@tehjh Not sure how well this is known. Branch history table can encode very long sequences (for the same branch); extractible similar to genome assembly from short reads. Managed to store 1000 length history in https://github.com/JuliaLang/julia/pull/29888 …. BranchScope for long sequences ;)
-
-
I am just as amazed; this came out of analyzing a spurious regression report in the ci-benchmarks for julialang. In retrospect, I feel stupid and the mechanism is obvious corollary of Agner Fog's description. CPUs were laptop broadwell and server haswell.
-
1/ But because obvious only in retrospect, let me explain. Setting: We branch on a sequence with density 0.5; periodic with period N. CPU wins if it manages to learn this random N-bit sequence, because then it can predict and is fast in badly written benchmarks.
- 7 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.