diff options
Diffstat (limited to 'src/runtime/NEON/functions')
-rw-r--r-- | src/runtime/NEON/functions/NEPoolingLayer.cpp | 27 | ||||
-rw-r--r-- | src/runtime/NEON/functions/NESoftmaxLayer.cpp | 83 |
2 files changed, 23 insertions, 87 deletions
diff --git a/src/runtime/NEON/functions/NEPoolingLayer.cpp b/src/runtime/NEON/functions/NEPoolingLayer.cpp index bbf3e7cc4e..8d267a32c0 100644 --- a/src/runtime/NEON/functions/NEPoolingLayer.cpp +++ b/src/runtime/NEON/functions/NEPoolingLayer.cpp @@ -26,6 +26,7 @@ #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/Tensor.h" +#include "src/core/helpers/MemoryHelpers.h" #include "src/runtime/cpu/operators/CpuPool2d.h" namespace arm_compute @@ -35,15 +36,18 @@ struct NEPoolingLayer::Impl ITensor *src{ nullptr }; ITensor *dst{ nullptr }; ITensor *indices{ nullptr }; - Tensor workspace{ nullptr }; std::unique_ptr<cpu::CpuPool2d> op{ nullptr }; + MemoryGroup memory_group{}; + ITensorPack run_pack{}; + WorkspaceData<Tensor> workspace_tensors{}; }; NEPoolingLayer::~NEPoolingLayer() = default; NEPoolingLayer::NEPoolingLayer(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(memory_manager), _impl(std::make_unique<Impl>()) + : _impl(std::make_unique<Impl>()) { + _impl->memory_group = MemoryGroup(std::move(memory_manager)); } void NEPoolingLayer::configure(ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices) @@ -54,14 +58,8 @@ void NEPoolingLayer::configure(ITensor *input, ITensor *output, const PoolingLay _impl->op = std::make_unique<cpu::CpuPool2d>(); _impl->op->configure(input->info(), output->info(), pool_info, (indices) ? indices->info() : nullptr); - // Allocate workspace based on kernel's memory requirements - const experimental::MemoryRequirements mem_req = _impl->op->workspace(); - if(!mem_req.empty()) - { - _impl->workspace.allocator()->init(TensorInfo(TensorShape{ (mem_req[0].size + mem_req[0].alignment) }, 1, DataType::S8), mem_req[0].alignment); - _memory_group.manage(&_impl->workspace); - _impl->workspace.allocator()->allocate(); - } + _impl->run_pack = { { TensorType::ACL_SRC, _impl->src }, { TensorType::ACL_DST_0, _impl->dst }, { TensorType::ACL_DST_1, _impl->indices } }; + _impl->workspace_tensors = manage_workspace<Tensor>(_impl->op->workspace(), _impl->memory_group, _impl->run_pack); } Status NEPoolingLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices) @@ -71,11 +69,8 @@ Status NEPoolingLayer::validate(const ITensorInfo *input, const ITensorInfo *out void NEPoolingLayer::run() { - 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); - pack.add_tensor(TensorType::ACL_INT_0, &_impl->workspace); - _impl->op->run(pack); + MemoryGroupResourceScope scope_mg(_impl->memory_group); + ARM_COMPUTE_ERROR_ON_NULLPTR(_impl->src, _impl->dst); + _impl->op->run(_impl->run_pack); } } // namespace arm_compute diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp index 3f1e43a8f2..af8546d4ca 100644 --- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp +++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp @@ -23,6 +23,7 @@ */ #include "arm_compute/runtime/NEON/functions/NESoftmaxLayer.h" #include "arm_compute/core/Validate.h" +#include "arm_compute/runtime/MemoryGroup.h" #include "arm_compute/runtime/Tensor.h" #include "src/core/cpu/kernels/CpuSoftmaxKernel.h" #include "src/core/helpers/SoftmaxHelpers.h" @@ -36,16 +37,17 @@ struct NESoftmaxLayerGeneric<IS_LOG>::Impl const ITensor *src{ nullptr }; ITensor *dst{ nullptr }; Tensor max{ nullptr }; - Tensor tmp{ nullptr }; - Tensor input_permuted{ nullptr }; - Tensor output_permuted{ nullptr }; std::unique_ptr<cpu::CpuSoftmaxGeneric<IS_LOG>> op{ nullptr }; + MemoryGroup memory_group{}; + ITensorPack run_pack{}; + WorkspaceData<Tensor> workspace_tensors{}; }; template <bool IS_LOG> NESoftmaxLayerGeneric<IS_LOG>::NESoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(std::move(memory_manager)), _impl(std::make_unique<Impl>()) + : _impl(std::make_unique<Impl>()) { + _impl->memory_group = MemoryGroup(std::move(memory_manager)); } template <bool IS_LOG> @@ -65,64 +67,8 @@ void NESoftmaxLayerGeneric<IS_LOG>::configure(ITensor *input, ITensor *output, f _impl->op = std::make_unique<cpu::CpuSoftmaxGeneric<IS_LOG>>(); _impl->op->configure(input->info(), output->info(), beta, axis); - const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(input->info()->num_dimensions()))); - const bool needs_permute = actual_axis > 0; - if(needs_permute) - { - // Add to the memory manager _input_permuted - auto permute_input = std::make_unique<cpu::CpuPermute>(); - _memory_group.manage(&_impl->input_permuted); - permute_input->configure(input->info(), _impl->input_permuted.info(), softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis)); - } - - // We want to deal with a 2D input. Either it is the permuted version of the original input (4D case) - // or it is the original input case (2D case) - ITensor *tmp_input = (needs_permute ? &_impl->input_permuted : input); - - // Create intermediate tensors shapes - const TensorInfo input_info = tmp_input->info()->clone()->reset_padding().set_is_resizable(true); - DataType tmp_data_type = is_data_type_quantized_asymmetric(tmp_input->info()->data_type()) ? DataType::F32 : tmp_input->info()->data_type(); - TensorInfo tensor_info_tmp(input_info.clone()->set_data_type(tmp_data_type)); - - // Init intermediate tensors - TensorShape max_sum_shape = tmp_input->info()->tensor_shape(); - max_sum_shape.set(0, 1); - _impl->max.allocator()->init(input_info.clone()->set_tensor_shape(max_sum_shape)); - _impl->tmp.allocator()->init(tensor_info_tmp); - - // Manage intermediate buffers - _memory_group.manage(&_impl->max); - _memory_group.manage(&_impl->tmp); - - // Configure kernels - auto max_kernel = std::make_unique<cpu::kernels::CpuLogits1DMaxKernel>(); - auto softmax_kernel = std::make_unique<cpu::kernels::CpuLogits1DSoftmaxKernel<IS_LOG>>(); - max_kernel->configure(tmp_input->info(), _impl->max.info()); - - if(needs_permute) - { - auto permute_output = std::make_unique<cpu::CpuPermute>(); - // Add to the memory manager _output_permuted - _memory_group.manage(&_impl->output_permuted); - - // The normalization kernel stores the result in a permuted output tensor - softmax_kernel->configure(tmp_input->info(), _impl->max.info(), _impl->output_permuted.info(), beta, _impl->tmp.info()); - _impl->input_permuted.allocator()->allocate(); - - // Re-permute the permuted output into the requested (4D) output - permute_output->configure(_impl->output_permuted.info(), output->info(), softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis)); - - // Allocate the intermediate permuted tensors - _impl->output_permuted.allocator()->allocate(); - } - else - { - softmax_kernel->configure(tmp_input->info(), _impl->max.info(), output->info(), beta, _impl->tmp.info()); - } - - // Allocate intermediate buffers - _impl->max.allocator()->allocate(); - _impl->tmp.allocator()->allocate(); + _impl->run_pack = { { TensorType::ACL_SRC, _impl->src }, { TensorType::ACL_DST, _impl->dst } }; + _impl->workspace_tensors = manage_workspace<Tensor>(_impl->op->workspace(), _impl->memory_group, _impl->run_pack); } template <bool IS_LOG> @@ -136,15 +82,10 @@ Status NESoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const I template <bool IS_LOG> void NESoftmaxLayerGeneric<IS_LOG>::run() { - MemoryGroupResourceScope scope_mg(_memory_group); - ITensorPack pack; - pack.add_tensor(TensorType::ACL_SRC, _impl->src); - pack.add_tensor(TensorType::ACL_DST, _impl->dst); - pack.add_tensor(TensorType::ACL_INT_0, &_impl->tmp); - pack.add_tensor(TensorType::ACL_INT_1, &_impl->max); - pack.add_tensor(TensorType::ACL_INT_2, &_impl->input_permuted); - pack.add_tensor(TensorType::ACL_INT_3, &_impl->output_permuted); - _impl->op->run(pack); + // Acquire all the temporaries + MemoryGroupResourceScope scope_mg(_impl->memory_group); + ARM_COMPUTE_ERROR_ON_NULLPTR(_impl->src, _impl->dst); + _impl->op->run(_impl->run_pack); } template class NESoftmaxLayerGeneric<false>; |