DT Performance Analyzer v0.6

gwbarn · February 3, 2026, 2:16pm

I’m going to upgrade my CPU from a Ryzen 3700x to a 9600x very soon, should I do these tests with both systems and upload results here?

Qor · February 3, 2026, 2:18pm

Yes, I’d love to!

sugarbravo · February 3, 2026, 3:25pm

DSC06065.txt (15.3 KB)
DSC07828.txt (16.1 KB)

$ clinfo -l
Platform #0: Intel(R) OpenCL Graphics
 `-- Device #0: Intel(R) Arc(TM) B580 Graphics

Qor · February 3, 2026, 5:28pm

Thank you very much! Very fast GPU, but it looks like you also have a VERY fast CPU.

Qor · February 3, 2026, 5:32pm

Thank you very much!

sugarbravo · February 3, 2026, 6:00pm

$ inxi -Cm
Memory:
  System RAM: total: 32 GiB available: 30.41 GiB used: 6.75 GiB (22.2%)
  Device-2: DIMMA2 type: DDR5 size: 16 GiB speed: 6400 MT/s
  Device-4: DIMMB2 type: DDR5 size: 16 GiB speed: 6400 MT/s
CPU:
  Info: 8-core model: AMD Ryzen 7 9700X bits: 64 type: MT MCP cache: L2: 8 MiB
  Speed (MHz): avg: 605 min/max: 605/5582

Debian Sid, Plasma desktop.

gwbarn · February 3, 2026, 7:21pm

First set of results. I’ve learned that one can use the --disable-opencl flag to easily and temporarily run Darktable without opencl (including in these testing commands!).
GPU is RTX 3070, system RAM is 16GB dual channel 3200MT/s
3700x-24mp-without-opencl.txt (9.1 KB)
3700x-61mp-without-opencl.txt (8.8 KB)
3700x-24mp-with-opencl.txt (17.5 KB)
3700x-61mp-with-opencl.txt (16.9 KB)

danny · February 4, 2026, 4:31pm

4080m:
4080m_dt_b.txt (16.7 KB)
13700H IrisXe:
iris_dt_b.txt (17.1 KB)
13700H cpu:
13700h_dt_b2.txt (11.7 KB)

Qor · February 4, 2026, 5:01pm

@danny, @gwbarn

Many thanks to both of you. I have inserted the logs!

gwbarn · February 5, 2026, 2:17pm

Upgrade is up and running. 50% time taken on CPU only seems even better than I was expecting! I’m hoping I didn’t have a RAM bottleneck with the 3700x (it didn’t seem to want to use anywhere near the available 10GB or so that I had?)
Now I have 32GB DDR5 6000MT/s.
9600x-24mp-with-opencl.txt (16.0 KB)
9600x-24mp-without-opencl.txt (9.1 KB)
9600x-61mp-with-opencl.txt (15.9 KB)
9600x-61mp-without-opencl.txt (8.8 KB)

Now I need to edit my previous post, but I don’t have that option.

Brian_Innes · February 5, 2026, 9:44pm

I’m currently in 32nd place with my Nvidia GTX 1660

Qor · February 5, 2026, 10:45pm

Please send me your log-file. thx.

europlatus · February 5, 2026, 10:59pm

RX6700XT-61.txt (17.7 KB)
RX6700XT-24.txt (16.7 KB)

Brian_Innes · February 5, 2026, 11:28pm

benchmark.txt (15.6 KB)

Qor · February 6, 2026, 12:11am

@europlatus , @Brian_Innes

Thank you both for the logs.

There is currently a small problem with masks + exposure. I think this should be fixed by the 5.4.1 update at the latest.

github.com/darktable-org/darktable

Internal tiling for OpenCL guided filter

master ← jenshannoschwalm:guided_cl_filter

offen 08:42AM - 29 Jan 26 UTC

jenshannoschwalm

+255 -169

The OpenCL guided filter (used in blending and hazeremoval) requires 21 floats/p…ixel internally thus often leads to CPU fallbacks (overall requirements are >30 floats/pixel) while processing large images in HQ darkroom or while exporting. This commit introduces internal tiling for improved performance. The overlap - thus the efficiacy of internal tiling - depends on the guided filter radius, still performance is much better. We use the device specific advantage hint if provided. With tiled OpenCL code we relax calculation of blending tiling requirements. - some formatting - updated kernels to access a specific part of the guide ________________________________________________________________________________________________________ BTW this work was inspired by the first proposed "benchmark" in https://discuss.pixls.us/t/dt-performance-analyzer-v0-6/55563/54

Macchiato17 · February 8, 2026, 9:45pm

I just updated to master [23b9e9af] and found, that the performance test for the larger file DSC07828 took significantly longer. Processing time raised from 19,8s (from a test I did with darktable 5.5.0+150~gcd56293fa7) to 27s (new test with darktable 5.5.0+259~g23b9e9af19).

Any ideas or anyone who experiences something similar?

Qor · February 8, 2026, 9:55pm

5.5.0+247~gef00fb611a.txt (15,8 KB)

I can confirm that. - RustiCL usually runs faster for me than ROCm (due to a bug in ROCm).

Macchiato17 · February 8, 2026, 10:14pm

Thanks for testing in your side, Chris.

Asking @hannoschwalm => I’m not sure, if this performance drop arose from latest changes in the OpenCL code or something different may have caused that. Should I track an issue on github for that?

Jens-Hanno_Schwalm · February 8, 2026, 10:35pm

i would need verbose opencl & pipe logs from 5.4.1 and master on the same system.

Qor · February 9, 2026, 12:18am

I checked everything again, and the power profile did not load correctly for me under Linux. So there is no deterioration for me.

Please sende me your Log, I have an idea what the problem might be.

@Jens-Hanno_Schwalm

ROCm has become significantly faster on 8060S iGPU. Very nice! If you need the individual logs: At the very end of the “Detailed Results” (click to expand).