LuxMark v3.1: LuxRender propaganda with OpenCL

Discussions related to GPU Acceleration in LuxRender

Moderators: Dade, jromang, tomb, coordinators

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby kroko » Wed Sep 16, 2015 12:45 pm

@pciccone
Feedback Assistant reports do not show in bugtracker! As said, this is some different end-users feedback system. However if you wish I can push this in bugreport?
However from what I've read about your struggle this was to no avail.
Therefore I did not try that, but used this special system to report on OS X E Capitan Public Beta (it allows also submitting things about iOS 9 Beta, and previously Yosemite Public Beta, iOS 8 Beta) and suggest this approach also to you. Complaining as end-user, explaining crippled OS X from a "simple" POV, not as developer.
EDIT: Maybe my entry has not been processed yet and when it does it will fall into bugtracker under the account I used to sign in Feedback Assistant, dunno. Have to wait for response (if any comes).

@jensverwiebe
Yep, GeForce GTX 980 Ti is a beast, single one giving 10317. When we bought 980s for a project (it was for multi-4k-screen-interactive installation @60fps where I coded everything to be run on GPUs (as no CPU would manage that), optimised GLSL shaders, cached VBOs, VAOs blabla, ran on Ubuntu) Ti was not out yet and next step was Titan X. Titan was (and is) mad in both in performance... and price. So much more that we stayed with 980s. Not saying that 980s are cheapos hehe.
Note that my 2xGTX980 on MSW gave 15372, thus 1.5x 1xGTX980Ti. Not that bad actually. Yep, dual D700 gave mere 9675, but I do not find a single place I state that I expect something more from these GPUs. I've been Mac user long enough, you know ;)

I patched, compiled and run a quick test http://pastebin.com/F2PG4Tgc, got
CL_DEVICE_EXTENSIONS: cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_byte_addressable_store cl_khr_image2d_from_buffer cl_khr_depth_images cl_APPLE_command_queue_priority cl_APPLE_command_queue_select_compute_units cl_khr_fp64
CL_DEVICE_PROFILE: FULL_PROFILE
Right.

Should I check for some other specific param_name in clGetDeviceInfo that could help in my discussion with Apple?

Is there any single thing that HAS TO work per spec (and is needed by Lux), but does not work and can be shown by simple C/C++ app plus OpenCL kernel (without showing the whole Lux package)? Like unit testing where unit per unit can be feeded down Apple OpenCL dev throats, or is this done already :)? Is there a repo somewhere of such tests?

Code: Select all
warp-node-03:opencl warpnode$ clang -framework OpenCL dumpcl.c -o dumpcl && ./dumpcl
Found Devices:      3

Device: 0

   Device Type:      CL_DEVICE_TYPE_CPU
   Name:          Intel(R) Xeon(R) CPU E5-1620 v2 @ 3.70GHz
   Vendor:       Intel
   Vendor ID:      -1
   Version:      OpenCL 1.2
   Profile:      FULL_PROFILE
   Driver:       1.1
   Extensions:      cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event cl_khr_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_byte_addressable_store cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_image2d_from_buffer cl_APPLE_fp64_basic_ops cl_APPLE_fixed_alpha_channel_orders cl_APPLE_biased_fixed_point_image_formats cl_APPLE_command_queue_priority
   Extensions:      1.1
   Max Work-Item Sizes:   (1024,1,1)
   Max Work Group Size:   1024
   Max Compute Units:   8
   Max Frequency (Mhz):   3700
   Cache Line (bytes):   10485760
   Global Memory (MB):   16384
   Local Memory (MB):   0
   Local Memory Type:   CL_GLOBAL
   Max Mem Alloc (MB):   4096
   Max Param Size (MB):   4096
   Base Mem Align (bits):   1024
   Address Space (bits):   64
   Image Support:      1
   Float Functionality:   INF and NaN support
   ECC Support:      0
   Exec Functionality:   CL_EXEC_KERNEL CL_EXEC_NATIVE_KERNEL
   Little Endian Device:   1
   Profiling Res (ns):   1
   Device Available:   1

Device: 1

   Device Type:      CL_DEVICE_TYPE_GPU
   Name:          AMD Radeon HD - FirePro D700 Compute Engine
   Vendor:       AMD
   Vendor ID:      16915456
   Version:      OpenCL 1.2
   Profile:      FULL_PROFILE
   Driver:       1.2 (Jul 29 2015 02:43:15)
   Extensions:      cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_byte_addressable_store cl_khr_image2d_from_buffer cl_khr_depth_images cl_APPLE_command_queue_priority cl_APPLE_command_queue_select_compute_units cl_khr_fp64
   Extensions:      1.2 (Jul 29 2015 02:43:15)
   Max Work-Item Sizes:   (256,256,256)
   Max Work Group Size:   256
   Max Compute Units:   32
   Max Frequency (Mhz):   850
   Cache Line (bytes):   0
   Global Memory (MB):   6144
   Local Memory (MB):   0
   Local Memory Type:   CL_LOCAL
   Max Mem Alloc (MB):   1536
   Max Param Size (MB):   1024
   Base Mem Align (bits):   32768
   Address Space (bits):   32
   Image Support:      1
   Float Functionality:   Round to nearest support
   ECC Support:      0
   Exec Functionality:   CL_EXEC_KERNEL
   Little Endian Device:   1
   Profiling Res (ns):   37
   Device Available:   1

Device: 2

   Device Type:      CL_DEVICE_TYPE_GPU
   Name:          AMD Radeon HD - FirePro D700 Compute Engine
   Vendor:       AMD
   Vendor ID:      33692672
   Version:      OpenCL 1.2
   Profile:      FULL_PROFILE
   Driver:       1.2 (Jul 29 2015 02:43:15)
   Extensions:      cl_APPLE_SetMemObjectDestructor cl_APPLE_ContextLoggingFunctions cl_APPLE_clut cl_APPLE_query_kernel_names cl_APPLE_gl_sharing cl_khr_gl_event cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_byte_addressable_store cl_khr_image2d_from_buffer cl_khr_depth_images cl_APPLE_command_queue_priority cl_APPLE_command_queue_select_compute_units cl_khr_fp64
   Extensions:      1.2 (Jul 29 2015 02:43:15)
   Max Work-Item Sizes:   (256,256,256)
   Max Work Group Size:   256
   Max Compute Units:   32
   Max Frequency (Mhz):   150
   Cache Line (bytes):   0
   Global Memory (MB):   6144
   Local Memory (MB):   0
   Local Memory Type:   CL_LOCAL
   Max Mem Alloc (MB):   1536
   Max Param Size (MB):   1024
   Base Mem Align (bits):   32768
   Address Space (bits):   32
   Image Support:      1
   Float Functionality:   Round to nearest support
   ECC Support:      0
   Exec Functionality:   CL_EXEC_KERNEL
   Little Endian Device:   1
   Profiling Res (ns):   37
   Device Available:   1
Last edited by kroko on Wed Sep 16, 2015 11:49 pm, edited 4 times in total.
firmitas utilitas venustas baby!
User avatar
kroko
 
Posts: 39
Joined: Mon Jan 31, 2011 5:58 am
Location: Riga, Latvia

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby jensverwiebe » Wed Sep 16, 2015 12:56 pm

Yep, i have also still 2 gtx 980 in my machine, additionally to the 980 ti.
Ti driving screens + computing, 980's for compute only.
So you have all 4 GB avavailable per gpu, 980 ti minus screen mem, which is often up to 500MB. Ergo the scene
can always take up to 4GB pushing all 3 gpu.

Considering Apple graps 1200€ for the D300->D700 upgrade, one D700 is around 650€ which is even
more expensive than a GTX 980. ( Assuming one D300 would be worth 50€ )

Not a good buy .....

Jens
User avatar
jensverwiebe
Developer
 
Posts: 3420
Joined: Wed Apr 02, 2008 4:34 pm

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby kroko » Wed Sep 16, 2015 1:17 pm

Just checked - CompuBenchCL fails T-Rex scene with CL_BUILD_PROGRAM_FAILURE (the only one with dual GPU support). However all single GPU scenes passes. They are not advanced though (optical flow, FFT).

@jensverwiebe I know. But I HAVE TO use specific toolset that exists on OS X only (i.e. Syphon, Spout is not an option as MSW is not an option, because i need *nix under the hood, as I said - toolset :) ) and at the time this mac pro was bought modding older one with flashed GPUs was not an option for various reasons. If these tools + Adobe + decend CAD existed on BSD/Linux... you know the answer. This is going offtopic, so we better stop here. :)
firmitas utilitas venustas baby!
User avatar
kroko
 
Posts: 39
Joined: Mon Jan 31, 2011 5:58 am
Location: Riga, Latvia

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby Dade » Mon Sep 21, 2015 9:31 am

Some serious fire power at work: http://luxmark.info/node/1156
User avatar
Dade
Developer
 
Posts: 8394
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby kroko » Thu Sep 24, 2015 11:50 am

1)
I pushed some additions to luxrays, lux and luxmark mercurial which allow building lux on OS X 10.11.1 beta with Xcode 7.1 beta 2. Some pull requests are still pending.

2)
Created an updated readme https://github.com/WARP-LAB/LuxRenderOSX
This readme "will build" only after all pull requests are accepted.
Somebody could revise it, if all ok it can be moved to http://www.luxrender.net/wiki/Building_on_OSX

3)
Built all, including luxmark.
No surprises as data found in AMD7000Controller.kext on 10.11.1 shows that no updates for AMD have been made in this minor El Capitan update.

Screen Shot 2015-09-24 at 19.38.29.png
osx10.11.1


4)
But also some good news for Blender comunity as newest Blender RC + Cycles can use OpenCL on OSX (my D700s work). LuxRender still sad panda, but the day will eventually come... :)
firmitas utilitas venustas baby!
User avatar
kroko
 
Posts: 39
Joined: Mon Jan 31, 2011 5:58 am
Location: Riga, Latvia

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby Fangio » Wed Nov 04, 2015 4:45 am

Good news everyone, at least with a Mac. There is progress with Apples OpenCL drivers, for the first time the v3.1 Hotel Lobby scene is rendering here in El Capitan.

It still didn't in OS X 10.11.1, now this is in 10.11.2 dev beta 2 build 15C31f.

LM 3.1 Hotel & cMP 5.1 & 280X.jpg

http://www.luxmark.info/node/1389
Fangio
 
Posts: 6
Joined: Thu Dec 11, 2014 12:57 pm

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby pciccone » Wed Nov 04, 2015 10:21 am

That's good news, at least for AMD drivers. I tested it on a MacBookPro with nVidia GPU and it still crashes.
User avatar
pciccone
Developer
 
Posts: 1579
Joined: Wed Jan 13, 2010 11:02 am
Location: South Carolina

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby kroko » Fri Nov 13, 2015 1:23 pm

I can confirm that all LuxMark scenes pass (Neumann does not has green text on the ring, Hotel also passes) on OS X 10.11.2 public beta 3 (15C40a).

1xD700 LuxBall http://www.luxmark.info/node/1512
1xD700 Microphone http://www.luxmark.info/node/1513
1xD700 Hotel http://www.luxmark.info/node/1514

2xD700 LuxBall http://www.luxmark.info/node/1509
2xD700 Microphone http://www.luxmark.info/node/1510
2xD700 Hotel http://www.luxmark.info/node/1511

AMD7000Controller.kext has pushed version from 1.36.46 to 1.40.15

Code: Select all
<dict>
   <key>BuildVersion</key>
   <string>8</string>
   <key>CFBundleShortVersionString</key>
   <string>1.40.15</string>
   <key>CFBundleVersion</key>
   <string>1.4.0</string>
   <key>ProjectName</key>
   <string>GPUDriversAMD</string>
   <key>SourceVersion</key>
   <string>1040015000000</string>
</dict>
Last edited by kroko on Tue Nov 17, 2015 6:15 am, edited 1 time in total.
firmitas utilitas venustas baby!
User avatar
kroko
 
Posts: 39
Joined: Mon Jan 31, 2011 5:58 am
Location: Riga, Latvia

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby pciccone » Fri Nov 13, 2015 1:40 pm

Good to know. Thank you for sharing the results.
User avatar
pciccone
Developer
 
Posts: 1579
Joined: Wed Jan 13, 2010 11:02 am
Location: South Carolina

Re: LuxMark v3.1: LuxRender propaganda with OpenCL

Postby Dade » Thu Nov 26, 2015 11:18 am

FYI, I was renewing the hosting and domain name for LuxMark and noticed that luxmark.org domain was free. I have bought it (i.e. looks like a better domain than luxmark.info) so now the LuxMark web site is available also at http://www.luxmark.org
User avatar
Dade
Developer
 
Posts: 8394
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

PreviousNext

Return to GPU Acceleration

Who is online

Users browsing this forum: No registered users and 1 guest