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 Libav.
   7  *
   8  * Libav 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  * Libav 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 Libav; 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
  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 ff_fdct_ifast (DCTELEM *data);
  42 void ff_fdct_ifast248 (DCTELEM *data);
  43 void ff_jpeg_fdct_islow_8(DCTELEM *data);
  44 void ff_jpeg_fdct_islow_10(DCTELEM *data);
  45 void ff_fdct248_islow_8(DCTELEM *data);
  46 void ff_fdct248_islow_10(DCTELEM *data);
  47
  48 void ff_j_rev_dct (DCTELEM *data);
  49 void ff_wmv2_idct_c(DCTELEM *data);
  50
  51 void ff_fdct_mmx(DCTELEM *block);
  52 void ff_fdct_mmx2(DCTELEM *block);
  53 void ff_fdct_sse2(DCTELEM *block);
  54
  55 #define H264_IDCT(depth) \
  56 void ff_h264_idct8_add_ ## depth ## _c(uint8_t *dst, DCTELEM *block, int stride);\
  57 void ff_h264_idct_add_ ## depth ## _c(uint8_t *dst, DCTELEM *block, int stride);\
  58 void ff_h264_idct8_dc_add_ ## depth ## _c(uint8_t *dst, DCTELEM *block, int stride);\
  59 void ff_h264_idct_dc_add_ ## depth ## _c(uint8_t *dst, DCTELEM *block, int stride);\
  60 void ff_h264_idct_add16_ ## depth ## _c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);\
  61 void ff_h264_idct_add16intra_ ## depth ## _c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);\
  62 void ff_h264_idct8_add4_ ## depth ## _c(uint8_t *dst, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);\
  63 void ff_h264_idct_add8_422_ ## depth ## _c(uint8_t **dest, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);\
  64 void ff_h264_idct_add8_ ## depth ## _c(uint8_t **dest, const int *blockoffset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]);\
  65 void ff_h264_luma_dc_dequant_idct_ ## depth ## _c(DCTELEM *output, DCTELEM *input, int qmul);\
  66 void ff_h264_chroma422_dc_dequant_idct_ ## depth ## _c(DCTELEM *block, int qmul);\
  67 void ff_h264_chroma_dc_dequant_idct_ ## depth ## _c(DCTELEM *block, int qmul);
  68
  69 H264_IDCT( 8)
  70 H264_IDCT( 9)
  71 H264_IDCT(10)
  72
  73 void ff_svq3_luma_dc_dequant_idct_c(DCTELEM *output, DCTELEM *input, int qp);
  74 void ff_svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
  75
  76 /* encoding scans */
  77 extern const uint8_t ff_alternate_horizontal_scan[64];
  78 extern const uint8_t ff_alternate_vertical_scan[64];
  79 extern const uint8_t ff_zigzag_direct[64];
  80 extern const uint8_t ff_zigzag248_direct[64];
  81
  82 /* pixel operations */
  83 #define MAX_NEG_CROP 1024
  84
  85 /* temporary */
  86 extern uint32_t ff_squareTbl[512];
  87 extern uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP];
  88
  89 #define PUTAVG_PIXELS(depth)\
  90 void ff_put_pixels8x8_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
  91 void ff_avg_pixels8x8_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
  92 void ff_put_pixels16x16_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);\
  93 void ff_avg_pixels16x16_ ## depth ## _c(uint8_t *dst, uint8_t *src, int stride);
  94
  95 PUTAVG_PIXELS( 8)
  96 PUTAVG_PIXELS( 9)
  97 PUTAVG_PIXELS(10)
  98
  99 #define ff_put_pixels8x8_c ff_put_pixels8x8_8_c
 100 #define ff_avg_pixels8x8_c ff_avg_pixels8x8_8_c
 101 #define ff_put_pixels16x16_c ff_put_pixels16x16_8_c
 102 #define ff_avg_pixels16x16_c ff_avg_pixels16x16_8_c
 103
 104 /* EA functions */
 105 void ff_ea_idct_put_c(uint8_t *dest, int linesize, DCTELEM *block);
 106
 107 /* RV40 functions */
 108 void ff_put_rv40_qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride);
 109 void ff_avg_rv40_qpel16_mc33_c(uint8_t *dst, uint8_t *src, int stride);
 110 void ff_put_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride);
 111 void ff_avg_rv40_qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride);
 112
 113 /* 1/2^n downscaling functions from imgconvert.c */
 114 void ff_shrink22(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 115 void ff_shrink44(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 116 void ff_shrink88(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 117
 118 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
 119               int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 120
 121 /* minimum alignment rules ;)
 122 If you notice errors in the align stuff, need more alignment for some ASM code
 123 for some CPU or need to use a function with less aligned data then send a mail
 124 to the libav-devel mailing list, ...
 125
 126 !warning These alignments might not match reality, (missing attribute((align))
 127 stuff somewhere possible).
 128 I (Michael) did not check them, these are just the alignments which I think
 129 could be reached easily ...
 130
 131 !future video codecs might need functions with less strict alignment
 132 */
 133
 134 /*
 135 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
 136 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
 137 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
 138 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
 139 void clear_blocks_c(DCTELEM *blocks);
 140 */
 141
 142 /* add and put pixel (decoding) */
 143 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
 144 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller than 4
 145 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
 146 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);
 147 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 148 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
 149
 150 typedef void (*op_fill_func)(uint8_t *block/*align width (8 or 16)*/, uint8_t value, int line_size, int h);
 151
 152 #define DEF_OLD_QPEL(name)\
 153 void ff_put_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 154 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 155 void ff_avg_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 156
 157 DEF_OLD_QPEL(qpel16_mc11_old_c)
 158 DEF_OLD_QPEL(qpel16_mc31_old_c)
 159 DEF_OLD_QPEL(qpel16_mc12_old_c)
 160 DEF_OLD_QPEL(qpel16_mc32_old_c)
 161 DEF_OLD_QPEL(qpel16_mc13_old_c)
 162 DEF_OLD_QPEL(qpel16_mc33_old_c)
 163 DEF_OLD_QPEL(qpel8_mc11_old_c)
 164 DEF_OLD_QPEL(qpel8_mc31_old_c)
 165 DEF_OLD_QPEL(qpel8_mc12_old_c)
 166 DEF_OLD_QPEL(qpel8_mc32_old_c)
 167 DEF_OLD_QPEL(qpel8_mc13_old_c)
 168 DEF_OLD_QPEL(qpel8_mc33_old_c)
 169
 170 #define CALL_2X_PIXELS(a, b, n)\
 171 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
 172     b(block  , pixels  , line_size, h);\
 173     b(block+n, pixels+n, line_size, h);\
 174 }
 175
 176 /* motion estimation */
 177 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller than 2
 178 // although currently h<4 is not used as functions with width <8 are neither used nor implemented
 179 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))*/;
 180
 181 /**
 182  * Scantable.
 183  */
 184 typedef struct ScanTable{
 185     const uint8_t *scantable;
 186     uint8_t permutated[64];
 187     uint8_t raster_end[64];
 188 } ScanTable;
 189
 190 void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
 191 void ff_init_scantable_permutation(uint8_t *idct_permutation,
 192                                    int idct_permutation_type);
 193
 194 #define EMULATED_EDGE(depth) \
 195 void ff_emulated_edge_mc_ ## depth (uint8_t *buf, const uint8_t *src, int linesize,\
 196                          int block_w, int block_h,\
 197                          int src_x, int src_y, int w, int h);
 198
 199 EMULATED_EDGE(8)
 200 EMULATED_EDGE(9)
 201 EMULATED_EDGE(10)
 202
 203 void ff_add_pixels_clamped_c(const DCTELEM *block, uint8_t *dest, int linesize);
 204 void ff_put_pixels_clamped_c(const DCTELEM *block, uint8_t *dest, int linesize);
 205 void ff_put_signed_pixels_clamped_c(const DCTELEM *block, uint8_t *dest, int linesize);
 206
 207 /**
 208  * DSPContext.
 209  */
 210 typedef struct DSPContext {
 211     /**
 212      * Size of DCT coefficients.
 213      */
 214     int dct_bits;
 215
 216     /* pixel ops : interface with DCT */
 217     void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
 218     void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
 219     void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 220     void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 221     void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 222     void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size);
 223     void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size);
 224     int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/);
 225     /**
 226      * Motion estimation with emulated edge values.
 227      * @param buf pointer to destination buffer (unaligned)
 228      * @param src pointer to pixel source (unaligned)
 229      * @param linesize width (in pixels) for src/buf
 230      * @param block_w number of pixels (per row) to copy to buf
 231      * @param block_h nummber of pixel rows to copy to buf
 232      * @param src_x offset of src to start of row - this may be negative
 233      * @param src_y offset of src to top of image - this may be negative
 234      * @param w width of src in pixels
 235      * @param h height of src in pixels
 236      */
 237     void (*emulated_edge_mc)(uint8_t *buf, const uint8_t *src, int linesize,
 238                              int block_w, int block_h,
 239                              int src_x, int src_y, int w, int h);
 240     /**
 241      * translational global motion compensation.
 242      */
 243     void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
 244     /**
 245      * global motion compensation.
 246      */
 247     void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
 248                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 249     void (*clear_block)(DCTELEM *block/*align 16*/);
 250     void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
 251     int (*pix_sum)(uint8_t * pix, int line_size);
 252     int (*pix_norm1)(uint8_t * pix, int line_size);
 253 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
 254
 255     me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
 256     me_cmp_func sse[6];
 257     me_cmp_func hadamard8_diff[6];
 258     me_cmp_func dct_sad[6];
 259     me_cmp_func quant_psnr[6];
 260     me_cmp_func bit[6];
 261     me_cmp_func rd[6];
 262     me_cmp_func vsad[6];
 263     me_cmp_func vsse[6];
 264     me_cmp_func nsse[6];
 265     me_cmp_func w53[6];
 266     me_cmp_func w97[6];
 267     me_cmp_func dct_max[6];
 268     me_cmp_func dct264_sad[6];
 269
 270     me_cmp_func me_pre_cmp[6];
 271     me_cmp_func me_cmp[6];
 272     me_cmp_func me_sub_cmp[6];
 273     me_cmp_func mb_cmp[6];
 274     me_cmp_func ildct_cmp[6]; //only width 16 used
 275     me_cmp_func frame_skip_cmp[6]; //only width 8 used
 276
 277     int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
 278                              int size);
 279
 280     /**
 281      * Halfpel motion compensation with rounding (a+b+1)>>1.
 282      * this is an array[4][4] of motion compensation functions for 4
 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 where the result is stored
 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 put_pixels_tab[4][4];
 291
 292     /**
 293      * Halfpel motion compensation with rounding (a+b+1)>>1.
 294      * This is an array[4][4] of motion compensation functions for 4
 295      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 296      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 297      * @param block destination into which the result is averaged (a+b+1)>>1
 298      * @param pixels source
 299      * @param line_size number of bytes in a horizontal line of block
 300      * @param h height
 301      */
 302     op_pixels_func avg_pixels_tab[4][4];
 303
 304     /**
 305      * Halfpel motion compensation with no rounding (a+b)>>1.
 306      * this is an array[2][4] of motion compensation functions for 2
 307      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 308      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 309      * @param block destination where the result is stored
 310      * @param pixels source
 311      * @param line_size number of bytes in a horizontal line of block
 312      * @param h height
 313      */
 314     op_pixels_func put_no_rnd_pixels_tab[4][4];
 315
 316     /**
 317      * Halfpel motion compensation with no rounding (a+b)>>1.
 318      * this is an array[2][4] of motion compensation functions for 2
 319      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 320      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 321      * @param block destination into which the result is averaged (a+b)>>1
 322      * @param pixels source
 323      * @param line_size number of bytes in a horizontal line of block
 324      * @param h height
 325      */
 326     op_pixels_func avg_no_rnd_pixels_tab[4][4];
 327
 328     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);
 329
 330     /**
 331      * Thirdpel motion compensation with rounding (a+b+1)>>1.
 332      * this is an array[12] of motion compensation functions for the 9 thirdpe
 333      * positions<br>
 334      * *pixels_tab[ xthirdpel + 4*ythirdpel ]
 335      * @param block destination where the result is stored
 336      * @param pixels source
 337      * @param line_size number of bytes in a horizontal line of block
 338      * @param h height
 339      */
 340     tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 341     tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 342
 343     qpel_mc_func put_qpel_pixels_tab[2][16];
 344     qpel_mc_func avg_qpel_pixels_tab[2][16];
 345     qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
 346     qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
 347     qpel_mc_func put_mspel_pixels_tab[8];
 348
 349     /**
 350      * h264 Chroma MC
 351      */
 352     h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
 353     h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
 354
 355     qpel_mc_func put_h264_qpel_pixels_tab[4][16];
 356     qpel_mc_func avg_h264_qpel_pixels_tab[4][16];
 357
 358     qpel_mc_func put_2tap_qpel_pixels_tab[4][16];
 359     qpel_mc_func avg_2tap_qpel_pixels_tab[4][16];
 360
 361     me_cmp_func pix_abs[2][4];
 362
 363     /* huffyuv specific */
 364     void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
 365     void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
 366     /**
 367      * subtract huffyuv's variant of median prediction
 368      * note, this might read from src1[-1], src2[-1]
 369      */
 370     void (*sub_hfyu_median_prediction)(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int w, int *left, int *left_top);
 371     void (*add_hfyu_median_prediction)(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
 372     int  (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src, int w, int left);
 373     void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha);
 374     void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
 375     void (*bswap16_buf)(uint16_t *dst, const uint16_t *src, int len);
 376
 377     void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
 378     void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
 379
 380     void (*h261_loop_filter)(uint8_t *src, int stride);
 381
 382     void (*x8_v_loop_filter)(uint8_t *src, int stride, int qscale);
 383     void (*x8_h_loop_filter)(uint8_t *src, int stride, int qscale);
 384
 385     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
 386     void (*vorbis_inverse_coupling)(float *mag, float *ang, int blocksize);
 387     void (*ac3_downmix)(float (*samples)[256], float (*matrix)[2], int out_ch, int in_ch, int len);
 388     /* assume len is a multiple of 16, and arrays are 32-byte aligned */
 389     void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len);
 390     /* assume len is a multiple of 8, and src arrays are 16-byte aligned */
 391     void (*vector_fmul_add)(float *dst, const float *src0, const float *src1, const float *src2, int len);
 392     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
 393     void (*vector_fmul_window)(float *dst, const float *src0, const float *src1, const float *win, int len);
 394     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
 395     void (*vector_clipf)(float *dst /* align 16 */, const float *src /* align 16 */, float min, float max, int len /* align 16 */);
 396     /**
 397      * Multiply a vector of floats by a scalar float.  Source and
 398      * destination vectors must overlap exactly or not at all.
 399      * @param dst result vector, 16-byte aligned
 400      * @param src input vector, 16-byte aligned
 401      * @param mul scalar value
 402      * @param len length of vector, multiple of 4
 403      */
 404     void (*vector_fmul_scalar)(float *dst, const float *src, float mul,
 405                                int len);
 406     /**
 407      * Calculate the scalar product of two vectors of floats.
 408      * @param v1  first vector, 16-byte aligned
 409      * @param v2  second vector, 16-byte aligned
 410      * @param len length of vectors, multiple of 4
 411      */
 412     float (*scalarproduct_float)(const float *v1, const float *v2, int len);
 413     /**
 414      * Calculate the sum and difference of two vectors of floats.
 415      * @param v1  first input vector, sum output, 16-byte aligned
 416      * @param v2  second input vector, difference output, 16-byte aligned
 417      * @param len length of vectors, multiple of 4
 418      */
 419     void (*butterflies_float)(float *restrict v1, float *restrict v2, int len);
 420
 421     /**
 422      * Calculate the sum and difference of two vectors of floats and interleave
 423      * results into a separate output vector of floats, with each sum
 424      * positioned before the corresponding difference.
 425      *
 426      * @param dst  output vector
 427      *             constraints: 16-byte aligned
 428      * @param src0 first input vector
 429      *             constraints: 32-byte aligned
 430      * @param src1 second input vector
 431      *             constraints: 32-byte aligned
 432      * @param len  number of elements in the input
 433      *             constraints: multiple of 8
 434      */
 435     void (*butterflies_float_interleave)(float *dst, const float *src0,
 436                                          const float *src1, int len);
 437
 438     /* (I)DCT */
 439     void (*fdct)(DCTELEM *block/* align 16*/);
 440     void (*fdct248)(DCTELEM *block/* align 16*/);
 441
 442     /* IDCT really*/
 443     void (*idct)(DCTELEM *block/* align 16*/);
 444
 445     /**
 446      * block -> idct -> clip to unsigned 8 bit -> dest.
 447      * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
 448      * @param line_size size in bytes of a horizontal line of dest
 449      */
 450     void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 451
 452     /**
 453      * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
 454      * @param line_size size in bytes of a horizontal line of dest
 455      */
 456     void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 457
 458     /**
 459      * idct input permutation.
 460      * several optimized IDCTs need a permutated input (relative to the normal order of the reference
 461      * IDCT)
 462      * this permutation must be performed before the idct_put/add, note, normally this can be merged
 463      * with the zigzag/alternate scan<br>
 464      * an example to avoid confusion:
 465      * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
 466      * - (x -> reference dct -> reference idct -> x)
 467      * - (x -> reference dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
 468      * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
 469      */
 470     uint8_t idct_permutation[64];
 471     int idct_permutation_type;
 472 #define FF_NO_IDCT_PERM 1
 473 #define FF_LIBMPEG2_IDCT_PERM 2
 474 #define FF_SIMPLE_IDCT_PERM 3
 475 #define FF_TRANSPOSE_IDCT_PERM 4
 476 #define FF_PARTTRANS_IDCT_PERM 5
 477 #define FF_SSE2_IDCT_PERM 6
 478
 479     int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
 480     void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
 481 #define BASIS_SHIFT 16
 482 #define RECON_SHIFT 6
 483
 484     void (*draw_edges)(uint8_t *buf, int wrap, int width, int height, int w, int h, int sides);
 485 #define EDGE_WIDTH 16
 486 #define EDGE_TOP    1
 487 #define EDGE_BOTTOM 2
 488
 489     void (*prefetch)(void *mem, int stride, int h);
 490
 491     void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 492
 493     /* mlp/truehd functions */
 494     void (*mlp_filter_channel)(int32_t *state, const int32_t *coeff,
 495                                int firorder, int iirorder,
 496                                unsigned int filter_shift, int32_t mask, int blocksize,
 497                                int32_t *sample_buffer);
 498
 499     /* intrax8 functions */
 500     void (*x8_spatial_compensation[12])(uint8_t *src , uint8_t *dst, int linesize);
 501     void (*x8_setup_spatial_compensation)(uint8_t *src, uint8_t *dst, int linesize,
 502            int * range, int * sum,  int edges);
 503
 504     /**
 505      * Calculate scalar product of two vectors.
 506      * @param len length of vectors, should be multiple of 16
 507      */
 508     int32_t (*scalarproduct_int16)(const int16_t *v1, const int16_t *v2/*align 16*/, int len);
 509     /* ape functions */
 510     /**
 511      * Calculate scalar product of v1 and v2,
 512      * and v1[i] += v3[i] * mul
 513      * @param len length of vectors, should be multiple of 16
 514      */
 515     int32_t (*scalarproduct_and_madd_int16)(int16_t *v1/*align 16*/, const int16_t *v2, const int16_t *v3, int len, int mul);
 516
 517     /**
 518      * Apply symmetric window in 16-bit fixed-point.
 519      * @param output destination array
 520      *               constraints: 16-byte aligned
 521      * @param input  source array
 522      *               constraints: 16-byte aligned
 523      * @param window window array
 524      *               constraints: 16-byte aligned, at least len/2 elements
 525      * @param len    full window length
 526      *               constraints: multiple of ? greater than zero
 527      */
 528     void (*apply_window_int16)(int16_t *output, const int16_t *input,
 529                                const int16_t *window, unsigned int len);
 530
 531     /**
 532      * Clip each element in an array of int32_t to a given minimum and maximum value.
 533      * @param dst  destination array
 534      *             constraints: 16-byte aligned
 535      * @param src  source array
 536      *             constraints: 16-byte aligned
 537      * @param min  minimum value
 538      *             constraints: must be in the range [-(1 << 24), 1 << 24]
 539      * @param max  maximum value
 540      *             constraints: must be in the range [-(1 << 24), 1 << 24]
 541      * @param len  number of elements in the array
 542      *             constraints: multiple of 32 greater than zero
 543      */
 544     void (*vector_clip_int32)(int32_t *dst, const int32_t *src, int32_t min,
 545                               int32_t max, unsigned int len);
 546
 547     op_fill_func fill_block_tab[2];
 548 } DSPContext;
 549
 550 void ff_dsputil_static_init(void);
 551 void ff_dsputil_init(DSPContext* p, AVCodecContext *avctx);
 552
 553 int ff_check_alignment(void);
 554
 555 /**
 556  * permute block according to permuatation.
 557  * @param last last non zero element in scantable order
 558  */
 559 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
 560
 561 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
 562
 563 #define         BYTE_VEC32(c)   ((c)*0x01010101UL)
 564 #define         BYTE_VEC64(c)   ((c)*0x0001000100010001UL)
 565
 566 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
 567 {
 568     return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 569 }
 570
 571 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
 572 {
 573     return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 574 }
 575
 576 static inline uint64_t rnd_avg64(uint64_t a, uint64_t b)
 577 {
 578     return (a | b) - (((a ^ b) & ~BYTE_VEC64(0x01)) >> 1);
 579 }
 580
 581 static inline uint64_t no_rnd_avg64(uint64_t a, uint64_t b)
 582 {
 583     return (a & b) + (((a ^ b) & ~BYTE_VEC64(0x01)) >> 1);
 584 }
 585
 586 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 587     switch(type&0xFF){
 588     default:
 589     case FF_CMP_SAD:
 590         return lambda>>FF_LAMBDA_SHIFT;
 591     case FF_CMP_DCT:
 592         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 593     case FF_CMP_W53:
 594         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 595     case FF_CMP_W97:
 596         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 597     case FF_CMP_SATD:
 598     case FF_CMP_DCT264:
 599         return (2*lambda)>>FF_LAMBDA_SHIFT;
 600     case FF_CMP_RD:
 601     case FF_CMP_PSNR:
 602     case FF_CMP_SSE:
 603     case FF_CMP_NSSE:
 604         return lambda2>>FF_LAMBDA_SHIFT;
 605     case FF_CMP_BIT:
 606         return 1;
 607     }
 608 }
 609
 610 void ff_dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
 611 void ff_dsputil_init_arm(DSPContext* c, AVCodecContext *avctx);
 612 void ff_dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx);
 613 void ff_dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
 614 void ff_dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
 615 void ff_dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
 616 void ff_dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
 617 void ff_dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
 618
 619 void ff_dsputil_init_dwt(DSPContext *c);
 620 void ff_intrax8dsp_init(DSPContext* c, AVCodecContext *avctx);
 621 void ff_mlp_init(DSPContext* c, AVCodecContext *avctx);
 622 void ff_mlp_init_x86(DSPContext* c, AVCodecContext *avctx);
 623
 624 #if (ARCH_ARM && HAVE_NEON) || ARCH_PPC || HAVE_MMI || HAVE_MMX
 625 #   define STRIDE_ALIGN 16
 626 #else
 627 #   define STRIDE_ALIGN 8
 628 #endif
 629
 630 #define LOCAL_ALIGNED_A(a, t, v, s, o, ...)             \
 631     uint8_t la_##v[sizeof(t s o) + (a)];                \
 632     t (*v) o = (void *)FFALIGN((uintptr_t)la_##v, a)
 633
 634 #define LOCAL_ALIGNED_D(a, t, v, s, o, ...) DECLARE_ALIGNED(a, t, v) s o
 635
 636 #define LOCAL_ALIGNED(a, t, v, ...) LOCAL_ALIGNED_A(a, t, v, __VA_ARGS__,,)
 637
 638 #if HAVE_LOCAL_ALIGNED_8
 639 #   define LOCAL_ALIGNED_8(t, v, ...) LOCAL_ALIGNED_D(8, t, v, __VA_ARGS__,,)
 640 #else
 641 #   define LOCAL_ALIGNED_8(t, v, ...) LOCAL_ALIGNED(8, t, v, __VA_ARGS__)
 642 #endif
 643
 644 #if HAVE_LOCAL_ALIGNED_16
 645 #   define LOCAL_ALIGNED_16(t, v, ...) LOCAL_ALIGNED_D(16, t, v, __VA_ARGS__,,)
 646 #else
 647 #   define LOCAL_ALIGNED_16(t, v, ...) LOCAL_ALIGNED(16, t, v, __VA_ARGS__)
 648 #endif
 649
 650 #define WRAPPER8_16(name8, name16)\
 651 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 652     return name8(s, dst           , src           , stride, h)\
 653           +name8(s, dst+8         , src+8         , stride, h);\
 654 }
 655
 656 #define WRAPPER8_16_SQ(name8, name16)\
 657 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 658     int score=0;\
 659     score +=name8(s, dst           , src           , stride, 8);\
 660     score +=name8(s, dst+8         , src+8         , stride, 8);\
 661     if(h==16){\
 662         dst += 8*stride;\
 663         src += 8*stride;\
 664         score +=name8(s, dst           , src           , stride, 8);\
 665         score +=name8(s, dst+8         , src+8         , stride, 8);\
 666     }\
 667     return score;\
 668 }
 669
 670
 671 static inline void copy_block2(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 672 {
 673     int i;
 674     for(i=0; i<h; i++)
 675     {
 676         AV_WN16(dst   , AV_RN16(src   ));
 677         dst+=dstStride;
 678         src+=srcStride;
 679     }
 680 }
 681
 682 static inline void copy_block4(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 683 {
 684     int i;
 685     for(i=0; i<h; i++)
 686     {
 687         AV_WN32(dst   , AV_RN32(src   ));
 688         dst+=dstStride;
 689         src+=srcStride;
 690     }
 691 }
 692
 693 static inline void copy_block8(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 694 {
 695     int i;
 696     for(i=0; i<h; i++)
 697     {
 698         AV_WN32(dst   , AV_RN32(src   ));
 699         AV_WN32(dst+4 , AV_RN32(src+4 ));
 700         dst+=dstStride;
 701         src+=srcStride;
 702     }
 703 }
 704
 705 static inline void copy_block9(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 706 {
 707     int i;
 708     for(i=0; i<h; i++)
 709     {
 710         AV_WN32(dst   , AV_RN32(src   ));
 711         AV_WN32(dst+4 , AV_RN32(src+4 ));
 712         dst[8]= src[8];
 713         dst+=dstStride;
 714         src+=srcStride;
 715     }
 716 }
 717
 718 static inline void copy_block16(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 719 {
 720     int i;
 721     for(i=0; i<h; i++)
 722     {
 723         AV_WN32(dst   , AV_RN32(src   ));
 724         AV_WN32(dst+4 , AV_RN32(src+4 ));
 725         AV_WN32(dst+8 , AV_RN32(src+8 ));
 726         AV_WN32(dst+12, AV_RN32(src+12));
 727         dst+=dstStride;
 728         src+=srcStride;
 729     }
 730 }
 731
 732 static inline void copy_block17(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 733 {
 734     int i;
 735     for(i=0; i<h; i++)
 736     {
 737         AV_WN32(dst   , AV_RN32(src   ));
 738         AV_WN32(dst+4 , AV_RN32(src+4 ));
 739         AV_WN32(dst+8 , AV_RN32(src+8 ));
 740         AV_WN32(dst+12, AV_RN32(src+12));
 741         dst[16]= src[16];
 742         dst+=dstStride;
 743         src+=srcStride;
 744     }
 745 }
 746
 747 #endif /* AVCODEC_DSPUTIL_H */