https://bugs.kde.org/show_bug.cgi?id=429354
Carl Love <c...@us.ibm.com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Attachment #133461|0 |1
is obsolete| |
--- Comment #5 from Carl Love <c...@us.ibm.com> ---
Created attachment 136162
--> https://bugs.kde.org/attachment.cgi?id=136162&action=edit
Functional support for ISA 3.1, VSX Mask manipulation operations
Updated the patch per comments.
I looked at the suggestions on how to make copy_MSB_bit_fields() more
efficient. PPC64 supports doing the basic algorithm that was outlined as a
single V128 value except for doing the sum across lanes. That would need to be
done using a clean helper. In studying the code, I realized that the
copy_MSB_bit_fields() is very similar to the ISA 3.1 instruction vgnb. The
difference is the result is stored in the low order bits not the high order
bits. I rewrote copy_MSB_bit_fields() using the vgnb implementation as a base
leveraging the existing clean helpers. The new function is much more
efficient. Although I didn't use the suggested algorithm, it was very helpful
to study. I realized that a number of the other instructions in the patch
could be done using the arithmetic right shift much more efficiently without
the use of the copy_MSB_bit_fields() or the "for(i = 0; i< max; i++)" which
Julian had concerns about.
Reimplemented the vexpandbm, vexpanddmn, vexpandhm, vexpandwm instructions
eliminating the call to copy_MSB_bit_fileds() and the need for the
for(i = 0; i< max; i++) loop. The new implementation is much more efficient.
Reimplemented the mtvsrbm, mtvsrhm, mtvsrwm, mtvsrdm instructions to remove the
need for the for(i = 0; i< max; i++) loop. The new implementation is much more
efficient.
The for(i = 0; i< max; i++) has been removed.
--
You are receiving this mail because:
You are watching all bug changes.
