annotate dsputil.h @ 2184:3378d0677903 libavcodec

4x4 SSE compare function wavelet based compare functions make epzs_motion_search() more flexible so it can be used for a wider range of block sizes make get_penalty_factor() independant of MpegEncContext
author michael
date Sun, 22 Aug 2004 17:16:03 +0000
parents 869805505b30
children 784c38a3ffc2
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.
1739
07a484280a82 copyright year update of the files i touched and remembered, things look annoyingly unmaintained otherwise
michael
parents: 1729
diff changeset
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
429
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
5 *
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
6 * This library is free software; you can redistribute it and/or
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
8 * License as published by the Free Software Foundation; either
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
9 * version 2 of the License, or (at your option) any later version.
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
10 *
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
11 * This library is distributed in the hope that it will be useful,
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
14 * Lesser General Public License for more details.
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
15 *
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
17 * License along with this library; if not, write to the Free Software
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
718a22dc121f license/copyright change
glantau
parents: 403
diff changeset
19 */
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
20
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
21 /**
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
22 * @file dsputil.h
1106
1e39f273ecd6 per file doxy
michaelni
parents: 1104
diff changeset
23 * DSP utils.
1213
a71fe70b428b doc fixes (forgot to commit)
michaelni
parents: 1201
diff changeset
24 * note, many functions in here may use MMX which trashes the FPU state, it is
a71fe70b428b doc fixes (forgot to commit)
michaelni
parents: 1201
diff changeset
25 * absolutely necessary to call emms_c() between dsp & float/double code
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
26 */
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
27
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
28 #ifndef DSPUTIL_H
986e461dc072 Initial revision
glantau
parents:
diff changeset
29 #define DSPUTIL_H
986e461dc072 Initial revision
glantau
parents:
diff changeset
30
986e461dc072 Initial revision
glantau
parents:
diff changeset
31 #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
32 #include "avcodec.h"
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
33
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
34
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
35 //#define DEBUG
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
36 /* dct code */
986e461dc072 Initial revision
glantau
parents:
diff changeset
37 typedef short DCTELEM;
986e461dc072 Initial revision
glantau
parents:
diff changeset
38
474
11dbd00682fc avoid name clash with libjpeg - added missing externs
bellard
parents: 429
diff changeset
39 void fdct_ifast (DCTELEM *data);
1571
aa4dc16c0f18 * adding integer/floating point AAN implementations for DCT 2-4-8
romansh
parents: 1567
diff changeset
40 void fdct_ifast248 (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
41 void ff_jpeg_fdct_islow (DCTELEM *data);
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1565
diff changeset
42 void ff_fdct248_islow (DCTELEM *data);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
43
986e461dc072 Initial revision
glantau
parents:
diff changeset
44 void j_rev_dct (DCTELEM *data);
986e461dc072 Initial revision
glantau
parents:
diff changeset
45
687
9abb13c21fbe fdct_mmx -> ff_fdct_mmx (renamed to avoid namespace conflict with xvid)
arpi_esp
parents: 675
diff changeset
46 void ff_fdct_mmx(DCTELEM *block);
1565
1a9a63f59849 minor mmx2 optimization if the dct
michael
parents: 1527
diff changeset
47 void ff_fdct_mmx2(DCTELEM *block);
1765
e31754bc5b65 SSE2 fdct by (Balatoni Denes <pnis at coder dot hu>)
michael
parents: 1739
diff changeset
48 void ff_fdct_sse2(DCTELEM *block);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
49
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
50 /* encoding scans */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
51 extern const uint8_t ff_alternate_horizontal_scan[64];
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
52 extern const uint8_t ff_alternate_vertical_scan[64];
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
53 extern const uint8_t ff_zigzag_direct[64];
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1565
diff changeset
54 extern const uint8_t ff_zigzag248_direct[64];
190
9e0e56869d05 fix for non-mmx runtimedetect encoding bugs - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents: 174
diff changeset
55
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
56 /* pixel operations */
2090
869805505b30 increase size of crop table
michael
parents: 2075
diff changeset
57 #define MAX_NEG_CROP 1024
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
58
986e461dc072 Initial revision
glantau
parents:
diff changeset
59 /* temporary */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
60 extern uint32_t squareTbl[512];
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
61 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
62
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
63 /* VP3 DSP functions */
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
64 void vp3_dsp_init_c(void);
1977
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
65 void vp3_idct_c(int16_t *input_data, int16_t *dequant_matrix,
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
66 int coeff_count, DCTELEM *output_data);
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
67
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
68 void vp3_dsp_init_mmx(void);
1977
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
69 void vp3_idct_mmx(int16_t *input_data, int16_t *dequant_matrix,
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
70 int coeff_count, DCTELEM *output_data);
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
71
1972
8a556283601d hook up support for SSE2-optimized VP3 IDCT
melanson
parents: 1959
diff changeset
72 void vp3_dsp_init_sse2(void);
1977
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
73 void vp3_idct_sse2(int16_t *input_data, int16_t *dequant_matrix,
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
74 int coeff_count, DCTELEM *output_data);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
75
675
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
76 /* minimum alignment rules ;)
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
77 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
78 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
79
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
80 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
81 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
82
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
83 !future video codecs might need functions with less strict alignment
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
84 */
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
85
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
86 /*
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
87 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
88 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
89 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
90 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
296
c1a8a1b4a24b sizeof(s->block) isnt 64*6*2 anymore bugfix
michaelni
parents: 294
diff changeset
91 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
92 */
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
93
986e461dc072 Initial revision
glantau
parents:
diff changeset
94 /* add and put pixel (decoding) */
675
9008cbe229ce alignment info
michaelni
parents: 651
diff changeset
95 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
1709
39a209b0c82c info about h and tpel
michael
parents: 1708
diff changeset
96 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
97 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
98 typedef void (*tpel_mc_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int w, int h);
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
99 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
100 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
101
984
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
102 #define DEF_OLD_QPEL(name)\
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
103 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
104 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
105 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
106
984
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
107 DEF_OLD_QPEL(qpel16_mc11_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
108 DEF_OLD_QPEL(qpel16_mc31_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
109 DEF_OLD_QPEL(qpel16_mc12_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
110 DEF_OLD_QPEL(qpel16_mc32_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
111 DEF_OLD_QPEL(qpel16_mc13_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
112 DEF_OLD_QPEL(qpel16_mc33_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
113 DEF_OLD_QPEL(qpel8_mc11_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
114 DEF_OLD_QPEL(qpel8_mc31_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
115 DEF_OLD_QPEL(qpel8_mc12_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
116 DEF_OLD_QPEL(qpel8_mc32_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
117 DEF_OLD_QPEL(qpel8_mc13_old_c)
e162c09efbe7 qpel fix
michaelni
parents: 975
diff changeset
118 DEF_OLD_QPEL(qpel8_mc33_old_c)
651
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
119
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
120 #define CALL_2X_PIXELS(a, b, n)\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
121 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
122 b(block , pixels , line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
123 b(block+n, pixels+n, line_size, h);\
45e8f39fda50 put/avg_pixels16
michaelni
parents: 638
diff changeset
124 }
255
db20b987c32d divx5-gmc support
michaelni
parents: 252
diff changeset
125
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
126 /* motion estimation */
1709
39a209b0c82c info about h and tpel
michael
parents: 1708
diff changeset
127 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
39a209b0c82c info about h and tpel
michael
parents: 1708
diff changeset
128 // allthough currently h<4 is not used as functions with width <8 are not used and neither implemented
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
129 typedef int (*me_cmp_func)(void /*MpegEncContext*/ *s, uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size, int h)/* __attribute__ ((const))*/;
936
caa77cd960c0 qpel encoding
michaelni
parents: 894
diff changeset
130
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
131
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
132 /**
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
133 * DSPContext.
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
134 */
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
135 typedef struct DSPContext {
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
136 /* pixel ops : interface with DCT */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
137 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
138 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
139 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
1984
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
140 void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
141 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
142 /**
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
143 * translational global motion compensation.
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
144 */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
145 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
146 /**
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
147 * global motion compensation.
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
148 */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
149 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
150 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
151 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
152 int (*pix_sum)(uint8_t * pix, int line_size);
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
153 int (*pix_norm1)(uint8_t * pix, int line_size);
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
154 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
155
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
156 me_cmp_func sad[5]; /* identical to pix_absAxA except additional void * */
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
157 me_cmp_func sse[5];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
158 me_cmp_func hadamard8_diff[5];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
159 me_cmp_func dct_sad[5];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
160 me_cmp_func quant_psnr[5];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
161 me_cmp_func bit[5];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
162 me_cmp_func rd[5];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
163 me_cmp_func vsad[5];
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
164 me_cmp_func vsse[5];
2065
9e4bebc39ade noise preserving sum of squares comparission function
michael
parents: 2045
diff changeset
165 me_cmp_func nsse[5];
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
166 me_cmp_func w53[5];
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
167 me_cmp_func w97[5];
936
caa77cd960c0 qpel encoding
michaelni
parents: 894
diff changeset
168
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
169 me_cmp_func me_pre_cmp[5];
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
170 me_cmp_func me_cmp[5];
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
171 me_cmp_func me_sub_cmp[5];
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
172 me_cmp_func mb_cmp[5];
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
173 me_cmp_func ildct_cmp[5]; //only width 16 used
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
174
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
175 /**
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
176 * Halfpel motion compensation with rounding (a+b+1)>>1.
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
177 * this is an array[4][4] of motion compensation funcions for 4
1713
49e46374baa4 put_pixels width comment fix
michael
parents: 1709
diff changeset
178 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
1213
a71fe70b428b doc fixes (forgot to commit)
michaelni
parents: 1201
diff changeset
179 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
180 * @param block destination where the result is stored
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
181 * @param pixels source
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
182 * @param line_size number of bytes in a horizontal line of block
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
183 * @param h height
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
184 */
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
185 op_pixels_func put_pixels_tab[4][4];
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
186
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
187 /**
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
188 * Halfpel motion compensation with rounding (a+b+1)>>1.
1320
7979b98470c8 4, not 2
tmmm
parents: 1319
diff changeset
189 * This is an array[4][4] of motion compensation functions for 4
1713
49e46374baa4 put_pixels width comment fix
michael
parents: 1709
diff changeset
190 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
1213
a71fe70b428b doc fixes (forgot to commit)
michaelni
parents: 1201
diff changeset
191 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
192 * @param block destination into which the result is averaged (a+b+1)>>1
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
193 * @param pixels source
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
194 * @param line_size number of bytes in a horizontal line of block
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
195 * @param h height
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
196 */
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
197 op_pixels_func avg_pixels_tab[4][4];
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
198
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
199 /**
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
200 * Halfpel motion compensation with no rounding (a+b)>>1.
1225
michaelni
parents: 1213
diff changeset
201 * this is an array[2][4] of motion compensation funcions for 2
michaelni
parents: 1213
diff changeset
202 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
1213
a71fe70b428b doc fixes (forgot to commit)
michaelni
parents: 1201
diff changeset
203 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
204 * @param block destination where the result is stored
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
205 * @param pixels source
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
206 * @param line_size number of bytes in a horizontal line of block
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
207 * @param h height
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
208 */
2075
aae3b4e66560 chroma me fix
michael
parents: 2065
diff changeset
209 op_pixels_func put_no_rnd_pixels_tab[4][4];
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
210
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
211 /**
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
212 * Halfpel motion compensation with no rounding (a+b)>>1.
1225
michaelni
parents: 1213
diff changeset
213 * this is an array[2][4] of motion compensation funcions for 2
michaelni
parents: 1213
diff changeset
214 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
1213
a71fe70b428b doc fixes (forgot to commit)
michaelni
parents: 1201
diff changeset
215 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
1136
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
216 * @param block destination into which the result is averaged (a+b)>>1
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
217 * @param pixels source
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
218 * @param line_size number of bytes in a horizontal line of block
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
219 * @param h height
9eb49e9e8dd9 halfpel MC doxy
michaelni
parents: 1111
diff changeset
220 */
2075
aae3b4e66560 chroma me fix
michael
parents: 2065
diff changeset
221 op_pixels_func avg_no_rnd_pixels_tab[4][4];
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
222
1864
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
223 void (*put_no_rnd_pixels_l2[2])(uint8_t *block/*align width (8 or 16)*/, const uint8_t *a/*align 1*/, const uint8_t *b/*align 1*/, int line_size, int h);
9a73e6dc5d68 fix motion compensation with (x+1/2,y+1/2) MVs
michael
parents: 1784
diff changeset
224
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
225 /**
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
226 * Thirdpel motion compensation with rounding (a+b+1)>>1.
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
227 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
228 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
229 * @param block destination where the result is stored
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
230 * @param pixels source
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
231 * @param line_size number of bytes in a horizontal line of block
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
232 * @param h height
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
233 */
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
234 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
1319
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
235 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
449f6e32b425 added support for B-frames and multiple slices
tmmm
parents: 1273
diff changeset
236
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
237 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
238 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
239 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
240 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
936
caa77cd960c0 qpel encoding
michaelni
parents: 894
diff changeset
241 qpel_mc_func put_mspel_pixels_tab[8];
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
242
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
243 /**
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
244 * h264 Chram MC
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
245 */
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
246 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
247 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
248
1168
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
249 qpel_mc_func put_h264_qpel_pixels_tab[3][16];
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
250 qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
5af9aeadbdc3 H264 decoder & demuxer
michaelni
parents: 1136
diff changeset
251
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
252 me_cmp_func pix_abs[2][4];
866
725ef4ea3ecc huffyuv
michaelni
parents: 862
diff changeset
253
725ef4ea3ecc huffyuv
michaelni
parents: 862
diff changeset
254 /* huffyuv specific */
725ef4ea3ecc huffyuv
michaelni
parents: 862
diff changeset
255 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
936
caa77cd960c0 qpel encoding
michaelni
parents: 894
diff changeset
256 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
1527
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1446
diff changeset
257 /**
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1446
diff changeset
258 * subtract huffyuv's variant of median prediction
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1446
diff changeset
259 * note, this might read from src1[-1], src2[-1]
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1446
diff changeset
260 */
8ffd0c00e6df mmx2 optimization of huffyuv median encoding
michael
parents: 1446
diff changeset
261 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
1273
a979fab41ed8 ASV1 codec
michaelni
parents: 1267
diff changeset
262 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
263
1644
835cf346975e h263 loop filter
michael
parents: 1603
diff changeset
264 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
835cf346975e h263 loop filter
michael
parents: 1603
diff changeset
265 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
835cf346975e h263 loop filter
michael
parents: 1603
diff changeset
266
2045
9447bbd8a7e9 rewrite h261 loop filter
michael
parents: 2044
diff changeset
267 void (*h261_loop_filter)(uint8_t *src, int stride);
2044
b6f2add2511e h261 decoder by (Maarten Daniels <maarten.daniels at student dot luc dot ac dot be>)
michael
parents: 2024
diff changeset
268
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
269 /* (I)DCT */
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
270 void (*fdct)(DCTELEM *block/* align 16*/);
1567
e08df4d22d27 * introducing dct248 into the DSP context.
romansh
parents: 1565
diff changeset
271 void (*fdct248)(DCTELEM *block/* align 16*/);
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
272
1324
7d328fd9d8a5 the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents: 1320
diff changeset
273 /* IDCT really*/
7d328fd9d8a5 the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents: 1320
diff changeset
274 void (*idct)(DCTELEM *block/* align 16*/);
7d328fd9d8a5 the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents: 1320
diff changeset
275
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
276 /**
1104
6a01ffc5351c cosmetics
michaelni
parents: 1102
diff changeset
277 * block -> idct -> clip to unsigned 8 bit -> dest.
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
278 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
1111
michaelni
parents: 1106
diff changeset
279 * @param line_size size in bytes of a horizotal line of dest
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
280 */
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
281 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
282
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
283 /**
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
284 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
1111
michaelni
parents: 1106
diff changeset
285 * @param line_size size in bytes of a horizotal line of dest
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
286 */
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
287 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
288
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
289 /**
1104
6a01ffc5351c cosmetics
michaelni
parents: 1102
diff changeset
290 * idct input permutation.
1241
6a873cad413e more idct_permute docs
michaelni
parents: 1225
diff changeset
291 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
6a873cad413e more idct_permute docs
michaelni
parents: 1225
diff changeset
292 * IDCT)
6a873cad413e more idct_permute docs
michaelni
parents: 1225
diff changeset
293 * this permutation must be performed before the idct_put/add, note, normally this can be merged
6a873cad413e more idct_permute docs
michaelni
parents: 1225
diff changeset
294 * with the zigzag/alternate scan<br>
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
295 * an example to avoid confusion:
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
296 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
297 * - (x -> referece dct -> reference idct -> x)
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
298 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
299 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
300 */
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
301 uint8_t idct_permutation[64];
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
302 int idct_permutation_type;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
303 #define FF_NO_IDCT_PERM 1
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
304 #define FF_LIBMPEG2_IDCT_PERM 2
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
305 #define FF_SIMPLE_IDCT_PERM 3
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
306 #define FF_TRANSPOSE_IDCT_PERM 4
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
307
1784
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1765
diff changeset
308 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1765
diff changeset
309 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1765
diff changeset
310 #define BASIS_SHIFT 16
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1765
diff changeset
311 #define RECON_SHIFT 6
65f7bd09f37b quantizer noise shaping optimization
michael
parents: 1765
diff changeset
312
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
313 /**
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
314 * This function handles any initialization for the VP3 DSP functions.
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
315 */
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
316 void (*vp3_dsp_init)(void);
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
317
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
318 /**
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
319 * This function is responsible for taking a block of zigzag'd,
1977
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
320 * quantized DCT coefficients and reconstructing the original block of
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
321 * samples.
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
322 * @param input_data 64 zigzag'd, quantized DCT coefficients
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
323 * @param dequant_matrix 64 zigzag'd quantizer coefficients
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
324 * @param coeff_count index of the last coefficient
1977
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
325 * @param output_samples space for 64 DCTELEMs where the transformed
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
326 * samples will be stored
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
327 */
1977
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
328 void (*vp3_idct)(int16_t *input_data, int16_t *dequant_matrix,
89422281f6f6 reorganize and simplify the VP3 IDCT stuff
melanson
parents: 1974
diff changeset
329 int coeff_count, DCTELEM *output_samples);
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents: 1864
diff changeset
330
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
331 } DSPContext;
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
332
1201
e0fc95a6eb4e fixed static init
bellard
parents: 1168
diff changeset
333 void dsputil_static_init(void);
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
334 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
335
764
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
336 /**
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
337 * permute block according to permuatation.
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
338 * @param last last non zero element in scantable order
d4726182dfd2 optimize block_permute()
michaelni
parents: 753
diff changeset
339 */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
340 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
34
2733a4c1c693 added block permutation functions
glantau
parents: 19
diff changeset
341
1729
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
342 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
a4a5e7521339 interlaced dct decision cleanup
michael
parents: 1713
diff changeset
343
1264
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
344 #define BYTE_VEC32(c) ((c)*0x01010101UL)
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
345
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
346 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
347 {
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
348 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
349 }
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
350
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
351 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
352 {
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
353 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
354 }
2fa34e615c76 cleanup
michaelni
parents: 1259
diff changeset
355
2184
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
356 static inline int get_penalty_factor(int lambda, int lambda2, int type){
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
357 switch(type&0xFF){
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
358 default:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
359 case FF_CMP_SAD:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
360 return lambda>>FF_LAMBDA_SHIFT;
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
361 case FF_CMP_DCT:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
362 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
363 case FF_CMP_W53:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
364 return (4*lambda)>>(FF_LAMBDA_SHIFT);
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
365 case FF_CMP_W97:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
366 return (2*lambda)>>(FF_LAMBDA_SHIFT);
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
367 case FF_CMP_SATD:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
368 return (2*lambda)>>FF_LAMBDA_SHIFT;
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
369 case FF_CMP_RD:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
370 case FF_CMP_PSNR:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
371 case FF_CMP_SSE:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
372 case FF_CMP_NSSE:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
373 return lambda2>>FF_LAMBDA_SHIFT;
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
374 case FF_CMP_BIT:
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
375 return 1;
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
376 }
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
377 }
3378d0677903 4x4 SSE compare function
michael
parents: 2090
diff changeset
378
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
379 /**
1104
6a01ffc5351c cosmetics
michaelni
parents: 1102
diff changeset
380 * Empty mmx state.
1102
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
381 * this must be called between any dsp function and float/double code.
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
382 * for example sin(); dsp->idct_put(); emms_c(); cos()
eeaa1e837efb more dox
michaelni
parents: 1092
diff changeset
383 */
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
384 #define emms_c()
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
385
995
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 984
diff changeset
386 /* should be defined by architectures supporting
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 984
diff changeset
387 one or more MultiMedia extension */
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 984
diff changeset
388 int mm_support(void);
edc10966b081 altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents: 984
diff changeset
389
1974
8c5489b2cf3e move __align16 some place where non-MMX machines can see it
melanson
parents: 1972
diff changeset
390 #define __align16 __attribute__ ((aligned (16)))
8c5489b2cf3e move __align16 some place where non-MMX machines can see it
melanson
parents: 1972
diff changeset
391
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
392 #if defined(HAVE_MMX)
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
393
862
058194d7ade6 * fixing some minor const warnings
kabi
parents: 853
diff changeset
394 #undef emms_c
853
eacc2dd8fd9d * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 838
diff changeset
395
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
396 #define MM_MMX 0x0001 /* standard MMX */
986e461dc072 Initial revision
glantau
parents:
diff changeset
397 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
986e461dc072 Initial revision
glantau
parents:
diff changeset
398 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
986e461dc072 Initial revision
glantau
parents:
diff changeset
399 #define MM_SSE 0x0008 /* SSE functions */
986e461dc072 Initial revision
glantau
parents:
diff changeset
400 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
986e461dc072 Initial revision
glantau
parents:
diff changeset
401
986e461dc072 Initial revision
glantau
parents:
diff changeset
402 extern int mm_flags;
986e461dc072 Initial revision
glantau
parents:
diff changeset
403
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
404 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
405 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
1984
ef919e9ef73e separate out put_signed_pixels_clamped() into its own function and
melanson
parents: 1977
diff changeset
406 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
407
986e461dc072 Initial revision
glantau
parents:
diff changeset
408 static inline void emms(void)
986e461dc072 Initial revision
glantau
parents:
diff changeset
409 {
6
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
410 __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
411 }
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
412
936
caa77cd960c0 qpel encoding
michaelni
parents: 894
diff changeset
413
6
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
414 #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
415 {\
ec4642daa6fe added emms_c() macro which should can used in c code in both mmx/non mmx cases
glantau
parents: 2
diff changeset
416 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
417 emms();\
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
418 }
986e461dc072 Initial revision
glantau
parents:
diff changeset
419
986e461dc072 Initial revision
glantau
parents:
diff changeset
420 #define __align8 __attribute__ ((aligned (8)))
986e461dc072 Initial revision
glantau
parents:
diff changeset
421
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
422 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
423 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
1065
92fb44eae6b6 * prototypes at one place
kabi
parents: 1064
diff changeset
424
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
425 #elif defined(ARCH_ARMV4L)
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
426
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
427 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
1974
8c5489b2cf3e move __align16 some place where non-MMX machines can see it
melanson
parents: 1972
diff changeset
428 line optimizations */
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
429 #define __align8 __attribute__ ((aligned (4)))
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
430
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
431 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
62
4bfc845cdfea arm optimizations
glantau
parents: 50
diff changeset
432
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
433 #elif defined(HAVE_MLIB)
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
434
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
435 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
436 #define __align8 __attribute__ ((aligned (8)))
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
437
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
438 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
88
06f63b58d2a8 mlib merge
glantau
parents: 62
diff changeset
439
1959
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1879
diff changeset
440 #elif defined(ARCH_SPARC)
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1879
diff changeset
441
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1879
diff changeset
442 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1879
diff changeset
443 #define __align8 __attribute__ ((aligned (8)))
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1879
diff changeset
444 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
55b7435c59b8 VIS optimized motion compensation code. by (David S. Miller <davem at redhat dot com>)
michael
parents: 1879
diff changeset
445
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
446 #elif defined(ARCH_ALPHA)
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
447
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
448 #define __align8 __attribute__ ((aligned (8)))
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
449
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
450 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
214
73df666cacc7 Alpha optimizations by Falk Hueffner <falk.hueffner@student.uni-tuebingen.de>
nickols_k
parents: 190
diff changeset
451
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
452 #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
453
894
a408778eff87 altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 884
diff changeset
454 #define MM_ALTIVEC 0x0001 /* standard AltiVec */
a408778eff87 altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 884
diff changeset
455
a408778eff87 altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 884
diff changeset
456 extern int mm_flags;
a408778eff87 altivec accelerated v-resample patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
michaelni
parents: 884
diff changeset
457
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1008
diff changeset
458 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
1653
11433ade9e06 altivec.h bug workaround (untested)
michael
parents: 1644
diff changeset
459 #define pixel altivec_pixel
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1008
diff changeset
460 #include <altivec.h>
1653
11433ade9e06 altivec.h bug workaround (untested)
michael
parents: 1644
diff changeset
461 #undef pixel
1033
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1008
diff changeset
462 #endif
b4172ff70d27 Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents: 1008
diff changeset
463
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
464 #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
465
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
466 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
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
467
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
468 #elif defined(HAVE_MMI)
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
469
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
470 #define __align8 __attribute__ ((aligned (16)))
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
471
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1065
diff changeset
472 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 687
diff changeset
473
1259
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1241
diff changeset
474 #elif defined(ARCH_SH4)
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1241
diff changeset
475
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1241
diff changeset
476 #define __align8 __attribute__ ((aligned (8)))
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1241
diff changeset
477
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1241
diff changeset
478 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
e8c3884f2c7e sh4 optimized idct & bswap patch by (BERO <bero at geocities dot co dot jp>)
michaelni
parents: 1241
diff changeset
479
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
480 #else
986e461dc072 Initial revision
glantau
parents:
diff changeset
481
986e461dc072 Initial revision
glantau
parents:
diff changeset
482 #define __align8
986e461dc072 Initial revision
glantau
parents:
diff changeset
483
986e461dc072 Initial revision
glantau
parents:
diff changeset
484 #endif
986e461dc072 Initial revision
glantau
parents:
diff changeset
485
493
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
486 #ifdef __GNUC__
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
487
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
488 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
489 struct unaligned_32 { uint32_t l; } __attribute__((packed));
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
490 struct unaligned_16 { uint16_t l; } __attribute__((packed));
493
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
491
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
492 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
493
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
493 #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
494 #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
495
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
496 #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
497
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
498 #else /* __GNUC__ */
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
499
1267
85b71f9f7450 moving the svq3 motion compensation stuff to dsputil (this also means that existing optimized halfpel code is used now ...)
michaelni
parents: 1264
diff changeset
500 #define LD16(a) (*((uint16_t*)(a)))
493
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
501 #define LD32(a) (*((uint32_t*)(a)))
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
502 #define LD64(a) (*((uint64_t*)(a)))
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
503
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
504 #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
505
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
506 #endif /* !__GNUC__ */
873b9075d853 move unaligned access macros to dsputil.h - added unaligned 32 bit store
bellard
parents: 480
diff changeset
507
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 214
diff changeset
508 /* PSNR */
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 1040
diff changeset
509 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
252
ddb1a0e94cf4 - Added PSNR feature to libavcodec and ffmpeg. By now just Y PSNR until I'm
pulento
parents: 214
diff changeset
510 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
511 AVCodecContext *avctx);
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
512
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
513 /* FFT computation */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
514
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
515 /* 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
516 FFTSample type */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
517 typedef float FFTSample;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
518
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
519 typedef struct FFTComplex {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
520 FFTSample re, im;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
521 } FFTComplex;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
522
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
523 typedef struct FFTContext {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
524 int nbits;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
525 int inverse;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
526 uint16_t *revtab;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
527 FFTComplex *exptab;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
528 FFTComplex *exptab1; /* only used by SSE code */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
529 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
530 } FFTContext;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
531
1879
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1866
diff changeset
532 int ff_fft_init(FFTContext *s, int nbits, int inverse);
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1866
diff changeset
533 void ff_fft_permute(FFTContext *s, FFTComplex *z);
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1866
diff changeset
534 void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1866
diff changeset
535 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z);
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1866
diff changeset
536 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
975
e05d525505c5 fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents: 954
diff changeset
537
1879
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1866
diff changeset
538 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z)
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
539 {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
540 s->fft_calc(s, z);
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
541 }
1879
dd63cb7e5080 fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents: 1866
diff changeset
542 void ff_fft_end(FFTContext *s);
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
543
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
544 /* MDCT computation */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
545
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
546 typedef struct MDCTContext {
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
547 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
548 int nbits; /* n = 2^nbits */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
549 /* pre/post rotation tables */
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
550 FFTSample *tcos;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
551 FFTSample *tsin;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
552 FFTContext fft;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
553 } MDCTContext;
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
554
794
670009af4fc2 avoid name clash
bellard
parents: 781
diff changeset
555 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
556 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
557 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
558 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
559 const FFTSample *input, FFTSample *tmp);
794
670009af4fc2 avoid name clash
bellard
parents: 781
diff changeset
560 void ff_mdct_end(MDCTContext *s);
781
6f5e87957bcb new generic FFT/MDCT code for audio codecs
bellard
parents: 764
diff changeset
561
1708
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
562 #define WARPER8_16(name8, name16)\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
563 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
564 return name8(s, dst , src , stride, h)\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
565 +name8(s, dst+8 , src+8 , stride, h);\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
566 }
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
567
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
568 #define WARPER8_16_SQ(name8, name16)\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
569 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
570 int score=0;\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
571 score +=name8(s, dst , src , stride, 8);\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
572 score +=name8(s, dst+8 , src+8 , stride, 8);\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
573 if(h==16){\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
574 dst += 8*stride;\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
575 src += 8*stride;\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
576 score +=name8(s, dst , src , stride, 8);\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
577 score +=name8(s, dst+8 , src+8 , stride, 8);\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
578 }\
dea5b2946999 interlaced motion estimation
michael
parents: 1653
diff changeset
579 return score;\
936
caa77cd960c0 qpel encoding
michaelni
parents: 894
diff changeset
580 }
caa77cd960c0 qpel encoding
michaelni
parents: 894
diff changeset
581
838
b78812db886f lrintf detection (based upon a patch by Franois Revol <revol at free dot fr>)
michaelni
parents: 802
diff changeset
582 #ifndef HAVE_LRINTF
796
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
583 /* 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
584 /* better than nothing implementation. */
802
99db75050d38 NetBSD fix by Bernd Ernesti <mplayer@lists.veego.de>
al3x
parents: 796
diff changeset
585 /* 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
586 static inline long int lrintf(float x)
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
587 {
1040
998d5035b15b win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents: 1033
diff changeset
588 #ifdef CONFIG_WIN32
998d5035b15b win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents: 1033
diff changeset
589 /* XXX: incorrect, but make it compile */
998d5035b15b win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents: 1033
diff changeset
590 return (int)(x);
998d5035b15b win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents: 1033
diff changeset
591 #else
796
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
592 return (int)(rint(x));
1040
998d5035b15b win32: rint() does not seem to be defined with mingw32-gcc 2.95 - do you have a better solution ?
bellard
parents: 1033
diff changeset
593 #endif
796
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
594 }
2024
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1984
diff changeset
595 #else
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1984
diff changeset
596 #ifndef _ISOC9X_SOURCE
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1984
diff changeset
597 #define _ISOC9X_SOURCE
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1984
diff changeset
598 #endif
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1984
diff changeset
599 #include <math.h>
0
986e461dc072 Initial revision
glantau
parents:
diff changeset
600 #endif
796
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
601
8a5b70c68fbd added lrintf for non ISOC libcs (fixme: find a better test)
bellard
parents: 794
diff changeset
602 #endif