Why is there still several CPU-to-GPU transfers between two consecutive kernel calls in CUDA?

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

I have already packed the data and transferred it to the GPU. The relationship between the two kernel functions is that the value computed by the first kernel function is used by the second kernel function. There shouldn’t be any data transfer between these two kernel functions, so why is there still several transfers from the CPU to the GPU?
This is the behavior between the first and second kernel functions.

1103
1103857×454 145 KB

Then, it is the data transfer behavior between these two kernel functions as measured by Nsight Systems.
1104
1104303×454 6.74 KB

Additionally, I want to mention that I have already allocated pinned memory, and all the arrays used have been allocated memory. I have only encapsulated the global subroutine

2

Difficult to say without a reproducer. Could be the Fortran descriptors being copied as we’re discussing in your other post.

3

Sure, let me clarify your meaning. What I previously mentioned about the small transfers being caused by not using a module to encapsulate memory allocation, data transfer, and kernel functions was incorrect. What I meant is that you might not be able to see the actual examples, and for now, we can only suspect that the descriptors in the transfer are affecting the transfer time.