Following the T1040DPAARM
Below is the description to to Enqueue Command
188.8.131.52.1 Writing a Command into the CR
To write a command into the CR, software must perform the following tasks:
2. Write all words of the command other than word 0.
4. Write word 0 (containing the command verb and alternating valid bit).
5. dcbf to flush the command from cache to the BMan
As far as I understand the QMAN has both CE and CI registers area's. From performance point of view should it make sense to use the CI area to enqueue command and therefor omit the dcbz and dcbf commands.
As to the performance that is reached with that mechanism we have much less then in p2020 ETSEC for sending frames. As far as I understand most of this performance issue is due to dcbz and dcbf.
Any idea how can it be improved ?
Is it possible (to reduce the ring size so less cache lines will be used ?