Timestamp query resets

[kainino0x]

I'm migrating this from an inline issue:

Note: The timestamp values may be zero if the physical device reset timestamp counter, please ignore it and the following values.

Issue: Write normative text about timestamp value resets.

I do not know what needs to be written here. More detail from #614 which I assume is about the same thing:

Compute the difference of the two timestamp queries. The time delta may be negative due to the timestamp counter may be reset after a long time on some platforms. We can suggest users to skip the invalid time delta if it’s negative.

Regardless since this is defined in the V1 spec (even though implementations may ship without it, as Chrome does), marking as V1.

[kainino0x]

@haoxli do you know what needs to be added here?

[kainino0x]

Thread where I requested this inline issue be added: #771 (comment)

[haoxli]

Sorry for my late reply, I missed this notification.

On Nivdia/AMD GPUs, their GPU timestamps from query API are nanoseconds since POSIX time (also known as Epoch time, Unix time), timestamps have always been positive growth.

To my knowledge, timestamp query resets will happen on Intel GPUs, its GPU counter restart from a big number every few hours. So we have to make user aware that we should skip and do not compare the timestamp values if any timestamp is less than the previous one.

On Apple GPUs, the GPU counter is similar to Intel GPUs, but its timestamps are much bigger than the values on Intel and Nividia GPUs. I'm not sure what will happen if the counter reaches the maximum limitation (may never reach if the counter is large enough).

Note: The timestamp values may be zero ....

This is out of date. We don't need to make the timestamps to be 0s when timestamp query resets, which will complicate the implementation of the compute shader for timestamp conversion. Talked with Corentin in sync meeting before, just notice users to take care of the timestamps which may be reset.

[kainino0x]

Resolved in 2023-05-24 meeting: just describe in the spec when/how timestamp resets can occur, don't try to guarantee anything more