Update the cuda API and enable tensor core for GEMM (#9622)

* change from hgemm to gemmEx

* fix cpplint

paddle/fluid/operators/math/math_function.cu

@@ -39,18 +39,33 @@ void gemm<platform::CUDADeviceContext, float16>(
   cublasOperation_t cuTransB =
       (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
 
-  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);
+  float h_alpha = static_cast<float>(alpha);
+  float h_beta = static_cast<float>(beta);
 
   // TODO(kexinzhao): add processing code for compute capability < 53 case
   PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53,
-                    "cublas Hgemm requires GPU compute capability >= 53");
-  PADDLE_ENFORCE(platform::dynload::cublasHgemm(
-      context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb,
-      h_A, lda, &h_beta, h_C, N));
+                    "cublas fp16 gemm requires GPU compute capability >= 53");
+
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT;
+#if CUDA_VERSION >= 9000
+  if (context.GetComputeCapability() >= 70) {
+    PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(),
+                                                        CUBLAS_TENSOR_OP_MATH));
+    algo = CUBLAS_GEMM_DFALT_TENSOR_OP;
+  } else {
+    PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(),
+                                                        CUBLAS_DEFAULT_MATH));
+  }
+#endif
+
+  // cublasHgemm does true FP16 computation which is slow for non-Volta
+  // GPUs. So use cublasGemmEx instead which does pesudo FP16 computation:
+  // input/output in fp16, computation in fp32, which can also be accelerated
+  // using tensor cores in volta GPUs.
+  PADDLE_ENFORCE(platform::dynload::cublasGemmEx(
+      context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, B,
+      CUDA_R_16F, ldb, A, CUDA_R_16F, lda, &h_beta, C, CUDA_R_16F, N,
+      CUDA_R_32F, algo));
 }
 
 template <>

paddle/fluid/platform/dynload/cublas.cc

@@ -24,6 +24,10 @@ void *cublas_dso_handle = nullptr;
 
 CUBLAS_BLAS_ROUTINE_EACH(DEFINE_WRAP);
 
+#ifdef CUBLAS_BLAS_ROUTINE_EACH_R2
+CUBLAS_BLAS_ROUTINE_EACH_R2(DEFINE_WRAP);
+#endif
+
 }  // namespace dynload
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/dynload/cublas.h

@@ -15,8 +15,9 @@ limitations under the License. */
 #pragma once
 
 #include <cublas_v2.h>
+#include <cuda.h>
 #include <dlfcn.h>
-#include <mutex>
+#include <mutex>  // NOLINT
 #include "paddle/fluid/platform/dynload/dynamic_loader.h"
 
 namespace paddle {
@@ -70,6 +71,7 @@ extern void *cublas_dso_handle;
   __macro(cublasDgemm_v2);                \
   __macro(cublasHgemm);                   \
   __macro(cublasSgemmEx);                 \
+  __macro(cublasGemmEx);                  \
   __macro(cublasSgeam_v2);                \
   __macro(cublasDgeam_v2);                \
   __macro(cublasCreate_v2);               \
@@ -89,9 +91,15 @@ extern void *cublas_dso_handle;
   __macro(cublasSgetrfBatched);           \
   __macro(cublasSgetriBatched);           \
   __macro(cublasDgetrfBatched);           \
-  __macro(cublasDgetriBatched)
+  __macro(cublasDgetriBatched);
 
-CUBLAS_BLAS_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP);
+CUBLAS_BLAS_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
+
+// APIs available after CUDA 9.0
+#if CUDA_VERSION >= 9000
+#define CUBLAS_BLAS_ROUTINE_EACH_R2(__macro) __macro(cublasSetMathMode);
+CUBLAS_BLAS_ROUTINE_EACH_R2(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
+#endif
 
 #undef DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
 }  // namespace dynload