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_add_add_c(float *dst, const float *src0, const float *src1,
  70                               const float *src2, int src3, int blocksize, int step);
  71 void ff_vector_fmul_window_c(float *dst, const float *src0, const float *src1,
  72                              const float *win, float add_bias, int len);
  73 void ff_float_to_int16_c(int16_t *dst, const float *src, long len);
  74 void ff_float_to_int16_interleave_c(int16_t *dst, const float **src, long len, int channels);
  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 /* VP3 DSP functions */
  90 void ff_vp3_idct_c(DCTELEM *block/* align 16*/);
  91 void ff_vp3_idct_put_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
  92 void ff_vp3_idct_add_c(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
  93
  94 void ff_vp3_v_loop_filter_c(uint8_t *src, int stride, int *bounding_values);
  95 void ff_vp3_h_loop_filter_c(uint8_t *src, int stride, int *bounding_values);
  96
  97 /* VP6 DSP functions */
  98 void ff_vp6_filter_diag4_c(uint8_t *dst, uint8_t *src, int stride,
  99                            const int16_t *h_weights, const int16_t *v_weights);
 100
 101 /* 1/2^n downscaling functions from imgconvert.c */
 102 void ff_img_copy_plane(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 103 void ff_shrink22(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 104 void ff_shrink44(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 105 void ff_shrink88(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 106
 107 void ff_gmc_c(uint8_t *dst, uint8_t *src, int stride, int h, int ox, int oy,
 108               int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 109
 110 /* minimum alignment rules ;)
 111 If you notice errors in the align stuff, need more alignment for some ASM code
 112 for some CPU or need to use a function with less aligned data then send a mail
 113 to the ffmpeg-devel mailing list, ...
 114
 115 !warning These alignments might not match reality, (missing attribute((align))
 116 stuff somewhere possible).
 117 I (Michael) did not check them, these are just the alignments which I think
 118 could be reached easily ...
 119
 120 !future video codecs might need functions with less strict alignment
 121 */
 122
 123 /*
 124 void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
 125 void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
 126 void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
 127 void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
 128 void clear_blocks_c(DCTELEM *blocks);
 129 */
 130
 131 /* add and put pixel (decoding) */
 132 // blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
 133 //h for op_pixels_func is limited to {width/2, width} but never larger than 16 and never smaller then 4
 134 typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
 135 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);
 136 typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 137 typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
 138 typedef void (*h264_weight_func)(uint8_t *block, int stride, int log2_denom, int weight, int offset);
 139 typedef void (*h264_biweight_func)(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset);
 140
 141 #define DEF_OLD_QPEL(name)\
 142 void ff_put_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 143 void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
 144 void ff_avg_        ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
 145
 146 DEF_OLD_QPEL(qpel16_mc11_old_c)
 147 DEF_OLD_QPEL(qpel16_mc31_old_c)
 148 DEF_OLD_QPEL(qpel16_mc12_old_c)
 149 DEF_OLD_QPEL(qpel16_mc32_old_c)
 150 DEF_OLD_QPEL(qpel16_mc13_old_c)
 151 DEF_OLD_QPEL(qpel16_mc33_old_c)
 152 DEF_OLD_QPEL(qpel8_mc11_old_c)
 153 DEF_OLD_QPEL(qpel8_mc31_old_c)
 154 DEF_OLD_QPEL(qpel8_mc12_old_c)
 155 DEF_OLD_QPEL(qpel8_mc32_old_c)
 156 DEF_OLD_QPEL(qpel8_mc13_old_c)
 157 DEF_OLD_QPEL(qpel8_mc33_old_c)
 158
 159 #define CALL_2X_PIXELS(a, b, n)\
 160 static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
 161     b(block  , pixels  , line_size, h);\
 162     b(block+n, pixels+n, line_size, h);\
 163 }
 164
 165 /* motion estimation */
 166 // h is limited to {width/2, width, 2*width} but never larger than 16 and never smaller then 2
 167 // although currently h<4 is not used as functions with width <8 are neither used nor implemented
 168 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))*/;
 169
 170
 171 // for snow slices
 172 typedef struct slice_buffer_s slice_buffer;
 173
 174 /**
 175  * Scantable.
 176  */
 177 typedef struct ScanTable{
 178     const uint8_t *scantable;
 179     uint8_t permutated[64];
 180     uint8_t raster_end[64];
 181 #if ARCH_PPC
 182                 /** Used by dct_quantize_altivec to find last-non-zero */
 183     DECLARE_ALIGNED(16, uint8_t, inverse[64]);
 184 #endif
 185 } ScanTable;
 186
 187 void ff_init_scantable(uint8_t *, ScanTable *st, const uint8_t *src_scantable);
 188
 189 void ff_emulated_edge_mc(uint8_t *buf, uint8_t *src, int linesize,
 190                          int block_w, int block_h,
 191                          int src_x, int src_y, int w, int h);
 192
 193 /**
 194  * DSPContext.
 195  */
 196 typedef struct DSPContext {
 197     /* pixel ops : interface with DCT */
 198     void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
 199     void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
 200     void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 201     void (*put_signed_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 202     void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
 203     void (*add_pixels8)(uint8_t *pixels, DCTELEM *block, int line_size);
 204     void (*add_pixels4)(uint8_t *pixels, DCTELEM *block, int line_size);
 205     int (*sum_abs_dctelem)(DCTELEM *block/*align 16*/);
 206     /**
 207      * translational global motion compensation.
 208      */
 209     void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
 210     /**
 211      * global motion compensation.
 212      */
 213     void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
 214                     int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
 215     void (*clear_block)(DCTELEM *block/*align 16*/);
 216     void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
 217     int (*pix_sum)(uint8_t * pix, int line_size);
 218     int (*pix_norm1)(uint8_t * pix, int line_size);
 219 // 16x16 8x8 4x4 2x2 16x8 8x4 4x2 8x16 4x8 2x4
 220
 221     me_cmp_func sad[6]; /* identical to pix_absAxA except additional void * */
 222     me_cmp_func sse[6];
 223     me_cmp_func hadamard8_diff[6];
 224     me_cmp_func dct_sad[6];
 225     me_cmp_func quant_psnr[6];
 226     me_cmp_func bit[6];
 227     me_cmp_func rd[6];
 228     me_cmp_func vsad[6];
 229     me_cmp_func vsse[6];
 230     me_cmp_func nsse[6];
 231     me_cmp_func w53[6];
 232     me_cmp_func w97[6];
 233     me_cmp_func dct_max[6];
 234     me_cmp_func dct264_sad[6];
 235
 236     me_cmp_func me_pre_cmp[6];
 237     me_cmp_func me_cmp[6];
 238     me_cmp_func me_sub_cmp[6];
 239     me_cmp_func mb_cmp[6];
 240     me_cmp_func ildct_cmp[6]; //only width 16 used
 241     me_cmp_func frame_skip_cmp[6]; //only width 8 used
 242
 243     int (*ssd_int8_vs_int16)(const int8_t *pix1, const int16_t *pix2,
 244                              int size);
 245
 246     /**
 247      * Halfpel motion compensation with rounding (a+b+1)>>1.
 248      * this is an array[4][4] of motion compensation functions for 4
 249      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 250      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 251      * @param block destination where the result is stored
 252      * @param pixels source
 253      * @param line_size number of bytes in a horizontal line of block
 254      * @param h height
 255      */
 256     op_pixels_func put_pixels_tab[4][4];
 257
 258     /**
 259      * Halfpel motion compensation with rounding (a+b+1)>>1.
 260      * This is an array[4][4] of motion compensation functions for 4
 261      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 262      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 263      * @param block destination into which the result is averaged (a+b+1)>>1
 264      * @param pixels source
 265      * @param line_size number of bytes in a horizontal line of block
 266      * @param h height
 267      */
 268     op_pixels_func avg_pixels_tab[4][4];
 269
 270     /**
 271      * Halfpel motion compensation with no rounding (a+b)>>1.
 272      * this is an array[2][4] of motion compensation functions for 2
 273      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 274      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 275      * @param block destination where the result is stored
 276      * @param pixels source
 277      * @param line_size number of bytes in a horizontal line of block
 278      * @param h height
 279      */
 280     op_pixels_func put_no_rnd_pixels_tab[4][4];
 281
 282     /**
 283      * Halfpel motion compensation with no rounding (a+b)>>1.
 284      * this is an array[2][4] of motion compensation functions for 2
 285      * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
 286      * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
 287      * @param block destination into which the result is averaged (a+b)>>1
 288      * @param pixels source
 289      * @param line_size number of bytes in a horizontal line of block
 290      * @param h height
 291      */
 292     op_pixels_func avg_no_rnd_pixels_tab[4][4];
 293
 294     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);
 295
 296     /**
 297      * Thirdpel motion compensation with rounding (a+b+1)>>1.
 298      * this is an array[12] of motion compensation functions for the 9 thirdpe
 299      * positions<br>
 300      * *pixels_tab[ xthirdpel + 4*ythirdpel ]
 301      * @param block destination where the result is stored
 302      * @param pixels source
 303      * @param line_size number of bytes in a horizontal line of block
 304      * @param h height
 305      */
 306     tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 307     tpel_mc_func avg_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
 308
 309     qpel_mc_func put_qpel_pixels_tab[2][16];
 310     qpel_mc_func avg_qpel_pixels_tab[2][16];
 311     qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
 312     qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
 313     qpel_mc_func put_mspel_pixels_tab[8];
 314
 315     /**
 316      * h264 Chroma MC
 317      */
 318     h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
 319     h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
 320     /* This is really one func used in VC-1 decoding */
 321     h264_chroma_mc_func put_no_rnd_vc1_chroma_pixels_tab[3];
 322     h264_chroma_mc_func avg_no_rnd_vc1_chroma_pixels_tab[3];
 323
 324     qpel_mc_func put_h264_qpel_pixels_tab[4][16];
 325     qpel_mc_func avg_h264_qpel_pixels_tab[4][16];
 326
 327     qpel_mc_func put_2tap_qpel_pixels_tab[4][16];
 328     qpel_mc_func avg_2tap_qpel_pixels_tab[4][16];
 329
 330     h264_weight_func weight_h264_pixels_tab[10];
 331     h264_biweight_func biweight_h264_pixels_tab[10];
 332
 333     /* AVS specific */
 334     qpel_mc_func put_cavs_qpel_pixels_tab[2][16];
 335     qpel_mc_func avg_cavs_qpel_pixels_tab[2][16];
 336     void (*cavs_filter_lv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 337     void (*cavs_filter_lh)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 338     void (*cavs_filter_cv)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 339     void (*cavs_filter_ch)(uint8_t *pix, int stride, int alpha, int beta, int tc, int bs1, int bs2);
 340     void (*cavs_idct8_add)(uint8_t *dst, DCTELEM *block, int stride);
 341
 342     me_cmp_func pix_abs[2][4];
 343
 344     /* huffyuv specific */
 345     void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
 346     void (*add_bytes_l2)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 16*/, int w);
 347     void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
 348     /**
 349      * subtract huffyuv's variant of median prediction
 350      * note, this might read from src1[-1], src2[-1]
 351      */
 352     void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
 353     void (*add_hfyu_median_prediction)(uint8_t *dst, uint8_t *top, uint8_t *diff, int w, int *left, int *left_top);
 354     /* this might write to dst[w] */
 355     void (*add_png_paeth_prediction)(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
 356     void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
 357
 358     void (*h264_v_loop_filter_luma)(uint8_t *pix/*align 16*/, int stride, int alpha, int beta, int8_t *tc0);
 359     void (*h264_h_loop_filter_luma)(uint8_t *pix/*align 4 */, int stride, int alpha, int beta, int8_t *tc0);
 360     /* v/h_loop_filter_luma_intra: align 16 */
 361     void (*h264_v_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
 362     void (*h264_h_loop_filter_luma_intra)(uint8_t *pix, int stride, int alpha, int beta);
 363     void (*h264_v_loop_filter_chroma)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta, int8_t *tc0);
 364     void (*h264_h_loop_filter_chroma)(uint8_t *pix/*align 4*/, int stride, int alpha, int beta, int8_t *tc0);
 365     void (*h264_v_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
 366     void (*h264_h_loop_filter_chroma_intra)(uint8_t *pix/*align 8*/, int stride, int alpha, int beta);
 367     // h264_loop_filter_strength: simd only. the C version is inlined in h264.c
 368     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],
 369                                       int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field);
 370
 371     void (*h263_v_loop_filter)(uint8_t *src, int stride, int qscale);
 372     void (*h263_h_loop_filter)(uint8_t *src, int stride, int qscale);
 373
 374     void (*h261_loop_filter)(uint8_t *src, int stride);
 375
 376     void (*x8_v_loop_filter)(uint8_t *src, int stride, int qscale);
 377     void (*x8_h_loop_filter)(uint8_t *src, int stride, int qscale);
 378
 379     void (*vp3_v_loop_filter)(uint8_t *src, int stride, int *bounding_values);
 380     void (*vp3_h_loop_filter)(uint8_t *src, int stride, int *bounding_values);
 381
 382     void (*vp6_filter_diag4)(uint8_t *dst, uint8_t *src, int stride,
 383                              const int16_t *h_weights,const int16_t *v_weights);
 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     /* no alignment needed */
 389     void (*flac_compute_autocorr)(const int32_t *data, int len, int lag, double *autoc);
 390     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
 391     void (*vector_fmul)(float *dst, const float *src, int len);
 392     void (*vector_fmul_reverse)(float *dst, const float *src0, const float *src1, int len);
 393     /* assume len is a multiple of 8, and src arrays are 16-byte aligned */
 394     void (*vector_fmul_add_add)(float *dst, const float *src0, const float *src1, const float *src2, int src3, int len, int step);
 395     /* assume len is a multiple of 4, and arrays are 16-byte aligned */
 396     void (*vector_fmul_window)(float *dst, const float *src0, const float *src1, const float *win, float add_bias, int len);
 397     /* assume len is a multiple of 8, and arrays are 16-byte aligned */
 398     void (*int32_to_float_fmul_scalar)(float *dst, const int *src, float mul, int len);
 399
 400     /* C version: convert floats from the range [384.0,386.0] to ints in [-32768,32767]
 401      * simd versions: convert floats from [-32768.0,32767.0] without rescaling and arrays are 16byte aligned */
 402     void (*float_to_int16)(int16_t *dst, const float *src, long len);
 403     void (*float_to_int16_interleave)(int16_t *dst, const float **src, long len, int channels);
 404
 405     /* (I)DCT */
 406     void (*fdct)(DCTELEM *block/* align 16*/);
 407     void (*fdct248)(DCTELEM *block/* align 16*/);
 408
 409     /* IDCT really*/
 410     void (*idct)(DCTELEM *block/* align 16*/);
 411
 412     /**
 413      * block -> idct -> clip to unsigned 8 bit -> dest.
 414      * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
 415      * @param line_size size in bytes of a horizontal line of dest
 416      */
 417     void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 418
 419     /**
 420      * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
 421      * @param line_size size in bytes of a horizontal line of dest
 422      */
 423     void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
 424
 425     /**
 426      * idct input permutation.
 427      * several optimized IDCTs need a permutated input (relative to the normal order of the reference
 428      * IDCT)
 429      * this permutation must be performed before the idct_put/add, note, normally this can be merged
 430      * with the zigzag/alternate scan<br>
 431      * an example to avoid confusion:
 432      * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
 433      * - (x -> referece dct -> reference idct -> x)
 434      * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
 435      * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
 436      */
 437     uint8_t idct_permutation[64];
 438     int idct_permutation_type;
 439 #define FF_NO_IDCT_PERM 1
 440 #define FF_LIBMPEG2_IDCT_PERM 2
 441 #define FF_SIMPLE_IDCT_PERM 3
 442 #define FF_TRANSPOSE_IDCT_PERM 4
 443 #define FF_PARTTRANS_IDCT_PERM 5
 444 #define FF_SSE2_IDCT_PERM 6
 445
 446     int (*try_8x8basis)(int16_t rem[64], int16_t weight[64], int16_t basis[64], int scale);
 447     void (*add_8x8basis)(int16_t rem[64], int16_t basis[64], int scale);
 448 #define BASIS_SHIFT 16
 449 #define RECON_SHIFT 6
 450
 451     void (*draw_edges)(uint8_t *buf, int wrap, int width, int height, int w);
 452 #define EDGE_WIDTH 16
 453
 454     /* h264 functions */
 455     /* NOTE!!! if you implement any of h264_idct8_add, h264_idct8_add4 then you must implement all of them
 456        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
 457         The reason for above, is that no 2 out of one list may use a different permutation.
 458     */
 459     void (*h264_idct_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
 460     void (*h264_idct8_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
 461     void (*h264_idct_dc_add)(uint8_t *dst/*align 4*/, DCTELEM *block/*align 16*/, int stride);
 462     void (*h264_idct8_dc_add)(uint8_t *dst/*align 8*/, DCTELEM *block/*align 16*/, int stride);
 463     void (*h264_dct)(DCTELEM block[4][4]);
 464     void (*h264_idct_add16)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 465     void (*h264_idct8_add4)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 466     void (*h264_idct_add8)(uint8_t **dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 467     void (*h264_idct_add16intra)(uint8_t *dst/*align 16*/, const int *blockoffset, DCTELEM *block/*align 16*/, int stride, const uint8_t nnzc[6*8]);
 468
 469     /* snow wavelet */
 470     void (*vertical_compose97i)(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5, int width);
 471     void (*horizontal_compose97i)(IDWTELEM *b, int width);
 472     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);
 473
 474     void (*prefetch)(void *mem, int stride, int h);
 475
 476     void (*shrink[4])(uint8_t *dst, int dst_wrap, const uint8_t *src, int src_wrap, int width, int height);
 477
 478     /* mlp/truehd functions */
 479     void (*mlp_filter_channel)(int32_t *state, const int32_t *coeff,
 480                                int firorder, int iirorder,
 481                                unsigned int filter_shift, int32_t mask, int blocksize,
 482                                int32_t *sample_buffer);
 483
 484     /* vc1 functions */
 485     void (*vc1_inv_trans_8x8)(DCTELEM *b);
 486     void (*vc1_inv_trans_8x4)(uint8_t *dest, int line_size, DCTELEM *block);
 487     void (*vc1_inv_trans_4x8)(uint8_t *dest, int line_size, DCTELEM *block);
 488     void (*vc1_inv_trans_4x4)(uint8_t *dest, int line_size, DCTELEM *block);
 489     void (*vc1_inv_trans_8x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 490     void (*vc1_inv_trans_8x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 491     void (*vc1_inv_trans_4x8_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 492     void (*vc1_inv_trans_4x4_dc)(uint8_t *dest, int line_size, DCTELEM *block);
 493     void (*vc1_v_overlap)(uint8_t* src, int stride);
 494     void (*vc1_h_overlap)(uint8_t* src, int stride);
 495     void (*vc1_v_loop_filter4)(uint8_t *src, int stride, int pq);
 496     void (*vc1_h_loop_filter4)(uint8_t *src, int stride, int pq);
 497     void (*vc1_v_loop_filter8)(uint8_t *src, int stride, int pq);
 498     void (*vc1_h_loop_filter8)(uint8_t *src, int stride, int pq);
 499     void (*vc1_v_loop_filter16)(uint8_t *src, int stride, int pq);
 500     void (*vc1_h_loop_filter16)(uint8_t *src, int stride, int pq);
 501     /* put 8x8 block with bicubic interpolation and quarterpel precision
 502      * last argument is actually round value instead of height
 503      */
 504     op_pixels_func put_vc1_mspel_pixels_tab[16];
 505     op_pixels_func avg_vc1_mspel_pixels_tab[16];
 506
 507     /* intrax8 functions */
 508     void (*x8_spatial_compensation[12])(uint8_t *src , uint8_t *dst, int linesize);
 509     void (*x8_setup_spatial_compensation)(uint8_t *src, uint8_t *dst, int linesize,
 510            int * range, int * sum,  int edges);
 511
 512     /* ape functions */
 513     /**
 514      * Add contents of the second vector to the first one.
 515      * @param len length of vectors, should be multiple of 16
 516      */
 517     void (*add_int16)(int16_t *v1/*align 16*/, int16_t *v2, int len);
 518     /**
 519      * Add contents of the second vector to the first one.
 520      * @param len length of vectors, should be multiple of 16
 521      */
 522     void (*sub_int16)(int16_t *v1/*align 16*/, int16_t *v2, int len);
 523     /**
 524      * Calculate scalar product of two vectors.
 525      * @param len length of vectors, should be multiple of 16
 526      * @param shift number of bits to discard from product
 527      */
 528     int32_t (*scalarproduct_int16)(int16_t *v1, int16_t *v2/*align 16*/, int len, int shift);
 529
 530     /* rv30 functions */
 531     qpel_mc_func put_rv30_tpel_pixels_tab[4][16];
 532     qpel_mc_func avg_rv30_tpel_pixels_tab[4][16];
 533
 534     /* rv40 functions */
 535     qpel_mc_func put_rv40_qpel_pixels_tab[4][16];
 536     qpel_mc_func avg_rv40_qpel_pixels_tab[4][16];
 537     h264_chroma_mc_func put_rv40_chroma_pixels_tab[3];
 538     h264_chroma_mc_func avg_rv40_chroma_pixels_tab[3];
 539 } DSPContext;
 540
 541 void dsputil_static_init(void);
 542 void dsputil_init(DSPContext* p, AVCodecContext *avctx);
 543
 544 int ff_check_alignment(void);
 545
 546 /**
 547  * permute block according to permuatation.
 548  * @param last last non zero element in scantable order
 549  */
 550 void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
 551
 552 void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type);
 553
 554 #define         BYTE_VEC32(c)   ((c)*0x01010101UL)
 555
 556 static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
 557 {
 558     return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 559 }
 560
 561 static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
 562 {
 563     return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
 564 }
 565
 566 static inline int get_penalty_factor(int lambda, int lambda2, int type){
 567     switch(type&0xFF){
 568     default:
 569     case FF_CMP_SAD:
 570         return lambda>>FF_LAMBDA_SHIFT;
 571     case FF_CMP_DCT:
 572         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
 573     case FF_CMP_W53:
 574         return (4*lambda)>>(FF_LAMBDA_SHIFT);
 575     case FF_CMP_W97:
 576         return (2*lambda)>>(FF_LAMBDA_SHIFT);
 577     case FF_CMP_SATD:
 578     case FF_CMP_DCT264:
 579         return (2*lambda)>>FF_LAMBDA_SHIFT;
 580     case FF_CMP_RD:
 581     case FF_CMP_PSNR:
 582     case FF_CMP_SSE:
 583     case FF_CMP_NSSE:
 584         return lambda2>>FF_LAMBDA_SHIFT;
 585     case FF_CMP_BIT:
 586         return 1;
 587     }
 588 }
 589
 590 /**
 591  * Empty mmx state.
 592  * this must be called between any dsp function and float/double code.
 593  * for example sin(); dsp->idct_put(); emms_c(); cos()
 594  */
 595 #define emms_c()
 596
 597 /* should be defined by architectures supporting
 598    one or more MultiMedia extension */
 599 int mm_support(void);
 600
 601 void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
 602 void dsputil_init_arm(DSPContext* c, AVCodecContext *avctx);
 603 void dsputil_init_bfin(DSPContext* c, AVCodecContext *avctx);
 604 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
 605 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
 606 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
 607 void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
 608 void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
 609 void dsputil_init_vis(DSPContext* c, AVCodecContext *avctx);
 610
 611 #define DECLARE_ALIGNED_16(t, v) DECLARE_ALIGNED(16, t, v)
 612
 613 #if HAVE_MMX
 614
 615 #undef emms_c
 616
 617 extern int mm_flags;
 618
 619 void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
 620 void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
 621 void put_signed_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
 622
 623 static inline void emms(void)
 624 {
 625     __asm__ volatile ("emms;":::"memory");
 626 }
 627
 628
 629 #define emms_c() \
 630 {\
 631     if (mm_flags & FF_MM_MMX)\
 632         emms();\
 633 }
 634
 635 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
 636
 637 #elif ARCH_ARM
 638
 639 extern int mm_flags;
 640
 641 #if HAVE_NEON
 642 #   define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v)
 643 #   define STRIDE_ALIGN 16
 644 #endif
 645
 646 #elif ARCH_PPC
 647
 648 extern int mm_flags;
 649
 650 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v)
 651 #define STRIDE_ALIGN 16
 652
 653 #elif HAVE_MMI
 654
 655 #define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(16, t, v)
 656 #define STRIDE_ALIGN 16
 657
 658 #else
 659
 660 #define mm_flags 0
 661 #define mm_support() 0
 662
 663 #endif
 664
 665 #ifndef DECLARE_ALIGNED_8
 666 #   define DECLARE_ALIGNED_8(t, v) DECLARE_ALIGNED(8, t, v)
 667 #endif
 668
 669 #ifndef STRIDE_ALIGN
 670 #   define STRIDE_ALIGN 8
 671 #endif
 672
 673 /* PSNR */
 674 void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
 675               int orig_linesize[3], int coded_linesize,
 676               AVCodecContext *avctx);
 677
 678 /* FFT computation */
 679
 680 /* NOTE: soon integer code will be added, so you must use the
 681    FFTSample type */
 682 typedef float FFTSample;
 683
 684 struct MDCTContext;
 685
 686 typedef struct FFTComplex {
 687     FFTSample re, im;
 688 } FFTComplex;
 689
 690 typedef struct FFTContext {
 691     int nbits;
 692     int inverse;
 693     uint16_t *revtab;
 694     FFTComplex *exptab;
 695     FFTComplex *exptab1; /* only used by SSE code */
 696     FFTComplex *tmp_buf;
 697     void (*fft_permute)(struct FFTContext *s, FFTComplex *z);
 698     void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
 699     void (*imdct_calc)(struct MDCTContext *s, FFTSample *output, const FFTSample *input);
 700     void (*imdct_half)(struct MDCTContext *s, FFTSample *output, const FFTSample *input);
 701 } FFTContext;
 702
 703 extern FFTSample* ff_cos_tabs[13];
 704
 705 /**
 706  * Sets up a complex FFT.
 707  * @param nbits           log2 of the length of the input array
 708  * @param inverse         if 0 perform the forward transform, if 1 perform the inverse
 709  */
 710 int ff_fft_init(FFTContext *s, int nbits, int inverse);
 711 void ff_fft_permute_c(FFTContext *s, FFTComplex *z);
 712 void ff_fft_permute_sse(FFTContext *s, FFTComplex *z);
 713 void ff_fft_calc_c(FFTContext *s, FFTComplex *z);
 714 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z);
 715 void ff_fft_calc_3dn(FFTContext *s, FFTComplex *z);
 716 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z);
 717 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z);
 718
 719 /**
 720  * Do the permutation needed BEFORE calling ff_fft_calc().
 721  */
 722 static inline void ff_fft_permute(FFTContext *s, FFTComplex *z)
 723 {
 724     s->fft_permute(s, z);
 725 }
 726 /**
 727  * Do a complex FFT with the parameters defined in ff_fft_init(). The
 728  * input data must be permuted before. No 1.0/sqrt(n) normalization is done.
 729  */
 730 static inline void ff_fft_calc(FFTContext *s, FFTComplex *z)
 731 {
 732     s->fft_calc(s, z);
 733 }
 734 void ff_fft_end(FFTContext *s);
 735
 736 /* MDCT computation */
 737
 738 typedef struct MDCTContext {
 739     int n;  /* size of MDCT (i.e. number of input data * 2) */
 740     int nbits; /* n = 2^nbits */
 741     /* pre/post rotation tables */
 742     FFTSample *tcos;
 743     FFTSample *tsin;
 744     FFTContext fft;
 745 } MDCTContext;
 746
 747 static inline void ff_imdct_calc(MDCTContext *s, FFTSample *output, const FFTSample *input)
 748 {
 749     s->fft.imdct_calc(s, output, input);
 750 }
 751 static inline void ff_imdct_half(MDCTContext *s, FFTSample *output, const FFTSample *input)
 752 {
 753     s->fft.imdct_half(s, output, input);
 754 }
 755
 756 /**
 757  * Generate a Kaiser-Bessel Derived Window.
 758  * @param   window  pointer to half window
 759  * @param   alpha   determines window shape
 760  * @param   n       size of half window
 761  */
 762 void ff_kbd_window_init(float *window, float alpha, int n);
 763
 764 /**
 765  * Generate a sine window.
 766  * @param   window  pointer to half window
 767  * @param   n       size of half window
 768  */
 769 void ff_sine_window_init(float *window, int n);
 770 extern float ff_sine_128 [ 128];
 771 extern float ff_sine_256 [ 256];
 772 extern float ff_sine_512 [ 512];
 773 extern float ff_sine_1024[1024];
 774 extern float ff_sine_2048[2048];
 775 extern float ff_sine_4096[4096];
 776 extern float *ff_sine_windows[6];
 777
 778 int ff_mdct_init(MDCTContext *s, int nbits, int inverse, double scale);
 779 void ff_imdct_calc_c(MDCTContext *s, FFTSample *output, const FFTSample *input);
 780 void ff_imdct_half_c(MDCTContext *s, FFTSample *output, const FFTSample *input);
 781 void ff_imdct_calc_3dn(MDCTContext *s, FFTSample *output, const FFTSample *input);
 782 void ff_imdct_half_3dn(MDCTContext *s, FFTSample *output, const FFTSample *input);
 783 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output, const FFTSample *input);
 784 void ff_imdct_half_3dn2(MDCTContext *s, FFTSample *output, const FFTSample *input);
 785 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, const FFTSample *input);
 786 void ff_imdct_half_sse(MDCTContext *s, FFTSample *output, const FFTSample *input);
 787 void ff_mdct_calc(MDCTContext *s, FFTSample *out, const FFTSample *input);
 788 void ff_mdct_end(MDCTContext *s);
 789
 790 /* Real Discrete Fourier Transform */
 791
 792 enum RDFTransformType {
 793     RDFT,
 794     IRDFT,
 795     RIDFT,
 796     IRIDFT,
 797 };
 798
 799 typedef struct {
 800     int nbits;
 801     int inverse;
 802     int sign_convention;
 803
 804     /* pre/post rotation tables */
 805     FFTSample *tcos;
 806     FFTSample *tsin;
 807     FFTContext fft;
 808 } RDFTContext;
 809
 810 /**
 811  * Sets up a real FFT.
 812  * @param nbits           log2 of the length of the input array
 813  * @param trans           the type of transform
 814  */
 815 int ff_rdft_init(RDFTContext *s, int nbits, enum RDFTransformType trans);
 816 void ff_rdft_calc(RDFTContext *s, FFTSample *data);
 817 void ff_rdft_end(RDFTContext *s);
 818
 819 #define WRAPPER8_16(name8, name16)\
 820 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 821     return name8(s, dst           , src           , stride, h)\
 822           +name8(s, dst+8         , src+8         , stride, h);\
 823 }
 824
 825 #define WRAPPER8_16_SQ(name8, name16)\
 826 static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride, int h){\
 827     int score=0;\
 828     score +=name8(s, dst           , src           , stride, 8);\
 829     score +=name8(s, dst+8         , src+8         , stride, 8);\
 830     if(h==16){\
 831         dst += 8*stride;\
 832         src += 8*stride;\
 833         score +=name8(s, dst           , src           , stride, 8);\
 834         score +=name8(s, dst+8         , src+8         , stride, 8);\
 835     }\
 836     return score;\
 837 }
 838
 839
 840 static inline void copy_block2(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 841 {
 842     int i;
 843     for(i=0; i<h; i++)
 844     {
 845         AV_WN16(dst   , AV_RN16(src   ));
 846         dst+=dstStride;
 847         src+=srcStride;
 848     }
 849 }
 850
 851 static inline void copy_block4(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 852 {
 853     int i;
 854     for(i=0; i<h; i++)
 855     {
 856         AV_WN32(dst   , AV_RN32(src   ));
 857         dst+=dstStride;
 858         src+=srcStride;
 859     }
 860 }
 861
 862 static inline void copy_block8(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 863 {
 864     int i;
 865     for(i=0; i<h; i++)
 866     {
 867         AV_WN32(dst   , AV_RN32(src   ));
 868         AV_WN32(dst+4 , AV_RN32(src+4 ));
 869         dst+=dstStride;
 870         src+=srcStride;
 871     }
 872 }
 873
 874 static inline void copy_block9(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 875 {
 876     int i;
 877     for(i=0; i<h; i++)
 878     {
 879         AV_WN32(dst   , AV_RN32(src   ));
 880         AV_WN32(dst+4 , AV_RN32(src+4 ));
 881         dst[8]= src[8];
 882         dst+=dstStride;
 883         src+=srcStride;
 884     }
 885 }
 886
 887 static inline void copy_block16(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 888 {
 889     int i;
 890     for(i=0; i<h; i++)
 891     {
 892         AV_WN32(dst   , AV_RN32(src   ));
 893         AV_WN32(dst+4 , AV_RN32(src+4 ));
 894         AV_WN32(dst+8 , AV_RN32(src+8 ));
 895         AV_WN32(dst+12, AV_RN32(src+12));
 896         dst+=dstStride;
 897         src+=srcStride;
 898     }
 899 }
 900
 901 static inline void copy_block17(uint8_t *dst, const uint8_t *src, int dstStride, int srcStride, int h)
 902 {
 903     int i;
 904     for(i=0; i<h; i++)
 905     {
 906         AV_WN32(dst   , AV_RN32(src   ));
 907         AV_WN32(dst+4 , AV_RN32(src+4 ));
 908         AV_WN32(dst+8 , AV_RN32(src+8 ));
 909         AV_WN32(dst+12, AV_RN32(src+12));
 910         dst[16]= src[16];
 911         dst+=dstStride;
 912         src+=srcStride;
 913     }
 914 }
 915
 916 #endif /* AVCODEC_DSPUTIL_H */