tree 7a67b6918337e722196f1ecea018d7e9493b2e7b
parent 025260095bfb8ca59fa11b7fd545adbed019c109
author Aniket Dhok <aniket.dhok@ittiam.com> 1539847932 +0530
committer Yunqing Wang <yunqingwang@google.com> 1539878909 +0000

Optimize highbd fwd_txfm modules

Added sse4_1 variant for identity tx_type of 4,8,16.

Enabled sse4_1 optimizations for tx_sizes 4x4, 4x8,
8x4, 8x8, 4x16, 16x4, 8x16, 16x8, 16x16.

Module level gains:
Tx_size    Gain w.r.t. C
4x4         7.14x
4x8         4.42x
4x16        4.25x
8x4         5.00x
8x8         4.57x
8x16        3.40x
16x16       3.00x
16x4        4.80x
16x8        3.28x

When tested for 20 frames of crowd_run_360p_10 at 1 mbps
for speed=1 preset, observed ~2.5% reduction in encoder time.

Change-Id: I4f2b2405a4becb02a07d991b491a01926892ec44