GPU kernels

GPU kernels are an experimental opt-in via defkernel, with a restricted kernel language. --emit ptx targets CUDA and --emit wgsl targets WebGPU. Host programs need a device driver from a separate repository; fib.gpu.sim checks memory/transfer protocols on the CPU and does not execute kernels. Atomics, shuffles and pinned-transfer contracts are documented in sections 12–14 of the GPU design. The CPU tensor API is independent of a driver and is useful as a reference for checking GPU results:

(ns main (:require [fib.tensor :as t]))
(defun main () -> i64 (do (println (t/sum (t/tensor [1.0 2.0 3.0]))) 0))

6.0
0