Am Mon, 23 Dec 2019 11:22:57 +0100
schrieb Ulrich Pegelow <[email protected]>:

> Am 23.12.19 um 10:22 schrieb Аl Воgnеr:
> > I tried "opencl_memory_headroom=100"
> > 
> > 10,485934 [dev_process_export] pixel pipeline processing took 3,299
> > secs (8,980 CPU)
> > 
> > So there is no big difference.
> 
> Assuming that the history you provided is a typical one

No, the files from https://math.dartmouth.edu/~sarunas/bench_raw/ are
files to compare with others. I use only a few modules, bench.srw uses a
lot more modules than I normally use.

Hard to say, if I found a typical fullformat excample, but with my own
fullformat example I see:

7,393864 [dev_pixelpipe] took 0,116 secs (0,300 CPU) processed `Gamma'
on CPU, blended on CPU [export]

All others are:
"blended on GPU [export]"

7,393968 [dev_process_export] pixel pipeline processing took 0,608 secs
(0,915 CPU)

Complete log is attached.

So my ff-36MP-photo is a lot faster calculated.

My goal is to avoid to copy stupid settings from the old pc.

Interesting is:
Spitzlicht-Rekonstruktion' on GPU, blended on GPU [export]

With the bench.srw it was calculated by the cpu. So the use of the cpu
depends on something and not on the coding of the module itself.

So my question, if some of these values should be changed, because the
new system is a lot more powerful:

cache_memory=1073741824
maximum_number_tiles=10000
metadata/resolution=300
opencl_checksum=3732205163
opencl_device_priority=*/!0,*/*/*
opencl_mandatory_timeout=200
opencl_memory_requirement=768
opencl_micro_nap=1000
opencl_number_event_handles=25
opencl_size_roundup=16
pixelpipe_synchronization_timeout=200
plugins/map/max_images_drawn=100
plugins/map/max_outline_nodes=10000


> then there is 
> only one direction where OpenCL memory optimization could help: if
> you can avoid the contrast equalizer module to use tiling.
> 
> But there is not much to gain. With tiling that module needs 0.43
> secs. The pure GPU processing time is 0.3 secs (eaw_decompose + 
> eaw_synthesize). Without tiling there is less overhead and less
> overlap so you might win in total about 0.2 secs. Judge by yourself
> if this is worth the effort.

1 sec difference does not matter a lot, 10 sec do, especially I don't
want to use a stupid configuration.

Al

____________________________________________________________________________
darktable user mailing list
to unsubscribe send a mail to [email protected]

Attachment: ff.darktable.log.tar.gz
Description: application/gzip

Reply via email to