annotate i386/idct_mmx.c @ 436:35de17dd6ed8 libavcodec

* undefine local defines when they are no longer needed
author kabi
date Mon, 27 May 2002 08:37:13 +0000
parents 0b09bd08ef4b
children f65d87bfdd5a
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
30
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
1 /*
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
2 * Note: For libavcodec, this code can also be used under the LGPL license
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
3 */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
4 /*
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
5 * idct_mmx.c
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
6 * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
7 *
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
8 * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
9 *
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
10 * mpeg2dec is free software; you can redistribute it and/or modify
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
11 * it under the terms of the GNU General Public License as published by
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
12 * the Free Software Foundation; either version 2 of the License, or
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
13 * (at your option) any later version.
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
14 *
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
15 * mpeg2dec is distributed in the hope that it will be useful,
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
18 * GNU General Public License for more details.
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
19 *
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
20 * You should have received a copy of the GNU General Public License
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
21 * along with this program; if not, write to the Free Software
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
22 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
23 */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
24
76
0b09bd08ef4b win32 fixes
glantau
parents: 30
diff changeset
25 #include "../common.h"
30
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
26
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
27 #include "mmx.h"
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
28
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
29 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
30
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
31 #define ROW_SHIFT 11
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
32 #define COL_SHIFT 6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
33
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
34 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
35 #define rounder(bias) {round (bias), round (bias)}
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
36
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
37 #if 0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
38 /* C row IDCT - its just here to document the MMXEXT and MMX versions */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
39 static inline void idct_row (int16_t * row, int offset,
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
40 int16_t * table, int32_t * rounder)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
41 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
42 int C1, C2, C3, C4, C5, C6, C7;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
43 int a0, a1, a2, a3, b0, b1, b2, b3;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
44
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
45 row += offset;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
46
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
47 C1 = table[1];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
48 C2 = table[2];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
49 C3 = table[3];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
50 C4 = table[4];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
51 C5 = table[5];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
52 C6 = table[6];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
53 C7 = table[7];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
54
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
55 a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
56 a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
57 a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
58 a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
59
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
60 b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
61 b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
62 b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
63 b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
64
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
65 row[0] = (a0 + b0) >> ROW_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
66 row[1] = (a1 + b1) >> ROW_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
67 row[2] = (a2 + b2) >> ROW_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
68 row[3] = (a3 + b3) >> ROW_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
69 row[4] = (a3 - b3) >> ROW_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
70 row[5] = (a2 - b2) >> ROW_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
71 row[6] = (a1 - b1) >> ROW_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
72 row[7] = (a0 - b0) >> ROW_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
73 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
74 #endif
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
75
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
76
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
77 /* MMXEXT row IDCT */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
78
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
79 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, -c4, -c2, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
80 c4, c6, c4, c6, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
81 c1, c3, -c1, -c5, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
82 c5, c7, c3, -c7, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
83 c4, -c6, c4, -c6, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
84 -c4, c2, c4, -c2, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
85 c5, -c1, c3, -c1, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
86 c7, c3, c7, -c5 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
87
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
88 static inline void mmxext_row_head (int16_t * row, int offset, int16_t * table)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
89 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
90 movq_m2r (*(row+offset), mm2); // mm2 = x6 x4 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
91
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
92 movq_m2r (*(row+offset+4), mm5); // mm5 = x7 x5 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
93 movq_r2r (mm2, mm0); // mm0 = x6 x4 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
94
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
95 movq_m2r (*table, mm3); // mm3 = -C2 -C4 C2 C4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
96 movq_r2r (mm5, mm6); // mm6 = x7 x5 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
97
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
98 movq_m2r (*(table+4), mm4); // mm4 = C6 C4 C6 C4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
99 pmaddwd_r2r (mm0, mm3); // mm3 = -C4*x4-C2*x6 C4*x0+C2*x2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
100
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
101 pshufw_r2r (mm2, mm2, 0x4e); // mm2 = x2 x0 x6 x4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
102 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
103
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
104 static inline void mmxext_row (int16_t * table, int32_t * rounder)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
105 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
106 movq_m2r (*(table+8), mm1); // mm1 = -C5 -C1 C3 C1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
107 pmaddwd_r2r (mm2, mm4); // mm4 = C4*x0+C6*x2 C4*x4+C6*x6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
108
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
109 pmaddwd_m2r (*(table+16), mm0); // mm0 = C4*x4-C6*x6 C4*x0-C6*x2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
110 pshufw_r2r (mm6, mm6, 0x4e); // mm6 = x3 x1 x7 x5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
111
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
112 movq_m2r (*(table+12), mm7); // mm7 = -C7 C3 C7 C5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
113 pmaddwd_r2r (mm5, mm1); // mm1 = -C1*x5-C5*x7 C1*x1+C3*x3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
114
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
115 paddd_m2r (*rounder, mm3); // mm3 += rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
116 pmaddwd_r2r (mm6, mm7); // mm7 = C3*x1-C7*x3 C5*x5+C7*x7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
117
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
118 pmaddwd_m2r (*(table+20), mm2); // mm2 = C4*x0-C2*x2 -C4*x4+C2*x6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
119 paddd_r2r (mm4, mm3); // mm3 = a1 a0 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
120
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
121 pmaddwd_m2r (*(table+24), mm5); // mm5 = C3*x5-C1*x7 C5*x1-C1*x3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
122 movq_r2r (mm3, mm4); // mm4 = a1 a0 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
123
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
124 pmaddwd_m2r (*(table+28), mm6); // mm6 = C7*x1-C5*x3 C7*x5+C3*x7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
125 paddd_r2r (mm7, mm1); // mm1 = b1 b0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
126
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
127 paddd_m2r (*rounder, mm0); // mm0 += rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
128 psubd_r2r (mm1, mm3); // mm3 = a1-b1 a0-b0 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
129
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
130 psrad_i2r (ROW_SHIFT, mm3); // mm3 = y6 y7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
131 paddd_r2r (mm4, mm1); // mm1 = a1+b1 a0+b0 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
132
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
133 paddd_r2r (mm2, mm0); // mm0 = a3 a2 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
134 psrad_i2r (ROW_SHIFT, mm1); // mm1 = y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
135
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
136 paddd_r2r (mm6, mm5); // mm5 = b3 b2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
137 movq_r2r (mm0, mm4); // mm4 = a3 a2 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
138
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
139 paddd_r2r (mm5, mm0); // mm0 = a3+b3 a2+b2 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
140 psubd_r2r (mm5, mm4); // mm4 = a3-b3 a2-b2 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
141 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
142
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
143 static inline void mmxext_row_tail (int16_t * row, int store)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
144 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
145 psrad_i2r (ROW_SHIFT, mm0); // mm0 = y3 y2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
146
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
147 psrad_i2r (ROW_SHIFT, mm4); // mm4 = y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
148
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
149 packssdw_r2r (mm0, mm1); // mm1 = y3 y2 y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
150
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
151 packssdw_r2r (mm3, mm4); // mm4 = y6 y7 y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
152
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
153 movq_r2m (mm1, *(row+store)); // save y3 y2 y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
154 pshufw_r2r (mm4, mm4, 0xb1); // mm4 = y7 y6 y5 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
155
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
156 /* slot */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
157
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
158 movq_r2m (mm4, *(row+store+4)); // save y7 y6 y5 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
159 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
160
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
161 static inline void mmxext_row_mid (int16_t * row, int store,
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
162 int offset, int16_t * table)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
163 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
164 movq_m2r (*(row+offset), mm2); // mm2 = x6 x4 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
165 psrad_i2r (ROW_SHIFT, mm0); // mm0 = y3 y2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
166
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
167 movq_m2r (*(row+offset+4), mm5); // mm5 = x7 x5 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
168 psrad_i2r (ROW_SHIFT, mm4); // mm4 = y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
169
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
170 packssdw_r2r (mm0, mm1); // mm1 = y3 y2 y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
171 movq_r2r (mm5, mm6); // mm6 = x7 x5 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
172
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
173 packssdw_r2r (mm3, mm4); // mm4 = y6 y7 y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
174 movq_r2r (mm2, mm0); // mm0 = x6 x4 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
175
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
176 movq_r2m (mm1, *(row+store)); // save y3 y2 y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
177 pshufw_r2r (mm4, mm4, 0xb1); // mm4 = y7 y6 y5 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
178
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
179 movq_m2r (*table, mm3); // mm3 = -C2 -C4 C2 C4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
180 movq_r2m (mm4, *(row+store+4)); // save y7 y6 y5 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
181
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
182 pmaddwd_r2r (mm0, mm3); // mm3 = -C4*x4-C2*x6 C4*x0+C2*x2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
183
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
184 movq_m2r (*(table+4), mm4); // mm4 = C6 C4 C6 C4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
185 pshufw_r2r (mm2, mm2, 0x4e); // mm2 = x2 x0 x6 x4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
186 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
187
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
188
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
189 /* MMX row IDCT */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
190
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
191 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) { c4, c2, c4, c6, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
192 c4, c6, -c4, -c2, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
193 c1, c3, c3, -c7, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
194 c5, c7, -c1, -c5, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
195 c4, -c6, c4, -c2, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
196 -c4, c2, c4, -c6, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
197 c5, -c1, c7, -c5, \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
198 c7, c3, c3, -c1 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
199
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
200 static inline void mmx_row_head (int16_t * row, int offset, int16_t * table)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
201 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
202 movq_m2r (*(row+offset), mm2); // mm2 = x6 x4 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
203
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
204 movq_m2r (*(row+offset+4), mm5); // mm5 = x7 x5 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
205 movq_r2r (mm2, mm0); // mm0 = x6 x4 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
206
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
207 movq_m2r (*table, mm3); // mm3 = C6 C4 C2 C4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
208 movq_r2r (mm5, mm6); // mm6 = x7 x5 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
209
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
210 punpckldq_r2r (mm0, mm0); // mm0 = x2 x0 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
211
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
212 movq_m2r (*(table+4), mm4); // mm4 = -C2 -C4 C6 C4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
213 pmaddwd_r2r (mm0, mm3); // mm3 = C4*x0+C6*x2 C4*x0+C2*x2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
214
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
215 movq_m2r (*(table+8), mm1); // mm1 = -C7 C3 C3 C1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
216 punpckhdq_r2r (mm2, mm2); // mm2 = x6 x4 x6 x4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
217 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
218
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
219 static inline void mmx_row (int16_t * table, int32_t * rounder)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
220 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
221 pmaddwd_r2r (mm2, mm4); // mm4 = -C4*x4-C2*x6 C4*x4+C6*x6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
222 punpckldq_r2r (mm5, mm5); // mm5 = x3 x1 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
223
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
224 pmaddwd_m2r (*(table+16), mm0); // mm0 = C4*x0-C2*x2 C4*x0-C6*x2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
225 punpckhdq_r2r (mm6, mm6); // mm6 = x7 x5 x7 x5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
226
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
227 movq_m2r (*(table+12), mm7); // mm7 = -C5 -C1 C7 C5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
228 pmaddwd_r2r (mm5, mm1); // mm1 = C3*x1-C7*x3 C1*x1+C3*x3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
229
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
230 paddd_m2r (*rounder, mm3); // mm3 += rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
231 pmaddwd_r2r (mm6, mm7); // mm7 = -C1*x5-C5*x7 C5*x5+C7*x7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
232
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
233 pmaddwd_m2r (*(table+20), mm2); // mm2 = C4*x4-C6*x6 -C4*x4+C2*x6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
234 paddd_r2r (mm4, mm3); // mm3 = a1 a0 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
235
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
236 pmaddwd_m2r (*(table+24), mm5); // mm5 = C7*x1-C5*x3 C5*x1-C1*x3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
237 movq_r2r (mm3, mm4); // mm4 = a1 a0 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
238
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
239 pmaddwd_m2r (*(table+28), mm6); // mm6 = C3*x5-C1*x7 C7*x5+C3*x7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
240 paddd_r2r (mm7, mm1); // mm1 = b1 b0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
241
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
242 paddd_m2r (*rounder, mm0); // mm0 += rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
243 psubd_r2r (mm1, mm3); // mm3 = a1-b1 a0-b0 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
244
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
245 psrad_i2r (ROW_SHIFT, mm3); // mm3 = y6 y7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
246 paddd_r2r (mm4, mm1); // mm1 = a1+b1 a0+b0 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
247
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
248 paddd_r2r (mm2, mm0); // mm0 = a3 a2 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
249 psrad_i2r (ROW_SHIFT, mm1); // mm1 = y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
250
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
251 paddd_r2r (mm6, mm5); // mm5 = b3 b2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
252 movq_r2r (mm0, mm7); // mm7 = a3 a2 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
253
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
254 paddd_r2r (mm5, mm0); // mm0 = a3+b3 a2+b2 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
255 psubd_r2r (mm5, mm7); // mm7 = a3-b3 a2-b2 + rounder
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
256 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
257
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
258 static inline void mmx_row_tail (int16_t * row, int store)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
259 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
260 psrad_i2r (ROW_SHIFT, mm0); // mm0 = y3 y2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
261
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
262 psrad_i2r (ROW_SHIFT, mm7); // mm7 = y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
263
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
264 packssdw_r2r (mm0, mm1); // mm1 = y3 y2 y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
265
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
266 packssdw_r2r (mm3, mm7); // mm7 = y6 y7 y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
267
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
268 movq_r2m (mm1, *(row+store)); // save y3 y2 y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
269 movq_r2r (mm7, mm4); // mm4 = y6 y7 y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
270
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
271 pslld_i2r (16, mm7); // mm7 = y7 0 y5 0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
272
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
273 psrld_i2r (16, mm4); // mm4 = 0 y6 0 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
274
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
275 por_r2r (mm4, mm7); // mm7 = y7 y6 y5 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
276
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
277 /* slot */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
278
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
279 movq_r2m (mm7, *(row+store+4)); // save y7 y6 y5 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
280 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
281
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
282 static inline void mmx_row_mid (int16_t * row, int store,
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
283 int offset, int16_t * table)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
284 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
285 movq_m2r (*(row+offset), mm2); // mm2 = x6 x4 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
286 psrad_i2r (ROW_SHIFT, mm0); // mm0 = y3 y2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
287
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
288 movq_m2r (*(row+offset+4), mm5); // mm5 = x7 x5 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
289 psrad_i2r (ROW_SHIFT, mm7); // mm7 = y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
290
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
291 packssdw_r2r (mm0, mm1); // mm1 = y3 y2 y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
292 movq_r2r (mm5, mm6); // mm6 = x7 x5 x3 x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
293
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
294 packssdw_r2r (mm3, mm7); // mm7 = y6 y7 y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
295 movq_r2r (mm2, mm0); // mm0 = x6 x4 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
296
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
297 movq_r2m (mm1, *(row+store)); // save y3 y2 y1 y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
298 movq_r2r (mm7, mm1); // mm1 = y6 y7 y4 y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
299
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
300 punpckldq_r2r (mm0, mm0); // mm0 = x2 x0 x2 x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
301 psrld_i2r (16, mm7); // mm7 = 0 y6 0 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
302
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
303 movq_m2r (*table, mm3); // mm3 = C6 C4 C2 C4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
304 pslld_i2r (16, mm1); // mm1 = y7 0 y5 0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
305
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
306 movq_m2r (*(table+4), mm4); // mm4 = -C2 -C4 C6 C4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
307 por_r2r (mm1, mm7); // mm7 = y7 y6 y5 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
308
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
309 movq_m2r (*(table+8), mm1); // mm1 = -C7 C3 C3 C1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
310 punpckhdq_r2r (mm2, mm2); // mm2 = x6 x4 x6 x4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
311
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
312 movq_r2m (mm7, *(row+store+4)); // save y7 y6 y5 y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
313 pmaddwd_r2r (mm0, mm3); // mm3 = C4*x0+C6*x2 C4*x0+C2*x2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
314 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
315
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
316
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
317 #if 0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
318 // C column IDCT - its just here to document the MMXEXT and MMX versions
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
319 static inline void idct_col (int16_t * col, int offset)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
320 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
321 /* multiplication - as implemented on mmx */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
322 #define F(c,x) (((c) * (x)) >> 16)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
323
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
324 /* saturation - it helps us handle torture test cases */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
325 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
326
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
327 int16_t x0, x1, x2, x3, x4, x5, x6, x7;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
328 int16_t y0, y1, y2, y3, y4, y5, y6, y7;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
329 int16_t a0, a1, a2, a3, b0, b1, b2, b3;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
330 int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
331
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
332 col += offset;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
333
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
334 x0 = col[0*8];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
335 x1 = col[1*8];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
336 x2 = col[2*8];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
337 x3 = col[3*8];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
338 x4 = col[4*8];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
339 x5 = col[5*8];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
340 x6 = col[6*8];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
341 x7 = col[7*8];
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
342
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
343 u04 = S (x0 + x4);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
344 v04 = S (x0 - x4);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
345 u26 = S (F (T2, x6) + x2);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
346 v26 = S (F (T2, x2) - x6);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
347
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
348 a0 = S (u04 + u26);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
349 a1 = S (v04 + v26);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
350 a2 = S (v04 - v26);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
351 a3 = S (u04 - u26);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
352
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
353 u17 = S (F (T1, x7) + x1);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
354 v17 = S (F (T1, x1) - x7);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
355 u35 = S (F (T3, x5) + x3);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
356 v35 = S (F (T3, x3) - x5);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
357
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
358 b0 = S (u17 + u35);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
359 b3 = S (v17 - v35);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
360 u12 = S (u17 - u35);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
361 v12 = S (v17 + v35);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
362 u12 = S (2 * F (C4, u12));
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
363 v12 = S (2 * F (C4, v12));
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
364 b1 = S (u12 + v12);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
365 b2 = S (u12 - v12);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
366
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
367 y0 = S (a0 + b0) >> COL_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
368 y1 = S (a1 + b1) >> COL_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
369 y2 = S (a2 + b2) >> COL_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
370 y3 = S (a3 + b3) >> COL_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
371
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
372 y4 = S (a3 - b3) >> COL_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
373 y5 = S (a2 - b2) >> COL_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
374 y6 = S (a1 - b1) >> COL_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
375 y7 = S (a0 - b0) >> COL_SHIFT;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
376
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
377 col[0*8] = y0;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
378 col[1*8] = y1;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
379 col[2*8] = y2;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
380 col[3*8] = y3;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
381 col[4*8] = y4;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
382 col[5*8] = y5;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
383 col[6*8] = y6;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
384 col[7*8] = y7;
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
385 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
386 #endif
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
387
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
388
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
389 // MMX column IDCT
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
390 static inline void idct_col (int16_t * col, int offset)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
391 {
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
392 #define T1 13036
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
393 #define T2 27146
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
394 #define T3 43790
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
395 #define C4 23170
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
396
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
397 static short _T1[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
398 static short _T2[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
399 static short _T3[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
400 static short _C4[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
401
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
402 /* column code adapted from peter gubanov */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
403 /* http://www.elecard.com/peter/idct.shtml */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
404
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
405 movq_m2r (*_T1, mm0); // mm0 = T1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
406
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
407 movq_m2r (*(col+offset+1*8), mm1); // mm1 = x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
408 movq_r2r (mm0, mm2); // mm2 = T1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
409
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
410 movq_m2r (*(col+offset+7*8), mm4); // mm4 = x7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
411 pmulhw_r2r (mm1, mm0); // mm0 = T1*x1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
412
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
413 movq_m2r (*_T3, mm5); // mm5 = T3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
414 pmulhw_r2r (mm4, mm2); // mm2 = T1*x7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
415
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
416 movq_m2r (*(col+offset+5*8), mm6); // mm6 = x5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
417 movq_r2r (mm5, mm7); // mm7 = T3-1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
418
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
419 movq_m2r (*(col+offset+3*8), mm3); // mm3 = x3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
420 psubsw_r2r (mm4, mm0); // mm0 = v17
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
421
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
422 movq_m2r (*_T2, mm4); // mm4 = T2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
423 pmulhw_r2r (mm3, mm5); // mm5 = (T3-1)*x3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
424
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
425 paddsw_r2r (mm2, mm1); // mm1 = u17
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
426 pmulhw_r2r (mm6, mm7); // mm7 = (T3-1)*x5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
427
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
428 /* slot */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
429
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
430 movq_r2r (mm4, mm2); // mm2 = T2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
431 paddsw_r2r (mm3, mm5); // mm5 = T3*x3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
432
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
433 pmulhw_m2r (*(col+offset+2*8), mm4);// mm4 = T2*x2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
434 paddsw_r2r (mm6, mm7); // mm7 = T3*x5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
435
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
436 psubsw_r2r (mm6, mm5); // mm5 = v35
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
437 paddsw_r2r (mm3, mm7); // mm7 = u35
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
438
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
439 movq_m2r (*(col+offset+6*8), mm3); // mm3 = x6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
440 movq_r2r (mm0, mm6); // mm6 = v17
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
441
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
442 pmulhw_r2r (mm3, mm2); // mm2 = T2*x6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
443 psubsw_r2r (mm5, mm0); // mm0 = b3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
444
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
445 psubsw_r2r (mm3, mm4); // mm4 = v26
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
446 paddsw_r2r (mm6, mm5); // mm5 = v12
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
447
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
448 movq_r2m (mm0, *(col+offset+3*8)); // save b3 in scratch0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
449 movq_r2r (mm1, mm6); // mm6 = u17
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
450
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
451 paddsw_m2r (*(col+offset+2*8), mm2);// mm2 = u26
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
452 paddsw_r2r (mm7, mm6); // mm6 = b0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
453
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
454 psubsw_r2r (mm7, mm1); // mm1 = u12
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
455 movq_r2r (mm1, mm7); // mm7 = u12
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
456
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
457 movq_m2r (*(col+offset+0*8), mm3); // mm3 = x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
458 paddsw_r2r (mm5, mm1); // mm1 = u12+v12
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
459
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
460 movq_m2r (*_C4, mm0); // mm0 = C4/2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
461 psubsw_r2r (mm5, mm7); // mm7 = u12-v12
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
462
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
463 movq_r2m (mm6, *(col+offset+5*8)); // save b0 in scratch1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
464 pmulhw_r2r (mm0, mm1); // mm1 = b1/2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
465
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
466 movq_r2r (mm4, mm6); // mm6 = v26
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
467 pmulhw_r2r (mm0, mm7); // mm7 = b2/2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
468
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
469 movq_m2r (*(col+offset+4*8), mm5); // mm5 = x4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
470 movq_r2r (mm3, mm0); // mm0 = x0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
471
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
472 psubsw_r2r (mm5, mm3); // mm3 = v04
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
473 paddsw_r2r (mm5, mm0); // mm0 = u04
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
474
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
475 paddsw_r2r (mm3, mm4); // mm4 = a1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
476 movq_r2r (mm0, mm5); // mm5 = u04
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
477
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
478 psubsw_r2r (mm6, mm3); // mm3 = a2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
479 paddsw_r2r (mm2, mm5); // mm5 = a0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
480
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
481 paddsw_r2r (mm1, mm1); // mm1 = b1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
482 psubsw_r2r (mm2, mm0); // mm0 = a3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
483
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
484 paddsw_r2r (mm7, mm7); // mm7 = b2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
485 movq_r2r (mm3, mm2); // mm2 = a2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
486
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
487 movq_r2r (mm4, mm6); // mm6 = a1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
488 paddsw_r2r (mm7, mm3); // mm3 = a2+b2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
489
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
490 psraw_i2r (COL_SHIFT, mm3); // mm3 = y2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
491 paddsw_r2r (mm1, mm4); // mm4 = a1+b1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
492
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
493 psraw_i2r (COL_SHIFT, mm4); // mm4 = y1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
494 psubsw_r2r (mm1, mm6); // mm6 = a1-b1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
495
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
496 movq_m2r (*(col+offset+5*8), mm1); // mm1 = b0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
497 psubsw_r2r (mm7, mm2); // mm2 = a2-b2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
498
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
499 psraw_i2r (COL_SHIFT, mm6); // mm6 = y6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
500 movq_r2r (mm5, mm7); // mm7 = a0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
501
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
502 movq_r2m (mm4, *(col+offset+1*8)); // save y1
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
503 psraw_i2r (COL_SHIFT, mm2); // mm2 = y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
504
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
505 movq_r2m (mm3, *(col+offset+2*8)); // save y2
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
506 paddsw_r2r (mm1, mm5); // mm5 = a0+b0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
507
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
508 movq_m2r (*(col+offset+3*8), mm4); // mm4 = b3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
509 psubsw_r2r (mm1, mm7); // mm7 = a0-b0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
510
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
511 psraw_i2r (COL_SHIFT, mm5); // mm5 = y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
512 movq_r2r (mm0, mm3); // mm3 = a3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
513
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
514 movq_r2m (mm2, *(col+offset+5*8)); // save y5
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
515 psubsw_r2r (mm4, mm3); // mm3 = a3-b3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
516
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
517 psraw_i2r (COL_SHIFT, mm7); // mm7 = y7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
518 paddsw_r2r (mm0, mm4); // mm4 = a3+b3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
519
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
520 movq_r2m (mm5, *(col+offset+0*8)); // save y0
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
521 psraw_i2r (COL_SHIFT, mm3); // mm3 = y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
522
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
523 movq_r2m (mm6, *(col+offset+6*8)); // save y6
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
524 psraw_i2r (COL_SHIFT, mm4); // mm4 = y3
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
525
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
526 movq_r2m (mm7, *(col+offset+7*8)); // save y7
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
527
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
528 movq_r2m (mm3, *(col+offset+4*8)); // save y4
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
529
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
530 movq_r2m (mm4, *(col+offset+3*8)); // save y3
436
35de17dd6ed8 * undefine local defines when they are no longer needed
kabi
parents: 76
diff changeset
531
35de17dd6ed8 * undefine local defines when they are no longer needed
kabi
parents: 76
diff changeset
532 #undef T1
35de17dd6ed8 * undefine local defines when they are no longer needed
kabi
parents: 76
diff changeset
533 #undef T2
35de17dd6ed8 * undefine local defines when they are no longer needed
kabi
parents: 76
diff changeset
534 #undef T3
35de17dd6ed8 * undefine local defines when they are no longer needed
kabi
parents: 76
diff changeset
535 #undef C4
30
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
536 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
537
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
538 static int32_t rounder0[] ATTR_ALIGN(8) =
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
539 rounder ((1 << (COL_SHIFT - 1)) - 0.5);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
540 static int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
541 static int32_t rounder1[] ATTR_ALIGN(8) =
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
542 rounder (1.25683487303); /* C1*(C1/C4+C1+C7)/2 */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
543 static int32_t rounder7[] ATTR_ALIGN(8) =
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
544 rounder (-0.25); /* C1*(C7/C4+C7-C1)/2 */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
545 static int32_t rounder2[] ATTR_ALIGN(8) =
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
546 rounder (0.60355339059); /* C2 * (C6+C2)/2 */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
547 static int32_t rounder6[] ATTR_ALIGN(8) =
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
548 rounder (-0.25); /* C2 * (C6-C2)/2 */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
549 static int32_t rounder3[] ATTR_ALIGN(8) =
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
550 rounder (0.087788325588); /* C3*(-C3/C4+C3+C5)/2 */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
551 static int32_t rounder5[] ATTR_ALIGN(8) =
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
552 rounder (-0.441341716183); /* C3*(-C5/C4+C5-C3)/2 */
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
553
436
35de17dd6ed8 * undefine local defines when they are no longer needed
kabi
parents: 76
diff changeset
554 #undef COL_SHIFT
35de17dd6ed8 * undefine local defines when they are no longer needed
kabi
parents: 76
diff changeset
555 #undef ROW_SHIFT
30
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
556
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
557 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
558 void idct (int16_t * block) \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
559 { \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
560 static int16_t table04[] ATTR_ALIGN(16) = \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
561 table (22725, 21407, 19266, 16384, 12873, 8867, 4520); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
562 static int16_t table17[] ATTR_ALIGN(16) = \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
563 table (31521, 29692, 26722, 22725, 17855, 12299, 6270); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
564 static int16_t table26[] ATTR_ALIGN(16) = \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
565 table (29692, 27969, 25172, 21407, 16819, 11585, 5906); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
566 static int16_t table35[] ATTR_ALIGN(16) = \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
567 table (26722, 25172, 22654, 19266, 15137, 10426, 5315); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
568 \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
569 idct_row_head (block, 0*8, table04); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
570 idct_row (table04, rounder0); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
571 idct_row_mid (block, 0*8, 4*8, table04); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
572 idct_row (table04, rounder4); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
573 idct_row_mid (block, 4*8, 1*8, table17); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
574 idct_row (table17, rounder1); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
575 idct_row_mid (block, 1*8, 7*8, table17); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
576 idct_row (table17, rounder7); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
577 idct_row_mid (block, 7*8, 2*8, table26); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
578 idct_row (table26, rounder2); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
579 idct_row_mid (block, 2*8, 6*8, table26); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
580 idct_row (table26, rounder6); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
581 idct_row_mid (block, 6*8, 3*8, table35); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
582 idct_row (table35, rounder3); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
583 idct_row_mid (block, 3*8, 5*8, table35); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
584 idct_row (table35, rounder5); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
585 idct_row_tail (block, 5*8); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
586 \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
587 idct_col (block, 0); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
588 idct_col (block, 4); \
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
589 }
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
590
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
591
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
592 declare_idct (ff_mmxext_idct, mmxext_table,
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
593 mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
594
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
595 declare_idct (ff_mmx_idct, mmx_table,
b8a290072f26 added idct mmx code
glantau
parents:
diff changeset
596 mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)