I want some test data for LU or QR

[winon]

I want to find some huge test data. They should be big enough, so that I can compare my CUDA program with CULA'S LU OR QR. If you know, please tell me. Thanks!

[john]

For the most part you should be able to use random data for testing purposes if your main interest is speed (it's a different story if you're examining accuracy). QR executes in a data-independent manner, so as long as you don't have any under/overflow then it will execute in the same amount of time whether the data is random or structured. LU's execution is data dependent, and the number of row swaps performed will (probably) increase with random data so there might be a small performance penalty but at least it will apply to both solvers.

I hope that helps,

John

[morris]

I'm finding that the R matrix (taken by copying the upper part of A) from:

culaDeviceSgeqrf((int)mrowsA, (int)ncolsA, A_d, (int)mrowsA, tau_d);

has rows that differ randomly in sign with the matlab [Q,R] = qr(X,0) call.
It could be the 3rd, 5th, 6th, 10th, etc. About half of the row exhibit this error.

I've tested my wrapper by using dgeqrf from glapack_090209 written by Volkov et. al and don't get these sign changes.

I'm using Cula Basic 1.1 (on Vista 64) that I downloaded this week (Dec. 16th).

dgetrf() is fine BTW (with the associated visible loss of precision on 64 x 64 input matrix of floats)

I'll check if culaSgeqrf() has the same issue.

[john]

Hi Morris,
Have you attempted to reconstruct the matrix and/or use the QR results? We tested against Matlab a few minutes ago to confirm and found that while there were differences, BOTH QR results were valid. Our answer matched MKL's as well.

The QR factorize isn't unique unless the diagonal values of R are all positive, which ours are (and Matlab's aren't). I think you're just observing two correct but different answers.

Regards,
John

[morris]

John,

Indeed, diag(R) from Matlab are not all positive, but absolutely they match the diag of gpu R matrix: norm(diag(R) - abs(diag(Rgpu))) is like 10e-6 (expected for a 64x64 matrix of randn floats).

I was hoping to run:

sorgqr((int)mrowsA, (int)min, (int)min, A_d, (int)mrowsA, tau_d);

which generates the orthogonal matrix Q from the compact representation returned by sgeqrf(), but sorgqr() is not part of Cala Basic.

I do believe that I'm indeed observing different factorizations. I'll have to dig a bit in our software lib for sorgqr.

Regards,

Morris