TypeTree speedup shiftindicies

[wsmoses]

Fixes #1743

[wsmoses]

note I left the slow version right now to do correctness tests, which already found a bug locally (which have fixed). Going to run it on the whole test suite, then disable the check once it comes back valid.

[wsmoses]

Also note this fix hopefully changes the asymptotic behavior of ShiftIndices from $O(n^2)$ to $O(n)$ [which it really should've been the whole time]

[vchuravy]

Seems like a speedup from 220s to 180s, but TypeTree::insert and TypeTree:checkedOrIn are still very hot

Before:

After:

Detailed analysis:

• checkedOrIn ~28s (42% of compilation time)
  • 17s from update Analysis
  • 10.5s from EnzymeMergeTypeTree
• CanonicalizeInPlace: 11s

[ZuseZ4]

@vchuravy If already at it, could you please also profile one from #1641 if you have a few minutes? My server never had the permissions for vtune. My new Laptop is reasonable otherwise, so I can also set it up myself

[wsmoses]

Considering I don’t see ShiftIndices more on the flame graph I’m going to call this a win. I’ll see what I can do about the Julia function. mergeTypeTree is the C API and primarily used in that function so I think it’s the next most meaningful place we should be able to reduce overhead from. Essentially because it should be not terrible to make it called only once. CheckedOrIn however doesn’t have any obvious speed ups I remember.

…

On Tue, Feb 20, 2024 at 2:02 AM Manuel Drehwald ***@***.***> wrote: @vchuravy <https://github.com/vchuravy> If already at it, could you please also profile #1641 <#1641> if you have a few minutes? My server never had the permissions for vtune. My new Laptop is reasonable otherwise, so I can also set it up myself — Reply to this email directly, view it on GitHub <#1744 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJTUXDS4ZNE23F3VAS4O43YURC7BAVCNFSM6AAAAABDQF76POVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNJTGU4TKMZVGQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>