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 ff_jpeg_fdct_islow (DCTELEM *data);
41 void j_rev_dct (DCTELEM *data);
43 void ff_fdct_mmx(DCTELEM *block);
46 extern const uint8_t ff_alternate_horizontal_scan[64];
47 extern const uint8_t ff_alternate_vertical_scan[64];
48 extern const uint8_t ff_zigzag_direct[64];
50 /* pixel operations */
51 #define MAX_NEG_CROP 384
54 extern uint32_t squareTbl[512];
55 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
58 /* minimum alignment rules ;)
59 if u notice errors in the align stuff, need more alignment for some asm code for some cpu
60 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
62 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
63 i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
65 !future video codecs might need functions with less strict alignment
69 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
70 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
71 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
72 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
73 void clear_blocks_c(DCTELEM *blocks);
76 /* add and put pixel (decoding) */
77 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
78 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
79 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);
80 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
81 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
83 #define DEF_OLD_QPEL(name)\
84 void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
85 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
86 void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
88 DEF_OLD_QPEL(qpel16_mc11_old_c)
89 DEF_OLD_QPEL(qpel16_mc31_old_c)
90 DEF_OLD_QPEL(qpel16_mc12_old_c)
91 DEF_OLD_QPEL(qpel16_mc32_old_c)
92 DEF_OLD_QPEL(qpel16_mc13_old_c)
93 DEF_OLD_QPEL(qpel16_mc33_old_c)
94 DEF_OLD_QPEL(qpel8_mc11_old_c)
95 DEF_OLD_QPEL(qpel8_mc31_old_c)
96 DEF_OLD_QPEL(qpel8_mc12_old_c)
97 DEF_OLD_QPEL(qpel8_mc32_old_c)
98 DEF_OLD_QPEL(qpel8_mc13_old_c)
99 DEF_OLD_QPEL(qpel8_mc33_old_c)
101 #define CALL_2X_PIXELS(a, b, n)\
102 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
103 b(block , pixels , line_size, h);\
104 b(block+n, pixels+n, line_size, h);\
107 /* motion estimation */
109 typedef int (*op_pixels_abs_func)(uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size)/* __attribute__ ((const))*/;
111 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))*/;
117 typedef struct DSPContext {
118 /* pixel ops : interface with DCT */
119 void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
120 void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
121 void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
122 void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
124 * translational global motion compensation.
126 void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
128 * global motion compensation.
130 void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
131 int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
132 void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
133 int (*pix_sum)(uint8_t * pix, int line_size);
134 int (*pix_norm1)(uint8_t * pix, int line_size);
135 me_cmp_func sad[2]; /* identical to pix_absAxA except additional void * */
137 me_cmp_func hadamard8_diff[2];
138 me_cmp_func dct_sad[2];
139 me_cmp_func quant_psnr[2];
142 int (*hadamard8_abs )(uint8_t *src, int stride, int mean);
144 me_cmp_func me_pre_cmp[11];
145 me_cmp_func me_cmp[11];
146 me_cmp_func me_sub_cmp[11];
147 me_cmp_func mb_cmp[11];
149 /* maybe create an array for 16/8/4/2 functions */
151 * Halfpel motion compensation with rounding (a+b+1)>>1.
152 * this is an array[4][4] of motion compensation funcions for 4
153 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br>
154 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
155 * @param block destination where the result is stored
156 * @param pixels source
157 * @param line_size number of bytes in a horizontal line of block
160 op_pixels_func put_pixels_tab[4][4];
163 * Halfpel motion compensation with rounding (a+b+1)>>1.
164 * This is an array[4][4] of motion compensation functions for 4
165 * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br>
166 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
167 * @param block destination into which the result is averaged (a+b+1)>>1
168 * @param pixels source
169 * @param line_size number of bytes in a horizontal line of block
172 op_pixels_func avg_pixels_tab[4][4];
175 * Halfpel motion compensation with no rounding (a+b)>>1.
176 * this is an array[2][4] of motion compensation funcions for 2
177 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
178 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
179 * @param block destination where the result is stored
180 * @param pixels source
181 * @param line_size number of bytes in a horizontal line of block
184 op_pixels_func put_no_rnd_pixels_tab[2][4];
187 * Halfpel motion compensation with no rounding (a+b)>>1.
188 * this is an array[2][4] of motion compensation funcions for 2
189 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
190 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
191 * @param block destination into which the result is averaged (a+b)>>1
192 * @param pixels source
193 * @param line_size number of bytes in a horizontal line of block
196 op_pixels_func avg_no_rnd_pixels_tab[2][4];
199 * Thirdpel motion compensation with rounding (a+b+1)>>1.
200 * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
201 * *pixels_tab[ xthirdpel + 4*ythirdpel ]
202 * @param block destination where the result is stored
203 * @param pixels source
204 * @param line_size number of bytes in a horizontal line of block
207 tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
208 tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
210 qpel_mc_func put_qpel_pixels_tab[2][16];
211 qpel_mc_func avg_qpel_pixels_tab[2][16];
212 qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
213 qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
214 qpel_mc_func put_mspel_pixels_tab[8];
219 h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
220 h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
222 qpel_mc_func put_h264_qpel_pixels_tab[3][16];
223 qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
225 op_pixels_abs_func pix_abs16x16;
226 op_pixels_abs_func pix_abs16x16_x2;
227 op_pixels_abs_func pix_abs16x16_y2;
228 op_pixels_abs_func pix_abs16x16_xy2;
229 op_pixels_abs_func pix_abs8x8;
230 op_pixels_abs_func pix_abs8x8_x2;
231 op_pixels_abs_func pix_abs8x8_y2;
232 op_pixels_abs_func pix_abs8x8_xy2;
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);
237 void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
240 void (*fdct)(DCTELEM *block/* align 16*/);
243 void (*idct)(DCTELEM *block/* align 16*/);
246 * block -> idct -> clip to unsigned 8 bit -> dest.
247 * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
248 * @param line_size size in bytes of a horizotal line of dest
250 void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
253 * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
254 * @param line_size size in bytes of a horizotal line of dest
256 void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
259 * idct input permutation.
260 * several optimized IDCTs need a permutated input (relative to the normal order of the reference
262 * this permutation must be performed before the idct_put/add, note, normally this can be merged
263 * with the zigzag/alternate scan<br>
264 * an example to avoid confusion:
265 * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
266 * - (x -> referece dct -> reference idct -> x)
267 * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
268 * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
270 uint8_t idct_permutation[64];
271 int idct_permutation_type;
272 #define FF_NO_IDCT_PERM 1
273 #define FF_LIBMPEG2_IDCT_PERM 2
274 #define FF_SIMPLE_IDCT_PERM 3
275 #define FF_TRANSPOSE_IDCT_PERM 4
279 void dsputil_static_init(void);
280 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
283 * permute block according to permuatation.
284 * @param last last non zero element in scantable order
286 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
288 #define BYTE_VEC32(c) ((c)*0x01010101UL)
290 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
292 return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
295 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
297 return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
302 * this must be called between any dsp function and float/double code.
303 * for example sin(); dsp->idct_put(); emms_c(); cos()
307 /* should be defined by architectures supporting
308 one or more MultiMedia extension */
309 int mm_support(void);
311 #if defined(HAVE_MMX)
315 #define MM_MMX 0x0001 /* standard MMX */
316 #define MM_3DNOW 0x0004 /* AMD 3DNOW */
317 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
318 #define MM_SSE 0x0008 /* SSE functions */
319 #define MM_SSE2 0x0010 /* PIV SSE2 functions */
323 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
324 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
326 static inline void emms(void)
328 __asm __volatile ("emms;":::"memory");
334 if (mm_flags & MM_MMX)\
338 #define __align8 __attribute__ ((aligned (8)))
340 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
341 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
343 #elif defined(ARCH_ARMV4L)
345 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
347 #define __align8 __attribute__ ((aligned (4)))
349 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
351 #elif defined(HAVE_MLIB)
353 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
354 #define __align8 __attribute__ ((aligned (8)))
356 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
358 #elif defined(ARCH_ALPHA)
360 #define __align8 __attribute__ ((aligned (8)))
362 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
364 #elif defined(ARCH_POWERPC)
366 #define MM_ALTIVEC 0x0001 /* standard AltiVec */
370 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
374 #define __align8 __attribute__ ((aligned (16)))
376 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
378 #elif defined(HAVE_MMI)
380 #define __align8 __attribute__ ((aligned (16)))
382 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
384 #elif defined(ARCH_SH4)
386 #define __align8 __attribute__ ((aligned (8)))
388 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
398 struct unaligned_64 { uint64_t l; } __attribute__((packed));
399 struct unaligned_32 { uint32_t l; } __attribute__((packed));
400 struct unaligned_16 { uint16_t l; } __attribute__((packed));
402 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
403 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
404 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
406 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
410 #define LD16(a) (*((uint16_t*)(a)))
411 #define LD32(a) (*((uint32_t*)(a)))
412 #define LD64(a) (*((uint64_t*)(a)))
414 #define ST32(a, b) *((uint32_t*)(a)) = (b)
416 #endif /* !__GNUC__ */
419 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
420 int orig_linesize[3], int coded_linesize,
421 AVCodecContext *avctx);
423 /* FFT computation */
425 /* NOTE: soon integer code will be added, so you must use the
427 typedef float FFTSample;
429 typedef struct FFTComplex {
433 typedef struct FFTContext {
438 FFTComplex *exptab1; /* only used by SSE code */
439 void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
442 int fft_init(FFTContext *s, int nbits, int inverse);
443 void fft_permute(FFTContext *s, FFTComplex *z);
444 void fft_calc_c(FFTContext *s, FFTComplex *z);
445 void fft_calc_sse(FFTContext *s, FFTComplex *z);
446 void fft_calc_altivec(FFTContext *s, FFTComplex *z);
448 static inline void fft_calc(FFTContext *s, FFTComplex *z)
452 void fft_end(FFTContext *s);
454 /* MDCT computation */
456 typedef struct MDCTContext {
457 int n; /* size of MDCT (i.e. number of input data * 2) */
458 int nbits; /* n = 2^nbits */
459 /* pre/post rotation tables */
465 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
466 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
467 const FFTSample *input, FFTSample *tmp);
468 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
469 const FFTSample *input, FFTSample *tmp);
470 void ff_mdct_end(MDCTContext *s);
472 #define WARPER88_1616(name8, name16)\
473 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride){\
474 return name8(s, dst , src , stride)\
475 +name8(s, dst+8 , src+8 , stride)\
476 +name8(s, dst +8*stride, src +8*stride, stride)\
477 +name8(s, dst+8+8*stride, src+8+8*stride, stride);\
481 /* XXX: add ISOC specific test to avoid specific BSD testing. */
482 /* better than nothing implementation. */
483 /* btw, rintf() is existing on fbsd too -- alex */
484 static inline long int lrintf(float x)
487 /* XXX: incorrect, but make it compile */
490 return (int)(rint(x));
495 #if defined(CONFIG_OS2) || defined(CONFIG_SUNOS)
496 static inline float floorf(float f) {