[XPU] Fix precision for paddle.Tensor.__sub__ complex64/complex128

[YqGe585]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

The XPU kernel for paddle.Tensor.__sub__ failed with PaddleError when operating on complex64/complex128 types (e.g. complex64 - float64 or float64 - complex64). The root cause was two missing capabilities:

1. XPU cast kernel lacked complex128 support: The CastKernel on XPU only handled DataType::COMPLEX64 (under PADDLE_WITH_XPU_FFT) but not DataType::COMPLEX128. When type promotion yielded complex128 from complex64 + float64, the cast kernel threw "Not supported cast float64 -> complex128".

2. XPU subtract kernel had no complex type support: The SubtractKernel and SubtractGradKernel registrations on XPU only included float, float16, bfloat16, int, int64_t — no complex64 or complex128. The xdnn library also lacks broadcast_sub<double> and broadcast_sub_grad<double>, so complex128 (whose real/imag parts are double) requires a float-cast workaround.

Fix for cast kernel

Added DataType::COMPLEX128 case in the switch statement (under PADDLE_WITH_XPU_FFT), following the same Real/Imag decomposition pattern used for COMPLEX64. Added CastKernel<phi::complex128, XPUContext> specialization and registered phi::complex128 in the kernel registration.

Fix for subtract kernel

Added SubtractKernel<phi::complex64> and SubtractKernel<phi::complex128> specializations under PADDLE_WITH_XPU_FFT, using Real/Imag decomposition. For complex128, since xdnn lacks broadcast_sub<double>, the fix casts real/imag double parts to float, performs subtraction in float, then casts back to double and recombines via ComplexKernel<double>. The same float-cast workaround applies to SubtractGradKernel<phi::complex128>. Registered both phi::complex64 and phi::complex128 in forward and grad kernel registrations.

Modified files

• paddle/phi/kernels/xpu/cast_kernel.cc: Added COMPLEX128 case, complex128 specialization, and registration
• paddle/phi/kernels/xpu/elementwise_subtract_kernel.cc: Added complex64/complex128 forward kernel specializations and registration
• paddle/phi/kernels/xpu/elementwise_subtract_grad_kernel.cc: Added complex64/complex128 grad kernel specializations and registration

Does this PR introduce a precision change?

Yes — XPU precision corrected to align with GPU for complex subtraction operations. Previously these cases threw kernel errors; now they produce correct results matching GPU output.

[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.