Chris Elrod

I do fear that adding indexing methods would encourage/steer people towards less flexible and less convenient APIs for whatever they're using it for. Anyone using this should write their structs...

That PR of VectorizationBase has a definite latency regression: ```julia chriselrod@chriselrod-XPS-13-9310 ~> jm --startup=no -e "@time @time using VectorizedRNG" 0.702441 seconds (1.86 M allocations: 121.201 MiB, 1.73% gc time, 7.57%...

Hmm, that makes sense I guess.

PRs welcome.

I think GCs would have to be implemented at the compiler level. Julia's GC isn't aware of/doesn't handle `alloca`, but IIRC Julia's compiler is what handles converting memory allocations into...

That looks like the error was from the `Random` standard library: ```julia ERROR: InitError: could not load library "libdSFMT" libdSFMT.so: cannot open shared object file: No such file or directory...

> Currently getting a segfault and the trace here is very overwhelming - I've looked at each of these addresses but they don't lead me to places where I've changed...

With AVX512, you may want to use [vscalefps](https://www.felixcloutier.com/x86/vscalefps). It computes `zmm0 = zmm1 * 2^{zmm2}`, where all are floats. It overflows and underflows properly, letting you remove checks + blends....

@jart Sure. If it helps, I just wrote a C++ implementation you can look at here: https://github.com/chriselrod/ExpAVX512 The source is in `include/ExpAVX512.hpp`. The README contains benchmarks. I didn't benchmark against...

> The `INFINITY` constant alone is used 83 times in the llama.cpp codebase, so compiling with `-ffinite-math-only` might not be a bright idea. If you want us to stop using...