Added Faiss #8572
Added Faiss #8572
Conversation
|
Replaces #8483 |
stemann
force-pushed
the
feature/faiss
branch
from
7e7b18a to
29ca836
Compare
stemann
force-pushed
the
feature/faiss
branch
2 times, most recently
from
5b63415 to
e8ac9ed
Compare
* Added mingw32 patch: To make posix_memalign etc. available on mingw32. * Dropped CUDA archs 60, 61 for CUDA 11.8+. * Update CMake wrt. CUDA arch detection: https://github.com/Kitware/CMake/blob/v3.23.3/Modules/CMakeDetermineCUDACompiler.cmake#L654 * Fixed CUDA cmake configure: Added symlink from $prefix/cuda/lib64 to $prefix/cuda/lib * Updated to v1.9.0 * Added header file products * Added additional patches for mingw32
stemann
force-pushed
the
feature/faiss
branch
from
e8ac9ed to
70c112e
Compare
F/Faiss/Faiss_GPU/build_tarballs.jl
Outdated
| include(joinpath(YGGDRASIL_DIR, "fancy_toys.jl")) | ||
| include(joinpath(YGGDRASIL_DIR, "platforms", "cuda.jl")) | ||
|
|
||
| name = "Faiss_GPU" |
There was a problem hiding this comment.
@maleadt Does it make more sense to name packages _CUDA, rather than _GPU?
There is also some level of AMD rocm support in faiss.
There was a problem hiding this comment.
How is multiple GPU backends implemented in this package? One advantage to splitting AMD and CUDA into separate packages would be to avoid downloading both sets of dependencies when only one would be needed.
There was a problem hiding this comment.
It's implemented via build flags (cf. below), so in addition to keeping dependencies separate, it makes a lot of sense with separate CUDA and AMD GPU/ROCM builds.
- GPU-related options:
-DCUDAToolkit_ROOT=/path/to/cuda-10.1in order to hint to the path of
the CUDA toolkit (for more information, see
CMake docs),-DCMAKE_CUDA_ARCHITECTURES="75;72"for specifying which GPU architectures
to build against (see CUDA docs to
determine which architecture(s) you should pick),-DFAISS_ENABLE_ROCM=ONin order to enable building GPU indices for AMD GPUs.
-DFAISS_ENABLE_GPUmust beONwhen using this option. (possible values areONandOFF),
Source: https://github.com/facebookresearch/faiss/blob/main/INSTALL.md#step-1-invoking-cmake
There was a problem hiding this comment.
I have renamed Faiss_GPU to Faiss_CUDA
No description provided.