diff options
author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-02-03 16:05:00 +0000 |
---|---|---|
committer | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-02-08 09:46:35 +0000 |
commit | 1928904316e80ba0549b94ae1f905d7e79bda812 (patch) | |
tree | ac44d4118f2beb6c6b454995abaeb76228ab54ab /src/runtime/NEON/functions/NEPoolingLayer.cpp | |
parent | dda6914c6e923187c2ca2c3bfd71677e9c9e5c68 (diff) | |
download | ComputeLibrary-1928904316e80ba0549b94ae1f905d7e79bda812.tar.gz |
Make NEON Pooling kernels and functions state-less
Partially resolves COMPMID-3999
Change-Id: Ib39d40694df5c5f0a9401488e0c3af3ac26e8c55
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4984
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/runtime/NEON/functions/NEPoolingLayer.cpp')
-rw-r--r-- | src/runtime/NEON/functions/NEPoolingLayer.cpp | 103 |
1 files changed, 24 insertions, 79 deletions
diff --git a/src/runtime/NEON/functions/NEPoolingLayer.cpp b/src/runtime/NEON/functions/NEPoolingLayer.cpp index 0c857b54dc..dd7a3a337e 100644 --- a/src/runtime/NEON/functions/NEPoolingLayer.cpp +++ b/src/runtime/NEON/functions/NEPoolingLayer.cpp @@ -23,103 +23,48 @@ */ #include "arm_compute/runtime/NEON/functions/NEPoolingLayer.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/runtime/NEON/NEScheduler.h" -#include "src/core/NEON/kernels/NEFillBorderKernel.h" -#include "src/core/NEON/kernels/NEPoolingLayerKernel.h" -#include "src/runtime/NEON/functions/NEPoolingAssemblyDispatch.h" +#include "arm_compute/core/Validate.h" +#include "src/runtime/cpu/operators/CpuPooling.h" namespace arm_compute { +struct NEPoolingLayer::Impl +{ + ITensor *src{ nullptr }; + ITensor *dst{ nullptr }; + ITensor *indices{ nullptr }; + std::shared_ptr<IMemoryManager> memory_manager{ nullptr }; + std::unique_ptr<cpu::CpuPooling> op{ nullptr }; +}; + NEPoolingLayer::~NEPoolingLayer() = default; NEPoolingLayer::NEPoolingLayer(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_manager(std::move(memory_manager)), _pooling_layer_kernel(), _border_handler(), _asm_glue(), _is_global_pooling_layer(false), _data_layout(DataLayout::NCHW) + : _impl(std::make_unique<Impl>()) { + _impl->memory_manager = std::move(memory_manager); } void NEPoolingLayer::configure(ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices) { - // Check if we can run assembly kernels. Currently, indices are not supported by those kernels - const bool run_optimised = bool(NEPoolingAssemblyDispatch::validate(input->info(), output->info(), pool_info)) && (indices == nullptr); - - if(run_optimised) - { - _asm_glue = std::make_unique<NEPoolingAssemblyDispatch>(_memory_manager); - _asm_glue->configure(input, output, pool_info); - ARM_COMPUTE_ERROR_ON(!_asm_glue->is_configured()); - } - else - { - // Check if we have Global Pooling Layer - _is_global_pooling_layer = (input->info()->dimension(0) == pool_info.pool_size.width) && (input->info()->dimension(1) == pool_info.pool_size.height); - - // Get data layout - _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->info()->data_layout() : pool_info.data_layout; - - // Configure pooling kernel - _pooling_layer_kernel = std::make_unique<NEPoolingLayerKernel>(); - _pooling_layer_kernel->configure(input, output, pool_info, indices); - - switch(_data_layout) - { - case DataLayout::NCHW: - { - // Configure border depending on operation required (quantize border in case of asymmetric data_type) - BorderMode border_mode = (!indices && pool_info.pool_type == PoolingType::MAX) ? BorderMode::REPLICATE : BorderMode::CONSTANT; - PixelValue zero_value((indices) ? std::numeric_limits<int>::min() : 0.f); - if(is_data_type_quantized_asymmetric(input->info()->data_type()) && !pool_info.exclude_padding) - { - zero_value = PixelValue(0, input->info()->data_type(), input->info()->quantization_info()); - } - _border_handler = std::make_unique<NEFillBorderKernel>(); - _border_handler->configure(input, _pooling_layer_kernel->border_size(), border_mode, zero_value); - break; - } - case DataLayout::NHWC: - break; - default: - ARM_COMPUTE_ERROR("Data layout not supported"); - } - } + _impl->src = input; + _impl->dst = output; + _impl->indices = indices; + _impl->op = std::make_unique<cpu::CpuPooling>(_impl->memory_manager); + _impl->op->configure(input->info(), output->info(), pool_info, (indices) ? indices->info() : nullptr); } Status NEPoolingLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) { - const bool run_optimised = bool(NEPoolingAssemblyDispatch::validate(input, output, pool_info)) && (indices == nullptr); - - if(run_optimised) - { - return Status{}; - } - - return NEPoolingLayerKernel::validate(input, output, pool_info, indices); + return cpu::CpuPooling::validate(input, output, pool_info, indices); } void NEPoolingLayer::run() { - if(_asm_glue && _asm_glue->is_configured()) - { - _asm_glue->run(); - } - else - { - switch(_data_layout) - { - case DataLayout::NCHW: - // Fill border - NEScheduler::get().schedule(_border_handler.get(), Window::DimY); - - // Run pooling layer - NEScheduler::get().schedule(_pooling_layer_kernel.get(), _is_global_pooling_layer ? Window::DimZ : Window::DimY); - break; - case DataLayout::NHWC: - // Run pooling layer - NEScheduler::get().schedule(_pooling_layer_kernel.get(), Window::DimX); - break; - default: - ARM_COMPUTE_ERROR("Data layout not supported"); - } - } + ITensorPack pack; + pack.add_tensor(TensorType::ACL_SRC, _impl->src); + pack.add_tensor(TensorType::ACL_DST_0, _impl->dst); + pack.add_tensor(TensorType::ACL_DST_1, _impl->indices); + _impl->op->run(pack); } } // namespace arm_compute |