annotate simple_idct.c @ 3025:8a7d694a21eb libavcodec

Make shared lib name handling consistent with the other Makefiles. patch by Jacob Meuser < jakemsr ..@.. jakemsr ... com >
author diego
date Mon, 09 Jan 2006 02:05:17 +0000
parents bfabfdf9ce55
children 0b546eab515d
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: 396
diff changeset
2 * Simple IDCT
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
3 *
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
4 * Copyright (c) 2001 Michael Niedermayer <michaelni@gmx.at>
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
5 *
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
6 * This library is free software; you can redistribute it and/or
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
8 * License as published by the Free Software Foundation; either
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
9 * version 2 of the License, or (at your option) any later version.
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
10 *
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
11 * This library is distributed in the hope that it will be useful,
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
14 * Lesser General Public License for more details.
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
15 *
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
17 * License along with this library; if not, write to the Free Software
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
19 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
20
1106
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
21 /**
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
22 * @file simple_idct.c
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
23 * simpleidct in C.
1e39f273ecd6 per file doxy
michaelni
parents: 1064
diff changeset
24 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
25
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
26 /*
429
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
27 based upon some outcommented c code from mpeg2dec (idct_mmx.c
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
28 written by Aaron Holtzman <aholtzma@ess.engr.uvic.ca>)
429
718a22dc121f license/copyright change
glantau
parents: 396
diff changeset
29 */
396
fce0a2520551 removed useless header includes - use av memory functions
glantau
parents: 352
diff changeset
30 #include "avcodec.h"
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
31 #include "dsputil.h"
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
32 #include "simple_idct.h"
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
33
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
34 #if 0
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
35 #define W1 2841 /* 2048*sqrt (2)*cos (1*pi/16) */
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
36 #define W2 2676 /* 2048*sqrt (2)*cos (2*pi/16) */
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
37 #define W3 2408 /* 2048*sqrt (2)*cos (3*pi/16) */
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
38 #define W4 2048 /* 2048*sqrt (2)*cos (4*pi/16) */
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
39 #define W5 1609 /* 2048*sqrt (2)*cos (5*pi/16) */
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
40 #define W6 1108 /* 2048*sqrt (2)*cos (6*pi/16) */
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
41 #define W7 565 /* 2048*sqrt (2)*cos (7*pi/16) */
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
42 #define ROW_SHIFT 8
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
43 #define COL_SHIFT 17
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
44 #else
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
45 #define W1 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
46 #define W2 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
47 #define W3 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: 215
diff changeset
48 #define W4 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
49 #define W5 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
50 #define W6 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
51 #define W7 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
52 #define ROW_SHIFT 11
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
53 #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
54 #endif
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
55
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
56 #if defined(ARCH_POWERPC_405)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
57
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
58 /* signed 16x16 -> 32 multiply add accumulate */
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
59 #define MAC16(rt, ra, rb) \
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
60 asm ("maclhw %0, %2, %3" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb));
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
61
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
62 /* signed 16x16 -> 32 multiply */
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
63 #define MUL16(rt, ra, rb) \
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
64 asm ("mullhw %0, %1, %2" : "=r" (rt) : "r" (ra), "r" (rb));
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
65
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
66 #else
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
67
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
68 /* signed 16x16 -> 32 multiply add accumulate */
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
69 #define MAC16(rt, ra, rb) rt += (ra) * (rb)
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
70
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
71 /* signed 16x16 -> 32 multiply */
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
72 #define MUL16(rt, ra, rb) rt = (ra) * (rb)
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
73
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
74 #endif
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
75
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
76 static inline void idctRowCondDC (DCTELEM * row)
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
77 {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
78 int a0, a1, a2, a3, b0, b1, b2, b3;
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
79 #ifdef FAST_64BIT
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
80 uint64_t temp;
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
81 #else
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
82 uint32_t temp;
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
83 #endif
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
84
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
85 #ifdef FAST_64BIT
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
86 #ifdef WORDS_BIGENDIAN
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
87 #define ROW0_MASK 0xffff000000000000LL
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
88 #else
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
89 #define ROW0_MASK 0xffffLL
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
90 #endif
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
91 if(sizeof(DCTELEM)==2){
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
92 if ( ((((uint64_t *)row)[0] & ~ROW0_MASK) |
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
93 ((uint64_t *)row)[1]) == 0) {
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
94 temp = (row[0] << 3) & 0xffff;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
95 temp += temp << 16;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
96 temp += temp << 32;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
97 ((uint64_t *)row)[0] = temp;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
98 ((uint64_t *)row)[1] = temp;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
99 return;
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
100 }
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
101 }else{
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
102 if (!(row[1]|row[2]|row[3]|row[4]|row[5]|row[6]|row[7])) {
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
103 row[0]=row[1]=row[2]=row[3]=row[4]=row[5]=row[6]=row[7]= row[0] << 3;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
104 return;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
105 }
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
106 }
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
107 #else
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
108 if(sizeof(DCTELEM)==2){
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
109 if (!(((uint32_t*)row)[1] |
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
110 ((uint32_t*)row)[2] |
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
111 ((uint32_t*)row)[3] |
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
112 row[1])) {
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
113 temp = (row[0] << 3) & 0xffff;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
114 temp += temp << 16;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
115 ((uint32_t*)row)[0]=((uint32_t*)row)[1] =
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
116 ((uint32_t*)row)[2]=((uint32_t*)row)[3] = temp;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
117 return;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
118 }
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
119 }else{
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
120 if (!(row[1]|row[2]|row[3]|row[4]|row[5]|row[6]|row[7])) {
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
121 row[0]=row[1]=row[2]=row[3]=row[4]=row[5]=row[6]=row[7]= row[0] << 3;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
122 return;
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
123 }
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
124 }
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
125 #endif
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
126
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
127 a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1));
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
128 a1 = a0;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
129 a2 = a0;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
130 a3 = a0;
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
131
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
132 /* no need to optimize : gcc does it */
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
133 a0 += W2 * row[2];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
134 a1 += W6 * row[2];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
135 a2 -= W6 * row[2];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
136 a3 -= W2 * row[2];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
137
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
138 MUL16(b0, W1, row[1]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
139 MAC16(b0, W3, row[3]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
140 MUL16(b1, W3, row[1]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
141 MAC16(b1, -W7, row[3]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
142 MUL16(b2, W5, row[1]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
143 MAC16(b2, -W1, row[3]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
144 MUL16(b3, W7, row[1]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
145 MAC16(b3, -W5, row[3]);
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
146
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
147 #ifdef FAST_64BIT
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
148 temp = ((uint64_t*)row)[1];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
149 #else
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
150 temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
151 #endif
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
152 if (temp != 0) {
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
153 a0 += W4*row[4] + W6*row[6];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
154 a1 += - W4*row[4] - W2*row[6];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
155 a2 += - W4*row[4] + W2*row[6];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
156 a3 += W4*row[4] - W6*row[6];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
157
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
158 MAC16(b0, W5, row[5]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
159 MAC16(b0, W7, row[7]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
160
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
161 MAC16(b1, -W1, row[5]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
162 MAC16(b1, -W5, row[7]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
163
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
164 MAC16(b2, W7, row[5]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
165 MAC16(b2, W3, row[7]);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
166
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
167 MAC16(b3, W3, row[5]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
168 MAC16(b3, -W1, row[7]);
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
169 }
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
170
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
171 row[0] = (a0 + b0) >> ROW_SHIFT;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
172 row[7] = (a0 - b0) >> ROW_SHIFT;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
173 row[1] = (a1 + b1) >> ROW_SHIFT;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
174 row[6] = (a1 - b1) >> ROW_SHIFT;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
175 row[2] = (a2 + b2) >> ROW_SHIFT;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
176 row[5] = (a2 - b2) >> ROW_SHIFT;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
177 row[3] = (a3 + b3) >> ROW_SHIFT;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
178 row[4] = (a3 - b3) >> ROW_SHIFT;
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
179 }
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
180
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
181 static inline void idctSparseColPut (uint8_t *dest, int line_size,
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
182 DCTELEM * col)
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
183 {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
184 int a0, a1, a2, a3, b0, b1, b2, b3;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
185 uint8_t *cm = cropTbl + MAX_NEG_CROP;
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
186
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
187 /* XXX: I did that only to give same values as previous code */
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
188 a0 = W4 * (col[8*0] + ((1<<(COL_SHIFT-1))/W4));
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
189 a1 = a0;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
190 a2 = a0;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
191 a3 = a0;
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
192
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
193 a0 += + W2*col[8*2];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
194 a1 += + W6*col[8*2];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
195 a2 += - W6*col[8*2];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
196 a3 += - W2*col[8*2];
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
197
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
198 MUL16(b0, W1, col[8*1]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
199 MUL16(b1, W3, col[8*1]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
200 MUL16(b2, W5, col[8*1]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
201 MUL16(b3, W7, col[8*1]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
202
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
203 MAC16(b0, + W3, col[8*3]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
204 MAC16(b1, - W7, col[8*3]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
205 MAC16(b2, - W1, col[8*3]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
206 MAC16(b3, - W5, col[8*3]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
207
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
208 if(col[8*4]){
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
209 a0 += + W4*col[8*4];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
210 a1 += - W4*col[8*4];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
211 a2 += - W4*col[8*4];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
212 a3 += + W4*col[8*4];
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
213 }
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
214
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
215 if (col[8*5]) {
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
216 MAC16(b0, + W5, col[8*5]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
217 MAC16(b1, - W1, col[8*5]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
218 MAC16(b2, + W7, col[8*5]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
219 MAC16(b3, + W3, col[8*5]);
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
220 }
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
221
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
222 if(col[8*6]){
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
223 a0 += + W6*col[8*6];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
224 a1 += - W2*col[8*6];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
225 a2 += + W2*col[8*6];
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
226 a3 += - W6*col[8*6];
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
227 }
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
228
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
229 if (col[8*7]) {
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
230 MAC16(b0, + W7, col[8*7]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
231 MAC16(b1, - W5, col[8*7]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
232 MAC16(b2, + W3, col[8*7]);
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
233 MAC16(b3, - W1, col[8*7]);
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
234 }
205
ccf36af385f3 (commit by michael)
arpi_esp
parents: 175
diff changeset
235
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
236 dest[0] = cm[(a0 + b0) >> COL_SHIFT];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
237 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
238 dest[0] = cm[(a1 + b1) >> COL_SHIFT];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
239 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
240 dest[0] = cm[(a2 + b2) >> COL_SHIFT];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
241 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
242 dest[0] = cm[(a3 + b3) >> COL_SHIFT];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
243 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
244 dest[0] = cm[(a3 - b3) >> COL_SHIFT];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
245 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
246 dest[0] = cm[(a2 - b2) >> COL_SHIFT];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
247 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
248 dest[0] = cm[(a1 - b1) >> COL_SHIFT];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
249 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
250 dest[0] = cm[(a0 - b0) >> COL_SHIFT];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
251 }
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
252
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
253 static inline void idctSparseColAdd (uint8_t *dest, int line_size,
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
254 DCTELEM * col)
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
255 {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
256 int a0, a1, a2, a3, b0, b1, b2, b3;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
257 uint8_t *cm = cropTbl + MAX_NEG_CROP;
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
258
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
259 /* XXX: I did that only to give same values as previous code */
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
260 a0 = W4 * (col[8*0] + ((1<<(COL_SHIFT-1))/W4));
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
261 a1 = a0;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
262 a2 = a0;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
263 a3 = a0;
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
264
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
265 a0 += + W2*col[8*2];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
266 a1 += + W6*col[8*2];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
267 a2 += - W6*col[8*2];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
268 a3 += - W2*col[8*2];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
269
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
270 MUL16(b0, W1, col[8*1]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
271 MUL16(b1, W3, col[8*1]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
272 MUL16(b2, W5, col[8*1]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
273 MUL16(b3, W7, col[8*1]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
274
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
275 MAC16(b0, + W3, col[8*3]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
276 MAC16(b1, - W7, col[8*3]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
277 MAC16(b2, - W1, col[8*3]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
278 MAC16(b3, - W5, col[8*3]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
279
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
280 if(col[8*4]){
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
281 a0 += + W4*col[8*4];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
282 a1 += - W4*col[8*4];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
283 a2 += - W4*col[8*4];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
284 a3 += + W4*col[8*4];
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
285 }
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
286
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
287 if (col[8*5]) {
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
288 MAC16(b0, + W5, col[8*5]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
289 MAC16(b1, - W1, col[8*5]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
290 MAC16(b2, + W7, col[8*5]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
291 MAC16(b3, + W3, col[8*5]);
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
292 }
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
293
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
294 if(col[8*6]){
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
295 a0 += + W6*col[8*6];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
296 a1 += - W2*col[8*6];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
297 a2 += + W2*col[8*6];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
298 a3 += - W6*col[8*6];
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
299 }
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
300
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
301 if (col[8*7]) {
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
302 MAC16(b0, + W7, col[8*7]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
303 MAC16(b1, - W5, col[8*7]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
304 MAC16(b2, + W3, col[8*7]);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
305 MAC16(b3, - W1, col[8*7]);
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
306 }
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
307
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
308 dest[0] = cm[dest[0] + ((a0 + b0) >> COL_SHIFT)];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
309 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
310 dest[0] = cm[dest[0] + ((a1 + b1) >> COL_SHIFT)];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
311 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
312 dest[0] = cm[dest[0] + ((a2 + b2) >> COL_SHIFT)];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
313 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
314 dest[0] = cm[dest[0] + ((a3 + b3) >> COL_SHIFT)];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
315 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
316 dest[0] = cm[dest[0] + ((a3 - b3) >> COL_SHIFT)];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
317 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
318 dest[0] = cm[dest[0] + ((a2 - b2) >> COL_SHIFT)];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
319 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
320 dest[0] = cm[dest[0] + ((a1 - b1) >> COL_SHIFT)];
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
321 dest += line_size;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
322 dest[0] = cm[dest[0] + ((a0 - b0) >> COL_SHIFT)];
175
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
323 }
bd77d3cbb233 new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff changeset
324
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
325 static inline void idctSparseCol (DCTELEM * col)
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
326 {
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
327 int a0, a1, a2, a3, b0, b1, b2, b3;
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
328
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
329 /* XXX: I did that only to give same values as previous code */
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
330 a0 = W4 * (col[8*0] + ((1<<(COL_SHIFT-1))/W4));
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
331 a1 = a0;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
332 a2 = a0;
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
333 a3 = a0;
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
334
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
335 a0 += + W2*col[8*2];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
336 a1 += + W6*col[8*2];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
337 a2 += - W6*col[8*2];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
338 a3 += - W2*col[8*2];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
339
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
340 MUL16(b0, W1, col[8*1]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
341 MUL16(b1, W3, col[8*1]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
342 MUL16(b2, W5, col[8*1]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
343 MUL16(b3, W7, col[8*1]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
344
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
345 MAC16(b0, + W3, col[8*3]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
346 MAC16(b1, - W7, col[8*3]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
347 MAC16(b2, - W1, col[8*3]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
348 MAC16(b3, - W5, col[8*3]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
349
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
350 if(col[8*4]){
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
351 a0 += + W4*col[8*4];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
352 a1 += - W4*col[8*4];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
353 a2 += - W4*col[8*4];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
354 a3 += + W4*col[8*4];
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
355 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
356
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
357 if (col[8*5]) {
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
358 MAC16(b0, + W5, col[8*5]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
359 MAC16(b1, - W1, col[8*5]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
360 MAC16(b2, + W7, col[8*5]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
361 MAC16(b3, + W3, col[8*5]);
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
362 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
363
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
364 if(col[8*6]){
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
365 a0 += + W6*col[8*6];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
366 a1 += - W2*col[8*6];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
367 a2 += + W2*col[8*6];
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
368 a3 += - W6*col[8*6];
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
369 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
370
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
371 if (col[8*7]) {
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
372 MAC16(b0, + W7, col[8*7]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
373 MAC16(b1, - W5, col[8*7]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
374 MAC16(b2, + W3, col[8*7]);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
375 MAC16(b3, - W1, col[8*7]);
2979
bfabfdf9ce55 COSMETICS: tabs --> spaces, some prettyprinting
diego
parents: 2967
diff changeset
376 }
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
377
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
378 col[0 ] = ((a0 + b0) >> COL_SHIFT);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
379 col[8 ] = ((a1 + b1) >> COL_SHIFT);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
380 col[16] = ((a2 + b2) >> COL_SHIFT);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
381 col[24] = ((a3 + b3) >> COL_SHIFT);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
382 col[32] = ((a3 - b3) >> COL_SHIFT);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
383 col[40] = ((a2 - b2) >> COL_SHIFT);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
384 col[48] = ((a1 - b1) >> COL_SHIFT);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
385 col[56] = ((a0 - b0) >> COL_SHIFT);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
386 }
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
387
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
388 void simple_idct_put(uint8_t *dest, int line_size, DCTELEM *block)
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
389 {
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
390 int i;
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
391 for(i=0; i<8; i++)
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
392 idctRowCondDC(block + i*8);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
393
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
394 for(i=0; i<8; i++)
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
395 idctSparseColPut(dest + i, line_size, block + i);
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
396 }
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
397
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
398 void simple_idct_add(uint8_t *dest, int line_size, DCTELEM *block)
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
399 {
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
400 int i;
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
401 for(i=0; i<8; i++)
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
402 idctRowCondDC(block + i*8);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
403
479
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
404 for(i=0; i<8; i++)
40ffce2cb6ef added inlined put/add functions
bellard
parents: 476
diff changeset
405 idctSparseColAdd(dest + i, line_size, block + i);
476
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
406 }
ec13b0a726c3 removed unused code - began to merge alpha specific stuff - added mac macros for suitable CPUs
bellard
parents: 464
diff changeset
407
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
408 void simple_idct(DCTELEM *block)
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
409 {
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
410 int i;
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
411 for(i=0; i<8; i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
412 idctRowCondDC(block + i*8);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
413
633
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
414 for(i=0; i<8; i++)
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
415 idctSparseCol(block + i);
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
416 }
e7b72c1dfa1b dct-test update
michaelni
parents: 503
diff changeset
417
719
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
418 /* 2x4x8 idct */
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
419
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
420 #define CN_SHIFT 12
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
421 #define C_FIX(x) ((int)((x) * (1 << CN_SHIFT) + 0.5))
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 719
diff changeset
422 #define C1 C_FIX(0.6532814824)
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 719
diff changeset
423 #define C2 C_FIX(0.2705980501)
719
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
424
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 719
diff changeset
425 /* row idct is multiple by 16 * sqrt(2.0), col idct4 is normalized,
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 719
diff changeset
426 and the butterfly must be multiplied by 0.5 * sqrt(2.0) */
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 719
diff changeset
427 #define C_SHIFT (4+1+12)
719
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
428
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
429 static inline void idct4col(uint8_t *dest, int line_size, const DCTELEM *col)
719
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
430 {
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
431 int c0, c1, c2, c3, a0, a1, a2, a3;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
432 const uint8_t *cm = cropTbl + MAX_NEG_CROP;
719
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
433
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
434 a0 = col[8*0];
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
435 a1 = col[8*2];
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
436 a2 = col[8*4];
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
437 a3 = col[8*6];
722
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 719
diff changeset
438 c0 = ((a0 + a2) << (CN_SHIFT - 1)) + (1 << (C_SHIFT - 1));
ff90043f4a2d in fact IDCT248 needs to be normalized as I suspected
bellard
parents: 719
diff changeset
439 c2 = ((a0 - a2) << (CN_SHIFT - 1)) + (1 << (C_SHIFT - 1));
719
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
440 c1 = a1 * C1 + a3 * C2;
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
441 c3 = a1 * C2 - a3 * C1;
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
442 dest[0] = cm[(c0 + c1) >> C_SHIFT];
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
443 dest += line_size;
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
444 dest[0] = cm[(c2 + c3) >> C_SHIFT];
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
445 dest += line_size;
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
446 dest[0] = cm[(c2 - c3) >> C_SHIFT];
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
447 dest += line_size;
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
448 dest[0] = cm[(c0 - c1) >> C_SHIFT];
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
449 }
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
450
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
451 #define BF(k) \
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
452 {\
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
453 int a0, a1;\
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
454 a0 = ptr[k];\
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
455 a1 = ptr[8 + k];\
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
456 ptr[k] = a0 + a1;\
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
457 ptr[8 + k] = a0 - a1;\
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
458 }
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
459
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
460 /* only used by DV codec. The input must be interlaced. 128 is added
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
461 to the pixels before clamping to avoid systematic error
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
462 (1024*sqrt(2)) offset would be needed otherwise. */
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
463 /* XXX: I think a 1.0/sqrt(2) normalization should be needed to
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
464 compensate the extra butterfly stage - I don't have the full DV
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
465 specification */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
466 void simple_idct248_put(uint8_t *dest, int line_size, DCTELEM *block)
719
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
467 {
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
468 int i;
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
469 DCTELEM *ptr;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 1106
diff changeset
470
719
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
471 /* butterfly */
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
472 ptr = block;
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
473 for(i=0;i<4;i++) {
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
474 BF(0);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
475 BF(1);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
476 BF(2);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
477 BF(3);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
478 BF(4);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
479 BF(5);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
480 BF(6);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
481 BF(7);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
482 ptr += 2 * 8;
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
483 }
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
484
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
485 /* IDCT8 on each line */
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
486 for(i=0; i<8; i++) {
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
487 idctRowCondDC(block + i*8);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
488 }
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
489
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
490 /* IDCT4 and store */
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
491 for(i=0;i<8;i++) {
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
492 idct4col(dest + i, 2 * line_size, block + i);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
493 idct4col(dest + line_size + i, 2 * line_size, block + 8 + i);
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
494 }
2b7ff6dfee35 first version of IDCT248 for DV decoding support
bellard
parents: 642
diff changeset
495 }
936
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
496
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
497 /* 8x4 & 4x8 WMV2 IDCT */
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
498 #undef CN_SHIFT
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
499 #undef C_SHIFT
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
500 #undef C_FIX
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
501 #undef C1
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
502 #undef C2
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
503 #define CN_SHIFT 12
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
504 #define C_FIX(x) ((int)((x) * 1.414213562 * (1 << CN_SHIFT) + 0.5))
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
505 #define C1 C_FIX(0.6532814824)
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
506 #define C2 C_FIX(0.2705980501)
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
507 #define C3 C_FIX(0.5)
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
508 #define C_SHIFT (4+1+12)
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
509 static inline void idct4col_add(uint8_t *dest, int line_size, const DCTELEM *col)
936
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
510 {
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
511 int c0, c1, c2, c3, a0, a1, a2, a3;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
512 const uint8_t *cm = cropTbl + MAX_NEG_CROP;
936
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
513
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
514 a0 = col[8*0];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
515 a1 = col[8*1];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
516 a2 = col[8*2];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
517 a3 = col[8*3];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
518 c0 = (a0 + a2)*C3 + (1 << (C_SHIFT - 1));
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
519 c2 = (a0 - a2)*C3 + (1 << (C_SHIFT - 1));
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
520 c1 = a1 * C1 + a3 * C2;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
521 c3 = a1 * C2 - a3 * C1;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
522 dest[0] = cm[dest[0] + ((c0 + c1) >> C_SHIFT)];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
523 dest += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
524 dest[0] = cm[dest[0] + ((c2 + c3) >> C_SHIFT)];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
525 dest += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
526 dest[0] = cm[dest[0] + ((c2 - c3) >> C_SHIFT)];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
527 dest += line_size;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
528 dest[0] = cm[dest[0] + ((c0 - c1) >> C_SHIFT)];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
529 }
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
530
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
531 #define RN_SHIFT 15
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
532 #define R_FIX(x) ((int)((x) * 1.414213562 * (1 << RN_SHIFT) + 0.5))
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
533 #define R1 R_FIX(0.6532814824)
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
534 #define R2 R_FIX(0.2705980501)
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
535 #define R3 R_FIX(0.5)
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
536 #define R_SHIFT 11
1008
fb6cbb8a04a3 fixing DCTELEM != short
michaelni
parents: 936
diff changeset
537 static inline void idct4row(DCTELEM *row)
936
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
538 {
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
539 int c0, c1, c2, c3, a0, a1, a2, a3;
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
540 //const uint8_t *cm = cropTbl + MAX_NEG_CROP;
936
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
541
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
542 a0 = row[0];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
543 a1 = row[1];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
544 a2 = row[2];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
545 a3 = row[3];
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
546 c0 = (a0 + a2)*R3 + (1 << (R_SHIFT - 1));
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
547 c2 = (a0 - a2)*R3 + (1 << (R_SHIFT - 1));
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
548 c1 = a1 * R1 + a3 * R2;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
549 c3 = a1 * R2 - a3 * R1;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
550 row[0]= (c0 + c1) >> R_SHIFT;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
551 row[1]= (c2 + c3) >> R_SHIFT;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
552 row[2]= (c2 - c3) >> R_SHIFT;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
553 row[3]= (c0 - c1) >> R_SHIFT;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
554 }
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
555
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
556 void simple_idct84_add(uint8_t *dest, int line_size, DCTELEM *block)
936
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
557 {
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
558 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
559
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
560 /* IDCT8 on each line */
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
561 for(i=0; i<4; i++) {
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
562 idctRowCondDC(block + i*8);
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
563 }
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
564
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
565 /* IDCT4 and store */
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
566 for(i=0;i<8;i++) {
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
567 idct4col_add(dest + i, line_size, block + i);
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
568 }
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
569 }
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
570
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1008
diff changeset
571 void simple_idct48_add(uint8_t *dest, int line_size, DCTELEM *block)
936
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
572 {
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
573 int i;
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
574
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
575 /* IDCT4 on each line */
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
576 for(i=0; i<8; i++) {
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
577 idct4row(block + i*8);
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
578 }
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
579
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
580 /* IDCT8 and store */
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
581 for(i=0; i<4; i++){
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
582 idctSparseColAdd(dest + i, line_size, block + i);
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
583 }
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
584 }
caa77cd960c0 qpel encoding
michaelni
parents: 744
diff changeset
585