diff --git a/paddle/fluid/operators/math/blas_impl.cu.h b/paddle/fluid/operators/math/blas_impl.cu.h
index 86e49469912256d305e6a5d19f1736aef51a754b..89935829ab35a52dd85bcaf906b53e41d576cf3f 100644
--- a/paddle/fluid/operators/math/blas_impl.cu.h
+++ b/paddle/fluid/operators/math/blas_impl.cu.h
@@ -126,14 +126,9 @@ inline void Blas<platform::CUDADeviceContext>::GEMM(
       CUDA_R_32F, algo));
 #else
   // CUDA 7.5 does not support cublasGemmEx, hence we fall back to use hgemm
-  const half h_alpha = static_cast<const half>(alpha);
-  const half h_beta = static_cast<const half>(beta);
-  const half *h_A = reinterpret_cast<const half *>(A);
-  const half *h_B = reinterpret_cast<const half *>(B);
-  half *h_C = reinterpret_cast<half *>(C);
-
-  CUBlas<platform::float16>(context_.cublas_handle(), cuTransB, cuTransA, N, M,
-                            K, &h_alpha, h_B, ldb, h_A, lda, &h_beta, h_C, N);
+  CUBlas<platform::float16>::GEMM(context_.cublas_handle(), cuTransB, cuTransA,
+                                  N, M, K, &h_alpha, h_B, ldb, h_A, lda,
+                                  &h_beta, h_C, N);
 #endif  // CUDA_VERSION >= 8000
 }