1
0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-02 03:06:28 +02:00
FFmpeg/libavcodec/arm/vp9lpf_16bpp_neon.S

1045 lines
38 KiB
ArmAsm
Raw Normal View History

arm: Add NEON optimizations for 10 and 12 bit vp9 loop filter This work is sponsored by, and copyright, Google. This is pretty much similar to the 8 bpp version, but in some senses simpler. All input pixels are 16 bits, and all intermediates also fit in 16 bits, so there's no lengthening/narrowing in the filter at all. For the full 16 pixel wide filter, we can only process 4 pixels at a time (using an implementation very much similar to the one for 8 bpp), but we can do 8 pixels at a time for the 4 and 8 pixel wide filters with a different implementation of the core filter. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_loop_filter_h_4_8_10bpp_neon: 1.83 2.16 1.40 2.09 vp9_loop_filter_h_8_8_10bpp_neon: 1.39 1.67 1.24 1.70 vp9_loop_filter_h_16_8_10bpp_neon: 1.56 1.47 1.10 1.81 vp9_loop_filter_h_16_16_10bpp_neon: 1.94 1.69 1.33 2.24 vp9_loop_filter_mix2_h_44_16_10bpp_neon: 2.01 2.27 1.67 2.39 vp9_loop_filter_mix2_h_48_16_10bpp_neon: 1.84 2.06 1.45 2.19 vp9_loop_filter_mix2_h_84_16_10bpp_neon: 1.89 2.20 1.47 2.29 vp9_loop_filter_mix2_h_88_16_10bpp_neon: 1.69 2.12 1.47 2.08 vp9_loop_filter_mix2_v_44_16_10bpp_neon: 3.16 3.98 2.50 4.05 vp9_loop_filter_mix2_v_48_16_10bpp_neon: 2.84 3.64 2.25 3.77 vp9_loop_filter_mix2_v_84_16_10bpp_neon: 2.65 3.45 2.16 3.54 vp9_loop_filter_mix2_v_88_16_10bpp_neon: 2.55 3.30 2.16 3.55 vp9_loop_filter_v_4_8_10bpp_neon: 2.85 3.97 2.24 3.68 vp9_loop_filter_v_8_8_10bpp_neon: 2.27 3.19 1.96 3.08 vp9_loop_filter_v_16_8_10bpp_neon: 3.42 2.74 2.26 4.40 vp9_loop_filter_v_16_16_10bpp_neon: 2.86 2.44 1.93 3.88 The speedup vs C code measured in checkasm is around 1.1-4x. These numbers are quite inconclusive though, since the checkasm test runs multiple filterings on top of each other, so later rounds might end up with different codepaths (different decisions on which filter to apply, based on input pixel differences). Based on START_TIMER/STOP_TIMER wrapping around a few individual functions, the speedup vs C code is around 2-4x. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-05 12:51:08 +02:00
/*
* Copyright (c) 2017 Google Inc.
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* 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.
*
* FFmpeg is distributed in the hope that it will be useful,
* 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
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/arm/asm.S"
.macro transpose16_q_8x8 rq0, rq1, rq2, rq3, rq4, rq5, rq6, rq7, r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15
vswp \r1, \r8 @ vtrn.64 \rq0, \rq4
vswp \r3, \r10 @ vtrn.64 \rq1, \rq5
vswp \r5, \r12 @ vtrn.64 \rq2, \rq6
vswp \r7, \r14 @ vtrn.64 \rq3, \rq7
vtrn.32 \rq0, \rq2
vtrn.32 \rq1, \rq3
vtrn.32 \rq4, \rq6
vtrn.32 \rq5, \rq7
vtrn.16 \rq0, \rq1
vtrn.16 \rq2, \rq3
vtrn.16 \rq4, \rq5
vtrn.16 \rq6, \rq7
.endm
.macro transpose16_4x4 r0, r1, r2, r3
vtrn.32 \r0, \r2
vtrn.32 \r1, \r3
vtrn.16 \r0, \r1
vtrn.16 \r2, \r3
.endm
@ Do a 4x4 transpose, using q registers for the subtransposes that don't
@ need to address the indiviudal d registers.
@ r0,r1 == rq0, r2,r3 == rq1
.macro transpose16_q_4x4 rq0, rq1, r0, r1, r2, r3
vtrn.32 \rq0, \rq1
vtrn.16 \r0, \r1
vtrn.16 \r2, \r3
.endm
@ The input to and output from this macro is in the registers q8-q15,
@ and q0-q7 are used as scratch registers.
@ p3 = q8, p0 = q11, q0 = q12, q3 = q15
.macro loop_filter_q wd
vdup.u16 q0, r2 @ E
vdup.u16 q1, r3 @ I
vabd.u16 q2, q8, q9 @ abs(p3 - p2)
vabd.u16 q3, q9, q10 @ abs(p2 - p1)
vabd.u16 q4, q10, q11 @ abs(p1 - p0)
vabd.u16 q5, q12, q13 @ abs(q0 - q1)
vabd.u16 q6, q13, q14 @ abs(q1 - q2)
vabd.u16 q7, q14, q15 @ abs(q2 - q3)
vmax.u16 q2, q2, q3
vmax.u16 q3, q4, q5
vmax.u16 q4, q6, q7
vabd.u16 q5, q11, q12 @ abs(p0 - q0)
vmax.u16 q2, q2, q3
vadd.u16 q5, q5, q5 @ abs(p0 - q0) * 2
vabd.u16 q6, q10, q13 @ abs(p1 - q1)
vmax.u16 q2, q2, q4 @ max(abs(p3 - p2), ..., abs(q2 - q3))
vshr.u16 q6, q6, #1
vcle.u16 q2, q2, q1 @ max(abs()) <= I
vadd.u16 q5, q5, q6 @ abs(p0 - q0) * 2 + abs(p1 - q1) >> 1
vcle.u16 q5, q5, q0
vand q2, q2, q5 @ fm
vmovn.u16 d10, q2
vmov r8, r9, d10
orrs r8, r8, r9
@ If no pixels need filtering, just exit as soon as possible
beq 9f
.if \wd >= 8
vdup.u16 q0, r5
vabd.u16 q1, q8, q11 @ abs(p3 - p0)
vabd.u16 q3, q9, q11 @ abs(p2 - p0)
vabd.u16 q4, q10, q11 @ abs(p1 - p0)
vabd.u16 q5, q13, q12 @ abs(q1 - q0)
vabd.u16 q6, q14, q12 @ abs(q2 - q0)
vabd.u16 q7, q15, q12 @ abs(q3 - q0)
vmax.u16 q1, q1, q3
vmax.u16 q4, q4, q5
vmax.u16 q6, q6, q7
@ The rest of the calculation of flat8in is interleaved below
.endif
@ Calculate the normal inner loop filter for 2 or 4 pixels
vabd.u16 q3, q10, q11 @ abs(p1 - p0)
.if \wd == 8
vmax.u16 q1, q1, q4
.endif
vabd.u16 q4, q13, q12 @ abs(q1 - q0)
.if \wd == 8
vmax.u16 q1, q1, q6
.endif
vsub.u16 q5, q10, q13 @ p1 - q1
vmax.u16 q3, q3, q4 @ max(abs(p1 - p0), abs(q1 - q0))
vdup.u16 q4, r4 @ H
vsub.u16 q6, q12, q11 @ q0 - p0
.if \wd == 8
vcle.u16 q1, q1, q0 @ flat8in
.endif
vdup.u16 q0, r6 @ left shift for saturation
vcle.u16 q3, q3, q4 @ !hev
.if \wd == 8
vand q1, q1, q2 @ flat8in && fm
.endif
vneg.s16 q4, q0 @ negative left shift after saturation
vqshl.s16 q5, q5, q0
.if \wd == 8
vbic q2, q2, q1 @ fm && !flat8in
.endif
vmov.s16 q7, #3
vand q3, q3, q2 @ !hev && fm && !flat8in
vshl.s16 q5, q5, q4 @ av_clip_int2p(p1 - q1, BIT_DEPTH - 1)
vmul.s16 q6, q6, q7 @ 3 * (q0 - p0)
vbic q5, q5, q3 @ if (!hev) av_clip_int2p = 0
vadd.s16 q6, q6, q5 @ 3 * (q0 - p0) [+ av_clip_int2p(p1 - q1)]
vmov.s16 q5, #4
vqshl.s16 q6, q6, q0
vmov.s16 q0, #3
vshl.s16 q6, q6, q4 @ av_clip_int2p(3 * (q0 - p0) [+ av_clip_int2p(p1 - q1)], BIT_DEPTH - 1) = f
vdup.u16 q4, r7 @ max pixel value
vshr.u16 q4, q4, #1 @ (1 << (BIT_DEPTH - 1)) - 1)
vadd.s16 q5, q6, q5 @ f + 4
vadd.s16 q0, q6, q0 @ f + 3
vmov.s16 q6, #0
vmin.s16 q5, q5, q4 @ FFMIN(f + 4, (1 << (BIT_DEPTH - 1)) - 1)
vmin.s16 q0, q0, q4 @ FFMIN(f + 3, (1 << (BIT_DEPTH - 1)) - 1)
vdup.u16 q4, r7 @ max pixel value
vshr.s16 q5, q5, #3 @ f1
vshr.s16 q0, q0, #3 @ f2
vadd.s16 q0, q11, q0 @ p0 + f2
vsub.s16 q7, q12, q5 @ q0 - f1
vmin.s16 q0, q0, q4
vmin.s16 q7, q7, q4
vrshr.s16 q5, q5, #1 @ f = (f1 + 1) >> 1
vmax.s16 q0, q0, q6 @ out p0
vmax.s16 q7, q7, q6 @ out q0
vbit q11, q0, q2 @ if (fm && !flat8in)
vbit q12, q7, q2
.if \wd >= 8
vmovn.u16 d4, q1
.endif
vadd.s16 q0, q10, q5 @ p1 + f
vsub.s16 q7, q13, q5 @ q1 - f
.if \wd >= 8
vmov r8, r9, d4
.endif
vmin.s16 q0, q0, q4
vmin.s16 q7, q7, q4
.if \wd >= 8
orrs r8, r8, r9
.endif
vmax.s16 q0, q0, q6 @ out p1
vmax.s16 q7, q7, q6 @ out q1
vbit q10, q0, q3 @ if (!hev && fm && !flat8in)
vbit q13, q7, q3
.if \wd >= 8
@ If no pixels need flat8in, jump to a writeout of the inner 4 pixels
beq 6f
@ flat8in
vadd.u16 q2, q8, q9
vadd.u16 q3, q10, q13
vadd.u16 q4, q8, q10
vadd.u16 q5, q11, q14
vadd.u16 q0, q2, q2
vadd.u16 q0, q0, q11
vadd.u16 q0, q0, q12
vadd.u16 q0, q0, q4
vsub.s16 q3, q3, q2
vsub.s16 q5, q5, q4
vrshr.u16 q6, q0, #3 @ out p2
vadd.u16 q0, q0, q3
vadd.u16 q2, q8, q11
vadd.u16 q3, q12, q15
vrshr.u16 q7, q0, #3 @ out p1
vadd.u16 q0, q0, q5
vsub.s16 q3, q3, q2
vadd.u16 q4, q9, q12
vbit q9, q6, q1
vadd.u16 q5, q13, q15
vrshr.u16 q6, q0, #3 @ out p0
vadd.u16 q0, q0, q3
vsub.s16 q5, q5, q4
vadd.u16 q2, q10, q13
vbit q10, q7, q1
vadd.u16 q3, q14, q15
vrshr.u16 q7, q0, #3 @ out q0
vadd.u16 q0, q0, q5
vsub.s16 q3, q3, q2
vbit q11, q6, q1
vrshr.u16 q6, q0, #3 @ out q1
vadd.u16 q0, q0, q3
vbit q12, q7, q1
vrshr.u16 q7, q0, #3 @ out q2
vbit q13, q6, q1
vbit q14, q7, q1
.endif
.endm
@ The input to and output from this macro is in the registers d16-d31,
@ and d0-d7 are used as scratch registers.
@ p7 = d16 .. p3 = d20, p0 = d23, q0 = d24, q3 = d27, q7 = d31
@ Depending on the width of the loop filter, we either use d16-d19
@ and d28-d31 as temp registers, or d8-d15.
@ In practice, this is only ever instantiated once, so the macro parameters
@ could be hardcoded, but keeping them as is, to keep similarities to the
@ 8 bpp and aarch64 versions.
.macro loop_filter wd, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8
vdup.u16 d0, r2 @ E
vdup.u16 d2, r3 @ I
vabd.u16 d4, d20, d21 @ abs(p3 - p2)
vabd.u16 d5, d21, d22 @ abs(p2 - p1)
vabd.u16 d6, d22, d23 @ abs(p1 - p0)
vabd.u16 d7, d24, d25 @ abs(q0 - q1)
vabd.u16 \tmp1, d25, d26 @ abs(q1 - q2)
vabd.u16 \tmp2, d26, d27 @ abs(q2 - q3)
vmax.u16 d4, d4, d5
vmax.u16 d5, d6, d7
vmax.u16 \tmp1, \tmp1, \tmp2
vabd.u16 d6, d23, d24 @ abs(p0 - q0)
vmax.u16 d4, d4, d5
vadd.u16 d6, d6, d6 @ abs(p0 - q0) * 2
vabd.u16 d5, d22, d25 @ abs(p1 - q1)
vmax.u16 d4, d4, \tmp1 @ max(abs(p3 - p2), ..., abs(q2 - q3))
vshr.u16 d5, d5, #1
vcle.u16 d4, d4, d2 @ max(abs()) <= I
vadd.u16 d6, d6, d5 @ abs(p0 - q0) * 2 + abs(p1 - q1) >> 1
vcle.u16 d6, d6, d0
vand d4, d4, d6 @ fm
vdup.u16 d3, r4 @ H
vmov r8, r9, d4
orrs r8, r8, r9
@ If no pixels need filtering, just exit as soon as possible
beq 9f
.if \wd >= 8
vdup.u16 d0, r5
vabd.u16 d6, d20, d23 @ abs(p3 - p0)
vabd.u16 d2, d21, d23 @ abs(p2 - p0)
vabd.u16 d1, d22, d23 @ abs(p1 - p0)
vabd.u16 \tmp1, d25, d24 @ abs(q1 - q0)
vabd.u16 \tmp2, d26, d24 @ abs(q2 - q0)
vabd.u16 \tmp3, d27, d24 @ abs(q3 - q0)
vmax.u16 d6, d6, d2
vmax.u16 d1, d1, \tmp1
vmax.u16 \tmp2, \tmp2, \tmp3
.if \wd == 16
vabd.u16 d7, d16, d23 @ abs(p7 - p0)
vmax.u16 d6, d6, d1
vabd.u16 d2, d17, d23 @ abs(p6 - p0)
vmax.u16 d6, d6, \tmp2
vabd.u16 d1, d18, d23 @ abs(p5 - p0)
vcle.u16 d6, d6, d0 @ flat8in
vabd.u16 d8, d19, d23 @ abs(p4 - p0)
vand d6, d6, d4 @ flat8in && fm
vabd.u16 d9, d28, d24 @ abs(q4 - q0)
vbic d4, d4, d6 @ fm && !flat8in
vabd.u16 d10, d29, d24 @ abs(q5 - q0)
vabd.u16 d11, d30, d24 @ abs(q6 - q0)
vabd.u16 d12, d31, d24 @ abs(q7 - q0)
vmax.u16 d7, d7, d2
vmax.u16 d1, d1, d8
vmax.u16 d9, d9, d10
vmax.u16 d11, d11, d12
@ The rest of the calculation of flat8out is interleaved below
.else
@ The rest of the calculation of flat8in is interleaved below
.endif
.endif
@ Calculate the normal inner loop filter for 2 or 4 pixels
vabd.u16 d5, d22, d23 @ abs(p1 - p0)
.if \wd == 16
vmax.u16 d7, d7, d1
vmax.u16 d9, d9, d11
.elseif \wd == 8
vmax.u16 d6, d6, d1
.endif
vabd.u16 d1, d25, d24 @ abs(q1 - q0)
.if \wd == 16
vmax.u16 d7, d7, d9
.elseif \wd == 8
vmax.u16 d6, d6, \tmp2
.endif
vdup.u16 \tmp2, r6 @ left shift for saturation
vsub.u16 \tmp1, d22, d25 @ p1 - q1
vneg.s16 \tmp6, \tmp2 @ negative left shift after saturation
vmax.u16 d5, d5, d1 @ max(abs(p1 - p0), abs(q1 - q0))
vsub.u16 \tmp3, d24, d23 @ q0 - p0
vmov.s16 \tmp5, #3
.if \wd == 8
vcle.u16 d6, d6, d0 @ flat8in
.endif
vcle.u16 d5, d5, d3 @ !hev
.if \wd == 8
vand d6, d6, d4 @ flat8in && fm
.endif
vqshl.s16 \tmp1, \tmp1, \tmp2
.if \wd == 16
vcle.u16 d7, d7, d0 @ flat8out
.elseif \wd == 8
vbic d4, d4, d6 @ fm && !flat8in
.endif
vand d5, d5, d4 @ !hev && fm && !flat8in
.if \wd == 16
vand d7, d7, d6 @ flat8out && flat8in && fm
.endif
vshl.s16 \tmp1, \tmp1, \tmp6 @ av_clip_int2p(p1 - q1, BIT_DEPTH - 1)
vmul.s16 \tmp3, \tmp3, \tmp5 @ 3 * (q0 - p0)
vbic \tmp1, \tmp1, d5 @ if (!hev) av_clip_int2p = 0
vmov.s16 d2, #4
vadd.s16 \tmp3, \tmp3, \tmp1 @ 3 * (q0 - p0) [+ av_clip_int2p(p1 - q1)]
vmov.s16 d3, #3
vqshl.s16 \tmp1, \tmp3, \tmp2
vmov.s16 \tmp5, #0
vshl.s16 \tmp1, \tmp1, \tmp6 @ av_clip_int2p(3 * (q0 - p0) [+ av_clip_int2p(p1 - q1)], BIT_DEPTH - 1) = f
vdup.u16 \tmp6, r7 @ max pixel value
.if \wd == 16
vbic d6, d6, d7 @ fm && flat8in && !flat8out
.endif
vshr.u16 \tmp2, \tmp6, #1 @ (1 << (BIT_DEPTH - 1)) - 1
vadd.s16 \tmp3, \tmp1, d2 @ f + 4
vadd.s16 \tmp4, \tmp1, d3 @ f + 3
vmin.s16 \tmp3, \tmp3, \tmp2 @ FFMIN(f + 4, (1 << (BIT_DEPTH - 1)) - 1)
vmin.s16 \tmp4, \tmp4, \tmp2 @ FFMIN(f + 3, (1 << (BIT_DEPTH - 1)) - 1)
vshr.s16 \tmp3, \tmp3, #3 @ f1
vshr.s16 \tmp4, \tmp4, #3 @ f2
vadd.s16 d0, d23, \tmp4 @ p0 + f2
vsub.s16 d2, d24, \tmp3 @ q0 - f1
vmin.s16 d0, d0, \tmp6
vmin.s16 d2, d2, \tmp6
vrshr.s16 \tmp3, \tmp3, #1 @ f = (f1 + 1) >> 1
vmax.s16 d0, d0, \tmp5 @ out p0
vmax.s16 d2, d2, \tmp5 @ out q0
vbit d23, d0, d4 @ if (fm && !flat8in)
vbit d24, d2, d4
vadd.s16 d0, d22, \tmp3 @ p1 + f
vsub.s16 d2, d25, \tmp3 @ q1 - f
.if \wd >= 8
vmov r8, r9, d6
.endif
vmin.s16 d0, d0, \tmp6
vmin.s16 d2, d2, \tmp6
.if \wd >= 8
orrs r8, r8, r9
.endif
vmax.s16 d0, d0, \tmp5 @ out p1
vmax.s16 d2, d2, \tmp5 @ out q1
vbit d22, d0, d5 @ if (!hev && fm && !flat8in)
vbit d25, d2, d5
.if \wd >= 8
@ If no pixels need flat8in, jump to flat8out
@ (or to a writeout of the inner 4 pixels, for wd=8)
beq 6f
@ flat8in
vadd.u16 \tmp1, d20, d21
vadd.u16 \tmp3, d22, d25
vadd.u16 \tmp5, d20, d22
vadd.u16 \tmp7, d23, d26
vadd.u16 d0, \tmp1, \tmp1
vadd.u16 d0, d0, d23
vadd.u16 d0, d0, d24
vadd.u16 d0, d0, \tmp5
vsub.s16 \tmp3, \tmp3, \tmp1
vsub.s16 \tmp7, \tmp7, \tmp5
vrshr.u16 d2, d0, #3 @ out p2
vadd.u16 d0, d0, \tmp3
vadd.u16 \tmp1, d20, d23
vadd.u16 \tmp3, d24, d27
vrshr.u16 d3, d0, #3 @ out p1
vadd.u16 d0, d0, \tmp7
vsub.s16 \tmp3, \tmp3, \tmp1
vadd.u16 \tmp5, d21, d24
vadd.u16 \tmp7, d25, d27
vrshr.u16 d4, d0, #3 @ out p0
vadd.u16 d0, d0, \tmp3
vsub.s16 \tmp7, \tmp7, \tmp5
vadd.u16 \tmp1, d22, d25
vadd.u16 \tmp3, d26, d27
vrshr.u16 d5, d0, #3 @ out d0
vadd.u16 d0, d0, \tmp7
vsub.s16 \tmp3, \tmp3, \tmp1
vrshr.u16 \tmp5, d0, #3 @ out q1
vadd.u16 d0, d0, \tmp3
@ The output here is written back into the input registers. This doesn't
@ matter for the flat8out part below, since we only update those pixels
@ which won't be touched below.
vbit d21, d2, d6
vbit d22, d3, d6
vbit d23, d4, d6
vrshr.u16 \tmp6, d0, #3 @ out q2
vbit d24, d5, d6
vbit d25, \tmp5, d6
vbit d26, \tmp6, d6
.endif
.if \wd == 16
6:
vorr d2, d6, d7
vmov r8, r9, d2
orrs r8, r8, r9
@ If no pixels needed flat8in nor flat8out, jump to a
@ writeout of the inner 4 pixels
beq 7f
vmov r8, r9, d7
orrs r8, r8, r9
@ If no pixels need flat8out, jump to a writeout of the inner 6 pixels
beq 8f
@ flat8out
@ This writes all outputs into d2-d17 (skipping d6 and d16).
@ If this part is skipped, the output is read from d21-d26 (which is the input
@ to this section).
vshl.u16 d0, d16, #3 @ 8 * d16
vsub.u16 d0, d0, d16 @ 7 * d16
vadd.u16 d0, d0, d17
vadd.u16 d8, d17, d18
vadd.u16 d10, d19, d20
vadd.s16 d0, d0, d8
vadd.u16 d8, d16, d17
vadd.u16 d12, d21, d22
vadd.s16 d0, d0, d10
vadd.u16 d10, d18, d25
vadd.u16 d14, d23, d24
vsub.s16 d10, d10, d8
vadd.s16 d0, d0, d12
vadd.s16 d0, d0, d14
vadd.u16 d12, d16, d18
vadd.u16 d14, d19, d26
vrshr.u16 d2, d0, #4
vadd.s16 d0, d0, d10
vadd.u16 d8, d16, d19
vadd.u16 d10, d20, d27
vsub.s16 d14, d14, d12
vbif d2, d17, d7
vrshr.u16 d3, d0, #4
vadd.s16 d0, d0, d14
vadd.u16 d12, d16, d20
vadd.u16 d14, d21, d28
vsub.s16 d10, d10, d8
vbif d3, d18, d7
vrshr.u16 d4, d0, #4
vadd.s16 d0, d0, d10
vadd.u16 d8, d16, d21
vadd.u16 d10, d22, d29
vsub.s16 d14, d14, d12
vbif d4, d19, d7
vrshr.u16 d5, d0, #4
vadd.s16 d0, d0, d14
vadd.u16 d12, d16, d22
vadd.u16 d14, d23, d30
vsub.s16 d10, d10, d8
vbif d5, d20, d7
vrshr.u16 d6, d0, #4
vadd.s16 d0, d0, d10
vadd.u16 d10, d16, d23
vsub.s16 d14, d14, d12
vadd.u16 d12, d24, d31
vbif d6, d21, d7
vrshr.u16 d8, d0, #4
vadd.s16 d0, d0, d14
vsub.s16 d10, d12, d10
vadd.u16 d12, d17, d24
vadd.u16 d14, d25, d31
vbif d8, d22, d7
vrshr.u16 d9, d0, #4
vadd.s16 d0, d0, d10
vsub.s16 d14, d14, d12
vadd.u16 d12, d26, d31
vbif d9, d23, d7
vrshr.u16 d10, d0, #4
vadd.s16 d0, d0, d14
vadd.u16 d14, d18, d25
vadd.u16 d18, d19, d26
vsub.s16 d12, d12, d14
vadd.u16 d14, d27, d31
vbif d10, d24, d7
vrshr.u16 d11, d0, #4
vadd.s16 d0, d0, d12
vadd.u16 d12, d20, d27
vsub.s16 d14, d14, d18
vadd.u16 d18, d28, d31
vbif d11, d25, d7
vsub.s16 d18, d18, d12
vrshr.u16 d12, d0, #4
vadd.s16 d0, d0, d14
vadd.u16 d14, d21, d28
vadd.u16 d20, d29, d31
vbif d12, d26, d7
vrshr.u16 d13, d0, #4
vadd.s16 d0, d0, d18
vsub.s16 d20, d20, d14
vadd.u16 d18, d22, d29
vadd.u16 d22, d30, d31
vbif d13, d27, d7
vrshr.u16 d14, d0, #4
vadd.s16 d0, d0, d20
vsub.s16 d22, d22, d18
vbif d14, d28, d7
vrshr.u16 d15, d0, #4
vadd.s16 d0, d0, d22
vbif d15, d29, d7
vrshr.u16 d17, d0, #4
vbif d17, d30, d7
.endif
.endm
.macro loop_filter_q_4
loop_filter_q 4
.endm
.macro loop_filter_q_8
loop_filter_q 8
.endm
.macro loop_filter_16
loop_filter 16, d8, d9, d10, d11, d12, d13, d14, d15
.endm
@ The public functions in this file have got the following signature:
@ void loop_filter(uint8_t *dst, ptrdiff_t stride, int mb_lim, int lim, int hev_thr);
.macro bpp_frontend func, bpp
function ff_\func\()_\bpp\()_neon, export=1
push {r4-r9,lr}
ldr r4, [sp, #28]
vpush {q4-q7}
lsl r2, r2, #\bpp - 8
lsl r3, r3, #\bpp - 8
lsl r4, r4, #\bpp - 8
mov r5, #1 << (\bpp - 8)
mov r6, #16 - \bpp
movw r7, #((1 << \bpp) - 1)
bl \func\()_16_neon
vpop {q4-q7}
pop {r4-r9,pc}
endfunc
.endm
.macro bpp_frontends func
bpp_frontend \func, 10
bpp_frontend \func, 12
.endm
.macro bpp_frontend_rep func, suffix, int_suffix, rep, dir, bpp
function ff_\func\()_\suffix\()_\bpp\()_neon, export=1
push {r4-r9,lr}
ldr r4, [sp, #28]
vpush {q4-q7}
lsl r2, r2, #\bpp - 8
lsl r3, r3, #\bpp - 8
lsl r4, r4, #\bpp - 8
mov r5, #1 << (\bpp - 8)
mov r6, #16 - \bpp
movw r7, #((1 << \bpp) - 1)
bl \func\()_\int_suffix\()_16_neon
.ifc \dir,h
add r0, r0, r1, lsl #2
.else
add r0, r0, #8
.endif
bl \func\()_\int_suffix\()_16_neon
.if \rep >= 4
.ifc \dir,h
add r0, r0, r1, lsl #2
bl \func\()_\int_suffix\()_16_neon
add r0, r0, r1, lsl #2
bl \func\()_\int_suffix\()_16_neon
.else
add r0, r0, #8
bl \func\()_\int_suffix\()_16_neon
add r0, r0, #8
bl \func\()_\int_suffix\()_16_neon
.endif
.endif
vpop {q4-q7}
pop {r4-r9,pc}
endfunc
.endm
.macro bpp_frontends_rep func, suffix, int_suffix, rep, dir
bpp_frontend_rep \func, \suffix, \int_suffix, \rep, \dir, 10
bpp_frontend_rep \func, \suffix, \int_suffix, \rep, \dir, 12
.endm
.macro bpp_frontend_mix2 wd1, wd2, dir, bpp
function ff_vp9_loop_filter_\dir\()_\wd1\()\wd2\()_16_\bpp\()_neon, export=1
push {r4-r9,lr}
ldr r4, [sp, #28]
vpush {q4-q7}
push {r2, r3, r4}
and r2, r2, #0xff
and r3, r3, #0xff
and r4, r4, #0xff
lsl r2, r2, #\bpp - 8
lsl r3, r3, #\bpp - 8
lsl r4, r4, #\bpp - 8
mov r5, #1 << (\bpp - 8)
mov r6, #16 - \bpp
movw r7, #((1 << \bpp) - 1)
bl vp9_loop_filter_\dir\()_\wd1\()_8_16_neon
.ifc \dir,h
add r0, r0, r1, lsl #3
.else
add r0, r0, #16
.endif
pop {r2, r3, r4}
lsr r2, r2, #8
lsr r3, r3, #8
lsr r4, r4, #8
lsl r2, r2, #\bpp - 8
lsl r3, r3, #\bpp - 8
lsl r4, r4, #\bpp - 8
bl vp9_loop_filter_\dir\()_\wd2\()_8_16_neon
vpop {q4-q7}
pop {r4-r9,pc}
endfunc
.endm
.macro bpp_frontends_mix2 wd1, wd2
bpp_frontend_mix2 \wd1, \wd2, v, 10
bpp_frontend_mix2 \wd1, \wd2, v, 12
bpp_frontend_mix2 \wd1, \wd2, h, 10
bpp_frontend_mix2 \wd1, \wd2, h, 12
.endm
function vp9_loop_filter_v_4_8_16_neon
sub r12, r0, r1, lsl #2
vld1.16 {q8}, [r12,:128], r1 @ p3
vld1.16 {q12}, [r0, :128], r1 @ q0
vld1.16 {q9}, [r12,:128], r1 @ p2
vld1.16 {q13}, [r0, :128], r1 @ q1
vld1.16 {q10}, [r12,:128], r1 @ p1
vld1.16 {q14}, [r0, :128], r1 @ q2
vld1.16 {q11}, [r12,:128], r1 @ p0
vld1.16 {q15}, [r0, :128], r1 @ q3
sub r0, r0, r1, lsl #2
sub r12, r12, r1, lsl #1
loop_filter_q_4
vst1.16 {q10}, [r12,:128], r1
vst1.16 {q12}, [r0, :128], r1
vst1.16 {q11}, [r12,:128], r1
vst1.16 {q13}, [r0, :128], r1
sub r0, r0, r1, lsl #1
9:
bx lr
endfunc
bpp_frontends vp9_loop_filter_v_4_8
function vp9_loop_filter_h_4_8_16_neon
sub r12, r0, #8
add r0, r12, r1, lsl #2
vld1.16 {q8}, [r12,:64], r1
vld1.16 {q12}, [r0, :64], r1
vld1.16 {q9}, [r12,:64], r1
vld1.16 {q13}, [r0, :64], r1
vld1.16 {q10}, [r12,:64], r1
vld1.16 {q14}, [r0, :64], r1
vld1.16 {q11}, [r12,:64], r1
vld1.16 {q15}, [r0, :64], r1
sub r12, r12, r1, lsl #2
sub r0, r0, r1, lsl #2
@ Move r0/r12 forward by 2 pixels; we don't need to rewrite the
@ outermost 2 pixels since they aren't changed.
add r12, r12, #4
add r0, r0, #4
transpose16_q_8x8 q8, q9, q10, q11, q12, q13, q14, q15, d16, d17, d18, d19, d20, d21, d22, d23, d24, d25, d26, d27, d28, d29, d30, d31
loop_filter_q_4
@ We only will write the mid 4 pixels back; after the loop filter,
@ these are in q10, q11, q12, q13, ordered as rows (8x4 pixels).
@ We need to transpose them to columns, done with a
@ 4x4 transpose (which in practice is two 4x4 transposes of the two
@ 4x4 halves of the 8x4 pixels; into 4x8 pixels).
transpose16_4x4 q10, q11, q12, q13
vst1.16 {d20}, [r12], r1
vst1.16 {d21}, [r0], r1
vst1.16 {d22}, [r12], r1
vst1.16 {d23}, [r0], r1
vst1.16 {d24}, [r12], r1
vst1.16 {d25}, [r0], r1
vst1.16 {d26}, [r12], r1
vst1.16 {d27}, [r0], r1
sub r12, r12, r1, lsl #2
9:
add r0, r12, #4
bx lr
endfunc
bpp_frontends vp9_loop_filter_h_4_8
function vp9_loop_filter_v_8_8_16_neon
sub r12, r0, r1, lsl #2
vld1.16 {q8}, [r12,:128], r1 @ p3
vld1.16 {q12}, [r0, :128], r1 @ q0
vld1.16 {q9}, [r12,:128], r1 @ p2
vld1.16 {q13}, [r0, :128], r1 @ q1
vld1.16 {q10}, [r12,:128], r1 @ p1
vld1.16 {q14}, [r0, :128], r1 @ q2
vld1.16 {q11}, [r12,:128], r1 @ p0
vld1.16 {q15}, [r0, :128], r1 @ q3
sub r12, r12, r1, lsl #2
sub r0, r0, r1, lsl #2
add r12, r12, r1
loop_filter_q_8
vst1.16 {q9}, [r12,:128], r1
vst1.16 {q12}, [r0, :128], r1
vst1.16 {q10}, [r12,:128], r1
vst1.16 {q13}, [r0, :128], r1
vst1.16 {q11}, [r12,:128], r1
vst1.16 {q14}, [r0, :128], r1
sub r0, r0, r1, lsl #1
sub r0, r0, r1
9:
bx lr
6:
sub r12, r0, r1, lsl #1
vst1.16 {q10}, [r12,:128], r1
vst1.16 {q12}, [r0, :128], r1
vst1.16 {q11}, [r12,:128], r1
vst1.16 {q13}, [r0, :128], r1
sub r0, r0, r1, lsl #1
bx lr
endfunc
bpp_frontends vp9_loop_filter_v_8_8
function vp9_loop_filter_h_8_8_16_neon
sub r12, r0, #8
add r0, r12, r1, lsl #2
vld1.16 {q8}, [r12,:64], r1
vld1.16 {q12}, [r0, :64], r1
vld1.16 {q9}, [r12,:64], r1
vld1.16 {q13}, [r0, :64], r1
vld1.16 {q10}, [r12,:64], r1
vld1.16 {q14}, [r0, :64], r1
vld1.16 {q11}, [r12,:64], r1
vld1.16 {q15}, [r0, :64], r1
sub r12, r12, r1, lsl #2
sub r0, r0, r1, lsl #2
transpose16_q_8x8 q8, q9, q10, q11, q12, q13, q14, q15, d16, d17, d18, d19, d20, d21, d22, d23, d24, d25, d26, d27, d28, d29, d30, d31
loop_filter_q_8
@ Even though only 6 pixels per row have been changed, we write the
@ full 8 pixel registers.
transpose16_q_8x8 q8, q9, q10, q11, q12, q13, q14, q15, d16, d17, d18, d19, d20, d21, d22, d23, d24, d25, d26, d27, d28, d29, d30, d31
vst1.16 {q8}, [r12,:64], r1
vst1.16 {q12}, [r0, :64], r1
vst1.16 {q9}, [r12,:64], r1
vst1.16 {q13}, [r0, :64], r1
vst1.16 {q10}, [r12,:64], r1
vst1.16 {q14}, [r0, :64], r1
vst1.16 {q11}, [r12,:64], r1
vst1.16 {q15}, [r0, :64], r1
sub r12, r12, r1, lsl #2
9:
add r0, r12, #8
bx lr
6:
@ If we didn't need to do the flat8in part, we use the same writeback
@ as in loop_filter_h_4_8.
add r12, r12, #4
add r0, r0, #4
transpose16_4x4 q10, q11, q12, q13
vst1.16 {d20}, [r12], r1
vst1.16 {d21}, [r0], r1
vst1.16 {d22}, [r12], r1
vst1.16 {d23}, [r0], r1
vst1.16 {d24}, [r12], r1
vst1.16 {d25}, [r0], r1
vst1.16 {d26}, [r12], r1
vst1.16 {d27}, [r0], r1
sub r12, r12, r1, lsl #2
add r0, r12, #4
bx lr
endfunc
bpp_frontends vp9_loop_filter_h_8_8
bpp_frontends_mix2 4, 4
bpp_frontends_mix2 4, 8
bpp_frontends_mix2 8, 4
bpp_frontends_mix2 8, 8
function vp9_loop_filter_v_16_4_16_neon
sub r12, r0, r1, lsl #3
@ Read p7-p0 using r12 and q0-q7 using r0
vld1.16 {d16}, [r12,:64], r1 @ p7
vld1.16 {d24}, [r0, :64], r1 @ q0
vld1.16 {d17}, [r12,:64], r1 @ p6
vld1.16 {d25}, [r0, :64], r1 @ q1
vld1.16 {d18}, [r12,:64], r1 @ p5
vld1.16 {d26}, [r0, :64], r1 @ q2
vld1.16 {d19}, [r12,:64], r1 @ p4
vld1.16 {d27}, [r0, :64], r1 @ q3
vld1.16 {d20}, [r12,:64], r1 @ p3
vld1.16 {d28}, [r0, :64], r1 @ q4
vld1.16 {d21}, [r12,:64], r1 @ p2
vld1.16 {d29}, [r0, :64], r1 @ q5
vld1.16 {d22}, [r12,:64], r1 @ p1
vld1.16 {d30}, [r0, :64], r1 @ q6
vld1.16 {d23}, [r12,:64], r1 @ p0
vld1.16 {d31}, [r0, :64], r1 @ q7
sub r12, r12, r1, lsl #3
sub r0, r0, r1, lsl #3
add r12, r12, r1
loop_filter_16
@ If we did the flat8out part, we get the output in
@ d2-d17 (skipping d7 and d16). r12 points to r0 - 7 * stride,
@ store d2-d9 there, and d10-d17 into r0.
vst1.16 {d2}, [r12,:64], r1
vst1.16 {d10}, [r0, :64], r1
vst1.16 {d3}, [r12,:64], r1
vst1.16 {d11}, [r0, :64], r1
vst1.16 {d4}, [r12,:64], r1
vst1.16 {d12}, [r0, :64], r1
vst1.16 {d5}, [r12,:64], r1
vst1.16 {d13}, [r0, :64], r1
vst1.16 {d6}, [r12,:64], r1
vst1.16 {d14}, [r0, :64], r1
vst1.16 {d8}, [r12,:64], r1
vst1.16 {d15}, [r0, :64], r1
vst1.16 {d9}, [r12,:64], r1
vst1.16 {d17}, [r0, :64], r1
sub r0, r0, r1, lsl #3
add r0, r0, r1
9:
bx lr
8:
add r12, r12, r1, lsl #2
@ If we didn't do the flat8out part, the output is left in the
@ input registers.
vst1.16 {d21}, [r12,:64], r1
vst1.16 {d24}, [r0, :64], r1
vst1.16 {d22}, [r12,:64], r1
vst1.16 {d25}, [r0, :64], r1
vst1.16 {d23}, [r12,:64], r1
vst1.16 {d26}, [r0, :64], r1
sub r0, r0, r1, lsl #1
sub r0, r0, r1
bx lr
7:
sub r12, r0, r1, lsl #1
vst1.16 {d22}, [r12,:64], r1
vst1.16 {d24}, [r0, :64], r1
vst1.16 {d23}, [r12,:64], r1
vst1.16 {d25}, [r0, :64], r1
sub r0, r0, r1, lsl #1
bx lr
endfunc
bpp_frontends_rep vp9_loop_filter_v_16, 8, 4, 2, v
bpp_frontends_rep vp9_loop_filter_v_16, 16, 4, 4, v
function vp9_loop_filter_h_16_4_16_neon
sub r12, r0, #16
sub r0, r0, #8
vld1.16 {d16}, [r12,:64], r1
vld1.16 {d20}, [r0, :64], r1
vld1.16 {d17}, [r12,:64], r1
vld1.16 {d21}, [r0, :64], r1
vld1.16 {d18}, [r12,:64], r1
vld1.16 {d22}, [r0, :64], r1
vld1.16 {d19}, [r12,:64], r1
vld1.16 {d23}, [r0, :64], r1
sub r12, r12, r1, lsl #2
sub r0, r0, r1, lsl #2
add r12, r12, #16
add r0, r0, #16
vld1.16 {d24}, [r12,:64], r1
vld1.16 {d28}, [r0, :64], r1
vld1.16 {d25}, [r12,:64], r1
vld1.16 {d29}, [r0, :64], r1
vld1.16 {d26}, [r12,:64], r1
vld1.16 {d30}, [r0, :64], r1
vld1.16 {d27}, [r12,:64], r1
vld1.16 {d31}, [r0, :64], r1
sub r0, r0, r1, lsl #2
sub r12, r12, r1, lsl #2
sub r12, r12, #16
sub r0, r0, #16
@ The 16x4 pixels read above is in four 4x4 blocks
transpose16_q_4x4 q8, q9, d16, d17, d18, d19
transpose16_q_4x4 q10, q11, d20, d21, d22, d23
transpose16_q_4x4 q12, q13, d24, d25, d26, d27
transpose16_q_4x4 q14, q15, d28, d29, d30, d31
loop_filter_16
@ Transpose back; this is the same transpose as above, but
@ we can't take advantage of q registers for the transpose, since
@ all d registers in the transpose aren't consecutive.
transpose16_4x4 d16, d2, d3, d4
transpose16_4x4 d5, d6, d8, d9
transpose16_4x4 d10, d11, d12, d13
transpose16_4x4 d14, d15, d17, d31
vst1.16 {d16}, [r12,:64], r1
vst1.16 {d5}, [r0, :64], r1
vst1.16 {d2}, [r12,:64], r1
vst1.16 {d6}, [r0, :64], r1
vst1.16 {d3}, [r12,:64], r1
vst1.16 {d8}, [r0, :64], r1
vst1.16 {d4}, [r12,:64], r1
vst1.16 {d9}, [r0, :64], r1
sub r12, r12, r1, lsl #2
sub r0, r0, r1, lsl #2
add r12, r12, #16
add r0, r0, #16
vst1.16 {d10}, [r12,:64], r1
vst1.16 {d14}, [r0, :64], r1
vst1.16 {d11}, [r12,:64], r1
vst1.16 {d15}, [r0, :64], r1
vst1.16 {d12}, [r12,:64], r1
vst1.16 {d17}, [r0, :64], r1
vst1.16 {d13}, [r12,:64], r1
vst1.16 {d31}, [r0, :64], r1
sub r0, r0, r1, lsl #2
sub r0, r0, #8
bx lr
9:
add r0, r0, #8
bx lr
8:
add r12, r12, #8
add r0, r0, #8
transpose16_q_4x4 q10, q11, d20, d21, d22, d23
transpose16_q_4x4 q12, q13, d24, d25, d26, d27
vst1.16 {d20}, [r12,:64], r1
vst1.16 {d24}, [r0, :64], r1
vst1.16 {d21}, [r12,:64], r1
vst1.16 {d25}, [r0, :64], r1
vst1.16 {d22}, [r12,:64], r1
vst1.16 {d26}, [r0, :64], r1
vst1.16 {d23}, [r12,:64], r1
vst1.16 {d27}, [r0, :64], r1
sub r0, r0, r1, lsl #2
bx lr
7:
add r12, r12, #12
add r0, r12, r1, lsl #1
transpose16_q_4x4 q11, q12, d22, d23, d24, d25
vst1.16 {d22}, [r12], r1
vst1.16 {d24}, [r0], r1
vst1.16 {d23}, [r12], r1
vst1.16 {d25}, [r0], r1
sub r0, r0, r1, lsl #2
add r0, r0, #4
bx lr
endfunc
bpp_frontends_rep vp9_loop_filter_h_16, 8, 4, 2, h
bpp_frontends_rep vp9_loop_filter_h_16, 16, 4, 4, h