GPU Program Support on RISC-V GPU

on YouTube ↗18m

" summary

A usable stack that runs CUDA on CPUs and RISC-V combines an LLVM-based kernel transform pipeline with an open-source CUDA runtime and target-specific backends. The system, called CUP framework, collapses CUDA thread grids into loops, turns thread-local storage into arrays, maps CUDA memory spaces to host memory, and implements floating-point atomics either via LR/SC lock conversions or by adding ISA support. A RISC-V GPU prototype Vortex uses minimal ISA extensions (spawn, thread‑mask, split/join, bar, texture); 30–40 kernels run end-to-end and manual intrinsics give measurable vectorization speedups, but proprietary libs like cuDNN block full ML coverage.

AI-written summary · about

" chapters (8)

1. 0:00Introduction and motivation
2. 2:08Vortex project overview
3. 7:47Vectorization challenges
4. 9:15SIMT architecture proposal
5. 11:23CUDA support on Vortex
6. 13:54ML workload limitations
7. 15:18Future work and updates
8. 16:00Q&A

" related talks

• 20mRISC V Software State of the UnionRandy Allen · RISC-V Summit 2019
• 19mRISC V ISA & Foundation OverviewRick O'Conner · 8th RISC-V Workshop Barcelona 2018
• 36mAccelerate RISC-V Instruction Set Simulation by Tiered JIT CompilationYen-Fu Chen, Ching-Chun (Jim) Huang, Meng-Hung Chen, Chia-Heng Tu · VMIL 2024
• 19mRISC V For Heterogeneous ComputingJustin Cormack · RISC-V Summit 2019
• 24mRISC V LLVM: Towards A Production-Ready LLVM-Based ToolchainAlex Bradbury · 7th RISC-V Workshop 2017
• 14mRISC-V for Gaming: Emulating X86 on RISC-VParis Oplopoios, felix86 · RISC-V Summit North America 2025
• 36mPorting ML to RISC‑V: How Not to Lose PerformanceKseniya Zaytseva · Saint HighLoad++ 2025
• 28mGHC’s RISC-V Native Code Generation BackendSven Tennie · 2025 Haskell Implementors' Workshop