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 file is part of FFmpeg.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with FFmpeg; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 /**
  24  * @file libavcodec/dsputil.h
  25  * DSP utils.
  26  * note, many functions in here may use MMX which trashes the FPU state, it is
  27  * absolutely necessary to call emms_c() between dsp & float/double code
  28  */
  29
  30 #ifndef AVCODEC_DSPUTIL_H
  31 #define AVCODEC_DSPUTIL_H
  32
  33 #include "libavutil/intreadwrite.h"
  34 #include "avcodec.h"
  35
  36
  37 //#define DEBUG
  38 /* dct code */
  39 typedef short DCTELEM;
  40 typedef int DWTELEM;
  41 typedef short IDWTELEM;
  42
  43 void fdct_ifast (DCTELEM *data);
  44 void fdct_ifast248 (DCTELEM *data);
  45 void ff_jpeg_fdct_islow (DCTELEM *data);
  46 void ff_fdct248_islow (DCTELEM *data);
  47
  48 void j_rev_dct (DCTELEM *data);
  49 void j_rev_dct4 (DCTELEM *data);
  50 void j_rev_dct2 (DCTELEM *data);
  51 void j_rev_dct1 (DCTELEM *data);
  52 void ff_wmv2_idct_c(DCTELEM *data);
  53
  54 void ff_fdct_mmx(DCTELEM *block);
  55 void ff_fdct_mmx2(DCTELEM *block);
  56 void ff_fdct_sse2(DCTELEM *block);
  57
  58 void ff_h264_idct8_add_c(uint8_t *dst, DCTELEM *block, int stride);
  59 void ff_h264_idct_add_c(uint8_t *dst, DCTELEM *block, int stride);
  60 void ff_h264_idct8_dc_add_c(uint8_t *dst, DCTELEM *block, int stride);
  61 void ff_h264_idct_dc_add_c(uint8_t *dst, DCTELEM *block, int stride);
  62 void ff_h264_lowres_idct_add_c(uint8_t *dst, int stride, DCTELEM *block);
  63 void ff_h264_lowres_idct_put_c(uint8_t *dst, int stride, DCTELEM *block);
  64 void ff_h264_idct_add16_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  65 void ff_h264_idct_add16intra_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  66 void ff_h264_idct8_add4_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  67 void ff_h264_idct_add8_c(uint8_t **dest, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  68
  69 void ff_vector_fmul_window_c(float *dst, const float *src0, const float *src1,
  70                              const float *win, float add_bias, int len);
  71 void ff_float_to_int16_c(int16_t *dst, const float *src, long len);
  72 void ff_float_to_int16_interleave_c(int16_t *dst, const float **src, long len, int channels);
  73
  74 /* encoding scans */
  75 extern const uint8_t ff_alternate_horizontal_scan[64];
  76 extern const uint8_t ff_alternate_vertical_scan[64];
  77 extern const uint8_t ff_zigzag_direct[64];
  78 extern const uint8_t ff_zigzag248_direct[64];
  79
  80 /* pixel operations */
  81 #define MAX_NEG_CROP 1024
  82
  83 /* temporary */
  84 extern uint32_t ff_squareTbl[512];
  85 extern uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP];
  86
  87 /* VP3 DSP functions */
  88 void ff_vp3_idct_c(DCTELEM *block/* align 16*/);
  89 void ff_vp3_idct_put_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
  90 void ff_vp3_idct_add_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
  91
  92 void ff_vp3_v_loop_filter_c(uint8_t *src, int stride, int *bounding_values);
  93 void ff_vp3_h_loop_filter_c(uint8_t *src, int stride, int *bounding_values);
  94
  95 /* VP6 DSP functions */
  96 void ff_vp6_filter_diag4_c(uint8_t *dst, uint8_t *src, int stride,
  97                            const int16_t *h_weights, const int16_t *v_weights);
  98
  99 /* 1/2^n downscaling functions from imgconvert.c */
 100 void ff_img_copy_plane(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 101 void ff_shrink22(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 102 void ff_shrink44(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 103 void ff_shrink88(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 104
 105 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
 106               int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 107
 108 /* minimum alignment rules ;)
 109 If you notice errors in the align stuff, need more alignment for some ASM code
 110 for some CPU or need to use a function with less aligned data then send a mail
 111 to the ffmpeg-devel mailing list, ...
 112
 113 !warning These alignments might not match reality, (missing attribute((align))
 114 stuff somewhere possible).
 115 I (Michael) did not check them, these are just the alignments which I think
 116 could be reached easily ...
 117
 118 !future video codecs might need functions with less strict alignment
 119 */
 120
 121 /*
 122 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
 123 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
 124 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
 125 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
 126 void clear_blocks_c(DCTELEM *blocks);
 127 */
 128
 129 /* add and put pixel (decoding) */
 130 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
 131 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
 132 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
 133 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);
 134 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 135 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
 136 typedef void (*h264_weight_func)(uint8_t *block, int stride, int log2_denom, int weight, int offset);
 137 typedef void (*h264_biweight_func)(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset);
 138
 139 #define DEF_OLD_QPEL(name)\
 140 void ff_put_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 141 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 142 void ff_avg_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 143
 144 DEF_OLD_QPEL(qpel16_mc11_old_c)
 145 DEF_OLD_QPEL(qpel16_mc31_old_c)
 146 DEF_OLD_QPEL(qpel16_mc12_old_c)
 147 DEF_OLD_QPEL(qpel16_mc32_old_c)
 148 DEF_OLD_QPEL(qpel16_mc13_old_c)
 149 DEF_OLD_QPEL(qpel16_mc33_old_c)
 150 DEF_OLD_QPEL(qpel8_mc11_old_c)
 151 DEF_OLD_QPEL(qpel8_mc31_old_c)
 152 DEF_OLD_QPEL(qpel8_mc12_old_c)
 153 DEF_OLD_QPEL(qpel8_mc32_old_c)
 154 DEF_OLD_QPEL(qpel8_mc13_old_c)
 155 DEF_OLD_QPEL(qpel8_mc33_old_c)
 156
 157 #define CALL_2X_PIXELS(a, b, n)\
 158 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
 159     b(block  , pixels  , line_size, h);\
 160     b(block+n, pixels+n, line_size, h);\
 161 }
 162
 163 /* motion estimation */
 164 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
 165 // although currently h<4 is not used as functions with width <8 are neither used nor implemented
 166 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))*/;
 167
 168
 169 // for snow slices
 170 typedef struct slice_buffer_s slice_buffer;
 171
 172 /**
 173  * Scantable.
 174  */
 175 typedef struct ScanTable{
 176     const uint8_t *scantable;
 177     uint8_t permutated[64];
 178     uint8_t raster_end[64];
 179 #if ARCH_PPC
 180                 /** Used by dct_quantize_altivec to find last-non-zero */
 181     DECLARE_ALIGNED(16, uint8_t, inverse[64]);
 182 #endif
 183 } ScanTable;
 184
 185 void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
 186
 187 void ff_emulated_edge_mc(uint8_t *buf, uint8_t *src, int linesize,
 188                          int block_w, int block_h,
 189                          int src_x, int src_y, int w, int h);
 190
 191 /**
 192  * DSPContext.
 193  */
 194 typedef struct DSPContext {
 195     /* pixel ops : interface with DCT */
 196     void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
 197     void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
 198     void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 199     void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 200     void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 201     void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size);
 202     void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size);
 203     int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/);
 204     /**
 205      * translational global motion compensation.
 206      */
 207     void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
 208     /**
 209      * global motion compensation.
 210      */
 211     void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
 212                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 213     void (*clear_block)(DCTELEM *block/*align 16*/);
 214     void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
 215     int (*pix_sum)(uint8_t * pix, int line_size);
 216     int (*pix_norm1)(uint8_t * pix, int line_size);
 217 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
 218
 219     me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
 220     me_cmp_func sse[6];
 221     me_cmp_func hadamard8_diff[6];
 222     me_cmp_func dct_sad[6];
 223     me_cmp_func quant_psnr[6];
 224     me_cmp_func bit[6];
 225     me_cmp_func rd[6];
 226     me_cmp_func vsad[6];
 227     me_cmp_func vsse[6];
 228     me_cmp_func nsse[6];
 229     me_cmp_func w53[6];
 230     me_cmp_func w97[6];
 231     me_cmp_func dct_max[6];
 232     me_cmp_func dct264_sad[6];
 233
 234     me_cmp_func me_pre_cmp[6];
 235     me_cmp_func me_cmp[6];
 236     me_cmp_func me_sub_cmp[6];
 237     me_cmp_func mb_cmp[6];
 238     me_cmp_func ildct_cmp[6]; //only width 16 used
 239     me_cmp_func frame_skip_cmp[6]; //only width 8 used
 240
 241     int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
 242                              int size);
 243
 244     /**
 245      * Halfpel motion compensation with rounding (a+b+1)>>1.
 246      * this is an array[4][4] of motion compensation functions for 4
 247      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 248      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 249      * @param block destination where the result is stored
 250      * @param pixels source
 251      * @param line_size number of bytes in a horizontal line of block
 252      * @param h height
 253      */
 254     op_pixels_func put_pixels_tab[4][4];
 255
 256     /**
 257      * Halfpel motion compensation with rounding (a+b+1)>>1.
 258      * This is an array[4][4] of motion compensation functions for 4
 259      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 260      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 261      * @param block destination into which the result is averaged (a+b+1)>>1
 262      * @param pixels source
 263      * @param line_size number of bytes in a horizontal line of block
 264      * @param h height
 265      */
 266     op_pixels_func avg_pixels_tab[4][4];
 267
 268     /**
 269      * Halfpel motion compensation with no rounding (a+b)>>1.
 270      * this is an array[2][4] of motion compensation functions for 2
 271      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 272      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 273      * @param block destination where the result is stored
 274      * @param pixels source
 275      * @param line_size number of bytes in a horizontal line of block
 276      * @param h height
 277      */
 278     op_pixels_func put_no_rnd_pixels_tab[4][4];
 279
 280     /**
 281      * Halfpel motion compensation with no rounding (a+b)>>1.
 282      * this is an array[2][4] of motion compensation functions for 2
 283      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 284      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 285      * @param block destination into which the result is averaged (a+b)>>1
 286      * @param pixels source
 287      * @param line_size number of bytes in a horizontal line of block
 288      * @param h height
 289      */
 290     op_pixels_func avg_no_rnd_pixels_tab[4][4];
 291
 292     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);
 293
 294     /**
 295      * Thirdpel motion compensation with rounding (a+b+1)>>1.
 296      * this is an array[12] of motion compensation functions for the 9 thirdpe
 297      * positions<br>
 298      * *pixels_tab[ xthirdpel + 4*ythirdpel ]
 299      * @param block destination where the result is stored
 300      * @param pixels source
 301      * @param line_size number of bytes in a horizontal line of block
 302      * @param h height
 303      */
 304     tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 305     tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 306
 307     qpel_mc_func put_qpel_pixels_tab[2][16];
 308     qpel_mc_func avg_qpel_pixels_tab[2][16];
 309     qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
 310     qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
 311     qpel_mc_func put_mspel_pixels_tab[8];
 312
 313     /**
 314      * h264 Chroma MC
 315      */
 316     h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
 317     h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
 318     /* This is really one func used in VC-1 decoding */
 319     h264_chroma_mc_func put_no_rnd_vc1_chroma_pixels_tab[3];
 320     h264_chroma_mc_func avg_no_rnd_vc1_chroma_pixels_tab[3];
 321
 322     qpel_mc_func put_h264_qpel_pixels_tab[4][16];
 323     qpel_mc_func avg_h264_qpel_pixels_tab[4][16];
 324
 325     qpel_mc_func put_2tap_qpel_pixels_tab[4][16];
 326     qpel_mc_func avg_2tap_qpel_pixels_tab[4][16];
 327
 328     h264_weight_func weight_h264_pixels_tab[10];
 329     h264_biweight_func biweight_h264_pixels_tab[10];
 330
 331     /* AVS specific */
 332     qpel_mc_func put_cavs_qpel_pixels_tab[2][16];
 333     qpel_mc_func avg_cavs_qpel_pixels_tab[2][16];
 334     void (*cavs_filter_lv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 335     void (*cavs_filter_lh)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 336     void (*cavs_filter_cv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 337     void (*cavs_filter_ch)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 338     void (*cavs_idct8_add)(uint8_t *dst, DCTELEM *block, int stride);
 339
 340     me_cmp_func pix_abs[2][4];
 341
 342     /* huffyuv specific */
 343     void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
 344     void (*add_bytes_l2)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 16*/, int w);
 345     void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
 346     /**
 347      * subtract huffyuv's variant of median prediction
 348      * note, this might read from src1[-1], src2[-1]
 349      */
 350     void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
 351     void (*add_hfyu_median_prediction)(uint8_t *dst, uint8_t *top, uint8_t *diff, int w, int *left, int *left_top);
 352     /* this might write to dst[w] */
 353     void (*add_png_paeth_prediction)(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
 354     void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
 355
 356     void (*h264_v_loop_filter_luma)(uint8_t *pix/*align 16*/, int stride, int alpha, int beta, int8_t *tc0);
 357     void (*h264_h_loop_filter_luma)(uint8_t *pix/*align 4 */, int stride, int alpha, int beta, int8_t *tc0);
 358     /* v/h_loop_filter_luma_intra: align 16 */
 359     void (*h264_v_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
 360     void (*h264_h_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
 361     void (*h264_v_loop_filter_chroma)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta, int8_t *tc0);
 362     void (*h264_h_loop_filter_chroma)(uint8_t *pix/*align 4*/, int stride, int alpha, int beta, int8_t *tc0);
 363     void (*h264_v_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
 364     void (*h264_h_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
 365     // h264_loop_filter_strength: simd only. the C version is inlined in h264.c
 366     void (*h264_loop_filter_strength)(int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
 367                                       int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field);
 368
 369     void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
 370     void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
 371
 372     void (*h261_loop_filter)(uint8_t *src, int stride);
 373
 374     void (*x8_v_loop_filter)(uint8_t *src, int stride, int qscale);
 375     void (*x8_h_loop_filter)(uint8_t *src, int stride, int qscale);
 376
 377     void (*vp3_v_loop_filter)(uint8_t *src, int stride, int *bounding_values);
 378     void (*vp3_h_loop_filter)(uint8_t *src, int stride, int *bounding_values);
 379
 380     void (*vp6_filter_diag4)(uint8_t *dst, uint8_t *src, int stride,
 381                              const int16_t *h_weights,const int16_t *v_weights);
 382
 383     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
 384     void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize);
 385     void (*ac3_downmix)(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len);
 386     /* no alignment needed */
 387     void (*flac_compute_autocorr)(const int32_t *data, int len, int lag, double *autoc);
 388     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
 389     void (*vector_fmul)(float *dst, const float *src, int len);
 390     void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len);
 391     /* assume len is a multiple of 8, and src arrays are 16-byte aligned */
 392     void (*vector_fmul_add)(float *dst, const float *src0, const float *src1, const float *src2, int len);
 393     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
 394     void (*vector_fmul_window)(float *dst, const float *src0, const float *src1, const float *win, float add_bias, int len);
 395     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
 396     void (*int32_to_float_fmul_scalar)(float *dst, const int *src, float mul, int len);
 397     void (*vector_clipf)(float *dst /* align 16 */, const float *src /* align 16 */, float min, float max, int len /* align 16 */);
 398     /**
 399      * Multiply a vector of floats by a scalar float.  Source and
 400      * destination vectors must overlap exactly or not at all.
 401      * @param dst result vector, 16-byte aligned
 402      * @param src input vector, 16-byte aligned
 403      * @param mul scalar value
 404      * @param len length of vector, multiple of 4
 405      */
 406     void (*vector_fmul_scalar)(float *dst, const float *src, float mul,
 407                                int len);
 408     /**
 409      * Multiply a vector of floats by concatenated short vectors of
 410      * floats and by a scalar float.  Source and destination vectors
 411      * must overlap exactly or not at all.
 412      * [0]: short vectors of length 2, 8-byte aligned
 413      * [1]: short vectors of length 4, 16-byte aligned
 414      * @param dst output vector, 16-byte aligned
 415      * @param src input vector, 16-byte aligned
 416      * @param sv  array of pointers to short vectors
 417      * @param mul scalar value
 418      * @param len number of elements in src and dst, multiple of 4
 419      */
 420     void (*vector_fmul_sv_scalar[2])(float *dst, const float *src,
 421                                      const float **sv, float mul, int len);
 422     /**
 423      * Multiply short vectors of floats by a scalar float, store
 424      * concatenated result.
 425      * [0]: short vectors of length 2, 8-byte aligned
 426      * [1]: short vectors of length 4, 16-byte aligned
 427      * @param dst output vector, 16-byte aligned
 428      * @param sv  array of pointers to short vectors
 429      * @param mul scalar value
 430      * @param len number of output elements, multiple of 4
 431      */
 432     void (*sv_fmul_scalar[2])(float *dst, const float **sv,
 433                               float mul, int len);
 434     /**
 435      * Calculate the scalar product of two vectors of floats.
 436      * @param v1  first vector, 16-byte aligned
 437      * @param v2  second vector, 16-byte aligned
 438      * @param len length of vectors, multiple of 4
 439      */
 440     float (*scalarproduct_float)(const float *v1, const float *v2, int len);
 441     /**
 442      * Calculate the sum and difference of two vectors of floats.
 443      * @param v1  first input vector, sum output, 16-byte aligned
 444      * @param v2  second input vector, difference output, 16-byte aligned
 445      * @param len length of vectors, multiple of 4
 446      */
 447     void (*butterflies_float)(float *restrict v1, float *restrict v2, int len);
 448
 449     /* C version: convert floats from the range [384.0,386.0] to ints in [-32768,32767]
 450      * simd versions: convert floats from [-32768.0,32767.0] without rescaling and arrays are 16byte aligned */
 451     void (*float_to_int16)(int16_t *dst, const float *src, long len);
 452     void (*float_to_int16_interleave)(int16_t *dst, const float **src, long len, int channels);
 453
 454     /* (I)DCT */
 455     void (*fdct)(DCTELEM *block/* align 16*/);
 456     void (*fdct248)(DCTELEM *block/* align 16*/);
 457
 458     /* IDCT really*/
 459     void (*idct)(DCTELEM *block/* align 16*/);
 460
 461     /**
 462      * block -> idct -> clip to unsigned 8 bit -> dest.
 463      * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
 464      * @param line_size size in bytes of a horizontal line of dest
 465      */
 466     void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 467
 468     /**
 469      * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
 470      * @param line_size size in bytes of a horizontal line of dest
 471      */
 472     void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 473
 474     /**
 475      * idct input permutation.
 476      * several optimized IDCTs need a permutated input (relative to the normal order of the reference
 477      * IDCT)
 478      * this permutation must be performed before the idct_put/add, note, normally this can be merged
 479      * with the zigzag/alternate scan<br>
 480      * an example to avoid confusion:
 481      * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
 482      * - (x -> referece dct -> reference idct -> x)
 483      * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
 484      * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
 485      */
 486     uint8_t idct_permutation[64];
 487     int idct_permutation_type;
 488 #define FF_NO_IDCT_PERM 1
 489 #define FF_LIBMPEG2_IDCT_PERM 2
 490 #define FF_SIMPLE_IDCT_PERM 3
 491 #define FF_TRANSPOSE_IDCT_PERM 4
 492 #define FF_PARTTRANS_IDCT_PERM 5
 493 #define FF_SSE2_IDCT_PERM 6
 494
 495     int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
 496     void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
 497 #define BASIS_SHIFT 16
 498 #define RECON_SHIFT 6
 499
 500     void (*draw_edges)(uint8_t *buf, int wrap, int width, int height, int w);
 501 #define EDGE_WIDTH 16
 502
 503     /* h264 functions */
 504     /* NOTE!!! if you implement any of h264_idct8_add, h264_idct8_add4 then you must implement all of them
 505        NOTE!!! if you implement any of h264_idct_add, h264_idct_add16, h264_idct_add16intra, h264_idct_add8 then you must implement all of them
 506         The reason for above, is that no 2 out of one list may use a different permutation.
 507     */
 508     void (*h264_idct_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
 509     void (*h264_idct8_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
 510     void (*h264_idct_dc_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
 511     void (*h264_idct8_dc_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
 512     void (*h264_dct)(DCTELEM block[4][4]);
 513     void (*h264_idct_add16)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 514     void (*h264_idct8_add4)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 515     void (*h264_idct_add8)(uint8_t **dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 516     void (*h264_idct_add16intra)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 517
 518     /* snow wavelet */
 519     void (*vertical_compose97i)(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width);
 520     void (*horizontal_compose97i)(IDWTELEM *b, int width);
 521     void (*inner_add_yblock)(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h, int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
 522
 523     void (*prefetch)(void *mem, int stride, int h);
 524
 525     void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 526
 527     /* mlp/truehd functions */
 528     void (*mlp_filter_channel)(int32_t *state, const int32_t *coeff,
 529                                int firorder, int iirorder,
 530                                unsigned int filter_shift, int32_t mask, int blocksize,
 531                                int32_t *sample_buffer);
 532
 533     /* vc1 functions */
 534     void (*vc1_inv_trans_8x8)(DCTELEM *b);
 535     void (*vc1_inv_trans_8x4)(uint8_t *dest, int line_size, DCTELEM *block);
 536     void (*vc1_inv_trans_4x8)(uint8_t *dest, int line_size, DCTELEM *block);
 537     void (*vc1_inv_trans_4x4)(uint8_t *dest, int line_size, DCTELEM *block);
 538     void (*vc1_inv_trans_8x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 539     void (*vc1_inv_trans_8x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 540     void (*vc1_inv_trans_4x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 541     void (*vc1_inv_trans_4x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 542     void (*vc1_v_overlap)(uint8_t* src, int stride);
 543     void (*vc1_h_overlap)(uint8_t* src, int stride);
 544     void (*vc1_v_loop_filter4)(uint8_t *src, int stride, int pq);
 545     void (*vc1_h_loop_filter4)(uint8_t *src, int stride, int pq);
 546     void (*vc1_v_loop_filter8)(uint8_t *src, int stride, int pq);
 547     void (*vc1_h_loop_filter8)(uint8_t *src, int stride, int pq);
 548     void (*vc1_v_loop_filter16)(uint8_t *src, int stride, int pq);
 549     void (*vc1_h_loop_filter16)(uint8_t *src, int stride, int pq);
 550     /* put 8x8 block with bicubic interpolation and quarterpel precision
 551      * last argument is actually round value instead of height
 552      */
 553     op_pixels_func put_vc1_mspel_pixels_tab[16];
 554     op_pixels_func avg_vc1_mspel_pixels_tab[16];
 555
 556     /* intrax8 functions */
 557     void (*x8_spatial_compensation[12])(uint8_t *src , uint8_t *dst, int linesize);
 558     void (*x8_setup_spatial_compensation)(uint8_t *src, uint8_t *dst, int linesize,
 559            int * range, int * sum,  int edges);
 560
 561     /* ape functions */
 562     /**
 563      * Add contents of the second vector to the first one.
 564      * @param len length of vectors, should be multiple of 16
 565      */
 566     void (*add_int16)(int16_t *v1/*align 16*/, int16_t *v2, int len);
 567     /**
 568      * Add contents of the second vector to the first one.
 569      * @param len length of vectors, should be multiple of 16
 570      */
 571     void (*sub_int16)(int16_t *v1/*align 16*/, int16_t *v2, int len);
 572     /**
 573      * Calculate scalar product of two vectors.
 574      * @param len length of vectors, should be multiple of 16
 575      * @param shift number of bits to discard from product
 576      */
 577     int32_t (*scalarproduct_int16)(int16_t *v1, int16_t *v2/*align 16*/, int len, int shift);
 578
 579     /* rv30 functions */
 580     qpel_mc_func put_rv30_tpel_pixels_tab[4][16];
 581     qpel_mc_func avg_rv30_tpel_pixels_tab[4][16];
 582
 583     /* rv40 functions */
 584     qpel_mc_func put_rv40_qpel_pixels_tab[4][16];
 585     qpel_mc_func avg_rv40_qpel_pixels_tab[4][16];
 586     h264_chroma_mc_func put_rv40_chroma_pixels_tab[3];
 587     h264_chroma_mc_func avg_rv40_chroma_pixels_tab[3];
 588 } DSPContext;
 589
 590 void dsputil_static_init(void);
 591 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
 592
 593 int ff_check_alignment(void);
 594
 595 /**
 596  * permute block according to permuatation.
 597  * @param last last non zero element in scantable order
 598  */
 599 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
 600
 601 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
 602
 603 #define         BYTE_VEC32(c)   ((c)*0x01010101UL)
 604
 605 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
 606 {
 607     return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 608 }
 609
 610 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
 611 {
 612     return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 613 }
 614
 615 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 616     switch(type&0xFF){
 617     default:
 618     case FF_CMP_SAD:
 619         return lambda>>FF_LAMBDA_SHIFT;
 620     case FF_CMP_DCT:
 621         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 622     case FF_CMP_W53:
 623         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 624     case FF_CMP_W97:
 625         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 626     case FF_CMP_SATD:
 627     case FF_CMP_DCT264:
 628         return (2*lambda)>>FF_LAMBDA_SHIFT;
 629     case FF_CMP_RD:
 630     case FF_CMP_PSNR:
 631     case FF_CMP_SSE:
 632     case FF_CMP_NSSE:
 633         return lambda2>>FF_LAMBDA_SHIFT;
 634     case FF_CMP_BIT:
 635         return 1;
 636     }
 637 }
 638
 639 /**
 640  * Empty mmx state.
 641  * this must be called between any dsp function and float/double code.
 642  * for example sin(); dsp->idct_put(); emms_c(); cos()
 643  */
 644 #define emms_c()
 645
 646 /* should be defined by architectures supporting
 647    one or more MultiMedia extension */
 648 int mm_support(void);
 649 extern int mm_flags;
 650
 651 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
 652 void dsputil_init_arm(DSPContext* c, AVCodecContext *avctx);
 653 void dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx);
 654 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
 655 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
 656 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
 657 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
 658 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
 659 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
 660
 661 #define DECLARE_ALIGNED_16(t, v) DECLARE_ALIGNED(16, t, v)
 662 #define DECLARE_ALIGNED_8(t, v)  DECLARE_ALIGNED(8, t, v)
 663
 664 #if HAVE_MMX
 665
 666 #undef emms_c
 667
 668 static inline void emms(void)
 669 {
 670     __asm__ volatile ("emms;":::"memory");
 671 }
 672
 673
 674 #define emms_c() \
 675 {\
 676     if (mm_flags & FF_MM_MMX)\
 677         emms();\
 678 }
 679
 680 #elif ARCH_ARM
 681
 682 #if HAVE_NEON
 683 #   define STRIDE_ALIGN 16
 684 #endif
 685
 686 #elif ARCH_PPC
 687
 688 #define STRIDE_ALIGN 16
 689
 690 #elif HAVE_MMI
 691
 692 #define STRIDE_ALIGN 16
 693
 694 #else
 695
 696 #define mm_flags 0
 697 #define mm_support() 0
 698
 699 #endif
 700
 701 #ifndef STRIDE_ALIGN
 702 #   define STRIDE_ALIGN 8
 703 #endif
 704
 705 /* PSNR */
 706 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
 707               int orig_linesize[3], int coded_linesize,
 708               AVCodecContext *avctx);
 709
 710 /* FFT computation */
 711
 712 /* NOTE: soon integer code will be added, so you must use the
 713    FFTSample type */
 714 typedef float FFTSample;
 715
 716 typedef struct FFTComplex {
 717     FFTSample re, im;
 718 } FFTComplex;
 719
 720 typedef struct FFTContext {
 721     int nbits;
 722     int inverse;
 723     uint16_t *revtab;
 724     FFTComplex *exptab;
 725     FFTComplex *exptab1; /* only used by SSE code */
 726     FFTComplex *tmp_buf;
 727     int mdct_size; /* size of MDCT (i.e. number of input data * 2) */
 728     int mdct_bits; /* n = 2^nbits */
 729     /* pre/post rotation tables */
 730     FFTSample *tcos;
 731     FFTSample *tsin;
 732     void (*fft_permute)(struct FFTContext *s, FFTComplex *z);
 733     void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
 734     void (*imdct_calc)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
 735     void (*imdct_half)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
 736     void (*mdct_calc)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
 737     int split_radix;
 738     int permutation;
 739 #define FF_MDCT_PERM_NONE       0
 740 #define FF_MDCT_PERM_INTERLEAVE 1
 741 } FFTContext;
 742
 743 extern FFTSample* const ff_cos_tabs[13];
 744
 745 /**
 746  * Sets up a complex FFT.
 747  * @param nbits           log2 of the length of the input array
 748  * @param inverse         if 0 perform the forward transform, if 1 perform the inverse
 749  */
 750 int ff_fft_init(FFTContext *s, int nbits, int inverse);
 751 void ff_fft_permute_c(FFTContext *s, FFTComplex *z);
 752 void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
 753
 754 void ff_fft_init_altivec(FFTContext *s);
 755 void ff_fft_init_mmx(FFTContext *s);
 756 void ff_fft_init_arm(FFTContext *s);
 757
 758 /**
 759  * Do the permutation needed BEFORE calling ff_fft_calc().
 760  */
 761 static inline void ff_fft_permute(FFTContext *s, FFTComplex *z)
 762 {
 763     s->fft_permute(s, z);
 764 }
 765 /**
 766  * Do a complex FFT with the parameters defined in ff_fft_init(). The
 767  * input data must be permuted before. No 1.0/sqrt(n) normalization is done.
 768  */
 769 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z)
 770 {
 771     s->fft_calc(s, z);
 772 }
 773 void ff_fft_end(FFTContext *s);
 774
 775 /* MDCT computation */
 776
 777 static inline void ff_imdct_calc(FFTContext *s, FFTSample *output, const FFTSample *input)
 778 {
 779     s->imdct_calc(s, output, input);
 780 }
 781 static inline void ff_imdct_half(FFTContext *s, FFTSample *output, const FFTSample *input)
 782 {
 783     s->imdct_half(s, output, input);
 784 }
 785
 786 static inline void ff_mdct_calc(FFTContext *s, FFTSample *output,
 787                                 const FFTSample *input)
 788 {
 789     s->mdct_calc(s, output, input);
 790 }
 791
 792 /**
 793  * Generate a Kaiser-Bessel Derived Window.
 794  * @param   window  pointer to half window
 795  * @param   alpha   determines window shape
 796  * @param   n       size of half window
 797  */
 798 void ff_kbd_window_init(float *window, float alpha, int n);
 799
 800 /**
 801  * Generate a sine window.
 802  * @param   window  pointer to half window
 803  * @param   n       size of half window
 804  */
 805 void ff_sine_window_init(float *window, int n);
 806 extern float ff_sine_32  [  32];
 807 extern float ff_sine_64  [  64];
 808 extern float ff_sine_128 [ 128];
 809 extern float ff_sine_256 [ 256];
 810 extern float ff_sine_512 [ 512];
 811 extern float ff_sine_1024[1024];
 812 extern float ff_sine_2048[2048];
 813 extern float ff_sine_4096[4096];
 814 extern float * const ff_sine_windows[13];
 815
 816 int ff_mdct_init(FFTContext *s, int nbits, int inverse, double scale);
 817 void ff_imdct_calc_c(FFTContext *s, FFTSample *output, const FFTSample *input);
 818 void ff_imdct_half_c(FFTContext *s, FFTSample *output, const FFTSample *input);
 819 void ff_mdct_calc_c(FFTContext *s, FFTSample *output, const FFTSample *input);
 820 void ff_mdct_end(FFTContext *s);
 821
 822 /* Real Discrete Fourier Transform */
 823
 824 enum RDFTransformType {
 825     RDFT,
 826     IRDFT,
 827     RIDFT,
 828     IRIDFT,
 829 };
 830
 831 typedef struct {
 832     int nbits;
 833     int inverse;
 834     int sign_convention;
 835
 836     /* pre/post rotation tables */
 837     FFTSample *tcos;
 838     FFTSample *tsin;
 839     FFTContext fft;
 840 } RDFTContext;
 841
 842 /**
 843  * Sets up a real FFT.
 844  * @param nbits           log2 of the length of the input array
 845  * @param trans           the type of transform
 846  */
 847 int ff_rdft_init(RDFTContext *s, int nbits, enum RDFTransformType trans);
 848 void ff_rdft_calc(RDFTContext *s, FFTSample *data);
 849 void ff_rdft_end(RDFTContext *s);
 850
 851 #define WRAPPER8_16(name8, name16)\
 852 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 853     return name8(s, dst           , src           , stride, h)\
 854           +name8(s, dst+8         , src+8         , stride, h);\
 855 }
 856
 857 #define WRAPPER8_16_SQ(name8, name16)\
 858 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 859     int score=0;\
 860     score +=name8(s, dst           , src           , stride, 8);\
 861     score +=name8(s, dst+8         , src+8         , stride, 8);\
 862     if(h==16){\
 863         dst += 8*stride;\
 864         src += 8*stride;\
 865         score +=name8(s, dst           , src           , stride, 8);\
 866         score +=name8(s, dst+8         , src+8         , stride, 8);\
 867     }\
 868     return score;\
 869 }
 870
 871
 872 static inline void copy_block2(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 873 {
 874     int i;
 875     for(i=0; i<h; i++)
 876     {
 877         AV_WN16(dst   , AV_RN16(src   ));
 878         dst+=dstStride;
 879         src+=srcStride;
 880     }
 881 }
 882
 883 static inline void copy_block4(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 884 {
 885     int i;
 886     for(i=0; i<h; i++)
 887     {
 888         AV_WN32(dst   , AV_RN32(src   ));
 889         dst+=dstStride;
 890         src+=srcStride;
 891     }
 892 }
 893
 894 static inline void copy_block8(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 895 {
 896     int i;
 897     for(i=0; i<h; i++)
 898     {
 899         AV_WN32(dst   , AV_RN32(src   ));
 900         AV_WN32(dst+4 , AV_RN32(src+4 ));
 901         dst+=dstStride;
 902         src+=srcStride;
 903     }
 904 }
 905
 906 static inline void copy_block9(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 907 {
 908     int i;
 909     for(i=0; i<h; i++)
 910     {
 911         AV_WN32(dst   , AV_RN32(src   ));
 912         AV_WN32(dst+4 , AV_RN32(src+4 ));
 913         dst[8]= src[8];
 914         dst+=dstStride;
 915         src+=srcStride;
 916     }
 917 }
 918
 919 static inline void copy_block16(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 920 {
 921     int i;
 922     for(i=0; i<h; i++)
 923     {
 924         AV_WN32(dst   , AV_RN32(src   ));
 925         AV_WN32(dst+4 , AV_RN32(src+4 ));
 926         AV_WN32(dst+8 , AV_RN32(src+8 ));
 927         AV_WN32(dst+12, AV_RN32(src+12));
 928         dst+=dstStride;
 929         src+=srcStride;
 930     }
 931 }
 932
 933 static inline void copy_block17(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 934 {
 935     int i;
 936     for(i=0; i<h; i++)
 937     {
 938         AV_WN32(dst   , AV_RN32(src   ));
 939         AV_WN32(dst+4 , AV_RN32(src+4 ));
 940         AV_WN32(dst+8 , AV_RN32(src+8 ));
 941         AV_WN32(dst+12, AV_RN32(src+12));
 942         dst[16]= src[16];
 943         dst+=dstStride;
 944         src+=srcStride;
 945     }
 946 }
 947
 948 #endif /* AVCODEC_DSPUTIL_H */