Using darktable master with Intel OpenCL graphics? Please help!

hannoschwalm · February 15, 2023, 7:13pm

Friends, there is a difficult issue on github HLR green artifacts with OpenCL · Issue #13632 · darktable-org/darktable · GitHub

In short: Using the highlights reconstruction inpaint opposed algorithm leads to severe color artefacts at least for one user. This issue couldn’t yet be reproduced on other (Nvidia, AMD) hardware. So we don’t know if this is

something specific in the installation of the specific user
some specific driver issue
specific for the hardware

If you are running darktable OpenCL on an Intel device could you check this?

Please remember you need an image with highlights blown out Like in clouds
Please report if issue or not
Please report OS and Intel driver version (darktable -d opencl will tell you) for both cases in (2)
The

g-man · February 15, 2023, 7:47pm

Could it be beneficial to pick one of the PlayRaw images so we can use the -d pipe and compare actual values?

I think @sugarbravo mentioned that he has an Intel GPU.

hannoschwalm · February 15, 2023, 8:07pm

Perfect

ayro · February 15, 2023, 11:58pm

I’m the user mentioned in the issue. I’m happy to upload the duck pond image + xmp from the pull request for people to test with.

This file is licensed Creative Commons, By-Attribution, Share-Alike.
20190822_003649.nef (28.8 MB)
20190822_003649.nef.xmp (1.2 KB)

IvanB · February 16, 2023, 3:20pm

Hi, no issues in my case.

Specs:

Darktable 4.3.0+640~g22c9eaf9d

Distributor ID:	Linuxmint
Description:	Linux Mint 21.1
Release:	21.1
Codename:	vera

[dt_opencl_device_init]
   DEVICE:                   0: 'Intel(R) UHD Graphics 620 [0x5917]'
   CANONICAL NAME:           intelruhdgraphics6200x5917
   PLATFORM NAME & VENDOR:   Intel(R) OpenCL HD Graphics, Intel(R) Corporation
   DRIVER VERSION:           1.0.0
   DEVICE VERSION:           OpenCL 3.0 NEO 
   DEVICE_TYPE:              GPU
   GLOBAL MEM SIZE:          6137 MB
   MAX MEM ALLOC:            3069 MB
   MAX IMAGE SIZE:           16384 x 16384
   MAX WORK GROUP SIZE:      256
   MAX WORK ITEM DIMENSIONS: 3
   MAX WORK ITEM SIZES:      [ 256 256 256 ]
   ASYNC PIXELPIPE:          NO
   PINNED MEMORY TRANSFER:   WANTED
   MEMORY TUNING:            NO
   FORCED HEADROOM:          400
   AVOID ATOMICS:            NO
   MICRO NAP:                250
   ROUNDUP WIDTH:            16
   ROUNDUP HEIGHT:           16
   CHECK EVENT HANDLES:      128
   PERFORMANCE:              1.640
   TILING ADVANTAGE:         0.000
   DEFAULT DEVICE:           NO
   KERNEL DIRECTORY:         /opt/darktable-test/share/darktable/kernels
   CL COMPILER OPTION:       -cl-fast-relaxed-math

g-man · February 16, 2023, 3:34pm

IvanB, can you post the -d pipe results in the GitHub PR for the inpaint?

github.com/darktable-org/darktable

Improve precision of HLR chroma corrections

darktable-org:master ← jenshannoschwalm:opposed_precision

opened 06:02AM - 15 Feb 23 UTC

jenshannoschwalm

+58 -37

There has been some difference between OpenCL / CPU code for the calculated chro…ma correction coeffs due to different floating point precision. In this pr we calculate the coeffs via `double` leading to almost identical results. Downsides: 1. requires `double` support for the device (all drivers should support that) 2. some performance penalty. That is in fact **very** small especially in darkroom as we have cached data now. Not sure if this fixes problems for certain Intel Neo drivers though. See #13632 @groutr could you test this on your intel device? @MStraeten could you please test on AMD and OSX?

hannoschwalm · February 16, 2023, 3:40pm

Just closed this pr. Will do some more homework (docs reading …) and come back also here. Thanks for your patience

sugarbravo · February 16, 2023, 5:08pm

I can see no artifacts while using OpenCL with Intel Arc A750 GPU. Kubuntu 22.10, i915 kernel module from 6.2.0-060200rc5-generic.

$ /opt/darktable/bin/darktable --version
this is darktable 4.3.0+647~g015c6fb01
copyright (c) 2009-2023 johannes hanika
https://github.com/darktable-org/darktable/issues/new/choose

compile options:
  bit depth is 64 bit
  normal build
  SSE2 optimized codepath enabled
  OpenMP support enabled
  OpenCL support enabled
  Lua support disabled
  Colord support enabled
  gPhoto2 support disabled
  GraphicsMagick support enabled
  ImageMagick support disabled
  libavif support disabled
  libheif support enabled
  libjxl support disabled
  OpenJPEG support enabled
  OpenEXR support disabled
  WebP support enabled

$ /opt/darktable/bin/darktable-cltest
     0.0172 [dt_get_sysresource_level] switched to 2 as `large'
     0.0172   total mem:       32003MB
     0.0172   mipmap cache:    4000MB
     0.0172   available mem:   21877MB
     0.0172   singlebuff:      500MB
     0.0172   OpenCL tune mem: WANTED
     0.0172   OpenCL pinned:   OFF
[opencl_init] opencl related configuration options:
[opencl_init] opencl: ON
[opencl_init] opencl_scheduling_profile: 'very fast GPU'
[opencl_init] opencl_library: 'default path'
[opencl_init] opencl_device_priority: '*/!0,*/*/*/!0,*'
[opencl_init] opencl_mandatory_timeout: 400
[opencl_init] opencl library 'libOpenCL' found on your system and loaded
[opencl_init] found 2 platforms
[opencl_init] no devices found for Advanced Micro Devices, Inc. (vendor) - AMD Accelerated Parallel Processing (name)
[opencl_init] found 1 device

[dt_opencl_device_init]
   DEVICE:                   0: 'Intel(R) Graphics [0x56a1]'
   CANONICAL NAME:           intelrgraphics0x56a1
   PLATFORM NAME & VENDOR:   Intel(R) OpenCL HD Graphics, Intel(R) Corporation
   DRIVER VERSION:           1.0.0
   DEVICE VERSION:           OpenCL 3.0 NEO 
   DEVICE_TYPE:              GPU
   GLOBAL MEM SIZE:          7722 MB
   MAX MEM ALLOC:            3861 MB
   MAX IMAGE SIZE:           16384 x 16384
   MAX WORK GROUP SIZE:      1024
   MAX WORK ITEM DIMENSIONS: 3
   MAX WORK ITEM SIZES:      [ 1024 1024 1024 ]
   ASYNC PIXELPIPE:          NO
   PINNED MEMORY TRANSFER:   NO
   MEMORY TUNING:            WANTED
   FORCED HEADROOM:          400
   AVOID ATOMICS:            NO
   MICRO NAP:                250
   ROUNDUP WIDTH:            16
   ROUNDUP HEIGHT:           16
   CHECK EVENT HANDLES:      128
   PERFORMANCE:              5.013
   TILING ADVANTAGE:         0.000
   DEFAULT DEVICE:           NO
   KERNEL DIRECTORY:         /opt/darktable/share/darktable/kernels
   CL COMPILER OPTION:       -cl-fast-relaxed-math
   KERNEL LOADING TIME:       0.0186 sec
[opencl_init] OpenCL successfully initialized.
[opencl_init] here are the internal numbers and names of OpenCL devices available to darktable:
[opencl_init]           0       'Intel(R) Graphics [0x56a1]'
[opencl_init] FINALLY: opencl is AVAILABLE on this system.
[opencl_init] initial status of opencl enabled flag is ON.
[dt_opencl_update_priorities] these are your device priorities:
[dt_opencl_update_priorities]           image   preview export  thumbs  preview2
[dt_opencl_update_priorities]           0       0       0       0       0
[dt_opencl_update_priorities] show if opencl use is mandatory for a given pixelpipe:
[dt_opencl_update_priorities]           image   preview export  thumbs  preview2
[dt_opencl_update_priorities]           1       1       1       1       1
[opencl_synchronization_timeout] synchronization timeout set to 0
[dt_opencl_update_priorities] these are your device priorities:
[dt_opencl_update_priorities]           image   preview export  thumbs  preview2
[dt_opencl_update_priorities]           0       0       0       0       0
[dt_opencl_update_priorities] show if opencl use is mandatory for a given pixelpipe:
[dt_opencl_update_priorities]           image   preview export  thumbs  preview2
[dt_opencl_update_priorities]           1       1       1       1       1
[opencl_synchronization_timeout] synchronization timeout set to 0

LebedevRI · February 16, 2023, 5:57pm

Is that a cache invalidation issue? Does removing the kernel cache (~/.cache/darktable/cached_kernels_for_*?) hide it?

hannoschwalm · February 16, 2023, 7:01pm

I think i have identified some problems, could you check this?

github.com/darktable-org/darktable

Bugfix: make sure correct HLR chroma corrections

darktable-org:master ← jenshannoschwalm:fix_opposed_opencl2

opened 06:53PM - 16 Feb 23 UTC

jenshannoschwalm

+98 -98

See #13632 and #13646 In some situations the calculated chroma corrections ar…e vastly wrong. The first idea was about precision errors because of using floats for summing up data and the counters, that could be ruled out. The reported errors could be related (this is my current understanding) to: 1. improper data sampling because of the relaxed global memory policy, this could still be the case although unlikely because the errors keep happening after correcting data to be always in a gpu cacheline. (It might be we have to intruce a `lock-per-line`) 2. Other issues on AMD often pinpointed to a) float data not being initialized or b) math errors like div by zero. (Nvidia never had problems on this) The code has been reviewed for strictly avoiding this. 3. The for_each_channel usage was wrong in many parts of the code as we don't calc 4-channel pixels but we do a loop over a "color-range" of 0-2. This could have resulted to a div-by-zero. 4. The dark-level was a code left-over, it's safe to use 0.2 of clip in all cases. (This is the only maths change in this pr. The dark clips lead to problems in very few images if there were a lot of black-to-white transitions) 5. It is very good to avoid chroma corrections derived from a very small number of border pixels, we make sure now to have at least a "sensible" number. Thanks to many people having tested #13646, could i ask you to check again @groutr @MStraeten @kofa73 @gi-man @sarunasb ? To test this: 1. Use the "duck-at-the-pond" image from #13632 2. Start dt either with `--disable-opencl` or not 3. use `-d pipe` and watch out for reports like `[opposed chroma cache GPU/CPU]` 4. Compare the values reported, they must be **"almost identical"** for CPU vs GPU, any difference must be concidered to be a bug. 5. If you made sure: rawprepare, temperature and clip in hlr are all set to default, and using opposed method your results should be ``` 13.2351 [opposed chroma cache GPU] red: 1.056767 (18406), green: -0.025674 (123257), blue: -0.394672 (16525) for opphash 11498776663581970802 3.7372 [opposed chroma cache CPU] red: 1.056766 (18406), green: -0.025674 (123257), blue: -0.394673 (16525) for opphash 11498776663581970802 ``` A sidenote, the algo has also changed slightly, the results might be slightly different from master but tested on this pr the reports must be identical for GPU and CPU