initial commit

paddle/fluid/operators/batch_norm_op.cu.cc

@@ -270,9 +270,9 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
 }  // namespace paddle
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(
-    batch_norm,
+    batch_norm, ops::BatchNormKernel<plat::CUDADeviceContext, float>,
-    ops::BatchNormKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::BatchNormKernel<plat::CUDADeviceContext, plat::float16>);
 REGISTER_OP_CUDA_KERNEL(
-    batch_norm_grad,
+    batch_norm_grad, ops::BatchNormGradKernel<plat::CUDADeviceContext, float>);
-    ops::BatchNormGradKernel<paddle::platform::CUDADeviceContext, float>);

paddle/fluid/operators/math/math_function.cc

@@ -278,6 +278,7 @@ void axpy<platform::CPUDeviceContext, double>(
   cblas_daxpy(n, alpha, x, 1, y, 1);
 }
 
+template struct SetConstant<platform::CPUDeviceContext, platform::float16>;
 template struct SetConstant<platform::CPUDeviceContext, float>;
 template struct SetConstant<platform::CPUDeviceContext, double>;
 template struct SetConstant<platform::CPUDeviceContext, int>;

paddle/fluid/operators/math/math_function.cu

@@ -348,6 +348,7 @@ void axpy<platform::CUDADeviceContext, double>(
                                                 &alpha, x, 1, y, 1));
 }
 
+template struct SetConstant<platform::CUDADeviceContext, platform::float16>;
 template struct SetConstant<platform::CUDADeviceContext, float>;
 template struct SetConstant<platform::CUDADeviceContext, double>;
 template struct SetConstant<platform::CUDADeviceContext, int>;