Profiler count cuda time or cpu time?

1

Hi, I am using profiler to count cuda kernels time, but I met a very strange question.
codes is below:

x = torch.randn((10, 3, 224, 224)).cuda()
a = torch.randn((5, 6, 7, 8))

with torch.autograd.profiler.profile(use_cuda=True) as prof:
    y = x ** 2
    y = y + 1.5
    y = y * 2.4
    y = torch.relu(y)
    b = a + 1.1
    y = y * 1/4
print("prof is:\n{}".format(prof))

and I find the kernel a is placed on cpu, but in profiler table, its time also shown, and its cpu time are neally nearly to cuda time. I am confused.

2

Here is image.

9_11
9_114032×1231 334 KB

I know the cuda time is get by cudaEvent series kernels, I think it is hardware time, but a is cpu kernel, its cpu time is shown in column of CUDA time.
I want to know, profiler used to count time, is the time including hardware time and cpu time which cpu kernel is inserted in cuda kernels?
Thank you very much!

1 Like