CUDA memory access recorder and visualizer

This is a simple tool that records all memory accesses and timestamps of the accesses in a CUDA program. It is done by writing every memory access index and streaming multiprocessor clock cycle value during the access into global device memory. Depending on the amount of memory accesses, this might require quite a lot of space and makes this tool usable only on very small datasets.

The example directory contains some screen recordings of access pattern animations for the examples.

Quickstart

To generate access patterns for the examples seen above, go to

cd examples/v0
make && ./bin/main

Start a local web server for the animation app:

cd ../../web && python3 -m http.server

Go to http://0.0.0.0:8000 and submit the generated examples/v0/access-patterns-v0.json file.

You should now see the access pattern from the first gif.

To use the v1 kernel, open examples/v0/main.cu and define the kernel_v1 macro instead of kernel_v0.

If it does not work

Some things to try:

• Make sure you are running wrapper AccessCounter first, then PatternRecorder.
• The wrapper objects support only array indexing, pointer arithmetic etc. is not available.
• Make sure the wrapper object calls enter_kernel once somewhere at the beginning of a kernel before the first memory access.
• Make sure you call cudaDeviceSynchronize after the kernel call so that the unified memory pointers are accessible at the host.
• Define the value of macro PR_VERBOSITY as 1 before including pattern_recorder.cuh. This will trigger some asserts.
• If you are getting a warning of possibly using too much device memory, try reducing the number of required memory accesses by using a smaller data sample.