[0-size Tensor Job2 No.11] Add 0-size Tensor support for paddle.expand_as

paddle/phi/kernels/gpu/expand_as_grad_kernel.cu

@@ -17,9 +17,9 @@
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/reduce_function.h"
 #include "paddle/phi/kernels/reduce_sum_kernel.h"
-
 namespace phi {
 
 template <typename T, typename Context>
@@ -28,6 +28,11 @@ void ExpandAsGradKernel(const Context& context,
  const DenseTensor& out_grad,
  const std::vector<int64_t>& target_shape,
  DenseTensor* in_grad) {
+ if (out_grad.numel() == 0) {
+ phi::Full<T, Context>(
+ context, phi::IntArray(common::vectorize(in_grad->dims())), 0, in_grad);
+ return;
+ }
  auto in_dims = x.dims();
  auto out_dims = out_grad.dims();
  int in_rank = in_dims.size();

paddle/phi/kernels/gpu/expand_as_kernel.cu

@@ -29,6 +29,10 @@ void ExpandAsKernel(const Context& dev_ctx,
  const paddle::optional<DenseTensor>& y,
  const std::vector<int64_t>& target_shape_t,
  DenseTensor* out) {
+ if (x.numel() == 0 || (y.get_ptr() && y.get_ptr()->numel() == 0)) {
+ dev_ctx.template Alloc<T>(out);
+ return;
+ }
  std::vector<int64_t> target_shape = target_shape_t;
 
  if (y.get_ptr()) {
@@ -43,10 +47,10 @@ void ExpandAsKernel(const Context& dev_ctx,
  vec_in_dims.insert(vec_in_dims.begin(), diff, 1);
 
  for (unsigned int i = 0; i < vec_in_dims.size(); ++i) {
- PADDLE_ENFORCE_NE(
-  target_shape[i],
-  0,
-  errors::InvalidArgument("The value of target shape cannot be zero."));
+ if (target_shape[i] == 0) {
+ dev_ctx.template Alloc<T>(out);
+ return;
+ }
  if (i < diff) {
  PADDLE_ENFORCE_GT(
  target_shape[i],

paddle/phi/kernels/impl/expand_as_grad_kernel_impl.h

@@ -15,8 +15,8 @@
 #pragma once
 
 #include "paddle/phi/core/tensor_utils.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/impl/expand_as_kernel_impl.h"
-
 namespace phi {
 template <typename Context, typename T, int Dims>
 void ExpandAsBackward(const Context& dev_ctx,
@@ -48,6 +48,11 @@ void ExpandAsGradKernel(const Context& context,
  const DenseTensor& out_grad,
  const std::vector<int64_t>& target_shape,
  DenseTensor* in_grad) {
+ if (out_grad.numel() == 0) {
+ phi::Full<T, Context>(
+ context, phi::IntArray(common::vectorize(in_grad->dims())), 0, in_grad);
+ return;
+ }
  auto x_dims = x.dims();
  auto out_grad_dims = out_grad.dims();
  std::vector<int64_t> real_target_shape =

paddle/phi/kernels/impl/expand_as_kernel_impl.h

@@ -19,7 +19,6 @@
 
 #include "paddle/phi/kernels/funcs/eigen/common.h"
 #include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
-
 #define MAX_RANK_SUPPORTED 8
 
 namespace phi {
@@ -39,10 +38,10 @@ void ExpandAs(const Context& context,
  return;
  }
  for (size_t i = 0; i < vec_in_dims.size(); ++i) {
- PADDLE_ENFORCE_NE(
-  target_shape[i],
-  0,
-  errors::InvalidArgument("The value of target shape cannot be zero."));
+ if (target_shape[i] == 0) {
+ context.template Alloc<T>(out);
+ return;
+ }
  if (i < diff) {
  PADDLE_ENFORCE_GT(
  target_shape[i],
@@ -100,6 +99,10 @@ void ExpandAsKernel(const Context& dev_ctx,
  const paddle::optional<DenseTensor>& y,
  const std::vector<int64_t>& target_shape,
  DenseTensor* out) {
+ if (x.numel() == 0 || (y.get_ptr() && y.get_ptr()->numel() == 0)) {
+ dev_ctx.template Alloc<T>(out);
+ return;
+ }
  auto rank = x.dims().size();
  auto target_rank = target_shape.size();
  PADDLE_ENFORCE_GE(target_rank,

paddle/phi/kernels/xpu/expand_as_kernel.cc

@@ -32,10 +32,10 @@ void ExpandAs(const Context& context,
  auto diff = target_shape.size() - vec_in_dims.size();
  vec_in_dims.insert(vec_in_dims.begin(), diff, 1);
  for (size_t i = 0; i < vec_in_dims.size(); ++i) {
- PADDLE_ENFORCE_NE(target_shape[i],
-  0,
-  common::errors::InvalidArgument(
-  "The value of target shape cannot be zero."));
+ if (target_shape[i] == 0) {
+ context.template Alloc<T>(out);
+ return;
+ }
  if (vec_in_dims[i] != 1) {
  PADDLE_ENFORCE_EQ(
  vec_in_dims[i],
@@ -89,6 +89,10 @@ void ExpandAsKernel(const Context& dev_ctx,
  const paddle::optional<DenseTensor>& y,
  const std::vector<int64_t>& target_shape,
  DenseTensor* out) {
+ if (x.numel() == 0 || (y.get_ptr() && y.get_ptr()->numel() == 0)) {
+ dev_ctx.template Alloc<T>(out);
+ return;
+ }
  auto rank = x.dims().size();
  auto target_rank = target_shape.size();
  PADDLE_ENFORCE_GE(target_rank,

test/legacy_test/test_expand_as_v2_op.py

@@ -79,6 +79,32 @@ def if_enable_cinn(self):
  self.enable_cinn = False
 
 
+class TestExpandAs_ZeroSize(TestExpandAsBasic):
+ def init_inputs_and_outputs(self):
+ x = np.random.random([2, 1]).astype(self.dtype)
+ target_tensor = np.random.random([2, 0]).astype(self.dtype)
+ self.inputs = {'X': x, "Y": target_tensor}
+ self.attrs = {'target_shape': target_tensor.shape}
+ output = np.random.random([2, 0]).astype(self.dtype)
+ self.outputs = {'Out': output}
+
+ def test_check_output(self):
+ self.check_output(check_pir=True)
+
+ def test_check_grad(self):
+ self.check_grad(['X'], 'Out', check_pir=True)
+
+
+class TestExpandAs_ZeroSize2(TestExpandAs_ZeroSize):
+ def init_inputs_and_outputs(self):
+ x = np.random.random([3, 0]).astype(self.dtype)
+ target_tensor = np.random.random([3, 0]).astype(self.dtype)
+ self.inputs = {'X': x, "Y": target_tensor}
+ self.attrs = {'target_shape': target_tensor.shape}
+ output = np.random.random([3, 0]).astype(self.dtype)
+ self.outputs = {'Out': output}
+
+
 @unittest.skipIf(
  not core.is_compiled_with_cuda()
  or not core.is_bfloat16_supported(core.CUDAPlace(0)),

test/legacy_test/test_solve_op.py

@@ -330,6 +330,7 @@ def setUp(self):
  self.place.append(paddle.CUDAPlace(0))
 
  def check_static_result(self, place):
+ paddle.enable_static()
  with base.program_guard(base.Program(), base.Program()):
  paddle_input_x = paddle.static.data(
  name="input_x", shape=[3, 3], dtype=self.dtype
@@ -968,7 +969,7 @@ def test_static(self):
  self.check_static_result(
  place=place,
  x_shape=[10, 0, 0],
- y_shape=[6, 0, 0],
+ y_shape=[10, 0, 0],
  np_y_shape=[10, 0, 0],
  )
  with self.assertRaises(ValueError) as context: