Accuracy problem in solution of linear equations Help Required

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hey all,

I am solving linear equations by Jacobi’s iterative method.

My code goes here:-

The code has been deleted

Problem:-

When i am running it on emulatio mode, i get correct answers,
but in vain when on nvidia’s gtx 8800.
The answers are slightly deviated for less variables(2,3,4),
But for 50,60 variables the answers go wrong!!

Any suggestions!!

#2

This usally means that you have a race condition. Since emulation run is syncornized you won’t see it there.

The problem might be here:

if(ty!=tx)

				temp[n+ty]-=(temp[tx]*a[ty*n+tx]);

You update the same temp[XX] for all tx values for a certain ty thread. Thus all tx threads (for example for ty==0 or ty == 1,…)

update the same shared memory element overriding everything…

eyal

#3

Thanks i’ll definietly try.

Here i want that it is not updated if ty==tx, can you please give the modofied code!

Your reply would be highly appreciated!