ggml-alloc : use virtual memory for measurement

[slaren]

Instead of using a fixed memory address, allocates an uncommitted virtual memory region that is guaranteed to not overlap with any other allocations. Should solve issues with most 32-bit platforms and OS X. Wasm may still be an issue, as I don't think it supports mmap.

[slaren]

I have tested this successfully under Linux and Windows, including 32-bit builds, but I would need some help testing with OS X.

[CoruNethron]

I have tested this successfully under Linux and Windows, including 32-bit builds, but I would need some help testing with OS X.

I'll do. What kind of test could I perform?

[slaren]

All you would need to do is make sure it builds without errors and generate a few tokens using main or any other example.

[CoruNethron]

@slaren, I've tested alloc-vmem (c031b6c) - all binaries built without errors or warnings. Generated few tokens without any problem on Mac M1.

[slaren]

Awesome, thanks!

[ggerganov]

Did some testing on M2 Ultra and it works.

I wonder, should we fallback to the old strategy if mmap is not available?
I'm planning soon to update whisper.cpp to use ggml-alloc and update the WASM examples.

[slaren]

I have added a fallback for systems without virtual memory, however I noticed that emscripten has a shim for mmap that just allocates memory, which is not great. It should still work, but the memory usage will be high, so we may have to add a check to disable mmap when building with emscripten.

[staviq]

As a sidenote

Considering mmap, this seems related: WebAssembly/WASI#304

There's also this project which I have used successfully before, and it's simple enough so extending it is not very complicated, and it uses clang to produce wasm: https://github.com/schellingb/wajic

Edit: It should be possible to get rid of emscripten in favour of clang, if that helps in any way.