On Mon, 25 Jun 2012, Jason Garrett-Glaser wrote:
>> + � �packssdw � m0, m3 � � � � � ; m0 = �0, �6, 12, 18, �3, �9, 15, 21
>> + � �packssdw � m1, m4 � � � � � ; m1 = �1, �7, 13, 19, �4, 10, 16, 22
>> + � �packssdw � m2, m5 � � � � � ; m2 = �2, �8, 14, 20, �5, 11, 17, 23
>> + � � � � � � � � � � � � � � � �; unpack words:
>> + � �movhlps � �m3, m0 � � � � � ; m3 = �3, �9, 15, 21, �x, �x, �x, �x
>> + � �punpcklwd �m0, m1 � � � � � ; m0 = �0, �1, �6, �7, 12, 13, 18, 19
>> + � �punpckhwd �m1, m2 � � � � � ; m1 = �4, �5, 10, 11, 16, 17, 22, 23
>> + � �punpcklwd �m2, m3 � � � � � ; m2 = �2, �3, �8, �9, 14, 15, 20, 21
>> + � � � � � � � � � � � � � � � �; unpack dwords:
>> + � �punpckhdq �m3, m0, m2 � � � ; m3 = 12, 13, 14, 15, 18, 19, 20, 21
If you swap the inputs here
>> + � �punpckldq �m2, m1 � � � � � ; m2 = �2, �3, �4, �5, �8, �9, 10, 11
>> + � � � � � � � � � � � � � � � �; shuffle qwords:
>> + � �movhlps � �m4, m2 � � � � � ; m4 = �8, �9, 10, 11, �x, �x, �x, �x
>> + � �movlhps � �m4, m3 � � � � � ; m4 = �8, �9, 10, 11, 12, 13, 14, 15
>> (final)
then you can get away with one less op here
>> + � �movlhps � �m2, m0 � � � � � ; m2 = �2, �3, �4, �5, �0, �1, �6, �7
>> + � �movhlps � �m3, m1 � � � � � ; m3 = 16, 17, 22, 23, 18, 19, 20, 21
at the cost of changing this to shufps or pblendw or vpperm.
Which should be a good tradeoff on nehalem+ and k10+, but bad on
conroe/penryn/k8.
>> + � � � � � � � � � � � � � � � �; shuffle dwords:
>> + � �pshufd � � m2, m2, q3102 � �; m2 = �0, �1, �2, �3, �4, �5, �6, �7
>> (final)
>> + � �pshufd � � m3, m3, q1320 � �; m3 = 16, 17, 18, 19, 20, 21, 22, 23
>> (final)
>
> Could some shuffps be used here instead?
No, shufps has the constraint that the bottom half must come from one
source reg and the top half from the other. vpperm would work.
--Loren Merritt
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
