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 1024
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];
169 me_cmp_func me_pre_cmp[5];
170 me_cmp_func me_cmp[5];
171 me_cmp_func me_sub_cmp[5];
172 me_cmp_func mb_cmp[5];
173 me_cmp_func ildct_cmp[5]; //only width 16 used
176 * Halfpel motion compensation with rounding (a+b+1)>>1.
177 * this is an array[4][4] of motion compensation funcions for 4
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
185 op_pixels_func put_pixels_tab[4][4];
188 * Halfpel motion compensation with rounding (a+b+1)>>1.
189 * This is an array[4][4] of motion compensation functions for 4
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)>>1
193 * @param pixels source
194 * @param line_size number of bytes in a horizontal line of block
197 op_pixels_func avg_pixels_tab[4][4];
200 * Halfpel motion compensation with no rounding (a+b)>>1.
201 * this is an array[2][4] of motion compensation funcions for 2
202 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
203 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
204 * @param block destination where the result is stored
205 * @param pixels source
206 * @param line_size number of bytes in a horizontal line of block
209 op_pixels_func put_no_rnd_pixels_tab[4][4];
212 * Halfpel motion compensation with no rounding (a+b)>>1.
213 * this is an array[2][4] of motion compensation funcions for 2
214 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
215 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
216 * @param block destination into which the result is averaged (a+b)>>1
217 * @param pixels source
218 * @param line_size number of bytes in a horizontal line of block
221 op_pixels_func avg_no_rnd_pixels_tab[4][4];
223 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);
226 * Thirdpel motion compensation with rounding (a+b+1)>>1.
227 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
228 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
229 * @param block destination where the result is stored
230 * @param pixels source
231 * @param line_size number of bytes in a horizontal line of block
234 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
235 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
237 qpel_mc_func put_qpel_pixels_tab[2][16];
238 qpel_mc_func avg_qpel_pixels_tab[2][16];
239 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
240 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
241 qpel_mc_func put_mspel_pixels_tab[8];
246 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
247 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
249 qpel_mc_func put_h264_qpel_pixels_tab[3][16];
250 qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
252 me_cmp_func pix_abs[2][4];
254 /* huffyuv specific */
255 void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
256 void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
258 * subtract huffyuv's variant of median prediction
259 * note, this might read from src1[-1], src2[-1]
261 void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
262 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
264 void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
265 void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
267 void (*h261_loop_filter)(uint8_t *src, int stride);
270 void (*fdct)(DCTELEM *block/* align 16*/);
271 void (*fdct248)(DCTELEM *block/* align 16*/);
274 void (*idct)(DCTELEM *block/* align 16*/);
277 * block -> idct -> clip to unsigned 8 bit -> dest.
278 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
279 * @param line_size size in bytes of a horizotal line of dest
281 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
284 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
285 * @param line_size size in bytes of a horizotal line of dest
287 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
290 * idct input permutation.
291 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
293 * this permutation must be performed before the idct_put/add, note, normally this can be merged
294 * with the zigzag/alternate scan<br>
295 * an example to avoid confusion:
296 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
297 * - (x -> referece dct -> reference idct -> x)
298 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
299 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
301 uint8_t idct_permutation[64];
302 int idct_permutation_type;
303 #define FF_NO_IDCT_PERM 1
304 #define FF_LIBMPEG2_IDCT_PERM 2
305 #define FF_SIMPLE_IDCT_PERM 3
306 #define FF_TRANSPOSE_IDCT_PERM 4
308 int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
309 void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
310 #define BASIS_SHIFT 16
311 #define RECON_SHIFT 6
314 * This function handles any initialization for the VP3 DSP functions.
316 void (*vp3_dsp_init)(void);
319 * This function is responsible for taking a block of zigzag'd,
320 * quantized DCT coefficients and reconstructing the original block of
322 * @param input_data 64 zigzag'd, quantized DCT coefficients
323 * @param dequant_matrix 64 zigzag'd quantizer coefficients
324 * @param coeff_count index of the last coefficient
325 * @param output_samples space for 64 DCTELEMs where the transformed
326 * samples will be stored
328 void (*vp3_idct)(int16_t *input_data, int16_t *dequant_matrix,
329 int coeff_count, DCTELEM *output_samples);
333 void dsputil_static_init(void);
334 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
337 * permute block according to permuatation.
338 * @param last last non zero element in scantable order
340 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
342 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
344 #define BYTE_VEC32(c) ((c)*0x01010101UL)
346 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
348 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
351 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
353 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
356 static inline int get_penalty_factor(int lambda, int lambda2, int type){
360 return lambda>>FF_LAMBDA_SHIFT;
362 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
364 return (4*lambda)>>(FF_LAMBDA_SHIFT);
366 return (2*lambda)>>(FF_LAMBDA_SHIFT);
368 return (2*lambda)>>FF_LAMBDA_SHIFT;
373 return lambda2>>FF_LAMBDA_SHIFT;
381 * this must be called between any dsp function and float/double code.
382 * for example sin(); dsp->idct_put(); emms_c(); cos()
386 /* should be defined by architectures supporting
387 one or more MultiMedia extension */
388 int mm_support(void);
390 #define __align16 __attribute__ ((aligned (16)))
392 #if defined(HAVE_MMX)
396 #define MM_MMX 0x0001 /* standard MMX */
397 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
398 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
399 #define MM_SSE 0x0008 /* SSE functions */
400 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
404 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
405 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
406 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
408 static inline void emms(void)
410 __asm __volatile ("emms;":::"memory");
416 if (mm_flags & MM_MMX)\
420 #define __align8 __attribute__ ((aligned (8)))
422 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
423 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
425 #elif defined(ARCH_ARMV4L)
427 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
428 line optimizations */
429 #define __align8 __attribute__ ((aligned (4)))
431 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
433 #elif defined(HAVE_MLIB)
435 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
436 #define __align8 __attribute__ ((aligned (8)))
438 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
440 #elif defined(ARCH_SPARC)
442 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
443 #define __align8 __attribute__ ((aligned (8)))
444 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
446 #elif defined(ARCH_ALPHA)
448 #define __align8 __attribute__ ((aligned (8)))
450 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
452 #elif defined(ARCH_POWERPC)
454 #define MM_ALTIVEC 0x0001 /* standard AltiVec */
458 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
459 #define pixel altivec_pixel
464 #define __align8 __attribute__ ((aligned (16)))
466 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
468 #elif defined(HAVE_MMI)
470 #define __align8 __attribute__ ((aligned (16)))
472 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
474 #elif defined(ARCH_SH4)
476 #define __align8 __attribute__ ((aligned (8)))
478 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
488 struct unaligned_64 { uint64_t l; } __attribute__((packed));
489 struct unaligned_32 { uint32_t l; } __attribute__((packed));
490 struct unaligned_16 { uint16_t l; } __attribute__((packed));
492 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
493 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
494 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
496 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
500 #define LD16(a) (*((uint16_t*)(a)))
501 #define LD32(a) (*((uint32_t*)(a)))
502 #define LD64(a) (*((uint64_t*)(a)))
504 #define ST32(a, b) *((uint32_t*)(a)) = (b)
506 #endif /* !__GNUC__ */
509 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
510 int orig_linesize[3], int coded_linesize,
511 AVCodecContext *avctx);
513 /* FFT computation */
515 /* NOTE: soon integer code will be added, so you must use the
517 typedef float FFTSample;
519 typedef struct FFTComplex {
523 typedef struct FFTContext {
528 FFTComplex *exptab1; /* only used by SSE code */
529 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
532 int ff_fft_init(FFTContext *s, int nbits, int inverse);
533 void ff_fft_permute(FFTContext *s, FFTComplex *z);
534 void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
535 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z);
536 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
538 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z)
542 void ff_fft_end(FFTContext *s);
544 /* MDCT computation */
546 typedef struct MDCTContext {
547 int n; /* size of MDCT (i.e. number of input data * 2) */
548 int nbits; /* n = 2^nbits */
549 /* pre/post rotation tables */
555 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
556 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
557 const FFTSample *input, FFTSample *tmp);
558 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
559 const FFTSample *input, FFTSample *tmp);
560 void ff_mdct_end(MDCTContext *s);
562 #define WARPER8_16(name8, name16)\
563 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
564 return name8(s, dst , src , stride, h)\
565 +name8(s, dst+8 , src+8 , stride, h);\
568 #define WARPER8_16_SQ(name8, name16)\
569 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
571 score +=name8(s, dst , src , stride, 8);\
572 score +=name8(s, dst+8 , src+8 , stride, 8);\
576 score +=name8(s, dst , src , stride, 8);\
577 score +=name8(s, dst+8 , src+8 , stride, 8);\
583 /* XXX: add ISOC specific test to avoid specific BSD testing. */
584 /* better than nothing implementation. */
585 /* btw, rintf() is existing on fbsd too -- alex */
586 static always_inline long int lrintf(float x)
593 : "=m" (i) : "t" (x) : "st"
597 /* XXX: incorrect, but make it compile */
598 return (int)(x + (x < 0 ? -0.5 : 0.5));
601 return (int)(rint(x));
605 #ifndef _ISOC9X_SOURCE
606 #define _ISOC9X_SOURCE