Mercurial > libavcodec.hg
annotate simple_idct.c @ 419:b71190bacce8 libavcodec
applied 64bit patch from Ulrich Hecht <uli at suse dot de>
author | alex |
---|---|
date | Thu, 23 May 2002 15:33:23 +0000 |
parents | fce0a2520551 |
children | 718a22dc121f |
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 /* |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
2 Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at) |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
3 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
4 This program is free software; you can redistribute it and/or modify |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
5 it under the terms of the GNU General Public License as published by |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
6 the Free Software Foundation; either version 2 of the License, or |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
7 (at your option) any later version. |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
8 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
9 This program is distributed in the hope that it will be useful, |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
10 but WITHOUT ANY WARRANTY; without even the implied warranty of |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
12 GNU General Public License for more details. |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
13 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
14 You should have received a copy of the GNU General Public License |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
15 along with this program; if not, write to the Free Software |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
17 */ |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
18 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
19 /* |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
20 based upon some outcommented c code from mpeg2dec (idct_mmx.c written by Aaron Holtzman <aholtzma@ess.engr.uvic.ca>) |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
21 */ |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
22 |
396
fce0a2520551
removed useless header includes - use av memory functions
glantau
parents:
352
diff
changeset
|
23 #include "avcodec.h" |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
24 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
25 #include "simple_idct.h" |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
26 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
27 #if 0 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
28 #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
|
29 #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
|
30 #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
|
31 #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
|
32 #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
|
33 #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
|
34 #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
|
35 #define ROW_SHIFT 8 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
36 #define COL_SHIFT 17 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
37 #else |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
38 #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
|
39 #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
|
40 #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
|
41 #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
|
42 #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
|
43 #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
|
44 #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
|
45 #define ROW_SHIFT 11 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
46 #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
|
47 #endif |
205 | 48 |
49 /* 8x8 Matrix used to do a trivial (slow) 8 point IDCT */ | |
50 static int coeff[64]={ | |
51 W4, W4, W4, W4, W4, W4, W4, W4, | |
52 W1, W3, W5, W7,-W7,-W5,-W3,-W1, | |
53 W2, W6,-W6,-W2,-W2,-W6, W6, W2, | |
54 W3,-W7,-W1,-W5, W5, W1, W7,-W3, | |
55 W4,-W4,-W4, W4, W4,-W4,-W4, W4, | |
56 W5,-W1, W7, W3,-W3,-W7, W1,-W5, | |
57 W6,-W2, W2,-W6,-W6, W2,-W2, W6, | |
58 W7,-W5, W3,-W1, W1,-W3, W5,-W7 | |
59 }; | |
60 | |
61 static int inline idctRowCondZ (int16_t * row) | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
62 { |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
63 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
|
64 |
205 | 65 if( !( ((uint32_t*)row)[0]|((uint32_t*)row)[1] |((uint32_t*)row)[2] |((uint32_t*)row)[3])) { |
66 /* row[0] = row[1] = row[2] = row[3] = row[4] = | |
67 row[5] = row[6] = row[7] = 0;*/ | |
68 return 0; | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
69 } |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
70 |
205 | 71 if(!( ((uint32_t*)row)[2] |((uint32_t*)row)[3] )){ |
72 a0 = W4*row[0] + W2*row[2] + (1<<(ROW_SHIFT-1)); | |
73 a1 = W4*row[0] + W6*row[2] + (1<<(ROW_SHIFT-1)); | |
74 a2 = W4*row[0] - W6*row[2] + (1<<(ROW_SHIFT-1)); | |
75 a3 = W4*row[0] - W2*row[2] + (1<<(ROW_SHIFT-1)); | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
76 |
205 | 77 b0 = W1*row[1] + W3*row[3]; |
78 b1 = W3*row[1] - W7*row[3]; | |
79 b2 = W5*row[1] - W1*row[3]; | |
80 b3 = W7*row[1] - W5*row[3]; | |
81 }else{ | |
82 a0 = W4*row[0] + W2*row[2] + W4*row[4] + W6*row[6] + (1<<(ROW_SHIFT-1)); | |
83 a1 = W4*row[0] + W6*row[2] - W4*row[4] - W2*row[6] + (1<<(ROW_SHIFT-1)); | |
84 a2 = W4*row[0] - W6*row[2] - W4*row[4] + W2*row[6] + (1<<(ROW_SHIFT-1)); | |
85 a3 = W4*row[0] - W2*row[2] + W4*row[4] - W6*row[6] + (1<<(ROW_SHIFT-1)); | |
86 | |
87 b0 = W1*row[1] + W3*row[3] + W5*row[5] + W7*row[7]; | |
88 b1 = W3*row[1] - W7*row[3] - W1*row[5] - W5*row[7]; | |
89 b2 = W5*row[1] - W1*row[3] + W7*row[5] + W3*row[7]; | |
90 b3 = W7*row[1] - W5*row[3] + W3*row[5] - W1*row[7]; | |
91 } | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
92 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
93 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
|
94 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
|
95 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
|
96 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
|
97 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
|
98 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
|
99 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
|
100 row[7] = (a0 - b0) >> ROW_SHIFT; |
205 | 101 |
102 return 1; | |
103 } | |
104 | |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
105 #ifdef ARCH_ALPHA |
215
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
106 /* 0: all entries 0, 1: only first entry nonzero, 2: otherwise */ |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
107 static int inline idctRowCondDC(int16_t *row) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
108 { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
109 int_fast32_t a0, a1, a2, a3, b0, b1, b2, b3; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
110 uint64_t *lrow = (uint64_t *) row; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
111 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
112 if (lrow[1] == 0) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
113 if (lrow[0] == 0) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
114 return 0; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
115 if ((lrow[0] & ~0xffffULL) == 0) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
116 uint64_t v; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
117 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
118 a0 = W4 * row[0]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
119 a0 += 1 << (ROW_SHIFT - 1); |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
120 a0 >>= ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
121 v = (uint16_t) a0; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
122 v += v << 16; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
123 v += v << 32; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
124 lrow[0] = v; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
125 lrow[1] = v; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
126 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
127 return 1; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
128 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
129 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
130 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
131 a0 = W4 * row[0]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
132 a1 = W4 * row[0]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
133 a2 = W4 * row[0]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
134 a3 = W4 * row[0]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
135 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
136 if (row[2]) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
137 a0 += W2 * row[2]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
138 a1 += W6 * row[2]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
139 a2 -= W6 * row[2]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
140 a3 -= W2 * row[2]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
141 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
142 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
143 if (row[4]) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
144 a0 += W4 * row[4]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
145 a1 -= W4 * row[4]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
146 a2 -= W4 * row[4]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
147 a3 += W4 * row[4]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
148 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
149 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
150 if (row[6]) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
151 a0 += W6 * row[6]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
152 a1 -= W2 * row[6]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
153 a2 += W2 * row[6]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
154 a3 -= W6 * row[6]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
155 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
156 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
157 a0 += 1 << (ROW_SHIFT - 1); |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
158 a1 += 1 << (ROW_SHIFT - 1); |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
159 a2 += 1 << (ROW_SHIFT - 1); |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
160 a3 += 1 << (ROW_SHIFT - 1); |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
161 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
162 if (row[1]) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
163 b0 = W1 * row[1]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
164 b1 = W3 * row[1]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
165 b2 = W5 * row[1]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
166 b3 = W7 * row[1]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
167 } else { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
168 b0 = 0; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
169 b1 = 0; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
170 b2 = 0; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
171 b3 = 0; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
172 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
173 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
174 if (row[3]) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
175 b0 += W3 * row[3]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
176 b1 -= W7 * row[3]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
177 b2 -= W1 * row[3]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
178 b3 -= W5 * row[3]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
179 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
180 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
181 if (row[5]) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
182 b0 += W5 * row[5]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
183 b1 -= W1 * row[5]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
184 b2 += W7 * row[5]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
185 b3 += W3 * row[5]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
186 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
187 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
188 if (row[7]) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
189 b0 += W7 * row[7]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
190 b1 -= W5 * row[7]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
191 b2 += W3 * row[7]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
192 b3 -= W1 * row[7]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
193 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
194 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
195 row[0] = (a0 + b0) >> ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
196 row[1] = (a1 + b1) >> ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
197 row[2] = (a2 + b2) >> ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
198 row[3] = (a3 + b3) >> ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
199 row[4] = (a3 - b3) >> ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
200 row[5] = (a2 - b2) >> ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
201 row[6] = (a1 - b1) >> ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
202 row[7] = (a0 - b0) >> ROW_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
203 |
215
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
204 return 2; |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
205 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
206 #else /* not ARCH_ALPHA */ |
205 | 207 static int inline idctRowCondDC (int16_t * row) |
208 { | |
209 int a0, a1, a2, a3, b0, b1, b2, b3; | |
210 | |
211 if( !( ((uint32_t*)row)[1] |((uint32_t*)row)[2] |((uint32_t*)row)[3]| row[1])) { | |
212 // row[0] = row[1] = row[2] = row[3] = row[4] = row[5] = row[6] = row[7] = row[0]<<3; | |
213 uint16_t temp= row[0]<<3; | |
214 ((uint32_t*)row)[0]=((uint32_t*)row)[1]= | |
215 ((uint32_t*)row)[2]=((uint32_t*)row)[3]= temp + (temp<<16); | |
216 return 0; | |
217 } | |
218 | |
219 if(!( ((uint32_t*)row)[2] |((uint32_t*)row)[3] )){ | |
220 a0 = W4*row[0] + W2*row[2] + (1<<(ROW_SHIFT-1)); | |
221 a1 = W4*row[0] + W6*row[2] + (1<<(ROW_SHIFT-1)); | |
222 a2 = W4*row[0] - W6*row[2] + (1<<(ROW_SHIFT-1)); | |
223 a3 = W4*row[0] - W2*row[2] + (1<<(ROW_SHIFT-1)); | |
224 | |
225 b0 = W1*row[1] + W3*row[3]; | |
226 b1 = W3*row[1] - W7*row[3]; | |
227 b2 = W5*row[1] - W1*row[3]; | |
228 b3 = W7*row[1] - W5*row[3]; | |
229 }else{ | |
230 a0 = W4*row[0] + W2*row[2] + W4*row[4] + W6*row[6] + (1<<(ROW_SHIFT-1)); | |
231 a1 = W4*row[0] + W6*row[2] - W4*row[4] - W2*row[6] + (1<<(ROW_SHIFT-1)); | |
232 a2 = W4*row[0] - W6*row[2] - W4*row[4] + W2*row[6] + (1<<(ROW_SHIFT-1)); | |
233 a3 = W4*row[0] - W2*row[2] + W4*row[4] - W6*row[6] + (1<<(ROW_SHIFT-1)); | |
234 | |
235 b0 = W1*row[1] + W3*row[3] + W5*row[5] + W7*row[7]; | |
236 b1 = W3*row[1] - W7*row[3] - W1*row[5] - W5*row[7]; | |
237 b2 = W5*row[1] - W1*row[3] + W7*row[5] + W3*row[7]; | |
238 b3 = W7*row[1] - W5*row[3] + W3*row[5] - W1*row[7]; | |
239 } | |
240 | |
241 row[0] = (a0 + b0) >> ROW_SHIFT; | |
242 row[7] = (a0 - b0) >> ROW_SHIFT; | |
243 row[1] = (a1 + b1) >> ROW_SHIFT; | |
244 row[6] = (a1 - b1) >> ROW_SHIFT; | |
245 row[2] = (a2 + b2) >> ROW_SHIFT; | |
246 row[5] = (a2 - b2) >> ROW_SHIFT; | |
247 row[3] = (a3 + b3) >> ROW_SHIFT; | |
248 row[4] = (a3 - b3) >> ROW_SHIFT; | |
249 | |
250 return 1; | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
251 } |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
252 #endif /* not ARCH_ALPHA */ |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
253 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
254 static void inline idctCol (int16_t * col) |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
255 { |
205 | 256 |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
257 /* |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
258 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
|
259 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
|
260 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
|
261 return; |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
262 }*/ |
205 | 263 |
264 int a0, a1, a2, a3, b0, b1, b2, b3; | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
265 col[0] += (1<<(COL_SHIFT-1))/W4; |
205 | 266 a0 = W4*col[8*0] + W2*col[8*2] + W4*col[8*4] + W6*col[8*6]; |
267 a1 = W4*col[8*0] + W6*col[8*2] - W4*col[8*4] - W2*col[8*6]; | |
268 a2 = W4*col[8*0] - W6*col[8*2] - W4*col[8*4] + W2*col[8*6]; | |
269 a3 = W4*col[8*0] - W2*col[8*2] + W4*col[8*4] - W6*col[8*6]; | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
270 |
205 | 271 b0 = W1*col[8*1] + W3*col[8*3] + W5*col[8*5] + W7*col[8*7]; |
272 b1 = W3*col[8*1] - W7*col[8*3] - W1*col[8*5] - W5*col[8*7]; | |
273 b2 = W5*col[8*1] - W1*col[8*3] + W7*col[8*5] + W3*col[8*7]; | |
274 b3 = W7*col[8*1] - W5*col[8*3] + W3*col[8*5] - W1*col[8*7]; | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
275 |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
276 col[8*0] = (a0 + b0) >> COL_SHIFT; |
205 | 277 col[8*7] = (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
|
278 col[8*1] = (a1 + b1) >> COL_SHIFT; |
205 | 279 col[8*6] = (a1 - b1) >> COL_SHIFT; |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
280 col[8*2] = (a2 + b2) >> COL_SHIFT; |
205 | 281 col[8*5] = (a2 - b2) >> COL_SHIFT; |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
282 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
|
283 col[8*4] = (a3 - b3) >> COL_SHIFT; |
205 | 284 } |
285 | |
286 static void inline idctSparseCol (int16_t * col) | |
287 { | |
288 int a0, a1, a2, a3, b0, b1, b2, b3; | |
289 col[0] += (1<<(COL_SHIFT-1))/W4; | |
290 a0 = W4*col[8*0]; | |
291 a1 = W4*col[8*0]; | |
292 a2 = W4*col[8*0]; | |
293 a3 = W4*col[8*0]; | |
294 | |
295 if(col[8*2]){ | |
296 a0 += + W2*col[8*2]; | |
297 a1 += + W6*col[8*2]; | |
298 a2 += - W6*col[8*2]; | |
299 a3 += - W2*col[8*2]; | |
300 } | |
301 | |
302 if(col[8*4]){ | |
303 a0 += + W4*col[8*4]; | |
304 a1 += - W4*col[8*4]; | |
305 a2 += - W4*col[8*4]; | |
306 a3 += + W4*col[8*4]; | |
307 } | |
308 | |
309 if(col[8*6]){ | |
310 a0 += + W6*col[8*6]; | |
311 a1 += - W2*col[8*6]; | |
312 a2 += + W2*col[8*6]; | |
313 a3 += - W6*col[8*6]; | |
314 } | |
315 | |
316 if(col[8*1]){ | |
317 b0 = W1*col[8*1]; | |
318 b1 = W3*col[8*1]; | |
319 b2 = W5*col[8*1]; | |
320 b3 = W7*col[8*1]; | |
321 }else{ | |
322 b0 = | |
323 b1 = | |
324 b2 = | |
325 b3 = 0; | |
326 } | |
327 | |
328 if(col[8*3]){ | |
329 b0 += + W3*col[8*3]; | |
330 b1 += - W7*col[8*3]; | |
331 b2 += - W1*col[8*3]; | |
332 b3 += - W5*col[8*3]; | |
333 } | |
334 | |
335 if(col[8*5]){ | |
336 b0 += + W5*col[8*5]; | |
337 b1 += - W1*col[8*5]; | |
338 b2 += + W7*col[8*5]; | |
339 b3 += + W3*col[8*5]; | |
340 } | |
341 | |
342 if(col[8*7]){ | |
343 b0 += + W7*col[8*7]; | |
344 b1 += - W5*col[8*7]; | |
345 b2 += + W3*col[8*7]; | |
346 b3 += - W1*col[8*7]; | |
347 } | |
348 | |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
349 #ifndef ARCH_ALPHA |
205 | 350 if(!(b0|b1|b2|b3)){ |
351 col[8*0] = (a0) >> COL_SHIFT; | |
352 col[8*7] = (a0) >> COL_SHIFT; | |
353 col[8*1] = (a1) >> COL_SHIFT; | |
354 col[8*6] = (a1) >> COL_SHIFT; | |
355 col[8*2] = (a2) >> COL_SHIFT; | |
356 col[8*5] = (a2) >> COL_SHIFT; | |
357 col[8*3] = (a3) >> COL_SHIFT; | |
358 col[8*4] = (a3) >> COL_SHIFT; | |
359 }else{ | |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
360 #endif |
205 | 361 col[8*0] = (a0 + b0) >> COL_SHIFT; |
362 col[8*7] = (a0 - b0) >> COL_SHIFT; | |
363 col[8*1] = (a1 + b1) >> COL_SHIFT; | |
364 col[8*6] = (a1 - b1) >> COL_SHIFT; | |
365 col[8*2] = (a2 + b2) >> COL_SHIFT; | |
366 col[8*5] = (a2 - b2) >> COL_SHIFT; | |
367 col[8*3] = (a3 + b3) >> COL_SHIFT; | |
368 col[8*4] = (a3 - b3) >> COL_SHIFT; | |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
369 #ifndef ARCH_ALPHA |
205 | 370 } |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
371 #endif |
205 | 372 } |
373 | |
374 static void inline idctSparse2Col (int16_t * col) | |
375 { | |
376 int a0, a1, a2, a3, b0, b1, b2, b3; | |
377 col[0] += (1<<(COL_SHIFT-1))/W4; | |
378 a0 = W4*col[8*0]; | |
379 a1 = W4*col[8*0]; | |
380 a2 = W4*col[8*0]; | |
381 a3 = W4*col[8*0]; | |
382 | |
383 if(col[8*2]){ | |
384 a0 += + W2*col[8*2]; | |
385 a1 += + W6*col[8*2]; | |
386 a2 += - W6*col[8*2]; | |
387 a3 += - W2*col[8*2]; | |
388 } | |
389 | |
390 if(col[8*4]){ | |
391 a0 += + W4*col[8*4]; | |
392 a1 += - W4*col[8*4]; | |
393 a2 += - W4*col[8*4]; | |
394 a3 += + W4*col[8*4]; | |
395 } | |
396 | |
397 if(col[8*6]){ | |
398 a0 += + W6*col[8*6]; | |
399 a1 += - W2*col[8*6]; | |
400 a2 += + W2*col[8*6]; | |
401 a3 += - W6*col[8*6]; | |
402 } | |
403 | |
404 if(col[8*1] || 1){ | |
405 b0 = W1*col[8*1]; | |
406 b1 = W3*col[8*1]; | |
407 b2 = W5*col[8*1]; | |
408 b3 = W7*col[8*1]; | |
409 }else{ | |
410 b0 = | |
411 b1 = | |
412 b2 = | |
413 b3 = 0; | |
414 } | |
415 | |
416 if(col[8*3]){ | |
417 b0 += + W3*col[8*3]; | |
418 b1 += - W7*col[8*3]; | |
419 b2 += - W1*col[8*3]; | |
420 b3 += - W5*col[8*3]; | |
421 } | |
422 | |
423 if(col[8*5]){ | |
424 b0 += + W5*col[8*5]; | |
425 b1 += - W1*col[8*5]; | |
426 b2 += + W7*col[8*5]; | |
427 b3 += + W3*col[8*5]; | |
428 } | |
429 | |
430 if(col[8*7]){ | |
431 b0 += + W7*col[8*7]; | |
432 b1 += - W5*col[8*7]; | |
433 b2 += + W3*col[8*7]; | |
434 b3 += - W1*col[8*7]; | |
435 } | |
436 | |
437 col[8*0] = (a0 + b0) >> COL_SHIFT; | |
438 col[8*7] = (a0 - b0) >> COL_SHIFT; | |
439 col[8*1] = (a1 + b1) >> COL_SHIFT; | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
440 col[8*6] = (a1 - b1) >> COL_SHIFT; |
205 | 441 col[8*2] = (a2 + b2) >> COL_SHIFT; |
442 col[8*5] = (a2 - b2) >> COL_SHIFT; | |
443 col[8*3] = (a3 + b3) >> COL_SHIFT; | |
444 col[8*4] = (a3 - b3) >> COL_SHIFT; | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
445 } |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
446 |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
447 #ifdef ARCH_ALPHA |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
448 /* If all rows but the first one are zero after row transformation, |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
449 all rows will be identical after column transformation. */ |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
450 static inline void idctCol2(int16_t *col) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
451 { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
452 int i; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
453 uint64_t l, r; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
454 uint64_t *lcol = (uint64_t *) col; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
455 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
456 for (i = 0; i < 8; ++i) { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
457 int a0 = col[0] + (1 << (COL_SHIFT - 1)) / W4; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
458 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
459 a0 *= W4; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
460 col[0] = a0 >> COL_SHIFT; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
461 ++col; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
462 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
463 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
464 l = lcol[0]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
465 r = lcol[1]; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
466 lcol[ 2] = l; lcol[ 3] = r; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
467 lcol[ 4] = l; lcol[ 5] = r; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
468 lcol[ 6] = l; lcol[ 7] = r; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
469 lcol[ 8] = l; lcol[ 9] = r; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
470 lcol[10] = l; lcol[11] = r; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
471 lcol[12] = l; lcol[13] = r; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
472 lcol[14] = l; lcol[15] = r; |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
473 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
474 #endif |
205 | 475 |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
476 void simple_idct (short *block) |
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
477 { |
205 | 478 |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
479 int i; |
205 | 480 |
481 #if 0 | |
482 int nonZero[8]; | |
483 int buffer[64]; | |
484 int nNonZero=0; | |
485 | |
486 idctRowCondDC(block); | |
487 | |
488 for(i=1; i<8; i++) | |
489 { | |
490 nonZero[nNonZero]=i; | |
491 nNonZero+= idctRowCondZ(block + i*8); | |
492 } | |
493 | |
494 if(nNonZero==0) | |
495 { | |
496 for(i=0; i<8; i++) | |
497 { | |
498 block[i ]= | |
499 block[i+8 ]= | |
500 block[i+16]= | |
501 block[i+24]= | |
502 block[i+32]= | |
503 block[i+40]= | |
504 block[i+48]= | |
505 block[i+56]= (W4*block[i] + (1<<(COL_SHIFT-1))) >> COL_SHIFT; | |
506 } | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
507 } |
205 | 508 else if(nNonZero==1) |
509 { | |
510 int index= nonZero[0]*8; | |
511 for(i=0; i<8; i++) | |
512 { | |
513 int bias= W4*block[i] + (1<<(COL_SHIFT-1)); | |
514 int c= block[i + index]; | |
515 block[i ]= (c*coeff[index ] + bias) >> COL_SHIFT; | |
516 block[i+8 ]= (c*coeff[index+1] + bias) >> COL_SHIFT; | |
517 block[i+16]= (c*coeff[index+2] + bias) >> COL_SHIFT; | |
518 block[i+24]= (c*coeff[index+3] + bias) >> COL_SHIFT; | |
519 block[i+32]= (c*coeff[index+4] + bias) >> COL_SHIFT; | |
520 block[i+40]= (c*coeff[index+5] + bias) >> COL_SHIFT; | |
521 block[i+48]= (c*coeff[index+6] + bias) >> COL_SHIFT; | |
522 block[i+56]= (c*coeff[index+7] + bias) >> COL_SHIFT; | |
523 } | |
524 } | |
525 /* else if(nNonZero==2) | |
526 { | |
527 int index1= nonZero[0]*8; | |
528 int index2= nonZero[1]*8; | |
529 for(i=0; i<8; i++) | |
530 { | |
531 int bias= W4*block[i] + (1<<(COL_SHIFT-1)); | |
532 int c1= block[i + index1]; | |
533 int c2= block[i + index2]; | |
534 block[i ]= (c1*coeff[index1 ] + c2*coeff[index2 ] + bias) >> COL_SHIFT; | |
535 block[i+8 ]= (c1*coeff[index1+1] + c2*coeff[index2+1] + bias) >> COL_SHIFT; | |
536 block[i+16]= (c1*coeff[index1+2] + c2*coeff[index2+2] + bias) >> COL_SHIFT; | |
537 block[i+24]= (c1*coeff[index1+3] + c2*coeff[index2+3] + bias) >> COL_SHIFT; | |
538 block[i+32]= (c1*coeff[index1+4] + c2*coeff[index2+4] + bias) >> COL_SHIFT; | |
539 block[i+40]= (c1*coeff[index1+5] + c2*coeff[index2+5] + bias) >> COL_SHIFT; | |
540 block[i+48]= (c1*coeff[index1+6] + c2*coeff[index2+6] + bias) >> COL_SHIFT; | |
541 block[i+56]= (c1*coeff[index1+7] + c2*coeff[index2+7] + bias) >> COL_SHIFT; | |
542 } | |
543 }*/ | |
544 else | |
545 { | |
546 for(i=0; i<8; i++) | |
547 idctSparse2Col(block + i); | |
548 } | |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
549 #elif defined(ARCH_ALPHA) |
215
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
550 int rowsZero = 1; /* all rows except row 0 zero */ |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
551 int rowsConstant = 1; /* all rows consist of a constant value */ |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
552 |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
553 for (i = 0; i < 8; i++) { |
215
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
554 int sparseness = idctRowCondDC(block + 8 * i); |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
555 |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
556 if (i > 0 && sparseness > 0) |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
557 rowsZero = 0; |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
558 if (sparseness == 2) |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
559 rowsConstant = 0; |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
560 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
561 |
215
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
562 if (rowsZero) { |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
563 idctCol2(block); |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
564 } else if (rowsConstant) { |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
565 uint64_t *lblock = (uint64_t *) block; |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
566 |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
567 idctSparseCol(block); |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
568 for (i = 0; i < 8; i++) { |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
569 uint64_t v = (uint16_t) block[i * 8]; |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
570 |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
571 v += v << 16; |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
572 v += v << 32; |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
573 lblock[0] = v; |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
574 lblock[1] = v; |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
575 lblock += 2; |
1fe6b64feefb
Small simple idct improvement for Alpha by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
214
diff
changeset
|
576 } |
214
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
577 } else { |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
578 for (i = 0; i < 8; i++) |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
579 idctSparseCol(block + i); |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
580 } |
73df666cacc7
Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents:
205
diff
changeset
|
581 #else |
205 | 582 for(i=0; i<8; i++) |
583 idctRowCondDC(block + i*8); | |
584 | |
585 for(i=0; i<8; i++) | |
586 idctSparseCol(block + i); | |
587 #endif | |
175
bd77d3cbb233
new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable
arpi_esp
parents:
diff
changeset
|
588 } |