fix test_elementwise_sub_op

[DrRyanHuang]

PR types

Others

PR changes

Others

Description

• 【PIR】 PIR算子推全和单测修复 #59382

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[kangguangli]

这个算子单测失败的原因应该不是 在PIR下缺少 elementwise_sub算子的定义，在op_compat.yaml 里面已经将elementwise_sub 映射到 subtract。这个问题可能较复杂一点，有兴趣的话可以先分析下单测失败原因。

[DrRyanHuang]

这是单测 test_elementwise_sub_op 报错的log:

W1221 13:44:05.297209 643880 pir_interpreter.cc:1665]  OP id:2 pd_op.mean_all raises an EnforceNotMet exception phi::enforce::EnforceNotMet, In user code:


    InvalidArgumentError: The type of data we are trying to retrieve (float64) does not match the type of data (complex128) currently contained in the container.
      [Hint: Expected dtype() == phi::CppTypeToDataType<T>::Type(), but received dtype():13 != phi::CppTypeToDataType<T>::Type():11.] (at /home/aistudio/Paddle/paddle/phi/core/dense_tensor.cc:171)
      [operator < pd_kernel.phi_kernel > error]

ERROR: test_check_grad_ingore_x (__main__.TestRealComplexElementwiseSubOp)
ERROR: test_check_grad_ingore_y (__main__.TestRealComplexElementwiseSubOp)
ERROR: test_check_grad_normal (__main__.TestRealComplexElementwiseSubOp)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test/legacy_test/test_elementwise_sub_op.py", line 856, in test_check_grad_ingore_x
    self.check_grad(
  File "/home/aistudio/Paddle/test/legacy_test/op_test.py", line 2969, in check_grad
    self.check_grad_with_place(
  File "/home/aistudio/Paddle/test/legacy_test/op_test.py", line 3207, in check_grad_with_place
    analytic_grads = self._get_gradient(
  File "/home/aistudio/Paddle/test/legacy_test/op_test.py", line 3671, in _get_gradient
    self._check_ir_grad_output(
  File "/home/aistudio/Paddle/test/legacy_test/op_test.py", line 3513, in _check_ir_grad_output
    executor.run(
  File "/home/aistudio/Paddle/build/python/paddle/base/executor.py", line 1772, in run
    res = self._run_impl(
  File "/home/aistudio/Paddle/build/python/paddle/base/executor.py", line 1978, in _run_impl
    ret = new_exe.run(
  File "/home/aistudio/Paddle/build/python/paddle/base/executor.py", line 827, in run
    tensors = self._new_exe.run(
ValueError: In user code:

    InvalidArgumentError: The type of data we are trying to retrieve (float64) does not match the type of data (complex128) currently contained in the container.
      [Hint: Expected dtype() == phi::CppTypeToDataType<T>::Type(), but received dtype():13 != phi::CppTypeToDataType<T>::Type():11.] (at /home/aistudio/Paddle/paddle/phi/core/dense_tensor.cc:171)
      [operator < pd_kernel.phi_kernel > error]

----------------------------------------------------------------------
Ran 8 tests in 3.073s

FAILED (errors=3)

貌似是 mean_all 和 subtract infer_meta 函数问题? 但我看他俩的 yaml 文件

- op : mean_all
  args : (Tensor x)
  output : Tensor
  infer_meta :
    func : MeanAllInferMeta
  kernel :
    func : mean_all
  backward : mean_all_grad

void MeanAllInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->set_dims(common::make_ddim({}));
  out->set_dtype(x.dtype());
  out->set_layout(x.layout());
}

都有 out->set_dtype(x.dtype()); 这句呀，输出类型指定了，也不像是 kernel 注册的问题

PD_REGISTER_KERNEL(mean_all,
                   GPU,
                   ALL_LAYOUT,
                   phi::MeanAllKernel,
                   float,
                   double,
                   phi::dtype::float16,
                   phi::dtype::complex<float>,
                   phi::dtype::complex<double>) {}

这里复数类型也都注册过，之前说算子的类型检查下沉到C++侧了, 是这个问题吗? 可以给点提示嘛 😆
@kangguangli

[kangguangli]

这里的问题是pd_op_to_kernel_pass里面在类型提升后没有处理输出的类型，解决方案是在类型提升后再做一次 infer_meta，重新推导输出类型，目前内部同事 @chen2016013 正在开发中，可以关注下。