RISC-V Vector: Current Status and Next?
summary
RISC-V currently runs Vector in user mode with eager save/restore and deferred restore before return to user space, and kernel-mode Vector support is in patches (V3/V4) adding vectorized memcpy/memmove and copy_to/from_user variants. Tests show blind Vector use gives 15% regression on many small (~100 B) copies but up to 54% improvement on large network copies. Work implements per-CPU tracking, a boot-time enable with an ST flag to cut ~100 cycles of overhead, and recommends thresholding (DT or runtime detect) to pick scalar vs Vector.
AI-written summary · about
chapters (8)
- 0:00Introduction to RISC-V Vector
- 1:01Current status of Vector
- 2:31Challenges with Kernel Vector
- 4:01Optimizing Kernel Vector performance
- 5:47Optimal size for Vector operations
- 7:15Memory bandwidth considerations
- 9:07User space Vector handling
- 10:31Runtime detection of Vector thresholds
related talks
32mRISC-V ftrace: working with preemption
27mTuning in-kernel routines on RISC-V
21mRISC V Hypervisor Support
25mPreparing RISC-V Linux for RVA23
50mInitializing RISC-V A Guided Tour for ARM Developers
28mGHC’s RISC-V Native Code Generation Backend
13mRISC-V 32-bit glibc port
26mThe Big-endian RISC-V Linux Adventure