Read (Download) the paper (PDF)

Slides (PDF) are here

To reproduce the results in this paper, you'll need GPU-FPX's old version with the hang bug fix (for Sec 3.1-3.3)

• Follow instructions on https://github.com/LLNL/GPU-FPX to install and build GPU-FPX. Then try out the examples in this paper as described below.
• The version without the hang is in the nvbit-1.7.5 branch of https://github.com/LLNL/GPU-FPX.

To try out the new Tensor-Core version of GPU-FPX that has merged the functionality into one tool called "detector.so", do this:

• Follow the general instructions on https://github.com/LLNL/GPU-FPX.git but with priority to instructions in https://github.com/parfloat/nixnan.git (new version).

• This version currently only supports compute capability 8.6, so programs should be compiled with the -arch=sm_86 flag.

To reproduce the SRU results of Sec 3.1, please do the following:

• As instructed in https://github.com/asappresearch/sru/, do pip install sru
• Then follow the script fully given in the paper.

To reproduce the results in Sec 3.2, please do the following:

• The version that does not have any issues is main.cu (given here on this Github)

• The version that has the issues is main1.cu (given here on this Github)

• Whichever you are running, compile the file as follows: for main.cu, it is as follows:

• nvcc -O3 -arch=sm_75 -lineinfo -o cuszp main.cu

• Here, set arch=... version before running to match your GPU)

• Then run LD_PRELOAD=detector.so ./cuszp

• change 'detector' to 'analyzer' to conduct analysis

• Change to nvcc -O3 -arch=sm_75 -lineinfo -o cuszp1 main1.cu

• Then run LD_PRELOAD=detector.so ./cuszp1 to see the bug reported in the paper

• main1.cu differs from main.cu in that its tests have higher coverage

• Each run of cuszp or cuszp1 will generate a test data file and save the compressed data into the file test_data.cuszp.bin

To reproduce the results in Sec 3.3 involving PyBlaz, do the following

• First install PyBlaz as instructed here:
• Next, you need to run script3.py which is included in this Github