-
Notifications
You must be signed in to change notification settings - Fork 167
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[enhancement] Support peek() op in XLS #1383
Comments
Note that peek() is generally requested to implement arbiters without the need for additional state within the block (or when the outside world uses a pop() as a signal that the arbiter chose a particular candidate). |
I'm not quite sure this is fully accurate as written, and I don't believe it's necessarily a problem. One way to think about this is that in both cases, it depends on scheduling. If the Currently, the DSLX interpreter, proc IR interpreter, and proc IR JIT don't ever simulate operations as happening simultaneously, so given the token edge, we can only get the ordering This is to be expected, since the interpreter serializes things, and is why logic that involves non-KPN operations can't be fully covered by simulations that don't take scheduling into account. FWIW, we also don't stress-test the results of changing execution order, so the interpreter doesn't even maximize the possible coverage for this scenario within these bounds. So - I'm not sure we shouldn't just go ahead and implement On the other hand, the proposed |
What's hard to do? (limit 100 words)
Often requested is a peek() op that given a streaming channel, will return the data at the head of the channel (if any) and a valid bit (true if there were data).
While both peek() and recv_non_blocking() bring us out of KPN land, peek() poses a challenge to simulation vs hardware as in simulation we do not backtrack executed instruction when blocking.
For example:
In hardware, the behavior depends on how peek and recv are scheduled. Let's say that they are scheduled in the same cycle. In this case, once the recv occurs, valid will be true and data will be equal to data0.
In an interpreter, one possible sequence of events is that the channel is empty, and peek executes --> data = 0 and valid = 0. The interpreter then blocks on recv. Once there is data on the channel, recv executes and returns data0.
In order to continue with supporting peek() operation, we should clarify and agree on the behavior of the interpreters, and how the operator is impacted by scheduling.
Current best alternative workaround (limit 100 words)
This is as opposed to the recv_non_blocking() op which will pop the data at the head of the channel if it exists.
Your view of the "best case XLS enhancement" (limit 100 words)
One idea that came out of the discussion was to implement peek and receive as a single op/expression -- so we don't need any roll-back functionality in case a receive blocks.
Proposal was something like like the following
The text was updated successfully, but these errors were encountered: