!15785 npu enable reusing input tensor

From: @zhaozhenlong
Reviewed-by: @zhang_xue_tong,@zhanghaibo5
Signed-off-by: @zhang_xue_tong
This commit is contained in:
mindspore-ci-bot 2021-04-28 09:35:58 +08:00 committed by Gitee
commit 0a44273530
3 changed files with 16 additions and 6 deletions

View File

@ -15,6 +15,7 @@
*/ */
#include "src/runtime/agent/npu/npu_executor.h" #include "src/runtime/agent/npu/npu_executor.h"
#include <unordered_map>
#include "include/errorcode.h" #include "include/errorcode.h"
#include "src/runtime/agent/npu/npu_manager.h" #include "src/runtime/agent/npu/npu_manager.h"
#include "nnacl/pack.h" #include "nnacl/pack.h"
@ -100,15 +101,24 @@ bool IsSameShapeOutTensor(Tensor *tensor, std::shared_ptr<hiai::AiTensor> npu_te
} }
int NPUExecutor::Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors, int NPUExecutor::Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors,
const std::vector<kernel::LiteKernel *> &out_kernels, const std::vector<kernel::LiteKernel *> &in_kernels,
const std::vector<kernel::LiteKernel *> &kernels, Allocator *allocator, const std::vector<kernel::LiteKernel *> &kernels, Allocator *allocator,
const KernelCallBack &before, const KernelCallBack &after) { const KernelCallBack &before, const KernelCallBack &after) {
hiai::AiContext context; hiai::AiContext context;
std::vector<bool> inputs_visited(in_tensors.size(), false); std::unordered_map<lite::Tensor *, int> tensor_uses;
for (const auto ker : in_kernels) {
for (const auto ker_input : ker->in_tensors()) {
if (tensor_uses.find(ker_input) == tensor_uses.end()) {
tensor_uses.insert({ker_input, 1});
} else {
tensor_uses[ker_input]++;
}
}
}
for (int i = 0; i < npu_input_tensors_.size(); ++i) { for (int i = 0; i < npu_input_tensors_.size(); ++i) {
int index = 0; int index = 0;
for (; index < in_tensors.size(); index++) { for (; index < in_tensors.size(); index++) {
if (!inputs_visited[index] && IsSameShapeInTensor(in_tensors[index], npu_input_tensors_[i])) { if (tensor_uses[in_tensors[index]] > 0 && IsSameShapeInTensor(in_tensors[index], npu_input_tensors_[i])) {
void *data = in_tensors[index]->data_c(); void *data = in_tensors[index]->data_c();
if (data == nullptr) { if (data == nullptr) {
MS_LOG(ERROR) << "For " << model_name_ << ", the " << i << "th input data is nullptr"; MS_LOG(ERROR) << "For " << model_name_ << ", the " << i << "th input data is nullptr";
@ -116,7 +126,7 @@ int NPUExecutor::Run(const std::vector<Tensor *> &in_tensors, const std::vector<
} }
memcpy(npu_input_tensors_[i]->GetBuffer(), data, in_tensors[index]->Size()); memcpy(npu_input_tensors_[i]->GetBuffer(), data, in_tensors[index]->Size());
inputs_visited[index] = true; tensor_uses[in_tensors[index]]--;
in_tensors[index]->DecRefCount(); in_tensors[index]->DecRefCount();
break; break;
} }

View File

@ -36,7 +36,7 @@ class NPUExecutor : public Executor {
int Prepare(const std::vector<kernel::LiteKernel *> &kernels) override; int Prepare(const std::vector<kernel::LiteKernel *> &kernels) override;
int Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors, int Run(const std::vector<Tensor *> &in_tensors, const std::vector<Tensor *> &out_tensors,
const std::vector<kernel::LiteKernel *> &out_kernels, const std::vector<kernel::LiteKernel *> &kernels, const std::vector<kernel::LiteKernel *> &in_kernels, const std::vector<kernel::LiteKernel *> &kernels,
Allocator *allocator = nullptr, const KernelCallBack &before = nullptr, Allocator *allocator = nullptr, const KernelCallBack &before = nullptr,
const KernelCallBack &after = nullptr); const KernelCallBack &after = nullptr);

View File

@ -90,7 +90,7 @@ std::shared_ptr<domi::ModelBufferData> SubGraphNpuKernel::BuildIRModel() {
int SubGraphNpuKernel::Run() { int SubGraphNpuKernel::Run() {
return reinterpret_cast<lite::NPUExecutor *>(this->executor_) return reinterpret_cast<lite::NPUExecutor *>(this->executor_)
->Run(in_tensors_, out_tensor_sorted_, out_nodes_, nodes_); ->Run(in_tensors_, out_tensor_sorted_, in_nodes_, nodes_);
} }
int SubGraphNpuKernel::BuildNPUInputOp() { int SubGraphNpuKernel::BuildNPUInputOp() {