484624 - improve math performance with doubles (matrix-mult)

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine, defect)

Description

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: javascript benchmark

Here's a simple matrix multiply benchmark that we should do much better on.  Sticking this here to keep the code around.

Current tm on my MBP is around 1800ms; interp is around 11000ms.  That's not a bad speedup, but a pure C version is around 65ms... so we still have a good amount to improve.

Comment 1

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: c++ benchmark

Comment 2

[Ryan VanderMeulen [:RyanVM]]

Current JS shell numbers for the attached testcase:
Interp: 9508.995 ms
-j:     9581.010 ms
-m:     2067.728 ms
-m -n:  268.358 ms

JM+TI is a huge win on this testcase. I don't have the pure C version compiled, but JM+TI appears to be within a factor of 4-5 of it. Not sure how fast this needs to be to consider it "fixed", though.

Comment 3

[André Bargull [:anba]]

No longer reproducible - Resolving as WFM.

With iteration count increased from 1.000.000 to 10.000.000, SpiderMonkey finishes the benchmark in ~68ms whereas the cpp version needs ~345ms.