FFmpeg: libavcodec/alpha/dsputil

00001 /*
00002  * Alpha optimized DSP utils
00003  * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
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 "dsputil_alpha.h"
00024 #include "asm.h"
00025 
00026 void (*put_pixels_clamped_axp_p)(const DCTELEM *block, uint8_t *pixels,
00027                                  int line_size);
00028 void (*add_pixels_clamped_axp_p)(const DCTELEM *block, uint8_t *pixels,
00029                                  int line_size);
00030 
00031 #if 0
00032 /* These functions were the base for the optimized assembler routines,
00033    and remain here for documentation purposes.  */
00034 static void put_pixels_clamped_mvi(const DCTELEM *block, uint8_t *pixels,
00035                                    int line_size)
00036 {
00037     int i = 8;
00038     uint64_t clampmask = zap(-1, 0xaa); /* 0x00ff00ff00ff00ff */
00039 
00040     do {
00041         uint64_t shorts0, shorts1;
00042 
00043         shorts0 = ldq(block);
00044         shorts0 = maxsw4(shorts0, 0);
00045         shorts0 = minsw4(shorts0, clampmask);
00046         stl(pkwb(shorts0), pixels);
00047 
00048         shorts1 = ldq(block + 4);
00049         shorts1 = maxsw4(shorts1, 0);
00050         shorts1 = minsw4(shorts1, clampmask);
00051         stl(pkwb(shorts1), pixels + 4);
00052 
00053         pixels += line_size;
00054         block += 8;
00055     } while (--i);
00056 }
00057 
00058 void add_pixels_clamped_mvi(const DCTELEM *block, uint8_t *pixels,
00059                             int line_size)
00060 {
00061     int h = 8;
00062     /* Keep this function a leaf function by generating the constants
00063        manually (mainly for the hack value ;-).  */
00064     uint64_t clampmask = zap(-1, 0xaa); /* 0x00ff00ff00ff00ff */
00065     uint64_t signmask  = zap(-1, 0x33);
00066     signmask ^= signmask >> 1;  /* 0x8000800080008000 */
00067 
00068     do {
00069         uint64_t shorts0, pix0, signs0;
00070         uint64_t shorts1, pix1, signs1;
00071 
00072         shorts0 = ldq(block);
00073         shorts1 = ldq(block + 4);
00074 
00075         pix0    = unpkbw(ldl(pixels));
00076         /* Signed subword add (MMX paddw).  */
00077         signs0  = shorts0 & signmask;
00078         shorts0 &= ~signmask;
00079         shorts0 += pix0;
00080         shorts0 ^= signs0;
00081         /* Clamp. */
00082         shorts0 = maxsw4(shorts0, 0);
00083         shorts0 = minsw4(shorts0, clampmask);
00084 
00085         /* Next 4.  */
00086         pix1    = unpkbw(ldl(pixels + 4));
00087         signs1  = shorts1 & signmask;
00088         shorts1 &= ~signmask;
00089         shorts1 += pix1;
00090         shorts1 ^= signs1;
00091         shorts1 = maxsw4(shorts1, 0);
00092         shorts1 = minsw4(shorts1, clampmask);
00093 
00094         stl(pkwb(shorts0), pixels);
00095         stl(pkwb(shorts1), pixels + 4);
00096 
00097         pixels += line_size;
00098         block += 8;
00099     } while (--h);
00100 }
00101 #endif
00102 
00103 static void clear_blocks_axp(DCTELEM *blocks) {
00104     uint64_t *p = (uint64_t *) blocks;
00105     int n = sizeof(DCTELEM) * 6 * 64;
00106 
00107     do {
00108         p[0] = 0;
00109         p[1] = 0;
00110         p[2] = 0;
00111         p[3] = 0;
00112         p[4] = 0;
00113         p[5] = 0;
00114         p[6] = 0;
00115         p[7] = 0;
00116         p += 8;
00117         n -= 8 * 8;
00118     } while (n);
00119 }
00120 
00121 static inline uint64_t avg2_no_rnd(uint64_t a, uint64_t b)
00122 {
00123     return (a & b) + (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
00124 }
00125 
00126 static inline uint64_t avg2(uint64_t a, uint64_t b)
00127 {
00128     return (a | b) - (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
00129 }
00130 
00131 #if 0
00132 /* The XY2 routines basically utilize this scheme, but reuse parts in
00133    each iteration.  */
00134 static inline uint64_t avg4(uint64_t l1, uint64_t l2, uint64_t l3, uint64_t l4)
00135 {
00136     uint64_t r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
00137                 + ((l2 & ~BYTE_VEC(0x03)) >> 2)
00138                 + ((l3 & ~BYTE_VEC(0x03)) >> 2)
00139                 + ((l4 & ~BYTE_VEC(0x03)) >> 2);
00140     uint64_t r2 = ((  (l1 & BYTE_VEC(0x03))
00141                     + (l2 & BYTE_VEC(0x03))
00142                     + (l3 & BYTE_VEC(0x03))
00143                     + (l4 & BYTE_VEC(0x03))
00144                     + BYTE_VEC(0x02)) >> 2) & BYTE_VEC(0x03);
00145     return r1 + r2;
00146 }
00147 #endif
00148 
00149 #define OP(LOAD, STORE)                         \
00150     do {                                        \
00151         STORE(LOAD(pixels), block);             \
00152         pixels += line_size;                    \
00153         block += line_size;                     \
00154     } while (--h)
00155 
00156 #define OP_X2(LOAD, STORE)                                      \
00157     do {                                                        \
00158         uint64_t pix1, pix2;                                    \
00159                                                                 \
00160         pix1 = LOAD(pixels);                                    \
00161         pix2 = pix1 >> 8 | ((uint64_t) pixels[8] << 56);        \
00162         STORE(AVG2(pix1, pix2), block);                         \
00163         pixels += line_size;                                    \
00164         block += line_size;                                     \
00165     } while (--h)
00166 
00167 #define OP_Y2(LOAD, STORE)                      \
00168     do {                                        \
00169         uint64_t pix = LOAD(pixels);            \
00170         do {                                    \
00171             uint64_t next_pix;                  \
00172                                                 \
00173             pixels += line_size;                \
00174             next_pix = LOAD(pixels);            \
00175             STORE(AVG2(pix, next_pix), block);  \
00176             block += line_size;                 \
00177             pix = next_pix;                     \
00178         } while (--h);                          \
00179     } while (0)
00180 
00181 #define OP_XY2(LOAD, STORE)                                                 \
00182     do {                                                                    \
00183         uint64_t pix1 = LOAD(pixels);                                       \
00184         uint64_t pix2 = pix1 >> 8 | ((uint64_t) pixels[8] << 56);           \
00185         uint64_t pix_l = (pix1 & BYTE_VEC(0x03))                            \
00186                        + (pix2 & BYTE_VEC(0x03));                           \
00187         uint64_t pix_h = ((pix1 & ~BYTE_VEC(0x03)) >> 2)                    \
00188                        + ((pix2 & ~BYTE_VEC(0x03)) >> 2);                   \
00189                                                                             \
00190         do {                                                                \
00191             uint64_t npix1, npix2;                                          \
00192             uint64_t npix_l, npix_h;                                        \
00193             uint64_t avg;                                                   \
00194                                                                             \
00195             pixels += line_size;                                            \
00196             npix1 = LOAD(pixels);                                           \
00197             npix2 = npix1 >> 8 | ((uint64_t) pixels[8] << 56);              \
00198             npix_l = (npix1 & BYTE_VEC(0x03))                               \
00199                    + (npix2 & BYTE_VEC(0x03));                              \
00200             npix_h = ((npix1 & ~BYTE_VEC(0x03)) >> 2)                       \
00201                    + ((npix2 & ~BYTE_VEC(0x03)) >> 2);                      \
00202             avg = (((pix_l + npix_l + AVG4_ROUNDER) >> 2) & BYTE_VEC(0x03)) \
00203                 + pix_h + npix_h;                                           \
00204             STORE(avg, block);                                              \
00205                                                                             \
00206             block += line_size;                                             \
00207             pix_l = npix_l;                                                 \
00208             pix_h = npix_h;                                                 \
00209         } while (--h);                                                      \
00210     } while (0)
00211 
00212 #define MAKE_OP(OPNAME, SUFF, OPKIND, STORE)                                \
00213 static void OPNAME ## _pixels ## SUFF ## _axp                               \
00214         (uint8_t *restrict block, const uint8_t *restrict pixels,           \
00215          int line_size, int h)                                              \
00216 {                                                                           \
00217     if ((size_t) pixels & 0x7) {                                            \
00218         OPKIND(uldq, STORE);                                                \
00219     } else {                                                                \
00220         OPKIND(ldq, STORE);                                                 \
00221     }                                                                       \
00222 }                                                                           \
00223                                                                             \
00224 static void OPNAME ## _pixels16 ## SUFF ## _axp                             \
00225         (uint8_t *restrict block, const uint8_t *restrict pixels,           \
00226          int line_size, int h)                                              \
00227 {                                                                           \
00228     OPNAME ## _pixels ## SUFF ## _axp(block,     pixels,     line_size, h); \
00229     OPNAME ## _pixels ## SUFF ## _axp(block + 8, pixels + 8, line_size, h); \
00230 }
00231 
00232 #define PIXOP(OPNAME, STORE)                    \
00233     MAKE_OP(OPNAME, ,     OP,     STORE)        \
00234     MAKE_OP(OPNAME, _x2,  OP_X2,  STORE)        \
00235     MAKE_OP(OPNAME, _y2,  OP_Y2,  STORE)        \
00236     MAKE_OP(OPNAME, _xy2, OP_XY2, STORE)
00237 
00238 /* Rounding primitives.  */
00239 #define AVG2 avg2
00240 #define AVG4 avg4
00241 #define AVG4_ROUNDER BYTE_VEC(0x02)
00242 #define STORE(l, b) stq(l, b)
00243 PIXOP(put, STORE);
00244 
00245 #undef STORE
00246 #define STORE(l, b) stq(AVG2(l, ldq(b)), b);
00247 PIXOP(avg, STORE);
00248 
00249 /* Not rounding primitives.  */
00250 #undef AVG2
00251 #undef AVG4
00252 #undef AVG4_ROUNDER
00253 #undef STORE
00254 #define AVG2 avg2_no_rnd
00255 #define AVG4 avg4_no_rnd
00256 #define AVG4_ROUNDER BYTE_VEC(0x01)
00257 #define STORE(l, b) stq(l, b)
00258 PIXOP(put_no_rnd, STORE);
00259 
00260 #undef STORE
00261 #define STORE(l, b) stq(AVG2(l, ldq(b)), b);
00262 PIXOP(avg_no_rnd, STORE);
00263 
00264 static void put_pixels16_axp_asm(uint8_t *block, const uint8_t *pixels,
00265                                  int line_size, int h)
00266 {
00267     put_pixels_axp_asm(block,     pixels,     line_size, h);
00268     put_pixels_axp_asm(block + 8, pixels + 8, line_size, h);
00269 }
00270 
00271 void ff_dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx)
00272 {
00273     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
00274 
00275     if (!high_bit_depth) {
00276     c->put_pixels_tab[0][0] = put_pixels16_axp_asm;
00277     c->put_pixels_tab[0][1] = put_pixels16_x2_axp;
00278     c->put_pixels_tab[0][2] = put_pixels16_y2_axp;
00279     c->put_pixels_tab[0][3] = put_pixels16_xy2_axp;
00280 
00281     c->put_no_rnd_pixels_tab[0][0] = put_pixels16_axp_asm;
00282     c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_axp;
00283     c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_axp;
00284     c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_axp;
00285 
00286     c->avg_pixels_tab[0][0] = avg_pixels16_axp;
00287     c->avg_pixels_tab[0][1] = avg_pixels16_x2_axp;
00288     c->avg_pixels_tab[0][2] = avg_pixels16_y2_axp;
00289     c->avg_pixels_tab[0][3] = avg_pixels16_xy2_axp;
00290 
00291     c->avg_no_rnd_pixels_tab[0][0] = avg_no_rnd_pixels16_axp;
00292     c->avg_no_rnd_pixels_tab[0][1] = avg_no_rnd_pixels16_x2_axp;
00293     c->avg_no_rnd_pixels_tab[0][2] = avg_no_rnd_pixels16_y2_axp;
00294     c->avg_no_rnd_pixels_tab[0][3] = avg_no_rnd_pixels16_xy2_axp;
00295 
00296     c->put_pixels_tab[1][0] = put_pixels_axp_asm;
00297     c->put_pixels_tab[1][1] = put_pixels_x2_axp;
00298     c->put_pixels_tab[1][2] = put_pixels_y2_axp;
00299     c->put_pixels_tab[1][3] = put_pixels_xy2_axp;
00300 
00301     c->put_no_rnd_pixels_tab[1][0] = put_pixels_axp_asm;
00302     c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels_x2_axp;
00303     c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels_y2_axp;
00304     c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels_xy2_axp;
00305 
00306     c->avg_pixels_tab[1][0] = avg_pixels_axp;
00307     c->avg_pixels_tab[1][1] = avg_pixels_x2_axp;
00308     c->avg_pixels_tab[1][2] = avg_pixels_y2_axp;
00309     c->avg_pixels_tab[1][3] = avg_pixels_xy2_axp;
00310 
00311     c->avg_no_rnd_pixels_tab[1][0] = avg_no_rnd_pixels_axp;
00312     c->avg_no_rnd_pixels_tab[1][1] = avg_no_rnd_pixels_x2_axp;
00313     c->avg_no_rnd_pixels_tab[1][2] = avg_no_rnd_pixels_y2_axp;
00314     c->avg_no_rnd_pixels_tab[1][3] = avg_no_rnd_pixels_xy2_axp;
00315 
00316     c->clear_blocks = clear_blocks_axp;
00317     }
00318 
00319     /* amask clears all bits that correspond to present features.  */
00320     if (amask(AMASK_MVI) == 0) {
00321         c->put_pixels_clamped = put_pixels_clamped_mvi_asm;
00322         c->add_pixels_clamped = add_pixels_clamped_mvi_asm;
00323 
00324         if (!high_bit_depth)
00325             c->get_pixels   = get_pixels_mvi;
00326         c->diff_pixels      = diff_pixels_mvi;
00327         c->sad[0]           = pix_abs16x16_mvi_asm;
00328         c->sad[1]           = pix_abs8x8_mvi;
00329         c->pix_abs[0][0]    = pix_abs16x16_mvi_asm;
00330         c->pix_abs[1][0]    = pix_abs8x8_mvi;
00331         c->pix_abs[0][1]    = pix_abs16x16_x2_mvi;
00332         c->pix_abs[0][2]    = pix_abs16x16_y2_mvi;
00333         c->pix_abs[0][3]    = pix_abs16x16_xy2_mvi;
00334     }
00335 
00336     put_pixels_clamped_axp_p = c->put_pixels_clamped;
00337     add_pixels_clamped_axp_p = c->add_pixels_clamped;
00338 
00339     if (!avctx->lowres && avctx->bits_per_raw_sample <= 8 &&
00340         (avctx->idct_algo == FF_IDCT_AUTO ||
00341          avctx->idct_algo == FF_IDCT_SIMPLEALPHA)) {
00342         c->idct_put = ff_simple_idct_put_axp;
00343         c->idct_add = ff_simple_idct_add_axp;
00344         c->idct =     ff_simple_idct_axp;
00345     }
00346 }
libavcodec/alpha/dsputil_alpha.c
