Many hyperlinks are disabled.
Use anonymous login
to enable hyperlinks.
7 check-ins related to "highway-simd-experiment"
|
2025-12-02
| ||
| 17:04 |
separating the scalar and simd-based loops from each other, i.e. use iteration over vector length as inner loop instead of outer has not changed the result in a material way. we now only know that it is the scalar loops mostly driving the cpu temperature, not the simd-loops. with this the simd experiment is closed. in the following I will work on unrolling more core loops as means of | |
| 12:25 |
benchmarked unrolled simd-loops to see if that gains us performance. result: no, it does not. each level of unrolling is has actually worse strongly suspecting that the simd loops encounter thermal throttling and some behaviour of the scalar loops may indicate that even they get | |
|
2025-12-01
| ||
| 14:23 |
fix: missed two, remove references to ops which were not compared. check-in: 2a9b817a3a user: aku tags: highway-simd-experiment | |
| 14:19 |
fix: remove references to ops which were not compared. check-in: 0444d7616c user: aku tags: highway-simd-experiment | |
| 14:08 |
benchmarking the new simd code, and comparing against the 4-unrolled super-scalar code the results heavily favor the latter, except for some ops and very large vectors. as noted in the associated README, this is likely because we have only | |
| 14:03 |
fixed all the issues in the pre-existing unused theoretical highway code (templates, generator, etc). with the fixes in place, and activated highway code is generated disabled the vector.bench(mark) file, and corrected a typo in the | |
|
2025-11-28
| ||
| 17:48 |
chore: put the benchmarks results for loop-unrolling of the vector math into the permanent record. note: even the readme is generated. check-in: 4c6a79be86 user: aku tags: trunk | |