diff --git a/src/operators/kernel/fpga/softmax_kernel.cpp b/src/operators/kernel/fpga/softmax_kernel.cpp
index 7c784ce474bbb2588dcf78ecded740777445fc80..fd84cb8e46c974d23816c0dd4c99a545d996c409 100644
--- a/src/operators/kernel/fpga/softmax_kernel.cpp
+++ b/src/operators/kernel/fpga/softmax_kernel.cpp
@@ -25,21 +25,22 @@ 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 output_ptr = param->Out();
-    fpga::BypassArgs args;
-    args.input_layout_type = fpga::LAYOUT_HWC;
-    args.output_layout_type = fpga::LAYOUT_CHW;
-    args.input_data_type = fpga::DATA_TYPE_FP16;
-    args.output_data_type = fpga::DATA_TYPE_FP32;
-    args.image.address = (void *)(input_ptr);
-    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;
-    param->SetFpgaArgs(args);
-  }
+  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;
+  args.input_data_type = fpga::DATA_TYPE_FP16;
+  args.output_data_type = fpga::DATA_TYPE_FP32;
+  args.image.address = (void *)(input_ptr);
+  args.image.height = (uint32_t)input->dims()[0];
+  args.image.width = (uint32_t)input->dims()[1];
+  args.image.channels = 1;
+  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);
diff --git a/src/operators/math/math_function.cpp b/src/operators/math/math_function.cpp
index 6ef9fb2a8252e82014ebebc22f82066eeb324c0d..14269817ededd097c4c9ade20be5ee773c02d692 100644
--- a/src/operators/math/math_function.cpp
+++ b/src/operators/math/math_function.cpp
@@ -36,13 +36,35 @@ void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
   int N = dim_out[1];
   int K = (!trans_a) ? dim_a[1] : dim_a[0];
 
+  if (trans_a) {
+    int numel = matrix_a.numel();
+    int m = matrix_a.dims()[0];
+    int n = matrix_a.dims()[1];
+    float *tmp = (float *)(matrix_a.data<float>());
+    float *a = static_cast<float *>(
+        paddle_mobile::memory::Alloc(sizeof(float) * numel));
+    int index = 0;
+    for (int j = 0; j < n; j++) {
+      for (int i = 0; i < m; i++) {
+        a[index++] = tmp[i * n + j];
+      }
+    }
+#ifdef _OPENMP
+    Sgemm_omp(M, N, K, alpha, a, K, matrix_b.data<float>(), N, beta,
+              matrix_out->data<float>(), N, relu, bias);
+#else
+    Sgemm(M, N, K, alpha, a, K, matrix_b.data<float>(), N, beta,
+          matrix_out->data<float>(), N, relu, bias);
+#endif
+  } else {
 #ifdef _OPENMP
-  Sgemm_omp(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(),
-            N, beta, matrix_out->data<float>(), N, relu, bias);
+    Sgemm_omp(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(),
+              N, beta, matrix_out->data<float>(), N, relu, bias);
 #else
-  Sgemm(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(), N,
-        beta, matrix_out->data<float>(), N, relu, bias);
+    Sgemm(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(), N,
+          beta, matrix_out->data<float>(), N, relu, bias);
 #endif
+  }
 }
 
 template <>
diff --git a/src/operators/op_param.h b/src/operators/op_param.h
index 1728e6a6cc778ec223c3f14c971404ba3a5cc0f7..1c5815c64236f1b67fb6ab7752d0c4caef7c2646 100644
--- a/src/operators/op_param.h
+++ b/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); }