!11173 [MSLITE] delete fp16 cast

From: @ling_qiao_min
Reviewed-by: 
Signed-off-by:

mindspore/lite/nnacl/base/gather_base.c

@@ -15,31 +15,22 @@
  */
 
 #include "nnacl/base/gather_base.h"
-int GatherFp32(const float *input, int outer_size, int inner_size, int limit, const int *indices,
-               int indices_element_size, float *output) {
-  for (int m = 0; m < outer_size; ++m) {
-    const float *inputm = input + inner_size * m * limit;
-    float *outputm = output + inner_size * m * indices_element_size;
-    for (int i = 0; i < indices_element_size; ++i) {
-      if (indices[i] < 0 || indices[i] > limit) {
-        return NNACL_ERR;
-      }
-      memcpy(outputm + i * inner_size, inputm + indices[i] * inner_size, sizeof(float) * inner_size);
-    }
-  }
-  return NNACL_OK;
-}
 
-int GatherInt32(const int32_t *input, int outer_size, int inner_size, int limit, const int *indices,
-                int indices_element_size, int32_t *output) {
+int Gather(const void *input, int outer_size, int inner_size, int limit, const int *indices, int indices_element_size,
+           void *output, int data_size) {
+  const int8_t *int8_in = (int8_t *)input;
+  int8_t *int8_out = (int8_t *)output;
+
   for (int m = 0; m < outer_size; ++m) {
-    const int32_t *inputm = input + inner_size * m * limit;
-    int32_t *outputm = output + inner_size * m * indices_element_size;
+    const int8_t *int8_in_m = int8_in + inner_size * m * limit * data_size;
+    int8_t *int8_out_m = int8_out + inner_size * m * indices_element_size * data_size;
+
     for (int i = 0; i < indices_element_size; ++i) {
       if (indices[i] < 0 || indices[i] > limit) {
         return NNACL_ERR;
       }
-      memcpy(outputm + i * inner_size, inputm + indices[i] * inner_size, sizeof(int32_t) * inner_size);
+      memcpy(int8_out_m + i * inner_size * data_size, int8_in_m + indices[i] * inner_size * data_size,
+             data_size * inner_size);
     }
   }
   return NNACL_OK;

mindspore/lite/nnacl/base/gather_base.h

@@ -24,10 +24,8 @@
 #ifdef __cplusplus
 extern "C" {
 #endif
-int GatherFp32(const float *input, int outer_size, int inner_size, int limit, const int *indices,
-               int indices_element_size, float *output);
-int GatherInt32(const int32_t *input, int outer_size, int inner_size, int limit, const int *indices,
-                int indices_element_size, int32_t *output);
+int Gather(const void *input, int outer_size, int inner_size, int limit, const int *indices, int indices_element_size,
+           void *output, int data_size);
 #ifdef __cplusplus
 }
 #endif

mindspore/lite/src/runtime/kernel/arm/fp16/activation_fp16.cc

@@ -20,7 +20,6 @@
 #include "src/runtime/runtime_api.h"
 #include "include/errorcode.h"
 #include "src/runtime/kernel/arm/fp16/common_fp16.h"
-#include "nnacl/fp16/cast_fp16.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
@@ -47,35 +46,6 @@ int ActivationFp16CPUKernel::Init() {
 
 int ActivationFp16CPUKernel::ReSize() { return RET_OK; }
 
-int ActivationFp16CPUKernel::MallocTmpBuffer() {
-  fp16_input_ = ConvertInputFp32toFp16(in_tensors_.at(0), context_);
-  if (fp16_input_ == nullptr) {
-    MS_LOG(ERROR) << "malloc data failed";
-    return RET_ERROR;
-  }
-  fp16_output_ = MallocOutputFp16(out_tensors_.at(0), context_);
-  if (fp16_output_ == nullptr) {
-    MS_LOG(ERROR) << "malloc data failed";
-    return RET_ERROR;
-  }
-  return RET_OK;
-}
-
-void ActivationFp16CPUKernel::FreeTmpBuffer() {
-  if (in_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    if (fp16_input_ != nullptr) {
-      context_->allocator->Free(fp16_input_);
-      fp16_input_ = nullptr;
-    }
-  }
-  if (out_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    if (fp16_output_ != nullptr) {
-      context_->allocator->Free(fp16_output_);
-      fp16_output_ = nullptr;
-    }
-  }
-}
-
 int ActivationFp16CPUKernel::DoActivation(int task_id) {
   auto length = in_tensors_.at(0)->ElementsNum();
 
@@ -115,24 +85,18 @@ int ActivationFp16Run(void *cdata, int task_id) {
 }
 
 int ActivationFp16CPUKernel::Run() {
-  auto ret = MallocTmpBuffer();
-  if (ret != RET_OK) {
-    FreeTmpBuffer();
-    return ret;
-  }
+  auto input_tensor = in_tensors_.at(0);
+  auto output_tensor = out_tensors_.at(0);
+
+  fp16_input_ = reinterpret_cast<float16_t *>(input_tensor->data_c());
+  fp16_output_ = reinterpret_cast<float16_t *>(output_tensor->data_c());
 
   int error_code = ParallelLaunch(this->context_->thread_pool_, ActivationFp16Run, this, thread_count_);
   if (error_code != RET_OK) {
     MS_LOG(ERROR) << "Activation function error error_code[" << error_code << "]";
-    FreeTmpBuffer();
     return RET_ERROR;
   }
 
-  auto out_tensor = out_tensors_.at(0);
-  if (out_tensor->data_type() == kNumberTypeFloat32) {
-    Float16ToFloat32(fp16_output_, reinterpret_cast<float *>(out_tensor->MutableData()), out_tensor->ElementsNum());
-  }
-  FreeTmpBuffer();
   return RET_OK;
 }
 

mindspore/lite/src/runtime/kernel/arm/fp16/activation_fp16.h

@@ -37,8 +37,6 @@ class ActivationFp16CPUKernel : public LiteKernel {
   int ReSize() override;
   int Run() override;
   int DoActivation(int task_id);
-  int MallocTmpBuffer();
-  void FreeTmpBuffer();
 
  private:
   int thread_count_;

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

@@ -209,6 +209,7 @@ int ArithmeticFP16CPUKernel::Run() {
     FreeTmpBuffer();
     return RET_ERROR;
   }
+
   auto ret = ParallelLaunch(this->context_->thread_pool_, ArithmeticsRunFp16, this, context_->thread_num_);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "ArithmeticsRunFp16 run error error_code[" << ret << "]";

mindspore/lite/src/runtime/kernel/arm/fp16/arithmetic_self_fp16.cc

@@ -16,7 +16,6 @@
 #include "src/runtime/kernel/arm/fp16/arithmetic_self_fp16.h"
 #include "src/runtime/kernel/arm/fp16/common_fp16.h"
 #include "src/kernel_registry.h"
-#include "nnacl/fp16/cast_fp16.h"
 #include "nnacl/fp16/arithmetic_self_fp16.h"
 
 using mindspore::lite::KernelRegistrar;
@@ -72,36 +71,17 @@ int ArithmeticSelfFp16CPUKernel::DoExecute(int task_id) {
   return ret;
 }
 
-void ArithmeticSelfFp16CPUKernel::FreeInputAndOutput() {
-  if (in_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    context_->allocator->Free(input_fp16_ptr_);
-    input_fp16_ptr_ = nullptr;
-  }
-  if (out_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    context_->allocator->Free(output_fp16_ptr_);
-    output_fp16_ptr_ = nullptr;
-  }
-}
-
 int ArithmeticSelfFp16CPUKernel::Run() {
   auto input_tensor = in_tensors_.at(0);
   auto output_tensor = out_tensors_.at(0);
-  input_fp16_ptr_ = ConvertInputFp32toFp16(input_tensor, context_);
-  output_fp16_ptr_ = MallocOutputFp16(output_tensor, context_);
-  if (input_fp16_ptr_ == nullptr || output_fp16_ptr_ == nullptr) {
-    FreeInputAndOutput();
-    MS_LOG(ERROR) << "input or output is nullptr";
-    return RET_ERROR;
-  }
+
+  input_fp16_ptr_ = reinterpret_cast<float16_t *>(input_tensor->data_c());
+  output_fp16_ptr_ = reinterpret_cast<float16_t *>(output_tensor->data_c());
+
   auto ret = ParallelLaunch(this->context_->thread_pool_, ArithmeticSelfRun, this, op_parameter_->thread_num_);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "ArithmeticSelfRun error error_code[" << ret << "]";
   }
-  if (out_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    Float16ToFloat32(output_fp16_ptr_, reinterpret_cast<float *>(output_tensor->MutableData()),
-                     output_tensor->ElementsNum());
-  }
-  FreeInputAndOutput();
   return ret;
 }
 

mindspore/lite/src/runtime/kernel/arm/fp16/arithmetic_self_fp16.h

@@ -35,7 +35,6 @@ class ArithmeticSelfFp16CPUKernel : public ArithmeticSelfCPUKernel {
   int DoExecute(int task_id) override;
 
  private:
-  void FreeInputAndOutput();
   ArithmeticSelfFp16Func GetArithmeticSelfFp16Fun(int primitive_type);
   ArithmeticSelfFp16Func fp16_func_ = nullptr;
   float16_t *input_fp16_ptr_ = nullptr;

mindspore/lite/src/runtime/kernel/arm/fp16/crop_fp16.cc

@@ -49,37 +49,18 @@ static int CropFp16Run(void *cdata, int task_id) {
 }
 
 int CropFp16CPUKernel::Run() {
-  input_ptr_ = ConvertInputFp32toFp16(in_tensors_.at(kInputIndex), context_);
-  output_ptr_ = MallocOutputFp16(out_tensors_.at(kOutputIndex), context_);
-  if (input_ptr_ == nullptr || output_ptr_ == nullptr) {
-    MS_LOG(ERROR) << "input or output is nullptr";
-    FreeInputAndOutput();
-    return RET_ERROR;
-  }
+  auto input_tensor = in_tensors_.at(0);
+  auto output_tensor = out_tensors_.at(0);
+
+  input_ptr_ = reinterpret_cast<float16_t *>(input_tensor->data_c());
+  output_ptr_ = reinterpret_cast<float16_t *>(output_tensor->data_c());
 
   auto ret = ParallelLaunch(this->context_->thread_pool_, CropFp16Run, this, crop_para_->thread_count_);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "ParallelLaunch failed: " << ret;
-    FreeInputAndOutput();
-  }
-  if (out_tensors_.at(kOutputIndex)->data_type() == kNumberTypeFloat32) {
-    Float16ToFloat32(output_ptr_, reinterpret_cast<float *>(out_tensors_.at(kOutputIndex)->data_c()),
-                     out_tensors_.at(kOutputIndex)->ElementsNum());
   }
-  FreeInputAndOutput();
   return ret;
 }
 
-void CropFp16CPUKernel::FreeInputAndOutput() {
-  if (in_tensors_.at(kInputIndex)->data_type() == kNumberTypeFloat32) {
-    context_->allocator->Free(input_ptr_);
-    input_ptr_ = nullptr;
-  }
-  if (out_tensors_.at(kOutputIndex)->data_type() == kNumberTypeFloat32) {
-    context_->allocator->Free(output_ptr_);
-    output_ptr_ = nullptr;
-  }
-}
-
 REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Crop, LiteKernelCreator<CropFp16CPUKernel>)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp16/crop_fp16.h

@@ -21,7 +21,6 @@
 #include <vector>
 #include "include/errorcode.h"
 #include "nnacl/crop_parameter.h"
-#include "nnacl/fp16/cast_fp16.h"
 #include "nnacl/fp16/crop_fp16.h"
 #include "src/lite_kernel.h"
 #include "src/runtime/kernel/arm/base/crop_base.h"
@@ -44,7 +43,6 @@ class CropFp16CPUKernel : public CropBaseCPUKernel {
  private:
   float16_t *input_ptr_ = nullptr;
   float16_t *output_ptr_ = nullptr;
-  void FreeInputAndOutput();
 };
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp16/pad_fp16.cc

@@ -16,7 +16,6 @@
 
 #include "src/runtime/kernel/arm/fp16/pad_fp16.h"
 #include "src/runtime/kernel/arm/fp16/common_fp16.h"
-#include "nnacl/fp16/cast_fp16.h"
 #include "src/kernel_registry.h"
 #include "src/runtime/runtime_api.h"
 
@@ -43,16 +42,10 @@ int PadFp16CPUKernel::RunMirrorPadImpl(int task_id) {
 int PadFp16CPUKernel::Run() {
   auto input_tensor = in_tensors_.at(0);
   auto output_tensor = out_tensors_.at(0);
-  is_input_fp32_ = input_tensor->data_type() == kNumberTypeFloat32;
-  is_output_fp32_ = output_tensor->data_type() == kNumberTypeFloat32;
 
-  input_ = ConvertInputFp32toFp16(input_tensor, context_);
-  output_ = MallocOutputFp16(output_tensor, context_);
-  if (input_ == nullptr || output_ == nullptr) {
-    FreeInputAndOutput();
-    MS_LOG(ERROR) << "input or output is nullptr";
-    return RET_ERROR;
-  }
+  input_ = reinterpret_cast<float16_t *>(input_tensor->data_c());
+  output_ = reinterpret_cast<float16_t *>(output_tensor->data_c());
+
   int ret = 0;
   if (pad_param_->pad_mode_ == static_cast<int>(schema::PaddingMode_CONSTANT)) {
     if (pad_param_->constant_value_ - 0.0f < 1e-5) {
@@ -73,22 +66,8 @@ int PadFp16CPUKernel::Run() {
       MS_LOG(ERROR) << "Pad Reflect or Symmetric mode run error, error_code[" << ret << "]";
     }
   }
-  if (is_output_fp32_) {
-    Float16ToFloat32(output_, reinterpret_cast<float *>(output_tensor->MutableData()), output_tensor->ElementsNum());
-  }
-  FreeInputAndOutput();
-  return ret;
-}
 
-void PadFp16CPUKernel::FreeInputAndOutput() {
-  if (is_input_fp32_) {
-    context_->allocator->Free(input_);
-    input_ = nullptr;
-  }
-  if (is_output_fp32_) {
-    context_->allocator->Free(output_);
-    output_ = nullptr;
-  }
+  return ret;
 }
 
 REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Pad, LiteKernelCreator<PadFp16CPUKernel>)

mindspore/lite/src/runtime/kernel/arm/fp16/pad_fp16.h

@@ -35,9 +35,6 @@ class PadFp16CPUKernel : public PadCPUKernel {
   int RunMirrorPadImpl(int task_id) override;
 
  private:
-  void FreeInputAndOutput();
-  bool is_input_fp32_ = false;
-  bool is_output_fp32_ = false;
   float16_t *input_ = nullptr;
   float16_t *output_ = nullptr;
 };

mindspore/lite/src/runtime/kernel/arm/fp16/pooling_fp16.cc

@@ -20,7 +20,6 @@
 #include "src/runtime/runtime_api.h"
 #include "include/errorcode.h"
 #include "nnacl/op_base.h"
-#include "nnacl/fp16/cast_fp16.h"
 #include "src/runtime/kernel/arm/fp16/common_fp16.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;
@@ -84,31 +83,17 @@ static int PoolingFp16Impl(void *cdata, int task_id) {
 }
 
 int PoolingFp16CPUKernel::Run() {
-  auto input_tensor = in_tensors_.at(kInputIndex);
-  auto in_data_type_ = input_tensor->data_type();
-  MS_ASSERT(in_data_type_ == kNumberTypeFloat32 || in_data_type_ == kNumberTypeFloat16);
-  fp16_input_ = ConvertInputFp32toFp16(input_tensor, context_);
+  auto input_tensor = in_tensors_.at(0);
+  auto output_tensor = out_tensors_.at(0);
 
-  auto out_tensor = out_tensors_.at(kOutputIndex);
-  auto out_data_type_ = out_tensor->data_type();
-  MS_ASSERT(out_data_type_ == kNumberTypeFloat32 || out_data_type_ == kNumberTypeFloat16);
-  fp16_output_ = MallocOutputFp16(out_tensor, context_);
+  fp16_input_ = reinterpret_cast<float16_t *>(input_tensor->data_c());
+  fp16_output_ = reinterpret_cast<float16_t *>(output_tensor->data_c());
 
   int error_code = ParallelLaunch(this->context_->thread_pool_, PoolingFp16Impl, this, thread_count_);
   if (error_code != RET_OK) {
     MS_LOG(ERROR) << "pooling error error_code[" << error_code << "]";
     return RET_ERROR;
   }
-
-  if (in_data_type_ == kNumberTypeFloat32) {
-    context_->allocator->Free(fp16_input_);
-  }
-  if (out_data_type_ == kNumberTypeFloat32) {
-    auto out_ele_num = out_tensor->ElementsNum();
-    auto output_addr = reinterpret_cast<float *>(out_tensor->MutableData());
-    Float16ToFloat32(fp16_output_, output_addr, out_ele_num);
-    context_->allocator->Free(fp16_output_);
-  }
   return RET_OK;
 }
 

mindspore/lite/src/runtime/kernel/arm/fp16/reshape_fp16.cc

@@ -30,39 +30,12 @@ namespace mindspore::kernel {
 int ReshapeFp16CPUKernel::Run() {
   auto in_tensor = in_tensors_.at(kInputIndex);
   auto out_tensor = out_tensors_.at(kOutputIndex);
-  auto input_ptr = in_tensor->MutableData();
-  auto output_ptr = out_tensor->MutableData();
-  size_t data_size = out_tensor->Size();
 
-  auto in_datatype = in_tensor->data_type();
-  auto out_datatype = out_tensor->data_type();
-  if (in_datatype != out_datatype) {
-    if (in_datatype == kNumberTypeFloat32 && out_datatype == kNumberTypeFloat16) {
-      input_ptr = context_->allocator->Malloc(in_tensor->ElementsNum() * sizeof(float16_t));
-      if (input_ptr == nullptr) {
-        MS_LOG(ERROR) << "malloc in tensor fail!";
-        return mindspore::lite::RET_MEMORY_FAILED;
-      }
-      Float32ToFloat16(reinterpret_cast<float *>(in_tensor->MutableData()), reinterpret_cast<float16_t *>(input_ptr),
-                       in_tensor->ElementsNum());
-    } else if ((in_datatype == kNumberTypeFloat16 && out_datatype == kNumberTypeFloat32)) {
-      input_ptr = context_->allocator->Malloc(in_tensor->ElementsNum() * sizeof(float));
-      if (input_ptr == nullptr) {
-        MS_LOG(ERROR) << "malloc in tensor fail!";
-        return mindspore::lite::RET_MEMORY_FAILED;
-      }
-      Float16ToFloat32(reinterpret_cast<float16_t *>(in_tensor->MutableData()), reinterpret_cast<float *>(input_ptr),
-                       in_tensor->ElementsNum());
-    } else {
-      MS_LOG(ERROR) << "unsupported data type, in_datatype: " << in_datatype << ",out_datatype: " << out_datatype;
-      return RET_ERROR;
-    }
-  }
+  float16_t *input_ptr = reinterpret_cast<float16_t *>(in_tensor->data_c());
+  float16_t *output_ptr = reinterpret_cast<float16_t *>(out_tensor->data_c());
+
+  Reshape(input_ptr, output_ptr, out_tensor->Size());
 
-  Reshape(input_ptr, output_ptr, data_size);
-  if (in_datatype != out_datatype) {
-    context_->allocator->Free(input_ptr);
-  }
   return RET_OK;
 }
 

mindspore/lite/src/runtime/kernel/arm/fp16/slice_fp16.cc

@@ -16,7 +16,6 @@
 #include "src/runtime/kernel/arm/fp16/slice_fp16.h"
 #include "src/runtime/kernel/arm/fp16/common_fp16.h"
 #include "src/kernel_registry.h"
-#include "nnacl/fp16/cast_fp16.h"
 #include "nnacl/fp16/slice_fp16.h"
 
 using mindspore::lite::KernelRegistrar;
@@ -31,13 +30,12 @@ int SliceFp16CPUKernel::SliceParallelRun(int thread_id) {
 }
 
 int SliceFp16CPUKernel::Run() {
-  input_fp16_ = ConvertInputFp32toFp16(in_tensors_.at(0), context_);
-  output_fp16_ = MallocOutputFp16(out_tensors_.at(0), context_);
-  if (input_fp16_ == nullptr || output_fp16_ == nullptr) {
-    FreeInputAndOutput();
-    MS_LOG(ERROR) << "input or output is nullptr";
-    return RET_ERROR;
-  }
+  auto input_tensor = in_tensors_.at(0);
+  auto output_tensor = out_tensors_.at(0);
+
+  input_fp16_ = reinterpret_cast<float16_t *>(input_tensor->data_c());
+  output_fp16_ = reinterpret_cast<float16_t *>(output_tensor->data_c());
+
   if (param_->size_[1] < op_parameter_->thread_num_) {
     DoSliceFp16NoParallel(input_fp16_, output_fp16_, param_);
     return RET_OK;
@@ -46,24 +44,8 @@ int SliceFp16CPUKernel::Run() {
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "slice launch fail!ret: " << ret;
   }
-  if (out_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    Float16ToFloat32(output_fp16_, reinterpret_cast<float *>(out_tensors_.at(0)->MutableData()),
-                     out_tensors_.at(0)->ElementsNum());
-  }
-  FreeInputAndOutput();
   return ret;
 }
 
-void SliceFp16CPUKernel::FreeInputAndOutput() {
-  if (in_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    context_->allocator->Free(input_fp16_);
-    input_fp16_ = nullptr;
-  }
-  if (out_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    context_->allocator->Free(output_fp16_);
-    output_fp16_ = nullptr;
-  }
-}
-
 REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Slice, LiteKernelCreator<SliceFp16CPUKernel>)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp16/slice_fp16.h

@@ -32,7 +32,6 @@ class SliceFp16CPUKernel : public SliceCPUKernel {
   int SliceParallelRun(int thread_id) override;
 
  protected:
-  void FreeInputAndOutput();
   float16_t *input_fp16_ = nullptr;
   float16_t *output_fp16_ = nullptr;
 };

mindspore/lite/src/runtime/kernel/arm/fp16/softmax_fp16.cc

@@ -69,17 +69,6 @@ int SoftmaxFp16CPUKernel::MallocTmpBuffer() {
     return RET_ERROR;
   }
   memset(sum_data_, 0, out_plane_size * in_plane_size * sizeof(float16_t));
-
-  input_fp16_ = ConvertInputFp32toFp16(in_tensors_.at(kInputIndex), context_);
-  if (input_fp16_ == nullptr) {
-    MS_LOG(ERROR) << "malloc data failed";
-    return RET_ERROR;
-  }
-  output_fp16_ = MallocOutputFp16(out_tensors_.at(kOutputIndex), context_);
-  if (output_fp16_ == nullptr) {
-    MS_LOG(ERROR) << "malloc data failed";
-    return RET_ERROR;
-  }
   return RET_OK;
 }
 
@@ -88,19 +77,6 @@ void SoftmaxFp16CPUKernel::FreeTmpBuffer() {
     context_->allocator->Free(sum_data_);
     sum_data_ = nullptr;
   }
-  if (in_tensors_.at(kInputIndex)->data_type() == kNumberTypeFloat32) {
-    if (input_fp16_ != nullptr) {
-      context_->allocator->Free(input_fp16_);
-      input_fp16_ = nullptr;
-    }
-  }
-
-  if (out_tensors_.at(kOutputIndex)->data_type() == kNumberTypeFloat32) {
-    if (output_fp16_ != nullptr) {
-      context_->allocator->Free(output_fp16_);
-      output_fp16_ = nullptr;
-    }
-  }
 }
 
 int SoftmaxFp16CPUKernel::Run() {
@@ -110,11 +86,15 @@ int SoftmaxFp16CPUKernel::Run() {
     MS_LOG(ERROR) << "MallocTmpBuffer failed";
     return RET_ERROR;
   }
+
+  auto input_tensor = in_tensors_.at(0);
+  auto output_tensor = out_tensors_.at(0);
+
+  input_fp16_ = reinterpret_cast<float16_t *>(input_tensor->data_c());
+  output_fp16_ = reinterpret_cast<float16_t *>(output_tensor->data_c());
+
   SoftmaxFp16(input_fp16_, output_fp16_, sum_data_, softmax_param_);
-  auto out_tensor = out_tensors_.at(kOutputIndex);
-  if (out_tensor->data_type() == kNumberTypeFloat32) {
-    Float16ToFloat32(output_fp16_, reinterpret_cast<float *>(out_tensor->MutableData()), out_tensor->ElementsNum());
-  }
+
   FreeTmpBuffer();
   return RET_OK;
 }

mindspore/lite/src/runtime/kernel/arm/fp16/split_fp16.cc

@@ -17,7 +17,6 @@
 #include "src/runtime/kernel/arm/fp16/common_fp16.h"
 #include "src/runtime/kernel/arm/base/split_base.h"
 #include "nnacl/fp16/split_fp16.h"
-#include "nnacl/fp16/cast_fp16.h"
 #include "nnacl/split.h"
 #include "nnacl/split_parameter.h"
 #include "src/kernel_registry.h"
@@ -31,7 +30,6 @@ using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_Split;
 
 namespace mindspore::kernel {
-
 int SplitFp16CPUKernel::Init() {
   auto ret = SplitBaseCPUKernel::Init();
   if (ret != RET_OK) {
@@ -76,45 +74,19 @@ static int SplitFp16Run(void *cdata, int task_id) {
 }
 
 int SplitFp16CPUKernel::Run() {
-  input_ptr_ = ConvertInputFp32toFp16(in_tensors_.at(0), context_);
-  if (input_ptr_ == nullptr) {
-    MS_LOG(ERROR) << "input or output is nullptr";
-    return RET_ERROR;
-  }
+  auto input_tensor = in_tensors_.at(0);
+  input_ptr_ = reinterpret_cast<float16_t *>(input_tensor->data_c());
+
   for (int i = 0; i < param->num_split_; i++) {
-    output_ptr_.at(i) = MallocOutputFp16(out_tensors_.at(i), context_);
-    if (output_ptr_.at(i) == nullptr) {
-      FreeInputAndOutput();
-      MS_LOG(ERROR) << "input or output is nullptr";
-      return RET_ERROR;
-    }
+    auto output_tensor = out_tensors_.at(i);
+    output_ptr_.at(i) = reinterpret_cast<float16_t *>(output_tensor->data_c());
   }
+
   auto ret = ParallelLaunch(this->context_->thread_pool_, SplitFp16Run, this, thread_n_num_);
-  for (int i = 0; i < param->num_split_; i++) {
-    if (out_tensors_.at(i)->data_type() == kNumberTypeFloat32) {
-      Float16ToFloat32(output_ptr_.at(i), reinterpret_cast<float *>(out_tensors_.at(i)->MutableData()),
-                       out_tensors_.at(i)->ElementsNum());
-    }
-  }
-  FreeInputAndOutput();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "split error error_code[" << ret << "]";
   }
   return ret;
 }
-
-void SplitFp16CPUKernel::FreeInputAndOutput() {
-  if (in_tensors_.at(0)->data_type() == kNumberTypeFloat32) {
-    context_->allocator->Free(input_ptr_);
-    input_ptr_ = nullptr;
-  }
-  for (int i = 0; i < param->num_split_; i++) {
-    if (out_tensors_.at(i)->data_type() == kNumberTypeFloat32) {
-      context_->allocator->Free(output_ptr_.at(i));
-      output_ptr_.at(i) = nullptr;
-    }
-  }
-}
-
 REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Split, LiteKernelCreator<SplitFp16CPUKernel>)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp16/split_fp16.h

@@ -39,7 +39,6 @@ class SplitFp16CPUKernel : public SplitBaseCPUKernel {
  private:
   float16_t *input_ptr_ = nullptr;
   std::vector<float16_t *> output_ptr_;
-  void FreeInputAndOutput();
 };
 }  // namespace mindspore::kernel
 

mindspore/lite/src/runtime/kernel/arm/fp32/gather_fp32.cc

@@ -41,12 +41,6 @@ int GatherCPUKernel::DoGather(int task_id) {
   auto indices_tensor = in_tensors_.at(1);
   auto out_tensor = out_tensors_.at(0);
 
-  auto input_ptr = reinterpret_cast<float *>(input_tensor->MutableData());
-  auto output_ptr = reinterpret_cast<float *>(out_tensor->MutableData());
-
-  auto input_int32 = reinterpret_cast<int32_t *>(input_tensor->MutableData());
-  auto output_int32 = reinterpret_cast<int32_t *>(out_tensor->MutableData());
-
   auto in_shape = input_tensor->shape();
   int in_rank = in_shape.size();
   int indices_element_size = indices_tensor->ElementsNum();
@@ -65,16 +59,15 @@ int GatherCPUKernel::DoGather(int task_id) {
   int count = MSMIN(stride, outer_size - stride * task_id);
   auto thread_stride = stride * task_id;
 
-  int error_code;
-  if (input_tensor->data_type() == kNumberTypeInt32) {
-    input_int32 += thread_stride * limit;
-    output_int32 += thread_stride * indices_element_size;
-    error_code = GatherInt32(input_int32, count, inner_size, limit, indices_data_, indices_element_size, output_int32);
-  } else {
-    input_ptr += thread_stride * limit;
-    output_ptr += thread_stride * indices_element_size;
-    error_code = GatherFp32(input_ptr, count, inner_size, limit, indices_data_, indices_element_size, output_ptr);
-  }
+  int8_t *int8_in = reinterpret_cast<int8_t *>(input_tensor->data_c());
+  int8_t *int8_out = reinterpret_cast<int8_t *>(out_tensor->data_c());
+
+  int data_size = lite::DataTypeSize(input_tensor->data_type());
+  int8_in += thread_stride * limit * data_size;
+  int8_out += thread_stride * indices_element_size * data_size;
+
+  int error_code = Gather(int8_in, count, inner_size, limit, indices_data_, indices_element_size, int8_out, data_size);
+
   return error_code;
 }