7 @@ void NEPadLayer::configure_reflect_symmetric_mode(ITensor *input, ITensor *outpu const int32_t end_mask_after = ends_after[i] < 0 ? ~0 : ~(1u << i); // Reflect the input values for the padding before and after the input. - std::vector concat_vector; + std::vector concat_vector; if(_padding[i].first > 0) { if(i < prev->info()->num_dimensions()) -- cgit v1.2.1