After this patch everything works:
--- demo_fft.py 2010-09-28 22:55:55.000000000 +0200
+++ ../demo_fft.py 2010-09-28 22:55:31.000000000 +0200
@@ -74,7 +74,6 @@
#add results to variable
timesCPU.append(secs/count)
- del plan
#print the data out
tbl = Table()It looks like garbage collector was not cleaning up "plan" objects and they clogged up device. Removing objects "by hand" solved problem. Here are my results: Power of 2|Size |Time GPU |Time CPU |GPU vs CPU speedup ----------+-------+-----------------+-----------------+------------------ 21 |2097152|0.00115163841248 |0.56857734375 | 493.711687272 20 |1048576|0.000566671991348|0.248464501953 | 438.46264814 19 |524288 |0.000293673610687|0.126432165527 | 430.51932801 18 |262144 |0.000173881602287|0.0582307922363 | 334.887598632 17 |131072 |9.17439997196e-05|0.0268174407959 | 292.307299419 16 |65536 |7.21791982651e-05|0.0116292510986 | 161.116379485 15 |32768 |5.14944016933e-05|0.00522409934998 | 101.449850434 14 |16384 |4.94143992662e-05|0.00230527687073 | 46.65192545 13 |8192 |4.74175989628e-05|0.00108720321655 | 22.9282637741 12 |4096 |3.08640003204e-05|0.000258844804764| 8.38662526168 11 |2048 |2.81504005194e-05|0.000108675205708| 3.86052076356 10 |1024 |2.38572788239e-05|4.93587207794e-05| 2.06891662473 9 |512 |2.39504003525e-05|2.50713610649e-05| 1.04680342274 8 |256 |2.37785601616e-05|1.36412799358e-05| 0.573679812534 7 |128 |2.59728002548e-05|9.44159984589e-06| 0.363518748585 6 |64 |2.3837120533e-05 |7.46496021748e-06| 0.313165350955 5 |32 |2.3746240139e-05 |6.40191972256e-06| 0.269597194549 4 |16 |2.36556792259e-05|6.12096011639e-06| 0.258752245409 3 |8 |2.37475204468e-05|5.9017598629e-06 | 0.248521098282 2 |4 |2.39055991173e-05|5.52640020847e-06|0.231175976028 -- Tomasz Rybak <bogom...@post.pl> GPG/PGP key ID: 2AD5 9860 Fingerprint A481 824E 7DD3 9C0E C40A 488E C654 FB33 2AD5 9860 http://member.acm.org/~tomaszrybak
signature.asc
Description: This is a digitally signed message part
_______________________________________________ PyCUDA mailing list PyCUDA@tiker.net http://lists.tiker.net/listinfo/pycuda
