!4602 [MS][LITE] fullconnection matmul A B matrix const node bug

Merge pull request !4602 from ling/fc

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

@@ -23,6 +23,11 @@ using mindspore::lite::RET_OK;
 
 namespace mindspore::kernel {
 FullconnectionCPUKernel::~FullconnectionCPUKernel() {
+  FreeBuf();
+  return;
+}
+
+void FullconnectionCPUKernel::FreeBuf() {
   if (a_c8_ptr_ != nullptr) {
     free(a_c8_ptr_);
     a_c8_ptr_ = nullptr;
@@ -41,7 +46,11 @@ FullconnectionCPUKernel::~FullconnectionCPUKernel() {
   }
 }
 
-int FullconnectionCPUKernel::ReSize() { return RET_OK; }
+int FullconnectionCPUKernel::ReSize() {
+  FreeBuf();
+  Init();
+  return RET_OK;
+}
 
 int FullconnectionCPUKernel::Init() {
   if (context_->infer_shape_interrupt_ && !context_->running_) {
@@ -75,16 +84,44 @@ int FullconnectionCPUKernel::Init() {
     return RET_MEMORY_FAILED;
   }
   memset(b_r8_ptr_, 0, fc_param_->col_8_ * fc_param_->deep_ * sizeof(float));
-  RowMajor2Col8Major(reinterpret_cast<float *>(in_tensors_[1]->Data()), b_r8_ptr_, fc_param_->col_, fc_param_->deep_);
 
   c_r8x8_ptr_ = reinterpret_cast<float *>(malloc(fc_param_->row_8_ * fc_param_->col_8_ * sizeof(float)));
   if (c_r8x8_ptr_ == nullptr) {
     return RET_MEMORY_FAILED;
   }
   memset(c_r8x8_ptr_, 0, fc_param_->row_8_ * fc_param_->col_8_ * sizeof(float));
+
+  fc_param_->a_const_ = false;
+  fc_param_->b_const_ = false;
+  InitMatrixA(reinterpret_cast<float *>(in_tensors_[0]->Data()), a_c8_ptr_);
+  InitMatrixB(reinterpret_cast<float *>(in_tensors_[1]->Data()), b_r8_ptr_);
   return RET_OK;
 }
 
+void FullconnectionCPUKernel::InitMatrixA(float *src_ptr, float *dst_ptr) {
+  if (fc_param_->a_const_ == true) {
+    return;
+  }
+  if (src_ptr == nullptr) {
+    return;
+  }
+  fc_param_->a_const_ = true;
+  RowMajor2Col8Major(src_ptr, a_c8_ptr_, fc_param_->row_, fc_param_->deep_);
+  return;
+}
+
+void FullconnectionCPUKernel::InitMatrixB(float *src_ptr, float *dst_ptr) {
+  if (fc_param_->b_const_ == true) {
+    return;
+  }
+  if (src_ptr == nullptr) {
+    return;
+  }
+  fc_param_->b_const_ = true;
+  RowMajor2Col8Major(src_ptr, dst_ptr, fc_param_->col_, fc_param_->deep_);
+  return;
+}
+
 int FcFp32MatmulRun(int task_id, LiteParallelGroupEnv *penv, void *cdata) {
   auto fc = reinterpret_cast<FullconnectionCPUKernel *>(cdata);
   auto error_code = fc->DoMatmul(task_id);
@@ -115,9 +152,11 @@ int FullconnectionCPUKernel::Run() {
     return prepare_ret;
   }
   auto a_ptr = reinterpret_cast<float *>(in_tensors_.at(0)->Data());
+  auto b_ptr = reinterpret_cast<float *>(in_tensors_.at(1)->Data());
   auto output_ptr = reinterpret_cast<float *>(out_tensors_.at(0)->Data());
 
-  RowMajor2Col8Major(a_ptr, a_c8_ptr_, fc_param_->row_, fc_param_->deep_);
+  InitMatrixA(a_ptr, a_c8_ptr_);
+  InitMatrixB(b_ptr, b_r8_ptr_);
 
   LiteBackendParallelLaunch(FcFp32MatmulRun, this, thread_count_);
 

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

@@ -40,6 +40,11 @@ class FullconnectionCPUKernel : public FullconnectionBaseCPUKernel {
 
  public:
   int DoMatmul(int task_id);
+  void FreeBuf();
+
+ private:
+  void InitMatrixA(float *src_ptr, float *dst_ptr);
+  void InitMatrixB(float *src_ptr, float *dst_ptr);
 
  private:
   float *a_c8_ptr_;

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

@@ -78,6 +78,11 @@ int MatmulCPUKernel::Init() {
   }
   memset(c_r8x8_ptr_, 0, params_->row_8_ * params_->col_8_ * sizeof(float));
 
+  params_->a_const_ = false;
+  params_->b_const_ = false;
+  InitMatrixA(reinterpret_cast<float *>(in_tensors_[0]->Data()), a_c8_ptr_);
+  InitMatrixB(reinterpret_cast<float *>(in_tensors_[1]->Data()), b_r8_ptr_);
+
   if (in_tensors_.size() == 3) {
     bias_ptr_ = reinterpret_cast<float *>(malloc(params_->col_8_ * sizeof(float)));
     memset(bias_ptr_, 0, params_->col_8_ * sizeof(float));
@@ -89,6 +94,40 @@ int MatmulCPUKernel::Init() {
   return RET_OK;
 }
 
+void MatmulCPUKernel::InitMatrixA(float *src_ptr, float *dst_ptr) {
+  if (params_->a_const_ == true) {
+    return;
+  }
+  if (src_ptr == nullptr) {
+    return;
+  }
+  params_->a_const_ = true;
+
+  if (params_->a_transpose_) {
+    RowMajor2Row8Major(src_ptr, dst_ptr, params_->deep_, params_->row_);
+  } else {
+    RowMajor2Col8Major(src_ptr, a_c8_ptr_, params_->row_, params_->deep_);
+  }
+  return;
+}
+
+void MatmulCPUKernel::InitMatrixB(float *src_ptr, float *dst_ptr) {
+  if (params_->b_const_ == true) {
+    return;
+  }
+  if (src_ptr == nullptr) {
+    return;
+  }
+  params_->b_const_ = true;
+
+  if (params_->b_transpose_) {
+    RowMajor2Col8Major(src_ptr, dst_ptr, params_->col_, params_->deep_);
+  } else {
+    RowMajor2Row8Major(src_ptr, dst_ptr, params_->deep_, params_->col_);
+  }
+  return;
+}
+
 int MatmulCPUKernel::RunImpl(int task_id) {
   int cur_oc = MSMIN(thread_stride_, UP_DIV(params_->col_8_, 8) - task_id * thread_stride_);
   if (cur_oc <= 0) {
@@ -131,16 +170,10 @@ int MatmulCPUKernel::Run() {
     auto cur_a_ptr = a_ptr + i * a_stride;
     auto cur_b_ptr = b_ptr + i * b_stride;
     auto cur_c_ptr = c_ptr + i * c_stride;
-    if (params_->a_transpose_) {
-      RowMajor2Row8Major(cur_a_ptr, a_c8_ptr_, params_->deep_, params_->row_);
-    } else {
-      RowMajor2Col8Major(cur_a_ptr, a_c8_ptr_, params_->row_, params_->deep_);
-    }
-    if (params_->b_transpose_) {
-      RowMajor2Col8Major(cur_b_ptr, b_r8_ptr_, params_->col_, params_->deep_);
-    } else {
-      RowMajor2Row8Major(cur_b_ptr, b_r8_ptr_, params_->deep_, params_->col_);
-    }
+
+    InitMatrixA(cur_a_ptr, a_c8_ptr_);
+    InitMatrixB(cur_b_ptr, b_r8_ptr_);
+
     LiteBackendParallelLaunch(MatmulFloatRun, this, thread_count_);
     Row8x8Major2RowMajor(c_r8x8_ptr_, cur_c_ptr, params_->row_, params_->col_, params_->col_);
   }

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

@@ -35,6 +35,10 @@ class MatmulCPUKernel : public MatmulBaseCPUKernel {
   int Run() override;
   int RunImpl(int task_id);
 
+ private:
+  void InitMatrixA(float *src_ptr, float *dst_ptr);
+  void InitMatrixB(float *src_ptr, float *dst_ptr);
+
  private:
   float *a_c8_ptr_;
   float *b_r8_ptr_;

mindspore/lite/src/runtime/kernel/arm/nnacl/matmul_parameter.h

@@ -33,6 +33,8 @@ typedef struct MatMulParameter {
   int batch;
   bool a_transpose_; /* false :  row-major  */
   bool b_transpose_; /* true  :  col-major  */
+  bool a_const_;
+  bool b_const_;
   ActType act_type_;
 } MatMulParameter;