9bb5fd2df566b663cb785702e88f340215d073ab
[libav.git] / libavcodec / dsputil.h
1 /*
2 * DSP utils
3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard.
4 *
5 * This library is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU Lesser General Public
7 * License as published by the Free Software Foundation; either
8 * version 2 of the License, or (at your option) any later version.
9 *
10 * This library is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * Lesser General Public License for more details.
14 *
15 * You should have received a copy of the GNU Lesser General Public
16 * License along with this library; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20 /**
21 * @file dsputil.h
22 * DSP utils.
23 * note, many functions in here may use MMX which trashes the FPU state, it is
24 * absolutely necessary to call emms_c() between dsp & float/double code
25 */
26
27 #ifndef DSPUTIL_H
28 #define DSPUTIL_H
29
30 #include "common.h"
31 #include "avcodec.h"
32
33
34 //#define DEBUG
35 /* dct code */
36 typedef short DCTELEM;
37
38 void fdct_ifast (DCTELEM *data);
39 void ff_jpeg_fdct_islow (DCTELEM *data);
40
41 void j_rev_dct (DCTELEM *data);
42
43 void ff_fdct_mmx(DCTELEM *block);
44 void ff_fdct_mmx2(DCTELEM *block);
45
46 /* encoding scans */
47 extern const uint8_t ff_alternate_horizontal_scan[64];
48 extern const uint8_t ff_alternate_vertical_scan[64];
49 extern const uint8_t ff_zigzag_direct[64];
50
51 /* pixel operations */
52 #define MAX_NEG_CROP 384
53
54 /* temporary */
55 extern uint32_t squareTbl[512];
56 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
57
58
59 /* minimum alignment rules ;)
60 if u notice errors in the align stuff, need more alignment for some asm code for some cpu
61 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
62
63 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
64 i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
65
66 !future video codecs might need functions with less strict alignment
67 */
68
69 /*
70 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
71 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
72 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
73 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
74 void clear_blocks_c(DCTELEM *blocks);
75 */
76
77 /* add and put pixel (decoding) */
78 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
79 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
80 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);
81 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
82 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
83
84 #define DEF_OLD_QPEL(name)\
85 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
86 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
87 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
88
89 DEF_OLD_QPEL(qpel16_mc11_old_c)
90 DEF_OLD_QPEL(qpel16_mc31_old_c)
91 DEF_OLD_QPEL(qpel16_mc12_old_c)
92 DEF_OLD_QPEL(qpel16_mc32_old_c)
93 DEF_OLD_QPEL(qpel16_mc13_old_c)
94 DEF_OLD_QPEL(qpel16_mc33_old_c)
95 DEF_OLD_QPEL(qpel8_mc11_old_c)
96 DEF_OLD_QPEL(qpel8_mc31_old_c)
97 DEF_OLD_QPEL(qpel8_mc12_old_c)
98 DEF_OLD_QPEL(qpel8_mc32_old_c)
99 DEF_OLD_QPEL(qpel8_mc13_old_c)
100 DEF_OLD_QPEL(qpel8_mc33_old_c)
101
102 #define CALL_2X_PIXELS(a, b, n)\
103 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
104 b(block , pixels , line_size, h);\
105 b(block+n, pixels+n, line_size, h);\
106 }
107
108 /* motion estimation */
109
110 typedef int (*op_pixels_abs_func)(uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size)/* __attribute__ ((const))*/;
111
112 typedef int (*me_cmp_func)(void /*MpegEncContext*/ *s, uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size)/* __attribute__ ((const))*/;
113
114
115 /**
116 * DSPContext.
117 */
118 typedef struct DSPContext {
119 /* pixel ops : interface with DCT */
120 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
121 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
122 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
123 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
124 /**
125 * translational global motion compensation.
126 */
127 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
128 /**
129 * global motion compensation.
130 */
131 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
132 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
133 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
134 int (*pix_sum)(uint8_t * pix, int line_size);
135 int (*pix_norm1)(uint8_t * pix, int line_size);
136 me_cmp_func sad[2]; /* identical to pix_absAxA except additional void * */
137 me_cmp_func sse[2];
138 me_cmp_func hadamard8_diff[2];
139 me_cmp_func dct_sad[2];
140 me_cmp_func quant_psnr[2];
141 me_cmp_func bit[2];
142 me_cmp_func rd[2];
143 int (*hadamard8_abs )(uint8_t *src, int stride, int mean);
144
145 me_cmp_func me_pre_cmp[11];
146 me_cmp_func me_cmp[11];
147 me_cmp_func me_sub_cmp[11];
148 me_cmp_func mb_cmp[11];
149
150 /* maybe create an array for 16/8/4/2 functions */
151 /**
152 * Halfpel motion compensation with rounding (a+b+1)>>1.
153 * this is an array[4][4] of motion compensation funcions for 4
154 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br>
155 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
156 * @param block destination where the result is stored
157 * @param pixels source
158 * @param line_size number of bytes in a horizontal line of block
159 * @param h height
160 */
161 op_pixels_func put_pixels_tab[4][4];
162
163 /**
164 * Halfpel motion compensation with rounding (a+b+1)>>1.
165 * This is an array[4][4] of motion compensation functions for 4
166 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br>
167 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
168 * @param block destination into which the result is averaged (a+b+1)>>1
169 * @param pixels source
170 * @param line_size number of bytes in a horizontal line of block
171 * @param h height
172 */
173 op_pixels_func avg_pixels_tab[4][4];
174
175 /**
176 * Halfpel motion compensation with no rounding (a+b)>>1.
177 * this is an array[2][4] of motion compensation funcions for 2
178 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
179 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
180 * @param block destination where the result is stored
181 * @param pixels source
182 * @param line_size number of bytes in a horizontal line of block
183 * @param h height
184 */
185 op_pixels_func put_no_rnd_pixels_tab[2][4];
186
187 /**
188 * Halfpel motion compensation with no rounding (a+b)>>1.
189 * this is an array[2][4] of motion compensation funcions for 2
190 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
191 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
192 * @param block destination into which the result is averaged (a+b)>>1
193 * @param pixels source
194 * @param line_size number of bytes in a horizontal line of block
195 * @param h height
196 */
197 op_pixels_func avg_no_rnd_pixels_tab[2][4];
198
199 /**
200 * Thirdpel motion compensation with rounding (a+b+1)>>1.
201 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
202 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
203 * @param block destination where the result is stored
204 * @param pixels source
205 * @param line_size number of bytes in a horizontal line of block
206 * @param h height
207 */
208 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
209 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
210
211 qpel_mc_func put_qpel_pixels_tab[2][16];
212 qpel_mc_func avg_qpel_pixels_tab[2][16];
213 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
214 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
215 qpel_mc_func put_mspel_pixels_tab[8];
216
217 /**
218 * h264 Chram MC
219 */
220 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
221 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
222
223 qpel_mc_func put_h264_qpel_pixels_tab[3][16];
224 qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
225
226 op_pixels_abs_func pix_abs16x16;
227 op_pixels_abs_func pix_abs16x16_x2;
228 op_pixels_abs_func pix_abs16x16_y2;
229 op_pixels_abs_func pix_abs16x16_xy2;
230 op_pixels_abs_func pix_abs8x8;
231 op_pixels_abs_func pix_abs8x8_x2;
232 op_pixels_abs_func pix_abs8x8_y2;
233 op_pixels_abs_func pix_abs8x8_xy2;
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 /* (I)DCT */
246 void (*fdct)(DCTELEM *block/* align 16*/);
247
248 /* IDCT really*/
249 void (*idct)(DCTELEM *block/* align 16*/);
250
251 /**
252 * block -> idct -> clip to unsigned 8 bit -> dest.
253 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
254 * @param line_size size in bytes of a horizotal line of dest
255 */
256 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
257
258 /**
259 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
260 * @param line_size size in bytes of a horizotal line of dest
261 */
262 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
263
264 /**
265 * idct input permutation.
266 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
267 * IDCT)
268 * this permutation must be performed before the idct_put/add, note, normally this can be merged
269 * with the zigzag/alternate scan<br>
270 * an example to avoid confusion:
271 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
272 * - (x -> referece dct -> reference idct -> x)
273 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
274 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
275 */
276 uint8_t idct_permutation[64];
277 int idct_permutation_type;
278 #define FF_NO_IDCT_PERM 1
279 #define FF_LIBMPEG2_IDCT_PERM 2
280 #define FF_SIMPLE_IDCT_PERM 3
281 #define FF_TRANSPOSE_IDCT_PERM 4
282
283 } DSPContext;
284
285 void dsputil_static_init(void);
286 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
287
288 /**
289 * permute block according to permuatation.
290 * @param last last non zero element in scantable order
291 */
292 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
293
294 #define BYTE_VEC32(c) ((c)*0x01010101UL)
295
296 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
297 {
298 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
299 }
300
301 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
302 {
303 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
304 }
305
306 /**
307 * Empty mmx state.
308 * this must be called between any dsp function and float/double code.
309 * for example sin(); dsp->idct_put(); emms_c(); cos()
310 */
311 #define emms_c()
312
313 /* should be defined by architectures supporting
314 one or more MultiMedia extension */
315 int mm_support(void);
316
317 #if defined(HAVE_MMX)
318
319 #undef emms_c
320
321 #define MM_MMX 0x0001 /* standard MMX */
322 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
323 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
324 #define MM_SSE 0x0008 /* SSE functions */
325 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
326
327 extern int mm_flags;
328
329 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
330 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
331
332 static inline void emms(void)
333 {
334 __asm __volatile ("emms;":::"memory");
335 }
336
337
338 #define emms_c() \
339 {\
340 if (mm_flags & MM_MMX)\
341 emms();\
342 }
343
344 #define __align8 __attribute__ ((aligned (8)))
345
346 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
347 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
348
349 #elif defined(ARCH_ARMV4L)
350
351 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
352 line ptimizations */
353 #define __align8 __attribute__ ((aligned (4)))
354
355 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
356
357 #elif defined(HAVE_MLIB)
358
359 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
360 #define __align8 __attribute__ ((aligned (8)))
361
362 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
363
364 #elif defined(ARCH_ALPHA)
365
366 #define __align8 __attribute__ ((aligned (8)))
367
368 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
369
370 #elif defined(ARCH_POWERPC)
371
372 #define MM_ALTIVEC 0x0001 /* standard AltiVec */
373
374 extern int mm_flags;
375
376 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
377 #include <altivec.h>
378 #endif
379
380 #define __align8 __attribute__ ((aligned (16)))
381
382 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
383
384 #elif defined(HAVE_MMI)
385
386 #define __align8 __attribute__ ((aligned (16)))
387
388 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
389
390 #elif defined(ARCH_SH4)
391
392 #define __align8 __attribute__ ((aligned (8)))
393
394 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
395
396 #else
397
398 #define __align8
399
400 #endif
401
402 #ifdef __GNUC__
403
404 struct unaligned_64 { uint64_t l; } __attribute__((packed));
405 struct unaligned_32 { uint32_t l; } __attribute__((packed));
406 struct unaligned_16 { uint16_t l; } __attribute__((packed));
407
408 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
409 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
410 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
411
412 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
413
414 #else /* __GNUC__ */
415
416 #define LD16(a) (*((uint16_t*)(a)))
417 #define LD32(a) (*((uint32_t*)(a)))
418 #define LD64(a) (*((uint64_t*)(a)))
419
420 #define ST32(a, b) *((uint32_t*)(a)) = (b)
421
422 #endif /* !__GNUC__ */
423
424 /* PSNR */
425 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
426 int orig_linesize[3], int coded_linesize,
427 AVCodecContext *avctx);
428
429 /* FFT computation */
430
431 /* NOTE: soon integer code will be added, so you must use the
432 FFTSample type */
433 typedef float FFTSample;
434
435 typedef struct FFTComplex {
436 FFTSample re, im;
437 } FFTComplex;
438
439 typedef struct FFTContext {
440 int nbits;
441 int inverse;
442 uint16_t *revtab;
443 FFTComplex *exptab;
444 FFTComplex *exptab1; /* only used by SSE code */
445 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
446 } FFTContext;
447
448 int fft_init(FFTContext *s, int nbits, int inverse);
449 void fft_permute(FFTContext *s, FFTComplex *z);
450 void fft_calc_c(FFTContext *s, FFTComplex *z);
451 void fft_calc_sse(FFTContext *s, FFTComplex *z);
452 void fft_calc_altivec(FFTContext *s, FFTComplex *z);
453
454 static inline void fft_calc(FFTContext *s, FFTComplex *z)
455 {
456 s->fft_calc(s, z);
457 }
458 void fft_end(FFTContext *s);
459
460 /* MDCT computation */
461
462 typedef struct MDCTContext {
463 int n; /* size of MDCT (i.e. number of input data * 2) */
464 int nbits; /* n = 2^nbits */
465 /* pre/post rotation tables */
466 FFTSample *tcos;
467 FFTSample *tsin;
468 FFTContext fft;
469 } MDCTContext;
470
471 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
472 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
473 const FFTSample *input, FFTSample *tmp);
474 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
475 const FFTSample *input, FFTSample *tmp);
476 void ff_mdct_end(MDCTContext *s);
477
478 #define WARPER88_1616(name8, name16)\
479 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride){\
480 return name8(s, dst , src , stride)\
481 +name8(s, dst+8 , src+8 , stride)\
482 +name8(s, dst +8*stride, src +8*stride, stride)\
483 +name8(s, dst+8+8*stride, src+8+8*stride, stride);\
484 }
485
486 #ifndef HAVE_LRINTF
487 /* XXX: add ISOC specific test to avoid specific BSD testing. */
488 /* better than nothing implementation. */
489 /* btw, rintf() is existing on fbsd too -- alex */
490 static inline long int lrintf(float x)
491 {
492 #ifdef CONFIG_WIN32
493 /* XXX: incorrect, but make it compile */
494 return (int)(x);
495 #else
496 return (int)(rint(x));
497 #endif
498 }
499 #endif
500
501 #if defined(CONFIG_OS2) || defined(CONFIG_SUNOS)
502 static inline float floorf(float f) {
503 return floor(f);
504 }
505 #endif
506
507 #endif