we have an issue with simple_matmult program from the samples package which we cannot explain up to now:
Enabling verification, it turns out, that the results of the matrix-matrix-multiplication by CPU and GPU do not match for certain matrix sizes.
For example, 64x64 and 128x128 always returns a PASSED, while a y-size of 65 or 129 returns a FAILED. This issue does not only depend on one of the sizes: While any x size with y size of 128 returns a PASSED, for y size of 1024 this is not true.
Have you any clue, from what this could arise?
Thanks for hints and answers,
we're currently using SDK v1.2.1-beta with a Radeon 3870 Firestream 9170 card on a SuSE Linux station.
Is the version 1.3 already available? Where can I download it?
Thanks, Micah. I will wait and try it with SDK 1.3.
However, I wonder if we are the only one recognized that bug(?) in SDK 1.2.1-beta. Is there anyone out there who can confirm, that simple_matmult is computing incorrect results with SDK 1.2.1-beta? I just want to clarify that it is caused by the SDK and not maybe by a failure of the graphics card or whatsoever.
rahulgarg, thanks for helping.
Following call does the job perfectly and returns a PASSED:
./simple_matmult -e -x 64 -y 64
However, following call returns a FAILED:
./simple_matmult -e -x 65 -y 65
I'm very interested in your results whether you can confirm or not.
Linux x86-64 (CentOS 5), amdstream-brook-1.2.1_beta-1, Radeon 4870
./simple_matmult -e -x 64 -y 64 : PASSED
./simple_matmult -e -x 64 -y 65 : PASSED
./simple_matmult -e -x 64 -y 129 : PASSED
./simple_matmult -e -x 65 -y 64 : PASSED
./simple_matmult -e -x 129 -y 64 : PASSED
./simple_matmult -e -x 65 -y 129 : PASSED
thank you for your answer! rahulgarg really confused me with his "x,y,h,w" parameters. Are there different versions of SDK 1.2.1-beta? In our version, the -h parameter displays the help screen, -w is not a valid parameter. The parameters are parsed by a function from the common files of the SDK, so I really don't understand where these differences come from.
Since 65x65 is working for Ceq, we have to seek for the cause elsewhere.
Can it be a compiler problem?
Thanks for further help,
The options may have changed. I do not recall what they were before, but for the current release (1.2.1), -h is height of the computation domain, -w is width of the domain, and -? is "help" and shows you all the options.