Mirrored from https://github.com/SNAPKITTYAGENT9NOVA/rust-opencl-gpu at commit 45d641d. Part of the SnapKitty October 2026 main drop.

rust-opencl-gpu

Rust + OpenCL (via the ocl crate) experiments in composable GPU kernels, plus a SystemVerilog bit-string accelerator.

Layout

Path	Contents
src/functor/	Library (gpu crate): composable element-wise GPU operations over a u8 Boolean buffer (0 = false, 1 = true)
tests/negate_gpu.rs	Device tests: every operation runs as an OpenCL kernel and is checked against a host oracle
src/main.rs	Binary: runs a small OpenCL kernel through a Kernel trait, then the Negate and agent-workflow demos, on the first available device
src/demo.rs	Standalone host-only demo of the same kernel design; needs no GPU and is not part of the Cargo build (rustc src/demo.rs)
benchmark.cu	CUDA benchmark of L1 cache pollution from streaming loads (separate from the Rust crate)
bit_accelerator/	SystemVerilog bit-string accelerator: RTL, testbench, Why3 proofs, ISA, docs (see bit_accelerator/README.md)
.github/workflows/ci.yml	CI: Rust fmt/clippy/tests on PoCL, minimum Rust version, accelerator lint/simulation/proofs
LICENSE	SnapKitty Bit Hardware Source License
OpenCL-SDK-v2025.07.23-Win-x64/	Vendored Windows OpenCL SDK
notes.txt	Dependency note

The functor library

• Kernels: Negate (output[i] = input[i] ^ 1) and Add (output[i] = a[i] + b[i], wrapping u8). Negate rejects any input byte other than 0 or 1; Add of two true values gives 2, so Negate(Add(x, y)) is NOR only where x and y are never both true.
• Expr: an expression tree such as Negate(Add(x, y)). Every node runs as an OpenCL kernel; Negate also reads its input back to the host to check it is canonical before launching.
• Agents: ValidationAgent, ExecutionAgent, VerificationAgent and Workflow split input validation, execution and output checking.
• Canonical form: bools_to_gpu, gpu_to_bools, is_canonical and validate_buffer convert and check the 0/1 representation.
• Kernel registry: KERNEL_REGISTRY and validate_registry declare each kernel's name, parameters, minimum work items and safety notes.
• Errors: GpuError covers OpenCL failures, invalid counts, short buffers, missing Expr inputs, empty buffers and zero local sizes.

Requirements

• Rust 1.85 or newer (edition 2024).
• An OpenCL runtime with at least one device. On Ubuntu 24.04, sudo apt-get install pocl-opencl-icd ocl-icd-opencl-dev provides a CPU device.
• On Windows, the vendored SDK provides the headers and OpenCL.lib.

Build, run and test

cargo build --release
cargo run --release                     # src/main.rs
cargo fmt --check
cargo clippy --all-targets -- -D warnings
RUST_TEST_THREADS=1 cargo test          # 8 host-only unit tests + 36 device tests

Device tests fail, rather than skip, when no OpenCL device is available.

Run the tests serially on PoCL. PoCL 5.0 can abort with a pocl_release_dlhandle_cache assertion when several OpenCL contexts are released from different threads at once. With one test thread the suite passed 40 of 40 runs, against 26 of 40 with parallel threads. Real GPU drivers do not use PoCL's code path.

benchmark.cu needs the CUDA toolkit and an NVIDIA GPU; it is not built by Cargo or CI.

Bit-string accelerator

The hardware design is documented in bit_accelerator/README.md. cd bit_accelerator && make test runs lint, the testbench under Verilator and Icarus, and the Why3 proofs (needs verilator iverilog why3 z3, then why3 config detect). A copy of this tree also lives in the bit-string-accelerator repository.

---

License

Licensed under SnapKitty Bit Hardware Source License 1.0. Full text: LICENSE.