FFmpeg: libavcodec/mlib/dsputil

00001 /*
00002  * Sun mediaLib optimized DSP utils
00003  * Copyright (c) 2001 Fabrice Bellard
00004  *
00005  * This file is part of FFmpeg.
00006  *
00007  * FFmpeg is free software; you can redistribute it and/or
00008  * modify it under the terms of the GNU Lesser General Public
00009  * License as published by the Free Software Foundation; either
00010  * version 2.1 of the License, or (at your option) any later version.
00011  *
00012  * FFmpeg is distributed in the hope that it will be useful,
00013  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00014  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00015  * Lesser General Public License for more details.
00016  *
00017  * You should have received a copy of the GNU Lesser General Public
00018  * License along with FFmpeg; if not, write to the Free Software
00019  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00020  */
00021 
00022 #include "libavcodec/dsputil.h"
00023 #include "libavcodec/mpegvideo.h"
00024 
00025 #include <mlib_types.h>
00026 #include <mlib_status.h>
00027 #include <mlib_sys.h>
00028 #include <mlib_algebra.h>
00029 #include <mlib_video.h>
00030 
00031 /* misc */
00032 
00033 static void get_pixels_mlib(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
00034 {
00035   int i;
00036 
00037   for (i=0;i<8;i++) {
00038     mlib_VectorConvert_S16_U8_Mod((mlib_s16 *)block, (mlib_u8 *)pixels, 8);
00039 
00040     pixels += line_size;
00041     block += 8;
00042   }
00043 }
00044 
00045 static void diff_pixels_mlib(DCTELEM *restrict block, const uint8_t *s1, const uint8_t *s2, int line_size)
00046 {
00047   int i;
00048 
00049   for (i=0;i<8;i++) {
00050     mlib_VectorSub_S16_U8_Mod((mlib_s16 *)block, (mlib_u8 *)s1, (mlib_u8 *)s2, 8);
00051 
00052     s1 += line_size;
00053     s2 += line_size;
00054     block += 8;
00055   }
00056 }
00057 
00058 static void add_pixels_clamped_mlib(const DCTELEM *block, uint8_t *pixels, int line_size)
00059 {
00060     mlib_VideoAddBlock_U8_S16(pixels, (mlib_s16 *)block, line_size);
00061 }
00062 
00063 /* put block, width 16 pixel, height 8/16 */
00064 
00065 static void put_pixels16_mlib (uint8_t * dest, const uint8_t * ref,
00066                                int stride, int height)
00067 {
00068   switch (height) {
00069     case 8:
00070       mlib_VideoCopyRef_U8_U8_16x8(dest, (uint8_t *)ref, stride);
00071     break;
00072 
00073     case 16:
00074       mlib_VideoCopyRef_U8_U8_16x16(dest, (uint8_t *)ref, stride);
00075     break;
00076 
00077     default:
00078       assert(0);
00079   }
00080 }
00081 
00082 static void put_pixels16_x2_mlib (uint8_t * dest, const uint8_t * ref,
00083                                   int stride, int height)
00084 {
00085   switch (height) {
00086     case 8:
00087       mlib_VideoInterpX_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
00088     break;
00089 
00090     case 16:
00091       mlib_VideoInterpX_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
00092     break;
00093 
00094     default:
00095       assert(0);
00096   }
00097 }
00098 
00099 static void put_pixels16_y2_mlib (uint8_t * dest, const uint8_t * ref,
00100                                   int stride, int height)
00101 {
00102   switch (height) {
00103     case 8:
00104       mlib_VideoInterpY_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
00105     break;
00106 
00107     case 16:
00108       mlib_VideoInterpY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
00109     break;
00110 
00111     default:
00112       assert(0);
00113   }
00114 }
00115 
00116 static void put_pixels16_xy2_mlib(uint8_t * dest, const uint8_t * ref,
00117                                   int stride, int height)
00118 {
00119   switch (height) {
00120     case 8:
00121       mlib_VideoInterpXY_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
00122     break;
00123 
00124     case 16:
00125       mlib_VideoInterpXY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
00126     break;
00127 
00128     default:
00129       assert(0);
00130   }
00131 }
00132 
00133 /* put block, width 8 pixel, height 4/8/16 */
00134 
00135 static void put_pixels8_mlib (uint8_t * dest, const uint8_t * ref,
00136                                int stride, int height)
00137 {
00138   switch (height) {
00139     case 4:
00140       mlib_VideoCopyRef_U8_U8_8x4(dest, (uint8_t *)ref, stride);
00141     break;
00142 
00143     case 8:
00144       mlib_VideoCopyRef_U8_U8_8x8(dest, (uint8_t *)ref, stride);
00145     break;
00146 
00147     case 16:
00148       mlib_VideoCopyRef_U8_U8_8x16(dest, (uint8_t *)ref, stride);
00149     break;
00150 
00151     default:
00152       assert(0);
00153   }
00154 }
00155 
00156 static void put_pixels8_x2_mlib (uint8_t * dest, const uint8_t * ref,
00157                                   int stride, int height)
00158 {
00159   switch (height) {
00160     case 4:
00161       mlib_VideoInterpX_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
00162     break;
00163 
00164     case 8:
00165       mlib_VideoInterpX_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
00166     break;
00167 
00168     case 16:
00169       mlib_VideoInterpX_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
00170     break;
00171 
00172     default:
00173       assert(0);
00174   }
00175 }
00176 
00177 static void put_pixels8_y2_mlib (uint8_t * dest, const uint8_t * ref,
00178                                   int stride, int height)
00179 {
00180   switch (height) {
00181     case 4:
00182       mlib_VideoInterpY_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
00183     break;
00184 
00185     case 8:
00186       mlib_VideoInterpY_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
00187     break;
00188 
00189     case 16:
00190       mlib_VideoInterpY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
00191     break;
00192 
00193     default:
00194       assert(0);
00195   }
00196 }
00197 
00198 static void put_pixels8_xy2_mlib(uint8_t * dest, const uint8_t * ref,
00199                                   int stride, int height)
00200 {
00201   switch (height) {
00202     case 4:
00203       mlib_VideoInterpXY_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
00204     break;
00205 
00206     case 8:
00207       mlib_VideoInterpXY_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
00208     break;
00209 
00210     case 16:
00211       mlib_VideoInterpXY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
00212     break;
00213 
00214     default:
00215       assert(0);
00216   }
00217 }
00218 
00219 /* average block, width 16 pixel, height 8/16 */
00220 
00221 static void avg_pixels16_mlib (uint8_t * dest, const uint8_t * ref,
00222                                int stride, int height)
00223 {
00224   switch (height) {
00225     case 8:
00226       mlib_VideoCopyRefAve_U8_U8_16x8(dest, (uint8_t *)ref, stride);
00227     break;
00228 
00229     case 16:
00230       mlib_VideoCopyRefAve_U8_U8_16x16(dest, (uint8_t *)ref, stride);
00231     break;
00232 
00233     default:
00234       assert(0);
00235   }
00236 }
00237 
00238 static void avg_pixels16_x2_mlib (uint8_t * dest, const uint8_t * ref,
00239                                   int stride, int height)
00240 {
00241   switch (height) {
00242     case 8:
00243       mlib_VideoInterpAveX_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
00244     break;
00245 
00246     case 16:
00247       mlib_VideoInterpAveX_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
00248     break;
00249 
00250     default:
00251       assert(0);
00252   }
00253 }
00254 
00255 static void avg_pixels16_y2_mlib (uint8_t * dest, const uint8_t * ref,
00256                                   int stride, int height)
00257 {
00258   switch (height) {
00259     case 8:
00260       mlib_VideoInterpAveY_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
00261     break;
00262 
00263     case 16:
00264       mlib_VideoInterpAveY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
00265     break;
00266 
00267     default:
00268       assert(0);
00269   }
00270 }
00271 
00272 static void avg_pixels16_xy2_mlib(uint8_t * dest, const uint8_t * ref,
00273                                   int stride, int height)
00274 {
00275   switch (height) {
00276     case 8:
00277       mlib_VideoInterpAveXY_U8_U8_16x8(dest, (uint8_t *)ref, stride, stride);
00278     break;
00279 
00280     case 16:
00281       mlib_VideoInterpAveXY_U8_U8_16x16(dest, (uint8_t *)ref, stride, stride);
00282     break;
00283 
00284     default:
00285       assert(0);
00286   }
00287 }
00288 
00289 /* average block, width 8 pixel, height 4/8/16 */
00290 
00291 static void avg_pixels8_mlib (uint8_t * dest, const uint8_t * ref,
00292                                int stride, int height)
00293 {
00294   switch (height) {
00295     case 4:
00296       mlib_VideoCopyRefAve_U8_U8_8x4(dest, (uint8_t *)ref, stride);
00297     break;
00298 
00299     case 8:
00300       mlib_VideoCopyRefAve_U8_U8_8x8(dest, (uint8_t *)ref, stride);
00301     break;
00302 
00303     case 16:
00304       mlib_VideoCopyRefAve_U8_U8_8x16(dest, (uint8_t *)ref, stride);
00305     break;
00306 
00307     default:
00308       assert(0);
00309   }
00310 }
00311 
00312 static void avg_pixels8_x2_mlib (uint8_t * dest, const uint8_t * ref,
00313                                   int stride, int height)
00314 {
00315   switch (height) {
00316     case 4:
00317       mlib_VideoInterpAveX_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
00318     break;
00319 
00320     case 8:
00321       mlib_VideoInterpAveX_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
00322     break;
00323 
00324     case 16:
00325       mlib_VideoInterpAveX_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
00326     break;
00327 
00328     default:
00329       assert(0);
00330   }
00331 }
00332 
00333 static void avg_pixels8_y2_mlib (uint8_t * dest, const uint8_t * ref,
00334                                   int stride, int height)
00335 {
00336   switch (height) {
00337     case 4:
00338       mlib_VideoInterpAveY_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
00339     break;
00340 
00341     case 8:
00342       mlib_VideoInterpAveY_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
00343     break;
00344 
00345     case 16:
00346       mlib_VideoInterpAveY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
00347     break;
00348 
00349     default:
00350       assert(0);
00351   }
00352 }
00353 
00354 static void avg_pixels8_xy2_mlib(uint8_t * dest, const uint8_t * ref,
00355                                   int stride, int height)
00356 {
00357   switch (height) {
00358     case 4:
00359       mlib_VideoInterpAveXY_U8_U8_8x4(dest, (uint8_t *)ref, stride, stride);
00360     break;
00361 
00362     case 8:
00363       mlib_VideoInterpAveXY_U8_U8_8x8(dest, (uint8_t *)ref, stride, stride);
00364     break;
00365 
00366     case 16:
00367       mlib_VideoInterpAveXY_U8_U8_8x16(dest, (uint8_t *)ref, stride, stride);
00368     break;
00369 
00370     default:
00371       assert(0);
00372   }
00373 }
00374 
00375 /* swap byte order of a buffer */
00376 
00377 static void bswap_buf_mlib(uint32_t *dst, const uint32_t *src, int w)
00378 {
00379   mlib_VectorReverseByteOrder_U32_U32(dst, src, w);
00380 }
00381 
00382 /* transformations */
00383 
00384 static void ff_idct_put_mlib(uint8_t *dest, int line_size, DCTELEM *data)
00385 {
00386     int i;
00387     uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
00388 
00389     mlib_VideoIDCT8x8_S16_S16 (data, data);
00390 
00391     for(i=0;i<8;i++) {
00392         dest[0] = cm[data[0]];
00393         dest[1] = cm[data[1]];
00394         dest[2] = cm[data[2]];
00395         dest[3] = cm[data[3]];
00396         dest[4] = cm[data[4]];
00397         dest[5] = cm[data[5]];
00398         dest[6] = cm[data[6]];
00399         dest[7] = cm[data[7]];
00400 
00401         dest += line_size;
00402         data += 8;
00403     }
00404 }
00405 
00406 static void ff_idct_add_mlib(uint8_t *dest, int line_size, DCTELEM *data)
00407 {
00408     mlib_VideoIDCT8x8_S16_S16 (data, data);
00409     mlib_VideoAddBlock_U8_S16(dest, (mlib_s16 *)data, line_size);
00410 }
00411 
00412 static void ff_idct_mlib(DCTELEM *data)
00413 {
00414     mlib_VideoIDCT8x8_S16_S16 (data, data);
00415 }
00416 
00417 static void ff_fdct_mlib(DCTELEM *data)
00418 {
00419     mlib_VideoDCT8x8_S16_S16 (data, data);
00420 }
00421 
00422 void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx)
00423 {
00424     const int high_bit_depth = avctx->codec_id == CODEC_ID_H264 && avctx->bits_per_raw_sample > 8;
00425 
00426     c->get_pixels  = get_pixels_mlib;
00427     c->diff_pixels = diff_pixels_mlib;
00428     c->add_pixels_clamped = add_pixels_clamped_mlib;
00429 
00430     if (!high_bit_depth) {
00431     c->put_pixels_tab[0][0] = put_pixels16_mlib;
00432     c->put_pixels_tab[0][1] = put_pixels16_x2_mlib;
00433     c->put_pixels_tab[0][2] = put_pixels16_y2_mlib;
00434     c->put_pixels_tab[0][3] = put_pixels16_xy2_mlib;
00435     c->put_pixels_tab[1][0] = put_pixels8_mlib;
00436     c->put_pixels_tab[1][1] = put_pixels8_x2_mlib;
00437     c->put_pixels_tab[1][2] = put_pixels8_y2_mlib;
00438     c->put_pixels_tab[1][3] = put_pixels8_xy2_mlib;
00439 
00440     c->avg_pixels_tab[0][0] = avg_pixels16_mlib;
00441     c->avg_pixels_tab[0][1] = avg_pixels16_x2_mlib;
00442     c->avg_pixels_tab[0][2] = avg_pixels16_y2_mlib;
00443     c->avg_pixels_tab[0][3] = avg_pixels16_xy2_mlib;
00444     c->avg_pixels_tab[1][0] = avg_pixels8_mlib;
00445     c->avg_pixels_tab[1][1] = avg_pixels8_x2_mlib;
00446     c->avg_pixels_tab[1][2] = avg_pixels8_y2_mlib;
00447     c->avg_pixels_tab[1][3] = avg_pixels8_xy2_mlib;
00448 
00449     c->put_no_rnd_pixels_tab[0][0] = put_pixels16_mlib;
00450     c->put_no_rnd_pixels_tab[1][0] = put_pixels8_mlib;
00451     }
00452 
00453     c->bswap_buf = bswap_buf_mlib;
00454 }
00455 
00456 void MPV_common_init_mlib(MpegEncContext *s)
00457 {
00458     if(s->avctx->dct_algo==FF_DCT_AUTO || s->avctx->dct_algo==FF_DCT_MLIB){
00459         s->dsp.fdct = ff_fdct_mlib;
00460     }
00461 
00462     if(s->avctx->idct_algo==FF_IDCT_MLIB){
00463         s->dsp.idct_put= ff_idct_put_mlib;
00464         s->dsp.idct_add= ff_idct_add_mlib;
00465         s->dsp.idct    = ff_idct_mlib;
00466         s->dsp.idct_permutation_type= FF_NO_IDCT_PERM;
00467     }
00468 }
libavcodec/mlib/dsputil_mlib.c
