annotate i386/simple_idct_mmx.c @ 2189:70b27300a496 libavcodec

quad tree based motion compensation (currently only 16x16 & 8x8 OBMC blocks, but can be extended to other block sizes easily) different intra block coding (previous was just an ugly hack) 1.8% bitrate reduction -0.01PSNR (foreman@352x288 qscale=8) 1.5% bitrate reduction +0.05PSNR (foreman@352x288 qscale=1)
author michael
date Thu, 26 Aug 2004 20:04:54 +0000
parents f65d87bfdd5a
children ef2149182f1c
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1 /*
429
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
2 * Simple IDCT MMX
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
3 *
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
4 * Copyright (c) 2001, 2002 Michael Niedermayer <michaelni@gmx.at>
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
5 *
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
6 * This library is free software; you can redistribute it and/or
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
8 * License as published by the Free Software Foundation; either
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
9 * version 2 of the License, or (at your option) any later version.
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
10 *
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
11 * This library is distributed in the hope that it will be useful,
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
14 * Lesser General Public License for more details.
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
15 *
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
17 * License along with this library; if not, write to the Free Software
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
19 */
718a22dc121f license/copyright change
glantau
parents: 352
diff changeset
20 #include "../dsputil.h"
2024
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1845
diff changeset
21 #include "../simple_idct.h"
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
22
351
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
23 /*
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
24 23170.475006
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
25 22725.260826
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
26 21406.727617
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
27 19265.545870
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
28 16384.000000
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
29 12872.826198
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
30 8866.956905
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
31 4520.335430
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
32 */
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
33 #define C0 23170 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
34 #define C1 22725 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
35 #define C2 21407 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
36 #define C3 19266 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
351
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
37 #if 0
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
38 #define C4 16384 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
351
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
39 #else
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
40 #define C4 16383 //cos(i*M_PI/16)*sqrt(2)*(1<<14) - 0.5
6cef8253faab set C4 to 16383
michaelni
parents: 350
diff changeset
41 #endif
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
42 #define C5 12873 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
43 #define C6 8867 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
44 #define C7 4520 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
45
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
46 #define ROW_SHIFT 11
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
47 #define COL_SHIFT 20 // 6
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
48
1845
3054613980a8 attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents: 1469
diff changeset
49 static const uint64_t attribute_used __attribute__((aligned(8))) wm1010= 0xFFFF0000FFFF0000ULL;
3054613980a8 attribute used patch by (mitya at school dot ioffe dot ru (Dmitry Baryshkov))
michael
parents: 1469
diff changeset
50 static const uint64_t attribute_used __attribute__((aligned(8))) d40000= 0x0000000000040000ULL;
1469
43ceb6e34b06 another non const static, maybe its thread save now
michaelni
parents: 1064
diff changeset
51
43ceb6e34b06 another non const static, maybe its thread save now
michaelni
parents: 1064
diff changeset
52 static const int16_t __attribute__((aligned(8))) coeffs[]= {
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
53 1<<(ROW_SHIFT-1), 0, 1<<(ROW_SHIFT-1), 0,
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
54 // 1<<(COL_SHIFT-1), 0, 1<<(COL_SHIFT-1), 0,
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
55 // 0, 1<<(COL_SHIFT-1-16), 0, 1<<(COL_SHIFT-1-16),
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
56 1<<(ROW_SHIFT-1), 1, 1<<(ROW_SHIFT-1), 0,
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
57 // the 1 = ((1<<(COL_SHIFT-1))/C4)<<ROW_SHIFT :)
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
58 // 0, 0, 0, 0,
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
59 // 0, 0, 0, 0,
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
60
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
61 C4, C4, C4, C4,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
62 C4, -C4, C4, -C4,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
63
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
64 C2, C6, C2, C6,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
65 C6, -C2, C6, -C2,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
66
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
67 C1, C3, C1, C3,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
68 C5, C7, C5, C7,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
69
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
70 C3, -C7, C3, -C7,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
71 -C1, -C5, -C1, -C5,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
72
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
73 C5, -C1, C5, -C1,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
74 C7, C3, C7, C3,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
75
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
76 C7, -C5, C7, -C5,
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
77 C3, -C1, C3, -C1
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
78 };
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
79
350
6ebbecc10063 - Advanced Intra Coding (AIC) support for H.263+ encoder, just DC by now.
pulento
parents: 213
diff changeset
80 #if 0
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
81 static void unused_var_killer(){
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
82 int a= wm1010 + d40000;
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
83 temp[0]=a;
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
84 }
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
85
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
86 static void inline idctCol (int16_t * col, int16_t *input)
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
87 {
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
88 #undef C0
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
89 #undef C1
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
90 #undef C2
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
91 #undef C3
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
92 #undef C4
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
93 #undef C5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
94 #undef C6
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
95 #undef C7
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
96 int a0, a1, a2, a3, b0, b1, b2, b3;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
97 const int C0 = 23170; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
98 const int C1 = 22725; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
99 const int C2 = 21407; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
100 const int C3 = 19266; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
352
5a8eb5cf9f92 C4=16383 for the c version too and even for some outcommented code
michaelni
parents: 351
diff changeset
101 const int C4 = 16383; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
102 const int C5 = 12873; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
103 const int C6 = 8867; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
104 const int C7 = 4520; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
105 /*
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
106 if( !(col[8*1] | col[8*2] |col[8*3] |col[8*4] |col[8*5] |col[8*6] | col[8*7])) {
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
107 col[8*0] = col[8*1] = col[8*2] = col[8*3] = col[8*4] =
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
108 col[8*5] = col[8*6] = col[8*7] = col[8*0]<<3;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
109 return;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
110 }*/
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
111
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
112 col[8*0] = input[8*0 + 0];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
113 col[8*1] = input[8*2 + 0];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
114 col[8*2] = input[8*0 + 1];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
115 col[8*3] = input[8*2 + 1];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
116 col[8*4] = input[8*4 + 0];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
117 col[8*5] = input[8*6 + 0];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
118 col[8*6] = input[8*4 + 1];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
119 col[8*7] = input[8*6 + 1];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
120
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
121 a0 = C4*col[8*0] + C2*col[8*2] + C4*col[8*4] + C6*col[8*6] + (1<<(COL_SHIFT-1));
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
122 a1 = C4*col[8*0] + C6*col[8*2] - C4*col[8*4] - C2*col[8*6] + (1<<(COL_SHIFT-1));
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
123 a2 = C4*col[8*0] - C6*col[8*2] - C4*col[8*4] + C2*col[8*6] + (1<<(COL_SHIFT-1));
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
124 a3 = C4*col[8*0] - C2*col[8*2] + C4*col[8*4] - C6*col[8*6] + (1<<(COL_SHIFT-1));
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
125
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
126 b0 = C1*col[8*1] + C3*col[8*3] + C5*col[8*5] + C7*col[8*7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
127 b1 = C3*col[8*1] - C7*col[8*3] - C1*col[8*5] - C5*col[8*7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
128 b2 = C5*col[8*1] - C1*col[8*3] + C7*col[8*5] + C3*col[8*7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
129 b3 = C7*col[8*1] - C5*col[8*3] + C3*col[8*5] - C1*col[8*7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
130
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
131 col[8*0] = (a0 + b0) >> COL_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
132 col[8*1] = (a1 + b1) >> COL_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
133 col[8*2] = (a2 + b2) >> COL_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
134 col[8*3] = (a3 + b3) >> COL_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
135 col[8*4] = (a3 - b3) >> COL_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
136 col[8*5] = (a2 - b2) >> COL_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
137 col[8*6] = (a1 - b1) >> COL_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
138 col[8*7] = (a0 - b0) >> COL_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
139 }
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
140
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
141 static void inline idctRow (int16_t * output, int16_t * input)
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
142 {
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
143 int16_t row[8];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
144
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
145 int a0, a1, a2, a3, b0, b1, b2, b3;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
146 const int C0 = 23170; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
147 const int C1 = 22725; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
148 const int C2 = 21407; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
149 const int C3 = 19266; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
352
5a8eb5cf9f92 C4=16383 for the c version too and even for some outcommented code
michaelni
parents: 351
diff changeset
150 const int C4 = 16383; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
151 const int C5 = 12873; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
152 const int C6 = 8867; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
153 const int C7 = 4520; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
154
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
155 row[0] = input[0];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
156 row[2] = input[1];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
157 row[4] = input[4];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
158 row[6] = input[5];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
159 row[1] = input[8];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
160 row[3] = input[9];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
161 row[5] = input[12];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
162 row[7] = input[13];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
163
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
164 if( !(row[1] | row[2] |row[3] |row[4] |row[5] |row[6] | row[7]) ) {
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
165 row[0] = row[1] = row[2] = row[3] = row[4] =
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
166 row[5] = row[6] = row[7] = row[0]<<3;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
167 output[0] = row[0];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
168 output[2] = row[1];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
169 output[4] = row[2];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
170 output[6] = row[3];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
171 output[8] = row[4];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
172 output[10] = row[5];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
173 output[12] = row[6];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
174 output[14] = row[7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
175 return;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
176 }
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
177
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
178 a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + (1<<(ROW_SHIFT-1));
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
179 a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + (1<<(ROW_SHIFT-1));
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
180 a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + (1<<(ROW_SHIFT-1));
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
181 a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + (1<<(ROW_SHIFT-1));
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
182
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
183 b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
184 b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
185 b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
186 b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
187
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
188 row[0] = (a0 + b0) >> ROW_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
189 row[1] = (a1 + b1) >> ROW_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
190 row[2] = (a2 + b2) >> ROW_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
191 row[3] = (a3 + b3) >> ROW_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
192 row[4] = (a3 - b3) >> ROW_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
193 row[5] = (a2 - b2) >> ROW_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
194 row[6] = (a1 - b1) >> ROW_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
195 row[7] = (a0 - b0) >> ROW_SHIFT;
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
196
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
197 output[0] = row[0];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
198 output[2] = row[1];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
199 output[4] = row[2];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
200 output[6] = row[3];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
201 output[8] = row[4];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
202 output[10] = row[5];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
203 output[12] = row[6];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
204 output[14] = row[7];
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
205 }
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
206 #endif
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
207
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
208 static inline void idct(int16_t *block)
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
209 {
1469
43ceb6e34b06 another non const static, maybe its thread save now
michaelni
parents: 1064
diff changeset
210 int64_t __attribute__((aligned(8))) align_tmp[16];
43ceb6e34b06 another non const static, maybe its thread save now
michaelni
parents: 1064
diff changeset
211 int16_t * const temp= (int16_t*)align_tmp;
43ceb6e34b06 another non const static, maybe its thread save now
michaelni
parents: 1064
diff changeset
212
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
213 asm volatile(
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
214 #if 0 //Alternative, simpler variant
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
215
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
216 #define ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
217 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
218 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
219 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
220 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
221 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
222 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
223 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
224 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
225 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
226 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
227 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
228 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
229 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
230 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
231 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
232 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
233 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
234 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
235 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
236 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
237 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
238 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
239 "paddd %%mm0, %%mm0 \n\t" \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
240 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
241 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
242 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
243 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
244 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
245 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
246 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
247 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
248 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
249 "psrad $" #shift ", %%mm7 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
250 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
251 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
252 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
253 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
254 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
255 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
256 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
257 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
258 "movq %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
259 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
260 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
261 "movq %%mm2, 24+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
262 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
263 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
264 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
265 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
266 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
267 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
268 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
269 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
270 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
271 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
272 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
273 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
274 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
275 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
276 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
277 "psrad $" #shift ", %%mm6 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
278 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
279 "movq %%mm2, 8+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
280 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
281 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
282 "movq %%mm4, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
283
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
284 #define COL_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
285 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
286 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
287 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
288 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
289 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
290 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
291 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
292 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
293 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
294 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
295 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
296 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
297 #rounder ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
298 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
299 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
300 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
301 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
302 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
303 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
304 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
305 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
306 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
307 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
308 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
309 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
310 "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
311 "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
312 "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
313 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
314 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
315 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
316 "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
317 "psrad $" #shift ", %%mm7 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
318 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
319 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
320 "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
321 "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
322 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
323 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
324 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
325 "movd %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
326 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
327 "movd %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
328 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
329 "movd %%mm2, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
330 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
331 "movd %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
332 "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
333 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
334 "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
335 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
336 "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
337 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
338 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
339 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
340 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
341 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
342 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
343 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
344 "psrad $" #shift ", %%mm5 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
345 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
346 "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
347 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
348 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
349 "psrad $" #shift ", %%mm6 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
350 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
351 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
352 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
353 "movd %%mm2, 32+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
354 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
355 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
356 "movd %%mm6, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
357 "movd %%mm4, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
358 "movd %%mm5, 80+" #dst " \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
359
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
360
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
361 #define DC_COND_ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
362 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
363 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
364 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
365 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
213
e80ad397d30e Cygwin's mangling by Felix Buenemann <atmosfear@users.sourceforge.net>
nickols_k
parents: 209
diff changeset
366 "movq "MANGLE(wm1010)", %%mm4 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
367 "pand %%mm0, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
368 "por %%mm1, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
369 "por %%mm2, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
370 "por %%mm3, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
371 "packssdw %%mm4,%%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
372 "movd %%mm4, %%eax \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
373 "orl %%eax, %%eax \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
374 "jz 1f \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
375 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
376 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
377 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
378 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
379 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
380 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
381 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
382 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
383 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
384 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
385 #rounder ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
386 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
387 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
388 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
389 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
390 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
391 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
392 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
393 "paddd %%mm0, %%mm0 \n\t" \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
394 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
395 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
396 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
397 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
398 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
399 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
400 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
401 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
402 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
403 "psrad $" #shift ", %%mm7 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
404 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
405 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
406 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
407 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
408 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
409 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
410 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
411 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
412 "movq %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
413 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
414 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
415 "movq %%mm2, 24+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
416 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
417 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
418 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
419 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
420 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
421 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
422 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
423 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
424 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
425 "psrad $" #shift ", %%mm2 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
426 "psrad $" #shift ", %%mm0 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
427 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
428 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
429 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
430 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
431 "psrad $" #shift ", %%mm6 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
432 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
433 "movq %%mm2, 8+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
434 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
435 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
436 "movq %%mm4, 16+" #dst " \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
437 "jmp 2f \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
438 "1: \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
439 "pslld $16, %%mm0 \n\t"\
213
e80ad397d30e Cygwin's mangling by Felix Buenemann <atmosfear@users.sourceforge.net>
nickols_k
parents: 209
diff changeset
440 "#paddd "MANGLE(d40000)", %%mm0 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
441 "psrad $13, %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
442 "packssdw %%mm0, %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
443 "movq %%mm0, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
444 "movq %%mm0, 8+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
445 "movq %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
446 "movq %%mm0, 24+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
447 "2: \n\t"
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
448
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
449
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
450 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
451 ROW_IDCT( (%0), 8(%0), 16(%0), 24(%0), 0(%1),paddd 8(%2), 11)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
452 /*ROW_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1), paddd (%2), 11)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
453 ROW_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1), paddd (%2), 11)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
454 ROW_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1), paddd (%2), 11)*/
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
455
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
456 DC_COND_ROW_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1),paddd (%2), 11)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
457 DC_COND_ROW_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
458 DC_COND_ROW_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
459
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
460
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
461 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
462 COL_IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
463 COL_IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
464 COL_IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
465 COL_IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
466
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
467 #else
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
468
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
469 #define DC_COND_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
470 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
471 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
472 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
473 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
213
e80ad397d30e Cygwin's mangling by Felix Buenemann <atmosfear@users.sourceforge.net>
nickols_k
parents: 209
diff changeset
474 "movq "MANGLE(wm1010)", %%mm4 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
475 "pand %%mm0, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
476 "por %%mm1, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
477 "por %%mm2, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
478 "por %%mm3, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
479 "packssdw %%mm4,%%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
480 "movd %%mm4, %%eax \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
481 "orl %%eax, %%eax \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
482 "jz 1f \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
483 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
484 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
485 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
486 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
487 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
488 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
489 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
490 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
491 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
492 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
493 #rounder ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
494 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
495 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
496 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
497 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
498 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
499 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
500 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
501 "paddd %%mm0, %%mm0 \n\t" \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
502 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
503 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
504 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
505 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
506 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
507 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
508 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
509 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
510 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
511 "psrad $" #shift ", %%mm7 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
512 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
513 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
514 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
515 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
516 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
517 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
518 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
519 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
520 "movq %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
521 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
522 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
523 "movq %%mm2, 24+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
524 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
525 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
526 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
527 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
528 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
529 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
530 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
531 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
532 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
533 "psrad $" #shift ", %%mm2 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
534 "psrad $" #shift ", %%mm0 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
535 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
536 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
537 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
538 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
539 "psrad $" #shift ", %%mm6 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
540 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
541 "movq %%mm2, 8+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
542 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
543 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
544 "movq %%mm4, 16+" #dst " \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
545 "jmp 2f \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
546 "1: \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
547 "pslld $16, %%mm0 \n\t"\
213
e80ad397d30e Cygwin's mangling by Felix Buenemann <atmosfear@users.sourceforge.net>
nickols_k
parents: 209
diff changeset
548 "paddd "MANGLE(d40000)", %%mm0 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
549 "psrad $13, %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
550 "packssdw %%mm0, %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
551 "movq %%mm0, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
552 "movq %%mm0, 8+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
553 "movq %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
554 "movq %%mm0, 24+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
555 "2: \n\t"
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
556
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
557 #define Z_COND_IDCT(src0, src4, src1, src5, dst, rounder, shift, bt) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
558 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
559 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
560 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
561 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
562 "movq %%mm0, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
563 "por %%mm1, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
564 "por %%mm2, %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
565 "por %%mm3, %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
566 "packssdw %%mm4,%%mm4 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
567 "movd %%mm4, %%eax \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
568 "orl %%eax, %%eax \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
569 "jz " #bt " \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
570 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
571 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
572 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
573 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
574 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
575 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
576 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
577 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
578 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
579 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
580 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
581 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
582 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
583 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
584 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
585 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
586 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
587 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
588 "paddd %%mm0, %%mm0 \n\t" \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
589 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
590 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
591 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
592 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
593 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
594 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
595 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
596 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
597 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
598 "psrad $" #shift ", %%mm7 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
599 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
600 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
601 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
602 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
603 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
604 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
605 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
606 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
607 "movq %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
608 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
609 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
610 "movq %%mm2, 24+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
611 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
612 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
613 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
614 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
615 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
616 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
617 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
618 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
619 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
620 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
621 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
622 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
623 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
624 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
625 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
626 "psrad $" #shift ", %%mm6 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
627 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
628 "movq %%mm2, 8+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
629 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
630 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
631 "movq %%mm4, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
632
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
633 #define ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
634 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
635 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
636 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
637 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
638 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
639 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
640 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
641 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
642 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
643 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
644 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
645 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
646 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
647 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
648 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
649 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
650 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
651 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
652 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
653 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
654 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
655 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
656 "paddd %%mm0, %%mm0 \n\t" \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
657 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
658 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
659 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
660 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
661 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
662 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
663 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
664 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
665 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
666 "psrad $" #shift ", %%mm7 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
667 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
668 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
669 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
670 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
671 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
672 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
673 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
674 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
675 "movq %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
676 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
677 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
678 "movq %%mm2, 24+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
679 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
680 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
681 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
682 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
683 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
684 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
685 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
686 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
687 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
688 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
689 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
690 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
691 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
692 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
693 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
694 "psrad $" #shift ", %%mm6 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
695 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
696 "movq %%mm2, 8+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
697 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
698 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
699 "movq %%mm4, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
700
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
701 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
702 DC_COND_IDCT( 0(%0), 8(%0), 16(%0), 24(%0), 0(%1),paddd 8(%2), 11)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
703 Z_COND_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1),paddd (%2), 11, 4f)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
704 Z_COND_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11, 2f)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
705 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 1f)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
706
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
707 #undef IDCT
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
708 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
709 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
710 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
711 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
712 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
713 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
714 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
715 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
716 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
717 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
718 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
719 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
720 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
721 #rounder ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
722 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
723 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
724 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
725 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
726 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
727 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
728 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
729 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
730 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
731 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
732 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
733 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
734 "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
735 "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
736 "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
737 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
738 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
739 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
740 "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
741 "psrad $" #shift ", %%mm7 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
742 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
743 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
744 "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
745 "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
746 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
747 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
748 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
749 "movd %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
750 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
751 "movd %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
752 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
753 "movd %%mm2, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
754 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
755 "movd %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
756 "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
757 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
758 "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
759 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
760 "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
761 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
762 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
763 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
764 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
765 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
766 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
767 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
768 "psrad $" #shift ", %%mm5 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
769 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
770 "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
771 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
772 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
773 "psrad $" #shift ", %%mm6 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
774 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
775 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
776 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
777 "movd %%mm2, 32+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
778 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
779 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
780 "movd %%mm6, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
781 "movd %%mm4, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
782 "movd %%mm5, 80+" #dst " \n\t"
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
783
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
784
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
785 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
786 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
787 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
788 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
789 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
790 "jmp 9f \n\t"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
791
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
792 "#.balign 16 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
793 "4: \n\t"
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
794 Z_COND_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11, 6f)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
795 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 5f)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
796
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
797 #undef IDCT
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
798 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
799 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
800 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
801 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
802 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
803 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
804 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
805 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
806 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
807 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
808 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
809 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
810 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
811 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
812 #rounder ", %%mm0 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
813 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
814 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
815 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
816 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
817 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
818 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
819 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
820 "movq 72(%2), %%mm7 \n\t" /* -C5 -C1 -C5 -C1 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
821 "pmaddwd %%mm3, %%mm7 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
822 "paddd %%mm4, %%mm1 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
823 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
824 "psubd %%mm1, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
825 "psrad $" #shift ", %%mm1 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
826 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
827 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
828 "paddd %%mm7, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
829 "psubd %%mm7, %%mm2 \n\t" /* A1-B1 a1-b1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
830 "psrad $" #shift ", %%mm0 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
831 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
832 "packssdw %%mm1, %%mm1 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
833 "movd %%mm1, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
834 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
835 "movd %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
836 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
837 "movd %%mm2, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
838 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
839 "movd %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
840 "movq 88(%2), %%mm1 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
841 "pmaddwd %%mm3, %%mm1 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
842 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
843 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
844 "paddd %%mm1, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
845 "psubd %%mm1, %%mm5 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
846 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
847 "psrad $" #shift ", %%mm5 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
848 "movq %%mm6, %%mm1 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
849 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
850 "psubd %%mm3, %%mm1 \n\t" /* a3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
851 "psrad $" #shift ", %%mm6 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
852 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
853 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
854 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
855 "movd %%mm2, 32+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
856 "packssdw %%mm1, %%mm1 \n\t" /* A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
857 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
858 "movd %%mm6, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
859 "movd %%mm1, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
860 "movd %%mm5, 80+" #dst " \n\t"
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
861
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
862 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
863 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
864 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
865 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
866 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
867 "jmp 9f \n\t"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
868
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
869 "#.balign 16 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
870 "6: \n\t"
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
871 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 7f)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
872
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
873 #undef IDCT
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
874 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
875 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
876 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
877 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
878 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
879 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
880 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
881 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
882 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
883 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
884 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
885 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
886 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
887 "movq 72(%2), %%mm7 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
888 "pmaddwd %%mm3, %%mm7 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
889 "paddd %%mm4, %%mm1 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
890 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
891 "psubd %%mm1, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
892 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
893 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
894 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
895 "paddd %%mm7, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
896 "psubd %%mm7, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
897 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
898 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
899 "packssdw %%mm1, %%mm1 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
900 "movd %%mm1, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
901 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
902 "movd %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
903 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
904 "movd %%mm2, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
905 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
906 "movd %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
907 "movq 88(%2), %%mm1 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
908 "pmaddwd %%mm3, %%mm1 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
909 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
910 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
911 "paddd %%mm1, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
912 "psubd %%mm1, %%mm5 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
913 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
914 "psrad $" #shift ", %%mm5 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
915 "movq %%mm6, %%mm1 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
916 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
917 "psubd %%mm3, %%mm1 \n\t" /* a3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
918 "psrad $" #shift ", %%mm6 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
919 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
920 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
921 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
922 "movd %%mm2, 32+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
923 "packssdw %%mm1, %%mm1 \n\t" /* A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
924 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
925 "movd %%mm6, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
926 "movd %%mm1, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
927 "movd %%mm5, 80+" #dst " \n\t"
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
928
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
929
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
930 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
931 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
932 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
933 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
934 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
935 "jmp 9f \n\t"
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
936
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
937 "#.balign 16 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
938 "2: \n\t"
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
939 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 3f)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
940
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
941 #undef IDCT
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
942 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
943 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
944 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
945 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
946 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
947 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
948 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
949 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
950 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
951 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
952 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
953 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
954 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
955 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
956 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
957 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
958 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
959 "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
960 "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
961 "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
962 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
963 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
964 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
965 "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
966 "psrad $" #shift ", %%mm7 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
967 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
968 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
969 "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
970 "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
971 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
972 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
973 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
974 "movd %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
975 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
976 "movd %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
977 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
978 "movd %%mm2, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
979 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
980 "movd %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
981 "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
982 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
983 "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
984 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
985 "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
986 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
987 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
988 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
989 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
990 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
991 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
992 "psrad $" #shift ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
993 "psrad $" #shift ", %%mm5 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
994 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
995 "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
996 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
997 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
998 "psrad $" #shift ", %%mm6 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
999 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1000 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1001 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1002 "movd %%mm2, 32+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1003 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1004 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1005 "movd %%mm6, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1006 "movd %%mm4, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1007 "movd %%mm5, 80+" #dst " \n\t"
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1008
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1009 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1010 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1011 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1012 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1013 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1014 "jmp 9f \n\t"
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1015
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1016 "#.balign 16 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1017 "3: \n\t"
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1018 #undef IDCT
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1019 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1020 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1021 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1022 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1023 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1024 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1025 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1026 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1027 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1028 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1029 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1030 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1031 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1032 "movq 64(%2), %%mm3 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1033 "pmaddwd %%mm2, %%mm3 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1034 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1035 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1036 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1037 "psrad $" #shift ", %%mm7 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1038 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1039 "movq %%mm0, %%mm1 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1040 "paddd %%mm3, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1041 "psubd %%mm3, %%mm1 \n\t" /* A1-B1 a1-b1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1042 "psrad $" #shift ", %%mm0 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1043 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1044 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1045 "movd %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1046 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1047 "movd %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1048 "packssdw %%mm1, %%mm1 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1049 "movd %%mm1, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1050 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1051 "movd %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1052 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1053 "pmaddwd %%mm2, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1054 "pmaddwd 96(%2), %%mm2 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1055 "movq %%mm5, %%mm1 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1056 "paddd %%mm4, %%mm1 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1057 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1058 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1059 "psrad $" #shift ", %%mm5 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1060 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1061 "paddd %%mm2, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1062 "psubd %%mm2, %%mm4 \n\t" /* a3-B3 a3-b3 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1063 "psrad $" #shift ", %%mm6 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1064 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1065 "packssdw %%mm1, %%mm1 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1066 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1067 "movd %%mm1, 32+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1068 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1069 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1070 "movd %%mm6, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1071 "movd %%mm4, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1072 "movd %%mm5, 80+" #dst " \n\t"
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1073
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1074
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1075 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1076 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1077 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1078 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1079 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1080 "jmp 9f \n\t"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1081
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1082 "#.balign 16 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1083 "5: \n\t"
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1084 #undef IDCT
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1085 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1086 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1087 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1088 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1089 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1090 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1091 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1092 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1093 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1094 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1095 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1096 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1097 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1098 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1099 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1100 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1101 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1102 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1103 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1104 "movq 8+" #src0 ", %%mm2 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1105 "movq 8+" #src4 ", %%mm3 \n\t" /* R6 R2 r6 r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1106 "movq 16(%2), %%mm1 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1107 "pmaddwd %%mm2, %%mm1 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1108 "movq 24(%2), %%mm7 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1109 "pmaddwd %%mm7, %%mm2 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1110 "movq 32(%2), %%mm7 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1111 "pmaddwd %%mm3, %%mm7 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1112 "pmaddwd 40(%2), %%mm3 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1113 #rounder ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1114 "paddd %%mm1, %%mm7 \n\t" /* A0 a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1115 "paddd %%mm1, %%mm1 \n\t" /* 2C0 2c0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1116 #rounder ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1117 "psubd %%mm7, %%mm1 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1118 "paddd %%mm2, %%mm3 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1119 "paddd %%mm2, %%mm2 \n\t" /* 2C1 2c1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1120 "psubd %%mm3, %%mm2 \n\t" /* A2 a2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1121 "psrad $" #shift ", %%mm4 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1122 "psrad $" #shift ", %%mm7 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1123 "psrad $" #shift ", %%mm3 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1124 "packssdw %%mm7, %%mm4 \n\t" /* A0 a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1125 "movq %%mm4, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1126 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1127 "packssdw %%mm3, %%mm0 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1128 "movq %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1129 "movq %%mm0, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1130 "movq %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1131 "psrad $" #shift ", %%mm5 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1132 "psrad $" #shift ", %%mm6 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1133 "psrad $" #shift ", %%mm2 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1134 "packssdw %%mm2, %%mm5 \n\t" /* A2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1135 "movq %%mm5, 32+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1136 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1137 "packssdw %%mm1, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1138 "movq %%mm6, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1139 "movq %%mm6, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1140 "movq %%mm5, 80+" #dst " \n\t"
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1141
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1142
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1143 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1144 IDCT( 0(%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1145 //IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1146 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1147 //IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1148 "jmp 9f \n\t"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1149
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1150
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1151 "#.balign 16 \n\t"\
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1152 "1: \n\t"
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1153 #undef IDCT
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1154 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1155 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1156 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1157 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1158 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1159 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1160 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1161 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1162 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1163 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1164 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1165 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1166 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1167 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1168 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1169 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1170 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1171 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1172 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1173 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1174 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1175 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1176 "movq 64(%2), %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1177 "pmaddwd %%mm2, %%mm1 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1178 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1179 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1180 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1181 "psrad $" #shift ", %%mm7 \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1182 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1183 "movq %%mm0, %%mm3 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1184 "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1185 "psubd %%mm1, %%mm3 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1186 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1187 "psrad $" #shift ", %%mm3 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1188 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1189 "movd %%mm7, " #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1190 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1191 "movd %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1192 "packssdw %%mm3, %%mm3 \n\t" /* A1-B1 a1-b1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1193 "movd %%mm3, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1194 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1195 "movd %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1196 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1197 "pmaddwd %%mm2, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1198 "pmaddwd 96(%2), %%mm2 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1199 "movq %%mm5, %%mm3 \n\t" /* A2 a2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1200 "paddd %%mm4, %%mm3 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1201 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1202 "psrad $" #shift ", %%mm3 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1203 "psrad $" #shift ", %%mm5 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1204 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1205 "paddd %%mm2, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1206 "psubd %%mm2, %%mm4 \n\t" /* a3-B3 a3-b3 */\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1207 "psrad $" #shift ", %%mm6 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1208 "packssdw %%mm3, %%mm3 \n\t" /* A2+B2 a2+b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1209 "movd %%mm3, 32+" #dst " \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1210 "psrad $" #shift ", %%mm4 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1211 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1212 "movd %%mm6, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1213 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1214 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1215 "movd %%mm4, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1216 "movd %%mm5, 80+" #dst " \n\t"
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1217
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1218
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1219 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1220 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1221 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1222 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1223 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1224 "jmp 9f \n\t"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1225
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1226
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1227 "#.balign 16 \n\t"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1228 "7: \n\t"
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1229 #undef IDCT
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1230 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1231 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1232 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1233 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1234 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1235 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1236 #rounder ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1237 #rounder ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1238 "psrad $" #shift ", %%mm4 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1239 "psrad $" #shift ", %%mm0 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1240 "movq 8+" #src0 ", %%mm2 \n\t" /* R4 R0 r4 r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1241 "movq 16(%2), %%mm1 \n\t" /* C4 C4 C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1242 "pmaddwd %%mm2, %%mm1 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1243 "movq 24(%2), %%mm7 \n\t" /* -C4 C4 -C4 C4 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1244 "pmaddwd %%mm7, %%mm2 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1245 "movq 32(%2), %%mm7 \n\t" /* C6 C2 C6 C2 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1246 #rounder ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1247 #rounder ", %%mm2 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1248 "psrad $" #shift ", %%mm1 \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1249 "packssdw %%mm1, %%mm4 \n\t" /* A0 a0 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1250 "movq %%mm4, " #dst " \n\t"\
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1251 "psrad $" #shift ", %%mm2 \n\t"\
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1252 "packssdw %%mm2, %%mm0 \n\t" /* A1 a1 */\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1253 "movq %%mm0, 16+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1254 "movq %%mm0, 96+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1255 "movq %%mm4, 112+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1256 "movq %%mm0, 32+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1257 "movq %%mm4, 48+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1258 "movq %%mm4, 64+" #dst " \n\t"\
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1259 "movq %%mm0, 80+" #dst " \n\t"
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1260
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1261 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1262 IDCT( 0(%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1263 //IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1264 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1265 //IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1266
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1267
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1268 #endif
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1269
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1270 /*
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1271 Input
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1272 00 40 04 44 20 60 24 64
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1273 10 30 14 34 50 70 54 74
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1274 01 41 03 43 21 61 23 63
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1275 11 31 13 33 51 71 53 73
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1276 02 42 06 46 22 62 26 66
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1277 12 32 16 36 52 72 56 76
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1278 05 45 07 47 25 65 27 67
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1279 15 35 17 37 55 75 57 77
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1280
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1281 Temp
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1282 00 04 10 14 20 24 30 34
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1283 40 44 50 54 60 64 70 74
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1284 01 03 11 13 21 23 31 33
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1285 41 43 51 53 61 63 71 73
209
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1286 02 06 12 16 22 26 32 36
c0d8ecae7ac5 (commit by michael)
arpi_esp
parents: 175
diff changeset
1287 42 46 52 56 62 66 72 76
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1288 05 07 15 17 25 27 35 37
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1289 45 47 55 57 65 67 75 77
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1290 */
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1291
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1292 "9: \n\t"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1293 :: "r" (block), "r" (temp), "r" (coeffs)
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1294 : "%eax"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1295 );
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1296 }
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1297
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1298 void ff_simple_idct_mmx(int16_t *block)
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1299 {
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1300 idct(block);
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1301 }
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1302
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1303 //FIXME merge add/put into the idct
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1304
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 854
diff changeset
1305 void ff_simple_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1306 {
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1307 idct(block);
854
3034f1816596 * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 706
diff changeset
1308 put_pixels_clamped_mmx(block, dest, line_size);
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
1309 }
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 854
diff changeset
1310 void ff_simple_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block)
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1311 {
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1312 idct(block);
854
3034f1816596 * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 706
diff changeset
1313 add_pixels_clamped_mmx(block, dest, line_size);
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 429
diff changeset
1314 }