Effect of NVGPU_IOCTL_CHANNEL_FORCE_RESET on GPU computation

[jay_hunter]

The documentation provides insufficient information regarding the impact of the NVGPU_IOCTL_CHANNEL_FORCE_RESET operation on the GPU. Specifically, it is unclear how the GPU responds to a hardware channel reset. For instance, does a reset interrupt further computations submitted through the reset kernel? And if there’s a computation workload with a ‘while (true)’ loop submitted through the reset channel, how does the GPU respond following the execution of NVGPU_IOCTL_CHANNEL_FORCE_RESET? Your clarification on these points would be highly appreciated. Thank you!

[AastaLLL]

Hi,

After resetting, you should be able to run a new task.
But we will check with our internal for more info and update to you later.

Thanks.

[AastaLLL]

Hi,

Here is some update from our internal team.

The mechanism is better described as a “context” reset.

When taking a channel, based on the data we can determine which context that channel belongs to.
Then removing the context from the GPU’s runlist such that it doesn’t further run.

This terminates any work pending on the channel.
If there is an actively working engine, it will be reset by halting.

Thanks.