gfx/cairo/libpixman/src/pixman-inlines.h

Thu, 22 Jan 2015 13:21:57 +0100

author
Michael Schloh von Bennewitz <michael@schloh.com>
date
Thu, 22 Jan 2015 13:21:57 +0100
branch
TOR_BUG_9701
changeset 15
b8a032363ba2
permissions
-rw-r--r--

Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6

michael@0 1 /* -*- Mode: c; c-basic-offset: 4; tab-width: 8; indent-tabs-mode: t; -*- */
michael@0 2 /*
michael@0 3 * Copyright © 2000 SuSE, Inc.
michael@0 4 * Copyright © 2007 Red Hat, Inc.
michael@0 5 *
michael@0 6 * Permission to use, copy, modify, distribute, and sell this software and its
michael@0 7 * documentation for any purpose is hereby granted without fee, provided that
michael@0 8 * the above copyright notice appear in all copies and that both that
michael@0 9 * copyright notice and this permission notice appear in supporting
michael@0 10 * documentation, and that the name of SuSE not be used in advertising or
michael@0 11 * publicity pertaining to distribution of the software without specific,
michael@0 12 * written prior permission. SuSE makes no representations about the
michael@0 13 * suitability of this software for any purpose. It is provided "as is"
michael@0 14 * without express or implied warranty.
michael@0 15 *
michael@0 16 * SuSE DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, INCLUDING ALL
michael@0 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN NO EVENT SHALL SuSE
michael@0 18 * BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
michael@0 19 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION
michael@0 20 * OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN
michael@0 21 * CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
michael@0 22 *
michael@0 23 * Author: Keith Packard, SuSE, Inc.
michael@0 24 */
michael@0 25
michael@0 26 #ifndef PIXMAN_FAST_PATH_H__
michael@0 27 #define PIXMAN_FAST_PATH_H__
michael@0 28
michael@0 29 #include "pixman-private.h"
michael@0 30
michael@0 31 #define PIXMAN_REPEAT_COVER -1
michael@0 32
michael@0 33 /* Flags describing input parameters to fast path macro template.
michael@0 34 * Turning on some flag values may indicate that
michael@0 35 * "some property X is available so template can use this" or
michael@0 36 * "some property X should be handled by template".
michael@0 37 *
michael@0 38 * FLAG_HAVE_SOLID_MASK
michael@0 39 * Input mask is solid so template should handle this.
michael@0 40 *
michael@0 41 * FLAG_HAVE_NON_SOLID_MASK
michael@0 42 * Input mask is bits mask so template should handle this.
michael@0 43 *
michael@0 44 * FLAG_HAVE_SOLID_MASK and FLAG_HAVE_NON_SOLID_MASK are mutually
michael@0 45 * exclusive. (It's not allowed to turn both flags on)
michael@0 46 */
michael@0 47 #define FLAG_NONE (0)
michael@0 48 #define FLAG_HAVE_SOLID_MASK (1 << 1)
michael@0 49 #define FLAG_HAVE_NON_SOLID_MASK (1 << 2)
michael@0 50
michael@0 51 /* To avoid too short repeated scanline function calls, extend source
michael@0 52 * scanlines having width less than below constant value.
michael@0 53 */
michael@0 54 #define REPEAT_NORMAL_MIN_WIDTH 64
michael@0 55
michael@0 56 static force_inline pixman_bool_t
michael@0 57 repeat (pixman_repeat_t repeat, int *c, int size)
michael@0 58 {
michael@0 59 if (repeat == PIXMAN_REPEAT_NONE)
michael@0 60 {
michael@0 61 if (*c < 0 || *c >= size)
michael@0 62 return FALSE;
michael@0 63 }
michael@0 64 else if (repeat == PIXMAN_REPEAT_NORMAL)
michael@0 65 {
michael@0 66 while (*c >= size)
michael@0 67 *c -= size;
michael@0 68 while (*c < 0)
michael@0 69 *c += size;
michael@0 70 }
michael@0 71 else if (repeat == PIXMAN_REPEAT_PAD)
michael@0 72 {
michael@0 73 *c = CLIP (*c, 0, size - 1);
michael@0 74 }
michael@0 75 else /* REFLECT */
michael@0 76 {
michael@0 77 *c = MOD (*c, size * 2);
michael@0 78 if (*c >= size)
michael@0 79 *c = size * 2 - *c - 1;
michael@0 80 }
michael@0 81 return TRUE;
michael@0 82 }
michael@0 83
michael@0 84 static force_inline int
michael@0 85 pixman_fixed_to_bilinear_weight (pixman_fixed_t x)
michael@0 86 {
michael@0 87 return (x >> (16 - BILINEAR_INTERPOLATION_BITS)) &
michael@0 88 ((1 << BILINEAR_INTERPOLATION_BITS) - 1);
michael@0 89 }
michael@0 90
michael@0 91 #if BILINEAR_INTERPOLATION_BITS <= 4
michael@0 92 /* Inspired by Filter_32_opaque from Skia */
michael@0 93 static force_inline uint32_t
michael@0 94 bilinear_interpolation (uint32_t tl, uint32_t tr,
michael@0 95 uint32_t bl, uint32_t br,
michael@0 96 int distx, int disty)
michael@0 97 {
michael@0 98 int distxy, distxiy, distixy, distixiy;
michael@0 99 uint32_t lo, hi;
michael@0 100
michael@0 101 distx <<= (4 - BILINEAR_INTERPOLATION_BITS);
michael@0 102 disty <<= (4 - BILINEAR_INTERPOLATION_BITS);
michael@0 103
michael@0 104 distxy = distx * disty;
michael@0 105 distxiy = (distx << 4) - distxy; /* distx * (16 - disty) */
michael@0 106 distixy = (disty << 4) - distxy; /* disty * (16 - distx) */
michael@0 107 distixiy =
michael@0 108 16 * 16 - (disty << 4) -
michael@0 109 (distx << 4) + distxy; /* (16 - distx) * (16 - disty) */
michael@0 110
michael@0 111 lo = (tl & 0xff00ff) * distixiy;
michael@0 112 hi = ((tl >> 8) & 0xff00ff) * distixiy;
michael@0 113
michael@0 114 lo += (tr & 0xff00ff) * distxiy;
michael@0 115 hi += ((tr >> 8) & 0xff00ff) * distxiy;
michael@0 116
michael@0 117 lo += (bl & 0xff00ff) * distixy;
michael@0 118 hi += ((bl >> 8) & 0xff00ff) * distixy;
michael@0 119
michael@0 120 lo += (br & 0xff00ff) * distxy;
michael@0 121 hi += ((br >> 8) & 0xff00ff) * distxy;
michael@0 122
michael@0 123 return ((lo >> 8) & 0xff00ff) | (hi & ~0xff00ff);
michael@0 124 }
michael@0 125
michael@0 126 #else
michael@0 127 #if SIZEOF_LONG > 4
michael@0 128
michael@0 129 static force_inline uint32_t
michael@0 130 bilinear_interpolation (uint32_t tl, uint32_t tr,
michael@0 131 uint32_t bl, uint32_t br,
michael@0 132 int distx, int disty)
michael@0 133 {
michael@0 134 uint64_t distxy, distxiy, distixy, distixiy;
michael@0 135 uint64_t tl64, tr64, bl64, br64;
michael@0 136 uint64_t f, r;
michael@0 137
michael@0 138 distx <<= (8 - BILINEAR_INTERPOLATION_BITS);
michael@0 139 disty <<= (8 - BILINEAR_INTERPOLATION_BITS);
michael@0 140
michael@0 141 distxy = distx * disty;
michael@0 142 distxiy = distx * (256 - disty);
michael@0 143 distixy = (256 - distx) * disty;
michael@0 144 distixiy = (256 - distx) * (256 - disty);
michael@0 145
michael@0 146 /* Alpha and Blue */
michael@0 147 tl64 = tl & 0xff0000ff;
michael@0 148 tr64 = tr & 0xff0000ff;
michael@0 149 bl64 = bl & 0xff0000ff;
michael@0 150 br64 = br & 0xff0000ff;
michael@0 151
michael@0 152 f = tl64 * distixiy + tr64 * distxiy + bl64 * distixy + br64 * distxy;
michael@0 153 r = f & 0x0000ff0000ff0000ull;
michael@0 154
michael@0 155 /* Red and Green */
michael@0 156 tl64 = tl;
michael@0 157 tl64 = ((tl64 << 16) & 0x000000ff00000000ull) | (tl64 & 0x0000ff00ull);
michael@0 158
michael@0 159 tr64 = tr;
michael@0 160 tr64 = ((tr64 << 16) & 0x000000ff00000000ull) | (tr64 & 0x0000ff00ull);
michael@0 161
michael@0 162 bl64 = bl;
michael@0 163 bl64 = ((bl64 << 16) & 0x000000ff00000000ull) | (bl64 & 0x0000ff00ull);
michael@0 164
michael@0 165 br64 = br;
michael@0 166 br64 = ((br64 << 16) & 0x000000ff00000000ull) | (br64 & 0x0000ff00ull);
michael@0 167
michael@0 168 f = tl64 * distixiy + tr64 * distxiy + bl64 * distixy + br64 * distxy;
michael@0 169 r |= ((f >> 16) & 0x000000ff00000000ull) | (f & 0xff000000ull);
michael@0 170
michael@0 171 return (uint32_t)(r >> 16);
michael@0 172 }
michael@0 173
michael@0 174 #else
michael@0 175
michael@0 176 #ifdef LOW_QUALITY_INTERPOLATION
michael@0 177 /* Based on Filter_32_opaque_portable from Skia */
michael@0 178 static force_inline uint32_t
michael@0 179 bilinear_interpolation(uint32_t a00, uint32_t a01,
michael@0 180 uint32_t a10, uint32_t a11,
michael@0 181 int x, int y)
michael@0 182 {
michael@0 183 int xy = x * y;
michael@0 184 static const uint32_t mask = 0xff00ff;
michael@0 185
michael@0 186 int scale = 256 - 16*y - 16*x + xy;
michael@0 187 uint32_t lo = (a00 & mask) * scale;
michael@0 188 uint32_t hi = ((a00 >> 8) & mask) * scale;
michael@0 189
michael@0 190 scale = 16*x - xy;
michael@0 191 lo += (a01 & mask) * scale;
michael@0 192 hi += ((a01 >> 8) & mask) * scale;
michael@0 193
michael@0 194 scale = 16*y - xy;
michael@0 195 lo += (a10 & mask) * scale;
michael@0 196 hi += ((a10 >> 8) & mask) * scale;
michael@0 197
michael@0 198 lo += (a11 & mask) * xy;
michael@0 199 hi += ((a11 >> 8) & mask) * xy;
michael@0 200
michael@0 201 return ((lo >> 8) & mask) | (hi & ~mask);
michael@0 202 }
michael@0 203 #else
michael@0 204 static force_inline uint32_t
michael@0 205 bilinear_interpolation (uint32_t tl, uint32_t tr,
michael@0 206 uint32_t bl, uint32_t br,
michael@0 207 int distx, int disty)
michael@0 208 {
michael@0 209 int distxy, distxiy, distixy, distixiy;
michael@0 210 uint32_t f, r;
michael@0 211
michael@0 212 distx <<= (8 - BILINEAR_INTERPOLATION_BITS);
michael@0 213 disty <<= (8 - BILINEAR_INTERPOLATION_BITS);
michael@0 214
michael@0 215 distxy = distx * disty;
michael@0 216 distxiy = (distx << 8) - distxy; /* distx * (256 - disty) */
michael@0 217 distixy = (disty << 8) - distxy; /* disty * (256 - distx) */
michael@0 218 distixiy =
michael@0 219 256 * 256 - (disty << 8) -
michael@0 220 (distx << 8) + distxy; /* (256 - distx) * (256 - disty) */
michael@0 221
michael@0 222 /* Blue */
michael@0 223 r = (tl & 0x000000ff) * distixiy + (tr & 0x000000ff) * distxiy
michael@0 224 + (bl & 0x000000ff) * distixy + (br & 0x000000ff) * distxy;
michael@0 225
michael@0 226 /* Green */
michael@0 227 f = (tl & 0x0000ff00) * distixiy + (tr & 0x0000ff00) * distxiy
michael@0 228 + (bl & 0x0000ff00) * distixy + (br & 0x0000ff00) * distxy;
michael@0 229 r |= f & 0xff000000;
michael@0 230
michael@0 231 tl >>= 16;
michael@0 232 tr >>= 16;
michael@0 233 bl >>= 16;
michael@0 234 br >>= 16;
michael@0 235 r >>= 16;
michael@0 236
michael@0 237 /* Red */
michael@0 238 f = (tl & 0x000000ff) * distixiy + (tr & 0x000000ff) * distxiy
michael@0 239 + (bl & 0x000000ff) * distixy + (br & 0x000000ff) * distxy;
michael@0 240 r |= f & 0x00ff0000;
michael@0 241
michael@0 242 /* Alpha */
michael@0 243 f = (tl & 0x0000ff00) * distixiy + (tr & 0x0000ff00) * distxiy
michael@0 244 + (bl & 0x0000ff00) * distixy + (br & 0x0000ff00) * distxy;
michael@0 245 r |= f & 0xff000000;
michael@0 246
michael@0 247 return r;
michael@0 248 }
michael@0 249 #endif
michael@0 250 #endif
michael@0 251 #endif // BILINEAR_INTERPOLATION_BITS <= 4
michael@0 252
michael@0 253 /*
michael@0 254 * For each scanline fetched from source image with PAD repeat:
michael@0 255 * - calculate how many pixels need to be padded on the left side
michael@0 256 * - calculate how many pixels need to be padded on the right side
michael@0 257 * - update width to only count pixels which are fetched from the image
michael@0 258 * All this information is returned via 'width', 'left_pad', 'right_pad'
michael@0 259 * arguments. The code is assuming that 'unit_x' is positive.
michael@0 260 *
michael@0 261 * Note: 64-bit math is used in order to avoid potential overflows, which
michael@0 262 * is probably excessive in many cases. This particular function
michael@0 263 * may need its own correctness test and performance tuning.
michael@0 264 */
michael@0 265 static force_inline void
michael@0 266 pad_repeat_get_scanline_bounds (int32_t source_image_width,
michael@0 267 pixman_fixed_t vx,
michael@0 268 pixman_fixed_t unit_x,
michael@0 269 int32_t * width,
michael@0 270 int32_t * left_pad,
michael@0 271 int32_t * right_pad)
michael@0 272 {
michael@0 273 int64_t max_vx = (int64_t) source_image_width << 16;
michael@0 274 int64_t tmp;
michael@0 275 if (vx < 0)
michael@0 276 {
michael@0 277 tmp = ((int64_t) unit_x - 1 - vx) / unit_x;
michael@0 278 if (tmp > *width)
michael@0 279 {
michael@0 280 *left_pad = *width;
michael@0 281 *width = 0;
michael@0 282 }
michael@0 283 else
michael@0 284 {
michael@0 285 *left_pad = (int32_t) tmp;
michael@0 286 *width -= (int32_t) tmp;
michael@0 287 }
michael@0 288 }
michael@0 289 else
michael@0 290 {
michael@0 291 *left_pad = 0;
michael@0 292 }
michael@0 293 tmp = ((int64_t) unit_x - 1 - vx + max_vx) / unit_x - *left_pad;
michael@0 294 if (tmp < 0)
michael@0 295 {
michael@0 296 *right_pad = *width;
michael@0 297 *width = 0;
michael@0 298 }
michael@0 299 else if (tmp >= *width)
michael@0 300 {
michael@0 301 *right_pad = 0;
michael@0 302 }
michael@0 303 else
michael@0 304 {
michael@0 305 *right_pad = *width - (int32_t) tmp;
michael@0 306 *width = (int32_t) tmp;
michael@0 307 }
michael@0 308 }
michael@0 309
michael@0 310 /* A macroified version of specialized nearest scalers for some
michael@0 311 * common 8888 and 565 formats. It supports SRC and OVER ops.
michael@0 312 *
michael@0 313 * There are two repeat versions, one that handles repeat normal,
michael@0 314 * and one without repeat handling that only works if the src region
michael@0 315 * used is completely covered by the pre-repeated source samples.
michael@0 316 *
michael@0 317 * The loops are unrolled to process two pixels per iteration for better
michael@0 318 * performance on most CPU architectures (superscalar processors
michael@0 319 * can issue several operations simultaneously, other processors can hide
michael@0 320 * instructions latencies by pipelining operations). Unrolling more
michael@0 321 * does not make much sense because the compiler will start running out
michael@0 322 * of spare registers soon.
michael@0 323 */
michael@0 324
michael@0 325 #define GET_8888_ALPHA(s) ((s) >> 24)
michael@0 326 /* This is not actually used since we don't have an OVER with
michael@0 327 565 source, but it is needed to build. */
michael@0 328 #define GET_0565_ALPHA(s) 0xff
michael@0 329 #define GET_x888_ALPHA(s) 0xff
michael@0 330
michael@0 331 #define FAST_NEAREST_SCANLINE(scanline_func_name, SRC_FORMAT, DST_FORMAT, \
michael@0 332 src_type_t, dst_type_t, OP, repeat_mode) \
michael@0 333 static force_inline void \
michael@0 334 scanline_func_name (dst_type_t *dst, \
michael@0 335 const src_type_t *src, \
michael@0 336 int32_t w, \
michael@0 337 pixman_fixed_t vx, \
michael@0 338 pixman_fixed_t unit_x, \
michael@0 339 pixman_fixed_t src_width_fixed, \
michael@0 340 pixman_bool_t fully_transparent_src) \
michael@0 341 { \
michael@0 342 uint32_t d; \
michael@0 343 src_type_t s1, s2; \
michael@0 344 uint8_t a1, a2; \
michael@0 345 int x1, x2; \
michael@0 346 \
michael@0 347 if (PIXMAN_OP_ ## OP == PIXMAN_OP_OVER && fully_transparent_src) \
michael@0 348 return; \
michael@0 349 \
michael@0 350 if (PIXMAN_OP_ ## OP != PIXMAN_OP_SRC && PIXMAN_OP_ ## OP != PIXMAN_OP_OVER) \
michael@0 351 abort(); \
michael@0 352 \
michael@0 353 while ((w -= 2) >= 0) \
michael@0 354 { \
michael@0 355 x1 = pixman_fixed_to_int (vx); \
michael@0 356 vx += unit_x; \
michael@0 357 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NORMAL) \
michael@0 358 { \
michael@0 359 /* This works because we know that unit_x is positive */ \
michael@0 360 while (vx >= 0) \
michael@0 361 vx -= src_width_fixed; \
michael@0 362 } \
michael@0 363 s1 = *(src + x1); \
michael@0 364 \
michael@0 365 x2 = pixman_fixed_to_int (vx); \
michael@0 366 vx += unit_x; \
michael@0 367 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NORMAL) \
michael@0 368 { \
michael@0 369 /* This works because we know that unit_x is positive */ \
michael@0 370 while (vx >= 0) \
michael@0 371 vx -= src_width_fixed; \
michael@0 372 } \
michael@0 373 s2 = *(src + x2); \
michael@0 374 \
michael@0 375 if (PIXMAN_OP_ ## OP == PIXMAN_OP_OVER) \
michael@0 376 { \
michael@0 377 a1 = GET_ ## SRC_FORMAT ## _ALPHA(s1); \
michael@0 378 a2 = GET_ ## SRC_FORMAT ## _ALPHA(s2); \
michael@0 379 \
michael@0 380 if (a1 == 0xff) \
michael@0 381 { \
michael@0 382 *dst = convert_ ## SRC_FORMAT ## _to_ ## DST_FORMAT (s1); \
michael@0 383 } \
michael@0 384 else if (s1) \
michael@0 385 { \
michael@0 386 d = convert_ ## DST_FORMAT ## _to_8888 (*dst); \
michael@0 387 s1 = convert_ ## SRC_FORMAT ## _to_8888 (s1); \
michael@0 388 a1 ^= 0xff; \
michael@0 389 UN8x4_MUL_UN8_ADD_UN8x4 (d, a1, s1); \
michael@0 390 *dst = convert_8888_to_ ## DST_FORMAT (d); \
michael@0 391 } \
michael@0 392 dst++; \
michael@0 393 \
michael@0 394 if (a2 == 0xff) \
michael@0 395 { \
michael@0 396 *dst = convert_ ## SRC_FORMAT ## _to_ ## DST_FORMAT (s2); \
michael@0 397 } \
michael@0 398 else if (s2) \
michael@0 399 { \
michael@0 400 d = convert_## DST_FORMAT ## _to_8888 (*dst); \
michael@0 401 s2 = convert_## SRC_FORMAT ## _to_8888 (s2); \
michael@0 402 a2 ^= 0xff; \
michael@0 403 UN8x4_MUL_UN8_ADD_UN8x4 (d, a2, s2); \
michael@0 404 *dst = convert_8888_to_ ## DST_FORMAT (d); \
michael@0 405 } \
michael@0 406 dst++; \
michael@0 407 } \
michael@0 408 else /* PIXMAN_OP_SRC */ \
michael@0 409 { \
michael@0 410 *dst++ = convert_ ## SRC_FORMAT ## _to_ ## DST_FORMAT (s1); \
michael@0 411 *dst++ = convert_ ## SRC_FORMAT ## _to_ ## DST_FORMAT (s2); \
michael@0 412 } \
michael@0 413 } \
michael@0 414 \
michael@0 415 if (w & 1) \
michael@0 416 { \
michael@0 417 x1 = pixman_fixed_to_int (vx); \
michael@0 418 s1 = *(src + x1); \
michael@0 419 \
michael@0 420 if (PIXMAN_OP_ ## OP == PIXMAN_OP_OVER) \
michael@0 421 { \
michael@0 422 a1 = GET_ ## SRC_FORMAT ## _ALPHA(s1); \
michael@0 423 \
michael@0 424 if (a1 == 0xff) \
michael@0 425 { \
michael@0 426 *dst = convert_ ## SRC_FORMAT ## _to_ ## DST_FORMAT (s1); \
michael@0 427 } \
michael@0 428 else if (s1) \
michael@0 429 { \
michael@0 430 d = convert_## DST_FORMAT ## _to_8888 (*dst); \
michael@0 431 s1 = convert_ ## SRC_FORMAT ## _to_8888 (s1); \
michael@0 432 a1 ^= 0xff; \
michael@0 433 UN8x4_MUL_UN8_ADD_UN8x4 (d, a1, s1); \
michael@0 434 *dst = convert_8888_to_ ## DST_FORMAT (d); \
michael@0 435 } \
michael@0 436 dst++; \
michael@0 437 } \
michael@0 438 else /* PIXMAN_OP_SRC */ \
michael@0 439 { \
michael@0 440 *dst++ = convert_ ## SRC_FORMAT ## _to_ ## DST_FORMAT (s1); \
michael@0 441 } \
michael@0 442 } \
michael@0 443 }
michael@0 444
michael@0 445 #define FAST_NEAREST_MAINLOOP_INT(scale_func_name, scanline_func, src_type_t, mask_type_t, \
michael@0 446 dst_type_t, repeat_mode, have_mask, mask_is_solid) \
michael@0 447 static void \
michael@0 448 fast_composite_scaled_nearest ## scale_func_name (pixman_implementation_t *imp, \
michael@0 449 pixman_composite_info_t *info) \
michael@0 450 { \
michael@0 451 PIXMAN_COMPOSITE_ARGS (info); \
michael@0 452 dst_type_t *dst_line; \
michael@0 453 mask_type_t *mask_line; \
michael@0 454 src_type_t *src_first_line; \
michael@0 455 int y; \
michael@0 456 pixman_fixed_t src_width_fixed = pixman_int_to_fixed (src_image->bits.width); \
michael@0 457 pixman_fixed_t max_vy; \
michael@0 458 pixman_vector_t v; \
michael@0 459 pixman_fixed_t vx, vy; \
michael@0 460 pixman_fixed_t unit_x, unit_y; \
michael@0 461 int32_t left_pad, right_pad; \
michael@0 462 \
michael@0 463 src_type_t *src; \
michael@0 464 dst_type_t *dst; \
michael@0 465 mask_type_t solid_mask; \
michael@0 466 const mask_type_t *mask = &solid_mask; \
michael@0 467 int src_stride, mask_stride, dst_stride; \
michael@0 468 \
michael@0 469 PIXMAN_IMAGE_GET_LINE (dest_image, dest_x, dest_y, dst_type_t, dst_stride, dst_line, 1); \
michael@0 470 if (have_mask) \
michael@0 471 { \
michael@0 472 if (mask_is_solid) \
michael@0 473 solid_mask = _pixman_image_get_solid (imp, mask_image, dest_image->bits.format); \
michael@0 474 else \
michael@0 475 PIXMAN_IMAGE_GET_LINE (mask_image, mask_x, mask_y, mask_type_t, \
michael@0 476 mask_stride, mask_line, 1); \
michael@0 477 } \
michael@0 478 /* pass in 0 instead of src_x and src_y because src_x and src_y need to be \
michael@0 479 * transformed from destination space to source space */ \
michael@0 480 PIXMAN_IMAGE_GET_LINE (src_image, 0, 0, src_type_t, src_stride, src_first_line, 1); \
michael@0 481 \
michael@0 482 /* reference point is the center of the pixel */ \
michael@0 483 v.vector[0] = pixman_int_to_fixed (src_x) + pixman_fixed_1 / 2; \
michael@0 484 v.vector[1] = pixman_int_to_fixed (src_y) + pixman_fixed_1 / 2; \
michael@0 485 v.vector[2] = pixman_fixed_1; \
michael@0 486 \
michael@0 487 if (!pixman_transform_point_3d (src_image->common.transform, &v)) \
michael@0 488 return; \
michael@0 489 \
michael@0 490 unit_x = src_image->common.transform->matrix[0][0]; \
michael@0 491 unit_y = src_image->common.transform->matrix[1][1]; \
michael@0 492 \
michael@0 493 /* Round down to closest integer, ensuring that 0.5 rounds to 0, not 1 */ \
michael@0 494 v.vector[0] -= pixman_fixed_e; \
michael@0 495 v.vector[1] -= pixman_fixed_e; \
michael@0 496 \
michael@0 497 vx = v.vector[0]; \
michael@0 498 vy = v.vector[1]; \
michael@0 499 \
michael@0 500 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NORMAL) \
michael@0 501 { \
michael@0 502 max_vy = pixman_int_to_fixed (src_image->bits.height); \
michael@0 503 \
michael@0 504 /* Clamp repeating positions inside the actual samples */ \
michael@0 505 repeat (PIXMAN_REPEAT_NORMAL, &vx, src_width_fixed); \
michael@0 506 repeat (PIXMAN_REPEAT_NORMAL, &vy, max_vy); \
michael@0 507 } \
michael@0 508 \
michael@0 509 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_PAD || \
michael@0 510 PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NONE) \
michael@0 511 { \
michael@0 512 pad_repeat_get_scanline_bounds (src_image->bits.width, vx, unit_x, \
michael@0 513 &width, &left_pad, &right_pad); \
michael@0 514 vx += left_pad * unit_x; \
michael@0 515 } \
michael@0 516 \
michael@0 517 while (--height >= 0) \
michael@0 518 { \
michael@0 519 dst = dst_line; \
michael@0 520 dst_line += dst_stride; \
michael@0 521 if (have_mask && !mask_is_solid) \
michael@0 522 { \
michael@0 523 mask = mask_line; \
michael@0 524 mask_line += mask_stride; \
michael@0 525 } \
michael@0 526 \
michael@0 527 y = pixman_fixed_to_int (vy); \
michael@0 528 vy += unit_y; \
michael@0 529 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NORMAL) \
michael@0 530 repeat (PIXMAN_REPEAT_NORMAL, &vy, max_vy); \
michael@0 531 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_PAD) \
michael@0 532 { \
michael@0 533 repeat (PIXMAN_REPEAT_PAD, &y, src_image->bits.height); \
michael@0 534 src = src_first_line + src_stride * y; \
michael@0 535 if (left_pad > 0) \
michael@0 536 { \
michael@0 537 scanline_func (mask, dst, \
michael@0 538 src + src_image->bits.width - src_image->bits.width + 1, \
michael@0 539 left_pad, -pixman_fixed_e, 0, src_width_fixed, FALSE); \
michael@0 540 } \
michael@0 541 if (width > 0) \
michael@0 542 { \
michael@0 543 scanline_func (mask + (mask_is_solid ? 0 : left_pad), \
michael@0 544 dst + left_pad, src + src_image->bits.width, width, \
michael@0 545 vx - src_width_fixed, unit_x, src_width_fixed, FALSE); \
michael@0 546 } \
michael@0 547 if (right_pad > 0) \
michael@0 548 { \
michael@0 549 scanline_func (mask + (mask_is_solid ? 0 : left_pad + width), \
michael@0 550 dst + left_pad + width, src + src_image->bits.width, \
michael@0 551 right_pad, -pixman_fixed_e, 0, src_width_fixed, FALSE); \
michael@0 552 } \
michael@0 553 } \
michael@0 554 else if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NONE) \
michael@0 555 { \
michael@0 556 static const src_type_t zero[1] = { 0 }; \
michael@0 557 if (y < 0 || y >= src_image->bits.height) \
michael@0 558 { \
michael@0 559 scanline_func (mask, dst, zero + 1, left_pad + width + right_pad, \
michael@0 560 -pixman_fixed_e, 0, src_width_fixed, TRUE); \
michael@0 561 continue; \
michael@0 562 } \
michael@0 563 src = src_first_line + src_stride * y; \
michael@0 564 if (left_pad > 0) \
michael@0 565 { \
michael@0 566 scanline_func (mask, dst, zero + 1, left_pad, \
michael@0 567 -pixman_fixed_e, 0, src_width_fixed, TRUE); \
michael@0 568 } \
michael@0 569 if (width > 0) \
michael@0 570 { \
michael@0 571 scanline_func (mask + (mask_is_solid ? 0 : left_pad), \
michael@0 572 dst + left_pad, src + src_image->bits.width, width, \
michael@0 573 vx - src_width_fixed, unit_x, src_width_fixed, FALSE); \
michael@0 574 } \
michael@0 575 if (right_pad > 0) \
michael@0 576 { \
michael@0 577 scanline_func (mask + (mask_is_solid ? 0 : left_pad + width), \
michael@0 578 dst + left_pad + width, zero + 1, right_pad, \
michael@0 579 -pixman_fixed_e, 0, src_width_fixed, TRUE); \
michael@0 580 } \
michael@0 581 } \
michael@0 582 else \
michael@0 583 { \
michael@0 584 src = src_first_line + src_stride * y; \
michael@0 585 scanline_func (mask, dst, src + src_image->bits.width, width, vx - src_width_fixed, \
michael@0 586 unit_x, src_width_fixed, FALSE); \
michael@0 587 } \
michael@0 588 } \
michael@0 589 }
michael@0 590
michael@0 591 /* A workaround for old sun studio, see: https://bugs.freedesktop.org/show_bug.cgi?id=32764 */
michael@0 592 #define FAST_NEAREST_MAINLOOP_COMMON(scale_func_name, scanline_func, src_type_t, mask_type_t, \
michael@0 593 dst_type_t, repeat_mode, have_mask, mask_is_solid) \
michael@0 594 FAST_NEAREST_MAINLOOP_INT(_ ## scale_func_name, scanline_func, src_type_t, mask_type_t, \
michael@0 595 dst_type_t, repeat_mode, have_mask, mask_is_solid)
michael@0 596
michael@0 597 #define FAST_NEAREST_MAINLOOP_NOMASK(scale_func_name, scanline_func, src_type_t, dst_type_t, \
michael@0 598 repeat_mode) \
michael@0 599 static force_inline void \
michael@0 600 scanline_func##scale_func_name##_wrapper ( \
michael@0 601 const uint8_t *mask, \
michael@0 602 dst_type_t *dst, \
michael@0 603 const src_type_t *src, \
michael@0 604 int32_t w, \
michael@0 605 pixman_fixed_t vx, \
michael@0 606 pixman_fixed_t unit_x, \
michael@0 607 pixman_fixed_t max_vx, \
michael@0 608 pixman_bool_t fully_transparent_src) \
michael@0 609 { \
michael@0 610 scanline_func (dst, src, w, vx, unit_x, max_vx, fully_transparent_src); \
michael@0 611 } \
michael@0 612 FAST_NEAREST_MAINLOOP_INT (scale_func_name, scanline_func##scale_func_name##_wrapper, \
michael@0 613 src_type_t, uint8_t, dst_type_t, repeat_mode, FALSE, FALSE)
michael@0 614
michael@0 615 #define FAST_NEAREST_MAINLOOP(scale_func_name, scanline_func, src_type_t, dst_type_t, \
michael@0 616 repeat_mode) \
michael@0 617 FAST_NEAREST_MAINLOOP_NOMASK(_ ## scale_func_name, scanline_func, src_type_t, \
michael@0 618 dst_type_t, repeat_mode)
michael@0 619
michael@0 620 #define FAST_NEAREST(scale_func_name, SRC_FORMAT, DST_FORMAT, \
michael@0 621 src_type_t, dst_type_t, OP, repeat_mode) \
michael@0 622 FAST_NEAREST_SCANLINE(scaled_nearest_scanline_ ## scale_func_name ## _ ## OP, \
michael@0 623 SRC_FORMAT, DST_FORMAT, src_type_t, dst_type_t, \
michael@0 624 OP, repeat_mode) \
michael@0 625 FAST_NEAREST_MAINLOOP_NOMASK(_ ## scale_func_name ## _ ## OP, \
michael@0 626 scaled_nearest_scanline_ ## scale_func_name ## _ ## OP, \
michael@0 627 src_type_t, dst_type_t, repeat_mode)
michael@0 628
michael@0 629
michael@0 630 #define SCALED_NEAREST_FLAGS \
michael@0 631 (FAST_PATH_SCALE_TRANSFORM | \
michael@0 632 FAST_PATH_NO_ALPHA_MAP | \
michael@0 633 FAST_PATH_NEAREST_FILTER | \
michael@0 634 FAST_PATH_NO_ACCESSORS | \
michael@0 635 FAST_PATH_NARROW_FORMAT)
michael@0 636
michael@0 637 #define SIMPLE_NEAREST_FAST_PATH_NORMAL(op,s,d,func) \
michael@0 638 { PIXMAN_OP_ ## op, \
michael@0 639 PIXMAN_ ## s, \
michael@0 640 (SCALED_NEAREST_FLAGS | \
michael@0 641 FAST_PATH_NORMAL_REPEAT | \
michael@0 642 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 643 PIXMAN_null, 0, \
michael@0 644 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 645 fast_composite_scaled_nearest_ ## func ## _normal ## _ ## op, \
michael@0 646 }
michael@0 647
michael@0 648 #define SIMPLE_NEAREST_FAST_PATH_PAD(op,s,d,func) \
michael@0 649 { PIXMAN_OP_ ## op, \
michael@0 650 PIXMAN_ ## s, \
michael@0 651 (SCALED_NEAREST_FLAGS | \
michael@0 652 FAST_PATH_PAD_REPEAT | \
michael@0 653 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 654 PIXMAN_null, 0, \
michael@0 655 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 656 fast_composite_scaled_nearest_ ## func ## _pad ## _ ## op, \
michael@0 657 }
michael@0 658
michael@0 659 #define SIMPLE_NEAREST_FAST_PATH_NONE(op,s,d,func) \
michael@0 660 { PIXMAN_OP_ ## op, \
michael@0 661 PIXMAN_ ## s, \
michael@0 662 (SCALED_NEAREST_FLAGS | \
michael@0 663 FAST_PATH_NONE_REPEAT | \
michael@0 664 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 665 PIXMAN_null, 0, \
michael@0 666 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 667 fast_composite_scaled_nearest_ ## func ## _none ## _ ## op, \
michael@0 668 }
michael@0 669
michael@0 670 #define SIMPLE_NEAREST_FAST_PATH_COVER(op,s,d,func) \
michael@0 671 { PIXMAN_OP_ ## op, \
michael@0 672 PIXMAN_ ## s, \
michael@0 673 SCALED_NEAREST_FLAGS | FAST_PATH_SAMPLES_COVER_CLIP_NEAREST, \
michael@0 674 PIXMAN_null, 0, \
michael@0 675 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 676 fast_composite_scaled_nearest_ ## func ## _cover ## _ ## op, \
michael@0 677 }
michael@0 678
michael@0 679 #define SIMPLE_NEAREST_A8_MASK_FAST_PATH_NORMAL(op,s,d,func) \
michael@0 680 { PIXMAN_OP_ ## op, \
michael@0 681 PIXMAN_ ## s, \
michael@0 682 (SCALED_NEAREST_FLAGS | \
michael@0 683 FAST_PATH_NORMAL_REPEAT | \
michael@0 684 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 685 PIXMAN_a8, MASK_FLAGS (a8, FAST_PATH_UNIFIED_ALPHA), \
michael@0 686 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 687 fast_composite_scaled_nearest_ ## func ## _normal ## _ ## op, \
michael@0 688 }
michael@0 689
michael@0 690 #define SIMPLE_NEAREST_A8_MASK_FAST_PATH_PAD(op,s,d,func) \
michael@0 691 { PIXMAN_OP_ ## op, \
michael@0 692 PIXMAN_ ## s, \
michael@0 693 (SCALED_NEAREST_FLAGS | \
michael@0 694 FAST_PATH_PAD_REPEAT | \
michael@0 695 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 696 PIXMAN_a8, MASK_FLAGS (a8, FAST_PATH_UNIFIED_ALPHA), \
michael@0 697 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 698 fast_composite_scaled_nearest_ ## func ## _pad ## _ ## op, \
michael@0 699 }
michael@0 700
michael@0 701 #define SIMPLE_NEAREST_A8_MASK_FAST_PATH_NONE(op,s,d,func) \
michael@0 702 { PIXMAN_OP_ ## op, \
michael@0 703 PIXMAN_ ## s, \
michael@0 704 (SCALED_NEAREST_FLAGS | \
michael@0 705 FAST_PATH_NONE_REPEAT | \
michael@0 706 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 707 PIXMAN_a8, MASK_FLAGS (a8, FAST_PATH_UNIFIED_ALPHA), \
michael@0 708 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 709 fast_composite_scaled_nearest_ ## func ## _none ## _ ## op, \
michael@0 710 }
michael@0 711
michael@0 712 #define SIMPLE_NEAREST_A8_MASK_FAST_PATH_COVER(op,s,d,func) \
michael@0 713 { PIXMAN_OP_ ## op, \
michael@0 714 PIXMAN_ ## s, \
michael@0 715 SCALED_NEAREST_FLAGS | FAST_PATH_SAMPLES_COVER_CLIP_NEAREST, \
michael@0 716 PIXMAN_a8, MASK_FLAGS (a8, FAST_PATH_UNIFIED_ALPHA), \
michael@0 717 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 718 fast_composite_scaled_nearest_ ## func ## _cover ## _ ## op, \
michael@0 719 }
michael@0 720
michael@0 721 #define SIMPLE_NEAREST_SOLID_MASK_FAST_PATH_NORMAL(op,s,d,func) \
michael@0 722 { PIXMAN_OP_ ## op, \
michael@0 723 PIXMAN_ ## s, \
michael@0 724 (SCALED_NEAREST_FLAGS | \
michael@0 725 FAST_PATH_NORMAL_REPEAT | \
michael@0 726 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 727 PIXMAN_solid, MASK_FLAGS (solid, FAST_PATH_UNIFIED_ALPHA), \
michael@0 728 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 729 fast_composite_scaled_nearest_ ## func ## _normal ## _ ## op, \
michael@0 730 }
michael@0 731
michael@0 732 #define SIMPLE_NEAREST_SOLID_MASK_FAST_PATH_PAD(op,s,d,func) \
michael@0 733 { PIXMAN_OP_ ## op, \
michael@0 734 PIXMAN_ ## s, \
michael@0 735 (SCALED_NEAREST_FLAGS | \
michael@0 736 FAST_PATH_PAD_REPEAT | \
michael@0 737 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 738 PIXMAN_solid, MASK_FLAGS (solid, FAST_PATH_UNIFIED_ALPHA), \
michael@0 739 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 740 fast_composite_scaled_nearest_ ## func ## _pad ## _ ## op, \
michael@0 741 }
michael@0 742
michael@0 743 #define SIMPLE_NEAREST_SOLID_MASK_FAST_PATH_NONE(op,s,d,func) \
michael@0 744 { PIXMAN_OP_ ## op, \
michael@0 745 PIXMAN_ ## s, \
michael@0 746 (SCALED_NEAREST_FLAGS | \
michael@0 747 FAST_PATH_NONE_REPEAT | \
michael@0 748 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 749 PIXMAN_solid, MASK_FLAGS (solid, FAST_PATH_UNIFIED_ALPHA), \
michael@0 750 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 751 fast_composite_scaled_nearest_ ## func ## _none ## _ ## op, \
michael@0 752 }
michael@0 753
michael@0 754 #define SIMPLE_NEAREST_SOLID_MASK_FAST_PATH_COVER(op,s,d,func) \
michael@0 755 { PIXMAN_OP_ ## op, \
michael@0 756 PIXMAN_ ## s, \
michael@0 757 SCALED_NEAREST_FLAGS | FAST_PATH_SAMPLES_COVER_CLIP_NEAREST, \
michael@0 758 PIXMAN_solid, MASK_FLAGS (solid, FAST_PATH_UNIFIED_ALPHA), \
michael@0 759 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 760 fast_composite_scaled_nearest_ ## func ## _cover ## _ ## op, \
michael@0 761 }
michael@0 762
michael@0 763 /* Prefer the use of 'cover' variant, because it is faster */
michael@0 764 #define SIMPLE_NEAREST_FAST_PATH(op,s,d,func) \
michael@0 765 SIMPLE_NEAREST_FAST_PATH_COVER (op,s,d,func), \
michael@0 766 SIMPLE_NEAREST_FAST_PATH_NONE (op,s,d,func), \
michael@0 767 SIMPLE_NEAREST_FAST_PATH_PAD (op,s,d,func), \
michael@0 768 SIMPLE_NEAREST_FAST_PATH_NORMAL (op,s,d,func)
michael@0 769
michael@0 770 #define SIMPLE_NEAREST_A8_MASK_FAST_PATH(op,s,d,func) \
michael@0 771 SIMPLE_NEAREST_A8_MASK_FAST_PATH_COVER (op,s,d,func), \
michael@0 772 SIMPLE_NEAREST_A8_MASK_FAST_PATH_NONE (op,s,d,func), \
michael@0 773 SIMPLE_NEAREST_A8_MASK_FAST_PATH_PAD (op,s,d,func)
michael@0 774
michael@0 775 #define SIMPLE_NEAREST_SOLID_MASK_FAST_PATH(op,s,d,func) \
michael@0 776 SIMPLE_NEAREST_SOLID_MASK_FAST_PATH_COVER (op,s,d,func), \
michael@0 777 SIMPLE_NEAREST_SOLID_MASK_FAST_PATH_NONE (op,s,d,func), \
michael@0 778 SIMPLE_NEAREST_SOLID_MASK_FAST_PATH_PAD (op,s,d,func)
michael@0 779
michael@0 780 /*****************************************************************************/
michael@0 781
michael@0 782 /*
michael@0 783 * Identify 5 zones in each scanline for bilinear scaling. Depending on
michael@0 784 * whether 2 pixels to be interpolated are fetched from the image itself,
michael@0 785 * from the padding area around it or from both image and padding area.
michael@0 786 */
michael@0 787 static force_inline void
michael@0 788 bilinear_pad_repeat_get_scanline_bounds (int32_t source_image_width,
michael@0 789 pixman_fixed_t vx,
michael@0 790 pixman_fixed_t unit_x,
michael@0 791 int32_t * left_pad,
michael@0 792 int32_t * left_tz,
michael@0 793 int32_t * width,
michael@0 794 int32_t * right_tz,
michael@0 795 int32_t * right_pad)
michael@0 796 {
michael@0 797 int width1 = *width, left_pad1, right_pad1;
michael@0 798 int width2 = *width, left_pad2, right_pad2;
michael@0 799
michael@0 800 pad_repeat_get_scanline_bounds (source_image_width, vx, unit_x,
michael@0 801 &width1, &left_pad1, &right_pad1);
michael@0 802 pad_repeat_get_scanline_bounds (source_image_width, vx + pixman_fixed_1,
michael@0 803 unit_x, &width2, &left_pad2, &right_pad2);
michael@0 804
michael@0 805 *left_pad = left_pad2;
michael@0 806 *left_tz = left_pad1 - left_pad2;
michael@0 807 *right_tz = right_pad2 - right_pad1;
michael@0 808 *right_pad = right_pad1;
michael@0 809 *width -= *left_pad + *left_tz + *right_tz + *right_pad;
michael@0 810 }
michael@0 811
michael@0 812 /*
michael@0 813 * Main loop template for single pass bilinear scaling. It needs to be
michael@0 814 * provided with 'scanline_func' which should do the compositing operation.
michael@0 815 * The needed function has the following prototype:
michael@0 816 *
michael@0 817 * scanline_func (dst_type_t * dst,
michael@0 818 * const mask_type_ * mask,
michael@0 819 * const src_type_t * src_top,
michael@0 820 * const src_type_t * src_bottom,
michael@0 821 * int32_t width,
michael@0 822 * int weight_top,
michael@0 823 * int weight_bottom,
michael@0 824 * pixman_fixed_t vx,
michael@0 825 * pixman_fixed_t unit_x,
michael@0 826 * pixman_fixed_t max_vx,
michael@0 827 * pixman_bool_t zero_src)
michael@0 828 *
michael@0 829 * Where:
michael@0 830 * dst - destination scanline buffer for storing results
michael@0 831 * mask - mask buffer (or single value for solid mask)
michael@0 832 * src_top, src_bottom - two source scanlines
michael@0 833 * width - number of pixels to process
michael@0 834 * weight_top - weight of the top row for interpolation
michael@0 835 * weight_bottom - weight of the bottom row for interpolation
michael@0 836 * vx - initial position for fetching the first pair of
michael@0 837 * pixels from the source buffer
michael@0 838 * unit_x - position increment needed to move to the next pair
michael@0 839 * of pixels
michael@0 840 * max_vx - image size as a fixed point value, can be used for
michael@0 841 * implementing NORMAL repeat (when it is supported)
michael@0 842 * zero_src - boolean hint variable, which is set to TRUE when
michael@0 843 * all source pixels are fetched from zero padding
michael@0 844 * zone for NONE repeat
michael@0 845 *
michael@0 846 * Note: normally the sum of 'weight_top' and 'weight_bottom' is equal to
michael@0 847 * BILINEAR_INTERPOLATION_RANGE, but sometimes it may be less than that
michael@0 848 * for NONE repeat when handling fuzzy antialiased top or bottom image
michael@0 849 * edges. Also both top and bottom weight variables are guaranteed to
michael@0 850 * have value, which is less than BILINEAR_INTERPOLATION_RANGE.
michael@0 851 * For example, the weights can fit into unsigned byte or be used
michael@0 852 * with 8-bit SIMD multiplication instructions for 8-bit interpolation
michael@0 853 * precision.
michael@0 854 */
michael@0 855
michael@0 856 /* Replace a single "scanline_func" with "fetch_func" & "op_func" to allow optional
michael@0 857 * two stage processing (bilinear fetch to a temp buffer, followed by unscaled
michael@0 858 * combine), "op_func" may be NULL, in this case we keep old behavior.
michael@0 859 * This is ugly and gcc issues some warnings, but works.
michael@0 860 *
michael@0 861 * An advice: clang has much better error reporting than gcc for deeply nested macros.
michael@0 862 */
michael@0 863
michael@0 864 #define scanline_func(dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 865 scanline_buf, mask, src_top, src_bottom, width, \
michael@0 866 weight_top, weight_bottom, vx, unit_x, max_vx, zero_src) \
michael@0 867 do { \
michael@0 868 if (op_func != NULL) \
michael@0 869 { \
michael@0 870 fetch_func ((void *)scanline_buf, (mask), (src_top), (src_bottom), (width), \
michael@0 871 (weight_top), (weight_bottom), (vx), (unit_x), (max_vx), (zero_src)); \
michael@0 872 ((void (*)(dst_type_t *, const mask_type_t *, const src_type_t *, int)) op_func)\
michael@0 873 ((dst), (mask), (src_type_t *)scanline_buf, (width)); \
michael@0 874 } \
michael@0 875 else \
michael@0 876 { \
michael@0 877 fetch_func ((void*)(dst), (mask), (src_top), (src_bottom), (width), (weight_top), \
michael@0 878 (weight_bottom), (vx), (unit_x), (max_vx), (zero_src)); \
michael@0 879 } \
michael@0 880 } while (0)
michael@0 881
michael@0 882
michael@0 883 #define SCANLINE_BUFFER_LENGTH 3072
michael@0 884
michael@0 885 #define FAST_BILINEAR_MAINLOOP_INT(scale_func_name, fetch_func, op_func, src_type_t, \
michael@0 886 mask_type_t, dst_type_t, repeat_mode, flags) \
michael@0 887 static void \
michael@0 888 fast_composite_scaled_bilinear ## scale_func_name (pixman_implementation_t *imp, \
michael@0 889 pixman_composite_info_t *info) \
michael@0 890 { \
michael@0 891 PIXMAN_COMPOSITE_ARGS (info); \
michael@0 892 dst_type_t *dst_line; \
michael@0 893 mask_type_t *mask_line; \
michael@0 894 src_type_t *src_first_line; \
michael@0 895 int y1, y2; \
michael@0 896 pixman_fixed_t max_vx = INT32_MAX; /* suppress uninitialized variable warning */ \
michael@0 897 pixman_vector_t v; \
michael@0 898 pixman_fixed_t vx, vy; \
michael@0 899 pixman_fixed_t unit_x, unit_y; \
michael@0 900 int32_t left_pad, left_tz, right_tz, right_pad; \
michael@0 901 \
michael@0 902 dst_type_t *dst; \
michael@0 903 mask_type_t solid_mask; \
michael@0 904 const mask_type_t *mask = &solid_mask; \
michael@0 905 int src_stride, mask_stride, dst_stride; \
michael@0 906 \
michael@0 907 int src_width; \
michael@0 908 pixman_fixed_t src_width_fixed; \
michael@0 909 int max_x; \
michael@0 910 pixman_bool_t need_src_extension; \
michael@0 911 \
michael@0 912 uint64_t stack_scanline_buffer[SCANLINE_BUFFER_LENGTH]; \
michael@0 913 uint8_t *scanline_buffer = (uint8_t *) stack_scanline_buffer; \
michael@0 914 \
michael@0 915 PIXMAN_IMAGE_GET_LINE (dest_image, dest_x, dest_y, dst_type_t, dst_stride, dst_line, 1); \
michael@0 916 if (flags & FLAG_HAVE_SOLID_MASK) \
michael@0 917 { \
michael@0 918 solid_mask = _pixman_image_get_solid (imp, mask_image, dest_image->bits.format); \
michael@0 919 mask_stride = 0; \
michael@0 920 } \
michael@0 921 else if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 922 { \
michael@0 923 PIXMAN_IMAGE_GET_LINE (mask_image, mask_x, mask_y, mask_type_t, \
michael@0 924 mask_stride, mask_line, 1); \
michael@0 925 } \
michael@0 926 \
michael@0 927 /* pass in 0 instead of src_x and src_y because src_x and src_y need to be \
michael@0 928 * transformed from destination space to source space */ \
michael@0 929 PIXMAN_IMAGE_GET_LINE (src_image, 0, 0, src_type_t, src_stride, src_first_line, 1); \
michael@0 930 \
michael@0 931 /* reference point is the center of the pixel */ \
michael@0 932 v.vector[0] = pixman_int_to_fixed (src_x) + pixman_fixed_1 / 2; \
michael@0 933 v.vector[1] = pixman_int_to_fixed (src_y) + pixman_fixed_1 / 2; \
michael@0 934 v.vector[2] = pixman_fixed_1; \
michael@0 935 \
michael@0 936 if (!pixman_transform_point_3d (src_image->common.transform, &v)) \
michael@0 937 return; \
michael@0 938 \
michael@0 939 unit_x = src_image->common.transform->matrix[0][0]; \
michael@0 940 unit_y = src_image->common.transform->matrix[1][1]; \
michael@0 941 \
michael@0 942 v.vector[0] -= pixman_fixed_1 / 2; \
michael@0 943 v.vector[1] -= pixman_fixed_1 / 2; \
michael@0 944 \
michael@0 945 vy = v.vector[1]; \
michael@0 946 \
michael@0 947 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_PAD || \
michael@0 948 PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NONE) \
michael@0 949 { \
michael@0 950 bilinear_pad_repeat_get_scanline_bounds (src_image->bits.width, v.vector[0], unit_x, \
michael@0 951 &left_pad, &left_tz, &width, &right_tz, &right_pad); \
michael@0 952 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_PAD) \
michael@0 953 { \
michael@0 954 /* PAD repeat does not need special handling for 'transition zones' and */ \
michael@0 955 /* they can be combined with 'padding zones' safely */ \
michael@0 956 left_pad += left_tz; \
michael@0 957 right_pad += right_tz; \
michael@0 958 left_tz = right_tz = 0; \
michael@0 959 } \
michael@0 960 v.vector[0] += left_pad * unit_x; \
michael@0 961 } \
michael@0 962 \
michael@0 963 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NORMAL) \
michael@0 964 { \
michael@0 965 vx = v.vector[0]; \
michael@0 966 repeat (PIXMAN_REPEAT_NORMAL, &vx, pixman_int_to_fixed(src_image->bits.width)); \
michael@0 967 max_x = pixman_fixed_to_int (vx + (width - 1) * (int64_t)unit_x) + 1; \
michael@0 968 \
michael@0 969 if (src_image->bits.width < REPEAT_NORMAL_MIN_WIDTH) \
michael@0 970 { \
michael@0 971 src_width = 0; \
michael@0 972 \
michael@0 973 while (src_width < REPEAT_NORMAL_MIN_WIDTH && src_width <= max_x) \
michael@0 974 src_width += src_image->bits.width; \
michael@0 975 \
michael@0 976 need_src_extension = TRUE; \
michael@0 977 } \
michael@0 978 else \
michael@0 979 { \
michael@0 980 src_width = src_image->bits.width; \
michael@0 981 need_src_extension = FALSE; \
michael@0 982 } \
michael@0 983 \
michael@0 984 src_width_fixed = pixman_int_to_fixed (src_width); \
michael@0 985 } \
michael@0 986 \
michael@0 987 if (op_func != NULL && width * sizeof(src_type_t) > sizeof(stack_scanline_buffer)) \
michael@0 988 { \
michael@0 989 scanline_buffer = pixman_malloc_ab (width, sizeof(src_type_t)); \
michael@0 990 \
michael@0 991 if (!scanline_buffer) \
michael@0 992 return; \
michael@0 993 } \
michael@0 994 \
michael@0 995 while (--height >= 0) \
michael@0 996 { \
michael@0 997 int weight1, weight2; \
michael@0 998 dst = dst_line; \
michael@0 999 dst_line += dst_stride; \
michael@0 1000 vx = v.vector[0]; \
michael@0 1001 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1002 { \
michael@0 1003 mask = mask_line; \
michael@0 1004 mask_line += mask_stride; \
michael@0 1005 } \
michael@0 1006 \
michael@0 1007 y1 = pixman_fixed_to_int (vy); \
michael@0 1008 weight2 = pixman_fixed_to_bilinear_weight (vy); \
michael@0 1009 if (weight2) \
michael@0 1010 { \
michael@0 1011 /* both weight1 and weight2 are smaller than BILINEAR_INTERPOLATION_RANGE */ \
michael@0 1012 y2 = y1 + 1; \
michael@0 1013 weight1 = BILINEAR_INTERPOLATION_RANGE - weight2; \
michael@0 1014 } \
michael@0 1015 else \
michael@0 1016 { \
michael@0 1017 /* set both top and bottom row to the same scanline and tweak weights */ \
michael@0 1018 y2 = y1; \
michael@0 1019 weight1 = weight2 = BILINEAR_INTERPOLATION_RANGE / 2; \
michael@0 1020 } \
michael@0 1021 vy += unit_y; \
michael@0 1022 if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_PAD) \
michael@0 1023 { \
michael@0 1024 src_type_t *src1, *src2; \
michael@0 1025 src_type_t buf1[2]; \
michael@0 1026 src_type_t buf2[2]; \
michael@0 1027 repeat (PIXMAN_REPEAT_PAD, &y1, src_image->bits.height); \
michael@0 1028 repeat (PIXMAN_REPEAT_PAD, &y2, src_image->bits.height); \
michael@0 1029 src1 = src_first_line + src_stride * y1; \
michael@0 1030 src2 = src_first_line + src_stride * y2; \
michael@0 1031 \
michael@0 1032 if (left_pad > 0) \
michael@0 1033 { \
michael@0 1034 buf1[0] = buf1[1] = src1[0]; \
michael@0 1035 buf2[0] = buf2[1] = src2[0]; \
michael@0 1036 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1037 scanline_buffer, mask, buf1, buf2, left_pad, weight1, weight2, \
michael@0 1038 0, 0, 0, FALSE); \
michael@0 1039 dst += left_pad; \
michael@0 1040 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1041 mask += left_pad; \
michael@0 1042 } \
michael@0 1043 if (width > 0) \
michael@0 1044 { \
michael@0 1045 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1046 scanline_buffer, mask, src1, src2, width, weight1, weight2, \
michael@0 1047 vx, unit_x, 0, FALSE); \
michael@0 1048 dst += width; \
michael@0 1049 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1050 mask += width; \
michael@0 1051 } \
michael@0 1052 if (right_pad > 0) \
michael@0 1053 { \
michael@0 1054 buf1[0] = buf1[1] = src1[src_image->bits.width - 1]; \
michael@0 1055 buf2[0] = buf2[1] = src2[src_image->bits.width - 1]; \
michael@0 1056 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1057 scanline_buffer, mask, buf1, buf2, right_pad, weight1, weight2, \
michael@0 1058 0, 0, 0, FALSE); \
michael@0 1059 } \
michael@0 1060 } \
michael@0 1061 else if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NONE) \
michael@0 1062 { \
michael@0 1063 src_type_t *src1, *src2; \
michael@0 1064 src_type_t buf1[2]; \
michael@0 1065 src_type_t buf2[2]; \
michael@0 1066 /* handle top/bottom zero padding by just setting weights to 0 if needed */ \
michael@0 1067 if (y1 < 0) \
michael@0 1068 { \
michael@0 1069 weight1 = 0; \
michael@0 1070 y1 = 0; \
michael@0 1071 } \
michael@0 1072 if (y1 >= src_image->bits.height) \
michael@0 1073 { \
michael@0 1074 weight1 = 0; \
michael@0 1075 y1 = src_image->bits.height - 1; \
michael@0 1076 } \
michael@0 1077 if (y2 < 0) \
michael@0 1078 { \
michael@0 1079 weight2 = 0; \
michael@0 1080 y2 = 0; \
michael@0 1081 } \
michael@0 1082 if (y2 >= src_image->bits.height) \
michael@0 1083 { \
michael@0 1084 weight2 = 0; \
michael@0 1085 y2 = src_image->bits.height - 1; \
michael@0 1086 } \
michael@0 1087 src1 = src_first_line + src_stride * y1; \
michael@0 1088 src2 = src_first_line + src_stride * y2; \
michael@0 1089 \
michael@0 1090 if (left_pad > 0) \
michael@0 1091 { \
michael@0 1092 buf1[0] = buf1[1] = 0; \
michael@0 1093 buf2[0] = buf2[1] = 0; \
michael@0 1094 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1095 scanline_buffer, mask, buf1, buf2, left_pad, weight1, weight2, \
michael@0 1096 0, 0, 0, TRUE); \
michael@0 1097 dst += left_pad; \
michael@0 1098 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1099 mask += left_pad; \
michael@0 1100 } \
michael@0 1101 if (left_tz > 0) \
michael@0 1102 { \
michael@0 1103 buf1[0] = 0; \
michael@0 1104 buf1[1] = src1[0]; \
michael@0 1105 buf2[0] = 0; \
michael@0 1106 buf2[1] = src2[0]; \
michael@0 1107 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1108 scanline_buffer, mask, buf1, buf2, left_tz, weight1, weight2, \
michael@0 1109 pixman_fixed_frac (vx), unit_x, 0, FALSE); \
michael@0 1110 dst += left_tz; \
michael@0 1111 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1112 mask += left_tz; \
michael@0 1113 vx += left_tz * unit_x; \
michael@0 1114 } \
michael@0 1115 if (width > 0) \
michael@0 1116 { \
michael@0 1117 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1118 scanline_buffer, mask, src1, src2, width, weight1, weight2, \
michael@0 1119 vx, unit_x, 0, FALSE); \
michael@0 1120 dst += width; \
michael@0 1121 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1122 mask += width; \
michael@0 1123 vx += width * unit_x; \
michael@0 1124 } \
michael@0 1125 if (right_tz > 0) \
michael@0 1126 { \
michael@0 1127 buf1[0] = src1[src_image->bits.width - 1]; \
michael@0 1128 buf1[1] = 0; \
michael@0 1129 buf2[0] = src2[src_image->bits.width - 1]; \
michael@0 1130 buf2[1] = 0; \
michael@0 1131 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1132 scanline_buffer, mask, buf1, buf2, right_tz, weight1, weight2, \
michael@0 1133 pixman_fixed_frac (vx), unit_x, 0, FALSE); \
michael@0 1134 dst += right_tz; \
michael@0 1135 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1136 mask += right_tz; \
michael@0 1137 } \
michael@0 1138 if (right_pad > 0) \
michael@0 1139 { \
michael@0 1140 buf1[0] = buf1[1] = 0; \
michael@0 1141 buf2[0] = buf2[1] = 0; \
michael@0 1142 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1143 scanline_buffer, mask, buf1, buf2, right_pad, weight1, weight2, \
michael@0 1144 0, 0, 0, TRUE); \
michael@0 1145 } \
michael@0 1146 } \
michael@0 1147 else if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NORMAL) \
michael@0 1148 { \
michael@0 1149 int32_t num_pixels; \
michael@0 1150 int32_t width_remain; \
michael@0 1151 src_type_t * src_line_top; \
michael@0 1152 src_type_t * src_line_bottom; \
michael@0 1153 src_type_t buf1[2]; \
michael@0 1154 src_type_t buf2[2]; \
michael@0 1155 src_type_t extended_src_line0[REPEAT_NORMAL_MIN_WIDTH*2]; \
michael@0 1156 src_type_t extended_src_line1[REPEAT_NORMAL_MIN_WIDTH*2]; \
michael@0 1157 int i, j; \
michael@0 1158 \
michael@0 1159 repeat (PIXMAN_REPEAT_NORMAL, &y1, src_image->bits.height); \
michael@0 1160 repeat (PIXMAN_REPEAT_NORMAL, &y2, src_image->bits.height); \
michael@0 1161 src_line_top = src_first_line + src_stride * y1; \
michael@0 1162 src_line_bottom = src_first_line + src_stride * y2; \
michael@0 1163 \
michael@0 1164 if (need_src_extension) \
michael@0 1165 { \
michael@0 1166 for (i=0; i<src_width;) \
michael@0 1167 { \
michael@0 1168 for (j=0; j<src_image->bits.width; j++, i++) \
michael@0 1169 { \
michael@0 1170 extended_src_line0[i] = src_line_top[j]; \
michael@0 1171 extended_src_line1[i] = src_line_bottom[j]; \
michael@0 1172 } \
michael@0 1173 } \
michael@0 1174 \
michael@0 1175 src_line_top = &extended_src_line0[0]; \
michael@0 1176 src_line_bottom = &extended_src_line1[0]; \
michael@0 1177 } \
michael@0 1178 \
michael@0 1179 /* Top & Bottom wrap around buffer */ \
michael@0 1180 buf1[0] = src_line_top[src_width - 1]; \
michael@0 1181 buf1[1] = src_line_top[0]; \
michael@0 1182 buf2[0] = src_line_bottom[src_width - 1]; \
michael@0 1183 buf2[1] = src_line_bottom[0]; \
michael@0 1184 \
michael@0 1185 width_remain = width; \
michael@0 1186 \
michael@0 1187 while (width_remain > 0) \
michael@0 1188 { \
michael@0 1189 /* We use src_width_fixed because it can make vx in original source range */ \
michael@0 1190 repeat (PIXMAN_REPEAT_NORMAL, &vx, src_width_fixed); \
michael@0 1191 \
michael@0 1192 /* Wrap around part */ \
michael@0 1193 if (pixman_fixed_to_int (vx) == src_width - 1) \
michael@0 1194 { \
michael@0 1195 /* for positive unit_x \
michael@0 1196 * num_pixels = max(n) + 1, where vx + n*unit_x < src_width_fixed \
michael@0 1197 * \
michael@0 1198 * vx is in range [0, src_width_fixed - pixman_fixed_e] \
michael@0 1199 * So we are safe from overflow. \
michael@0 1200 */ \
michael@0 1201 num_pixels = ((src_width_fixed - vx - pixman_fixed_e) / unit_x) + 1; \
michael@0 1202 \
michael@0 1203 if (num_pixels > width_remain) \
michael@0 1204 num_pixels = width_remain; \
michael@0 1205 \
michael@0 1206 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, \
michael@0 1207 dst, scanline_buffer, mask, buf1, buf2, num_pixels, \
michael@0 1208 weight1, weight2, pixman_fixed_frac(vx), \
michael@0 1209 unit_x, src_width_fixed, FALSE); \
michael@0 1210 \
michael@0 1211 width_remain -= num_pixels; \
michael@0 1212 vx += num_pixels * unit_x; \
michael@0 1213 dst += num_pixels; \
michael@0 1214 \
michael@0 1215 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1216 mask += num_pixels; \
michael@0 1217 \
michael@0 1218 repeat (PIXMAN_REPEAT_NORMAL, &vx, src_width_fixed); \
michael@0 1219 } \
michael@0 1220 \
michael@0 1221 /* Normal scanline composite */ \
michael@0 1222 if (pixman_fixed_to_int (vx) != src_width - 1 && width_remain > 0) \
michael@0 1223 { \
michael@0 1224 /* for positive unit_x \
michael@0 1225 * num_pixels = max(n) + 1, where vx + n*unit_x < (src_width_fixed - 1) \
michael@0 1226 * \
michael@0 1227 * vx is in range [0, src_width_fixed - pixman_fixed_e] \
michael@0 1228 * So we are safe from overflow here. \
michael@0 1229 */ \
michael@0 1230 num_pixels = ((src_width_fixed - pixman_fixed_1 - vx - pixman_fixed_e) \
michael@0 1231 / unit_x) + 1; \
michael@0 1232 \
michael@0 1233 if (num_pixels > width_remain) \
michael@0 1234 num_pixels = width_remain; \
michael@0 1235 \
michael@0 1236 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, \
michael@0 1237 dst, scanline_buffer, mask, src_line_top, src_line_bottom, \
michael@0 1238 num_pixels, weight1, weight2, vx, unit_x, src_width_fixed, \
michael@0 1239 FALSE); \
michael@0 1240 \
michael@0 1241 width_remain -= num_pixels; \
michael@0 1242 vx += num_pixels * unit_x; \
michael@0 1243 dst += num_pixels; \
michael@0 1244 \
michael@0 1245 if (flags & FLAG_HAVE_NON_SOLID_MASK) \
michael@0 1246 mask += num_pixels; \
michael@0 1247 } \
michael@0 1248 } \
michael@0 1249 } \
michael@0 1250 else \
michael@0 1251 { \
michael@0 1252 scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst, \
michael@0 1253 scanline_buffer, mask, \
michael@0 1254 src_first_line + src_stride * y1, \
michael@0 1255 src_first_line + src_stride * y2, width, \
michael@0 1256 weight1, weight2, vx, unit_x, max_vx, FALSE); \
michael@0 1257 } \
michael@0 1258 } \
michael@0 1259 if (scanline_buffer != (uint8_t *) stack_scanline_buffer) \
michael@0 1260 free (scanline_buffer); \
michael@0 1261 }
michael@0 1262
michael@0 1263 /* A workaround for old sun studio, see: https://bugs.freedesktop.org/show_bug.cgi?id=32764 */
michael@0 1264 #define FAST_BILINEAR_MAINLOOP_COMMON(scale_func_name, fetch_func, op_func, src_type_t, mask_type_t,\
michael@0 1265 dst_type_t, repeat_mode, flags) \
michael@0 1266 FAST_BILINEAR_MAINLOOP_INT(_ ## scale_func_name, fetch_func, op_func, src_type_t, mask_type_t,\
michael@0 1267 dst_type_t, repeat_mode, flags)
michael@0 1268
michael@0 1269 #define SCALED_BILINEAR_FLAGS \
michael@0 1270 (FAST_PATH_SCALE_TRANSFORM | \
michael@0 1271 FAST_PATH_NO_ALPHA_MAP | \
michael@0 1272 FAST_PATH_BILINEAR_FILTER | \
michael@0 1273 FAST_PATH_NO_ACCESSORS | \
michael@0 1274 FAST_PATH_NARROW_FORMAT)
michael@0 1275
michael@0 1276 #define SIMPLE_BILINEAR_FAST_PATH_PAD(op,s,d,func) \
michael@0 1277 { PIXMAN_OP_ ## op, \
michael@0 1278 PIXMAN_ ## s, \
michael@0 1279 (SCALED_BILINEAR_FLAGS | \
michael@0 1280 FAST_PATH_PAD_REPEAT | \
michael@0 1281 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1282 PIXMAN_null, 0, \
michael@0 1283 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1284 fast_composite_scaled_bilinear_ ## func ## _pad ## _ ## op, \
michael@0 1285 }
michael@0 1286
michael@0 1287 #define SIMPLE_BILINEAR_FAST_PATH_NONE(op,s,d,func) \
michael@0 1288 { PIXMAN_OP_ ## op, \
michael@0 1289 PIXMAN_ ## s, \
michael@0 1290 (SCALED_BILINEAR_FLAGS | \
michael@0 1291 FAST_PATH_NONE_REPEAT | \
michael@0 1292 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1293 PIXMAN_null, 0, \
michael@0 1294 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1295 fast_composite_scaled_bilinear_ ## func ## _none ## _ ## op, \
michael@0 1296 }
michael@0 1297
michael@0 1298 #define SIMPLE_BILINEAR_FAST_PATH_COVER(op,s,d,func) \
michael@0 1299 { PIXMAN_OP_ ## op, \
michael@0 1300 PIXMAN_ ## s, \
michael@0 1301 SCALED_BILINEAR_FLAGS | FAST_PATH_SAMPLES_COVER_CLIP_BILINEAR, \
michael@0 1302 PIXMAN_null, 0, \
michael@0 1303 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1304 fast_composite_scaled_bilinear_ ## func ## _cover ## _ ## op, \
michael@0 1305 }
michael@0 1306
michael@0 1307 #define SIMPLE_BILINEAR_FAST_PATH_NORMAL(op,s,d,func) \
michael@0 1308 { PIXMAN_OP_ ## op, \
michael@0 1309 PIXMAN_ ## s, \
michael@0 1310 (SCALED_BILINEAR_FLAGS | \
michael@0 1311 FAST_PATH_NORMAL_REPEAT | \
michael@0 1312 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1313 PIXMAN_null, 0, \
michael@0 1314 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1315 fast_composite_scaled_bilinear_ ## func ## _normal ## _ ## op, \
michael@0 1316 }
michael@0 1317
michael@0 1318 #define SIMPLE_BILINEAR_A8_MASK_FAST_PATH_PAD(op,s,d,func) \
michael@0 1319 { PIXMAN_OP_ ## op, \
michael@0 1320 PIXMAN_ ## s, \
michael@0 1321 (SCALED_BILINEAR_FLAGS | \
michael@0 1322 FAST_PATH_PAD_REPEAT | \
michael@0 1323 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1324 PIXMAN_a8, MASK_FLAGS (a8, FAST_PATH_UNIFIED_ALPHA), \
michael@0 1325 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1326 fast_composite_scaled_bilinear_ ## func ## _pad ## _ ## op, \
michael@0 1327 }
michael@0 1328
michael@0 1329 #define SIMPLE_BILINEAR_A8_MASK_FAST_PATH_NONE(op,s,d,func) \
michael@0 1330 { PIXMAN_OP_ ## op, \
michael@0 1331 PIXMAN_ ## s, \
michael@0 1332 (SCALED_BILINEAR_FLAGS | \
michael@0 1333 FAST_PATH_NONE_REPEAT | \
michael@0 1334 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1335 PIXMAN_a8, MASK_FLAGS (a8, FAST_PATH_UNIFIED_ALPHA), \
michael@0 1336 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1337 fast_composite_scaled_bilinear_ ## func ## _none ## _ ## op, \
michael@0 1338 }
michael@0 1339
michael@0 1340 #define SIMPLE_BILINEAR_A8_MASK_FAST_PATH_COVER(op,s,d,func) \
michael@0 1341 { PIXMAN_OP_ ## op, \
michael@0 1342 PIXMAN_ ## s, \
michael@0 1343 SCALED_BILINEAR_FLAGS | FAST_PATH_SAMPLES_COVER_CLIP_BILINEAR, \
michael@0 1344 PIXMAN_a8, MASK_FLAGS (a8, FAST_PATH_UNIFIED_ALPHA), \
michael@0 1345 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1346 fast_composite_scaled_bilinear_ ## func ## _cover ## _ ## op, \
michael@0 1347 }
michael@0 1348
michael@0 1349 #define SIMPLE_BILINEAR_A8_MASK_FAST_PATH_NORMAL(op,s,d,func) \
michael@0 1350 { PIXMAN_OP_ ## op, \
michael@0 1351 PIXMAN_ ## s, \
michael@0 1352 (SCALED_BILINEAR_FLAGS | \
michael@0 1353 FAST_PATH_NORMAL_REPEAT | \
michael@0 1354 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1355 PIXMAN_a8, MASK_FLAGS (a8, FAST_PATH_UNIFIED_ALPHA), \
michael@0 1356 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1357 fast_composite_scaled_bilinear_ ## func ## _normal ## _ ## op, \
michael@0 1358 }
michael@0 1359
michael@0 1360 #define SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH_PAD(op,s,d,func) \
michael@0 1361 { PIXMAN_OP_ ## op, \
michael@0 1362 PIXMAN_ ## s, \
michael@0 1363 (SCALED_BILINEAR_FLAGS | \
michael@0 1364 FAST_PATH_PAD_REPEAT | \
michael@0 1365 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1366 PIXMAN_solid, MASK_FLAGS (solid, FAST_PATH_UNIFIED_ALPHA), \
michael@0 1367 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1368 fast_composite_scaled_bilinear_ ## func ## _pad ## _ ## op, \
michael@0 1369 }
michael@0 1370
michael@0 1371 #define SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH_NONE(op,s,d,func) \
michael@0 1372 { PIXMAN_OP_ ## op, \
michael@0 1373 PIXMAN_ ## s, \
michael@0 1374 (SCALED_BILINEAR_FLAGS | \
michael@0 1375 FAST_PATH_NONE_REPEAT | \
michael@0 1376 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1377 PIXMAN_solid, MASK_FLAGS (solid, FAST_PATH_UNIFIED_ALPHA), \
michael@0 1378 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1379 fast_composite_scaled_bilinear_ ## func ## _none ## _ ## op, \
michael@0 1380 }
michael@0 1381
michael@0 1382 #define SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH_COVER(op,s,d,func) \
michael@0 1383 { PIXMAN_OP_ ## op, \
michael@0 1384 PIXMAN_ ## s, \
michael@0 1385 SCALED_BILINEAR_FLAGS | FAST_PATH_SAMPLES_COVER_CLIP_BILINEAR, \
michael@0 1386 PIXMAN_solid, MASK_FLAGS (solid, FAST_PATH_UNIFIED_ALPHA), \
michael@0 1387 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1388 fast_composite_scaled_bilinear_ ## func ## _cover ## _ ## op, \
michael@0 1389 }
michael@0 1390
michael@0 1391 #define SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH_NORMAL(op,s,d,func) \
michael@0 1392 { PIXMAN_OP_ ## op, \
michael@0 1393 PIXMAN_ ## s, \
michael@0 1394 (SCALED_BILINEAR_FLAGS | \
michael@0 1395 FAST_PATH_NORMAL_REPEAT | \
michael@0 1396 FAST_PATH_X_UNIT_POSITIVE), \
michael@0 1397 PIXMAN_solid, MASK_FLAGS (solid, FAST_PATH_UNIFIED_ALPHA), \
michael@0 1398 PIXMAN_ ## d, FAST_PATH_STD_DEST_FLAGS, \
michael@0 1399 fast_composite_scaled_bilinear_ ## func ## _normal ## _ ## op, \
michael@0 1400 }
michael@0 1401
michael@0 1402 /* Prefer the use of 'cover' variant, because it is faster */
michael@0 1403 #define SIMPLE_BILINEAR_FAST_PATH(op,s,d,func) \
michael@0 1404 SIMPLE_BILINEAR_FAST_PATH_COVER (op,s,d,func), \
michael@0 1405 SIMPLE_BILINEAR_FAST_PATH_NONE (op,s,d,func), \
michael@0 1406 SIMPLE_BILINEAR_FAST_PATH_PAD (op,s,d,func), \
michael@0 1407 SIMPLE_BILINEAR_FAST_PATH_NORMAL (op,s,d,func)
michael@0 1408
michael@0 1409 #define SIMPLE_BILINEAR_A8_MASK_FAST_PATH(op,s,d,func) \
michael@0 1410 SIMPLE_BILINEAR_A8_MASK_FAST_PATH_COVER (op,s,d,func), \
michael@0 1411 SIMPLE_BILINEAR_A8_MASK_FAST_PATH_NONE (op,s,d,func), \
michael@0 1412 SIMPLE_BILINEAR_A8_MASK_FAST_PATH_PAD (op,s,d,func), \
michael@0 1413 SIMPLE_BILINEAR_A8_MASK_FAST_PATH_NORMAL (op,s,d,func)
michael@0 1414
michael@0 1415 #define SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH(op,s,d,func) \
michael@0 1416 SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH_COVER (op,s,d,func), \
michael@0 1417 SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH_NONE (op,s,d,func), \
michael@0 1418 SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH_PAD (op,s,d,func), \
michael@0 1419 SIMPLE_BILINEAR_SOLID_MASK_FAST_PATH_NORMAL (op,s,d,func)
michael@0 1420
michael@0 1421 #endif

mercurial