remove redundant preprocess in arithmetic operator and fix some fp16 bugs

mindspore/lite/src/runtime/kernel/arm/fp16/arithmetic_fp16.cc

@@ -21,8 +21,8 @@
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
 #include "src/runtime/runtime_api.h"
-#include "src/ops/populate/arithmetic_populate.h"
 #include "include/errorcode.h"
+#include "src/ops/arithmetic.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
@@ -100,44 +100,26 @@ int ArithmeticFP16CPUKernel::Init() {
   return ReSize();
 }
 
-int ArithmeticFP16CPUKernel::PreProcess() {
-  if (!InferShapeDone()) {
-    (const_cast<mindspore::lite::PrimitiveC *>(primitive_))->set_infer_flag(true);
-    auto ret = (const_cast<mindspore::lite::PrimitiveC *>(primitive_))->InferShape(in_tensors_, out_tensors_);
-    if (ret != 0) {
-      (const_cast<mindspore::lite::PrimitiveC *>(primitive_))->set_infer_flag(false);
-      MS_LOG(ERROR) << "InferShape fail!";
-      return ret;
-    }
-    if (op_parameter_ != nullptr) {
-      free(op_parameter_);
-      op_parameter_ = nullptr;
-    }
-    op_parameter_ = PopulateArithmetic(primitive_);
-    if (op_parameter_ == nullptr) {
-      MS_LOG(ERROR) << "Malloc parameter failed";
-      return RET_ERROR;
-    }
-    param_ = reinterpret_cast<ArithmeticParameter *>(op_parameter_);
-    ret = ReSize();
-    if (ret != 0) {
-      MS_LOG(ERROR) << "ReSize fail!ret: " << ret;
-      return ret;
-    }
-  }
+void ArithmeticFP16CPUKernel::InitParam() {
+  auto arithmetic_lite_primitive = (lite::Arithmetic *)primitive_;
+  param_->broadcasting_ = arithmetic_lite_primitive->Broadcasting();
+  param_->ndim_ = arithmetic_lite_primitive->NDims();
 
-  auto outputs = this->out_tensors();
-  for (auto *output : outputs) {
-    MS_ASSERT(output != nullptr);
-    output->MallocData();
-  }
-  return RET_OK;
+  param_->in_elements_num0_ = in_tensors_[0]->ElementsNum();
+  param_->in_elements_num1_ = in_tensors_[1]->ElementsNum();
+  param_->out_elements_num_ = out_tensors_[0]->ElementsNum();
+  memcpy(param_->in_shape0_, reinterpret_cast<const lite::Arithmetic *>(primitive_)->InShape0().data(),
+         reinterpret_cast<const lite::Arithmetic *>(primitive_)->InShape0().size() * sizeof(int));
+  memcpy(param_->in_shape1_, reinterpret_cast<const lite::Arithmetic *>(primitive_)->InShape1().data(),
+         reinterpret_cast<const lite::Arithmetic *>(primitive_)->InShape1().size() * sizeof(int));
+  memcpy(param_->out_shape_, reinterpret_cast<const lite::Arithmetic *>(primitive_)->OutputShape().data(),
+         reinterpret_cast<const lite::Arithmetic *>(primitive_)->OutputShape().size() * sizeof(int));
+
+  return;
 }
 
 int ArithmeticFP16CPUKernel::ReSize() {
-  param_->in_elements_num0_ = in_tensors_.at(0)->ElementsNum();
-  param_->in_elements_num1_ = in_tensors_.at(1)->ElementsNum();
-  param_->out_elements_num_ = out_tensors_.at(0)->ElementsNum();
+  InitParam();
 
   if (param_->in_elements_num0_ == 1 || param_->in_elements_num1_ == 1) {
     param_->broadcasting_ = false;

mindspore/lite/src/runtime/kernel/arm/fp16/arithmetic_fp16.h

@@ -44,7 +44,6 @@ class ArithmeticFP16CPUKernel : public LiteKernel {
   ~ArithmeticFP16CPUKernel() = default;
 
   int Init() override;
-  int PreProcess() override;
   int ReSize() override;
   int Run() override;
   int DoArithmetic(int task_id);
@@ -52,6 +51,7 @@ class ArithmeticFP16CPUKernel : public LiteKernel {
                    int out_thread_stride);
 
  private:
+  void InitParam();
   void FreeTmpBuffer();
   int outside_;
   int break_pos_;

mindspore/lite/src/runtime/kernel/arm/fp32/arithmetic_fp32.cc

@@ -20,7 +20,7 @@
 #include "src/kernel_registry.h"
 #include "src/runtime/kernel/arm/int8/add_int8.h"
 #include "src/runtime/runtime_api.h"
-#include "src/ops/populate/arithmetic_populate.h"
+#include "src/ops/arithmetic.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
@@ -94,40 +94,6 @@ int ArithmeticCPUKernel::InitBroadCastCase() {
   return RET_OK;
 }
 
-int ArithmeticCPUKernel::PreProcess() {
-  if (!InferShapeDone()) {
-    (const_cast<mindspore::lite::PrimitiveC *>(primitive_))->set_infer_flag(true);
-    auto ret = (const_cast<mindspore::lite::PrimitiveC *>(primitive_))->InferShape(in_tensors_, out_tensors_);
-    if (ret != 0) {
-      (const_cast<mindspore::lite::PrimitiveC *>(primitive_))->set_infer_flag(false);
-      MS_LOG(ERROR) << "InferShape fail!";
-      return ret;
-    }
-    if (op_parameter_ != nullptr) {
-      free(op_parameter_);
-      op_parameter_ = nullptr;
-    }
-    op_parameter_ = PopulateArithmetic(primitive_);
-    if (op_parameter_ == nullptr) {
-      MS_LOG(ERROR) << "Malloc parameter failed";
-      return RET_ERROR;
-    }
-    arithmeticParameter_ = reinterpret_cast<ArithmeticParameter *>(op_parameter_);
-    ret = ReSize();
-    if (ret != 0) {
-      MS_LOG(ERROR) << "ReSize fail!ret: " << ret;
-      return ret;
-    }
-  }
-
-  auto outputs = this->out_tensors();
-  for (auto *output : outputs) {
-    MS_ASSERT(output != nullptr);
-    output->MallocData();
-  }
-  return RET_OK;
-}
-
 void ArithmeticCPUKernel::InitRunFunction() {
   switch (op_parameter_->type_) {
     case PrimitiveType_Mul:

mindspore/lite/src/runtime/kernel/arm/fp32/arithmetic_fp32.h

@@ -62,7 +62,6 @@ class ArithmeticCPUKernel : public LiteKernel {
   ~ArithmeticCPUKernel() override;
 
   int Init() override;
-  int PreProcess() override;
   int ReSize() override;
   int Run() override;
   virtual int DoArithmetic(int task_id);

mindspore/lite/test/models_onnx.cfg

@@ -7,7 +7,7 @@ efficientnet-lite4-11.onnx
 mobilenetv2-7.onnx
 shufflenet-v2-10.onnx
 squeezenet1.1-7.onnx
-#densenet-9.onnx
+densenet-9.onnx
 ml_table_detection_fp32.onnx
 ml_table_segment.onnx
 googlenet-9.onnx

mindspore/lite/test/models_onnx_fp16.cfg

@@ -7,7 +7,7 @@ efficientnet-lite4-11.onnx 2
 mobilenetv2-7.onnx 8
 shufflenet-v2-10.onnx 5
 squeezenet1.1-7.onnx 1
-#densenet-9.onnx 6
+densenet-9.onnx 6
 ml_table_detection_fp32.onnx 2
 ml_table_segment.onnx 2
 googlenet-9.onnx 3

mindspore/lite/tools/benchmark/benchmark.cc

@@ -38,20 +38,20 @@ int Benchmark::GenerateRandomData(size_t size, void *data, TypeId data_type) {
   switch (data_type) {
     case kNumberTypeFloat32:
     case kNumberTypeFloat:
-      FillInputData<float>(size, data, std::uniform_real_distribution<float>(-0.5f, 0.5f));
+      FillInputData<float>(size, data, std::uniform_real_distribution<float>(0.1f, 1.0f));
       break;
     case kNumberTypeFloat64:
-      FillInputData<double>(size, data, std::uniform_real_distribution<double>(-0.5, 0.5));
+      FillInputData<double>(size, data, std::uniform_real_distribution<double>(0.1, 1.0));
       break;
     case kNumberTypeInt64:
-      FillInputData<int64_t>(size, data, std::uniform_int_distribution<int64_t>(0, 99));
+      FillInputData<int64_t>(size, data, std::uniform_int_distribution<int64_t>(0, 1));
       break;
     case kNumberTypeInt:
     case kNumberTypeInt32:
-      FillInputData<int32_t>(size, data, std::uniform_int_distribution<int32_t>(0, 99));
+      FillInputData<int32_t>(size, data, std::uniform_int_distribution<int32_t>(0, 1));
       break;
     case kNumberTypeInt16:
-      FillInputData<int16_t>(size, data, std::uniform_int_distribution<int16_t>(0, 99));
+      FillInputData<int16_t>(size, data, std::uniform_int_distribution<int16_t>(0, 1));
       break;
     case kNumberTypeInt8:
       FillInputData<int8_t>(size, data, std::uniform_int_distribution<int8_t>(-127, 127));