git.sesse.net Git - ffmpeg/blob - libavcodec/alpha/dsputil_alpha.c

   1 /*
   2  * Alpha optimized DSP utils
   3  * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #include "libavutil/attributes.h"
  23 #include "libavcodec/dsputil.h"
  24 #include "dsputil_alpha.h"
  25 #include "asm.h"
  26
  27 void (*put_pixels_clamped_axp_p)(const int16_t *block, uint8_t *pixels,
  28                                  int line_size);
  29 void (*add_pixels_clamped_axp_p)(const int16_t *block, uint8_t *pixels,
  30                                  int line_size);
  31
  32 #if 0
  33 /* These functions were the base for the optimized assembler routines,
  34    and remain here for documentation purposes.  */
  35 static void put_pixels_clamped_mvi(const int16_t *block, uint8_t *pixels,
  36                                    ptrdiff_t line_size)
  37 {
  38     int i = 8;
  39     uint64_t clampmask = zap(-1, 0xaa); /* 0x00ff00ff00ff00ff */
  40
  41     do {
  42         uint64_t shorts0, shorts1;
  43
  44         shorts0 = ldq(block);
  45         shorts0 = maxsw4(shorts0, 0);
  46         shorts0 = minsw4(shorts0, clampmask);
  47         stl(pkwb(shorts0), pixels);
  48
  49         shorts1 = ldq(block + 4);
  50         shorts1 = maxsw4(shorts1, 0);
  51         shorts1 = minsw4(shorts1, clampmask);
  52         stl(pkwb(shorts1), pixels + 4);
  53
  54         pixels += line_size;
  55         block += 8;
  56     } while (--i);
  57 }
  58
  59 void add_pixels_clamped_mvi(const int16_t *block, uint8_t *pixels,
  60                             ptrdiff_t line_size)
  61 {
  62     int h = 8;
  63     /* Keep this function a leaf function by generating the constants
  64        manually (mainly for the hack value ;-).  */
  65     uint64_t clampmask = zap(-1, 0xaa); /* 0x00ff00ff00ff00ff */
  66     uint64_t signmask  = zap(-1, 0x33);
  67     signmask ^= signmask >> 1;  /* 0x8000800080008000 */
  68
  69     do {
  70         uint64_t shorts0, pix0, signs0;
  71         uint64_t shorts1, pix1, signs1;
  72
  73         shorts0 = ldq(block);
  74         shorts1 = ldq(block + 4);
  75
  76         pix0    = unpkbw(ldl(pixels));
  77         /* Signed subword add (MMX paddw).  */
  78         signs0  = shorts0 & signmask;
  79         shorts0 &= ~signmask;
  80         shorts0 += pix0;
  81         shorts0 ^= signs0;
  82         /* Clamp. */
  83         shorts0 = maxsw4(shorts0, 0);
  84         shorts0 = minsw4(shorts0, clampmask);
  85
  86         /* Next 4.  */
  87         pix1    = unpkbw(ldl(pixels + 4));
  88         signs1  = shorts1 & signmask;
  89         shorts1 &= ~signmask;
  90         shorts1 += pix1;
  91         shorts1 ^= signs1;
  92         shorts1 = maxsw4(shorts1, 0);
  93         shorts1 = minsw4(shorts1, clampmask);
  94
  95         stl(pkwb(shorts0), pixels);
  96         stl(pkwb(shorts1), pixels + 4);
  97
  98         pixels += line_size;
  99         block += 8;
 100     } while (--h);
 101 }
 102 #endif
 103
 104 static void clear_blocks_axp(int16_t *blocks) {
 105     uint64_t *p = (uint64_t *) blocks;
 106     int n = sizeof(int16_t) * 6 * 64;
 107
 108     do {
 109         p[0] = 0;
 110         p[1] = 0;
 111         p[2] = 0;
 112         p[3] = 0;
 113         p[4] = 0;
 114         p[5] = 0;
 115         p[6] = 0;
 116         p[7] = 0;
 117         p += 8;
 118         n -= 8 * 8;
 119     } while (n);
 120 }
 121
 122 static inline uint64_t avg2_no_rnd(uint64_t a, uint64_t b)
 123 {
 124     return (a & b) + (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
 125 }
 126
 127 static inline uint64_t avg2(uint64_t a, uint64_t b)
 128 {
 129     return (a | b) - (((a ^ b) & BYTE_VEC(0xfe)) >> 1);
 130 }
 131
 132 #if 0
 133 /* The XY2 routines basically utilize this scheme, but reuse parts in
 134    each iteration.  */
 135 static inline uint64_t avg4(uint64_t l1, uint64_t l2, uint64_t l3, uint64_t l4)
 136 {
 137     uint64_t r1 = ((l1 & ~BYTE_VEC(0x03)) >> 2)
 138                 + ((l2 & ~BYTE_VEC(0x03)) >> 2)
 139                 + ((l3 & ~BYTE_VEC(0x03)) >> 2)
 140                 + ((l4 & ~BYTE_VEC(0x03)) >> 2);
 141     uint64_t r2 = ((  (l1 & BYTE_VEC(0x03))
 142                     + (l2 & BYTE_VEC(0x03))
 143                     + (l3 & BYTE_VEC(0x03))
 144                     + (l4 & BYTE_VEC(0x03))
 145                     + BYTE_VEC(0x02)) >> 2) & BYTE_VEC(0x03);
 146     return r1 + r2;
 147 }
 148 #endif
 149
 150 #define OP(LOAD, STORE)                         \
 151     do {                                        \
 152         STORE(LOAD(pixels), block);             \
 153         pixels += line_size;                    \
 154         block += line_size;                     \
 155     } while (--h)
 156
 157 #define OP_X2(LOAD, STORE)                                      \
 158     do {                                                        \
 159         uint64_t pix1, pix2;                                    \
 160                                                                 \
 161         pix1 = LOAD(pixels);                                    \
 162         pix2 = pix1 >> 8 | ((uint64_t) pixels[8] << 56);        \
 163         STORE(AVG2(pix1, pix2), block);                         \
 164         pixels += line_size;                                    \
 165         block += line_size;                                     \
 166     } while (--h)
 167
 168 #define OP_Y2(LOAD, STORE)                      \
 169     do {                                        \
 170         uint64_t pix = LOAD(pixels);            \
 171         do {                                    \
 172             uint64_t next_pix;                  \
 173                                                 \
 174             pixels += line_size;                \
 175             next_pix = LOAD(pixels);            \
 176             STORE(AVG2(pix, next_pix), block);  \
 177             block += line_size;                 \
 178             pix = next_pix;                     \
 179         } while (--h);                          \
 180     } while (0)
 181
 182 #define OP_XY2(LOAD, STORE)                                                 \
 183     do {                                                                    \
 184         uint64_t pix1 = LOAD(pixels);                                       \
 185         uint64_t pix2 = pix1 >> 8 | ((uint64_t) pixels[8] << 56);           \
 186         uint64_t pix_l = (pix1 & BYTE_VEC(0x03))                            \
 187                        + (pix2 & BYTE_VEC(0x03));                           \
 188         uint64_t pix_h = ((pix1 & ~BYTE_VEC(0x03)) >> 2)                    \
 189                        + ((pix2 & ~BYTE_VEC(0x03)) >> 2);                   \
 190                                                                             \
 191         do {                                                                \
 192             uint64_t npix1, npix2;                                          \
 193             uint64_t npix_l, npix_h;                                        \
 194             uint64_t avg;                                                   \
 195                                                                             \
 196             pixels += line_size;                                            \
 197             npix1 = LOAD(pixels);                                           \
 198             npix2 = npix1 >> 8 | ((uint64_t) pixels[8] << 56);              \
 199             npix_l = (npix1 & BYTE_VEC(0x03))                               \
 200                    + (npix2 & BYTE_VEC(0x03));                              \
 201             npix_h = ((npix1 & ~BYTE_VEC(0x03)) >> 2)                       \
 202                    + ((npix2 & ~BYTE_VEC(0x03)) >> 2);                      \
 203             avg = (((pix_l + npix_l + AVG4_ROUNDER) >> 2) & BYTE_VEC(0x03)) \
 204                 + pix_h + npix_h;                                           \
 205             STORE(avg, block);                                              \
 206                                                                             \
 207             block += line_size;                                             \
 208             pix_l = npix_l;                                                 \
 209             pix_h = npix_h;                                                 \
 210         } while (--h);                                                      \
 211     } while (0)
 212
 213 #define MAKE_OP(OPNAME, SUFF, OPKIND, STORE)                                \
 214 static void OPNAME ## _pixels ## SUFF ## _axp                               \
 215         (uint8_t *restrict block, const uint8_t *restrict pixels,           \
 216          ptrdiff_t line_size, int h)                                        \
 217 {                                                                           \
 218     if ((size_t) pixels & 0x7) {                                            \
 219         OPKIND(uldq, STORE);                                                \
 220     } else {                                                                \
 221         OPKIND(ldq, STORE);                                                 \
 222     }                                                                       \
 223 }                                                                           \
 224                                                                             \
 225 static void OPNAME ## _pixels16 ## SUFF ## _axp                             \
 226         (uint8_t *restrict block, const uint8_t *restrict pixels,           \
 227          ptrdiff_t line_size, int h)                                        \
 228 {                                                                           \
 229     OPNAME ## _pixels ## SUFF ## _axp(block,     pixels,     line_size, h); \
 230     OPNAME ## _pixels ## SUFF ## _axp(block + 8, pixels + 8, line_size, h); \
 231 }
 232
 233 #define PIXOP(OPNAME, STORE)                    \
 234     MAKE_OP(OPNAME, ,     OP,     STORE)        \
 235     MAKE_OP(OPNAME, _x2,  OP_X2,  STORE)        \
 236     MAKE_OP(OPNAME, _y2,  OP_Y2,  STORE)        \
 237     MAKE_OP(OPNAME, _xy2, OP_XY2, STORE)
 238
 239 /* Rounding primitives.  */
 240 #define AVG2 avg2
 241 #define AVG4 avg4
 242 #define AVG4_ROUNDER BYTE_VEC(0x02)
 243 #define STORE(l, b) stq(l, b)
 244 PIXOP(put, STORE);
 245
 246 #undef STORE
 247 #define STORE(l, b) stq(AVG2(l, ldq(b)), b);
 248 PIXOP(avg, STORE);
 249
 250 /* Not rounding primitives.  */
 251 #undef AVG2
 252 #undef AVG4
 253 #undef AVG4_ROUNDER
 254 #undef STORE
 255 #define AVG2 avg2_no_rnd
 256 #define AVG4 avg4_no_rnd
 257 #define AVG4_ROUNDER BYTE_VEC(0x01)
 258 #define STORE(l, b) stq(l, b)
 259 PIXOP(put_no_rnd, STORE);
 260
 261 #undef STORE
 262 #define STORE(l, b) stq(AVG2(l, ldq(b)), b);
 263 PIXOP(avg_no_rnd, STORE);
 264
 265 static void put_pixels16_axp_asm(uint8_t *block, const uint8_t *pixels,
 266                                  ptrdiff_t line_size, int h)
 267 {
 268     put_pixels_axp_asm(block,     pixels,     line_size, h);
 269     put_pixels_axp_asm(block + 8, pixels + 8, line_size, h);
 270 }
 271
 272 av_cold void ff_dsputil_init_alpha(DSPContext *c, AVCodecContext *avctx)
 273 {
 274     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
 275
 276     if (!high_bit_depth) {
 277     c->put_pixels_tab[0][0] = put_pixels16_axp_asm;
 278     c->put_pixels_tab[0][1] = put_pixels16_x2_axp;
 279     c->put_pixels_tab[0][2] = put_pixels16_y2_axp;
 280     c->put_pixels_tab[0][3] = put_pixels16_xy2_axp;
 281
 282     c->put_no_rnd_pixels_tab[0][0] = put_pixels16_axp_asm;
 283     c->put_no_rnd_pixels_tab[0][1] = put_no_rnd_pixels16_x2_axp;
 284     c->put_no_rnd_pixels_tab[0][2] = put_no_rnd_pixels16_y2_axp;
 285     c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_axp;
 286
 287     c->avg_pixels_tab[0][0] = avg_pixels16_axp;
 288     c->avg_pixels_tab[0][1] = avg_pixels16_x2_axp;
 289     c->avg_pixels_tab[0][2] = avg_pixels16_y2_axp;
 290     c->avg_pixels_tab[0][3] = avg_pixels16_xy2_axp;
 291
 292     c->avg_no_rnd_pixels_tab[0] = avg_no_rnd_pixels16_axp;
 293     c->avg_no_rnd_pixels_tab[1] = avg_no_rnd_pixels16_x2_axp;
 294     c->avg_no_rnd_pixels_tab[2] = avg_no_rnd_pixels16_y2_axp;
 295     c->avg_no_rnd_pixels_tab[3] = avg_no_rnd_pixels16_xy2_axp;
 296
 297     c->put_pixels_tab[1][0] = put_pixels_axp_asm;
 298     c->put_pixels_tab[1][1] = put_pixels_x2_axp;
 299     c->put_pixels_tab[1][2] = put_pixels_y2_axp;
 300     c->put_pixels_tab[1][3] = put_pixels_xy2_axp;
 301
 302     c->put_no_rnd_pixels_tab[1][0] = put_pixels_axp_asm;
 303     c->put_no_rnd_pixels_tab[1][1] = put_no_rnd_pixels_x2_axp;
 304     c->put_no_rnd_pixels_tab[1][2] = put_no_rnd_pixels_y2_axp;
 305     c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels_xy2_axp;
 306
 307     c->avg_pixels_tab[1][0] = avg_pixels_axp;
 308     c->avg_pixels_tab[1][1] = avg_pixels_x2_axp;
 309     c->avg_pixels_tab[1][2] = avg_pixels_y2_axp;
 310     c->avg_pixels_tab[1][3] = avg_pixels_xy2_axp;
 311
 312     c->clear_blocks = clear_blocks_axp;
 313     }
 314
 315     /* amask clears all bits that correspond to present features.  */
 316     if (amask(AMASK_MVI) == 0) {
 317         c->put_pixels_clamped = put_pixels_clamped_mvi_asm;
 318         c->add_pixels_clamped = add_pixels_clamped_mvi_asm;
 319
 320         if (!high_bit_depth)
 321             c->get_pixels   = get_pixels_mvi;
 322         c->diff_pixels      = diff_pixels_mvi;
 323         c->sad[0]           = pix_abs16x16_mvi_asm;
 324         c->sad[1]           = pix_abs8x8_mvi;
 325         c->pix_abs[0][0]    = pix_abs16x16_mvi_asm;
 326         c->pix_abs[1][0]    = pix_abs8x8_mvi;
 327         c->pix_abs[0][1]    = pix_abs16x16_x2_mvi;
 328         c->pix_abs[0][2]    = pix_abs16x16_y2_mvi;
 329         c->pix_abs[0][3]    = pix_abs16x16_xy2_mvi;
 330     }
 331
 332     put_pixels_clamped_axp_p = c->put_pixels_clamped;
 333     add_pixels_clamped_axp_p = c->add_pixels_clamped;
 334
 335     if (!avctx->lowres && avctx->bits_per_raw_sample <= 8 &&
 336         (avctx->idct_algo == FF_IDCT_AUTO ||
 337          avctx->idct_algo == FF_IDCT_SIMPLEALPHA)) {
 338         c->idct_put = ff_simple_idct_put_axp;
 339         c->idct_add = ff_simple_idct_add_axp;
 340         c->idct =     ff_simple_idct_axp;
 341     }
 342 }