diff options
11 files changed, 88 insertions, 18 deletions
diff --git a/runtime/neurun/backend/acl_cl/TensorRegister.h b/runtime/neurun/backend/acl_cl/TensorRegister.h index b78c9b547..a523f2d47 100644 --- a/runtime/neurun/backend/acl_cl/TensorRegister.h +++ b/runtime/neurun/backend/acl_cl/TensorRegister.h @@ -18,6 +18,8 @@ #define __NEURUN_BACKEND_ACL_CL_TENSOR_REGISTER_H__ #include <AclTensorRegister.h> +#include <misc/polymorphic_downcast.h> +#include "TensorBuilder.h" namespace neurun { @@ -26,7 +28,22 @@ namespace backend namespace acl_cl { -using TensorRegister = acl_common::AclTensorRegister; +class TensorRegister : public acl_common::AclTensorRegister +{ +public: + TensorRegister(const model::Operands &operands, + const std::shared_ptr<TensorBuilder> &tensor_builder) + : acl_common::AclTensorRegister{operands, tensor_builder} + { + // DO NOTHING + } + + void setUsesCount(const model::OperandIndex &ind, size_t num_uses) const override + { + nnfw::misc::polymorphic_downcast<TensorBuilder *>(tensor_builder().get()) + ->setUsesCount(ind, num_uses); + } +}; } // namespace acl_cl } // namespace backend diff --git a/runtime/neurun/backend/acl_cl/operand/CLTensor.cc b/runtime/neurun/backend/acl_cl/operand/CLTensor.cc index 6153fc2e4..dab74e65f 100644 --- a/runtime/neurun/backend/acl_cl/operand/CLTensor.cc +++ b/runtime/neurun/backend/acl_cl/operand/CLTensor.cc @@ -31,8 +31,8 @@ namespace acl_cl namespace operand { -CLTensor::CLTensor(const arm_compute::TensorInfo &info, size_t rank) - : _cl_tensor(std::make_shared<arm_compute::CLTensor>()), _rank{rank} +CLTensor::CLTensor(const arm_compute::TensorInfo &info, size_t rank, size_t num_uses) + : _cl_tensor(std::make_shared<arm_compute::CLTensor>()), _rank{rank}, _num_uses{num_uses} { allocator()->init(info); } diff --git a/runtime/neurun/backend/acl_cl/operand/CLTensor.h b/runtime/neurun/backend/acl_cl/operand/CLTensor.h index 952851623..8518bf0c3 100644 --- a/runtime/neurun/backend/acl_cl/operand/CLTensor.h +++ b/runtime/neurun/backend/acl_cl/operand/CLTensor.h @@ -38,7 +38,7 @@ public: CLTensor() = delete; public: - CLTensor(const arm_compute::TensorInfo &info, size_t rank); + CLTensor(const arm_compute::TensorInfo &info, size_t rank, size_t num_uses); public: size_t num_dimensions() const final { return _rank; } @@ -46,6 +46,7 @@ public: public: const arm_compute::CLTensor *handle() const override; arm_compute::CLTensor *handle() override; + size_t num_uses() const { return _num_uses; } public: arm_compute::CLTensorAllocator *allocator(); @@ -63,6 +64,7 @@ public: private: std::shared_ptr<arm_compute::CLTensor> _cl_tensor; size_t _rank; + size_t _num_uses; }; } // namespace operand diff --git a/runtime/neurun/backend/acl_common/AclMemoryManager.h b/runtime/neurun/backend/acl_common/AclMemoryManager.h index 60d2c5fe1..39662fc08 100644 --- a/runtime/neurun/backend/acl_common/AclMemoryManager.h +++ b/runtime/neurun/backend/acl_common/AclMemoryManager.h @@ -66,9 +66,9 @@ public: virtual void finishLifetime(const model::OperandIndex &) { /* DO NOTHING */} void buildTensor(const model::OperandIndex &ind, const ::arm_compute::TensorInfo &info, - size_t rank) + size_t rank, size_t num_uses) { - auto tensor = std::make_shared<T_Tensor>(info, rank); + auto tensor = std::make_shared<T_Tensor>(info, rank, num_uses); _tensors[ind] = tensor; } diff --git a/runtime/neurun/backend/acl_common/AclTensorManager.h b/runtime/neurun/backend/acl_common/AclTensorManager.h index 872966365..3849f18dd 100644 --- a/runtime/neurun/backend/acl_common/AclTensorManager.h +++ b/runtime/neurun/backend/acl_common/AclTensorManager.h @@ -51,7 +51,7 @@ public: void deallocateInternalBufferManager(void); void buildTensor(const model::OperandIndex &ind, const ::arm_compute::TensorInfo &info, - size_t rank, bool as_const); + size_t rank, bool as_const, size_t num_uses); void buildSubtensor(const model::OperandIndex &parent, const model::OperandIndex &child, const ::arm_compute::TensorShape &shape, const ::arm_compute::Coordinates &coordinates, size_t rank, @@ -145,17 +145,17 @@ void AclTensorManager<T_ITensor, T_Tensor, T_SubTensor, T_Object>::deallocateInt template <typename T_ITensor, typename T_Tensor, typename T_SubTensor, typename T_Object> void AclTensorManager<T_ITensor, T_Tensor, T_SubTensor, T_Object>::buildTensor( const model::OperandIndex &ind, const ::arm_compute::TensorInfo &info, size_t rank, - bool as_const) + bool as_const, size_t num_uses) { assert(_ind_to_mgr.find(ind) == _ind_to_mgr.end()); if (as_const) { - _const_mgr->buildTensor(ind, info, rank); + _const_mgr->buildTensor(ind, info, rank, num_uses); _ind_to_mgr.insert({ind, *_const_mgr}); } else { - _nonconst_mgr->buildTensor(ind, info, rank); + _nonconst_mgr->buildTensor(ind, info, rank, num_uses); _ind_to_mgr.insert({ind, *_nonconst_mgr}); } } @@ -281,7 +281,9 @@ void AclTensorManager<T_ITensor, T_Tensor, T_SubTensor, T_Object>::tryDeallocCon { const auto &ind = it->first; auto tensor = it->second; - if (tensor->handle() && !tensor->handle()->is_used()) + // NOTE The condition "tensor->num_uses() < 2" is used to prevent deallocating a constant tensor + // used in several nodes. + if (tensor->handle() && !tensor->handle()->is_used() && tensor->num_uses() < 2) { VERBOSE(AclTensorManager) << "Tensor #" << ind.value() << " will be deallocated as an unused constant tensor" << std::endl; diff --git a/runtime/neurun/backend/acl_common/AclTensorRegister.cc b/runtime/neurun/backend/acl_common/AclTensorRegister.cc index 14691a6fb..b85cfe72a 100644 --- a/runtime/neurun/backend/acl_common/AclTensorRegister.cc +++ b/runtime/neurun/backend/acl_common/AclTensorRegister.cc @@ -30,6 +30,20 @@ AclTensorRegister::AclTensorRegister(const model::Operands &operands, assert(tensor_builder != nullptr); } +void AclTensorRegister::visit(const model::Subgraph &subgraph) +{ + for (const auto &e : subgraph.operations()) + { + const auto &node = *(e.node); + node.accept(*this); + // Set count of nodes to use operand + for (const auto &input : node.getInputs()) + { + setUsesCount(input, _operands.at(input).getUses().size()); + } + } +} + } // namespace acl_common } // namespace backend } // namespace neurun diff --git a/runtime/neurun/backend/acl_common/AclTensorRegister.h b/runtime/neurun/backend/acl_common/AclTensorRegister.h index ff215072f..1c316255c 100644 --- a/runtime/neurun/backend/acl_common/AclTensorRegister.h +++ b/runtime/neurun/backend/acl_common/AclTensorRegister.h @@ -28,11 +28,18 @@ namespace acl_common class AclTensorRegister : public ITensorRegister { -public: +protected: AclTensorRegister(const model::Operands &operands, const std::shared_ptr<ITensorBuilder> &tensor_builder); -private: +public: + virtual ~AclTensorRegister() = default; + +protected: + void visit(const model::Subgraph &subgraph); + virtual void setUsesCount(const model::OperandIndex &ind, size_t num_uses) const = 0; + +protected: const model::Operands &operands() const override { return _operands; } std::shared_ptr<ITensorBuilder> tensor_builder() const override { return _tensor_builder; } bool supportSubTensor() const final { return true; } diff --git a/runtime/neurun/backend/acl_common/TemplTensorBuilder.h b/runtime/neurun/backend/acl_common/TemplTensorBuilder.h index ef70ee2d6..c770b52e5 100644 --- a/runtime/neurun/backend/acl_common/TemplTensorBuilder.h +++ b/runtime/neurun/backend/acl_common/TemplTensorBuilder.h @@ -97,6 +97,13 @@ public: T_AclTensorManager *acl_tensor_manager(void) { return _tensor_mgr.get(); } + void setUsesCount(const model::OperandIndex &index, size_t num_uses) + { + assert(_uses_count_map.find(index) != _uses_count_map.end() ? _uses_count_map[index] == num_uses + : true); + _uses_count_map[index] = num_uses; + } + private: void buildTensors(void); void buildSubtensors(void); @@ -108,6 +115,7 @@ private: model::OperandIndexMap<compiler::SubTensorInfo> _subtensor_info_map; model::OperandIndexMap<bool> _apply_dim_correction_map; model::OperandIndexMap<model::Layout> _tensor_layout_map; + model::OperandIndexMap<size_t> _uses_count_map; std::unique_ptr<T_AclTensorManager> _tensor_mgr; model::OperandIndexSequence _constants; @@ -325,7 +333,8 @@ void TemplTensorBuilder<T_ITensor, T_Tensor, T_SubTensor, T_Object>::buildTensor const auto &backend_layout = _tensor_layout_map[root_parent]; auto tensor_info = asTensorInfo(info.shape(), info.typeInfo(), model::Layout::UNKNOWN, backend_layout, _apply_dim_correction_map[ind]); - _tensor_mgr->buildTensor(ind, tensor_info, info.shape().rank(), _constants.contains(ind)); + _tensor_mgr->buildTensor(ind, tensor_info, info.shape().rank(), _constants.contains(ind), + _uses_count_map[ind]); } } diff --git a/runtime/neurun/backend/acl_neon/TensorRegister.h b/runtime/neurun/backend/acl_neon/TensorRegister.h index 388cb9172..708beb46e 100644 --- a/runtime/neurun/backend/acl_neon/TensorRegister.h +++ b/runtime/neurun/backend/acl_neon/TensorRegister.h @@ -18,6 +18,8 @@ #define __NEURUN_BACKEND_ACL_NEON_TENSOR_REGISTER_H__ #include <AclTensorRegister.h> +#include <misc/polymorphic_downcast.h> +#include "TensorBuilder.h" namespace neurun { @@ -26,7 +28,22 @@ namespace backend namespace acl_neon { -using TensorRegister = acl_common::AclTensorRegister; +class TensorRegister : public acl_common::AclTensorRegister +{ +public: + TensorRegister(const model::Operands &operands, + const std::shared_ptr<TensorBuilder> &tensor_builder) + : acl_common::AclTensorRegister{operands, tensor_builder} + { + // DO NOTHING + } + + void setUsesCount(const model::OperandIndex &ind, size_t num_uses) const override + { + nnfw::misc::polymorphic_downcast<TensorBuilder *>(tensor_builder().get()) + ->setUsesCount(ind, num_uses); + } +}; } // namespace acl_neon } // namespace backend diff --git a/runtime/neurun/backend/acl_neon/operand/NETensor.cc b/runtime/neurun/backend/acl_neon/operand/NETensor.cc index 756403ef1..8a9ece88f 100644 --- a/runtime/neurun/backend/acl_neon/operand/NETensor.cc +++ b/runtime/neurun/backend/acl_neon/operand/NETensor.cc @@ -27,8 +27,8 @@ namespace acl_neon namespace operand { -NETensor::NETensor(const arm_compute::TensorInfo &info, size_t rank) - : _ne_tensor(std::make_shared<arm_compute::Tensor>()), _rank{rank} +NETensor::NETensor(const arm_compute::TensorInfo &info, size_t rank, size_t num_uses) + : _ne_tensor(std::make_shared<arm_compute::Tensor>()), _rank{rank}, _num_uses{num_uses} { allocator()->init(info); } diff --git a/runtime/neurun/backend/acl_neon/operand/NETensor.h b/runtime/neurun/backend/acl_neon/operand/NETensor.h index 298a82054..3de4695e9 100644 --- a/runtime/neurun/backend/acl_neon/operand/NETensor.h +++ b/runtime/neurun/backend/acl_neon/operand/NETensor.h @@ -37,7 +37,7 @@ public: NETensor() = delete; public: - NETensor(const arm_compute::TensorInfo &info, size_t rank); + NETensor(const arm_compute::TensorInfo &info, size_t rank, size_t num_uses); public: size_t num_dimensions() const final { return _rank; } @@ -45,6 +45,7 @@ public: public: const arm_compute::Tensor *handle() const override; arm_compute::Tensor *handle() override; + size_t num_uses() const { return _num_uses; } public: arm_compute::TensorAllocator *allocator(); @@ -52,6 +53,7 @@ public: private: std::shared_ptr<arm_compute::Tensor> _ne_tensor; size_t _rank; + size_t _num_uses; }; } // namespace operand |