You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The spatial array in Gemmini is double-buffered; you can load in new weights into one set of registers, while continuing to perform matmuls with the old weights in the other set of registers.
compute_accumulated and compute_preloaded in WS mode determine whether your new matmul will use the new weights that you just loaded in, or whether they should use the old weights. Tbh, these types of decisions probably shouldn't have been exposed to the programmer; it would have been better if the hardware had handled these kinds of decisions under-the-hood instead.
So you compute on Register A until a compute_accumulated instruction is encountered, after which you compute on Register B?
And the preload instruction will load into the currently inactive register?
Hello,
from the documentation, they seem to do exactly the same operation for the WS dataflow. Is there a difference here that I am not seeing? Thank you!
The text was updated successfully, but these errors were encountered: