2008-12-16 00:12:41 +02:00
|
|
|
/*
|
|
|
|
|
* Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
|
|
|
|
|
*
|
2011-03-18 19:35:10 +02:00
|
|
|
* This file is part of Libav.
|
2008-12-16 00:12:41 +02:00
|
|
|
*
|
2011-03-18 19:35:10 +02:00
|
|
|
* Libav is free software; you can redistribute it and/or
|
2008-12-16 00:12:41 +02:00
|
|
|
* 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.
|
|
|
|
|
*
|
2011-03-18 19:35:10 +02:00
|
|
|
* Libav is distributed in the hope that it will be useful,
|
2008-12-16 00:12:41 +02:00
|
|
|
* 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
|
2011-03-18 19:35:10 +02:00
|
|
|
* License along with Libav; if not, write to the Free Software
|
2008-12-16 00:12:41 +02:00
|
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
|
|
|
*/
|
|
|
|
|
|
2012-05-21 22:46:23 +03:00
|
|
|
#include "libavutil/arm/asm.S"
|
2011-11-29 16:38:10 +03:00
|
|
|
#include "neon.S"
|
2008-12-16 00:12:47 +02:00
|
|
|
|
2008-12-16 00:12:44 +02:00
|
|
|
/* H.264 loop filter */
|
|
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro h264_loop_filter_start
|
|
|
|
|
ldr r12, [sp]
|
2008-12-16 00:12:44 +02:00
|
|
|
tst r2, r2
|
2011-12-02 19:45:00 +03:00
|
|
|
ldr r12, [r12]
|
2011-06-14 13:29:48 +03:00
|
|
|
it ne
|
2008-12-16 00:12:44 +02:00
|
|
|
tstne r3, r3
|
2011-12-02 19:45:00 +03:00
|
|
|
vmov.32 d24[0], r12
|
|
|
|
|
and r12, r12, r12, lsl #16
|
2011-06-14 13:29:48 +03:00
|
|
|
it eq
|
2008-12-16 00:12:44 +02:00
|
|
|
bxeq lr
|
2011-12-02 19:45:00 +03:00
|
|
|
ands r12, r12, r12, lsl #8
|
2011-06-14 13:29:48 +03:00
|
|
|
it lt
|
2008-12-16 00:12:44 +02:00
|
|
|
bxlt lr
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2008-12-16 00:12:44 +02:00
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro h264_loop_filter_luma
|
2008-12-16 00:12:44 +02:00
|
|
|
vdup.8 q11, r2 @ alpha
|
|
|
|
|
vmovl.u8 q12, d24
|
|
|
|
|
vabd.u8 q6, q8, q0 @ abs(p0 - q0)
|
|
|
|
|
vmovl.u16 q12, d24
|
|
|
|
|
vabd.u8 q14, q9, q8 @ abs(p1 - p0)
|
|
|
|
|
vsli.16 q12, q12, #8
|
|
|
|
|
vabd.u8 q15, q1, q0 @ abs(q1 - q0)
|
|
|
|
|
vsli.32 q12, q12, #16
|
|
|
|
|
vclt.u8 q6, q6, q11 @ < alpha
|
|
|
|
|
vdup.8 q11, r3 @ beta
|
|
|
|
|
vclt.s8 q7, q12, #0
|
|
|
|
|
vclt.u8 q14, q14, q11 @ < beta
|
|
|
|
|
vclt.u8 q15, q15, q11 @ < beta
|
|
|
|
|
vbic q6, q6, q7
|
|
|
|
|
vabd.u8 q4, q10, q8 @ abs(p2 - p0)
|
|
|
|
|
vand q6, q6, q14
|
|
|
|
|
vabd.u8 q5, q2, q0 @ abs(q2 - q0)
|
|
|
|
|
vclt.u8 q4, q4, q11 @ < beta
|
|
|
|
|
vand q6, q6, q15
|
|
|
|
|
vclt.u8 q5, q5, q11 @ < beta
|
|
|
|
|
vand q4, q4, q6
|
|
|
|
|
vand q5, q5, q6
|
|
|
|
|
vand q12, q12, q6
|
|
|
|
|
vrhadd.u8 q14, q8, q0
|
|
|
|
|
vsub.i8 q6, q12, q4
|
|
|
|
|
vqadd.u8 q7, q9, q12
|
|
|
|
|
vhadd.u8 q10, q10, q14
|
|
|
|
|
vsub.i8 q6, q6, q5
|
|
|
|
|
vhadd.u8 q14, q2, q14
|
|
|
|
|
vmin.u8 q7, q7, q10
|
|
|
|
|
vqsub.u8 q11, q9, q12
|
|
|
|
|
vqadd.u8 q2, q1, q12
|
|
|
|
|
vmax.u8 q7, q7, q11
|
|
|
|
|
vqsub.u8 q11, q1, q12
|
|
|
|
|
vmin.u8 q14, q2, q14
|
|
|
|
|
vmovl.u8 q2, d0
|
|
|
|
|
vmax.u8 q14, q14, q11
|
|
|
|
|
vmovl.u8 q10, d1
|
|
|
|
|
vsubw.u8 q2, q2, d16
|
|
|
|
|
vsubw.u8 q10, q10, d17
|
|
|
|
|
vshl.i16 q2, q2, #2
|
|
|
|
|
vshl.i16 q10, q10, #2
|
|
|
|
|
vaddw.u8 q2, q2, d18
|
|
|
|
|
vaddw.u8 q10, q10, d19
|
|
|
|
|
vsubw.u8 q2, q2, d2
|
|
|
|
|
vsubw.u8 q10, q10, d3
|
|
|
|
|
vrshrn.i16 d4, q2, #3
|
|
|
|
|
vrshrn.i16 d5, q10, #3
|
|
|
|
|
vbsl q4, q7, q9
|
|
|
|
|
vbsl q5, q14, q1
|
|
|
|
|
vneg.s8 q7, q6
|
|
|
|
|
vmovl.u8 q14, d16
|
|
|
|
|
vmin.s8 q2, q2, q6
|
|
|
|
|
vmovl.u8 q6, d17
|
|
|
|
|
vmax.s8 q2, q2, q7
|
|
|
|
|
vmovl.u8 q11, d0
|
|
|
|
|
vmovl.u8 q12, d1
|
|
|
|
|
vaddw.s8 q14, q14, d4
|
|
|
|
|
vaddw.s8 q6, q6, d5
|
|
|
|
|
vsubw.s8 q11, q11, d4
|
|
|
|
|
vsubw.s8 q12, q12, d5
|
|
|
|
|
vqmovun.s16 d16, q14
|
|
|
|
|
vqmovun.s16 d17, q6
|
|
|
|
|
vqmovun.s16 d0, q11
|
|
|
|
|
vqmovun.s16 d1, q12
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
function ff_h264_v_loop_filter_luma_neon, export=1
|
|
|
|
|
h264_loop_filter_start
|
|
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
vld1.8 {d0, d1}, [r0,:128], r1
|
|
|
|
|
vld1.8 {d2, d3}, [r0,:128], r1
|
|
|
|
|
vld1.8 {d4, d5}, [r0,:128], r1
|
2008-12-16 00:12:44 +02:00
|
|
|
sub r0, r0, r1, lsl #2
|
|
|
|
|
sub r0, r0, r1, lsl #1
|
2011-12-02 19:45:00 +03:00
|
|
|
vld1.8 {d20,d21}, [r0,:128], r1
|
|
|
|
|
vld1.8 {d18,d19}, [r0,:128], r1
|
|
|
|
|
vld1.8 {d16,d17}, [r0,:128], r1
|
2008-12-16 00:12:44 +02:00
|
|
|
|
2011-06-14 13:29:48 +03:00
|
|
|
vpush {d8-d15}
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
h264_loop_filter_luma
|
|
|
|
|
|
|
|
|
|
sub r0, r0, r1, lsl #1
|
2011-12-02 19:45:00 +03:00
|
|
|
vst1.8 {d8, d9}, [r0,:128], r1
|
|
|
|
|
vst1.8 {d16,d17}, [r0,:128], r1
|
|
|
|
|
vst1.8 {d0, d1}, [r0,:128], r1
|
|
|
|
|
vst1.8 {d10,d11}, [r0,:128]
|
2008-12-16 00:12:44 +02:00
|
|
|
|
2011-06-14 13:29:48 +03:00
|
|
|
vpop {d8-d15}
|
2008-12-16 00:12:44 +02:00
|
|
|
bx lr
|
2010-03-09 18:17:56 +02:00
|
|
|
endfunc
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
function ff_h264_h_loop_filter_luma_neon, export=1
|
|
|
|
|
h264_loop_filter_start
|
|
|
|
|
|
|
|
|
|
sub r0, r0, #4
|
2011-12-02 19:45:00 +03:00
|
|
|
vld1.8 {d6}, [r0], r1
|
|
|
|
|
vld1.8 {d20}, [r0], r1
|
|
|
|
|
vld1.8 {d18}, [r0], r1
|
|
|
|
|
vld1.8 {d16}, [r0], r1
|
|
|
|
|
vld1.8 {d0}, [r0], r1
|
|
|
|
|
vld1.8 {d2}, [r0], r1
|
|
|
|
|
vld1.8 {d4}, [r0], r1
|
|
|
|
|
vld1.8 {d26}, [r0], r1
|
|
|
|
|
vld1.8 {d7}, [r0], r1
|
|
|
|
|
vld1.8 {d21}, [r0], r1
|
|
|
|
|
vld1.8 {d19}, [r0], r1
|
|
|
|
|
vld1.8 {d17}, [r0], r1
|
|
|
|
|
vld1.8 {d1}, [r0], r1
|
|
|
|
|
vld1.8 {d3}, [r0], r1
|
|
|
|
|
vld1.8 {d5}, [r0], r1
|
|
|
|
|
vld1.8 {d27}, [r0], r1
|
2008-12-16 00:12:44 +02:00
|
|
|
|
2008-12-16 00:12:47 +02:00
|
|
|
transpose_8x8 q3, q10, q9, q8, q0, q1, q2, q13
|
2008-12-16 00:12:44 +02:00
|
|
|
|
2011-06-14 13:29:48 +03:00
|
|
|
vpush {d8-d15}
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
h264_loop_filter_luma
|
|
|
|
|
|
2009-06-18 01:33:04 +03:00
|
|
|
transpose_4x4 q4, q8, q0, q5
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
sub r0, r0, r1, lsl #4
|
2009-06-18 01:33:04 +03:00
|
|
|
add r0, r0, #2
|
|
|
|
|
vst1.32 {d8[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d16[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d0[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d10[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d8[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d16[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d0[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d10[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d9[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d17[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d1[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d11[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d9[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d17[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d1[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d11[1]}, [r0], r1
|
2008-12-16 00:12:44 +02:00
|
|
|
|
2011-06-14 13:29:48 +03:00
|
|
|
vpop {d8-d15}
|
2008-12-16 00:12:44 +02:00
|
|
|
bx lr
|
2010-03-09 18:17:56 +02:00
|
|
|
endfunc
|
2008-12-16 00:12:44 +02:00
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro h264_loop_filter_chroma
|
2008-12-16 00:12:44 +02:00
|
|
|
vdup.8 d22, r2 @ alpha
|
|
|
|
|
vmovl.u8 q12, d24
|
|
|
|
|
vabd.u8 d26, d16, d0 @ abs(p0 - q0)
|
|
|
|
|
vmovl.u8 q2, d0
|
|
|
|
|
vabd.u8 d28, d18, d16 @ abs(p1 - p0)
|
|
|
|
|
vsubw.u8 q2, q2, d16
|
|
|
|
|
vsli.16 d24, d24, #8
|
|
|
|
|
vshl.i16 q2, q2, #2
|
|
|
|
|
vabd.u8 d30, d2, d0 @ abs(q1 - q0)
|
|
|
|
|
vaddw.u8 q2, q2, d18
|
|
|
|
|
vclt.u8 d26, d26, d22 @ < alpha
|
|
|
|
|
vsubw.u8 q2, q2, d2
|
|
|
|
|
vdup.8 d22, r3 @ beta
|
|
|
|
|
vrshrn.i16 d4, q2, #3
|
|
|
|
|
vclt.u8 d28, d28, d22 @ < beta
|
|
|
|
|
vclt.u8 d30, d30, d22 @ < beta
|
|
|
|
|
vmin.s8 d4, d4, d24
|
2010-07-16 00:59:27 +03:00
|
|
|
vneg.s8 d25, d24
|
|
|
|
|
vand d26, d26, d28
|
2008-12-16 00:12:44 +02:00
|
|
|
vmax.s8 d4, d4, d25
|
2010-07-16 00:59:27 +03:00
|
|
|
vand d26, d26, d30
|
2008-12-16 00:12:44 +02:00
|
|
|
vmovl.u8 q11, d0
|
2010-07-16 00:59:27 +03:00
|
|
|
vand d4, d4, d26
|
|
|
|
|
vmovl.u8 q14, d16
|
2008-12-16 00:12:44 +02:00
|
|
|
vaddw.s8 q14, q14, d4
|
|
|
|
|
vsubw.s8 q11, q11, d4
|
|
|
|
|
vqmovun.s16 d16, q14
|
|
|
|
|
vqmovun.s16 d0, q11
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
function ff_h264_v_loop_filter_chroma_neon, export=1
|
|
|
|
|
h264_loop_filter_start
|
|
|
|
|
|
|
|
|
|
sub r0, r0, r1, lsl #1
|
2011-12-02 19:45:00 +03:00
|
|
|
vld1.8 {d18}, [r0,:64], r1
|
|
|
|
|
vld1.8 {d16}, [r0,:64], r1
|
|
|
|
|
vld1.8 {d0}, [r0,:64], r1
|
|
|
|
|
vld1.8 {d2}, [r0,:64]
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
h264_loop_filter_chroma
|
|
|
|
|
|
|
|
|
|
sub r0, r0, r1, lsl #1
|
2011-12-02 19:45:00 +03:00
|
|
|
vst1.8 {d16}, [r0,:64], r1
|
|
|
|
|
vst1.8 {d0}, [r0,:64], r1
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
bx lr
|
2010-03-09 18:17:56 +02:00
|
|
|
endfunc
|
2008-12-16 00:12:44 +02:00
|
|
|
|
|
|
|
|
function ff_h264_h_loop_filter_chroma_neon, export=1
|
|
|
|
|
h264_loop_filter_start
|
|
|
|
|
|
|
|
|
|
sub r0, r0, #2
|
|
|
|
|
vld1.32 {d18[0]}, [r0], r1
|
|
|
|
|
vld1.32 {d16[0]}, [r0], r1
|
|
|
|
|
vld1.32 {d0[0]}, [r0], r1
|
|
|
|
|
vld1.32 {d2[0]}, [r0], r1
|
|
|
|
|
vld1.32 {d18[1]}, [r0], r1
|
|
|
|
|
vld1.32 {d16[1]}, [r0], r1
|
|
|
|
|
vld1.32 {d0[1]}, [r0], r1
|
|
|
|
|
vld1.32 {d2[1]}, [r0], r1
|
|
|
|
|
|
|
|
|
|
vtrn.16 d18, d0
|
|
|
|
|
vtrn.16 d16, d2
|
|
|
|
|
vtrn.8 d18, d16
|
|
|
|
|
vtrn.8 d0, d2
|
|
|
|
|
|
|
|
|
|
h264_loop_filter_chroma
|
|
|
|
|
|
|
|
|
|
vtrn.16 d18, d0
|
|
|
|
|
vtrn.16 d16, d2
|
|
|
|
|
vtrn.8 d18, d16
|
|
|
|
|
vtrn.8 d0, d2
|
|
|
|
|
|
|
|
|
|
sub r0, r0, r1, lsl #3
|
|
|
|
|
vst1.32 {d18[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d16[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d0[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d2[0]}, [r0], r1
|
|
|
|
|
vst1.32 {d18[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d16[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d0[1]}, [r0], r1
|
|
|
|
|
vst1.32 {d2[1]}, [r0], r1
|
|
|
|
|
|
|
|
|
|
bx lr
|
2010-03-09 18:17:56 +02:00
|
|
|
endfunc
|
2008-12-16 00:12:47 +02:00
|
|
|
|
2009-01-25 15:04:41 +02:00
|
|
|
@ Biweighted prediction
|
|
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro biweight_16 macs, macd
|
2009-01-25 15:04:41 +02:00
|
|
|
vdup.8 d0, r4
|
|
|
|
|
vdup.8 d1, r5
|
|
|
|
|
vmov q2, q8
|
|
|
|
|
vmov q3, q8
|
2011-10-21 10:00:39 +03:00
|
|
|
1: subs r3, r3, #2
|
2009-01-25 15:04:41 +02:00
|
|
|
vld1.8 {d20-d21},[r0,:128], r2
|
|
|
|
|
\macd q2, d0, d20
|
|
|
|
|
pld [r0]
|
|
|
|
|
\macd q3, d0, d21
|
|
|
|
|
vld1.8 {d22-d23},[r1,:128], r2
|
|
|
|
|
\macs q2, d1, d22
|
|
|
|
|
pld [r1]
|
|
|
|
|
\macs q3, d1, d23
|
|
|
|
|
vmov q12, q8
|
|
|
|
|
vld1.8 {d28-d29},[r0,:128], r2
|
|
|
|
|
vmov q13, q8
|
|
|
|
|
\macd q12, d0, d28
|
|
|
|
|
pld [r0]
|
|
|
|
|
\macd q13, d0, d29
|
|
|
|
|
vld1.8 {d30-d31},[r1,:128], r2
|
|
|
|
|
\macs q12, d1, d30
|
|
|
|
|
pld [r1]
|
|
|
|
|
\macs q13, d1, d31
|
|
|
|
|
vshl.s16 q2, q2, q9
|
|
|
|
|
vshl.s16 q3, q3, q9
|
|
|
|
|
vqmovun.s16 d4, q2
|
|
|
|
|
vqmovun.s16 d5, q3
|
|
|
|
|
vshl.s16 q12, q12, q9
|
|
|
|
|
vshl.s16 q13, q13, q9
|
|
|
|
|
vqmovun.s16 d24, q12
|
|
|
|
|
vqmovun.s16 d25, q13
|
|
|
|
|
vmov q3, q8
|
|
|
|
|
vst1.8 {d4- d5}, [r6,:128], r2
|
|
|
|
|
vmov q2, q8
|
|
|
|
|
vst1.8 {d24-d25},[r6,:128], r2
|
|
|
|
|
bne 1b
|
|
|
|
|
pop {r4-r6, pc}
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2009-01-25 15:04:41 +02:00
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro biweight_8 macs, macd
|
2009-01-25 15:04:41 +02:00
|
|
|
vdup.8 d0, r4
|
|
|
|
|
vdup.8 d1, r5
|
|
|
|
|
vmov q1, q8
|
|
|
|
|
vmov q10, q8
|
2011-10-21 10:00:39 +03:00
|
|
|
1: subs r3, r3, #2
|
2009-01-25 15:04:41 +02:00
|
|
|
vld1.8 {d4},[r0,:64], r2
|
|
|
|
|
\macd q1, d0, d4
|
|
|
|
|
pld [r0]
|
|
|
|
|
vld1.8 {d5},[r1,:64], r2
|
|
|
|
|
\macs q1, d1, d5
|
|
|
|
|
pld [r1]
|
|
|
|
|
vld1.8 {d6},[r0,:64], r2
|
|
|
|
|
\macd q10, d0, d6
|
|
|
|
|
pld [r0]
|
|
|
|
|
vld1.8 {d7},[r1,:64], r2
|
|
|
|
|
\macs q10, d1, d7
|
|
|
|
|
pld [r1]
|
|
|
|
|
vshl.s16 q1, q1, q9
|
|
|
|
|
vqmovun.s16 d2, q1
|
|
|
|
|
vshl.s16 q10, q10, q9
|
|
|
|
|
vqmovun.s16 d4, q10
|
|
|
|
|
vmov q10, q8
|
|
|
|
|
vst1.8 {d2},[r6,:64], r2
|
|
|
|
|
vmov q1, q8
|
|
|
|
|
vst1.8 {d4},[r6,:64], r2
|
|
|
|
|
bne 1b
|
|
|
|
|
pop {r4-r6, pc}
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2009-01-25 15:04:41 +02:00
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro biweight_4 macs, macd
|
2009-01-25 15:04:41 +02:00
|
|
|
vdup.8 d0, r4
|
|
|
|
|
vdup.8 d1, r5
|
|
|
|
|
vmov q1, q8
|
|
|
|
|
vmov q10, q8
|
2011-10-21 10:00:39 +03:00
|
|
|
1: subs r3, r3, #4
|
2009-01-25 15:04:41 +02:00
|
|
|
vld1.32 {d4[0]},[r0,:32], r2
|
|
|
|
|
vld1.32 {d4[1]},[r0,:32], r2
|
|
|
|
|
\macd q1, d0, d4
|
|
|
|
|
pld [r0]
|
|
|
|
|
vld1.32 {d5[0]},[r1,:32], r2
|
|
|
|
|
vld1.32 {d5[1]},[r1,:32], r2
|
|
|
|
|
\macs q1, d1, d5
|
|
|
|
|
pld [r1]
|
|
|
|
|
blt 2f
|
|
|
|
|
vld1.32 {d6[0]},[r0,:32], r2
|
|
|
|
|
vld1.32 {d6[1]},[r0,:32], r2
|
|
|
|
|
\macd q10, d0, d6
|
|
|
|
|
pld [r0]
|
|
|
|
|
vld1.32 {d7[0]},[r1,:32], r2
|
|
|
|
|
vld1.32 {d7[1]},[r1,:32], r2
|
|
|
|
|
\macs q10, d1, d7
|
|
|
|
|
pld [r1]
|
|
|
|
|
vshl.s16 q1, q1, q9
|
|
|
|
|
vqmovun.s16 d2, q1
|
|
|
|
|
vshl.s16 q10, q10, q9
|
|
|
|
|
vqmovun.s16 d4, q10
|
|
|
|
|
vmov q10, q8
|
|
|
|
|
vst1.32 {d2[0]},[r6,:32], r2
|
|
|
|
|
vst1.32 {d2[1]},[r6,:32], r2
|
|
|
|
|
vmov q1, q8
|
|
|
|
|
vst1.32 {d4[0]},[r6,:32], r2
|
|
|
|
|
vst1.32 {d4[1]},[r6,:32], r2
|
|
|
|
|
bne 1b
|
|
|
|
|
pop {r4-r6, pc}
|
|
|
|
|
2: vshl.s16 q1, q1, q9
|
|
|
|
|
vqmovun.s16 d2, q1
|
|
|
|
|
vst1.32 {d2[0]},[r6,:32], r2
|
|
|
|
|
vst1.32 {d2[1]},[r6,:32], r2
|
|
|
|
|
pop {r4-r6, pc}
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2009-01-25 15:04:41 +02:00
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro biweight_func w
|
2011-10-21 10:00:39 +03:00
|
|
|
function ff_biweight_h264_pixels_\w\()_neon, export=1
|
2009-01-25 15:04:41 +02:00
|
|
|
push {r4-r6, lr}
|
2011-10-21 10:00:39 +03:00
|
|
|
ldr r12, [sp, #16]
|
|
|
|
|
add r4, sp, #20
|
2009-01-25 15:04:41 +02:00
|
|
|
ldm r4, {r4-r6}
|
|
|
|
|
lsr lr, r4, #31
|
|
|
|
|
add r6, r6, #1
|
|
|
|
|
eors lr, lr, r5, lsr #30
|
|
|
|
|
orr r6, r6, #1
|
2011-10-21 10:00:39 +03:00
|
|
|
vdup.16 q9, r12
|
|
|
|
|
lsl r6, r6, r12
|
2009-01-25 15:04:41 +02:00
|
|
|
vmvn q9, q9
|
|
|
|
|
vdup.16 q8, r6
|
|
|
|
|
mov r6, r0
|
|
|
|
|
beq 10f
|
|
|
|
|
subs lr, lr, #1
|
|
|
|
|
beq 20f
|
|
|
|
|
subs lr, lr, #1
|
|
|
|
|
beq 30f
|
|
|
|
|
b 40f
|
|
|
|
|
10: biweight_\w vmlal.u8, vmlal.u8
|
|
|
|
|
20: rsb r4, r4, #0
|
|
|
|
|
biweight_\w vmlal.u8, vmlsl.u8
|
|
|
|
|
30: rsb r4, r4, #0
|
|
|
|
|
rsb r5, r5, #0
|
|
|
|
|
biweight_\w vmlsl.u8, vmlsl.u8
|
|
|
|
|
40: rsb r5, r5, #0
|
|
|
|
|
biweight_\w vmlsl.u8, vmlal.u8
|
2010-03-09 18:17:56 +02:00
|
|
|
endfunc
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2009-01-25 15:04:41 +02:00
|
|
|
|
|
|
|
|
biweight_func 16
|
|
|
|
|
biweight_func 8
|
|
|
|
|
biweight_func 4
|
2009-01-25 15:04:45 +02:00
|
|
|
|
|
|
|
|
@ Weighted prediction
|
|
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro weight_16 add
|
2011-10-21 10:00:39 +03:00
|
|
|
vdup.8 d0, r12
|
|
|
|
|
1: subs r2, r2, #2
|
2009-01-25 15:04:45 +02:00
|
|
|
vld1.8 {d20-d21},[r0,:128], r1
|
2009-02-28 15:48:54 +02:00
|
|
|
vmull.u8 q2, d0, d20
|
2009-01-25 15:04:45 +02:00
|
|
|
pld [r0]
|
2009-02-28 15:48:54 +02:00
|
|
|
vmull.u8 q3, d0, d21
|
2009-01-25 15:04:45 +02:00
|
|
|
vld1.8 {d28-d29},[r0,:128], r1
|
2009-02-28 15:48:54 +02:00
|
|
|
vmull.u8 q12, d0, d28
|
2009-01-25 15:04:45 +02:00
|
|
|
pld [r0]
|
2009-02-28 15:48:54 +02:00
|
|
|
vmull.u8 q13, d0, d29
|
|
|
|
|
\add q2, q8, q2
|
|
|
|
|
vrshl.s16 q2, q2, q9
|
|
|
|
|
\add q3, q8, q3
|
|
|
|
|
vrshl.s16 q3, q3, q9
|
2009-01-25 15:04:45 +02:00
|
|
|
vqmovun.s16 d4, q2
|
|
|
|
|
vqmovun.s16 d5, q3
|
2009-02-28 15:48:54 +02:00
|
|
|
\add q12, q8, q12
|
|
|
|
|
vrshl.s16 q12, q12, q9
|
|
|
|
|
\add q13, q8, q13
|
|
|
|
|
vrshl.s16 q13, q13, q9
|
2009-01-25 15:04:45 +02:00
|
|
|
vqmovun.s16 d24, q12
|
|
|
|
|
vqmovun.s16 d25, q13
|
|
|
|
|
vst1.8 {d4- d5}, [r4,:128], r1
|
|
|
|
|
vst1.8 {d24-d25},[r4,:128], r1
|
|
|
|
|
bne 1b
|
|
|
|
|
pop {r4, pc}
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2009-01-25 15:04:45 +02:00
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro weight_8 add
|
2011-10-21 10:00:39 +03:00
|
|
|
vdup.8 d0, r12
|
|
|
|
|
1: subs r2, r2, #2
|
2009-01-25 15:04:45 +02:00
|
|
|
vld1.8 {d4},[r0,:64], r1
|
2009-02-28 15:48:54 +02:00
|
|
|
vmull.u8 q1, d0, d4
|
2009-01-25 15:04:45 +02:00
|
|
|
pld [r0]
|
|
|
|
|
vld1.8 {d6},[r0,:64], r1
|
2009-02-28 15:48:54 +02:00
|
|
|
vmull.u8 q10, d0, d6
|
|
|
|
|
\add q1, q8, q1
|
2009-01-25 15:04:45 +02:00
|
|
|
pld [r0]
|
2009-02-28 15:48:54 +02:00
|
|
|
vrshl.s16 q1, q1, q9
|
2009-01-25 15:04:45 +02:00
|
|
|
vqmovun.s16 d2, q1
|
2009-02-28 15:48:54 +02:00
|
|
|
\add q10, q8, q10
|
|
|
|
|
vrshl.s16 q10, q10, q9
|
2009-01-25 15:04:45 +02:00
|
|
|
vqmovun.s16 d4, q10
|
|
|
|
|
vst1.8 {d2},[r4,:64], r1
|
|
|
|
|
vst1.8 {d4},[r4,:64], r1
|
|
|
|
|
bne 1b
|
|
|
|
|
pop {r4, pc}
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2009-01-25 15:04:45 +02:00
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro weight_4 add
|
2011-10-21 10:00:39 +03:00
|
|
|
vdup.8 d0, r12
|
2009-01-25 15:04:45 +02:00
|
|
|
vmov q1, q8
|
|
|
|
|
vmov q10, q8
|
2011-10-21 10:00:39 +03:00
|
|
|
1: subs r2, r2, #4
|
2009-01-25 15:04:45 +02:00
|
|
|
vld1.32 {d4[0]},[r0,:32], r1
|
|
|
|
|
vld1.32 {d4[1]},[r0,:32], r1
|
2009-02-28 15:48:54 +02:00
|
|
|
vmull.u8 q1, d0, d4
|
2009-01-25 15:04:45 +02:00
|
|
|
pld [r0]
|
|
|
|
|
blt 2f
|
|
|
|
|
vld1.32 {d6[0]},[r0,:32], r1
|
|
|
|
|
vld1.32 {d6[1]},[r0,:32], r1
|
2009-02-28 15:48:54 +02:00
|
|
|
vmull.u8 q10, d0, d6
|
2009-01-25 15:04:45 +02:00
|
|
|
pld [r0]
|
2009-02-28 15:48:54 +02:00
|
|
|
\add q1, q8, q1
|
|
|
|
|
vrshl.s16 q1, q1, q9
|
2009-01-25 15:04:45 +02:00
|
|
|
vqmovun.s16 d2, q1
|
2009-02-28 15:48:54 +02:00
|
|
|
\add q10, q8, q10
|
|
|
|
|
vrshl.s16 q10, q10, q9
|
2009-01-25 15:04:45 +02:00
|
|
|
vqmovun.s16 d4, q10
|
|
|
|
|
vmov q10, q8
|
|
|
|
|
vst1.32 {d2[0]},[r4,:32], r1
|
|
|
|
|
vst1.32 {d2[1]},[r4,:32], r1
|
|
|
|
|
vmov q1, q8
|
|
|
|
|
vst1.32 {d4[0]},[r4,:32], r1
|
|
|
|
|
vst1.32 {d4[1]},[r4,:32], r1
|
|
|
|
|
bne 1b
|
|
|
|
|
pop {r4, pc}
|
2009-02-28 15:48:54 +02:00
|
|
|
2: \add q1, q8, q1
|
|
|
|
|
vrshl.s16 q1, q1, q9
|
2009-01-25 15:04:45 +02:00
|
|
|
vqmovun.s16 d2, q1
|
|
|
|
|
vst1.32 {d2[0]},[r4,:32], r1
|
|
|
|
|
vst1.32 {d2[1]},[r4,:32], r1
|
|
|
|
|
pop {r4, pc}
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2009-01-25 15:04:45 +02:00
|
|
|
|
2011-12-02 19:45:00 +03:00
|
|
|
.macro weight_func w
|
2011-10-21 10:00:39 +03:00
|
|
|
function ff_weight_h264_pixels_\w\()_neon, export=1
|
2009-01-25 15:04:45 +02:00
|
|
|
push {r4, lr}
|
2011-10-21 10:00:39 +03:00
|
|
|
ldr r12, [sp, #8]
|
|
|
|
|
ldr r4, [sp, #12]
|
|
|
|
|
cmp r3, #1
|
|
|
|
|
lsl r4, r4, r3
|
2009-01-25 15:04:45 +02:00
|
|
|
vdup.16 q8, r4
|
|
|
|
|
mov r4, r0
|
2009-02-28 15:48:54 +02:00
|
|
|
ble 20f
|
2011-10-21 10:00:39 +03:00
|
|
|
rsb lr, r3, #1
|
2009-02-28 15:48:54 +02:00
|
|
|
vdup.16 q9, lr
|
2011-10-21 10:00:39 +03:00
|
|
|
cmp r12, #0
|
2009-02-28 15:48:54 +02:00
|
|
|
blt 10f
|
|
|
|
|
weight_\w vhadd.s16
|
2011-10-21 10:00:39 +03:00
|
|
|
10: rsb r12, r12, #0
|
2009-02-28 15:48:54 +02:00
|
|
|
weight_\w vhsub.s16
|
2011-10-21 10:00:39 +03:00
|
|
|
20: rsb lr, r3, #0
|
2009-02-28 15:48:54 +02:00
|
|
|
vdup.16 q9, lr
|
2011-10-21 10:00:39 +03:00
|
|
|
cmp r12, #0
|
2009-01-25 15:04:45 +02:00
|
|
|
blt 10f
|
2009-02-28 15:48:54 +02:00
|
|
|
weight_\w vadd.s16
|
2011-10-21 10:00:39 +03:00
|
|
|
10: rsb r12, r12, #0
|
2009-02-28 15:48:54 +02:00
|
|
|
weight_\w vsub.s16
|
2010-03-09 18:17:56 +02:00
|
|
|
endfunc
|
2011-12-02 19:45:00 +03:00
|
|
|
.endm
|
2009-01-25 15:04:45 +02:00
|
|
|
|
|
|
|
|
weight_func 16
|
|
|
|
|
weight_func 8
|
|
|
|
|
weight_func 4
|
