matmul support float16/double

paddle/fluid/operators/math/blas_impl.cu.h

@@ -96,10 +96,22 @@ struct CUBlas<platform::float16> {
                                        reinterpret_cast<__half *>(C), ldc));
   }
 
-  template <typename... ARGS>
-  static void GEMM_BATCH(ARGS... args) {
+  static void GEMM_BATCH(cublasHandle_t handle, cublasOperation_t transa,
+                         cublasOperation_t transb, int m, int n, int k,
+                         const float16 *alpha, const float16 *A, int lda,
+                         long long int strideA, const float16 *B,  // NOLINT
+                         int ldb, long long int strideB,           // NOLINT
+                         const float16 *beta, float16 *C, int ldc,
+                         long long int strideC,  // NOLINT
+                         int batchCount) {
 #if CUDA_VERSION >= 8000
-    PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched(args...));
+    PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched(
+        handle, transa, transb, m, n, k,
+        reinterpret_cast<const __half *>(alpha),
+        reinterpret_cast<const __half *>(A), lda, strideA,
+        reinterpret_cast<const __half *>(B), ldb, strideB,
+        reinterpret_cast<const __half *>(beta), reinterpret_cast<__half *>(C),
+        ldc, strideC, batchCount));
 #else
     PADDLE_THROW("HgemmStridedBatched is not supported on cuda <= 7.5");
 #endif

paddle/fluid/operators/math/math_function.cu

@@ -35,7 +35,8 @@ template struct SetConstant<platform::CUDADeviceContext, bool>;
 
 #define DEFINE_GPU_TRANS(RANK)                                          \
   template struct Transpose<platform::CUDADeviceContext, float, RANK>;  \
-  template struct Transpose<platform::CUDADeviceContext, double, RANK>;
+  template struct Transpose<platform::CUDADeviceContext, double, RANK>; \
+  template struct Transpose<platform::CUDADeviceContext, float16, RANK>;
 
 DEFINE_GPU_TRANS(1);
 DEFINE_GPU_TRANS(2);

paddle/fluid/operators/matmul_op.cc

@@ -421,15 +421,27 @@ REGISTER_OPERATOR(matmul, ops::MatMulOp, ops::MatMulOpMaker,
                   ops::MatMulOpGradMaker);
 REGISTER_OPERATOR(matmul_grad, ops::MatMulOpGrad);
 REGISTER_OP_CPU_KERNEL(
-    matmul, ops::MatMulKernel<paddle::platform::CPUDeviceContext, float>);
+    matmul, ops::MatMulKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::MatMulKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::MatMulKernel<paddle::platform::CPUDeviceContext,
+                      paddle::platform::float16>);
 REGISTER_OP_CPU_KERNEL(
     matmul_grad,
-    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext,
+                          paddle::platform::float16>);
 
 #ifdef PADDLE_WITH_CUDA
 REGISTER_OP_CUDA_KERNEL(
-    matmul, ops::MatMulKernel<paddle::platform::CUDADeviceContext, float>);
+    matmul, ops::MatMulKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MatMulKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::MatMulKernel<paddle::platform::CUDADeviceContext,
+                      paddle::platform::float16>);
 REGISTER_OP_CUDA_KERNEL(
     matmul_grad,
-    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext,
+                          paddle::platform::float16>);
 #endif