RadixSort/VK
RadixSort/VK is a high-performance sorting library for Vulkan 1.2.
Features include:
- Ultra-fast stable sorting of 32‑bit or 64‑bit keyvals
- Key size is declared at sort time
- Indirectly dispatchable
- Simple to integrate in a Vulkan 1.2 environment
Usage
See radix_sort_vk.h.
Device Support
The following architectures are supported:
| Vendor | Architecture | 32‑bit Keyvals | 64‑bit Keyvals | Notes |
|---|
| NVIDIA | sm_35+ | ✔ | ✔ | |
| AMD | GCN3+ | ✔ | ✔ | |
| AMD | RDNA+ | ✔ | ✔ | 64-wide subgroup |
| ARM | Bifrost4 | ✔ | ✔ | |
| ARM | Bifrost8 | ✔ | ✔ | |
| Intel | GEN8+ | ✔ | ✔ | |
| Intel | Xe+ | ✔ | ✔ | |
Benchmarks
NVIDIA GeForce RTX 2060
NVIDIA Quadro K2200
AMD Radeon RX 6600 XT
AMD Radeon RX 560
Intel HD Graphics 615
ARM Mali G52
ARM Mali G31
Future Improvements
- 96-bit and 128-bit keyval support.
- Possibly reduce GMEM transactions by accumulating stores that share the same transaction boundary.
References
- Merrill, Duane and Michael Garland. “Single-pass Parallel Prefix Scan with Decoupled Lookback.” (2016).
- Adinets, Andy. “A Faster Radix Sort Implementation.” NVIDIA GTC 2020.
