Check remaining credits when fencing in Dromajo

[drichmond]

Issue for tracking completion of #721

At the moment, fence just checks if the BP/Dromajo TX fifos are empty and does not check credits. Both checks are required, neither is sufficient for a true fencing.

[sripathi-muralitharan]

I am not sure if this already solves the issue but I do check for credits here -

https://github.com/sripathi-muralitharan/bsg_replicant/blob/3c0ac03e9b9f768eaae98676c791ad0fba98c4de/libraries/platforms/dromajo-vcs/bsg_manycore_platform.cpp#L368-L385

Although, in simulator.cpp, I will have to move the location of the call to update the credits in the Dromajo FIFOs. Basically, move this piece of code to outside the do-while loop -

https://github.com/sripathi-muralitharan/bsg_replicant/blob/3c0ac03e9b9f768eaae98676c791ad0fba98c4de/libraries/platforms/dromajo-vcs/bsg_manycore_simulator.cpp#L409-L411

[drichmond]

Maybe, but I'm still a little confused. I left comments that I think will clear the fence method up.

Can you explain why you need to move that code?

[sripathi-muralitharan]

Because it updates the credits for Dromajo only on packets transmitted to the manycore hardware. In the case of fence, no packets are transmitted to the hardware but the DPI FIFOs are also draining as packets are accepted into the manycore hardware. Therefore, the credits should be updated regardless even of a packet transmission to the manycore hardware or not to get accurate estimates of the credits used by the hardware.

But I think instead of making a call to the function that updates credits in the transmit function, doing it in the eval function makes more sense.

[drichmond]

Ah. Yes, doing it in the eval function sounds like a good idea.

The credits in hardware are updated independently of sending a packet, right?