CALL instructions in the trace

[christindbose]

Hi,

I've been trying to trace NCCL kernels (ReduceScatter in this example).

I'm seeing many invocations of a CALL instruction and it seems like that the trace generated is incomplete.

What's weird is that this only happens after a vector size of 128kB. I can trace reducescatter kernels for vector sizes smaller than 128kB with no issues and correlate them fairly well on hardware.

Any advice would help on interpreting this and possible solutions.