From 2aad21a900a21f467b3ec6b37420f892f0d80221 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Thu, 2 Jul 2020 12:43:53 +0100 Subject: COMPMID-3388: Async support to CLReshapeLayerKernel kernels/functions Signed-off-by: Michalis Spyrou Change-Id: I141a943dfd691069317860e852ecdd0ba7391604 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3501 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas Comments-Addressed: Arm Jenkins --- src/runtime/CL/functions/CLArgMinMaxLayer.cpp | 8 +-- .../CL/functions/CLGenerateProposalsLayer.cpp | 20 ++++---- src/runtime/CL/functions/CLReductionOperation.cpp | 8 +-- src/runtime/CL/functions/CLReshapeLayer.cpp | 59 ++++++++++++++++++++-- src/runtime/CL/functions/CLSoftmaxLayer.cpp | 24 ++++----- 5 files changed, 84 insertions(+), 35 deletions(-) (limited to 'src/runtime/CL/functions') diff --git a/src/runtime/CL/functions/CLArgMinMaxLayer.cpp b/src/runtime/CL/functions/CLArgMinMaxLayer.cpp index cb2b290adf..8fcd04f2fa 100644 --- a/src/runtime/CL/functions/CLArgMinMaxLayer.cpp +++ b/src/runtime/CL/functions/CLArgMinMaxLayer.cpp @@ -35,7 +35,7 @@ namespace arm_compute { CLArgMinMaxLayer::CLArgMinMaxLayer(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _results_vector(), _not_reshaped_output(), _reduction_kernels_vector(), _reshape_kernel(), _num_of_stages(), _reduction_axis() + : _memory_group(std::move(memory_manager)), _results_vector(), _not_reshaped_output(), _reduction_kernels_vector(), _reshape(), _num_of_stages(), _reduction_axis() { } @@ -103,7 +103,7 @@ Status CLArgMinMaxLayer::validate(const ITensorInfo *input, int axis, const ITen const unsigned int last_stage = num_of_stages - 1; ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[last_stage - 1], ¬_reshaped_output, axis, op)); } - ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(¬_reshaped_output, output)); + ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(¬_reshaped_output, output)); return Status{}; } @@ -158,7 +158,7 @@ void CLArgMinMaxLayer::configure(const CLCompileContext &compile_context, const _reduction_kernels_vector[last_stage].configure(compile_context, input, &_results_vector[last_stage - 1], &_not_reshaped_output, axis, op); _results_vector[last_stage - 1].allocator()->allocate(); } - _reshape_kernel.configure(compile_context, &_not_reshaped_output, output); + _reshape.configure(compile_context, &_not_reshaped_output, output); _not_reshaped_output.allocator()->allocate(); } @@ -170,6 +170,6 @@ void CLArgMinMaxLayer::run() { CLScheduler::get().enqueue(_reduction_kernels_vector[i], false); } - CLScheduler::get().enqueue(_reshape_kernel, false); + _reshape.run(); } } // namespace arm_compute \ No newline at end of file diff --git a/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp b/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp index 7f037fc51f..1b89bb4cfe 100644 --- a/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp +++ b/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp @@ -31,9 +31,9 @@ namespace arm_compute CLGenerateProposalsLayer::CLGenerateProposalsLayer(std::shared_ptr memory_manager) : _memory_group(memory_manager), _permute_deltas_kernel(), - _flatten_deltas_kernel(), + _flatten_deltas(), _permute_scores_kernel(), - _flatten_scores_kernel(), + _flatten_scores(), _compute_anchors_kernel(), _bounding_box_kernel(), _pad_kernel(), @@ -102,12 +102,12 @@ void CLGenerateProposalsLayer::configure(const CLCompileContext &compile_context { _memory_group.manage(&_deltas_permuted); _permute_deltas_kernel.configure(compile_context, deltas, &_deltas_permuted, PermutationVector{ 2, 0, 1 }); - _flatten_deltas_kernel.configure(compile_context, &_deltas_permuted, &_deltas_flattened); + _flatten_deltas.configure(compile_context, &_deltas_permuted, &_deltas_flattened); _deltas_permuted.allocator()->allocate(); } else { - _flatten_deltas_kernel.configure(compile_context, deltas, &_deltas_flattened); + _flatten_deltas.configure(compile_context, deltas, &_deltas_flattened); } const TensorShape flatten_shape_scores(1, total_num_anchors); @@ -119,12 +119,12 @@ void CLGenerateProposalsLayer::configure(const CLCompileContext &compile_context { _memory_group.manage(&_scores_permuted); _permute_scores_kernel.configure(compile_context, scores, &_scores_permuted, PermutationVector{ 2, 0, 1 }); - _flatten_scores_kernel.configure(compile_context, &_scores_permuted, &_scores_flattened); + _flatten_scores.configure(compile_context, &_scores_permuted, &_scores_flattened); _scores_permuted.allocator()->allocate(); } else { - _flatten_scores_kernel.configure(compile_context, scores, &_scores_flattened); + _flatten_scores.configure(compile_context, scores, &_scores_flattened); } CLTensor *anchors_to_use = &_all_anchors; @@ -240,12 +240,12 @@ Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITens } TensorInfo deltas_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true)); - ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&deltas_permuted_info, &deltas_flattened_info)); + ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(&deltas_permuted_info, &deltas_flattened_info)); TensorInfo scores_flattened_info(scores->clone()->set_tensor_shape(TensorShape(1, total_num_anchors)).set_is_resizable(true)); TensorInfo proposals_4_roi_values(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true)); - ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&scores_permuted_info, &scores_flattened_info)); + ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(&scores_permuted_info, &scores_flattened_info)); TensorInfo *proposals_4_roi_values_to_use = &proposals_4_roi_values; TensorInfo proposals_4_roi_values_quantized(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true)); @@ -350,8 +350,8 @@ void CLGenerateProposalsLayer::run() CLScheduler::get().enqueue(_permute_deltas_kernel, false); CLScheduler::get().enqueue(_permute_scores_kernel, false); } - CLScheduler::get().enqueue(_flatten_deltas_kernel, false); - CLScheduler::get().enqueue(_flatten_scores_kernel, false); + _flatten_deltas.run(); + _flatten_scores.run(); if(_is_qasymm8) { diff --git a/src/runtime/CL/functions/CLReductionOperation.cpp b/src/runtime/CL/functions/CLReductionOperation.cpp index b659ecfaf6..2d7db0a20a 100644 --- a/src/runtime/CL/functions/CLReductionOperation.cpp +++ b/src/runtime/CL/functions/CLReductionOperation.cpp @@ -39,7 +39,7 @@ namespace arm_compute { CLReductionOperation::CLReductionOperation(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _results_vector(), _reduction_kernels_vector(), _border_handlers_vector(), _reshape_kernel(), _op(), _num_of_stages(), _reduction_axis(), _is_serial(), + : _memory_group(std::move(memory_manager)), _results_vector(), _reduction_kernels_vector(), _border_handlers_vector(), _reshape(), _op(), _num_of_stages(), _reduction_axis(), _is_serial(), _is_reshape_required(false) { } @@ -152,7 +152,7 @@ Status CLReductionOperation::validate(const ITensorInfo *input, const ITensorInf if(is_reshape_required) { - ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(output_internal, output)); + ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(output_internal, output)); } return Status{}; @@ -351,7 +351,7 @@ void CLReductionOperation::configure(const CLCompileContext &compile_context, IC if(_is_reshape_required) { - _reshape_kernel.configure(compile_context, &_results_vector.back(), output); + _reshape.configure(compile_context, &_results_vector.back(), output); _results_vector.back().allocator()->allocate(); } } @@ -375,7 +375,7 @@ void CLReductionOperation::run() if(_is_reshape_required) { - CLScheduler::get().enqueue(_reshape_kernel, false); + _reshape.run(); } } } // namespace arm_compute diff --git a/src/runtime/CL/functions/CLReshapeLayer.cpp b/src/runtime/CL/functions/CLReshapeLayer.cpp index 13baedb3f9..6fc8608552 100644 --- a/src/runtime/CL/functions/CLReshapeLayer.cpp +++ b/src/runtime/CL/functions/CLReshapeLayer.cpp @@ -28,7 +28,43 @@ #include "support/MemorySupport.h" /** [CLReshapeLayer snippet] **/ -using namespace arm_compute; +namespace arm_compute +{ +namespace experimental +{ +void CLReshapeLayer::configure(const CLCompileContext &compile_context, const ITensorInfo *input, ITensorInfo *output) +{ + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(compile_context, input, output); + _kernel = std::move(k); +} + +Status CLReshapeLayer::validate(const ITensorInfo *input, const ITensorInfo *output) +{ + return arm_compute::CLReshapeLayerKernel::validate(input, output); +} + +MemoryRequirements CLReshapeLayer::workspace() const +{ + return MemoryRequirements{}; +} +} // namespace experimental + +struct CLReshapeLayer::Impl +{ + const ICLTensor *src{ nullptr }; + ICLTensor *dst{ nullptr }; + std::unique_ptr op{ nullptr }; +}; + +CLReshapeLayer::CLReshapeLayer() + : _impl(support::cpp14::make_unique()) +{ +} + +CLReshapeLayer::CLReshapeLayer(CLReshapeLayer &&) = default; +CLReshapeLayer &CLReshapeLayer::operator=(CLReshapeLayer &&) = default; +CLReshapeLayer::~CLReshapeLayer() = default; void CLReshapeLayer::configure(const ICLTensor *input, ICLTensor *output) { @@ -37,13 +73,26 @@ void CLReshapeLayer::configure(const ICLTensor *input, ICLTensor *output) void CLReshapeLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output) { - auto k = arm_compute::support::cpp14::make_unique(); - k->configure(compile_context, input, output); - _kernel = std::move(k); + _impl->src = input; + _impl->dst = output; + _impl->op = arm_compute::support::cpp14::make_unique(); + _impl->op->configure(compile_context, input->info(), output->info()); } Status CLReshapeLayer::validate(const ITensorInfo *input, const ITensorInfo *output) { - return CLReshapeLayerKernel::validate(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ON_ERROR(experimental::CLReshapeLayer::validate(input, output)); + + return Status{}; +} + +void CLReshapeLayer::run() +{ + const InputTensorMap src{ { TensorType::ACL_SRC, _impl->src } }; + const OutputTensorMap dst{ { TensorType::ACL_DST, _impl->dst } }; + + _impl->op->run(src, dst, {}); } +} // namespace arm_compute /** [CLReshapeLayer snippet] **/ diff --git a/src/runtime/CL/functions/CLSoftmaxLayer.cpp b/src/runtime/CL/functions/CLSoftmaxLayer.cpp index 71ccf9fa01..52fac4f846 100644 --- a/src/runtime/CL/functions/CLSoftmaxLayer.cpp +++ b/src/runtime/CL/functions/CLSoftmaxLayer.cpp @@ -36,7 +36,7 @@ namespace arm_compute { template CLSoftmaxLayerGeneric::CLSoftmaxLayerGeneric(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _max_shift_exp_sum_kernel(), _norm_kernel(), _flatten_kernel_ptr(), _reshape_kernel(), _max(), _sum(), _tmp(), _input_flattened(), _output_flattened(), + : _memory_group(std::move(memory_manager)), _max_shift_exp_sum_kernel(), _norm_kernel(), _flatten_ptr(), _reshape(), _max(), _sum(), _tmp(), _input_flattened(), _output_flattened(), _needs_flattening(false) { } @@ -64,15 +64,15 @@ void CLSoftmaxLayerGeneric::configure_reshape_input_kernel(const CLCompi // 2. first_n_reduce_axes == 4: Reduce all 4 dimensions. This can only be handled by CLReshapeKernel instead of CLFlattenKernel. if(first_n_reduce_axes == 3) { - auto flatten_kernel_ptr = support::cpp14::make_unique(); - flatten_kernel_ptr->configure(compile_context, input, &_input_flattened); - _flatten_kernel_ptr = std::move(flatten_kernel_ptr); + auto flatten = support::cpp14::make_unique(); + flatten->configure(compile_context, input, &_input_flattened); + _flatten_ptr = std::move(flatten); } else { - auto reshape_kernel_ptr = support::cpp14::make_unique(); - reshape_kernel_ptr->configure(compile_context, input, &_input_flattened); - _flatten_kernel_ptr = std::move(reshape_kernel_ptr); + auto reshape_ptr = support::cpp14::make_unique(); + reshape_ptr->configure(compile_context, input, &_input_flattened); + _flatten_ptr = std::move(reshape_ptr); } // We need to init the output tensor here. Indeed, the reshape kernel expects @@ -152,7 +152,7 @@ void CLSoftmaxLayerGeneric::configure(const CLCompileContext &compile_co _norm_kernel.configure(compile_context, &_tmp, &_sum, &_output_flattened, softmax_info); // Reshape the flat output into a the requested (4D) output - _reshape_kernel.configure(compile_context, &_output_flattened, output); + _reshape.configure(compile_context, &_output_flattened, output); // Allocate the intermediate flat tensors _input_flattened.allocator()->allocate(); @@ -199,11 +199,11 @@ Status CLSoftmaxLayerGeneric::validate(const ITensorInfo *input, const I if(first_n_reduce_axes == 3) { - ARM_COMPUTE_RETURN_ON_ERROR(CLFlattenLayerKernel::validate(input, &tensor_info_flat)); + ARM_COMPUTE_RETURN_ON_ERROR(CLFlattenLayer::validate(input, &tensor_info_flat)); } else { - ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(input, &tensor_info_flat)); + ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(input, &tensor_info_flat)); } } @@ -231,7 +231,7 @@ void CLSoftmaxLayerGeneric::run() if(_needs_flattening) { - CLScheduler::get().enqueue(*_flatten_kernel_ptr, false); + _flatten_ptr->run(); } CLScheduler::get().enqueue(_max_shift_exp_sum_kernel, false); @@ -239,7 +239,7 @@ void CLSoftmaxLayerGeneric::run() if(_needs_flattening) { - CLScheduler::get().enqueue(_reshape_kernel, true); + _reshape.run(); } } -- cgit v1.2.1