Upstreaming SVE in QEMU
summary
Upstream work adds a new TCG vector API and host backends, demonstrates SVE support by converting some AArch64 LV64 ops, and lands softfloat refactoring; the infrastructure totals 7,000 lines for the API and the SVE series is 11,000 lines across 67 patches. Linux userland SVE state is 99% implemented; the first‑fault register remains untested. DBT-style benchmarks drop from 54s → 38s for an SVE binary versus NEON in QEMU using softfloat, and later host-FPU fallbacks show 2–3× speedups.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 13:26TCG vector size challenges
- 22:21Half precision floating point
- 34:15Modular upstreaming approach
- 36:45Demo of SVE benchmarks
- 41:10Future work and Q&A
related talks
42mThe Big-endian RISC-V Linux Adventure
20mEnhance the Performance of QEMU RVV Load/Store Implementation
19mPorting SLEEF to RISC-V
32mDeveloping the RISC-V Hypervisor Extensions in QEMU
19mThe RISC-V Software Ecosystem: Primed for the Latest ISA Extensions
15mPushing the Packed SIMD Extension Over the Line: An Update on the Progress of Key RISC...
17mUnlocking 15% More Performance: A Case Study in LLVM Optimization for RISC-V
17mAccelerating Shared Library Execution in a DBT