[0-size Tensor No.248] Add 0-size Tensor support for reshape

paddle/phi/kernels/reshape_grad_kernel.cc

@@ -28,6 +28,10 @@ void ReshapeGradKernel(const Context& dev_ctx,
                        const DenseTensor& x,
                        const DenseTensor& out_grad,
                        DenseTensor* x_grad) {
+  if (x_grad->numel() == 0) {
+    dev_ctx.Alloc(x_grad, x_grad->dtype());
+    return;
+  }
   // NOTE: [Why not to use x.dims() ?]
   // Because inplace strategy is different between old IR and PIR,
   // we need fix it into x.dims() after cleaning old IR system.

paddle/phi/kernels/reshape_kernel.cc

@@ -35,6 +35,8 @@ void ReshapeKernel(const Context& dev_ctx,
     return;
   }
   dev_ctx.Alloc(out, x.dtype());
+  if (out->numel() == 0) return;
+
   // TODO(chenweihang): the output dims are overwrite after copying,
   // here we need to use copy method that only copy data
   auto dims = out->dims();

test/legacy_test/test_reshape_op.py

@@ -92,6 +92,35 @@ def init_data(self):
         self.inferred_shape = ()
 
 
+class TestReshapeOp_ZeroSize(OpTest):
+    def init_data(self):
+        self.ori_shape = (0, 2)
+        self.new_shape = (2, 0)
+        self.inferred_shape = (2, 0)
+
+    def setUp(self):
+        self.init_data()
+        self.op_type = "reshape2"
+        self.python_api = paddle.tensor.reshape
+        self.public_python_api = paddle.tensor.reshape
+        self.python_out_sig = ['Out']
+        self.inputs = {"X": np.random.random(self.ori_shape).astype("float32")}
+        self.attrs = {"shape": self.new_shape}
+        self.outputs = {
+            "Out": self.inputs["X"].reshape(self.inferred_shape),
+            'XShape': np.random.random(self.ori_shape).astype("float32"),
+        }
+
+    def test_check_output(self):
+        self.check_output(no_check_set=['XShape'])
+
+    def test_check_grad(self):
+        self.check_grad(
+            ["X"],
+            "Out",
+        )
+
+
 @unittest.skipIf(
     not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
     "BFP16 test runs only on CUDA",