5.5x faster, half as many divisions, recovering accuracy #99
Conversation
PallHaraldsson
changed the title
|
@PallHaraldsson I appreciate your effort trying to improve the Julia code. Maybe this is related to the blocker in #57? |
I don't know if this line is needed, those are the (latest) package versions I used (assume 0.5 works too): ⌃ [81625895] StaticCompiler v0.4.12 [86c06d3c] StaticTools v0.8.8 I used Julia 1.9.0, not (that should also work and 1.8.5, doubt any have speed advantage): ``` FROM julia:1.8.2 ```
|
Hi @niklas-heer, I believe this is done now finally after a long wait, and the speed maintained hopefully, after a fix for StaticCompiler.jl. A "sufficiently smart compiler" could in theory figure out my transformation, but I very much doubt any will (then would do already), or will come soon, since the optimization isn't very general-purpose? Why is "Julia (ux4)" marked specially in the graph? Was it at some point the fastest Julia version? Or some other reason for it? Maybe just drop it, in a later PR? |
|
Please hold of on merging. I believe I timed 5x faster for sure originally (in Julia REPL), but StaticCompiler.jl couldn't handle that code. Now after fixing it seems slower, I need to see why, and for sure compare to all the right versions. |
No description provided.