annotate arm/dsputil_neon_s.S @ 9345:e0a7a6338526 libavcodec

ARM: NEON optimized put_signed_pixels_clamped
author conrad
date Sat, 04 Apr 2009 21:02:48 +0000
parents 9ea1ea6db616
children 93c20dd3da43
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
8334
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
1 /*
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
2 * ARM NEON optimised DSP functions
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
3 * Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
4 *
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
5 * This file is part of FFmpeg.
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
6 *
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
9 * License as published by the Free Software Foundation; either
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
11 *
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
15 * Lesser General Public License for more details.
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
16 *
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
17 * You should have received a copy of the GNU Lesser General Public
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
18 * License along with FFmpeg; if not, write to the Free Software
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
20 */
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
21
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
22 #include "asm.S"
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
23
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
24 preserve8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
25 .fpu neon
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
26 .text
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
27
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
28 .macro pixels16 avg=0
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
29 .if \avg
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
30 mov ip, r0
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
31 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
32 1: vld1.64 {d0, d1}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
33 vld1.64 {d2, d3}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
34 vld1.64 {d4, d5}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
35 pld [r1, r2, lsl #2]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
36 vld1.64 {d6, d7}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
37 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
38 pld [r1, r2]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
39 pld [r1, r2, lsl #1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
40 .if \avg
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
41 vld1.64 {d16,d17}, [ip], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
42 vrhadd.u8 q0, q0, q8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
43 vld1.64 {d18,d19}, [ip], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
44 vrhadd.u8 q1, q1, q9
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
45 vld1.64 {d20,d21}, [ip], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
46 vrhadd.u8 q2, q2, q10
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
47 vld1.64 {d22,d23}, [ip], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
48 vrhadd.u8 q3, q3, q11
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
49 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
50 subs r3, r3, #4
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
51 vst1.64 {d0, d1}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
52 vst1.64 {d2, d3}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
53 vst1.64 {d4, d5}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
54 vst1.64 {d6, d7}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
55 bne 1b
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
56 bx lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
57 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
58
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
59 .macro pixels16_x2 vhadd=vrhadd.u8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
60 1: vld1.64 {d0-d2}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
61 vld1.64 {d4-d6}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
62 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
63 pld [r1, r2]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
64 subs r3, r3, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
65 vext.8 q1, q0, q1, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
66 \vhadd q0, q0, q1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
67 vext.8 q3, q2, q3, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
68 \vhadd q2, q2, q3
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
69 vst1.64 {d0, d1}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
70 vst1.64 {d4, d5}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
71 bne 1b
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
72 bx lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
73 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
74
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
75 .macro pixels16_y2 vhadd=vrhadd.u8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
76 push {lr}
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
77 add ip, r1, r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
78 lsl lr, r2, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
79 vld1.64 {d0, d1}, [r1], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
80 vld1.64 {d2, d3}, [ip], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
81 1: subs r3, r3, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
82 \vhadd q2, q0, q1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
83 vld1.64 {d0, d1}, [r1], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
84 \vhadd q3, q0, q1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
85 vld1.64 {d2, d3}, [ip], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
86 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
87 pld [ip]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
88 vst1.64 {d4, d5}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
89 vst1.64 {d6, d7}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
90 bne 1b
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
91 pop {pc}
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
92 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
93
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
94 .macro pixels16_xy2 vshrn=vrshrn.u16 no_rnd=0
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
95 push {lr}
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
96 lsl lr, r2, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
97 add ip, r1, r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
98 vld1.64 {d0-d2}, [r1], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
99 vld1.64 {d4-d6}, [ip], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
100 .if \no_rnd
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
101 vmov.i16 q13, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
102 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
103 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
104 pld [ip]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
105 vext.8 q1, q0, q1, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
106 vext.8 q3, q2, q3, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
107 vaddl.u8 q8, d0, d2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
108 vaddl.u8 q10, d1, d3
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
109 vaddl.u8 q9, d4, d6
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
110 vaddl.u8 q11, d5, d7
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
111 1: subs r3, r3, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
112 vld1.64 {d0-d2}, [r1], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
113 vadd.u16 q12, q8, q9
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
114 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
115 .if \no_rnd
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
116 vadd.u16 q12, q12, q13
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
117 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
118 vext.8 q15, q0, q1, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
119 vadd.u16 q1 , q10, q11
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
120 \vshrn d28, q12, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
121 .if \no_rnd
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
122 vadd.u16 q1, q1, q13
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
123 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
124 \vshrn d29, q1, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
125 vaddl.u8 q8, d0, d30
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
126 vld1.64 {d2-d4}, [ip], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
127 vaddl.u8 q10, d1, d31
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
128 vst1.64 {d28,d29}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
129 vadd.u16 q12, q8, q9
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
130 pld [ip]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
131 .if \no_rnd
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
132 vadd.u16 q12, q12, q13
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
133 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
134 vext.8 q2, q1, q2, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
135 vadd.u16 q0, q10, q11
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
136 \vshrn d30, q12, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
137 .if \no_rnd
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
138 vadd.u16 q0, q0, q13
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
139 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
140 \vshrn d31, q0, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
141 vaddl.u8 q9, d2, d4
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
142 vaddl.u8 q11, d3, d5
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
143 vst1.64 {d30,d31}, [r0,:128], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
144 bgt 1b
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
145 pop {pc}
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
146 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
147
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
148 .macro pixels8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
149 1: vld1.64 {d0}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
150 vld1.64 {d1}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
151 vld1.64 {d2}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
152 pld [r1, r2, lsl #2]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
153 vld1.64 {d3}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
154 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
155 pld [r1, r2]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
156 pld [r1, r2, lsl #1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
157 subs r3, r3, #4
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
158 vst1.64 {d0}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
159 vst1.64 {d1}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
160 vst1.64 {d2}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
161 vst1.64 {d3}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
162 bne 1b
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
163 bx lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
164 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
165
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
166 .macro pixels8_x2 vhadd=vrhadd.u8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
167 1: vld1.64 {d0, d1}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
168 vext.8 d1, d0, d1, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
169 vld1.64 {d2, d3}, [r1], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
170 vext.8 d3, d2, d3, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
171 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
172 pld [r1, r2]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
173 subs r3, r3, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
174 vswp d1, d2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
175 \vhadd q0, q0, q1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
176 vst1.64 {d0}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
177 vst1.64 {d1}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
178 bne 1b
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
179 bx lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
180 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
181
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
182 .macro pixels8_y2 vhadd=vrhadd.u8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
183 push {lr}
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
184 add ip, r1, r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
185 lsl lr, r2, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
186 vld1.64 {d0}, [r1], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
187 vld1.64 {d1}, [ip], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
188 1: subs r3, r3, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
189 \vhadd d4, d0, d1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
190 vld1.64 {d0}, [r1], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
191 \vhadd d5, d0, d1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
192 vld1.64 {d1}, [ip], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
193 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
194 pld [ip]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
195 vst1.64 {d4}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
196 vst1.64 {d5}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
197 bne 1b
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
198 pop {pc}
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
199 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
200
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
201 .macro pixels8_xy2 vshrn=vrshrn.u16 no_rnd=0
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
202 push {lr}
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
203 lsl lr, r2, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
204 add ip, r1, r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
205 vld1.64 {d0, d1}, [r1], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
206 vld1.64 {d2, d3}, [ip], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
207 .if \no_rnd
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
208 vmov.i16 q11, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
209 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
210 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
211 pld [ip]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
212 vext.8 d4, d0, d1, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
213 vext.8 d6, d2, d3, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
214 vaddl.u8 q8, d0, d4
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
215 vaddl.u8 q9, d2, d6
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
216 1: subs r3, r3, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
217 vld1.64 {d0, d1}, [r1], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
218 pld [r1]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
219 vadd.u16 q10, q8, q9
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
220 vext.8 d4, d0, d1, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
221 .if \no_rnd
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
222 vadd.u16 q10, q10, q11
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
223 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
224 vaddl.u8 q8, d0, d4
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
225 \vshrn d5, q10, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
226 vld1.64 {d2, d3}, [ip], lr
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
227 vadd.u16 q10, q8, q9
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
228 pld [ip]
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
229 .if \no_rnd
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
230 vadd.u16 q10, q10, q11
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
231 .endif
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
232 vst1.64 {d5}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
233 \vshrn d7, q10, #2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
234 vext.8 d6, d2, d3, #1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
235 vaddl.u8 q9, d2, d6
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
236 vst1.64 {d7}, [r0,:64], r2
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
237 bgt 1b
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
238 pop {pc}
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
239 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
240
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
241 .macro pixfunc pfx name suf rnd_op args:vararg
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
242 function ff_\pfx\name\suf\()_neon, export=1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
243 \name \rnd_op \args
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
244 .endfunc
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
245 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
246
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
247 .macro pixfunc2 pfx name args:vararg
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
248 pixfunc \pfx \name
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
249 pixfunc \pfx \name \args
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
250 .endm
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
251
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
252 function ff_put_h264_qpel16_mc00_neon, export=1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
253 mov r3, #16
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
254 .endfunc
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
255
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
256 pixfunc put_ pixels16
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
257 pixfunc2 put_ pixels16_x2, _no_rnd, vhadd.u8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
258 pixfunc2 put_ pixels16_y2, _no_rnd, vhadd.u8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
259 pixfunc2 put_ pixels16_xy2, _no_rnd, vshrn.u16, 1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
260
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
261 function ff_avg_h264_qpel16_mc00_neon, export=1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
262 mov r3, #16
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
263 .endfunc
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
264
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
265 pixfunc avg_ pixels16,, 1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
266
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
267 function ff_put_h264_qpel8_mc00_neon, export=1
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
268 mov r3, #8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
269 .endfunc
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
270
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
271 pixfunc put_ pixels8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
272 pixfunc2 put_ pixels8_x2, _no_rnd, vhadd.u8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
273 pixfunc2 put_ pixels8_y2, _no_rnd, vhadd.u8
6bdd6dfc3574 ARM: NEON optimised put_pixels functions
mru
parents:
diff changeset
274 pixfunc2 put_ pixels8_xy2, _no_rnd, vshrn.u16, 1
8492
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
275
9345
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
276 function ff_put_signed_pixels_clamped_neon, export=1
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
277 vmov.u8 d31, #128
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
278 vld1.64 {d16-d17}, [r0,:128]!
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
279 vqmovn.s16 d0, q8
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
280 vld1.64 {d18-d19}, [r0,:128]!
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
281 vqmovn.s16 d1, q9
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
282 vld1.64 {d16-d17}, [r0,:128]!
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
283 vqmovn.s16 d2, q8
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
284 vld1.64 {d18-d19}, [r0,:128]!
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
285 vadd.u8 d0, d0, d31
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
286 vld1.64 {d20-d21}, [r0,:128]!
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
287 vadd.u8 d1, d1, d31
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
288 vld1.64 {d22-d23}, [r0,:128]!
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
289 vadd.u8 d2, d2, d31
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
290 vst1.64 {d0}, [r1,:64], r2
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
291 vqmovn.s16 d3, q9
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
292 vst1.64 {d1}, [r1,:64], r2
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
293 vqmovn.s16 d4, q10
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
294 vst1.64 {d2}, [r1,:64], r2
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
295 vqmovn.s16 d5, q11
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
296 vld1.64 {d24-d25}, [r0,:128]!
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
297 vadd.u8 d3, d3, d31
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
298 vld1.64 {d26-d27}, [r0,:128]!
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
299 vadd.u8 d4, d4, d31
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
300 vadd.u8 d5, d5, d31
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
301 vst1.64 {d3}, [r1,:64], r2
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
302 vqmovn.s16 d6, q12
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
303 vst1.64 {d4}, [r1,:64], r2
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
304 vqmovn.s16 d7, q13
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
305 vst1.64 {d5}, [r1,:64], r2
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
306 vadd.u8 d6, d6, d31
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
307 vadd.u8 d7, d7, d31
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
308 vst1.64 {d6}, [r1,:64], r2
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
309 vst1.64 {d7}, [r1,:64], r2
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
310 bx lr
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
311 .endfunc
e0a7a6338526 ARM: NEON optimized put_signed_pixels_clamped
conrad
parents: 9344
diff changeset
312
9344
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
313 function ff_add_pixels_clamped_neon, export=1
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
314 mov r3, r1
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
315 vld1.64 {d16}, [r1,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
316 vld1.64 {d0-d1}, [r0,:128]!
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
317 vaddw.u8 q0, q0, d16
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
318 vld1.64 {d17}, [r1,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
319 vld1.64 {d2-d3}, [r0,:128]!
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
320 vqmovun.s16 d0, q0
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
321 vld1.64 {d18}, [r1,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
322 vaddw.u8 q1, q1, d17
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
323 vld1.64 {d4-d5}, [r0,:128]!
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
324 vaddw.u8 q2, q2, d18
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
325 vst1.64 {d0}, [r3,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
326 vqmovun.s16 d2, q1
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
327 vld1.64 {d19}, [r1,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
328 vld1.64 {d6-d7}, [r0,:128]!
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
329 vaddw.u8 q3, q3, d19
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
330 vqmovun.s16 d4, q2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
331 vst1.64 {d2}, [r3,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
332 vld1.64 {d16}, [r1,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
333 vqmovun.s16 d6, q3
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
334 vld1.64 {d0-d1}, [r0,:128]!
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
335 vaddw.u8 q0, q0, d16
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
336 vst1.64 {d4}, [r3,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
337 vld1.64 {d17}, [r1,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
338 vld1.64 {d2-d3}, [r0,:128]!
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
339 vaddw.u8 q1, q1, d17
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
340 vst1.64 {d6}, [r3,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
341 vqmovun.s16 d0, q0
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
342 vld1.64 {d18}, [r1,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
343 vld1.64 {d4-d5}, [r0,:128]!
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
344 vaddw.u8 q2, q2, d18
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
345 vst1.64 {d0}, [r3,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
346 vqmovun.s16 d2, q1
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
347 vld1.64 {d19}, [r1,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
348 vqmovun.s16 d4, q2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
349 vld1.64 {d6-d7}, [r0,:128]!
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
350 vaddw.u8 q3, q3, d19
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
351 vst1.64 {d2}, [r3,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
352 vqmovun.s16 d6, q3
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
353 vst1.64 {d4}, [r3,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
354 vst1.64 {d6}, [r3,:64], r2
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
355 bx lr
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
356 .endfunc
9ea1ea6db616 ARM: NEON optimised add_pixels_clamped
mru
parents: 8698
diff changeset
357
8492
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
358 function ff_float_to_int16_neon, export=1
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
359 subs r2, r2, #8
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
360 vld1.64 {d0-d1}, [r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
361 vcvt.s32.f32 q8, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
362 vld1.64 {d2-d3}, [r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
363 vcvt.s32.f32 q9, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
364 beq 3f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
365 bics ip, r2, #15
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
366 beq 2f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
367 1: subs ip, ip, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
368 vshrn.s32 d4, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
369 vld1.64 {d0-d1}, [r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
370 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
371 vshrn.s32 d5, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
372 vld1.64 {d2-d3}, [r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
373 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
374 vshrn.s32 d6, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
375 vst1.64 {d4-d5}, [r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
376 vshrn.s32 d7, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
377 vld1.64 {d16-d17},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
378 vcvt.s32.f32 q8, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
379 vld1.64 {d18-d19},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
380 vcvt.s32.f32 q9, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
381 vst1.64 {d6-d7}, [r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
382 bne 1b
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
383 ands r2, r2, #15
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
384 beq 3f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
385 2: vld1.64 {d0-d1}, [r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
386 vshrn.s32 d4, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
387 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
388 vld1.64 {d2-d3}, [r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
389 vshrn.s32 d5, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
390 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
391 vshrn.s32 d6, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
392 vst1.64 {d4-d5}, [r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
393 vshrn.s32 d7, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
394 vst1.64 {d6-d7}, [r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
395 bx lr
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
396 3: vshrn.s32 d4, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
397 vshrn.s32 d5, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
398 vst1.64 {d4-d5}, [r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
399 bx lr
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
400 .endfunc
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
401
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
402 function ff_float_to_int16_interleave_neon, export=1
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
403 cmp r3, #2
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
404 ldrlt r1, [r1]
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
405 blt ff_float_to_int16_neon
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
406 bne 4f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
407
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
408 ldr r3, [r1]
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
409 ldr r1, [r1, #4]
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
410
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
411 subs r2, r2, #8
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
412 vld1.64 {d0-d1}, [r3,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
413 vcvt.s32.f32 q8, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
414 vld1.64 {d2-d3}, [r3,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
415 vcvt.s32.f32 q9, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
416 vld1.64 {d20-d21},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
417 vcvt.s32.f32 q10, q10, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
418 vld1.64 {d22-d23},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
419 vcvt.s32.f32 q11, q11, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
420 beq 3f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
421 bics ip, r2, #15
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
422 beq 2f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
423 1: subs ip, ip, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
424 vld1.64 {d0-d1}, [r3,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
425 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
426 vsri.32 q10, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
427 vld1.64 {d2-d3}, [r3,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
428 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
429 vld1.64 {d24-d25},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
430 vcvt.s32.f32 q12, q12, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
431 vld1.64 {d26-d27},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
432 vsri.32 q11, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
433 vst1.64 {d20-d21},[r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
434 vcvt.s32.f32 q13, q13, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
435 vst1.64 {d22-d23},[r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
436 vsri.32 q12, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
437 vld1.64 {d16-d17},[r3,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
438 vsri.32 q13, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
439 vst1.64 {d24-d25},[r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
440 vcvt.s32.f32 q8, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
441 vld1.64 {d18-d19},[r3,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
442 vcvt.s32.f32 q9, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
443 vld1.64 {d20-d21},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
444 vcvt.s32.f32 q10, q10, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
445 vld1.64 {d22-d23},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
446 vcvt.s32.f32 q11, q11, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
447 vst1.64 {d26-d27},[r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
448 bne 1b
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
449 ands r2, r2, #15
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
450 beq 3f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
451 2: vsri.32 q10, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
452 vld1.64 {d0-d1}, [r3,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
453 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
454 vld1.64 {d2-d3}, [r3,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
455 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
456 vld1.64 {d24-d25},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
457 vcvt.s32.f32 q12, q12, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
458 vsri.32 q11, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
459 vld1.64 {d26-d27},[r1,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
460 vcvt.s32.f32 q13, q13, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
461 vst1.64 {d20-d21},[r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
462 vsri.32 q12, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
463 vst1.64 {d22-d23},[r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
464 vsri.32 q13, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
465 vst1.64 {d24-d27},[r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
466 bx lr
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
467 3: vsri.32 q10, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
468 vsri.32 q11, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
469 vst1.64 {d20-d23},[r0,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
470 bx lr
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
471
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
472 4: push {r4-r8,lr}
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
473 cmp r3, #4
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
474 lsl ip, r3, #1
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
475 blt 4f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
476
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
477 @ 4 channels
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
478 5: ldmia r1!, {r4-r7}
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
479 mov lr, r2
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
480 mov r8, r0
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
481 vld1.64 {d16-d17},[r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
482 vcvt.s32.f32 q8, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
483 vld1.64 {d18-d19},[r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
484 vcvt.s32.f32 q9, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
485 vld1.64 {d20-d21},[r6,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
486 vcvt.s32.f32 q10, q10, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
487 vld1.64 {d22-d23},[r7,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
488 vcvt.s32.f32 q11, q11, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
489 6: subs lr, lr, #8
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
490 vld1.64 {d0-d1}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
491 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
492 vsri.32 q9, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
493 vld1.64 {d2-d3}, [r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
494 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
495 vsri.32 q11, q10, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
496 vld1.64 {d4-d5}, [r6,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
497 vcvt.s32.f32 q2, q2, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
498 vzip.32 d18, d22
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
499 vld1.64 {d6-d7}, [r7,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
500 vcvt.s32.f32 q3, q3, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
501 vzip.32 d19, d23
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
502 vst1.64 {d18}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
503 vsri.32 q1, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
504 vst1.64 {d22}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
505 vsri.32 q3, q2, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
506 vst1.64 {d19}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
507 vzip.32 d2, d6
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
508 vst1.64 {d23}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
509 vzip.32 d3, d7
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
510 beq 7f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
511 vld1.64 {d16-d17},[r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
512 vcvt.s32.f32 q8, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
513 vst1.64 {d2}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
514 vld1.64 {d18-d19},[r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
515 vcvt.s32.f32 q9, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
516 vst1.64 {d6}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
517 vld1.64 {d20-d21},[r6,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
518 vcvt.s32.f32 q10, q10, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
519 vst1.64 {d3}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
520 vld1.64 {d22-d23},[r7,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
521 vcvt.s32.f32 q11, q11, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
522 vst1.64 {d7}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
523 b 6b
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
524 7: vst1.64 {d2}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
525 vst1.64 {d6}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
526 vst1.64 {d3}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
527 vst1.64 {d7}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
528 subs r3, r3, #4
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
529 popeq {r4-r8,pc}
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
530 cmp r3, #4
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
531 add r0, r0, #8
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
532 bge 5b
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
533
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
534 @ 2 channels
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
535 4: cmp r3, #2
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
536 blt 4f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
537 ldmia r1!, {r4-r5}
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
538 mov lr, r2
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
539 mov r8, r0
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
540 tst lr, #8
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
541 vld1.64 {d16-d17},[r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
542 vcvt.s32.f32 q8, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
543 vld1.64 {d18-d19},[r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
544 vcvt.s32.f32 q9, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
545 vld1.64 {d20-d21},[r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
546 vcvt.s32.f32 q10, q10, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
547 vld1.64 {d22-d23},[r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
548 vcvt.s32.f32 q11, q11, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
549 beq 6f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
550 subs lr, lr, #8
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
551 beq 7f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
552 vsri.32 d18, d16, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
553 vsri.32 d19, d17, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
554 vld1.64 {d16-d17},[r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
555 vcvt.s32.f32 q8, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
556 vst1.32 {d18[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
557 vsri.32 d22, d20, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
558 vst1.32 {d18[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
559 vsri.32 d23, d21, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
560 vst1.32 {d19[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
561 vst1.32 {d19[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
562 vld1.64 {d18-d19},[r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
563 vcvt.s32.f32 q9, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
564 vst1.32 {d22[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
565 vst1.32 {d22[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
566 vld1.64 {d20-d21},[r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
567 vcvt.s32.f32 q10, q10, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
568 vst1.32 {d23[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
569 vst1.32 {d23[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
570 vld1.64 {d22-d23},[r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
571 vcvt.s32.f32 q11, q11, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
572 6: subs lr, lr, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
573 vld1.64 {d0-d1}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
574 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
575 vsri.32 d18, d16, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
576 vld1.64 {d2-d3}, [r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
577 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
578 vsri.32 d19, d17, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
579 vld1.64 {d4-d5}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
580 vcvt.s32.f32 q2, q2, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
581 vld1.64 {d6-d7}, [r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
582 vcvt.s32.f32 q3, q3, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
583 vst1.32 {d18[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
584 vsri.32 d22, d20, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
585 vst1.32 {d18[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
586 vsri.32 d23, d21, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
587 vst1.32 {d19[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
588 vsri.32 d2, d0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
589 vst1.32 {d19[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
590 vsri.32 d3, d1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
591 vst1.32 {d22[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
592 vsri.32 d6, d4, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
593 vst1.32 {d22[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
594 vsri.32 d7, d5, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
595 vst1.32 {d23[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
596 vst1.32 {d23[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
597 beq 6f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
598 vld1.64 {d16-d17},[r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
599 vcvt.s32.f32 q8, q8, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
600 vst1.32 {d2[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
601 vst1.32 {d2[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
602 vld1.64 {d18-d19},[r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
603 vcvt.s32.f32 q9, q9, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
604 vst1.32 {d3[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
605 vst1.32 {d3[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
606 vld1.64 {d20-d21},[r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
607 vcvt.s32.f32 q10, q10, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
608 vst1.32 {d6[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
609 vst1.32 {d6[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
610 vld1.64 {d22-d23},[r5,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
611 vcvt.s32.f32 q11, q11, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
612 vst1.32 {d7[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
613 vst1.32 {d7[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
614 bgt 6b
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
615 6: vst1.32 {d2[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
616 vst1.32 {d2[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
617 vst1.32 {d3[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
618 vst1.32 {d3[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
619 vst1.32 {d6[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
620 vst1.32 {d6[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
621 vst1.32 {d7[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
622 vst1.32 {d7[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
623 b 8f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
624 7: vsri.32 d18, d16, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
625 vsri.32 d19, d17, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
626 vst1.32 {d18[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
627 vsri.32 d22, d20, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
628 vst1.32 {d18[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
629 vsri.32 d23, d21, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
630 vst1.32 {d19[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
631 vst1.32 {d19[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
632 vst1.32 {d22[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
633 vst1.32 {d22[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
634 vst1.32 {d23[0]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
635 vst1.32 {d23[1]}, [r8], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
636 8: subs r3, r3, #2
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
637 add r0, r0, #4
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
638 popeq {r4-r8,pc}
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
639
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
640 @ 1 channel
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
641 4: ldr r4, [r1],#4
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
642 tst r2, #8
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
643 mov lr, r2
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
644 mov r5, r0
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
645 vld1.64 {d0-d1}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
646 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
647 vld1.64 {d2-d3}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
648 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
649 bne 8f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
650 6: subs lr, lr, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
651 vld1.64 {d4-d5}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
652 vcvt.s32.f32 q2, q2, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
653 vld1.64 {d6-d7}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
654 vcvt.s32.f32 q3, q3, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
655 vst1.16 {d0[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
656 vst1.16 {d0[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
657 vst1.16 {d1[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
658 vst1.16 {d1[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
659 vst1.16 {d2[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
660 vst1.16 {d2[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
661 vst1.16 {d3[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
662 vst1.16 {d3[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
663 beq 7f
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
664 vld1.64 {d0-d1}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
665 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
666 vld1.64 {d2-d3}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
667 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
668 7: vst1.16 {d4[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
669 vst1.16 {d4[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
670 vst1.16 {d5[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
671 vst1.16 {d5[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
672 vst1.16 {d6[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
673 vst1.16 {d6[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
674 vst1.16 {d7[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
675 vst1.16 {d7[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
676 bgt 6b
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
677 pop {r4-r8,pc}
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
678 8: subs lr, lr, #8
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
679 vst1.16 {d0[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
680 vst1.16 {d0[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
681 vst1.16 {d1[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
682 vst1.16 {d1[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
683 vst1.16 {d2[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
684 vst1.16 {d2[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
685 vst1.16 {d3[1]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
686 vst1.16 {d3[3]}, [r5,:16], ip
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
687 popeq {r4-r8,pc}
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
688 vld1.64 {d0-d1}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
689 vcvt.s32.f32 q0, q0, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
690 vld1.64 {d2-d3}, [r4,:128]!
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
691 vcvt.s32.f32 q1, q1, #16
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
692 b 6b
639169d7fad5 ARM: NEON optimised float_to_int16
mru
parents: 8359
diff changeset
693 .endfunc
8697
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
694
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
695 function ff_vector_fmul_neon, export=1
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
696 mov r3, r0
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
697 subs r2, r2, #8
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
698 vld1.64 {d0-d3}, [r0,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
699 vld1.64 {d4-d7}, [r1,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
700 vmul.f32 q8, q0, q2
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
701 vmul.f32 q9, q1, q3
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
702 beq 3f
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
703 bics ip, r2, #15
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
704 beq 2f
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
705 1: subs ip, ip, #16
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
706 vld1.64 {d0-d1}, [r0,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
707 vld1.64 {d4-d5}, [r1,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
708 vmul.f32 q10, q0, q2
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
709 vld1.64 {d2-d3}, [r0,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
710 vld1.64 {d6-d7}, [r1,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
711 vmul.f32 q11, q1, q3
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
712 vst1.64 {d16-d19},[r3,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
713 vld1.64 {d0-d1}, [r0,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
714 vld1.64 {d4-d5}, [r1,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
715 vmul.f32 q8, q0, q2
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
716 vld1.64 {d2-d3}, [r0,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
717 vld1.64 {d6-d7}, [r1,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
718 vmul.f32 q9, q1, q3
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
719 vst1.64 {d20-d23},[r3,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
720 bne 1b
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
721 ands r2, r2, #15
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
722 beq 3f
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
723 2: vld1.64 {d0-d1}, [r0,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
724 vld1.64 {d4-d5}, [r1,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
725 vst1.64 {d16-d17},[r3,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
726 vmul.f32 q8, q0, q2
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
727 vld1.64 {d2-d3}, [r0,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
728 vld1.64 {d6-d7}, [r1,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
729 vst1.64 {d18-d19},[r3,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
730 vmul.f32 q9, q1, q3
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
731 3: vst1.64 {d16-d19},[r3,:128]!
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
732 bx lr
307b176f91e7 ARM: NEON optimised vector_fmul
mru
parents: 8492
diff changeset
733 .endfunc
8698
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
734
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
735 function ff_vector_fmul_window_neon, export=1
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
736 vld1.32 {d16[],d17[]}, [sp,:32]
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
737 push {r4,r5,lr}
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
738 ldr lr, [sp, #16]
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
739 sub r2, r2, #8
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
740 sub r5, lr, #2
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
741 add r2, r2, r5, lsl #2
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
742 add r4, r3, r5, lsl #3
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
743 add ip, r0, r5, lsl #3
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
744 mov r5, #-16
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
745 vld1.64 {d0,d1}, [r1,:128]!
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
746 vld1.64 {d2,d3}, [r2,:128], r5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
747 vld1.64 {d4,d5}, [r3,:128]!
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
748 vld1.64 {d6,d7}, [r4,:128], r5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
749 1: subs lr, lr, #4
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
750 vmov q11, q8
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
751 vmla.f32 d22, d0, d4
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
752 vmov q10, q8
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
753 vmla.f32 d23, d1, d5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
754 vrev64.32 q3, q3
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
755 vmla.f32 d20, d0, d7
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
756 vrev64.32 q1, q1
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
757 vmla.f32 d21, d1, d6
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
758 beq 2f
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
759 vmla.f32 d22, d3, d7
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
760 vld1.64 {d0,d1}, [r1,:128]!
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
761 vmla.f32 d23, d2, d6
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
762 vld1.64 {d18,d19},[r2,:128], r5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
763 vmls.f32 d20, d3, d4
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
764 vld1.64 {d24,d25},[r3,:128]!
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
765 vmls.f32 d21, d2, d5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
766 vld1.64 {d6,d7}, [r4,:128], r5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
767 vmov q1, q9
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
768 vrev64.32 q11, q11
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
769 vmov q2, q12
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
770 vswp d22, d23
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
771 vst1.64 {d20,d21},[r0,:128]!
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
772 vst1.64 {d22,d23},[ip,:128], r5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
773 b 1b
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
774 2: vmla.f32 d22, d3, d7
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
775 vmla.f32 d23, d2, d6
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
776 vmls.f32 d20, d3, d4
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
777 vmls.f32 d21, d2, d5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
778 vrev64.32 q11, q11
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
779 vswp d22, d23
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
780 vst1.64 {d20,d21},[r0,:128]!
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
781 vst1.64 {d22,d23},[ip,:128], r5
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
782 pop {r4,r5,pc}
24a7b5d0eb27 ARM: NEON optimised vector_fmul_window
mru
parents: 8697
diff changeset
783 .endfunc