ZAzackmorris·May 12, 2025
I've blabbered on about this too much, but I decided to see if there's been any progress on multithreading libraries for GPU and found some for lisp:
https://repository.lboro.ac.uk/articles/conference_contribut... (has pdf download)
https://github.com/michelp/hillisp
Unfortunately neither of these seem to do any heavy lifting, they just transpile lisp statements into GPU kernels like WebMonkey does. But it's a start.
To get real work done in concurrency with SISD/MIMD to SIMD transpiling, we'd need a small team with a modest budget to work for perhaps 2 years on figuring out how to statically analyze unmodified lisp programs and break them down into independent fragments that can execute in parallel without side effects. Then use libraries like these to run those fragments as kernels and aggregate the results. Conceptually this would work
Read the rest
like a spreadsheet, watching for changes in inputs and calculating the result across as many cores as possible unsupervised.
Once that runtime reached MVP, work could begin on transpiling common languages like C and Javascript to lisp. This is also an open problem, but straightforward to solve. The difficulty appears to be that transpiling imperative code to lisp has only been done as a hack with impure or monadic logic that defeats the whole purpose of using functional programming. It also tends to lose human readability by using too much glue code through bindings, rather than naming variables common to both runtimes like how we name registers in C++ that can be used within asm() blocks:
https://gcc.gnu.org/onlinedocs/gcc/Local-Register-Variables....