virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-11-21 10:55:51 +02:00
Code Issues Releases Activity
8e5a2495a8
FFmpeg/libavcodec/aarch64/h264cmc_neon.S

451 lines
15 KiB
ArmAsm
Raw Normal View History

aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
/*
* Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
* Copyright (c) 2013 Janne Grunau <janne-libav@jannau.net>
*
Merge commit '71617884a2a673908bd5c0f73d4f91fdca3da82a' * commit '71617884a2a673908bd5c0f73d4f91fdca3da82a': aarch64: h264 chroma motion compensation NEON optimizations Merged-by: Michael Niedermayer <michaelni@gmx.at>
2014-01-15 16:59:44 +03:00
* This file is part of FFmpeg.
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
*
Merge commit '71617884a2a673908bd5c0f73d4f91fdca3da82a' * commit '71617884a2a673908bd5c0f73d4f91fdca3da82a': aarch64: h264 chroma motion compensation NEON optimizations Merged-by: Michael Niedermayer <michaelni@gmx.at>
2014-01-15 16:59:44 +03:00
* FFmpeg is free software; you can redistribute it and/or
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
Merge commit '71617884a2a673908bd5c0f73d4f91fdca3da82a' * commit '71617884a2a673908bd5c0f73d4f91fdca3da82a': aarch64: h264 chroma motion compensation NEON optimizations Merged-by: Michael Niedermayer <michaelni@gmx.at>
2014-01-15 16:59:44 +03:00
* FFmpeg is distributed in the hope that it will be useful,
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
Merge commit '71617884a2a673908bd5c0f73d4f91fdca3da82a' * commit '71617884a2a673908bd5c0f73d4f91fdca3da82a': aarch64: h264 chroma motion compensation NEON optimizations Merged-by: Michael Niedermayer <michaelni@gmx.at>
2014-01-15 16:59:44 +03:00
* License along with FFmpeg; if not, write to the Free Software
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/aarch64/asm.S"
h264chroma: Change type of stride parameters to ptrdiff_t This avoids SIMD-optimized functions having to sign-extend their stride argument manually to be able to do pointer arithmetic.
2016-09-01 21:41:01 +02:00
/* chroma_mc8(uint8_t *dst, uint8_t *src, ptrdiff_t stride, int h, int x, int y) */
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
.macro h264_chroma_mc8 type, codec=h264
function ff_\type\()_\codec\()_chroma_mc8_neon, export=1
.ifc \type,avg
mov x8, x0
.endif
prfm pldl1strm, [x1]
prfm pldl1strm, [x1, x2]
.ifc \codec,rv40
movrel x6, rv40bias
lsr w9, w5, #1
lsr w10, w4, #1
lsl w9, w9, #3
lsl w10, w10, #1
add w9, w9, w10
add x6, x6, w9, UXTW
ld1r {v22.8H}, [x6]
.endif
.ifc \codec,vc1
movi v22.8H, #28
.endif
mul w7, w4, w5
lsl w14, w5, #3
lsl w13, w4, #3
cmp w7, #0
sub w6, w14, w7
sub w12, w13, w7
sub w4, w7, w13
sub w4, w4, w14
add w4, w4, #64
b.eq 2f
dup v0.8B, w4
dup v1.8B, w12
ld1 {v4.8B, v5.8B}, [x1], x2
dup v2.8B, w6
dup v3.8B, w7
ext v5.8B, v4.8B, v5.8B, #1
1: ld1 {v6.8B, v7.8B}, [x1], x2
umull v16.8H, v4.8B, v0.8B
umlal v16.8H, v5.8B, v1.8B
ext v7.8B, v6.8B, v7.8B, #1
ld1 {v4.8B, v5.8B}, [x1], x2
umlal v16.8H, v6.8B, v2.8B
prfm pldl1strm, [x1]
ext v5.8B, v4.8B, v5.8B, #1
umlal v16.8H, v7.8B, v3.8B
umull v17.8H, v6.8B, v0.8B
subs w3, w3, #2
umlal v17.8H, v7.8B, v1.8B
umlal v17.8H, v4.8B, v2.8B
umlal v17.8H, v5.8B, v3.8B
prfm pldl1strm, [x1, x2]
.ifc \codec,h264
rshrn v16.8B, v16.8H, #6
rshrn v17.8B, v17.8H, #6
.else
add v16.8H, v16.8H, v22.8H
add v17.8H, v17.8H, v22.8H
shrn v16.8B, v16.8H, #6
shrn v17.8B, v17.8H, #6
.endif
.ifc \type,avg
ld1 {v20.8B}, [x8], x2
ld1 {v21.8B}, [x8], x2
urhadd v16.8B, v16.8B, v20.8B
urhadd v17.8B, v17.8B, v21.8B
.endif
st1 {v16.8B}, [x0], x2
st1 {v17.8B}, [x0], x2
b.gt 1b
ret
h264: avoid using uninitialized memory in NEON chroma mc Adapt commit 982b596ea6640bfe218a31f6c3fc542d9fe61c31 for the arm and aarch64 NEON asm. 5-10% faster on Cortex-A9.
2014-06-18 18:10:42 +03:00
2: adds w12, w12, w6
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
dup v0.8B, w4
h264: avoid using uninitialized memory in NEON chroma mc Adapt commit 982b596ea6640bfe218a31f6c3fc542d9fe61c31 for the arm and aarch64 NEON asm. 5-10% faster on Cortex-A9.
2014-06-18 18:10:42 +03:00
b.eq 5f
tst w6, w6
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
dup v1.8B, w12
b.eq 4f
ld1 {v4.8B}, [x1], x2
3: ld1 {v6.8B}, [x1], x2
umull v16.8H, v4.8B, v0.8B
umlal v16.8H, v6.8B, v1.8B
ld1 {v4.8B}, [x1], x2
umull v17.8H, v6.8B, v0.8B
umlal v17.8H, v4.8B, v1.8B
prfm pldl1strm, [x1]
.ifc \codec,h264
rshrn v16.8B, v16.8H, #6
rshrn v17.8B, v17.8H, #6
.else
add v16.8H, v16.8H, v22.8H
add v17.8H, v17.8H, v22.8H
shrn v16.8B, v16.8H, #6
shrn v17.8B, v17.8H, #6
.endif
prfm pldl1strm, [x1, x2]
.ifc \type,avg
ld1 {v20.8B}, [x8], x2
ld1 {v21.8B}, [x8], x2
urhadd v16.8B, v16.8B, v20.8B
urhadd v17.8B, v17.8B, v21.8B
.endif
subs w3, w3, #2
st1 {v16.8B}, [x0], x2
st1 {v17.8B}, [x0], x2
b.gt 3b
ret
4: ld1 {v4.8B, v5.8B}, [x1], x2
ld1 {v6.8B, v7.8B}, [x1], x2
ext v5.8B, v4.8B, v5.8B, #1
ext v7.8B, v6.8B, v7.8B, #1
prfm pldl1strm, [x1]
subs w3, w3, #2
umull v16.8H, v4.8B, v0.8B
umlal v16.8H, v5.8B, v1.8B
umull v17.8H, v6.8B, v0.8B
umlal v17.8H, v7.8B, v1.8B
prfm pldl1strm, [x1, x2]
.ifc \codec,h264
rshrn v16.8B, v16.8H, #6
rshrn v17.8B, v17.8H, #6
.else
add v16.8H, v16.8H, v22.8H
add v17.8H, v17.8H, v22.8H
shrn v16.8B, v16.8H, #6
shrn v17.8B, v17.8H, #6
.endif
.ifc \type,avg
ld1 {v20.8B}, [x8], x2
ld1 {v21.8B}, [x8], x2
urhadd v16.8B, v16.8B, v20.8B
urhadd v17.8B, v17.8B, v21.8B
.endif
st1 {v16.8B}, [x0], x2
st1 {v17.8B}, [x0], x2
b.gt 4b
ret
h264: avoid using uninitialized memory in NEON chroma mc Adapt commit 982b596ea6640bfe218a31f6c3fc542d9fe61c31 for the arm and aarch64 NEON asm. 5-10% faster on Cortex-A9.
2014-06-18 18:10:42 +03:00
5: ld1 {v4.8B}, [x1], x2
ld1 {v5.8B}, [x1], x2
prfm pldl1strm, [x1]
subs w3, w3, #2
umull v16.8H, v4.8B, v0.8B
umull v17.8H, v5.8B, v0.8B
prfm pldl1strm, [x1, x2]
.ifc \codec,h264
rshrn v16.8B, v16.8H, #6
rshrn v17.8B, v17.8H, #6
.else
add v16.8H, v16.8H, v22.8H
add v17.8H, v17.8H, v22.8H
shrn v16.8B, v16.8H, #6
shrn v17.8B, v17.8H, #6
.endif
.ifc \type,avg
ld1 {v20.8B}, [x8], x2
ld1 {v21.8B}, [x8], x2
urhadd v16.8B, v16.8B, v20.8B
urhadd v17.8B, v17.8B, v21.8B
.endif
st1 {v16.8B}, [x0], x2
st1 {v17.8B}, [x0], x2
b.gt 5b
ret
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
endfunc
.endm
h264chroma: Change type of stride parameters to ptrdiff_t This avoids SIMD-optimized functions having to sign-extend their stride argument manually to be able to do pointer arithmetic.
2016-09-01 21:41:01 +02:00
/* chroma_mc4(uint8_t *dst, uint8_t *src, ptrdiff_t stride, int h, int x, int y) */
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
.macro h264_chroma_mc4 type, codec=h264
function ff_\type\()_\codec\()_chroma_mc4_neon, export=1
.ifc \type,avg
mov x8, x0
.endif
prfm pldl1strm, [x1]
prfm pldl1strm, [x1, x2]
.ifc \codec,rv40
movrel x6, rv40bias
lsr w9, w5, #1
lsr w10, w4, #1
lsl w9, w9, #3
lsl w10, w10, #1
add w9, w9, w10
add x6, x6, w9, UXTW
ld1r {v22.8H}, [x6]
.endif
.ifc \codec,vc1
movi v22.8H, #28
.endif
mul w7, w4, w5
lsl w14, w5, #3
lsl w13, w4, #3
cmp w7, #0
sub w6, w14, w7
sub w12, w13, w7
sub w4, w7, w13
sub w4, w4, w14
add w4, w4, #64
b.eq 2f
dup v24.8B, w4
dup v25.8B, w12
ld1 {v4.8B}, [x1], x2
dup v26.8B, w6
dup v27.8B, w7
ext v5.8B, v4.8B, v5.8B, #1
trn1 v0.2S, v24.2S, v25.2S
trn1 v2.2S, v26.2S, v27.2S
trn1 v4.2S, v4.2S, v5.2S
1: ld1 {v6.8B}, [x1], x2
ext v7.8B, v6.8B, v7.8B, #1
trn1 v6.2S, v6.2S, v7.2S
umull v18.8H, v4.8B, v0.8B
umlal v18.8H, v6.8B, v2.8B
ld1 {v4.8B}, [x1], x2
ext v5.8B, v4.8B, v5.8B, #1
trn1 v4.2S, v4.2S, v5.2S
prfm pldl1strm, [x1]
umull v19.8H, v6.8B, v0.8B
umlal v19.8H, v4.8B, v2.8B
trn1 v30.2D, v18.2D, v19.2D
trn2 v31.2D, v18.2D, v19.2D
add v18.8H, v30.8H, v31.8H
.ifc \codec,h264
rshrn v16.8B, v18.8H, #6
.else
add v18.8H, v18.8H, v22.8H
shrn v16.8B, v18.8H, #6
.endif
subs w3, w3, #2
prfm pldl1strm, [x1, x2]
.ifc \type,avg
ld1 {v20.S}[0], [x8], x2
ld1 {v20.S}[1], [x8], x2
urhadd v16.8B, v16.8B, v20.8B
.endif
st1 {v16.S}[0], [x0], x2
st1 {v16.S}[1], [x0], x2
b.gt 1b
ret
h264: avoid using uninitialized memory in NEON chroma mc Adapt commit 982b596ea6640bfe218a31f6c3fc542d9fe61c31 for the arm and aarch64 NEON asm. 5-10% faster on Cortex-A9.
2014-06-18 18:10:42 +03:00
2: adds w12, w12, w6
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
dup v30.8B, w4
h264: avoid using uninitialized memory in NEON chroma mc Adapt commit 982b596ea6640bfe218a31f6c3fc542d9fe61c31 for the arm and aarch64 NEON asm. 5-10% faster on Cortex-A9.
2014-06-18 18:10:42 +03:00
b.eq 5f
tst w6, w6
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
dup v31.8B, w12
trn1 v0.2S, v30.2S, v31.2S
trn2 v1.2S, v30.2S, v31.2S
b.eq 4f
ext v1.8B, v0.8B, v1.8B, #4
ld1 {v4.S}[0], [x1], x2
3: ld1 {v4.S}[1], [x1], x2
umull v18.8H, v4.8B, v0.8B
ld1 {v4.S}[0], [x1], x2
umull v19.8H, v4.8B, v1.8B
trn1 v30.2D, v18.2D, v19.2D
trn2 v31.2D, v18.2D, v19.2D
add v18.8H, v30.8H, v31.8H
prfm pldl1strm, [x1]
.ifc \codec,h264
rshrn v16.8B, v18.8H, #6
.else
add v18.8H, v18.8H, v22.8H
shrn v16.8B, v18.8H, #6
.endif
.ifc \type,avg
ld1 {v20.S}[0], [x8], x2
ld1 {v20.S}[1], [x8], x2
urhadd v16.8B, v16.8B, v20.8B
.endif
subs w3, w3, #2
prfm pldl1strm, [x1, x2]
st1 {v16.S}[0], [x0], x2
st1 {v16.S}[1], [x0], x2
b.gt 3b
ret
4: ld1 {v4.8B}, [x1], x2
ld1 {v6.8B}, [x1], x2
ext v5.8B, v4.8B, v5.8B, #1
ext v7.8B, v6.8B, v7.8B, #1
trn1 v4.2S, v4.2S, v5.2S
trn1 v6.2S, v6.2S, v7.2S
umull v18.8H, v4.8B, v0.8B
umull v19.8H, v6.8B, v0.8B
subs w3, w3, #2
trn1 v30.2D, v18.2D, v19.2D
trn2 v31.2D, v18.2D, v19.2D
add v18.8H, v30.8H, v31.8H
prfm pldl1strm, [x1]
.ifc \codec,h264
rshrn v16.8B, v18.8H, #6
.else
add v18.8H, v18.8H, v22.8H
shrn v16.8B, v18.8H, #6
.endif
.ifc \type,avg
ld1 {v20.S}[0], [x8], x2
ld1 {v20.S}[1], [x8], x2
urhadd v16.8B, v16.8B, v20.8B
.endif
prfm pldl1strm, [x1]
st1 {v16.S}[0], [x0], x2
st1 {v16.S}[1], [x0], x2
b.gt 4b
ret
h264: avoid using uninitialized memory in NEON chroma mc Adapt commit 982b596ea6640bfe218a31f6c3fc542d9fe61c31 for the arm and aarch64 NEON asm. 5-10% faster on Cortex-A9.
2014-06-18 18:10:42 +03:00
5: ld1 {v4.S}[0], [x1], x2
ld1 {v4.S}[1], [x1], x2
umull v18.8H, v4.8B, v30.8B
subs w3, w3, #2
prfm pldl1strm, [x1]
.ifc \codec,h264
rshrn v16.8B, v18.8H, #6
.else
add v18.8H, v18.8H, v22.8H
shrn v16.8B, v18.8H, #6
.endif
.ifc \type,avg
ld1 {v20.S}[0], [x8], x2
ld1 {v20.S}[1], [x8], x2
urhadd v16.8B, v16.8B, v20.8B
.endif
prfm pldl1strm, [x1]
st1 {v16.S}[0], [x0], x2
st1 {v16.S}[1], [x0], x2
b.gt 5b
ret
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
endfunc
.endm
.macro h264_chroma_mc2 type
function ff_\type\()_h264_chroma_mc2_neon, export=1
prfm pldl1strm, [x1]
prfm pldl1strm, [x1, x2]
orr w7, w4, w5
cbz w7, 2f
mul w7, w4, w5
lsl w14, w5, #3
lsl w13, w4, #3
sub w6, w14, w7
sub w12, w13, w7
sub w4, w7, w13
sub w4, w4, w14
add w4, w4, #64
dup v0.8B, w4
dup v2.8B, w12
dup v1.8B, w6
dup v3.8B, w7
trn1 v0.4H, v0.4H, v2.4H
trn1 v1.4H, v1.4H, v3.4H
1:
ld1 {v4.S}[0], [x1], x2
ld1 {v4.S}[1], [x1], x2
rev64 v5.2S, v4.2S
ld1 {v5.S}[1], [x1]
ext v6.8B, v4.8B, v5.8B, #1
ext v7.8B, v5.8B, v4.8B, #1
trn1 v4.4H, v4.4H, v6.4H
trn1 v5.4H, v5.4H, v7.4H
umull v16.8H, v4.8B, v0.8B
umlal v16.8H, v5.8B, v1.8B
.ifc \type,avg
ld1 {v18.H}[0], [x0], x2
ld1 {v18.H}[2], [x0]
sub x0, x0, x2
.endif
rev64 v17.4S, v16.4S
add v16.8H, v16.8H, v17.8H
rshrn v16.8B, v16.8H, #6
.ifc \type,avg
urhadd v16.8B, v16.8B, v18.8B
.endif
st1 {v16.H}[0], [x0], x2
st1 {v16.H}[2], [x0], x2
subs w3, w3, #2
b.gt 1b
ret
2:
ld1 {v16.H}[0], [x1], x2
ld1 {v16.H}[1], [x1], x2
.ifc \type,avg
ld1 {v18.H}[0], [x0], x2
ld1 {v18.H}[1], [x0]
sub x0, x0, x2
urhadd v16.8B, v16.8B, v18.8B
.endif
st1 {v16.H}[0], [x0], x2
st1 {v16.H}[1], [x0], x2
subs w3, w3, #2
b.gt 2b
ret
endfunc
.endm
h264_chroma_mc8 put
h264_chroma_mc8 avg
h264_chroma_mc4 put
h264_chroma_mc4 avg
h264_chroma_mc2 put
h264_chroma_mc2 avg
#if CONFIG_RV40_DECODER
const rv40bias
.short 0, 16, 32, 16
.short 32, 28, 32, 28
.short 0, 32, 16, 32
.short 32, 28, 32, 28
endconst
h264_chroma_mc8 put, rv40
h264_chroma_mc8 avg, rv40
h264_chroma_mc4 put, rv40
h264_chroma_mc4 avg, rv40
#endif
avcodec: fix vc1dsp dependencies
2016-09-25 12:56:55 +02:00
#if CONFIG_VC1DSP
aarch64: h264 chroma motion compensation NEON optimizations Since RV40 and VC-1 use almost the same algorithm so optimizations for those two decoders are easy to do and included.
2013-12-10 22:16:08 +03:00
h264_chroma_mc8 put, vc1
h264_chroma_mc8 avg, vc1
h264_chroma_mc4 put, vc1
h264_chroma_mc4 avg, vc1
#endif
Reference in New Issue Copy Permalink