annotate dsputil.h @ 884:2cef5c4c0ca6 libavcodec

* altivec and pix_norm patch by Brian Foley
author kabi
date Fri, 22 Nov 2002 07:53:06 +0000
parents 725ef4ea3ecc
children a408778eff87
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
1 /*
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
2 * DSP utils
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard.
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
4 *
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
5 * This library is free software; you can redistribute it and/or
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
6 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
7 * License as published by the Free Software Foundation; either
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
8 * version 2 of the License, or (at your option) any later version.
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
9 *
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
10 * This library is distributed in the hope that it will be useful,
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
13 * Lesser General Public License for more details.
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
14 *
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
15 * You should have received a copy of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
16 * License along with this library; if not, write to the Free Software
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
18 */
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
19 #ifndef DSPUTIL_H
986e461dc072 Initial revision
glantau
parents:
diff changeset
20 #define DSPUTIL_H
986e461dc072 Initial revision
glantau
parents:
diff changeset
21
986e461dc072 Initial revision
glantau
parents:
diff changeset
22 #include "common.h"
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 214
diff changeset
23 #include "avcodec.h"
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
24
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
25 //#define DEBUG
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
26 /* dct code */
986e461dc072 Initial revision
glantau
parents:
diff changeset
27 typedef short DCTELEM;
986e461dc072 Initial revision
glantau
parents:
diff changeset
28
474
11dbd00682fc avoid name clash with libjpeg - added missing externs
bellard
parents: 429
diff changeset
29 void fdct_ifast (DCTELEM *data);
625
bb6a69f9d409 slow but accurate integer dct from IJG (should be ok with the LGPL as the old DCT is the fast integer DCT from IJG)
michaelni
parents: 623
diff changeset
30 void ff_jpeg_fdct_islow (DCTELEM *data);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
31
986e461dc072 Initial revision
glantau
parents:
diff changeset
32 void j_rev_dct (DCTELEM *data);
986e461dc072 Initial revision
glantau
parents:
diff changeset
33
687
9abb13c21fbe fdct_mmx -> ff_fdct_mmx (renamed to avoid namespace conflict with xvid)
arpi_esp
parents: 675
diff changeset
34 void ff_fdct_mmx(DCTELEM *block);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
35
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
36 /* encoding scans */
706
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
37 extern const UINT8 ff_alternate_horizontal_scan[64];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
38 extern const UINT8 ff_alternate_vertical_scan[64];
e65798d228ea idct permutation cleanup, idct can be selected per context now
michaelni
parents: 689
diff changeset
39 extern const UINT8 ff_zigzag_direct[64];
190
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
40
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
41 /* pixel operations */
986e461dc072 Initial revision
glantau
parents:
diff changeset
42 #define MAX_NEG_CROP 384
986e461dc072 Initial revision
glantau
parents:
diff changeset
43
986e461dc072 Initial revision
glantau
parents:
diff changeset
44 /* temporary */
986e461dc072 Initial revision
glantau
parents:
diff changeset
45 extern UINT32 squareTbl[512];
50
4ea4c10d03d8 export for imgconvert usage
glantau
parents: 34
diff changeset
46 extern UINT8 cropTbl[256 + 2 * MAX_NEG_CROP];
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
47
986e461dc072 Initial revision
glantau
parents:
diff changeset
48
675
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
49 /* minimum alignment rules ;)
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
50 if u notice errors in the align stuff, need more alignment for some asm code for some cpu
675
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
51 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
52
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
53 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
54 i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
55
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
56 !future video codecs might need functions with less strict alignment
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
57 */
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
58
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
59 /*
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
60 void get_pixels_c(DCTELEM *block, const UINT8 *pixels, int line_size);
324
9c6f056f0e41 fixed mpeg4 time stuff on encoding
michaelni
parents: 296
diff changeset
61 void diff_pixels_c(DCTELEM *block, const UINT8 *s1, const UINT8 *s2, int stride);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
62 void put_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size);
986e461dc072 Initial revision
glantau
parents:
diff changeset
63 void add_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size);
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
64 void clear_blocks_c(DCTELEM *blocks);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
65 */
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
66
986e461dc072 Initial revision
glantau
parents:
diff changeset
67 /* add and put pixel (decoding) */
675
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
68 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
69 typedef void (*op_pixels_func)(UINT8 *block/*align width (8 or 16)*/, const UINT8 *pixels/*align 1*/, int line_size, int h);
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
70 typedef void (*qpel_mc_func)(UINT8 *dst/*align width (8 or 16)*/, UINT8 *src/*align 1*/, int stride);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
71
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
72
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
73
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
74 #define CALL_2X_PIXELS(a, b, n)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
75 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
76 b(block , pixels , line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
77 b(block+n, pixels+n, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
78 }
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
79
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
80 /* motion estimation */
986e461dc072 Initial revision
glantau
parents:
diff changeset
81
675
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
82 typedef int (*op_pixels_abs_func)(UINT8 *blk1/*align width (8 or 16)*/, UINT8 *blk2/*align 1*/, int line_size);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
83 /*
294
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 255
diff changeset
84 int pix_abs16x16_c(UINT8 *blk1, UINT8 *blk2, int lx);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 255
diff changeset
85 int pix_abs16x16_x2_c(UINT8 *blk1, UINT8 *blk2, int lx);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 255
diff changeset
86 int pix_abs16x16_y2_c(UINT8 *blk1, UINT8 *blk2, int lx);
944632089814 4MV motion estimation (not finished yet)
michaelni
parents: 255
diff changeset
87 int pix_abs16x16_xy2_c(UINT8 *blk1, UINT8 *blk2, int lx);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
88 */
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
89 typedef struct DSPContext {
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
90 /* pixel ops : interface with DCT */
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
91 void (*get_pixels)(DCTELEM *block/*align 16*/, const UINT8 *pixels/*align 8*/, int line_size);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
92 void (*diff_pixels)(DCTELEM *block/*align 16*/, const UINT8 *s1/*align 8*/, const UINT8 *s2/*align 8*/, int stride);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
93 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, UINT8 *pixels/*align 8*/, int line_size);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
94 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, UINT8 *pixels/*align 8*/, int line_size);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
95 void (*gmc1)(UINT8 *dst/*align 8*/, UINT8 *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
96 void (*gmc )(UINT8 *dst/*align 8*/, UINT8 *src/*align 1*/, int stride, int h, int ox, int oy,
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
97 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
98 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
99 int (*pix_sum)(UINT8 * pix, int line_size);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
100 int (*pix_norm1)(UINT8 * pix, int line_size);
884
2cef5c4c0ca6 * altivec and pix_norm patch by Brian Foley
kabi
parents: 866
diff changeset
101 int (*pix_norm)(UINT8 * pix1, UINT8 * pix2, int line_size);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
102
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
103 /* maybe create an array for 16/8 functions */
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
104 op_pixels_func put_pixels_tab[2][4];
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
105 op_pixels_func avg_pixels_tab[2][4];
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
106 op_pixels_func put_no_rnd_pixels_tab[2][4];
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
107 op_pixels_func avg_no_rnd_pixels_tab[2][4];
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
108 qpel_mc_func put_qpel_pixels_tab[2][16];
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
109 qpel_mc_func avg_qpel_pixels_tab[2][16];
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
110 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
111 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
112
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
113 op_pixels_abs_func pix_abs16x16;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
114 op_pixels_abs_func pix_abs16x16_x2;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
115 op_pixels_abs_func pix_abs16x16_y2;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
116 op_pixels_abs_func pix_abs16x16_xy2;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
117 op_pixels_abs_func pix_abs8x8;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
118 op_pixels_abs_func pix_abs8x8_x2;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
119 op_pixels_abs_func pix_abs8x8_y2;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
120 op_pixels_abs_func pix_abs8x8_xy2;
866
725ef4ea3ecc huffyuv
michaelni
parents: 862
diff changeset
121
725ef4ea3ecc huffyuv
michaelni
parents: 862
diff changeset
122 /* huffyuv specific */
725ef4ea3ecc huffyuv
michaelni
parents: 862
diff changeset
123 //FIXME note: alignment isnt guranteed currently but could be if needed
725ef4ea3ecc huffyuv
michaelni
parents: 862
diff changeset
124 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
725ef4ea3ecc huffyuv
michaelni
parents: 862
diff changeset
125 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 16*/,int w);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
126 } DSPContext;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
127
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
128 void dsputil_init(DSPContext* p, unsigned mask);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
129
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
130 /**
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
131 * permute block according to permuatation.
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
132 * @param last last non zero element in scantable order
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
133 */
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
134 void ff_block_permute(INT16 *block, UINT8 *permutation, const UINT8 *scantable, int last);
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
135
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
136 #define emms_c()
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
137
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
138 #if defined(HAVE_MMX)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
139
862
058194d7ade6 * fixing some minor const warnings
kabi
parents: 853
diff changeset
140 #undef emms_c
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
141
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
142 #define MM_MMX 0x0001 /* standard MMX */
986e461dc072 Initial revision
glantau
parents:
diff changeset
143 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
986e461dc072 Initial revision
glantau
parents:
diff changeset
144 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
986e461dc072 Initial revision
glantau
parents:
diff changeset
145 #define MM_SSE 0x0008 /* SSE functions */
986e461dc072 Initial revision
glantau
parents:
diff changeset
146 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
986e461dc072 Initial revision
glantau
parents:
diff changeset
147
986e461dc072 Initial revision
glantau
parents:
diff changeset
148 extern int mm_flags;
986e461dc072 Initial revision
glantau
parents:
diff changeset
149
986e461dc072 Initial revision
glantau
parents:
diff changeset
150 int mm_support(void);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
151 void add_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line_size);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
152 void put_pixels_clamped_mmx(const DCTELEM *block, UINT8 *pixels, int line_size);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
153
986e461dc072 Initial revision
glantau
parents:
diff changeset
154 static inline void emms(void)
986e461dc072 Initial revision
glantau
parents:
diff changeset
155 {
6
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
156 __asm __volatile ("emms;":::"memory");
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
157 }
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
158
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
159 #define emms_c() \
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
160 {\
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
161 if (mm_flags & MM_MMX)\
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
162 emms();\
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
163 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
164
986e461dc072 Initial revision
glantau
parents:
diff changeset
165 #define __align8 __attribute__ ((aligned (8)))
986e461dc072 Initial revision
glantau
parents:
diff changeset
166
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
167 void dsputil_init_mmx(DSPContext* c, unsigned mask);
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
168 void dsputil_set_bit_exact_mmx(DSPContext* c, unsigned mask);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
169
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
170 #elif defined(ARCH_ARMV4L)
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
171
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
172 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
173 line ptimizations */
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
174 #define __align8 __attribute__ ((aligned (4)))
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
175
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
176 void dsputil_init_armv4l(DSPContext* c, unsigned mask);
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
177
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
178 #elif defined(HAVE_MLIB)
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
179
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
180 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
181 #define __align8 __attribute__ ((aligned (8)))
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
182
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
183 void dsputil_init_mlib(DSPContext* c, unsigned mask);
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
184
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
185 #elif defined(ARCH_ALPHA)
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
186
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
187 #define __align8 __attribute__ ((aligned (8)))
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
188
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
189 void dsputil_init_alpha(DSPContext* c, unsigned mask);
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
190
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 612
diff changeset
191 #elif defined(ARCH_POWERPC)
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 612
diff changeset
192
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 612
diff changeset
193 #define __align8 __attribute__ ((aligned (16)))
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 612
diff changeset
194
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
195 void dsputil_init_ppc(DSPContext* c, unsigned mask);
623
92e99e506920 first cut at altivec support on darwin patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 612
diff changeset
196
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
197 #elif defined(HAVE_MMI)
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
198
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
199 #define __align8 __attribute__ ((aligned (16)))
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
200
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
201 void dsputil_init_mmi(DSPContext* c, unsigned mask);
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
202
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
203 #else
986e461dc072 Initial revision
glantau
parents:
diff changeset
204
986e461dc072 Initial revision
glantau
parents:
diff changeset
205 #define __align8
986e461dc072 Initial revision
glantau
parents:
diff changeset
206
986e461dc072 Initial revision
glantau
parents:
diff changeset
207 #endif
986e461dc072 Initial revision
glantau
parents:
diff changeset
208
493
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
209 #ifdef __GNUC__
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
210
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
211 struct unaligned_64 { uint64_t l; } __attribute__((packed));
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
212 struct unaligned_32 { uint32_t l; } __attribute__((packed));
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
213
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
214 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
215 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
216
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
217 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
218
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
219 #else /* __GNUC__ */
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
220
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
221 #define LD32(a) (*((uint32_t*)(a)))
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
222 #define LD64(a) (*((uint64_t*)(a)))
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
223
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
224 #define ST32(a, b) *((uint32_t*)(a)) = (b)
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
225
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
226 #endif /* !__GNUC__ */
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
227
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 214
diff changeset
228 /* PSNR */
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 214
diff changeset
229 void get_psnr(UINT8 *orig_image[3], UINT8 *coded_image[3],
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 214
diff changeset
230 int orig_linesize[3], int coded_linesize,
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 214
diff changeset
231 AVCodecContext *avctx);
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
232
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
233 /* FFT computation */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
234
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
235 /* NOTE: soon integer code will be added, so you must use the
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
236 FFTSample type */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
237 typedef float FFTSample;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
238
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
239 typedef struct FFTComplex {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
240 FFTSample re, im;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
241 } FFTComplex;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
242
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
243 typedef struct FFTContext {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
244 int nbits;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
245 int inverse;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
246 uint16_t *revtab;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
247 FFTComplex *exptab;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
248 FFTComplex *exptab1; /* only used by SSE code */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
249 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
250 } FFTContext;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
251
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
252 int fft_init(FFTContext *s, int nbits, int inverse);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
253 void fft_permute(FFTContext *s, FFTComplex *z);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
254 void fft_calc_c(FFTContext *s, FFTComplex *z);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
255 void fft_calc_sse(FFTContext *s, FFTComplex *z);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
256 static inline void fft_calc(FFTContext *s, FFTComplex *z)
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
257 {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
258 s->fft_calc(s, z);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
259 }
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
260 void fft_end(FFTContext *s);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
261
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
262 /* MDCT computation */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
263
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
264 typedef struct MDCTContext {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
265 int n; /* size of MDCT (i.e. number of input data * 2) */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
266 int nbits; /* n = 2^nbits */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
267 /* pre/post rotation tables */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
268 FFTSample *tcos;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
269 FFTSample *tsin;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
270 FFTContext fft;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
271 } MDCTContext;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
272
794
670009af4fc2 avoid name clash
bellard
parents: 781
diff changeset
273 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
274 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
275 const FFTSample *input, FFTSample *tmp);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
276 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
277 const FFTSample *input, FFTSample *tmp);
794
670009af4fc2 avoid name clash
bellard
parents: 781
diff changeset
278 void ff_mdct_end(MDCTContext *s);
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
279
838
b78812db886f lrintf detection (based upon a patch by Franois Revol <revol at free dot fr>)
michaelni
parents: 802
diff changeset
280 #ifndef HAVE_LRINTF
796
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
281 /* XXX: add ISOC specific test to avoid specific BSD testing. */
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
282 /* better than nothing implementation. */
802
99db75050d38 NetBSD fix by Bernd Ernesti <mplayer@lists.veego.de>
al3x
parents: 796
diff changeset
283 /* btw, rintf() is existing on fbsd too -- alex */
796
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
284 static inline long int lrintf(float x)
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
285 {
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
286 return (int)(rint(x));
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
287 }
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
288 #endif
796
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
289
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
290 #endif