3 * Copyright (c) 2000, 2001, 2002 Fabrice Bellard.
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
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.
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.
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
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
37 typedef short DCTELEM;
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);
44 void j_rev_dct (DCTELEM *data);
46 void ff_fdct_mmx(DCTELEM *block);
47 void ff_fdct_mmx2(DCTELEM *block);
48 void ff_fdct_sse2(DCTELEM *block);
51 extern const uint8_t ff_alternate_horizontal_scan[64];
52 extern const uint8_t ff_alternate_vertical_scan[64];
53 extern const uint8_t ff_zigzag_direct[64];
54 extern const uint8_t ff_zigzag248_direct[64];
56 /* pixel operations */
57 #define MAX_NEG_CROP 384
60 extern uint32_t squareTbl[512];
61 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
63 /* VP3 DSP functions */
64 void vp3_dsp_init_c(void);
65 void vp3_idct_c(int16_t *input_data, int16_t *dequant_matrix,
66 int coeff_count, DCTELEM *output_data);
68 void vp3_dsp_init_mmx(void);
69 void vp3_idct_mmx(int16_t *input_data, int16_t *dequant_matrix,
70 int coeff_count, DCTELEM *output_data);
72 void vp3_dsp_init_sse2(void);
73 void vp3_idct_sse2(int16_t *input_data, int16_t *dequant_matrix,
74 int coeff_count, DCTELEM *output_data);
76 /* minimum alignment rules ;)
77 if u notice errors in the align stuff, need more alignment for some asm code for some cpu
78 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
80 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
81 i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
83 !future video codecs might need functions with less strict alignment
87 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
88 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
89 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
90 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
91 void clear_blocks_c(DCTELEM *blocks);
94 /* add and put pixel (decoding) */
95 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
96 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
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);
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);
99 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
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);
102 #define DEF_OLD_QPEL(name)\
103 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
104 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
105 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
107 DEF_OLD_QPEL(qpel16_mc11_old_c)
108 DEF_OLD_QPEL(qpel16_mc31_old_c)
109 DEF_OLD_QPEL(qpel16_mc12_old_c)
110 DEF_OLD_QPEL(qpel16_mc32_old_c)
111 DEF_OLD_QPEL(qpel16_mc13_old_c)
112 DEF_OLD_QPEL(qpel16_mc33_old_c)
113 DEF_OLD_QPEL(qpel8_mc11_old_c)
114 DEF_OLD_QPEL(qpel8_mc31_old_c)
115 DEF_OLD_QPEL(qpel8_mc12_old_c)
116 DEF_OLD_QPEL(qpel8_mc32_old_c)
117 DEF_OLD_QPEL(qpel8_mc13_old_c)
118 DEF_OLD_QPEL(qpel8_mc33_old_c)
120 #define CALL_2X_PIXELS(a, b, n)\
121 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
122 b(block , pixels , line_size, h);\
123 b(block+n, pixels+n, line_size, h);\
126 /* motion estimation */
127 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
128 // allthough currently h<4 is not used as functions with width <8 are not used and neither implemented
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))*/;
135 typedef struct DSPContext {
136 /* pixel ops : interface with DCT */
137 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
138 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
139 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
140 void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
141 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
143 * translational global motion compensation.
145 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
147 * global motion compensation.
149 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
150 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
151 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
152 int (*pix_sum)(uint8_t * pix, int line_size);
153 int (*pix_norm1)(uint8_t * pix, int line_size);
154 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
156 me_cmp_func sad[5]; /* identical to pix_absAxA except additional void * */
158 me_cmp_func hadamard8_diff[5];
159 me_cmp_func dct_sad[5];
160 me_cmp_func quant_psnr[5];
166 me_cmp_func me_pre_cmp[5];
167 me_cmp_func me_cmp[5];
168 me_cmp_func me_sub_cmp[5];
169 me_cmp_func mb_cmp[5];
170 me_cmp_func ildct_cmp[5]; //only width 16 used
173 * Halfpel motion compensation with rounding (a+b+1)>>1.
174 * this is an array[4][4] of motion compensation funcions for 4
175 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
176 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
177 * @param block destination where the result is stored
178 * @param pixels source
179 * @param line_size number of bytes in a horizontal line of block
182 op_pixels_func put_pixels_tab[4][4];
185 * Halfpel motion compensation with rounding (a+b+1)>>1.
186 * This is an array[4][4] of motion compensation functions for 4
187 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
188 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
189 * @param block destination into which the result is averaged (a+b+1)>>1
190 * @param pixels source
191 * @param line_size number of bytes in a horizontal line of block
194 op_pixels_func avg_pixels_tab[4][4];
197 * Halfpel motion compensation with no rounding (a+b)>>1.
198 * this is an array[2][4] of motion compensation funcions for 2
199 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
200 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
201 * @param block destination where the result is stored
202 * @param pixels source
203 * @param line_size number of bytes in a horizontal line of block
206 op_pixels_func put_no_rnd_pixels_tab[2][4];
209 * Halfpel motion compensation with no rounding (a+b)>>1.
210 * this is an array[2][4] of motion compensation funcions for 2
211 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
212 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
213 * @param block destination into which the result is averaged (a+b)>>1
214 * @param pixels source
215 * @param line_size number of bytes in a horizontal line of block
218 op_pixels_func avg_no_rnd_pixels_tab[2][4];
220 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);
223 * Thirdpel motion compensation with rounding (a+b+1)>>1.
224 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
225 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
226 * @param block destination where the result is stored
227 * @param pixels source
228 * @param line_size number of bytes in a horizontal line of block
231 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
232 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
234 qpel_mc_func put_qpel_pixels_tab[2][16];
235 qpel_mc_func avg_qpel_pixels_tab[2][16];
236 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
237 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
238 qpel_mc_func put_mspel_pixels_tab[8];
243 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
244 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
246 qpel_mc_func put_h264_qpel_pixels_tab[3][16];
247 qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
249 me_cmp_func pix_abs[2][4];
251 /* huffyuv specific */
252 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
253 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
255 * subtract huffyuv's variant of median prediction
256 * note, this might read from src1[-1], src2[-1]
258 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
259 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
261 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
262 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
265 void (*fdct)(DCTELEM *block/* align 16*/);
266 void (*fdct248)(DCTELEM *block/* align 16*/);
269 void (*idct)(DCTELEM *block/* align 16*/);
272 * block -> idct -> clip to unsigned 8 bit -> dest.
273 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
274 * @param line_size size in bytes of a horizotal line of dest
276 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
279 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
280 * @param line_size size in bytes of a horizotal line of dest
282 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
285 * idct input permutation.
286 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
288 * this permutation must be performed before the idct_put/add, note, normally this can be merged
289 * with the zigzag/alternate scan<br>
290 * an example to avoid confusion:
291 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
292 * - (x -> referece dct -> reference idct -> x)
293 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
294 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
296 uint8_t idct_permutation[64];
297 int idct_permutation_type;
298 #define FF_NO_IDCT_PERM 1
299 #define FF_LIBMPEG2_IDCT_PERM 2
300 #define FF_SIMPLE_IDCT_PERM 3
301 #define FF_TRANSPOSE_IDCT_PERM 4
303 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
304 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
305 #define BASIS_SHIFT 16
306 #define RECON_SHIFT 6
309 * This function handles any initialization for the VP3 DSP functions.
311 void (*vp3_dsp_init)(void);
314 * This function is responsible for taking a block of zigzag'd,
315 * quantized DCT coefficients and reconstructing the original block of
317 * @param input_data 64 zigzag'd, quantized DCT coefficients
318 * @param dequant_matrix 64 zigzag'd quantizer coefficients
319 * @param coeff_count index of the last coefficient
320 * @param output_samples space for 64 DCTELEMs where the transformed
321 * samples will be stored
323 void (*vp3_idct)(int16_t *input_data, int16_t *dequant_matrix,
324 int coeff_count, DCTELEM *output_samples);
328 void dsputil_static_init(void);
329 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
332 * permute block according to permuatation.
333 * @param last last non zero element in scantable order
335 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
337 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
339 #define BYTE_VEC32(c) ((c)*0x01010101UL)
341 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
343 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
346 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
348 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
353 * this must be called between any dsp function and float/double code.
354 * for example sin(); dsp->idct_put(); emms_c(); cos()
358 /* should be defined by architectures supporting
359 one or more MultiMedia extension */
360 int mm_support(void);
362 #define __align16 __attribute__ ((aligned (16)))
364 #if defined(HAVE_MMX)
368 #define MM_MMX 0x0001 /* standard MMX */
369 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
370 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
371 #define MM_SSE 0x0008 /* SSE functions */
372 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
376 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
377 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
378 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
380 static inline void emms(void)
382 __asm __volatile ("emms;":::"memory");
388 if (mm_flags & MM_MMX)\
392 #define __align8 __attribute__ ((aligned (8)))
394 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
395 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
397 #elif defined(ARCH_ARMV4L)
399 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
400 line optimizations */
401 #define __align8 __attribute__ ((aligned (4)))
403 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
405 #elif defined(HAVE_MLIB)
407 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
408 #define __align8 __attribute__ ((aligned (8)))
410 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
412 #elif defined(ARCH_SPARC)
414 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
415 #define __align8 __attribute__ ((aligned (8)))
416 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
418 #elif defined(ARCH_ALPHA)
420 #define __align8 __attribute__ ((aligned (8)))
422 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
424 #elif defined(ARCH_POWERPC)
426 #define MM_ALTIVEC 0x0001 /* standard AltiVec */
430 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
431 #define pixel altivec_pixel
436 #define __align8 __attribute__ ((aligned (16)))
438 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
440 #elif defined(HAVE_MMI)
442 #define __align8 __attribute__ ((aligned (16)))
444 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
446 #elif defined(ARCH_SH4)
448 #define __align8 __attribute__ ((aligned (8)))
450 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
460 struct unaligned_64 { uint64_t l; } __attribute__((packed));
461 struct unaligned_32 { uint32_t l; } __attribute__((packed));
462 struct unaligned_16 { uint16_t l; } __attribute__((packed));
464 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
465 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
466 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
468 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
472 #define LD16(a) (*((uint16_t*)(a)))
473 #define LD32(a) (*((uint32_t*)(a)))
474 #define LD64(a) (*((uint64_t*)(a)))
476 #define ST32(a, b) *((uint32_t*)(a)) = (b)
478 #endif /* !__GNUC__ */
481 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
482 int orig_linesize[3], int coded_linesize,
483 AVCodecContext *avctx);
485 /* FFT computation */
487 /* NOTE: soon integer code will be added, so you must use the
489 typedef float FFTSample;
491 typedef struct FFTComplex {
495 typedef struct FFTContext {
500 FFTComplex *exptab1; /* only used by SSE code */
501 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
504 int ff_fft_init(FFTContext *s, int nbits, int inverse);
505 void ff_fft_permute(FFTContext *s, FFTComplex *z);
506 void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
507 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z);
508 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
510 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z)
514 void ff_fft_end(FFTContext *s);
516 /* MDCT computation */
518 typedef struct MDCTContext {
519 int n; /* size of MDCT (i.e. number of input data * 2) */
520 int nbits; /* n = 2^nbits */
521 /* pre/post rotation tables */
527 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
528 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
529 const FFTSample *input, FFTSample *tmp);
530 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
531 const FFTSample *input, FFTSample *tmp);
532 void ff_mdct_end(MDCTContext *s);
534 #define WARPER8_16(name8, name16)\
535 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
536 return name8(s, dst , src , stride, h)\
537 +name8(s, dst+8 , src+8 , stride, h);\
540 #define WARPER8_16_SQ(name8, name16)\
541 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
543 score +=name8(s, dst , src , stride, 8);\
544 score +=name8(s, dst+8 , src+8 , stride, 8);\
548 score +=name8(s, dst , src , stride, 8);\
549 score +=name8(s, dst+8 , src+8 , stride, 8);\
555 /* XXX: add ISOC specific test to avoid specific BSD testing. */
556 /* better than nothing implementation. */
557 /* btw, rintf() is existing on fbsd too -- alex */
558 static inline long int lrintf(float x)
561 /* XXX: incorrect, but make it compile */
564 return (int)(rint(x));
568 #ifndef _ISOC9X_SOURCE
569 #define _ISOC9X_SOURCE