Profiling allocator GPU memory usage with GPGMM

Why use GPGMM for GPU memory profiling?

GPU applications very typically rely on the ability to recycle larger OS allocated memory blocks by either sub-dividing them into smaller blocks or re-using the entire memory block once not being used. Because this happens by the application, above the OS, existing OS-level profiling tools are unable to accurately report memory usage without application-specific instrumentation. GPGMM provides this instrumentation to application developers.

Profiling memory usage through GPGMM event tracing.

GPGMM event tracing is able to report:

1. Allocator performance and memory utilization metrics.
2. Allocation lifetimes (static vs transient).
3. Detailed API timings.
4. Captured API calls for later (offline) playback.

To enable:

1. Specify ALLOCATOR_RECORD_FLAGS option when creating the allocator.
2. Run the app, where upon termination, a captured trace file will be dumped to the current working directory.
3. Navigate to chrome://tracing, click the load button, then select the trace file to view.