[0-size Tensor No.41、125] Add 0-size Tensor support for cumsum (PaddlePaddle#72799)

Author: co63oc

paddle/phi/kernels/cpu/cum_kernel.cc

@@ -57,6 +57,10 @@ void ScanKernel(const Context& dev_ctx,
                 bool reverse,
                 Reducer reducer,
                 DenseTensor* out) {
+  if (out && out->numel() == 0) {
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   dev_ctx.template Alloc<T>(out);
 
   if (x.numel() == 1) {

paddle/phi/kernels/gpu/cum_kernel.cu

@@ -274,6 +274,10 @@ void ScanKernel(const Context& dev_ctx,
                 bool reverse,
                 Op op,
                 DenseTensor* out) {
+  if (out && out->numel() == 0) {
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   T* out_data = dev_ctx.template Alloc<T>(out);
 
   // For 0D Tensor

paddle/phi/kernels/impl/logcumsumexp_grad_impl.h

@@ -50,6 +50,10 @@ void LogcumsumexpGradKernel(const Context& dev_ctx,
                             bool exclusive,
                             bool reverse,
                             DenseTensor* d_x) {
+  if (d_x && d_x->numel() == 0) {
+    dev_ctx.template Alloc<T>(d_x);
+    return;
+  }
   reverse = !reverse;
   dev_ctx.template Alloc<T>(d_x);
 

paddle/phi/kernels/xpu/cum_kernel.cc

@@ -28,6 +28,10 @@ void CumsumKernel(const Context& dev_ctx,
                   bool reverse,
                   DenseTensor* out) {
   using XPUType = typename XPUTypeTrait<T>::Type;
+  if (out && out->numel() == 0) {
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   dev_ctx.template Alloc<T>(out);
 
   if (x.numel() == 1) {

test/legacy_test/test_cumsum_op.py

@@ -747,5 +747,39 @@ def test_fp16(self):
             paddle.disable_static()
 
 
+def create_test_class(op_type, dtype, shape, axis):
+    class Cls(unittest.TestCase):
+        def test_zero_size(self):
+            paddle.disable_static()
+            numpy_tensor_1 = np.random.rand(*shape).astype(dtype)
+            paddle_x = paddle.to_tensor(numpy_tensor_1)
+            paddle_x.stop_gradient = False
+
+            paddle_api = eval(f"paddle.{op_type}")
+            paddle_out = paddle_api(paddle_x, axis=axis)
+            numpy_api = eval(f"np.{op_type}")
+            numpy_out = numpy_api(numpy_tensor_1, axis=axis)
+
+            np.testing.assert_allclose(
+                paddle_out.numpy(),
+                numpy_out,
+                1e-2,
+                1e-2,
+            )
+            np.testing.assert_allclose(
+                paddle_out.shape,
+                numpy_out.shape,
+            )
+
+    cls_name = f"{op_type}{dtype}_0SizeTest"
+    Cls.__name__ = cls_name
+    globals()[cls_name] = Cls
+
+
+create_test_class("cumsum", "float32", [3, 4, 0], 0)
+create_test_class("cumsum", "float64", [3, 4, 0, 3, 4], -2)
+create_test_class("cumsum", "int32", [3, 4, 0], 0)
+create_test_class("cumsum", "int64", [3, 4, 0, 3, 4], -1)
+
 if __name__ == '__main__':
     unittest.main()

test/legacy_test/test_logcumsumexp_op.py

@@ -375,5 +375,40 @@ def test_check_grad(self):
         )
 
 
+def create_test_class(op_type, dtype, shape, axis):
+    class Cls(unittest.TestCase):
+        def test_zero_size(self):
+            paddle.disable_static()
+            numpy_tensor_1 = np.random.rand(*shape).astype(dtype)
+            paddle_x = paddle.to_tensor(numpy_tensor_1)
+            paddle_x.stop_gradient = False
+
+            paddle_api = eval(f"paddle.{op_type}")
+            paddle_out = paddle_api(paddle_x, axis=axis)
+            numpy_out = np.log(
+                np.cumsum(np.exp(numpy_tensor_1), axis=axis)
+            )  # Numpy does not have logcumsumexp
+
+            np.testing.assert_allclose(
+                paddle_out.numpy(),
+                numpy_out,
+                1e-2,
+                1e-2,
+            )
+            np.testing.assert_allclose(
+                paddle_out.shape,
+                numpy_out.shape,
+            )
+
+    cls_name = f"{op_type}{dtype}_0SizeTest"
+    Cls.__name__ = cls_name
+    globals()[cls_name] = Cls
+
+
+create_test_class("logcumsumexp", "float32", [3, 4, 0], 0)
+create_test_class("logcumsumexp", "float64", [3, 4, 0, 3, 4], -2)
+create_test_class("logcumsumexp", "int32", [3, 4, 0], 0)
+create_test_class("logcumsumexp", "int64", [3, 4, 0, 3, 4], -1)
+
 if __name__ == '__main__':
     unittest.main()