!5274 [MS][LITE] arm cpu op conv depthwise, deconv depthwise: malloc input and output temp buffer using allocator

Merge pull request !5274 from yangruoqi713/lite

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc

@@ -109,15 +109,15 @@ static int ConvDwFp16Run(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
 }
 
 int ConvolutionDepthwiseFp16CPUKernel::Run() {
+  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
+    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+    return RET_ERROR;
+  }
   auto ret = Prepare();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Prepare failed.";
     return RET_ERROR;
   }
-  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
-    MS_LOG(ERROR) << "Only support input channel equals output channel.";
-    return RET_ERROR;
-  }
 
   ret = ConvolutionBaseFP16CPUKernel::GetExecuteTensor();
   if (ret != RET_OK) {

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.cc

@@ -38,20 +38,6 @@ ConvolutionDepthwiseSWFp16CPUKernel::~ConvolutionDepthwiseSWFp16CPUKernel() {
     delete packed_weight_;
     packed_weight_ = nullptr;
   }
-  FreeTmpBuffer();
-}
-
-void ConvolutionDepthwiseSWFp16CPUKernel::FreeTmpBuffer() {
-  if (need_align_) {
-    if (packed_input_ != nullptr) {
-      delete packed_input_;
-      packed_input_ = nullptr;
-    }
-    if (packed_output_ != nullptr) {
-      delete packed_output_;
-      packed_output_ = nullptr;
-    }
-  }
 }
 
 int ConvolutionDepthwiseSWFp16CPUKernel::InitBuffer() {
@@ -59,14 +45,14 @@ int ConvolutionDepthwiseSWFp16CPUKernel::InitBuffer() {
     need_align_ = true;
     int C8 = UP_DIV(conv_param_->input_channel_, C8NUM);
     int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C8NUM * C8;
-    packed_input_ = reinterpret_cast<float16_t *>(malloc(pack_input_size * sizeof(float16_t)));
+    packed_input_ = reinterpret_cast<float16_t *>(context_->allocator->Malloc(pack_input_size * sizeof(float16_t)));
     if (packed_input_ == nullptr) {
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
     }
 
     int pack_output_size = conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * C8NUM * C8;
-    packed_output_ = reinterpret_cast<float16_t *>(malloc(pack_output_size * sizeof(float16_t)));
+    packed_output_ = reinterpret_cast<float16_t *>(context_->allocator->Malloc(pack_output_size * sizeof(float16_t)));
     if (packed_output_ == nullptr) {
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
@@ -129,18 +115,11 @@ int ConvolutionDepthwiseSWFp16CPUKernel::Init() {
 }
 
 int ConvolutionDepthwiseSWFp16CPUKernel::ReSize() {
-  FreeTmpBuffer();
   auto ret = ConvolutionBaseCPUKernel::Init();
   if (ret != RET_OK) {
     return ret;
   }
   InitSlidingParamConvDw(sliding_, conv_param_, C8NUM);
-
-  ret = InitBuffer();
-  if (ret != 0) {
-    MS_LOG(ERROR) << "Convolution depthwise fp16 InitBuffer failed.";
-    return RET_ERROR;
-  }
   return RET_OK;
 }
 
@@ -161,13 +140,19 @@ static int ConvDwSWFp16Run(int task_id, LiteParallelGroupEnv *penv, void *cdata)
 }
 
 int ConvolutionDepthwiseSWFp16CPUKernel::Run() {
+  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
+    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+    return RET_ERROR;
+  }
+
   auto ret = Prepare();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Prepare failed.";
     return RET_ERROR;
   }
-  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
-    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+  ret = InitBuffer();
+  if (ret != 0) {
+    MS_LOG(ERROR) << "Convolution depthwise fp16 InitBuffer failed.";
     return RET_ERROR;
   }
 
@@ -194,6 +179,8 @@ int ConvolutionDepthwiseSWFp16CPUKernel::Run() {
   if (need_align_) {
     PackNHWC8ToNHWCFp16(packed_output_, execute_output_, conv_param_->output_batch_,
                         conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
+    context_->allocator->Free(packed_input_);
+    context_->allocator->Free(packed_output_);
   }
   ConvolutionBaseFP16CPUKernel::IfCastOutput();
   ConvolutionBaseFP16CPUKernel::FreeTmpBuffer();

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.h

@@ -50,7 +50,6 @@ class ConvolutionDepthwiseSWFp16CPUKernel : public ConvolutionBaseFP16CPUKernel
   int Execute(int task_id);
 
  private:
-  void FreeTmpBuffer();
   SlidingWindowParam *sliding_ = nullptr;
   float16_t *packed_weight_ = nullptr;
   float16_t *packed_input_ = nullptr;

mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.cc

@@ -37,20 +37,6 @@ DeconvolutionDepthwiseFp16CPUKernel::~DeconvolutionDepthwiseFp16CPUKernel() {
     delete packed_weight_;
     packed_weight_ = nullptr;
   }
-  FreeTmpBuffer();
-}
-
-void DeconvolutionDepthwiseFp16CPUKernel::FreeTmpBuffer() {
-  if (need_align_) {
-    if (packed_input_ != nullptr) {
-      delete packed_input_;
-      packed_input_ = nullptr;
-    }
-    if (packed_output_ != nullptr) {
-      delete packed_output_;
-      packed_output_ = nullptr;
-    }
-  }
 }
 
 int DeconvolutionDepthwiseFp16CPUKernel::InitSlideParam() {
@@ -69,14 +55,14 @@ int DeconvolutionDepthwiseFp16CPUKernel::InitSlideParam() {
 int DeconvolutionDepthwiseFp16CPUKernel::InitBuffer() {
   int C8 = UP_DIV(conv_param_->input_channel_, C8NUM);
   int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C8NUM * C8;
-  packed_input_ = reinterpret_cast<float16_t *>(malloc(pack_input_size * sizeof(float16_t)));
+  packed_input_ = reinterpret_cast<float16_t *>(context_->allocator->Malloc(pack_input_size * sizeof(float16_t)));
   if (packed_input_ == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
   }
 
   int pack_output_size = conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * C8NUM * C8;
-  packed_output_ = reinterpret_cast<float16_t *>(malloc(pack_output_size * sizeof(float16_t)));
+  packed_output_ = reinterpret_cast<float16_t *>(context_->allocator->Malloc(pack_output_size * sizeof(float16_t)));
   if (packed_output_ == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
@@ -137,17 +123,11 @@ int DeconvolutionDepthwiseFp16CPUKernel::Init() {
 }
 
 int DeconvolutionDepthwiseFp16CPUKernel::ReSize() {
-  FreeTmpBuffer();
   InitSlideParam();
   auto ret = ConvolutionBaseCPUKernel::Init();
   if (ret != RET_OK) {
     return ret;
   }
-  ret = InitBuffer();
-  if (ret != 0) {
-    MS_LOG(ERROR) << "Deconvolution depthwise fp16 InitBuffer failed.";
-    return RET_ERROR;
-  }
   return RET_OK;
 }
 
@@ -168,13 +148,18 @@ static int DeconvDwFp16Run(int task_id, LiteParallelGroupEnv *penv, void *cdata)
 }
 
 int DeconvolutionDepthwiseFp16CPUKernel::Run() {
+  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
+    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+    return RET_ERROR;
+  }
   auto ret = Prepare();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Prepare failed.";
     return RET_ERROR;
   }
-  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
-    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+  ret = InitBuffer();
+  if (ret != 0) {
+    MS_LOG(ERROR) << "Deconvolution depthwise fp16 InitBuffer failed.";
     return RET_ERROR;
   }
 
@@ -202,6 +187,8 @@ int DeconvolutionDepthwiseFp16CPUKernel::Run() {
   if (need_align_) {
     PackNHWC8ToNHWCFp16(packed_output_, execute_output_, conv_param_->output_batch_,
                         conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
+    context_->allocator->Free(packed_input_);
+    context_->allocator->Free(packed_output_);
   }
   ConvolutionBaseFP16CPUKernel::IfCastOutput();
   ConvolutionBaseFP16CPUKernel::FreeTmpBuffer();

mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_depthwise_fp16.h

@@ -52,7 +52,6 @@ class DeconvolutionDepthwiseFp16CPUKernel : public ConvolutionBaseFP16CPUKernel
   int Execute(int task_id);
 
  private:
-  void FreeTmpBuffer();
   SlidingWindowParam *sliding_ = nullptr;
   float16_t *packed_weight_ = nullptr;
   float16_t *packed_input_ = nullptr;

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

@@ -100,15 +100,16 @@ int ConvDwRun(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
 }
 
 int ConvolutionDepthwiseCPUKernel::Run() {
+  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
+    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+    return RET_ERROR;
+  }
   auto ret = Prepare();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Prepare failed.";
     return ret;
   }
-  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
-    MS_LOG(ERROR) << "Only support input channel equals output channel.";
-    return RET_ERROR;
-  }
+
   auto input_tensor = in_tensors_.at(kInputIndex);
   input_ptr_ = reinterpret_cast<float *>(input_tensor->Data());
 

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

@@ -36,20 +36,6 @@ ConvolutionDepthwiseSWCPUKernel::~ConvolutionDepthwiseSWCPUKernel() {
     delete packed_weight_;
     packed_weight_ = nullptr;
   }
-  FreeTmpBuffer();
-}
-
-void ConvolutionDepthwiseSWCPUKernel::FreeTmpBuffer() {
-  if (need_align_) {
-    if (packed_input_ != nullptr) {
-      delete packed_input_;
-      packed_input_ = nullptr;
-    }
-    if (packed_output_ != nullptr) {
-      delete packed_output_;
-      packed_output_ = nullptr;
-    }
-  }
 }
 
 int ConvolutionDepthwiseSWCPUKernel::InitWeightBias() {
@@ -89,7 +75,7 @@ int ConvolutionDepthwiseSWCPUKernel::InitBuffer() {
     need_align_ = true;
     int IC4 = UP_DIV(conv_param_->input_channel_, C4NUM);
     int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C4NUM * IC4;
-    packed_input_ = reinterpret_cast<float *>(malloc(pack_input_size * sizeof(float)));
+    packed_input_ = reinterpret_cast<float *>(context_->allocator->Malloc(pack_input_size * sizeof(float)));
     if (packed_input_ == nullptr) {
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
@@ -97,7 +83,7 @@ int ConvolutionDepthwiseSWCPUKernel::InitBuffer() {
 
     int OC4 = UP_DIV(conv_param_->output_channel_, C4NUM);
     int pack_output_size = conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * C4NUM * OC4;
-    packed_output_ = reinterpret_cast<float *>(malloc(pack_output_size * sizeof(float)));
+    packed_output_ = reinterpret_cast<float *>(context_->allocator->Malloc(pack_output_size * sizeof(float)));
     if (packed_output_ == nullptr) {
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
@@ -125,16 +111,9 @@ int ConvolutionDepthwiseSWCPUKernel::Init() {
 }
 
 int ConvolutionDepthwiseSWCPUKernel::ReSize() {
-  FreeTmpBuffer();
   ConvolutionBaseCPUKernel::Init();
   InitSlidingParamConvDw(sliding_, conv_param_, C4NUM);
   conv_param_->thread_num_ = MSMIN(thread_count_, conv_param_->output_h_);
-
-  auto ret = InitBuffer();
-  if (ret != 0) {
-    MS_LOG(ERROR) << "Convolution depthwise fp32 InitBuffer failed.";
-    return RET_ERROR;
-  }
   return RET_OK;
 }
 
@@ -155,13 +134,20 @@ int ConvDwSWRun(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
 }
 
 int ConvolutionDepthwiseSWCPUKernel::Run() {
+  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
+    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+    return RET_ERROR;
+  }
+
   auto ret = Prepare();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Prepare failed.";
     return ret;
   }
-  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
-    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+
+  ret = InitBuffer();
+  if (ret != 0) {
+    MS_LOG(ERROR) << "Convolution depthwise fp32 InitBuffer failed.";
     return RET_ERROR;
   }
   auto input_tensor = in_tensors_.at(kInputIndex);
@@ -190,7 +176,10 @@ int ConvolutionDepthwiseSWCPUKernel::Run() {
   if (need_align_) {
     PackNHWC4ToNHWCFp32(packed_output_, output_ptr, conv_param_->output_batch_,
                         conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
+    context_->allocator->Free(packed_input_);
+    context_->allocator->Free(packed_output_);
   }
+
   return RET_OK;
 }
 }  // namespace mindspore::kernel

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

@@ -40,7 +40,6 @@ class ConvolutionDepthwiseSWCPUKernel : public ConvolutionBaseCPUKernel {
   int Execute(int task_id);
 
  private:
-  void FreeTmpBuffer();
   SlidingWindowParam *sliding_ = nullptr;
   float *packed_weight_ = nullptr;
   float *packed_input_ = nullptr;

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

@@ -36,20 +36,6 @@ DeconvolutionDepthwiseCPUKernel::~DeconvolutionDepthwiseCPUKernel() {
     delete packed_weight_;
     packed_weight_ = nullptr;
   }
-  FreeTmpBuffer();
-}
-
-void DeconvolutionDepthwiseCPUKernel::FreeTmpBuffer() {
-  if (need_align_) {
-    if (packed_input_ != nullptr) {
-      delete packed_input_;
-      packed_input_ = nullptr;
-    }
-    if (packed_output_ != nullptr) {
-      delete packed_output_;
-      packed_output_ = nullptr;
-    }
-  }
 }
 
 int DeconvolutionDepthwiseCPUKernel::InitSlideParam() {
@@ -100,7 +86,7 @@ int DeconvolutionDepthwiseCPUKernel::InitBuffer() {
     need_align_ = true;
     int IC4 = UP_DIV(conv_param_->input_channel_, C4NUM);
     int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C4NUM * IC4;
-    packed_input_ = reinterpret_cast<float *>(malloc(pack_input_size * sizeof(float)));
+    packed_input_ = reinterpret_cast<float *>(context_->allocator->Malloc(pack_input_size * sizeof(float)));
     if (packed_input_ == nullptr) {
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
@@ -108,7 +94,7 @@ int DeconvolutionDepthwiseCPUKernel::InitBuffer() {
 
     int OC4 = UP_DIV(conv_param_->output_channel_, C4NUM);
     int pack_output_size = conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * C4NUM * OC4;
-    packed_output_ = reinterpret_cast<float *>(malloc(pack_output_size * sizeof(float)));
+    packed_output_ = reinterpret_cast<float *>(context_->allocator->Malloc(pack_output_size * sizeof(float)));
     if (packed_output_ == nullptr) {
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
@@ -137,15 +123,8 @@ int DeconvolutionDepthwiseCPUKernel::Init() {
 }
 
 int DeconvolutionDepthwiseCPUKernel::ReSize() {
-  FreeTmpBuffer();
   InitSlideParam();
   ConvolutionBaseCPUKernel::Init();
-
-  auto ret = InitBuffer();
-  if (ret != 0) {
-    MS_LOG(ERROR) << "Deconvolution depthwise fp32 InitBuffer failed.ret: " << ret;
-    return ret;
-  }
   return RET_OK;
 }
 
@@ -166,15 +145,23 @@ int DeconvDwRun(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
 }
 
 int DeconvolutionDepthwiseCPUKernel::Run() {
-  auto prepare_ret = Prepare();
-  if (prepare_ret != RET_OK) {
-    MS_LOG(ERROR) << "Prepare fail!ret: " << prepare_ret;
-    return prepare_ret;
-  }
   if (conv_param_->input_channel_ != conv_param_->output_channel_) {
     MS_LOG(ERROR) << "Only support input channel equals output channel.";
     return RET_ERROR;
   }
+
+  auto ret = Prepare();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare fail!ret: " << ret;
+    return ret;
+  }
+
+  ret = InitBuffer();
+  if (ret != 0) {
+    MS_LOG(ERROR) << "Deconvolution depthwise fp32 InitBuffer failed.ret: " << ret;
+    return ret;
+  }
+
   auto input_tensor = in_tensors_.at(kInputIndex);
   auto input_addr = reinterpret_cast<float *>(input_tensor->Data());
 
@@ -191,7 +178,7 @@ int DeconvolutionDepthwiseCPUKernel::Run() {
     packed_output_ = output_addr;
   }
 
-  auto ret = LiteBackendParallelLaunch(DeconvDwRun, this, conv_param_->thread_num_);
+  ret = LiteBackendParallelLaunch(DeconvDwRun, this, conv_param_->thread_num_);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "DeconvDwRun error: error_code[" << ret << "]";
     return RET_ERROR;
@@ -200,6 +187,8 @@ int DeconvolutionDepthwiseCPUKernel::Run() {
   if (need_align_) {
     PackNHWC4ToNHWCFp32(packed_output_, output_addr, conv_param_->output_batch_,
                         conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
+    context_->allocator->Free(packed_input_);
+    context_->allocator->Free(packed_output_);
   }
   return RET_OK;
 }

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

@@ -41,7 +41,6 @@ class DeconvolutionDepthwiseCPUKernel : public ConvolutionBaseCPUKernel {
   int Execute(int task_id);
 
  private:
-  void FreeTmpBuffer();
   SlidingWindowParam *sliding_ = nullptr;
   float *packed_weight_ = nullptr;
   float *packed_input_ = nullptr;

mindspore/lite/src/runtime/kernel/arm/int8/convolution_depthwise_int8.cc

@@ -28,19 +28,6 @@ using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_DepthwiseConv2D;
 
 namespace mindspore::kernel {
-void ConvolutionDepthwiseInt8CPUKernel::FreeTmpBuffer() {
-  if (packed_input_ != nullptr) {
-    free(packed_input_);
-    packed_input_ = nullptr;
-  }
-  if (need_align_) {
-    if (packed_output_ != nullptr) {
-      free(packed_output_);
-      packed_output_ = nullptr;
-    }
-  }
-}
-
 ConvolutionDepthwiseInt8CPUKernel::~ConvolutionDepthwiseInt8CPUKernel() {
   if (sliding != nullptr) {
     delete sliding;
@@ -50,7 +37,6 @@ ConvolutionDepthwiseInt8CPUKernel::~ConvolutionDepthwiseInt8CPUKernel() {
     free(packed_weight_);
     packed_weight_ = nullptr;
   }
-  FreeTmpBuffer();
   FreeQuantParam();
 }
 
@@ -88,7 +74,7 @@ int ConvolutionDepthwiseInt8CPUKernel::InitWeightBias() {
 int ConvolutionDepthwiseInt8CPUKernel::InitBuffer() {
   int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C4NUM *
                         UP_DIV(conv_param_->input_channel_, 4);
-  packed_input_ = reinterpret_cast<int16_t *>(malloc(pack_input_size * sizeof(int16_t)));
+  packed_input_ = reinterpret_cast<int16_t *>(context_->allocator->Malloc(pack_input_size * sizeof(int16_t)));
   if (packed_input_ == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
@@ -98,7 +84,7 @@ int ConvolutionDepthwiseInt8CPUKernel::InitBuffer() {
     need_align_ = true;
     int pack_output_size = conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * C4NUM *
                            UP_DIV(conv_param_->output_channel_, C4NUM);
-    packed_output_ = reinterpret_cast<int8_t *>(malloc(pack_output_size * sizeof(int8_t)));
+    packed_output_ = reinterpret_cast<int8_t *>(context_->allocator->Malloc(pack_output_size * sizeof(int8_t)));
     if (packed_input_ == nullptr) {
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
@@ -120,7 +106,6 @@ int ConvolutionDepthwiseInt8CPUKernel::Init() {
 }
 
 int ConvolutionDepthwiseInt8CPUKernel::ReSize() {
-  FreeTmpBuffer();
   ConvolutionBaseCPUKernel::Init();
   InitSlidingParamConvDw(sliding, conv_param_, C4NUM);
 
@@ -134,11 +119,6 @@ int ConvolutionDepthwiseInt8CPUKernel::ReSize() {
     MS_LOG(ERROR) << "Depthwise int8 InitWeightBias error!";
     return ret;
   }
-  ret = InitBuffer();
-  if (ret != RET_OK) {
-    MS_LOG(ERROR) << "Depthwise int8 ReSize error!";
-    return ret;
-  }
   return RET_OK;
 }
 
@@ -159,14 +139,20 @@ int ConvDwInt8Run(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
 }
 
 int ConvolutionDepthwiseInt8CPUKernel::Run() {
+  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
+    MS_LOG(ERROR) << "Only support input channel equals output channel.";
+    return RET_ERROR;
+  }
   auto ret = Prepare();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Prepare failed.";
     return RET_ERROR;
   }
-  if (conv_param_->input_channel_ != conv_param_->output_channel_) {
-    MS_LOG(ERROR) << "Only support input channel equals output channel.";
-    return RET_ERROR;
+
+  ret = InitBuffer();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Depthwise int8 ReSize error!";
+    return ret;
   }
 
   auto input_tensor = in_tensors_.at(kInputIndex);
@@ -187,7 +173,9 @@ int ConvolutionDepthwiseInt8CPUKernel::Run() {
   if (need_align_) {
     PackNHWC4ToNHWCInt8(packed_output_, output_addr, conv_param_->output_batch_,
                         conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
+    context_->allocator->Free(packed_output_);
   }
+  context_->allocator->Free(packed_input_);
   return RET_OK;
 }
 

mindspore/lite/src/runtime/kernel/arm/int8/convolution_depthwise_int8.h

@@ -40,7 +40,6 @@ class ConvolutionDepthwiseInt8CPUKernel : public ConvolutionBaseCPUKernel {
   int Execute(int task_id);
 
  private:
-  void FreeTmpBuffer();
   SlidingWindowParam *sliding = nullptr;
   int16_t *packed_weight_ = nullptr;
   int16_t *packed_input_ = nullptr;

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

@@ -37,27 +37,9 @@ DeconvolutionDepthwiseInt8CPUKernel::~DeconvolutionDepthwiseInt8CPUKernel() {
     delete packed_weight_;
     packed_weight_ = nullptr;
   }
-  FreeTmpBuffer();
   FreeQuantParam();
 }
 
-void DeconvolutionDepthwiseInt8CPUKernel::FreeTmpBuffer() {
-  if (packed_input_ != nullptr) {
-    delete packed_input_;
-    packed_input_ = nullptr;
-  }
-  if (need_align_) {
-    if (packed_output_ != nullptr) {
-      delete packed_output_;
-      packed_output_ = nullptr;
-    }
-  }
-  if (output_buffer_ != nullptr) {
-    delete output_buffer_;
-    output_buffer_ = nullptr;
-  }
-}
-
 int DeconvolutionDepthwiseInt8CPUKernel::InitWeightBias() {
   // init weight: int8 -> int16
   // o, h, w, i -> o/8, h, w, i, 8; o == group, i == 1
@@ -111,7 +93,7 @@ int DeconvolutionDepthwiseInt8CPUKernel::InitSlideParam() {
 int DeconvolutionDepthwiseInt8CPUKernel::InitBuffer() {
   int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C4NUM *
                         UP_DIV(conv_param_->input_channel_, 4);
-  packed_input_ = reinterpret_cast<int16_t *>(malloc(pack_input_size * sizeof(int16_t)));
+  packed_input_ = reinterpret_cast<int16_t *>(context_->allocator->Malloc(pack_input_size * sizeof(int16_t)));
   if (packed_input_ == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
@@ -121,7 +103,7 @@ int DeconvolutionDepthwiseInt8CPUKernel::InitBuffer() {
     need_align_ = true;
     int pack_output_size = conv_param_->output_batch_ * conv_param_->output_h_ * conv_param_->output_w_ * C4NUM *
                            UP_DIV(conv_param_->output_channel_, C4NUM);
-    packed_output_ = reinterpret_cast<int8_t *>(malloc(pack_output_size * sizeof(int8_t)));
+    packed_output_ = reinterpret_cast<int8_t *>(context_->allocator->Malloc(pack_output_size * sizeof(int8_t)));
     if (packed_output_ == nullptr) {
       MS_LOG(ERROR) << "Malloc buffer failed.";
       return RET_ERROR;
@@ -129,8 +111,8 @@ int DeconvolutionDepthwiseInt8CPUKernel::InitBuffer() {
     memset(packed_output_, 0, pack_output_size * sizeof(int8_t));
   }
 
-  output_buffer_ =
-    reinterpret_cast<int32_t *>(malloc(conv_param_->output_h_ * conv_param_->output_w_ * C4NUM * sizeof(int32_t)));
+  output_buffer_ = reinterpret_cast<int32_t *>(
+    context_->allocator->Malloc(conv_param_->output_h_ * conv_param_->output_w_ * C4NUM * sizeof(int32_t)));
   if (output_buffer_ == nullptr) {
     MS_LOG(ERROR) << "Malloc buffer failed.";
     return RET_ERROR;
@@ -165,7 +147,6 @@ int DeconvolutionDepthwiseInt8CPUKernel::Init() {
 }
 
 int DeconvolutionDepthwiseInt8CPUKernel::ReSize() {
-  FreeTmpBuffer();
   InitSlideParam();
   ConvolutionBaseCPUKernel::Init();
 
@@ -224,7 +205,10 @@ int DeconvolutionDepthwiseInt8CPUKernel::Run() {
   if (need_align_) {
     PackNHWC4ToNHWCInt8(packed_output_, output_addr, conv_param_->output_batch_,
                         conv_param_->output_h_ * conv_param_->output_w_, conv_param_->output_channel_);
+    context_->allocator->Free(packed_output_);
   }
+  context_->allocator->Free(packed_input_);
+  context_->allocator->Free(output_buffer_);
   return RET_OK;
 }
 

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

@@ -41,7 +41,6 @@ class DeconvolutionDepthwiseInt8CPUKernel : public ConvolutionBaseCPUKernel {
   int Execute(int task_id);
 
  private:
-  void FreeTmpBuffer();
   SlidingWindowParam *sliding = nullptr;
   int16_t *packed_weight_ = nullptr;
   int16_t *packed_input_ = nullptr;