annotate arm/h264dsp_neon.S @ 9956:0448219eb490 libavcodec

IIR filter test program.
author alexc
date Thu, 16 Jul 2009 22:17:20 +0000
parents f5ffd813dc7f
children be725249ea67
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
1 /*
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
2 * Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
3 *
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
4 * This file is part of FFmpeg.
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
5 *
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
6 * FFmpeg is free software; you can redistribute it and/or
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
8 * License as published by the Free Software Foundation; either
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
9 * version 2.1 of the License, or (at your option) any later version.
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
10 *
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
11 * FFmpeg is distributed in the hope that it will be useful,
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
14 * Lesser General Public License for more details.
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
15 *
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
17 * License along with FFmpeg; if not, write to the Free Software
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
19 */
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
20
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
21 #include "asm.S"
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
22
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
23 .fpu neon
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
24
8338
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
25 .macro transpose_8x8 r0 r1 r2 r3 r4 r5 r6 r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
26 vtrn.32 \r0, \r4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
27 vtrn.32 \r1, \r5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
28 vtrn.32 \r2, \r6
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
29 vtrn.32 \r3, \r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
30 vtrn.16 \r0, \r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
31 vtrn.16 \r1, \r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
32 vtrn.16 \r4, \r6
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
33 vtrn.16 \r5, \r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
34 vtrn.8 \r0, \r1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
35 vtrn.8 \r2, \r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
36 vtrn.8 \r4, \r5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
37 vtrn.8 \r6, \r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
38 .endm
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
39
9864
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
40 .macro transpose_4x4 r0 r1 r2 r3
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
41 vtrn.16 \r0, \r2
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
42 vtrn.16 \r1, \r3
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
43 vtrn.8 \r0, \r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
44 vtrn.8 \r2, \r3
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
45 .endm
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
46
8338
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
47 .macro swap4 r0 r1 r2 r3 r4 r5 r6 r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
48 vswp \r0, \r4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
49 vswp \r1, \r5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
50 vswp \r2, \r6
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
51 vswp \r3, \r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
52 .endm
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
53
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
54 .macro transpose16_4x4 r0 r1 r2 r3 r4 r5 r6 r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
55 vtrn.32 \r0, \r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
56 vtrn.32 \r1, \r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
57 vtrn.32 \r4, \r6
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
58 vtrn.32 \r5, \r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
59 vtrn.16 \r0, \r1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
60 vtrn.16 \r2, \r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
61 vtrn.16 \r4, \r5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
62 vtrn.16 \r6, \r7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
63 .endm
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
64
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
65 /* chroma_mc8(uint8_t *dst, uint8_t *src, int stride, int h, int x, int y) */
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
66 .macro h264_chroma_mc8 type
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
67 function ff_\type\()_h264_chroma_mc8_neon, export=1
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
68 push {r4-r7, lr}
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
69 ldrd r4, [sp, #20]
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
70 .ifc \type,avg
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
71 mov lr, r0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
72 .endif
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
73 pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
74 pld [r1, r2]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
75
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
76 muls r7, r4, r5
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
77 rsb r6, r7, r5, lsl #3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
78 rsb ip, r7, r4, lsl #3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
79 sub r4, r7, r4, lsl #3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
80 sub r4, r4, r5, lsl #3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
81 add r4, r4, #64
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
82
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
83 beq 2f
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
84
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
85 add r5, r1, r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
86
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
87 vdup.8 d0, r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
88 lsl r4, r2, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
89 vdup.8 d1, ip
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
90 vld1.64 {d4, d5}, [r1], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
91 vdup.8 d2, r6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
92 vld1.64 {d6, d7}, [r5], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
93 vdup.8 d3, r7
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
94
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
95 vext.8 d5, d4, d5, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
96 vext.8 d7, d6, d7, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
97
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
98 1: pld [r5]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
99 vmull.u8 q8, d4, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
100 vmlal.u8 q8, d5, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
101 vld1.64 {d4, d5}, [r1], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
102 vmlal.u8 q8, d6, d2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
103 vext.8 d5, d4, d5, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
104 vmlal.u8 q8, d7, d3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
105 vmull.u8 q9, d6, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
106 subs r3, r3, #2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
107 vmlal.u8 q9, d7, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
108 vmlal.u8 q9, d4, d2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
109 vmlal.u8 q9, d5, d3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
110 vrshrn.u16 d16, q8, #6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
111 vld1.64 {d6, d7}, [r5], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
112 pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
113 vrshrn.u16 d17, q9, #6
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
114 .ifc \type,avg
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
115 vld1.64 {d20}, [lr,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
116 vld1.64 {d21}, [lr,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
117 vrhadd.u8 q8, q8, q10
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
118 .endif
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
119 vext.8 d7, d6, d7, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
120 vst1.64 {d16}, [r0,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
121 vst1.64 {d17}, [r0,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
122 bgt 1b
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
123
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
124 pop {r4-r7, pc}
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
125
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
126 2: tst r6, r6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
127 add ip, ip, r6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
128 vdup.8 d0, r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
129 vdup.8 d1, ip
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
130
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
131 beq 4f
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
132
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
133 add r5, r1, r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
134 lsl r4, r2, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
135 vld1.64 {d4}, [r1], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
136 vld1.64 {d6}, [r5], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
137
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
138 3: pld [r5]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
139 vmull.u8 q8, d4, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
140 vmlal.u8 q8, d6, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
141 vld1.64 {d4}, [r1], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
142 vmull.u8 q9, d6, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
143 vmlal.u8 q9, d4, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
144 vld1.64 {d6}, [r5], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
145 vrshrn.u16 d16, q8, #6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
146 vrshrn.u16 d17, q9, #6
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
147 .ifc \type,avg
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
148 vld1.64 {d20}, [lr,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
149 vld1.64 {d21}, [lr,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
150 vrhadd.u8 q8, q8, q10
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
151 .endif
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
152 subs r3, r3, #2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
153 pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
154 vst1.64 {d16}, [r0,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
155 vst1.64 {d17}, [r0,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
156 bgt 3b
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
157
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
158 pop {r4-r7, pc}
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
159
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
160 4: vld1.64 {d4, d5}, [r1], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
161 vld1.64 {d6, d7}, [r1], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
162 vext.8 d5, d4, d5, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
163 vext.8 d7, d6, d7, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
164
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
165 5: pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
166 subs r3, r3, #2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
167 vmull.u8 q8, d4, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
168 vmlal.u8 q8, d5, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
169 vld1.64 {d4, d5}, [r1], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
170 vmull.u8 q9, d6, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
171 vmlal.u8 q9, d7, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
172 pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
173 vext.8 d5, d4, d5, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
174 vrshrn.u16 d16, q8, #6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
175 vrshrn.u16 d17, q9, #6
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
176 .ifc \type,avg
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
177 vld1.64 {d20}, [lr,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
178 vld1.64 {d21}, [lr,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
179 vrhadd.u8 q8, q8, q10
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
180 .endif
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
181 vld1.64 {d6, d7}, [r1], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
182 vext.8 d7, d6, d7, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
183 vst1.64 {d16}, [r0,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
184 vst1.64 {d17}, [r0,:64], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
185 bgt 5b
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
186
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
187 pop {r4-r7, pc}
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
188 .endfunc
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
189 .endm
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
190
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
191 /* chroma_mc4(uint8_t *dst, uint8_t *src, int stride, int h, int x, int y) */
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
192 .macro h264_chroma_mc4 type
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
193 function ff_\type\()_h264_chroma_mc4_neon, export=1
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
194 push {r4-r7, lr}
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
195 ldrd r4, [sp, #20]
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
196 .ifc \type,avg
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
197 mov lr, r0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
198 .endif
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
199 pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
200 pld [r1, r2]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
201
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
202 muls r7, r4, r5
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
203 rsb r6, r7, r5, lsl #3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
204 rsb ip, r7, r4, lsl #3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
205 sub r4, r7, r4, lsl #3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
206 sub r4, r4, r5, lsl #3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
207 add r4, r4, #64
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
208
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
209 beq 2f
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
210
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
211 add r5, r1, r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
212
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
213 vdup.8 d0, r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
214 lsl r4, r2, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
215 vdup.8 d1, ip
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
216 vld1.64 {d4}, [r1], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
217 vdup.8 d2, r6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
218 vld1.64 {d6}, [r5], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
219 vdup.8 d3, r7
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
220
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
221 vext.8 d5, d4, d5, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
222 vext.8 d7, d6, d7, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
223 vtrn.32 d4, d5
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
224 vtrn.32 d6, d7
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
225
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
226 vtrn.32 d0, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
227 vtrn.32 d2, d3
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
228
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
229 1: pld [r5]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
230 vmull.u8 q8, d4, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
231 vmlal.u8 q8, d6, d2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
232 vld1.64 {d4}, [r1], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
233 vext.8 d5, d4, d5, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
234 vtrn.32 d4, d5
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
235 vmull.u8 q9, d6, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
236 vmlal.u8 q9, d4, d2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
237 vld1.64 {d6}, [r5], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
238 vadd.i16 d16, d16, d17
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
239 vadd.i16 d17, d18, d19
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
240 vrshrn.u16 d16, q8, #6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
241 subs r3, r3, #2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
242 pld [r1]
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
243 .ifc \type,avg
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
244 vld1.32 {d20[0]}, [lr,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
245 vld1.32 {d20[1]}, [lr,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
246 vrhadd.u8 d16, d16, d20
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
247 .endif
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
248 vext.8 d7, d6, d7, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
249 vtrn.32 d6, d7
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
250 vst1.32 {d16[0]}, [r0,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
251 vst1.32 {d16[1]}, [r0,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
252 bgt 1b
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
253
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
254 pop {r4-r7, pc}
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
255
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
256 2: tst r6, r6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
257 add ip, ip, r6
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
258 vdup.8 d0, r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
259 vdup.8 d1, ip
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
260 vtrn.32 d0, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
261
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
262 beq 4f
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
263
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
264 vext.32 d1, d0, d1, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
265 add r5, r1, r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
266 lsl r4, r2, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
267 vld1.32 {d4[0]}, [r1], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
268 vld1.32 {d4[1]}, [r5], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
269
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
270 3: pld [r5]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
271 vmull.u8 q8, d4, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
272 vld1.32 {d4[0]}, [r1], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
273 vmull.u8 q9, d4, d1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
274 vld1.32 {d4[1]}, [r5], r4
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
275 vadd.i16 d16, d16, d17
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
276 vadd.i16 d17, d18, d19
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
277 vrshrn.u16 d16, q8, #6
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
278 .ifc \type,avg
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
279 vld1.32 {d20[0]}, [lr,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
280 vld1.32 {d20[1]}, [lr,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
281 vrhadd.u8 d16, d16, d20
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
282 .endif
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
283 subs r3, r3, #2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
284 pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
285 vst1.32 {d16[0]}, [r0,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
286 vst1.32 {d16[1]}, [r0,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
287 bgt 3b
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
288
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
289 pop {r4-r7, pc}
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
290
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
291 4: vld1.64 {d4}, [r1], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
292 vld1.64 {d6}, [r1], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
293 vext.8 d5, d4, d5, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
294 vext.8 d7, d6, d7, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
295 vtrn.32 d4, d5
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
296 vtrn.32 d6, d7
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
297
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
298 5: vmull.u8 q8, d4, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
299 vmull.u8 q9, d6, d0
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
300 subs r3, r3, #2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
301 vld1.64 {d4}, [r1], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
302 vext.8 d5, d4, d5, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
303 vtrn.32 d4, d5
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
304 vadd.i16 d16, d16, d17
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
305 vadd.i16 d17, d18, d19
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
306 pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
307 vrshrn.u16 d16, q8, #6
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
308 .ifc \type,avg
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
309 vld1.32 {d20[0]}, [lr,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
310 vld1.32 {d20[1]}, [lr,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
311 vrhadd.u8 d16, d16, d20
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
312 .endif
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
313 vld1.64 {d6}, [r1], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
314 vext.8 d7, d6, d7, #1
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
315 vtrn.32 d6, d7
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
316 pld [r1]
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
317 vst1.32 {d16[0]}, [r0,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
318 vst1.32 {d16[1]}, [r0,:32], r2
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
319 bgt 5b
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
320
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
321 pop {r4-r7, pc}
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
322 .endfunc
8336
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
323 .endm
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
324
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
325 .text
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
326 .align
c8401acb05d1 ARM: NEON optimised {put,avg}_h264_chroma_mc[48]
mru
parents:
diff changeset
327
8626
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
328 h264_chroma_mc8 put
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
329 h264_chroma_mc8 avg
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
330 h264_chroma_mc4 put
8d425ee85ddb ARM: simplify ff_put/avg_h264_chroma_mc4/8_neon definitions, no code change
mru
parents: 8359
diff changeset
331 h264_chroma_mc4 avg
8337
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
332
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
333 /* H.264 loop filter */
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
334
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
335 .macro h264_loop_filter_start
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
336 ldr ip, [sp]
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
337 tst r2, r2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
338 ldr ip, [ip]
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
339 tstne r3, r3
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
340 vmov.32 d24[0], ip
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
341 and ip, ip, ip, lsl #16
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
342 bxeq lr
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
343 ands ip, ip, ip, lsl #8
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
344 bxlt lr
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
345 .endm
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
346
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
347 .macro align_push_regs
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
348 and ip, sp, #15
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
349 add ip, ip, #32
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
350 sub sp, sp, ip
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
351 vst1.64 {d12-d15}, [sp,:128]
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
352 sub sp, sp, #32
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
353 vst1.64 {d8-d11}, [sp,:128]
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
354 .endm
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
355
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
356 .macro align_pop_regs
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
357 vld1.64 {d8-d11}, [sp,:128]!
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
358 vld1.64 {d12-d15}, [sp,:128], ip
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
359 .endm
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
360
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
361 .macro h264_loop_filter_luma
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
362 vdup.8 q11, r2 @ alpha
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
363 vmovl.u8 q12, d24
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
364 vabd.u8 q6, q8, q0 @ abs(p0 - q0)
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
365 vmovl.u16 q12, d24
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
366 vabd.u8 q14, q9, q8 @ abs(p1 - p0)
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
367 vsli.16 q12, q12, #8
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
368 vabd.u8 q15, q1, q0 @ abs(q1 - q0)
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
369 vsli.32 q12, q12, #16
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
370 vclt.u8 q6, q6, q11 @ < alpha
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
371 vdup.8 q11, r3 @ beta
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
372 vclt.s8 q7, q12, #0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
373 vclt.u8 q14, q14, q11 @ < beta
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
374 vclt.u8 q15, q15, q11 @ < beta
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
375 vbic q6, q6, q7
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
376 vabd.u8 q4, q10, q8 @ abs(p2 - p0)
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
377 vand q6, q6, q14
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
378 vabd.u8 q5, q2, q0 @ abs(q2 - q0)
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
379 vclt.u8 q4, q4, q11 @ < beta
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
380 vand q6, q6, q15
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
381 vclt.u8 q5, q5, q11 @ < beta
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
382 vand q4, q4, q6
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
383 vand q5, q5, q6
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
384 vand q12, q12, q6
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
385 vrhadd.u8 q14, q8, q0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
386 vsub.i8 q6, q12, q4
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
387 vqadd.u8 q7, q9, q12
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
388 vhadd.u8 q10, q10, q14
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
389 vsub.i8 q6, q6, q5
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
390 vhadd.u8 q14, q2, q14
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
391 vmin.u8 q7, q7, q10
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
392 vqsub.u8 q11, q9, q12
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
393 vqadd.u8 q2, q1, q12
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
394 vmax.u8 q7, q7, q11
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
395 vqsub.u8 q11, q1, q12
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
396 vmin.u8 q14, q2, q14
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
397 vmovl.u8 q2, d0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
398 vmax.u8 q14, q14, q11
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
399 vmovl.u8 q10, d1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
400 vsubw.u8 q2, q2, d16
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
401 vsubw.u8 q10, q10, d17
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
402 vshl.i16 q2, q2, #2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
403 vshl.i16 q10, q10, #2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
404 vaddw.u8 q2, q2, d18
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
405 vaddw.u8 q10, q10, d19
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
406 vsubw.u8 q2, q2, d2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
407 vsubw.u8 q10, q10, d3
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
408 vrshrn.i16 d4, q2, #3
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
409 vrshrn.i16 d5, q10, #3
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
410 vbsl q4, q7, q9
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
411 vbsl q5, q14, q1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
412 vneg.s8 q7, q6
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
413 vmovl.u8 q14, d16
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
414 vmin.s8 q2, q2, q6
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
415 vmovl.u8 q6, d17
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
416 vmax.s8 q2, q2, q7
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
417 vmovl.u8 q11, d0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
418 vmovl.u8 q12, d1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
419 vaddw.s8 q14, q14, d4
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
420 vaddw.s8 q6, q6, d5
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
421 vsubw.s8 q11, q11, d4
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
422 vsubw.s8 q12, q12, d5
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
423 vqmovun.s16 d16, q14
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
424 vqmovun.s16 d17, q6
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
425 vqmovun.s16 d0, q11
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
426 vqmovun.s16 d1, q12
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
427 .endm
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
428
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
429 function ff_h264_v_loop_filter_luma_neon, export=1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
430 h264_loop_filter_start
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
431
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
432 vld1.64 {d0, d1}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
433 vld1.64 {d2, d3}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
434 vld1.64 {d4, d5}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
435 sub r0, r0, r1, lsl #2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
436 sub r0, r0, r1, lsl #1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
437 vld1.64 {d20,d21}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
438 vld1.64 {d18,d19}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
439 vld1.64 {d16,d17}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
440
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
441 align_push_regs
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
442
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
443 h264_loop_filter_luma
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
444
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
445 sub r0, r0, r1, lsl #1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
446 vst1.64 {d8, d9}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
447 vst1.64 {d16,d17}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
448 vst1.64 {d0, d1}, [r0,:128], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
449 vst1.64 {d10,d11}, [r0,:128]
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
450
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
451 align_pop_regs
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
452 bx lr
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
453 .endfunc
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
454
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
455 function ff_h264_h_loop_filter_luma_neon, export=1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
456 h264_loop_filter_start
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
457
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
458 sub r0, r0, #4
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
459 vld1.64 {d6}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
460 vld1.64 {d20}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
461 vld1.64 {d18}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
462 vld1.64 {d16}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
463 vld1.64 {d0}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
464 vld1.64 {d2}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
465 vld1.64 {d4}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
466 vld1.64 {d26}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
467 vld1.64 {d7}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
468 vld1.64 {d21}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
469 vld1.64 {d19}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
470 vld1.64 {d17}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
471 vld1.64 {d1}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
472 vld1.64 {d3}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
473 vld1.64 {d5}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
474 vld1.64 {d27}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
475
8338
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
476 transpose_8x8 q3, q10, q9, q8, q0, q1, q2, q13
8337
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
477
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
478 align_push_regs
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
479
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
480 h264_loop_filter_luma
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
481
9864
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
482 transpose_4x4 q4, q8, q0, q5
8337
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
483
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
484 sub r0, r0, r1, lsl #4
9864
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
485 add r0, r0, #2
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
486 vst1.32 {d8[0]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
487 vst1.32 {d16[0]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
488 vst1.32 {d0[0]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
489 vst1.32 {d10[0]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
490 vst1.32 {d8[1]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
491 vst1.32 {d16[1]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
492 vst1.32 {d0[1]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
493 vst1.32 {d10[1]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
494 vst1.32 {d9[0]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
495 vst1.32 {d17[0]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
496 vst1.32 {d1[0]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
497 vst1.32 {d11[0]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
498 vst1.32 {d9[1]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
499 vst1.32 {d17[1]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
500 vst1.32 {d1[1]}, [r0], r1
f5ffd813dc7f ARM: slightly faster NEON H264 horizontal loop filter
mru
parents: 9072
diff changeset
501 vst1.32 {d11[1]}, [r0], r1
8337
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
502
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
503 align_pop_regs
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
504 bx lr
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
505 .endfunc
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
506
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
507 .macro h264_loop_filter_chroma
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
508 vdup.8 d22, r2 @ alpha
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
509 vmovl.u8 q12, d24
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
510 vabd.u8 d26, d16, d0 @ abs(p0 - q0)
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
511 vmovl.u8 q2, d0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
512 vabd.u8 d28, d18, d16 @ abs(p1 - p0)
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
513 vsubw.u8 q2, q2, d16
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
514 vsli.16 d24, d24, #8
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
515 vshl.i16 q2, q2, #2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
516 vabd.u8 d30, d2, d0 @ abs(q1 - q0)
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
517 vaddw.u8 q2, q2, d18
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
518 vclt.u8 d26, d26, d22 @ < alpha
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
519 vsubw.u8 q2, q2, d2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
520 vdup.8 d22, r3 @ beta
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
521 vclt.s8 d25, d24, #0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
522 vrshrn.i16 d4, q2, #3
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
523 vclt.u8 d28, d28, d22 @ < beta
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
524 vbic d26, d26, d25
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
525 vclt.u8 d30, d30, d22 @ < beta
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
526 vand d26, d26, d28
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
527 vneg.s8 d25, d24
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
528 vand d26, d26, d30
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
529 vmin.s8 d4, d4, d24
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
530 vmovl.u8 q14, d16
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
531 vand d4, d4, d26
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
532 vmax.s8 d4, d4, d25
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
533 vmovl.u8 q11, d0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
534 vaddw.s8 q14, q14, d4
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
535 vsubw.s8 q11, q11, d4
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
536 vqmovun.s16 d16, q14
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
537 vqmovun.s16 d0, q11
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
538 .endm
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
539
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
540 function ff_h264_v_loop_filter_chroma_neon, export=1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
541 h264_loop_filter_start
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
542
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
543 sub r0, r0, r1, lsl #1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
544 vld1.64 {d18}, [r0,:64], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
545 vld1.64 {d16}, [r0,:64], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
546 vld1.64 {d0}, [r0,:64], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
547 vld1.64 {d2}, [r0,:64]
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
548
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
549 h264_loop_filter_chroma
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
550
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
551 sub r0, r0, r1, lsl #1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
552 vst1.64 {d16}, [r0,:64], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
553 vst1.64 {d0}, [r0,:64], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
554
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
555 bx lr
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
556 .endfunc
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
557
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
558 function ff_h264_h_loop_filter_chroma_neon, export=1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
559 h264_loop_filter_start
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
560
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
561 sub r0, r0, #2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
562 vld1.32 {d18[0]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
563 vld1.32 {d16[0]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
564 vld1.32 {d0[0]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
565 vld1.32 {d2[0]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
566 vld1.32 {d18[1]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
567 vld1.32 {d16[1]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
568 vld1.32 {d0[1]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
569 vld1.32 {d2[1]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
570
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
571 vtrn.16 d18, d0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
572 vtrn.16 d16, d2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
573 vtrn.8 d18, d16
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
574 vtrn.8 d0, d2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
575
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
576 h264_loop_filter_chroma
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
577
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
578 vtrn.16 d18, d0
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
579 vtrn.16 d16, d2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
580 vtrn.8 d18, d16
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
581 vtrn.8 d0, d2
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
582
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
583 sub r0, r0, r1, lsl #3
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
584 vst1.32 {d18[0]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
585 vst1.32 {d16[0]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
586 vst1.32 {d0[0]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
587 vst1.32 {d2[0]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
588 vst1.32 {d18[1]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
589 vst1.32 {d16[1]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
590 vst1.32 {d0[1]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
591 vst1.32 {d2[1]}, [r0], r1
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
592
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
593 bx lr
d43b7f4c5c1c ARM: NEON optimised H.264 loop filter
mru
parents: 8336
diff changeset
594 .endfunc
8338
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
595
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
596 /* H.264 qpel MC */
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
597
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
598 .macro lowpass_const r
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
599 movw \r, #5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
600 movt \r, #20
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
601 vmov.32 d6[0], \r
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
602 .endm
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
603
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
604 .macro lowpass_8 r0, r1, r2, r3, d0, d1, narrow=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
605 .if \narrow
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
606 t0 .req q0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
607 t1 .req q8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
608 .else
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
609 t0 .req \d0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
610 t1 .req \d1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
611 .endif
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
612 vext.8 d2, \r0, \r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
613 vext.8 d3, \r0, \r1, #3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
614 vaddl.u8 q1, d2, d3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
615 vext.8 d4, \r0, \r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
616 vext.8 d5, \r0, \r1, #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
617 vaddl.u8 q2, d4, d5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
618 vext.8 d30, \r0, \r1, #5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
619 vaddl.u8 t0, \r0, d30
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
620 vext.8 d18, \r2, \r3, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
621 vmla.i16 t0, q1, d6[1]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
622 vext.8 d19, \r2, \r3, #3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
623 vaddl.u8 q9, d18, d19
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
624 vext.8 d20, \r2, \r3, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
625 vmls.i16 t0, q2, d6[0]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
626 vext.8 d21, \r2, \r3, #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
627 vaddl.u8 q10, d20, d21
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
628 vext.8 d31, \r2, \r3, #5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
629 vaddl.u8 t1, \r2, d31
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
630 vmla.i16 t1, q9, d6[1]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
631 vmls.i16 t1, q10, d6[0]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
632 .if \narrow
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
633 vqrshrun.s16 \d0, t0, #5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
634 vqrshrun.s16 \d1, t1, #5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
635 .endif
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
636 .unreq t0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
637 .unreq t1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
638 .endm
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
639
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
640 .macro lowpass_8_1 r0, r1, d0, narrow=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
641 .if \narrow
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
642 t0 .req q0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
643 .else
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
644 t0 .req \d0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
645 .endif
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
646 vext.8 d2, \r0, \r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
647 vext.8 d3, \r0, \r1, #3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
648 vaddl.u8 q1, d2, d3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
649 vext.8 d4, \r0, \r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
650 vext.8 d5, \r0, \r1, #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
651 vaddl.u8 q2, d4, d5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
652 vext.8 d30, \r0, \r1, #5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
653 vaddl.u8 t0, \r0, d30
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
654 vmla.i16 t0, q1, d6[1]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
655 vmls.i16 t0, q2, d6[0]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
656 .if \narrow
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
657 vqrshrun.s16 \d0, t0, #5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
658 .endif
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
659 .unreq t0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
660 .endm
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
661
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
662 .macro lowpass_8.16 r0, r1, l0, h0, l1, h1, d
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
663 vext.16 q1, \r0, \r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
664 vext.16 q0, \r0, \r1, #3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
665 vaddl.s16 q9, d2, d0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
666 vext.16 q2, \r0, \r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
667 vaddl.s16 q1, d3, d1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
668 vext.16 q3, \r0, \r1, #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
669 vaddl.s16 q10, d4, d6
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
670 vext.16 \r1, \r0, \r1, #5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
671 vaddl.s16 q2, d5, d7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
672 vaddl.s16 q0, \h0, \h1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
673 vaddl.s16 q8, \l0, \l1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
674
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
675 vshl.i32 q3, q9, #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
676 vshl.i32 q9, q9, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
677 vshl.i32 q15, q10, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
678 vadd.i32 q9, q9, q3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
679 vadd.i32 q10, q10, q15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
680
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
681 vshl.i32 q3, q1, #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
682 vshl.i32 q1, q1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
683 vshl.i32 q15, q2, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
684 vadd.i32 q1, q1, q3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
685 vadd.i32 q2, q2, q15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
686
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
687 vadd.i32 q9, q9, q8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
688 vsub.i32 q9, q9, q10
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
689
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
690 vadd.i32 q1, q1, q0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
691 vsub.i32 q1, q1, q2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
692
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
693 vrshrn.s32 d18, q9, #10
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
694 vrshrn.s32 d19, q1, #10
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
695
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
696 vqmovun.s16 \d, q9
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
697 .endm
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
698
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
699 function put_h264_qpel16_h_lowpass_neon_packed
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
700 mov r4, lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
701 mov ip, #16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
702 mov r3, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
703 bl put_h264_qpel8_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
704 sub r1, r1, r2, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
705 add r1, r1, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
706 mov ip, #16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
707 mov lr, r4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
708 b put_h264_qpel8_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
709 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
710
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
711 function put_h264_qpel16_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
712 push {lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
713 mov ip, #16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
714 bl put_h264_qpel8_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
715 sub r0, r0, r3, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
716 sub r1, r1, r2, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
717 add r0, r0, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
718 add r1, r1, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
719 mov ip, #16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
720 pop {lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
721 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
722
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
723 function put_h264_qpel8_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
724 1: vld1.64 {d0, d1}, [r1], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
725 vld1.64 {d16,d17}, [r1], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
726 subs ip, ip, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
727 lowpass_8 d0, d1, d16, d17, d0, d16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
728 vst1.64 {d0}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
729 vst1.64 {d16}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
730 bne 1b
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
731 bx lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
732 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
733
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
734 function put_h264_qpel16_h_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
735 push {lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
736 mov ip, #16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
737 bl put_h264_qpel8_h_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
738 sub r0, r0, r2, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
739 sub r1, r1, r2, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
740 sub r3, r3, r2, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
741 add r0, r0, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
742 add r1, r1, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
743 add r3, r3, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
744 mov ip, #16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
745 pop {lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
746 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
747
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
748 function put_h264_qpel8_h_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
749 1: vld1.64 {d0, d1}, [r1], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
750 vld1.64 {d16,d17}, [r1], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
751 vld1.64 {d28}, [r3], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
752 vld1.64 {d29}, [r3], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
753 subs ip, ip, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
754 lowpass_8 d0, d1, d16, d17, d0, d1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
755 vrhadd.u8 q0, q0, q14
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
756 vst1.64 {d0}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
757 vst1.64 {d1}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
758 bne 1b
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
759 bx lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
760 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
761
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
762 function put_h264_qpel16_v_lowpass_neon_packed
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
763 mov r4, lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
764 mov r2, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
765 bl put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
766 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
767 bl put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
768 sub r1, r1, r3, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
769 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
770 add r1, r1, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
771 bl put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
772 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
773 mov lr, r4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
774 b put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
775 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
776
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
777 function put_h264_qpel16_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
778 mov r4, lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
779 bl put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
780 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
781 bl put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
782 sub r0, r0, r2, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
783 add r0, r0, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
784 sub r1, r1, r3, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
785 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
786 add r1, r1, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
787 bl put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
788 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
789 mov lr, r4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
790 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
791
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
792 function put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
793 vld1.64 {d8}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
794 vld1.64 {d10}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
795 vld1.64 {d12}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
796 vld1.64 {d14}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
797 vld1.64 {d22}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
798 vld1.64 {d24}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
799 vld1.64 {d26}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
800 vld1.64 {d28}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
801 vld1.64 {d9}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
802 vld1.64 {d11}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
803 vld1.64 {d13}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
804 vld1.64 {d15}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
805 vld1.64 {d23}, [r1]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
806
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
807 transpose_8x8 q4, q5, q6, q7, q11, q12, q13, q14
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
808 lowpass_8 d8, d9, d10, d11, d8, d10
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
809 lowpass_8 d12, d13, d14, d15, d12, d14
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
810 lowpass_8 d22, d23, d24, d25, d22, d24
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
811 lowpass_8 d26, d27, d28, d29, d26, d28
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
812 transpose_8x8 d8, d10, d12, d14, d22, d24, d26, d28
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
813
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
814 vst1.64 {d8}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
815 vst1.64 {d10}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
816 vst1.64 {d12}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
817 vst1.64 {d14}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
818 vst1.64 {d22}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
819 vst1.64 {d24}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
820 vst1.64 {d26}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
821 vst1.64 {d28}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
822
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
823 bx lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
824 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
825
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
826 function put_h264_qpel16_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
827 mov r4, lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
828 bl put_h264_qpel8_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
829 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
830 bl put_h264_qpel8_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
831 sub r0, r0, r3, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
832 sub ip, ip, r2, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
833 add r0, r0, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
834 add ip, ip, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
835 sub r1, r1, r3, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
836 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
837 add r1, r1, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
838 bl put_h264_qpel8_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
839 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
840 mov lr, r4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
841 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
842
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
843 function put_h264_qpel8_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
844 vld1.64 {d8}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
845 vld1.64 {d10}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
846 vld1.64 {d12}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
847 vld1.64 {d14}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
848 vld1.64 {d22}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
849 vld1.64 {d24}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
850 vld1.64 {d26}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
851 vld1.64 {d28}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
852 vld1.64 {d9}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
853 vld1.64 {d11}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
854 vld1.64 {d13}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
855 vld1.64 {d15}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
856 vld1.64 {d23}, [r1]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
857
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
858 transpose_8x8 q4, q5, q6, q7, q11, q12, q13, q14
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
859 lowpass_8 d8, d9, d10, d11, d8, d9
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
860 lowpass_8 d12, d13, d14, d15, d12, d13
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
861 lowpass_8 d22, d23, d24, d25, d22, d23
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
862 lowpass_8 d26, d27, d28, d29, d26, d27
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
863 transpose_8x8 d8, d9, d12, d13, d22, d23, d26, d27
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
864
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
865 vld1.64 {d0}, [ip], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
866 vld1.64 {d1}, [ip], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
867 vld1.64 {d2}, [ip], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
868 vld1.64 {d3}, [ip], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
869 vld1.64 {d4}, [ip], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
870 vrhadd.u8 q0, q0, q4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
871 vld1.64 {d5}, [ip], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
872 vrhadd.u8 q1, q1, q6
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
873 vld1.64 {d10}, [ip], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
874 vrhadd.u8 q2, q2, q11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
875 vld1.64 {d11}, [ip], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
876
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
877 vst1.64 {d0}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
878 vst1.64 {d1}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
879 vrhadd.u8 q5, q5, q13
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
880 vst1.64 {d2}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
881 vst1.64 {d3}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
882 vst1.64 {d4}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
883 vst1.64 {d5}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
884 vst1.64 {d10}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
885 vst1.64 {d11}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
886
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
887 bx lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
888 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
889
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
890 function put_h264_qpel8_hv_lowpass_neon_top
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
891 lowpass_const ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
892 mov ip, #12
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
893 1: vld1.64 {d0, d1}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
894 vld1.64 {d16,d17}, [r1], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
895 subs ip, ip, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
896 lowpass_8 d0, d1, d16, d17, q11, q12, narrow=0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
897 vst1.64 {d22-d25}, [r4,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
898 bne 1b
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
899
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
900 vld1.64 {d0, d1}, [r1]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
901 lowpass_8_1 d0, d1, q12, narrow=0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
902
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
903 mov ip, #-16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
904 add r4, r4, ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
905 vld1.64 {d30,d31}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
906 vld1.64 {d20,d21}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
907 vld1.64 {d18,d19}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
908 vld1.64 {d16,d17}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
909 vld1.64 {d14,d15}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
910 vld1.64 {d12,d13}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
911 vld1.64 {d10,d11}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
912 vld1.64 {d8, d9}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
913 vld1.64 {d6, d7}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
914 vld1.64 {d4, d5}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
915 vld1.64 {d2, d3}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
916 vld1.64 {d0, d1}, [r4,:128]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
917
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
918 swap4 d1, d3, d5, d7, d8, d10, d12, d14
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
919 transpose16_4x4 q0, q1, q2, q3, q4, q5, q6, q7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
920
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
921 swap4 d17, d19, d21, d31, d24, d26, d28, d22
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
922 transpose16_4x4 q8, q9, q10, q15, q12, q13, q14, q11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
923
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
924 vst1.64 {d30,d31}, [r4,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
925 vst1.64 {d6, d7}, [r4,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
926 vst1.64 {d20,d21}, [r4,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
927 vst1.64 {d4, d5}, [r4,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
928 vst1.64 {d18,d19}, [r4,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
929 vst1.64 {d2, d3}, [r4,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
930 vst1.64 {d16,d17}, [r4,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
931 vst1.64 {d0, d1}, [r4,:128]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
932
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
933 lowpass_8.16 q4, q12, d8, d9, d24, d25, d8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
934 lowpass_8.16 q5, q13, d10, d11, d26, d27, d9
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
935 lowpass_8.16 q6, q14, d12, d13, d28, d29, d10
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
936 lowpass_8.16 q7, q11, d14, d15, d22, d23, d11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
937
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
938 vld1.64 {d16,d17}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
939 vld1.64 {d30,d31}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
940 lowpass_8.16 q8, q15, d16, d17, d30, d31, d12
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
941 vld1.64 {d16,d17}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
942 vld1.64 {d30,d31}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
943 lowpass_8.16 q8, q15, d16, d17, d30, d31, d13
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
944 vld1.64 {d16,d17}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
945 vld1.64 {d30,d31}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
946 lowpass_8.16 q8, q15, d16, d17, d30, d31, d14
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
947 vld1.64 {d16,d17}, [r4,:128], ip
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
948 vld1.64 {d30,d31}, [r4,:128]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
949 lowpass_8.16 q8, q15, d16, d17, d30, d31, d15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
950
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
951 transpose_8x8 d12, d13, d14, d15, d8, d9, d10, d11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
952
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
953 bx lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
954 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
955
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
956 function put_h264_qpel8_hv_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
957 mov r10, lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
958 bl put_h264_qpel8_hv_lowpass_neon_top
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
959 vst1.64 {d12}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
960 vst1.64 {d13}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
961 vst1.64 {d14}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
962 vst1.64 {d15}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
963 vst1.64 {d8}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
964 vst1.64 {d9}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
965 vst1.64 {d10}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
966 vst1.64 {d11}, [r0,:64], r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
967
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
968 mov lr, r10
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
969 bx lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
970 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
971
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
972 function put_h264_qpel8_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
973 mov r10, lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
974 bl put_h264_qpel8_hv_lowpass_neon_top
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
975
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
976 vld1.64 {d0, d1}, [r2,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
977 vld1.64 {d2, d3}, [r2,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
978 vrhadd.u8 q0, q0, q6
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
979 vld1.64 {d4, d5}, [r2,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
980 vrhadd.u8 q1, q1, q7
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
981 vld1.64 {d6, d7}, [r2,:128]!
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
982 vrhadd.u8 q2, q2, q4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
983
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
984 vst1.64 {d0}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
985 vrhadd.u8 q3, q3, q5
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
986 vst1.64 {d1}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
987 vst1.64 {d2}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
988 vst1.64 {d3}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
989 vst1.64 {d4}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
990 vst1.64 {d5}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
991 vst1.64 {d6}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
992 vst1.64 {d7}, [r0,:64], r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
993
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
994 mov lr, r10
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
995 bx lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
996 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
997
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
998 function put_h264_qpel16_hv_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
999 mov r9, lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1000 bl put_h264_qpel8_hv_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1001 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1002 bl put_h264_qpel8_hv_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1003 sub r1, r1, r3, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1004 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1005 add r1, r1, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1006 sub r0, r0, r2, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1007 add r0, r0, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1008 bl put_h264_qpel8_hv_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1009 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1010 mov lr, r9
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1011 b put_h264_qpel8_hv_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1012 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1013
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1014 function put_h264_qpel16_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1015 mov r9, lr
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1016 sub r2, r4, #256
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1017 bl put_h264_qpel8_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1018 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1019 bl put_h264_qpel8_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1020 sub r1, r1, r3, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1021 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1022 add r1, r1, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1023 sub r0, r0, r3, lsl #4
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1024 add r0, r0, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1025 bl put_h264_qpel8_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1026 sub r1, r1, r3, lsl #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1027 mov lr, r9
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1028 b put_h264_qpel8_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1029 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1030
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1031 function ff_put_h264_qpel8_mc10_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1032 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1033 mov r3, r1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1034 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1035 mov ip, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1036 b put_h264_qpel8_h_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1037 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1038
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1039 function ff_put_h264_qpel8_mc20_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1040 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1041 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1042 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1043 mov ip, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1044 b put_h264_qpel8_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1045 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1046
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1047 function ff_put_h264_qpel8_mc30_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1048 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1049 add r3, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1050 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1051 mov ip, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1052 b put_h264_qpel8_h_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1053 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1054
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1055 function ff_put_h264_qpel8_mc01_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1056 push {lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1057 mov ip, r1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1058 put_h264_qpel8_mc01:
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1059 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1060 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1061 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1062 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1063 bl put_h264_qpel8_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1064 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1065 pop {pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1066 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1067
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1068 function ff_put_h264_qpel8_mc11_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1069 push {r0, r1, r2, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1070 put_h264_qpel8_mc11:
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1071 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1072 sub sp, sp, #64
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1073 mov r0, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1074 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1075 mov r3, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1076 mov ip, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1077 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1078 bl put_h264_qpel8_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1079 ldrd r0, [sp, #128]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1080 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1081 add ip, sp, #64
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1082 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1083 mov r2, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1084 bl put_h264_qpel8_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1085 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1086 add sp, sp, #76
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1087 pop {pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1088 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1089
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1090 function ff_put_h264_qpel8_mc21_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1091 push {r0, r1, r4, r10, r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1092 put_h264_qpel8_mc21:
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1093 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1094 mov r11, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1095 bic sp, sp, #15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1096 sub sp, sp, #(8*8+16*12)
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1097 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1098 mov r3, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1099 mov r0, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1100 mov ip, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1101 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1102 bl put_h264_qpel8_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1103 mov r4, r0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1104 ldrd r0, [r11]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1105 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1106 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1107 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1108 sub r2, r4, #64
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1109 bl put_h264_qpel8_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1110 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1111 add sp, r11, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1112 pop {r4, r10, r11, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1113 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1114
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1115 function ff_put_h264_qpel8_mc31_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1116 add r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1117 push {r0, r1, r2, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1118 sub r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1119 b put_h264_qpel8_mc11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1120 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1121
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1122 function ff_put_h264_qpel8_mc02_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1123 push {lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1124 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1125 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1126 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1127 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1128 bl put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1129 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1130 pop {pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1131 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1132
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1133 function ff_put_h264_qpel8_mc12_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1134 push {r0, r1, r4, r10, r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1135 put_h264_qpel8_mc12:
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1136 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1137 mov r11, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1138 bic sp, sp, #15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1139 sub sp, sp, #(8*8+16*12)
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1140 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1141 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1142 mov r2, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1143 mov r0, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1144 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1145 bl put_h264_qpel8_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1146 mov r4, r0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1147 ldrd r0, [r11]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1148 sub r1, r1, r3, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1149 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1150 sub r2, r4, #64
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1151 bl put_h264_qpel8_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1152 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1153 add sp, r11, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1154 pop {r4, r10, r11, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1155 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1156
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1157 function ff_put_h264_qpel8_mc22_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1158 push {r4, r10, r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1159 mov r11, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1160 bic sp, sp, #15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1161 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1162 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1163 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1164 sub sp, sp, #(16*12)
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1165 mov r4, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1166 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1167 bl put_h264_qpel8_hv_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1168 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1169 mov sp, r11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1170 pop {r4, r10, r11, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1171 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1172
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1173 function ff_put_h264_qpel8_mc32_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1174 push {r0, r1, r4, r10, r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1175 add r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1176 b put_h264_qpel8_mc12
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1177 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1178
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1179 function ff_put_h264_qpel8_mc03_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1180 push {lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1181 add ip, r1, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1182 b put_h264_qpel8_mc01
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1183 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1184
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1185 function ff_put_h264_qpel8_mc13_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1186 push {r0, r1, r2, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1187 add r1, r1, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1188 b put_h264_qpel8_mc11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1189 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1190
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1191 function ff_put_h264_qpel8_mc23_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1192 push {r0, r1, r4, r10, r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1193 add r1, r1, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1194 b put_h264_qpel8_mc21
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1195 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1196
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1197 function ff_put_h264_qpel8_mc33_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1198 add r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1199 push {r0, r1, r2, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1200 add r1, r1, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1201 sub r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1202 b put_h264_qpel8_mc11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1203 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1204
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1205 function ff_put_h264_qpel16_mc10_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1206 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1207 mov r3, r1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1208 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1209 b put_h264_qpel16_h_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1210 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1211
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1212 function ff_put_h264_qpel16_mc20_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1213 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1214 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1215 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1216 b put_h264_qpel16_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1217 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1218
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1219 function ff_put_h264_qpel16_mc30_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1220 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1221 add r3, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1222 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1223 b put_h264_qpel16_h_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1224 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1225
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1226 function ff_put_h264_qpel16_mc01_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1227 push {r4, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1228 mov ip, r1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1229 put_h264_qpel16_mc01:
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1230 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1231 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1232 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1233 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1234 bl put_h264_qpel16_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1235 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1236 pop {r4, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1237 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1238
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1239 function ff_put_h264_qpel16_mc11_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1240 push {r0, r1, r4, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1241 put_h264_qpel16_mc11:
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1242 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1243 sub sp, sp, #256
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1244 mov r0, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1245 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1246 mov r3, #16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1247 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1248 bl put_h264_qpel16_h_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1249 add r0, sp, #256
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1250 ldrd r0, [r0, #64]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1251 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1252 add ip, sp, #64
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1253 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1254 mov r2, #16
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1255 bl put_h264_qpel16_v_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1256 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1257 add sp, sp, #(256+8)
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1258 pop {r4, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1259 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1260
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1261 function ff_put_h264_qpel16_mc21_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1262 push {r0, r1, r4-r5, r9-r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1263 put_h264_qpel16_mc21:
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1264 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1265 mov r11, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1266 bic sp, sp, #15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1267 sub sp, sp, #(16*16+16*12)
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1268 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1269 mov r0, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1270 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1271 bl put_h264_qpel16_h_lowpass_neon_packed
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1272 mov r4, r0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1273 ldrd r0, [r11]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1274 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1275 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1276 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1277 bl put_h264_qpel16_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1278 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1279 add sp, r11, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1280 pop {r4-r5, r9-r11, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1281 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1282
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1283 function ff_put_h264_qpel16_mc31_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1284 add r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1285 push {r0, r1, r4, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1286 sub r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1287 b put_h264_qpel16_mc11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1288 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1289
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1290 function ff_put_h264_qpel16_mc02_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1291 push {r4, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1292 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1293 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1294 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1295 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1296 bl put_h264_qpel16_v_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1297 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1298 pop {r4, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1299 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1300
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1301 function ff_put_h264_qpel16_mc12_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1302 push {r0, r1, r4-r5, r9-r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1303 put_h264_qpel16_mc12:
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1304 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1305 mov r11, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1306 bic sp, sp, #15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1307 sub sp, sp, #(16*16+16*12)
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1308 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1309 mov r0, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1310 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1311 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1312 bl put_h264_qpel16_v_lowpass_neon_packed
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1313 mov r4, r0
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1314 ldrd r0, [r11]
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1315 sub r1, r1, r3, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1316 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1317 mov r2, r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1318 bl put_h264_qpel16_hv_lowpass_l2_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1319 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1320 add sp, r11, #8
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1321 pop {r4-r5, r9-r11, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1322 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1323
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1324 function ff_put_h264_qpel16_mc22_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1325 push {r4, r9-r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1326 lowpass_const r3
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1327 mov r11, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1328 bic sp, sp, #15
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1329 sub r1, r1, r2, lsl #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1330 sub r1, r1, #2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1331 mov r3, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1332 sub sp, sp, #(16*12)
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1333 mov r4, sp
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1334 vpush {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1335 bl put_h264_qpel16_hv_lowpass_neon
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1336 vpop {d8-d15}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1337 mov sp, r11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1338 pop {r4, r9-r11, pc}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1339 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1340
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1341 function ff_put_h264_qpel16_mc32_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1342 push {r0, r1, r4-r5, r9-r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1343 add r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1344 b put_h264_qpel16_mc12
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1345 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1346
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1347 function ff_put_h264_qpel16_mc03_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1348 push {r4, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1349 add ip, r1, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1350 b put_h264_qpel16_mc01
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1351 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1352
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1353 function ff_put_h264_qpel16_mc13_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1354 push {r0, r1, r4, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1355 add r1, r1, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1356 b put_h264_qpel16_mc11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1357 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1358
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1359 function ff_put_h264_qpel16_mc23_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1360 push {r0, r1, r4-r5, r9-r11, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1361 add r1, r1, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1362 b put_h264_qpel16_mc21
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1363 .endfunc
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1364
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1365 function ff_put_h264_qpel16_mc33_neon, export=1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1366 add r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1367 push {r0, r1, r4, lr}
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1368 add r1, r1, r2
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1369 sub r1, r1, #1
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1370 b put_h264_qpel16_mc11
b294a0d5bc50 ARM: NEON optimised H.264 8x8 and 16x16 qpel MC
mru
parents: 8337
diff changeset
1371 .endfunc
8663
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1372
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1373 @ Biweighted prediction
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1374
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1375 .macro biweight_16 macs, macd
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1376 vdup.8 d0, r4
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1377 vdup.8 d1, r5
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1378 vmov q2, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1379 vmov q3, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1380 1: subs ip, ip, #2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1381 vld1.8 {d20-d21},[r0,:128], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1382 \macd q2, d0, d20
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1383 pld [r0]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1384 \macd q3, d0, d21
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1385 vld1.8 {d22-d23},[r1,:128], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1386 \macs q2, d1, d22
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1387 pld [r1]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1388 \macs q3, d1, d23
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1389 vmov q12, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1390 vld1.8 {d28-d29},[r0,:128], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1391 vmov q13, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1392 \macd q12, d0, d28
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1393 pld [r0]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1394 \macd q13, d0, d29
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1395 vld1.8 {d30-d31},[r1,:128], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1396 \macs q12, d1, d30
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1397 pld [r1]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1398 \macs q13, d1, d31
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1399 vshl.s16 q2, q2, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1400 vshl.s16 q3, q3, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1401 vqmovun.s16 d4, q2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1402 vqmovun.s16 d5, q3
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1403 vshl.s16 q12, q12, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1404 vshl.s16 q13, q13, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1405 vqmovun.s16 d24, q12
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1406 vqmovun.s16 d25, q13
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1407 vmov q3, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1408 vst1.8 {d4- d5}, [r6,:128], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1409 vmov q2, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1410 vst1.8 {d24-d25},[r6,:128], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1411 bne 1b
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1412 pop {r4-r6, pc}
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1413 .endm
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1414
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1415 .macro biweight_8 macs, macd
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1416 vdup.8 d0, r4
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1417 vdup.8 d1, r5
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1418 vmov q1, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1419 vmov q10, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1420 1: subs ip, ip, #2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1421 vld1.8 {d4},[r0,:64], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1422 \macd q1, d0, d4
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1423 pld [r0]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1424 vld1.8 {d5},[r1,:64], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1425 \macs q1, d1, d5
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1426 pld [r1]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1427 vld1.8 {d6},[r0,:64], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1428 \macd q10, d0, d6
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1429 pld [r0]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1430 vld1.8 {d7},[r1,:64], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1431 \macs q10, d1, d7
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1432 pld [r1]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1433 vshl.s16 q1, q1, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1434 vqmovun.s16 d2, q1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1435 vshl.s16 q10, q10, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1436 vqmovun.s16 d4, q10
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1437 vmov q10, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1438 vst1.8 {d2},[r6,:64], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1439 vmov q1, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1440 vst1.8 {d4},[r6,:64], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1441 bne 1b
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1442 pop {r4-r6, pc}
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1443 .endm
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1444
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1445 .macro biweight_4 macs, macd
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1446 vdup.8 d0, r4
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1447 vdup.8 d1, r5
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1448 vmov q1, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1449 vmov q10, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1450 1: subs ip, ip, #4
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1451 vld1.32 {d4[0]},[r0,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1452 vld1.32 {d4[1]},[r0,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1453 \macd q1, d0, d4
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1454 pld [r0]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1455 vld1.32 {d5[0]},[r1,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1456 vld1.32 {d5[1]},[r1,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1457 \macs q1, d1, d5
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1458 pld [r1]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1459 blt 2f
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1460 vld1.32 {d6[0]},[r0,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1461 vld1.32 {d6[1]},[r0,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1462 \macd q10, d0, d6
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1463 pld [r0]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1464 vld1.32 {d7[0]},[r1,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1465 vld1.32 {d7[1]},[r1,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1466 \macs q10, d1, d7
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1467 pld [r1]
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1468 vshl.s16 q1, q1, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1469 vqmovun.s16 d2, q1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1470 vshl.s16 q10, q10, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1471 vqmovun.s16 d4, q10
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1472 vmov q10, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1473 vst1.32 {d2[0]},[r6,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1474 vst1.32 {d2[1]},[r6,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1475 vmov q1, q8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1476 vst1.32 {d4[0]},[r6,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1477 vst1.32 {d4[1]},[r6,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1478 bne 1b
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1479 pop {r4-r6, pc}
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1480 2: vshl.s16 q1, q1, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1481 vqmovun.s16 d2, q1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1482 vst1.32 {d2[0]},[r6,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1483 vst1.32 {d2[1]},[r6,:32], r2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1484 pop {r4-r6, pc}
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1485 .endm
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1486
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1487 .macro biweight_func w
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1488 function biweight_h264_pixels_\w\()_neon
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1489 push {r4-r6, lr}
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1490 add r4, sp, #16
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1491 ldm r4, {r4-r6}
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1492 lsr lr, r4, #31
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1493 add r6, r6, #1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1494 eors lr, lr, r5, lsr #30
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1495 orr r6, r6, #1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1496 vdup.16 q9, r3
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1497 lsl r6, r6, r3
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1498 vmvn q9, q9
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1499 vdup.16 q8, r6
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1500 mov r6, r0
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1501 beq 10f
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1502 subs lr, lr, #1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1503 beq 20f
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1504 subs lr, lr, #1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1505 beq 30f
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1506 b 40f
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1507 10: biweight_\w vmlal.u8, vmlal.u8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1508 20: rsb r4, r4, #0
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1509 biweight_\w vmlal.u8, vmlsl.u8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1510 30: rsb r4, r4, #0
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1511 rsb r5, r5, #0
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1512 biweight_\w vmlsl.u8, vmlsl.u8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1513 40: rsb r5, r5, #0
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1514 biweight_\w vmlsl.u8, vmlal.u8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1515 .endfunc
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1516 .endm
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1517
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1518 .macro biweight_entry w, h, b=1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1519 function ff_biweight_h264_pixels_\w\()x\h\()_neon, export=1
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1520 mov ip, #\h
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1521 .if \b
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1522 b biweight_h264_pixels_\w\()_neon
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1523 .endif
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1524 .endfunc
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1525 .endm
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1526
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1527 biweight_entry 16, 8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1528 biweight_entry 16, 16, b=0
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1529 biweight_func 16
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1530
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1531 biweight_entry 8, 16
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1532 biweight_entry 8, 4
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1533 biweight_entry 8, 8, b=0
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1534 biweight_func 8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1535
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1536 biweight_entry 4, 8
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1537 biweight_entry 4, 2
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1538 biweight_entry 4, 4, b=0
23f7711e777e ARM: NEON optimised H.264 biweighted prediction
mru
parents: 8626
diff changeset
1539 biweight_func 4
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1540
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1541 @ Weighted prediction
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1542
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1543 .macro weight_16 add
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1544 vdup.8 d0, r3
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1545 1: subs ip, ip, #2
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1546 vld1.8 {d20-d21},[r0,:128], r1
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1547 vmull.u8 q2, d0, d20
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1548 pld [r0]
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1549 vmull.u8 q3, d0, d21
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1550 vld1.8 {d28-d29},[r0,:128], r1
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1551 vmull.u8 q12, d0, d28
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1552 pld [r0]
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1553 vmull.u8 q13, d0, d29
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1554 \add q2, q8, q2
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1555 vrshl.s16 q2, q2, q9
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1556 \add q3, q8, q3
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1557 vrshl.s16 q3, q3, q9
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1558 vqmovun.s16 d4, q2
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1559 vqmovun.s16 d5, q3
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1560 \add q12, q8, q12
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1561 vrshl.s16 q12, q12, q9
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1562 \add q13, q8, q13
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1563 vrshl.s16 q13, q13, q9
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1564 vqmovun.s16 d24, q12
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1565 vqmovun.s16 d25, q13
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1566 vst1.8 {d4- d5}, [r4,:128], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1567 vst1.8 {d24-d25},[r4,:128], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1568 bne 1b
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1569 pop {r4, pc}
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1570 .endm
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1571
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1572 .macro weight_8 add
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1573 vdup.8 d0, r3
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1574 1: subs ip, ip, #2
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1575 vld1.8 {d4},[r0,:64], r1
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1576 vmull.u8 q1, d0, d4
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1577 pld [r0]
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1578 vld1.8 {d6},[r0,:64], r1
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1579 vmull.u8 q10, d0, d6
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1580 \add q1, q8, q1
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1581 pld [r0]
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1582 vrshl.s16 q1, q1, q9
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1583 vqmovun.s16 d2, q1
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1584 \add q10, q8, q10
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1585 vrshl.s16 q10, q10, q9
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1586 vqmovun.s16 d4, q10
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1587 vst1.8 {d2},[r4,:64], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1588 vst1.8 {d4},[r4,:64], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1589 bne 1b
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1590 pop {r4, pc}
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1591 .endm
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1592
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1593 .macro weight_4 add
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1594 vdup.8 d0, r3
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1595 vmov q1, q8
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1596 vmov q10, q8
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1597 1: subs ip, ip, #4
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1598 vld1.32 {d4[0]},[r0,:32], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1599 vld1.32 {d4[1]},[r0,:32], r1
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1600 vmull.u8 q1, d0, d4
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1601 pld [r0]
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1602 blt 2f
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1603 vld1.32 {d6[0]},[r0,:32], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1604 vld1.32 {d6[1]},[r0,:32], r1
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1605 vmull.u8 q10, d0, d6
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1606 pld [r0]
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1607 \add q1, q8, q1
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1608 vrshl.s16 q1, q1, q9
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1609 vqmovun.s16 d2, q1
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1610 \add q10, q8, q10
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1611 vrshl.s16 q10, q10, q9
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1612 vqmovun.s16 d4, q10
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1613 vmov q10, q8
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1614 vst1.32 {d2[0]},[r4,:32], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1615 vst1.32 {d2[1]},[r4,:32], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1616 vmov q1, q8
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1617 vst1.32 {d4[0]},[r4,:32], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1618 vst1.32 {d4[1]},[r4,:32], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1619 bne 1b
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1620 pop {r4, pc}
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1621 2: \add q1, q8, q1
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1622 vrshl.s16 q1, q1, q9
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1623 vqmovun.s16 d2, q1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1624 vst1.32 {d2[0]},[r4,:32], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1625 vst1.32 {d2[1]},[r4,:32], r1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1626 pop {r4, pc}
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1627 .endm
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1628
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1629 .macro weight_func w
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1630 function weight_h264_pixels_\w\()_neon
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1631 push {r4, lr}
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1632 ldr r4, [sp, #8]
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1633 cmp r2, #1
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1634 lsl r4, r4, r2
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1635 vdup.16 q8, r4
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1636 mov r4, r0
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1637 ble 20f
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1638 rsb lr, r2, #1
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1639 vdup.16 q9, lr
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1640 cmp r3, #0
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1641 blt 10f
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1642 weight_\w vhadd.s16
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1643 10: rsb r3, r3, #0
9072
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1644 weight_\w vhsub.s16
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1645 20: rsb lr, r2, #0
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1646 vdup.16 q9, lr
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1647 cmp r3, #0
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1648 blt 10f
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1649 weight_\w vadd.s16
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1650 10: rsb r3, r3, #0
d56b711c6c5d ARM: fix corner-case overflow in H.264 weighted prediction
mru
parents: 8664
diff changeset
1651 weight_\w vsub.s16
8664
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1652 .endfunc
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1653 .endm
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1654
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1655 .macro weight_entry w, h, b=1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1656 function ff_weight_h264_pixels_\w\()x\h\()_neon, export=1
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1657 mov ip, #\h
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1658 .if \b
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1659 b weight_h264_pixels_\w\()_neon
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1660 .endif
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1661 .endfunc
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1662 .endm
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1663
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1664 weight_entry 16, 8
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1665 weight_entry 16, 16, b=0
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1666 weight_func 16
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1667
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1668 weight_entry 8, 16
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1669 weight_entry 8, 4
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1670 weight_entry 8, 8, b=0
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1671 weight_func 8
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1672
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1673 weight_entry 4, 8
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1674 weight_entry 4, 2
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1675 weight_entry 4, 4, b=0
882c351e69c2 ARM: NEON optimised H.264 weighted prediction
mru
parents: 8663
diff changeset
1676 weight_func 4