search

google / XNNPACK

High-efficiency floating-point neural network inference operators for mobile, server, and Web
Other
1.89k stars 376 forks source link

[PyTorch] XNNPACK `static_assert` not supported on MSVC 2019 #7534

Open xuhancn opened 3 days ago

xuhancn commented 3 days ago

My pervious PR to add build option to turn off libm, link: https://github.com/google/XNNPACK/pull/7522

After that, I tried to upgrade XNNPACK of PyTorch, but I found static_assert was not supported on MSVC 2019, error CI link: https://hud.pytorch.org/pr/141754.

I read the code current XNNPACK code use some static_assert, such as https://github.com/google/XNNPACK/blob/668a518ebcfd5ae0024ebd9d540868e26ff93171/src/xnnpack/buffer.h#L55

static_assert is not supported well by MSVC 2019, reference: https://learn.microsoft.com/en-us/cpp/build/reference/zc-static-assert?view=msvc-160

So, static_assert will bring in some compatibility issue on MSVC 2019. It also limited XNNPACK upgrade by PyTorch. If we can downgrade static_assert to assert, and let XNNPACK has better compatibility.

CC: @dsharletg

gonnet commented 2 days ago

Hi Xu,

I think the problem is that the two-parameter static_assert(bool, const char *) is supported, but the single-parameter `static_assert(bool) is not. The former is part of the C++11 standard, whereas the latter was only added to C++17 (link).

Does it work if you add a second parameter "", e.g. as is done in src/xnnpack/buffer.h?

Cheers, Pedro

gonnet commented 2 days ago

Fixed this in #7535.

xuhancn commented 2 days ago

Hi @gonnet new error reported from PyTorch CI: https://hud.pytorch.org/pr/141754 I will setup a local MSVC 2019 environment to do more debug work next week. Thanks.