Page 835 of 837 FirstFirst ... 335735785825833834835836837 LastLast
Results 16,681 to 16,700 of 16726
  1. #16681
    These recent tile changes makes CUDA work similar to OpenCL. If there's differences in the optimal tile size that can be due to a different graphics card, scene, number of samples, denoising, etc, it's not so much a CUDA vs. OpenCL thing. The goal is to make render time independent of tile size, but there's a way to go still.

    Regarding memory usage increases with the daily builds, seems like it might be building without jemalloc.
    Last edited by brecht; 07-Oct-17 at 14:54.



  2. #16682
    Originally Posted by brecht View Post
    This recent tile changes makes CUDA work similar to OpenCL. If there's differences in the optimal tile size that can be due to a different graphics card, scene, number of samples, denoising, etc, it's not so much a CUDA vs. OpenCL thing. The goal is to make render time independent of tile size, but there's a way to go still.

    Regarding memory usage increases with the daily builds, seems like it might be building without jemalloc.
    For the record: i always compile with jemalloc ( @YAFU also )

    And btw2: Brecht, you are a monster Again such a big win in renderspeed and usability, especially when using multiple gpu.
    Victor in 7 minutes on 3 gpu now ... Woooow.

    Jens



  3. #16683
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,886
    @jensverwiebe. I am not sure if you had read my message #16678. Problem was from master, but I had not clarified that I was using daily builds from buildbots. Sorry.
    When compiled by myself, I do not get the excess of RAM problem. I guess I'm compiling with that "jemalloc" thing
    Be patient, English is not my language.



  4. #16684
    Originally Posted by YAFU View Post
    @jensverwiebe. I am not sure if you had read my message #16678. Problem was from master, but I had not clarified that I was using daily builds from buildbots. Sorry.
    When compiled by myself, I do not get the excess of RAM problem. I guess I'm compiling with that "jemalloc" thing
    Did read #16678, just wanted clarify then Brecht hit the nail on the head now.

    Jens



  5. #16685
    Member
    Join Date
    Sep 2012
    Posts
    2,685
    32 seconds "Tile Size TEST" w/ BMW27.blend // W7x64 /// GTX 1060 6GB

    Official 2.79 VS daily 83ce028 (07. 10. 2017)




    PS
    How can one set Cycles to stop rendering after certain amount of time? Is it impossible? or am i loosing my mind...


    *ATS 256 - Auto Tile Size addon set to: Po2 256 (240x180) || also used for finals
    Last edited by burnin; 07-Oct-17 at 16:28. Reason: info added



  6. #16686
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,886
    Brecht has given us a new thing to play with (CPU + GPU render!):
    https://developer.blender.org/D2873

    BMW27.blend scene from this forum. GTX 960. i7-3770. 32x32 tiles size. Linux 64bits:

    Master (buildbot):
    GPU 01:31.63
    CPU 02:47.15

    Master + Patch compiled by myself, CUDA 8
    GPU 01:52.56
    CPU 02:54.17
    GPU + CPU 01:13.40

    My best render time ever!


    Edit:
    I have not quite understood if volumetrics are disabled with the patch. Anyway I have done a test and the final result looks the same. This is, in those heavy scenes for GPU, where you have similar render times with CPU and with GPU, render with CPU + GPU saves a lot of time!
    Example of times in this kind of scene (with denoiser):
    GPU 04:22.38
    CPU 05:24.06
    GPU+CPU 02:24.50

    Edit 2:
    jensverwiebe explained to me in IRC that volumes are not disabled, but an inefficient bvh must be used to render across devices. < Thanks.
    Last edited by YAFU; 08-Oct-17 at 09:21.
    Be patient, English is not my language.



  7. #16687
    That would be nice.. if it works in my case it would double my speed of rendering. (my old GPU and CPU run in about equal times)



  8. #16688
    Originally Posted by Razorblade View Post
    That would be nice.. if it works in my case it would double my speed of rendering. (my old GPU and CPU run in about equal times)
    with some small code changes, OpenCL, CUDA and CPU can render together. BMW at 1225spp renders in 52sec on Vega64+1080Ti at 256x256. Using CPU makes it slower however at the moment, I'm investigating the bottleneck here.
    If you want cool new features for Blender and Cycles ( like 1.5x to 2x faster renderings ), you can donate here https://www.paypal.me/matmenu and https://www.patreon.com/matmenu



  9. #16689
    Member
    Join Date
    Oct 2011
    Location
    Stockholm
    Posts
    213
    Smaller tiles should also work well for adaptive sampling, shouldn't it? Looks like we are gonna get a lovely speedup with all of these, thanks Brecht.



  10. #16690
    Member
    Join Date
    Apr 2012
    Posts
    1,792
    Originally Posted by YAFU View Post
    Brecht has given us a new thing to play with (CPU + GPU render!):
    https://developer.blender.org/D2873

    BMW27.blend scene from this forum. GTX 960. i7-3770. 32x32 tiles size. Linux 64bits:

    Master (buildbot):
    GPU 01:31.63
    CPU 02:47.15

    Master + Patch compiled by myself, CUDA 8
    GPU 01:52.56
    CPU 02:54.17
    GPU + CPU 01:13.40

    My best render time ever!
    Is this working in a manner I suggested a few posts back?



  11. #16691
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,886
    Hi moony.
    As I said before, I'm not sure if you in your calculations took into account that Cycles rendering on GPU needs a dedicated CPU thread per GPU. So if you have two GPUs, you calculate CPU render times as if it had two less threads.

    Edit:
    Talking about CUDA. I think "bliblubli" said in IRC that OpenCL render does not use CPU threads.

    Edit 2:
    Mmm, testing OpenCL in nvidia and it uses one CPU thread at 100%. So I'm not sure now. Maybe ATI only feature.

    Edit 3:
    Here is a comparison of CPU usage rendering with GPU only (32x32), between Buildbot (small tiles feature only), and compiled with the CPU+GPU patch (BMW27.blend from this forum):
    Buildbot (GPU render only):


    GPU+CPU patch (GPU render only):


    Perhaps this explains the slowness I'm having when rendering in patched version with only one device (BMW27.blend)?

    Or does patched version no longer make use of one CPU thread per GPU?
    Last edited by YAFU; 09-Oct-17 at 05:35.
    Be patient, English is not my language.



  12. #16692
    @YAFU can you upload the patch? For CPU+GPU rendering?
    Thanks



  13. #16693
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,886
    Originally Posted by drgci View Post
    @YAFU can you upload the patch? For CPU+GPU rendering?
    Thanks
    Here is the patch:
    https://developer.blender.org/D2873
    (on the right, Download Raw Diff)

    If you are referring to a build, I use Linux and I do not know how to build portable builds.
    Be patient, English is not my language.



  14. #16694
    Member
    Join Date
    Apr 2012
    Posts
    1,792
    Originally Posted by YAFU View Post
    Hi moony.
    As I said before, I'm not sure if you in your calculations took into account that Cycles rendering on GPU needs a dedicated CPU thread per GPU. So if you have two GPUs, you calculate CPU render times as if it had two less threads.
    You are correct - my calculations didn't take that into account. I assumed all 12 threads my CPU and 1 thread each for the two GPUs.

    Taking into account what you said - i'd still get 10 threads from the CPU and 1 each from the GPU which would still result in a significant time saving overall.



  15. #16695
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,886
    Here my devices with the latest version of the patch and enabling debug mode:

    intel Core i7-3770 CPU @ 3.4GHz (OpenCL) appears twice because I also have driver installed for HD 4000 on Linux, anyway it never worked in iGPU.

    I suppose it does not make sense (or I will not save time) by having two items enabled at the same time occupying the same physical device.

    Edit:
    Regarding OpenCL for iGPU, it is very likely that I have not installed OpenCL driver correctly. I'm not even sure what it was I had installed. At that time, intel made it difficult to install on Linux. I should clean all that garbage out of my system.
    Last edited by YAFU; 11-Oct-17 at 21:47.
    Be patient, English is not my language.



  16. #16696
    How do you actually patch it? I am on OSX is it for version 2.79 or 2.8?



  17. #16697
    Originally Posted by animani View Post
    How do you actually patch it? I am on OSX is it for version 2.79 or 2.8?
    You download the .diff file and in the same folder where you enter "make" command enter "git apply /Users/.../D2873.diff" to the real location of the diff file.



  18. #16698
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,886
    Originally Posted by animani View Post
    is it for version 2.79 or 2.8?
    I tested it on Master (2.79)
    Be patient, English is not my language.



  19. #16699
    Originally Posted by DcVertice View Post
    You download the .diff file and in the same folder where you enter "make" command enter "git apply /Users/.../D2873.diff" to the real location of the diff file.
    Is this the right link to clone from?

    http://git.blender.org/blender.git



  20. #16700
    Originally Posted by YAFU View Post
    I tested it on Master (2.79)
    Are you on Windows or Mac? It seems Mac compiling is not as easy as on Windows. For me Make does not work and I am trying to use CMake but I get some other errors.

    Not sure if I even have the 2.79 version. How could I know that?

    If I try to apply the DIFF file I get this:

    error: intern/cycles/blender/addon/properties.py: No such file or directory
    error: intern/cycles/blender/addon/ui.py: No such file or directory
    error: intern/cycles/blender/addon/version_update.py: No such file or directory
    error: intern/cycles/blender/blender_sync.cpp: No such file or directory
    error: intern/cycles/device/device.cpp: No such file or directory
    error: intern/cycles/device/device_cuda.cpp: No such file or directory
    error: intern/cycles/device/opencl/opencl_util.cpp: No such file or directory
    error: source/blender/blenkernel/BKE_blender_version.h: No such file or directory



Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •