Closed wimrijnders closed 3 years ago
Actually, I believe it would be better to have separate loops for the add alu and mul alu, so that the test is complete.
Same goes for the next loop, where the rotate value comes from register r5
instead of a small immediate.
Thank you for your reporting. We will fix it.
My pleasure!
That was quick! Pulling and testing right away.
Working!
However, I see something interesting: the assembled code output for rotate(r1, r0, i)
(add alias) is the same as nop().rotate(r1, r0, i)
(mul alias). In both cases, the mul ALU is used.
Perhaps it is the case that rotate only works via the mul ALU?
Yes, rotate only works on mul ALU as in VC4 QPU.
Thank you for the confirmation! And thank you for the extremely quick fix.
Hi there,
I believe the following is incorrect:
Source:
i % 1
will always return 1, therefore the condition is always false and the mul alias branch is always selected. I think it should bei % 2
, so that add alias and mul alias are alternately outputted, which seems to be the intent here.Do you agree?