860794 - OdinMonkey: investigate slowdown on matrix multiplication

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine, defect)

Description

[Luke Wagner [:luke]]

For the attached code, the handwritten (non-asm.js) code is slightly faster than the asm.js on OdinMonkey.  It's a tiny kernel doing double math and float32 loads/stores so I'd expect a slight speedup.

Comment 1

[Luke Wagner [:luke]]

Attachment: handwritten

Comment 2

[Luke Wagner [:luke]]

Attachment: equivalent asm.js

Comment 3

[Luke Wagner [:luke]]

Actually, on my faster desktop, they are about the same.  My MBP showed the slowdown.  Perhaps misaligned loads?

Comment 4

[Nicolas B. Pierron [:nbp]]

(In reply to Luke Wagner [:luke] from comment #3)
> Actually, on my faster desktop, they are about the same.  My MBP showed the
> slowdown.  Perhaps misaligned loads?

Isn't the hand written one only doing integer arithmetic, because there is only identity matrix flowing in, and operations such as multiplications and additions are monitored as being numbers?

Comment 5

[Luke Wagner [:luke]]

Great question; I tried that last night and I didn't see a change.  For one thing, the multiplied values are flowing out of and into a Float32Array.

Comment 6

[Luke Wagner [:luke]]

I checked again and asm.js does seem slightly faster, both in the browser and in the shell.  valgrind shows less instructions and less data references.