Gallium3D's LLVMpipe Is Much Faster With Mesa 9.2

[bridgman]

You jumped from OpenGL (which the fixed-function blocks are designed to accelerate) to OpenCL (where those blocks won't make much difference except in cases where you can use the texture engines to pre-filter values from an input array before the kernel uses them). Did you mean OpenGL or OpenCL ?

[jakubo]

apparently i know too little of all that so it was just a wild stab in the dark.

[bridgman]

No worries... keep asking questions, but maybe use more general words to describe what you're thinking about ?

[droidhacker]

Utterly USELESS test that indicates absolutely nothing useful AT ALL. i7? Seriously? When will you ever find one of those without some usable GPU?

I will say this again; if you want to test llvmpipe opengl, use one of those old or weak systems that doesn't have a usable GPU, like, oh, I don't know... intel Z5xx.


And don't test games. Test their ability to run a composited desktop fluently, without significantly impacting background processes or maxing out the crappy CPU.


There are precisely TWO places where this fallback driver will be used; crappy and/or old hardware without GPU, or servers-run-by-morons, which have plenty of CPU to spare on compositing.

[Kivada]

Anyone can simply reproduce the tests I have done and investigate further. Everything is automated down to the performance Git bisecting. But I don't have the time to run all of that when I have other things to benchmark and other work.

Thats not your reader's job, it's yours.

You want to play reviewer, you do those investigations. Reviewers for Mac and Windows do this, with OSS it should be even easier for you to do it.

[Michael]

Thats not your reader's job, it's yours.

You want to play reviewer, you do those investigations. Reviewers for Mac and Windows do this, with OSS it should be even easier for you to do it.

I provide the content as-is, how I want, and what I can do with the resources I have for running this site by myself. I don't care how other sites do it. If you don't like it, you can leave. Those that want to dig deeper are free to. I provide plenty of guidance for digging deeper for my commercial customers and those developers wishing to dig deeper into problems shown on Phoronix within their code.

[jrch2k8]

Utterly USELESS test that indicates absolutely nothing useful AT ALL. i7? Seriously? When will you ever find one of those without some usable GPU?

I will say this again; if you want to test llvmpipe opengl, use one of those old or weak systems that doesn't have a usable GPU, like, oh, I don't know... intel Z5xx.


And don't test games. Test their ability to run a composited desktop fluently, without significantly impacting background processes or maxing out the crappy CPU.


There are precisely TWO places where this fallback driver will be used; crappy and/or old hardware without GPU, or servers-run-by-morons, which have plenty of CPU to spare on compositing.

well the answer is could be, i mean no matter how wonderful an llvm renderer could be and is irrelevant how much optimization you do in LLVM, the performance ceil will scale very predictably based on how powerful and wide your SIMD block are and how much thread you have avalailable to distribute the work, so the difference between an prescott P4 and an i7 haswell will be massive as can be easily expected.

so in a single threaded MMX/SSE CPU you can expect a very very basic composite to be usable but the moment you try to use shaders all will go to hell because you can't parallelize operations and your SIMD units are too narrow, too few and too slow and well an Atom is even worse.

why an LLVM backend then, LOL? because they can mostly and is LOOOOT better than the old software fallback and with something like wayland provide a nice fallback, i mean for wayland it has enough punch to run smooth in small CPU's even ARM as long your DE push down the eyecandy but for fast fluid window management with some nice shadows and video LLVM is efficient enough but for example color managent or complex shading effect will test your patience.[as a side note LLVM force Kwin to software render and turn off all heavy effects].

So, it make sense to test old CPU? no, not yet X11 will eat it alive[specially Xrender] but it will make more sense once wayland is here since you can manage more efficiently the resources and you remove the massive roundtrips but sure Micheal could do it and show a bunch 1FPS benches and yeah it show the desktop but if you move a window the CPU jump to 100% and lag

[jrch2k8]

I provide the content as-is, how I want, and what I can do with the resources I have for running this site by myself. I don't care how other sites do it. If you don't like it, you can leave. Those that want to dig deeper are free to. I provide plenty of guidance for digging deeper for my commercial customers and those developers wishing to dig deeper into problems shown on Phoronix within their code.

well for the purpose of this test is done good enough Michael[show the relative performance grow between LLVM and Mesa releases] but sometimes people have misunderstanding or lack of know how that guide them to believe tests are miraculous events that should test scenario where the outcome is ovbious like in this case and based on those spectations they get frustated thinking their scanario is valid.

for those amazed with the performance the answer is:

1.) autovectorization is active[driver/LLVM backend] and it supports now AVX/AVX2/FMA3-4
2.) older CPUs can hardly run a decent[they miss too much silicon] composite, wait for wayland since X11 is not friendly for CPU rendering
3.) LLVM improvements in something like a P4 are almost noticeable
4.) There is no magic algorithm that make any CPU go twice fast, is just using properly new technologies available on the current CPUs + some improvements

i would only recommend Michael is to make a wider test once the first wayland desktop enviroment is available for public testing, this way is easier to appreciate why this backed exists