Ragnar {Groot Koerkamp} @curiouscoding.nl · Mar 13

Lastly, regions that pass this suffix-filter with cost <=k are verified fully in a 2nd full-width stage that uses SIMD with fewer but larger lanes. The result: throughput that is independent of text length, 2-4x higher than v1 (with AVX2/512) for many patterns and long texts, up to 13x for 150bp.

0 likes 1 replies

?

Replies

Ragnar {Groot Koerkamp} · Mar 13

Outside of Barbell, we're already seeing lots of applications for CRISPR off-target detection. Searching 23bp patterns with k=3 gives 105Gbp/s total throughput over all patterns on 16 threads (as in, 32 patterns are searched at the same time at ~3Gbp/s of text each).