About the service
I develop high-performance numerical kernels in Futhark — a purely functional data-parallel language that compiles to CUDA and OpenCL. If you have a hot loop, simulation or large volume of array operations that is hitting the CPU ceiling, I will rewrite it to run on the GPU and you will not have to write a single line of low-level device code by hand. Futhark provides strong guarantees: no side effects, no aliasing surprises, and the compiler itself fuses map, reduce, scan and segmented operations into tight kernels. The typical workflow is this. First I profile your current pipeline and find sections that are truly parallel. Then I express them in Futhark through purely functional combinators, preserving readability and determinism. I compare the result against your baseline version, tune block sizes and memory access patterns and iterate until the numbers are convincing. Finally I embed everything back into your stack: Futhark generates a C library and I wrap it so you can call it from Python via ctypes or a small extension, or link it directly into a host program in C or C++. Over the years I have accelerated Monte Carlo option pricing, image and signal filters, physical stencils, linear algebra procedures and preprocessing for machine learning. You get clean source, a build script, a test suite that checks GPU output against a baseline, benchmark results and a short note on how to rebuild and extend the code. Describe the task and target hardware — and I will estimate real acceleration before we start.
