On 6/5/25 10:15 PM, Vineet Gupta wrote:
On 6/5/25 20:09, 钟居哲 wrote:
Hi. Vineet. The series of patches LGTM from myside.
Thx.
But I wonder whether you would like to optimize VXRM which is using
mode-switching too.
Was not planning to :-)
I saw in spec 2017 spec 624 x264.
csrwi vxrm is calling multiples times.
Is it now ? I think in gcc-15 Jeff added a trick to make it hoist out of loop,
all the way to function top - I think it was pixel_avg in x264.
2024-10-30 a65e1487cda9 [RISC-V] Aggressively hoist VXRM assignments
I haven't checked recently though !
Right. We're already optimizing vxrm and it's also a notably different
animal than frm.
Under the right circumstances we add a fake "need" for vxrm on the first
insn. That handles the hot cases in x264.
It would be interesting to see if the recent work makes us find more
vaaddu opportunities (I was aware of one we missed in x264, but again
the code was pretty cold so it never got on the todo list).
Jeff
