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 basic verification idea of T1 IP is based on the online difftest. However when we are getting bigger and bigger, verification becomes a new problem, which we need to think about in a new way:
Does online difftest really necessary?
Answer is NO. Basically online difftest has no difference between offline. It just stop verification at early as possible, but does this really necessary? I don't think so.
difftest only helps for those lazy guys don't wanna to re-implement SVA for each instructions.
What's the interface between RTL and DV?
Originally we use DPI to probe each signals from designs.
But problem is, when we have different configurations, DPI call is different. C codegen is a nightmare.
Thus what we are looking for is a decoupled verification framework:
There are two Verification frameworks to provide trade-off among: simulation performance, verification target, and delivery.
IP Level: forcing on the vector performance, avoiding and overhead from scalar part, it should skip the scalar execution with emulating by spike. The DPI input is Scalar->Vector IO and Memory IO. This signals are driven by DPI.
SoC Level: forcing on the Tile-level verification with Rocket as scalar code, it will fully test the performance of the vector tile. The interface is the front-pront for ITCM and DTCM and all vector memory ports
TBD, too tired.
The text was updated successfully, but these errors were encountered:
The basic verification idea of T1 IP is based on the online difftest. However when we are getting bigger and bigger, verification becomes a new problem, which we need to think about in a new way:
Thus what we are looking for is a decoupled verification framework:
TBD, too tired.
The text was updated successfully, but these errors were encountered: