Cannot run same script twice.

[inducer]

Cannot run same script twice.

[yuyichao]

Can you try a lower version of beignet? e.g. 0.9

[decabyte]

Yes, my fault. I didn't installed the right beignet version, with 0.8.1 no problems with the examples. First time without a warming Nvidia under PyOpenCL. :)

It is worth noticing that the version shipped with Ubuntu 14.04 is very old, 0.3.1-1.

Any recommendation for using the right version on different machines? Better to compile beignet from source checking out a specific version everytime or relying on other packages? Maybe the Debian / Ubuntu+1 ones?

$ python benchmark.py 
Execution time of test without OpenCL:  0.0580661296844 s
===============================================================
Platform name: Experiment Intel Gen OCL Driver
Platform profile: FULL_PROFILE
Platform vendor: Intel
Platform version: OpenCL 1.1 beignet 0.8.0
---------------------------------------------------------------
Device name: Intel(R) HD Graphics IvyBridge M GT2
Device type: GPU
Device memory:  128 MB
Device max clock speed: 1000 MHz
Device compute units: 128
Device max work group size: 1024
Device max work item sizes: [512, 512, 512]
Data points: 8388608
Workers: 256
Preferred work group size multiple: 16
Execution time of test: 0.191544 s
Results OK

[yuyichao]

My local beignet version is 3 commit ahead of 0.9.1 and it runs without problems. If you found a commit that breaks pyopencl I think you should report to beignet with the bad commit you are on.

[yuyichao]

Also, I'm supprised that beignet <= 0.9.1 works at all from a binary package since it is only recently that beignet can work on a different architecture it is compiled on. (i.e. if you want to run it on IveBridge CPU, beignet had to be compiled on that CPU as well).
I would recommand using the latest git master or at least >= 0.9.2 since it is the first release which use llvm bytecode at compile time.

[decabyte]

Thanks @yuyichao I'll compile the latest git master asap. Given the OS release I'll try the LLVM/clang 3.4 stack as it is shipped by default on 14.04.

What about the git mesa local repo? It is a required dependency for using pyopencl and beignet successfully?

[yuyichao]

There's a recent thread on a compiling problem on the beignet list and the reply seems to indicate that 10.1 should be fine. I'm not sure what's the officially supported versions but at least git master is not necessary. I have just compiled the latest beignet master with mesa 10.2.5 and it looks OK. I guess you should just test it with whatever mesa version you have and report to beignet list if it fails to figure out whether it is possible to support the version you have (I guess any recent versions should be fine).

[decabyte]

Yes, the 0.9.2+ seems to works much better. I've compiled with the latest Intel Graphics stack for Ubuntu 14.04 which brings Mesa 10.2.2. Compilation is fine and the only missing part is the cl_khr_gl_sharing cause I didn't rebuild Mesa.

$ python benchmark.py 
Execution time of test without OpenCL:  0.0574040412903 s
===============================================================
Platform name: Intel Gen OCL Driver
Platform profile: FULL_PROFILE
Platform vendor: Intel
Platform version: OpenCL 1.2 beignet 0.9
---------------------------------------------------------------
Device name: Intel(R) HD Graphics IvyBridge M GT2
Device type: GPU
Device memory:  1024 MB
Device max clock speed: 1000 MHz
Device compute units: 16
Device max work group size: 1024
Device max work item sizes: [1024, 1024, 1024]
Data points: 8388608
Workers: 256
Preferred work group size multiple: 16
Execution time of test: 0.0121039 s
Results OK