3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard.
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.
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.
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
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
36 typedef short DCTELEM;
38 void fdct_ifast (DCTELEM *data);
39 void fdct_ifast248 (DCTELEM *data);
40 void ff_jpeg_fdct_islow (DCTELEM *data);
41 void ff_fdct248_islow (DCTELEM *data);
43 void j_rev_dct (DCTELEM *data);
45 void ff_fdct_mmx(DCTELEM *block);
46 void ff_fdct_mmx2(DCTELEM *block);
49 extern const uint8_t ff_alternate_horizontal_scan[64];
50 extern const uint8_t ff_alternate_vertical_scan[64];
51 extern const uint8_t ff_zigzag_direct[64];
52 extern const uint8_t ff_zigzag248_direct[64];
54 /* pixel operations */
55 #define MAX_NEG_CROP 384
58 extern uint32_t squareTbl[512];
59 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
62 /* minimum alignment rules ;)
63 if u notice errors in the align stuff, need more alignment for some asm code for some cpu
64 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
66 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
67 i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
69 !future video codecs might need functions with less strict alignment
73 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
74 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
75 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
76 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
77 void clear_blocks_c(DCTELEM *blocks);
80 /* add and put pixel (decoding) */
81 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
82 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
83 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
84 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);
85 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
86 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 #define DEF_OLD_QPEL(name)\
89 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
90 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
91 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
93 DEF_OLD_QPEL(qpel16_mc11_old_c)
94 DEF_OLD_QPEL(qpel16_mc31_old_c)
95 DEF_OLD_QPEL(qpel16_mc12_old_c)
96 DEF_OLD_QPEL(qpel16_mc32_old_c)
97 DEF_OLD_QPEL(qpel16_mc13_old_c)
98 DEF_OLD_QPEL(qpel16_mc33_old_c)
99 DEF_OLD_QPEL(qpel8_mc11_old_c)
100 DEF_OLD_QPEL(qpel8_mc31_old_c)
101 DEF_OLD_QPEL(qpel8_mc12_old_c)
102 DEF_OLD_QPEL(qpel8_mc32_old_c)
103 DEF_OLD_QPEL(qpel8_mc13_old_c)
104 DEF_OLD_QPEL(qpel8_mc33_old_c)
106 #define CALL_2X_PIXELS(a, b, n)\
107 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
108 b(block , pixels , line_size, h);\
109 b(block+n, pixels+n, line_size, h);\
112 /* motion estimation */
113 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
114 // allthough currently h<4 is not used as functions with width <8 are not used and neither implemented
115 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))*/;
121 typedef struct DSPContext {
122 /* pixel ops : interface with DCT */
123 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
124 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
125 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
126 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
128 * translational global motion compensation.
130 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
132 * global motion compensation.
134 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
135 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
136 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
137 int (*pix_sum)(uint8_t * pix, int line_size);
138 int (*pix_norm1)(uint8_t * pix, int line_size);
139 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
141 me_cmp_func sad[5]; /* identical to pix_absAxA except additional void * */
143 me_cmp_func hadamard8_diff[5];
144 me_cmp_func dct_sad[5];
145 me_cmp_func quant_psnr[5];
151 me_cmp_func me_pre_cmp[5];
152 me_cmp_func me_cmp[5];
153 me_cmp_func me_sub_cmp[5];
154 me_cmp_func mb_cmp[5];
155 me_cmp_func ildct_cmp[5]; //only width 16 used
158 * Halfpel motion compensation with rounding (a+b+1)>>1.
159 * this is an array[4][4] of motion compensation funcions for 4
160 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
161 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
162 * @param block destination where the result is stored
163 * @param pixels source
164 * @param line_size number of bytes in a horizontal line of block
167 op_pixels_func put_pixels_tab[4][4];
170 * Halfpel motion compensation with rounding (a+b+1)>>1.
171 * This is an array[4][4] of motion compensation functions for 4
172 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
173 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
174 * @param block destination into which the result is averaged (a+b+1)>>1
175 * @param pixels source
176 * @param line_size number of bytes in a horizontal line of block
179 op_pixels_func avg_pixels_tab[4][4];
182 * Halfpel motion compensation with no rounding (a+b)>>1.
183 * this is an array[2][4] of motion compensation funcions for 2
184 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
185 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
186 * @param block destination where the result is stored
187 * @param pixels source
188 * @param line_size number of bytes in a horizontal line of block
191 op_pixels_func put_no_rnd_pixels_tab[2][4];
194 * Halfpel motion compensation with no rounding (a+b)>>1.
195 * this is an array[2][4] of motion compensation funcions for 2
196 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
197 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
198 * @param block destination into which the result is averaged (a+b)>>1
199 * @param pixels source
200 * @param line_size number of bytes in a horizontal line of block
203 op_pixels_func avg_no_rnd_pixels_tab[2][4];
206 * Thirdpel motion compensation with rounding (a+b+1)>>1.
207 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
208 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
209 * @param block destination where the result is stored
210 * @param pixels source
211 * @param line_size number of bytes in a horizontal line of block
214 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
215 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
217 qpel_mc_func put_qpel_pixels_tab[2][16];
218 qpel_mc_func avg_qpel_pixels_tab[2][16];
219 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
220 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
221 qpel_mc_func put_mspel_pixels_tab[8];
226 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
227 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
229 qpel_mc_func put_h264_qpel_pixels_tab[3][16];
230 qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
232 me_cmp_func pix_abs[2][4];
234 /* huffyuv specific */
235 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
236 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
238 * subtract huffyuv's variant of median prediction
239 * note, this might read from src1[-1], src2[-1]
241 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
242 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
244 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
245 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
248 void (*fdct)(DCTELEM *block/* align 16*/);
249 void (*fdct248)(DCTELEM *block/* align 16*/);
252 void (*idct)(DCTELEM *block/* align 16*/);
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
259 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
262 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
263 * @param line_size size in bytes of a horizotal line of dest
265 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
268 * idct input permutation.
269 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
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 ->...)
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
288 void dsputil_static_init(void);
289 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
292 * permute block according to permuatation.
293 * @param last last non zero element in scantable order
295 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
297 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
299 #define BYTE_VEC32(c) ((c)*0x01010101UL)
301 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
303 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
306 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
308 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
313 * this must be called between any dsp function and float/double code.
314 * for example sin(); dsp->idct_put(); emms_c(); cos()
318 /* should be defined by architectures supporting
319 one or more MultiMedia extension */
320 int mm_support(void);
322 #if defined(HAVE_MMX)
326 #define MM_MMX 0x0001 /* standard MMX */
327 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
328 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
329 #define MM_SSE 0x0008 /* SSE functions */
330 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
334 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
335 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
337 static inline void emms(void)
339 __asm __volatile ("emms;":::"memory");
345 if (mm_flags & MM_MMX)\
349 #define __align8 __attribute__ ((aligned (8)))
351 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
352 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
354 #elif defined(ARCH_ARMV4L)
356 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
358 #define __align8 __attribute__ ((aligned (4)))
360 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
362 #elif defined(HAVE_MLIB)
364 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
365 #define __align8 __attribute__ ((aligned (8)))
367 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
369 #elif defined(ARCH_ALPHA)
371 #define __align8 __attribute__ ((aligned (8)))
373 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
375 #elif defined(ARCH_POWERPC)
377 #define MM_ALTIVEC 0x0001 /* standard AltiVec */
381 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
382 #define pixel altivec_pixel
387 #define __align8 __attribute__ ((aligned (16)))
389 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
391 #elif defined(HAVE_MMI)
393 #define __align8 __attribute__ ((aligned (16)))
395 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
397 #elif defined(ARCH_SH4)
399 #define __align8 __attribute__ ((aligned (8)))
401 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
411 struct unaligned_64 { uint64_t l; } __attribute__((packed));
412 struct unaligned_32 { uint32_t l; } __attribute__((packed));
413 struct unaligned_16 { uint16_t l; } __attribute__((packed));
415 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
416 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
417 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
419 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
423 #define LD16(a) (*((uint16_t*)(a)))
424 #define LD32(a) (*((uint32_t*)(a)))
425 #define LD64(a) (*((uint64_t*)(a)))
427 #define ST32(a, b) *((uint32_t*)(a)) = (b)
429 #endif /* !__GNUC__ */
432 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
433 int orig_linesize[3], int coded_linesize,
434 AVCodecContext *avctx);
436 /* FFT computation */
438 /* NOTE: soon integer code will be added, so you must use the
440 typedef float FFTSample;
442 typedef struct FFTComplex {
446 typedef struct FFTContext {
451 FFTComplex *exptab1; /* only used by SSE code */
452 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
455 int fft_init(FFTContext *s, int nbits, int inverse);
456 void fft_permute(FFTContext *s, FFTComplex *z);
457 void fft_calc_c(FFTContext *s, FFTComplex *z);
458 void fft_calc_sse(FFTContext *s, FFTComplex *z);
459 void fft_calc_altivec(FFTContext *s, FFTComplex *z);
461 static inline void fft_calc(FFTContext *s, FFTComplex *z)
465 void fft_end(FFTContext *s);
467 /* MDCT computation */
469 typedef struct MDCTContext {
470 int n; /* size of MDCT (i.e. number of input data * 2) */
471 int nbits; /* n = 2^nbits */
472 /* pre/post rotation tables */
478 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
479 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
480 const FFTSample *input, FFTSample *tmp);
481 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
482 const FFTSample *input, FFTSample *tmp);
483 void ff_mdct_end(MDCTContext *s);
485 #define WARPER8_16(name8, name16)\
486 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
487 return name8(s, dst , src , stride, h)\
488 +name8(s, dst+8 , src+8 , stride, h);\
491 #define WARPER8_16_SQ(name8, name16)\
492 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
494 score +=name8(s, dst , src , stride, 8);\
495 score +=name8(s, dst+8 , src+8 , stride, 8);\
499 score +=name8(s, dst , src , stride, 8);\
500 score +=name8(s, dst+8 , src+8 , stride, 8);\
506 /* XXX: add ISOC specific test to avoid specific BSD testing. */
507 /* better than nothing implementation. */
508 /* btw, rintf() is existing on fbsd too -- alex */
509 static inline long int lrintf(float x)
512 /* XXX: incorrect, but make it compile */
515 return (int)(rint(x));