It's a progress bar whose completion is controlled from a running CUDA kernel compiled with https://numba.pydata.org.
python kernel_progress.py
- The latest Numba master branch
- A Compute Capability 6.0 or greater CUDA GPU (e.g. GTX 1xxx, RTX 2/3xxx, Pascal / Volta / Turing / Ampere Quadro GPUs).
- Linux (Managed Memory in Numba is experimental on Windows, and this use of it seems to crash it).
- ... Maybe other things I forgot to mention.
A managed array visible to both the host and device holds an integer representing current progress:
progress = cuda.managed_array(1, dtype=np.uint64)The running kernel increments this in a loop (with some sleeping, so it doesn't run too fast):
for i in range(MAX_VAL):
cuda.atomic.inc(progress, 0, MAX_VAL)
cuda.nanosleep(KERNEL_SLEEP)Simultaneously, the host reads the progress value and updates the progress bar in a loop:
while val < MAX_VAL:
sleep(0.001)
val = progress[0]
bar.goto(val)Maybe...
- I have not thought hard about synchronization issues between the host and device,
- I have not tested this extensively,
- YMMV!
Maybe not...
- If you're going to, you should probably read and understand Section M.2.2 of the CUDA Programming Guide, which discusses coherency and concurrency of managed memory.
- You should also examine whether this technique has a performance impact,
- and think about whether breaking a long-running kernel launch into multiple shorter kernels would provide a better way of synchronizing progress with the host.