feat: add detection template for cuda and avx. #268
Conversation
Just merge cuda and avx detection and change layout.
| #endif | ||
| ?? TryLoad("avx2/libllama.so", System.Runtime.Intrinsics.X86.Avx2.IsSupported) | ||
| ?? TryLoad("avx/libllama.so", System.Runtime.Intrinsics.X86.Avx.IsSupported) |
There was a problem hiding this comment.
AVX/AVX2/AVX512 detection should already be handled by this
| // Try to load a preferred library, based on CPU feature detection | ||
| TryLoadLibrary(); | ||
| #if NET6_0_OR_GREATER | ||
| NativeLibrary.SetDllImportResolver(typeof(NativeApi).Assembly, LLamaImportResolver); |
There was a problem hiding this comment.
I deliberately avoided using NativeLibrary.SetDllImportResolver when I originally set this up, since it overwrites global configuration that may already be in use by the application.
|
The intention when I originally introduced this was that new dependencies could easily be added (in order of preference, most preferred first) something like this: return TryLoad("magic/libllama.dll", IsMagicSupported())
?? TryLoad("unicorns/libllama.dll", IsUnicornsSupported())
?? TryLoad("rainbows/libllama.dll", IsRainbowsSupported())
?? etc...i.e. any new platform can simply be chained in, so long as you can write a boolean method that prechecks if that particular thing is supported. I deliberately avoided using |
Just for review, please don't merge it.