Profiling render times with 'cuda_rgb'

[AakashKT]

Hi,

Thank you for the awesome work with Mitsuba 3!

I was wondering how we can profile render times.
In my case, I am calling mi.render() from python three times, for three different integrators.

time_ref = time.time()
ref_img = mi.render(scene, integrator='direct' spp=100)
time_ref = time.time() - time_ref

time_1 = time.time()
img_1= mi.render(scene, integrator=integrator_1, spp=100)
time_1= time.time() - time_1

time_2= time.time()
img_2= mi.render(scene, integrator=integrator_2, spp=100)
time_2= time.time() - time_2

I know for a fact that time_1 and time_2 should be almost equal, but the latter is larger. In fact, when I evaluate integrator_2 first and then integrator_1, I get opposite effect.

integrator_1 and integrator_2 are both witten in python.

Is there a better way to do this profiling?

Thanks!

[njroussel]

Hi @AakashKT

There's a feature called KernelHistory which allows us to get the actual runtimes of the CUDA/Optix kernels. I've given a brief explanation on how to use it in this comment.

If the kernel runtimes are similiar for the two mi.render calls then the only difference is the time it takes to trace through the Python integrator to build the JIT graph.(Note: there should typically be two kernels per mi.render call)
If the kernel runtimes differ widely, it most likely has to do with the nature of your integrators. The recommended steps to profile this any deeper would be to remove parts of you integrators to understand the runtime costs that they entail.

[AakashKT]

Hi @njroussel ,

Thanks for the quick reply, works like a charm!