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
  41 void fdct_ifast (DCTELEM *data);
  42 void fdct_ifast248 (DCTELEM *data);
  43 void ff_jpeg_fdct_islow (DCTELEM *data);
  44 void ff_fdct248_islow (DCTELEM *data);
  45
  46 void j_rev_dct (DCTELEM *data);
  47 void j_rev_dct4 (DCTELEM *data);
  48 void j_rev_dct2 (DCTELEM *data);
  49 void j_rev_dct1 (DCTELEM *data);
  50 void ff_wmv2_idct_c(DCTELEM *data);
  51
  52 void ff_fdct_mmx(DCTELEM *block);
  53 void ff_fdct_mmx2(DCTELEM *block);
  54 void ff_fdct_sse2(DCTELEM *block);
  55
  56 void ff_h264_idct8_add_c(uint8_t *dst, DCTELEM *block, int stride);
  57 void ff_h264_idct_add_c(uint8_t *dst, DCTELEM *block, int stride);
  58 void ff_h264_idct8_dc_add_c(uint8_t *dst, DCTELEM *block, int stride);
  59 void ff_h264_idct_dc_add_c(uint8_t *dst, DCTELEM *block, int stride);
  60 void ff_h264_lowres_idct_add_c(uint8_t *dst, int stride, DCTELEM *block);
  61 void ff_h264_lowres_idct_put_c(uint8_t *dst, int stride, DCTELEM *block);
  62 void ff_h264_idct_add16_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  63 void ff_h264_idct_add16intra_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  64 void ff_h264_idct8_add4_c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  65 void ff_h264_idct_add8_c(uint8_t **dest, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  66
  67 void ff_vector_fmul_window_c(float *dst, const float *src0, const float *src1,
  68                              const float *win, float add_bias, int len);
  69 void ff_float_to_int16_c(int16_t *dst, const float *src, long len);
  70 void ff_float_to_int16_interleave_c(int16_t *dst, const float **src, long len, int channels);
  71
  72 /* encoding scans */
  73 extern const uint8_t ff_alternate_horizontal_scan[64];
  74 extern const uint8_t ff_alternate_vertical_scan[64];
  75 extern const uint8_t ff_zigzag_direct[64];
  76 extern const uint8_t ff_zigzag248_direct[64];
  77
  78 /* pixel operations */
  79 #define MAX_NEG_CROP 1024
  80
  81 /* temporary */
  82 extern uint32_t ff_squareTbl[512];
  83 extern uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP];
  84
  85 /* VP3 DSP functions */
  86 void ff_vp3_idct_c(DCTELEM *block/* align 16*/);
  87 void ff_vp3_idct_put_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
  88 void ff_vp3_idct_add_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
  89
  90 void ff_vp3_v_loop_filter_c(uint8_t *src, int stride, int *bounding_values);
  91 void ff_vp3_h_loop_filter_c(uint8_t *src, int stride, int *bounding_values);
  92
  93 /* VP6 DSP functions */
  94 void ff_vp6_filter_diag4_c(uint8_t *dst, uint8_t *src, int stride,
  95                            const int16_t *h_weights, const int16_t *v_weights);
  96
  97 /* Bink functions */
  98 void ff_bink_idct_c    (DCTELEM *block);
  99 void ff_bink_idct_add_c(uint8_t *dest, int linesize, DCTELEM *block);
 100 void ff_bink_idct_put_c(uint8_t *dest, int linesize, DCTELEM *block);
 101
 102 /* CAVS functions */
 103 void ff_put_cavs_qpel8_mc00_c(uint8_t *dst, uint8_t *src, int stride);
 104 void ff_avg_cavs_qpel8_mc00_c(uint8_t *dst, uint8_t *src, int stride);
 105 void ff_put_cavs_qpel16_mc00_c(uint8_t *dst, uint8_t *src, int stride);
 106 void ff_avg_cavs_qpel16_mc00_c(uint8_t *dst, uint8_t *src, int stride);
 107
 108 /* VC1 functions */
 109 void ff_put_vc1_mspel_mc00_c(uint8_t *dst, const uint8_t *src, int stride, int rnd);
 110 void ff_avg_vc1_mspel_mc00_c(uint8_t *dst, const uint8_t *src, int stride, int rnd);
 111
 112 /* EA functions */
 113 void ff_ea_idct_put_c(uint8_t *dest, int linesize, DCTELEM *block);
 114
 115 /* 1/2^n downscaling functions from imgconvert.c */
 116 void ff_img_copy_plane(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 117 void ff_shrink22(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 118 void ff_shrink44(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 119 void ff_shrink88(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 120
 121 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
 122               int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 123
 124 /* minimum alignment rules ;)
 125 If you notice errors in the align stuff, need more alignment for some ASM code
 126 for some CPU or need to use a function with less aligned data then send a mail
 127 to the ffmpeg-devel mailing list, ...
 128
 129 !warning These alignments might not match reality, (missing attribute((align))
 130 stuff somewhere possible).
 131 I (Michael) did not check them, these are just the alignments which I think
 132 could be reached easily ...
 133
 134 !future video codecs might need functions with less strict alignment
 135 */
 136
 137 /*
 138 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
 139 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
 140 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
 141 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
 142 void clear_blocks_c(DCTELEM *blocks);
 143 */
 144
 145 /* add and put pixel (decoding) */
 146 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
 147 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
 148 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
 149 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);
 150 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 151 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
 152 typedef void (*h264_weight_func)(uint8_t *block, int stride, int log2_denom, int weight, int offset);
 153 typedef void (*h264_biweight_func)(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset);
 154
 155 typedef void (*op_fill_func)(uint8_t *block/*align width (8 or 16)*/, uint8_t value, int line_size, int h);
 156
 157 #define DEF_OLD_QPEL(name)\
 158 void ff_put_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 159 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 160 void ff_avg_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 161
 162 DEF_OLD_QPEL(qpel16_mc11_old_c)
 163 DEF_OLD_QPEL(qpel16_mc31_old_c)
 164 DEF_OLD_QPEL(qpel16_mc12_old_c)
 165 DEF_OLD_QPEL(qpel16_mc32_old_c)
 166 DEF_OLD_QPEL(qpel16_mc13_old_c)
 167 DEF_OLD_QPEL(qpel16_mc33_old_c)
 168 DEF_OLD_QPEL(qpel8_mc11_old_c)
 169 DEF_OLD_QPEL(qpel8_mc31_old_c)
 170 DEF_OLD_QPEL(qpel8_mc12_old_c)
 171 DEF_OLD_QPEL(qpel8_mc32_old_c)
 172 DEF_OLD_QPEL(qpel8_mc13_old_c)
 173 DEF_OLD_QPEL(qpel8_mc33_old_c)
 174
 175 #define CALL_2X_PIXELS(a, b, n)\
 176 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
 177     b(block  , pixels  , line_size, h);\
 178     b(block+n, pixels+n, line_size, h);\
 179 }
 180
 181 /* motion estimation */
 182 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
 183 // although currently h<4 is not used as functions with width <8 are neither used nor implemented
 184 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))*/;
 185
 186 /**
 187  * Scantable.
 188  */
 189 typedef struct ScanTable{
 190     const uint8_t *scantable;
 191     uint8_t permutated[64];
 192     uint8_t raster_end[64];
 193 #if ARCH_PPC
 194                 /** Used by dct_quantize_altivec to find last-non-zero */
 195     DECLARE_ALIGNED(16, uint8_t, inverse)[64];
 196 #endif
 197 } ScanTable;
 198
 199 void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
 200
 201 void ff_emulated_edge_mc(uint8_t *buf, uint8_t *src, int linesize,
 202                          int block_w, int block_h,
 203                          int src_x, int src_y, int w, int h);
 204
 205 /**
 206  * DSPContext.
 207  */
 208 typedef struct DSPContext {
 209     /* pixel ops : interface with DCT */
 210     void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
 211     void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
 212     void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 213     void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 214     void (*put_pixels_nonclamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 215     void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 216     void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size);
 217     void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size);
 218     int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/);
 219     /**
 220      * translational global motion compensation.
 221      */
 222     void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
 223     /**
 224      * global motion compensation.
 225      */
 226     void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
 227                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 228     void (*clear_block)(DCTELEM *block/*align 16*/);
 229     void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
 230     int (*pix_sum)(uint8_t * pix, int line_size);
 231     int (*pix_norm1)(uint8_t * pix, int line_size);
 232 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
 233
 234     me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
 235     me_cmp_func sse[6];
 236     me_cmp_func hadamard8_diff[6];
 237     me_cmp_func dct_sad[6];
 238     me_cmp_func quant_psnr[6];
 239     me_cmp_func bit[6];
 240     me_cmp_func rd[6];
 241     me_cmp_func vsad[6];
 242     me_cmp_func vsse[6];
 243     me_cmp_func nsse[6];
 244     me_cmp_func w53[6];
 245     me_cmp_func w97[6];
 246     me_cmp_func dct_max[6];
 247     me_cmp_func dct264_sad[6];
 248
 249     me_cmp_func me_pre_cmp[6];
 250     me_cmp_func me_cmp[6];
 251     me_cmp_func me_sub_cmp[6];
 252     me_cmp_func mb_cmp[6];
 253     me_cmp_func ildct_cmp[6]; //only width 16 used
 254     me_cmp_func frame_skip_cmp[6]; //only width 8 used
 255
 256     int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
 257                              int size);
 258
 259     /**
 260      * Halfpel motion compensation with rounding (a+b+1)>>1.
 261      * this is an array[4][4] of motion compensation functions for 4
 262      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 263      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 264      * @param block destination where the result is stored
 265      * @param pixels source
 266      * @param line_size number of bytes in a horizontal line of block
 267      * @param h height
 268      */
 269     op_pixels_func put_pixels_tab[4][4];
 270
 271     /**
 272      * Halfpel motion compensation with rounding (a+b+1)>>1.
 273      * This is an array[4][4] of motion compensation functions for 4
 274      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 275      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 276      * @param block destination into which the result is averaged (a+b+1)>>1
 277      * @param pixels source
 278      * @param line_size number of bytes in a horizontal line of block
 279      * @param h height
 280      */
 281     op_pixels_func avg_pixels_tab[4][4];
 282
 283     /**
 284      * Halfpel motion compensation with no rounding (a+b)>>1.
 285      * this is an array[2][4] of motion compensation functions for 2
 286      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 287      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 288      * @param block destination where the result is stored
 289      * @param pixels source
 290      * @param line_size number of bytes in a horizontal line of block
 291      * @param h height
 292      */
 293     op_pixels_func put_no_rnd_pixels_tab[4][4];
 294
 295     /**
 296      * Halfpel motion compensation with no rounding (a+b)>>1.
 297      * this is an array[2][4] of motion compensation functions for 2
 298      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 299      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 300      * @param block destination into which the result is averaged (a+b)>>1
 301      * @param pixels source
 302      * @param line_size number of bytes in a horizontal line of block
 303      * @param h height
 304      */
 305     op_pixels_func avg_no_rnd_pixels_tab[4][4];
 306
 307     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);
 308
 309     /**
 310      * Thirdpel motion compensation with rounding (a+b+1)>>1.
 311      * this is an array[12] of motion compensation functions for the 9 thirdpe
 312      * positions<br>
 313      * *pixels_tab[ xthirdpel + 4*ythirdpel ]
 314      * @param block destination where the result is stored
 315      * @param pixels source
 316      * @param line_size number of bytes in a horizontal line of block
 317      * @param h height
 318      */
 319     tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 320     tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 321
 322     qpel_mc_func put_qpel_pixels_tab[2][16];
 323     qpel_mc_func avg_qpel_pixels_tab[2][16];
 324     qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
 325     qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
 326     qpel_mc_func put_mspel_pixels_tab[8];
 327
 328     /**
 329      * h264 Chroma MC
 330      */
 331     h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
 332     h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
 333     /* This is really one func used in VC-1 decoding */
 334     h264_chroma_mc_func put_no_rnd_vc1_chroma_pixels_tab[3];
 335     h264_chroma_mc_func avg_no_rnd_vc1_chroma_pixels_tab[3];
 336
 337     qpel_mc_func put_h264_qpel_pixels_tab[4][16];
 338     qpel_mc_func avg_h264_qpel_pixels_tab[4][16];
 339
 340     qpel_mc_func put_2tap_qpel_pixels_tab[4][16];
 341     qpel_mc_func avg_2tap_qpel_pixels_tab[4][16];
 342
 343     h264_weight_func weight_h264_pixels_tab[10];
 344     h264_biweight_func biweight_h264_pixels_tab[10];
 345
 346     /* AVS specific */
 347     qpel_mc_func put_cavs_qpel_pixels_tab[2][16];
 348     qpel_mc_func avg_cavs_qpel_pixels_tab[2][16];
 349     void (*cavs_filter_lv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 350     void (*cavs_filter_lh)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 351     void (*cavs_filter_cv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 352     void (*cavs_filter_ch)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 353     void (*cavs_idct8_add)(uint8_t *dst, DCTELEM *block, int stride);
 354
 355     me_cmp_func pix_abs[2][4];
 356
 357     /* huffyuv specific */
 358     void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
 359     void (*add_bytes_l2)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 16*/, int w);
 360     void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
 361     /**
 362      * subtract huffyuv's variant of median prediction
 363      * note, this might read from src1[-1], src2[-1]
 364      */
 365     void (*sub_hfyu_median_prediction)(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top);
 366     void (*add_hfyu_median_prediction)(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
 367     int  (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src, int w, int left);
 368     void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha);
 369     /* this might write to dst[w] */
 370     void (*add_png_paeth_prediction)(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
 371     void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
 372
 373     void (*h264_v_loop_filter_luma)(uint8_t *pix/*align 16*/, int stride, int alpha, int beta, int8_t *tc0);
 374     void (*h264_h_loop_filter_luma)(uint8_t *pix/*align 4 */, int stride, int alpha, int beta, int8_t *tc0);
 375     /* v/h_loop_filter_luma_intra: align 16 */
 376     void (*h264_v_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
 377     void (*h264_h_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
 378     void (*h264_v_loop_filter_chroma)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta, int8_t *tc0);
 379     void (*h264_h_loop_filter_chroma)(uint8_t *pix/*align 4*/, int stride, int alpha, int beta, int8_t *tc0);
 380     void (*h264_v_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
 381     void (*h264_h_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
 382     // h264_loop_filter_strength: simd only. the C version is inlined in h264.c
 383     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],
 384                                       int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field);
 385
 386     void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
 387     void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
 388
 389     void (*h261_loop_filter)(uint8_t *src, int stride);
 390
 391     void (*x8_v_loop_filter)(uint8_t *src, int stride, int qscale);
 392     void (*x8_h_loop_filter)(uint8_t *src, int stride, int qscale);
 393
 394     void (*vp3_v_loop_filter)(uint8_t *src, int stride, int *bounding_values);
 395     void (*vp3_h_loop_filter)(uint8_t *src, int stride, int *bounding_values);
 396
 397     void (*vp6_filter_diag4)(uint8_t *dst, uint8_t *src, int stride,
 398                              const int16_t *h_weights,const int16_t *v_weights);
 399
 400     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
 401     void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize);
 402     void (*ac3_downmix)(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len);
 403     /* no alignment needed */
 404     void (*lpc_compute_autocorr)(const int32_t *data, int len, int lag, double *autoc);
 405     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
 406     void (*vector_fmul)(float *dst, const float *src, int len);
 407     void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len);
 408     /* assume len is a multiple of 8, and src arrays are 16-byte aligned */
 409     void (*vector_fmul_add)(float *dst, const float *src0, const float *src1, const float *src2, int len);
 410     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
 411     void (*vector_fmul_window)(float *dst, const float *src0, const float *src1, const float *win, float add_bias, int len);
 412     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
 413     void (*int32_to_float_fmul_scalar)(float *dst, const int *src, float mul, int len);
 414     void (*vector_clipf)(float *dst /* align 16 */, const float *src /* align 16 */, float min, float max, int len /* align 16 */);
 415     /**
 416      * Multiply a vector of floats by a scalar float.  Source and
 417      * destination vectors must overlap exactly or not at all.
 418      * @param dst result vector, 16-byte aligned
 419      * @param src input vector, 16-byte aligned
 420      * @param mul scalar value
 421      * @param len length of vector, multiple of 4
 422      */
 423     void (*vector_fmul_scalar)(float *dst, const float *src, float mul,
 424                                int len);
 425     /**
 426      * Multiply a vector of floats by concatenated short vectors of
 427      * floats and by a scalar float.  Source and destination vectors
 428      * must overlap exactly or not at all.
 429      * [0]: short vectors of length 2, 8-byte aligned
 430      * [1]: short vectors of length 4, 16-byte aligned
 431      * @param dst output vector, 16-byte aligned
 432      * @param src input vector, 16-byte aligned
 433      * @param sv  array of pointers to short vectors
 434      * @param mul scalar value
 435      * @param len number of elements in src and dst, multiple of 4
 436      */
 437     void (*vector_fmul_sv_scalar[2])(float *dst, const float *src,
 438                                      const float **sv, float mul, int len);
 439     /**
 440      * Multiply short vectors of floats by a scalar float, store
 441      * concatenated result.
 442      * [0]: short vectors of length 2, 8-byte aligned
 443      * [1]: short vectors of length 4, 16-byte aligned
 444      * @param dst output vector, 16-byte aligned
 445      * @param sv  array of pointers to short vectors
 446      * @param mul scalar value
 447      * @param len number of output elements, multiple of 4
 448      */
 449     void (*sv_fmul_scalar[2])(float *dst, const float **sv,
 450                               float mul, int len);
 451     /**
 452      * Calculate the scalar product of two vectors of floats.
 453      * @param v1  first vector, 16-byte aligned
 454      * @param v2  second vector, 16-byte aligned
 455      * @param len length of vectors, multiple of 4
 456      */
 457     float (*scalarproduct_float)(const float *v1, const float *v2, int len);
 458     /**
 459      * Calculate the sum and difference of two vectors of floats.
 460      * @param v1  first input vector, sum output, 16-byte aligned
 461      * @param v2  second input vector, difference output, 16-byte aligned
 462      * @param len length of vectors, multiple of 4
 463      */
 464     void (*butterflies_float)(float *restrict v1, float *restrict v2, int len);
 465
 466     /* C version: convert floats from the range [384.0,386.0] to ints in [-32768,32767]
 467      * simd versions: convert floats from [-32768.0,32767.0] without rescaling and arrays are 16byte aligned */
 468     void (*float_to_int16)(int16_t *dst, const float *src, long len);
 469     void (*float_to_int16_interleave)(int16_t *dst, const float **src, long len, int channels);
 470
 471     /* (I)DCT */
 472     void (*fdct)(DCTELEM *block/* align 16*/);
 473     void (*fdct248)(DCTELEM *block/* align 16*/);
 474
 475     /* IDCT really*/
 476     void (*idct)(DCTELEM *block/* align 16*/);
 477
 478     /**
 479      * block -> idct -> clip to unsigned 8 bit -> dest.
 480      * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
 481      * @param line_size size in bytes of a horizontal line of dest
 482      */
 483     void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 484
 485     /**
 486      * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
 487      * @param line_size size in bytes of a horizontal line of dest
 488      */
 489     void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 490
 491     /**
 492      * idct input permutation.
 493      * several optimized IDCTs need a permutated input (relative to the normal order of the reference
 494      * IDCT)
 495      * this permutation must be performed before the idct_put/add, note, normally this can be merged
 496      * with the zigzag/alternate scan<br>
 497      * an example to avoid confusion:
 498      * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
 499      * - (x -> referece dct -> reference idct -> x)
 500      * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
 501      * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
 502      */
 503     uint8_t idct_permutation[64];
 504     int idct_permutation_type;
 505 #define FF_NO_IDCT_PERM 1
 506 #define FF_LIBMPEG2_IDCT_PERM 2
 507 #define FF_SIMPLE_IDCT_PERM 3
 508 #define FF_TRANSPOSE_IDCT_PERM 4
 509 #define FF_PARTTRANS_IDCT_PERM 5
 510 #define FF_SSE2_IDCT_PERM 6
 511
 512     int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
 513     void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
 514 #define BASIS_SHIFT 16
 515 #define RECON_SHIFT 6
 516
 517     void (*draw_edges)(uint8_t *buf, int wrap, int width, int height, int w);
 518 #define EDGE_WIDTH 16
 519
 520     /* h264 functions */
 521     /* NOTE!!! if you implement any of h264_idct8_add, h264_idct8_add4 then you must implement all of them
 522        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
 523         The reason for above, is that no 2 out of one list may use a different permutation.
 524     */
 525     void (*h264_idct_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
 526     void (*h264_idct8_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
 527     void (*h264_idct_dc_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
 528     void (*h264_idct8_dc_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
 529     void (*h264_dct)(DCTELEM block[4][4]);
 530     void (*h264_idct_add16)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 531     void (*h264_idct8_add4)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 532     void (*h264_idct_add8)(uint8_t **dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 533     void (*h264_idct_add16intra)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 534
 535     void (*prefetch)(void *mem, int stride, int h);
 536
 537     void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 538
 539     /* mlp/truehd functions */
 540     void (*mlp_filter_channel)(int32_t *state, const int32_t *coeff,
 541                                int firorder, int iirorder,
 542                                unsigned int filter_shift, int32_t mask, int blocksize,
 543                                int32_t *sample_buffer);
 544
 545     /* vc1 functions */
 546     void (*vc1_inv_trans_8x8)(DCTELEM *b);
 547     void (*vc1_inv_trans_8x4)(uint8_t *dest, int line_size, DCTELEM *block);
 548     void (*vc1_inv_trans_4x8)(uint8_t *dest, int line_size, DCTELEM *block);
 549     void (*vc1_inv_trans_4x4)(uint8_t *dest, int line_size, DCTELEM *block);
 550     void (*vc1_inv_trans_8x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 551     void (*vc1_inv_trans_8x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 552     void (*vc1_inv_trans_4x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 553     void (*vc1_inv_trans_4x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 554     void (*vc1_v_overlap)(uint8_t* src, int stride);
 555     void (*vc1_h_overlap)(uint8_t* src, int stride);
 556     void (*vc1_v_loop_filter4)(uint8_t *src, int stride, int pq);
 557     void (*vc1_h_loop_filter4)(uint8_t *src, int stride, int pq);
 558     void (*vc1_v_loop_filter8)(uint8_t *src, int stride, int pq);
 559     void (*vc1_h_loop_filter8)(uint8_t *src, int stride, int pq);
 560     void (*vc1_v_loop_filter16)(uint8_t *src, int stride, int pq);
 561     void (*vc1_h_loop_filter16)(uint8_t *src, int stride, int pq);
 562     /* put 8x8 block with bicubic interpolation and quarterpel precision
 563      * last argument is actually round value instead of height
 564      */
 565     op_pixels_func put_vc1_mspel_pixels_tab[16];
 566     op_pixels_func avg_vc1_mspel_pixels_tab[16];
 567
 568     /* intrax8 functions */
 569     void (*x8_spatial_compensation[12])(uint8_t *src , uint8_t *dst, int linesize);
 570     void (*x8_setup_spatial_compensation)(uint8_t *src, uint8_t *dst, int linesize,
 571            int * range, int * sum,  int edges);
 572
 573     /**
 574      * Calculate scalar product of two vectors.
 575      * @param len length of vectors, should be multiple of 16
 576      * @param shift number of bits to discard from product
 577      */
 578     int32_t (*scalarproduct_int16)(int16_t *v1, int16_t *v2/*align 16*/, int len, int shift);
 579     /* ape functions */
 580     /**
 581      * Calculate scalar product of v1 and v2,
 582      * and v1[i] += v3[i] * mul
 583      * @param len length of vectors, should be multiple of 16
 584      */
 585     int32_t (*scalarproduct_and_madd_int16)(int16_t *v1/*align 16*/, int16_t *v2, int16_t *v3, int len, int mul);
 586
 587     /* rv30 functions */
 588     qpel_mc_func put_rv30_tpel_pixels_tab[4][16];
 589     qpel_mc_func avg_rv30_tpel_pixels_tab[4][16];
 590
 591     /* rv40 functions */
 592     qpel_mc_func put_rv40_qpel_pixels_tab[4][16];
 593     qpel_mc_func avg_rv40_qpel_pixels_tab[4][16];
 594     h264_chroma_mc_func put_rv40_chroma_pixels_tab[3];
 595     h264_chroma_mc_func avg_rv40_chroma_pixels_tab[3];
 596
 597     /* bink functions */
 598     op_fill_func fill_block_tab[2];
 599     void (*scale_block)(const uint8_t src[64]/*align 8*/, uint8_t *dst/*align 8*/, int linesize);
 600 } DSPContext;
 601
 602 void dsputil_static_init(void);
 603 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
 604
 605 int ff_check_alignment(void);
 606
 607 /**
 608  * permute block according to permuatation.
 609  * @param last last non zero element in scantable order
 610  */
 611 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
 612
 613 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
 614
 615 #define         BYTE_VEC32(c)   ((c)*0x01010101UL)
 616
 617 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
 618 {
 619     return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 620 }
 621
 622 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
 623 {
 624     return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 625 }
 626
 627 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 628     switch(type&0xFF){
 629     default:
 630     case FF_CMP_SAD:
 631         return lambda>>FF_LAMBDA_SHIFT;
 632     case FF_CMP_DCT:
 633         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 634     case FF_CMP_W53:
 635         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 636     case FF_CMP_W97:
 637         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 638     case FF_CMP_SATD:
 639     case FF_CMP_DCT264:
 640         return (2*lambda)>>FF_LAMBDA_SHIFT;
 641     case FF_CMP_RD:
 642     case FF_CMP_PSNR:
 643     case FF_CMP_SSE:
 644     case FF_CMP_NSSE:
 645         return lambda2>>FF_LAMBDA_SHIFT;
 646     case FF_CMP_BIT:
 647         return 1;
 648     }
 649 }
 650
 651 /**
 652  * Empty mmx state.
 653  * this must be called between any dsp function and float/double code.
 654  * for example sin(); dsp->idct_put(); emms_c(); cos()
 655  */
 656 #define emms_c()
 657
 658 /* should be defined by architectures supporting
 659    one or more MultiMedia extension */
 660 int mm_support(void);
 661 extern int mm_flags;
 662
 663 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
 664 void dsputil_init_arm(DSPContext* c, AVCodecContext *avctx);
 665 void dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx);
 666 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
 667 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
 668 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
 669 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
 670 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
 671 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
 672
 673 void ff_dsputil_init_dwt(DSPContext *c);
 674 void ff_cavsdsp_init(DSPContext* c, AVCodecContext *avctx);
 675 void ff_rv30dsp_init(DSPContext* c, AVCodecContext *avctx);
 676 void ff_rv40dsp_init(DSPContext* c, AVCodecContext *avctx);
 677 void ff_vc1dsp_init(DSPContext* c, AVCodecContext *avctx);
 678 void ff_intrax8dsp_init(DSPContext* c, AVCodecContext *avctx);
 679 void ff_mlp_init(DSPContext* c, AVCodecContext *avctx);
 680 void ff_mlp_init_x86(DSPContext* c, AVCodecContext *avctx);
 681
 682 #if HAVE_MMX
 683
 684 #undef emms_c
 685
 686 static inline void emms(void)
 687 {
 688     __asm__ volatile ("emms;":::"memory");
 689 }
 690
 691
 692 #define emms_c() \
 693 {\
 694     if (mm_flags & FF_MM_MMX)\
 695         emms();\
 696 }
 697
 698 #elif ARCH_ARM
 699
 700 #if HAVE_NEON
 701 #   define STRIDE_ALIGN 16
 702 #endif
 703
 704 #elif ARCH_PPC
 705
 706 #define STRIDE_ALIGN 16
 707
 708 #elif HAVE_MMI
 709
 710 #define STRIDE_ALIGN 16
 711
 712 #else
 713
 714 #define mm_flags 0
 715 #define mm_support() 0
 716
 717 #endif
 718
 719 #ifndef STRIDE_ALIGN
 720 #   define STRIDE_ALIGN 8
 721 #endif
 722
 723 #define LOCAL_ALIGNED(a, t, v, s, ...)                          \
 724     uint8_t la_##v[sizeof(t s __VA_ARGS__) + (a)];              \
 725     t (*v) __VA_ARGS__ = (void *)FFALIGN((uintptr_t)la_##v, a)
 726
 727 #if HAVE_LOCAL_ALIGNED_8
 728 #   define LOCAL_ALIGNED_8(t, v, s, ...) DECLARE_ALIGNED(8, t, v) s __VA_ARGS__
 729 #else
 730 #   define LOCAL_ALIGNED_8(t, v, s, ...) LOCAL_ALIGNED(8, t, v, s, __VA_ARGS__)
 731 #endif
 732
 733 #if HAVE_LOCAL_ALIGNED_16
 734 #   define LOCAL_ALIGNED_16(t, v, s, ...) DECLARE_ALIGNED(16, t, v) s __VA_ARGS__
 735 #else
 736 #   define LOCAL_ALIGNED_16(t, v, s, ...) LOCAL_ALIGNED(16, t, v, s, __VA_ARGS__)
 737 #endif
 738
 739 /* PSNR */
 740 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
 741               int orig_linesize[3], int coded_linesize,
 742               AVCodecContext *avctx);
 743
 744 #define WRAPPER8_16(name8, name16)\
 745 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 746     return name8(s, dst           , src           , stride, h)\
 747           +name8(s, dst+8         , src+8         , stride, h);\
 748 }
 749
 750 #define WRAPPER8_16_SQ(name8, name16)\
 751 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 752     int score=0;\
 753     score +=name8(s, dst           , src           , stride, 8);\
 754     score +=name8(s, dst+8         , src+8         , stride, 8);\
 755     if(h==16){\
 756         dst += 8*stride;\
 757         src += 8*stride;\
 758         score +=name8(s, dst           , src           , stride, 8);\
 759         score +=name8(s, dst+8         , src+8         , stride, 8);\
 760     }\
 761     return score;\
 762 }
 763
 764
 765 static inline void copy_block2(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 766 {
 767     int i;
 768     for(i=0; i<h; i++)
 769     {
 770         AV_WN16(dst   , AV_RN16(src   ));
 771         dst+=dstStride;
 772         src+=srcStride;
 773     }
 774 }
 775
 776 static inline void copy_block4(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 777 {
 778     int i;
 779     for(i=0; i<h; i++)
 780     {
 781         AV_WN32(dst   , AV_RN32(src   ));
 782         dst+=dstStride;
 783         src+=srcStride;
 784     }
 785 }
 786
 787 static inline void copy_block8(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 788 {
 789     int i;
 790     for(i=0; i<h; i++)
 791     {
 792         AV_WN32(dst   , AV_RN32(src   ));
 793         AV_WN32(dst+4 , AV_RN32(src+4 ));
 794         dst+=dstStride;
 795         src+=srcStride;
 796     }
 797 }
 798
 799 static inline void copy_block9(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 800 {
 801     int i;
 802     for(i=0; i<h; i++)
 803     {
 804         AV_WN32(dst   , AV_RN32(src   ));
 805         AV_WN32(dst+4 , AV_RN32(src+4 ));
 806         dst[8]= src[8];
 807         dst+=dstStride;
 808         src+=srcStride;
 809     }
 810 }
 811
 812 static inline void copy_block16(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 813 {
 814     int i;
 815     for(i=0; i<h; i++)
 816     {
 817         AV_WN32(dst   , AV_RN32(src   ));
 818         AV_WN32(dst+4 , AV_RN32(src+4 ));
 819         AV_WN32(dst+8 , AV_RN32(src+8 ));
 820         AV_WN32(dst+12, AV_RN32(src+12));
 821         dst+=dstStride;
 822         src+=srcStride;
 823     }
 824 }
 825
 826 static inline void copy_block17(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 827 {
 828     int i;
 829     for(i=0; i<h; i++)
 830     {
 831         AV_WN32(dst   , AV_RN32(src   ));
 832         AV_WN32(dst+4 , AV_RN32(src+4 ));
 833         AV_WN32(dst+8 , AV_RN32(src+8 ));
 834         AV_WN32(dst+12, AV_RN32(src+12));
 835         dst[16]= src[16];
 836         dst+=dstStride;
 837         src+=srcStride;
 838     }
 839 }
 840
 841 #endif /* AVCODEC_DSPUTIL_H */