OSX Multibeam OpenCL question

[Joe Fox]

Oh, and as far as execution time with the reference result set. I will go back through and build CPU only, then GPU and report back here with the processing times.

-Joe

[Urs Echternacht]

Your better of to check which hardware the new Mac Pro workstations contain :
https://www.apple.com/mac-pro/

So far we have build separate apps, one per device vendor.

[Joe Fox]

Ok, so boinc has the ability to distribute the correct client for not just the OS platform, but other parameters like gpu vendor? That's cool

-Joe

[Joe Fox]

oh, and looks like the mac pro has dual AMD D300s or D500s, not nvidia. My mistake.

[Joe Fox]

Looks like there may be another problem. My binary runs successfully, but doesn't produce valid results. Attached is the results.sah and stderr.txt from my latest run. For what it is worth this ran in 14 minutes and 10 seconds.

[Raistmer]

Ah, I examined the svn diff, and looks like the code is now wrapped with an ifdef USE_OPENCL_INTEL. When I set that macro in the CFLAGS, then my binary works again. Thank you!

That leads me to a question. What should the USE_OPENCL_??? macros be set at for a wide distribution build, like a beta build? For my local machine, which has an Intel GPU, the USE_OPENCL_INTEL makes sense, but what about for the new Mac Pro desktop workstations? They have an nvidia GPU if I recall correctly. How would a widely distributed binary that was compiled with USE_OPENCL_INTEL work for the nvidia system?

Thanks for the assistance!
-Joe

Currently we use 3 separate binaries  for different GPU types. There are some pecularities for each GPU vendor so I consider easier to maintain 3 separate optimized builds instead of universal under-optimized one.

[Raistmer]

Looks like there may be another problem. My binary runs successfully, but doesn't produce valid results. Attached is the results.sah and stderr.txt from my latest run. For what it is worth this ran in 14 minutes and 10 seconds.

Glancing on your first log it seems you missed correct set of defines.
I'll post what I use for Windows here, maybe Urs will correct it for OS X specific.

[Joe Fox]

Regarding maintaining the different binaries, that makes sense. Regarding the incorrect defines, thanks. I'll go back to the OSX list, as well as compare to the windows build list that you send.

Thanks
-Joe

[Urs Echternacht]

Looks like there may be another problem. My binary runs successfully, but doesn't produce valid results. Attached is the results.sah and stderr.txt from my latest run. For what it is worth this ran in 14 minutes and 10 seconds.

The "reference_work_unit.sah" in the "test_workunits" directory seems to be an old v6 workunit without autocorr-signal detection. Instead, try to use "refquick.wu" or "reftiny.wu" from "DynWUs" directory for quick testing.
Additionally i've tried to adapt the readme (see attached) you have been referring your work to for Intel GPU builds. You may want to compare the "-D" commandline switches (you used the term "macro" before) , also called "preprocessor directives", from your build to what is in there. Especially "-DSETI7" is important for good results.

[Joe Fox]

Thank you Urs, I will go through these and report back.

[Joe Fox]

The "reference_work_unit.sah" in the "test_workunits" directory seems to be an old v6 workunit without autocorr-signal detection. Instead, try to use "refquick.wu" or "reftiny.wu" from "DynWUs" directory for quick testing.

I don't see those files inside of DynWUs. I only have the following:
PG0009.wu
PG0395.wu
PG0444.wu
PG1327.wu
init_data.xml
PG0009_v7.wu
PG0395_v7.wu
PG0444_v7.wu
PG1327_v7.wu

Any advice?

-Joe

[Joe Fox]

Here is what I found with using one of the DynWUs.

I used "PG1327_v7.wu"

and got the attached results.

Elapsed time was 6 minutes and seconds 37

How can I test the results for accuracy?

Thanks
-Joe

[Raistmer]

list of defines I used in windows build:
OCL_CHIRP3;USE_OPENCL;ATI_OS_WIN;SETI7;USE_FFTW;WIN32;_WIN32;_MT;NDEBUG;_WINDOWS;CLIENT;_CONSOLE;USE_I386_OPTIMIZATIONS;USE_I386_XEON;USE_SSE;USE_AMD

OCL_CHIRP3 is required  for best precision, USE_FFTW - for CPU FFT when needed, SETI7 as Urs already described required too.

[Urs Echternacht]

In lunatics Downloads you will find some Tools, like more testworkunits and a benchmark-script for "Terminal" usage. That benchmark-package contains another tool for result comparison (rescmp4_d, d for Darwin) That comparison tool you could use also standalone if you have two results to compare (e.g. result_a.sah, result_b.sah).
If you start the comparison tool without arguments you get some howto use.

[Joe Fox]

One interesting thing that I have found in my testing of the new builds. Only the build with USE_OPENCL_INTEL works correctly, regardless of the video card manufacturer used for the mac. I have tested on a new Mac Pro with dual  AMD D700s, a MacBook Pro with an Intel Iris, and an older MacBook Pro with an nvidia. The USE_OPENCL_HD5xxx build would fail to compike the .cl files, and the USE_OPENCL_NV build would fail with the original error I ran in to (CL error -49). My conclusion is that the macs are using intel opencl drivers. Does this seem reasonable?