]> granicus.if.org Git - libvpx/commit
vp9: neon: optimise loads in horiz convolve functions
authorMans Rullgard <mans@mansr.com>
Tue, 30 Jul 2013 17:11:06 +0000 (18:11 +0100)
committerJohann <johann.koenig@duck.com>
Fri, 2 Aug 2013 15:11:08 +0000 (08:11 -0700)
commit5e2e78d024455a60dfe509a25cf1d715d7de03c3
treef12594e43530270ffc47279173378593be958ed0
parentd85ae8718312f36bd43da423ee22dc098e5f2c6c
vp9: neon: optimise loads in horiz convolve functions

Loading to single lanes in multiple registers is expensive since
it requires a read and write of each register which saturates
the register file access.  Loading to single registers followed
by a separate transpose reduces this pressure.

Change-Id: I4cc35887ddbca80e5e635b50d2b1d158de9668ee
vp9/common/arm/neon/vp9_convolve8_avg_neon.asm
vp9/common/arm/neon/vp9_convolve8_neon.asm