annotate libmpeg2/idct_mmx.c @ 14360:ed9c07a9915a

Cumulative sync with 1.840: 1.840: just typos. 1.839: alphabetical order 1.838: small fixes/improvements (by Sebastian Kraemer, except the hyphen stuff)
author gpoirier
date Tue, 04 Jan 2005 21:13:05 +0000
parents 96259a2f2142
children 1385ec491ffb
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
1 /*
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
2 * idct_mmx.c
10303
2c0b6ec77d39 sync with libmpeg2 cvs
alex
parents: 9852
diff changeset
3 * Copyright (C) 2000-2003 Michel Lespinasse <walken@zoy.org>
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
4 * Copyright (C) 1999-2000 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
5 *
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
6 * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
7 * See http://libmpeg2.sourceforge.net/ for updates.
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
8 *
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
9 * mpeg2dec is free software; you can redistribute it and/or modify
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
10 * it under the terms of the GNU General Public License as published by
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
11 * the Free Software Foundation; either version 2 of the License, or
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
12 * (at your option) any later version.
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
13 *
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
14 * mpeg2dec is distributed in the hope that it will be useful,
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
17 * GNU General Public License for more details.
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
18 *
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
19 * You should have received a copy of the GNU General Public License
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
20 * along with this program; if not, write to the Free Software
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
22 */
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
23
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
24 #include "config.h"
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
25
13864
96259a2f2142 enable mmx support on x86_64 in libmpeg2
aurel
parents: 12932
diff changeset
26 #if defined(ARCH_X86) || defined(ARCH_X86_64)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
27
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
28 #include <inttypes.h>
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
29
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
30 #include "mpeg2.h"
12932
d0a8810e155c Importing libmpeg2 from mpeg2dec-0.4.0b
henry
parents: 10303
diff changeset
31 #include "attributes.h"
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
32 #include "mpeg2_internal.h"
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
33 #include "mmx.h"
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
34
12932
d0a8810e155c Importing libmpeg2 from mpeg2dec-0.4.0b
henry
parents: 10303
diff changeset
35 #define ROW_SHIFT 15
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
36 #define COL_SHIFT 6
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
37
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
38 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
39 #define rounder(bias) {round (bias), round (bias)}
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
40
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
41
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
42 #if 0
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
43 /* C row IDCT - its just here to document the MMXEXT and MMX versions */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
44 static inline void idct_row (int16_t * row, int offset,
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
45 int16_t * table, int32_t * rounder)
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
46 {
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
47 int C1, C2, C3, C4, C5, C6, C7;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
48 int a0, a1, a2, a3, b0, b1, b2, b3;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
49
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
50 row += offset;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
51
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
52 C1 = table[1];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
53 C2 = table[2];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
54 C3 = table[3];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
55 C4 = table[4];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
56 C5 = table[5];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
57 C6 = table[6];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
58 C7 = table[7];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
59
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
60 a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
61 a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
62 a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
63 a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
64
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
65 b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
66 b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
67 b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
68 b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
69
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
70 row[0] = (a0 + b0) >> ROW_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
71 row[1] = (a1 + b1) >> ROW_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
72 row[2] = (a2 + b2) >> ROW_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
73 row[3] = (a3 + b3) >> ROW_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
74 row[4] = (a3 - b3) >> ROW_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
75 row[5] = (a2 - b2) >> ROW_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
76 row[6] = (a1 - b1) >> ROW_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
77 row[7] = (a0 - b0) >> ROW_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
78 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
79 #endif
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
80
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
81
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
82 /* MMXEXT row IDCT */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
83
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
84 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, -c4, -c2, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
85 c4, c6, c4, c6, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
86 c1, c3, -c1, -c5, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
87 c5, c7, c3, -c7, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
88 c4, -c6, c4, -c6, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
89 -c4, c2, c4, -c2, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
90 c5, -c1, c3, -c1, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
91 c7, c3, c7, -c5 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
92
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
93 static inline void mmxext_row_head (int16_t * const row, const int offset,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
94 const int16_t * const table)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
95 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
96 movq_m2r (*(row+offset), mm2); /* mm2 = x6 x4 x2 x0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
97
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
98 movq_m2r (*(row+offset+4), mm5); /* mm5 = x7 x5 x3 x1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
99 movq_r2r (mm2, mm0); /* mm0 = x6 x4 x2 x0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
100
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
101 movq_m2r (*table, mm3); /* mm3 = -C2 -C4 C2 C4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
102 movq_r2r (mm5, mm6); /* mm6 = x7 x5 x3 x1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
103
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
104 movq_m2r (*(table+4), mm4); /* mm4 = C6 C4 C6 C4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
105 pmaddwd_r2r (mm0, mm3); /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
106
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
107 pshufw_r2r (mm2, mm2, 0x4e); /* mm2 = x2 x0 x6 x4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
108 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
109
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
110 static inline void mmxext_row (const int16_t * const table,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
111 const int32_t * const rounder)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
112 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
113 movq_m2r (*(table+8), mm1); /* mm1 = -C5 -C1 C3 C1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
114 pmaddwd_r2r (mm2, mm4); /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
115
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
116 pmaddwd_m2r (*(table+16), mm0); /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
117 pshufw_r2r (mm6, mm6, 0x4e); /* mm6 = x3 x1 x7 x5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
118
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
119 movq_m2r (*(table+12), mm7); /* mm7 = -C7 C3 C7 C5 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
120 pmaddwd_r2r (mm5, mm1); /* mm1 = -C1*x5-C5*x7 C1*x1+C3*x3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
121
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
122 paddd_m2r (*rounder, mm3); /* mm3 += rounder */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
123 pmaddwd_r2r (mm6, mm7); /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
124
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
125 pmaddwd_m2r (*(table+20), mm2); /* mm2 = C4*x0-C2*x2 -C4*x4+C2*x6 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
126 paddd_r2r (mm4, mm3); /* mm3 = a1 a0 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
127
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
128 pmaddwd_m2r (*(table+24), mm5); /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
129 movq_r2r (mm3, mm4); /* mm4 = a1 a0 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
130
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
131 pmaddwd_m2r (*(table+28), mm6); /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
132 paddd_r2r (mm7, mm1); /* mm1 = b1 b0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
133
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
134 paddd_m2r (*rounder, mm0); /* mm0 += rounder */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
135 psubd_r2r (mm1, mm3); /* mm3 = a1-b1 a0-b0 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
136
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
137 psrad_i2r (ROW_SHIFT, mm3); /* mm3 = y6 y7 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
138 paddd_r2r (mm4, mm1); /* mm1 = a1+b1 a0+b0 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
139
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
140 paddd_r2r (mm2, mm0); /* mm0 = a3 a2 + rounder */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
141 psrad_i2r (ROW_SHIFT, mm1); /* mm1 = y1 y0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
142
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
143 paddd_r2r (mm6, mm5); /* mm5 = b3 b2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
144 movq_r2r (mm0, mm4); /* mm4 = a3 a2 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
145
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
146 paddd_r2r (mm5, mm0); /* mm0 = a3+b3 a2+b2 + rounder */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
147 psubd_r2r (mm5, mm4); /* mm4 = a3-b3 a2-b2 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
148 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
149
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
150 static inline void mmxext_row_tail (int16_t * const row, const int store)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
151 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
152 psrad_i2r (ROW_SHIFT, mm0); /* mm0 = y3 y2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
153
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
154 psrad_i2r (ROW_SHIFT, mm4); /* mm4 = y4 y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
155
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
156 packssdw_r2r (mm0, mm1); /* mm1 = y3 y2 y1 y0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
157
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
158 packssdw_r2r (mm3, mm4); /* mm4 = y6 y7 y4 y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
159
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
160 movq_r2m (mm1, *(row+store)); /* save y3 y2 y1 y0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
161 pshufw_r2r (mm4, mm4, 0xb1); /* mm4 = y7 y6 y5 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
162
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
163 /* slot */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
164
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
165 movq_r2m (mm4, *(row+store+4)); /* save y7 y6 y5 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
166 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
167
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
168 static inline void mmxext_row_mid (int16_t * const row, const int store,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
169 const int offset,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
170 const int16_t * const table)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
171 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
172 movq_m2r (*(row+offset), mm2); /* mm2 = x6 x4 x2 x0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
173 psrad_i2r (ROW_SHIFT, mm0); /* mm0 = y3 y2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
174
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
175 movq_m2r (*(row+offset+4), mm5); /* mm5 = x7 x5 x3 x1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
176 psrad_i2r (ROW_SHIFT, mm4); /* mm4 = y4 y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
177
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
178 packssdw_r2r (mm0, mm1); /* mm1 = y3 y2 y1 y0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
179 movq_r2r (mm5, mm6); /* mm6 = x7 x5 x3 x1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
180
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
181 packssdw_r2r (mm3, mm4); /* mm4 = y6 y7 y4 y5 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
182 movq_r2r (mm2, mm0); /* mm0 = x6 x4 x2 x0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
183
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
184 movq_r2m (mm1, *(row+store)); /* save y3 y2 y1 y0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
185 pshufw_r2r (mm4, mm4, 0xb1); /* mm4 = y7 y6 y5 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
186
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
187 movq_m2r (*table, mm3); /* mm3 = -C2 -C4 C2 C4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
188 movq_r2m (mm4, *(row+store+4)); /* save y7 y6 y5 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
189
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
190 pmaddwd_r2r (mm0, mm3); /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
191
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
192 movq_m2r (*(table+4), mm4); /* mm4 = C6 C4 C6 C4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
193 pshufw_r2r (mm2, mm2, 0x4e); /* mm2 = x2 x0 x6 x4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
194 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
195
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
196
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
197 /* MMX row IDCT */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
198
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
199 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
200 c4, c6, -c4, -c2, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
201 c1, c3, c3, -c7, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
202 c5, c7, -c1, -c5, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
203 c4, -c6, c4, -c2, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
204 -c4, c2, c4, -c6, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
205 c5, -c1, c7, -c5, \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
206 c7, c3, c3, -c1 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
207
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
208 static inline void mmx_row_head (int16_t * const row, const int offset,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
209 const int16_t * const table)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
210 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
211 movq_m2r (*(row+offset), mm2); /* mm2 = x6 x4 x2 x0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
212
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
213 movq_m2r (*(row+offset+4), mm5); /* mm5 = x7 x5 x3 x1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
214 movq_r2r (mm2, mm0); /* mm0 = x6 x4 x2 x0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
215
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
216 movq_m2r (*table, mm3); /* mm3 = C6 C4 C2 C4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
217 movq_r2r (mm5, mm6); /* mm6 = x7 x5 x3 x1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
218
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
219 punpckldq_r2r (mm0, mm0); /* mm0 = x2 x0 x2 x0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
220
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
221 movq_m2r (*(table+4), mm4); /* mm4 = -C2 -C4 C6 C4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
222 pmaddwd_r2r (mm0, mm3); /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
223
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
224 movq_m2r (*(table+8), mm1); /* mm1 = -C7 C3 C3 C1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
225 punpckhdq_r2r (mm2, mm2); /* mm2 = x6 x4 x6 x4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
226 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
227
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
228 static inline void mmx_row (const int16_t * const table,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
229 const int32_t * const rounder)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
230 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
231 pmaddwd_r2r (mm2, mm4); /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
232 punpckldq_r2r (mm5, mm5); /* mm5 = x3 x1 x3 x1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
233
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
234 pmaddwd_m2r (*(table+16), mm0); /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
235 punpckhdq_r2r (mm6, mm6); /* mm6 = x7 x5 x7 x5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
236
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
237 movq_m2r (*(table+12), mm7); /* mm7 = -C5 -C1 C7 C5 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
238 pmaddwd_r2r (mm5, mm1); /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
239
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
240 paddd_m2r (*rounder, mm3); /* mm3 += rounder */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
241 pmaddwd_r2r (mm6, mm7); /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
242
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
243 pmaddwd_m2r (*(table+20), mm2); /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
244 paddd_r2r (mm4, mm3); /* mm3 = a1 a0 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
245
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
246 pmaddwd_m2r (*(table+24), mm5); /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
247 movq_r2r (mm3, mm4); /* mm4 = a1 a0 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
248
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
249 pmaddwd_m2r (*(table+28), mm6); /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
250 paddd_r2r (mm7, mm1); /* mm1 = b1 b0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
251
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
252 paddd_m2r (*rounder, mm0); /* mm0 += rounder */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
253 psubd_r2r (mm1, mm3); /* mm3 = a1-b1 a0-b0 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
254
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
255 psrad_i2r (ROW_SHIFT, mm3); /* mm3 = y6 y7 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
256 paddd_r2r (mm4, mm1); /* mm1 = a1+b1 a0+b0 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
257
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
258 paddd_r2r (mm2, mm0); /* mm0 = a3 a2 + rounder */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
259 psrad_i2r (ROW_SHIFT, mm1); /* mm1 = y1 y0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
260
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
261 paddd_r2r (mm6, mm5); /* mm5 = b3 b2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
262 movq_r2r (mm0, mm7); /* mm7 = a3 a2 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
263
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
264 paddd_r2r (mm5, mm0); /* mm0 = a3+b3 a2+b2 + rounder */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
265 psubd_r2r (mm5, mm7); /* mm7 = a3-b3 a2-b2 + rounder */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
266 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
267
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
268 static inline void mmx_row_tail (int16_t * const row, const int store)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
269 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
270 psrad_i2r (ROW_SHIFT, mm0); /* mm0 = y3 y2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
271
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
272 psrad_i2r (ROW_SHIFT, mm7); /* mm7 = y4 y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
273
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
274 packssdw_r2r (mm0, mm1); /* mm1 = y3 y2 y1 y0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
275
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
276 packssdw_r2r (mm3, mm7); /* mm7 = y6 y7 y4 y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
277
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
278 movq_r2m (mm1, *(row+store)); /* save y3 y2 y1 y0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
279 movq_r2r (mm7, mm4); /* mm4 = y6 y7 y4 y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
280
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
281 pslld_i2r (16, mm7); /* mm7 = y7 0 y5 0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
282
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
283 psrld_i2r (16, mm4); /* mm4 = 0 y6 0 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
284
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
285 por_r2r (mm4, mm7); /* mm7 = y7 y6 y5 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
286
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
287 /* slot */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
288
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
289 movq_r2m (mm7, *(row+store+4)); /* save y7 y6 y5 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
290 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
291
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
292 static inline void mmx_row_mid (int16_t * const row, const int store,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
293 const int offset, const int16_t * const table)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
294 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
295 movq_m2r (*(row+offset), mm2); /* mm2 = x6 x4 x2 x0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
296 psrad_i2r (ROW_SHIFT, mm0); /* mm0 = y3 y2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
297
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
298 movq_m2r (*(row+offset+4), mm5); /* mm5 = x7 x5 x3 x1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
299 psrad_i2r (ROW_SHIFT, mm7); /* mm7 = y4 y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
300
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
301 packssdw_r2r (mm0, mm1); /* mm1 = y3 y2 y1 y0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
302 movq_r2r (mm5, mm6); /* mm6 = x7 x5 x3 x1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
303
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
304 packssdw_r2r (mm3, mm7); /* mm7 = y6 y7 y4 y5 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
305 movq_r2r (mm2, mm0); /* mm0 = x6 x4 x2 x0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
306
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
307 movq_r2m (mm1, *(row+store)); /* save y3 y2 y1 y0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
308 movq_r2r (mm7, mm1); /* mm1 = y6 y7 y4 y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
309
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
310 punpckldq_r2r (mm0, mm0); /* mm0 = x2 x0 x2 x0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
311 psrld_i2r (16, mm7); /* mm7 = 0 y6 0 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
312
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
313 movq_m2r (*table, mm3); /* mm3 = C6 C4 C2 C4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
314 pslld_i2r (16, mm1); /* mm1 = y7 0 y5 0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
315
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
316 movq_m2r (*(table+4), mm4); /* mm4 = -C2 -C4 C6 C4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
317 por_r2r (mm1, mm7); /* mm7 = y7 y6 y5 y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
318
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
319 movq_m2r (*(table+8), mm1); /* mm1 = -C7 C3 C3 C1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
320 punpckhdq_r2r (mm2, mm2); /* mm2 = x6 x4 x6 x4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
321
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
322 movq_r2m (mm7, *(row+store+4)); /* save y7 y6 y5 y4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
323 pmaddwd_r2r (mm0, mm3); /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
324 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
325
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
326
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
327 #if 0
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
328 /* C column IDCT - its just here to document the MMXEXT and MMX versions */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
329 static inline void idct_col (int16_t * col, int offset)
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
330 {
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
331 /* multiplication - as implemented on mmx */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
332 #define F(c,x) (((c) * (x)) >> 16)
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
333
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
334 /* saturation - it helps us handle torture test cases */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
335 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
336
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
337 int16_t x0, x1, x2, x3, x4, x5, x6, x7;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
338 int16_t y0, y1, y2, y3, y4, y5, y6, y7;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
339 int16_t a0, a1, a2, a3, b0, b1, b2, b3;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
340 int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
341
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
342 col += offset;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
343
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
344 x0 = col[0*8];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
345 x1 = col[1*8];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
346 x2 = col[2*8];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
347 x3 = col[3*8];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
348 x4 = col[4*8];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
349 x5 = col[5*8];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
350 x6 = col[6*8];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
351 x7 = col[7*8];
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
352
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
353 u04 = S (x0 + x4);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
354 v04 = S (x0 - x4);
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
355 u26 = S (F (T2, x6) + x2);
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
356 v26 = S (F (T2, x2) - x6);
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
357
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
358 a0 = S (u04 + u26);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
359 a1 = S (v04 + v26);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
360 a2 = S (v04 - v26);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
361 a3 = S (u04 - u26);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
362
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
363 u17 = S (F (T1, x7) + x1);
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
364 v17 = S (F (T1, x1) - x7);
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
365 u35 = S (F (T3, x5) + x3);
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
366 v35 = S (F (T3, x3) - x5);
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
367
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
368 b0 = S (u17 + u35);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
369 b3 = S (v17 - v35);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
370 u12 = S (u17 - u35);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
371 v12 = S (v17 + v35);
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
372 u12 = S (2 * F (C4, u12));
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
373 v12 = S (2 * F (C4, v12));
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
374 b1 = S (u12 + v12);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
375 b2 = S (u12 - v12);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
376
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
377 y0 = S (a0 + b0) >> COL_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
378 y1 = S (a1 + b1) >> COL_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
379 y2 = S (a2 + b2) >> COL_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
380 y3 = S (a3 + b3) >> COL_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
381
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
382 y4 = S (a3 - b3) >> COL_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
383 y5 = S (a2 - b2) >> COL_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
384 y6 = S (a1 - b1) >> COL_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
385 y7 = S (a0 - b0) >> COL_SHIFT;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
386
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
387 col[0*8] = y0;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
388 col[1*8] = y1;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
389 col[2*8] = y2;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
390 col[3*8] = y3;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
391 col[4*8] = y4;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
392 col[5*8] = y5;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
393 col[6*8] = y6;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
394 col[7*8] = y7;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
395 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
396 #endif
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
397
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
398
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
399 /* MMX column IDCT */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
400 static inline void idct_col (int16_t * const col, const int offset)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
401 {
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
402 #define T1 13036
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
403 #define T2 27146
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
404 #define T3 43790
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
405 #define C4 23170
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
406
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
407 static const short _T1[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
408 static const short _T2[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
409 static const short _T3[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
410 static const short _C4[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
411
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
412 /* column code adapted from peter gubanov */
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
413 /* http://www.elecard.com/peter/idct.shtml */
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
414
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
415 movq_m2r (*_T1, mm0); /* mm0 = T1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
416
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
417 movq_m2r (*(col+offset+1*8), mm1); /* mm1 = x1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
418 movq_r2r (mm0, mm2); /* mm2 = T1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
419
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
420 movq_m2r (*(col+offset+7*8), mm4); /* mm4 = x7 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
421 pmulhw_r2r (mm1, mm0); /* mm0 = T1*x1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
422
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
423 movq_m2r (*_T3, mm5); /* mm5 = T3 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
424 pmulhw_r2r (mm4, mm2); /* mm2 = T1*x7 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
425
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
426 movq_m2r (*(col+offset+5*8), mm6); /* mm6 = x5 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
427 movq_r2r (mm5, mm7); /* mm7 = T3-1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
428
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
429 movq_m2r (*(col+offset+3*8), mm3); /* mm3 = x3 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
430 psubsw_r2r (mm4, mm0); /* mm0 = v17 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
431
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
432 movq_m2r (*_T2, mm4); /* mm4 = T2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
433 pmulhw_r2r (mm3, mm5); /* mm5 = (T3-1)*x3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
434
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
435 paddsw_r2r (mm2, mm1); /* mm1 = u17 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
436 pmulhw_r2r (mm6, mm7); /* mm7 = (T3-1)*x5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
437
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
438 /* slot */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
439
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
440 movq_r2r (mm4, mm2); /* mm2 = T2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
441 paddsw_r2r (mm3, mm5); /* mm5 = T3*x3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
442
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
443 pmulhw_m2r (*(col+offset+2*8), mm4);/* mm4 = T2*x2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
444 paddsw_r2r (mm6, mm7); /* mm7 = T3*x5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
445
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
446 psubsw_r2r (mm6, mm5); /* mm5 = v35 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
447 paddsw_r2r (mm3, mm7); /* mm7 = u35 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
448
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
449 movq_m2r (*(col+offset+6*8), mm3); /* mm3 = x6 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
450 movq_r2r (mm0, mm6); /* mm6 = v17 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
451
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
452 pmulhw_r2r (mm3, mm2); /* mm2 = T2*x6 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
453 psubsw_r2r (mm5, mm0); /* mm0 = b3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
454
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
455 psubsw_r2r (mm3, mm4); /* mm4 = v26 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
456 paddsw_r2r (mm6, mm5); /* mm5 = v12 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
457
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
458 movq_r2m (mm0, *(col+offset+3*8)); /* save b3 in scratch0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
459 movq_r2r (mm1, mm6); /* mm6 = u17 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
460
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
461 paddsw_m2r (*(col+offset+2*8), mm2);/* mm2 = u26 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
462 paddsw_r2r (mm7, mm6); /* mm6 = b0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
463
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
464 psubsw_r2r (mm7, mm1); /* mm1 = u12 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
465 movq_r2r (mm1, mm7); /* mm7 = u12 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
466
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
467 movq_m2r (*(col+offset+0*8), mm3); /* mm3 = x0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
468 paddsw_r2r (mm5, mm1); /* mm1 = u12+v12 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
469
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
470 movq_m2r (*_C4, mm0); /* mm0 = C4/2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
471 psubsw_r2r (mm5, mm7); /* mm7 = u12-v12 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
472
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
473 movq_r2m (mm6, *(col+offset+5*8)); /* save b0 in scratch1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
474 pmulhw_r2r (mm0, mm1); /* mm1 = b1/2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
475
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
476 movq_r2r (mm4, mm6); /* mm6 = v26 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
477 pmulhw_r2r (mm0, mm7); /* mm7 = b2/2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
478
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
479 movq_m2r (*(col+offset+4*8), mm5); /* mm5 = x4 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
480 movq_r2r (mm3, mm0); /* mm0 = x0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
481
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
482 psubsw_r2r (mm5, mm3); /* mm3 = v04 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
483 paddsw_r2r (mm5, mm0); /* mm0 = u04 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
484
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
485 paddsw_r2r (mm3, mm4); /* mm4 = a1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
486 movq_r2r (mm0, mm5); /* mm5 = u04 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
487
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
488 psubsw_r2r (mm6, mm3); /* mm3 = a2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
489 paddsw_r2r (mm2, mm5); /* mm5 = a0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
490
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
491 paddsw_r2r (mm1, mm1); /* mm1 = b1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
492 psubsw_r2r (mm2, mm0); /* mm0 = a3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
493
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
494 paddsw_r2r (mm7, mm7); /* mm7 = b2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
495 movq_r2r (mm3, mm2); /* mm2 = a2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
496
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
497 movq_r2r (mm4, mm6); /* mm6 = a1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
498 paddsw_r2r (mm7, mm3); /* mm3 = a2+b2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
499
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
500 psraw_i2r (COL_SHIFT, mm3); /* mm3 = y2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
501 paddsw_r2r (mm1, mm4); /* mm4 = a1+b1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
502
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
503 psraw_i2r (COL_SHIFT, mm4); /* mm4 = y1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
504 psubsw_r2r (mm1, mm6); /* mm6 = a1-b1 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
505
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
506 movq_m2r (*(col+offset+5*8), mm1); /* mm1 = b0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
507 psubsw_r2r (mm7, mm2); /* mm2 = a2-b2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
508
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
509 psraw_i2r (COL_SHIFT, mm6); /* mm6 = y6 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
510 movq_r2r (mm5, mm7); /* mm7 = a0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
511
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
512 movq_r2m (mm4, *(col+offset+1*8)); /* save y1 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
513 psraw_i2r (COL_SHIFT, mm2); /* mm2 = y5 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
514
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
515 movq_r2m (mm3, *(col+offset+2*8)); /* save y2 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
516 paddsw_r2r (mm1, mm5); /* mm5 = a0+b0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
517
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
518 movq_m2r (*(col+offset+3*8), mm4); /* mm4 = b3 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
519 psubsw_r2r (mm1, mm7); /* mm7 = a0-b0 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
520
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
521 psraw_i2r (COL_SHIFT, mm5); /* mm5 = y0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
522 movq_r2r (mm0, mm3); /* mm3 = a3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
523
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
524 movq_r2m (mm2, *(col+offset+5*8)); /* save y5 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
525 psubsw_r2r (mm4, mm3); /* mm3 = a3-b3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
526
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
527 psraw_i2r (COL_SHIFT, mm7); /* mm7 = y7 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
528 paddsw_r2r (mm0, mm4); /* mm4 = a3+b3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
529
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
530 movq_r2m (mm5, *(col+offset+0*8)); /* save y0 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
531 psraw_i2r (COL_SHIFT, mm3); /* mm3 = y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
532
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
533 movq_r2m (mm6, *(col+offset+6*8)); /* save y6 */
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
534 psraw_i2r (COL_SHIFT, mm4); /* mm4 = y3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
535
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
536 movq_r2m (mm7, *(col+offset+7*8)); /* save y7 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
537
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
538 movq_r2m (mm3, *(col+offset+4*8)); /* save y4 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
539
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
540 movq_r2m (mm4, *(col+offset+3*8)); /* save y3 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
541 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
542
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
543
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
544 static const int32_t rounder0[] ATTR_ALIGN(8) =
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
545 rounder ((1 << (COL_SHIFT - 1)) - 0.5);
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
546 static const int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
547 static const int32_t rounder1[] ATTR_ALIGN(8) =
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
548 rounder (1.25683487303); /* C1*(C1/C4+C1+C7)/2 */
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
549 static const int32_t rounder7[] ATTR_ALIGN(8) =
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
550 rounder (-0.25); /* C1*(C7/C4+C7-C1)/2 */
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
551 static const int32_t rounder2[] ATTR_ALIGN(8) =
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
552 rounder (0.60355339059); /* C2 * (C6+C2)/2 */
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
553 static const int32_t rounder6[] ATTR_ALIGN(8) =
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
554 rounder (-0.25); /* C2 * (C6-C2)/2 */
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
555 static const int32_t rounder3[] ATTR_ALIGN(8) =
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
556 rounder (0.087788325588); /* C3*(-C3/C4+C3+C5)/2 */
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
557 static const int32_t rounder5[] ATTR_ALIGN(8) =
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
558 rounder (-0.441341716183); /* C3*(-C5/C4+C5-C3)/2 */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
559
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
560
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
561 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
562 static inline void idct (int16_t * const block) \
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
563 { \
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
564 static const int16_t table04[] ATTR_ALIGN(16) = \
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
565 table (22725, 21407, 19266, 16384, 12873, 8867, 4520); \
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
566 static const int16_t table17[] ATTR_ALIGN(16) = \
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
567 table (31521, 29692, 26722, 22725, 17855, 12299, 6270); \
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
568 static const int16_t table26[] ATTR_ALIGN(16) = \
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
569 table (29692, 27969, 25172, 21407, 16819, 11585, 5906); \
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
570 static const int16_t table35[] ATTR_ALIGN(16) = \
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
571 table (26722, 25172, 22654, 19266, 15137, 10426, 5315); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
572 \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
573 idct_row_head (block, 0*8, table04); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
574 idct_row (table04, rounder0); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
575 idct_row_mid (block, 0*8, 4*8, table04); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
576 idct_row (table04, rounder4); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
577 idct_row_mid (block, 4*8, 1*8, table17); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
578 idct_row (table17, rounder1); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
579 idct_row_mid (block, 1*8, 7*8, table17); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
580 idct_row (table17, rounder7); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
581 idct_row_mid (block, 7*8, 2*8, table26); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
582 idct_row (table26, rounder2); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
583 idct_row_mid (block, 2*8, 6*8, table26); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
584 idct_row (table26, rounder6); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
585 idct_row_mid (block, 6*8, 3*8, table35); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
586 idct_row (table35, rounder3); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
587 idct_row_mid (block, 3*8, 5*8, table35); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
588 idct_row (table35, rounder5); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
589 idct_row_tail (block, 5*8); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
590 \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
591 idct_col (block, 0); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
592 idct_col (block, 4); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
593 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
594
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
595
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
596 #define COPY_MMX(offset,r0,r1,r2) \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
597 do { \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
598 movq_m2r (*(block+offset), r0); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
599 dest += stride; \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
600 movq_m2r (*(block+offset+4), r1); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
601 movq_r2m (r2, *dest); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
602 packuswb_r2r (r1, r0); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
603 } while (0)
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
604
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
605 static inline void block_copy (int16_t * const block, uint8_t * dest,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
606 const int stride)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
607 {
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
608 movq_m2r (*(block+0*8), mm0);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
609 movq_m2r (*(block+0*8+4), mm1);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
610 movq_m2r (*(block+1*8), mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
611 packuswb_r2r (mm1, mm0);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
612 movq_m2r (*(block+1*8+4), mm3);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
613 movq_r2m (mm0, *dest);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
614 packuswb_r2r (mm3, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
615 COPY_MMX (2*8, mm0, mm1, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
616 COPY_MMX (3*8, mm2, mm3, mm0);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
617 COPY_MMX (4*8, mm0, mm1, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
618 COPY_MMX (5*8, mm2, mm3, mm0);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
619 COPY_MMX (6*8, mm0, mm1, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
620 COPY_MMX (7*8, mm2, mm3, mm0);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
621 movq_r2m (mm2, *(dest+stride));
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
622 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
623
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
624
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
625 #define ADD_MMX(offset,r1,r2,r3,r4) \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
626 do { \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
627 movq_m2r (*(dest+2*stride), r1); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
628 packuswb_r2r (r4, r3); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
629 movq_r2r (r1, r2); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
630 dest += stride; \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
631 movq_r2m (r3, *dest); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
632 punpcklbw_r2r (mm0, r1); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
633 paddsw_m2r (*(block+offset), r1); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
634 punpckhbw_r2r (mm0, r2); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
635 paddsw_m2r (*(block+offset+4), r2); \
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
636 } while (0)
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
637
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
638 static inline void block_add (int16_t * const block, uint8_t * dest,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
639 const int stride)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
640 {
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
641 movq_m2r (*dest, mm1);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
642 pxor_r2r (mm0, mm0);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
643 movq_m2r (*(dest+stride), mm3);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
644 movq_r2r (mm1, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
645 punpcklbw_r2r (mm0, mm1);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
646 movq_r2r (mm3, mm4);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
647 paddsw_m2r (*(block+0*8), mm1);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
648 punpckhbw_r2r (mm0, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
649 paddsw_m2r (*(block+0*8+4), mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
650 punpcklbw_r2r (mm0, mm3);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
651 paddsw_m2r (*(block+1*8), mm3);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
652 packuswb_r2r (mm2, mm1);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
653 punpckhbw_r2r (mm0, mm4);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
654 movq_r2m (mm1, *dest);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
655 paddsw_m2r (*(block+1*8+4), mm4);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
656 ADD_MMX (2*8, mm1, mm2, mm3, mm4);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
657 ADD_MMX (3*8, mm3, mm4, mm1, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
658 ADD_MMX (4*8, mm1, mm2, mm3, mm4);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
659 ADD_MMX (5*8, mm3, mm4, mm1, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
660 ADD_MMX (6*8, mm1, mm2, mm3, mm4);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
661 ADD_MMX (7*8, mm3, mm4, mm1, mm2);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
662 packuswb_r2r (mm4, mm3);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
663 movq_r2m (mm3, *(dest+stride));
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
664 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
665
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
666
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
667 static inline void block_zero (int16_t * const block)
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
668 {
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
669 pxor_r2r (mm0, mm0);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
670 movq_r2m (mm0, *(block+0*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
671 movq_r2m (mm0, *(block+1*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
672 movq_r2m (mm0, *(block+2*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
673 movq_r2m (mm0, *(block+3*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
674 movq_r2m (mm0, *(block+4*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
675 movq_r2m (mm0, *(block+5*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
676 movq_r2m (mm0, *(block+6*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
677 movq_r2m (mm0, *(block+7*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
678 movq_r2m (mm0, *(block+8*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
679 movq_r2m (mm0, *(block+9*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
680 movq_r2m (mm0, *(block+10*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
681 movq_r2m (mm0, *(block+11*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
682 movq_r2m (mm0, *(block+12*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
683 movq_r2m (mm0, *(block+13*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
684 movq_r2m (mm0, *(block+14*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
685 movq_r2m (mm0, *(block+15*4));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
686 }
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
687
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
688
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
689 #define CPU_MMXEXT 0
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
690 #define CPU_MMX 1
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
691
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
692 #define dup4(reg) \
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
693 do { \
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
694 if (cpu != CPU_MMXEXT) { \
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
695 punpcklwd_r2r (reg, reg); \
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
696 punpckldq_r2r (reg, reg); \
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
697 } else \
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
698 pshufw_r2r (reg, reg, 0x00); \
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
699 } while (0)
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
700
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
701 static inline void block_add_DC (int16_t * const block, uint8_t * dest,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
702 const int stride, const int cpu)
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
703 {
12932
d0a8810e155c Importing libmpeg2 from mpeg2dec-0.4.0b
henry
parents: 10303
diff changeset
704 movd_v2r ((block[0] + 64) >> 7, mm0);
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
705 pxor_r2r (mm1, mm1);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
706 movq_m2r (*dest, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
707 dup4 (mm0);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
708 psubsw_r2r (mm0, mm1);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
709 packuswb_r2r (mm0, mm0);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
710 paddusb_r2r (mm0, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
711 packuswb_r2r (mm1, mm1);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
712 movq_m2r (*(dest + stride), mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
713 psubusb_r2r (mm1, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
714 block[0] = 0;
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
715 paddusb_r2r (mm0, mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
716 movq_r2m (mm2, *dest);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
717 psubusb_r2r (mm1, mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
718 movq_m2r (*(dest + 2*stride), mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
719 dest += stride;
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
720 movq_r2m (mm3, *dest);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
721 paddusb_r2r (mm0, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
722 movq_m2r (*(dest + 2*stride), mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
723 psubusb_r2r (mm1, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
724 dest += stride;
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
725 paddusb_r2r (mm0, mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
726 movq_r2m (mm2, *dest);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
727 psubusb_r2r (mm1, mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
728 movq_m2r (*(dest + 2*stride), mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
729 dest += stride;
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
730 movq_r2m (mm3, *dest);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
731 paddusb_r2r (mm0, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
732 movq_m2r (*(dest + 2*stride), mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
733 psubusb_r2r (mm1, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
734 dest += stride;
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
735 paddusb_r2r (mm0, mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
736 movq_r2m (mm2, *dest);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
737 psubusb_r2r (mm1, mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
738 movq_m2r (*(dest + 2*stride), mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
739 dest += stride;
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
740 movq_r2m (mm3, *dest);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
741 paddusb_r2r (mm0, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
742 movq_m2r (*(dest + 2*stride), mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
743 psubusb_r2r (mm1, mm2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
744 block[63] = 0;
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
745 paddusb_r2r (mm0, mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
746 movq_r2m (mm2, *(dest + stride));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
747 psubusb_r2r (mm1, mm3);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
748 movq_r2m (mm3, *(dest + 2*stride));
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
749 }
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
750
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
751
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
752 declare_idct (mmxext_idct, mmxext_table,
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
753 mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
754
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
755 void mpeg2_idct_copy_mmxext (int16_t * const block, uint8_t * const dest,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
756 const int stride)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
757 {
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
758 mmxext_idct (block);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
759 block_copy (block, dest, stride);
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
760 block_zero (block);
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
761 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
762
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
763 void mpeg2_idct_add_mmxext (const int last, int16_t * const block,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
764 uint8_t * const dest, const int stride)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
765 {
12932
d0a8810e155c Importing libmpeg2 from mpeg2dec-0.4.0b
henry
parents: 10303
diff changeset
766 if (last != 129 || (block[0] & (7 << 4)) == (4 << 4)) {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
767 mmxext_idct (block);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
768 block_add (block, dest, stride);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
769 block_zero (block);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
770 } else
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
771 block_add_DC (block, dest, stride, CPU_MMXEXT);
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
772 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
773
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
774
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
775 declare_idct (mmx_idct, mmx_table,
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
776 mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
777
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
778 void mpeg2_idct_copy_mmx (int16_t * const block, uint8_t * const dest,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
779 const int stride)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
780 {
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
781 mmx_idct (block);
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
782 block_copy (block, dest, stride);
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
783 block_zero (block);
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
784 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
785
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
786 void mpeg2_idct_add_mmx (const int last, int16_t * const block,
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
787 uint8_t * const dest, const int stride)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
788 {
12932
d0a8810e155c Importing libmpeg2 from mpeg2dec-0.4.0b
henry
parents: 10303
diff changeset
789 if (last != 129 || (block[0] & (7 << 4)) == (4 << 4)) {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
790 mmx_idct (block);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
791 block_add (block, dest, stride);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
792 block_zero (block);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
793 } else
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
794 block_add_DC (block, dest, stride, CPU_MMX);
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
795 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
796
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
797
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
798 void mpeg2_idct_mmx_init (void)
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
799 {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
800 extern uint8_t mpeg2_scan_norm[64];
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
801 extern uint8_t mpeg2_scan_alt[64];
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
802 int i, j;
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
803
36
846535ace7a2 libmpeg2-0.2.0 merge
arpi_esp
parents: 1
diff changeset
804 /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
805
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
806 for (i = 0; i < 64; i++) {
9852
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
807 j = mpeg2_scan_norm[i];
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
808 mpeg2_scan_norm[i] = (j & 0x38) | ((j & 6) >> 1) | ((j & 1) << 2);
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
809 j = mpeg2_scan_alt[i];
47984e3f54ce Importing libmpeg2 from mpeg2dec-0.3.1
arpi
parents: 1437
diff changeset
810 mpeg2_scan_alt[i] = (j & 0x38) | ((j & 6) >> 1) | ((j & 1) << 2);
1
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
811 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
812 }
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
813
3b5f5d1c5041 Initial revision
arpi_esp
parents:
diff changeset
814 #endif