From 14cf420ec2801531f506216e2e8353b05a97499f Mon Sep 17 00:00:00 2001
From: Zhang Ting <zhangting_2017@163.com>
Date: Tue, 8 Dec 2020 08:20:44 +0800
Subject: [PATCH] revert cast eigen kernel (#29445)

---
 paddle/fluid/operators/cast_op.h              | 49 +++----------------
 .../fluid/tests/unittests/test_cast_op.py     | 12 -----
 2 files changed, 7 insertions(+), 54 deletions(-)

diff --git a/paddle/fluid/operators/cast_op.h b/paddle/fluid/operators/cast_op.h
index 91276ba6e8b..8fa0416049f 100644
--- a/paddle/fluid/operators/cast_op.h
+++ b/paddle/fluid/operators/cast_op.h
@@ -48,52 +48,17 @@ struct CastOpFunctor {
   }
 };
 
-template <typename DeviceContext, typename InT, typename OutT>
-static void CastFunction(const framework::ExecutionContext& context) {
-  auto* in = context.Input<framework::Tensor>("X");
-  auto* out = context.Output<framework::Tensor>("Out");
-
-  auto in_t = framework::EigenVector<InT>::Flatten(*in);
-  out->mutable_data<OutT>(context.GetPlace());
-  auto out_t = framework::EigenVector<OutT>::Flatten(*out);
-  auto& place =
-      *context.template device_context<DeviceContext>().eigen_device();
-  out_t.device(place) = in_t.template cast<OutT>();
-}
-
 template <typename DeviceContext, typename InT>
 class CastOpKernel : public framework::OpKernel<InT> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto out_type = static_cast<framework::proto::VarType::Type>(
-        context.Attr<int>("out_dtype"));
-
-    if (out_type == paddle::framework::proto::VarType::FP64) {
-      CastFunction<DeviceContext, InT, double>(context);
-    } else if (out_type == paddle::framework::proto::VarType::FP32) {
-      CastFunction<DeviceContext, InT, float>(context);
-    } else if (out_type == paddle::framework::proto::VarType::FP16) {
-      CastFunction<DeviceContext, InT, paddle::platform::float16>(context);
-    } else if (out_type == paddle::framework::proto::VarType::INT64) {
-      CastFunction<DeviceContext, InT, int64_t>(context);
-    } else if (out_type == paddle::framework::proto::VarType::INT32) {
-      CastFunction<DeviceContext, InT, int>(context);
-    } else if (out_type == paddle::framework::proto::VarType::UINT8) {
-      CastFunction<DeviceContext, InT, uint8_t>(context);
-    } else if (out_type == paddle::framework::proto::VarType::BOOL) {
-      CastFunction<DeviceContext, InT, bool>(context);
-    } else if (out_type == paddle::framework::proto::VarType::COMPLEX64) {
-      CastFunction<DeviceContext, InT, paddle::platform::complex64>(context);
-    } else if (out_type == paddle::framework::proto::VarType::COMPLEX128) {
-      CastFunction<DeviceContext, InT, paddle::platform::complex128>(context);
-    } else {
-      // NOTE(chenweihang): if else branch do nothing, the output var will
-      // be non-initialized in dygraph, which will throw error if the
-      // non-initialized var is used as the next op's input
-      PADDLE_THROW(platform::errors::Unimplemented(
-          "Now does not support casting Tensor to `%s` data type.",
-          framework::DataTypeToString(out_type)));
-    }
+    auto* in = context.Input<framework::Tensor>("X");
+    auto* out = context.Output<framework::Tensor>("Out");
+    framework::VisitDataType(
+        static_cast<framework::proto::VarType::Type>(
+            context.Attr<int>("out_dtype")),
+        CastOpFunctor<DeviceContext, InT>(
+            in, out, context.template device_context<DeviceContext>()));
   }
 };
 
diff --git a/python/paddle/fluid/tests/unittests/test_cast_op.py b/python/paddle/fluid/tests/unittests/test_cast_op.py
index 44fdd8c74bf..0fc3dccab4a 100644
--- a/python/paddle/fluid/tests/unittests/test_cast_op.py
+++ b/python/paddle/fluid/tests/unittests/test_cast_op.py
@@ -90,18 +90,6 @@ class TestCastOpError(unittest.TestCase):
             self.assertRaises(TypeError, test_dtype_type)
 
 
-class TestCastOpErrorInDygraph(unittest.TestCase):
-    def test_non_support_out_dtype(self):
-        paddle.disable_static()
-
-        with self.assertRaises(NotImplementedError):
-            tensor = paddle.randn([10, 10], 'float32')
-            core.ops.cast(tensor, 'in_dtype', core.VarDesc.VarType.FP32,
-                          'out_dtype', core.VarDesc.VarType.INT16)
-
-        paddle.enable_static()
-
-
 if __name__ == '__main__':
     paddle.enable_static()
     unittest.main()
-- 
GitLab