search

ucb-bar / chisel2-deprecated

chisel.eecs.berkeley.edu
387 stars 89 forks source link

Fixed AsyncFifo so reset not tied to 0 #690

Open da-steve101 opened 8 years ago

da-steve101 commented 8 years ago

Previous logic was

if reset = 1
  s1_rst = 0
else
  s1_rst = reset // reset = 0

changed to

s1_rst = reset
ghost commented 8 years ago

Can one of the admins verify this patch?

ucbjrl commented 8 years ago

ok to test

aswaterman commented 8 years ago

That looks correct to me. However, I'm also wondering about the other Regs that are reset with init=. They appear to be using the module's implicit reset, which is probably incorrect for either the read side or the write side... right?

sdtwigg commented 8 years ago

No, assuming things haven't changed too much since I last peered into the recesses of chisel MCD code: If a register defines an explicit clock, then the reset signal attached to that clock is used to reset that register. AFAIK, nothing in that module uses the implicit reset.

I think, on a more philosophical level, if this AsyncFIFO is going to get some more exposure, its 'contract' (particular, reset behavior across clock domains) should be more explicitly explained spelled out. If I recall, this AsyncFIFO relies on higher-level reset synchronization between the clock domains. For example, you will get incorrect behavior if one side goes in and out of reset and the other side does not reset. Also, the module assumes that it is OK to use only 2 registers to construct a synchronizer between the two clock domains.