CUDA Error, too many resources requested for launch

[Nannew]

I think there are something needs to be fixed on working set size. I use a Window 10 machine with a NVIDIA Quadro 4000 and CUDA 8.0.

When I tried to train a set with more features, I got following errors. The data set "a9a"used here is from LibSVM site:https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html

PS O:\thundersvm\thundersvm\build\bin\Release> ./thundersvm-train -g 0.125 -c 10 a9a.txt svm.model 2018-05-09 14:06:59,499 INFO [default] loading dataset from file "a9a.txt" 2018-05-09 14:07:00,168 INFO [default] #instances = 32561, #features = 123 2018-05-09 14:07:00,171 INFO [default] training C-SVC 2018-05-09 14:07:00,174 INFO [default] C = 10 2018-05-09 14:07:00,176 INFO [default] gamma = 0.125 2018-05-09 14:07:00,228 INFO [default] #classes = 2 2018-05-09 14:07:00,347 INFO [default] working set size = 1024 2018-05-09 14:07:00,374 INFO [default] training start 2018-05-09 14:07:00,773 FATAL [default] Check failed: [error == cudaSuccess] too many resources requested for launch 2018-05-09 14:07:00,778 WARNING [default] Aborting application. Reason: Fatal log at [O:/thundersvm/thundersvm/src/thundersvm/kernel/smo_kernel.cu:309]

[zeyiwen]

Would you use the -m option, say "-m 1500" (i.e., using ~1.5GB memory as your GPU has only 2GB memory), to set the memory usage for ThunderSVM?

We will investigate and improve the code to address this issue. Thanks!

[Nannew]

same error with "-m 1500" or "-m 1000"

[zeyiwen]

I've added a new branch called dev-old-gpu. Please checkout the code and see if it solves your issue.

We can't reproduce your issue, as we don't have Quadro 4000 GPUs which are relatively old...