Added Faiss #8483
Added Faiss #8483
Conversation
| -#ifdef _MSC_VER | ||
| +#ifdef _MSC_VER || defined(__MINGW32__) |
There was a problem hiding this comment.
This should simply check _WIN32 if the relevant thing is the target, rather than the compiler: https://learn.microsoft.com/en-us/cpp/preprocessor/predefined-macros?view=msvc-170
Building version v1.7.4 to get support for aarch64-linux-gnu - avoids ARM NEON optimization (which uses vld1q_f32_x2 etc.) included in v1.8.0: facebookresearch/faiss#3166
To make posix_memalign etc. available on mingw32.
|
How do we want this? v1.7.4, without windows, musl and freebsd, builds just fine: 3fe4eb8 Complications:
To-do (future?)
|
Open an issue? Edit: ah, I guess that may require the neon extension. |
I might just have gotten too lazy at first - too lazy to create a patch for the aarch64/neon bits...; it was easier rolling back to v1.7.4 - and skipping musl and freebsd at first - better than not having aarch64-linux support. |
As there is already CUDA-support in Faiss.jl (using the Faiss Python API via PythonCall), CUDA-support might be worth giving a go in the first iteration of a Faiss JLL (this PR): zsz00/Faiss.jl#33 (comment) Cf. https://github.com/zsz00/Faiss.jl/blob/v0.3.5/src/Faiss.jl#L43 |
* Excluded Jetson CUDA archs * Dropped CUDA archs 60, 61 for CUDA 11.8+
No description provided.