From eed05e1adb7f704b3e1bfececfedb8476af9d33a Mon Sep 17 00:00:00 2001
From: wangchaochaohu <wangchao66@baidu.com>
Date: Tue, 8 Sep 2020 00:08:19 +0800
Subject: [PATCH] fix gpu kernel for numel Op (#27085) (#27130)

---
 paddle/fluid/operators/size_op.cc    |  2 +-
 paddle/fluid/operators/size_op.cu    |  2 +-
 paddle/fluid/operators/size_op.h     | 14 ++++++++++++--
 python/paddle/tensor/manipulation.py |  2 +-
 4 files changed, 15 insertions(+), 5 deletions(-)

diff --git a/paddle/fluid/operators/size_op.cc b/paddle/fluid/operators/size_op.cc
index b45fa7c791f..70733d64367 100644
--- a/paddle/fluid/operators/size_op.cc
+++ b/paddle/fluid/operators/size_op.cc
@@ -53,7 +53,7 @@ REGISTER_OPERATOR(
     size, ops::SizeOp, ops::SizeOpMaker,
     paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>,
     paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>);
-REGISTER_OP_CPU_KERNEL(size, ops::SizeKernel<int>, ops::SizeKernel<int32_t>,
+REGISTER_OP_CPU_KERNEL(size, ops::SizeKernel<int>, ops::SizeKernel<int64_t>,
                        ops::SizeKernel<paddle::platform::float16>,
                        ops::SizeKernel<float>, ops::SizeKernel<double>,
                        ops::SizeKernel<bool>);
diff --git a/paddle/fluid/operators/size_op.cu b/paddle/fluid/operators/size_op.cu
index 3ea30326932..de56ecd9527 100644
--- a/paddle/fluid/operators/size_op.cu
+++ b/paddle/fluid/operators/size_op.cu
@@ -16,7 +16,7 @@ limitations under the License. */
 
 REGISTER_OP_CUDA_KERNEL(
     size, paddle::operators::SizeKernel<int>,
-    paddle::operators::SizeKernel<int32_t>,
+    paddle::operators::SizeKernel<int64_t>,
     paddle::operators::SizeKernel<paddle::platform::float16>,
     paddle::operators::SizeKernel<float>, paddle::operators::SizeKernel<bool>,
     paddle::operators::SizeKernel<double>);
diff --git a/paddle/fluid/operators/size_op.h b/paddle/fluid/operators/size_op.h
index fb440708971..e8c53d6e683 100644
--- a/paddle/fluid/operators/size_op.h
+++ b/paddle/fluid/operators/size_op.h
@@ -26,8 +26,18 @@ class SizeKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* in_t = ctx.Input<Tensor>("Input");
     auto* out_t = ctx.Output<Tensor>("Out");
-    auto out_data = out_t->mutable_data<int64_t>(platform::CPUPlace());
-    out_data[0] = in_t->numel();
+    auto place = ctx.GetPlace();
+    auto out_data = out_t->mutable_data<int64_t>(place);
+    auto cpu_place = platform::CPUPlace();
+    if (place == cpu_place) {
+      out_data[0] = in_t->numel();
+    } else {
+      Tensor cpu_tensor;
+      auto cpu_data =
+          cpu_tensor.mutable_data<int64_t>(out_t->dims(), cpu_place);
+      cpu_data[0] = in_t->numel();
+      TensorCopy(cpu_tensor, place, out_t);
+    }
   }
 };
 }  // namespace operators
diff --git a/python/paddle/tensor/manipulation.py b/python/paddle/tensor/manipulation.py
index 1a7da590c0e..b42a423f399 100644
--- a/python/paddle/tensor/manipulation.py
+++ b/python/paddle/tensor/manipulation.py
@@ -1001,7 +1001,7 @@ def chunk(x, chunks, axis=0, name=None):
             x_np = np.random.random([3, 9, 5]).astype("int32")
             x = paddle.to_tensor(x_np)
 
-            out0, out1, out22 = paddle.chunk(x, chunks=3, axis=1)
+            out0, out1, out2 = paddle.chunk(x, chunks=3, axis=1)
             # out0.shape [3, 3, 5]
             # out1.shape [3, 3, 5]
             # out2.shape [3, 3, 5]
-- 
GitLab