Gitiles
Code Review Sign In
gerrit-public.fairphone.software / fp2-dev / platform / external / chromium_org / third_party / libyuv / bb6bddc9fb4aea694ef26d7761d9fbcba8f5b6c1 / . / source / row_mips.cc
blob: 3e1a1faea6826d18cd3d2b0cf83072503af10031 [file] [log] [blame]
fbarchard@google.comca410052012-10-14 06:01:19 +0000[diff] [blame]1/*
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000[diff] [blame^]2 * Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
fbarchard@google.comca410052012-10-14 06:01:19 +0000[diff] [blame]3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11#include "libyuv/row.h"
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000[diff] [blame^]12
fbarchard@google.comca410052012-10-14 06:01:19 +0000[diff] [blame]13#ifdef __cplusplus
14namespace libyuv {
15extern "C" {
16#endif
17
18#if !defined(YUV_DISABLE_ASM) && defined(__mips__)
19#ifdef HAS_SPLITUV_MIPS_DSPR2
fbarchard@google.comca410052012-10-14 06:01:19 +0000[diff] [blame]20void SplitUV_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
21 int width) {
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000[diff] [blame^]22
23 __asm__ __volatile__ (
24 ".set push \n"
25 ".set noreorder \n"
26
27 "srl $t4, %[width], 4 \n" // multiplies of 16
28 "blez $t4, 2f \n"
29 " andi %[width], %[width], 0xf \n" // residual
30 "andi $t0, %[src_uv], 0x3 \n"
31 "andi $t1, %[dst_u], 0x3 \n"
32 "andi $t2, %[dst_v], 0x3 \n"
33 "or $t0, $t0, $t1 \n"
34 "or $t0, $t0, $t2 \n"
35
36 "beqz $t0, 12f \n" // test if aligned
37 " nop \n"
38
39 // src and dst are unaligned
40 "1: \n"
41 "addiu $t4, $t4, -1 \n"
42 "lwr $t0, 0(%[src_uv]) \n"
43 "lwl $t0, 3(%[src_uv]) \n" // V1 | U1 | V0 | U0
44 "lwr $t1, 4(%[src_uv]) \n"
45 "lwl $t1, 7(%[src_uv]) \n" // V3 | U3 | V2 | U2
46 "lwr $t2, 8(%[src_uv]) \n"
47 "lwl $t2, 11(%[src_uv]) \n" // V5 | U5 | V4 | U4
48 "lwr $t3, 12(%[src_uv]) \n"
49 "lwl $t3, 15(%[src_uv]) \n" // V7 | U7 | V6 | U6
50 "lwr $t5, 16(%[src_uv]) \n"
51 "lwl $t5, 19(%[src_uv]) \n" // V9 | U9 | V8 | U8
52 "lwr $t6, 20(%[src_uv]) \n"
53 "lwl $t6, 23(%[src_uv]) \n" // V11 | U11 | V10 | U10
54 "lwr $t7, 24(%[src_uv]) \n"
55 "lwl $t7, 27(%[src_uv]) \n" // V13 | U13 | V12 | U12
56 "lwr $t8, 28(%[src_uv]) \n"
57 "lwl $t8, 31(%[src_uv]) \n" // V15 | U15 | V14 | U14
58
59 "precrq.qb.ph $t9, $t1, $t0 \n" // V3 | V2 | V1 | V0
60 "precr.qb.ph $t0, $t1, $t0 \n" // U3 | U2 | U1 | U0
61 "precrq.qb.ph $t1, $t3, $t2 \n" // V7 | V6 | V5 | V4
62 "precr.qb.ph $t2, $t3, $t2 \n" // U7 | U6 | U5 | U4
63 "precrq.qb.ph $t3, $t6, $t5 \n" // V11 | V10 | V9 | V8
64 "precr.qb.ph $t5, $t6, $t5 \n" // U11 | U10 | U9 | U8
65 "precrq.qb.ph $t6, $t8, $t7 \n" // V15 | V14 | V13 | V12
66 "precr.qb.ph $t7, $t8, $t7 \n" // U15 | U14 | U13 | U12
67 "addiu %[src_uv], %[src_uv], 32 \n"
68
69 "swr $t9, 0(%[dst_v]) \n"
70 "swl $t9, 3(%[dst_v]) \n"
71 "swr $t0, 0(%[dst_u]) \n"
72 "swl $t0, 3(%[dst_u]) \n"
73 "swr $t1, 4(%[dst_v]) \n"
74 "swl $t1, 7(%[dst_v]) \n"
75 "swr $t2, 4(%[dst_u]) \n"
76 "swl $t2, 7(%[dst_u]) \n"
77 "swr $t3, 8(%[dst_v]) \n"
78 "swl $t3, 11(%[dst_v]) \n"
79 "swr $t5, 8(%[dst_u]) \n"
80 "swl $t5, 11(%[dst_u]) \n"
81 "swr $t6, 12(%[dst_v]) \n"
82 "swl $t6, 15(%[dst_v]) \n"
83 "swr $t7, 12(%[dst_u]) \n"
84 "swl $t7, 15(%[dst_u]) \n"
85 "addiu %[dst_u], %[dst_u], 16 \n"
86 "bgtz $t4, 1b \n"
87 " addiu %[dst_v], %[dst_v], 16 \n"
88
89 "beqz %[width], 3f \n"
90 " nop \n"
91 "b 2f \n"
92 " nop \n"
93
94 // src and dst are aligned
95 "12: \n"
96 "addiu $t4, $t4, -1 \n"
97 "lw $t0, 0(%[src_uv]) \n" // V1 | U1 | V0 | U0
98 "lw $t1, 4(%[src_uv]) \n" // V3 | U3 | V2 | U2
99 "lw $t2, 8(%[src_uv]) \n" // V5 | U5 | V4 | U4
100 "lw $t3, 12(%[src_uv]) \n" // V7 | U7 | V6 | U6
101 "lw $t5, 16(%[src_uv]) \n" // V9 | U9 | V8 | U8
102 "lw $t6, 20(%[src_uv]) \n" // V11 | U11 | V10 | U10
103 "lw $t7, 24(%[src_uv]) \n" // V13 | U13 | V12 | U12
104 "lw $t8, 28(%[src_uv]) \n" // V15 | U15 | V14 | U14
105
106 "addiu %[src_uv], %[src_uv], 32 \n"
107 "precrq.qb.ph $t9, $t1, $t0 \n" // V3 | V2 | V1 | V0
108 "precr.qb.ph $t0, $t1, $t0 \n" // U3 | U2 | U1 | U0
109 "precrq.qb.ph $t1, $t3, $t2 \n" // V7 | V6 | V5 | V4
110 "precr.qb.ph $t2, $t3, $t2 \n" // U7 | U6 | U5 | U4
111 "precrq.qb.ph $t3, $t6, $t5 \n" // V11 | V10 | V9 | V8
112 "precr.qb.ph $t5, $t6, $t5 \n" // U11 | U10 | U9 | U8
113 "precrq.qb.ph $t6, $t8, $t7 \n" // V15 | V14 | V13 | V12
114 "precr.qb.ph $t7, $t8, $t7 \n" // U15 | U14 | U13 | U12
115
116 "sw $t9, 0(%[dst_v]) \n"
117 "sw $t0, 0(%[dst_u]) \n"
118 "sw $t1, 4(%[dst_v]) \n"
119 "sw $t2, 4(%[dst_u]) \n"
120 "sw $t3, 8(%[dst_v]) \n"
121 "sw $t5, 8(%[dst_u]) \n"
122 "sw $t6, 12(%[dst_v]) \n"
123 "sw $t7, 12(%[dst_u]) \n"
124 "addiu %[dst_v], %[dst_v], 16 \n"
125 "bgtz $t4, 12b \n"
126 " addiu %[dst_u], %[dst_u], 16 \n"
127
128 "beqz %[width], 3f \n"
129 " nop \n"
130
131 "2: \n"
132 "lbu $t0, 0(%[src_uv]) \n"
133 "lbu $t1, 1(%[src_uv]) \n"
134 "addiu %[src_uv], %[src_uv], 2 \n"
135 "addiu %[width], %[width], -1 \n"
136 "sb $t0, 0(%[dst_u]) \n"
137 "sb $t1, 0(%[dst_v]) \n"
138 "addiu %[dst_u], %[dst_u], 1 \n"
139 "bgtz %[width], 2b \n"
140 " addiu %[dst_v], %[dst_v], 1 \n"
141
142 "3: \n"
143 ".set pop \n"
144 : [src_uv] "+r" (src_uv),
145 [width] "+r" (width),
146 [dst_u] "+r" (dst_u),
147 [dst_v] "+r" (dst_v)
148 :
149 : "t0", "t1", "t2", "t3",
150 "t4", "t5", "t6", "t7", "t8", "t9"
fbarchard@google.comca410052012-10-14 06:01:19 +0000[diff] [blame]151 );
152}
153#endif // HAS_SPLITUV_MIPS_DSPR2
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000[diff] [blame^]154
fbarchard@google.comca410052012-10-14 06:01:19 +0000[diff] [blame]155#endif // __mips__
156
157#ifdef __cplusplus
158} // extern "C"
159} // namespace libyuv
160#endif