Show HN: Compile the same source for GPUs and CPUs
by spacedcowboy on 11/13/2026, 5:53:42 PM
I'm writing a language called xc and its compiler, xcc for the last 6 months or so (full disclosure: with a lot of help from Claude). I've finally got version 1 (actually v1.0.1) out, for Linux (x86-64 and arm64, including the Pi 5), macOS and Windows.<p>The idea: you write an ordinary loop, the compiler works out whether it can run in parallel, and it emits both a multi-core CPU version and GPU kernels (MSL, PTX, SPIR-V or WGSL, all generated by xcc itself without needing an SDK installed).<p>At run time the program measures each block on that machine and keeps the faster placement.<p>Examples:<p>- On an M3 Max a Mandelbrot block takes 13.7 ms on all 16 cores and 1.3 ms on the GPU<p>- On a Ryzen 9955HX's 2-CU integrated Radeon the GPU loses to the 32 CPU threads, so it stays on the CPU.<p>The compiler also reads plain old C via clang's front end and does the same thing to C loops, with no pragmas. Comparing xcc to OpenMP for both gcc and clang (with the helping hand of "#pragma omp parallel for" gives (times in ms):<p><pre><code> xcc gcc+omp clang+omp mandel_deep 371 2170 2231 mandelbrot 2.45 12.2 12.8 md5 9.39 16.1 13.4 md5_mask 8.17 12.1 13.4 nbody 0.45 1.29 9.92 perlin 5.87 4.38 7.05 saxpy 2.26 2.18 2.32 </code></pre> Geometric mean: xcc is 2.1x faster than GCC+OpenMP and 3.0x faster than clang+OpenMP.<p>All the builds are checked to make sure they give the same result - no point having a benchmark that doesn't produce the same result! xcc has a notable win for nbody, because its vectorisation is better (across bodies) but perlin has been a bit of a bugbear throughout the leading-up-to-v1 timespan.<p>Platform-specific binary tarballs, the sources, OpenMP versions and method are on the site (linked from github), and there are Phoronix Test Suite profiles if you'd like to run them yourself.<p>It's only just made v1 so there's a few rough edges: C++ through clang is in-progress (but not official yet), and the Pi 5's GPU is slow. Not sure if that's the Pi or the code yet...<p>There's more to the language but I don't want this to be an essay! The real, honest-to-$deity cool thing is the <i>automatic</i> parallelising of standard C/xc loops and being able to run that code on the GPU with <i>no boilerplate</i> - "ARC for Algorithms" (what Objective-C's ARC did for memory management), it "just works".<p>Very keen to get feedback and fix any gaps. I think it's generally useful right now, but it is "v1" - Bug reports welcome as GitHub issues...
https://github.com/ThrudTheBarbarian/xc
Comments
by:
11/13/2026, 5:53:42 PM