RepoMirrors
/
ffmpeg
mirror of https://git.ffmpeg.org/ffmpeg.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

aarch64: vp9itxfm: Optimize 16x16 and 32x32 idct dc by unrolling 

This work is sponsored by, and copyright, Google.

Before:                           Cortex A53
vp9_inv_dct_dct_16x16_sub1_add_neon:   235.3
vp9_inv_dct_dct_32x32_sub1_add_neon:   555.1
After:
vp9_inv_dct_dct_16x16_sub1_add_neon:   180.2
vp9_inv_dct_dct_32x32_sub1_add_neon:   475.3

This is cherrypicked from libav commit
3fcf788fbb.

Signed-off-by: Martin Storsjö <martin@martin.st>
This commit is contained in:
Martin Storsjö 2017-01-04 12:57:56 +02:00
parent 758302e4bc
commit 148cc0bb89
1 changed files with 36 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
libavcodec/aarch64/vp9itxfm_neon.S
View File

@ -495,16 +495,23 @@ function idct16x16_dc_add_neon
srshr v2.8h, v2.8h, #6 srshr v2.8h, v2.8h, #6
mov x3, x0
mov x4, #16 mov x4, #16
1: 1:
// Loop to add the constant from v2 into all 16x16 outputs // Loop to add the constant from v2 into all 16x16 outputs
ld1 {v3.16b}, [x0] subs x4, x4, #2
uaddw v4.8h, v2.8h, v3.8b ld1 {v3.16b}, [x0], x1
uaddw2 v5.8h, v2.8h, v3.16b ld1 {v4.16b}, [x0], x1
sqxtun v4.8b, v4.8h uaddw v16.8h, v2.8h, v3.8b
sqxtun2 v4.16b, v5.8h uaddw2 v17.8h, v2.8h, v3.16b
st1 {v4.16b}, [x0], x1 uaddw v18.8h, v2.8h, v4.8b
subs x4, x4, #1 uaddw2 v19.8h, v2.8h, v4.16b
sqxtun v3.8b, v16.8h
sqxtun2 v3.16b, v17.8h
sqxtun v4.8b, v18.8h
sqxtun2 v4.16b, v19.8h
st1 {v3.16b}, [x3], x1
st1 {v4.16b}, [x3], x1
b.ne 1b b.ne 1b
ret ret
@ -1054,20 +1061,31 @@ function idct32x32_dc_add_neon
srshr v0.8h, v2.8h, #6 srshr v0.8h, v2.8h, #6
mov x3, x0
mov x4, #32 mov x4, #32
1: 1:
// Loop to add the constant v0 into all 32x32 outputs // Loop to add the constant v0 into all 32x32 outputs
ld1 {v1.16b,v2.16b}, [x0] subs x4, x4, #2
uaddw v3.8h, v0.8h, v1.8b ld1 {v1.16b,v2.16b}, [x0], x1
uaddw2 v4.8h, v0.8h, v1.16b uaddw v16.8h, v0.8h, v1.8b
uaddw v5.8h, v0.8h, v2.8b uaddw2 v17.8h, v0.8h, v1.16b
uaddw2 v6.8h, v0.8h, v2.16b ld1 {v3.16b,v4.16b}, [x0], x1
sqxtun v3.8b, v3.8h uaddw v18.8h, v0.8h, v2.8b
sqxtun2 v3.16b, v4.8h uaddw2 v19.8h, v0.8h, v2.16b
sqxtun v4.8b, v5.8h uaddw v20.8h, v0.8h, v3.8b
sqxtun2 v4.16b, v6.8h uaddw2 v21.8h, v0.8h, v3.16b
st1 {v3.16b,v4.16b}, [x0], x1 uaddw v22.8h, v0.8h, v4.8b
subs x4, x4, #1 uaddw2 v23.8h, v0.8h, v4.16b
sqxtun v1.8b, v16.8h
sqxtun2 v1.16b, v17.8h
sqxtun v2.8b, v18.8h
sqxtun2 v2.16b, v19.8h
sqxtun v3.8b, v20.8h
sqxtun2 v3.16b, v21.8h
st1 {v1.16b,v2.16b}, [x3], x1
sqxtun v4.8b, v22.8h
sqxtun2 v4.16b, v23.8h
st1 {v3.16b,v4.16b}, [x3], x1
b.ne 1b b.ne 1b
ret ret