cudnn samples are provided to verify if the operator's calculation results are correct on other gpu products compatible to nvidia cuda/cudnn.
for example, we try to generate data(e.g, input/weight/bias/outout) with tensorflow api, such as tf.nn.maxpool(), and then using this data to verify the correctness of cudnnMaxpoolForward api.
# generate input/output data
$ python tf-maxpooling.py
# modify float32.h to replace input/output data.
dtype input[IN_SIZE] ={...}
dtype output[OUT_SIZE] = {...}
# verify cudnnMaxpoolForward api.
$ make
$ ./test