search

guagua-pamcn / a-dda

Automatically exported from code.google.com/p/a-dda
0 stars 0 forks source link

Issue in OpenCL implementation #207

Open GoogleCodeExporter opened 8 years ago

GoogleCodeExporter commented 8 years ago 
Hi, 

there seems to be an issue in the current OpenCL implementation from the SVN:

I use svn r1406 with OCL_BLAS enabled on an AMD GPU. For example with
adda_ocl -orient 49 43 25 -shape ellipsoid 1.4 2.2 -eq_rad 10 -iter bicg
the calculation for the second incident polarization fails. The relative norm 
of the residual becomes extremely large and no convergence is achieved. When I 
add "-init_field inc" to the command line, I get convergence.

However, the results from the OpenCL version in the SVN substantially deviate 
from the results obtained using other versions of ADDA. I have attached a table 
with some numbers for C_ext in various cases. Results for r1406/seq, 1.3b4/ocl, 
1.3b4/seq always agree, but r1406/ocl always deviates from this. The last block 
(where the qmr solver is used) suggests that the issue probably is in the 
current opencl implementation, not directly in the new OpenCL-bicg 
implementation.

This issue does not depend much on the geometry (I tested some other 
geometries).

Please let me know if you need more information on this.

Original issue reported on code.google.com by jgasteiger80 on 8 May 2015 at 8:59

Attachments:

GoogleCodeExporter commented 8 years ago

Original comment by yurkin on 8 May 2015 at 5:14

GoogleCodeExporter commented 8 years ago 
While trying to reproduce the issue using an Nvidia GTX 970, I figured out that 
the manual work group size change (from r1351 on) breaks on the execution of 
such grid size combinations. With removing the manual work group size settings, 
execution worked and gave the same result with any solver as the seq version. 
Please try if r1407 solves the issue also for you.

Original comment by Marcus.H...@gmail.com on 10 May 2015 at 5:46

GoogleCodeExporter commented 8 years ago 
Unfortunately your fix does not work in my configuration: Here on Ubuntu with 
the AMD W8100 GPU, r1407 has the same issues as described above for r1406.

Original comment by jgasteiger80 on 12 May 2015 at 8:04