From dc111d3476c2a6e6cb8147011bf946b43719719d Mon Sep 17 00:00:00 2001
From: fengjiayi <fengjiayi@baidu.com>
Date: Wed, 1 Aug 2018 10:30:40 +0800
Subject: [PATCH] update softmax_cudnn_op

---
 paddle/fluid/operators/softmax_cudnn_op.cu.cc | 23 ++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/paddle/fluid/operators/softmax_cudnn_op.cu.cc b/paddle/fluid/operators/softmax_cudnn_op.cu.cc
index 5596fa0648..2bdb23e999 100644
--- a/paddle/fluid/operators/softmax_cudnn_op.cu.cc
+++ b/paddle/fluid/operators/softmax_cudnn_op.cu.cc
@@ -30,8 +30,16 @@ class SoftmaxCUDNNKernel : public framework::OpKernel<T> {
     // allocate memory on device.
     Out->mutable_data<T>(context.GetPlace());
 
+    auto dims = X->dims();
+    auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1);
+    framework::LoDTensor flattened_x;
+    framework::LoDTensor flattened_out;
+    flattened_x.ShareDataWith(*X).Resize(flattened_dims);
+    flattened_out.ShareDataWith(*Out).Resize(flattened_dims);
+
     math::SoftmaxCUDNNFunctor<T>()(
-        context.template device_context<platform::CUDADeviceContext>(), X, Out);
+        context.template device_context<platform::CUDADeviceContext>(),
+        &flattened_x, &flattened_out);
   }
 };
 
@@ -46,9 +54,18 @@ class SoftmaxGradCUDNNKernel : public framework::OpKernel<T> {
     // allocate memory on device.
     dX->mutable_data<T>(context.GetPlace());
 
+    auto dims = Out->dims();
+    auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1);
+    framework::LoDTensor flattened_out;
+    framework::LoDTensor flattened_d_out;
+    framework::LoDTensor flattened_d_x;
+    flattened_out.ShareDataWith(*Out).Resize(flattened_dims);
+    flattened_d_out.ShareDataWith(*dOut).Resize(flattened_dims);
+    flattened_d_x.ShareDataWith(*dX).Resize(flattened_dims);
+
     math::SoftmaxGradCUDNNFunctor<T>()(
-        context.template device_context<platform::CUDADeviceContext>(), Out,
-        dOut, dX);
+        context.template device_context<platform::CUDADeviceContext>(),
+        &flattened_out, &flattened_d_out, &flattened_d_x);
   }
 };
 
-- 
GitLab