vetter/shoc

Failures with NeuralNet tests

Opened this issue · 4 comments

I'm able to build & run SHOC from the current "trunk" image (as of 2020-12-19) except for this case:

Running benchmark NeuralNet
    result for nn_learning:                     BenchmarkError
    result for nn_learning_pcie:                BenchmarkError

which gives the message in Logs/dev0_NeuralNet.log

Chose device: name='Graphics Device' index=0
Input training file not found - please check data directory!
test    atts    units   median  mean    stddev  min max trial0  

Note: Any results marked with (*) had missing values.
      This can occur on systems with a mixture of
      device types or architectural capabilities.

It looks like the file

src/cuda/level1/neuralnet/nn_data.zip

is not being unpacked as part of the configure/make/install process.
If I unzip the file into the top-level directory alongside the bin etc,, it looks like the benchmark is picking up the input files.

After the above fix, the file Logs/dev0_NeuralNet.err says

double free or corruption (out)
Aborted (core dumped)

Given the above installation issue, is this test expected to work? Also I don't see an OpenCL version of it.

@cponder : thanks for your note. i have been out for the holidays, but will be replying on these issues over the next few days. sorry for the delay. (It has been a strange year, so I took a long vacation!)

I've been encountered the same error, is it solved now? Appreciate it for you.