|
Message
From: michael scott<mike.scott@j...>
Date: Thu Jul 21 11:59:23 CEST 2005
Subject: [openrisc] MAC Instructions
Hi Damjan,
I forget to mention in my e-mail that I'm running the code
from QMEM in order to get 16MegaMACs throughput at 16Mhz cpu clock
Running code in WISHBONE memory (which in my case take 3 cycles)
hides the problem as the multiplier pipeline always completes in time
Interestingly, when executing l.div/l.ldivu in QMEM you can clearly
see a stall for approx 32 cycles while the op completes
I've attached some waveforms (zipped postcript)to show the effect
If you look on the qm_rdata bus you can see the five mac instructions
comprising altenating l.mac r2,r6 (0xc4023001) and l.mac r2,r7 (0xc4023801)
the corresponing MAc operation is observed on mult_mac_result
and follows the sequence 32,96,128,192,224 as expected by the code (also
attached)
However, when executing l.macrc r4 the mac pipline has only reached the
value 128 -there are still two cycles to go.
just before time 2696 us on the waveforms it can be seen that this
'intermediate'
value in copied to r4.
I notice the line mac_stall_r is only high for 1 singe cycle. I belive it
should
high for a further two cycles for correct operation
Hope this helps
Regards
Mike
-------------- next part --------------
A non-text attachment was scrubbed...
Name: MacTest.S
Type: application/octet-stream
Size: 1615 bytes
Desc: not available
Url : openrisc/attachments/20050721/c4d1e004wave-0001.bin
|
 |
Browse
