pixman-mips-dspr2-asm.S (120733B)
1 /* 2 * Copyright (c) 2012 3 * MIPS Technologies, Inc., California. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its 14 * contributors may be used to endorse or promote products derived from 15 * this software without specific prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 * 29 * Author: Nemanja Lukic (nemanja.lukic@rt-rk.com) 30 */ 31 32 #include "pixman-private.h" 33 #include "pixman-mips-dspr2-asm.h" 34 35 LEAF_MIPS_DSPR2(pixman_fill_buff16_mips) 36 /* 37 * a0 - *dest 38 * a1 - count (bytes) 39 * a2 - value to fill buffer with 40 */ 41 42 beqz a1, 3f 43 andi t1, a0, 0x0002 44 beqz t1, 0f /* check if address is 4-byte aligned */ 45 nop 46 sh a2, 0(a0) 47 addiu a0, a0, 2 48 addiu a1, a1, -2 49 0: 50 srl t1, a1, 5 /* t1 how many multiples of 32 bytes */ 51 replv.ph a2, a2 /* replicate fill value (16bit) in a2 */ 52 beqz t1, 2f 53 nop 54 1: 55 addiu t1, t1, -1 56 beqz t1, 11f 57 addiu a1, a1, -32 58 pref 30, 32(a0) 59 sw a2, 0(a0) 60 sw a2, 4(a0) 61 sw a2, 8(a0) 62 sw a2, 12(a0) 63 sw a2, 16(a0) 64 sw a2, 20(a0) 65 sw a2, 24(a0) 66 sw a2, 28(a0) 67 b 1b 68 addiu a0, a0, 32 69 11: 70 sw a2, 0(a0) 71 sw a2, 4(a0) 72 sw a2, 8(a0) 73 sw a2, 12(a0) 74 sw a2, 16(a0) 75 sw a2, 20(a0) 76 sw a2, 24(a0) 77 sw a2, 28(a0) 78 addiu a0, a0, 32 79 2: 80 blez a1, 3f 81 addiu a1, a1, -2 82 sh a2, 0(a0) 83 b 2b 84 addiu a0, a0, 2 85 3: 86 jr ra 87 nop 88 89 END(pixman_fill_buff16_mips) 90 91 LEAF_MIPS32R2(pixman_fill_buff32_mips) 92 /* 93 * a0 - *dest 94 * a1 - count (bytes) 95 * a2 - value to fill buffer with 96 */ 97 98 beqz a1, 3f 99 nop 100 srl t1, a1, 5 /* t1 how many multiples of 32 bytes */ 101 beqz t1, 2f 102 nop 103 1: 104 addiu t1, t1, -1 105 beqz t1, 11f 106 addiu a1, a1, -32 107 pref 30, 32(a0) 108 sw a2, 0(a0) 109 sw a2, 4(a0) 110 sw a2, 8(a0) 111 sw a2, 12(a0) 112 sw a2, 16(a0) 113 sw a2, 20(a0) 114 sw a2, 24(a0) 115 sw a2, 28(a0) 116 b 1b 117 addiu a0, a0, 32 118 11: 119 sw a2, 0(a0) 120 sw a2, 4(a0) 121 sw a2, 8(a0) 122 sw a2, 12(a0) 123 sw a2, 16(a0) 124 sw a2, 20(a0) 125 sw a2, 24(a0) 126 sw a2, 28(a0) 127 addiu a0, a0, 32 128 2: 129 blez a1, 3f 130 addiu a1, a1, -4 131 sw a2, 0(a0) 132 b 2b 133 addiu a0, a0, 4 134 3: 135 jr ra 136 nop 137 138 END(pixman_fill_buff32_mips) 139 140 LEAF_MIPS_DSPR2(pixman_composite_src_8888_0565_asm_mips) 141 /* 142 * a0 - dst (r5g6b5) 143 * a1 - src (a8r8g8b8) 144 * a2 - w 145 */ 146 147 beqz a2, 3f 148 nop 149 addiu t1, a2, -1 150 beqz t1, 2f 151 nop 152 li t4, 0xf800f800 153 li t5, 0x07e007e0 154 li t6, 0x001f001f 155 1: 156 lw t0, 0(a1) 157 lw t1, 4(a1) 158 addiu a1, a1, 8 159 addiu a2, a2, -2 160 161 CONVERT_2x8888_TO_2x0565 t0, t1, t2, t3, t4, t5, t6, t7, t8 162 163 sh t2, 0(a0) 164 sh t3, 2(a0) 165 166 addiu t2, a2, -1 167 bgtz t2, 1b 168 addiu a0, a0, 4 169 2: 170 beqz a2, 3f 171 nop 172 lw t0, 0(a1) 173 174 CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3 175 176 sh t1, 0(a0) 177 3: 178 j ra 179 nop 180 181 END(pixman_composite_src_8888_0565_asm_mips) 182 183 LEAF_MIPS_DSPR2(pixman_composite_src_0565_8888_asm_mips) 184 /* 185 * a0 - dst (a8r8g8b8) 186 * a1 - src (r5g6b5) 187 * a2 - w 188 */ 189 190 beqz a2, 3f 191 nop 192 addiu t1, a2, -1 193 beqz t1, 2f 194 nop 195 li t4, 0x07e007e0 196 li t5, 0x001F001F 197 1: 198 lhu t0, 0(a1) 199 lhu t1, 2(a1) 200 addiu a1, a1, 4 201 addiu a2, a2, -2 202 203 CONVERT_2x0565_TO_2x8888 t0, t1, t2, t3, t4, t5, t6, t7, t8, t9 204 205 sw t2, 0(a0) 206 sw t3, 4(a0) 207 208 addiu t2, a2, -1 209 bgtz t2, 1b 210 addiu a0, a0, 8 211 2: 212 beqz a2, 3f 213 nop 214 lhu t0, 0(a1) 215 216 CONVERT_1x0565_TO_1x8888 t0, t1, t2, t3 217 218 sw t1, 0(a0) 219 3: 220 j ra 221 nop 222 223 END(pixman_composite_src_0565_8888_asm_mips) 224 225 LEAF_MIPS_DSPR2(pixman_composite_src_x888_8888_asm_mips) 226 /* 227 * a0 - dst (a8r8g8b8) 228 * a1 - src (x8r8g8b8) 229 * a2 - w 230 */ 231 232 beqz a2, 4f 233 nop 234 li t9, 0xff000000 235 srl t8, a2, 3 /* t1 = how many multiples of 8 src pixels */ 236 beqz t8, 3f /* branch if less than 8 src pixels */ 237 nop 238 1: 239 addiu t8, t8, -1 240 beqz t8, 2f 241 addiu a2, a2, -8 242 pref 0, 32(a1) 243 lw t0, 0(a1) 244 lw t1, 4(a1) 245 lw t2, 8(a1) 246 lw t3, 12(a1) 247 lw t4, 16(a1) 248 lw t5, 20(a1) 249 lw t6, 24(a1) 250 lw t7, 28(a1) 251 addiu a1, a1, 32 252 or t0, t0, t9 253 or t1, t1, t9 254 or t2, t2, t9 255 or t3, t3, t9 256 or t4, t4, t9 257 or t5, t5, t9 258 or t6, t6, t9 259 or t7, t7, t9 260 pref 30, 32(a0) 261 sw t0, 0(a0) 262 sw t1, 4(a0) 263 sw t2, 8(a0) 264 sw t3, 12(a0) 265 sw t4, 16(a0) 266 sw t5, 20(a0) 267 sw t6, 24(a0) 268 sw t7, 28(a0) 269 b 1b 270 addiu a0, a0, 32 271 2: 272 lw t0, 0(a1) 273 lw t1, 4(a1) 274 lw t2, 8(a1) 275 lw t3, 12(a1) 276 lw t4, 16(a1) 277 lw t5, 20(a1) 278 lw t6, 24(a1) 279 lw t7, 28(a1) 280 addiu a1, a1, 32 281 or t0, t0, t9 282 or t1, t1, t9 283 or t2, t2, t9 284 or t3, t3, t9 285 or t4, t4, t9 286 or t5, t5, t9 287 or t6, t6, t9 288 or t7, t7, t9 289 sw t0, 0(a0) 290 sw t1, 4(a0) 291 sw t2, 8(a0) 292 sw t3, 12(a0) 293 sw t4, 16(a0) 294 sw t5, 20(a0) 295 sw t6, 24(a0) 296 sw t7, 28(a0) 297 beqz a2, 4f 298 addiu a0, a0, 32 299 3: 300 lw t0, 0(a1) 301 addiu a1, a1, 4 302 addiu a2, a2, -1 303 or t1, t0, t9 304 sw t1, 0(a0) 305 bnez a2, 3b 306 addiu a0, a0, 4 307 4: 308 jr ra 309 nop 310 311 END(pixman_composite_src_x888_8888_asm_mips) 312 313 #if defined(__MIPSEL__) || defined(__MIPSEL) || defined(_MIPSEL) || defined(MIPSEL) 314 LEAF_MIPS_DSPR2(pixman_composite_src_0888_8888_rev_asm_mips) 315 /* 316 * a0 - dst (a8r8g8b8) 317 * a1 - src (b8g8r8) 318 * a2 - w 319 */ 320 321 beqz a2, 6f 322 nop 323 324 lui t8, 0xff00; 325 srl t9, a2, 2 /* t9 = how many multiples of 4 src pixels */ 326 beqz t9, 4f /* branch if less than 4 src pixels */ 327 nop 328 329 li t0, 0x1 330 li t1, 0x2 331 li t2, 0x3 332 andi t3, a1, 0x3 333 beq t3, t0, 1f 334 nop 335 beq t3, t1, 2f 336 nop 337 beq t3, t2, 3f 338 nop 339 340 0: 341 beqz t9, 4f 342 addiu t9, t9, -1 343 lw t0, 0(a1) /* t0 = R2 | B1 | G1 | R1 */ 344 lw t1, 4(a1) /* t1 = G3 | R3 | B2 | G2 */ 345 lw t2, 8(a1) /* t2 = B4 | G4 | R4 | B3 */ 346 347 addiu a1, a1, 12 348 addiu a2, a2, -4 349 350 wsbh t0, t0 /* t0 = B1 | R2 | R1 | G1 */ 351 wsbh t1, t1 /* t1 = R3 | G3 | G2 | B2 */ 352 wsbh t2, t2 /* t2 = G4 | B4 | B3 | R4 */ 353 354 packrl.ph t3, t1, t0 /* t3 = G2 | B2 | B1 | R2 */ 355 packrl.ph t4, t0, t0 /* t4 = R1 | G1 | B1 | R2 */ 356 rotr t3, t3, 16 /* t3 = B1 | R2 | G2 | B2 */ 357 or t3, t3, t8 /* t3 = FF | R2 | G2 | B2 */ 358 srl t4, t4, 8 /* t4 = 0 | R1 | G1 | B1 */ 359 or t4, t4, t8 /* t4 = FF | R1 | G1 | B1 */ 360 packrl.ph t5, t2, t1 /* t5 = B3 | R4 | R3 | G3 */ 361 rotr t5, t5, 24 /* t5 = R4 | R3 | G3 | B3 */ 362 or t5, t5, t8 /* t5 = FF | R3 | G3 | B3 */ 363 rotr t2, t2, 16 /* t2 = B3 | R4 | G4 | B4 */ 364 or t2, t2, t8 /* t5 = FF | R3 | G3 | B3 */ 365 366 sw t4, 0(a0) 367 sw t3, 4(a0) 368 sw t5, 8(a0) 369 sw t2, 12(a0) 370 b 0b 371 addiu a0, a0, 16 372 373 1: 374 lbu t6, 0(a1) /* t6 = 0 | 0 | 0 | R1 */ 375 lhu t7, 1(a1) /* t7 = 0 | 0 | B1 | G1 */ 376 sll t6, t6, 16 /* t6 = 0 | R1 | 0 | 0 */ 377 wsbh t7, t7 /* t7 = 0 | 0 | G1 | B1 */ 378 or t7, t6, t7 /* t7 = 0 | R1 | G1 | B1 */ 379 11: 380 beqz t9, 4f 381 addiu t9, t9, -1 382 lw t0, 3(a1) /* t0 = R3 | B2 | G2 | R2 */ 383 lw t1, 7(a1) /* t1 = G4 | R4 | B3 | G3 */ 384 lw t2, 11(a1) /* t2 = B5 | G5 | R5 | B4 */ 385 386 addiu a1, a1, 12 387 addiu a2, a2, -4 388 389 wsbh t0, t0 /* t0 = B2 | R3 | R2 | G2 */ 390 wsbh t1, t1 /* t1 = R4 | G4 | G3 | B3 */ 391 wsbh t2, t2 /* t2 = G5 | B5 | B4 | R5 */ 392 393 packrl.ph t3, t1, t0 /* t3 = G3 | B3 | B2 | R3 */ 394 packrl.ph t4, t2, t1 /* t4 = B4 | R5 | R4 | G4 */ 395 rotr t0, t0, 24 /* t0 = R3 | R2 | G2 | B2 */ 396 rotr t3, t3, 16 /* t3 = B2 | R3 | G3 | B3 */ 397 rotr t4, t4, 24 /* t4 = R5 | R4 | G4 | B4 */ 398 or t7, t7, t8 /* t7 = FF | R1 | G1 | B1 */ 399 or t0, t0, t8 /* t0 = FF | R2 | G2 | B2 */ 400 or t3, t3, t8 /* t1 = FF | R3 | G3 | B3 */ 401 or t4, t4, t8 /* t3 = FF | R4 | G4 | B4 */ 402 403 sw t7, 0(a0) 404 sw t0, 4(a0) 405 sw t3, 8(a0) 406 sw t4, 12(a0) 407 rotr t7, t2, 16 /* t7 = xx | R5 | G5 | B5 */ 408 b 11b 409 addiu a0, a0, 16 410 411 2: 412 lhu t7, 0(a1) /* t7 = 0 | 0 | G1 | R1 */ 413 wsbh t7, t7 /* t7 = 0 | 0 | R1 | G1 */ 414 21: 415 beqz t9, 4f 416 addiu t9, t9, -1 417 lw t0, 2(a1) /* t0 = B2 | G2 | R2 | B1 */ 418 lw t1, 6(a1) /* t1 = R4 | B3 | G3 | R3 */ 419 lw t2, 10(a1) /* t2 = G5 | R5 | B4 | G4 */ 420 421 addiu a1, a1, 12 422 addiu a2, a2, -4 423 424 wsbh t0, t0 /* t0 = G2 | B2 | B1 | R2 */ 425 wsbh t1, t1 /* t1 = B3 | R4 | R3 | G3 */ 426 wsbh t2, t2 /* t2 = R5 | G5 | G4 | B4 */ 427 428 precr_sra.ph.w t7, t0, 0 /* t7 = R1 | G1 | B1 | R2 */ 429 rotr t0, t0, 16 /* t0 = B1 | R2 | G2 | B2 */ 430 packrl.ph t3, t2, t1 /* t3 = G4 | B4 | B3 | R4 */ 431 rotr t1, t1, 24 /* t1 = R4 | R3 | G3 | B3 */ 432 srl t7, t7, 8 /* t7 = 0 | R1 | G1 | B1 */ 433 rotr t3, t3, 16 /* t3 = B3 | R4 | G4 | B4 */ 434 or t7, t7, t8 /* t7 = FF | R1 | G1 | B1 */ 435 or t0, t0, t8 /* t0 = FF | R2 | G2 | B2 */ 436 or t1, t1, t8 /* t1 = FF | R3 | G3 | B3 */ 437 or t3, t3, t8 /* t3 = FF | R4 | G4 | B4 */ 438 439 sw t7, 0(a0) 440 sw t0, 4(a0) 441 sw t1, 8(a0) 442 sw t3, 12(a0) 443 srl t7, t2, 16 /* t7 = 0 | 0 | R5 | G5 */ 444 b 21b 445 addiu a0, a0, 16 446 447 3: 448 lbu t7, 0(a1) /* t7 = 0 | 0 | 0 | R1 */ 449 31: 450 beqz t9, 4f 451 addiu t9, t9, -1 452 lw t0, 1(a1) /* t0 = G2 | R2 | B1 | G1 */ 453 lw t1, 5(a1) /* t1 = B3 | G3 | R3 | B2 */ 454 lw t2, 9(a1) /* t2 = R5 | B4 | G4 | R4 */ 455 456 addiu a1, a1, 12 457 addiu a2, a2, -4 458 459 wsbh t0, t0 /* t0 = R2 | G2 | G1 | B1 */ 460 wsbh t1, t1 /* t1 = G3 | B3 | B2 | R3 */ 461 wsbh t2, t2 /* t2 = B4 | R5 | R4 | G4 */ 462 463 precr_sra.ph.w t7, t0, 0 /* t7 = xx | R1 | G1 | B1 */ 464 packrl.ph t3, t1, t0 /* t3 = B2 | R3 | R2 | G2 */ 465 rotr t1, t1, 16 /* t1 = B2 | R3 | G3 | B3 */ 466 rotr t4, t2, 24 /* t4 = R5 | R4 | G4 | B4 */ 467 rotr t3, t3, 24 /* t3 = R3 | R2 | G2 | B2 */ 468 or t7, t7, t8 /* t7 = FF | R1 | G1 | B1 */ 469 or t3, t3, t8 /* t3 = FF | R2 | G2 | B2 */ 470 or t1, t1, t8 /* t1 = FF | R3 | G3 | B3 */ 471 or t4, t4, t8 /* t4 = FF | R4 | G4 | B4 */ 472 473 sw t7, 0(a0) 474 sw t3, 4(a0) 475 sw t1, 8(a0) 476 sw t4, 12(a0) 477 srl t7, t2, 16 /* t7 = 0 | 0 | xx | R5 */ 478 b 31b 479 addiu a0, a0, 16 480 481 4: 482 beqz a2, 6f 483 nop 484 5: 485 lbu t0, 0(a1) /* t0 = 0 | 0 | 0 | R */ 486 lbu t1, 1(a1) /* t1 = 0 | 0 | 0 | G */ 487 lbu t2, 2(a1) /* t2 = 0 | 0 | 0 | B */ 488 addiu a1, a1, 3 489 490 sll t0, t0, 16 /* t2 = 0 | R | 0 | 0 */ 491 sll t1, t1, 8 /* t1 = 0 | 0 | G | 0 */ 492 493 or t2, t2, t1 /* t2 = 0 | 0 | G | B */ 494 or t2, t2, t0 /* t2 = 0 | R | G | B */ 495 or t2, t2, t8 /* t2 = FF | R | G | B */ 496 497 sw t2, 0(a0) 498 addiu a2, a2, -1 499 bnez a2, 5b 500 addiu a0, a0, 4 501 6: 502 j ra 503 nop 504 505 END(pixman_composite_src_0888_8888_rev_asm_mips) 506 507 LEAF_MIPS_DSPR2(pixman_composite_src_0888_0565_rev_asm_mips) 508 /* 509 * a0 - dst (r5g6b5) 510 * a1 - src (b8g8r8) 511 * a2 - w 512 */ 513 514 SAVE_REGS_ON_STACK 0, v0, v1 515 beqz a2, 6f 516 nop 517 518 li t6, 0xf800f800 519 li t7, 0x07e007e0 520 li t8, 0x001F001F 521 srl t9, a2, 2 /* t9 = how many multiples of 4 src pixels */ 522 beqz t9, 4f /* branch if less than 4 src pixels */ 523 nop 524 525 li t0, 0x1 526 li t1, 0x2 527 li t2, 0x3 528 andi t3, a1, 0x3 529 beq t3, t0, 1f 530 nop 531 beq t3, t1, 2f 532 nop 533 beq t3, t2, 3f 534 nop 535 536 0: 537 beqz t9, 4f 538 addiu t9, t9, -1 539 lw t0, 0(a1) /* t0 = R2 | B1 | G1 | R1 */ 540 lw t1, 4(a1) /* t1 = G3 | R3 | B2 | G2 */ 541 lw t2, 8(a1) /* t2 = B4 | G4 | R4 | B3 */ 542 543 addiu a1, a1, 12 544 addiu a2, a2, -4 545 546 wsbh t0, t0 /* t0 = B1 | R2 | R1 | G1 */ 547 wsbh t1, t1 /* t1 = R3 | G3 | G2 | B2 */ 548 wsbh t2, t2 /* t2 = G4 | B4 | B3 | R4 */ 549 550 packrl.ph t3, t1, t0 /* t3 = G2 | B2 | B1 | R2 */ 551 packrl.ph t4, t0, t0 /* t4 = R1 | G1 | B1 | R2 */ 552 rotr t3, t3, 16 /* t3 = B1 | R2 | G2 | B2 */ 553 srl t4, t4, 8 /* t4 = 0 | R1 | G1 | B1 */ 554 packrl.ph t5, t2, t1 /* t5 = B3 | R4 | R3 | G3 */ 555 rotr t5, t5, 24 /* t5 = R4 | R3 | G3 | B3 */ 556 rotr t2, t2, 16 /* t2 = B3 | R4 | G4 | B4 */ 557 558 CONVERT_2x8888_TO_2x0565 t4, t3, t4, t3, t6, t7, t8, v0, v1 559 CONVERT_2x8888_TO_2x0565 t5, t2, t5, t2, t6, t7, t8, v0, v1 560 561 sh t4, 0(a0) 562 sh t3, 2(a0) 563 sh t5, 4(a0) 564 sh t2, 6(a0) 565 b 0b 566 addiu a0, a0, 8 567 568 1: 569 lbu t4, 0(a1) /* t4 = 0 | 0 | 0 | R1 */ 570 lhu t5, 1(a1) /* t5 = 0 | 0 | B1 | G1 */ 571 sll t4, t4, 16 /* t4 = 0 | R1 | 0 | 0 */ 572 wsbh t5, t5 /* t5 = 0 | 0 | G1 | B1 */ 573 or t5, t4, t5 /* t5 = 0 | R1 | G1 | B1 */ 574 11: 575 beqz t9, 4f 576 addiu t9, t9, -1 577 lw t0, 3(a1) /* t0 = R3 | B2 | G2 | R2 */ 578 lw t1, 7(a1) /* t1 = G4 | R4 | B3 | G3 */ 579 lw t2, 11(a1) /* t2 = B5 | G5 | R5 | B4 */ 580 581 addiu a1, a1, 12 582 addiu a2, a2, -4 583 584 wsbh t0, t0 /* t0 = B2 | R3 | R2 | G2 */ 585 wsbh t1, t1 /* t1 = R4 | G4 | G3 | B3 */ 586 wsbh t2, t2 /* t2 = G5 | B5 | B4 | R5 */ 587 588 packrl.ph t3, t1, t0 /* t3 = G3 | B3 | B2 | R3 */ 589 packrl.ph t4, t2, t1 /* t4 = B4 | R5 | R4 | G4 */ 590 rotr t0, t0, 24 /* t0 = R3 | R2 | G2 | B2 */ 591 rotr t3, t3, 16 /* t3 = B2 | R3 | G3 | B3 */ 592 rotr t4, t4, 24 /* t4 = R5 | R4 | G4 | B4 */ 593 594 CONVERT_2x8888_TO_2x0565 t5, t0, t5, t0, t6, t7, t8, v0, v1 595 CONVERT_2x8888_TO_2x0565 t3, t4, t3, t4, t6, t7, t8, v0, v1 596 597 sh t5, 0(a0) 598 sh t0, 2(a0) 599 sh t3, 4(a0) 600 sh t4, 6(a0) 601 rotr t5, t2, 16 /* t5 = xx | R5 | G5 | B5 */ 602 b 11b 603 addiu a0, a0, 8 604 605 2: 606 lhu t5, 0(a1) /* t5 = 0 | 0 | G1 | R1 */ 607 wsbh t5, t5 /* t5 = 0 | 0 | R1 | G1 */ 608 21: 609 beqz t9, 4f 610 addiu t9, t9, -1 611 lw t0, 2(a1) /* t0 = B2 | G2 | R2 | B1 */ 612 lw t1, 6(a1) /* t1 = R4 | B3 | G3 | R3 */ 613 lw t2, 10(a1) /* t2 = G5 | R5 | B4 | G4 */ 614 615 addiu a1, a1, 12 616 addiu a2, a2, -4 617 618 wsbh t0, t0 /* t0 = G2 | B2 | B1 | R2 */ 619 wsbh t1, t1 /* t1 = B3 | R4 | R3 | G3 */ 620 wsbh t2, t2 /* t2 = R5 | G5 | G4 | B4 */ 621 622 precr_sra.ph.w t5, t0, 0 /* t5 = R1 | G1 | B1 | R2 */ 623 rotr t0, t0, 16 /* t0 = B1 | R2 | G2 | B2 */ 624 packrl.ph t3, t2, t1 /* t3 = G4 | B4 | B3 | R4 */ 625 rotr t1, t1, 24 /* t1 = R4 | R3 | G3 | B3 */ 626 srl t5, t5, 8 /* t5 = 0 | R1 | G1 | B1 */ 627 rotr t3, t3, 16 /* t3 = B3 | R4 | G4 | B4 */ 628 629 CONVERT_2x8888_TO_2x0565 t5, t0, t5, t0, t6, t7, t8, v0, v1 630 CONVERT_2x8888_TO_2x0565 t1, t3, t1, t3, t6, t7, t8, v0, v1 631 632 sh t5, 0(a0) 633 sh t0, 2(a0) 634 sh t1, 4(a0) 635 sh t3, 6(a0) 636 srl t5, t2, 16 /* t5 = 0 | 0 | R5 | G5 */ 637 b 21b 638 addiu a0, a0, 8 639 640 3: 641 lbu t5, 0(a1) /* t5 = 0 | 0 | 0 | R1 */ 642 31: 643 beqz t9, 4f 644 addiu t9, t9, -1 645 lw t0, 1(a1) /* t0 = G2 | R2 | B1 | G1 */ 646 lw t1, 5(a1) /* t1 = B3 | G3 | R3 | B2 */ 647 lw t2, 9(a1) /* t2 = R5 | B4 | G4 | R4 */ 648 649 addiu a1, a1, 12 650 addiu a2, a2, -4 651 652 wsbh t0, t0 /* t0 = R2 | G2 | G1 | B1 */ 653 wsbh t1, t1 /* t1 = G3 | B3 | B2 | R3 */ 654 wsbh t2, t2 /* t2 = B4 | R5 | R4 | G4 */ 655 656 precr_sra.ph.w t5, t0, 0 /* t5 = xx | R1 | G1 | B1 */ 657 packrl.ph t3, t1, t0 /* t3 = B2 | R3 | R2 | G2 */ 658 rotr t1, t1, 16 /* t1 = B2 | R3 | G3 | B3 */ 659 rotr t4, t2, 24 /* t4 = R5 | R4 | G4 | B4 */ 660 rotr t3, t3, 24 /* t3 = R3 | R2 | G2 | B2 */ 661 662 CONVERT_2x8888_TO_2x0565 t5, t3, t5, t3, t6, t7, t8, v0, v1 663 CONVERT_2x8888_TO_2x0565 t1, t4, t1, t4, t6, t7, t8, v0, v1 664 665 sh t5, 0(a0) 666 sh t3, 2(a0) 667 sh t1, 4(a0) 668 sh t4, 6(a0) 669 srl t5, t2, 16 /* t5 = 0 | 0 | xx | R5 */ 670 b 31b 671 addiu a0, a0, 8 672 673 4: 674 beqz a2, 6f 675 nop 676 5: 677 lbu t0, 0(a1) /* t0 = 0 | 0 | 0 | R */ 678 lbu t1, 1(a1) /* t1 = 0 | 0 | 0 | G */ 679 lbu t2, 2(a1) /* t2 = 0 | 0 | 0 | B */ 680 addiu a1, a1, 3 681 682 sll t0, t0, 16 /* t2 = 0 | R | 0 | 0 */ 683 sll t1, t1, 8 /* t1 = 0 | 0 | G | 0 */ 684 685 or t2, t2, t1 /* t2 = 0 | 0 | G | B */ 686 or t2, t2, t0 /* t2 = 0 | R | G | B */ 687 688 CONVERT_1x8888_TO_1x0565 t2, t3, t4, t5 689 690 sh t3, 0(a0) 691 addiu a2, a2, -1 692 bnez a2, 5b 693 addiu a0, a0, 2 694 6: 695 RESTORE_REGS_FROM_STACK 0, v0, v1 696 j ra 697 nop 698 699 END(pixman_composite_src_0888_0565_rev_asm_mips) 700 #endif 701 702 LEAF_MIPS_DSPR2(pixman_composite_src_pixbuf_8888_asm_mips) 703 /* 704 * a0 - dst (a8b8g8r8) 705 * a1 - src (a8r8g8b8) 706 * a2 - w 707 */ 708 709 SAVE_REGS_ON_STACK 0, v0 710 li v0, 0x00ff00ff 711 712 beqz a2, 3f 713 nop 714 addiu t1, a2, -1 715 beqz t1, 2f 716 nop 717 1: 718 lw t0, 0(a1) 719 lw t1, 4(a1) 720 addiu a1, a1, 8 721 addiu a2, a2, -2 722 srl t2, t0, 24 723 srl t3, t1, 24 724 725 MIPS_2xUN8x4_MUL_2xUN8 t0, t1, t2, t3, t0, t1, v0, t4, t5, t6, t7, t8, t9 726 727 sll t0, t0, 8 728 sll t1, t1, 8 729 andi t2, t2, 0xff 730 andi t3, t3, 0xff 731 or t0, t0, t2 732 or t1, t1, t3 733 wsbh t0, t0 734 wsbh t1, t1 735 rotr t0, t0, 16 736 rotr t1, t1, 16 737 sw t0, 0(a0) 738 sw t1, 4(a0) 739 740 addiu t2, a2, -1 741 bgtz t2, 1b 742 addiu a0, a0, 8 743 2: 744 beqz a2, 3f 745 nop 746 lw t0, 0(a1) 747 srl t1, t0, 24 748 749 MIPS_UN8x4_MUL_UN8 t0, t1, t0, v0, t3, t4, t5 750 751 sll t0, t0, 8 752 andi t1, t1, 0xff 753 or t0, t0, t1 754 wsbh t0, t0 755 rotr t0, t0, 16 756 sw t0, 0(a0) 757 3: 758 RESTORE_REGS_FROM_STACK 0, v0 759 j ra 760 nop 761 762 END(pixman_composite_src_pixbuf_8888_asm_mips) 763 764 LEAF_MIPS_DSPR2(pixman_composite_src_rpixbuf_8888_asm_mips) 765 /* 766 * a0 - dst (a8r8g8b8) 767 * a1 - src (a8r8g8b8) 768 * a2 - w 769 */ 770 771 SAVE_REGS_ON_STACK 0, v0 772 li v0, 0x00ff00ff 773 774 beqz a2, 3f 775 nop 776 addiu t1, a2, -1 777 beqz t1, 2f 778 nop 779 1: 780 lw t0, 0(a1) 781 lw t1, 4(a1) 782 addiu a1, a1, 8 783 addiu a2, a2, -2 784 srl t2, t0, 24 785 srl t3, t1, 24 786 787 MIPS_2xUN8x4_MUL_2xUN8 t0, t1, t2, t3, t0, t1, v0, t4, t5, t6, t7, t8, t9 788 789 sll t0, t0, 8 790 sll t1, t1, 8 791 andi t2, t2, 0xff 792 andi t3, t3, 0xff 793 or t0, t0, t2 794 or t1, t1, t3 795 rotr t0, t0, 8 796 rotr t1, t1, 8 797 sw t0, 0(a0) 798 sw t1, 4(a0) 799 800 addiu t2, a2, -1 801 bgtz t2, 1b 802 addiu a0, a0, 8 803 2: 804 beqz a2, 3f 805 nop 806 lw t0, 0(a1) 807 srl t1, t0, 24 808 809 MIPS_UN8x4_MUL_UN8 t0, t1, t0, v0, t3, t4, t5 810 811 sll t0, t0, 8 812 andi t1, t1, 0xff 813 or t0, t0, t1 814 rotr t0, t0, 8 815 sw t0, 0(a0) 816 3: 817 RESTORE_REGS_FROM_STACK 0, v0 818 j ra 819 nop 820 821 END(pixman_composite_src_rpixbuf_8888_asm_mips) 822 823 LEAF_MIPS_DSPR2(pixman_composite_src_n_8_8888_asm_mips) 824 /* 825 * a0 - dst (a8r8g8b8) 826 * a1 - src (32bit constant) 827 * a2 - mask (a8) 828 * a3 - w 829 */ 830 831 832 SAVE_REGS_ON_STACK 0, v0 833 li v0, 0x00ff00ff 834 835 beqz a3, 3f 836 nop 837 addiu t1, a3, -1 838 beqz t1, 2f 839 nop 840 841 1: 842 /* a1 = source (32bit constant) */ 843 lbu t0, 0(a2) /* t2 = mask (a8) */ 844 lbu t1, 1(a2) /* t3 = mask (a8) */ 845 addiu a2, a2, 2 846 847 MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, t2, t3, v0, t4, t5, t6, t7, t8, t9 848 849 sw t2, 0(a0) 850 sw t3, 4(a0) 851 addiu a3, a3, -2 852 addiu t2, a3, -1 853 bgtz t2, 1b 854 addiu a0, a0, 8 855 856 beqz a3, 3f 857 nop 858 859 2: 860 lbu t0, 0(a2) 861 addiu a2, a2, 1 862 863 MIPS_UN8x4_MUL_UN8 a1, t0, t1, v0, t3, t4, t5 864 865 sw t1, 0(a0) 866 addiu a3, a3, -1 867 addiu a0, a0, 4 868 869 3: 870 RESTORE_REGS_FROM_STACK 0, v0 871 j ra 872 nop 873 874 END(pixman_composite_src_n_8_8888_asm_mips) 875 876 LEAF_MIPS_DSPR2(pixman_composite_src_n_8_8_asm_mips) 877 /* 878 * a0 - dst (a8) 879 * a1 - src (32bit constant) 880 * a2 - mask (a8) 881 * a3 - w 882 */ 883 884 li t9, 0x00ff00ff 885 beqz a3, 3f 886 nop 887 srl t7, a3, 2 /* t7 = how many multiples of 4 dst pixels */ 888 beqz t7, 1f /* branch if less than 4 src pixels */ 889 nop 890 891 srl t8, a1, 24 892 replv.ph t8, t8 893 894 0: 895 beqz t7, 1f 896 addiu t7, t7, -1 897 lbu t0, 0(a2) 898 lbu t1, 1(a2) 899 lbu t2, 2(a2) 900 lbu t3, 3(a2) 901 902 addiu a2, a2, 4 903 904 precr_sra.ph.w t1, t0, 0 905 precr_sra.ph.w t3, t2, 0 906 precr.qb.ph t0, t3, t1 907 908 muleu_s.ph.qbl t2, t0, t8 909 muleu_s.ph.qbr t3, t0, t8 910 shra_r.ph t4, t2, 8 911 shra_r.ph t5, t3, 8 912 and t4, t4, t9 913 and t5, t5, t9 914 addq.ph t2, t2, t4 915 addq.ph t3, t3, t5 916 shra_r.ph t2, t2, 8 917 shra_r.ph t3, t3, 8 918 precr.qb.ph t2, t2, t3 919 920 sb t2, 0(a0) 921 srl t2, t2, 8 922 sb t2, 1(a0) 923 srl t2, t2, 8 924 sb t2, 2(a0) 925 srl t2, t2, 8 926 sb t2, 3(a0) 927 addiu a3, a3, -4 928 b 0b 929 addiu a0, a0, 4 930 931 1: 932 beqz a3, 3f 933 nop 934 srl t8, a1, 24 935 2: 936 lbu t0, 0(a2) 937 addiu a2, a2, 1 938 939 mul t2, t0, t8 940 shra_r.ph t3, t2, 8 941 andi t3, t3, 0x00ff 942 addq.ph t2, t2, t3 943 shra_r.ph t2, t2, 8 944 945 sb t2, 0(a0) 946 addiu a3, a3, -1 947 bnez a3, 2b 948 addiu a0, a0, 1 949 950 3: 951 j ra 952 nop 953 954 END(pixman_composite_src_n_8_8_asm_mips) 955 956 LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_8888_ca_asm_mips) 957 /* 958 * a0 - dst (a8r8g8b8) 959 * a1 - src (32bit constant) 960 * a2 - mask (a8r8g8b8) 961 * a3 - w 962 */ 963 964 beqz a3, 8f 965 nop 966 SAVE_REGS_ON_STACK 8, s0, s1, s2, s3, s4, s5 967 968 li t6, 0xff 969 addiu t7, zero, -1 /* t7 = 0xffffffff */ 970 srl t8, a1, 24 /* t8 = srca */ 971 li t9, 0x00ff00ff 972 973 addiu t1, a3, -1 974 beqz t1, 4f /* last pixel */ 975 nop 976 977 0: 978 lw t0, 0(a2) /* t0 = mask */ 979 lw t1, 4(a2) /* t1 = mask */ 980 addiu a3, a3, -2 /* w = w - 2 */ 981 or t2, t0, t1 982 beqz t2, 3f /* if (t0 == 0) && (t1 == 0) */ 983 addiu a2, a2, 8 984 and t2, t0, t1 985 beq t2, t7, 1f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */ 986 nop 987 988 //if(ma) 989 lw t2, 0(a0) /* t2 = dst */ 990 lw t3, 4(a0) /* t3 = dst */ 991 MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, t4, t5, t9, s0, s1, s2, s3, s4, s5 992 MIPS_2xUN8x4_MUL_2xUN8 t0, t1, t8, t8, t0, t1, t9, s0, s1, s2, s3, s4, s5 993 not t0, t0 994 not t1, t1 995 MIPS_2xUN8x4_MUL_2xUN8x4 t2, t3, t0, t1, t2, t3, t9, s0, s1, s2, s3, s4, s5 996 addu_s.qb t2, t4, t2 997 addu_s.qb t3, t5, t3 998 sw t2, 0(a0) 999 sw t3, 4(a0) 1000 addiu t1, a3, -1 1001 bgtz t1, 0b 1002 addiu a0, a0, 8 1003 b 4f 1004 nop 1005 1: 1006 //if (t0 == 0xffffffff) && (t1 == 0xffffffff): 1007 beq t8, t6, 2f /* if (srca == 0xff) */ 1008 nop 1009 lw t2, 0(a0) /* t2 = dst */ 1010 lw t3, 4(a0) /* t3 = dst */ 1011 not t0, a1 1012 not t1, a1 1013 srl t0, t0, 24 1014 srl t1, t1, 24 1015 MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t0, t1, t2, t3, t9, s0, s1, s2, s3, s4, s5 1016 addu_s.qb t2, a1, t2 1017 addu_s.qb t3, a1, t3 1018 sw t2, 0(a0) 1019 sw t3, 4(a0) 1020 addiu t1, a3, -1 1021 bgtz t1, 0b 1022 addiu a0, a0, 8 1023 b 4f 1024 nop 1025 2: 1026 sw a1, 0(a0) 1027 sw a1, 4(a0) 1028 3: 1029 addiu t1, a3, -1 1030 bgtz t1, 0b 1031 addiu a0, a0, 8 1032 1033 4: 1034 beqz a3, 7f 1035 nop 1036 /* a1 = src */ 1037 lw t0, 0(a2) /* t0 = mask */ 1038 beqz t0, 7f /* if (t0 == 0) */ 1039 nop 1040 beq t0, t7, 5f /* if (t0 == 0xffffffff) */ 1041 nop 1042 //if(ma) 1043 lw t1, 0(a0) /* t1 = dst */ 1044 MIPS_UN8x4_MUL_UN8x4 a1, t0, t2, t9, t3, t4, t5, s0 1045 MIPS_UN8x4_MUL_UN8 t0, t8, t0, t9, t3, t4, t5 1046 not t0, t0 1047 MIPS_UN8x4_MUL_UN8x4 t1, t0, t1, t9, t3, t4, t5, s0 1048 addu_s.qb t1, t2, t1 1049 sw t1, 0(a0) 1050 RESTORE_REGS_FROM_STACK 8, s0, s1, s2, s3, s4, s5 1051 j ra 1052 nop 1053 5: 1054 //if (t0 == 0xffffffff) 1055 beq t8, t6, 6f /* if (srca == 0xff) */ 1056 nop 1057 lw t1, 0(a0) /* t1 = dst */ 1058 not t0, a1 1059 srl t0, t0, 24 1060 MIPS_UN8x4_MUL_UN8 t1, t0, t1, t9, t2, t3, t4 1061 addu_s.qb t1, a1, t1 1062 sw t1, 0(a0) 1063 RESTORE_REGS_FROM_STACK 8, s0, s1, s2, s3, s4, s5 1064 j ra 1065 nop 1066 6: 1067 sw a1, 0(a0) 1068 7: 1069 RESTORE_REGS_FROM_STACK 8, s0, s1, s2, s3, s4, s5 1070 8: 1071 j ra 1072 nop 1073 1074 END(pixman_composite_over_n_8888_8888_ca_asm_mips) 1075 1076 LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_0565_ca_asm_mips) 1077 /* 1078 * a0 - dst (r5g6b5) 1079 * a1 - src (32bit constant) 1080 * a2 - mask (a8r8g8b8) 1081 * a3 - w 1082 */ 1083 1084 beqz a3, 8f 1085 nop 1086 SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7, s8 1087 1088 li t6, 0xff 1089 addiu t7, zero, -1 /* t7 = 0xffffffff */ 1090 srl t8, a1, 24 /* t8 = srca */ 1091 li t9, 0x00ff00ff 1092 li s6, 0xf800f800 1093 li s7, 0x07e007e0 1094 li s8, 0x001F001F 1095 1096 addiu t1, a3, -1 1097 beqz t1, 4f /* last pixel */ 1098 nop 1099 1100 0: 1101 lw t0, 0(a2) /* t0 = mask */ 1102 lw t1, 4(a2) /* t1 = mask */ 1103 addiu a3, a3, -2 /* w = w - 2 */ 1104 or t2, t0, t1 1105 beqz t2, 3f /* if (t0 == 0) && (t1 == 0) */ 1106 addiu a2, a2, 8 1107 and t2, t0, t1 1108 beq t2, t7, 1f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */ 1109 nop 1110 1111 //if(ma) 1112 lhu t2, 0(a0) /* t2 = dst */ 1113 lhu t3, 2(a0) /* t3 = dst */ 1114 MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, t4, t5, t9, s0, s1, s2, s3, s4, s5 1115 MIPS_2xUN8x4_MUL_2xUN8 t0, t1, t8, t8, t0, t1, t9, s0, s1, s2, s3, s4, s5 1116 not t0, t0 1117 not t1, t1 1118 CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, s7, s8, s0, s1, s2, s3 1119 MIPS_2xUN8x4_MUL_2xUN8x4 t2, t3, t0, t1, t2, t3, t9, s0, s1, s2, s3, s4, s5 1120 addu_s.qb t2, t4, t2 1121 addu_s.qb t3, t5, t3 1122 CONVERT_2x8888_TO_2x0565 t2, t3, t2, t3, s6, s7, s8, s0, s1 1123 sh t2, 0(a0) 1124 sh t3, 2(a0) 1125 addiu t1, a3, -1 1126 bgtz t1, 0b 1127 addiu a0, a0, 4 1128 b 4f 1129 nop 1130 1: 1131 //if (t0 == 0xffffffff) && (t1 == 0xffffffff): 1132 beq t8, t6, 2f /* if (srca == 0xff) */ 1133 nop 1134 lhu t2, 0(a0) /* t2 = dst */ 1135 lhu t3, 2(a0) /* t3 = dst */ 1136 not t0, a1 1137 not t1, a1 1138 srl t0, t0, 24 1139 srl t1, t1, 24 1140 CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, s7, s8, s0, s1, s2, s3 1141 MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t0, t1, t2, t3, t9, s0, s1, s2, s3, s4, s5 1142 addu_s.qb t2, a1, t2 1143 addu_s.qb t3, a1, t3 1144 CONVERT_2x8888_TO_2x0565 t2, t3, t2, t3, s6, s7, s8, s0, s1 1145 sh t2, 0(a0) 1146 sh t3, 2(a0) 1147 addiu t1, a3, -1 1148 bgtz t1, 0b 1149 addiu a0, a0, 4 1150 b 4f 1151 nop 1152 2: 1153 CONVERT_1x8888_TO_1x0565 a1, t2, s0, s1 1154 sh t2, 0(a0) 1155 sh t2, 2(a0) 1156 3: 1157 addiu t1, a3, -1 1158 bgtz t1, 0b 1159 addiu a0, a0, 4 1160 1161 4: 1162 beqz a3, 7f 1163 nop 1164 /* a1 = src */ 1165 lw t0, 0(a2) /* t0 = mask */ 1166 beqz t0, 7f /* if (t0 == 0) */ 1167 nop 1168 beq t0, t7, 5f /* if (t0 == 0xffffffff) */ 1169 nop 1170 //if(ma) 1171 lhu t1, 0(a0) /* t1 = dst */ 1172 MIPS_UN8x4_MUL_UN8x4 a1, t0, t2, t9, t3, t4, t5, s0 1173 MIPS_UN8x4_MUL_UN8 t0, t8, t0, t9, t3, t4, t5 1174 not t0, t0 1175 CONVERT_1x0565_TO_1x8888 t1, s1, s2, s3 1176 MIPS_UN8x4_MUL_UN8x4 s1, t0, s1, t9, t3, t4, t5, s0 1177 addu_s.qb s1, t2, s1 1178 CONVERT_1x8888_TO_1x0565 s1, t1, s0, s2 1179 sh t1, 0(a0) 1180 RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7, s8 1181 j ra 1182 nop 1183 5: 1184 //if (t0 == 0xffffffff) 1185 beq t8, t6, 6f /* if (srca == 0xff) */ 1186 nop 1187 lhu t1, 0(a0) /* t1 = dst */ 1188 not t0, a1 1189 srl t0, t0, 24 1190 CONVERT_1x0565_TO_1x8888 t1, s1, s2, s3 1191 MIPS_UN8x4_MUL_UN8 s1, t0, s1, t9, t2, t3, t4 1192 addu_s.qb s1, a1, s1 1193 CONVERT_1x8888_TO_1x0565 s1, t1, s0, s2 1194 sh t1, 0(a0) 1195 RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7, s8 1196 j ra 1197 nop 1198 6: 1199 CONVERT_1x8888_TO_1x0565 a1, t1, s0, s2 1200 sh t1, 0(a0) 1201 7: 1202 RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7, s8 1203 8: 1204 j ra 1205 nop 1206 1207 END(pixman_composite_over_n_8888_0565_ca_asm_mips) 1208 1209 LEAF_MIPS_DSPR2(pixman_composite_over_n_8_8_asm_mips) 1210 /* 1211 * a0 - dst (a8) 1212 * a1 - src (32bit constant) 1213 * a2 - mask (a8) 1214 * a3 - w 1215 */ 1216 1217 SAVE_REGS_ON_STACK 0, v0 1218 li t9, 0x00ff00ff 1219 beqz a3, 3f 1220 nop 1221 srl v0, a3, 2 /* v0 = how many multiples of 4 dst pixels */ 1222 beqz v0, 1f /* branch if less than 4 src pixels */ 1223 nop 1224 1225 srl t8, a1, 24 1226 replv.ph t8, t8 1227 1228 0: 1229 beqz v0, 1f 1230 addiu v0, v0, -1 1231 lbu t0, 0(a2) 1232 lbu t1, 1(a2) 1233 lbu t2, 2(a2) 1234 lbu t3, 3(a2) 1235 lbu t4, 0(a0) 1236 lbu t5, 1(a0) 1237 lbu t6, 2(a0) 1238 lbu t7, 3(a0) 1239 1240 addiu a2, a2, 4 1241 1242 precr_sra.ph.w t1, t0, 0 1243 precr_sra.ph.w t3, t2, 0 1244 precr_sra.ph.w t5, t4, 0 1245 precr_sra.ph.w t7, t6, 0 1246 1247 precr.qb.ph t0, t3, t1 1248 precr.qb.ph t1, t7, t5 1249 1250 muleu_s.ph.qbl t2, t0, t8 1251 muleu_s.ph.qbr t3, t0, t8 1252 shra_r.ph t4, t2, 8 1253 shra_r.ph t5, t3, 8 1254 and t4, t4, t9 1255 and t5, t5, t9 1256 addq.ph t2, t2, t4 1257 addq.ph t3, t3, t5 1258 shra_r.ph t2, t2, 8 1259 shra_r.ph t3, t3, 8 1260 precr.qb.ph t0, t2, t3 1261 not t6, t0 1262 1263 preceu.ph.qbl t7, t6 1264 preceu.ph.qbr t6, t6 1265 1266 muleu_s.ph.qbl t2, t1, t7 1267 muleu_s.ph.qbr t3, t1, t6 1268 shra_r.ph t4, t2, 8 1269 shra_r.ph t5, t3, 8 1270 and t4, t4, t9 1271 and t5, t5, t9 1272 addq.ph t2, t2, t4 1273 addq.ph t3, t3, t5 1274 shra_r.ph t2, t2, 8 1275 shra_r.ph t3, t3, 8 1276 precr.qb.ph t1, t2, t3 1277 1278 addu_s.qb t2, t0, t1 1279 1280 sb t2, 0(a0) 1281 srl t2, t2, 8 1282 sb t2, 1(a0) 1283 srl t2, t2, 8 1284 sb t2, 2(a0) 1285 srl t2, t2, 8 1286 sb t2, 3(a0) 1287 addiu a3, a3, -4 1288 b 0b 1289 addiu a0, a0, 4 1290 1291 1: 1292 beqz a3, 3f 1293 nop 1294 srl t8, a1, 24 1295 2: 1296 lbu t0, 0(a2) 1297 lbu t1, 0(a0) 1298 addiu a2, a2, 1 1299 1300 mul t2, t0, t8 1301 shra_r.ph t3, t2, 8 1302 andi t3, t3, 0x00ff 1303 addq.ph t2, t2, t3 1304 shra_r.ph t2, t2, 8 1305 not t3, t2 1306 andi t3, t3, 0x00ff 1307 1308 1309 mul t4, t1, t3 1310 shra_r.ph t5, t4, 8 1311 andi t5, t5, 0x00ff 1312 addq.ph t4, t4, t5 1313 shra_r.ph t4, t4, 8 1314 andi t4, t4, 0x00ff 1315 1316 addu_s.qb t2, t2, t4 1317 sb t2, 0(a0) 1318 addiu a3, a3, -1 1319 bnez a3, 2b 1320 addiu a0, a0, 1 1321 1322 3: 1323 RESTORE_REGS_FROM_STACK 0, v0 1324 j ra 1325 nop 1326 1327 END(pixman_composite_over_n_8_8_asm_mips) 1328 1329 LEAF_MIPS_DSPR2(pixman_composite_over_n_8_8888_asm_mips) 1330 /* 1331 * a0 - dst (a8r8g8b8) 1332 * a1 - src (32bit constant) 1333 * a2 - mask (a8) 1334 * a3 - w 1335 */ 1336 1337 SAVE_REGS_ON_STACK 4, s0, s1, s2, s3, s4 1338 beqz a3, 4f 1339 nop 1340 li t4, 0x00ff00ff 1341 li t5, 0xff 1342 addiu t0, a3, -1 1343 beqz t0, 3f /* last pixel */ 1344 srl t6, a1, 24 /* t6 = srca */ 1345 not s4, a1 1346 beq t5, t6, 2f /* if (srca == 0xff) */ 1347 srl s4, s4, 24 1348 1: 1349 /* a1 = src */ 1350 lbu t0, 0(a2) /* t0 = mask */ 1351 lbu t1, 1(a2) /* t1 = mask */ 1352 or t2, t0, t1 1353 beqz t2, 111f /* if (t0 == 0) && (t1 == 0) */ 1354 addiu a2, a2, 2 1355 and t3, t0, t1 1356 1357 lw t2, 0(a0) /* t2 = dst */ 1358 beq t3, t5, 11f /* if (t0 == 0xff) && (t1 == 0xff) */ 1359 lw t3, 4(a0) /* t3 = dst */ 1360 1361 MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, s0, s1, t4, t6, t7, t8, t9, s2, s3 1362 not s2, s0 1363 not s3, s1 1364 srl s2, s2, 24 1365 srl s3, s3, 24 1366 MIPS_2xUN8x4_MUL_2xUN8 t2, t3, s2, s3, t2, t3, t4, t0, t1, t6, t7, t8, t9 1367 addu_s.qb s2, t2, s0 1368 addu_s.qb s3, t3, s1 1369 sw s2, 0(a0) 1370 b 111f 1371 sw s3, 4(a0) 1372 11: 1373 MIPS_2xUN8x4_MUL_2xUN8 t2, t3, s4, s4, t2, t3, t4, t0, t1, t6, t7, t8, t9 1374 addu_s.qb s2, t2, a1 1375 addu_s.qb s3, t3, a1 1376 sw s2, 0(a0) 1377 sw s3, 4(a0) 1378 1379 111: 1380 addiu a3, a3, -2 1381 addiu t0, a3, -1 1382 bgtz t0, 1b 1383 addiu a0, a0, 8 1384 b 3f 1385 nop 1386 2: 1387 /* a1 = src */ 1388 lbu t0, 0(a2) /* t0 = mask */ 1389 lbu t1, 1(a2) /* t1 = mask */ 1390 or t2, t0, t1 1391 beqz t2, 222f /* if (t0 == 0) && (t1 == 0) */ 1392 addiu a2, a2, 2 1393 and t3, t0, t1 1394 beq t3, t5, 22f /* if (t0 == 0xff) && (t1 == 0xff) */ 1395 nop 1396 lw t2, 0(a0) /* t2 = dst */ 1397 lw t3, 4(a0) /* t3 = dst */ 1398 1399 OVER_2x8888_2x8_2x8888 a1, a1, t0, t1, t2, t3, \ 1400 t6, t7, t4, t8, t9, s0, s1, s2, s3 1401 sw t6, 0(a0) 1402 b 222f 1403 sw t7, 4(a0) 1404 22: 1405 sw a1, 0(a0) 1406 sw a1, 4(a0) 1407 222: 1408 addiu a3, a3, -2 1409 addiu t0, a3, -1 1410 bgtz t0, 2b 1411 addiu a0, a0, 8 1412 3: 1413 blez a3, 4f 1414 nop 1415 /* a1 = src */ 1416 lbu t0, 0(a2) /* t0 = mask */ 1417 beqz t0, 4f /* if (t0 == 0) */ 1418 addiu a2, a2, 1 1419 move t3, a1 1420 beq t0, t5, 31f /* if (t0 == 0xff) */ 1421 lw t1, 0(a0) /* t1 = dst */ 1422 1423 MIPS_UN8x4_MUL_UN8 a1, t0, t3, t4, t6, t7, t8 1424 31: 1425 not t2, t3 1426 srl t2, t2, 24 1427 MIPS_UN8x4_MUL_UN8 t1, t2, t1, t4, t6, t7, t8 1428 addu_s.qb t2, t1, t3 1429 sw t2, 0(a0) 1430 4: 1431 RESTORE_REGS_FROM_STACK 4, s0, s1, s2, s3, s4 1432 j ra 1433 nop 1434 1435 END(pixman_composite_over_n_8_8888_asm_mips) 1436 1437 LEAF_MIPS_DSPR2(pixman_composite_over_n_8_0565_asm_mips) 1438 /* 1439 * a0 - dst (r5g6b5) 1440 * a1 - src (32bit constant) 1441 * a2 - mask (a8) 1442 * a3 - w 1443 */ 1444 SAVE_REGS_ON_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8 1445 beqz a3, 4f 1446 nop 1447 li t4, 0x00ff00ff 1448 li t5, 0xff 1449 li t6, 0xf800f800 1450 li t7, 0x07e007e0 1451 li t8, 0x001F001F 1452 addiu t1, a3, -1 1453 beqz t1, 3f /* last pixel */ 1454 srl t0, a1, 24 /* t0 = srca */ 1455 not v0, a1 1456 beq t0, t5, 2f /* if (srca == 0xff) */ 1457 srl v0, v0, 24 1458 1: 1459 /* a1 = src */ 1460 lbu t0, 0(a2) /* t0 = mask */ 1461 lbu t1, 1(a2) /* t1 = mask */ 1462 or t2, t0, t1 1463 beqz t2, 111f /* if (t0 == 0) && (t1 == 0) */ 1464 addiu a2, a2, 2 1465 lhu t2, 0(a0) /* t2 = dst */ 1466 lhu t3, 2(a0) /* t3 = dst */ 1467 CONVERT_2x0565_TO_2x8888 t2, t3, s0, s1, t7, t8, t9, s2, s3, s4 1468 and t9, t0, t1 1469 beq t9, t5, 11f /* if (t0 == 0xff) && (t1 == 0xff) */ 1470 nop 1471 1472 MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, s2, s3, t4, t9, s4, s5, s6, s7, s8 1473 not s4, s2 1474 not s5, s3 1475 srl s4, s4, 24 1476 srl s5, s5, 24 1477 MIPS_2xUN8x4_MUL_2xUN8 s0, s1, s4, s5, s0, s1, t4, t9, t0, t1, s6, s7, s8 1478 addu_s.qb s4, s2, s0 1479 addu_s.qb s5, s3, s1 1480 CONVERT_2x8888_TO_2x0565 s4, s5, t2, t3, t6, t7, t8, s0, s1 1481 sh t2, 0(a0) 1482 b 111f 1483 sh t3, 2(a0) 1484 11: 1485 MIPS_2xUN8x4_MUL_2xUN8 s0, s1, v0, v0, s0, s1, t4, t9, t0, t1, s6, s7, s8 1486 addu_s.qb s4, a1, s0 1487 addu_s.qb s5, a1, s1 1488 CONVERT_2x8888_TO_2x0565 s4, s5, t2, t3, t6, t7, t8, s0, s1 1489 sh t2, 0(a0) 1490 sh t3, 2(a0) 1491 111: 1492 addiu a3, a3, -2 1493 addiu t0, a3, -1 1494 bgtz t0, 1b 1495 addiu a0, a0, 4 1496 b 3f 1497 nop 1498 2: 1499 CONVERT_1x8888_TO_1x0565 a1, s0, s1, s2 1500 21: 1501 /* a1 = src */ 1502 lbu t0, 0(a2) /* t0 = mask */ 1503 lbu t1, 1(a2) /* t1 = mask */ 1504 or t2, t0, t1 1505 beqz t2, 222f /* if (t0 == 0) && (t1 == 0) */ 1506 addiu a2, a2, 2 1507 and t9, t0, t1 1508 move s2, s0 1509 beq t9, t5, 22f /* if (t0 == 0xff) && (t2 == 0xff) */ 1510 move s3, s0 1511 lhu t2, 0(a0) /* t2 = dst */ 1512 lhu t3, 2(a0) /* t3 = dst */ 1513 1514 CONVERT_2x0565_TO_2x8888 t2, t3, s2, s3, t7, t8, s4, s5, s6, s7 1515 OVER_2x8888_2x8_2x8888 a1, a1, t0, t1, s2, s3, \ 1516 t2, t3, t4, t9, s4, s5, s6, s7, s8 1517 CONVERT_2x8888_TO_2x0565 t2, t3, s2, s3, t6, t7, t8, s4, s5 1518 22: 1519 sh s2, 0(a0) 1520 sh s3, 2(a0) 1521 222: 1522 addiu a3, a3, -2 1523 addiu t0, a3, -1 1524 bgtz t0, 21b 1525 addiu a0, a0, 4 1526 3: 1527 blez a3, 4f 1528 nop 1529 /* a1 = src */ 1530 lbu t0, 0(a2) /* t0 = mask */ 1531 beqz t0, 4f /* if (t0 == 0) */ 1532 nop 1533 lhu t1, 0(a0) /* t1 = dst */ 1534 CONVERT_1x0565_TO_1x8888 t1, t2, t3, t7 1535 beq t0, t5, 31f /* if (t0 == 0xff) */ 1536 move t3, a1 1537 1538 MIPS_UN8x4_MUL_UN8 a1, t0, t3, t4, t7, t8, t9 1539 31: 1540 not t6, t3 1541 srl t6, t6, 24 1542 MIPS_UN8x4_MUL_UN8 t2, t6, t2, t4, t7, t8, t9 1543 addu_s.qb t1, t2, t3 1544 CONVERT_1x8888_TO_1x0565 t1, t2, t3, t7 1545 sh t2, 0(a0) 1546 4: 1547 RESTORE_REGS_FROM_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8 1548 j ra 1549 nop 1550 1551 END(pixman_composite_over_n_8_0565_asm_mips) 1552 1553 LEAF_MIPS_DSPR2(pixman_composite_over_8888_n_8888_asm_mips) 1554 /* 1555 * a0 - dst (a8r8g8b8) 1556 * a1 - src (a8r8g8b8) 1557 * a2 - mask (32bit constant) 1558 * a3 - w 1559 */ 1560 1561 SAVE_REGS_ON_STACK 0, s0 1562 li t4, 0x00ff00ff 1563 beqz a3, 3f 1564 nop 1565 addiu t1, a3, -1 1566 srl a2, a2, 24 1567 beqz t1, 2f 1568 nop 1569 1570 1: 1571 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1572 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 1573 /* a2 = mask (32bit constant) */ 1574 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 1575 lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */ 1576 addiu a1, a1, 8 1577 1578 OVER_2x8888_2x8_2x8888 t0, t1, a2, a2, t2, t3, \ 1579 t5, t6, t4, t7, t8, t9, t0, t1, s0 1580 1581 sw t5, 0(a0) 1582 sw t6, 4(a0) 1583 addiu a3, a3, -2 1584 addiu t1, a3, -1 1585 bgtz t1, 1b 1586 addiu a0, a0, 8 1587 2: 1588 beqz a3, 3f 1589 nop 1590 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1591 /* a2 = mask (32bit constant) */ 1592 lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */ 1593 1594 OVER_8888_8_8888 t0, a2, t1, t3, t4, t5, t6, t7, t8 1595 1596 sw t3, 0(a0) 1597 3: 1598 RESTORE_REGS_FROM_STACK 0, s0 1599 j ra 1600 nop 1601 1602 END(pixman_composite_over_8888_n_8888_asm_mips) 1603 1604 LEAF_MIPS_DSPR2(pixman_composite_over_8888_n_0565_asm_mips) 1605 /* 1606 * a0 - dst (r5g6b5) 1607 * a1 - src (a8r8g8b8) 1608 * a2 - mask (32bit constant) 1609 * a3 - w 1610 */ 1611 1612 SAVE_REGS_ON_STACK 0, s0, s1, s2, s3 1613 li t6, 0x00ff00ff 1614 li t7, 0xf800f800 1615 li t8, 0x07e007e0 1616 li t9, 0x001F001F 1617 beqz a3, 3f 1618 nop 1619 srl a2, a2, 24 1620 addiu t1, a3, -1 1621 beqz t1, 2f 1622 nop 1623 1: 1624 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1625 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 1626 /* a2 = mask (32bit constant) */ 1627 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 1628 lhu t3, 2(a0) /* t2 = destination (r5g6b5) */ 1629 addiu a1, a1, 8 1630 1631 CONVERT_2x0565_TO_2x8888 t2, t3, t4, t5, t8, t9, s0, s1, t2, t3 1632 OVER_2x8888_2x8_2x8888 t0, t1, a2, a2, t4, t5, \ 1633 t2, t3, t6, t0, t1, s0, s1, s2, s3 1634 CONVERT_2x8888_TO_2x0565 t2, t3, t4, t5, t7, t8, t9, s0, s1 1635 1636 sh t4, 0(a0) 1637 sh t5, 2(a0) 1638 addiu a3, a3, -2 1639 addiu t1, a3, -1 1640 bgtz t1, 1b 1641 addiu a0, a0, 4 1642 2: 1643 beqz a3, 3f 1644 nop 1645 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1646 /* a2 = mask (32bit constant) */ 1647 lhu t1, 0(a0) /* t1 = destination (r5g6b5) */ 1648 1649 CONVERT_1x0565_TO_1x8888 t1, t2, t4, t5 1650 OVER_8888_8_8888 t0, a2, t2, t1, t6, t3, t4, t5, t7 1651 CONVERT_1x8888_TO_1x0565 t1, t3, t4, t5 1652 1653 sh t3, 0(a0) 1654 3: 1655 RESTORE_REGS_FROM_STACK 0, s0, s1, s2, s3 1656 j ra 1657 nop 1658 1659 END(pixman_composite_over_8888_n_0565_asm_mips) 1660 1661 LEAF_MIPS_DSPR2(pixman_composite_over_0565_n_0565_asm_mips) 1662 /* 1663 * a0 - dst (r5g6b5) 1664 * a1 - src (r5g6b5) 1665 * a2 - mask (32bit constant) 1666 * a3 - w 1667 */ 1668 1669 SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5 1670 li t6, 0x00ff00ff 1671 li t7, 0xf800f800 1672 li t8, 0x07e007e0 1673 li t9, 0x001F001F 1674 beqz a3, 3f 1675 nop 1676 srl a2, a2, 24 1677 addiu t1, a3, -1 1678 beqz t1, 2f 1679 nop 1680 1: 1681 lhu t0, 0(a1) /* t0 = source (r5g6b5) */ 1682 lhu t1, 2(a1) /* t1 = source (r5g6b5) */ 1683 /* a2 = mask (32bit constant) */ 1684 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 1685 lhu t3, 2(a0) /* t3 = destination (r5g6b5) */ 1686 addiu a1, a1, 4 1687 1688 CONVERT_2x0565_TO_2x8888 t0, t1, t4, t5, t8, t9, s0, s1, s2, s3 1689 CONVERT_2x0565_TO_2x8888 t2, t3, s0, s1, t8, t9, s2, s3, s4, s5 1690 OVER_2x8888_2x8_2x8888 t4, t5, a2, a2, s0, s1, \ 1691 t0, t1, t6, s2, s3, s4, s5, t4, t5 1692 CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t7, t8, t9, s2, s3 1693 1694 sh s0, 0(a0) 1695 sh s1, 2(a0) 1696 addiu a3, a3, -2 1697 addiu t1, a3, -1 1698 bgtz t1, 1b 1699 addiu a0, a0, 4 1700 2: 1701 beqz a3, 3f 1702 nop 1703 lhu t0, 0(a1) /* t0 = source (r5g6b5) */ 1704 /* a2 = mask (32bit constant) */ 1705 lhu t1, 0(a0) /* t1 = destination (r5g6b5) */ 1706 1707 CONVERT_1x0565_TO_1x8888 t0, t2, t4, t5 1708 CONVERT_1x0565_TO_1x8888 t1, t3, t4, t5 1709 OVER_8888_8_8888 t2, a2, t3, t0, t6, t1, t4, t5, t7 1710 CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5 1711 1712 sh t3, 0(a0) 1713 3: 1714 RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5 1715 j ra 1716 nop 1717 1718 END(pixman_composite_over_0565_n_0565_asm_mips) 1719 1720 LEAF_MIPS_DSPR2(pixman_composite_over_8888_8_8888_asm_mips) 1721 /* 1722 * a0 - dst (a8r8g8b8) 1723 * a1 - src (a8r8g8b8) 1724 * a2 - mask (a8) 1725 * a3 - w 1726 */ 1727 1728 SAVE_REGS_ON_STACK 0, s0, s1 1729 li t4, 0x00ff00ff 1730 beqz a3, 3f 1731 nop 1732 addiu t1, a3, -1 1733 beqz t1, 2f 1734 nop 1735 1: 1736 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1737 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 1738 lbu t2, 0(a2) /* t2 = mask (a8) */ 1739 lbu t3, 1(a2) /* t3 = mask (a8) */ 1740 lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */ 1741 lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */ 1742 addiu a1, a1, 8 1743 addiu a2, a2, 2 1744 1745 OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, t5, t6, \ 1746 t7, t8, t4, t9, s0, s1, t0, t1, t2 1747 1748 sw t7, 0(a0) 1749 sw t8, 4(a0) 1750 addiu a3, a3, -2 1751 addiu t1, a3, -1 1752 bgtz t1, 1b 1753 addiu a0, a0, 8 1754 2: 1755 beqz a3, 3f 1756 nop 1757 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1758 lbu t1, 0(a2) /* t1 = mask (a8) */ 1759 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 1760 1761 OVER_8888_8_8888 t0, t1, t2, t3, t4, t5, t6, t7, t8 1762 1763 sw t3, 0(a0) 1764 3: 1765 RESTORE_REGS_FROM_STACK 0, s0, s1 1766 j ra 1767 nop 1768 1769 END(pixman_composite_over_8888_8_8888_asm_mips) 1770 1771 LEAF_MIPS_DSPR2(pixman_composite_over_8888_8_0565_asm_mips) 1772 /* 1773 * a0 - dst (r5g6b5) 1774 * a1 - src (a8r8g8b8) 1775 * a2 - mask (a8) 1776 * a3 - w 1777 */ 1778 1779 SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5 1780 li t6, 0x00ff00ff 1781 li t7, 0xf800f800 1782 li t8, 0x07e007e0 1783 li t9, 0x001F001F 1784 beqz a3, 3f 1785 nop 1786 addiu t1, a3, -1 1787 beqz t1, 2f 1788 nop 1789 1: 1790 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1791 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 1792 lbu t2, 0(a2) /* t2 = mask (a8) */ 1793 lbu t3, 1(a2) /* t3 = mask (a8) */ 1794 lhu t4, 0(a0) /* t4 = destination (r5g6b5) */ 1795 lhu t5, 2(a0) /* t5 = destination (r5g6b5) */ 1796 addiu a1, a1, 8 1797 addiu a2, a2, 2 1798 1799 CONVERT_2x0565_TO_2x8888 t4, t5, s0, s1, t8, t9, s2, s3, s4, s5 1800 OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, s0, s1, \ 1801 t4, t5, t6, s2, s3, s4, s5, t0, t1 1802 CONVERT_2x8888_TO_2x0565 t4, t5, s0, s1, t7, t8, t9, s2, s3 1803 1804 sh s0, 0(a0) 1805 sh s1, 2(a0) 1806 addiu a3, a3, -2 1807 addiu t1, a3, -1 1808 bgtz t1, 1b 1809 addiu a0, a0, 4 1810 2: 1811 beqz a3, 3f 1812 nop 1813 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1814 lbu t1, 0(a2) /* t1 = mask (a8) */ 1815 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 1816 1817 CONVERT_1x0565_TO_1x8888 t2, t3, t4, t5 1818 OVER_8888_8_8888 t0, t1, t3, t2, t6, t4, t5, t7, t8 1819 CONVERT_1x8888_TO_1x0565 t2, t3, t4, t5 1820 1821 sh t3, 0(a0) 1822 3: 1823 RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5 1824 j ra 1825 nop 1826 1827 END(pixman_composite_over_8888_8_0565_asm_mips) 1828 1829 LEAF_MIPS_DSPR2(pixman_composite_over_0565_8_0565_asm_mips) 1830 /* 1831 * a0 - dst (r5g6b5) 1832 * a1 - src (r5g6b5) 1833 * a2 - mask (a8) 1834 * a3 - w 1835 */ 1836 1837 SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5 1838 li t4, 0xf800f800 1839 li t5, 0x07e007e0 1840 li t6, 0x001F001F 1841 li t7, 0x00ff00ff 1842 beqz a3, 3f 1843 nop 1844 addiu t1, a3, -1 1845 beqz t1, 2f 1846 nop 1847 1: 1848 lhu t0, 0(a1) /* t0 = source (r5g6b5) */ 1849 lhu t1, 2(a1) /* t1 = source (r5g6b5) */ 1850 lbu t2, 0(a2) /* t2 = mask (a8) */ 1851 lbu t3, 1(a2) /* t3 = mask (a8) */ 1852 lhu t8, 0(a0) /* t8 = destination (r5g6b5) */ 1853 lhu t9, 2(a0) /* t9 = destination (r5g6b5) */ 1854 addiu a1, a1, 4 1855 addiu a2, a2, 2 1856 1857 CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5 1858 CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, t0, t1 1859 OVER_2x8888_2x8_2x8888 s0, s1, t2, t3, s2, s3, \ 1860 t0, t1, t7, s4, s5, t8, t9, s0, s1 1861 CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3 1862 1863 sh s0, 0(a0) 1864 sh s1, 2(a0) 1865 addiu a3, a3, -2 1866 addiu t1, a3, -1 1867 bgtz t1, 1b 1868 addiu a0, a0, 4 1869 2: 1870 beqz a3, 3f 1871 nop 1872 lhu t0, 0(a1) /* t0 = source (r5g6b5) */ 1873 lbu t1, 0(a2) /* t1 = mask (a8) */ 1874 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 1875 1876 CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5 1877 CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6 1878 OVER_8888_8_8888 t3, t1, t4, t0, t7, t2, t5, t6, t8 1879 CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5 1880 1881 sh t3, 0(a0) 1882 3: 1883 RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5 1884 j ra 1885 nop 1886 1887 END(pixman_composite_over_0565_8_0565_asm_mips) 1888 1889 LEAF_MIPS_DSPR2(pixman_composite_over_8888_8888_8888_asm_mips) 1890 /* 1891 * a0 - dst (a8r8g8b8) 1892 * a1 - src (a8r8g8b8) 1893 * a2 - mask (a8r8g8b8) 1894 * a3 - w 1895 */ 1896 1897 SAVE_REGS_ON_STACK 0, s0, s1, s2 1898 li t4, 0x00ff00ff 1899 beqz a3, 3f 1900 nop 1901 addiu t1, a3, -1 1902 beqz t1, 2f 1903 nop 1904 1: 1905 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1906 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 1907 lw t2, 0(a2) /* t2 = mask (a8r8g8b8) */ 1908 lw t3, 4(a2) /* t3 = mask (a8r8g8b8) */ 1909 lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */ 1910 lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */ 1911 addiu a1, a1, 8 1912 addiu a2, a2, 8 1913 srl t2, t2, 24 1914 srl t3, t3, 24 1915 1916 OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, t5, t6, t7, t8, t4, t9, s0, s1, s2, t0, t1 1917 1918 sw t7, 0(a0) 1919 sw t8, 4(a0) 1920 addiu a3, a3, -2 1921 addiu t1, a3, -1 1922 bgtz t1, 1b 1923 addiu a0, a0, 8 1924 2: 1925 beqz a3, 3f 1926 nop 1927 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1928 lw t1, 0(a2) /* t1 = mask (a8r8g8b8) */ 1929 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 1930 srl t1, t1, 24 1931 1932 OVER_8888_8_8888 t0, t1, t2, t3, t4, t5, t6, t7, t8 1933 1934 sw t3, 0(a0) 1935 3: 1936 RESTORE_REGS_FROM_STACK 0, s0, s1, s2 1937 j ra 1938 nop 1939 1940 END(pixman_composite_over_8888_8888_8888_asm_mips) 1941 1942 LEAF_MIPS_DSPR2(pixman_composite_over_8888_8888_asm_mips) 1943 /* 1944 * a0 - dst (a8r8g8b8) 1945 * a1 - src (a8r8g8b8) 1946 * a2 - w 1947 */ 1948 1949 SAVE_REGS_ON_STACK 0, s0, s1, s2 1950 li t4, 0x00ff00ff 1951 beqz a2, 3f 1952 nop 1953 addiu t1, a2, -1 1954 beqz t1, 2f 1955 nop 1956 1: 1957 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1958 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 1959 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 1960 lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */ 1961 addiu a1, a1, 8 1962 1963 not t5, t0 1964 srl t5, t5, 24 1965 not t6, t1 1966 srl t6, t6, 24 1967 1968 or t7, t5, t6 1969 beqz t7, 11f 1970 or t8, t0, t1 1971 beqz t8, 12f 1972 1973 MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t5, t6, t7, t8, t4, t9, s0, s1, s2, t2, t3 1974 1975 addu_s.qb t0, t7, t0 1976 addu_s.qb t1, t8, t1 1977 11: 1978 sw t0, 0(a0) 1979 sw t1, 4(a0) 1980 12: 1981 addiu a2, a2, -2 1982 addiu t1, a2, -1 1983 bgtz t1, 1b 1984 addiu a0, a0, 8 1985 2: 1986 beqz a2, 3f 1987 nop 1988 1989 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 1990 lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */ 1991 addiu a1, a1, 4 1992 1993 not t2, t0 1994 srl t2, t2, 24 1995 1996 beqz t2, 21f 1997 nop 1998 beqz t0, 3f 1999 2000 MIPS_UN8x4_MUL_UN8 t1, t2, t3, t4, t5, t6, t7 2001 2002 addu_s.qb t0, t3, t0 2003 21: 2004 sw t0, 0(a0) 2005 2006 3: 2007 RESTORE_REGS_FROM_STACK 0, s0, s1, s2 2008 j ra 2009 nop 2010 2011 END(pixman_composite_over_8888_8888_asm_mips) 2012 2013 LEAF_MIPS_DSPR2(pixman_composite_over_8888_0565_asm_mips) 2014 /* 2015 * a0 - dst (r5g6b5) 2016 * a1 - src (a8r8g8b8) 2017 * a2 - w 2018 */ 2019 2020 SAVE_REGS_ON_STACK 8, s0, s1, s2, s3, s4, s5 2021 li t4, 0x00ff00ff 2022 li s3, 0xf800f800 2023 li s4, 0x07e007e0 2024 li s5, 0x001F001F 2025 beqz a2, 3f 2026 nop 2027 addiu t1, a2, -1 2028 beqz t1, 2f 2029 nop 2030 1: 2031 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 2032 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 2033 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 2034 lhu t3, 2(a0) /* t3 = destination (r5g6b5) */ 2035 addiu a1, a1, 8 2036 2037 not t5, t0 2038 srl t5, t5, 24 2039 not t6, t1 2040 srl t6, t6, 24 2041 2042 or t7, t5, t6 2043 beqz t7, 11f 2044 or t8, t0, t1 2045 beqz t8, 12f 2046 2047 CONVERT_2x0565_TO_2x8888 t2, t3, s0, s1, s4, s5, t7, t8, t9, s2 2048 MIPS_2xUN8x4_MUL_2xUN8 s0, s1, t5, t6, t7, t8, t4, t9, t2, t3, s2, s0, s1 2049 2050 addu_s.qb t0, t7, t0 2051 addu_s.qb t1, t8, t1 2052 11: 2053 CONVERT_2x8888_TO_2x0565 t0, t1, t7, t8, s3, s4, s5, t2, t3 2054 sh t7, 0(a0) 2055 sh t8, 2(a0) 2056 12: 2057 addiu a2, a2, -2 2058 addiu t1, a2, -1 2059 bgtz t1, 1b 2060 addiu a0, a0, 4 2061 2: 2062 beqz a2, 3f 2063 nop 2064 2065 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 2066 lhu t1, 0(a0) /* t1 = destination (r5g6b5) */ 2067 addiu a1, a1, 4 2068 2069 not t2, t0 2070 srl t2, t2, 24 2071 2072 beqz t2, 21f 2073 nop 2074 beqz t0, 3f 2075 2076 CONVERT_1x0565_TO_1x8888 t1, s0, t8, t9 2077 MIPS_UN8x4_MUL_UN8 s0, t2, t3, t4, t5, t6, t7 2078 2079 addu_s.qb t0, t3, t0 2080 21: 2081 CONVERT_1x8888_TO_1x0565 t0, s0, t8, t9 2082 sh s0, 0(a0) 2083 2084 3: 2085 RESTORE_REGS_FROM_STACK 8, s0, s1, s2, s3, s4, s5 2086 j ra 2087 nop 2088 2089 END(pixman_composite_over_8888_0565_asm_mips) 2090 2091 LEAF_MIPS_DSPR2(pixman_composite_over_n_0565_asm_mips) 2092 /* 2093 * a0 - dst (r5g6b5) 2094 * a1 - src (32bit constant) 2095 * a2 - w 2096 */ 2097 2098 beqz a2, 5f 2099 nop 2100 2101 not t0, a1 2102 srl t0, t0, 24 2103 bgtz t0, 1f 2104 nop 2105 CONVERT_1x8888_TO_1x0565 a1, t1, t2, t3 2106 0: 2107 sh t1, 0(a0) 2108 addiu a2, a2, -1 2109 bgtz a2, 0b 2110 addiu a0, a0, 2 2111 j ra 2112 nop 2113 2114 1: 2115 SAVE_REGS_ON_STACK 0, s0, s1, s2 2116 li t4, 0x00ff00ff 2117 li t5, 0xf800f800 2118 li t6, 0x07e007e0 2119 li t7, 0x001F001F 2120 addiu t1, a2, -1 2121 beqz t1, 3f 2122 nop 2123 2: 2124 lhu t1, 0(a0) /* t1 = destination (r5g6b5) */ 2125 lhu t2, 2(a0) /* t2 = destination (r5g6b5) */ 2126 2127 CONVERT_2x0565_TO_2x8888 t1, t2, t3, t8, t6, t7, t9, s0, s1, s2 2128 MIPS_2xUN8x4_MUL_2xUN8 t3, t8, t0, t0, t1, t2, t4, t9, s0, s1, s2, t3, t8 2129 addu_s.qb t1, t1, a1 2130 addu_s.qb t2, t2, a1 2131 CONVERT_2x8888_TO_2x0565 t1, t2, t3, t8, t5, t6, t7, s0, s1 2132 2133 sh t3, 0(a0) 2134 sh t8, 2(a0) 2135 2136 addiu a2, a2, -2 2137 addiu t1, a2, -1 2138 bgtz t1, 2b 2139 addiu a0, a0, 4 2140 3: 2141 beqz a2, 4f 2142 nop 2143 2144 lhu t1, 0(a0) /* t1 = destination (r5g6b5) */ 2145 2146 CONVERT_1x0565_TO_1x8888 t1, t2, s0, s1 2147 MIPS_UN8x4_MUL_UN8 t2, t0, t1, t4, s0, s1, s2 2148 addu_s.qb t1, t1, a1 2149 CONVERT_1x8888_TO_1x0565 t1, t2, s0, s1 2150 2151 sh t2, 0(a0) 2152 2153 4: 2154 RESTORE_REGS_FROM_STACK 0, s0, s1, s2 2155 5: 2156 j ra 2157 nop 2158 2159 END(pixman_composite_over_n_0565_asm_mips) 2160 2161 LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_asm_mips) 2162 /* 2163 * a0 - dst (a8r8g8b8) 2164 * a1 - src (32bit constant) 2165 * a2 - w 2166 */ 2167 2168 beqz a2, 5f 2169 nop 2170 2171 not t0, a1 2172 srl t0, t0, 24 2173 bgtz t0, 1f 2174 nop 2175 0: 2176 sw a1, 0(a0) 2177 addiu a2, a2, -1 2178 bgtz a2, 0b 2179 addiu a0, a0, 4 2180 j ra 2181 nop 2182 2183 1: 2184 SAVE_REGS_ON_STACK 0, s0, s1, s2 2185 li t4, 0x00ff00ff 2186 addiu t1, a2, -1 2187 beqz t1, 3f 2188 nop 2189 2: 2190 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 2191 lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */ 2192 2193 MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t0, t0, t7, t8, t4, t9, s0, s1, s2, t2, t3 2194 2195 addu_s.qb t7, t7, a1 2196 addu_s.qb t8, t8, a1 2197 2198 sw t7, 0(a0) 2199 sw t8, 4(a0) 2200 2201 addiu a2, a2, -2 2202 addiu t1, a2, -1 2203 bgtz t1, 2b 2204 addiu a0, a0, 8 2205 3: 2206 beqz a2, 4f 2207 nop 2208 2209 lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */ 2210 2211 MIPS_UN8x4_MUL_UN8 t1, t0, t3, t4, t5, t6, t7 2212 2213 addu_s.qb t3, t3, a1 2214 2215 sw t3, 0(a0) 2216 2217 4: 2218 RESTORE_REGS_FROM_STACK 0, s0, s1, s2 2219 5: 2220 j ra 2221 nop 2222 2223 END(pixman_composite_over_n_8888_asm_mips) 2224 2225 LEAF_MIPS_DSPR2(pixman_composite_add_8_8_8_asm_mips) 2226 /* 2227 * a0 - dst (a8) 2228 * a1 - src (a8) 2229 * a2 - mask (a8) 2230 * a3 - w 2231 */ 2232 2233 SAVE_REGS_ON_STACK 0, v0, v1 2234 li t9, 0x00ff00ff 2235 beqz a3, 3f 2236 nop 2237 2238 srl v0, a3, 2 /* v0 = how many multiples of 4 dst pixels */ 2239 beqz v0, 1f /* branch if less than 4 src pixels */ 2240 nop 2241 2242 0: 2243 beqz v0, 1f 2244 addiu v0, v0, -1 2245 lbu t0, 0(a2) 2246 lbu t1, 1(a2) 2247 lbu t2, 2(a2) 2248 lbu t3, 3(a2) 2249 lbu t4, 0(a0) 2250 lbu t5, 1(a0) 2251 lbu t6, 2(a0) 2252 lbu t7, 3(a0) 2253 2254 addiu a2, a2, 4 2255 2256 precr_sra.ph.w t1, t0, 0 2257 precr_sra.ph.w t3, t2, 0 2258 precr_sra.ph.w t5, t4, 0 2259 precr_sra.ph.w t7, t6, 0 2260 2261 precr.qb.ph t0, t3, t1 2262 precr.qb.ph t1, t7, t5 2263 2264 lbu t4, 0(a1) 2265 lbu v1, 1(a1) 2266 lbu t7, 2(a1) 2267 lbu t8, 3(a1) 2268 2269 addiu a1, a1, 4 2270 2271 precr_sra.ph.w v1, t4, 0 2272 precr_sra.ph.w t8, t7, 0 2273 2274 muleu_s.ph.qbl t2, t0, t8 2275 muleu_s.ph.qbr t3, t0, v1 2276 shra_r.ph t4, t2, 8 2277 shra_r.ph t5, t3, 8 2278 and t4, t4, t9 2279 and t5, t5, t9 2280 addq.ph t2, t2, t4 2281 addq.ph t3, t3, t5 2282 shra_r.ph t2, t2, 8 2283 shra_r.ph t3, t3, 8 2284 precr.qb.ph t0, t2, t3 2285 2286 addu_s.qb t2, t0, t1 2287 2288 sb t2, 0(a0) 2289 srl t2, t2, 8 2290 sb t2, 1(a0) 2291 srl t2, t2, 8 2292 sb t2, 2(a0) 2293 srl t2, t2, 8 2294 sb t2, 3(a0) 2295 addiu a3, a3, -4 2296 b 0b 2297 addiu a0, a0, 4 2298 2299 1: 2300 beqz a3, 3f 2301 nop 2302 2: 2303 lbu t8, 0(a1) 2304 lbu t0, 0(a2) 2305 lbu t1, 0(a0) 2306 addiu a1, a1, 1 2307 addiu a2, a2, 1 2308 2309 mul t2, t0, t8 2310 shra_r.ph t3, t2, 8 2311 andi t3, t3, 0xff 2312 addq.ph t2, t2, t3 2313 shra_r.ph t2, t2, 8 2314 andi t2, t2, 0xff 2315 2316 addu_s.qb t2, t2, t1 2317 sb t2, 0(a0) 2318 addiu a3, a3, -1 2319 bnez a3, 2b 2320 addiu a0, a0, 1 2321 2322 3: 2323 RESTORE_REGS_FROM_STACK 0, v0, v1 2324 j ra 2325 nop 2326 2327 END(pixman_composite_add_8_8_8_asm_mips) 2328 2329 LEAF_MIPS_DSPR2(pixman_composite_add_n_8_8_asm_mips) 2330 /* 2331 * a0 - dst (a8) 2332 * a1 - src (32bit constant) 2333 * a2 - mask (a8) 2334 * a3 - w 2335 */ 2336 2337 SAVE_REGS_ON_STACK 0, v0 2338 li t9, 0x00ff00ff 2339 beqz a3, 3f 2340 nop 2341 2342 srl v0, a3, 2 /* v0 = how many multiples of 4 dst pixels */ 2343 beqz v0, 1f /* branch if less than 4 src pixels */ 2344 nop 2345 2346 srl t8, a1, 24 2347 replv.ph t8, t8 2348 2349 0: 2350 beqz v0, 1f 2351 addiu v0, v0, -1 2352 lbu t0, 0(a2) 2353 lbu t1, 1(a2) 2354 lbu t2, 2(a2) 2355 lbu t3, 3(a2) 2356 lbu t4, 0(a0) 2357 lbu t5, 1(a0) 2358 lbu t6, 2(a0) 2359 lbu t7, 3(a0) 2360 2361 addiu a2, a2, 4 2362 2363 precr_sra.ph.w t1, t0, 0 2364 precr_sra.ph.w t3, t2, 0 2365 precr_sra.ph.w t5, t4, 0 2366 precr_sra.ph.w t7, t6, 0 2367 2368 precr.qb.ph t0, t3, t1 2369 precr.qb.ph t1, t7, t5 2370 2371 muleu_s.ph.qbl t2, t0, t8 2372 muleu_s.ph.qbr t3, t0, t8 2373 shra_r.ph t4, t2, 8 2374 shra_r.ph t5, t3, 8 2375 and t4, t4, t9 2376 and t5, t5, t9 2377 addq.ph t2, t2, t4 2378 addq.ph t3, t3, t5 2379 shra_r.ph t2, t2, 8 2380 shra_r.ph t3, t3, 8 2381 precr.qb.ph t0, t2, t3 2382 2383 addu_s.qb t2, t0, t1 2384 2385 sb t2, 0(a0) 2386 srl t2, t2, 8 2387 sb t2, 1(a0) 2388 srl t2, t2, 8 2389 sb t2, 2(a0) 2390 srl t2, t2, 8 2391 sb t2, 3(a0) 2392 addiu a3, a3, -4 2393 b 0b 2394 addiu a0, a0, 4 2395 2396 1: 2397 beqz a3, 3f 2398 nop 2399 srl t8, a1, 24 2400 2: 2401 lbu t0, 0(a2) 2402 lbu t1, 0(a0) 2403 addiu a2, a2, 1 2404 2405 mul t2, t0, t8 2406 shra_r.ph t3, t2, 8 2407 andi t3, t3, 0xff 2408 addq.ph t2, t2, t3 2409 shra_r.ph t2, t2, 8 2410 andi t2, t2, 0xff 2411 2412 addu_s.qb t2, t2, t1 2413 sb t2, 0(a0) 2414 addiu a3, a3, -1 2415 bnez a3, 2b 2416 addiu a0, a0, 1 2417 2418 3: 2419 RESTORE_REGS_FROM_STACK 0, v0 2420 j ra 2421 nop 2422 2423 END(pixman_composite_add_n_8_8_asm_mips) 2424 2425 LEAF_MIPS_DSPR2(pixman_composite_add_n_8_8888_asm_mips) 2426 /* 2427 * a0 - dst (a8r8g8b8) 2428 * a1 - src (32bit constant) 2429 * a2 - mask (a8) 2430 * a3 - w 2431 */ 2432 2433 SAVE_REGS_ON_STACK 0, s0, s1, s2 2434 li t4, 0x00ff00ff 2435 beqz a3, 3f 2436 nop 2437 addiu t1, a3, -1 2438 beqz t1, 2f 2439 nop 2440 1: 2441 /* a1 = source (32bit constant) */ 2442 lbu t0, 0(a2) /* t0 = mask (a8) */ 2443 lbu t1, 1(a2) /* t1 = mask (a8) */ 2444 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 2445 lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */ 2446 addiu a2, a2, 2 2447 2448 MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 a1, a1, \ 2449 t0, t1, \ 2450 t2, t3, \ 2451 t5, t6, \ 2452 t4, t7, t8, t9, s0, s1, s2 2453 2454 sw t5, 0(a0) 2455 sw t6, 4(a0) 2456 addiu a3, a3, -2 2457 addiu t1, a3, -1 2458 bgtz t1, 1b 2459 addiu a0, a0, 8 2460 2: 2461 beqz a3, 3f 2462 nop 2463 /* a1 = source (32bit constant) */ 2464 lbu t0, 0(a2) /* t0 = mask (a8) */ 2465 lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */ 2466 2467 MIPS_UN8x4_MUL_UN8_ADD_UN8x4 a1, t0, t1, t2, t4, t3, t5, t6 2468 2469 sw t2, 0(a0) 2470 3: 2471 RESTORE_REGS_FROM_STACK 0, s0, s1, s2 2472 j ra 2473 nop 2474 2475 END(pixman_composite_add_n_8_8888_asm_mips) 2476 2477 LEAF_MIPS_DSPR2(pixman_composite_add_0565_8_0565_asm_mips) 2478 /* 2479 * a0 - dst (r5g6b5) 2480 * a1 - src (r5g6b5) 2481 * a2 - mask (a8) 2482 * a3 - w 2483 */ 2484 2485 SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7 2486 li t4, 0xf800f800 2487 li t5, 0x07e007e0 2488 li t6, 0x001F001F 2489 li t7, 0x00ff00ff 2490 beqz a3, 3f 2491 nop 2492 addiu t1, a3, -1 2493 beqz t1, 2f 2494 nop 2495 1: 2496 lhu t0, 0(a1) /* t0 = source (r5g6b5) */ 2497 lhu t1, 2(a1) /* t1 = source (r5g6b5) */ 2498 lbu t2, 0(a2) /* t2 = mask (a8) */ 2499 lbu t3, 1(a2) /* t3 = mask (a8) */ 2500 lhu t8, 0(a0) /* t8 = destination (r5g6b5) */ 2501 lhu t9, 2(a0) /* t9 = destination (r5g6b5) */ 2502 addiu a1, a1, 4 2503 addiu a2, a2, 2 2504 2505 CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5 2506 CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, s6, s7 2507 MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 s0, s1, \ 2508 t2, t3, \ 2509 s2, s3, \ 2510 t0, t1, \ 2511 t7, s4, s5, s6, s7, t8, t9 2512 CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3 2513 2514 sh s0, 0(a0) 2515 sh s1, 2(a0) 2516 addiu a3, a3, -2 2517 addiu t1, a3, -1 2518 bgtz t1, 1b 2519 addiu a0, a0, 4 2520 2: 2521 beqz a3, 3f 2522 nop 2523 lhu t0, 0(a1) /* t0 = source (r5g6b5) */ 2524 lbu t1, 0(a2) /* t1 = mask (a8) */ 2525 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 2526 2527 CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5 2528 CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6 2529 MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t3, t1, t4, t0, t7, t2, t5, t6 2530 CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5 2531 2532 sh t3, 0(a0) 2533 3: 2534 RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7 2535 j ra 2536 nop 2537 2538 END(pixman_composite_add_0565_8_0565_asm_mips) 2539 2540 LEAF_MIPS_DSPR2(pixman_composite_add_8888_8_8888_asm_mips) 2541 /* 2542 * a0 - dst (a8r8g8b8) 2543 * a1 - src (a8r8g8b8) 2544 * a2 - mask (a8) 2545 * a3 - w 2546 */ 2547 2548 SAVE_REGS_ON_STACK 0, s0, s1, s2 2549 li t4, 0x00ff00ff 2550 beqz a3, 3f 2551 nop 2552 addiu t1, a3, -1 2553 beqz t1, 2f 2554 nop 2555 1: 2556 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 2557 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 2558 lbu t2, 0(a2) /* t2 = mask (a8) */ 2559 lbu t3, 1(a2) /* t3 = mask (a8) */ 2560 lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */ 2561 lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */ 2562 addiu a1, a1, 8 2563 addiu a2, a2, 2 2564 2565 MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \ 2566 t2, t3, \ 2567 t5, t6, \ 2568 t7, t8, \ 2569 t4, t9, s0, s1, s2, t0, t1 2570 2571 sw t7, 0(a0) 2572 sw t8, 4(a0) 2573 addiu a3, a3, -2 2574 addiu t1, a3, -1 2575 bgtz t1, 1b 2576 addiu a0, a0, 8 2577 2: 2578 beqz a3, 3f 2579 nop 2580 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 2581 lbu t1, 0(a2) /* t1 = mask (a8) */ 2582 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 2583 2584 MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t3, t4, t5, t6, t7 2585 2586 sw t3, 0(a0) 2587 3: 2588 RESTORE_REGS_FROM_STACK 0, s0, s1, s2 2589 j ra 2590 nop 2591 2592 END(pixman_composite_add_8888_8_8888_asm_mips) 2593 2594 LEAF_MIPS_DSPR2(pixman_composite_add_8888_n_8888_asm_mips) 2595 /* 2596 * a0 - dst (a8r8g8b8) 2597 * a1 - src (a8r8g8b8) 2598 * a2 - mask (32bit constant) 2599 * a3 - w 2600 */ 2601 2602 SAVE_REGS_ON_STACK 0, s0, s1, s2 2603 li t4, 0x00ff00ff 2604 beqz a3, 3f 2605 nop 2606 srl a2, a2, 24 2607 addiu t1, a3, -1 2608 beqz t1, 2f 2609 nop 2610 1: 2611 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 2612 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 2613 /* a2 = mask (32bit constant) */ 2614 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 2615 lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */ 2616 addiu a1, a1, 8 2617 2618 MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \ 2619 a2, a2, \ 2620 t2, t3, \ 2621 t5, t6, \ 2622 t4, t7, t8, t9, s0, s1, s2 2623 2624 sw t5, 0(a0) 2625 sw t6, 4(a0) 2626 addiu a3, a3, -2 2627 addiu t1, a3, -1 2628 bgtz t1, 1b 2629 addiu a0, a0, 8 2630 2: 2631 beqz a3, 3f 2632 nop 2633 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 2634 /* a2 = mask (32bit constant) */ 2635 lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */ 2636 2637 MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, a2, t1, t3, t4, t5, t6, t7 2638 2639 sw t3, 0(a0) 2640 3: 2641 RESTORE_REGS_FROM_STACK 0, s0, s1, s2 2642 j ra 2643 nop 2644 2645 END(pixman_composite_add_8888_n_8888_asm_mips) 2646 2647 LEAF_MIPS_DSPR2(pixman_composite_add_8888_8888_8888_asm_mips) 2648 /* 2649 * a0 - dst (a8r8g8b8) 2650 * a1 - src (a8r8g8b8) 2651 * a2 - mask (a8r8g8b8) 2652 * a3 - w 2653 */ 2654 2655 SAVE_REGS_ON_STACK 0, s0, s1, s2 2656 li t4, 0x00ff00ff 2657 beqz a3, 3f 2658 nop 2659 addiu t1, a3, -1 2660 beqz t1, 2f 2661 nop 2662 1: 2663 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 2664 lw t1, 4(a1) /* t1 = source (a8r8g8b8) */ 2665 lw t2, 0(a2) /* t2 = mask (a8r8g8b8) */ 2666 lw t3, 4(a2) /* t3 = mask (a8r8g8b8) */ 2667 lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */ 2668 lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */ 2669 addiu a1, a1, 8 2670 addiu a2, a2, 8 2671 srl t2, t2, 24 2672 srl t3, t3, 24 2673 2674 MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \ 2675 t2, t3, \ 2676 t5, t6, \ 2677 t7, t8, \ 2678 t4, t9, s0, s1, s2, t0, t1 2679 2680 sw t7, 0(a0) 2681 sw t8, 4(a0) 2682 addiu a3, a3, -2 2683 addiu t1, a3, -1 2684 bgtz t1, 1b 2685 addiu a0, a0, 8 2686 2: 2687 beqz a3, 3f 2688 nop 2689 lw t0, 0(a1) /* t0 = source (a8r8g8b8) */ 2690 lw t1, 0(a2) /* t1 = mask (a8r8g8b8) */ 2691 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 2692 srl t1, t1, 24 2693 2694 MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t3, t4, t5, t6, t7 2695 2696 sw t3, 0(a0) 2697 3: 2698 RESTORE_REGS_FROM_STACK 0, s0, s1, s2 2699 j ra 2700 nop 2701 2702 END(pixman_composite_add_8888_8888_8888_asm_mips) 2703 2704 LEAF_MIPS_DSPR2(pixman_composite_add_8_8_asm_mips) 2705 /* 2706 * a0 - dst (a8) 2707 * a1 - src (a8) 2708 * a2 - w 2709 */ 2710 2711 beqz a2, 3f 2712 nop 2713 srl t9, a2, 2 /* t9 = how many multiples of 4 dst pixels */ 2714 beqz t9, 1f /* branch if less than 4 src pixels */ 2715 nop 2716 2717 0: 2718 beqz t9, 1f 2719 addiu t9, t9, -1 2720 lbu t0, 0(a1) 2721 lbu t1, 1(a1) 2722 lbu t2, 2(a1) 2723 lbu t3, 3(a1) 2724 lbu t4, 0(a0) 2725 lbu t5, 1(a0) 2726 lbu t6, 2(a0) 2727 lbu t7, 3(a0) 2728 2729 addiu a1, a1, 4 2730 2731 precr_sra.ph.w t1, t0, 0 2732 precr_sra.ph.w t3, t2, 0 2733 precr_sra.ph.w t5, t4, 0 2734 precr_sra.ph.w t7, t6, 0 2735 2736 precr.qb.ph t0, t3, t1 2737 precr.qb.ph t1, t7, t5 2738 2739 addu_s.qb t2, t0, t1 2740 2741 sb t2, 0(a0) 2742 srl t2, t2, 8 2743 sb t2, 1(a0) 2744 srl t2, t2, 8 2745 sb t2, 2(a0) 2746 srl t2, t2, 8 2747 sb t2, 3(a0) 2748 addiu a2, a2, -4 2749 b 0b 2750 addiu a0, a0, 4 2751 2752 1: 2753 beqz a2, 3f 2754 nop 2755 2: 2756 lbu t0, 0(a1) 2757 lbu t1, 0(a0) 2758 addiu a1, a1, 1 2759 2760 addu_s.qb t2, t0, t1 2761 sb t2, 0(a0) 2762 addiu a2, a2, -1 2763 bnez a2, 2b 2764 addiu a0, a0, 1 2765 2766 3: 2767 j ra 2768 nop 2769 2770 END(pixman_composite_add_8_8_asm_mips) 2771 2772 LEAF_MIPS_DSPR2(pixman_composite_add_8888_8888_asm_mips) 2773 /* 2774 * a0 - dst (a8r8g8b8) 2775 * a1 - src (a8r8g8b8) 2776 * a2 - w 2777 */ 2778 2779 beqz a2, 4f 2780 nop 2781 2782 srl t9, a2, 2 /* t1 = how many multiples of 4 src pixels */ 2783 beqz t9, 3f /* branch if less than 4 src pixels */ 2784 nop 2785 1: 2786 addiu t9, t9, -1 2787 beqz t9, 2f 2788 addiu a2, a2, -4 2789 2790 lw t0, 0(a1) 2791 lw t1, 4(a1) 2792 lw t2, 8(a1) 2793 lw t3, 12(a1) 2794 lw t4, 0(a0) 2795 lw t5, 4(a0) 2796 lw t6, 8(a0) 2797 lw t7, 12(a0) 2798 addiu a1, a1, 16 2799 2800 addu_s.qb t4, t4, t0 2801 addu_s.qb t5, t5, t1 2802 addu_s.qb t6, t6, t2 2803 addu_s.qb t7, t7, t3 2804 2805 sw t4, 0(a0) 2806 sw t5, 4(a0) 2807 sw t6, 8(a0) 2808 sw t7, 12(a0) 2809 b 1b 2810 addiu a0, a0, 16 2811 2: 2812 lw t0, 0(a1) 2813 lw t1, 4(a1) 2814 lw t2, 8(a1) 2815 lw t3, 12(a1) 2816 lw t4, 0(a0) 2817 lw t5, 4(a0) 2818 lw t6, 8(a0) 2819 lw t7, 12(a0) 2820 addiu a1, a1, 16 2821 2822 addu_s.qb t4, t4, t0 2823 addu_s.qb t5, t5, t1 2824 addu_s.qb t6, t6, t2 2825 addu_s.qb t7, t7, t3 2826 2827 sw t4, 0(a0) 2828 sw t5, 4(a0) 2829 sw t6, 8(a0) 2830 sw t7, 12(a0) 2831 2832 beqz a2, 4f 2833 addiu a0, a0, 16 2834 3: 2835 lw t0, 0(a1) 2836 lw t1, 0(a0) 2837 addiu a1, a1, 4 2838 addiu a2, a2, -1 2839 addu_s.qb t1, t1, t0 2840 sw t1, 0(a0) 2841 bnez a2, 3b 2842 addiu a0, a0, 4 2843 4: 2844 jr ra 2845 nop 2846 2847 END(pixman_composite_add_8888_8888_asm_mips) 2848 2849 LEAF_MIPS_DSPR2(pixman_composite_out_reverse_8_0565_asm_mips) 2850 /* 2851 * a0 - dst (r5g6b5) 2852 * a1 - src (a8) 2853 * a2 - w 2854 */ 2855 2856 beqz a2, 4f 2857 nop 2858 2859 SAVE_REGS_ON_STACK 0, s0, s1, s2, s3 2860 li t2, 0xf800f800 2861 li t3, 0x07e007e0 2862 li t4, 0x001F001F 2863 li t5, 0x00ff00ff 2864 2865 addiu t1, a2, -1 2866 beqz t1, 2f 2867 nop 2868 1: 2869 lbu t0, 0(a1) /* t0 = source (a8) */ 2870 lbu t1, 1(a1) /* t1 = source (a8) */ 2871 lhu t6, 0(a0) /* t6 = destination (r5g6b5) */ 2872 lhu t7, 2(a0) /* t7 = destination (r5g6b5) */ 2873 addiu a1, a1, 2 2874 2875 not t0, t0 2876 not t1, t1 2877 andi t0, 0xff /* t0 = neg source1 */ 2878 andi t1, 0xff /* t1 = neg source2 */ 2879 CONVERT_2x0565_TO_2x8888 t6, t7, t8, t9, t3, t4, s0, s1, s2, s3 2880 MIPS_2xUN8x4_MUL_2xUN8 t8, t9, t0, t1, t6, t7, t5, s0, s1, s2, s3, t8, t9 2881 CONVERT_2x8888_TO_2x0565 t6, t7, t8, t9, t2, t3, t4, s0, s1 2882 2883 sh t8, 0(a0) 2884 sh t9, 2(a0) 2885 addiu a2, a2, -2 2886 addiu t1, a2, -1 2887 bgtz t1, 1b 2888 addiu a0, a0, 4 2889 2: 2890 beqz a2, 3f 2891 nop 2892 lbu t0, 0(a1) /* t0 = source (a8) */ 2893 lhu t1, 0(a0) /* t1 = destination (r5g6b5) */ 2894 2895 not t0, t0 2896 andi t0, 0xff /* t0 = neg source */ 2897 CONVERT_1x0565_TO_1x8888 t1, t2, t3, t4 2898 MIPS_UN8x4_MUL_UN8 t2, t0, t1, t5, t3, t4, t6 2899 CONVERT_1x8888_TO_1x0565 t1, t2, t3, t4 2900 2901 sh t2, 0(a0) 2902 3: 2903 RESTORE_REGS_FROM_STACK 0, s0, s1, s2, s3 2904 4: 2905 j ra 2906 nop 2907 2908 END(pixman_composite_out_reverse_8_0565_asm_mips) 2909 2910 LEAF_MIPS_DSPR2(pixman_composite_out_reverse_8_8888_asm_mips) 2911 /* 2912 * a0 - dst (a8r8g8b8) 2913 * a1 - src (a8) 2914 * a2 - w 2915 */ 2916 2917 beqz a2, 3f 2918 nop 2919 li t4, 0x00ff00ff 2920 addiu t1, a2, -1 2921 beqz t1, 2f 2922 nop 2923 1: 2924 lbu t0, 0(a1) /* t0 = source (a8) */ 2925 lbu t1, 1(a1) /* t1 = source (a8) */ 2926 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 2927 lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */ 2928 addiu a1, a1, 2 2929 not t0, t0 2930 not t1, t1 2931 andi t0, 0xff /* t0 = neg source */ 2932 andi t1, 0xff /* t1 = neg source */ 2933 2934 MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t0, t1, t5, t6, t4, t7, t8, t9, t2, t3, t0 2935 2936 sw t5, 0(a0) 2937 sw t6, 4(a0) 2938 addiu a2, a2, -2 2939 addiu t1, a2, -1 2940 bgtz t1, 1b 2941 addiu a0, a0, 8 2942 2: 2943 beqz a2, 3f 2944 nop 2945 lbu t0, 0(a1) /* t0 = source (a8) */ 2946 lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */ 2947 not t0, t0 2948 andi t0, 0xff /* t0 = neg source */ 2949 2950 MIPS_UN8x4_MUL_UN8 t1, t0, t2, t4, t3, t5, t6 2951 2952 sw t2, 0(a0) 2953 3: 2954 j ra 2955 nop 2956 2957 END(pixman_composite_out_reverse_8_8888_asm_mips) 2958 2959 LEAF_MIPS_DSPR2(pixman_composite_over_reverse_n_8888_asm_mips) 2960 /* 2961 * a0 - dst (a8r8g8b8) 2962 * a1 - src (32bit constant) 2963 * a2 - w 2964 */ 2965 2966 beqz a2, 5f 2967 nop 2968 2969 SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7 2970 li t0, 0x00ff00ff 2971 srl t9, a2, 2 /* t9 = how many multiples of 4 src pixels */ 2972 beqz t9, 2f /* branch if less than 4 src pixels */ 2973 nop 2974 1: 2975 beqz t9, 2f 2976 addiu t9, t9, -1 2977 2978 lw t1, 0(a0) 2979 lw t2, 4(a0) 2980 lw t3, 8(a0) 2981 lw t4, 12(a0) 2982 2983 addiu a2, a2, -4 2984 2985 not t5, t1 2986 not t6, t2 2987 not t7, t3 2988 not t8, t4 2989 srl t5, t5, 24 2990 srl t6, t6, 24 2991 srl t7, t7, 24 2992 srl t8, t8, 24 2993 replv.ph t5, t5 2994 replv.ph t6, t6 2995 replv.ph t7, t7 2996 replv.ph t8, t8 2997 muleu_s.ph.qbl s0, a1, t5 2998 muleu_s.ph.qbr s1, a1, t5 2999 muleu_s.ph.qbl s2, a1, t6 3000 muleu_s.ph.qbr s3, a1, t6 3001 muleu_s.ph.qbl s4, a1, t7 3002 muleu_s.ph.qbr s5, a1, t7 3003 muleu_s.ph.qbl s6, a1, t8 3004 muleu_s.ph.qbr s7, a1, t8 3005 3006 shra_r.ph t5, s0, 8 3007 shra_r.ph t6, s1, 8 3008 shra_r.ph t7, s2, 8 3009 shra_r.ph t8, s3, 8 3010 and t5, t5, t0 3011 and t6, t6, t0 3012 and t7, t7, t0 3013 and t8, t8, t0 3014 addq.ph s0, s0, t5 3015 addq.ph s1, s1, t6 3016 addq.ph s2, s2, t7 3017 addq.ph s3, s3, t8 3018 shra_r.ph s0, s0, 8 3019 shra_r.ph s1, s1, 8 3020 shra_r.ph s2, s2, 8 3021 shra_r.ph s3, s3, 8 3022 shra_r.ph t5, s4, 8 3023 shra_r.ph t6, s5, 8 3024 shra_r.ph t7, s6, 8 3025 shra_r.ph t8, s7, 8 3026 and t5, t5, t0 3027 and t6, t6, t0 3028 and t7, t7, t0 3029 and t8, t8, t0 3030 addq.ph s4, s4, t5 3031 addq.ph s5, s5, t6 3032 addq.ph s6, s6, t7 3033 addq.ph s7, s7, t8 3034 shra_r.ph s4, s4, 8 3035 shra_r.ph s5, s5, 8 3036 shra_r.ph s6, s6, 8 3037 shra_r.ph s7, s7, 8 3038 3039 precr.qb.ph t5, s0, s1 3040 precr.qb.ph t6, s2, s3 3041 precr.qb.ph t7, s4, s5 3042 precr.qb.ph t8, s6, s7 3043 addu_s.qb t5, t1, t5 3044 addu_s.qb t6, t2, t6 3045 addu_s.qb t7, t3, t7 3046 addu_s.qb t8, t4, t8 3047 3048 sw t5, 0(a0) 3049 sw t6, 4(a0) 3050 sw t7, 8(a0) 3051 sw t8, 12(a0) 3052 b 1b 3053 addiu a0, a0, 16 3054 3055 2: 3056 beqz a2, 4f 3057 nop 3058 3: 3059 lw t1, 0(a0) 3060 3061 not t2, t1 3062 srl t2, t2, 24 3063 replv.ph t2, t2 3064 3065 muleu_s.ph.qbl t4, a1, t2 3066 muleu_s.ph.qbr t5, a1, t2 3067 shra_r.ph t6, t4, 8 3068 shra_r.ph t7, t5, 8 3069 3070 and t6,t6,t0 3071 and t7,t7,t0 3072 3073 addq.ph t8, t4, t6 3074 addq.ph t9, t5, t7 3075 3076 shra_r.ph t8, t8, 8 3077 shra_r.ph t9, t9, 8 3078 3079 precr.qb.ph t9, t8, t9 3080 3081 addu_s.qb t9, t1, t9 3082 sw t9, 0(a0) 3083 3084 addiu a2, a2, -1 3085 bnez a2, 3b 3086 addiu a0, a0, 4 3087 4: 3088 RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7 3089 5: 3090 j ra 3091 nop 3092 3093 END(pixman_composite_over_reverse_n_8888_asm_mips) 3094 3095 LEAF_MIPS_DSPR2(pixman_composite_in_n_8_asm_mips) 3096 /* 3097 * a0 - dst (a8) 3098 * a1 - src (32bit constant) 3099 * a2 - w 3100 */ 3101 3102 li t9, 0x00ff00ff 3103 beqz a2, 3f 3104 nop 3105 srl t7, a2, 2 /* t7 = how many multiples of 4 dst pixels */ 3106 beqz t7, 1f /* branch if less than 4 src pixels */ 3107 nop 3108 3109 srl t8, a1, 24 3110 replv.ph t8, t8 3111 3112 0: 3113 beqz t7, 1f 3114 addiu t7, t7, -1 3115 lbu t0, 0(a0) 3116 lbu t1, 1(a0) 3117 lbu t2, 2(a0) 3118 lbu t3, 3(a0) 3119 3120 precr_sra.ph.w t1, t0, 0 3121 precr_sra.ph.w t3, t2, 0 3122 precr.qb.ph t0, t3, t1 3123 3124 muleu_s.ph.qbl t2, t0, t8 3125 muleu_s.ph.qbr t3, t0, t8 3126 shra_r.ph t4, t2, 8 3127 shra_r.ph t5, t3, 8 3128 and t4, t4, t9 3129 and t5, t5, t9 3130 addq.ph t2, t2, t4 3131 addq.ph t3, t3, t5 3132 shra_r.ph t2, t2, 8 3133 shra_r.ph t3, t3, 8 3134 precr.qb.ph t2, t2, t3 3135 3136 sb t2, 0(a0) 3137 srl t2, t2, 8 3138 sb t2, 1(a0) 3139 srl t2, t2, 8 3140 sb t2, 2(a0) 3141 srl t2, t2, 8 3142 sb t2, 3(a0) 3143 addiu a2, a2, -4 3144 b 0b 3145 addiu a0, a0, 4 3146 3147 1: 3148 beqz a2, 3f 3149 nop 3150 srl t8, a1, 24 3151 2: 3152 lbu t0, 0(a0) 3153 3154 mul t2, t0, t8 3155 shra_r.ph t3, t2, 8 3156 andi t3, t3, 0x00ff 3157 addq.ph t2, t2, t3 3158 shra_r.ph t2, t2, 8 3159 3160 sb t2, 0(a0) 3161 addiu a2, a2, -1 3162 bnez a2, 2b 3163 addiu a0, a0, 1 3164 3165 3: 3166 j ra 3167 nop 3168 3169 END(pixman_composite_in_n_8_asm_mips) 3170 3171 LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_8888_8888_OVER_asm_mips) 3172 /* 3173 * a0 - dst (a8r8g8b8) 3174 * a1 - src (a8r8g8b8) 3175 * a2 - w 3176 * a3 - vx 3177 * 16(sp) - unit_x 3178 */ 3179 3180 SAVE_REGS_ON_STACK 0, s0, s1, s2, s3 3181 lw t8, 16(sp) /* t8 = unit_x */ 3182 li t6, 0x00ff00ff 3183 beqz a2, 3f 3184 nop 3185 addiu t1, a2, -1 3186 beqz t1, 2f 3187 nop 3188 1: 3189 sra t0, a3, 16 /* t0 = vx >> 16 */ 3190 sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */ 3191 addu t0, a1, t0 3192 lw t0, 0(t0) /* t0 = source (a8r8g8b8) */ 3193 addu a3, a3, t8 /* a3 = vx + unit_x */ 3194 3195 sra t1, a3, 16 /* t0 = vx >> 16 */ 3196 sll t1, t1, 2 /* t0 = t0 * 4 (a8r8g8b8) */ 3197 addu t1, a1, t1 3198 lw t1, 0(t1) /* t1 = source (a8r8g8b8) */ 3199 addu a3, a3, t8 /* a3 = vx + unit_x */ 3200 3201 lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */ 3202 lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */ 3203 3204 OVER_2x8888_2x8888 t0, t1, t2, t3, t4, t5, t6, t7, t9, s0, s1, s2, s3 3205 3206 sw t4, 0(a0) 3207 sw t5, 4(a0) 3208 addiu a2, a2, -2 3209 addiu t1, a2, -1 3210 bgtz t1, 1b 3211 addiu a0, a0, 8 3212 2: 3213 beqz a2, 3f 3214 nop 3215 sra t0, a3, 16 /* t0 = vx >> 16 */ 3216 sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */ 3217 addu t0, a1, t0 3218 lw t0, 0(t0) /* t0 = source (a8r8g8b8) */ 3219 lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */ 3220 addu a3, a3, t8 /* a3 = vx + unit_x */ 3221 3222 OVER_8888_8888 t0, t1, t2, t6, t4, t5, t3, t7 3223 3224 sw t2, 0(a0) 3225 3: 3226 RESTORE_REGS_FROM_STACK 0, s0, s1, s2, s3 3227 j ra 3228 nop 3229 3230 END(pixman_scaled_nearest_scanline_8888_8888_OVER_asm_mips) 3231 3232 LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_8888_0565_OVER_asm_mips) 3233 /* 3234 * a0 - dst (r5g6b5) 3235 * a1 - src (a8r8g8b8) 3236 * a2 - w 3237 * a3 - vx 3238 * 16(sp) - unit_x 3239 */ 3240 3241 SAVE_REGS_ON_STACK 24, s0, s1, s2, s3, s4, v0, v1 3242 lw t8, 40(sp) /* t8 = unit_x */ 3243 li t4, 0x00ff00ff 3244 li t5, 0xf800f800 3245 li t6, 0x07e007e0 3246 li t7, 0x001F001F 3247 beqz a2, 3f 3248 nop 3249 addiu t1, a2, -1 3250 beqz t1, 2f 3251 nop 3252 1: 3253 sra t0, a3, 16 /* t0 = vx >> 16 */ 3254 sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */ 3255 addu t0, a1, t0 3256 lw t0, 0(t0) /* t0 = source (a8r8g8b8) */ 3257 addu a3, a3, t8 /* a3 = vx + unit_x */ 3258 sra t1, a3, 16 /* t0 = vx >> 16 */ 3259 sll t1, t1, 2 /* t0 = t0 * 4 (a8r8g8b8) */ 3260 addu t1, a1, t1 3261 lw t1, 0(t1) /* t1 = source (a8r8g8b8) */ 3262 addu a3, a3, t8 /* a3 = vx + unit_x */ 3263 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 3264 lhu t3, 2(a0) /* t3 = destination (r5g6b5) */ 3265 3266 CONVERT_2x0565_TO_2x8888 t2, t3, v0, v1, t6, t7, s0, s1, s2, s3 3267 OVER_2x8888_2x8888 t0, t1, v0, v1, t2, t3, t4, t9, s0, s1, s2, s3, s4 3268 CONVERT_2x8888_TO_2x0565 t2, t3, v0, v1, t5, t6, t7, t9, s2 3269 3270 sh v0, 0(a0) 3271 sh v1, 2(a0) 3272 addiu a2, a2, -2 3273 addiu t1, a2, -1 3274 bgtz t1, 1b 3275 addiu a0, a0, 4 3276 2: 3277 beqz a2, 3f 3278 nop 3279 sra t0, a3, 16 /* t0 = vx >> 16 */ 3280 sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */ 3281 addu t0, a1, t0 3282 lw t0, 0(t0) /* t0 = source (a8r8g8b8) */ 3283 lhu t1, 0(a0) /* t1 = destination (r5g6b5) */ 3284 addu a3, a3, t8 /* a3 = vx + unit_x */ 3285 3286 CONVERT_1x0565_TO_1x8888 t1, t2, t5, t6 3287 OVER_8888_8888 t0, t2, t1, t4, t3, t5, t6, t7 3288 CONVERT_1x8888_TO_1x0565 t1, t2, t5, t6 3289 3290 sh t2, 0(a0) 3291 3: 3292 RESTORE_REGS_FROM_STACK 24, s0, s1, s2, s3, s4, v0, v1 3293 j ra 3294 nop 3295 3296 END(pixman_scaled_nearest_scanline_8888_0565_OVER_asm_mips) 3297 3298 LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_0565_8888_SRC_asm_mips) 3299 /* 3300 * a0 - dst (a8r8g8b8) 3301 * a1 - src (r5g6b5) 3302 * a2 - w 3303 * a3 - vx 3304 * 16(sp) - unit_x 3305 */ 3306 3307 SAVE_REGS_ON_STACK 0, v0 3308 beqz a2, 3f 3309 nop 3310 3311 lw v0, 16(sp) /* v0 = unit_x */ 3312 addiu t1, a2, -1 3313 beqz t1, 2f 3314 nop 3315 3316 li t4, 0x07e007e0 3317 li t5, 0x001F001F 3318 1: 3319 sra t0, a3, 16 /* t0 = vx >> 16 */ 3320 sll t0, t0, 1 /* t0 = t0 * 2 ((r5g6b5)) */ 3321 addu t0, a1, t0 3322 lhu t0, 0(t0) /* t0 = source ((r5g6b5)) */ 3323 addu a3, a3, v0 /* a3 = vx + unit_x */ 3324 sra t1, a3, 16 /* t1 = vx >> 16 */ 3325 sll t1, t1, 1 /* t1 = t1 * 2 ((r5g6b5)) */ 3326 addu t1, a1, t1 3327 lhu t1, 0(t1) /* t1 = source ((r5g6b5)) */ 3328 addu a3, a3, v0 /* a3 = vx + unit_x */ 3329 addiu a2, a2, -2 3330 3331 CONVERT_2x0565_TO_2x8888 t0, t1, t2, t3, t4, t5, t6, t7, t8, t9 3332 3333 sw t2, 0(a0) 3334 sw t3, 4(a0) 3335 3336 addiu t2, a2, -1 3337 bgtz t2, 1b 3338 addiu a0, a0, 8 3339 2: 3340 beqz a2, 3f 3341 nop 3342 sra t0, a3, 16 /* t0 = vx >> 16 */ 3343 sll t0, t0, 1 /* t0 = t0 * 2 ((r5g6b5)) */ 3344 addu t0, a1, t0 3345 lhu t0, 0(t0) /* t0 = source ((r5g6b5)) */ 3346 3347 CONVERT_1x0565_TO_1x8888 t0, t1, t2, t3 3348 3349 sw t1, 0(a0) 3350 3: 3351 RESTORE_REGS_FROM_STACK 0, v0 3352 j ra 3353 nop 3354 3355 END(pixman_scaled_nearest_scanline_0565_8888_SRC_asm_mips) 3356 3357 LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_8888_8_0565_OVER_asm_mips) 3358 /* 3359 * a0 - dst (r5g6b5) 3360 * a1 - src (a8r8g8b8) 3361 * a2 - mask (a8) 3362 * a3 - w 3363 * 16(sp) - vx 3364 * 20(sp) - unit_x 3365 */ 3366 beqz a3, 4f 3367 nop 3368 3369 SAVE_REGS_ON_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5 3370 lw v0, 36(sp) /* v0 = vx */ 3371 lw v1, 40(sp) /* v1 = unit_x */ 3372 li t6, 0x00ff00ff 3373 li t7, 0xf800f800 3374 li t8, 0x07e007e0 3375 li t9, 0x001F001F 3376 3377 addiu t1, a3, -1 3378 beqz t1, 2f 3379 nop 3380 1: 3381 sra t0, v0, 16 /* t0 = vx >> 16 */ 3382 sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */ 3383 addu t0, a1, t0 3384 lw t0, 0(t0) /* t0 = source (a8r8g8b8) */ 3385 addu v0, v0, v1 /* v0 = vx + unit_x */ 3386 sra t1, v0, 16 /* t1 = vx >> 16 */ 3387 sll t1, t1, 2 /* t1 = t1 * 4 (a8r8g8b8) */ 3388 addu t1, a1, t1 3389 lw t1, 0(t1) /* t1 = source (a8r8g8b8) */ 3390 addu v0, v0, v1 /* v0 = vx + unit_x */ 3391 lbu t2, 0(a2) /* t2 = mask (a8) */ 3392 lbu t3, 1(a2) /* t3 = mask (a8) */ 3393 lhu t4, 0(a0) /* t4 = destination (r5g6b5) */ 3394 lhu t5, 2(a0) /* t5 = destination (r5g6b5) */ 3395 addiu a2, a2, 2 3396 3397 CONVERT_2x0565_TO_2x8888 t4, t5, s0, s1, t8, t9, s2, s3, s4, s5 3398 OVER_2x8888_2x8_2x8888 t0, t1, \ 3399 t2, t3, \ 3400 s0, s1, \ 3401 t4, t5, \ 3402 t6, s2, s3, s4, s5, t2, t3 3403 CONVERT_2x8888_TO_2x0565 t4, t5, s0, s1, t7, t8, t9, s2, s3 3404 3405 sh s0, 0(a0) 3406 sh s1, 2(a0) 3407 addiu a3, a3, -2 3408 addiu t1, a3, -1 3409 bgtz t1, 1b 3410 addiu a0, a0, 4 3411 2: 3412 beqz a3, 3f 3413 nop 3414 sra t0, v0, 16 /* t0 = vx >> 16 */ 3415 sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */ 3416 addu t0, a1, t0 3417 lw t0, 0(t0) /* t0 = source (a8r8g8b8) */ 3418 lbu t1, 0(a2) /* t1 = mask (a8) */ 3419 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 3420 3421 CONVERT_1x0565_TO_1x8888 t2, t3, t4, t5 3422 OVER_8888_8_8888 t0, t1, t3, t2, t6, t4, t5, t7, t8 3423 CONVERT_1x8888_TO_1x0565 t2, t3, t4, t5 3424 3425 sh t3, 0(a0) 3426 3: 3427 RESTORE_REGS_FROM_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5 3428 4: 3429 j ra 3430 nop 3431 3432 END(pixman_scaled_nearest_scanline_8888_8_0565_OVER_asm_mips) 3433 3434 LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_0565_8_0565_OVER_asm_mips) 3435 /* 3436 * a0 - dst (r5g6b5) 3437 * a1 - src (r5g6b5) 3438 * a2 - mask (a8) 3439 * a3 - w 3440 * 16(sp) - vx 3441 * 20(sp) - unit_x 3442 */ 3443 3444 beqz a3, 4f 3445 nop 3446 SAVE_REGS_ON_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5 3447 lw v0, 36(sp) /* v0 = vx */ 3448 lw v1, 40(sp) /* v1 = unit_x */ 3449 li t4, 0xf800f800 3450 li t5, 0x07e007e0 3451 li t6, 0x001F001F 3452 li t7, 0x00ff00ff 3453 3454 addiu t1, a3, -1 3455 beqz t1, 2f 3456 nop 3457 1: 3458 sra t0, v0, 16 /* t0 = vx >> 16 */ 3459 sll t0, t0, 1 /* t0 = t0 * 2 (r5g6b5) */ 3460 addu t0, a1, t0 3461 lhu t0, 0(t0) /* t0 = source (r5g6b5) */ 3462 addu v0, v0, v1 /* v0 = vx + unit_x */ 3463 sra t1, v0, 16 /* t1 = vx >> 16 */ 3464 sll t1, t1, 1 /* t1 = t1 * 2 (r5g6b5) */ 3465 addu t1, a1, t1 3466 lhu t1, 0(t1) /* t1 = source (r5g6b5) */ 3467 addu v0, v0, v1 /* v0 = vx + unit_x */ 3468 lbu t2, 0(a2) /* t2 = mask (a8) */ 3469 lbu t3, 1(a2) /* t3 = mask (a8) */ 3470 lhu t8, 0(a0) /* t8 = destination (r5g6b5) */ 3471 lhu t9, 2(a0) /* t9 = destination (r5g6b5) */ 3472 addiu a2, a2, 2 3473 3474 CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5 3475 CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, t0, t1 3476 OVER_2x8888_2x8_2x8888 s0, s1, \ 3477 t2, t3, \ 3478 s2, s3, \ 3479 t0, t1, \ 3480 t7, t8, t9, s4, s5, s0, s1 3481 CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3 3482 3483 sh s0, 0(a0) 3484 sh s1, 2(a0) 3485 addiu a3, a3, -2 3486 addiu t1, a3, -1 3487 bgtz t1, 1b 3488 addiu a0, a0, 4 3489 2: 3490 beqz a3, 3f 3491 nop 3492 sra t0, v0, 16 /* t0 = vx >> 16 */ 3493 sll t0, t0, 1 /* t0 = t0 * 2 (r5g6b5) */ 3494 addu t0, a1, t0 3495 3496 lhu t0, 0(t0) /* t0 = source (r5g6b5) */ 3497 lbu t1, 0(a2) /* t1 = mask (a8) */ 3498 lhu t2, 0(a0) /* t2 = destination (r5g6b5) */ 3499 3500 CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5 3501 CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6 3502 OVER_8888_8_8888 t3, t1, t4, t0, t7, t2, t5, t6, t8 3503 CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5 3504 3505 sh t3, 0(a0) 3506 3: 3507 RESTORE_REGS_FROM_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5 3508 4: 3509 j ra 3510 nop 3511 3512 END(pixman_scaled_nearest_scanline_0565_8_0565_OVER_asm_mips) 3513 3514 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_SRC_asm_mips) 3515 /* 3516 * a0 - *dst 3517 * a1 - *src_top 3518 * a2 - *src_bottom 3519 * a3 - w 3520 * 16(sp) - wt 3521 * 20(sp) - wb 3522 * 24(sp) - vx 3523 * 28(sp) - unit_x 3524 */ 3525 3526 beqz a3, 1f 3527 nop 3528 3529 SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7 3530 3531 lw s0, 36(sp) /* s0 = wt */ 3532 lw s1, 40(sp) /* s1 = wb */ 3533 lw s2, 44(sp) /* s2 = vx */ 3534 lw s3, 48(sp) /* s3 = unit_x */ 3535 li v0, BILINEAR_INTERPOLATION_RANGE 3536 3537 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3538 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3539 0: 3540 andi t4, s2, 0xffff /* t4 = (short)vx */ 3541 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 3542 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 3543 3544 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 3545 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 3546 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 3547 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 3548 3549 sra t9, s2, 16 3550 sll t9, t9, 2 3551 addiu t8, t9, 4 3552 lwx t0, t9(a1) /* t0 = tl */ 3553 lwx t1, t8(a1) /* t1 = tr */ 3554 addiu a3, a3, -1 3555 lwx t2, t9(a2) /* t2 = bl */ 3556 lwx t3, t8(a2) /* t3 = br */ 3557 3558 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 3559 3560 addu s2, s2, s3 /* vx += unit_x; */ 3561 sw t0, 0(a0) 3562 bnez a3, 0b 3563 addiu a0, a0, 4 3564 3565 RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7 3566 1: 3567 j ra 3568 nop 3569 3570 END(pixman_scaled_bilinear_scanline_8888_8888_SRC_asm_mips) 3571 3572 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_0565_SRC_asm_mips) 3573 /* 3574 * a0 - *dst 3575 * a1 - *src_top 3576 * a2 - *src_bottom 3577 * a3 - w 3578 * 16(sp) - wt 3579 * 20(sp) - wb 3580 * 24(sp) - vx 3581 * 28(sp) - unit_x 3582 */ 3583 3584 beqz a3, 1f 3585 nop 3586 3587 SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7 3588 3589 lw s0, 36(sp) /* s0 = wt */ 3590 lw s1, 40(sp) /* s1 = wb */ 3591 lw s2, 44(sp) /* s2 = vx */ 3592 lw s3, 48(sp) /* s3 = unit_x */ 3593 li v0, BILINEAR_INTERPOLATION_RANGE 3594 3595 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3596 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3597 0: 3598 andi t4, s2, 0xffff /* t4 = (short)vx */ 3599 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 3600 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 3601 3602 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 3603 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 3604 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 3605 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 3606 3607 sra t9, s2, 16 3608 sll t9, t9, 2 3609 addiu t8, t9, 4 3610 lwx t0, t9(a1) /* t0 = tl */ 3611 lwx t1, t8(a1) /* t1 = tr */ 3612 addiu a3, a3, -1 3613 lwx t2, t9(a2) /* t2 = bl */ 3614 lwx t3, t8(a2) /* t3 = br */ 3615 3616 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 3617 CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3 3618 3619 addu s2, s2, s3 /* vx += unit_x; */ 3620 sh t1, 0(a0) 3621 bnez a3, 0b 3622 addiu a0, a0, 2 3623 3624 RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7 3625 1: 3626 j ra 3627 nop 3628 3629 END(pixman_scaled_bilinear_scanline_8888_0565_SRC_asm_mips) 3630 3631 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8888_SRC_asm_mips) 3632 /* 3633 * a0 - *dst 3634 * a1 - *src_top 3635 * a2 - *src_bottom 3636 * a3 - w 3637 * 16(sp) - wt 3638 * 20(sp) - wb 3639 * 24(sp) - vx 3640 * 28(sp) - unit_x 3641 */ 3642 3643 beqz a3, 1f 3644 nop 3645 3646 SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 3647 3648 lw s0, 44(sp) /* s0 = wt */ 3649 lw s1, 48(sp) /* s1 = wb */ 3650 lw s2, 52(sp) /* s2 = vx */ 3651 lw s3, 56(sp) /* s3 = unit_x */ 3652 li v0, BILINEAR_INTERPOLATION_RANGE 3653 li v1, 0x07e007e0 3654 li s8, 0x001f001f 3655 3656 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3657 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3658 0: 3659 andi t4, s2, 0xffff /* t4 = (short)vx */ 3660 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 3661 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 3662 3663 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 3664 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 3665 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 3666 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 3667 3668 sra t9, s2, 16 3669 sll t9, t9, 1 3670 addiu t8, t9, 2 3671 lhx t0, t9(a1) /* t0 = tl */ 3672 lhx t1, t8(a1) /* t1 = tr */ 3673 andi t1, t1, 0xffff 3674 addiu a3, a3, -1 3675 lhx t2, t9(a2) /* t2 = bl */ 3676 lhx t3, t8(a2) /* t3 = br */ 3677 andi t3, t3, 0xffff 3678 3679 CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7 3680 CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7 3681 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 3682 3683 addu s2, s2, s3 /* vx += unit_x; */ 3684 sw t0, 0(a0) 3685 bnez a3, 0b 3686 addiu a0, a0, 4 3687 3688 RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 3689 1: 3690 j ra 3691 nop 3692 3693 END(pixman_scaled_bilinear_scanline_0565_8888_SRC_asm_mips) 3694 3695 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_0565_SRC_asm_mips) 3696 /* 3697 * a0 - *dst 3698 * a1 - *src_top 3699 * a2 - *src_bottom 3700 * a3 - w 3701 * 16(sp) - wt 3702 * 20(sp) - wb 3703 * 24(sp) - vx 3704 * 28(sp) - unit_x 3705 */ 3706 3707 beqz a3, 1f 3708 nop 3709 3710 SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 3711 3712 lw s0, 44(sp) /* s0 = wt */ 3713 lw s1, 48(sp) /* s1 = wb */ 3714 lw s2, 52(sp) /* s2 = vx */ 3715 lw s3, 56(sp) /* s3 = unit_x */ 3716 li v0, BILINEAR_INTERPOLATION_RANGE 3717 li v1, 0x07e007e0 3718 li s8, 0x001f001f 3719 3720 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3721 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3722 0: 3723 andi t4, s2, 0xffff /* t4 = (short)vx */ 3724 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 3725 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 3726 3727 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 3728 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 3729 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 3730 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 3731 3732 sra t9, s2, 16 3733 sll t9, t9, 1 3734 addiu t8, t9, 2 3735 lhx t0, t9(a1) /* t0 = tl */ 3736 lhx t1, t8(a1) /* t1 = tr */ 3737 andi t1, t1, 0xffff 3738 addiu a3, a3, -1 3739 lhx t2, t9(a2) /* t2 = bl */ 3740 lhx t3, t8(a2) /* t3 = br */ 3741 andi t3, t3, 0xffff 3742 3743 CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7 3744 CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7 3745 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 3746 CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3 3747 3748 addu s2, s2, s3 /* vx += unit_x; */ 3749 sh t1, 0(a0) 3750 bnez a3, 0b 3751 addiu a0, a0, 2 3752 3753 RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 3754 1: 3755 j ra 3756 nop 3757 3758 END(pixman_scaled_bilinear_scanline_0565_0565_SRC_asm_mips) 3759 3760 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_OVER_asm_mips) 3761 /* 3762 * a0 - *dst 3763 * a1 - *src_top 3764 * a2 - *src_bottom 3765 * a3 - w 3766 * 16(sp) - wt 3767 * 20(sp) - wb 3768 * 24(sp) - vx 3769 * 28(sp) - unit_x 3770 */ 3771 3772 beqz a3, 1f 3773 nop 3774 3775 SAVE_REGS_ON_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8 3776 3777 lw s0, 40(sp) /* s0 = wt */ 3778 lw s1, 44(sp) /* s1 = wb */ 3779 lw s2, 48(sp) /* s2 = vx */ 3780 lw s3, 52(sp) /* s3 = unit_x */ 3781 li v0, BILINEAR_INTERPOLATION_RANGE 3782 li s8, 0x00ff00ff 3783 3784 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3785 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3786 0: 3787 andi t4, s2, 0xffff /* t4 = (short)vx */ 3788 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 3789 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 3790 3791 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 3792 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 3793 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 3794 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 3795 3796 sra t9, s2, 16 3797 sll t9, t9, 2 3798 addiu t8, t9, 4 3799 lwx t0, t9(a1) /* t0 = tl */ 3800 lwx t1, t8(a1) /* t1 = tr */ 3801 addiu a3, a3, -1 3802 lwx t2, t9(a2) /* t2 = bl */ 3803 lwx t3, t8(a2) /* t3 = br */ 3804 3805 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 3806 lw t1, 0(a0) /* t1 = dest */ 3807 OVER_8888_8888 t0, t1, t2, s8, t3, t4, t5, t6 3808 3809 addu s2, s2, s3 /* vx += unit_x; */ 3810 sw t2, 0(a0) 3811 bnez a3, 0b 3812 addiu a0, a0, 4 3813 3814 RESTORE_REGS_FROM_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8 3815 1: 3816 j ra 3817 nop 3818 3819 END(pixman_scaled_bilinear_scanline_8888_8888_OVER_asm_mips) 3820 3821 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_ADD_asm_mips) 3822 /* 3823 * a0 - *dst 3824 * a1 - *src_top 3825 * a2 - *src_bottom 3826 * a3 - w 3827 * 16(sp) - wt 3828 * 20(sp) - wb 3829 * 24(sp) - vx 3830 * 28(sp) - unit_x 3831 */ 3832 3833 beqz a3, 1f 3834 nop 3835 3836 SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7 3837 3838 lw s0, 36(sp) /* s0 = wt */ 3839 lw s1, 40(sp) /* s1 = wb */ 3840 lw s2, 44(sp) /* s2 = vx */ 3841 lw s3, 48(sp) /* s3 = unit_x */ 3842 li v0, BILINEAR_INTERPOLATION_RANGE 3843 3844 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3845 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3846 0: 3847 andi t4, s2, 0xffff /* t4 = (short)vx */ 3848 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 3849 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 3850 3851 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 3852 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 3853 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 3854 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 3855 3856 sra t9, s2, 16 3857 sll t9, t9, 2 3858 addiu t8, t9, 4 3859 lwx t0, t9(a1) /* t0 = tl */ 3860 lwx t1, t8(a1) /* t1 = tr */ 3861 addiu a3, a3, -1 3862 lwx t2, t9(a2) /* t2 = bl */ 3863 lwx t3, t8(a2) /* t3 = br */ 3864 3865 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 3866 lw t1, 0(a0) 3867 addu_s.qb t2, t0, t1 3868 3869 addu s2, s2, s3 /* vx += unit_x; */ 3870 sw t2, 0(a0) 3871 bnez a3, 0b 3872 addiu a0, a0, 4 3873 3874 RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7 3875 1: 3876 j ra 3877 nop 3878 3879 END(pixman_scaled_bilinear_scanline_8888_8888_ADD_asm_mips) 3880 3881 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_SRC_asm_mips) 3882 /* 3883 * a0 - *dst 3884 * a1 - *mask 3885 * a2 - *src_top 3886 * a3 - *src_bottom 3887 * 16(sp) - wt 3888 * 20(sp) - wb 3889 * 24(sp) - vx 3890 * 28(sp) - unit_x 3891 * 32(sp) - w 3892 */ 3893 3894 lw v1, 32(sp) 3895 beqz v1, 1f 3896 nop 3897 3898 SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 3899 3900 lw s0, 44(sp) /* s0 = wt */ 3901 lw s1, 48(sp) /* s1 = wb */ 3902 lw s2, 52(sp) /* s2 = vx */ 3903 lw s3, 56(sp) /* s3 = unit_x */ 3904 li v0, BILINEAR_INTERPOLATION_RANGE 3905 li s8, 0x00ff00ff 3906 3907 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3908 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3909 0: 3910 andi t4, s2, 0xffff /* t4 = (short)vx */ 3911 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 3912 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 3913 3914 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 3915 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 3916 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 3917 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 3918 3919 sra t9, s2, 16 3920 sll t9, t9, 2 3921 addiu t8, t9, 4 3922 lwx t0, t9(a2) /* t0 = tl */ 3923 lwx t1, t8(a2) /* t1 = tr */ 3924 addiu v1, v1, -1 3925 lwx t2, t9(a3) /* t2 = bl */ 3926 lwx t3, t8(a3) /* t3 = br */ 3927 3928 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 3929 lbu t1, 0(a1) /* t1 = mask */ 3930 addiu a1, a1, 1 3931 MIPS_UN8x4_MUL_UN8 t0, t1, t0, s8, t2, t3, t4 3932 3933 addu s2, s2, s3 /* vx += unit_x; */ 3934 sw t0, 0(a0) 3935 bnez v1, 0b 3936 addiu a0, a0, 4 3937 3938 RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 3939 1: 3940 j ra 3941 nop 3942 3943 END(pixman_scaled_bilinear_scanline_8888_8_8888_SRC_asm_mips) 3944 3945 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_0565_SRC_asm_mips) 3946 /* 3947 * a0 - *dst 3948 * a1 - *mask 3949 * a2 - *src_top 3950 * a3 - *src_bottom 3951 * 16(sp) - wt 3952 * 20(sp) - wb 3953 * 24(sp) - vx 3954 * 28(sp) - unit_x 3955 * 32(sp) - w 3956 */ 3957 3958 lw v1, 32(sp) 3959 beqz v1, 1f 3960 nop 3961 3962 SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 3963 3964 lw s0, 44(sp) /* s0 = wt */ 3965 lw s1, 48(sp) /* s1 = wb */ 3966 lw s2, 52(sp) /* s2 = vx */ 3967 lw s3, 56(sp) /* s3 = unit_x */ 3968 li v0, BILINEAR_INTERPOLATION_RANGE 3969 li s8, 0x00ff00ff 3970 3971 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3972 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 3973 0: 3974 andi t4, s2, 0xffff /* t4 = (short)vx */ 3975 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 3976 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 3977 3978 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 3979 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 3980 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 3981 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 3982 3983 sra t9, s2, 16 3984 sll t9, t9, 2 3985 addiu t8, t9, 4 3986 lwx t0, t9(a2) /* t0 = tl */ 3987 lwx t1, t8(a2) /* t1 = tr */ 3988 addiu v1, v1, -1 3989 lwx t2, t9(a3) /* t2 = bl */ 3990 lwx t3, t8(a3) /* t3 = br */ 3991 3992 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 3993 lbu t1, 0(a1) /* t1 = mask */ 3994 addiu a1, a1, 1 3995 MIPS_UN8x4_MUL_UN8 t0, t1, t0, s8, t2, t3, t4 3996 CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3 3997 3998 addu s2, s2, s3 /* vx += unit_x; */ 3999 sh t1, 0(a0) 4000 bnez v1, 0b 4001 addiu a0, a0, 2 4002 4003 RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 4004 1: 4005 j ra 4006 nop 4007 4008 END(pixman_scaled_bilinear_scanline_8888_8_0565_SRC_asm_mips) 4009 4010 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8_x888_SRC_asm_mips) 4011 /* 4012 * a0 - *dst 4013 * a1 - *mask 4014 * a2 - *src_top 4015 * a3 - *src_bottom 4016 * 16(sp) - wt 4017 * 20(sp) - wb 4018 * 24(sp) - vx 4019 * 28(sp) - unit_x 4020 * 32(sp) - w 4021 */ 4022 4023 lw t0, 32(sp) 4024 beqz t0, 1f 4025 nop 4026 4027 SAVE_REGS_ON_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra 4028 4029 lw s0, 48(sp) /* s0 = wt */ 4030 lw s1, 52(sp) /* s1 = wb */ 4031 lw s2, 56(sp) /* s2 = vx */ 4032 lw s3, 60(sp) /* s3 = unit_x */ 4033 lw ra, 64(sp) /* ra = w */ 4034 li v0, 0x00ff00ff 4035 li v1, 0x07e007e0 4036 li s8, 0x001f001f 4037 4038 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 4039 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 4040 0: 4041 andi t4, s2, 0xffff /* t4 = (short)vx */ 4042 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 4043 li t5, BILINEAR_INTERPOLATION_RANGE 4044 subu t5, t5, t4 /* t5 = ( 256 - (vx>>8)) */ 4045 4046 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 4047 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 4048 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 4049 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 4050 4051 sra t9, s2, 16 4052 sll t9, t9, 1 4053 addiu t8, t9, 2 4054 lhx t0, t9(a2) /* t0 = tl */ 4055 lhx t1, t8(a2) /* t1 = tr */ 4056 andi t1, t1, 0xffff 4057 addiu ra, ra, -1 4058 lhx t2, t9(a3) /* t2 = bl */ 4059 lhx t3, t8(a3) /* t3 = br */ 4060 andi t3, t3, 0xffff 4061 4062 CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7 4063 CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7 4064 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 4065 lbu t1, 0(a1) /* t1 = mask */ 4066 addiu a1, a1, 1 4067 MIPS_UN8x4_MUL_UN8 t0, t1, t0, v0, t2, t3, t4 4068 4069 addu s2, s2, s3 /* vx += unit_x; */ 4070 sw t0, 0(a0) 4071 bnez ra, 0b 4072 addiu a0, a0, 4 4073 4074 RESTORE_REGS_FROM_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra 4075 1: 4076 j ra 4077 nop 4078 4079 END(pixman_scaled_bilinear_scanline_0565_8_x888_SRC_asm_mips) 4080 4081 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8_0565_SRC_asm_mips) 4082 /* 4083 * a0 - *dst 4084 * a1 - *mask 4085 * a2 - *src_top 4086 * a3 - *src_bottom 4087 * 16(sp) - wt 4088 * 20(sp) - wb 4089 * 24(sp) - vx 4090 * 28(sp) - unit_x 4091 * 32(sp) - w 4092 */ 4093 4094 lw t0, 32(sp) 4095 beqz t0, 1f 4096 nop 4097 4098 SAVE_REGS_ON_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra 4099 4100 lw s0, 48(sp) /* s0 = wt */ 4101 lw s1, 52(sp) /* s1 = wb */ 4102 lw s2, 56(sp) /* s2 = vx */ 4103 lw s3, 60(sp) /* s3 = unit_x */ 4104 lw ra, 64(sp) /* ra = w */ 4105 li v0, 0x00ff00ff 4106 li v1, 0x07e007e0 4107 li s8, 0x001f001f 4108 4109 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 4110 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 4111 0: 4112 andi t4, s2, 0xffff /* t4 = (short)vx */ 4113 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 4114 li t5, BILINEAR_INTERPOLATION_RANGE 4115 subu t5, t5, t4 /* t5 = ( 256 - (vx>>8)) */ 4116 4117 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 4118 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 4119 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 4120 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 4121 4122 sra t9, s2, 16 4123 sll t9, t9, 1 4124 addiu t8, t9, 2 4125 lhx t0, t9(a2) /* t0 = tl */ 4126 lhx t1, t8(a2) /* t1 = tr */ 4127 andi t1, t1, 0xffff 4128 addiu ra, ra, -1 4129 lhx t2, t9(a3) /* t2 = bl */ 4130 lhx t3, t8(a3) /* t3 = br */ 4131 andi t3, t3, 0xffff 4132 4133 CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7 4134 CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7 4135 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 4136 lbu t1, 0(a1) /* t1 = mask */ 4137 addiu a1, a1, 1 4138 MIPS_UN8x4_MUL_UN8 t0, t1, t0, v0, t2, t3, t4 4139 CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3 4140 4141 addu s2, s2, s3 /* vx += unit_x; */ 4142 sh t1, 0(a0) 4143 bnez ra, 0b 4144 addiu a0, a0, 2 4145 4146 RESTORE_REGS_FROM_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra 4147 1: 4148 j ra 4149 nop 4150 4151 END(pixman_scaled_bilinear_scanline_0565_8_0565_SRC_asm_mips) 4152 4153 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_OVER_asm_mips) 4154 /* 4155 * a0 - dst (a8r8g8b8) 4156 * a1 - mask (a8) 4157 * a2 - src_top (a8r8g8b8) 4158 * a3 - src_bottom (a8r8g8b8) 4159 * 16(sp) - wt 4160 * 20(sp) - wb 4161 * 24(sp) - vx 4162 * 28(sp) - unit_x 4163 * 32(sp) - w 4164 */ 4165 4166 SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 4167 4168 lw v1, 60(sp) /* v1 = w(sp + 32 + 28 save regs stack offset)*/ 4169 beqz v1, 1f 4170 nop 4171 4172 lw s0, 44(sp) /* s0 = wt */ 4173 lw s1, 48(sp) /* s1 = wb */ 4174 lw s2, 52(sp) /* s2 = vx */ 4175 lw s3, 56(sp) /* s3 = unit_x */ 4176 li v0, BILINEAR_INTERPOLATION_RANGE 4177 li s8, 0x00ff00ff 4178 4179 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 4180 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 4181 4182 0: 4183 andi t4, s2, 0xffff /* t4 = (short)vx */ 4184 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 4185 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 4186 4187 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 4188 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 4189 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 4190 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 4191 4192 sra t9, s2, 16 4193 sll t9, t9, 2 4194 addiu t8, t9, 4 4195 lwx t0, t9(a2) /* t0 = tl */ 4196 lwx t1, t8(a2) /* t1 = tr */ 4197 addiu v1, v1, -1 4198 lwx t2, t9(a3) /* t2 = bl */ 4199 lwx t3, t8(a3) /* t3 = br */ 4200 4201 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, \ 4202 t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 4203 lbu t1, 0(a1) /* t1 = mask */ 4204 lw t2, 0(a0) /* t2 = dst */ 4205 addiu a1, a1, 1 4206 OVER_8888_8_8888 t0, t1, t2, t0, s8, t3, t4, t5, t6 4207 4208 addu s2, s2, s3 /* vx += unit_x; */ 4209 sw t0, 0(a0) 4210 bnez v1, 0b 4211 addiu a0, a0, 4 4212 4213 1: 4214 RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 4215 j ra 4216 nop 4217 4218 END(pixman_scaled_bilinear_scanline_8888_8_8888_OVER_asm_mips) 4219 4220 LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_ADD_asm_mips) 4221 /* 4222 * a0 - *dst 4223 * a1 - *mask 4224 * a2 - *src_top 4225 * a3 - *src_bottom 4226 * 16(sp) - wt 4227 * 20(sp) - wb 4228 * 24(sp) - vx 4229 * 28(sp) - unit_x 4230 * 32(sp) - w 4231 */ 4232 4233 lw v1, 32(sp) 4234 beqz v1, 1f 4235 nop 4236 4237 SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 4238 4239 lw s0, 44(sp) /* s0 = wt */ 4240 lw s1, 48(sp) /* s1 = wb */ 4241 lw s2, 52(sp) /* s2 = vx */ 4242 lw s3, 56(sp) /* s3 = unit_x */ 4243 li v0, BILINEAR_INTERPOLATION_RANGE 4244 li s8, 0x00ff00ff 4245 4246 sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 4247 sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS)) 4248 0: 4249 andi t4, s2, 0xffff /* t4 = (short)vx */ 4250 srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */ 4251 subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */ 4252 4253 mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */ 4254 mul s5, s0, t4 /* s5 = wt*(vx>>8) */ 4255 mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */ 4256 mul s7, s1, t4 /* s7 = wb*(vx>>8) */ 4257 4258 sra t9, s2, 16 4259 sll t9, t9, 2 4260 addiu t8, t9, 4 4261 lwx t0, t9(a2) /* t0 = tl */ 4262 lwx t1, t8(a2) /* t1 = tr */ 4263 addiu v1, v1, -1 4264 lwx t2, t9(a3) /* t2 = bl */ 4265 lwx t3, t8(a3) /* t3 = br */ 4266 4267 BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7 4268 lbu t1, 0(a1) /* t1 = mask */ 4269 lw t2, 0(a0) /* t2 = dst */ 4270 addiu a1, a1, 1 4271 MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t0, s8, t3, t4, t5 4272 4273 addu s2, s2, s3 /* vx += unit_x; */ 4274 sw t0, 0(a0) 4275 bnez v1, 0b 4276 addiu a0, a0, 4 4277 4278 RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8 4279 1: 4280 j ra 4281 nop 4282 4283 END(pixman_scaled_bilinear_scanline_8888_8_8888_ADD_asm_mips)