Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
83228 Commits
37 Branches
400 Tags
465 MiB
Tag: Branch: Tree: cda9a3e80b
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from 'cda9a3e80b'
${ noResults }
FFmpeg/libavcodec/aarch64/h264idct_neon.S

409 lines
16 KiB
Raw Normal View History Unescape

aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
/*
* Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
* Copyright (c) 2013 Janne Grunau <janne-libav@jannau.net>
*
Merge commit '8438b3f09f6b225d0886cc385117c38eb44ca0c1' * commit '8438b3f09f6b225d0886cc385117c38eb44ca0c1': aarch64: h264 idct NEON assembler optimizations Merged-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
* This file is part of FFmpeg.
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
*
Merge commit '8438b3f09f6b225d0886cc385117c38eb44ca0c1' * commit '8438b3f09f6b225d0886cc385117c38eb44ca0c1': aarch64: h264 idct NEON assembler optimizations Merged-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
* FFmpeg is free software; you can redistribute it and/or
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
* 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 '8438b3f09f6b225d0886cc385117c38eb44ca0c1' * commit '8438b3f09f6b225d0886cc385117c38eb44ca0c1': aarch64: h264 idct NEON assembler optimizations Merged-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
* FFmpeg is distributed in the hope that it will be useful,
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
* 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 '8438b3f09f6b225d0886cc385117c38eb44ca0c1' * commit '8438b3f09f6b225d0886cc385117c38eb44ca0c1': aarch64: h264 idct NEON assembler optimizations Merged-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
* License along with FFmpeg; if not, write to the Free Software
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/aarch64/asm.S"
#include "neon.S"
function ff_h264_idct_add_neon, export=1
ld1 {v0.4H, v1.4H, v2.4H, v3.4H}, [x1]
sxtw x2, w2
movi v30.8H, #0
add v4.4H, v0.4H, v2.4H
sshr v16.4H, v1.4H, #1
st1 {v30.8H}, [x1], #16
sshr v17.4H, v3.4H, #1
st1 {v30.8H}, [x1], #16
sub v5.4H, v0.4H, v2.4H
aarch64: Make transpose_4x4H do a regular transpose Previously, ff_h264_idct_add_neon (originally in the arm version) used a non-regular transpose in order to be able to use more instructions that deal with registers as 128 bit register pairs. The aarch64 translation doesn't do it to the same extent, but brought along the same structure since it was a straight translation. This reshuffles ff_h264_idct_add_neon, bringing it closer to the C implementation, making the transpose_4x4H macro do a regular transpose, usable for other algorithms as well. Previously, the third and fourth output from transpose_4x4H were swapped, and prior to cc29d96d5a, the same inputs as well. In addition to just swapping the outputs, also renumber the intermediate registers for better readability (making the register order match transpose_4x8B). This runs with the same number of cycles as before. Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
sub v6.4H, v16.4H, v3.4H
add v7.4H, v1.4H, v17.4H
add v0.4H, v4.4H, v7.4H
add v1.4H, v5.4H, v6.4H
sub v2.4H, v5.4H, v6.4H
sub v3.4H, v4.4H, v7.4H
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
transpose_4x4H v0, v1, v2, v3, v4, v5, v6, v7
aarch64: Make transpose_4x4H do a regular transpose Previously, ff_h264_idct_add_neon (originally in the arm version) used a non-regular transpose in order to be able to use more instructions that deal with registers as 128 bit register pairs. The aarch64 translation doesn't do it to the same extent, but brought along the same structure since it was a straight translation. This reshuffles ff_h264_idct_add_neon, bringing it closer to the C implementation, making the transpose_4x4H macro do a regular transpose, usable for other algorithms as well. Previously, the third and fourth output from transpose_4x4H were swapped, and prior to cc29d96d5a, the same inputs as well. In addition to just swapping the outputs, also renumber the intermediate registers for better readability (making the register order match transpose_4x8B). This runs with the same number of cycles as before. Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
add v4.4H, v0.4H, v2.4H
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
ld1 {v18.S}[0], [x0], x2
aarch64: Make transpose_4x4H do a regular transpose Previously, ff_h264_idct_add_neon (originally in the arm version) used a non-regular transpose in order to be able to use more instructions that deal with registers as 128 bit register pairs. The aarch64 translation doesn't do it to the same extent, but brought along the same structure since it was a straight translation. This reshuffles ff_h264_idct_add_neon, bringing it closer to the C implementation, making the transpose_4x4H macro do a regular transpose, usable for other algorithms as well. Previously, the third and fourth output from transpose_4x4H were swapped, and prior to cc29d96d5a, the same inputs as well. In addition to just swapping the outputs, also renumber the intermediate registers for better readability (making the register order match transpose_4x8B). This runs with the same number of cycles as before. Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
sshr v16.4H, v3.4H, #1
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
sshr v17.4H, v1.4H, #1
ld1 {v18.S}[1], [x0], x2
aarch64: Make transpose_4x4H do a regular transpose Previously, ff_h264_idct_add_neon (originally in the arm version) used a non-regular transpose in order to be able to use more instructions that deal with registers as 128 bit register pairs. The aarch64 translation doesn't do it to the same extent, but brought along the same structure since it was a straight translation. This reshuffles ff_h264_idct_add_neon, bringing it closer to the C implementation, making the transpose_4x4H macro do a regular transpose, usable for other algorithms as well. Previously, the third and fourth output from transpose_4x4H were swapped, and prior to cc29d96d5a, the same inputs as well. In addition to just swapping the outputs, also renumber the intermediate registers for better readability (making the register order match transpose_4x8B). This runs with the same number of cycles as before. Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
sub v5.4H, v0.4H, v2.4H
ld1 {v19.S}[1], [x0], x2
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
add v6.4H, v16.4H, v1.4H
ins v4.D[1], v5.D[0]
aarch64: Make transpose_4x4H do a regular transpose Previously, ff_h264_idct_add_neon (originally in the arm version) used a non-regular transpose in order to be able to use more instructions that deal with registers as 128 bit register pairs. The aarch64 translation doesn't do it to the same extent, but brought along the same structure since it was a straight translation. This reshuffles ff_h264_idct_add_neon, bringing it closer to the C implementation, making the transpose_4x4H macro do a regular transpose, usable for other algorithms as well. Previously, the third and fourth output from transpose_4x4H were swapped, and prior to cc29d96d5a, the same inputs as well. In addition to just swapping the outputs, also renumber the intermediate registers for better readability (making the register order match transpose_4x8B). This runs with the same number of cycles as before. Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
sub v7.4H, v17.4H, v3.4H
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
ld1 {v19.S}[0], [x0], x2
ins v6.D[1], v7.D[0]
sub x0, x0, x2, lsl #2
add v0.8H, v4.8H, v6.8H
sub v1.8H, v4.8H, v6.8H
srshr v0.8H, v0.8H, #6
srshr v1.8H, v1.8H, #6
uaddw v0.8H, v0.8H, v18.8B
uaddw v1.8H, v1.8H, v19.8B
sqxtun v0.8B, v0.8H
sqxtun v1.8B, v1.8H
st1 {v0.S}[0], [x0], x2
st1 {v0.S}[1], [x0], x2
aarch64: Make transpose_4x4H do a regular transpose Previously, ff_h264_idct_add_neon (originally in the arm version) used a non-regular transpose in order to be able to use more instructions that deal with registers as 128 bit register pairs. The aarch64 translation doesn't do it to the same extent, but brought along the same structure since it was a straight translation. This reshuffles ff_h264_idct_add_neon, bringing it closer to the C implementation, making the transpose_4x4H macro do a regular transpose, usable for other algorithms as well. Previously, the third and fourth output from transpose_4x4H were swapped, and prior to cc29d96d5a, the same inputs as well. In addition to just swapping the outputs, also renumber the intermediate registers for better readability (making the register order match transpose_4x8B). This runs with the same number of cycles as before. Signed-off-by: Martin Storsjö <martin@martin.st>
9 years ago
st1 {v1.S}[1], [x0], x2
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
st1 {v1.S}[0], [x0], x2
sub x1, x1, #32
ret
endfunc
function ff_h264_idct_dc_add_neon, export=1
sxtw x2, w2
mov w3, #0
ld1r {v2.8H}, [x1]
strh w3, [x1]
srshr v2.8H, v2.8H, #6
ld1 {v0.S}[0], [x0], x2
ld1 {v0.S}[1], [x0], x2
uaddw v3.8H, v2.8H, v0.8B
ld1 {v1.S}[0], [x0], x2
ld1 {v1.S}[1], [x0], x2
uaddw v4.8H, v2.8H, v1.8B
sqxtun v0.8B, v3.8H
sqxtun v1.8B, v4.8H
sub x0, x0, x2, lsl #2
st1 {v0.S}[0], [x0], x2
st1 {v0.S}[1], [x0], x2
st1 {v1.S}[0], [x0], x2
st1 {v1.S}[1], [x0], x2
ret
endfunc
function ff_h264_idct_add16_neon, export=1
mov x12, x30
mov x6, x0 // dest
mov x5, x1 // block_offset
mov x1, x2 // block
mov w9, w3 // stride
movrel x7, scan8
mov x10, #16
aarch64: use EXTERN_ASM consistently for exported symbols Based on e3fec3f095ab5ea08ee662942d98526aaf5e3635 for arm.
11 years ago
movrel x13, X(ff_h264_idct_dc_add_neon)
movrel x14, X(ff_h264_idct_add_neon)
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
1: mov w2, w9
ldrb w3, [x7], #1
ldrsw x0, [x5], #4
ldrb w3, [x4, w3, uxtw]
subs w3, w3, #1
b.lt 2f
ldrsh w3, [x1]
add x0, x0, x6
ccmp w3, #0, #4, eq
csel x15, x13, x14, ne
blr x15
2: subs x10, x10, #1
add x1, x1, #32
b.ne 1b
ret x12
endfunc
function ff_h264_idct_add16intra_neon, export=1
mov x12, x30
mov x6, x0 // dest
mov x5, x1 // block_offset
mov x1, x2 // block
mov w9, w3 // stride
movrel x7, scan8
mov x10, #16
aarch64: use EXTERN_ASM consistently for exported symbols Based on e3fec3f095ab5ea08ee662942d98526aaf5e3635 for arm.
11 years ago
movrel x13, X(ff_h264_idct_dc_add_neon)
movrel x14, X(ff_h264_idct_add_neon)
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
1: mov w2, w9
ldrb w3, [x7], #1
ldrsw x0, [x5], #4
ldrb w3, [x4, w3, uxtw]
add x0, x0, x6
cmp w3, #0
ldrsh w3, [x1]
csel x15, x13, x14, eq
ccmp w3, #0, #0, eq
b.eq 2f
blr x15
2: subs x10, x10, #1
add x1, x1, #32
b.ne 1b
ret x12
endfunc
function ff_h264_idct_add8_neon, export=1
sub sp, sp, #0x40
stp x19, x20, [sp]
mov x12, x30
ldp x6, x15, [x0] // dest[0], dest[1]
add x5, x1, #16*4 // block_offset
add x9, x2, #16*32 // block
mov w19, w3 // stride
aarch64: use EXTERN_ASM consistently for exported symbols Based on e3fec3f095ab5ea08ee662942d98526aaf5e3635 for arm.
11 years ago
movrel x13, X(ff_h264_idct_dc_add_neon)
movrel x14, X(ff_h264_idct_add_neon)
aarch64: h264idct: Use the offset parameter to movrel Signed-off-by: Martin Storsjö <martin@martin.st> (cherry picked from commit 6a62795d4051f435a9a2c59395d96913693922f8) Cherry pick Suggested-by: Martin Storsjö This should fix the build failure on macosx Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
8 years ago
movrel x7, scan8, 16
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
mov x10, #0
mov x11, #16
1: mov w2, w19
ldrb w3, [x7, x10] // scan8[i]
ldrsw x0, [x5, x10, lsl #2] // block_offset[i]
ldrb w3, [x4, w3, uxtw] // nnzc[ scan8[i] ]
add x0, x0, x6 // block_offset[i] + dst[j-1]
add x1, x9, x10, lsl #5 // block + i * 16
cmp w3, #0
ldrsh w3, [x1] // block[i*16]
csel x20, x13, x14, eq
ccmp w3, #0, #0, eq
b.eq 2f
blr x20
2: add x10, x10, #1
cmp x10, #4
csel x10, x11, x10, eq // mov x10, #16
csel x6, x15, x6, eq
cmp x10, #20
b.lt 1b
ldp x19, x20, [sp]
add sp, sp, #0x40
ret x12
endfunc
.macro idct8x8_cols pass
.if \pass == 0
va .req v18
vb .req v30
sshr v18.8H, v26.8H, #1
add v16.8H, v24.8H, v28.8H
ld1 {v30.8H, v31.8H}, [x1]
st1 {v19.8H}, [x1], #16
st1 {v19.8H}, [x1], #16
sub v17.8H, v24.8H, v28.8H
sshr v19.8H, v30.8H, #1
sub v18.8H, v18.8H, v30.8H
add v19.8H, v19.8H, v26.8H
.else
va .req v30
vb .req v18
sshr v30.8H, v26.8H, #1
sshr v19.8H, v18.8H, #1
add v16.8H, v24.8H, v28.8H
sub v17.8H, v24.8H, v28.8H
sub v30.8H, v30.8H, v18.8H
add v19.8H, v19.8H, v26.8H
.endif
add v26.8H, v17.8H, va.8H
sub v28.8H, v17.8H, va.8H
add v24.8H, v16.8H, v19.8H
sub vb.8H, v16.8H, v19.8H
sub v16.8H, v29.8H, v27.8H
add v17.8H, v31.8H, v25.8H
sub va.8H, v31.8H, v25.8H
add v19.8H, v29.8H, v27.8H
sub v16.8H, v16.8H, v31.8H
sub v17.8H, v17.8H, v27.8H
add va.8H, va.8H, v29.8H
add v19.8H, v19.8H, v25.8H
sshr v25.8H, v25.8H, #1
sshr v27.8H, v27.8H, #1
sshr v29.8H, v29.8H, #1
sshr v31.8H, v31.8H, #1
sub v16.8H, v16.8H, v31.8H
sub v17.8H, v17.8H, v27.8H
add va.8H, va.8H, v29.8H
add v19.8H, v19.8H, v25.8H
sshr v25.8H, v16.8H, #2
sshr v27.8H, v17.8H, #2
sshr v29.8H, va.8H, #2
sshr v31.8H, v19.8H, #2
sub v19.8H, v19.8H, v25.8H
sub va.8H, v27.8H, va.8H
add v17.8H, v17.8H, v29.8H
add v16.8H, v16.8H, v31.8H
.if \pass == 0
sub v31.8H, v24.8H, v19.8H
add v24.8H, v24.8H, v19.8H
add v25.8H, v26.8H, v18.8H
sub v18.8H, v26.8H, v18.8H
add v26.8H, v28.8H, v17.8H
add v27.8H, v30.8H, v16.8H
sub v29.8H, v28.8H, v17.8H
sub v28.8H, v30.8H, v16.8H
.else
sub v31.8H, v24.8H, v19.8H
add v24.8H, v24.8H, v19.8H
add v25.8H, v26.8H, v30.8H
sub v30.8H, v26.8H, v30.8H
add v26.8H, v28.8H, v17.8H
sub v29.8H, v28.8H, v17.8H
add v27.8H, v18.8H, v16.8H
sub v28.8H, v18.8H, v16.8H
.endif
.unreq va
.unreq vb
.endm
function ff_h264_idct8_add_neon, export=1
movi v19.8H, #0
ld1 {v24.8H, v25.8H}, [x1]
st1 {v19.8H}, [x1], #16
st1 {v19.8H}, [x1], #16
ld1 {v26.8H, v27.8H}, [x1]
st1 {v19.8H}, [x1], #16
st1 {v19.8H}, [x1], #16
ld1 {v28.8H, v29.8H}, [x1]
st1 {v19.8H}, [x1], #16
st1 {v19.8H}, [x1], #16
idct8x8_cols 0
transpose_8x8H v24, v25, v26, v27, v28, v29, v18, v31, v6, v7
idct8x8_cols 1
mov x3, x0
srshr v24.8H, v24.8H, #6
ld1 {v0.8B}, [x0], x2
srshr v25.8H, v25.8H, #6
ld1 {v1.8B}, [x0], x2
srshr v26.8H, v26.8H, #6
ld1 {v2.8B}, [x0], x2
srshr v27.8H, v27.8H, #6
ld1 {v3.8B}, [x0], x2
srshr v28.8H, v28.8H, #6
ld1 {v4.8B}, [x0], x2
srshr v29.8H, v29.8H, #6
ld1 {v5.8B}, [x0], x2
srshr v30.8H, v30.8H, #6
ld1 {v6.8B}, [x0], x2
srshr v31.8H, v31.8H, #6
ld1 {v7.8B}, [x0], x2
uaddw v24.8H, v24.8H, v0.8B
uaddw v25.8H, v25.8H, v1.8B
uaddw v26.8H, v26.8H, v2.8B
sqxtun v0.8B, v24.8H
uaddw v27.8H, v27.8H, v3.8B
sqxtun v1.8B, v25.8H
uaddw v28.8H, v28.8H, v4.8B
sqxtun v2.8B, v26.8H
st1 {v0.8B}, [x3], x2
uaddw v29.8H, v29.8H, v5.8B
sqxtun v3.8B, v27.8H
st1 {v1.8B}, [x3], x2
uaddw v30.8H, v30.8H, v6.8B
sqxtun v4.8B, v28.8H
st1 {v2.8B}, [x3], x2
uaddw v31.8H, v31.8H, v7.8B
sqxtun v5.8B, v29.8H
st1 {v3.8B}, [x3], x2
sqxtun v6.8B, v30.8H
sqxtun v7.8B, v31.8H
st1 {v4.8B}, [x3], x2
st1 {v5.8B}, [x3], x2
st1 {v6.8B}, [x3], x2
st1 {v7.8B}, [x3], x2
sub x1, x1, #128
ret
endfunc
function ff_h264_idct8_dc_add_neon, export=1
mov w3, #0
sxtw x2, w2
ld1r {v31.8H}, [x1]
strh w3, [x1]
ld1 {v0.8B}, [x0], x2
srshr v31.8H, v31.8H, #6
ld1 {v1.8B}, [x0], x2
ld1 {v2.8B}, [x0], x2
uaddw v24.8H, v31.8H, v0.8B
ld1 {v3.8B}, [x0], x2
uaddw v25.8H, v31.8H, v1.8B
ld1 {v4.8B}, [x0], x2
uaddw v26.8H, v31.8H, v2.8B
ld1 {v5.8B}, [x0], x2
uaddw v27.8H, v31.8H, v3.8B
ld1 {v6.8B}, [x0], x2
uaddw v28.8H, v31.8H, v4.8B
ld1 {v7.8B}, [x0], x2
uaddw v29.8H, v31.8H, v5.8B
uaddw v30.8H, v31.8H, v6.8B
uaddw v31.8H, v31.8H, v7.8B
sqxtun v0.8B, v24.8H
sqxtun v1.8B, v25.8H
sqxtun v2.8B, v26.8H
sqxtun v3.8B, v27.8H
sub x0, x0, x2, lsl #3
st1 {v0.8B}, [x0], x2
sqxtun v4.8B, v28.8H
st1 {v1.8B}, [x0], x2
sqxtun v5.8B, v29.8H
st1 {v2.8B}, [x0], x2
sqxtun v6.8B, v30.8H
st1 {v3.8B}, [x0], x2
sqxtun v7.8B, v31.8H
st1 {v4.8B}, [x0], x2
st1 {v5.8B}, [x0], x2
st1 {v6.8B}, [x0], x2
st1 {v7.8B}, [x0], x2
ret
endfunc
function ff_h264_idct8_add4_neon, export=1
mov x12, x30
mov x6, x0
mov x5, x1
mov x1, x2
mov w2, w3
movrel x7, scan8
mov w10, #16
aarch64: use EXTERN_ASM consistently for exported symbols Based on e3fec3f095ab5ea08ee662942d98526aaf5e3635 for arm.
11 years ago
movrel x13, X(ff_h264_idct8_dc_add_neon)
movrel x14, X(ff_h264_idct8_add_neon)
aarch64: h264 idct NEON assembler optimizations Ported from ARMv7 NEON.
11 years ago
1: ldrb w9, [x7], #4
ldrsw x0, [x5], #16
ldrb w9, [x4, w9, UXTW]
subs w9, w9, #1
b.lt 2f
ldrsh w11, [x1]
add x0, x6, x0
ccmp w11, #0, #4, eq
csel x15, x13, x14, ne
blr x15
2: subs w10, w10, #4
add x1, x1, #128
b.ne 1b
ret x12
endfunc
const scan8
.byte 4+ 1*8, 5+ 1*8, 4+ 2*8, 5+ 2*8
.byte 6+ 1*8, 7+ 1*8, 6+ 2*8, 7+ 2*8
.byte 4+ 3*8, 5+ 3*8, 4+ 4*8, 5+ 4*8
.byte 6+ 3*8, 7+ 3*8, 6+ 4*8, 7+ 4*8
.byte 4+ 6*8, 5+ 6*8, 4+ 7*8, 5+ 7*8
.byte 6+ 6*8, 7+ 6*8, 6+ 7*8, 7+ 7*8
.byte 4+ 8*8, 5+ 8*8, 4+ 9*8, 5+ 9*8
.byte 6+ 8*8, 7+ 8*8, 6+ 9*8, 7+ 9*8
.byte 4+11*8, 5+11*8, 4+12*8, 5+12*8
.byte 6+11*8, 7+11*8, 6+12*8, 7+12*8
.byte 4+13*8, 5+13*8, 4+14*8, 5+14*8
.byte 6+13*8, 7+13*8, 6+14*8, 7+14*8
endconst