libavcodec/dsputil.h

   1 /*
   2  * DSP utils
   3  * Copyright (c) 2000, 2001, 2002 Fabrice Bellard.
   4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
   5  *
   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.
  10  *
  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.
  15  *
  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
  19  */
  20
  21 /**
  22  * @file dsputil.h
  23  * DSP utils.
  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
  26  */
  27
  28 #ifndef DSPUTIL_H
  29 #define DSPUTIL_H
  30
  31 #include "common.h"
  32 #include "avcodec.h"
  33
  34
  35 //#define DEBUG
  36 /* dct code */
  37 typedef short DCTELEM;
  38
  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);
  43
  44 void j_rev_dct (DCTELEM *data);
  45 void j_rev_dct4 (DCTELEM *data);
  46
  47 void ff_fdct_mmx(DCTELEM *block);
  48 void ff_fdct_mmx2(DCTELEM *block);
  49 void ff_fdct_sse2(DCTELEM *block);
  50
  51 /* encoding scans */
  52 extern const uint8_t ff_alternate_horizontal_scan[64];
  53 extern const uint8_t ff_alternate_vertical_scan[64];
  54 extern const uint8_t ff_zigzag_direct[64];
  55 extern const uint8_t ff_zigzag248_direct[64];
  56
  57 /* pixel operations */
  58 #define MAX_NEG_CROP 1024
  59
  60 /* temporary */
  61 extern uint32_t squareTbl[512];
  62 extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
  63
  64 /* VP3 DSP functions */
  65 void vp3_dsp_init_c(void);
  66 void vp3_idct_c(int16_t *input_data, int16_t *dequant_matrix,
  67     int coeff_count, DCTELEM *output_data);
  68
  69 void vp3_dsp_init_mmx(void);
  70 void vp3_idct_mmx(int16_t *input_data, int16_t *dequant_matrix,
  71     int coeff_count, DCTELEM *output_data);
  72
  73 void vp3_dsp_init_sse2(void);
  74 void vp3_idct_sse2(int16_t *input_data, int16_t *dequant_matrix,
  75     int coeff_count, DCTELEM *output_data);
  76
  77 /* minimum alignment rules ;)
  78 if u notice errors in the align stuff, need more alignment for some asm code for some cpu
  79 or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
  80
  81 !warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
  82 i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
  83
  84 !future video codecs might need functions with less strict alignment
  85 */
  86
  87 /*
  88 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
  89 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
  90 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
  91 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
  92 void clear_blocks_c(DCTELEM *blocks);
  93 */
  94
  95 /* add and put pixel (decoding) */
  96 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
  97 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
  98 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
  99 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);
 100 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 101 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
 103 #define DEF_OLD_QPEL(name)\
 104 void ff_put_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 105 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 106 void ff_avg_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 107
 108 DEF_OLD_QPEL(qpel16_mc11_old_c)
 109 DEF_OLD_QPEL(qpel16_mc31_old_c)
 110 DEF_OLD_QPEL(qpel16_mc12_old_c)
 111 DEF_OLD_QPEL(qpel16_mc32_old_c)
 112 DEF_OLD_QPEL(qpel16_mc13_old_c)
 113 DEF_OLD_QPEL(qpel16_mc33_old_c)
 114 DEF_OLD_QPEL(qpel8_mc11_old_c)
 115 DEF_OLD_QPEL(qpel8_mc31_old_c)
 116 DEF_OLD_QPEL(qpel8_mc12_old_c)
 117 DEF_OLD_QPEL(qpel8_mc32_old_c)
 118 DEF_OLD_QPEL(qpel8_mc13_old_c)
 119 DEF_OLD_QPEL(qpel8_mc33_old_c)
 120
 121 #define CALL_2X_PIXELS(a, b, n)\
 122 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
 123     b(block  , pixels  , line_size, h);\
 124     b(block+n, pixels+n, line_size, h);\
 125 }
 126
 127 /* motion estimation */
 128 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
 129 // allthough currently h<4 is not used as functions with width <8 are not used and neither implemented
 130 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))*/;
 131
 132
 133 /**
 134  * DSPContext.
 135  */
 136 typedef struct DSPContext {
 137     /* pixel ops : interface with DCT */
 138     void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
 139     void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
 140     void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 141     void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 142     void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 143     /**
 144      * translational global motion compensation.
 145      */
 146     void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
 147     /**
 148      * global motion compensation.
 149      */
 150     void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
 151                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 152     void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
 153     int (*pix_sum)(uint8_t * pix, int line_size);
 154     int (*pix_norm1)(uint8_t * pix, int line_size);
 155 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
 156
 157     me_cmp_func sad[5]; /* identical to pix_absAxA except additional void * */
 158     me_cmp_func sse[5];
 159     me_cmp_func hadamard8_diff[5];
 160     me_cmp_func dct_sad[5];
 161     me_cmp_func quant_psnr[5];
 162     me_cmp_func bit[5];
 163     me_cmp_func rd[5];
 164     me_cmp_func vsad[5];
 165     me_cmp_func vsse[5];
 166     me_cmp_func nsse[5];
 167     me_cmp_func w53[5];
 168     me_cmp_func w97[5];
 169
 170     me_cmp_func me_pre_cmp[5];
 171     me_cmp_func me_cmp[5];
 172     me_cmp_func me_sub_cmp[5];
 173     me_cmp_func mb_cmp[5];
 174     me_cmp_func ildct_cmp[5]; //only width 16 used
 175
 176     /**
 177      * Halfpel motion compensation with rounding (a+b+1)>>1.
 178      * this is an array[4][4] of motion compensation funcions for 4
 179      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 180      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 181      * @param block destination where the result is stored
 182      * @param pixels source
 183      * @param line_size number of bytes in a horizontal line of block
 184      * @param h height
 185      */
 186     op_pixels_func put_pixels_tab[4][4];
 187
 188     /**
 189      * Halfpel motion compensation with rounding (a+b+1)>>1.
 190      * This is an array[4][4] of motion compensation functions for 4
 191      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 192      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 193      * @param block destination into which the result is averaged (a+b+1)>>1
 194      * @param pixels source
 195      * @param line_size number of bytes in a horizontal line of block
 196      * @param h height
 197      */
 198     op_pixels_func avg_pixels_tab[4][4];
 199
 200     /**
 201      * Halfpel motion compensation with no rounding (a+b)>>1.
 202      * this is an array[2][4] of motion compensation funcions for 2
 203      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 204      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 205      * @param block destination where the result is stored
 206      * @param pixels source
 207      * @param line_size number of bytes in a horizontal line of block
 208      * @param h height
 209      */
 210     op_pixels_func put_no_rnd_pixels_tab[4][4];
 211
 212     /**
 213      * Halfpel motion compensation with no rounding (a+b)>>1.
 214      * this is an array[2][4] of motion compensation funcions for 2
 215      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 216      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 217      * @param block destination into which the result is averaged (a+b)>>1
 218      * @param pixels source
 219      * @param line_size number of bytes in a horizontal line of block
 220      * @param h height
 221      */
 222     op_pixels_func avg_no_rnd_pixels_tab[4][4];
 223
 224     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);
 225
 226     /**
 227      * Thirdpel motion compensation with rounding (a+b+1)>>1.
 228      * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
 229      * *pixels_tab[ xthirdpel + 4*ythirdpel ]
 230      * @param block destination where the result is stored
 231      * @param pixels source
 232      * @param line_size number of bytes in a horizontal line of block
 233      * @param h height
 234      */
 235     tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 236     tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 237
 238     qpel_mc_func put_qpel_pixels_tab[2][16];
 239     qpel_mc_func avg_qpel_pixels_tab[2][16];
 240     qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
 241     qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
 242     qpel_mc_func put_mspel_pixels_tab[8];
 243
 244     /**
 245      * h264 Chram MC
 246      */
 247     h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
 248     h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
 249
 250     qpel_mc_func put_h264_qpel_pixels_tab[3][16];
 251     qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
 252
 253     me_cmp_func pix_abs[2][4];
 254
 255     /* huffyuv specific */
 256     void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
 257     void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
 258     /**
 259      * subtract huffyuv's variant of median prediction
 260      * note, this might read from src1[-1], src2[-1]
 261      */
 262     void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
 263     void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
 264
 265     void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
 266     void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
 267
 268     void (*h261_loop_filter)(uint8_t *src, int stride);
 269
 270     /* (I)DCT */
 271     void (*fdct)(DCTELEM *block/* align 16*/);
 272     void (*fdct248)(DCTELEM *block/* align 16*/);
 273
 274     /* IDCT really*/
 275     void (*idct)(DCTELEM *block/* align 16*/);
 276
 277     /**
 278      * block -> idct -> clip to unsigned 8 bit -> dest.
 279      * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
 280      * @param line_size size in bytes of a horizotal line of dest
 281      */
 282     void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 283
 284     /**
 285      * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
 286      * @param line_size size in bytes of a horizotal line of dest
 287      */
 288     void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 289
 290     /**
 291      * idct input permutation.
 292      * several optimized IDCTs need a permutated input (relative to the normal order of the reference
 293      * IDCT)
 294      * this permutation must be performed before the idct_put/add, note, normally this can be merged
 295      * with the zigzag/alternate scan<br>
 296      * an example to avoid confusion:
 297      * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
 298      * - (x -> referece dct -> reference idct -> x)
 299      * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
 300      * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
 301      */
 302     uint8_t idct_permutation[64];
 303     int idct_permutation_type;
 304 #define FF_NO_IDCT_PERM 1
 305 #define FF_LIBMPEG2_IDCT_PERM 2
 306 #define FF_SIMPLE_IDCT_PERM 3
 307 #define FF_TRANSPOSE_IDCT_PERM 4
 308
 309     int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
 310     void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
 311 #define BASIS_SHIFT 16
 312 #define RECON_SHIFT 6
 313
 314     /**
 315      * This function handles any initialization for the VP3 DSP functions.
 316      */
 317     void (*vp3_dsp_init)(void);
 318
 319     /**
 320      * This function is responsible for taking a block of zigzag'd,
 321      * quantized DCT coefficients and reconstructing the original block of
 322      * samples.
 323      * @param input_data 64 zigzag'd, quantized DCT coefficients
 324      * @param dequant_matrix 64 zigzag'd quantizer coefficients
 325      * @param coeff_count index of the last coefficient
 326      * @param output_samples space for 64 DCTELEMs where the transformed
 327      * samples will be stored
 328      */
 329     void (*vp3_idct)(int16_t *input_data, int16_t *dequant_matrix,
 330         int coeff_count, DCTELEM *output_samples);
 331
 332 } DSPContext;
 333
 334 void dsputil_static_init(void);
 335 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
 336
 337 /**
 338  * permute block according to permuatation.
 339  * @param last last non zero element in scantable order
 340  */
 341 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
 342
 343 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
 344
 345 #define BYTE_VEC32(c)   ((c)*0x01010101UL)
 346
 347 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
 348 {
 349     return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 350 }
 351
 352 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
 353 {
 354     return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 355 }
 356
 357 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 358     switch(type&0xFF){
 359     default:
 360     case FF_CMP_SAD:
 361         return lambda>>FF_LAMBDA_SHIFT;
 362     case FF_CMP_DCT:
 363         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 364     case FF_CMP_W53:
 365         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 366     case FF_CMP_W97:
 367         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 368     case FF_CMP_SATD:
 369         return (2*lambda)>>FF_LAMBDA_SHIFT;
 370     case FF_CMP_RD:
 371     case FF_CMP_PSNR:
 372     case FF_CMP_SSE:
 373     case FF_CMP_NSSE:
 374         return lambda2>>FF_LAMBDA_SHIFT;
 375     case FF_CMP_BIT:
 376         return 1;
 377     }
 378 }
 379
 380 /**
 381  * Empty mmx state.
 382  * this must be called between any dsp function and float/double code.
 383  * for example sin(); dsp->idct_put(); emms_c(); cos()
 384  */
 385 #define emms_c()
 386
 387 /* should be defined by architectures supporting
 388    one or more MultiMedia extension */
 389 int mm_support(void);
 390
 391 #define __align16 __attribute__ ((aligned (16)))
 392
 393 #if defined(HAVE_MMX)
 394
 395 #undef emms_c
 396
 397 #define MM_MMX    0x0001 /* standard MMX */
 398 #define MM_3DNOW  0x0004 /* AMD 3DNOW */
 399 #define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
 400 #define MM_SSE    0x0008 /* SSE functions */
 401 #define MM_SSE2   0x0010 /* PIV SSE2 functions */
 402
 403 extern int mm_flags;
 404
 405 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
 406 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
 407 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
 408
 409 static inline void emms(void)
 410 {
 411     __asm __volatile ("emms;":::"memory");
 412 }
 413
 414
 415 #define emms_c() \
 416 {\
 417     if (mm_flags & MM_MMX)\
 418         emms();\
 419 }
 420
 421 #define __align8 __attribute__ ((aligned (8)))
 422
 423 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
 424 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
 425
 426 #elif defined(ARCH_ARMV4L)
 427
 428 /* This is to use 4 bytes read to the IDCT pointers for some 'zero'
 429    line optimizations */
 430 #define __align8 __attribute__ ((aligned (4)))
 431
 432 void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
 433
 434 #elif defined(HAVE_MLIB)
 435
 436 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
 437 #define __align8 __attribute__ ((aligned (8)))
 438
 439 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
 440
 441 #elif defined(ARCH_SPARC)
 442
 443 /* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
 444 #define __align8 __attribute__ ((aligned (8)))
 445 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
 446
 447 #elif defined(ARCH_ALPHA)
 448
 449 #define __align8 __attribute__ ((aligned (8)))
 450
 451 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
 452
 453 #elif defined(ARCH_POWERPC)
 454
 455 #define MM_ALTIVEC    0x0001 /* standard AltiVec */
 456
 457 extern int mm_flags;
 458
 459 #if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
 460 #define pixel altivec_pixel
 461 #include <altivec.h>
 462 #undef pixel
 463 #endif
 464
 465 #define __align8 __attribute__ ((aligned (16)))
 466
 467 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
 468
 469 #elif defined(HAVE_MMI)
 470
 471 #define __align8 __attribute__ ((aligned (16)))
 472
 473 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
 474
 475 #elif defined(ARCH_SH4)
 476
 477 #define __align8 __attribute__ ((aligned (8)))
 478
 479 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
 480
 481 #else
 482
 483 #define __align8
 484
 485 #endif
 486
 487 #ifdef __GNUC__
 488
 489 struct unaligned_64 { uint64_t l; } __attribute__((packed));
 490 struct unaligned_32 { uint32_t l; } __attribute__((packed));
 491 struct unaligned_16 { uint16_t l; } __attribute__((packed));
 492
 493 #define LD16(a) (((const struct unaligned_16 *) (a))->l)
 494 #define LD32(a) (((const struct unaligned_32 *) (a))->l)
 495 #define LD64(a) (((const struct unaligned_64 *) (a))->l)
 496
 497 #define ST32(a, b) (((struct unaligned_32 *) (a))->l) = (b)
 498
 499 #else /* __GNUC__ */
 500
 501 #define LD16(a) (*((uint16_t*)(a)))
 502 #define LD32(a) (*((uint32_t*)(a)))
 503 #define LD64(a) (*((uint64_t*)(a)))
 504
 505 #define ST32(a, b) *((uint32_t*)(a)) = (b)
 506
 507 #endif /* !__GNUC__ */
 508
 509 /* PSNR */
 510 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
 511               int orig_linesize[3], int coded_linesize,
 512               AVCodecContext *avctx);
 513
 514 /* FFT computation */
 515
 516 /* NOTE: soon integer code will be added, so you must use the
 517    FFTSample type */
 518 typedef float FFTSample;
 519
 520 typedef struct FFTComplex {
 521     FFTSample re, im;
 522 } FFTComplex;
 523
 524 typedef struct FFTContext {
 525     int nbits;
 526     int inverse;
 527     uint16_t *revtab;
 528     FFTComplex *exptab;
 529     FFTComplex *exptab1; /* only used by SSE code */
 530     void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
 531 } FFTContext;
 532
 533 int ff_fft_init(FFTContext *s, int nbits, int inverse);
 534 void ff_fft_permute(FFTContext *s, FFTComplex *z);
 535 void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
 536 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z);
 537 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
 538
 539 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z)
 540 {
 541     s->fft_calc(s, z);
 542 }
 543 void ff_fft_end(FFTContext *s);
 544
 545 /* MDCT computation */
 546
 547 typedef struct MDCTContext {
 548     int n;  /* size of MDCT (i.e. number of input data * 2) */
 549     int nbits; /* n = 2^nbits */
 550     /* pre/post rotation tables */
 551     FFTSample *tcos;
 552     FFTSample *tsin;
 553     FFTContext fft;
 554 } MDCTContext;
 555
 556 int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
 557 void ff_imdct_calc(MDCTContext *s, FFTSample *output,
 558                 const FFTSample *input, FFTSample *tmp);
 559 void ff_mdct_calc(MDCTContext *s, FFTSample *out,
 560                const FFTSample *input, FFTSample *tmp);
 561 void ff_mdct_end(MDCTContext *s);
 562
 563 #define WARPER8_16(name8, name16)\
 564 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 565     return name8(s, dst           , src           , stride, h)\
 566           +name8(s, dst+8         , src+8         , stride, h);\
 567 }
 568
 569 #define WARPER8_16_SQ(name8, name16)\
 570 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 571     int score=0;\
 572     score +=name8(s, dst           , src           , stride, 8);\
 573     score +=name8(s, dst+8         , src+8         , stride, 8);\
 574     if(h==16){\
 575         dst += 8*stride;\
 576         src += 8*stride;\
 577         score +=name8(s, dst           , src           , stride, 8);\
 578         score +=name8(s, dst+8         , src+8         , stride, 8);\
 579     }\
 580     return score;\
 581 }
 582
 583 #ifndef HAVE_LRINTF
 584 /* XXX: add ISOC specific test to avoid specific BSD testing. */
 585 /* better than nothing implementation. */
 586 /* btw, rintf() is existing on fbsd too -- alex */
 587 static always_inline long int lrintf(float x)
 588 {
 589 #ifdef CONFIG_WIN32
 590 #  ifdef ARCH_X86
 591     int32_t i;
 592     asm volatile(
 593         "fistpl %0\n\t"
 594         : "=m" (i) : "t" (x) : "st"
 595     );
 596     return i;
 597 #  else
 598     /* XXX: incorrect, but make it compile */
 599     return (int)(x + (x < 0 ? -0.5 : 0.5));
 600 #  endif
 601 #else
 602     return (int)(rint(x));
 603 #endif
 604 }
 605 #else
 606 #ifndef _ISOC9X_SOURCE
 607 #define _ISOC9X_SOURCE
 608 #endif
 609 #include <math.h>
 610 #endif
 611
 612 #endif