[MSLITE] Optimize gather fp16 op's performance

mindspore/lite/src/runtime/kernel/arm/fp16/gather_fp16.cc

@@ -30,7 +30,22 @@ using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_Gather;
 
 namespace mindspore::kernel {
+GatherFp16CPUKernel::~GatherFp16CPUKernel() {
+  if (input_data_) {
+    context_->allocator->Free(input_data_);
+    input_data_ = nullptr;
+  }
+}
+
 int GatherFp16CPUKernel::Init() {
+  auto input_tensor = in_tensors_.at(0);
+  if (input_tensor->data_type() == kNumberTypeFloat32 && input_tensor->data_c() != nullptr) {
+    const_input_ = true;
+    input_data_ =
+      reinterpret_cast<float16_t *>(context_->allocator->Malloc(input_tensor->ElementsNum() * sizeof(float16_t)));
+    Float32ToFloat16(reinterpret_cast<float *>(input_tensor->data_c()), input_data_, input_tensor->ElementsNum());
+  }
+
   if (!InferShapeDone()) {
     return RET_OK;
   }
@@ -128,11 +143,13 @@ int GatherFp16CPUKernel::Run() {
     MS_LOG(ERROR) << "AssignIndicesData failed, error_code[" << ret << "]";
     return ret;
   }
-  auto input_tensor = in_tensors_.at(0);
+  if (!const_input_) {
-  if (input_tensor->data_type() == kNumberTypeFloat32) {
+    auto input_tensor = in_tensors_.at(0);
-    input_data_ =
+    if (input_tensor->data_type() == kNumberTypeFloat32) {
-      reinterpret_cast<float16_t *>(context_->allocator->Malloc(input_tensor->ElementsNum() * sizeof(float16_t)));
+      input_data_ =
-    Float32ToFloat16(reinterpret_cast<float *>(input_tensor->data_c()), input_data_, input_tensor->ElementsNum());
+        reinterpret_cast<float16_t *>(context_->allocator->Malloc(input_tensor->ElementsNum() * sizeof(float16_t)));
+      Float32ToFloat16(reinterpret_cast<float *>(input_tensor->data_c()), input_data_, input_tensor->ElementsNum());
+    }
   }
   ret = ParallelLaunch(this->context_->thread_pool_, GatherRunFp16, this, op_parameter_->thread_num_);
   if (ret != RET_OK) {
@@ -142,7 +159,7 @@ int GatherFp16CPUKernel::Run() {
     context_->allocator->Free(indices_data_);
     indices_data_ = nullptr;
   }
-  if (input_data_) {
+  if (!const_input_ && input_data_) {
     context_->allocator->Free(input_data_);
     input_data_ = nullptr;
   }

mindspore/lite/src/runtime/kernel/arm/fp16/gather_fp16.h

@@ -30,7 +30,7 @@ class GatherFp16CPUKernel : public LiteKernel {
   GatherFp16CPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs,
                       const std::vector<lite::Tensor *> &outputs, const lite::InnerContext *ctx)
       : LiteKernel(parameter, inputs, outputs, ctx) {}
-  ~GatherFp16CPUKernel() = default;
+  ~GatherFp16CPUKernel() override;
 
   int Init() override;
   int ReSize() override;
@@ -42,6 +42,7 @@ class GatherFp16CPUKernel : public LiteKernel {
   int *indices_data_ = nullptr;
   int AssignIndicesData(bool isIndicesInt32, int indices_num, lite::Tensor *indices_tensor);
   float16_t *input_data_ = nullptr;
+  bool const_input_ = false;
 };
 }  // namespace mindspore::kernel