Statistics
| Branch: | Revision:

ffmpeg / libavcodec / dsputil.h @ 8f2ab833

History | View | Annotate | Download (17.6 KB)

1
/*
2
 * DSP utils
3
 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard.
4
 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5
 *
6
 * This library is free software; you can redistribute it and/or
7
 * modify it under the terms of the GNU Lesser General Public
8
 * License as published by the Free Software Foundation; either
9
 * version 2 of the License, or (at your option) any later version.
10
 *
11
 * This library is distributed in the hope that it will be useful,
12
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14
 * Lesser General Public License for more details.
15
 *
16
 * You should have received a copy of the GNU Lesser General Public
17
 * License along with this library; if not, write to the Free Software
18
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
19
 */
20

    
21
/**
22
 * @file dsputil.h
23
 * DSP utils.
24
 * note, many functions in here may use MMX which trashes the FPU state, it is
25
 * absolutely necessary to call emms_c() between dsp & float/double code
26
 */
27

    
28
#ifndef DSPUTIL_H
29
#define DSPUTIL_H
30

    
31
#include "common.h"
32
#include "avcodec.h"
33

    
34

    
35
//#define DEBUG
36
/* dct code */
37
typedef short DCTELEM;
38

    
39
void fdct_ifast (DCTELEM *data);
40
void fdct_ifast248 (DCTELEM *data);
41
void ff_jpeg_fdct_islow (DCTELEM *data);
42
void ff_fdct248_islow (DCTELEM *data);
43

    
44
void j_rev_dct (DCTELEM *data);
45

    
46
void ff_fdct_mmx(DCTELEM *block);
47
void ff_fdct_mmx2(DCTELEM *block);
48

    
49
/* encoding scans */
50
extern const uint8_t ff_alternate_horizontal_scan[64];
51
extern const uint8_t ff_alternate_vertical_scan[64];
52
extern const uint8_t ff_zigzag_direct[64];
53
extern const uint8_t ff_zigzag248_direct[64];
54

    
55
/* pixel operations */
56
#define MAX_NEG_CROP 384
57

    
58
/* temporary */
59
extern uint32_t squareTbl[512];
60
extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
61

    
62

    
63
/* minimum alignment rules ;)
64
if u notice errors in the align stuff, need more alignment for some asm code for some cpu
65
or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
66

67
!warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
68
i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
69

70
!future video codecs might need functions with less strict alignment
71
*/
72

    
73
/*
74
void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
75
void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
76
void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
77
void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
78
void clear_blocks_c(DCTELEM *blocks);
79
*/
80

    
81
/* add and put pixel (decoding) */
82
// blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
83
//h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
84
typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
85
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);
86
typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
87
typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
88

    
89
#define DEF_OLD_QPEL(name)\
90
void ff_put_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
91
void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
92
void ff_avg_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
93

    
94
DEF_OLD_QPEL(qpel16_mc11_old_c)
95
DEF_OLD_QPEL(qpel16_mc31_old_c)
96
DEF_OLD_QPEL(qpel16_mc12_old_c)
97
DEF_OLD_QPEL(qpel16_mc32_old_c)
98
DEF_OLD_QPEL(qpel16_mc13_old_c)
99
DEF_OLD_QPEL(qpel16_mc33_old_c)
100
DEF_OLD_QPEL(qpel8_mc11_old_c)
101
DEF_OLD_QPEL(qpel8_mc31_old_c)
102
DEF_OLD_QPEL(qpel8_mc12_old_c)
103
DEF_OLD_QPEL(qpel8_mc32_old_c)
104
DEF_OLD_QPEL(qpel8_mc13_old_c)
105
DEF_OLD_QPEL(qpel8_mc33_old_c)
106

    
107
#define CALL_2X_PIXELS(a, b, n)\
108
static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
109
    b(block  , pixels  , line_size, h);\
110
    b(block+n, pixels+n, line_size, h);\
111
}
112

    
113
/* motion estimation */
114
// h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
115
// allthough currently h<4 is not used as functions with width <8 are not used and neither implemented
116
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))*/;
117

    
118

    
119
/**
120
 * DSPContext.
121
 */
122
typedef struct DSPContext {
123
    /* pixel ops : interface with DCT */
124
    void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
125
    void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
126
    void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
127
    void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
128
    /**
129
     * translational global motion compensation.
130
     */
131
    void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
132
    /**
133
     * global motion compensation.
134
     */
135
    void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
136
                    int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
137
    void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
138
    int (*pix_sum)(uint8_t * pix, int line_size);
139
    int (*pix_norm1)(uint8_t * pix, int line_size);
140
// 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
141
    
142
    me_cmp_func sad[5]; /* identical to pix_absAxA except additional void * */
143
    me_cmp_func sse[5];
144
    me_cmp_func hadamard8_diff[5];
145
    me_cmp_func dct_sad[5];
146
    me_cmp_func quant_psnr[5];
147
    me_cmp_func bit[5];
148
    me_cmp_func rd[5];
149
    me_cmp_func vsad[5];
150
    me_cmp_func vsse[5];
151

    
152
    me_cmp_func me_pre_cmp[5];
153
    me_cmp_func me_cmp[5];
154
    me_cmp_func me_sub_cmp[5];
155
    me_cmp_func mb_cmp[5];
156
    me_cmp_func ildct_cmp[5]; //only width 16 used
157

    
158
    /**
159
     * Halfpel motion compensation with rounding (a+b+1)>>1.
160
     * this is an array[4][4] of motion compensation funcions for 4 
161
     * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
162
     * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
163
     * @param block destination where the result is stored
164
     * @param pixels source
165
     * @param line_size number of bytes in a horizontal line of block
166
     * @param h height
167
     */
168
    op_pixels_func put_pixels_tab[4][4];
169

    
170
    /**
171
     * Halfpel motion compensation with rounding (a+b+1)>>1.
172
     * This is an array[4][4] of motion compensation functions for 4 
173
     * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
174
     * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
175
     * @param block destination into which the result is averaged (a+b+1)>>1
176
     * @param pixels source
177
     * @param line_size number of bytes in a horizontal line of block
178
     * @param h height
179
     */
180
    op_pixels_func avg_pixels_tab[4][4];
181

    
182
    /**
183
     * Halfpel motion compensation with no rounding (a+b)>>1.
184
     * this is an array[2][4] of motion compensation funcions for 2 
185
     * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
186
     * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
187
     * @param block destination where the result is stored
188
     * @param pixels source
189
     * @param line_size number of bytes in a horizontal line of block
190
     * @param h height
191
     */
192
    op_pixels_func put_no_rnd_pixels_tab[2][4];
193

    
194
    /**
195
     * Halfpel motion compensation with no rounding (a+b)>>1.
196
     * this is an array[2][4] of motion compensation funcions for 2 
197
     * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
198
     * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
199
     * @param block destination into which the result is averaged (a+b)>>1
200
     * @param pixels source
201
     * @param line_size number of bytes in a horizontal line of block
202
     * @param h height
203
     */
204
    op_pixels_func avg_no_rnd_pixels_tab[2][4];
205
    
206
    /**
207
     * Thirdpel motion compensation with rounding (a+b+1)>>1.
208
     * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
209
     * *pixels_tab[ xthirdpel + 4*ythirdpel ]
210
     * @param block destination where the result is stored
211
     * @param pixels source
212
     * @param line_size number of bytes in a horizontal line of block
213
     * @param h height
214
     */
215
    tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
216
    tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
217

    
218
    qpel_mc_func put_qpel_pixels_tab[2][16];
219
    qpel_mc_func avg_qpel_pixels_tab[2][16];
220
    qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
221
    qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
222
    qpel_mc_func put_mspel_pixels_tab[8];
223
    
224
    /**
225
     * h264 Chram MC
226
     */
227
    h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
228
    h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
229

    
230
    qpel_mc_func put_h264_qpel_pixels_tab[3][16];
231
    qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
232
    
233
    me_cmp_func pix_abs[2][4];
234
    
235
    /* huffyuv specific */
236
    void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
237
    void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
238
    /**
239
     * subtract huffyuv's variant of median prediction
240
     * note, this might read from src1[-1], src2[-1]
241
     */
242
    void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
243
    void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
244
    
245
    void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
246
    void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
247

    
248
    /* (I)DCT */
249
    void (*fdct)(DCTELEM *block/* align 16*/);
250
    void (*fdct248)(DCTELEM *block/* align 16*/);
251
    
252
    /* IDCT really*/
253
    void (*idct)(DCTELEM *block/* align 16*/);
254
    
255
    /**
256
     * block -> idct -> clip to unsigned 8 bit -> dest.
257
     * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
258
     * @param line_size size in bytes of a horizotal line of dest
259
     */
260
    void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
261
    
262
    /**
263
     * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
264
     * @param line_size size in bytes of a horizotal line of dest
265
     */
266
    void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
267
    
268
    /**
269
     * idct input permutation.
270
     * several optimized IDCTs need a permutated input (relative to the normal order of the reference
271
     * IDCT)
272
     * this permutation must be performed before the idct_put/add, note, normally this can be merged
273
     * with the zigzag/alternate scan<br>
274
     * an example to avoid confusion:
275
     * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
276
     * - (x -> referece dct -> reference idct -> x)
277
     * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
278
     * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
279
     */
280
    uint8_t idct_permutation[64];
281
    int idct_permutation_type;
282
#define FF_NO_IDCT_PERM 1
283
#define FF_LIBMPEG2_IDCT_PERM 2
284
#define FF_SIMPLE_IDCT_PERM 3
285
#define FF_TRANSPOSE_IDCT_PERM 4
286

    
287
} DSPContext;
288

    
289
void dsputil_static_init(void);
290
void dsputil_init(DSPContext* p, AVCodecContext *avctx);
291

    
292
/**
293
 * permute block according to permuatation.
294
 * @param last last non zero element in scantable order
295
 */
296
void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
297

    
298
void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
299

    
300
#define        BYTE_VEC32(c)        ((c)*0x01010101UL)
301

    
302
static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
303
{
304
    return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
305
}
306

    
307
static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
308
{
309
    return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
310
}
311

    
312
/**
313
 * Empty mmx state.
314
 * this must be called between any dsp function and float/double code.
315
 * for example sin(); dsp->idct_put(); emms_c(); cos()
316
 */
317
#define emms_c()
318

    
319
/* should be defined by architectures supporting
320
   one or more MultiMedia extension */
321
int mm_support(void);
322

    
323
#if defined(HAVE_MMX)
324

    
325
#undef emms_c
326

    
327
#define MM_MMX    0x0001 /* standard MMX */
328
#define MM_3DNOW  0x0004 /* AMD 3DNOW */
329
#define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
330
#define MM_SSE    0x0008 /* SSE functions */
331
#define MM_SSE2   0x0010 /* PIV SSE2 functions */
332

    
333
extern int mm_flags;
334

    
335
void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
336
void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
337

    
338
static inline void emms(void)
339
{
340
    __asm __volatile ("emms;":::"memory");
341
}
342

    
343

    
344
#define emms_c() \
345
{\
346
    if (mm_flags & MM_MMX)\
347
        emms();\
348
}
349

    
350
#define __align8 __attribute__ ((aligned (8)))
351

    
352
void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
353
void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
354

    
355
#elif defined(ARCH_ARMV4L)
356

    
357
/* This is to use 4 bytes read to the IDCT pointers for some 'zero'
358
   line ptimizations */
359
#define __align8 __attribute__ ((aligned (4)))
360

    
361
void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
362

    
363
#elif defined(HAVE_MLIB)
364

    
365
/* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
366
#define __align8 __attribute__ ((aligned (8)))
367

    
368
void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
369

    
370
#elif defined(ARCH_ALPHA)
371

    
372
#define __align8 __attribute__ ((aligned (8)))
373

    
374
void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
375

    
376
#elif defined(ARCH_POWERPC)
377

    
378
#define MM_ALTIVEC    0x0001 /* standard AltiVec */
379

    
380
extern int mm_flags;
381

    
382
#if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
383
#define pixel altivec_pixel
384
#include <altivec.h>
385
#undef pixel
386
#endif
387

    
388
#define __align8 __attribute__ ((aligned (16)))
389

    
390
void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
391

    
392
#elif defined(HAVE_MMI)
393

    
394
#define __align8 __attribute__ ((aligned (16)))
395

    
396
void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
397

    
398
#elif defined(ARCH_SH4)
399

    
400
#define __align8 __attribute__ ((aligned (8)))
401

    
402
void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
403

    
404
#else
405

    
406
#define __align8
407

    
408
#endif
409

    
410
#ifdef __GNUC__
411

    
412
struct unaligned_64 { uint64_t l; } __attribute__((packed));
413
struct unaligned_32 { uint32_t l; } __attribute__((packed));
414
struct unaligned_16 { uint16_t l; } __attribute__((packed));
415

    
416
#define LD16(a) (((const struct unaligned_16 *) (a))->l)
417
#define LD32(a) (((const struct unaligned_32 *) (a))->l)
418
#define LD64(a) (((const struct unaligned_64 *) (a))->l)
419

    
420
#define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
421

    
422
#else /* __GNUC__ */
423

    
424
#define LD16(a) (*((uint16_t*)(a)))
425
#define LD32(a) (*((uint32_t*)(a)))
426
#define LD64(a) (*((uint64_t*)(a)))
427

    
428
#define ST32(a, b) *((uint32_t*)(a)) = (b)
429

    
430
#endif /* !__GNUC__ */
431

    
432
/* PSNR */
433
void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
434
              int orig_linesize[3], int coded_linesize,
435
              AVCodecContext *avctx);
436

    
437
/* FFT computation */
438

    
439
/* NOTE: soon integer code will be added, so you must use the
440
   FFTSample type */
441
typedef float FFTSample;
442

    
443
typedef struct FFTComplex {
444
    FFTSample re, im;
445
} FFTComplex;
446

    
447
typedef struct FFTContext {
448
    int nbits;
449
    int inverse;
450
    uint16_t *revtab;
451
    FFTComplex *exptab;
452
    FFTComplex *exptab1; /* only used by SSE code */
453
    void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
454
} FFTContext;
455

    
456
int fft_init(FFTContext *s, int nbits, int inverse);
457
void fft_permute(FFTContext *s, FFTComplex *z);
458
void fft_calc_c(FFTContext *s, FFTComplex *z);
459
void fft_calc_sse(FFTContext *s, FFTComplex *z);
460
void fft_calc_altivec(FFTContext *s, FFTComplex *z);
461

    
462
static inline void fft_calc(FFTContext *s, FFTComplex *z)
463
{
464
    s->fft_calc(s, z);
465
}
466
void fft_end(FFTContext *s);
467

    
468
/* MDCT computation */
469

    
470
typedef struct MDCTContext {
471
    int n;  /* size of MDCT (i.e. number of input data * 2) */
472
    int nbits; /* n = 2^nbits */
473
    /* pre/post rotation tables */
474
    FFTSample *tcos;
475
    FFTSample *tsin;
476
    FFTContext fft;
477
} MDCTContext;
478

    
479
int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
480
void ff_imdct_calc(MDCTContext *s, FFTSample *output,
481
                const FFTSample *input, FFTSample *tmp);
482
void ff_mdct_calc(MDCTContext *s, FFTSample *out,
483
               const FFTSample *input, FFTSample *tmp);
484
void ff_mdct_end(MDCTContext *s);
485

    
486
#define WARPER8_16(name8, name16)\
487
static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
488
    return name8(s, dst           , src           , stride, h)\
489
          +name8(s, dst+8         , src+8         , stride, h);\
490
}
491

    
492
#define WARPER8_16_SQ(name8, name16)\
493
static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
494
    int score=0;\
495
    score +=name8(s, dst           , src           , stride, 8);\
496
    score +=name8(s, dst+8         , src+8         , stride, 8);\
497
    if(h==16){\
498
        dst += 8*stride;\
499
        src += 8*stride;\
500
        score +=name8(s, dst           , src           , stride, 8);\
501
        score +=name8(s, dst+8         , src+8         , stride, 8);\
502
    }\
503
    return score;\
504
}
505

    
506
#ifndef HAVE_LRINTF
507
/* XXX: add ISOC specific test to avoid specific BSD testing. */
508
/* better than nothing implementation. */
509
/* btw, rintf() is existing on fbsd too -- alex */
510
static inline long int lrintf(float x)
511
{
512
#ifdef CONFIG_WIN32
513
    /* XXX: incorrect, but make it compile */
514
    return (int)(x);
515
#else
516
    return (int)(rint(x));
517
#endif
518
}
519
#endif
520

    
521
#endif