Beating the Benchmarks: Co-evolving the ISA and Development Tools
summary
An open workflow uses QEMU, GCC and LLVM to trace and profile EEMBC CoreMark, SPEC CPU 2017 floating‑point workloads and ECMAScript, then drives ISA changes backed by numbers. New rounding ops produce 9–19% speedups on hot FP paths; a tiny load‑FP‑immediate encoding uses about 5 bits to cover common immediates; conditional operation extensions (Ventana condops / XVentanaCondOps) save roughly two instructions per branchless pattern and reduce overall instruction count by about 1%, at the cost of ~16 opcode points.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:50Benchmarks
- 3:06Premature optimization
- 4:04Evolution
- 10:13Evolution Process
- 14:16Examples
- 18:01Case Studies
- 22:00Conclusion
related talks
01h 00mFastware
56mWrite Fast Code Like a Native
21mDriving Innovation: Evolving the Role of Software in the RISC-V Ecosystem Beyond...
01h 01mGoing Nowhere Faster
44mFastware
18mRISC-V Customization After a Tape-out
01h 04mUnlocking Performance Through Reverse Engineering
17mUnleashing the Power of RISC-V E-Trace with a Highly Efficient Software Decoder