Mentions légales du service

Skip to content

Fix overflow in flop computation in testings

Philippe SWARTVAGHER requested to merge pswartva/chameleon:fix-flops-zlacpy into master

With big matrices, testing_zlacpy reports negative Gflop/s:

% ./testing/chameleon_stesting -o lacpy --n 3200:60000:6400 -H
 Id Function     threads gpus  P  Q mtxfmt  nb uplo        m     n   lda   ldb       seedA          tsub          time        gflops
  0 slacpy            36    0  1  1      0 320 Upper    3200  3200  3200  3200   846930886  0.000000e+00  1.656305e-03  1.236874e+01
  1 slacpy            36    0  1  1      0 320 Upper    9600  9600  9600  9600  1681692777  0.000000e+00  2.398928e-02  7.684232e+00
  2 slacpy            36    0  1  1      0 320 Upper   16000 16000 16000 16000  1714636915  0.000000e+00  5.037857e-02  1.016369e+01
  3 slacpy            36    0  1  1      0 320 Upper   22400 22400 22400 22400  1957747793  0.000000e+00  1.102050e-01  9.106347e+00
  4 slacpy            36    0  1  1      0 320 Upper   28800 28800 28800 28800   424238335  0.000000e+00  2.285272e-01  7.259258e+00
  5 slacpy            36    0  1  1      0 320 Upper   35200 35200 35200 35200   719885386  0.000000e+00  2.889825e-01  8.575434e+00
  6 slacpy            36    0  1  1      0 320 Upper   41600 41600 41600 41600  1649760492  0.000000e+00  4.286759e-01  8.074172e+00
  7 slacpy            36    0  1  1      0 320 Upper   48000 48000 48000 48000   596516649  0.000000e+00  5.470726e-01 -7.278446e+00
  8 slacpy            36    0  1  1      0 320 Upper   54400 54400 54400 54400  1189641421  0.000000e+00  6.930918e-01 -3.853899e+00

The problem comes from an overflow in the N*(N+1) multiplication here.

I fixed this with the following MR. Is it the best way to solve this, so should I apply the change to all other flops_* functions ?

Edited by Philippe SWARTVAGHER

Merge request reports