* introducing dct248 into the DSP context.
[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 void ff_fdct248_islow (DCTELEM *data);
41
42 void j_rev_dct (DCTELEM *data);
43
44 void ff_fdct_mmx(DCTELEM *block);
45 void ff_fdct_mmx2(DCTELEM *block);
46
47 /* encoding scans */
48 extern const uint8_t ff_alternate_horizontal_scan[64];
49 extern const uint8_t ff_alternate_vertical_scan[64];
50 extern const uint8_t ff_zigzag_direct[64];
51 extern const uint8_t ff_zigzag248_direct[64];
52
53 /* pixel operations */
54 #define MAX_NEG_CROP 384
55
56 /* temporary */
57 extern uint32_t squareTbl[512];
58 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
59
60
61 /* minimum alignment rules ;)
62 if u notice errors in the align stuff, need more alignment for some asm code for some cpu
63 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
64
65 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
66 i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
67
68 !future video codecs might need functions with less strict alignment
69 */
70
71 /*
72 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
73 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
74 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
75 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
76 void clear_blocks_c(DCTELEM *blocks);
77 */
78
79 /* add and put pixel (decoding) */
80 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
81 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
82 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);
83 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
84 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
85
86 #define DEF_OLD_QPEL(name)\
87 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
88 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
89 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
90
91 DEF_OLD_QPEL(qpel16_mc11_old_c)
92 DEF_OLD_QPEL(qpel16_mc31_old_c)
93 DEF_OLD_QPEL(qpel16_mc12_old_c)
94 DEF_OLD_QPEL(qpel16_mc32_old_c)
95 DEF_OLD_QPEL(qpel16_mc13_old_c)
96 DEF_OLD_QPEL(qpel16_mc33_old_c)
97 DEF_OLD_QPEL(qpel8_mc11_old_c)
98 DEF_OLD_QPEL(qpel8_mc31_old_c)
99 DEF_OLD_QPEL(qpel8_mc12_old_c)
100 DEF_OLD_QPEL(qpel8_mc32_old_c)
101 DEF_OLD_QPEL(qpel8_mc13_old_c)
102 DEF_OLD_QPEL(qpel8_mc33_old_c)
103
104 #define CALL_2X_PIXELS(a, b, n)\
105 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
106 b(block , pixels , line_size, h);\
107 b(block+n, pixels+n, line_size, h);\
108 }
109
110 /* motion estimation */
111
112 typedef int (*op_pixels_abs_func)(uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size)/* __attribute__ ((const))*/;
113
114 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))*/;
115
116
117 /**
118 * DSPContext.
119 */
120 typedef struct DSPContext {
121 /* pixel ops : interface with DCT */
122 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
123 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
124 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
125 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
126 /**
127 * translational global motion compensation.
128 */
129 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
130 /**
131 * global motion compensation.
132 */
133 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
134 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
135 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
136 int (*pix_sum)(uint8_t * pix, int line_size);
137 int (*pix_norm1)(uint8_t * pix, int line_size);
138 me_cmp_func sad[2]; /* identical to pix_absAxA except additional void * */
139 me_cmp_func sse[2];
140 me_cmp_func hadamard8_diff[2];
141 me_cmp_func dct_sad[2];
142 me_cmp_func quant_psnr[2];
143 me_cmp_func bit[2];
144 me_cmp_func rd[2];
145 int (*hadamard8_abs )(uint8_t *src, int stride, int mean);
146
147 me_cmp_func me_pre_cmp[11];
148 me_cmp_func me_cmp[11];
149 me_cmp_func me_sub_cmp[11];
150 me_cmp_func mb_cmp[11];
151
152 /* maybe create an array for 16/8/4/2 functions */
153 /**
154 * Halfpel motion compensation with rounding (a+b+1)>>1.
155 * this is an array[4][4] of motion compensation funcions for 4
156 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br>
157 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
158 * @param block destination where the result is stored
159 * @param pixels source
160 * @param line_size number of bytes in a horizontal line of block
161 * @param h height
162 */
163 op_pixels_func put_pixels_tab[4][4];
164
165 /**
166 * Halfpel motion compensation with rounding (a+b+1)>>1.
167 * This is an array[4][4] of motion compensation functions for 4
168 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br>
169 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
170 * @param block destination into which the result is averaged (a+b+1)>>1
171 * @param pixels source
172 * @param line_size number of bytes in a horizontal line of block
173 * @param h height
174 */
175 op_pixels_func avg_pixels_tab[4][4];
176
177 /**
178 * Halfpel motion compensation with no rounding (a+b)>>1.
179 * this is an array[2][4] of motion compensation funcions for 2
180 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
181 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
182 * @param block destination where the result is stored
183 * @param pixels source
184 * @param line_size number of bytes in a horizontal line of block
185 * @param h height
186 */
187 op_pixels_func put_no_rnd_pixels_tab[2][4];
188
189 /**
190 * Halfpel motion compensation with no rounding (a+b)>>1.
191 * this is an array[2][4] of motion compensation funcions for 2
192 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
193 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
194 * @param block destination into which the result is averaged (a+b)>>1
195 * @param pixels source
196 * @param line_size number of bytes in a horizontal line of block
197 * @param h height
198 */
199 op_pixels_func avg_no_rnd_pixels_tab[2][4];
200
201 /**
202 * Thirdpel motion compensation with rounding (a+b+1)>>1.
203 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
204 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
205 * @param block destination where the result is stored
206 * @param pixels source
207 * @param line_size number of bytes in a horizontal line of block
208 * @param h height
209 */
210 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
211 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
212
213 qpel_mc_func put_qpel_pixels_tab[2][16];
214 qpel_mc_func avg_qpel_pixels_tab[2][16];
215 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
216 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
217 qpel_mc_func put_mspel_pixels_tab[8];
218
219 /**
220 * h264 Chram MC
221 */
222 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
223 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
224
225 qpel_mc_func put_h264_qpel_pixels_tab[3][16];
226 qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
227
228 op_pixels_abs_func pix_abs16x16;
229 op_pixels_abs_func pix_abs16x16_x2;
230 op_pixels_abs_func pix_abs16x16_y2;
231 op_pixels_abs_func pix_abs16x16_xy2;
232 op_pixels_abs_func pix_abs8x8;
233 op_pixels_abs_func pix_abs8x8_x2;
234 op_pixels_abs_func pix_abs8x8_y2;
235 op_pixels_abs_func pix_abs8x8_xy2;
236
237 /* huffyuv specific */
238 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
239 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
240 /**
241 * subtract huffyuv's variant of median prediction
242 * note, this might read from src1[-1], src2[-1]
243 */
244 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
245 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
246
247 /* (I)DCT */
248 void (*fdct)(DCTELEM *block/* align 16*/);
249 void (*fdct248)(DCTELEM *block/* align 16*/);
250
251 /* IDCT really*/
252 void (*idct)(DCTELEM *block/* align 16*/);
253
254 /**
255 * block -> idct -> clip to unsigned 8 bit -> dest.
256 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
257 * @param line_size size in bytes of a horizotal line of dest
258 */
259 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
260
261 /**
262 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
263 * @param line_size size in bytes of a horizotal line of dest
264 */
265 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
266
267 /**
268 * idct input permutation.
269 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
270 * IDCT)
271 * this permutation must be performed before the idct_put/add, note, normally this can be merged
272 * with the zigzag/alternate scan<br>
273 * an example to avoid confusion:
274 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
275 * - (x -> referece dct -> reference idct -> x)
276 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
277 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
278 */
279 uint8_t idct_permutation[64];
280 int idct_permutation_type;
281 #define FF_NO_IDCT_PERM 1
282 #define FF_LIBMPEG2_IDCT_PERM 2
283 #define FF_SIMPLE_IDCT_PERM 3
284 #define FF_TRANSPOSE_IDCT_PERM 4
285
286 } DSPContext;
287
288 void dsputil_static_init(void);
289 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
290
291 /**
292 * permute block according to permuatation.
293 * @param last last non zero element in scantable order
294 */
295 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
296
297 #define BYTE_VEC32(c) ((c)*0x01010101UL)
298
299 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
300 {
301 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
302 }
303
304 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
305 {
306 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
307 }
308
309 /**
310 * Empty mmx state.
311 * this must be called between any dsp function and float/double code.
312 * for example sin(); dsp->idct_put(); emms_c(); cos()
313 */
314 #define emms_c()
315
316 /* should be defined by architectures supporting
317 one or more MultiMedia extension */
318 int mm_support(void);
319
320 #if defined(HAVE_MMX)
321
322 #undef emms_c
323
324 #define MM_MMX 0x0001 /* standard MMX */
325 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
326 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
327 #define MM_SSE 0x0008 /* SSE functions */
328 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
329
330 extern int mm_flags;
331
332 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
333 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
334
335 static inline void emms(void)
336 {
337 __asm __volatile ("emms;":::"memory");
338 }
339
340
341 #define emms_c() \
342 {\
343 if (mm_flags & MM_MMX)\
344 emms();\
345 }
346
347 #define __align8 __attribute__ ((aligned (8)))
348
349 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
350 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
351
352 #elif defined(ARCH_ARMV4L)
353
354 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
355 line ptimizations */
356 #define __align8 __attribute__ ((aligned (4)))
357
358 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
359
360 #elif defined(HAVE_MLIB)
361
362 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
363 #define __align8 __attribute__ ((aligned (8)))
364
365 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
366
367 #elif defined(ARCH_ALPHA)
368
369 #define __align8 __attribute__ ((aligned (8)))
370
371 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
372
373 #elif defined(ARCH_POWERPC)
374
375 #define MM_ALTIVEC 0x0001 /* standard AltiVec */
376
377 extern int mm_flags;
378
379 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
380 #include <altivec.h>
381 #endif
382
383 #define __align8 __attribute__ ((aligned (16)))
384
385 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
386
387 #elif defined(HAVE_MMI)
388
389 #define __align8 __attribute__ ((aligned (16)))
390
391 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
392
393 #elif defined(ARCH_SH4)
394
395 #define __align8 __attribute__ ((aligned (8)))
396
397 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
398
399 #else
400
401 #define __align8
402
403 #endif
404
405 #ifdef __GNUC__
406
407 struct unaligned_64 { uint64_t l; } __attribute__((packed));
408 struct unaligned_32 { uint32_t l; } __attribute__((packed));
409 struct unaligned_16 { uint16_t l; } __attribute__((packed));
410
411 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
412 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
413 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
414
415 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
416
417 #else /* __GNUC__ */
418
419 #define LD16(a) (*((uint16_t*)(a)))
420 #define LD32(a) (*((uint32_t*)(a)))
421 #define LD64(a) (*((uint64_t*)(a)))
422
423 #define ST32(a, b) *((uint32_t*)(a)) = (b)
424
425 #endif /* !__GNUC__ */
426
427 /* PSNR */
428 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
429 int orig_linesize[3], int coded_linesize,
430 AVCodecContext *avctx);
431
432 /* FFT computation */
433
434 /* NOTE: soon integer code will be added, so you must use the
435 FFTSample type */
436 typedef float FFTSample;
437
438 typedef struct FFTComplex {
439 FFTSample re, im;
440 } FFTComplex;
441
442 typedef struct FFTContext {
443 int nbits;
444 int inverse;
445 uint16_t *revtab;
446 FFTComplex *exptab;
447 FFTComplex *exptab1; /* only used by SSE code */
448 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
449 } FFTContext;
450
451 int fft_init(FFTContext *s, int nbits, int inverse);
452 void fft_permute(FFTContext *s, FFTComplex *z);
453 void fft_calc_c(FFTContext *s, FFTComplex *z);
454 void fft_calc_sse(FFTContext *s, FFTComplex *z);
455 void fft_calc_altivec(FFTContext *s, FFTComplex *z);
456
457 static inline void fft_calc(FFTContext *s, FFTComplex *z)
458 {
459 s->fft_calc(s, z);
460 }
461 void fft_end(FFTContext *s);
462
463 /* MDCT computation */
464
465 typedef struct MDCTContext {
466 int n; /* size of MDCT (i.e. number of input data * 2) */
467 int nbits; /* n = 2^nbits */
468 /* pre/post rotation tables */
469 FFTSample *tcos;
470 FFTSample *tsin;
471 FFTContext fft;
472 } MDCTContext;
473
474 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
475 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
476 const FFTSample *input, FFTSample *tmp);
477 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
478 const FFTSample *input, FFTSample *tmp);
479 void ff_mdct_end(MDCTContext *s);
480
481 #define WARPER88_1616(name8, name16)\
482 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride){\
483 return name8(s, dst , src , stride)\
484 +name8(s, dst+8 , src+8 , stride)\
485 +name8(s, dst +8*stride, src +8*stride, stride)\
486 +name8(s, dst+8+8*stride, src+8+8*stride, stride);\
487 }
488
489 #ifndef HAVE_LRINTF
490 /* XXX: add ISOC specific test to avoid specific BSD testing. */
491 /* better than nothing implementation. */
492 /* btw, rintf() is existing on fbsd too -- alex */
493 static inline long int lrintf(float x)
494 {
495 #ifdef CONFIG_WIN32
496 /* XXX: incorrect, but make it compile */
497 return (int)(x);
498 #else
499 return (int)(rint(x));
500 #endif
501 }
502 #endif
503
504 #if defined(CONFIG_OS2) || defined(CONFIG_SUNOS)
505 static inline float floorf(float f) {
506 return floor(f);
507 }
508 #endif
509
510 #endif