Expand description
Portable vs host-ISA vs Silicera selection harness.
Identical inputs, sample counts, warmups, and improvement thresholds across
three arms. The integer domain uses a real runtime AVX2 path when detected
(HostIsaReduce); other domains still use algorithmic stand-ins.
Separately compiled -march=native binaries remain a future comparison track.
Domains: integer, memory (multi size-class), float, branch, concurrency.
Structs§
- ArmResult
- One arm result.
- Arms
Harness Config - Shared harness knobs (must be identical across arms).
- Arms
Report - Full three-arm report for one workload family.
- Harness
Suite Report - Aggregate harness run across domains.
Functions§
- run_
branch_ arms - Run branch-domain three-arm harness.
- run_
concurrency_ arms - Run concurrency-domain three-arm harness.
- run_
domain_ harness - Run harness for a domain:
all|integer|memory|float|branch|concurrency. - run_
float_ arms - Run float-domain three-arm harness.
- run_
full_ harness - Run both integer and memory harnesses (legacy).
- run_
integer_ arms - Run integer-domain three-arm harness.
- run_
memory_ arms - Run memory-domain three-arm harness (scan vs prefetch-friendly).
- summarize_
size_ winner_ changes - Summarize whether memory size-class winners differ across L1/L2/L3/DRAM.