d14b0ll0s wrote:thanks for checking; so with CPU multi-thread support only (w/o GPU) this would be somewhere in 25-30% gain range, according to the numbers from your previous post;
Since you mentioned "Open CL unchecked = 3min 05sec" and "Open CL checked and selected GTX465 = 1min 38sec" only for stage 3 (almost double the time w/o GPU), and then "stage 3,4 & 6 = 1 min 56 sec" w/GPU, I assumed this would be almost double as well just w/CPU = closer to 4 mins now vs around 5 in ST911. Perhaps I misunderstood something.
Despite that, I'll try to do some tests myself and post results soon, now there's too much going on, so I'll start when the Mini arrives (any hints on compiling STE under OS X are will be very welcome).
What is important at this point, however, is that for some users this update is a life changer. I'm very much looking forward to deskewing improvements, but what has been done is itself quite remarkable.
Building it on OSX won't be easy, as you would be the 1st person ever to try that. There were some significant changes under the hood, most notably parts of ST being moved to shared libraries.
Scan Tailor experimental doesn't output 96 DPI images. It's just what your software shows when DPI information is missing. Usually what you get is input DPI times the resolution enhancement factor.
That's good to know, Tulon. I guess I won't spend too much time trying to do that if it fails. I'll see if Mac Ports can be of any help.
Has anyone ever compared the performance of ST911 running natively under OS X with running it in Parallels? I've heard it's usually around 20% loss with various programmes.
Scan Tailor experimental doesn't output 96 DPI images. It's just what your software shows when DPI information is missing. Usually what you get is input DPI times the resolution enhancement factor.
I'm running it in Parallels on OS X Yosemite now with 4GB RAM in VM and all 8 CPU threads on. The task which takes 1m30s on i7-3537U in STE running natively in Win10, on i7-3615QM in VM takes exactly 1min. Not the 100% speed bump I was expecting when moving from 2 cores to 4, but still a significant improvement. OpenCL unfortunately isn't supported in Parallels.
I'll try compiling from sources soon.
Now the weekend is here i decided once and for all to test out speed performance of Tulon's new STE verses his old ST 9.11.1. Additionally i wanted to test the speed performance differences selecting either/or OpenGL and Open CL (with GPU and CPU alternatively). My system is Windows 7-64bit / 16GB ram / i7-3770 CPU / Nvidia GTX465 GPU. I used a 241 page scanned book, all 24bit colour scans at 300 dpi
1st table is only a straight output benchmark. Here you will see that STE is 526% faster than ST911
2nd table i chose to test the curved line feature. Unfortunately when testing ST911 with auto dewarping it crashed out out of memory. When inclined and have gathered some patience i may try again with another sets of scans.
The results are interesting and at the same time baffling. e.g. leaving both accelerators unchecked yielded pretty good results compared to when actually having both accelerators checked.
I had assumed that having the accelerators checked no matter what hardware you had (i.e. onboard GPU or whatever slow CPU would at least have been better. But the benchmarks don't show this to be the case.
See next post for tabled results.
Last edited by xerum on 22 Aug 2015, 00:07, edited 1 time in total.
Here is the table of results followed on from previous post. What is interesting is that for OpenCL; In my case: CPU checked in OpenCL slowed me down. Each person will have to determine this for themselves.
Definitely a fast GPU could make a difference. However STE using all parallel CPU cores and both accelerators unchecked STE still slaughtered ST911. With both options checked and GPU selected for OpenCL it was a bloodbath. 386% faster and 526% respectively.
What i was attempting to highlight in previous posts a few days ago is that relatively speaking using the above results comparing dewarping in the new STE verses ST911 should be 300%-500% faster. Not the mere 20-30% as reported. In any case the new curved-lines function and keystone function are far superior in STE than ST911. For this alone it is well worth the move to STE.
STE Benchmark Results
stebench.jpg (127.28 KiB) Viewed 7314 times
Unfortunately when testing ST911.1 on the same book with auto Dewarp it kept crashing. Oh well with Tulon's new STE am not looking back. Rest in peace ST911.
ST911 out of memory crash when auto
st911oomerror.jpg (68.69 KiB) Viewed 7314 times
Last edited by xerum on 22 Aug 2015, 06:24, edited 3 times in total.
xerum wrote:if you don't have a GPU do not check and select your CPU as it will slow down STE considerably.
True for some CPUs, false for others. My Intel i7-4700HQ performs better in OpenCL mode than without.
xerum wrote:Alternatively just select your GPU no matter how fast or slow it is as it will make a beneficial difference.
While my integrated Intel 4600 GPU beats my CPU, d14b0ll0s reported his Intel 4000 GPU was losing to his CPU.
Scan Tailor experimental doesn't output 96 DPI images. It's just what your software shows when DPI information is missing. Usually what you get is input DPI times the resolution enhancement factor.
Tulon wrote:True for some CPUs, false for others. My Intel i7-4700HQ performs better in OpenCL mode than without.
Tulon wrote:While my integrated Intel 4600 GPU beats my CPU, d14b0ll0s reported his Intel 4000 GPU was losing to his CPU.
I edited my previous benchmark post to be subjective rather than objective since Tulon has different experience to mine. As there is a myriad of systems out there each person will need to do their own testing to determine the optimum STE performance settings for their system.