GPU Programming

Sounio aims to make GPU compute a first-class part of the language, not an afterthought.

At the surface level, this shows up as:

• kernel fn ... for GPU kernels
• with GPU for host functions that perform GPU operations (launch, device memory, sync, …)

What Works Today (Compiler Reality)

The compiler accepts kernel fn syntax and type-checks kernels as functions.

For example, this is valid and should type-check:

kernel fn noop(a: &[i32]) {
    // A minimal kernel (no GPU intrinsics used).
}

fn main() with IO {
    println("GPU docs: kernel syntax parses and type-checks.")
}

Spec vs Implementation

The full GPU programming model in Sounio includes:

• kernel launch APIs
• device memory allocation and transfers
• GPU intrinsics like thread/block IDs and synchronization

Some of these are feature-dependent and may require external toolchains (CUDA, Metal, SPIR-V) to actually execute kernels on real hardware.

When in doubt:

• read the stdlib GPU modules under stdlib/gpu/
• consult spec/LANGUAGE_SPECIFICATION.md (GPU section)
• prefer compiler fixtures and tests as the source of “what compiles today”