OpenCL Nightly Builds

[s55]

Performance hasn't changed. The scaling was only ever 5~6% CPU time, so that's only the kind of speedup you can expect. You get that in some situations, no improvement in others and slowdowns in somecases too. Not a lot more can be done with it (bar fixing the crash on Nvidia hardware which hasn't happened as other things have now taken priority)

Note, Intel Haswell GPU's get no real benifit at the moment, although if zero-copy ever gets enabled, it may be able to achieve the same small speedup as AMD cards.

Not sure much has changed with the x264 lookahead code. Afaik it works, but last I checked it only really benefited if it wasn't looking far ahead. I.e a value of around 10 rathern than the better default of 40.
(Can sometimes provide a few fps faster)

[AAT]

Thanks for the update, I will run some benchmarks to see how is working with AMD APU and Intel processor. I think 5-6% also better than nothing, but back than I was hoping much more performance increase.

[Smithcraft]

I think we were all hoping for more of an increase. And not just when downscaling.

SC

[AAT]

I've just run my regular Dallas tests (no down scaling) to see how it performs on different platforms. Unfortunately I could not use svn5938 nighty build for both OpenCl and non-OpenCl tests, since it does not have OpenCL switch anymore (I guess it runs default with OpenCL acceleration) and I compared performance with the non accelerated 0.9.9-1 version.


0.9.9-1 version;
AMD E2-3000M APU: 26minutes 38sec
Intel I5-3337U: 10minutes 41sec

svn5938 nighty build (OpenCL);
AMD E2-3000M APU: 23minutes 08sec
Intel I5-3337U: 9minutes 17sec
Intel I5-3337U + Radeon 8750M: 10minutes 23sec


Consequences;
If the performance boost are gained due to OpenCL acceleration (and not because I've used two different version Handbrake), than that is mean for AMD E2-3000M brings 15% boost, for Intel I5-3337U also 15%. 15% is actually modest improvement, wonder if this all come from OpenCL acceleration? The output file size stayed all the case 521,9xMB and I've used same settings in Handbrakes. It is also Interesting when I paired I5-3337U with Radeon 8750M, the test run slower as with the default HD4000 IGP...

[s55]

The choice of scaler is now in preferences. Lanczos is still default. Bicubic-CL is optional.

There are performance and setting tweaks in software from 0.9.9 to current, so it's not comparable.
There is no performance benefit when no scaling is applied.

As I've already said, the code that is being accelerated, was only ever around 5% CPU time and isn't normally a bottleneck so you can't expect to gain anymore than that.

[Deleted User 11865]

OpenCL is no longer a switch, it's now a scaling algorithm (you can choose between lanczos and opencl somewhere in Tools > Options).

[AAT]

I see, than 15% is nice improvement from "performance and setting tweaks" and unfortunately nothing extra from OpenCL at least for me because I'm not doing down-scaling. I thought the OpenCl look-ahead patch will work in non down scaling situations too.

[Deleted User 11865]

You need to enable it - opencl=1 in your advanced x264 options, IIRC.

[Stella]

OpenCL is no longer a switch, it's now a scaling algorithm (you can choose between lanczos and opencl somewhere in Tools > Options).

Are we talking about the latest "generic" nightly build? Since there is no longer a specific OpenCL nightly download.

[Deleted User 11865]

OpenCL is no longer a switch, it's now a scaling algorithm (you can choose between lanczos and opencl somewhere in Tools > Options).

Are we talking about the latest "generic" nightly build? Since there is no longer a specific OpenCL nightly download.

Yes.