tree c6f9aa3da811bf4cea3d66a169873b6d59fea1ba
parent 6f295fd0936ba0ed5231eba17f5004ea73472177
author Ruiling Song <ruiling.song@intel.com> 1568365228 +0800
committer Yunqing Wang <yunqingwang@google.com> 1569509962 +0000

Minor improvement to warp_affine_avx2

This patch basically includes two minor improvement:

1.) Through inserting the 128bit data to high half of ymm,
    we are reducing register pressure, this would avoid
    register spill/fill for gcc, and also save some unpacking
    instructions.

2.) The '(unsigned)' convert is telling the compiler that we
    know the array offset is non-negative, thus one 'movslq'
    instruction could be avoided for array offset calculation.

Profiling below command:
./aomenc 1080p_park_joy.y4m --cpu-used=3 --cq-level=32 --limit=50 --passes=1 -o parkjoy.webm

Before the patch:
     5.81%     5.80%  aomenc   aomenc            [.] av1_warp_affine_avx2
            |
             --5.79%--av1_warp_plane
                       av1_warp_affine_avx2
With the patch:
     5.53%     5.52%  aomenc   aomenc            [.] av1_warp_affine_avx2
            |
             --5.51%--av1_warp_plane
                       av1_warp_affine_avx2

Change-Id: Ic245e4da3f0f39f95f69c16fffce01fdd9fd9117