blob: 7e4125d21cbac13e30a4998a084e60bfdc2acd86 [file] [log] [blame]
fbarchard@google.comca410052012-10-14 06:01:19 +00001/*
fbarchard@google.combb6bddc2012-10-14 06:41:17 +00002 * Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
fbarchard@google.comca410052012-10-14 06:01:19 +00003 *
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 +000012
fbarchard@google.comca410052012-10-14 06:01:19 +000013#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 +000020void SplitUV_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
21 int width) {
fbarchard@google.combb6bddc2012-10-14 06:41:17 +000022
23 __asm__ __volatile__ (
24 ".set push \n"
25 ".set noreorder \n"
fbarchard@google.combb6bddc2012-10-14 06:41:17 +000026 "srl $t4, %[width], 4 \n" // multiplies of 16
27 "blez $t4, 2f \n"
28 " andi %[width], %[width], 0xf \n" // residual
fbarchard@google.combb6bddc2012-10-14 06:41:17 +000029
fbarchard@google.comdb694ed2012-10-17 21:54:04 +000030 "1: \n"
31 "addiu $t4, $t4, -1 \n"
32 "lw $t0, 0(%[src_uv]) \n" // V1 | U1 | V0 | U0
33 "lw $t1, 4(%[src_uv]) \n" // V3 | U3 | V2 | U2
34 "lw $t2, 8(%[src_uv]) \n" // V5 | U5 | V4 | U4
35 "lw $t3, 12(%[src_uv]) \n" // V7 | U7 | V6 | U6
36 "lw $t5, 16(%[src_uv]) \n" // V9 | U9 | V8 | U8
37 "lw $t6, 20(%[src_uv]) \n" // V11 | U11 | V10 | U10
38 "lw $t7, 24(%[src_uv]) \n" // V13 | U13 | V12 | U12
39 "lw $t8, 28(%[src_uv]) \n" // V15 | U15 | V14 | U14
40 "addiu %[src_uv], %[src_uv], 32 \n"
41 "precrq.qb.ph $t9, $t1, $t0 \n" // V3 | V2 | V1 | V0
42 "precr.qb.ph $t0, $t1, $t0 \n" // U3 | U2 | U1 | U0
43 "precrq.qb.ph $t1, $t3, $t2 \n" // V7 | V6 | V5 | V4
44 "precr.qb.ph $t2, $t3, $t2 \n" // U7 | U6 | U5 | U4
45 "precrq.qb.ph $t3, $t6, $t5 \n" // V11 | V10 | V9 | V8
46 "precr.qb.ph $t5, $t6, $t5 \n" // U11 | U10 | U9 | U8
47 "precrq.qb.ph $t6, $t8, $t7 \n" // V15 | V14 | V13 | V12
48 "precr.qb.ph $t7, $t8, $t7 \n" // U15 | U14 | U13 | U12
49 "sw $t9, 0(%[dst_v]) \n"
50 "sw $t0, 0(%[dst_u]) \n"
51 "sw $t1, 4(%[dst_v]) \n"
52 "sw $t2, 4(%[dst_u]) \n"
53 "sw $t3, 8(%[dst_v]) \n"
54 "sw $t5, 8(%[dst_u]) \n"
55 "sw $t6, 12(%[dst_v]) \n"
56 "sw $t7, 12(%[dst_u]) \n"
57 "addiu %[dst_v], %[dst_v], 16 \n"
58 "bgtz $t4, 1b \n"
59 " addiu %[dst_u], %[dst_u], 16 \n"
60
61 "beqz %[width], 3f \n"
fbarchard@google.combb6bddc2012-10-14 06:41:17 +000062 " nop \n"
63
fbarchard@google.comdb694ed2012-10-17 21:54:04 +000064 "2: \n"
65 "lbu $t0, 0(%[src_uv]) \n"
66 "lbu $t1, 1(%[src_uv]) \n"
67 "addiu %[src_uv], %[src_uv], 2 \n"
68 "addiu %[width], %[width], -1 \n"
69 "sb $t0, 0(%[dst_u]) \n"
70 "sb $t1, 0(%[dst_v]) \n"
71 "addiu %[dst_u], %[dst_u], 1 \n"
72 "bgtz %[width], 2b \n"
73 " addiu %[dst_v], %[dst_v], 1 \n"
74
75 "3: \n"
76 ".set pop \n"
77 : [src_uv] "+r" (src_uv),
78 [width] "+r" (width),
79 [dst_u] "+r" (dst_u),
80 [dst_v] "+r" (dst_v)
81 :
82 : "t0", "t1", "t2", "t3",
83 "t4", "t5", "t6", "t7", "t8", "t9"
84 );
85}
86
87void SplitUV_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
88 uint8* dst_v, int width) {
89 __asm__ __volatile__ (
90 ".set push \n"
91 ".set noreorder \n"
92 "srl $t4, %[width], 4 \n" // multiplies of 16
93 "blez $t4, 2f \n"
94 " andi %[width], %[width], 0xf \n" // residual
95
96 "1: \n"
fbarchard@google.combb6bddc2012-10-14 06:41:17 +000097 "addiu $t4, $t4, -1 \n"
98 "lwr $t0, 0(%[src_uv]) \n"
99 "lwl $t0, 3(%[src_uv]) \n" // V1 | U1 | V0 | U0
100 "lwr $t1, 4(%[src_uv]) \n"
101 "lwl $t1, 7(%[src_uv]) \n" // V3 | U3 | V2 | U2
102 "lwr $t2, 8(%[src_uv]) \n"
103 "lwl $t2, 11(%[src_uv]) \n" // V5 | U5 | V4 | U4
104 "lwr $t3, 12(%[src_uv]) \n"
105 "lwl $t3, 15(%[src_uv]) \n" // V7 | U7 | V6 | U6
106 "lwr $t5, 16(%[src_uv]) \n"
107 "lwl $t5, 19(%[src_uv]) \n" // V9 | U9 | V8 | U8
108 "lwr $t6, 20(%[src_uv]) \n"
109 "lwl $t6, 23(%[src_uv]) \n" // V11 | U11 | V10 | U10
110 "lwr $t7, 24(%[src_uv]) \n"
111 "lwl $t7, 27(%[src_uv]) \n" // V13 | U13 | V12 | U12
112 "lwr $t8, 28(%[src_uv]) \n"
113 "lwl $t8, 31(%[src_uv]) \n" // V15 | U15 | V14 | U14
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000114 "precrq.qb.ph $t9, $t1, $t0 \n" // V3 | V2 | V1 | V0
115 "precr.qb.ph $t0, $t1, $t0 \n" // U3 | U2 | U1 | U0
116 "precrq.qb.ph $t1, $t3, $t2 \n" // V7 | V6 | V5 | V4
117 "precr.qb.ph $t2, $t3, $t2 \n" // U7 | U6 | U5 | U4
118 "precrq.qb.ph $t3, $t6, $t5 \n" // V11 | V10 | V9 | V8
119 "precr.qb.ph $t5, $t6, $t5 \n" // U11 | U10 | U9 | U8
120 "precrq.qb.ph $t6, $t8, $t7 \n" // V15 | V14 | V13 | V12
121 "precr.qb.ph $t7, $t8, $t7 \n" // U15 | U14 | U13 | U12
122 "addiu %[src_uv], %[src_uv], 32 \n"
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000123 "swr $t9, 0(%[dst_v]) \n"
124 "swl $t9, 3(%[dst_v]) \n"
125 "swr $t0, 0(%[dst_u]) \n"
126 "swl $t0, 3(%[dst_u]) \n"
127 "swr $t1, 4(%[dst_v]) \n"
128 "swl $t1, 7(%[dst_v]) \n"
129 "swr $t2, 4(%[dst_u]) \n"
130 "swl $t2, 7(%[dst_u]) \n"
131 "swr $t3, 8(%[dst_v]) \n"
132 "swl $t3, 11(%[dst_v]) \n"
133 "swr $t5, 8(%[dst_u]) \n"
134 "swl $t5, 11(%[dst_u]) \n"
135 "swr $t6, 12(%[dst_v]) \n"
136 "swl $t6, 15(%[dst_v]) \n"
137 "swr $t7, 12(%[dst_u]) \n"
138 "swl $t7, 15(%[dst_u]) \n"
139 "addiu %[dst_u], %[dst_u], 16 \n"
140 "bgtz $t4, 1b \n"
141 " addiu %[dst_v], %[dst_v], 16 \n"
142
143 "beqz %[width], 3f \n"
144 " nop \n"
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000145
fbarchard@google.comdb694ed2012-10-17 21:54:04 +0000146 "2: \n"
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000147 "lbu $t0, 0(%[src_uv]) \n"
148 "lbu $t1, 1(%[src_uv]) \n"
149 "addiu %[src_uv], %[src_uv], 2 \n"
150 "addiu %[width], %[width], -1 \n"
151 "sb $t0, 0(%[dst_u]) \n"
152 "sb $t1, 0(%[dst_v]) \n"
153 "addiu %[dst_u], %[dst_u], 1 \n"
154 "bgtz %[width], 2b \n"
155 " addiu %[dst_v], %[dst_v], 1 \n"
156
fbarchard@google.comdb694ed2012-10-17 21:54:04 +0000157 "3: \n"
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000158 ".set pop \n"
159 : [src_uv] "+r" (src_uv),
160 [width] "+r" (width),
161 [dst_u] "+r" (dst_u),
162 [dst_v] "+r" (dst_v)
163 :
164 : "t0", "t1", "t2", "t3",
165 "t4", "t5", "t6", "t7", "t8", "t9"
fbarchard@google.comca410052012-10-14 06:01:19 +0000166 );
167}
168#endif // HAS_SPLITUV_MIPS_DSPR2
fbarchard@google.combb6bddc2012-10-14 06:41:17 +0000169
fbarchard@google.comca410052012-10-14 06:01:19 +0000170#endif // __mips__
171
172#ifdef __cplusplus
173} // extern "C"
174} // namespace libyuv
175#endif