1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/cairo/libpixman/src/pixman-mips-dspr2-asm.S Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,3373 @@
1.4 +/*
1.5 + * Copyright (c) 2012
1.6 + * MIPS Technologies, Inc., California.
1.7 + *
1.8 + * Redistribution and use in source and binary forms, with or without
1.9 + * modification, are permitted provided that the following conditions
1.10 + * are met:
1.11 + * 1. Redistributions of source code must retain the above copyright
1.12 + * notice, this list of conditions and the following disclaimer.
1.13 + * 2. Redistributions in binary form must reproduce the above copyright
1.14 + * notice, this list of conditions and the following disclaimer in the
1.15 + * documentation and/or other materials provided with the distribution.
1.16 + * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
1.17 + * contributors may be used to endorse or promote products derived from
1.18 + * this software without specific prior written permission.
1.19 + *
1.20 + * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
1.21 + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
1.22 + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
1.23 + * ARE DISCLAIMED. IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
1.24 + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
1.25 + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
1.26 + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
1.27 + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
1.28 + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
1.29 + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
1.30 + * SUCH DAMAGE.
1.31 + *
1.32 + * Author: Nemanja Lukic (nlukic@mips.com)
1.33 + */
1.34 +
1.35 +#include "pixman-private.h"
1.36 +#include "pixman-mips-dspr2-asm.h"
1.37 +
1.38 +LEAF_MIPS_DSPR2(pixman_fill_buff16_mips)
1.39 +/*
1.40 + * a0 - *dest
1.41 + * a1 - count (bytes)
1.42 + * a2 - value to fill buffer with
1.43 + */
1.44 +
1.45 + beqz a1, 3f
1.46 + andi t1, a0, 0x0002
1.47 + beqz t1, 0f /* check if address is 4-byte aligned */
1.48 + nop
1.49 + sh a2, 0(a0)
1.50 + addiu a0, a0, 2
1.51 + addiu a1, a1, -2
1.52 +0:
1.53 + srl t1, a1, 5 /* t1 how many multiples of 32 bytes */
1.54 + replv.ph a2, a2 /* replicate fill value (16bit) in a2 */
1.55 + beqz t1, 2f
1.56 + nop
1.57 +1:
1.58 + addiu t1, t1, -1
1.59 + beqz t1, 11f
1.60 + addiu a1, a1, -32
1.61 + pref 30, 32(a0)
1.62 + sw a2, 0(a0)
1.63 + sw a2, 4(a0)
1.64 + sw a2, 8(a0)
1.65 + sw a2, 12(a0)
1.66 + sw a2, 16(a0)
1.67 + sw a2, 20(a0)
1.68 + sw a2, 24(a0)
1.69 + sw a2, 28(a0)
1.70 + b 1b
1.71 + addiu a0, a0, 32
1.72 +11:
1.73 + sw a2, 0(a0)
1.74 + sw a2, 4(a0)
1.75 + sw a2, 8(a0)
1.76 + sw a2, 12(a0)
1.77 + sw a2, 16(a0)
1.78 + sw a2, 20(a0)
1.79 + sw a2, 24(a0)
1.80 + sw a2, 28(a0)
1.81 + addiu a0, a0, 32
1.82 +2:
1.83 + blez a1, 3f
1.84 + addiu a1, a1, -2
1.85 + sh a2, 0(a0)
1.86 + b 2b
1.87 + addiu a0, a0, 2
1.88 +3:
1.89 + jr ra
1.90 + nop
1.91 +
1.92 +END(pixman_fill_buff16_mips)
1.93 +
1.94 +LEAF_MIPS32R2(pixman_fill_buff32_mips)
1.95 +/*
1.96 + * a0 - *dest
1.97 + * a1 - count (bytes)
1.98 + * a2 - value to fill buffer with
1.99 + */
1.100 +
1.101 + beqz a1, 3f
1.102 + nop
1.103 + srl t1, a1, 5 /* t1 how many multiples of 32 bytes */
1.104 + beqz t1, 2f
1.105 + nop
1.106 +1:
1.107 + addiu t1, t1, -1
1.108 + beqz t1, 11f
1.109 + addiu a1, a1, -32
1.110 + pref 30, 32(a0)
1.111 + sw a2, 0(a0)
1.112 + sw a2, 4(a0)
1.113 + sw a2, 8(a0)
1.114 + sw a2, 12(a0)
1.115 + sw a2, 16(a0)
1.116 + sw a2, 20(a0)
1.117 + sw a2, 24(a0)
1.118 + sw a2, 28(a0)
1.119 + b 1b
1.120 + addiu a0, a0, 32
1.121 +11:
1.122 + sw a2, 0(a0)
1.123 + sw a2, 4(a0)
1.124 + sw a2, 8(a0)
1.125 + sw a2, 12(a0)
1.126 + sw a2, 16(a0)
1.127 + sw a2, 20(a0)
1.128 + sw a2, 24(a0)
1.129 + sw a2, 28(a0)
1.130 + addiu a0, a0, 32
1.131 +2:
1.132 + blez a1, 3f
1.133 + addiu a1, a1, -4
1.134 + sw a2, 0(a0)
1.135 + b 2b
1.136 + addiu a0, a0, 4
1.137 +3:
1.138 + jr ra
1.139 + nop
1.140 +
1.141 +END(pixman_fill_buff32_mips)
1.142 +
1.143 +LEAF_MIPS_DSPR2(pixman_composite_src_8888_0565_asm_mips)
1.144 +/*
1.145 + * a0 - dst (r5g6b5)
1.146 + * a1 - src (a8r8g8b8)
1.147 + * a2 - w
1.148 + */
1.149 +
1.150 + beqz a2, 3f
1.151 + nop
1.152 + addiu t1, a2, -1
1.153 + beqz t1, 2f
1.154 + nop
1.155 + li t4, 0xf800f800
1.156 + li t5, 0x07e007e0
1.157 + li t6, 0x001f001f
1.158 +1:
1.159 + lw t0, 0(a1)
1.160 + lw t1, 4(a1)
1.161 + addiu a1, a1, 8
1.162 + addiu a2, a2, -2
1.163 +
1.164 + CONVERT_2x8888_TO_2x0565 t0, t1, t2, t3, t4, t5, t6, t7, t8
1.165 +
1.166 + sh t2, 0(a0)
1.167 + sh t3, 2(a0)
1.168 +
1.169 + addiu t2, a2, -1
1.170 + bgtz t2, 1b
1.171 + addiu a0, a0, 4
1.172 +2:
1.173 + beqz a2, 3f
1.174 + nop
1.175 + lw t0, 0(a1)
1.176 +
1.177 + CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
1.178 +
1.179 + sh t1, 0(a0)
1.180 +3:
1.181 + j ra
1.182 + nop
1.183 +
1.184 +END(pixman_composite_src_8888_0565_asm_mips)
1.185 +
1.186 +LEAF_MIPS_DSPR2(pixman_composite_src_0565_8888_asm_mips)
1.187 +/*
1.188 + * a0 - dst (a8r8g8b8)
1.189 + * a1 - src (r5g6b5)
1.190 + * a2 - w
1.191 + */
1.192 +
1.193 + beqz a2, 3f
1.194 + nop
1.195 + addiu t1, a2, -1
1.196 + beqz t1, 2f
1.197 + nop
1.198 + li t4, 0x07e007e0
1.199 + li t5, 0x001F001F
1.200 +1:
1.201 + lhu t0, 0(a1)
1.202 + lhu t1, 2(a1)
1.203 + addiu a1, a1, 4
1.204 + addiu a2, a2, -2
1.205 +
1.206 + CONVERT_2x0565_TO_2x8888 t0, t1, t2, t3, t4, t5, t6, t7, t8, t9
1.207 +
1.208 + sw t2, 0(a0)
1.209 + sw t3, 4(a0)
1.210 +
1.211 + addiu t2, a2, -1
1.212 + bgtz t2, 1b
1.213 + addiu a0, a0, 8
1.214 +2:
1.215 + beqz a2, 3f
1.216 + nop
1.217 + lhu t0, 0(a1)
1.218 +
1.219 + CONVERT_1x0565_TO_1x8888 t0, t1, t2, t3
1.220 +
1.221 + sw t1, 0(a0)
1.222 +3:
1.223 + j ra
1.224 + nop
1.225 +
1.226 +END(pixman_composite_src_0565_8888_asm_mips)
1.227 +
1.228 +LEAF_MIPS_DSPR2(pixman_composite_src_x888_8888_asm_mips)
1.229 +/*
1.230 + * a0 - dst (a8r8g8b8)
1.231 + * a1 - src (x8r8g8b8)
1.232 + * a2 - w
1.233 + */
1.234 +
1.235 + beqz a2, 4f
1.236 + nop
1.237 + li t9, 0xff000000
1.238 + srl t8, a2, 3 /* t1 = how many multiples of 8 src pixels */
1.239 + beqz t8, 3f /* branch if less than 8 src pixels */
1.240 + nop
1.241 +1:
1.242 + addiu t8, t8, -1
1.243 + beqz t8, 2f
1.244 + addiu a2, a2, -8
1.245 + pref 0, 32(a1)
1.246 + lw t0, 0(a1)
1.247 + lw t1, 4(a1)
1.248 + lw t2, 8(a1)
1.249 + lw t3, 12(a1)
1.250 + lw t4, 16(a1)
1.251 + lw t5, 20(a1)
1.252 + lw t6, 24(a1)
1.253 + lw t7, 28(a1)
1.254 + addiu a1, a1, 32
1.255 + or t0, t0, t9
1.256 + or t1, t1, t9
1.257 + or t2, t2, t9
1.258 + or t3, t3, t9
1.259 + or t4, t4, t9
1.260 + or t5, t5, t9
1.261 + or t6, t6, t9
1.262 + or t7, t7, t9
1.263 + pref 30, 32(a0)
1.264 + sw t0, 0(a0)
1.265 + sw t1, 4(a0)
1.266 + sw t2, 8(a0)
1.267 + sw t3, 12(a0)
1.268 + sw t4, 16(a0)
1.269 + sw t5, 20(a0)
1.270 + sw t6, 24(a0)
1.271 + sw t7, 28(a0)
1.272 + b 1b
1.273 + addiu a0, a0, 32
1.274 +2:
1.275 + lw t0, 0(a1)
1.276 + lw t1, 4(a1)
1.277 + lw t2, 8(a1)
1.278 + lw t3, 12(a1)
1.279 + lw t4, 16(a1)
1.280 + lw t5, 20(a1)
1.281 + lw t6, 24(a1)
1.282 + lw t7, 28(a1)
1.283 + addiu a1, a1, 32
1.284 + or t0, t0, t9
1.285 + or t1, t1, t9
1.286 + or t2, t2, t9
1.287 + or t3, t3, t9
1.288 + or t4, t4, t9
1.289 + or t5, t5, t9
1.290 + or t6, t6, t9
1.291 + or t7, t7, t9
1.292 + sw t0, 0(a0)
1.293 + sw t1, 4(a0)
1.294 + sw t2, 8(a0)
1.295 + sw t3, 12(a0)
1.296 + sw t4, 16(a0)
1.297 + sw t5, 20(a0)
1.298 + sw t6, 24(a0)
1.299 + sw t7, 28(a0)
1.300 + beqz a2, 4f
1.301 + addiu a0, a0, 32
1.302 +3:
1.303 + lw t0, 0(a1)
1.304 + addiu a1, a1, 4
1.305 + addiu a2, a2, -1
1.306 + or t1, t0, t9
1.307 + sw t1, 0(a0)
1.308 + bnez a2, 3b
1.309 + addiu a0, a0, 4
1.310 +4:
1.311 + jr ra
1.312 + nop
1.313 +
1.314 +END(pixman_composite_src_x888_8888_asm_mips)
1.315 +
1.316 +LEAF_MIPS_DSPR2(pixman_composite_src_n_8_8888_asm_mips)
1.317 +/*
1.318 + * a0 - dst (a8r8g8b8)
1.319 + * a1 - src (32bit constant)
1.320 + * a2 - mask (a8)
1.321 + * a3 - w
1.322 + */
1.323 +
1.324 +
1.325 + SAVE_REGS_ON_STACK 0, v0
1.326 + li v0, 0x00ff00ff
1.327 +
1.328 + beqz a3, 3f
1.329 + nop
1.330 + addiu t1, a3, -1
1.331 + beqz t1, 2f
1.332 + nop
1.333 +
1.334 +1:
1.335 + /* a1 = source (32bit constant) */
1.336 + lbu t0, 0(a2) /* t2 = mask (a8) */
1.337 + lbu t1, 1(a2) /* t3 = mask (a8) */
1.338 + addiu a2, a2, 2
1.339 +
1.340 + MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, t2, t3, v0, t4, t5, t6, t7, t8, t9
1.341 +
1.342 + sw t2, 0(a0)
1.343 + sw t3, 4(a0)
1.344 + addiu a3, a3, -2
1.345 + addiu t2, a3, -1
1.346 + bgtz t2, 1b
1.347 + addiu a0, a0, 8
1.348 +
1.349 + beqz a3, 3f
1.350 + nop
1.351 +
1.352 +2:
1.353 + lbu t0, 0(a2)
1.354 + addiu a2, a2, 1
1.355 +
1.356 + MIPS_UN8x4_MUL_UN8 a1, t0, t1, v0, t3, t4, t5
1.357 +
1.358 + sw t1, 0(a0)
1.359 + addiu a3, a3, -1
1.360 + addiu a0, a0, 4
1.361 +
1.362 +3:
1.363 + RESTORE_REGS_FROM_STACK 0, v0
1.364 + j ra
1.365 + nop
1.366 +
1.367 +END(pixman_composite_src_n_8_8888_asm_mips)
1.368 +
1.369 +LEAF_MIPS_DSPR2(pixman_composite_src_n_8_8_asm_mips)
1.370 +/*
1.371 + * a0 - dst (a8)
1.372 + * a1 - src (32bit constant)
1.373 + * a2 - mask (a8)
1.374 + * a3 - w
1.375 + */
1.376 +
1.377 + li t9, 0x00ff00ff
1.378 + beqz a3, 3f
1.379 + nop
1.380 + srl t7, a3, 2 /* t7 = how many multiples of 4 dst pixels */
1.381 + beqz t7, 1f /* branch if less than 4 src pixels */
1.382 + nop
1.383 +
1.384 + srl t8, a1, 24
1.385 + replv.ph t8, t8
1.386 +
1.387 +0:
1.388 + beqz t7, 1f
1.389 + addiu t7, t7, -1
1.390 + lbu t0, 0(a2)
1.391 + lbu t1, 1(a2)
1.392 + lbu t2, 2(a2)
1.393 + lbu t3, 3(a2)
1.394 +
1.395 + addiu a2, a2, 4
1.396 +
1.397 + precr_sra.ph.w t1, t0, 0
1.398 + precr_sra.ph.w t3, t2, 0
1.399 + precr.qb.ph t0, t3, t1
1.400 +
1.401 + muleu_s.ph.qbl t2, t0, t8
1.402 + muleu_s.ph.qbr t3, t0, t8
1.403 + shra_r.ph t4, t2, 8
1.404 + shra_r.ph t5, t3, 8
1.405 + and t4, t4, t9
1.406 + and t5, t5, t9
1.407 + addq.ph t2, t2, t4
1.408 + addq.ph t3, t3, t5
1.409 + shra_r.ph t2, t2, 8
1.410 + shra_r.ph t3, t3, 8
1.411 + precr.qb.ph t2, t2, t3
1.412 +
1.413 + sb t2, 0(a0)
1.414 + srl t2, t2, 8
1.415 + sb t2, 1(a0)
1.416 + srl t2, t2, 8
1.417 + sb t2, 2(a0)
1.418 + srl t2, t2, 8
1.419 + sb t2, 3(a0)
1.420 + addiu a3, a3, -4
1.421 + b 0b
1.422 + addiu a0, a0, 4
1.423 +
1.424 +1:
1.425 + beqz a3, 3f
1.426 + nop
1.427 + srl t8, a1, 24
1.428 +2:
1.429 + lbu t0, 0(a2)
1.430 + addiu a2, a2, 1
1.431 +
1.432 + mul t2, t0, t8
1.433 + shra_r.ph t3, t2, 8
1.434 + andi t3, t3, 0x00ff
1.435 + addq.ph t2, t2, t3
1.436 + shra_r.ph t2, t2, 8
1.437 +
1.438 + sb t2, 0(a0)
1.439 + addiu a3, a3, -1
1.440 + bnez a3, 2b
1.441 + addiu a0, a0, 1
1.442 +
1.443 +3:
1.444 + j ra
1.445 + nop
1.446 +
1.447 +END(pixman_composite_src_n_8_8_asm_mips)
1.448 +
1.449 +LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_8888_ca_asm_mips)
1.450 +/*
1.451 + * a0 - dst (a8r8g8b8)
1.452 + * a1 - src (32bit constant)
1.453 + * a2 - mask (a8r8g8b8)
1.454 + * a3 - w
1.455 + */
1.456 +
1.457 + SAVE_REGS_ON_STACK 8, s0, s1, s2, s3, s4, s5
1.458 + beqz a3, 4f
1.459 + nop
1.460 + li t6, 0xff
1.461 + addiu t7, zero, -1 /* t7 = 0xffffffff */
1.462 + srl t8, a1, 24 /* t8 = srca */
1.463 + li t9, 0x00ff00ff
1.464 + addiu t1, a3, -1
1.465 + beqz t1, 3f /* last pixel */
1.466 + nop
1.467 + beq t8, t6, 2f /* if (srca == 0xff) */
1.468 + nop
1.469 +1:
1.470 + /* a1 = src */
1.471 + lw t0, 0(a2) /* t0 = mask */
1.472 + lw t1, 4(a2) /* t1 = mask */
1.473 + or t2, t0, t1
1.474 + beqz t2, 12f /* if (t0 == 0) && (t1 == 0) */
1.475 + addiu a2, a2, 8
1.476 + and t3, t0, t1
1.477 + move t4, a1 /* t4 = src */
1.478 + move t5, a1 /* t5 = src */
1.479 + lw t2, 0(a0) /* t2 = dst */
1.480 + beq t3, t7, 11f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */
1.481 + lw t3, 4(a0) /* t3 = dst */
1.482 + MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, t4, t5, t9, s0, s1, s2, s3, s4, s5
1.483 + MIPS_2xUN8x4_MUL_2xUN8 t0, t1, t8, t8, t0, t1, t9, s0, s1, s2, s3, s4, s5
1.484 +11:
1.485 + not t0, t0
1.486 + not t1, t1
1.487 + MIPS_2xUN8x4_MUL_2xUN8x4 t2, t3, t0, t1, t2, t3, t9, s0, s1, s2, s3, s4, s5
1.488 + addu_s.qb t2, t4, t2
1.489 + addu_s.qb t3, t5, t3
1.490 + sw t2, 0(a0)
1.491 + sw t3, 4(a0)
1.492 +12:
1.493 + addiu a3, a3, -2
1.494 + addiu t1, a3, -1
1.495 + bgtz t1, 1b
1.496 + addiu a0, a0, 8
1.497 + b 3f
1.498 + nop
1.499 +2:
1.500 + /* a1 = src */
1.501 + lw t0, 0(a2) /* t0 = mask */
1.502 + lw t1, 4(a2) /* t1 = mask */
1.503 + or t2, t0, t1
1.504 + beqz t2, 22f /* if (t0 == 0) & (t1 == 0) */
1.505 + addiu a2, a2, 8
1.506 + and t2, t0, t1
1.507 + move t4, a1
1.508 + beq t2, t7, 21f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */
1.509 + move t5, a1
1.510 + lw t2, 0(a0) /* t2 = dst */
1.511 + lw t3, 4(a0) /* t3 = dst */
1.512 + MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, t4, t5, t9, s0, s1, s2, s3, s4, s5
1.513 + not t0, t0
1.514 + not t1, t1
1.515 + MIPS_2xUN8x4_MUL_2xUN8x4 t2, t3, t0, t1, t2, t3, t9, s0, s1, s2, s3, s4, s5
1.516 + addu_s.qb t4, t4, t2
1.517 + addu_s.qb t5, t5, t3
1.518 +21:
1.519 + sw t4, 0(a0)
1.520 + sw t5, 4(a0)
1.521 +22:
1.522 + addiu a3, a3, -2
1.523 + addiu t1, a3, -1
1.524 + bgtz t1, 2b
1.525 + addiu a0, a0, 8
1.526 +3:
1.527 + blez a3, 4f
1.528 + nop
1.529 + /* a1 = src */
1.530 + lw t1, 0(a2) /* t1 = mask */
1.531 + beqz t1, 4f
1.532 + nop
1.533 + move t2, a1 /* t2 = src */
1.534 + beq t1, t7, 31f
1.535 + lw t0, 0(a0) /* t0 = dst */
1.536 +
1.537 + MIPS_UN8x4_MUL_UN8x4 a1, t1, t2, t9, t3, t4, t5, t6
1.538 + MIPS_UN8x4_MUL_UN8 t1, t8, t1, t9, t3, t4, t5
1.539 +31:
1.540 + not t1, t1
1.541 + MIPS_UN8x4_MUL_UN8x4 t0, t1, t0, t9, t3, t4, t5, t6
1.542 + addu_s.qb t0, t2, t0
1.543 + sw t0, 0(a0)
1.544 +4:
1.545 + RESTORE_REGS_FROM_STACK 8, s0, s1, s2, s3, s4, s5
1.546 + j ra
1.547 + nop
1.548 +
1.549 +END(pixman_composite_over_n_8888_8888_ca_asm_mips)
1.550 +
1.551 +LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_0565_ca_asm_mips)
1.552 +/*
1.553 + * a0 - dst (r5g6b5)
1.554 + * a1 - src (32bit constant)
1.555 + * a2 - mask (a8r8g8b8)
1.556 + * a3 - w
1.557 + */
1.558 +
1.559 + SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.560 + beqz a3, 4f
1.561 + nop
1.562 + li t5, 0xf800f800
1.563 + li t6, 0x07e007e0
1.564 + li t7, 0x001F001F
1.565 + li t9, 0x00ff00ff
1.566 +
1.567 + srl t8, a1, 24 /* t8 = srca */
1.568 + addiu t1, a3, -1
1.569 + beqz t1, 3f /* last pixel */
1.570 + nop
1.571 + li s0, 0xff /* s0 = 0xff */
1.572 + addiu s1, zero, -1 /* s1 = 0xffffffff */
1.573 +
1.574 + beq t8, s0, 2f /* if (srca == 0xff) */
1.575 + nop
1.576 +1:
1.577 + /* a1 = src */
1.578 + lw t0, 0(a2) /* t0 = mask */
1.579 + lw t1, 4(a2) /* t1 = mask */
1.580 + or t2, t0, t1
1.581 + beqz t2, 12f /* if (t0 == 0) && (t1 == 0) */
1.582 + addiu a2, a2, 8
1.583 + and t3, t0, t1
1.584 + move s2, a1 /* s2 = src */
1.585 + move s3, a1 /* s3 = src */
1.586 + lhu t2, 0(a0) /* t2 = dst */
1.587 + beq t3, s1, 11f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */
1.588 + lhu t3, 2(a0) /* t3 = dst */
1.589 + MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, s2, s3, t9, t4, s4, s5, s6, s7, s8
1.590 + MIPS_2xUN8x4_MUL_2xUN8 t0, t1, t8, t8, t0, t1, t9, t4, s4, s5, s6, s7, s8
1.591 +11:
1.592 + not t0, t0
1.593 + not t1, t1
1.594 + CONVERT_2x0565_TO_2x8888 t2, t3, s4, s5, t6, t7, t4, s6, s7, s8
1.595 + MIPS_2xUN8x4_MUL_2xUN8x4 s4, s5, t0, t1, s4, s5, t9, t4, s6, s7, s8, t0, t1
1.596 + addu_s.qb s2, s2, s4
1.597 + addu_s.qb s3, s3, s5
1.598 + CONVERT_2x8888_TO_2x0565 s2, s3, t2, t3, t5, t6, t7, s4, s5
1.599 + sh t2, 0(a0)
1.600 + sh t3, 2(a0)
1.601 +12:
1.602 + addiu a3, a3, -2
1.603 + addiu t1, a3, -1
1.604 + bgtz t1, 1b
1.605 + addiu a0, a0, 4
1.606 + b 3f
1.607 + nop
1.608 +2:
1.609 + /* a1 = src */
1.610 + lw t0, 0(a2) /* t0 = mask */
1.611 + lw t1, 4(a2) /* t1 = mask */
1.612 + or t2, t0, t1
1.613 + beqz t2, 22f /* if (t0 == 0) & (t1 == 0) */
1.614 + addiu a2, a2, 8
1.615 + and t3, t0, t1
1.616 + move t2, a1
1.617 + beq t3, s1, 21f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */
1.618 + move t3, a1
1.619 + lhu t2, 0(a0) /* t2 = dst */
1.620 + lhu t3, 2(a0) /* t3 = dst */
1.621 + MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, s2, s3, t9, t4, s4, s5, s6, s7, s8
1.622 + not t0, t0
1.623 + not t1, t1
1.624 + CONVERT_2x0565_TO_2x8888 t2, t3, s4, s5, t6, t7, t4, s6, s7, s8
1.625 + MIPS_2xUN8x4_MUL_2xUN8x4 s4, s5, t0, t1, s4, s5, t9, t4, s6, s7, s8, t2, t3
1.626 + addu_s.qb t2, s2, s4
1.627 + addu_s.qb t3, s3, s5
1.628 +21:
1.629 + CONVERT_2x8888_TO_2x0565 t2, t3, t0, t1, t5, t6, t7, s2, s3
1.630 + sh t0, 0(a0)
1.631 + sh t1, 2(a0)
1.632 +22:
1.633 + addiu a3, a3, -2
1.634 + addiu t1, a3, -1
1.635 + bgtz t1, 2b
1.636 + addiu a0, a0, 4
1.637 +3:
1.638 + blez a3, 4f
1.639 + nop
1.640 + /* a1 = src */
1.641 + lw t1, 0(a2) /* t1 = mask */
1.642 + beqz t1, 4f
1.643 + nop
1.644 + move t2, a1 /* t2 = src */
1.645 + beq t1, t7, 31f
1.646 + lhu t0, 0(a0) /* t0 = dst */
1.647 +
1.648 + MIPS_UN8x4_MUL_UN8x4 a1, t1, t2, t9, t3, t4, t5, t6
1.649 + MIPS_UN8x4_MUL_UN8 t1, t8, t1, t9, t3, t4, t5
1.650 +31:
1.651 + not t1, t1
1.652 + CONVERT_1x0565_TO_1x8888 t0, s1, s2, s3
1.653 + MIPS_UN8x4_MUL_UN8x4 s1, t1, t3, t9, t4, t5, t6, t7
1.654 + addu_s.qb t0, t2, t3
1.655 + CONVERT_1x8888_TO_1x0565 t0, s1, s2, s3
1.656 + sh s1, 0(a0)
1.657 +4:
1.658 + RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.659 + j ra
1.660 + nop
1.661 +
1.662 +END(pixman_composite_over_n_8888_0565_ca_asm_mips)
1.663 +
1.664 +LEAF_MIPS_DSPR2(pixman_composite_over_n_8_8888_asm_mips)
1.665 +/*
1.666 + * a0 - dst (a8r8g8b8)
1.667 + * a1 - src (32bit constant)
1.668 + * a2 - mask (a8)
1.669 + * a3 - w
1.670 + */
1.671 +
1.672 + SAVE_REGS_ON_STACK 4, s0, s1, s2, s3, s4
1.673 + beqz a3, 4f
1.674 + nop
1.675 + li t4, 0x00ff00ff
1.676 + li t5, 0xff
1.677 + addiu t0, a3, -1
1.678 + beqz t0, 3f /* last pixel */
1.679 + srl t6, a1, 24 /* t6 = srca */
1.680 + not s4, a1
1.681 + beq t5, t6, 2f /* if (srca == 0xff) */
1.682 + srl s4, s4, 24
1.683 +1:
1.684 + /* a1 = src */
1.685 + lbu t0, 0(a2) /* t0 = mask */
1.686 + lbu t1, 1(a2) /* t1 = mask */
1.687 + or t2, t0, t1
1.688 + beqz t2, 111f /* if (t0 == 0) && (t1 == 0) */
1.689 + addiu a2, a2, 2
1.690 + and t3, t0, t1
1.691 +
1.692 + lw t2, 0(a0) /* t2 = dst */
1.693 + beq t3, t5, 11f /* if (t0 == 0xff) && (t1 == 0xff) */
1.694 + lw t3, 4(a0) /* t3 = dst */
1.695 +
1.696 + MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, s0, s1, t4, t6, t7, t8, t9, s2, s3
1.697 + not s2, s0
1.698 + not s3, s1
1.699 + srl s2, s2, 24
1.700 + srl s3, s3, 24
1.701 + MIPS_2xUN8x4_MUL_2xUN8 t2, t3, s2, s3, t2, t3, t4, t0, t1, t6, t7, t8, t9
1.702 + addu_s.qb s2, t2, s0
1.703 + addu_s.qb s3, t3, s1
1.704 + sw s2, 0(a0)
1.705 + b 111f
1.706 + sw s3, 4(a0)
1.707 +11:
1.708 + MIPS_2xUN8x4_MUL_2xUN8 t2, t3, s4, s4, t2, t3, t4, t0, t1, t6, t7, t8, t9
1.709 + addu_s.qb s2, t2, a1
1.710 + addu_s.qb s3, t3, a1
1.711 + sw s2, 0(a0)
1.712 + sw s3, 4(a0)
1.713 +
1.714 +111:
1.715 + addiu a3, a3, -2
1.716 + addiu t0, a3, -1
1.717 + bgtz t0, 1b
1.718 + addiu a0, a0, 8
1.719 + b 3f
1.720 + nop
1.721 +2:
1.722 + /* a1 = src */
1.723 + lbu t0, 0(a2) /* t0 = mask */
1.724 + lbu t1, 1(a2) /* t1 = mask */
1.725 + or t2, t0, t1
1.726 + beqz t2, 222f /* if (t0 == 0) && (t1 == 0) */
1.727 + addiu a2, a2, 2
1.728 + and t3, t0, t1
1.729 + beq t3, t5, 22f /* if (t0 == 0xff) && (t1 == 0xff) */
1.730 + nop
1.731 + lw t2, 0(a0) /* t2 = dst */
1.732 + lw t3, 4(a0) /* t3 = dst */
1.733 +
1.734 + OVER_2x8888_2x8_2x8888 a1, a1, t0, t1, t2, t3, \
1.735 + t6, t7, t4, t8, t9, s0, s1, s2, s3
1.736 + sw t6, 0(a0)
1.737 + b 222f
1.738 + sw t7, 4(a0)
1.739 +22:
1.740 + sw a1, 0(a0)
1.741 + sw a1, 4(a0)
1.742 +222:
1.743 + addiu a3, a3, -2
1.744 + addiu t0, a3, -1
1.745 + bgtz t0, 2b
1.746 + addiu a0, a0, 8
1.747 +3:
1.748 + blez a3, 4f
1.749 + nop
1.750 + /* a1 = src */
1.751 + lbu t0, 0(a2) /* t0 = mask */
1.752 + beqz t0, 4f /* if (t0 == 0) */
1.753 + addiu a2, a2, 1
1.754 + move t3, a1
1.755 + beq t0, t5, 31f /* if (t0 == 0xff) */
1.756 + lw t1, 0(a0) /* t1 = dst */
1.757 +
1.758 + MIPS_UN8x4_MUL_UN8 a1, t0, t3, t4, t6, t7, t8
1.759 +31:
1.760 + not t2, t3
1.761 + srl t2, t2, 24
1.762 + MIPS_UN8x4_MUL_UN8 t1, t2, t1, t4, t6, t7, t8
1.763 + addu_s.qb t2, t1, t3
1.764 + sw t2, 0(a0)
1.765 +4:
1.766 + RESTORE_REGS_FROM_STACK 4, s0, s1, s2, s3, s4
1.767 + j ra
1.768 + nop
1.769 +
1.770 +END(pixman_composite_over_n_8_8888_asm_mips)
1.771 +
1.772 +LEAF_MIPS_DSPR2(pixman_composite_over_n_8_0565_asm_mips)
1.773 +/*
1.774 + * a0 - dst (r5g6b5)
1.775 + * a1 - src (32bit constant)
1.776 + * a2 - mask (a8)
1.777 + * a3 - w
1.778 + */
1.779 + SAVE_REGS_ON_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.780 + beqz a3, 4f
1.781 + nop
1.782 + li t4, 0x00ff00ff
1.783 + li t5, 0xff
1.784 + li t6, 0xf800f800
1.785 + li t7, 0x07e007e0
1.786 + li t8, 0x001F001F
1.787 + addiu t1, a3, -1
1.788 + beqz t1, 3f /* last pixel */
1.789 + srl t0, a1, 24 /* t0 = srca */
1.790 + not v0, a1
1.791 + beq t0, t5, 2f /* if (srca == 0xff) */
1.792 + srl v0, v0, 24
1.793 +1:
1.794 + /* a1 = src */
1.795 + lbu t0, 0(a2) /* t0 = mask */
1.796 + lbu t1, 1(a2) /* t1 = mask */
1.797 + or t2, t0, t1
1.798 + beqz t2, 111f /* if (t0 == 0) && (t1 == 0) */
1.799 + addiu a2, a2, 2
1.800 + lhu t2, 0(a0) /* t2 = dst */
1.801 + lhu t3, 2(a0) /* t3 = dst */
1.802 + CONVERT_2x0565_TO_2x8888 t2, t3, s0, s1, t7, t8, t9, s2, s3, s4
1.803 + and t9, t0, t1
1.804 + beq t9, t5, 11f /* if (t0 == 0xff) && (t1 == 0xff) */
1.805 + nop
1.806 +
1.807 + MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, s2, s3, t4, t9, s4, s5, s6, s7, s8
1.808 + not s4, s2
1.809 + not s5, s3
1.810 + srl s4, s4, 24
1.811 + srl s5, s5, 24
1.812 + MIPS_2xUN8x4_MUL_2xUN8 s0, s1, s4, s5, s0, s1, t4, t9, t0, t1, s6, s7, s8
1.813 + addu_s.qb s4, s2, s0
1.814 + addu_s.qb s5, s3, s1
1.815 + CONVERT_2x8888_TO_2x0565 s4, s5, t2, t3, t6, t7, t8, s0, s1
1.816 + sh t2, 0(a0)
1.817 + b 111f
1.818 + sh t3, 2(a0)
1.819 +11:
1.820 + MIPS_2xUN8x4_MUL_2xUN8 s0, s1, v0, v0, s0, s1, t4, t9, t0, t1, s6, s7, s8
1.821 + addu_s.qb s4, a1, s0
1.822 + addu_s.qb s5, a1, s1
1.823 + CONVERT_2x8888_TO_2x0565 s4, s5, t2, t3, t6, t7, t8, s0, s1
1.824 + sh t2, 0(a0)
1.825 + sh t3, 2(a0)
1.826 +111:
1.827 + addiu a3, a3, -2
1.828 + addiu t0, a3, -1
1.829 + bgtz t0, 1b
1.830 + addiu a0, a0, 4
1.831 + b 3f
1.832 + nop
1.833 +2:
1.834 + CONVERT_1x8888_TO_1x0565 a1, s0, s1, s2
1.835 +21:
1.836 + /* a1 = src */
1.837 + lbu t0, 0(a2) /* t0 = mask */
1.838 + lbu t1, 1(a2) /* t1 = mask */
1.839 + or t2, t0, t1
1.840 + beqz t2, 222f /* if (t0 == 0) && (t1 == 0) */
1.841 + addiu a2, a2, 2
1.842 + and t9, t0, t1
1.843 + move s2, s0
1.844 + beq t9, t5, 22f /* if (t0 == 0xff) && (t2 == 0xff) */
1.845 + move s3, s0
1.846 + lhu t2, 0(a0) /* t2 = dst */
1.847 + lhu t3, 2(a0) /* t3 = dst */
1.848 +
1.849 + CONVERT_2x0565_TO_2x8888 t2, t3, s2, s3, t7, t8, s4, s5, s6, s7
1.850 + OVER_2x8888_2x8_2x8888 a1, a1, t0, t1, s2, s3, \
1.851 + t2, t3, t4, t9, s4, s5, s6, s7, s8
1.852 + CONVERT_2x8888_TO_2x0565 t2, t3, s2, s3, t6, t7, t8, s4, s5
1.853 +22:
1.854 + sh s2, 0(a0)
1.855 + sh s3, 2(a0)
1.856 +222:
1.857 + addiu a3, a3, -2
1.858 + addiu t0, a3, -1
1.859 + bgtz t0, 21b
1.860 + addiu a0, a0, 4
1.861 +3:
1.862 + blez a3, 4f
1.863 + nop
1.864 + /* a1 = src */
1.865 + lbu t0, 0(a2) /* t0 = mask */
1.866 + beqz t0, 4f /* if (t0 == 0) */
1.867 + nop
1.868 + lhu t1, 0(a0) /* t1 = dst */
1.869 + CONVERT_1x0565_TO_1x8888 t1, t2, t3, t7
1.870 + beq t0, t5, 31f /* if (t0 == 0xff) */
1.871 + move t3, a1
1.872 +
1.873 + MIPS_UN8x4_MUL_UN8 a1, t0, t3, t4, t7, t8, t9
1.874 +31:
1.875 + not t6, t3
1.876 + srl t6, t6, 24
1.877 + MIPS_UN8x4_MUL_UN8 t2, t6, t2, t4, t7, t8, t9
1.878 + addu_s.qb t1, t2, t3
1.879 + CONVERT_1x8888_TO_1x0565 t1, t2, t3, t7
1.880 + sh t2, 0(a0)
1.881 +4:
1.882 + RESTORE_REGS_FROM_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.883 + j ra
1.884 + nop
1.885 +
1.886 +END(pixman_composite_over_n_8_0565_asm_mips)
1.887 +
1.888 +LEAF_MIPS_DSPR2(pixman_composite_over_8888_n_8888_asm_mips)
1.889 +/*
1.890 + * a0 - dst (a8r8g8b8)
1.891 + * a1 - src (a8r8g8b8)
1.892 + * a2 - mask (32bit constant)
1.893 + * a3 - w
1.894 + */
1.895 +
1.896 + SAVE_REGS_ON_STACK 0, s0
1.897 + li t4, 0x00ff00ff
1.898 + beqz a3, 3f
1.899 + nop
1.900 + addiu t1, a3, -1
1.901 + srl a2, a2, 24
1.902 + beqz t1, 2f
1.903 + nop
1.904 +
1.905 +1:
1.906 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.907 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.908 + /* a2 = mask (32bit constant) */
1.909 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.910 + lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
1.911 + addiu a1, a1, 8
1.912 +
1.913 + OVER_2x8888_2x8_2x8888 t0, t1, a2, a2, t2, t3, \
1.914 + t5, t6, t4, t7, t8, t9, t0, t1, s0
1.915 +
1.916 + sw t5, 0(a0)
1.917 + sw t6, 4(a0)
1.918 + addiu a3, a3, -2
1.919 + addiu t1, a3, -1
1.920 + bgtz t1, 1b
1.921 + addiu a0, a0, 8
1.922 +2:
1.923 + beqz a3, 3f
1.924 + nop
1.925 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.926 + /* a2 = mask (32bit constant) */
1.927 + lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
1.928 +
1.929 + OVER_8888_8_8888 t0, a2, t1, t3, t4, t5, t6, t7, t8
1.930 +
1.931 + sw t3, 0(a0)
1.932 +3:
1.933 + RESTORE_REGS_FROM_STACK 0, s0
1.934 + j ra
1.935 + nop
1.936 +
1.937 +END(pixman_composite_over_8888_n_8888_asm_mips)
1.938 +
1.939 +LEAF_MIPS_DSPR2(pixman_composite_over_8888_n_0565_asm_mips)
1.940 +/*
1.941 + * a0 - dst (r5g6b5)
1.942 + * a1 - src (a8r8g8b8)
1.943 + * a2 - mask (32bit constant)
1.944 + * a3 - w
1.945 + */
1.946 +
1.947 + SAVE_REGS_ON_STACK 0, s0, s1, s2, s3
1.948 + li t6, 0x00ff00ff
1.949 + li t7, 0xf800f800
1.950 + li t8, 0x07e007e0
1.951 + li t9, 0x001F001F
1.952 + beqz a3, 3f
1.953 + nop
1.954 + srl a2, a2, 24
1.955 + addiu t1, a3, -1
1.956 + beqz t1, 2f
1.957 + nop
1.958 +1:
1.959 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.960 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.961 + /* a2 = mask (32bit constant) */
1.962 + lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
1.963 + lhu t3, 2(a0) /* t2 = destination (r5g6b5) */
1.964 + addiu a1, a1, 8
1.965 +
1.966 + CONVERT_2x0565_TO_2x8888 t2, t3, t4, t5, t8, t9, s0, s1, t2, t3
1.967 + OVER_2x8888_2x8_2x8888 t0, t1, a2, a2, t4, t5, \
1.968 + t2, t3, t6, t0, t1, s0, s1, s2, s3
1.969 + CONVERT_2x8888_TO_2x0565 t2, t3, t4, t5, t7, t8, t9, s0, s1
1.970 +
1.971 + sh t4, 0(a0)
1.972 + sh t5, 2(a0)
1.973 + addiu a3, a3, -2
1.974 + addiu t1, a3, -1
1.975 + bgtz t1, 1b
1.976 + addiu a0, a0, 4
1.977 +2:
1.978 + beqz a3, 3f
1.979 + nop
1.980 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.981 + /* a2 = mask (32bit constant) */
1.982 + lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
1.983 +
1.984 + CONVERT_1x0565_TO_1x8888 t1, t2, t4, t5
1.985 + OVER_8888_8_8888 t0, a2, t2, t1, t6, t3, t4, t5, t7
1.986 + CONVERT_1x8888_TO_1x0565 t1, t3, t4, t5
1.987 +
1.988 + sh t3, 0(a0)
1.989 +3:
1.990 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2, s3
1.991 + j ra
1.992 + nop
1.993 +
1.994 +END(pixman_composite_over_8888_n_0565_asm_mips)
1.995 +
1.996 +LEAF_MIPS_DSPR2(pixman_composite_over_0565_n_0565_asm_mips)
1.997 +/*
1.998 + * a0 - dst (r5g6b5)
1.999 + * a1 - src (r5g6b5)
1.1000 + * a2 - mask (32bit constant)
1.1001 + * a3 - w
1.1002 + */
1.1003 +
1.1004 + SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5
1.1005 + li t6, 0x00ff00ff
1.1006 + li t7, 0xf800f800
1.1007 + li t8, 0x07e007e0
1.1008 + li t9, 0x001F001F
1.1009 + beqz a3, 3f
1.1010 + nop
1.1011 + srl a2, a2, 24
1.1012 + addiu t1, a3, -1
1.1013 + beqz t1, 2f
1.1014 + nop
1.1015 +1:
1.1016 + lhu t0, 0(a1) /* t0 = source (r5g6b5) */
1.1017 + lhu t1, 2(a1) /* t1 = source (r5g6b5) */
1.1018 + /* a2 = mask (32bit constant) */
1.1019 + lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
1.1020 + lhu t3, 2(a0) /* t3 = destination (r5g6b5) */
1.1021 + addiu a1, a1, 4
1.1022 +
1.1023 + CONVERT_2x0565_TO_2x8888 t0, t1, t4, t5, t8, t9, s0, s1, s2, s3
1.1024 + CONVERT_2x0565_TO_2x8888 t2, t3, s0, s1, t8, t9, s2, s3, s4, s5
1.1025 + OVER_2x8888_2x8_2x8888 t4, t5, a2, a2, s0, s1, \
1.1026 + t0, t1, t6, s2, s3, s4, s5, t4, t5
1.1027 + CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t7, t8, t9, s2, s3
1.1028 +
1.1029 + sh s0, 0(a0)
1.1030 + sh s1, 2(a0)
1.1031 + addiu a3, a3, -2
1.1032 + addiu t1, a3, -1
1.1033 + bgtz t1, 1b
1.1034 + addiu a0, a0, 4
1.1035 +2:
1.1036 + beqz a3, 3f
1.1037 + nop
1.1038 + lhu t0, 0(a1) /* t0 = source (r5g6b5) */
1.1039 + /* a2 = mask (32bit constant) */
1.1040 + lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
1.1041 +
1.1042 + CONVERT_1x0565_TO_1x8888 t0, t2, t4, t5
1.1043 + CONVERT_1x0565_TO_1x8888 t1, t3, t4, t5
1.1044 + OVER_8888_8_8888 t2, a2, t3, t0, t6, t1, t4, t5, t7
1.1045 + CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5
1.1046 +
1.1047 + sh t3, 0(a0)
1.1048 +3:
1.1049 + RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5
1.1050 + j ra
1.1051 + nop
1.1052 +
1.1053 +END(pixman_composite_over_0565_n_0565_asm_mips)
1.1054 +
1.1055 +LEAF_MIPS_DSPR2(pixman_composite_over_8888_8_8888_asm_mips)
1.1056 +/*
1.1057 + * a0 - dst (a8r8g8b8)
1.1058 + * a1 - src (a8r8g8b8)
1.1059 + * a2 - mask (a8)
1.1060 + * a3 - w
1.1061 + */
1.1062 +
1.1063 + SAVE_REGS_ON_STACK 0, s0, s1
1.1064 + li t4, 0x00ff00ff
1.1065 + beqz a3, 3f
1.1066 + nop
1.1067 + addiu t1, a3, -1
1.1068 + beqz t1, 2f
1.1069 + nop
1.1070 +1:
1.1071 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1072 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.1073 + lbu t2, 0(a2) /* t2 = mask (a8) */
1.1074 + lbu t3, 1(a2) /* t3 = mask (a8) */
1.1075 + lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */
1.1076 + lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */
1.1077 + addiu a1, a1, 8
1.1078 + addiu a2, a2, 2
1.1079 +
1.1080 + OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, t5, t6, \
1.1081 + t7, t8, t4, t9, s0, s1, t0, t1, t2
1.1082 +
1.1083 + sw t7, 0(a0)
1.1084 + sw t8, 4(a0)
1.1085 + addiu a3, a3, -2
1.1086 + addiu t1, a3, -1
1.1087 + bgtz t1, 1b
1.1088 + addiu a0, a0, 8
1.1089 +2:
1.1090 + beqz a3, 3f
1.1091 + nop
1.1092 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1093 + lbu t1, 0(a2) /* t1 = mask (a8) */
1.1094 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.1095 +
1.1096 + OVER_8888_8_8888 t0, t1, t2, t3, t4, t5, t6, t7, t8
1.1097 +
1.1098 + sw t3, 0(a0)
1.1099 +3:
1.1100 + RESTORE_REGS_FROM_STACK 0, s0, s1
1.1101 + j ra
1.1102 + nop
1.1103 +
1.1104 +END(pixman_composite_over_8888_8_8888_asm_mips)
1.1105 +
1.1106 +LEAF_MIPS_DSPR2(pixman_composite_over_8888_8_0565_asm_mips)
1.1107 +/*
1.1108 + * a0 - dst (r5g6b5)
1.1109 + * a1 - src (a8r8g8b8)
1.1110 + * a2 - mask (a8)
1.1111 + * a3 - w
1.1112 + */
1.1113 +
1.1114 + SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5
1.1115 + li t6, 0x00ff00ff
1.1116 + li t7, 0xf800f800
1.1117 + li t8, 0x07e007e0
1.1118 + li t9, 0x001F001F
1.1119 + beqz a3, 3f
1.1120 + nop
1.1121 + addiu t1, a3, -1
1.1122 + beqz t1, 2f
1.1123 + nop
1.1124 +1:
1.1125 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1126 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.1127 + lbu t2, 0(a2) /* t2 = mask (a8) */
1.1128 + lbu t3, 1(a2) /* t3 = mask (a8) */
1.1129 + lhu t4, 0(a0) /* t4 = destination (r5g6b5) */
1.1130 + lhu t5, 2(a0) /* t5 = destination (r5g6b5) */
1.1131 + addiu a1, a1, 8
1.1132 + addiu a2, a2, 2
1.1133 +
1.1134 + CONVERT_2x0565_TO_2x8888 t4, t5, s0, s1, t8, t9, s2, s3, s4, s5
1.1135 + OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, s0, s1, \
1.1136 + t4, t5, t6, s2, s3, s4, s5, t0, t1
1.1137 + CONVERT_2x8888_TO_2x0565 t4, t5, s0, s1, t7, t8, t9, s2, s3
1.1138 +
1.1139 + sh s0, 0(a0)
1.1140 + sh s1, 2(a0)
1.1141 + addiu a3, a3, -2
1.1142 + addiu t1, a3, -1
1.1143 + bgtz t1, 1b
1.1144 + addiu a0, a0, 4
1.1145 +2:
1.1146 + beqz a3, 3f
1.1147 + nop
1.1148 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1149 + lbu t1, 0(a2) /* t1 = mask (a8) */
1.1150 + lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
1.1151 +
1.1152 + CONVERT_1x0565_TO_1x8888 t2, t3, t4, t5
1.1153 + OVER_8888_8_8888 t0, t1, t3, t2, t6, t4, t5, t7, t8
1.1154 + CONVERT_1x8888_TO_1x0565 t2, t3, t4, t5
1.1155 +
1.1156 + sh t3, 0(a0)
1.1157 +3:
1.1158 + RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5
1.1159 + j ra
1.1160 + nop
1.1161 +
1.1162 +END(pixman_composite_over_8888_8_0565_asm_mips)
1.1163 +
1.1164 +LEAF_MIPS_DSPR2(pixman_composite_over_0565_8_0565_asm_mips)
1.1165 +/*
1.1166 + * a0 - dst (r5g6b5)
1.1167 + * a1 - src (r5g6b5)
1.1168 + * a2 - mask (a8)
1.1169 + * a3 - w
1.1170 + */
1.1171 +
1.1172 + SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5
1.1173 + li t4, 0xf800f800
1.1174 + li t5, 0x07e007e0
1.1175 + li t6, 0x001F001F
1.1176 + li t7, 0x00ff00ff
1.1177 + beqz a3, 3f
1.1178 + nop
1.1179 + addiu t1, a3, -1
1.1180 + beqz t1, 2f
1.1181 + nop
1.1182 +1:
1.1183 + lhu t0, 0(a1) /* t0 = source (r5g6b5) */
1.1184 + lhu t1, 2(a1) /* t1 = source (r5g6b5) */
1.1185 + lbu t2, 0(a2) /* t2 = mask (a8) */
1.1186 + lbu t3, 1(a2) /* t3 = mask (a8) */
1.1187 + lhu t8, 0(a0) /* t8 = destination (r5g6b5) */
1.1188 + lhu t9, 2(a0) /* t9 = destination (r5g6b5) */
1.1189 + addiu a1, a1, 4
1.1190 + addiu a2, a2, 2
1.1191 +
1.1192 + CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5
1.1193 + CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, t0, t1
1.1194 + OVER_2x8888_2x8_2x8888 s0, s1, t2, t3, s2, s3, \
1.1195 + t0, t1, t7, s4, s5, t8, t9, s0, s1
1.1196 + CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3
1.1197 +
1.1198 + sh s0, 0(a0)
1.1199 + sh s1, 2(a0)
1.1200 + addiu a3, a3, -2
1.1201 + addiu t1, a3, -1
1.1202 + bgtz t1, 1b
1.1203 + addiu a0, a0, 4
1.1204 +2:
1.1205 + beqz a3, 3f
1.1206 + nop
1.1207 + lhu t0, 0(a1) /* t0 = source (r5g6b5) */
1.1208 + lbu t1, 0(a2) /* t1 = mask (a8) */
1.1209 + lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
1.1210 +
1.1211 + CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5
1.1212 + CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6
1.1213 + OVER_8888_8_8888 t3, t1, t4, t0, t7, t2, t5, t6, t8
1.1214 + CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5
1.1215 +
1.1216 + sh t3, 0(a0)
1.1217 +3:
1.1218 + RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5
1.1219 + j ra
1.1220 + nop
1.1221 +
1.1222 +END(pixman_composite_over_0565_8_0565_asm_mips)
1.1223 +
1.1224 +LEAF_MIPS_DSPR2(pixman_composite_over_8888_8888_8888_asm_mips)
1.1225 +/*
1.1226 + * a0 - dst (a8r8g8b8)
1.1227 + * a1 - src (a8r8g8b8)
1.1228 + * a2 - mask (a8r8g8b8)
1.1229 + * a3 - w
1.1230 + */
1.1231 +
1.1232 + SAVE_REGS_ON_STACK 0, s0, s1, s2
1.1233 + li t4, 0x00ff00ff
1.1234 + beqz a3, 3f
1.1235 + nop
1.1236 + addiu t1, a3, -1
1.1237 + beqz t1, 2f
1.1238 + nop
1.1239 +1:
1.1240 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1241 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.1242 + lw t2, 0(a2) /* t2 = mask (a8r8g8b8) */
1.1243 + lw t3, 4(a2) /* t3 = mask (a8r8g8b8) */
1.1244 + lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */
1.1245 + lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */
1.1246 + addiu a1, a1, 8
1.1247 + addiu a2, a2, 8
1.1248 + srl t2, t2, 24
1.1249 + srl t3, t3, 24
1.1250 +
1.1251 + OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, t5, t6, t7, t8, t4, t9, s0, s1, s2, t0, t1
1.1252 +
1.1253 + sw t7, 0(a0)
1.1254 + sw t8, 4(a0)
1.1255 + addiu a3, a3, -2
1.1256 + addiu t1, a3, -1
1.1257 + bgtz t1, 1b
1.1258 + addiu a0, a0, 8
1.1259 +2:
1.1260 + beqz a3, 3f
1.1261 + nop
1.1262 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1263 + lw t1, 0(a2) /* t1 = mask (a8r8g8b8) */
1.1264 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.1265 + srl t1, t1, 24
1.1266 +
1.1267 + OVER_8888_8_8888 t0, t1, t2, t3, t4, t5, t6, t7, t8
1.1268 +
1.1269 + sw t3, 0(a0)
1.1270 +3:
1.1271 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2
1.1272 + j ra
1.1273 + nop
1.1274 +
1.1275 +END(pixman_composite_over_8888_8888_8888_asm_mips)
1.1276 +
1.1277 +LEAF_MIPS_DSPR2(pixman_composite_over_8888_8888_asm_mips)
1.1278 +/*
1.1279 + * a0 - dst (a8r8g8b8)
1.1280 + * a1 - src (a8r8g8b8)
1.1281 + * a2 - w
1.1282 + */
1.1283 +
1.1284 + SAVE_REGS_ON_STACK 0, s0, s1, s2
1.1285 + li t4, 0x00ff00ff
1.1286 + beqz a2, 3f
1.1287 + nop
1.1288 + addiu t1, a2, -1
1.1289 + beqz t1, 2f
1.1290 + nop
1.1291 +1:
1.1292 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1293 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.1294 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.1295 + lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
1.1296 + addiu a1, a1, 8
1.1297 +
1.1298 + not t5, t0
1.1299 + srl t5, t5, 24
1.1300 + not t6, t1
1.1301 + srl t6, t6, 24
1.1302 +
1.1303 + or t7, t5, t6
1.1304 + beqz t7, 11f
1.1305 + or t8, t0, t1
1.1306 + beqz t8, 12f
1.1307 +
1.1308 + MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t5, t6, t7, t8, t4, t9, s0, s1, s2, t2, t3
1.1309 +
1.1310 + addu_s.qb t0, t7, t0
1.1311 + addu_s.qb t1, t8, t1
1.1312 +11:
1.1313 + sw t0, 0(a0)
1.1314 + sw t1, 4(a0)
1.1315 +12:
1.1316 + addiu a2, a2, -2
1.1317 + addiu t1, a2, -1
1.1318 + bgtz t1, 1b
1.1319 + addiu a0, a0, 8
1.1320 +2:
1.1321 + beqz a2, 3f
1.1322 + nop
1.1323 +
1.1324 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1325 + lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
1.1326 + addiu a1, a1, 4
1.1327 +
1.1328 + not t2, t0
1.1329 + srl t2, t2, 24
1.1330 +
1.1331 + beqz t2, 21f
1.1332 + nop
1.1333 + beqz t0, 3f
1.1334 +
1.1335 + MIPS_UN8x4_MUL_UN8 t1, t2, t3, t4, t5, t6, t7
1.1336 +
1.1337 + addu_s.qb t0, t3, t0
1.1338 +21:
1.1339 + sw t0, 0(a0)
1.1340 +
1.1341 +3:
1.1342 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2
1.1343 + j ra
1.1344 + nop
1.1345 +
1.1346 +END(pixman_composite_over_8888_8888_asm_mips)
1.1347 +
1.1348 +LEAF_MIPS_DSPR2(pixman_composite_over_n_0565_asm_mips)
1.1349 +/*
1.1350 + * a0 - dst (r5g6b5)
1.1351 + * a1 - src (32bit constant)
1.1352 + * a2 - w
1.1353 + */
1.1354 +
1.1355 + beqz a2, 5f
1.1356 + nop
1.1357 +
1.1358 + not t0, a1
1.1359 + srl t0, t0, 24
1.1360 + bgtz t0, 1f
1.1361 + nop
1.1362 + CONVERT_1x8888_TO_1x0565 a1, t1, t2, t3
1.1363 +0:
1.1364 + sh t1, 0(a0)
1.1365 + addiu a2, a2, -1
1.1366 + bgtz a2, 0b
1.1367 + addiu a0, a0, 2
1.1368 + j ra
1.1369 + nop
1.1370 +
1.1371 +1:
1.1372 + SAVE_REGS_ON_STACK 0, s0, s1, s2
1.1373 + li t4, 0x00ff00ff
1.1374 + li t5, 0xf800f800
1.1375 + li t6, 0x07e007e0
1.1376 + li t7, 0x001F001F
1.1377 + addiu t1, a2, -1
1.1378 + beqz t1, 3f
1.1379 + nop
1.1380 +2:
1.1381 + lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
1.1382 + lhu t2, 2(a0) /* t2 = destination (r5g6b5) */
1.1383 +
1.1384 + CONVERT_2x0565_TO_2x8888 t1, t2, t3, t8, t6, t7, t9, s0, s1, s2
1.1385 + MIPS_2xUN8x4_MUL_2xUN8 t3, t8, t0, t0, t1, t2, t4, t9, s0, s1, s2, t3, t8
1.1386 + addu_s.qb t1, t1, a1
1.1387 + addu_s.qb t2, t2, a1
1.1388 + CONVERT_2x8888_TO_2x0565 t1, t2, t3, t8, t5, t6, t7, s0, s1
1.1389 +
1.1390 + sh t3, 0(a0)
1.1391 + sh t8, 2(a0)
1.1392 +
1.1393 + addiu a2, a2, -2
1.1394 + addiu t1, a2, -1
1.1395 + bgtz t1, 2b
1.1396 + addiu a0, a0, 4
1.1397 +3:
1.1398 + beqz a2, 4f
1.1399 + nop
1.1400 +
1.1401 + lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
1.1402 +
1.1403 + CONVERT_1x0565_TO_1x8888 t1, t2, s0, s1
1.1404 + MIPS_UN8x4_MUL_UN8 t2, t0, t1, t4, s0, s1, s2
1.1405 + addu_s.qb t1, t1, a1
1.1406 + CONVERT_1x8888_TO_1x0565 t1, t2, s0, s1
1.1407 +
1.1408 + sh t2, 0(a0)
1.1409 +
1.1410 +4:
1.1411 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2
1.1412 +5:
1.1413 + j ra
1.1414 + nop
1.1415 +
1.1416 +END(pixman_composite_over_n_0565_asm_mips)
1.1417 +
1.1418 +LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_asm_mips)
1.1419 +/*
1.1420 + * a0 - dst (a8r8g8b8)
1.1421 + * a1 - src (32bit constant)
1.1422 + * a2 - w
1.1423 + */
1.1424 +
1.1425 + beqz a2, 5f
1.1426 + nop
1.1427 +
1.1428 + not t0, a1
1.1429 + srl t0, t0, 24
1.1430 + bgtz t0, 1f
1.1431 + nop
1.1432 +0:
1.1433 + sw a1, 0(a0)
1.1434 + addiu a2, a2, -1
1.1435 + bgtz a2, 0b
1.1436 + addiu a0, a0, 4
1.1437 + j ra
1.1438 + nop
1.1439 +
1.1440 +1:
1.1441 + SAVE_REGS_ON_STACK 0, s0, s1, s2
1.1442 + li t4, 0x00ff00ff
1.1443 + addiu t1, a2, -1
1.1444 + beqz t1, 3f
1.1445 + nop
1.1446 +2:
1.1447 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.1448 + lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
1.1449 +
1.1450 + MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t0, t0, t7, t8, t4, t9, s0, s1, s2, t2, t3
1.1451 +
1.1452 + addu_s.qb t7, t7, a1
1.1453 + addu_s.qb t8, t8, a1
1.1454 +
1.1455 + sw t7, 0(a0)
1.1456 + sw t8, 4(a0)
1.1457 +
1.1458 + addiu a2, a2, -2
1.1459 + addiu t1, a2, -1
1.1460 + bgtz t1, 2b
1.1461 + addiu a0, a0, 8
1.1462 +3:
1.1463 + beqz a2, 4f
1.1464 + nop
1.1465 +
1.1466 + lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
1.1467 +
1.1468 + MIPS_UN8x4_MUL_UN8 t1, t0, t3, t4, t5, t6, t7
1.1469 +
1.1470 + addu_s.qb t3, t3, a1
1.1471 +
1.1472 + sw t3, 0(a0)
1.1473 +
1.1474 +4:
1.1475 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2
1.1476 +5:
1.1477 + j ra
1.1478 + nop
1.1479 +
1.1480 +END(pixman_composite_over_n_8888_asm_mips)
1.1481 +
1.1482 +LEAF_MIPS_DSPR2(pixman_composite_add_8_8_8_asm_mips)
1.1483 +/*
1.1484 + * a0 - dst (a8)
1.1485 + * a1 - src (a8)
1.1486 + * a2 - mask (a8)
1.1487 + * a3 - w
1.1488 + */
1.1489 +
1.1490 + SAVE_REGS_ON_STACK 0, v0, v1
1.1491 + li t9, 0x00ff00ff
1.1492 + beqz a3, 3f
1.1493 + nop
1.1494 +
1.1495 + srl v0, a3, 2 /* v0 = how many multiples of 4 dst pixels */
1.1496 + beqz v0, 1f /* branch if less than 4 src pixels */
1.1497 + nop
1.1498 +
1.1499 +0:
1.1500 + beqz v0, 1f
1.1501 + addiu v0, v0, -1
1.1502 + lbu t0, 0(a2)
1.1503 + lbu t1, 1(a2)
1.1504 + lbu t2, 2(a2)
1.1505 + lbu t3, 3(a2)
1.1506 + lbu t4, 0(a0)
1.1507 + lbu t5, 1(a0)
1.1508 + lbu t6, 2(a0)
1.1509 + lbu t7, 3(a0)
1.1510 +
1.1511 + addiu a2, a2, 4
1.1512 +
1.1513 + precr_sra.ph.w t1, t0, 0
1.1514 + precr_sra.ph.w t3, t2, 0
1.1515 + precr_sra.ph.w t5, t4, 0
1.1516 + precr_sra.ph.w t7, t6, 0
1.1517 +
1.1518 + precr.qb.ph t0, t3, t1
1.1519 + precr.qb.ph t1, t7, t5
1.1520 +
1.1521 + lbu t4, 0(a1)
1.1522 + lbu v1, 1(a1)
1.1523 + lbu t7, 2(a1)
1.1524 + lbu t8, 3(a1)
1.1525 +
1.1526 + addiu a1, a1, 4
1.1527 +
1.1528 + precr_sra.ph.w v1, t4, 0
1.1529 + precr_sra.ph.w t8, t7, 0
1.1530 +
1.1531 + muleu_s.ph.qbl t2, t0, t8
1.1532 + muleu_s.ph.qbr t3, t0, v1
1.1533 + shra_r.ph t4, t2, 8
1.1534 + shra_r.ph t5, t3, 8
1.1535 + and t4, t4, t9
1.1536 + and t5, t5, t9
1.1537 + addq.ph t2, t2, t4
1.1538 + addq.ph t3, t3, t5
1.1539 + shra_r.ph t2, t2, 8
1.1540 + shra_r.ph t3, t3, 8
1.1541 + precr.qb.ph t0, t2, t3
1.1542 +
1.1543 + addu_s.qb t2, t0, t1
1.1544 +
1.1545 + sb t2, 0(a0)
1.1546 + srl t2, t2, 8
1.1547 + sb t2, 1(a0)
1.1548 + srl t2, t2, 8
1.1549 + sb t2, 2(a0)
1.1550 + srl t2, t2, 8
1.1551 + sb t2, 3(a0)
1.1552 + addiu a3, a3, -4
1.1553 + b 0b
1.1554 + addiu a0, a0, 4
1.1555 +
1.1556 +1:
1.1557 + beqz a3, 3f
1.1558 + nop
1.1559 +2:
1.1560 + lbu t8, 0(a1)
1.1561 + lbu t0, 0(a2)
1.1562 + lbu t1, 0(a0)
1.1563 + addiu a1, a1, 1
1.1564 + addiu a2, a2, 1
1.1565 +
1.1566 + mul t2, t0, t8
1.1567 + shra_r.ph t3, t2, 8
1.1568 + andi t3, t3, 0xff
1.1569 + addq.ph t2, t2, t3
1.1570 + shra_r.ph t2, t2, 8
1.1571 + andi t2, t2, 0xff
1.1572 +
1.1573 + addu_s.qb t2, t2, t1
1.1574 + sb t2, 0(a0)
1.1575 + addiu a3, a3, -1
1.1576 + bnez a3, 2b
1.1577 + addiu a0, a0, 1
1.1578 +
1.1579 +3:
1.1580 + RESTORE_REGS_FROM_STACK 0, v0, v1
1.1581 + j ra
1.1582 + nop
1.1583 +
1.1584 +END(pixman_composite_add_8_8_8_asm_mips)
1.1585 +
1.1586 +LEAF_MIPS_DSPR2(pixman_composite_add_n_8_8_asm_mips)
1.1587 +/*
1.1588 + * a0 - dst (a8)
1.1589 + * a1 - src (32bit constant)
1.1590 + * a2 - mask (a8)
1.1591 + * a3 - w
1.1592 + */
1.1593 +
1.1594 + SAVE_REGS_ON_STACK 0, v0
1.1595 + li t9, 0x00ff00ff
1.1596 + beqz a3, 3f
1.1597 + nop
1.1598 +
1.1599 + srl v0, a3, 2 /* v0 = how many multiples of 4 dst pixels */
1.1600 + beqz v0, 1f /* branch if less than 4 src pixels */
1.1601 + nop
1.1602 +
1.1603 + srl t8, a1, 24
1.1604 + replv.ph t8, t8
1.1605 +
1.1606 +0:
1.1607 + beqz v0, 1f
1.1608 + addiu v0, v0, -1
1.1609 + lbu t0, 0(a2)
1.1610 + lbu t1, 1(a2)
1.1611 + lbu t2, 2(a2)
1.1612 + lbu t3, 3(a2)
1.1613 + lbu t4, 0(a0)
1.1614 + lbu t5, 1(a0)
1.1615 + lbu t6, 2(a0)
1.1616 + lbu t7, 3(a0)
1.1617 +
1.1618 + addiu a2, a2, 4
1.1619 +
1.1620 + precr_sra.ph.w t1, t0, 0
1.1621 + precr_sra.ph.w t3, t2, 0
1.1622 + precr_sra.ph.w t5, t4, 0
1.1623 + precr_sra.ph.w t7, t6, 0
1.1624 +
1.1625 + precr.qb.ph t0, t3, t1
1.1626 + precr.qb.ph t1, t7, t5
1.1627 +
1.1628 + muleu_s.ph.qbl t2, t0, t8
1.1629 + muleu_s.ph.qbr t3, t0, t8
1.1630 + shra_r.ph t4, t2, 8
1.1631 + shra_r.ph t5, t3, 8
1.1632 + and t4, t4, t9
1.1633 + and t5, t5, t9
1.1634 + addq.ph t2, t2, t4
1.1635 + addq.ph t3, t3, t5
1.1636 + shra_r.ph t2, t2, 8
1.1637 + shra_r.ph t3, t3, 8
1.1638 + precr.qb.ph t0, t2, t3
1.1639 +
1.1640 + addu_s.qb t2, t0, t1
1.1641 +
1.1642 + sb t2, 0(a0)
1.1643 + srl t2, t2, 8
1.1644 + sb t2, 1(a0)
1.1645 + srl t2, t2, 8
1.1646 + sb t2, 2(a0)
1.1647 + srl t2, t2, 8
1.1648 + sb t2, 3(a0)
1.1649 + addiu a3, a3, -4
1.1650 + b 0b
1.1651 + addiu a0, a0, 4
1.1652 +
1.1653 +1:
1.1654 + beqz a3, 3f
1.1655 + nop
1.1656 + srl t8, a1, 24
1.1657 +2:
1.1658 + lbu t0, 0(a2)
1.1659 + lbu t1, 0(a0)
1.1660 + addiu a2, a2, 1
1.1661 +
1.1662 + mul t2, t0, t8
1.1663 + shra_r.ph t3, t2, 8
1.1664 + andi t3, t3, 0xff
1.1665 + addq.ph t2, t2, t3
1.1666 + shra_r.ph t2, t2, 8
1.1667 + andi t2, t2, 0xff
1.1668 +
1.1669 + addu_s.qb t2, t2, t1
1.1670 + sb t2, 0(a0)
1.1671 + addiu a3, a3, -1
1.1672 + bnez a3, 2b
1.1673 + addiu a0, a0, 1
1.1674 +
1.1675 +3:
1.1676 + RESTORE_REGS_FROM_STACK 0, v0
1.1677 + j ra
1.1678 + nop
1.1679 +
1.1680 +END(pixman_composite_add_n_8_8_asm_mips)
1.1681 +
1.1682 +LEAF_MIPS_DSPR2(pixman_composite_add_n_8_8888_asm_mips)
1.1683 +/*
1.1684 + * a0 - dst (a8r8g8b8)
1.1685 + * a1 - src (32bit constant)
1.1686 + * a2 - mask (a8)
1.1687 + * a3 - w
1.1688 + */
1.1689 +
1.1690 + SAVE_REGS_ON_STACK 0, s0, s1, s2
1.1691 + li t4, 0x00ff00ff
1.1692 + beqz a3, 3f
1.1693 + nop
1.1694 + addiu t1, a3, -1
1.1695 + beqz t1, 2f
1.1696 + nop
1.1697 +1:
1.1698 + /* a1 = source (32bit constant) */
1.1699 + lbu t0, 0(a2) /* t0 = mask (a8) */
1.1700 + lbu t1, 1(a2) /* t1 = mask (a8) */
1.1701 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.1702 + lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
1.1703 + addiu a2, a2, 2
1.1704 +
1.1705 + MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 a1, a1, \
1.1706 + t0, t1, \
1.1707 + t2, t3, \
1.1708 + t5, t6, \
1.1709 + t4, t7, t8, t9, s0, s1, s2
1.1710 +
1.1711 + sw t5, 0(a0)
1.1712 + sw t6, 4(a0)
1.1713 + addiu a3, a3, -2
1.1714 + addiu t1, a3, -1
1.1715 + bgtz t1, 1b
1.1716 + addiu a0, a0, 8
1.1717 +2:
1.1718 + beqz a3, 3f
1.1719 + nop
1.1720 + /* a1 = source (32bit constant) */
1.1721 + lbu t0, 0(a2) /* t0 = mask (a8) */
1.1722 + lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
1.1723 +
1.1724 + MIPS_UN8x4_MUL_UN8_ADD_UN8x4 a1, t0, t1, t2, t4, t3, t5, t6
1.1725 +
1.1726 + sw t2, 0(a0)
1.1727 +3:
1.1728 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2
1.1729 + j ra
1.1730 + nop
1.1731 +
1.1732 +END(pixman_composite_add_n_8_8888_asm_mips)
1.1733 +
1.1734 +LEAF_MIPS_DSPR2(pixman_composite_add_0565_8_0565_asm_mips)
1.1735 +/*
1.1736 + * a0 - dst (r5g6b5)
1.1737 + * a1 - src (r5g6b5)
1.1738 + * a2 - mask (a8)
1.1739 + * a3 - w
1.1740 + */
1.1741 +
1.1742 + SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
1.1743 + li t4, 0xf800f800
1.1744 + li t5, 0x07e007e0
1.1745 + li t6, 0x001F001F
1.1746 + li t7, 0x00ff00ff
1.1747 + beqz a3, 3f
1.1748 + nop
1.1749 + addiu t1, a3, -1
1.1750 + beqz t1, 2f
1.1751 + nop
1.1752 +1:
1.1753 + lhu t0, 0(a1) /* t0 = source (r5g6b5) */
1.1754 + lhu t1, 2(a1) /* t1 = source (r5g6b5) */
1.1755 + lbu t2, 0(a2) /* t2 = mask (a8) */
1.1756 + lbu t3, 1(a2) /* t3 = mask (a8) */
1.1757 + lhu t8, 0(a0) /* t8 = destination (r5g6b5) */
1.1758 + lhu t9, 2(a0) /* t9 = destination (r5g6b5) */
1.1759 + addiu a1, a1, 4
1.1760 + addiu a2, a2, 2
1.1761 +
1.1762 + CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5
1.1763 + CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, s6, s7
1.1764 + MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 s0, s1, \
1.1765 + t2, t3, \
1.1766 + s2, s3, \
1.1767 + t0, t1, \
1.1768 + t7, s4, s5, s6, s7, t8, t9
1.1769 + CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3
1.1770 +
1.1771 + sh s0, 0(a0)
1.1772 + sh s1, 2(a0)
1.1773 + addiu a3, a3, -2
1.1774 + addiu t1, a3, -1
1.1775 + bgtz t1, 1b
1.1776 + addiu a0, a0, 4
1.1777 +2:
1.1778 + beqz a3, 3f
1.1779 + nop
1.1780 + lhu t0, 0(a1) /* t0 = source (r5g6b5) */
1.1781 + lbu t1, 0(a2) /* t1 = mask (a8) */
1.1782 + lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
1.1783 +
1.1784 + CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5
1.1785 + CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6
1.1786 + MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t3, t1, t4, t0, t7, t2, t5, t6
1.1787 + CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5
1.1788 +
1.1789 + sh t3, 0(a0)
1.1790 +3:
1.1791 + RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
1.1792 + j ra
1.1793 + nop
1.1794 +
1.1795 +END(pixman_composite_add_0565_8_0565_asm_mips)
1.1796 +
1.1797 +LEAF_MIPS_DSPR2(pixman_composite_add_8888_8_8888_asm_mips)
1.1798 +/*
1.1799 + * a0 - dst (a8r8g8b8)
1.1800 + * a1 - src (a8r8g8b8)
1.1801 + * a2 - mask (a8)
1.1802 + * a3 - w
1.1803 + */
1.1804 +
1.1805 + SAVE_REGS_ON_STACK 0, s0, s1, s2
1.1806 + li t4, 0x00ff00ff
1.1807 + beqz a3, 3f
1.1808 + nop
1.1809 + addiu t1, a3, -1
1.1810 + beqz t1, 2f
1.1811 + nop
1.1812 +1:
1.1813 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1814 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.1815 + lbu t2, 0(a2) /* t2 = mask (a8) */
1.1816 + lbu t3, 1(a2) /* t3 = mask (a8) */
1.1817 + lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */
1.1818 + lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */
1.1819 + addiu a1, a1, 8
1.1820 + addiu a2, a2, 2
1.1821 +
1.1822 + MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \
1.1823 + t2, t3, \
1.1824 + t5, t6, \
1.1825 + t7, t8, \
1.1826 + t4, t9, s0, s1, s2, t0, t1
1.1827 +
1.1828 + sw t7, 0(a0)
1.1829 + sw t8, 4(a0)
1.1830 + addiu a3, a3, -2
1.1831 + addiu t1, a3, -1
1.1832 + bgtz t1, 1b
1.1833 + addiu a0, a0, 8
1.1834 +2:
1.1835 + beqz a3, 3f
1.1836 + nop
1.1837 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1838 + lbu t1, 0(a2) /* t1 = mask (a8) */
1.1839 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.1840 +
1.1841 + MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t3, t4, t5, t6, t7
1.1842 +
1.1843 + sw t3, 0(a0)
1.1844 +3:
1.1845 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2
1.1846 + j ra
1.1847 + nop
1.1848 +
1.1849 +END(pixman_composite_add_8888_8_8888_asm_mips)
1.1850 +
1.1851 +LEAF_MIPS_DSPR2(pixman_composite_add_8888_n_8888_asm_mips)
1.1852 +/*
1.1853 + * a0 - dst (a8r8g8b8)
1.1854 + * a1 - src (a8r8g8b8)
1.1855 + * a2 - mask (32bit constant)
1.1856 + * a3 - w
1.1857 + */
1.1858 +
1.1859 + SAVE_REGS_ON_STACK 0, s0, s1, s2
1.1860 + li t4, 0x00ff00ff
1.1861 + beqz a3, 3f
1.1862 + nop
1.1863 + srl a2, a2, 24
1.1864 + addiu t1, a3, -1
1.1865 + beqz t1, 2f
1.1866 + nop
1.1867 +1:
1.1868 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1869 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.1870 + /* a2 = mask (32bit constant) */
1.1871 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.1872 + lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
1.1873 + addiu a1, a1, 8
1.1874 +
1.1875 + MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \
1.1876 + a2, a2, \
1.1877 + t2, t3, \
1.1878 + t5, t6, \
1.1879 + t4, t7, t8, t9, s0, s1, s2
1.1880 +
1.1881 + sw t5, 0(a0)
1.1882 + sw t6, 4(a0)
1.1883 + addiu a3, a3, -2
1.1884 + addiu t1, a3, -1
1.1885 + bgtz t1, 1b
1.1886 + addiu a0, a0, 8
1.1887 +2:
1.1888 + beqz a3, 3f
1.1889 + nop
1.1890 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1891 + /* a2 = mask (32bit constant) */
1.1892 + lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
1.1893 +
1.1894 + MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, a2, t1, t3, t4, t5, t6, t7
1.1895 +
1.1896 + sw t3, 0(a0)
1.1897 +3:
1.1898 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2
1.1899 + j ra
1.1900 + nop
1.1901 +
1.1902 +END(pixman_composite_add_8888_n_8888_asm_mips)
1.1903 +
1.1904 +LEAF_MIPS_DSPR2(pixman_composite_add_8888_8888_8888_asm_mips)
1.1905 +/*
1.1906 + * a0 - dst (a8r8g8b8)
1.1907 + * a1 - src (a8r8g8b8)
1.1908 + * a2 - mask (a8r8g8b8)
1.1909 + * a3 - w
1.1910 + */
1.1911 +
1.1912 + SAVE_REGS_ON_STACK 0, s0, s1, s2
1.1913 + li t4, 0x00ff00ff
1.1914 + beqz a3, 3f
1.1915 + nop
1.1916 + addiu t1, a3, -1
1.1917 + beqz t1, 2f
1.1918 + nop
1.1919 +1:
1.1920 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1921 + lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
1.1922 + lw t2, 0(a2) /* t2 = mask (a8r8g8b8) */
1.1923 + lw t3, 4(a2) /* t3 = mask (a8r8g8b8) */
1.1924 + lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */
1.1925 + lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */
1.1926 + addiu a1, a1, 8
1.1927 + addiu a2, a2, 8
1.1928 + srl t2, t2, 24
1.1929 + srl t3, t3, 24
1.1930 +
1.1931 + MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \
1.1932 + t2, t3, \
1.1933 + t5, t6, \
1.1934 + t7, t8, \
1.1935 + t4, t9, s0, s1, s2, t0, t1
1.1936 +
1.1937 + sw t7, 0(a0)
1.1938 + sw t8, 4(a0)
1.1939 + addiu a3, a3, -2
1.1940 + addiu t1, a3, -1
1.1941 + bgtz t1, 1b
1.1942 + addiu a0, a0, 8
1.1943 +2:
1.1944 + beqz a3, 3f
1.1945 + nop
1.1946 + lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
1.1947 + lw t1, 0(a2) /* t1 = mask (a8r8g8b8) */
1.1948 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.1949 + srl t1, t1, 24
1.1950 +
1.1951 + MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t3, t4, t5, t6, t7
1.1952 +
1.1953 + sw t3, 0(a0)
1.1954 +3:
1.1955 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2
1.1956 + j ra
1.1957 + nop
1.1958 +
1.1959 +END(pixman_composite_add_8888_8888_8888_asm_mips)
1.1960 +
1.1961 +LEAF_MIPS_DSPR2(pixman_composite_add_8_8_asm_mips)
1.1962 +/*
1.1963 + * a0 - dst (a8)
1.1964 + * a1 - src (a8)
1.1965 + * a2 - w
1.1966 + */
1.1967 +
1.1968 + beqz a2, 3f
1.1969 + nop
1.1970 + srl t9, a2, 2 /* t9 = how many multiples of 4 dst pixels */
1.1971 + beqz t9, 1f /* branch if less than 4 src pixels */
1.1972 + nop
1.1973 +
1.1974 +0:
1.1975 + beqz t9, 1f
1.1976 + addiu t9, t9, -1
1.1977 + lbu t0, 0(a1)
1.1978 + lbu t1, 1(a1)
1.1979 + lbu t2, 2(a1)
1.1980 + lbu t3, 3(a1)
1.1981 + lbu t4, 0(a0)
1.1982 + lbu t5, 1(a0)
1.1983 + lbu t6, 2(a0)
1.1984 + lbu t7, 3(a0)
1.1985 +
1.1986 + addiu a1, a1, 4
1.1987 +
1.1988 + precr_sra.ph.w t1, t0, 0
1.1989 + precr_sra.ph.w t3, t2, 0
1.1990 + precr_sra.ph.w t5, t4, 0
1.1991 + precr_sra.ph.w t7, t6, 0
1.1992 +
1.1993 + precr.qb.ph t0, t3, t1
1.1994 + precr.qb.ph t1, t7, t5
1.1995 +
1.1996 + addu_s.qb t2, t0, t1
1.1997 +
1.1998 + sb t2, 0(a0)
1.1999 + srl t2, t2, 8
1.2000 + sb t2, 1(a0)
1.2001 + srl t2, t2, 8
1.2002 + sb t2, 2(a0)
1.2003 + srl t2, t2, 8
1.2004 + sb t2, 3(a0)
1.2005 + addiu a2, a2, -4
1.2006 + b 0b
1.2007 + addiu a0, a0, 4
1.2008 +
1.2009 +1:
1.2010 + beqz a2, 3f
1.2011 + nop
1.2012 +2:
1.2013 + lbu t0, 0(a1)
1.2014 + lbu t1, 0(a0)
1.2015 + addiu a1, a1, 1
1.2016 +
1.2017 + addu_s.qb t2, t0, t1
1.2018 + sb t2, 0(a0)
1.2019 + addiu a2, a2, -1
1.2020 + bnez a2, 2b
1.2021 + addiu a0, a0, 1
1.2022 +
1.2023 +3:
1.2024 + j ra
1.2025 + nop
1.2026 +
1.2027 +END(pixman_composite_add_8_8_asm_mips)
1.2028 +
1.2029 +LEAF_MIPS_DSPR2(pixman_composite_add_8888_8888_asm_mips)
1.2030 +/*
1.2031 + * a0 - dst (a8r8g8b8)
1.2032 + * a1 - src (a8r8g8b8)
1.2033 + * a2 - w
1.2034 + */
1.2035 +
1.2036 + beqz a2, 4f
1.2037 + nop
1.2038 +
1.2039 + srl t9, a2, 2 /* t1 = how many multiples of 4 src pixels */
1.2040 + beqz t9, 3f /* branch if less than 4 src pixels */
1.2041 + nop
1.2042 +1:
1.2043 + addiu t9, t9, -1
1.2044 + beqz t9, 2f
1.2045 + addiu a2, a2, -4
1.2046 +
1.2047 + lw t0, 0(a1)
1.2048 + lw t1, 4(a1)
1.2049 + lw t2, 8(a1)
1.2050 + lw t3, 12(a1)
1.2051 + lw t4, 0(a0)
1.2052 + lw t5, 4(a0)
1.2053 + lw t6, 8(a0)
1.2054 + lw t7, 12(a0)
1.2055 + addiu a1, a1, 16
1.2056 +
1.2057 + addu_s.qb t4, t4, t0
1.2058 + addu_s.qb t5, t5, t1
1.2059 + addu_s.qb t6, t6, t2
1.2060 + addu_s.qb t7, t7, t3
1.2061 +
1.2062 + sw t4, 0(a0)
1.2063 + sw t5, 4(a0)
1.2064 + sw t6, 8(a0)
1.2065 + sw t7, 12(a0)
1.2066 + b 1b
1.2067 + addiu a0, a0, 16
1.2068 +2:
1.2069 + lw t0, 0(a1)
1.2070 + lw t1, 4(a1)
1.2071 + lw t2, 8(a1)
1.2072 + lw t3, 12(a1)
1.2073 + lw t4, 0(a0)
1.2074 + lw t5, 4(a0)
1.2075 + lw t6, 8(a0)
1.2076 + lw t7, 12(a0)
1.2077 + addiu a1, a1, 16
1.2078 +
1.2079 + addu_s.qb t4, t4, t0
1.2080 + addu_s.qb t5, t5, t1
1.2081 + addu_s.qb t6, t6, t2
1.2082 + addu_s.qb t7, t7, t3
1.2083 +
1.2084 + sw t4, 0(a0)
1.2085 + sw t5, 4(a0)
1.2086 + sw t6, 8(a0)
1.2087 + sw t7, 12(a0)
1.2088 +
1.2089 + beqz a2, 4f
1.2090 + addiu a0, a0, 16
1.2091 +3:
1.2092 + lw t0, 0(a1)
1.2093 + lw t1, 0(a0)
1.2094 + addiu a1, a1, 4
1.2095 + addiu a2, a2, -1
1.2096 + addu_s.qb t1, t1, t0
1.2097 + sw t1, 0(a0)
1.2098 + bnez a2, 3b
1.2099 + addiu a0, a0, 4
1.2100 +4:
1.2101 + jr ra
1.2102 + nop
1.2103 +
1.2104 +END(pixman_composite_add_8888_8888_asm_mips)
1.2105 +
1.2106 +LEAF_MIPS_DSPR2(pixman_composite_out_reverse_8_0565_asm_mips)
1.2107 +/*
1.2108 + * a0 - dst (r5g6b5)
1.2109 + * a1 - src (a8)
1.2110 + * a2 - w
1.2111 + */
1.2112 +
1.2113 + beqz a2, 4f
1.2114 + nop
1.2115 +
1.2116 + SAVE_REGS_ON_STACK 0, s0, s1, s2, s3
1.2117 + li t2, 0xf800f800
1.2118 + li t3, 0x07e007e0
1.2119 + li t4, 0x001F001F
1.2120 + li t5, 0x00ff00ff
1.2121 +
1.2122 + addiu t1, a2, -1
1.2123 + beqz t1, 2f
1.2124 + nop
1.2125 +1:
1.2126 + lbu t0, 0(a1) /* t0 = source (a8) */
1.2127 + lbu t1, 1(a1) /* t1 = source (a8) */
1.2128 + lhu t6, 0(a0) /* t6 = destination (r5g6b5) */
1.2129 + lhu t7, 2(a0) /* t7 = destination (r5g6b5) */
1.2130 + addiu a1, a1, 2
1.2131 +
1.2132 + not t0, t0
1.2133 + not t1, t1
1.2134 + andi t0, 0xff /* t0 = neg source1 */
1.2135 + andi t1, 0xff /* t1 = neg source2 */
1.2136 + CONVERT_2x0565_TO_2x8888 t6, t7, t8, t9, t3, t4, s0, s1, s2, s3
1.2137 + MIPS_2xUN8x4_MUL_2xUN8 t8, t9, t0, t1, t6, t7, t5, s0, s1, s2, s3, t8, t9
1.2138 + CONVERT_2x8888_TO_2x0565 t6, t7, t8, t9, t2, t3, t4, s0, s1
1.2139 +
1.2140 + sh t8, 0(a0)
1.2141 + sh t9, 2(a0)
1.2142 + addiu a2, a2, -2
1.2143 + addiu t1, a2, -1
1.2144 + bgtz t1, 1b
1.2145 + addiu a0, a0, 4
1.2146 +2:
1.2147 + beqz a2, 3f
1.2148 + nop
1.2149 + lbu t0, 0(a1) /* t0 = source (a8) */
1.2150 + lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
1.2151 +
1.2152 + not t0, t0
1.2153 + andi t0, 0xff /* t0 = neg source */
1.2154 + CONVERT_1x0565_TO_1x8888 t1, t2, t3, t4
1.2155 + MIPS_UN8x4_MUL_UN8 t2, t0, t1, t5, t3, t4, t6
1.2156 + CONVERT_1x8888_TO_1x0565 t1, t2, t3, t4
1.2157 +
1.2158 + sh t2, 0(a0)
1.2159 +3:
1.2160 + RESTORE_REGS_FROM_STACK 0, s0, s1, s2, s3
1.2161 +4:
1.2162 + j ra
1.2163 + nop
1.2164 +
1.2165 +END(pixman_composite_out_reverse_8_0565_asm_mips)
1.2166 +
1.2167 +LEAF_MIPS_DSPR2(pixman_composite_out_reverse_8_8888_asm_mips)
1.2168 +/*
1.2169 + * a0 - dst (a8r8g8b8)
1.2170 + * a1 - src (a8)
1.2171 + * a2 - w
1.2172 + */
1.2173 +
1.2174 + beqz a2, 3f
1.2175 + nop
1.2176 + li t4, 0x00ff00ff
1.2177 + addiu t1, a2, -1
1.2178 + beqz t1, 2f
1.2179 + nop
1.2180 +1:
1.2181 + lbu t0, 0(a1) /* t0 = source (a8) */
1.2182 + lbu t1, 1(a1) /* t1 = source (a8) */
1.2183 + lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
1.2184 + lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
1.2185 + addiu a1, a1, 2
1.2186 + not t0, t0
1.2187 + not t1, t1
1.2188 + andi t0, 0xff /* t0 = neg source */
1.2189 + andi t1, 0xff /* t1 = neg source */
1.2190 +
1.2191 + MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t0, t1, t5, t6, t4, t7, t8, t9, t2, t3, t0
1.2192 +
1.2193 + sw t5, 0(a0)
1.2194 + sw t6, 4(a0)
1.2195 + addiu a2, a2, -2
1.2196 + addiu t1, a2, -1
1.2197 + bgtz t1, 1b
1.2198 + addiu a0, a0, 8
1.2199 +2:
1.2200 + beqz a2, 3f
1.2201 + nop
1.2202 + lbu t0, 0(a1) /* t0 = source (a8) */
1.2203 + lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
1.2204 + not t0, t0
1.2205 + andi t0, 0xff /* t0 = neg source */
1.2206 +
1.2207 + MIPS_UN8x4_MUL_UN8 t1, t0, t2, t4, t3, t5, t6
1.2208 +
1.2209 + sw t2, 0(a0)
1.2210 +3:
1.2211 + j ra
1.2212 + nop
1.2213 +
1.2214 +END(pixman_composite_out_reverse_8_8888_asm_mips)
1.2215 +
1.2216 +LEAF_MIPS_DSPR2(pixman_composite_over_reverse_n_8888_asm_mips)
1.2217 +/*
1.2218 + * a0 - dst (a8r8g8b8)
1.2219 + * a1 - src (32bit constant)
1.2220 + * a2 - w
1.2221 + */
1.2222 +
1.2223 + beqz a2, 5f
1.2224 + nop
1.2225 +
1.2226 + SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
1.2227 + li t0, 0x00ff00ff
1.2228 + srl t9, a2, 2 /* t9 = how many multiples of 4 src pixels */
1.2229 + beqz t9, 2f /* branch if less than 4 src pixels */
1.2230 + nop
1.2231 +1:
1.2232 + beqz t9, 2f
1.2233 + addiu t9, t9, -1
1.2234 +
1.2235 + lw t1, 0(a0)
1.2236 + lw t2, 4(a0)
1.2237 + lw t3, 8(a0)
1.2238 + lw t4, 12(a0)
1.2239 +
1.2240 + addiu a2, a2, -4
1.2241 +
1.2242 + not t5, t1
1.2243 + not t6, t2
1.2244 + not t7, t3
1.2245 + not t8, t4
1.2246 + srl t5, t5, 24
1.2247 + srl t6, t6, 24
1.2248 + srl t7, t7, 24
1.2249 + srl t8, t8, 24
1.2250 + replv.ph t5, t5
1.2251 + replv.ph t6, t6
1.2252 + replv.ph t7, t7
1.2253 + replv.ph t8, t8
1.2254 + muleu_s.ph.qbl s0, a1, t5
1.2255 + muleu_s.ph.qbr s1, a1, t5
1.2256 + muleu_s.ph.qbl s2, a1, t6
1.2257 + muleu_s.ph.qbr s3, a1, t6
1.2258 + muleu_s.ph.qbl s4, a1, t7
1.2259 + muleu_s.ph.qbr s5, a1, t7
1.2260 + muleu_s.ph.qbl s6, a1, t8
1.2261 + muleu_s.ph.qbr s7, a1, t8
1.2262 +
1.2263 + shra_r.ph t5, s0, 8
1.2264 + shra_r.ph t6, s1, 8
1.2265 + shra_r.ph t7, s2, 8
1.2266 + shra_r.ph t8, s3, 8
1.2267 + and t5, t5, t0
1.2268 + and t6, t6, t0
1.2269 + and t7, t7, t0
1.2270 + and t8, t8, t0
1.2271 + addq.ph s0, s0, t5
1.2272 + addq.ph s1, s1, t6
1.2273 + addq.ph s2, s2, t7
1.2274 + addq.ph s3, s3, t8
1.2275 + shra_r.ph s0, s0, 8
1.2276 + shra_r.ph s1, s1, 8
1.2277 + shra_r.ph s2, s2, 8
1.2278 + shra_r.ph s3, s3, 8
1.2279 + shra_r.ph t5, s4, 8
1.2280 + shra_r.ph t6, s5, 8
1.2281 + shra_r.ph t7, s6, 8
1.2282 + shra_r.ph t8, s7, 8
1.2283 + and t5, t5, t0
1.2284 + and t6, t6, t0
1.2285 + and t7, t7, t0
1.2286 + and t8, t8, t0
1.2287 + addq.ph s4, s4, t5
1.2288 + addq.ph s5, s5, t6
1.2289 + addq.ph s6, s6, t7
1.2290 + addq.ph s7, s7, t8
1.2291 + shra_r.ph s4, s4, 8
1.2292 + shra_r.ph s5, s5, 8
1.2293 + shra_r.ph s6, s6, 8
1.2294 + shra_r.ph s7, s7, 8
1.2295 +
1.2296 + precr.qb.ph t5, s0, s1
1.2297 + precr.qb.ph t6, s2, s3
1.2298 + precr.qb.ph t7, s4, s5
1.2299 + precr.qb.ph t8, s6, s7
1.2300 + addu_s.qb t5, t1, t5
1.2301 + addu_s.qb t6, t2, t6
1.2302 + addu_s.qb t7, t3, t7
1.2303 + addu_s.qb t8, t4, t8
1.2304 +
1.2305 + sw t5, 0(a0)
1.2306 + sw t6, 4(a0)
1.2307 + sw t7, 8(a0)
1.2308 + sw t8, 12(a0)
1.2309 + b 1b
1.2310 + addiu a0, a0, 16
1.2311 +
1.2312 +2:
1.2313 + beqz a2, 4f
1.2314 + nop
1.2315 +3:
1.2316 + lw t1, 0(a0)
1.2317 +
1.2318 + not t2, t1
1.2319 + srl t2, t2, 24
1.2320 + replv.ph t2, t2
1.2321 +
1.2322 + muleu_s.ph.qbl t4, a1, t2
1.2323 + muleu_s.ph.qbr t5, a1, t2
1.2324 + shra_r.ph t6, t4, 8
1.2325 + shra_r.ph t7, t5, 8
1.2326 +
1.2327 + and t6,t6,t0
1.2328 + and t7,t7,t0
1.2329 +
1.2330 + addq.ph t8, t4, t6
1.2331 + addq.ph t9, t5, t7
1.2332 +
1.2333 + shra_r.ph t8, t8, 8
1.2334 + shra_r.ph t9, t9, 8
1.2335 +
1.2336 + precr.qb.ph t9, t8, t9
1.2337 +
1.2338 + addu_s.qb t9, t1, t9
1.2339 + sw t9, 0(a0)
1.2340 +
1.2341 + addiu a2, a2, -1
1.2342 + bnez a2, 3b
1.2343 + addiu a0, a0, 4
1.2344 +4:
1.2345 + RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
1.2346 +5:
1.2347 + j ra
1.2348 + nop
1.2349 +
1.2350 +END(pixman_composite_over_reverse_n_8888_asm_mips)
1.2351 +
1.2352 +LEAF_MIPS_DSPR2(pixman_composite_in_n_8_asm_mips)
1.2353 +/*
1.2354 + * a0 - dst (a8)
1.2355 + * a1 - src (a8r8g8b8)
1.2356 + * a2 - w
1.2357 + */
1.2358 +
1.2359 + beqz a2, 5f
1.2360 + nop
1.2361 +
1.2362 + SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
1.2363 + move t7, a1
1.2364 + srl t5, t7, 24
1.2365 + replv.ph t5, t5
1.2366 + srl t9, a2, 2 /* t1 = how many multiples of 4 src pixels */
1.2367 + beqz t9, 2f /* branch if less than 4 src pixels */
1.2368 + nop
1.2369 +
1.2370 +1:
1.2371 + addiu t9, t9, -1
1.2372 + addiu a2, a2, -4
1.2373 + lbu t0, 0(a0)
1.2374 + lbu t1, 1(a0)
1.2375 + lbu t2, 2(a0)
1.2376 + lbu t3, 3(a0)
1.2377 +
1.2378 + muleu_s.ph.qbl s0, t0, t5
1.2379 + muleu_s.ph.qbr s1, t0, t5
1.2380 + muleu_s.ph.qbl s2, t1, t5
1.2381 + muleu_s.ph.qbr s3, t1, t5
1.2382 + muleu_s.ph.qbl s4, t2, t5
1.2383 + muleu_s.ph.qbr s5, t2, t5
1.2384 + muleu_s.ph.qbl s6, t3, t5
1.2385 + muleu_s.ph.qbr s7, t3, t5
1.2386 +
1.2387 + shrl.ph t4, s0, 8
1.2388 + shrl.ph t6, s1, 8
1.2389 + shrl.ph t7, s2, 8
1.2390 + shrl.ph t8, s3, 8
1.2391 + addq.ph t0, s0, t4
1.2392 + addq.ph t1, s1, t6
1.2393 + addq.ph t2, s2, t7
1.2394 + addq.ph t3, s3, t8
1.2395 + shra_r.ph t0, t0, 8
1.2396 + shra_r.ph t1, t1, 8
1.2397 + shra_r.ph t2, t2, 8
1.2398 + shra_r.ph t3, t3, 8
1.2399 + shrl.ph t4, s4, 8
1.2400 + shrl.ph t6, s5, 8
1.2401 + shrl.ph t7, s6, 8
1.2402 + shrl.ph t8, s7, 8
1.2403 + addq.ph s0, s4, t4
1.2404 + addq.ph s1, s5, t6
1.2405 + addq.ph s2, s6, t7
1.2406 + addq.ph s3, s7, t8
1.2407 + shra_r.ph t4, s0, 8
1.2408 + shra_r.ph t6, s1, 8
1.2409 + shra_r.ph t7, s2, 8
1.2410 + shra_r.ph t8, s3, 8
1.2411 +
1.2412 + precr.qb.ph s0, t0, t1
1.2413 + precr.qb.ph s1, t2, t3
1.2414 + precr.qb.ph s2, t4, t6
1.2415 + precr.qb.ph s3, t7, t8
1.2416 +
1.2417 + sb s0, 0(a0)
1.2418 + sb s1, 1(a0)
1.2419 + sb s2, 2(a0)
1.2420 + sb s3, 3(a0)
1.2421 + bgtz t9, 1b
1.2422 + addiu a0, a0, 4
1.2423 +2:
1.2424 + beqz a2, 4f
1.2425 + nop
1.2426 +3:
1.2427 + lbu t1, 0(a0)
1.2428 +
1.2429 + muleu_s.ph.qbl t4, t1, t5
1.2430 + muleu_s.ph.qbr t7, t1, t5
1.2431 + shrl.ph t6, t4, 8
1.2432 + shrl.ph t0, t7, 8
1.2433 + addq.ph t8, t4, t6
1.2434 + addq.ph t9, t7, t0
1.2435 + shra_r.ph t8, t8, 8
1.2436 + shra_r.ph t9, t9, 8
1.2437 + precr.qb.ph t2, t8, t9
1.2438 + sb t2, 0(a0)
1.2439 + addiu a2, a2, -1
1.2440 + bnez a2, 3b
1.2441 + addiu a0, a0, 1
1.2442 +4:
1.2443 + RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
1.2444 +5:
1.2445 + j ra
1.2446 + nop
1.2447 +
1.2448 +END(pixman_composite_in_n_8_asm_mips)
1.2449 +
1.2450 +LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_8888_8_0565_OVER_asm_mips)
1.2451 +/*
1.2452 + * a0 - dst (r5g6b5)
1.2453 + * a1 - src (a8r8g8b8)
1.2454 + * a2 - mask (a8)
1.2455 + * a3 - w
1.2456 + * 16(sp) - vx
1.2457 + * 20(sp) - unit_x
1.2458 + */
1.2459 + beqz a3, 4f
1.2460 + nop
1.2461 +
1.2462 + SAVE_REGS_ON_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5
1.2463 + lw v0, 36(sp) /* v0 = vx */
1.2464 + lw v1, 40(sp) /* v1 = unit_x */
1.2465 + li t6, 0x00ff00ff
1.2466 + li t7, 0xf800f800
1.2467 + li t8, 0x07e007e0
1.2468 + li t9, 0x001F001F
1.2469 +
1.2470 + addiu t1, a3, -1
1.2471 + beqz t1, 2f
1.2472 + nop
1.2473 +1:
1.2474 + sra t0, v0, 16 /* t0 = vx >> 16 */
1.2475 + sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */
1.2476 + addu t0, a1, t0
1.2477 + lw t0, 0(t0) /* t0 = source (a8r8g8b8) */
1.2478 + addu v0, v0, v1 /* v0 = vx + unit_x */
1.2479 + sra t1, v0, 16 /* t1 = vx >> 16 */
1.2480 + sll t1, t1, 2 /* t1 = t1 * 4 (a8r8g8b8) */
1.2481 + addu t1, a1, t1
1.2482 + lw t1, 0(t1) /* t1 = source (a8r8g8b8) */
1.2483 + addu v0, v0, v1 /* v0 = vx + unit_x */
1.2484 + lbu t2, 0(a2) /* t2 = mask (a8) */
1.2485 + lbu t3, 1(a2) /* t3 = mask (a8) */
1.2486 + lhu t4, 0(a0) /* t4 = destination (r5g6b5) */
1.2487 + lhu t5, 2(a0) /* t5 = destination (r5g6b5) */
1.2488 + addiu a2, a2, 2
1.2489 +
1.2490 + CONVERT_2x0565_TO_2x8888 t4, t5, s0, s1, t8, t9, s2, s3, s4, s5
1.2491 + OVER_2x8888_2x8_2x8888 t0, t1, \
1.2492 + t2, t3, \
1.2493 + s0, s1, \
1.2494 + t4, t5, \
1.2495 + t6, s2, s3, s4, s5, t2, t3
1.2496 + CONVERT_2x8888_TO_2x0565 t4, t5, s0, s1, t7, t8, t9, s2, s3
1.2497 +
1.2498 + sh s0, 0(a0)
1.2499 + sh s1, 2(a0)
1.2500 + addiu a3, a3, -2
1.2501 + addiu t1, a3, -1
1.2502 + bgtz t1, 1b
1.2503 + addiu a0, a0, 4
1.2504 +2:
1.2505 + beqz a3, 3f
1.2506 + nop
1.2507 + sra t0, v0, 16 /* t0 = vx >> 16 */
1.2508 + sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */
1.2509 + addu t0, a1, t0
1.2510 + lw t0, 0(t0) /* t0 = source (a8r8g8b8) */
1.2511 + lbu t1, 0(a2) /* t1 = mask (a8) */
1.2512 + lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
1.2513 +
1.2514 + CONVERT_1x0565_TO_1x8888 t2, t3, t4, t5
1.2515 + OVER_8888_8_8888 t0, t1, t3, t2, t6, t4, t5, t7, t8
1.2516 + CONVERT_1x8888_TO_1x0565 t2, t3, t4, t5
1.2517 +
1.2518 + sh t3, 0(a0)
1.2519 +3:
1.2520 + RESTORE_REGS_FROM_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5
1.2521 +4:
1.2522 + j ra
1.2523 + nop
1.2524 +
1.2525 +END(pixman_scaled_nearest_scanline_8888_8_0565_OVER_asm_mips)
1.2526 +
1.2527 +LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_0565_8_0565_OVER_asm_mips)
1.2528 +/*
1.2529 + * a0 - dst (r5g6b5)
1.2530 + * a1 - src (r5g6b5)
1.2531 + * a2 - mask (a8)
1.2532 + * a3 - w
1.2533 + * 16(sp) - vx
1.2534 + * 20(sp) - unit_x
1.2535 + */
1.2536 +
1.2537 + beqz a3, 4f
1.2538 + nop
1.2539 + SAVE_REGS_ON_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5
1.2540 + lw v0, 36(sp) /* v0 = vx */
1.2541 + lw v1, 40(sp) /* v1 = unit_x */
1.2542 + li t4, 0xf800f800
1.2543 + li t5, 0x07e007e0
1.2544 + li t6, 0x001F001F
1.2545 + li t7, 0x00ff00ff
1.2546 +
1.2547 + addiu t1, a3, -1
1.2548 + beqz t1, 2f
1.2549 + nop
1.2550 +1:
1.2551 + sra t0, v0, 16 /* t0 = vx >> 16 */
1.2552 + sll t0, t0, 1 /* t0 = t0 * 2 (r5g6b5) */
1.2553 + addu t0, a1, t0
1.2554 + lhu t0, 0(t0) /* t0 = source (r5g6b5) */
1.2555 + addu v0, v0, v1 /* v0 = vx + unit_x */
1.2556 + sra t1, v0, 16 /* t1 = vx >> 16 */
1.2557 + sll t1, t1, 1 /* t1 = t1 * 2 (r5g6b5) */
1.2558 + addu t1, a1, t1
1.2559 + lhu t1, 0(t1) /* t1 = source (r5g6b5) */
1.2560 + addu v0, v0, v1 /* v0 = vx + unit_x */
1.2561 + lbu t2, 0(a2) /* t2 = mask (a8) */
1.2562 + lbu t3, 1(a2) /* t3 = mask (a8) */
1.2563 + lhu t8, 0(a0) /* t8 = destination (r5g6b5) */
1.2564 + lhu t9, 2(a0) /* t9 = destination (r5g6b5) */
1.2565 + addiu a2, a2, 2
1.2566 +
1.2567 + CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5
1.2568 + CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, t0, t1
1.2569 + OVER_2x8888_2x8_2x8888 s0, s1, \
1.2570 + t2, t3, \
1.2571 + s2, s3, \
1.2572 + t0, t1, \
1.2573 + t7, t8, t9, s4, s5, s0, s1
1.2574 + CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3
1.2575 +
1.2576 + sh s0, 0(a0)
1.2577 + sh s1, 2(a0)
1.2578 + addiu a3, a3, -2
1.2579 + addiu t1, a3, -1
1.2580 + bgtz t1, 1b
1.2581 + addiu a0, a0, 4
1.2582 +2:
1.2583 + beqz a3, 3f
1.2584 + nop
1.2585 + sra t0, v0, 16 /* t0 = vx >> 16 */
1.2586 + sll t0, t0, 1 /* t0 = t0 * 2 (r5g6b5) */
1.2587 + addu t0, a1, t0
1.2588 +
1.2589 + lhu t0, 0(t0) /* t0 = source (r5g6b5) */
1.2590 + lbu t1, 0(a2) /* t1 = mask (a8) */
1.2591 + lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
1.2592 +
1.2593 + CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5
1.2594 + CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6
1.2595 + OVER_8888_8_8888 t3, t1, t4, t0, t7, t2, t5, t6, t8
1.2596 + CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5
1.2597 +
1.2598 + sh t3, 0(a0)
1.2599 +3:
1.2600 + RESTORE_REGS_FROM_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5
1.2601 +4:
1.2602 + j ra
1.2603 + nop
1.2604 +
1.2605 +END(pixman_scaled_nearest_scanline_0565_8_0565_OVER_asm_mips)
1.2606 +
1.2607 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_SRC_asm_mips)
1.2608 +/*
1.2609 + * a0 - *dst
1.2610 + * a1 - *src_top
1.2611 + * a2 - *src_bottom
1.2612 + * a3 - w
1.2613 + * 16(sp) - wt
1.2614 + * 20(sp) - wb
1.2615 + * 24(sp) - vx
1.2616 + * 28(sp) - unit_x
1.2617 + */
1.2618 +
1.2619 + beqz a3, 1f
1.2620 + nop
1.2621 +
1.2622 + SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
1.2623 +
1.2624 + lw s0, 36(sp) /* s0 = wt */
1.2625 + lw s1, 40(sp) /* s1 = wb */
1.2626 + lw s2, 44(sp) /* s2 = vx */
1.2627 + lw s3, 48(sp) /* s3 = unit_x */
1.2628 + li v0, BILINEAR_INTERPOLATION_RANGE
1.2629 +
1.2630 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2631 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2632 +0:
1.2633 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.2634 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.2635 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.2636 +
1.2637 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.2638 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.2639 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.2640 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.2641 +
1.2642 + sra t9, s2, 16
1.2643 + sll t9, t9, 2
1.2644 + addiu t8, t9, 4
1.2645 + lwx t0, t9(a1) /* t0 = tl */
1.2646 + lwx t1, t8(a1) /* t1 = tr */
1.2647 + addiu a3, a3, -1
1.2648 + lwx t2, t9(a2) /* t2 = bl */
1.2649 + lwx t3, t8(a2) /* t3 = br */
1.2650 +
1.2651 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.2652 +
1.2653 + addu s2, s2, s3 /* vx += unit_x; */
1.2654 + sw t0, 0(a0)
1.2655 + bnez a3, 0b
1.2656 + addiu a0, a0, 4
1.2657 +
1.2658 + RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
1.2659 +1:
1.2660 + j ra
1.2661 + nop
1.2662 +
1.2663 +END(pixman_scaled_bilinear_scanline_8888_8888_SRC_asm_mips)
1.2664 +
1.2665 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_0565_SRC_asm_mips)
1.2666 +/*
1.2667 + * a0 - *dst
1.2668 + * a1 - *src_top
1.2669 + * a2 - *src_bottom
1.2670 + * a3 - w
1.2671 + * 16(sp) - wt
1.2672 + * 20(sp) - wb
1.2673 + * 24(sp) - vx
1.2674 + * 28(sp) - unit_x
1.2675 + */
1.2676 +
1.2677 + beqz a3, 1f
1.2678 + nop
1.2679 +
1.2680 + SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
1.2681 +
1.2682 + lw s0, 36(sp) /* s0 = wt */
1.2683 + lw s1, 40(sp) /* s1 = wb */
1.2684 + lw s2, 44(sp) /* s2 = vx */
1.2685 + lw s3, 48(sp) /* s3 = unit_x */
1.2686 + li v0, BILINEAR_INTERPOLATION_RANGE
1.2687 +
1.2688 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2689 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2690 +0:
1.2691 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.2692 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.2693 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.2694 +
1.2695 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.2696 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.2697 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.2698 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.2699 +
1.2700 + sra t9, s2, 16
1.2701 + sll t9, t9, 2
1.2702 + addiu t8, t9, 4
1.2703 + lwx t0, t9(a1) /* t0 = tl */
1.2704 + lwx t1, t8(a1) /* t1 = tr */
1.2705 + addiu a3, a3, -1
1.2706 + lwx t2, t9(a2) /* t2 = bl */
1.2707 + lwx t3, t8(a2) /* t3 = br */
1.2708 +
1.2709 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.2710 + CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
1.2711 +
1.2712 + addu s2, s2, s3 /* vx += unit_x; */
1.2713 + sh t1, 0(a0)
1.2714 + bnez a3, 0b
1.2715 + addiu a0, a0, 2
1.2716 +
1.2717 + RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
1.2718 +1:
1.2719 + j ra
1.2720 + nop
1.2721 +
1.2722 +END(pixman_scaled_bilinear_scanline_8888_0565_SRC_asm_mips)
1.2723 +
1.2724 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8888_SRC_asm_mips)
1.2725 +/*
1.2726 + * a0 - *dst
1.2727 + * a1 - *src_top
1.2728 + * a2 - *src_bottom
1.2729 + * a3 - w
1.2730 + * 16(sp) - wt
1.2731 + * 20(sp) - wb
1.2732 + * 24(sp) - vx
1.2733 + * 28(sp) - unit_x
1.2734 + */
1.2735 +
1.2736 + beqz a3, 1f
1.2737 + nop
1.2738 +
1.2739 + SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.2740 +
1.2741 + lw s0, 44(sp) /* s0 = wt */
1.2742 + lw s1, 48(sp) /* s1 = wb */
1.2743 + lw s2, 52(sp) /* s2 = vx */
1.2744 + lw s3, 56(sp) /* s3 = unit_x */
1.2745 + li v0, BILINEAR_INTERPOLATION_RANGE
1.2746 + li v1, 0x07e007e0
1.2747 + li s8, 0x001f001f
1.2748 +
1.2749 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2750 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2751 +0:
1.2752 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.2753 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.2754 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.2755 +
1.2756 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.2757 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.2758 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.2759 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.2760 +
1.2761 + sra t9, s2, 16
1.2762 + sll t9, t9, 1
1.2763 + addiu t8, t9, 2
1.2764 + lhx t0, t9(a1) /* t0 = tl */
1.2765 + lhx t1, t8(a1) /* t1 = tr */
1.2766 + andi t1, t1, 0xffff
1.2767 + addiu a3, a3, -1
1.2768 + lhx t2, t9(a2) /* t2 = bl */
1.2769 + lhx t3, t8(a2) /* t3 = br */
1.2770 + andi t3, t3, 0xffff
1.2771 +
1.2772 + CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7
1.2773 + CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7
1.2774 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.2775 +
1.2776 + addu s2, s2, s3 /* vx += unit_x; */
1.2777 + sw t0, 0(a0)
1.2778 + bnez a3, 0b
1.2779 + addiu a0, a0, 4
1.2780 +
1.2781 + RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.2782 +1:
1.2783 + j ra
1.2784 + nop
1.2785 +
1.2786 +END(pixman_scaled_bilinear_scanline_0565_8888_SRC_asm_mips)
1.2787 +
1.2788 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_0565_SRC_asm_mips)
1.2789 +/*
1.2790 + * a0 - *dst
1.2791 + * a1 - *src_top
1.2792 + * a2 - *src_bottom
1.2793 + * a3 - w
1.2794 + * 16(sp) - wt
1.2795 + * 20(sp) - wb
1.2796 + * 24(sp) - vx
1.2797 + * 28(sp) - unit_x
1.2798 + */
1.2799 +
1.2800 + beqz a3, 1f
1.2801 + nop
1.2802 +
1.2803 + SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.2804 +
1.2805 + lw s0, 44(sp) /* s0 = wt */
1.2806 + lw s1, 48(sp) /* s1 = wb */
1.2807 + lw s2, 52(sp) /* s2 = vx */
1.2808 + lw s3, 56(sp) /* s3 = unit_x */
1.2809 + li v0, BILINEAR_INTERPOLATION_RANGE
1.2810 + li v1, 0x07e007e0
1.2811 + li s8, 0x001f001f
1.2812 +
1.2813 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2814 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2815 +0:
1.2816 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.2817 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.2818 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.2819 +
1.2820 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.2821 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.2822 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.2823 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.2824 +
1.2825 + sra t9, s2, 16
1.2826 + sll t9, t9, 1
1.2827 + addiu t8, t9, 2
1.2828 + lhx t0, t9(a1) /* t0 = tl */
1.2829 + lhx t1, t8(a1) /* t1 = tr */
1.2830 + andi t1, t1, 0xffff
1.2831 + addiu a3, a3, -1
1.2832 + lhx t2, t9(a2) /* t2 = bl */
1.2833 + lhx t3, t8(a2) /* t3 = br */
1.2834 + andi t3, t3, 0xffff
1.2835 +
1.2836 + CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7
1.2837 + CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7
1.2838 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.2839 + CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
1.2840 +
1.2841 + addu s2, s2, s3 /* vx += unit_x; */
1.2842 + sh t1, 0(a0)
1.2843 + bnez a3, 0b
1.2844 + addiu a0, a0, 2
1.2845 +
1.2846 + RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.2847 +1:
1.2848 + j ra
1.2849 + nop
1.2850 +
1.2851 +END(pixman_scaled_bilinear_scanline_0565_0565_SRC_asm_mips)
1.2852 +
1.2853 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_OVER_asm_mips)
1.2854 +/*
1.2855 + * a0 - *dst
1.2856 + * a1 - *src_top
1.2857 + * a2 - *src_bottom
1.2858 + * a3 - w
1.2859 + * 16(sp) - wt
1.2860 + * 20(sp) - wb
1.2861 + * 24(sp) - vx
1.2862 + * 28(sp) - unit_x
1.2863 + */
1.2864 +
1.2865 + beqz a3, 1f
1.2866 + nop
1.2867 +
1.2868 + SAVE_REGS_ON_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.2869 +
1.2870 + lw s0, 40(sp) /* s0 = wt */
1.2871 + lw s1, 44(sp) /* s1 = wb */
1.2872 + lw s2, 48(sp) /* s2 = vx */
1.2873 + lw s3, 52(sp) /* s3 = unit_x */
1.2874 + li v0, BILINEAR_INTERPOLATION_RANGE
1.2875 + li s8, 0x00ff00ff
1.2876 +
1.2877 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2878 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2879 +0:
1.2880 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.2881 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.2882 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.2883 +
1.2884 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.2885 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.2886 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.2887 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.2888 +
1.2889 + sra t9, s2, 16
1.2890 + sll t9, t9, 2
1.2891 + addiu t8, t9, 4
1.2892 + lwx t0, t9(a1) /* t0 = tl */
1.2893 + lwx t1, t8(a1) /* t1 = tr */
1.2894 + addiu a3, a3, -1
1.2895 + lwx t2, t9(a2) /* t2 = bl */
1.2896 + lwx t3, t8(a2) /* t3 = br */
1.2897 +
1.2898 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.2899 + lw t1, 0(a0) /* t1 = dest */
1.2900 + OVER_8888_8888 t0, t1, t2, s8, t3, t4, t5, t6
1.2901 +
1.2902 + addu s2, s2, s3 /* vx += unit_x; */
1.2903 + sw t2, 0(a0)
1.2904 + bnez a3, 0b
1.2905 + addiu a0, a0, 4
1.2906 +
1.2907 + RESTORE_REGS_FROM_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.2908 +1:
1.2909 + j ra
1.2910 + nop
1.2911 +
1.2912 +END(pixman_scaled_bilinear_scanline_8888_8888_OVER_asm_mips)
1.2913 +
1.2914 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_ADD_asm_mips)
1.2915 +/*
1.2916 + * a0 - *dst
1.2917 + * a1 - *src_top
1.2918 + * a2 - *src_bottom
1.2919 + * a3 - w
1.2920 + * 16(sp) - wt
1.2921 + * 20(sp) - wb
1.2922 + * 24(sp) - vx
1.2923 + * 28(sp) - unit_x
1.2924 + */
1.2925 +
1.2926 + beqz a3, 1f
1.2927 + nop
1.2928 +
1.2929 + SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
1.2930 +
1.2931 + lw s0, 36(sp) /* s0 = wt */
1.2932 + lw s1, 40(sp) /* s1 = wb */
1.2933 + lw s2, 44(sp) /* s2 = vx */
1.2934 + lw s3, 48(sp) /* s3 = unit_x */
1.2935 + li v0, BILINEAR_INTERPOLATION_RANGE
1.2936 +
1.2937 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2938 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.2939 +0:
1.2940 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.2941 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.2942 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.2943 +
1.2944 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.2945 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.2946 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.2947 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.2948 +
1.2949 + sra t9, s2, 16
1.2950 + sll t9, t9, 2
1.2951 + addiu t8, t9, 4
1.2952 + lwx t0, t9(a1) /* t0 = tl */
1.2953 + lwx t1, t8(a1) /* t1 = tr */
1.2954 + addiu a3, a3, -1
1.2955 + lwx t2, t9(a2) /* t2 = bl */
1.2956 + lwx t3, t8(a2) /* t3 = br */
1.2957 +
1.2958 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.2959 + lw t1, 0(a0)
1.2960 + addu_s.qb t2, t0, t1
1.2961 +
1.2962 + addu s2, s2, s3 /* vx += unit_x; */
1.2963 + sw t2, 0(a0)
1.2964 + bnez a3, 0b
1.2965 + addiu a0, a0, 4
1.2966 +
1.2967 + RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
1.2968 +1:
1.2969 + j ra
1.2970 + nop
1.2971 +
1.2972 +END(pixman_scaled_bilinear_scanline_8888_8888_ADD_asm_mips)
1.2973 +
1.2974 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_SRC_asm_mips)
1.2975 +/*
1.2976 + * a0 - *dst
1.2977 + * a1 - *mask
1.2978 + * a2 - *src_top
1.2979 + * a3 - *src_bottom
1.2980 + * 16(sp) - wt
1.2981 + * 20(sp) - wb
1.2982 + * 24(sp) - vx
1.2983 + * 28(sp) - unit_x
1.2984 + * 32(sp) - w
1.2985 + */
1.2986 +
1.2987 + lw v1, 32(sp)
1.2988 + beqz v1, 1f
1.2989 + nop
1.2990 +
1.2991 + SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.2992 +
1.2993 + lw s0, 44(sp) /* s0 = wt */
1.2994 + lw s1, 48(sp) /* s1 = wb */
1.2995 + lw s2, 52(sp) /* s2 = vx */
1.2996 + lw s3, 56(sp) /* s3 = unit_x */
1.2997 + li v0, BILINEAR_INTERPOLATION_RANGE
1.2998 + li s8, 0x00ff00ff
1.2999 +
1.3000 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3001 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3002 +0:
1.3003 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.3004 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.3005 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.3006 +
1.3007 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.3008 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.3009 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.3010 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.3011 +
1.3012 + sra t9, s2, 16
1.3013 + sll t9, t9, 2
1.3014 + addiu t8, t9, 4
1.3015 + lwx t0, t9(a2) /* t0 = tl */
1.3016 + lwx t1, t8(a2) /* t1 = tr */
1.3017 + addiu v1, v1, -1
1.3018 + lwx t2, t9(a3) /* t2 = bl */
1.3019 + lwx t3, t8(a3) /* t3 = br */
1.3020 +
1.3021 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.3022 + lbu t1, 0(a1) /* t1 = mask */
1.3023 + addiu a1, a1, 1
1.3024 + MIPS_UN8x4_MUL_UN8 t0, t1, t0, s8, t2, t3, t4
1.3025 +
1.3026 + addu s2, s2, s3 /* vx += unit_x; */
1.3027 + sw t0, 0(a0)
1.3028 + bnez v1, 0b
1.3029 + addiu a0, a0, 4
1.3030 +
1.3031 + RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.3032 +1:
1.3033 + j ra
1.3034 + nop
1.3035 +
1.3036 +END(pixman_scaled_bilinear_scanline_8888_8_8888_SRC_asm_mips)
1.3037 +
1.3038 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_0565_SRC_asm_mips)
1.3039 +/*
1.3040 + * a0 - *dst
1.3041 + * a1 - *mask
1.3042 + * a2 - *src_top
1.3043 + * a3 - *src_bottom
1.3044 + * 16(sp) - wt
1.3045 + * 20(sp) - wb
1.3046 + * 24(sp) - vx
1.3047 + * 28(sp) - unit_x
1.3048 + * 32(sp) - w
1.3049 + */
1.3050 +
1.3051 + lw v1, 32(sp)
1.3052 + beqz v1, 1f
1.3053 + nop
1.3054 +
1.3055 + SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.3056 +
1.3057 + lw s0, 44(sp) /* s0 = wt */
1.3058 + lw s1, 48(sp) /* s1 = wb */
1.3059 + lw s2, 52(sp) /* s2 = vx */
1.3060 + lw s3, 56(sp) /* s3 = unit_x */
1.3061 + li v0, BILINEAR_INTERPOLATION_RANGE
1.3062 + li s8, 0x00ff00ff
1.3063 +
1.3064 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3065 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3066 +0:
1.3067 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.3068 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.3069 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.3070 +
1.3071 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.3072 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.3073 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.3074 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.3075 +
1.3076 + sra t9, s2, 16
1.3077 + sll t9, t9, 2
1.3078 + addiu t8, t9, 4
1.3079 + lwx t0, t9(a2) /* t0 = tl */
1.3080 + lwx t1, t8(a2) /* t1 = tr */
1.3081 + addiu v1, v1, -1
1.3082 + lwx t2, t9(a3) /* t2 = bl */
1.3083 + lwx t3, t8(a3) /* t3 = br */
1.3084 +
1.3085 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.3086 + lbu t1, 0(a1) /* t1 = mask */
1.3087 + addiu a1, a1, 1
1.3088 + MIPS_UN8x4_MUL_UN8 t0, t1, t0, s8, t2, t3, t4
1.3089 + CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
1.3090 +
1.3091 + addu s2, s2, s3 /* vx += unit_x; */
1.3092 + sh t1, 0(a0)
1.3093 + bnez v1, 0b
1.3094 + addiu a0, a0, 2
1.3095 +
1.3096 + RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.3097 +1:
1.3098 + j ra
1.3099 + nop
1.3100 +
1.3101 +END(pixman_scaled_bilinear_scanline_8888_8_0565_SRC_asm_mips)
1.3102 +
1.3103 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8_x888_SRC_asm_mips)
1.3104 +/*
1.3105 + * a0 - *dst
1.3106 + * a1 - *mask
1.3107 + * a2 - *src_top
1.3108 + * a3 - *src_bottom
1.3109 + * 16(sp) - wt
1.3110 + * 20(sp) - wb
1.3111 + * 24(sp) - vx
1.3112 + * 28(sp) - unit_x
1.3113 + * 32(sp) - w
1.3114 + */
1.3115 +
1.3116 + lw t0, 32(sp)
1.3117 + beqz t0, 1f
1.3118 + nop
1.3119 +
1.3120 + SAVE_REGS_ON_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra
1.3121 +
1.3122 + lw s0, 48(sp) /* s0 = wt */
1.3123 + lw s1, 52(sp) /* s1 = wb */
1.3124 + lw s2, 56(sp) /* s2 = vx */
1.3125 + lw s3, 60(sp) /* s3 = unit_x */
1.3126 + lw ra, 64(sp) /* ra = w */
1.3127 + li v0, 0x00ff00ff
1.3128 + li v1, 0x07e007e0
1.3129 + li s8, 0x001f001f
1.3130 +
1.3131 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3132 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3133 +0:
1.3134 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.3135 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.3136 + li t5, BILINEAR_INTERPOLATION_RANGE
1.3137 + subu t5, t5, t4 /* t5 = ( 256 - (vx>>8)) */
1.3138 +
1.3139 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.3140 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.3141 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.3142 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.3143 +
1.3144 + sra t9, s2, 16
1.3145 + sll t9, t9, 1
1.3146 + addiu t8, t9, 2
1.3147 + lhx t0, t9(a2) /* t0 = tl */
1.3148 + lhx t1, t8(a2) /* t1 = tr */
1.3149 + andi t1, t1, 0xffff
1.3150 + addiu ra, ra, -1
1.3151 + lhx t2, t9(a3) /* t2 = bl */
1.3152 + lhx t3, t8(a3) /* t3 = br */
1.3153 + andi t3, t3, 0xffff
1.3154 +
1.3155 + CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7
1.3156 + CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7
1.3157 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.3158 + lbu t1, 0(a1) /* t1 = mask */
1.3159 + addiu a1, a1, 1
1.3160 + MIPS_UN8x4_MUL_UN8 t0, t1, t0, v0, t2, t3, t4
1.3161 +
1.3162 + addu s2, s2, s3 /* vx += unit_x; */
1.3163 + sw t0, 0(a0)
1.3164 + bnez ra, 0b
1.3165 + addiu a0, a0, 4
1.3166 +
1.3167 + RESTORE_REGS_FROM_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra
1.3168 +1:
1.3169 + j ra
1.3170 + nop
1.3171 +
1.3172 +END(pixman_scaled_bilinear_scanline_0565_8_x888_SRC_asm_mips)
1.3173 +
1.3174 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8_0565_SRC_asm_mips)
1.3175 +/*
1.3176 + * a0 - *dst
1.3177 + * a1 - *mask
1.3178 + * a2 - *src_top
1.3179 + * a3 - *src_bottom
1.3180 + * 16(sp) - wt
1.3181 + * 20(sp) - wb
1.3182 + * 24(sp) - vx
1.3183 + * 28(sp) - unit_x
1.3184 + * 32(sp) - w
1.3185 + */
1.3186 +
1.3187 + lw t0, 32(sp)
1.3188 + beqz t0, 1f
1.3189 + nop
1.3190 +
1.3191 + SAVE_REGS_ON_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra
1.3192 +
1.3193 + lw s0, 48(sp) /* s0 = wt */
1.3194 + lw s1, 52(sp) /* s1 = wb */
1.3195 + lw s2, 56(sp) /* s2 = vx */
1.3196 + lw s3, 60(sp) /* s3 = unit_x */
1.3197 + lw ra, 64(sp) /* ra = w */
1.3198 + li v0, 0x00ff00ff
1.3199 + li v1, 0x07e007e0
1.3200 + li s8, 0x001f001f
1.3201 +
1.3202 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3203 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3204 +0:
1.3205 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.3206 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.3207 + li t5, BILINEAR_INTERPOLATION_RANGE
1.3208 + subu t5, t5, t4 /* t5 = ( 256 - (vx>>8)) */
1.3209 +
1.3210 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.3211 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.3212 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.3213 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.3214 +
1.3215 + sra t9, s2, 16
1.3216 + sll t9, t9, 1
1.3217 + addiu t8, t9, 2
1.3218 + lhx t0, t9(a2) /* t0 = tl */
1.3219 + lhx t1, t8(a2) /* t1 = tr */
1.3220 + andi t1, t1, 0xffff
1.3221 + addiu ra, ra, -1
1.3222 + lhx t2, t9(a3) /* t2 = bl */
1.3223 + lhx t3, t8(a3) /* t3 = br */
1.3224 + andi t3, t3, 0xffff
1.3225 +
1.3226 + CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7
1.3227 + CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7
1.3228 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.3229 + lbu t1, 0(a1) /* t1 = mask */
1.3230 + addiu a1, a1, 1
1.3231 + MIPS_UN8x4_MUL_UN8 t0, t1, t0, v0, t2, t3, t4
1.3232 + CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
1.3233 +
1.3234 + addu s2, s2, s3 /* vx += unit_x; */
1.3235 + sh t1, 0(a0)
1.3236 + bnez ra, 0b
1.3237 + addiu a0, a0, 2
1.3238 +
1.3239 + RESTORE_REGS_FROM_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra
1.3240 +1:
1.3241 + j ra
1.3242 + nop
1.3243 +
1.3244 +END(pixman_scaled_bilinear_scanline_0565_8_0565_SRC_asm_mips)
1.3245 +
1.3246 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_OVER_asm_mips)
1.3247 +/*
1.3248 + * a0 - dst (a8r8g8b8)
1.3249 + * a1 - mask (a8)
1.3250 + * a2 - src_top (a8r8g8b8)
1.3251 + * a3 - src_bottom (a8r8g8b8)
1.3252 + * 16(sp) - wt
1.3253 + * 20(sp) - wb
1.3254 + * 24(sp) - vx
1.3255 + * 28(sp) - unit_x
1.3256 + * 32(sp) - w
1.3257 + */
1.3258 +
1.3259 + SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.3260 +
1.3261 + lw v1, 60(sp) /* v1 = w(sp + 32 + 28 save regs stack offset)*/
1.3262 + beqz v1, 1f
1.3263 + nop
1.3264 +
1.3265 + lw s0, 44(sp) /* s0 = wt */
1.3266 + lw s1, 48(sp) /* s1 = wb */
1.3267 + lw s2, 52(sp) /* s2 = vx */
1.3268 + lw s3, 56(sp) /* s3 = unit_x */
1.3269 + li v0, BILINEAR_INTERPOLATION_RANGE
1.3270 + li s8, 0x00ff00ff
1.3271 +
1.3272 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3273 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3274 +
1.3275 +0:
1.3276 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.3277 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.3278 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.3279 +
1.3280 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.3281 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.3282 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.3283 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.3284 +
1.3285 + sra t9, s2, 16
1.3286 + sll t9, t9, 2
1.3287 + addiu t8, t9, 4
1.3288 + lwx t0, t9(a2) /* t0 = tl */
1.3289 + lwx t1, t8(a2) /* t1 = tr */
1.3290 + addiu v1, v1, -1
1.3291 + lwx t2, t9(a3) /* t2 = bl */
1.3292 + lwx t3, t8(a3) /* t3 = br */
1.3293 +
1.3294 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, \
1.3295 + t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.3296 + lbu t1, 0(a1) /* t1 = mask */
1.3297 + lw t2, 0(a0) /* t2 = dst */
1.3298 + addiu a1, a1, 1
1.3299 + OVER_8888_8_8888 t0, t1, t2, t0, s8, t3, t4, t5, t6
1.3300 +
1.3301 + addu s2, s2, s3 /* vx += unit_x; */
1.3302 + sw t0, 0(a0)
1.3303 + bnez v1, 0b
1.3304 + addiu a0, a0, 4
1.3305 +
1.3306 +1:
1.3307 + RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.3308 + j ra
1.3309 + nop
1.3310 +
1.3311 +END(pixman_scaled_bilinear_scanline_8888_8_8888_OVER_asm_mips)
1.3312 +
1.3313 +LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_ADD_asm_mips)
1.3314 +/*
1.3315 + * a0 - *dst
1.3316 + * a1 - *mask
1.3317 + * a2 - *src_top
1.3318 + * a3 - *src_bottom
1.3319 + * 16(sp) - wt
1.3320 + * 20(sp) - wb
1.3321 + * 24(sp) - vx
1.3322 + * 28(sp) - unit_x
1.3323 + * 32(sp) - w
1.3324 + */
1.3325 +
1.3326 + lw v1, 32(sp)
1.3327 + beqz v1, 1f
1.3328 + nop
1.3329 +
1.3330 + SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.3331 +
1.3332 + lw s0, 44(sp) /* s0 = wt */
1.3333 + lw s1, 48(sp) /* s1 = wb */
1.3334 + lw s2, 52(sp) /* s2 = vx */
1.3335 + lw s3, 56(sp) /* s3 = unit_x */
1.3336 + li v0, BILINEAR_INTERPOLATION_RANGE
1.3337 + li s8, 0x00ff00ff
1.3338 +
1.3339 + sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3340 + sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
1.3341 +0:
1.3342 + andi t4, s2, 0xffff /* t4 = (short)vx */
1.3343 + srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
1.3344 + subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
1.3345 +
1.3346 + mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
1.3347 + mul s5, s0, t4 /* s5 = wt*(vx>>8) */
1.3348 + mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
1.3349 + mul s7, s1, t4 /* s7 = wb*(vx>>8) */
1.3350 +
1.3351 + sra t9, s2, 16
1.3352 + sll t9, t9, 2
1.3353 + addiu t8, t9, 4
1.3354 + lwx t0, t9(a2) /* t0 = tl */
1.3355 + lwx t1, t8(a2) /* t1 = tr */
1.3356 + addiu v1, v1, -1
1.3357 + lwx t2, t9(a3) /* t2 = bl */
1.3358 + lwx t3, t8(a3) /* t3 = br */
1.3359 +
1.3360 + BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
1.3361 + lbu t1, 0(a1) /* t1 = mask */
1.3362 + lw t2, 0(a0) /* t2 = dst */
1.3363 + addiu a1, a1, 1
1.3364 + MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t0, s8, t3, t4, t5
1.3365 +
1.3366 + addu s2, s2, s3 /* vx += unit_x; */
1.3367 + sw t0, 0(a0)
1.3368 + bnez v1, 0b
1.3369 + addiu a0, a0, 4
1.3370 +
1.3371 + RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
1.3372 +1:
1.3373 + j ra
1.3374 + nop
1.3375 +
1.3376 +END(pixman_scaled_bilinear_scanline_8888_8_8888_ADD_asm_mips)
