vp8/common/x86/vp8_asm_stubs.c

   1 /*
   2  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
   3  *
   4  *  Use of this source code is governed by a BSD-style license
   5  *  that can be found in the LICENSE file in the root of the source
   6  *  tree. An additional intellectual property rights grant can be found
   7  *  in the file PATENTS.  All contributing project authors may
   8  *  be found in the AUTHORS file in the root of the source tree.
   9  */
  10
  11
  12 #include "vpx_ports/config.h"
  13 #include "vpx_ports/mem.h"
  14 #include "vp8/common/subpixel.h"
  15
  16 extern const short vp8_six_tap_mmx[8][6*8];
  17 extern const short vp8_bilinear_filters_mmx[8][2*8];
  18
  19 extern void vp8_filter_block1d_h6_mmx
  20 (
  21     unsigned char   *src_ptr,
  22     unsigned short  *output_ptr,
  23     unsigned int    src_pixels_per_line,
  24     unsigned int    pixel_step,
  25     unsigned int    output_height,
  26     unsigned int    output_width,
  27     const short      *vp8_filter
  28 );
  29 extern void vp8_filter_block1dc_v6_mmx
  30 (
  31     unsigned short *src_ptr,
  32     unsigned char  *output_ptr,
  33     int             output_pitch,
  34     unsigned int    pixels_per_line,
  35     unsigned int    pixel_step,
  36     unsigned int    output_height,
  37     unsigned int    output_width,
  38     const short    *vp8_filter
  39 );
  40 extern void vp8_filter_block1d8_h6_sse2
  41 (
  42     unsigned char  *src_ptr,
  43     unsigned short *output_ptr,
  44     unsigned int    src_pixels_per_line,
  45     unsigned int    pixel_step,
  46     unsigned int    output_height,
  47     unsigned int    output_width,
  48     const short    *vp8_filter
  49 );
  50 extern void vp8_filter_block1d16_h6_sse2
  51 (
  52     unsigned char  *src_ptr,
  53     unsigned short *output_ptr,
  54     unsigned int    src_pixels_per_line,
  55     unsigned int    pixel_step,
  56     unsigned int    output_height,
  57     unsigned int    output_width,
  58     const short    *vp8_filter
  59 );
  60 extern void vp8_filter_block1d8_v6_sse2
  61 (
  62     unsigned short *src_ptr,
  63     unsigned char *output_ptr,
  64     int dst_ptich,
  65     unsigned int pixels_per_line,
  66     unsigned int pixel_step,
  67     unsigned int output_height,
  68     unsigned int output_width,
  69     const short    *vp8_filter
  70 );
  71 extern void vp8_filter_block1d16_v6_sse2
  72 (
  73     unsigned short *src_ptr,
  74     unsigned char *output_ptr,
  75     int dst_ptich,
  76     unsigned int pixels_per_line,
  77     unsigned int pixel_step,
  78     unsigned int output_height,
  79     unsigned int output_width,
  80     const short    *vp8_filter
  81 );
  82 extern void vp8_unpack_block1d16_h6_sse2
  83 (
  84     unsigned char  *src_ptr,
  85     unsigned short *output_ptr,
  86     unsigned int    src_pixels_per_line,
  87     unsigned int    output_height,
  88     unsigned int    output_width
  89 );
  90 extern void vp8_filter_block1d8_h6_only_sse2
  91 (
  92     unsigned char  *src_ptr,
  93     unsigned int    src_pixels_per_line,
  94     unsigned char  *output_ptr,
  95     int dst_ptich,
  96     unsigned int    output_height,
  97     const short    *vp8_filter
  98 );
  99 extern void vp8_filter_block1d16_h6_only_sse2
 100 (
 101     unsigned char  *src_ptr,
 102     unsigned int    src_pixels_per_line,
 103     unsigned char  *output_ptr,
 104     int dst_ptich,
 105     unsigned int    output_height,
 106     const short    *vp8_filter
 107 );
 108 extern void vp8_filter_block1d8_v6_only_sse2
 109 (
 110     unsigned char *src_ptr,
 111     unsigned int   src_pixels_per_line,
 112     unsigned char *output_ptr,
 113     int dst_ptich,
 114     unsigned int   output_height,
 115     const short   *vp8_filter
 116 );
 117 extern prototype_subpixel_predict(vp8_bilinear_predict8x8_mmx);
 118
 119
 120 #if HAVE_MMX
 121 void vp8_sixtap_predict4x4_mmx
 122 (
 123     unsigned char  *src_ptr,
 124     int   src_pixels_per_line,
 125     int  xoffset,
 126     int  yoffset,
 127     unsigned char *dst_ptr,
 128     int dst_pitch
 129 )
 130 {
 131     DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 16*16);  /* Temp data bufffer used in filtering */
 132     const short *HFilter, *VFilter;
 133     HFilter = vp8_six_tap_mmx[xoffset];
 134     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line), FData2, src_pixels_per_line, 1, 9, 8, HFilter);
 135     VFilter = vp8_six_tap_mmx[yoffset];
 136     vp8_filter_block1dc_v6_mmx(FData2 + 8, dst_ptr, dst_pitch, 8, 4 , 4, 4, VFilter);
 137
 138 }
 139
 140
 141 void vp8_sixtap_predict16x16_mmx
 142 (
 143     unsigned char  *src_ptr,
 144     int   src_pixels_per_line,
 145     int  xoffset,
 146     int  yoffset,
 147     unsigned char *dst_ptr,
 148     int dst_pitch
 149 )
 150 {
 151
 152     DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 24*24);  /* Temp data bufffer used in filtering */
 153
 154     const short *HFilter, *VFilter;
 155
 156
 157     HFilter = vp8_six_tap_mmx[xoffset];
 158
 159     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 21, 32, HFilter);
 160     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 21, 32, HFilter);
 161     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 8,  FData2 + 8, src_pixels_per_line, 1, 21, 32, HFilter);
 162     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 12, FData2 + 12, src_pixels_per_line, 1, 21, 32, HFilter);
 163
 164     VFilter = vp8_six_tap_mmx[yoffset];
 165     vp8_filter_block1dc_v6_mmx(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, 16, VFilter);
 166     vp8_filter_block1dc_v6_mmx(FData2 + 36, dst_ptr + 4, dst_pitch, 32, 16 , 16, 16, VFilter);
 167     vp8_filter_block1dc_v6_mmx(FData2 + 40, dst_ptr + 8, dst_pitch, 32, 16 , 16, 16, VFilter);
 168     vp8_filter_block1dc_v6_mmx(FData2 + 44, dst_ptr + 12, dst_pitch, 32, 16 , 16, 16, VFilter);
 169
 170 }
 171
 172
 173 void vp8_sixtap_predict8x8_mmx
 174 (
 175     unsigned char  *src_ptr,
 176     int   src_pixels_per_line,
 177     int  xoffset,
 178     int  yoffset,
 179     unsigned char *dst_ptr,
 180     int dst_pitch
 181 )
 182 {
 183
 184     DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 256);    /* Temp data bufffer used in filtering */
 185
 186     const short *HFilter, *VFilter;
 187
 188     HFilter = vp8_six_tap_mmx[xoffset];
 189     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 13, 16, HFilter);
 190     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 13, 16, HFilter);
 191
 192     VFilter = vp8_six_tap_mmx[yoffset];
 193     vp8_filter_block1dc_v6_mmx(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 8, 8, VFilter);
 194     vp8_filter_block1dc_v6_mmx(FData2 + 20, dst_ptr + 4, dst_pitch, 16, 8 , 8, 8, VFilter);
 195
 196 }
 197
 198
 199 void vp8_sixtap_predict8x4_mmx
 200 (
 201     unsigned char  *src_ptr,
 202     int   src_pixels_per_line,
 203     int  xoffset,
 204     int  yoffset,
 205     unsigned char *dst_ptr,
 206     int dst_pitch
 207 )
 208 {
 209
 210     DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 256);    /* Temp data bufffer used in filtering */
 211
 212     const short *HFilter, *VFilter;
 213
 214     HFilter = vp8_six_tap_mmx[xoffset];
 215     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 9, 16, HFilter);
 216     vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 9, 16, HFilter);
 217
 218     VFilter = vp8_six_tap_mmx[yoffset];
 219     vp8_filter_block1dc_v6_mmx(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 4, 8, VFilter);
 220     vp8_filter_block1dc_v6_mmx(FData2 + 20, dst_ptr + 4, dst_pitch, 16, 8 , 4, 8, VFilter);
 221
 222 }
 223
 224
 225
 226 void vp8_bilinear_predict16x16_mmx
 227 (
 228     unsigned char  *src_ptr,
 229     int   src_pixels_per_line,
 230     int  xoffset,
 231     int  yoffset,
 232     unsigned char *dst_ptr,
 233     int dst_pitch
 234 )
 235 {
 236     vp8_bilinear_predict8x8_mmx(src_ptr,   src_pixels_per_line, xoffset, yoffset, dst_ptr,   dst_pitch);
 237     vp8_bilinear_predict8x8_mmx(src_ptr + 8, src_pixels_per_line, xoffset, yoffset, dst_ptr + 8, dst_pitch);
 238     vp8_bilinear_predict8x8_mmx(src_ptr + 8 * src_pixels_per_line,   src_pixels_per_line, xoffset, yoffset, dst_ptr + dst_pitch * 8,   dst_pitch);
 239     vp8_bilinear_predict8x8_mmx(src_ptr + 8 * src_pixels_per_line + 8, src_pixels_per_line, xoffset, yoffset, dst_ptr + dst_pitch * 8 + 8, dst_pitch);
 240 }
 241 #endif
 242
 243
 244 #if HAVE_SSE2
 245 void vp8_sixtap_predict16x16_sse2
 246 (
 247     unsigned char  *src_ptr,
 248     int   src_pixels_per_line,
 249     int  xoffset,
 250     int  yoffset,
 251     unsigned char *dst_ptr,
 252     int dst_pitch
 253
 254 )
 255 {
 256     DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 24*24);    /* Temp data bufffer used in filtering */
 257
 258     const short *HFilter, *VFilter;
 259
 260     if (xoffset)
 261     {
 262         if (yoffset)
 263         {
 264             HFilter = vp8_six_tap_mmx[xoffset];
 265             vp8_filter_block1d16_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 21, 32, HFilter);
 266             VFilter = vp8_six_tap_mmx[yoffset];
 267             vp8_filter_block1d16_v6_sse2(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, dst_pitch, VFilter);
 268         }
 269         else
 270         {
 271             /* First-pass only */
 272             HFilter = vp8_six_tap_mmx[xoffset];
 273             vp8_filter_block1d16_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 16, HFilter);
 274         }
 275     }
 276     else
 277     {
 278         /* Second-pass only */
 279         VFilter = vp8_six_tap_mmx[yoffset];
 280         vp8_unpack_block1d16_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 21, 32);
 281         vp8_filter_block1d16_v6_sse2(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, dst_pitch, VFilter);
 282     }
 283 }
 284
 285
 286 void vp8_sixtap_predict8x8_sse2
 287 (
 288     unsigned char  *src_ptr,
 289     int   src_pixels_per_line,
 290     int  xoffset,
 291     int  yoffset,
 292     unsigned char *dst_ptr,
 293     int dst_pitch
 294 )
 295 {
 296     DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 256);  /* Temp data bufffer used in filtering */
 297     const short *HFilter, *VFilter;
 298
 299     if (xoffset)
 300     {
 301         if (yoffset)
 302         {
 303             HFilter = vp8_six_tap_mmx[xoffset];
 304             vp8_filter_block1d8_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 13, 16, HFilter);
 305             VFilter = vp8_six_tap_mmx[yoffset];
 306             vp8_filter_block1d8_v6_sse2(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 8, dst_pitch, VFilter);
 307         }
 308         else
 309         {
 310             /* First-pass only */
 311             HFilter = vp8_six_tap_mmx[xoffset];
 312             vp8_filter_block1d8_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 8, HFilter);
 313         }
 314     }
 315     else
 316     {
 317         /* Second-pass only */
 318         VFilter = vp8_six_tap_mmx[yoffset];
 319         vp8_filter_block1d8_v6_only_sse2(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 8, VFilter);
 320     }
 321 }
 322
 323
 324 void vp8_sixtap_predict8x4_sse2
 325 (
 326     unsigned char  *src_ptr,
 327     int   src_pixels_per_line,
 328     int  xoffset,
 329     int  yoffset,
 330     unsigned char *dst_ptr,
 331     int dst_pitch
 332 )
 333 {
 334     DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 256);  /* Temp data bufffer used in filtering */
 335     const short *HFilter, *VFilter;
 336
 337     if (xoffset)
 338     {
 339         if (yoffset)
 340         {
 341             HFilter = vp8_six_tap_mmx[xoffset];
 342             vp8_filter_block1d8_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 9, 16, HFilter);
 343             VFilter = vp8_six_tap_mmx[yoffset];
 344             vp8_filter_block1d8_v6_sse2(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 4, dst_pitch, VFilter);
 345         }
 346         else
 347         {
 348             /* First-pass only */
 349             HFilter = vp8_six_tap_mmx[xoffset];
 350             vp8_filter_block1d8_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 4, HFilter);
 351         }
 352     }
 353     else
 354     {
 355         /* Second-pass only */
 356         VFilter = vp8_six_tap_mmx[yoffset];
 357         vp8_filter_block1d8_v6_only_sse2(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 4, VFilter);
 358     }
 359 }
 360
 361 #endif
 362
 363 #if HAVE_SSSE3
 364
 365 extern void vp8_filter_block1d8_h6_ssse3
 366 (
 367     unsigned char  *src_ptr,
 368     unsigned int    src_pixels_per_line,
 369     unsigned char  *output_ptr,
 370     unsigned int    output_pitch,
 371     unsigned int    output_height,
 372     unsigned int    vp8_filter_index
 373 );
 374
 375 extern void vp8_filter_block1d16_h6_ssse3
 376 (
 377     unsigned char  *src_ptr,
 378     unsigned int    src_pixels_per_line,
 379     unsigned char  *output_ptr,
 380     unsigned int    output_pitch,
 381     unsigned int    output_height,
 382     unsigned int    vp8_filter_index
 383 );
 384
 385 extern void vp8_filter_block1d16_v6_ssse3
 386 (
 387     unsigned char *src_ptr,
 388     unsigned int   src_pitch,
 389     unsigned char *output_ptr,
 390     unsigned int   out_pitch,
 391     unsigned int   output_height,
 392     unsigned int   vp8_filter_index
 393 );
 394
 395 extern void vp8_filter_block1d8_v6_ssse3
 396 (
 397     unsigned char *src_ptr,
 398     unsigned int   src_pitch,
 399     unsigned char *output_ptr,
 400     unsigned int   out_pitch,
 401     unsigned int   output_height,
 402     unsigned int   vp8_filter_index
 403 );
 404
 405 extern void vp8_filter_block1d4_h6_ssse3
 406 (
 407     unsigned char  *src_ptr,
 408     unsigned int    src_pixels_per_line,
 409     unsigned char  *output_ptr,
 410     unsigned int    output_pitch,
 411     unsigned int    output_height,
 412     unsigned int    vp8_filter_index
 413 );
 414
 415 extern void vp8_filter_block1d4_v6_ssse3
 416 (
 417     unsigned char *src_ptr,
 418     unsigned int   src_pitch,
 419     unsigned char *output_ptr,
 420     unsigned int   out_pitch,
 421     unsigned int   output_height,
 422     unsigned int   vp8_filter_index
 423 );
 424
 425 void vp8_sixtap_predict16x16_ssse3
 426 (
 427     unsigned char  *src_ptr,
 428     int   src_pixels_per_line,
 429     int  xoffset,
 430     int  yoffset,
 431     unsigned char *dst_ptr,
 432     int dst_pitch
 433
 434 )
 435 {
 436     DECLARE_ALIGNED_ARRAY(16, unsigned char, FData2, 24*24);
 437
 438     if (xoffset)
 439     {
 440         if (yoffset)
 441         {
 442             vp8_filter_block1d16_h6_ssse3(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, FData2, 16, 21, xoffset);
 443             vp8_filter_block1d16_v6_ssse3(FData2 , 16, dst_ptr, dst_pitch, 16, yoffset);
 444         }
 445         else
 446         {
 447             /* First-pass only */
 448             vp8_filter_block1d16_h6_ssse3(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 16, xoffset);
 449         }
 450     }
 451     else
 452     {
 453         /* Second-pass only */
 454         vp8_filter_block1d16_v6_ssse3(src_ptr - (2 * src_pixels_per_line) , src_pixels_per_line, dst_ptr, dst_pitch, 16, yoffset);
 455     }
 456 }
 457
 458 void vp8_sixtap_predict8x8_ssse3
 459 (
 460     unsigned char  *src_ptr,
 461     int   src_pixels_per_line,
 462     int  xoffset,
 463     int  yoffset,
 464     unsigned char *dst_ptr,
 465     int dst_pitch
 466 )
 467 {
 468     DECLARE_ALIGNED_ARRAY(16, unsigned char, FData2, 256);
 469
 470     if (xoffset)
 471     {
 472         if (yoffset)
 473         {
 474             vp8_filter_block1d8_h6_ssse3(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, FData2, 8, 13, xoffset);
 475             vp8_filter_block1d8_v6_ssse3(FData2, 8, dst_ptr, dst_pitch, 8, yoffset);
 476         }
 477         else
 478         {
 479             vp8_filter_block1d8_h6_ssse3(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 8, xoffset);
 480         }
 481     }
 482     else
 483     {
 484         /* Second-pass only */
 485         vp8_filter_block1d8_v6_ssse3(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 8, yoffset);
 486     }
 487 }
 488
 489
 490 void vp8_sixtap_predict8x4_ssse3
 491 (
 492     unsigned char  *src_ptr,
 493     int   src_pixels_per_line,
 494     int  xoffset,
 495     int  yoffset,
 496     unsigned char *dst_ptr,
 497     int dst_pitch
 498 )
 499 {
 500     DECLARE_ALIGNED_ARRAY(16, unsigned char, FData2, 256);
 501
 502     if (xoffset)
 503     {
 504         if (yoffset)
 505         {
 506             vp8_filter_block1d8_h6_ssse3(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, FData2, 8, 9, xoffset);
 507             vp8_filter_block1d8_v6_ssse3(FData2, 8, dst_ptr, dst_pitch, 4, yoffset);
 508         }
 509         else
 510         {
 511             /* First-pass only */
 512             vp8_filter_block1d8_h6_ssse3(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 4, xoffset);
 513         }
 514     }
 515     else
 516     {
 517         /* Second-pass only */
 518         vp8_filter_block1d8_v6_ssse3(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 4, yoffset);
 519     }
 520 }
 521
 522 void vp8_sixtap_predict4x4_ssse3
 523 (
 524     unsigned char  *src_ptr,
 525     int   src_pixels_per_line,
 526     int  xoffset,
 527     int  yoffset,
 528     unsigned char *dst_ptr,
 529     int dst_pitch
 530 )
 531 {
 532   DECLARE_ALIGNED_ARRAY(16, unsigned char, FData2, 4*9);
 533
 534   if (xoffset)
 535   {
 536       if (yoffset)
 537       {
 538           vp8_filter_block1d4_h6_ssse3(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, FData2, 4, 9, xoffset);
 539           vp8_filter_block1d4_v6_ssse3(FData2, 4, dst_ptr, dst_pitch, 4, yoffset);
 540       }
 541       else
 542       {
 543           vp8_filter_block1d4_h6_ssse3(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 4, xoffset);
 544       }
 545   }
 546   else
 547   {
 548       vp8_filter_block1d4_v6_ssse3(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 4, yoffset);
 549   }
 550
 551 }
 552
 553 #endif