Merge pull request PaddlePaddle#33 from carlushuang/0.15.0_1b7e38

add fp16 support for conv/bn/pool/softmax

Author: sabreshao

paddle/fluid/operators/batch_norm_op.cu.cc

@@ -317,6 +317,7 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(
-    batch_norm, ops::BatchNormKernel<plat::CUDADeviceContext, float>);
+    batch_norm, ops::BatchNormKernel<plat::CUDADeviceContext, float>,
+    ops::BatchNormKernel<plat::CUDADeviceContext, plat::float16>);
 REGISTER_OP_CUDA_KERNEL(
     batch_norm_grad, ops::BatchNormGradKernel<plat::CUDADeviceContext, float>);

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -326,7 +326,8 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
 
 namespace plat = paddle::platform;
 REGISTER_OP_KERNEL(conv2d, CUDNN, plat::CUDAPlace,
-                   paddle::operators::CUDNNConvOpKernel<float>);
+                   paddle::operators::CUDNNConvOpKernel<float>,
+                   paddle::operators::CUDNNConvOpKernel<plat::float16>);
 REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace,
                    paddle::operators::CUDNNConvGradOpKernel<float>);
 

paddle/fluid/operators/elementwise_add_op.cu

@@ -22,9 +22,12 @@ namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(
     elementwise_add, ops::ElementwiseAddKernel<plat::CUDADeviceContext, float>,
     ops::ElementwiseAddKernel<plat::CUDADeviceContext, double>,
-    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int>);
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int64_t>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, paddle::platform::float16>);
 REGISTER_OP_CUDA_KERNEL(
     elementwise_add_grad,
     ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, float>,
     ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, double>,
-    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int>);
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int>,
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int64_t>);

paddle/fluid/operators/math/softmax.cu

@@ -90,13 +90,16 @@ void SoftmaxGradCUDNNFunctor<T>::operator()(
 }
 
 template class SoftmaxCUDNNFunctor<float>;
+template class SoftmaxCUDNNFunctor<platform::float16>;
 template class SoftmaxGradCUDNNFunctor<float>;
+template class SoftmaxGradCUDNNFunctor<platform::float16>;
 
 template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16>;
 template class SoftmaxFunctor<platform::CUDADeviceContext, float>;
 template class SoftmaxFunctor<platform::CUDADeviceContext, double>;
 template class SoftmaxGradFunctor<platform::CUDADeviceContext, float>;
 template class SoftmaxGradFunctor<platform::CUDADeviceContext, double>;
+template class SoftmaxGradFunctor<platform::CUDADeviceContext, platform::float16>;
 
 }  // namespace math
 }  // namespace operators

paddle/fluid/operators/pool_cudnn_op.cu.cc

@@ -188,9 +188,11 @@ namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
 REGISTER_OP_KERNEL(pool2d, CUDNN, plat::CUDAPlace,
-                   ops::PoolCUDNNOpKernel<float>);
+                   ops::PoolCUDNNOpKernel<float>,
+                   ops::PoolCUDNNOpKernel<plat::float16>);
 REGISTER_OP_KERNEL(pool2d_grad, CUDNN, plat::CUDAPlace,
-                   ops::PoolCUDNNGradOpKernel<float>);
+                   ops::PoolCUDNNGradOpKernel<float>,
+                   ops::PoolCUDNNGradOpKernel<plat::float16>);
 
 REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace,
                    ops::PoolCUDNNOpKernel<float>);

paddle/fluid/operators/softmax_cudnn_op.cu.cc

@@ -75,6 +75,8 @@ class SoftmaxGradCUDNNKernel : public framework::OpKernel<T> {
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 REGISTER_OP_KERNEL(softmax, CUDNN, plat::CUDAPlace,
-                   ops::SoftmaxCUDNNKernel<float>);
+                   ops::SoftmaxCUDNNKernel<float>,
+                   ops::SoftmaxCUDNNKernel<plat::float16>);
 REGISTER_OP_KERNEL(softmax_grad, CUDNN, plat::CUDAPlace,
-                   ops::SoftmaxGradCUDNNKernel<float>);
+                   ops::SoftmaxGradCUDNNKernel<float>,
+                   ops::SoftmaxCUDNNKernel<plat::float16>);