arthurxlw / cytonrl Goto Github PK

The problem when I train a model after compiling will be like this:

A.L.E: Arcade Learning Environment (version 0.6.0)
[Powered by Stella]
Use -help for help screen.
Warning: couldn't load settings file: ./ale.cfg
version 1.0
mode: train
batchSize: 32
dueling: 1
env: roms/breakout.bin
gamma: 0.99
inputFrames: 4
learningRate: 0.0000625
learnStart: 50000
loadModel:
maxEpisodeSteps: 18000
maxSteps: 100000000
optimizer: RMSprop
priorityAlpha: 0.6
priorityBeta: 0.4
progPeriod: 10000
replayMemory: 1000000
saveModel: model/model
savePeriod: 1000000
showScreen: 0
targetQ: 30000
testEGreedy: 0.001
testEpisodes: 100
testMaxEpisodeSteps: 18000
testPeriod: 5000000
updatePeriod: 4
networkSize: 32:64:64:512
eGreedy: 1.0:0.01:5000000
cuDNN Error: CUDNN_STATUS_NOT_INITIALIZED src/cytonLib/Global.cu 67
cytonRl: src/cytonLib/basicHeads.cu:52: cudnnStatus_t cytonLib::checkError_(cudnnStatus_t, const char*, int): Assertion `false' failed.
Aborted (core dumped)

And my Driver Version is: NVIDIA-SMI 396.26, cuda=9.2.88, cudnn=7.1.4
What's wrong with it?

Well, I will really appreciate your answer. @arthurxlw

Thanks.

DeviceMatrix.cu
line 381
if(tid < warpSize){
for (size_t shift = warpSIze; shift>0;shift>>=1)
sdata[tid] += sdata[tid +shift];

without ”__syncthreads()“ here, the shared memory sdata should be set volatile.

reference:
https://stackoverflow.com/questions/21205471/cuda-in-warp-reduction-and-volatile-keyword