Merge pull request #790 from zhangyang0701/develop

Correct FPGA kernels. close  #789

src/operators/kernel/fpga/conv_add_bn_kernel.cpp

@@ -37,11 +37,11 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam *param) {
   auto bn_scale_ptr = param->InputScale()->data<float>();
   auto bn_bias_ptr = param->InputBias()->data<float>();
   const float epsilon = param->Epsilon();
-  PADDLE_MOBILE_ENFORCE(input->dims()[1] == bias->dims()[0] &&
+  PADDLE_MOBILE_ENFORCE(out->dims()[1] == bias->dims()[0] &&
                             bias->dims()[0] == param->InputBias()->dims()[0],
-                        "Image channel should be equal to bias number");
+                        "Output channel should be equal to bias number");
 
-  const int channel = input->dims()[1];
+  const int channel = out->dims()[1];
   float *bs_ptr =
       reinterpret_cast<float *>(fpga::fpga_malloc(2 * channel * sizeof(float)));
   Tensor *new_scale = new Tensor();

src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp

@@ -35,11 +35,11 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(FusionConvAddBNReluParam *param) {
   auto bn_scale_ptr = param->InputScale()->data<float>();
   auto bn_bias_ptr = param->InputBias()->data<float>();
   const float epsilon = param->Epsilon();
-  PADDLE_MOBILE_ENFORCE(input->dims()[1] == bias->dims()[0] &&
+  PADDLE_MOBILE_ENFORCE(out->dims()[1] == bias->dims()[0] &&
                             bias->dims()[0] == param->InputBias()->dims()[0],
-                        "Image channel should be equal to bias number");
+                        "Output channel should be equal to bias number");
 
-  const int channel = input->dims()[1];
+  const int channel = out->dims()[1];
   float *bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
   Tensor *new_scale = new Tensor();
   Tensor *new_bias = new Tensor();

src/operators/kernel/fpga/conv_add_relu_kernel.cpp

@@ -31,9 +31,9 @@ bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam *param) {
   Tensor *out = param->Output();
   auto out_ptr = out->mutable_data<half>();
 
-  PADDLE_MOBILE_ENFORCE(input->dims()[1] == bias->dims()[0],
-                        "Image channel should be equal to bias number");
-  int channel = input->dims()[1];
+  PADDLE_MOBILE_ENFORCE(out->dims()[1] == bias->dims()[0],
+                        "Output channel should be equal to bias number");
+  int channel = out->dims()[1];
   float *bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
   for (int i = 0; i < channel; i++) {
     bs_ptr[i * 2] = 1;

src/operators/kernel/fpga/conv_bn_kernel.cpp

@@ -35,10 +35,10 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam *param) {
   auto bn_scale_ptr = param->InputScale()->data<float>();
   auto bn_bias_ptr = param->InputBias()->data<float>();
   const float epsilon = param->Epsilon();
-  PADDLE_MOBILE_ENFORCE(input->dims()[1] == param->InputBias()->dims()[0],
-                        "Image channel should be equal to bias number");
+  PADDLE_MOBILE_ENFORCE(out->dims()[1] == param->InputBias()->dims()[0],
+                        "Output channel should be equal to bias number");
 
-  const int channel = input->dims()[1];
+  const int channel = out->dims()[1];
   float *bs_ptr =
       reinterpret_cast<float *>(fpga::fpga_malloc(2 * channel * sizeof(float)));
   Tensor *new_scale = new Tensor();

src/operators/kernel/fpga/conv_bn_relu_kernel.cpp

@@ -33,10 +33,10 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam *param) {
   auto bn_scale_ptr = param->InputScale()->data<float>();
   auto bn_bias_ptr = param->InputBias()->data<float>();
   const float epsilon = param->Epsilon();
-  PADDLE_MOBILE_ENFORCE(input->dims()[1] == param->InputBias()->dims()[0],
-                        "Image channel should be equal to bias number");
+  PADDLE_MOBILE_ENFORCE(out->dims()[1] == param->InputBias()->dims()[0],
+                        "Output channel should be equal to bias number");
 
-  const int channel = input->dims()[1];
+  const int channel = out->dims()[1];
   float *bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
   Tensor *new_scale = new Tensor();
   Tensor *new_bias = new Tensor();

src/operators/kernel/fpga/fc_relu_kernel.cpp

@@ -33,7 +33,7 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam *param) {
 
   PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == input_y->dims()[0],
                         "Image channel should be equal to weight number");
-  int channel = input_x->dims()[1];
+  int channel = out->dims()[1];
   float *bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
   for (int i = 0; i < channel; i++) {
     bs_ptr[i * 2] = 1;

src/operators/kernel/fpga/fusion_fc_kernel.cpp

@@ -32,7 +32,7 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam *param) {
 
   PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == input_y->dims()[0],
                         "Image channel should be equal to weight number");
-  int channel = input_x->dims()[1];
+  int channel = out->dims()[1];
   float *bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
   for (int i = 0; i < channel; i++) {
     bs_ptr[i * 2] = 1;

test/net/test_resnet.cpp

@@ -17,7 +17,13 @@ limitations under the License. */
 #include "../test_include.h"
 
 int main() {
+#ifdef PADDLE_MOBILE_FPGA
+  paddle_mobile::PaddleMobile<paddle_mobile::FPGA> paddle_mobile;
+#endif
+
+#ifdef PADDLE_MOBILE_CPU
   paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
+#endif
   paddle_mobile.SetThreadNum(4);
   auto time1 = time();
   if (paddle_mobile.Load(g_resnet, true)) {