From e8efba5787158c69daf38b946124569d93d46f05 Mon Sep 17 00:00:00 2001
From: Guoxia Wang <mingzilaochongtu@gmail.com>
Date: Thu, 30 Sep 2021 14:10:46 +0800
Subject: [PATCH] fix bug of reduce_sum when src_dtype != dst_dtype and
 reduce_num == 1 (#36123) (#36193)

---
 paddle/fluid/operators/reduce_ops/reduce_op.cu.h | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/paddle/fluid/operators/reduce_ops/reduce_op.cu.h b/paddle/fluid/operators/reduce_ops/reduce_op.cu.h
index 4760270caa..28b6ebc243 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_op.cu.h
+++ b/paddle/fluid/operators/reduce_ops/reduce_op.cu.h
@@ -34,6 +34,7 @@ namespace cub = hipcub;
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/operators/amp/fp16_type_traits.h"
+#include "paddle/fluid/operators/cast_op.h"
 #include "paddle/fluid/operators/kernel_primitives/kernel_primitives.h"
 #include "paddle/fluid/platform/cuda_device_function.h"
 #include "paddle/fluid/platform/fast_divmod.h"
@@ -705,8 +706,16 @@ void TensorReduceFunctorImpl(const framework::Tensor& x, framework::Tensor* y,
 
   if (config.reduce_num == 1) {
     auto out_dims = y->dims();
-    framework::TensorCopy(x, y->place(), y);
-    y->Resize(out_dims);
+    if (x.type() == y->type()) {
+      framework::TensorCopy(x, y->place(), y);
+      y->Resize(out_dims);
+    } else {
+      auto* dev_ctx = static_cast<platform::CUDADeviceContext*>(
+          paddle::platform::DeviceContextPool::Instance().Get(x.place()));
+      framework::VisitDataType(
+          static_cast<framework::proto::VarType::Type>(y->type()),
+          CastOpFunctor<platform::CUDADeviceContext, Tx>(&x, y, *dev_ctx));
+    }
     return;
   }
 
-- 
GitLab