Preparing weights to run in nengo from cuda


#1

i am using cuda 6.5 on linux 12.04. runung theconvnet.py(on my dummy dataset ) from hunse repo getting the following error.i have nvidia GTX 750 ti 5.0 graphic card and my system ram is 4gb. why this error is appear?

1.1 (0.00%)...python: src/nvmatrix.cu:1550: virtual cudaTextureObject_t NVMatrix::getTextureObject(): Assertion `_texObj != 0' failed.
Error signal 6:
/home/assad/Desktop/finalcode/mycuda/cudaconvnet/_ConvNet.so(_Z13signalHandleri+0x20)[0x7fa2dca0c740]
/lib/x86_64-linux-gnu/libc.so.6(+0x36570)[0x7fa2ec0da570]
/lib/x86_64-linux-gnu/libc.so.6(gsignal+0x35)[0x7fa2ec0da4f5]
/lib/x86_64-linux-gnu/libc.so.6(abort+0x17b)[0x7fa2ec0ddc5b]
/lib/x86_64-linux-gnu/libc.so.6(+0x2f1be)[0x7fa2ec0d31be]
/lib/x86_64-linux-gnu/libc.so.6(+0x2f262)[0x7fa2ec0d3262]
./nvmatrix/libnvmatrix.so(_ZN8NVMatrix16getTextureObjectEv+0x121)[0x7fa2dbd60941]
./nvmatrix/libnvmatrix.so(_ZN8NVMatrix10_aggregateIN12NVMatrixAggs3MaxEN11NVMatrixOps8IdentityEN17NVMatrixBinaryOps6SecondEEEviRS_T_T0_T1_P11CUstream_stPS_+0x312)[0x7fa2dbd7a9b2]
./nvmatrix/libnvmatrix.so(_ZN8NVMatrix3maxEiRS_+0x41)[0x7fa2dbd6c9f1]
/home/assad/Desktop/finalcode/mycuda/cudaconvnet/_ConvNet.so(_ZN12SoftmaxLayer9fpropActsEifji+0x7d)[0x7fa2dca3674d]
/home/assad/Desktop/finalcode/mycuda/cudaconvnet/_ConvNet.so(_ZN5Layer5fpropERSt3mapIiP8NVMatrixSt4lessIiESaISt4pairIKiS2_EEEji+0x230)[0x7fa2dca3ab00]
/home/assad/Desktop/finalcode/mycuda/cudaconvnet/_ConvNet.so(_ZN5Layer5fpropEji+0x21c)[0x7fa2dca3583c]
/home/assad/Desktop/finalcode/mycuda/cudaconvnet/_ConvNet.so(_ZN13ConvNetThread3runEv+0x185)[0x7fa2dca5c2f5]
/home/assad/Desktop/finalcode/mycuda/cudaconvnet/_ConvNet.so(_ZN6Thread18start_pthread_funcEPv+0x9)[0x7fa2dca00179]
/lib/x86_64-linux-gnu/libpthread.so.0(+0x7e9a)[0x7fa2ecb6ee9a]
/lib/x86_64-linux-gnu/libc.so.6(clone+0x6d)[0x7fa2ec19873d]
assad@assad:~/Desktop/finalcode/mycuda$

#2

That repository (https://github.com/hunse/cuda-convnet2) is based off of https://github.com/akrizhevsky/cuda-convnet2. Have you been able to get that original repository to work on your GPU? If not, you should direct your questions there. There are many intricacies to getting GPU code to compile and run, which I am not very familiar with.


#3

solved by commenting the line
#GENCODE_SM35="-gencode arch=compute_35,code=sm_35"
GENCODE_SM50="-gencode arch=compute_50,code=sm_50"
export GENCODE_ARCH="$GENCODE_SM50"