Merge pull request #962 from chonwhite/develop

fix:#961

src/operators/kernel/fpga/softmax_kernel.cpp

@@ -25,9 +25,9 @@ namespace operators {
 template <>
 bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   const Tensor *input = param->InputX();
-  if (input->type() == typeid(half)) {
-    auto input_ptr = input->data<half>();
+  auto input_ptr = input->data<float>();
   auto output_ptr = param->Out();
+  Tensor *floatInput = new Tensor(*input);
   fpga::BypassArgs args;
   args.input_layout_type = fpga::LAYOUT_HWC;
   args.output_layout_type = fpga::LAYOUT_CHW;
@@ -37,9 +37,10 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   args.image.height = (uint32_t)input->dims()[0];
   args.image.width = (uint32_t)input->dims()[1];
   args.image.channels = 1;
-    args.output.address = output_ptr;
+  args.output.address = (void *)floatInput->mutable_data<float>();
+
+  param->SetFloatInput(floatInput);
   param->SetFpgaArgs(args);
-  }
   return true;
 }
 
@@ -48,8 +49,12 @@ void SoftmaxKernel<FPGA, float>::Compute(
     const SoftmaxParam<FPGA> &param) const {
   DLOG << "======================================= FPGA SoftMAX "
           "===============================================";
-  const Tensor *in_x = param.InputX();
+  const Tensor *in_x = param.FloatInput();
   Tensor *out = param.Out();
+  fpga::fpga_flush((void *)in_x->data<float>(), in_x->memory_size());
+  fpga::PerformBypass(param.FpgaArgs());
+  fpga::fpga_invalidate(out->data<float>(), out->memory_size());
+
   auto x_dims = in_x->dims();
   out->Resize(x_dims);
   math::SoftmaxFuntor<CPU, float>()(in_x, out);

src/operators/op_param.h

@@ -785,7 +785,7 @@ class SoftmaxParam : public OpParam {
   fpga::BypassArgs fpga_bypass_args;
 
  public:
-  RType *FloatInput() {
+  RType *FloatInput() const {
     return float_input_x_ == nullptr ? input_x_ : float_input_x_.get();
   }
   void SetFloatInput(Tensor *input) { float_input_x_.reset(input); }