Phoronix is a leading source for Linux hardware reviews, benchmarking, and open-source news. With  performance analysis, hardware testing, and coverage of the latest developments in the Linux ecosystem, Phoronix provides  insights for Linux enthusiasts, developers, and system administrators, helping them make informed decisions about hardware compatibility, performance optimization, and software selection.

Phoronix

AMD is implementing device-side Profile Guided Optimization (PGO) for ROCm/HIP to improve GPU kernel performance. The new approach includes uniformity-aware PGO that detects whether GPU branches are uniform or divergent at runtime, preventing performance regressions that standard CPU PGO techniques can cause on GPUs. The implementation uses wave-aggregated counter increments and uniformity detection to achieve 12-14% speedups on uniform branches while avoiding regressions on divergent branches, where naive PGO could cause up to 3.7x slowdowns due to poor memory coalescing.

AMD Squeezing Out More More ROCm/HIP Performance With New Device-Side PGO