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);
45 void j_rev_dct4 (DCTELEM *data);
46 void j_rev_dct2 (DCTELEM *data);
47 void j_rev_dct1 (DCTELEM *data);
49 void ff_fdct_mmx(DCTELEM *block);
50 void ff_fdct_mmx2(DCTELEM *block);
51 void ff_fdct_sse2(DCTELEM *block);
53 void ff_h264_idct_add_c(uint8_t *dst, DCTELEM *block, int stride);
54 void ff_h264_lowres_idct_add_c(uint8_t *dst, int stride, DCTELEM *block);
55 void ff_h264_lowres_idct_put_c(uint8_t *dst, int stride, DCTELEM *block);
58 extern const uint8_t ff_alternate_horizontal_scan[64];
59 extern const uint8_t ff_alternate_vertical_scan[64];
60 extern const uint8_t ff_zigzag_direct[64];
61 extern const uint8_t ff_zigzag248_direct[64];
63 /* pixel operations */
64 #define MAX_NEG_CROP 1024
67 extern uint32_t squareTbl[512];
68 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
70 /* VP3 DSP functions */
71 void vp3_dsp_init_c(void);
72 void vp3_idct_c(int16_t *input_data, int16_t *dequant_matrix,
73 int coeff_count, DCTELEM *output_data);
75 void vp3_dsp_init_mmx(void);
76 void vp3_idct_mmx(int16_t *input_data, int16_t *dequant_matrix,
77 int coeff_count, DCTELEM *output_data);
79 void vp3_dsp_init_sse2(void);
80 void vp3_idct_sse2(int16_t *input_data, int16_t *dequant_matrix,
81 int coeff_count, DCTELEM *output_data);
83 /* minimum alignment rules ;)
84 if u notice errors in the align stuff, need more alignment for some asm code for some cpu
85 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
87 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
88 i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
90 !future video codecs might need functions with less strict alignment
94 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
95 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
96 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
97 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
98 void clear_blocks_c(DCTELEM *blocks);
101 /* add and put pixel (decoding) */
102 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
103 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
104 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
105 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);
106 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
107 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
109 #define DEF_OLD_QPEL(name)\
110 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
111 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
112 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
114 DEF_OLD_QPEL(qpel16_mc11_old_c)
115 DEF_OLD_QPEL(qpel16_mc31_old_c)
116 DEF_OLD_QPEL(qpel16_mc12_old_c)
117 DEF_OLD_QPEL(qpel16_mc32_old_c)
118 DEF_OLD_QPEL(qpel16_mc13_old_c)
119 DEF_OLD_QPEL(qpel16_mc33_old_c)
120 DEF_OLD_QPEL(qpel8_mc11_old_c)
121 DEF_OLD_QPEL(qpel8_mc31_old_c)
122 DEF_OLD_QPEL(qpel8_mc12_old_c)
123 DEF_OLD_QPEL(qpel8_mc32_old_c)
124 DEF_OLD_QPEL(qpel8_mc13_old_c)
125 DEF_OLD_QPEL(qpel8_mc33_old_c)
127 #define CALL_2X_PIXELS(a, b, n)\
128 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
129 b(block , pixels , line_size, h);\
130 b(block+n, pixels+n, line_size, h);\
133 /* motion estimation */
134 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
135 // allthough currently h<4 is not used as functions with width <8 are not used and neither implemented
136 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))*/;
142 typedef struct DSPContext {
143 /* pixel ops : interface with DCT */
144 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
145 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
146 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
147 void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
148 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
150 * translational global motion compensation.
152 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
154 * global motion compensation.
156 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
157 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
158 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
159 int (*pix_sum)(uint8_t * pix, int line_size);
160 int (*pix_norm1)(uint8_t * pix, int line_size);
161 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
163 me_cmp_func sad[5]; /* identical to pix_absAxA except additional void * */
165 me_cmp_func hadamard8_diff[5];
166 me_cmp_func dct_sad[5];
167 me_cmp_func quant_psnr[5];
175 me_cmp_func dct_max[5];
177 me_cmp_func me_pre_cmp[5];
178 me_cmp_func me_cmp[5];
179 me_cmp_func me_sub_cmp[5];
180 me_cmp_func mb_cmp[5];
181 me_cmp_func ildct_cmp[5]; //only width 16 used
182 me_cmp_func frame_skip_cmp[5]; //only width 8 used
185 * Halfpel motion compensation with rounding (a+b+1)>>1.
186 * this is an array[4][4] of motion compensation funcions 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 where the result is stored
190 * @param pixels source
191 * @param line_size number of bytes in a horizontal line of block
194 op_pixels_func put_pixels_tab[4][4];
197 * Halfpel motion compensation with rounding (a+b+1)>>1.
198 * This is an array[4][4] of motion compensation functions for 4
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 into which the result is averaged (a+b+1)>>1
202 * @param pixels source
203 * @param line_size number of bytes in a horizontal line of block
206 op_pixels_func avg_pixels_tab[4][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 where the result is stored
214 * @param pixels source
215 * @param line_size number of bytes in a horizontal line of block
218 op_pixels_func put_no_rnd_pixels_tab[4][4];
221 * Halfpel motion compensation with no rounding (a+b)>>1.
222 * this is an array[2][4] of motion compensation funcions for 2
223 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
224 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
225 * @param block destination into which the result is averaged (a+b)>>1
226 * @param pixels source
227 * @param line_size number of bytes in a horizontal line of block
230 op_pixels_func avg_no_rnd_pixels_tab[4][4];
232 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);
235 * Thirdpel motion compensation with rounding (a+b+1)>>1.
236 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
237 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
238 * @param block destination where the result is stored
239 * @param pixels source
240 * @param line_size number of bytes in a horizontal line of block
243 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
244 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
246 qpel_mc_func put_qpel_pixels_tab[2][16];
247 qpel_mc_func avg_qpel_pixels_tab[2][16];
248 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
249 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
250 qpel_mc_func put_mspel_pixels_tab[8];
255 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
256 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
258 qpel_mc_func put_h264_qpel_pixels_tab[3][16];
259 qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
261 me_cmp_func pix_abs[2][4];
263 /* huffyuv specific */
264 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
265 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
267 * subtract huffyuv's variant of median prediction
268 * note, this might read from src1[-1], src2[-1]
270 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
271 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
273 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
274 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
276 void (*h261_loop_filter)(uint8_t *src, int stride);
279 void (*fdct)(DCTELEM *block/* align 16*/);
280 void (*fdct248)(DCTELEM *block/* align 16*/);
283 void (*idct)(DCTELEM *block/* align 16*/);
286 * block -> idct -> clip to unsigned 8 bit -> dest.
287 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
288 * @param line_size size in bytes of a horizotal line of dest
290 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
293 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
294 * @param line_size size in bytes of a horizotal line of dest
296 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
299 * idct input permutation.
300 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
302 * this permutation must be performed before the idct_put/add, note, normally this can be merged
303 * with the zigzag/alternate scan<br>
304 * an example to avoid confusion:
305 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
306 * - (x -> referece dct -> reference idct -> x)
307 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
308 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
310 uint8_t idct_permutation[64];
311 int idct_permutation_type;
312 #define FF_NO_IDCT_PERM 1
313 #define FF_LIBMPEG2_IDCT_PERM 2
314 #define FF_SIMPLE_IDCT_PERM 3
315 #define FF_TRANSPOSE_IDCT_PERM 4
317 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
318 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
319 #define BASIS_SHIFT 16
320 #define RECON_SHIFT 6
323 * This function handles any initialization for the VP3 DSP functions.
325 void (*vp3_dsp_init)(void);
328 * This function is responsible for taking a block of zigzag'd,
329 * quantized DCT coefficients and reconstructing the original block of
331 * @param input_data 64 zigzag'd, quantized DCT coefficients
332 * @param dequant_matrix 64 zigzag'd quantizer coefficients
333 * @param coeff_count index of the last coefficient
334 * @param output_samples space for 64 DCTELEMs where the transformed
335 * samples will be stored
337 void (*vp3_idct)(int16_t *input_data, int16_t *dequant_matrix,
338 int coeff_count, DCTELEM *output_samples);
340 void (*h264_idct_add)(uint8_t *dst, DCTELEM *block, int stride);
343 void dsputil_static_init(void);
344 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
347 * permute block according to permuatation.
348 * @param last last non zero element in scantable order
350 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
352 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
354 #define BYTE_VEC32(c) ((c)*0x01010101UL)
356 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
358 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
361 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
363 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
366 static inline int get_penalty_factor(int lambda, int lambda2, int type){
370 return lambda>>FF_LAMBDA_SHIFT;
372 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
374 return (4*lambda)>>(FF_LAMBDA_SHIFT);
376 return (2*lambda)>>(FF_LAMBDA_SHIFT);
378 return (2*lambda)>>FF_LAMBDA_SHIFT;
383 return lambda2>>FF_LAMBDA_SHIFT;
391 * this must be called between any dsp function and float/double code.
392 * for example sin(); dsp->idct_put(); emms_c(); cos()
396 /* should be defined by architectures supporting
397 one or more MultiMedia extension */
398 int mm_support(void);
400 #define __align16 __attribute__ ((aligned (16)))
402 #if defined(HAVE_MMX)
406 #define MM_MMX 0x0001 /* standard MMX */
407 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
408 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
409 #define MM_SSE 0x0008 /* SSE functions */
410 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
411 #define MM_3DNOWEXT 0x0020 /* AMD 3DNowExt */
415 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
416 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
417 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
419 static inline void emms(void)
421 __asm __volatile ("emms;":::"memory");
427 if (mm_flags & MM_MMX)\
431 #define __align8 __attribute__ ((aligned (8)))
432 #define STRIDE_ALIGN 8
434 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
435 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
437 #elif defined(ARCH_ARMV4L)
439 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
440 line optimizations */
441 #define __align8 __attribute__ ((aligned (4)))
442 #define STRIDE_ALIGN 4
444 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
446 #elif defined(HAVE_MLIB)
448 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
449 #define __align8 __attribute__ ((aligned (8)))
450 #define STRIDE_ALIGN 8
452 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
454 #elif defined(ARCH_SPARC)
456 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
457 #define __align8 __attribute__ ((aligned (8)))
458 #define STRIDE_ALIGN 8
459 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
461 #elif defined(ARCH_ALPHA)
463 #define __align8 __attribute__ ((aligned (8)))
464 #define STRIDE_ALIGN 8
466 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
468 #elif defined(ARCH_POWERPC)
470 #define MM_ALTIVEC 0x0001 /* standard AltiVec */
474 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
475 #define pixel altivec_pixel
480 #define __align8 __attribute__ ((aligned (16)))
481 #define STRIDE_ALIGN 16
483 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
485 #elif defined(HAVE_MMI)
487 #define __align8 __attribute__ ((aligned (16)))
488 #define STRIDE_ALIGN 16
490 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
492 #elif defined(ARCH_SH4)
494 #define __align8 __attribute__ ((aligned (8)))
495 #define STRIDE_ALIGN 8
497 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
501 #define __align8 __attribute__ ((aligned (8)))
502 #define STRIDE_ALIGN 8
508 struct unaligned_64 { uint64_t l; } __attribute__((packed));
509 struct unaligned_32 { uint32_t l; } __attribute__((packed));
510 struct unaligned_16 { uint16_t l; } __attribute__((packed));
512 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
513 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
514 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
516 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
520 #define LD16(a) (*((uint16_t*)(a)))
521 #define LD32(a) (*((uint32_t*)(a)))
522 #define LD64(a) (*((uint64_t*)(a)))
524 #define ST32(a, b) *((uint32_t*)(a)) = (b)
526 #endif /* !__GNUC__ */
529 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
530 int orig_linesize[3], int coded_linesize,
531 AVCodecContext *avctx);
533 /* FFT computation */
535 /* NOTE: soon integer code will be added, so you must use the
537 typedef float FFTSample;
539 typedef struct FFTComplex {
543 typedef struct FFTContext {
548 FFTComplex *exptab1; /* only used by SSE code */
549 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
552 int ff_fft_init(FFTContext *s, int nbits, int inverse);
553 void ff_fft_permute(FFTContext *s, FFTComplex *z);
554 void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
555 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z);
556 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
558 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z)
562 void ff_fft_end(FFTContext *s);
564 /* MDCT computation */
566 typedef struct MDCTContext {
567 int n; /* size of MDCT (i.e. number of input data * 2) */
568 int nbits; /* n = 2^nbits */
569 /* pre/post rotation tables */
575 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
576 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
577 const FFTSample *input, FFTSample *tmp);
578 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
579 const FFTSample *input, FFTSample *tmp);
580 void ff_mdct_end(MDCTContext *s);
582 #define WARPER8_16(name8, name16)\
583 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
584 return name8(s, dst , src , stride, h)\
585 +name8(s, dst+8 , src+8 , stride, h);\
588 #define WARPER8_16_SQ(name8, name16)\
589 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
591 score +=name8(s, dst , src , stride, 8);\
592 score +=name8(s, dst+8 , src+8 , stride, 8);\
596 score +=name8(s, dst , src , stride, 8);\
597 score +=name8(s, dst+8 , src+8 , stride, 8);\
603 /* XXX: add ISOC specific test to avoid specific BSD testing. */
604 /* better than nothing implementation. */
605 /* btw, rintf() is existing on fbsd too -- alex */
606 static always_inline long int lrintf(float x)
613 : "=m" (i) : "t" (x) : "st"
617 /* XXX: incorrect, but make it compile */
618 return (int)(x + (x < 0 ? -0.5 : 0.5));
621 return (int)(rint(x));
625 #ifndef _ISOC9X_SOURCE
626 #define _ISOC9X_SOURCE