pixman/pixman-arm-neon.c

   1 /*
   2  * Copyright © 2009 ARM Ltd, Movial Creative Technologies Oy
   3  *
   4  * Permission to use, copy, modify, distribute, and sell this software and its
   5  * documentation for any purpose is hereby granted without fee, provided that
   6  * the above copyright notice appear in all copies and that both that
   7  * copyright notice and this permission notice appear in supporting
   8  * documentation, and that the name of ARM Ltd not be used in
   9  * advertising or publicity pertaining to distribution of the software without
  10  * specific, written prior permission.  ARM Ltd makes no
  11  * representations about the suitability of this software for any purpose.  It
  12  * is provided "as is" without express or implied warranty.
  13  *
  14  * THE COPYRIGHT HOLDERS DISCLAIM ALL WARRANTIES WITH REGARD TO THIS
  15  * SOFTWARE, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND
  16  * FITNESS, IN NO EVENT SHALL THE COPYRIGHT HOLDERS BE LIABLE FOR ANY
  17  * SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
  18  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN
  19  * AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING
  20  * OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS
  21  * SOFTWARE.
  22  *
  23  * Author:  Ian Rickards (ian.rickards@arm.com)
  24  * Author:  Jonathan Morton (jonathan.morton@movial.com)
  25  * Author:  Markku Vire (markku.vire@movial.com)
  26  *
  27  */
  28
  29 #ifdef HAVE_CONFIG_H
  30 #include <config.h>
  31 #endif
  32
  33 #include <string.h>
  34 #include "pixman-private.h"
  35
  36 #define BIND_SRC_NULL_DST(name, src_type, src_cnt, dst_type, dst_cnt)   \
  37 void                                                                    \
  38 pixman_composite_##name##_asm_neon (int32_t   w,                        \
  39                                     int32_t   h,                        \
  40                                     dst_type *dst,                      \
  41                                     int32_t   dst_stride,               \
  42                                     src_type *src,                      \
  43                                     int32_t   src_stride);              \
  44                                                                         \
  45 static void                                                             \
  46 neon_composite_##name (pixman_implementation_t *imp,                    \
  47                        pixman_op_t              op,                     \
  48                        pixman_image_t *         src_image,              \
  49                        pixman_image_t *         mask_image,             \
  50                        pixman_image_t *         dst_image,              \
  51                        int32_t                  src_x,                  \
  52                        int32_t                  src_y,                  \
  53                        int32_t                  mask_x,                 \
  54                        int32_t                  mask_y,                 \
  55                        int32_t                  dest_x,                 \
  56                        int32_t                  dest_y,                 \
  57                        int32_t                  width,                  \
  58                        int32_t                  height)                 \
  59 {                                                                       \
  60     dst_type *dst_line;                                                 \
  61     src_type *src_line;                                                 \
  62     int32_t dst_stride, src_stride;                                     \
  63                                                                         \
  64     PIXMAN_IMAGE_GET_LINE (src_image, src_x, src_y, src_type,           \
  65                            src_stride, src_line, src_cnt);              \
  66     PIXMAN_IMAGE_GET_LINE (dst_image, dest_x, dest_y, dst_type,         \
  67                            dst_stride, dst_line, dst_cnt);              \
  68                                                                         \
  69     pixman_composite_##name##_asm_neon (width, height,                  \
  70                                         dst_line, dst_stride,           \
  71                                         src_line, src_stride);          \
  72 }
  73
  74 #define BIND_N_MASK_DST(name, mask_type, mask_cnt, dst_type, dst_cnt)   \
  75 void                                                                    \
  76 pixman_composite_##name##_asm_neon (int32_t    w,                       \
  77                                     int32_t    h,                       \
  78                                     dst_type  *dst,                     \
  79                                     int32_t    dst_stride,              \
  80                                     uint32_t   src,                     \
  81                                     int32_t    unused,                  \
  82                                     mask_type *mask,                    \
  83                                     int32_t    mask_stride);            \
  84                                                                         \
  85 static void                                                             \
  86 neon_composite_##name (pixman_implementation_t *imp,                    \
  87                        pixman_op_t              op,                     \
  88                        pixman_image_t *         src_image,              \
  89                        pixman_image_t *         mask_image,             \
  90                        pixman_image_t *         dst_image,              \
  91                        int32_t                  src_x,                  \
  92                        int32_t                  src_y,                  \
  93                        int32_t                  mask_x,                 \
  94                        int32_t                  mask_y,                 \
  95                        int32_t                  dest_x,                 \
  96                        int32_t                  dest_y,                 \
  97                        int32_t                  width,                  \
  98                        int32_t                  height)                 \
  99 {                                                                       \
 100     dst_type  *dst_line;                                                \
 101     mask_type *mask_line;                                               \
 102     int32_t    dst_stride, mask_stride;                                 \
 103     uint32_t   src;                                                     \
 104                                                                         \
 105     src = _pixman_image_get_solid (src_image, dst_image->bits.format);  \
 106                                                                         \
 107     if (src == 0)                                                       \
 108         return;                                                         \
 109                                                                         \
 110     PIXMAN_IMAGE_GET_LINE (dst_image, dest_x, dest_y, dst_type,         \
 111                            dst_stride, dst_line, dst_cnt);              \
 112     PIXMAN_IMAGE_GET_LINE (mask_image, mask_x, mask_y, mask_type,       \
 113                            mask_stride, mask_line, mask_cnt);           \
 114                                                                         \
 115     pixman_composite_##name##_asm_neon (width, height,                  \
 116                                         dst_line, dst_stride,           \
 117                                         src, 0,                         \
 118                                         mask_line, mask_stride);        \
 119 }
 120
 121 #define BIND_SRC_N_DST(name, src_type, src_cnt, dst_type, dst_cnt)      \
 122 void                                                                    \
 123 pixman_composite_##name##_asm_neon (int32_t    w,                       \
 124                                     int32_t    h,                       \
 125                                     dst_type  *dst,                     \
 126                                     int32_t    dst_stride,              \
 127                                     src_type  *src,                     \
 128                                     int32_t    src_stride,              \
 129                                     uint32_t   mask);                   \
 130                                                                         \
 131 static void                                                             \
 132 neon_composite_##name (pixman_implementation_t *imp,                    \
 133                        pixman_op_t              op,                     \
 134                        pixman_image_t *         src_image,              \
 135                        pixman_image_t *         mask_image,             \
 136                        pixman_image_t *         dst_image,              \
 137                        int32_t                  src_x,                  \
 138                        int32_t                  src_y,                  \
 139                        int32_t                  mask_x,                 \
 140                        int32_t                  mask_y,                 \
 141                        int32_t                  dest_x,                 \
 142                        int32_t                  dest_y,                 \
 143                        int32_t                  width,                  \
 144                        int32_t                  height)                 \
 145 {                                                                       \
 146     dst_type  *dst_line;                                                \
 147     src_type  *src_line;                                                \
 148     int32_t    dst_stride, src_stride;                                  \
 149     uint32_t   mask;                                                    \
 150                                                                         \
 151     mask = _pixman_image_get_solid (mask_image, dst_image->bits.format);\
 152                                                                         \
 153     if (mask == 0)                                                      \
 154         return;                                                         \
 155                                                                         \
 156     PIXMAN_IMAGE_GET_LINE (dst_image, dest_x, dest_y, dst_type,         \
 157                            dst_stride, dst_line, dst_cnt);              \
 158     PIXMAN_IMAGE_GET_LINE (src_image, src_x, src_y, src_type,           \
 159                            src_stride, src_line, src_cnt);              \
 160                                                                         \
 161     pixman_composite_##name##_asm_neon (width, height,                  \
 162                                         dst_line, dst_stride,           \
 163                                         src_line, src_stride,           \
 164                                         mask);                          \
 165 }
 166
 167 #define BIND_SRC_MASK_DST(name, src_type, src_cnt, mask_type, mask_cnt, \
 168                           dst_type, dst_cnt)                            \
 169 void                                                                    \
 170 pixman_composite_##name##_asm_neon (int32_t    w,                       \
 171                                     int32_t    h,                       \
 172                                     dst_type  *dst,                     \
 173                                     int32_t    dst_stride,              \
 174                                     src_type  *src,                     \
 175                                     int32_t    src_stride,              \
 176                                     mask_type *mask,                    \
 177                                     int32_t    mask_stride);            \
 178                                                                         \
 179 static void                                                             \
 180 neon_composite_##name (pixman_implementation_t *imp,                    \
 181                        pixman_op_t              op,                     \
 182                        pixman_image_t *         src_image,              \
 183                        pixman_image_t *         mask_image,             \
 184                        pixman_image_t *         dst_image,              \
 185                        int32_t                  src_x,                  \
 186                        int32_t                  src_y,                  \
 187                        int32_t                  mask_x,                 \
 188                        int32_t                  mask_y,                 \
 189                        int32_t                  dest_x,                 \
 190                        int32_t                  dest_y,                 \
 191                        int32_t                  width,                  \
 192                        int32_t                  height)                 \
 193 {                                                                       \
 194     dst_type  *dst_line;                                                \
 195     src_type  *src_line;                                                \
 196     mask_type *mask_line;                                               \
 197     int32_t    dst_stride, src_stride, mask_stride;                     \
 198                                                                         \
 199     PIXMAN_IMAGE_GET_LINE (dst_image, dest_x, dest_y, dst_type,         \
 200                            dst_stride, dst_line, dst_cnt);              \
 201     PIXMAN_IMAGE_GET_LINE (src_image, src_x, src_y, src_type,           \
 202                            src_stride, src_line, src_cnt);              \
 203     PIXMAN_IMAGE_GET_LINE (mask_image, mask_x, mask_y, mask_type,       \
 204                            mask_stride, mask_line, mask_cnt);           \
 205                                                                         \
 206     pixman_composite_##name##_asm_neon (width, height,                  \
 207                                         dst_line, dst_stride,           \
 208                                         src_line, src_stride,           \
 209                                         mask_line, mask_stride);        \
 210 }
 211
 212
 213 BIND_SRC_NULL_DST(src_8888_8888, uint32_t, 1, uint32_t, 1)
 214 BIND_SRC_NULL_DST(src_0565_0565, uint16_t, 1, uint16_t, 1)
 215 BIND_SRC_NULL_DST(src_0888_0888, uint8_t, 3, uint8_t, 3)
 216 BIND_SRC_NULL_DST(src_8888_0565, uint32_t, 1, uint16_t, 1)
 217 BIND_SRC_NULL_DST(add_8000_8000, uint8_t, 1, uint8_t, 1)
 218
 219 BIND_SRC_NULL_DST(over_8888_0565, uint32_t, 1, uint16_t, 1)
 220 BIND_SRC_NULL_DST(over_8888_8888, uint32_t, 1, uint32_t, 1)
 221
 222 BIND_N_MASK_DST(over_n_8_0565, uint8_t, 1, uint16_t, 1)
 223 BIND_N_MASK_DST(over_n_8_8888, uint8_t, 1, uint32_t, 1)
 224 BIND_N_MASK_DST(add_n_8_8, uint8_t, 1, uint8_t, 1)
 225
 226 BIND_SRC_N_DST(over_8888_n_8888, uint32_t, 1, uint32_t, 1)
 227
 228 BIND_SRC_MASK_DST(add_8_8_8, uint8_t, 1, uint8_t, 1, uint8_t, 1)
 229
 230 void
 231 pixman_composite_src_n_8_asm_neon (int32_t   w,
 232                                    int32_t   h,
 233                                    uint8_t  *dst,
 234                                    int32_t   dst_stride,
 235                                    uint8_t   src);
 236
 237 void
 238 pixman_composite_src_n_0565_asm_neon (int32_t   w,
 239                                       int32_t   h,
 240                                       uint16_t *dst,
 241                                       int32_t   dst_stride,
 242                                       uint16_t  src);
 243
 244 void
 245 pixman_composite_src_n_8888_asm_neon (int32_t   w,
 246                                       int32_t   h,
 247                                       uint32_t *dst,
 248                                       int32_t   dst_stride,
 249                                       uint32_t  src);
 250
 251 static pixman_bool_t
 252 pixman_fill_neon (uint32_t *bits,
 253                   int       stride,
 254                   int       bpp,
 255                   int       x,
 256                   int       y,
 257                   int       width,
 258                   int       height,
 259                   uint32_t  _xor)
 260 {
 261     /* stride is always multiple of 32bit units in pixman */
 262     uint32_t byte_stride = stride * sizeof(uint32_t);
 263
 264     switch (bpp)
 265     {
 266     case 8:
 267         pixman_composite_src_n_8_asm_neon (
 268                 width,
 269                 height,
 270                 (uint8_t *)(((char *) bits) + y * byte_stride + x),
 271                 byte_stride,
 272                 _xor & 0xff);
 273         return TRUE;
 274     case 16:
 275         pixman_composite_src_n_0565_asm_neon (
 276                 width,
 277                 height,
 278                 (uint16_t *)(((char *) bits) + y * byte_stride + x * 2),
 279                 byte_stride / 2,
 280                 _xor & 0xffff);
 281         return TRUE;
 282     case 32:
 283         pixman_composite_src_n_8888_asm_neon (
 284                 width,
 285                 height,
 286                 (uint32_t *)(((char *) bits) + y * byte_stride + x * 4),
 287                 byte_stride / 4,
 288                 _xor);
 289         return TRUE;
 290     default:
 291         return FALSE;
 292     }
 293 }
 294
 295 static pixman_bool_t
 296 pixman_blt_neon (uint32_t *src_bits,
 297                  uint32_t *dst_bits,
 298                  int       src_stride,
 299                  int       dst_stride,
 300                  int       src_bpp,
 301                  int       dst_bpp,
 302                  int       src_x,
 303                  int       src_y,
 304                  int       dst_x,
 305                  int       dst_y,
 306                  int       width,
 307                  int       height)
 308 {
 309     if (src_bpp != dst_bpp)
 310         return FALSE;
 311
 312     switch (src_bpp)
 313     {
 314     case 16:
 315         pixman_composite_src_0565_0565_asm_neon (
 316                 width, height,
 317                 (uint16_t *)(((char *) dst_bits) +
 318                 dst_y * dst_stride * 4 + dst_x * 2), dst_stride * 2,
 319                 (uint16_t *)(((char *) src_bits) +
 320                 src_y * src_stride * 4 + src_x * 2), src_stride * 2);
 321         return TRUE;
 322     case 32:
 323         pixman_composite_src_8888_8888_asm_neon (
 324                 width, height,
 325                 (uint32_t *)(((char *) dst_bits) +
 326                 dst_y * dst_stride * 4 + dst_x * 4), dst_stride,
 327                 (uint32_t *)(((char *) src_bits) +
 328                 src_y * src_stride * 4 + src_x * 4), src_stride);
 329         return TRUE;
 330     default:
 331         return FALSE;
 332     }
 333 }
 334
 335 static const pixman_fast_path_t arm_neon_fast_path_array[] =
 336 {
 337     { PIXMAN_OP_SRC,  PIXMAN_r5g6b5,   PIXMAN_null,     PIXMAN_r5g6b5,   neon_composite_src_0565_0565    },
 338     { PIXMAN_OP_SRC,  PIXMAN_b5g6r5,   PIXMAN_null,     PIXMAN_b5g6r5,   neon_composite_src_0565_0565    },
 339     { PIXMAN_OP_SRC,  PIXMAN_a8r8g8b8, PIXMAN_null,     PIXMAN_r5g6b5,   neon_composite_src_8888_0565    },
 340     { PIXMAN_OP_SRC,  PIXMAN_x8r8g8b8, PIXMAN_null,     PIXMAN_r5g6b5,   neon_composite_src_8888_0565    },
 341     { PIXMAN_OP_SRC,  PIXMAN_a8b8g8r8, PIXMAN_null,     PIXMAN_b5g6r5,   neon_composite_src_8888_0565    },
 342     { PIXMAN_OP_SRC,  PIXMAN_x8b8g8r8, PIXMAN_null,     PIXMAN_b5g6r5,   neon_composite_src_8888_0565    },
 343     { PIXMAN_OP_SRC,  PIXMAN_a8r8g8b8, PIXMAN_null,     PIXMAN_x8r8g8b8, neon_composite_src_8888_8888    },
 344     { PIXMAN_OP_SRC,  PIXMAN_x8r8g8b8, PIXMAN_null,     PIXMAN_x8r8g8b8, neon_composite_src_8888_8888    },
 345     { PIXMAN_OP_SRC,  PIXMAN_a8b8g8r8, PIXMAN_null,     PIXMAN_x8b8g8r8, neon_composite_src_8888_8888    },
 346     { PIXMAN_OP_SRC,  PIXMAN_x8b8g8r8, PIXMAN_null,     PIXMAN_x8b8g8r8, neon_composite_src_8888_8888    },
 347     { PIXMAN_OP_SRC,  PIXMAN_r8g8b8,   PIXMAN_null,     PIXMAN_r8g8b8,   neon_composite_src_0888_0888    },
 348     { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_a8,       PIXMAN_r5g6b5,   neon_composite_over_n_8_0565    },
 349     { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_a8,       PIXMAN_b5g6r5,   neon_composite_over_n_8_0565    },
 350     { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_a8,       PIXMAN_a8r8g8b8, neon_composite_over_n_8_8888    },
 351     { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_a8,       PIXMAN_x8r8g8b8, neon_composite_over_n_8_8888    },
 352     { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_a8,       PIXMAN_a8b8g8r8, neon_composite_over_n_8_8888    },
 353     { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_a8,       PIXMAN_x8b8g8r8, neon_composite_over_n_8_8888    },
 354     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_solid,    PIXMAN_a8r8g8b8, neon_composite_over_8888_n_8888 },
 355     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_solid,    PIXMAN_x8r8g8b8, neon_composite_over_8888_n_8888 },
 356     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_null,     PIXMAN_r5g6b5,   neon_composite_over_8888_0565   },
 357     { PIXMAN_OP_OVER, PIXMAN_a8b8g8r8, PIXMAN_null,     PIXMAN_b5g6r5,   neon_composite_over_8888_0565   },
 358     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_null,     PIXMAN_a8r8g8b8, neon_composite_over_8888_8888   },
 359     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_null,     PIXMAN_x8r8g8b8, neon_composite_over_8888_8888   },
 360     { PIXMAN_OP_OVER, PIXMAN_a8b8g8r8, PIXMAN_null,     PIXMAN_a8b8g8r8, neon_composite_over_8888_8888   },
 361     { PIXMAN_OP_OVER, PIXMAN_a8b8g8r8, PIXMAN_null,     PIXMAN_x8b8g8r8, neon_composite_over_8888_8888   },
 362     { PIXMAN_OP_ADD,  PIXMAN_solid,    PIXMAN_a8,       PIXMAN_a8,       neon_composite_add_n_8_8        },
 363     { PIXMAN_OP_ADD,  PIXMAN_a8,       PIXMAN_a8,       PIXMAN_a8,       neon_composite_add_8_8_8        },
 364     { PIXMAN_OP_ADD,  PIXMAN_a8,       PIXMAN_null,     PIXMAN_a8,       neon_composite_add_8000_8000    },
 365     { PIXMAN_OP_NONE },
 366 };
 367
 368 const pixman_fast_path_t *const arm_neon_fast_paths = arm_neon_fast_path_array;
 369
 370 static void
 371 arm_neon_composite (pixman_implementation_t *imp,
 372                     pixman_op_t              op,
 373                     pixman_image_t *         src,
 374                     pixman_image_t *         mask,
 375                     pixman_image_t *         dest,
 376                     int32_t                  src_x,
 377                     int32_t                  src_y,
 378                     int32_t                  mask_x,
 379                     int32_t                  mask_y,
 380                     int32_t                  dest_x,
 381                     int32_t                  dest_y,
 382                     int32_t                  width,
 383                     int32_t                  height)
 384 {
 385     if (_pixman_run_fast_path (arm_neon_fast_paths, imp,
 386                                op, src, mask, dest,
 387                                src_x, src_y,
 388                                mask_x, mask_y,
 389                                dest_x, dest_y,
 390                                width, height))
 391     {
 392         return;
 393     }
 394
 395     _pixman_implementation_composite (imp->delegate, op,
 396                                       src, mask, dest,
 397                                       src_x, src_y,
 398                                       mask_x, mask_y,
 399                                       dest_x, dest_y,
 400                                       width, height);
 401 }
 402
 403 static pixman_bool_t
 404 arm_neon_blt (pixman_implementation_t *imp,
 405               uint32_t *               src_bits,
 406               uint32_t *               dst_bits,
 407               int                      src_stride,
 408               int                      dst_stride,
 409               int                      src_bpp,
 410               int                      dst_bpp,
 411               int                      src_x,
 412               int                      src_y,
 413               int                      dst_x,
 414               int                      dst_y,
 415               int                      width,
 416               int                      height)
 417 {
 418     if (!pixman_blt_neon (
 419             src_bits, dst_bits, src_stride, dst_stride, src_bpp, dst_bpp,
 420             src_x, src_y, dst_x, dst_y, width, height))
 421
 422     {
 423         return _pixman_implementation_blt (
 424             imp->delegate,
 425             src_bits, dst_bits, src_stride, dst_stride, src_bpp, dst_bpp,
 426             src_x, src_y, dst_x, dst_y, width, height);
 427     }
 428
 429     return TRUE;
 430 }
 431
 432 static pixman_bool_t
 433 arm_neon_fill (pixman_implementation_t *imp,
 434                uint32_t *               bits,
 435                int                      stride,
 436                int                      bpp,
 437                int                      x,
 438                int                      y,
 439                int                      width,
 440                int                      height,
 441                uint32_t xor)
 442 {
 443     if (pixman_fill_neon (bits, stride, bpp, x, y, width, height, xor))
 444         return TRUE;
 445
 446     return _pixman_implementation_fill (
 447         imp->delegate, bits, stride, bpp, x, y, width, height, xor);
 448 }
 449
 450 pixman_implementation_t *
 451 _pixman_implementation_create_arm_neon (void)
 452 {
 453     pixman_implementation_t *general = _pixman_implementation_create_fast_path ();
 454     pixman_implementation_t *imp = _pixman_implementation_create (general);
 455
 456     imp->composite = arm_neon_composite;
 457     imp->blt = arm_neon_blt;
 458     imp->fill = arm_neon_fill;
 459
 460     return imp;
 461 }