619761 - Google blob WebGL sample 2x faster in Chromium

Status: RESOLVED INCOMPLETE

(Core :: JavaScript Engine, defect)

Description

[David Humphrey (:humph)]

I'm not sure if this belongs here or in Canvas:WebGL, but running this demo, I'm about 2x slower on trunk than in Chromium (based on fps readings on that page).

Shark shows a bunch of js at the top of the profile instead of gl:

8.4%	8.4%	XUL	js::mjit::stubs::GetElem(js::VMFrame&)
8.2%	8.2%	XUL	TypedArrayTemplate<float>::obj_getProperty(JSContext*, JSObject*, JSObject*, long, js::Value*)
4.6%	4.6%	XUL	void js::mjit::stubs::SetElem<0>(js::VMFrame&)
2.8%	2.8%	XUL	TypedArrayTemplate<float>::obj_setProperty(JSContext*, JSObject*, long, js::Value*, int)
2.7%	2.7%	XUL	NameOp(js::VMFrame&, JSObject*, bool)
2.6%	2.6%	libSystem.B.dylib	__memcpy
2.0%	2.0%	XUL	js_NativeGet(JSContext*, JSObject*, JSObject*, s::Shape const*, unsigned int, js::Value*)
1.3%	1.3%	XUL	js::mjit::stubs::Equal(js::VMFrame&)
1.1%	1.1%	mach_kernel	vm_map_lookup_entry
1.0%	1.0%	XUL	TypedArrayTemplate<int>::obj_getProperty(JSContext*, JSObject*, JSObject*, long, js::Value*)
1.0%	1.0%	Unknown Library	0x1193172c0 [unknown]
1.0%	1.0%	XUL	EqualityCompiler::update()
1.0%	1.0%	XUL	CallPropertyOp(JSContext*, JSObject*, long, js::Value*, JSCallPropertyKind, int)
0.8%	0.8%	XUL	DisabledGetElem(js::VMFrame&, js::mjit::ic::GetElementIC*)
0.7%	0.7%	XUL	js::mjit::stubs::CallName(js::VMFrame&)
0.7%	0.7%	mach_kernel	ml_set_interrupts_enabled
0.7%	0.7%	IOACPIFamily	IOACPIPlatformDevice::ioRead8(unsigned short, IOMemoryMap*)
0.4%	0.4%	XUL	js::mjit::ic::Equality(js::VMFrame&, js::mjit::ic::EqualityICInfo*)
0.3%	0.3%	libobjc.A.dylib	objc_msgSend
0.3%	0.3%	GeForceGLDriver	glrCompExecuteKernel
0.2%	0.2%	Unknown Library	0x119317275 [unknown]
0.2%	0.2%	Unknown Library	0x119317282 [unknown]
0.1%	0.1%	Unknown Library	0x1193172a4 [unknown]
0.1%	0.1%	Unknown Library	0x1193172a8 [unknown]
0.1%	0.1%	Unknown Library	0x1193172bc [unknown]
0.1%	0.1%	XUL	js::GetCallVar(JSContext*, JSObject*, long, js::Value*)
0.1%	0.1%	XUL	nsIFrame::BuildDisplayListForChild(nsDisplayListBuilder*, nsIFrame*, nsRect const&, nsDisplayListSet const&, unsigned int)
0.1%	0.1%	Unknown Library	0x1196bb250 [unknown]
0.1%	0.1%	Unknown Library	0x1193172b2 [unknown]
0.1%	0.1%	GeForceGLDriver	gldCopyTexSubImage
0.1%	0.1%	Unknown Library	0x1196bb110 [unknown]

Comment 1

[Mike Shaver (:shaver emeritus)]

that profile section is < 30% of our time, though, so if we're 2x slower then there is a bunch of other stuff that's missing.  wonder if it's hidden kernel/system time or something.

Comment 2

[Boris Zbarsky [:bzbarsky]]

So my profile shows:

  44% in mjit-generated code (symbol-less, so wouldn't show up in David's
      "heavy" output unless he did some blaming of "Unknown Library" to
      callers).
  20% under stubs::GetElem (half self time, half typed array float stuff).
      This build doesn't have the in-progress patches for bug 594247, though.
   7% under stubs::SetElem (again, on typed arrays).
   7% under stubs::CallName (mostly self time in NameOp and js_NativeGet)
   5% in tjit-generated code (again, symbol-less)
   3% under ic::Equality
   2% in DisabledGetElem

and a long tail of GL stuff and stubs::GetProp and such, but it all adds up to just 12% given the numbers above.

Comment 3

[Boris Zbarsky [:bzbarsky]]

We should probably remeasure with the patches for bug 619761.

Comment 4

[Boris Zbarsky [:bzbarsky]]

Which, btw, don't apply to TM tip, making it difficult to test.  ;)

Comment 5

[Nicholas Nethercote [inactive]]

(In reply to comment #3)
> We should probably remeasure with the patches for bug 619761.

You mean bug 594247?  Bug 619761 is this bug.

Comment 6

[Boris Zbarsky [:bzbarsky]]

> You mean bug 594247?

Er, yes.

Comment 7

[David Mandelin [:dmandelin]]

I tried to recheck this just now. In Chrome 10, the fps meter displays "fps: 1", so I can't compare that way. The animation looks similar to me in the 2 browsers, though. Bug in Chrome?

Comment 8

[José Jeria]

Latest stable Chrome on Windows 7 gives me 60 fps and latest Firefox trunk give me around 28 fps. I don't see what comment 7 sees.

Comment 9

[Boris Zbarsky [:bzbarsky]]

José, clearly there's a new bug in Chrome 10 (which is not the stable version, note).

Comment 10

[David Mandelin [:dmandelin]]

Much of the time is in gfx, but there is a substantial amount in JS. Most of the slow calls seem to be NameOp, and they are for code like this:

function top() {
  function f() {
  }

  function g() {
    f();   // lookup of f takes a slow path
  }
}

Comment 11

[Guilherme Lima]

Just compared Nightly and Chrome 29. Both ran 100 blobs at the 48^3 resolution around 15 fps. Chrome was very stable while Firefox wasn't, but I think that's due to garbage collection.

Comment 12

[Mayank Bansal]

Link doesnt work. Calling this fixed.