!7686 [MS][LITE][CPU]fix bug of deDepthwise int8 bug

Merge pull request !7686 from fuzhiye/tmp

mindspore/lite/nnacl/int8/conv_int8.c

@@ -62,9 +62,9 @@ void Conv3x3Int8Gemm(int32_t *dst, const int16_t *src, const int16_t *weight, in
 #endif
 }
 
-void ConvInt8Opt(int8_t *input_data, int8_t *packed_input, int8_t *matmul_input, int8_t *packed_weight,
-                 const int32_t *bias_data, int8_t *output_data, int32_t *filter_zp, int32_t *input_sum, int task_id,
-                 ConvParameter *conv_param, MATMUL_OPT_R_FUNC matmul_func, bool is_optimize) {
+void ConvInt8(int8_t *input_data, int8_t *packed_input, int8_t *matmul_input, int8_t *packed_weight,
+              const int32_t *bias_data, int8_t *output_data, int32_t *filter_zp, int32_t *input_sum, int task_id,
+              ConvParameter *conv_param, MATMUL_OPT_R_FUNC matmul_func, bool is_optimize) {
   int kernel_h = conv_param->kernel_h_;
   int kernel_w = conv_param->kernel_w_;
   int in_batch = conv_param->input_batch_;

mindspore/lite/nnacl/int8/conv_int8.h

@@ -32,9 +32,9 @@
 extern "C" {
 #endif
 // int8 conv common
-void ConvInt8Opt(int8_t *input_data, int8_t *packed_input, int8_t *matmul_input, int8_t *packed_weight,
-                 const int32_t *bias_data, int8_t *output_data, int32_t *filter_zp, int32_t *input_sum, int task_id,
-                 ConvParameter *conv_param, MATMUL_OPT_R_FUNC matmul_func, bool is_optimize);
+void ConvInt8(int8_t *input_data, int8_t *packed_input, int8_t *matmul_input, int8_t *packed_weight,
+              const int32_t *bias_data, int8_t *output_data, int32_t *filter_zp, int32_t *input_sum, int task_id,
+              ConvParameter *conv_param, MATMUL_OPT_R_FUNC matmul_func, bool is_optimize);
 
 // int8 convolution 1x1
 void Conv1x1PreOptPeroc(const int8_t *src_input, int8_t *packed_input, int32_t *input_sum, size_t input_channel,

mindspore/lite/src/runtime/kernel/arm/base/convolution_base.cc

@@ -141,34 +141,6 @@ int ConvolutionBaseCPUKernel::SetIfPerChannel() {
   return RET_OK;
 }
 
-int ConvolutionBaseCPUKernel::SetIfAsymmetric() {
-  uint8_t asymmetric = 0b0;
-  auto filter_tensor = in_tensors_.at(kWeightIndex);
-  auto filter_ele_num = filter_tensor->ElementsNum();
-  auto filter_data = reinterpret_cast<int8_t *>(filter_tensor->MutableData());
-  int min_value = INT8_MAX;
-  int max_value = INT8_MIN;
-  for (int i = 0; i < filter_ele_num; ++i) {
-    min_value = min_value < filter_data[i] ? min_value : filter_data[i];
-    max_value = max_value > filter_data[i] ? max_value : filter_data[i];
-  }
-  if (conv_quant_arg_->filter_arg_num_ == kPerTensor) {
-    auto filter_zp = conv_quant_arg_->filter_quant_args_[0].zp_;
-    if (filter_zp != 0 && min_value >= -128 && max_value <= 127) {
-      asymmetric = asymmetric | FILTER_ASYMMETRIC;
-    }
-  } else {
-    auto filter_arg = conv_quant_arg_->filter_quant_args_;
-    for (int i = 0; i < conv_param_->output_channel_; ++i) {
-      if (filter_arg[i].zp_ != 0 && min_value >= -128 && max_value <= 127) {
-        asymmetric = asymmetric | FILTER_ASYMMETRIC;
-      }
-    }
-  }
-  conv_quant_arg_->asymmetric_ = asymmetric;
-  return RET_OK;
-}
-
 int ConvolutionBaseCPUKernel::MallocQuantParam() {
   conv_quant_arg_ = &conv_param_->conv_quant_arg_;
   auto input_tensor = in_tensors_.at(kInputIndex);

mindspore/lite/src/runtime/kernel/arm/base/convolution_base.h

@@ -48,7 +48,6 @@ class ConvolutionBaseCPUKernel : public LiteKernel {
   int Init() override;
   int ReSize() override { return 0; }
   int Run() override { return 0; }
-  int SetIfAsymmetric();
   int SetIfPerChannel();
   int MallocQuantParam();
   int SetQuantParam();

mindspore/lite/src/runtime/kernel/arm/int8/convolution_3x3_int8.cc

@@ -156,14 +156,7 @@ int Convolution3x3Int8CPUKernel::InitTmpBuffer() {
   return RET_OK;
 }
 
-void Convolution3x3Int8CPUKernel::ConfigInputOutput() {
-  auto output_tensor = out_tensors_.at(kOutputIndex);
-  output_tensor->SetFormat(schema::Format::Format_NHWC);
-}
-
 int Convolution3x3Int8CPUKernel::Init() {
-  // config input output
-  ConfigInputOutput();
   auto ret = SetQuantParam();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Set quant param failed.";

mindspore/lite/src/runtime/kernel/arm/int8/convolution_3x3_int8.h

@@ -38,7 +38,6 @@ class Convolution3x3Int8CPUKernel : public ConvolutionBaseCPUKernel {
   int RunImpl(int task_id);
   int InitWeightBias();
   int InitTmpBuffer();
-  void ConfigInputOutput();
 
  private:
   void FreeTmpBuffer();

mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.cc

@@ -59,7 +59,7 @@ void ConvolutionInt8CPUKernel::CheckSupportOptimize() {
   conv_param_->tile_num_ = tile_num_;
 }
 
-int ConvolutionInt8CPUKernel::InitWeightBiasOpt() {
+int ConvolutionInt8CPUKernel::InitWeightBias() {
   auto filter_tensor = in_tensors_.at(kWeightIndex);
   auto input_channel = filter_tensor->Channel();
   auto output_channel = filter_tensor->Batch();
@@ -83,7 +83,7 @@ int ConvolutionInt8CPUKernel::InitWeightBiasOpt() {
   }
 #endif
   int pack_weight_size = up_round_oc * up_round_deep;
-  int bias_size = up_round_oc * sizeof(int32_t);
+  size_t bias_size = up_round_oc * sizeof(int32_t);
   int32_t input_zp = conv_param_->conv_quant_arg_.input_quant_args_[0].zp_;
 
   // init weight
@@ -150,7 +150,7 @@ int ConvolutionInt8CPUKernel::InitWeightBiasOpt() {
   return RET_OK;
 }
 
-int ConvolutionInt8CPUKernel::InitTmpBufferOpt() {
+int ConvolutionInt8CPUKernel::InitTmpBuffer() {
   MS_ASSERT(ctx_->allocator != nullptr);
   int kernel_plane = conv_param_->kernel_h_ * conv_param_->kernel_w_;
   int tmp_size;
@@ -181,7 +181,7 @@ int ConvolutionInt8CPUKernel::Init() {
     return ret;
   }
 
-  ret = InitWeightBiasOpt();
+  ret = InitWeightBias();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Initialization for optimized int8 conv failed.";
     return RET_ERROR;
@@ -212,9 +212,8 @@ int ConvolutionInt8CPUKernel::RunImpl(int task_id) {
   auto input_tensor = in_tensors_.at(kInputIndex);
   auto ori_input_data = reinterpret_cast<int8_t *>(input_tensor->MutableData());
   auto output_addr = reinterpret_cast<int8_t *>(out_tensors_.at(kOutputIndex)->MutableData());
-  ConvInt8Opt(ori_input_data, packed_input_, matmul_packed_input_, packed_weight_,
-              reinterpret_cast<int32_t *>(bias_data_), output_addr, filter_zp_ptr_, input_sum_, task_id, conv_param_,
-              matmul_func_, support_optimize_);
+  ConvInt8(ori_input_data, packed_input_, matmul_packed_input_, packed_weight_, reinterpret_cast<int32_t *>(bias_data_),
+           output_addr, filter_zp_ptr_, input_sum_, task_id, conv_param_, matmul_func_, support_optimize_);
   return RET_OK;
 }
 
@@ -229,7 +228,7 @@ int ConvolutionInt8Impl(void *cdata, int task_id) {
 }
 
 int ConvolutionInt8CPUKernel::Run() {
-  auto ret = InitTmpBufferOpt();
+  auto ret = InitTmpBuffer();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Init tmp buffer failed.";
     return RET_ERROR;

mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.h

@@ -51,8 +51,8 @@ class ConvolutionInt8CPUKernel : public ConvolutionBaseCPUKernel {
   int Run() override;
   int RunImpl(int task_id);
   void CheckSupportOptimize();
-  int InitWeightBiasOpt();
-  int InitTmpBufferOpt();
+  int InitWeightBias();
+  int InitTmpBuffer();
 
  private:
   void FreeTmpBuffer() {

mindspore/lite/src/runtime/kernel/arm/int8/deconvolution_depthwise_int8.cc

@@ -149,12 +149,6 @@ int DeconvolutionDepthwiseInt8CPUKernel::Init() {
 int DeconvolutionDepthwiseInt8CPUKernel::ReSize() {
   InitSlideParam();
   ConvolutionBaseCPUKernel::Init();
-
-  auto ret = InitBuffer();
-  if (ret != RET_OK) {
-    MS_LOG(ERROR) << "Deconv Depthwise int8 InitBuffer error!";
-    return ret;
-  }
   return RET_OK;
 }
 
@@ -179,6 +173,11 @@ int DeconvolutionDepthwiseInt8CPUKernel::Run() {
     MS_LOG(ERROR) << "Only support input channel equals output channel.";
     return RET_ERROR;
   }
+  auto ret = InitBuffer();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Deconv Depthwise int8 InitBuffer error!";
+    return ret;
+  }
 
   // pack input, assume input format: NHWC -> NHWC4
   auto input_tensor = in_tensors_.at(kInputIndex);
@@ -191,7 +190,7 @@ int DeconvolutionDepthwiseInt8CPUKernel::Run() {
     packed_output_ = output_addr;
   }
 
-  auto ret = ParallelLaunch(this->context_->thread_pool_, DeconvDwInt8Run, this, conv_param_->thread_num_);
+  ret = ParallelLaunch(this->context_->thread_pool_, DeconvDwInt8Run, this, conv_param_->thread_num_);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "DeconvDwInt8Run error: error_code[" << ret << "]";
     return RET_ERROR;