Question: --skip-double flag disables float64 kernel support — intentional hardware limitation?

[PlumBlossomMaid]

Background

In build_inc.sh the CUDA compile flags include -mllvm --skip-double, which prevents all double-precision (float64) kernels from being compiled into the plugin.

Impact

As a result:

• contiguous(float64) is not registered (missing double in PD_CUSTOM_KERNEL_REGISTER)
• Float64 operations on Iluvatar GPU produce NaN or crash with "kernel not registered"
• This blocks many PaddlePaddle workflows that rely on float64 (e.g. STFT, audio processing)

Question

1. Is the --skip-double flag due to a hardware limitation (no float64 ALUs on BI-V150S)?
2. Or is it a compiler/driver issue that could be resolved in a future release?
3. If float64 is supported in hardware but slow, would enabling it with a build flag be acceptable?

We'd like to understand the roadmap so we can plan float64 support in our downstream projects like PaddleOcean (a high-level PaddlePaddle framework).

[PlumBlossomMaid]

我知道为什么了 我做了个实验发现 这玩意硬件不支持double计算 写的double Kernel在计算的时候会给你砍成float，但是存储的时候还是按照double存储的…… 确实没有必要编译了……