illinois-impact / ece408_project Goto Github PK

Show how to read various performance metrics and use them to understand effect of optimizations.

I verified that module load cuda-toolkit/10 on EWS will allow you to run nvvp from the command line instantly. This saves a lot of trouble installing the functionality.

If it works, maybe you want to modify https://github.com/illinois-impact/ece408_project#nvvp-on-ews accordingly.

In fall 2019, some students report seeing the following error occasionally

* Running /usr/bin/time python m2.1.py
Loading fashion-mnist data...done
Loading model...done
New Inference
Op Time: <snip>
python: malloc.c:3079: __libc_malloc: Assertion `!victim || chunk_is_mmapped (mem2chunk (victim)) || ar_ptr == arena_for_chunk (mem2chunk (victim))' failed.
Command terminated by signal 6

1. Guidelines to contact course staff early if there is trouble.
2. Add section of teammate work division in final report.

"Skeleton Code Description
custom/cpu-new-forward.cc and custom/new-forward.cu containes skeleton implementations for the CPU and CPU convolutions respectively. You can complete the project by modifying these two files only. custom/cpu-new-forward.h and custom/gpu-new-forward.h are the respective header files. You need not modify these files unless you need to declare your own functions.

The code in m2.cc, m3.cc, m4.cc and final.cc are the top level files that are executed for each milestone. You should not be modifying these files."

should be CPU and GPU