Merge pull request #1430 from hjchen2/ocr_attention

Fix generic fusion conv bugs

src/operators/kernel/central-arm-func/conv_add_add_prelu_arm_func.h

@@ -32,11 +32,11 @@ void ConvAddAddPReluCompute(const FusionConvAddAddPReluParam<CPU> &param) {
   Tensor filter = *param.Filter();
   Tensor bias = *param.Bias();
   Tensor bias1 = *param.Bias1();
-  int axis = param.Axis();
   Tensor *output = param.Output();
-
+  output->mutable_data<float>();
   float *biase_data = bias.data<float>();
 
+  int axis = param.Axis();
   int groups = param.Groups();
   std::vector<int> strides = param.Strides();
   std::vector<int> paddings = param.Paddings();

src/operators/kernel/central-arm-func/conv_add_arm_func.h

@@ -30,10 +30,11 @@ void ConvAddBasic(const FusionConvAddParam<CPU> &param) {
   const Tensor *input = param.Input();
   Tensor filter = *param.Filter();
   Tensor bias = *param.Bias();
-  int axis = param.Axis();
   Tensor *output = param.Output();
+  output->mutable_data<float>();
   float *biase_data = bias.data<float>();
 
+  int axis = param.Axis();
   int groups = param.Groups();
   std::vector<int> strides = param.Strides();
   std::vector<int> paddings = param.Paddings();

src/operators/kernel/central-arm-func/conv_add_bn_relu_arm_func.h

@@ -32,6 +32,8 @@ void ConvAddBNReluBasic(const FusionConvAddBNReluParam<CPU> &param) {
   Tensor new_bias = *param.NewBias();
   Tensor new_scale = *param.NewScale();
   Tensor *output = param.Output();
+  output->mutable_data<float>();
+
   int groups = param.Groups();
   std::vector<int> strides = param.Strides();
   std::vector<int> paddings = param.Paddings();
@@ -115,8 +117,6 @@ void ConvAddBNReluBasic(const FusionConvAddBNReluParam<CPU> &param) {
 
 template <typename P>
 void ConvAddBNReluCompute(const FusionConvAddBNReluParam<CPU> &param) {
-  Tensor Bias;
-  Bias.mutable_data<float>({param.Groups()});
   if (param.Groups() == param.Input()->dims()[1] &&
       param.Input()->dims()[1] == param.Output()->dims()[1] &&
       param.Filter()->dims()[2] == param.Filter()->dims()[3] &&

src/operators/kernel/central-arm-func/conv_add_prelu_arm_func.h

@@ -31,10 +31,11 @@ void ConvAddPReluCompute(const FusionConvAddPReluParam<CPU> &param) {
   const Tensor *input = param.Input();
   Tensor filter = *param.Filter();
   Tensor bias = *param.Bias();
-  int axis = param.Axis();
   Tensor *output = param.Output();
+  output->mutable_data<float>();
   float *biase_data = bias.data<float>();
 
+  int axis = param.Axis();
   int groups = param.Groups();
   std::vector<int> strides = param.Strides();
   std::vector<int> paddings = param.Paddings();

src/operators/kernel/central-arm-func/conv_bn_add_relu_arm_func.h

@@ -30,11 +30,11 @@ void ConvBNAddReluBasic(const FusionConvBNAddReluParam<CPU> &param) {
   Tensor filter = *param.Filter();
   Tensor new_bias = *param.NewBias();
   Tensor new_scale = *param.NewScale();
-  Tensor *output = param.Output();
   Tensor *bias1 = param.Bias();
+  Tensor *output = param.Output();
+  output->mutable_data<float>();
+
   int groups = param.Groups();
-  DLOG << "yangfei2";
-  DLOG << bias1->dims();
   std::vector<int> strides = param.Strides();
   std::vector<int> paddings = param.Paddings();
   std::vector<int> dilations = param.Dilations();

src/operators/kernel/central-arm-func/conv_transpose_arm_func.h

@@ -31,6 +31,7 @@ void ConvTransposeCompute(const ConvTransposeParam<CPU> &param) {
   const Tensor *input = param.Input();
   Tensor filter = *param.Filter();
   Tensor *output = param.Output();
+  output->mutable_data<P>();
 
   auto strides = param.Strides();
   auto paddings = param.Paddings();
@@ -76,8 +77,6 @@ void ConvTransposeCompute(const ConvTransposeParam<CPU> &param) {
   framework::DDim filter_matrix_shape = {input->dims()[1], col_matrix_shape[0]};
   filter.Resize(filter_matrix_shape);
 
-  output->mutable_data<P>();
-
   int in_step = static_cast<int>(input->dims()[1]) / groups;
   int out_step = static_cast<int>(output->dims()[1]) / groups;