pad_fp16_support_reflect

mindspore/lite/nnacl/fp16/pad_fp16.c

@@ -33,3 +33,10 @@ void PadFp16(const float16_t *input_data, float16_t *output_data, const int *inp
     }
   }
 }
+
+void MirrorPadFp16(const float16_t *input_data, float16_t *output_data, const int *input_shape,
+                   const PadParameter *pad_param, int begin, int end) {
+  for (int i = begin; i < end; ++i) {
+    output_data[i] = input_data[GetInputFlattenIndex(i, input_shape, pad_param)];
+  }
+}

mindspore/lite/nnacl/fp16/pad_fp16.h

@@ -19,12 +19,15 @@
 #ifdef ENABLE_NEON
 #include <arm_neon.h>
 #endif
+#include "nnacl/fp32/pad.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 void PadFp16(const float16_t *input_data, float16_t *output_data, const int *input_shape, const int *output_shape,
              const int *paddings, const int tid, const int thread_num);
+void MirrorPadFp16(const float16_t *input_data, float16_t *output_data, const int *input_shape,
+                   const PadParameter *pad_param, int begin, int end);
 #ifdef __cplusplus
 }
 #endif

mindspore/lite/nnacl/fp32/pad.h

@@ -31,6 +31,9 @@ void Pad(const float *input_data, float *output_data, const int *input_shape, co
          const int *paddings, const int tid, const int thread_num);
 void MirrorPad(const float *input_data, float *output_data, const int *input_shape, const PadParameter *pad_param,
                int begin, int end);
+
+int TransOut2InputDimIndex(int out_dim_index, int left_pad, int in_dim, int offset);
+int GetInputFlattenIndex(int out_flatten_index, const int *input_shape, const PadParameter *pad_param);
 #ifdef __cplusplus
 }
 #endif

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

@@ -32,6 +32,14 @@ int PadFp16CPUKernel::RunImpl(int task_id) {
   return RET_OK;
 }
 
+int PadFp16CPUKernel::RunMirrorPadImpl(int task_id) {
+  int unit = UP_DIV(out_tensors_.at(0)->ElementsNum(), context_->thread_num_);
+  int begin = unit * task_id;
+  int end = MSMIN(begin + unit, out_tensors_.at(0)->ElementsNum());
+  MirrorPadFp16(input_, output_, in_, pad_param_, begin, end);
+  return RET_OK;
+}
+
 int PadFp16CPUKernel::Run() {
   auto input_tensor = in_tensors_.at(0);
   auto output_tensor = out_tensors_.at(0);
@@ -45,18 +53,26 @@ int PadFp16CPUKernel::Run() {
     MS_LOG(ERROR) << "input or output is nullptr";
     return RET_ERROR;
   }
-
-  if (pad_param_->constant_value_ - 0.0f < 1e-5) {
-    memset(output_, 0, output_tensor->ElementsNum() * sizeof(float16_t));
+  int ret = 0;
+  if (pad_param_->pad_mode_ == static_cast<int>(schema::PaddingMode_CONSTANT)) {
+    if (pad_param_->constant_value_ - 0.0f < 1e-5) {
+      memset(output_, 0, output_tensor->ElementsNum() * sizeof(float16_t));
+    } else {
+      for (int i = 0; i < output_tensor->ElementsNum(); ++i) {
+        output_[i] = pad_param_->constant_value_;
+      }
+    }
+    ret = ParallelLaunch(this->context_->thread_pool_, PadImpl, this, op_parameter_->thread_num_);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "BatchnormRun error error_code[" << ret << "]";
+    }
   } else {
-    for (int i = 0; i < output_tensor->ElementsNum(); ++i) {
-      output_[i] = pad_param_->constant_value_;
+    HandleMirrorPad();
+    ret = ParallelLaunch(this->context_->thread_pool_, MirrorPadImpl, this, context_->thread_num_);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "Pad Reflect or Symmetric mode run error, error_code[" << ret << "]";
     }
   }
-  auto ret = ParallelLaunch(this->context_->thread_pool_, PadImpl, this, op_parameter_->thread_num_);
-  if (ret != RET_OK) {
-    MS_LOG(ERROR) << "BatchnormRun error error_code[" << ret << "]";
-  }
   if (is_output_fp32_) {
     Float16ToFloat32(output_, reinterpret_cast<float *>(output_tensor->MutableData()), output_tensor->ElementsNum());
   }

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

@@ -32,6 +32,7 @@ class PadFp16CPUKernel : public PadCPUKernel {
 
   int Run() override;
   int RunImpl(int task_id) override;
+  int RunMirrorPadImpl(int task_id) override;
 
  private:
   void FreeInputAndOutput();

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

@@ -38,10 +38,9 @@ class PadCPUKernel : public LiteKernel {
   int ReSize() override;
   int Run() override;
   virtual int RunImpl(int task_id);
-  int RunMirrorPadImpl(int task_id);
+  virtual int RunMirrorPadImpl(int task_id);
 
  private:
-  int HandleMirrorPad();
   int CheckPaddings(int *paddings, int length, int *input_shape, int mode);
   int CopyPaddingFromInput();
   void CalculateStrides();
@@ -49,6 +48,7 @@ class PadCPUKernel : public LiteKernel {
   int ExtendPaddings(int *paddings, int length, const int *ori_paddings, int ori_length);
 
  protected:
+  int HandleMirrorPad();
   PadParameter *pad_param_ = nullptr;
   int in_[4] = {0};
   int out_[4] = {0};