On Tue, May 6, 2014 at 7:32 PM, Eric Anholt <e...@anholt.net> wrote: > Juha-Pekka Heikkila <juhapekka.heikk...@gmail.com> writes: > >> These patches allow MAD opcode to be used with pre gen6 hardware. >> Instead of failing on emitting MAD there will be emitted MUL and ADD >> to simulate MAD. >> >> I tried this with piglit on ILK (gen5) and did not see regression. > > This hides the MUL and ADD from instruction scheduling, which I expect > to make performance worse. What was the motivation for this?
When talking about LRP with MAC there was brief comment about doing similar for MAD. Now that you mentioned I realize this first try was not so useful. I at first tried to go using accumulator as an attempt to benefit one register worth here but that turned out to be crash'n'burn, literally. I'll look at Matt's version with LINE next for this. /Juha-Pekka _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
