diff options
Diffstat (limited to 'src/runtime/NEON/functions/NERNNLayer.cpp')
| -rw-r--r-- | src/runtime/NEON/functions/NERNNLayer.cpp | 66 |
1 files changed, 39 insertions, 27 deletions
diff --git a/src/runtime/NEON/functions/NERNNLayer.cpp b/src/runtime/NEON/functions/NERNNLayer.cpp index a8e10482a7..2824693800 100644 --- a/src/runtime/NEON/functions/NERNNLayer.cpp +++ b/src/runtime/NEON/functions/NERNNLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -27,35 +27,37 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" -#include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/Validate.h" #include "arm_compute/runtime/NEON/NEScheduler.h" -#include "src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h" -#include "src/core/NEON/kernels/NEConvertQuantizedSignednessKernel.h" -#include "src/core/NEON/kernels/NECopyKernel.h" -#include "src/core/NEON/kernels/NEFlattenLayerKernel.h" -#include "src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" -#include "src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h" -#include "src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h" -#include "src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h" -#include "src/core/NEON/kernels/NEGEMMLowpReductionKernel.h" -#include "src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h" -#include "src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h" -#include "src/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" -#include "support/MemorySupport.h" + +#include "src/common/utils/Log.h" namespace arm_compute { NERNNLayer::~NERNNLayer() = default; NERNNLayer::NERNNLayer(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(std::move(memory_manager)), _gemm_state_f(), _add_f(), _activation(), _fully_connected(memory_manager), _copy_kernel(), _fully_connected_out(), _gemm_output(), _add_output(), + : _memory_group(std::move(memory_manager)), + _gemm_state_f(), + _add_f(), + _activation(), + _fully_connected(memory_manager), + _copy_f(), + _fully_connected_out(), + _gemm_output(), + _add_output(), _is_prepared(false) { } -Status NERNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *recurrent_weights, const ITensorInfo *bias, const ITensorInfo *hidden_state, - const ITensorInfo *output, const ActivationLayerInfo &info) +Status NERNNLayer::validate(const ITensorInfo *input, + const ITensorInfo *weights, + const ITensorInfo *recurrent_weights, + const ITensorInfo *bias, + const ITensorInfo *hidden_state, + const ITensorInfo *output, + const ActivationLayerInfo &info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::F16, DataType::F32); @@ -72,23 +74,34 @@ Status NERNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_height) != input->dimension(idx_height)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), hidden_state->tensor_shape()); - auto shape_info = TensorInfo(misc::shape_calculator::compute_rnn_shape(recurrent_weights, hidden_state->dimension(idx_height)), 1, input->data_type()); + auto shape_info = + TensorInfo(misc::shape_calculator::compute_rnn_shape(recurrent_weights, hidden_state->dimension(idx_height)), 1, + input->data_type()); ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, weights, bias, &shape_info)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEArithmeticAddition::validate(&shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE)); ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&shape_info, &shape_info, info)); return Status{}; } -void NERNNLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *recurrent_weights, const ITensor *bias, ITensor *hidden_state, ITensor *output, +void NERNNLayer::configure(const ITensor *input, + const ITensor *weights, + const ITensor *recurrent_weights, + const ITensor *bias, + ITensor *hidden_state, + ITensor *output, ActivationLayerInfo &info) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output); - ARM_COMPUTE_ERROR_THROW_ON(NERNNLayer::validate(input->info(), weights->info(), recurrent_weights->info(), bias->info(), hidden_state->info(), output->info(), info)); + ARM_COMPUTE_ERROR_THROW_ON(NERNNLayer::validate(input->info(), weights->info(), recurrent_weights->info(), + bias->info(), hidden_state->info(), output->info(), info)); + ARM_COMPUTE_LOG_PARAMS(input, weights, recurrent_weights, bias, hidden_state, output, info); const int idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT); - TensorShape shape = misc::shape_calculator::compute_rnn_shape(recurrent_weights->info(), hidden_state->info()->dimension(idx_height)); + TensorShape shape = misc::shape_calculator::compute_rnn_shape(recurrent_weights->info(), + hidden_state->info()->dimension(idx_height)); _is_prepared = false; @@ -114,8 +127,7 @@ void NERNNLayer::configure(const ITensor *input, const ITensor *weights, const I _activation.configure(&_add_output, hidden_state, info); _add_output.allocator()->allocate(); - _copy_kernel = arm_compute::support::cpp14::make_unique<NECopyKernel>(); - _copy_kernel->configure(hidden_state, output); + _copy_f.configure(hidden_state, output); } void NERNNLayer::run() @@ -132,12 +144,12 @@ void NERNNLayer::run() _activation.run(); // copy hidden out to output - NEScheduler::get().schedule(_copy_kernel.get(), Window::DimY); + _copy_f.run(); } void NERNNLayer::prepare() { - if(!_is_prepared) + if (!_is_prepared) { _fully_connected.prepare(); _gemm_state_f.prepare(); |