prgpu

GPU-accelerated rendering utilities for Adobe Premiere Pro and After Effects plugins, powered by Slang shaders that compile once and dispatch uniformly across Metal (macOS), CUDA (Windows), and a Rust/rayon-based CPU fallback.

What you get

• Single shader, three targets. Author compute shaders in Slang; prgpu's build::compile_shaders("./shaders") pipeline emits Metal .metallib, CUDA .cu / PTX, and C++ for the CPU path in one go. Constant buffers and struct layouts are verified byte-for-byte across all three.
• Zero-cost dispatch abstraction. prgpu::declare_kernel!(name, Params) generates typed entry points so kernel code in the host looks the same whether the dispatch lands on MTLComputeCommandEncoder, cuLaunchKernel, or a rayon tile loop.
• CPU fallback that isn't a toy. The cpu::render module batches work into row-strip tiles, sizes them to the host's rayon pool, and re-uses allocations across frames via a keyed buffer cache. Real production effects (radialblur, vignette, etc.) hit realtime-preview budgets at 1080p on the CPU path.
• Mip-chain pyramids built in. pipeline::mip::prepare_mip_source + generate_mips turn a flat input buffer into a multi-lod source in a couple of lines of host code. Effects whose sample radius grows with a parameter (radial blur, bokeh, glow, motion blur) can sample from downsampled levels for multi-x speedups with no user-visible quality loss. See docs/mip_chain.md for the design.
• Benchmarking out of the box. With the bench feature, a kernel author gets prgpu::bench::{Scene, Resolution, PixelFormat, KernelBenchmark} — a criterion harness that runs the real CPU dispatch against synthetic frames, so you measure the kernel and nothing else.

Crate layout

prgpu
├── src
│   ├── types/            Configuration, TextureDesc, FrameParams, Pixel
│   ├── cpu/              render_cpu / render_cpu_direct, tile scheduler,
│   │                     bounded rayon pool, keyed buffer cache
│   ├── gpu/              backend dispatch (Metal / CUDA), hot-reload hooks,
│   │                     device-agnostic buffer cache
│   ├── kernels/          declare_kernel!, built-in mip_downsample, helpers
│   │                     for common patterns (blur downsample, sweep samples)
│   ├── params/           FromParam, SetupParams, CpuParams traits + kernel
│   │                     params DSL macros
│   ├── ui/               BlendMode popup, shared UI parameter helpers
│   ├── build/            slangc driver, reflection-driven Rust binding gen,
│   │                     cpp → static lib compile (feature = "build")
│   └── bench.rs          criterion harness (feature = "bench")
├── shaders/              built-in Slang shaders (mip_downsample, etc)
├── docs/                 design notes: mip_chain, timing, xframe
└── prgpu-macro/          #[gpu_struct] proc-macro crate (transitive dep)

Usage

# Cargo.toml of a Premiere / AE plugin crate
[dependencies]
prgpu = "0.1"

[build-dependencies]
prgpu = { version = "0.1", features = ["build"] }

[dev-dependencies]
prgpu = { version = "0.1", features = ["timing", "bench"] }

// build.rs
fn main() {
    prgpu::build::compile_shaders("./shaders").unwrap();
    // ...
}

// src/kernel.rs
prgpu::kernel_params! {
    MyParams for crate::params::Params {
        strength: f32 = [float(Strength) / 100.0];
    }
}
prgpu::declare_kernel!(my_effect, MyParams);

Features

Slang SDK

prgpu's build feature auto-downloads the Slang SDK into target/.slang-sdk/<version>/ on first build. No manual setup. The version is pinned in src/build/sdk.rs.

Shader include path (vekl or your own)

When you call compile_shaders("./shaders"), prgpu looks for a sibling directory named vekl (relative to prgpu's own checkout) and adds it to the slangc include path if found. If your effect crate isn't laid out next to a vekl/ checkout, call the explicit form instead:

use std::path::PathBuf;
prgpu::build::compile_shaders_with(
    "./shaders",
    &[PathBuf::from("path/to/your/shader-lib")],
).unwrap();

prgpu is happy to run without any extra include path — your shaders just have to resolve their own import / #include statements against the directories you pass.

Documentation

• docs/mip_chain.md — host-side pyramid blur recipe
• docs/timing.md — instrumentation model
• docs/xframe.md — cross-frame state contract

License

Dual-licensed under MIT OR Apache-2.0.

Status

APIs are stabilising; the 0.1.x series may still see breaking changes in minor bumps. Pinning an exact =0.1.N version is reasonable if you're shipping a release binary.