• file
• revisions
• annotate
• diff
• comparison
• raw

gfx/cairo/libpixman/src/pixman-mips-dspr2-asm.S@6474c204b198 (annotated)

gfx/cairo/libpixman/src/pixman-mips-dspr2-asm.S

Wed, 31 Dec 2014 06:09:35 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Wed, 31 Dec 2014 06:09:35 +0100

changeset 0

6474c204b198

permissions

-rw-r--r--

Cloned upstream origin tor-browser at tor-browser-31.3.0esr-4.5-1-build1
revision ID fc1c9ff7c1b2defdbc039f12214767608f46423f for hacking purpose.

michael@0	1	/*
michael@0	2	* Copyright (c) 2012
michael@0	3	* MIPS Technologies, Inc., California.
michael@0	4	*
michael@0	5	* Redistribution and use in source and binary forms, with or without
michael@0	6	* modification, are permitted provided that the following conditions
michael@0	7	* are met:
michael@0	8	* 1. Redistributions of source code must retain the above copyright
michael@0	9	* notice, this list of conditions and the following disclaimer.
michael@0	10	* 2. Redistributions in binary form must reproduce the above copyright
michael@0	11	* notice, this list of conditions and the following disclaimer in the
michael@0	12	* documentation and/or other materials provided with the distribution.
michael@0	13	* 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
michael@0	14	* contributors may be used to endorse or promote products derived from
michael@0	15	* this software without specific prior written permission.
michael@0	16	*
michael@0	17	* THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
michael@0	18	* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
michael@0	19	* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
michael@0	20	* ARE DISCLAIMED. IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
michael@0	21	* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
michael@0	22	* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
michael@0	23	* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
michael@0	24	* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
michael@0	25	* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
michael@0	26	* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
michael@0	27	* SUCH DAMAGE.
michael@0	28	*
michael@0	29	* Author: Nemanja Lukic (nlukic@mips.com)
michael@0	30	*/
michael@0	31
michael@0	32	#include "pixman-private.h"
michael@0	33	#include "pixman-mips-dspr2-asm.h"
michael@0	34
michael@0	35	LEAF_MIPS_DSPR2(pixman_fill_buff16_mips)
michael@0	36	/*
michael@0	37	* a0 - *dest
michael@0	38	* a1 - count (bytes)
michael@0	39	* a2 - value to fill buffer with
michael@0	40	*/
michael@0	41
michael@0	42	beqz a1, 3f
michael@0	43	andi t1, a0, 0x0002
michael@0	44	beqz t1, 0f /* check if address is 4-byte aligned */
michael@0	45	nop
michael@0	46	sh a2, 0(a0)
michael@0	47	addiu a0, a0, 2
michael@0	48	addiu a1, a1, -2
michael@0	49	0:
michael@0	50	srl t1, a1, 5 /* t1 how many multiples of 32 bytes */
michael@0	51	replv.ph a2, a2 /* replicate fill value (16bit) in a2 */
michael@0	52	beqz t1, 2f
michael@0	53	nop
michael@0	54	1:
michael@0	55	addiu t1, t1, -1
michael@0	56	beqz t1, 11f
michael@0	57	addiu a1, a1, -32
michael@0	58	pref 30, 32(a0)
michael@0	59	sw a2, 0(a0)
michael@0	60	sw a2, 4(a0)
michael@0	61	sw a2, 8(a0)
michael@0	62	sw a2, 12(a0)
michael@0	63	sw a2, 16(a0)
michael@0	64	sw a2, 20(a0)
michael@0	65	sw a2, 24(a0)
michael@0	66	sw a2, 28(a0)
michael@0	67	b 1b
michael@0	68	addiu a0, a0, 32
michael@0	69	11:
michael@0	70	sw a2, 0(a0)
michael@0	71	sw a2, 4(a0)
michael@0	72	sw a2, 8(a0)
michael@0	73	sw a2, 12(a0)
michael@0	74	sw a2, 16(a0)
michael@0	75	sw a2, 20(a0)
michael@0	76	sw a2, 24(a0)
michael@0	77	sw a2, 28(a0)
michael@0	78	addiu a0, a0, 32
michael@0	79	2:
michael@0	80	blez a1, 3f
michael@0	81	addiu a1, a1, -2
michael@0	82	sh a2, 0(a0)
michael@0	83	b 2b
michael@0	84	addiu a0, a0, 2
michael@0	85	3:
michael@0	86	jr ra
michael@0	87	nop
michael@0	88
michael@0	89	END(pixman_fill_buff16_mips)
michael@0	90
michael@0	91	LEAF_MIPS32R2(pixman_fill_buff32_mips)
michael@0	92	/*
michael@0	93	* a0 - *dest
michael@0	94	* a1 - count (bytes)
michael@0	95	* a2 - value to fill buffer with
michael@0	96	*/
michael@0	97
michael@0	98	beqz a1, 3f
michael@0	99	nop
michael@0	100	srl t1, a1, 5 /* t1 how many multiples of 32 bytes */
michael@0	101	beqz t1, 2f
michael@0	102	nop
michael@0	103	1:
michael@0	104	addiu t1, t1, -1
michael@0	105	beqz t1, 11f
michael@0	106	addiu a1, a1, -32
michael@0	107	pref 30, 32(a0)
michael@0	108	sw a2, 0(a0)
michael@0	109	sw a2, 4(a0)
michael@0	110	sw a2, 8(a0)
michael@0	111	sw a2, 12(a0)
michael@0	112	sw a2, 16(a0)
michael@0	113	sw a2, 20(a0)
michael@0	114	sw a2, 24(a0)
michael@0	115	sw a2, 28(a0)
michael@0	116	b 1b
michael@0	117	addiu a0, a0, 32
michael@0	118	11:
michael@0	119	sw a2, 0(a0)
michael@0	120	sw a2, 4(a0)
michael@0	121	sw a2, 8(a0)
michael@0	122	sw a2, 12(a0)
michael@0	123	sw a2, 16(a0)
michael@0	124	sw a2, 20(a0)
michael@0	125	sw a2, 24(a0)
michael@0	126	sw a2, 28(a0)
michael@0	127	addiu a0, a0, 32
michael@0	128	2:
michael@0	129	blez a1, 3f
michael@0	130	addiu a1, a1, -4
michael@0	131	sw a2, 0(a0)
michael@0	132	b 2b
michael@0	133	addiu a0, a0, 4
michael@0	134	3:
michael@0	135	jr ra
michael@0	136	nop
michael@0	137
michael@0	138	END(pixman_fill_buff32_mips)
michael@0	139
michael@0	140	LEAF_MIPS_DSPR2(pixman_composite_src_8888_0565_asm_mips)
michael@0	141	/*
michael@0	142	* a0 - dst (r5g6b5)
michael@0	143	* a1 - src (a8r8g8b8)
michael@0	144	* a2 - w
michael@0	145	*/
michael@0	146
michael@0	147	beqz a2, 3f
michael@0	148	nop
michael@0	149	addiu t1, a2, -1
michael@0	150	beqz t1, 2f
michael@0	151	nop
michael@0	152	li t4, 0xf800f800
michael@0	153	li t5, 0x07e007e0
michael@0	154	li t6, 0x001f001f
michael@0	155	1:
michael@0	156	lw t0, 0(a1)
michael@0	157	lw t1, 4(a1)
michael@0	158	addiu a1, a1, 8
michael@0	159	addiu a2, a2, -2
michael@0	160
michael@0	161	CONVERT_2x8888_TO_2x0565 t0, t1, t2, t3, t4, t5, t6, t7, t8
michael@0	162
michael@0	163	sh t2, 0(a0)
michael@0	164	sh t3, 2(a0)
michael@0	165
michael@0	166	addiu t2, a2, -1
michael@0	167	bgtz t2, 1b
michael@0	168	addiu a0, a0, 4
michael@0	169	2:
michael@0	170	beqz a2, 3f
michael@0	171	nop
michael@0	172	lw t0, 0(a1)
michael@0	173
michael@0	174	CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
michael@0	175
michael@0	176	sh t1, 0(a0)
michael@0	177	3:
michael@0	178	j ra
michael@0	179	nop
michael@0	180
michael@0	181	END(pixman_composite_src_8888_0565_asm_mips)
michael@0	182
michael@0	183	LEAF_MIPS_DSPR2(pixman_composite_src_0565_8888_asm_mips)
michael@0	184	/*
michael@0	185	* a0 - dst (a8r8g8b8)
michael@0	186	* a1 - src (r5g6b5)
michael@0	187	* a2 - w
michael@0	188	*/
michael@0	189
michael@0	190	beqz a2, 3f
michael@0	191	nop
michael@0	192	addiu t1, a2, -1
michael@0	193	beqz t1, 2f
michael@0	194	nop
michael@0	195	li t4, 0x07e007e0
michael@0	196	li t5, 0x001F001F
michael@0	197	1:
michael@0	198	lhu t0, 0(a1)
michael@0	199	lhu t1, 2(a1)
michael@0	200	addiu a1, a1, 4
michael@0	201	addiu a2, a2, -2
michael@0	202
michael@0	203	CONVERT_2x0565_TO_2x8888 t0, t1, t2, t3, t4, t5, t6, t7, t8, t9
michael@0	204
michael@0	205	sw t2, 0(a0)
michael@0	206	sw t3, 4(a0)
michael@0	207
michael@0	208	addiu t2, a2, -1
michael@0	209	bgtz t2, 1b
michael@0	210	addiu a0, a0, 8
michael@0	211	2:
michael@0	212	beqz a2, 3f
michael@0	213	nop
michael@0	214	lhu t0, 0(a1)
michael@0	215
michael@0	216	CONVERT_1x0565_TO_1x8888 t0, t1, t2, t3
michael@0	217
michael@0	218	sw t1, 0(a0)
michael@0	219	3:
michael@0	220	j ra
michael@0	221	nop
michael@0	222
michael@0	223	END(pixman_composite_src_0565_8888_asm_mips)
michael@0	224
michael@0	225	LEAF_MIPS_DSPR2(pixman_composite_src_x888_8888_asm_mips)
michael@0	226	/*
michael@0	227	* a0 - dst (a8r8g8b8)
michael@0	228	* a1 - src (x8r8g8b8)
michael@0	229	* a2 - w
michael@0	230	*/
michael@0	231
michael@0	232	beqz a2, 4f
michael@0	233	nop
michael@0	234	li t9, 0xff000000
michael@0	235	srl t8, a2, 3 /* t1 = how many multiples of 8 src pixels */
michael@0	236	beqz t8, 3f /* branch if less than 8 src pixels */
michael@0	237	nop
michael@0	238	1:
michael@0	239	addiu t8, t8, -1
michael@0	240	beqz t8, 2f
michael@0	241	addiu a2, a2, -8
michael@0	242	pref 0, 32(a1)
michael@0	243	lw t0, 0(a1)
michael@0	244	lw t1, 4(a1)
michael@0	245	lw t2, 8(a1)
michael@0	246	lw t3, 12(a1)
michael@0	247	lw t4, 16(a1)
michael@0	248	lw t5, 20(a1)
michael@0	249	lw t6, 24(a1)
michael@0	250	lw t7, 28(a1)
michael@0	251	addiu a1, a1, 32
michael@0	252	or t0, t0, t9
michael@0	253	or t1, t1, t9
michael@0	254	or t2, t2, t9
michael@0	255	or t3, t3, t9
michael@0	256	or t4, t4, t9
michael@0	257	or t5, t5, t9
michael@0	258	or t6, t6, t9
michael@0	259	or t7, t7, t9
michael@0	260	pref 30, 32(a0)
michael@0	261	sw t0, 0(a0)
michael@0	262	sw t1, 4(a0)
michael@0	263	sw t2, 8(a0)
michael@0	264	sw t3, 12(a0)
michael@0	265	sw t4, 16(a0)
michael@0	266	sw t5, 20(a0)
michael@0	267	sw t6, 24(a0)
michael@0	268	sw t7, 28(a0)
michael@0	269	b 1b
michael@0	270	addiu a0, a0, 32
michael@0	271	2:
michael@0	272	lw t0, 0(a1)
michael@0	273	lw t1, 4(a1)
michael@0	274	lw t2, 8(a1)
michael@0	275	lw t3, 12(a1)
michael@0	276	lw t4, 16(a1)
michael@0	277	lw t5, 20(a1)
michael@0	278	lw t6, 24(a1)
michael@0	279	lw t7, 28(a1)
michael@0	280	addiu a1, a1, 32
michael@0	281	or t0, t0, t9
michael@0	282	or t1, t1, t9
michael@0	283	or t2, t2, t9
michael@0	284	or t3, t3, t9
michael@0	285	or t4, t4, t9
michael@0	286	or t5, t5, t9
michael@0	287	or t6, t6, t9
michael@0	288	or t7, t7, t9
michael@0	289	sw t0, 0(a0)
michael@0	290	sw t1, 4(a0)
michael@0	291	sw t2, 8(a0)
michael@0	292	sw t3, 12(a0)
michael@0	293	sw t4, 16(a0)
michael@0	294	sw t5, 20(a0)
michael@0	295	sw t6, 24(a0)
michael@0	296	sw t7, 28(a0)
michael@0	297	beqz a2, 4f
michael@0	298	addiu a0, a0, 32
michael@0	299	3:
michael@0	300	lw t0, 0(a1)
michael@0	301	addiu a1, a1, 4
michael@0	302	addiu a2, a2, -1
michael@0	303	or t1, t0, t9
michael@0	304	sw t1, 0(a0)
michael@0	305	bnez a2, 3b
michael@0	306	addiu a0, a0, 4
michael@0	307	4:
michael@0	308	jr ra
michael@0	309	nop
michael@0	310
michael@0	311	END(pixman_composite_src_x888_8888_asm_mips)
michael@0	312
michael@0	313	LEAF_MIPS_DSPR2(pixman_composite_src_n_8_8888_asm_mips)
michael@0	314	/*
michael@0	315	* a0 - dst (a8r8g8b8)
michael@0	316	* a1 - src (32bit constant)
michael@0	317	* a2 - mask (a8)
michael@0	318	* a3 - w
michael@0	319	*/
michael@0	320
michael@0	321
michael@0	322	SAVE_REGS_ON_STACK 0, v0
michael@0	323	li v0, 0x00ff00ff
michael@0	324
michael@0	325	beqz a3, 3f
michael@0	326	nop
michael@0	327	addiu t1, a3, -1
michael@0	328	beqz t1, 2f
michael@0	329	nop
michael@0	330
michael@0	331	1:
michael@0	332	/* a1 = source (32bit constant) */
michael@0	333	lbu t0, 0(a2) /* t2 = mask (a8) */
michael@0	334	lbu t1, 1(a2) /* t3 = mask (a8) */
michael@0	335	addiu a2, a2, 2
michael@0	336
michael@0	337	MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, t2, t3, v0, t4, t5, t6, t7, t8, t9
michael@0	338
michael@0	339	sw t2, 0(a0)
michael@0	340	sw t3, 4(a0)
michael@0	341	addiu a3, a3, -2
michael@0	342	addiu t2, a3, -1
michael@0	343	bgtz t2, 1b
michael@0	344	addiu a0, a0, 8
michael@0	345
michael@0	346	beqz a3, 3f
michael@0	347	nop
michael@0	348
michael@0	349	2:
michael@0	350	lbu t0, 0(a2)
michael@0	351	addiu a2, a2, 1
michael@0	352
michael@0	353	MIPS_UN8x4_MUL_UN8 a1, t0, t1, v0, t3, t4, t5
michael@0	354
michael@0	355	sw t1, 0(a0)
michael@0	356	addiu a3, a3, -1
michael@0	357	addiu a0, a0, 4
michael@0	358
michael@0	359	3:
michael@0	360	RESTORE_REGS_FROM_STACK 0, v0
michael@0	361	j ra
michael@0	362	nop
michael@0	363
michael@0	364	END(pixman_composite_src_n_8_8888_asm_mips)
michael@0	365
michael@0	366	LEAF_MIPS_DSPR2(pixman_composite_src_n_8_8_asm_mips)
michael@0	367	/*
michael@0	368	* a0 - dst (a8)
michael@0	369	* a1 - src (32bit constant)
michael@0	370	* a2 - mask (a8)
michael@0	371	* a3 - w
michael@0	372	*/
michael@0	373
michael@0	374	li t9, 0x00ff00ff
michael@0	375	beqz a3, 3f
michael@0	376	nop
michael@0	377	srl t7, a3, 2 /* t7 = how many multiples of 4 dst pixels */
michael@0	378	beqz t7, 1f /* branch if less than 4 src pixels */
michael@0	379	nop
michael@0	380
michael@0	381	srl t8, a1, 24
michael@0	382	replv.ph t8, t8
michael@0	383
michael@0	384	0:
michael@0	385	beqz t7, 1f
michael@0	386	addiu t7, t7, -1
michael@0	387	lbu t0, 0(a2)
michael@0	388	lbu t1, 1(a2)
michael@0	389	lbu t2, 2(a2)
michael@0	390	lbu t3, 3(a2)
michael@0	391
michael@0	392	addiu a2, a2, 4
michael@0	393
michael@0	394	precr_sra.ph.w t1, t0, 0
michael@0	395	precr_sra.ph.w t3, t2, 0
michael@0	396	precr.qb.ph t0, t3, t1
michael@0	397
michael@0	398	muleu_s.ph.qbl t2, t0, t8
michael@0	399	muleu_s.ph.qbr t3, t0, t8
michael@0	400	shra_r.ph t4, t2, 8
michael@0	401	shra_r.ph t5, t3, 8
michael@0	402	and t4, t4, t9
michael@0	403	and t5, t5, t9
michael@0	404	addq.ph t2, t2, t4
michael@0	405	addq.ph t3, t3, t5
michael@0	406	shra_r.ph t2, t2, 8
michael@0	407	shra_r.ph t3, t3, 8
michael@0	408	precr.qb.ph t2, t2, t3
michael@0	409
michael@0	410	sb t2, 0(a0)
michael@0	411	srl t2, t2, 8
michael@0	412	sb t2, 1(a0)
michael@0	413	srl t2, t2, 8
michael@0	414	sb t2, 2(a0)
michael@0	415	srl t2, t2, 8
michael@0	416	sb t2, 3(a0)
michael@0	417	addiu a3, a3, -4
michael@0	418	b 0b
michael@0	419	addiu a0, a0, 4
michael@0	420
michael@0	421	1:
michael@0	422	beqz a3, 3f
michael@0	423	nop
michael@0	424	srl t8, a1, 24
michael@0	425	2:
michael@0	426	lbu t0, 0(a2)
michael@0	427	addiu a2, a2, 1
michael@0	428
michael@0	429	mul t2, t0, t8
michael@0	430	shra_r.ph t3, t2, 8
michael@0	431	andi t3, t3, 0x00ff
michael@0	432	addq.ph t2, t2, t3
michael@0	433	shra_r.ph t2, t2, 8
michael@0	434
michael@0	435	sb t2, 0(a0)
michael@0	436	addiu a3, a3, -1
michael@0	437	bnez a3, 2b
michael@0	438	addiu a0, a0, 1
michael@0	439
michael@0	440	3:
michael@0	441	j ra
michael@0	442	nop
michael@0	443
michael@0	444	END(pixman_composite_src_n_8_8_asm_mips)
michael@0	445
michael@0	446	LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_8888_ca_asm_mips)
michael@0	447	/*
michael@0	448	* a0 - dst (a8r8g8b8)
michael@0	449	* a1 - src (32bit constant)
michael@0	450	* a2 - mask (a8r8g8b8)
michael@0	451	* a3 - w
michael@0	452	*/
michael@0	453
michael@0	454	SAVE_REGS_ON_STACK 8, s0, s1, s2, s3, s4, s5
michael@0	455	beqz a3, 4f
michael@0	456	nop
michael@0	457	li t6, 0xff
michael@0	458	addiu t7, zero, -1 /* t7 = 0xffffffff */
michael@0	459	srl t8, a1, 24 /* t8 = srca */
michael@0	460	li t9, 0x00ff00ff
michael@0	461	addiu t1, a3, -1
michael@0	462	beqz t1, 3f /* last pixel */
michael@0	463	nop
michael@0	464	beq t8, t6, 2f /* if (srca == 0xff) */
michael@0	465	nop
michael@0	466	1:
michael@0	467	/* a1 = src */
michael@0	468	lw t0, 0(a2) /* t0 = mask */
michael@0	469	lw t1, 4(a2) /* t1 = mask */
michael@0	470	or t2, t0, t1
michael@0	471	beqz t2, 12f /* if (t0 == 0) && (t1 == 0) */
michael@0	472	addiu a2, a2, 8
michael@0	473	and t3, t0, t1
michael@0	474	move t4, a1 /* t4 = src */
michael@0	475	move t5, a1 /* t5 = src */
michael@0	476	lw t2, 0(a0) /* t2 = dst */
michael@0	477	beq t3, t7, 11f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */
michael@0	478	lw t3, 4(a0) /* t3 = dst */
michael@0	479	MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, t4, t5, t9, s0, s1, s2, s3, s4, s5
michael@0	480	MIPS_2xUN8x4_MUL_2xUN8 t0, t1, t8, t8, t0, t1, t9, s0, s1, s2, s3, s4, s5
michael@0	481	11:
michael@0	482	not t0, t0
michael@0	483	not t1, t1
michael@0	484	MIPS_2xUN8x4_MUL_2xUN8x4 t2, t3, t0, t1, t2, t3, t9, s0, s1, s2, s3, s4, s5
michael@0	485	addu_s.qb t2, t4, t2
michael@0	486	addu_s.qb t3, t5, t3
michael@0	487	sw t2, 0(a0)
michael@0	488	sw t3, 4(a0)
michael@0	489	12:
michael@0	490	addiu a3, a3, -2
michael@0	491	addiu t1, a3, -1
michael@0	492	bgtz t1, 1b
michael@0	493	addiu a0, a0, 8
michael@0	494	b 3f
michael@0	495	nop
michael@0	496	2:
michael@0	497	/* a1 = src */
michael@0	498	lw t0, 0(a2) /* t0 = mask */
michael@0	499	lw t1, 4(a2) /* t1 = mask */
michael@0	500	or t2, t0, t1
michael@0	501	beqz t2, 22f /* if (t0 == 0) & (t1 == 0) */
michael@0	502	addiu a2, a2, 8
michael@0	503	and t2, t0, t1
michael@0	504	move t4, a1
michael@0	505	beq t2, t7, 21f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */
michael@0	506	move t5, a1
michael@0	507	lw t2, 0(a0) /* t2 = dst */
michael@0	508	lw t3, 4(a0) /* t3 = dst */
michael@0	509	MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, t4, t5, t9, s0, s1, s2, s3, s4, s5
michael@0	510	not t0, t0
michael@0	511	not t1, t1
michael@0	512	MIPS_2xUN8x4_MUL_2xUN8x4 t2, t3, t0, t1, t2, t3, t9, s0, s1, s2, s3, s4, s5
michael@0	513	addu_s.qb t4, t4, t2
michael@0	514	addu_s.qb t5, t5, t3
michael@0	515	21:
michael@0	516	sw t4, 0(a0)
michael@0	517	sw t5, 4(a0)
michael@0	518	22:
michael@0	519	addiu a3, a3, -2
michael@0	520	addiu t1, a3, -1
michael@0	521	bgtz t1, 2b
michael@0	522	addiu a0, a0, 8
michael@0	523	3:
michael@0	524	blez a3, 4f
michael@0	525	nop
michael@0	526	/* a1 = src */
michael@0	527	lw t1, 0(a2) /* t1 = mask */
michael@0	528	beqz t1, 4f
michael@0	529	nop
michael@0	530	move t2, a1 /* t2 = src */
michael@0	531	beq t1, t7, 31f
michael@0	532	lw t0, 0(a0) /* t0 = dst */
michael@0	533
michael@0	534	MIPS_UN8x4_MUL_UN8x4 a1, t1, t2, t9, t3, t4, t5, t6
michael@0	535	MIPS_UN8x4_MUL_UN8 t1, t8, t1, t9, t3, t4, t5
michael@0	536	31:
michael@0	537	not t1, t1
michael@0	538	MIPS_UN8x4_MUL_UN8x4 t0, t1, t0, t9, t3, t4, t5, t6
michael@0	539	addu_s.qb t0, t2, t0
michael@0	540	sw t0, 0(a0)
michael@0	541	4:
michael@0	542	RESTORE_REGS_FROM_STACK 8, s0, s1, s2, s3, s4, s5
michael@0	543	j ra
michael@0	544	nop
michael@0	545
michael@0	546	END(pixman_composite_over_n_8888_8888_ca_asm_mips)
michael@0	547
michael@0	548	LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_0565_ca_asm_mips)
michael@0	549	/*
michael@0	550	* a0 - dst (r5g6b5)
michael@0	551	* a1 - src (32bit constant)
michael@0	552	* a2 - mask (a8r8g8b8)
michael@0	553	* a3 - w
michael@0	554	*/
michael@0	555
michael@0	556	SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	557	beqz a3, 4f
michael@0	558	nop
michael@0	559	li t5, 0xf800f800
michael@0	560	li t6, 0x07e007e0
michael@0	561	li t7, 0x001F001F
michael@0	562	li t9, 0x00ff00ff
michael@0	563
michael@0	564	srl t8, a1, 24 /* t8 = srca */
michael@0	565	addiu t1, a3, -1
michael@0	566	beqz t1, 3f /* last pixel */
michael@0	567	nop
michael@0	568	li s0, 0xff /* s0 = 0xff */
michael@0	569	addiu s1, zero, -1 /* s1 = 0xffffffff */
michael@0	570
michael@0	571	beq t8, s0, 2f /* if (srca == 0xff) */
michael@0	572	nop
michael@0	573	1:
michael@0	574	/* a1 = src */
michael@0	575	lw t0, 0(a2) /* t0 = mask */
michael@0	576	lw t1, 4(a2) /* t1 = mask */
michael@0	577	or t2, t0, t1
michael@0	578	beqz t2, 12f /* if (t0 == 0) && (t1 == 0) */
michael@0	579	addiu a2, a2, 8
michael@0	580	and t3, t0, t1
michael@0	581	move s2, a1 /* s2 = src */
michael@0	582	move s3, a1 /* s3 = src */
michael@0	583	lhu t2, 0(a0) /* t2 = dst */
michael@0	584	beq t3, s1, 11f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */
michael@0	585	lhu t3, 2(a0) /* t3 = dst */
michael@0	586	MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, s2, s3, t9, t4, s4, s5, s6, s7, s8
michael@0	587	MIPS_2xUN8x4_MUL_2xUN8 t0, t1, t8, t8, t0, t1, t9, t4, s4, s5, s6, s7, s8
michael@0	588	11:
michael@0	589	not t0, t0
michael@0	590	not t1, t1
michael@0	591	CONVERT_2x0565_TO_2x8888 t2, t3, s4, s5, t6, t7, t4, s6, s7, s8
michael@0	592	MIPS_2xUN8x4_MUL_2xUN8x4 s4, s5, t0, t1, s4, s5, t9, t4, s6, s7, s8, t0, t1
michael@0	593	addu_s.qb s2, s2, s4
michael@0	594	addu_s.qb s3, s3, s5
michael@0	595	CONVERT_2x8888_TO_2x0565 s2, s3, t2, t3, t5, t6, t7, s4, s5
michael@0	596	sh t2, 0(a0)
michael@0	597	sh t3, 2(a0)
michael@0	598	12:
michael@0	599	addiu a3, a3, -2
michael@0	600	addiu t1, a3, -1
michael@0	601	bgtz t1, 1b
michael@0	602	addiu a0, a0, 4
michael@0	603	b 3f
michael@0	604	nop
michael@0	605	2:
michael@0	606	/* a1 = src */
michael@0	607	lw t0, 0(a2) /* t0 = mask */
michael@0	608	lw t1, 4(a2) /* t1 = mask */
michael@0	609	or t2, t0, t1
michael@0	610	beqz t2, 22f /* if (t0 == 0) & (t1 == 0) */
michael@0	611	addiu a2, a2, 8
michael@0	612	and t3, t0, t1
michael@0	613	move t2, a1
michael@0	614	beq t3, s1, 21f /* if (t0 == 0xffffffff) && (t1 == 0xffffffff) */
michael@0	615	move t3, a1
michael@0	616	lhu t2, 0(a0) /* t2 = dst */
michael@0	617	lhu t3, 2(a0) /* t3 = dst */
michael@0	618	MIPS_2xUN8x4_MUL_2xUN8x4 a1, a1, t0, t1, s2, s3, t9, t4, s4, s5, s6, s7, s8
michael@0	619	not t0, t0
michael@0	620	not t1, t1
michael@0	621	CONVERT_2x0565_TO_2x8888 t2, t3, s4, s5, t6, t7, t4, s6, s7, s8
michael@0	622	MIPS_2xUN8x4_MUL_2xUN8x4 s4, s5, t0, t1, s4, s5, t9, t4, s6, s7, s8, t2, t3
michael@0	623	addu_s.qb t2, s2, s4
michael@0	624	addu_s.qb t3, s3, s5
michael@0	625	21:
michael@0	626	CONVERT_2x8888_TO_2x0565 t2, t3, t0, t1, t5, t6, t7, s2, s3
michael@0	627	sh t0, 0(a0)
michael@0	628	sh t1, 2(a0)
michael@0	629	22:
michael@0	630	addiu a3, a3, -2
michael@0	631	addiu t1, a3, -1
michael@0	632	bgtz t1, 2b
michael@0	633	addiu a0, a0, 4
michael@0	634	3:
michael@0	635	blez a3, 4f
michael@0	636	nop
michael@0	637	/* a1 = src */
michael@0	638	lw t1, 0(a2) /* t1 = mask */
michael@0	639	beqz t1, 4f
michael@0	640	nop
michael@0	641	move t2, a1 /* t2 = src */
michael@0	642	beq t1, t7, 31f
michael@0	643	lhu t0, 0(a0) /* t0 = dst */
michael@0	644
michael@0	645	MIPS_UN8x4_MUL_UN8x4 a1, t1, t2, t9, t3, t4, t5, t6
michael@0	646	MIPS_UN8x4_MUL_UN8 t1, t8, t1, t9, t3, t4, t5
michael@0	647	31:
michael@0	648	not t1, t1
michael@0	649	CONVERT_1x0565_TO_1x8888 t0, s1, s2, s3
michael@0	650	MIPS_UN8x4_MUL_UN8x4 s1, t1, t3, t9, t4, t5, t6, t7
michael@0	651	addu_s.qb t0, t2, t3
michael@0	652	CONVERT_1x8888_TO_1x0565 t0, s1, s2, s3
michael@0	653	sh s1, 0(a0)
michael@0	654	4:
michael@0	655	RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	656	j ra
michael@0	657	nop
michael@0	658
michael@0	659	END(pixman_composite_over_n_8888_0565_ca_asm_mips)
michael@0	660
michael@0	661	LEAF_MIPS_DSPR2(pixman_composite_over_n_8_8888_asm_mips)
michael@0	662	/*
michael@0	663	* a0 - dst (a8r8g8b8)
michael@0	664	* a1 - src (32bit constant)
michael@0	665	* a2 - mask (a8)
michael@0	666	* a3 - w
michael@0	667	*/
michael@0	668
michael@0	669	SAVE_REGS_ON_STACK 4, s0, s1, s2, s3, s4
michael@0	670	beqz a3, 4f
michael@0	671	nop
michael@0	672	li t4, 0x00ff00ff
michael@0	673	li t5, 0xff
michael@0	674	addiu t0, a3, -1
michael@0	675	beqz t0, 3f /* last pixel */
michael@0	676	srl t6, a1, 24 /* t6 = srca */
michael@0	677	not s4, a1
michael@0	678	beq t5, t6, 2f /* if (srca == 0xff) */
michael@0	679	srl s4, s4, 24
michael@0	680	1:
michael@0	681	/* a1 = src */
michael@0	682	lbu t0, 0(a2) /* t0 = mask */
michael@0	683	lbu t1, 1(a2) /* t1 = mask */
michael@0	684	or t2, t0, t1
michael@0	685	beqz t2, 111f /* if (t0 == 0) && (t1 == 0) */
michael@0	686	addiu a2, a2, 2
michael@0	687	and t3, t0, t1
michael@0	688
michael@0	689	lw t2, 0(a0) /* t2 = dst */
michael@0	690	beq t3, t5, 11f /* if (t0 == 0xff) && (t1 == 0xff) */
michael@0	691	lw t3, 4(a0) /* t3 = dst */
michael@0	692
michael@0	693	MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, s0, s1, t4, t6, t7, t8, t9, s2, s3
michael@0	694	not s2, s0
michael@0	695	not s3, s1
michael@0	696	srl s2, s2, 24
michael@0	697	srl s3, s3, 24
michael@0	698	MIPS_2xUN8x4_MUL_2xUN8 t2, t3, s2, s3, t2, t3, t4, t0, t1, t6, t7, t8, t9
michael@0	699	addu_s.qb s2, t2, s0
michael@0	700	addu_s.qb s3, t3, s1
michael@0	701	sw s2, 0(a0)
michael@0	702	b 111f
michael@0	703	sw s3, 4(a0)
michael@0	704	11:
michael@0	705	MIPS_2xUN8x4_MUL_2xUN8 t2, t3, s4, s4, t2, t3, t4, t0, t1, t6, t7, t8, t9
michael@0	706	addu_s.qb s2, t2, a1
michael@0	707	addu_s.qb s3, t3, a1
michael@0	708	sw s2, 0(a0)
michael@0	709	sw s3, 4(a0)
michael@0	710
michael@0	711	111:
michael@0	712	addiu a3, a3, -2
michael@0	713	addiu t0, a3, -1
michael@0	714	bgtz t0, 1b
michael@0	715	addiu a0, a0, 8
michael@0	716	b 3f
michael@0	717	nop
michael@0	718	2:
michael@0	719	/* a1 = src */
michael@0	720	lbu t0, 0(a2) /* t0 = mask */
michael@0	721	lbu t1, 1(a2) /* t1 = mask */
michael@0	722	or t2, t0, t1
michael@0	723	beqz t2, 222f /* if (t0 == 0) && (t1 == 0) */
michael@0	724	addiu a2, a2, 2
michael@0	725	and t3, t0, t1
michael@0	726	beq t3, t5, 22f /* if (t0 == 0xff) && (t1 == 0xff) */
michael@0	727	nop
michael@0	728	lw t2, 0(a0) /* t2 = dst */
michael@0	729	lw t3, 4(a0) /* t3 = dst */
michael@0	730
michael@0	731	OVER_2x8888_2x8_2x8888 a1, a1, t0, t1, t2, t3, \
michael@0	732	t6, t7, t4, t8, t9, s0, s1, s2, s3
michael@0	733	sw t6, 0(a0)
michael@0	734	b 222f
michael@0	735	sw t7, 4(a0)
michael@0	736	22:
michael@0	737	sw a1, 0(a0)
michael@0	738	sw a1, 4(a0)
michael@0	739	222:
michael@0	740	addiu a3, a3, -2
michael@0	741	addiu t0, a3, -1
michael@0	742	bgtz t0, 2b
michael@0	743	addiu a0, a0, 8
michael@0	744	3:
michael@0	745	blez a3, 4f
michael@0	746	nop
michael@0	747	/* a1 = src */
michael@0	748	lbu t0, 0(a2) /* t0 = mask */
michael@0	749	beqz t0, 4f /* if (t0 == 0) */
michael@0	750	addiu a2, a2, 1
michael@0	751	move t3, a1
michael@0	752	beq t0, t5, 31f /* if (t0 == 0xff) */
michael@0	753	lw t1, 0(a0) /* t1 = dst */
michael@0	754
michael@0	755	MIPS_UN8x4_MUL_UN8 a1, t0, t3, t4, t6, t7, t8
michael@0	756	31:
michael@0	757	not t2, t3
michael@0	758	srl t2, t2, 24
michael@0	759	MIPS_UN8x4_MUL_UN8 t1, t2, t1, t4, t6, t7, t8
michael@0	760	addu_s.qb t2, t1, t3
michael@0	761	sw t2, 0(a0)
michael@0	762	4:
michael@0	763	RESTORE_REGS_FROM_STACK 4, s0, s1, s2, s3, s4
michael@0	764	j ra
michael@0	765	nop
michael@0	766
michael@0	767	END(pixman_composite_over_n_8_8888_asm_mips)
michael@0	768
michael@0	769	LEAF_MIPS_DSPR2(pixman_composite_over_n_8_0565_asm_mips)
michael@0	770	/*
michael@0	771	* a0 - dst (r5g6b5)
michael@0	772	* a1 - src (32bit constant)
michael@0	773	* a2 - mask (a8)
michael@0	774	* a3 - w
michael@0	775	*/
michael@0	776	SAVE_REGS_ON_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	777	beqz a3, 4f
michael@0	778	nop
michael@0	779	li t4, 0x00ff00ff
michael@0	780	li t5, 0xff
michael@0	781	li t6, 0xf800f800
michael@0	782	li t7, 0x07e007e0
michael@0	783	li t8, 0x001F001F
michael@0	784	addiu t1, a3, -1
michael@0	785	beqz t1, 3f /* last pixel */
michael@0	786	srl t0, a1, 24 /* t0 = srca */
michael@0	787	not v0, a1
michael@0	788	beq t0, t5, 2f /* if (srca == 0xff) */
michael@0	789	srl v0, v0, 24
michael@0	790	1:
michael@0	791	/* a1 = src */
michael@0	792	lbu t0, 0(a2) /* t0 = mask */
michael@0	793	lbu t1, 1(a2) /* t1 = mask */
michael@0	794	or t2, t0, t1
michael@0	795	beqz t2, 111f /* if (t0 == 0) && (t1 == 0) */
michael@0	796	addiu a2, a2, 2
michael@0	797	lhu t2, 0(a0) /* t2 = dst */
michael@0	798	lhu t3, 2(a0) /* t3 = dst */
michael@0	799	CONVERT_2x0565_TO_2x8888 t2, t3, s0, s1, t7, t8, t9, s2, s3, s4
michael@0	800	and t9, t0, t1
michael@0	801	beq t9, t5, 11f /* if (t0 == 0xff) && (t1 == 0xff) */
michael@0	802	nop
michael@0	803
michael@0	804	MIPS_2xUN8x4_MUL_2xUN8 a1, a1, t0, t1, s2, s3, t4, t9, s4, s5, s6, s7, s8
michael@0	805	not s4, s2
michael@0	806	not s5, s3
michael@0	807	srl s4, s4, 24
michael@0	808	srl s5, s5, 24
michael@0	809	MIPS_2xUN8x4_MUL_2xUN8 s0, s1, s4, s5, s0, s1, t4, t9, t0, t1, s6, s7, s8
michael@0	810	addu_s.qb s4, s2, s0
michael@0	811	addu_s.qb s5, s3, s1
michael@0	812	CONVERT_2x8888_TO_2x0565 s4, s5, t2, t3, t6, t7, t8, s0, s1
michael@0	813	sh t2, 0(a0)
michael@0	814	b 111f
michael@0	815	sh t3, 2(a0)
michael@0	816	11:
michael@0	817	MIPS_2xUN8x4_MUL_2xUN8 s0, s1, v0, v0, s0, s1, t4, t9, t0, t1, s6, s7, s8
michael@0	818	addu_s.qb s4, a1, s0
michael@0	819	addu_s.qb s5, a1, s1
michael@0	820	CONVERT_2x8888_TO_2x0565 s4, s5, t2, t3, t6, t7, t8, s0, s1
michael@0	821	sh t2, 0(a0)
michael@0	822	sh t3, 2(a0)
michael@0	823	111:
michael@0	824	addiu a3, a3, -2
michael@0	825	addiu t0, a3, -1
michael@0	826	bgtz t0, 1b
michael@0	827	addiu a0, a0, 4
michael@0	828	b 3f
michael@0	829	nop
michael@0	830	2:
michael@0	831	CONVERT_1x8888_TO_1x0565 a1, s0, s1, s2
michael@0	832	21:
michael@0	833	/* a1 = src */
michael@0	834	lbu t0, 0(a2) /* t0 = mask */
michael@0	835	lbu t1, 1(a2) /* t1 = mask */
michael@0	836	or t2, t0, t1
michael@0	837	beqz t2, 222f /* if (t0 == 0) && (t1 == 0) */
michael@0	838	addiu a2, a2, 2
michael@0	839	and t9, t0, t1
michael@0	840	move s2, s0
michael@0	841	beq t9, t5, 22f /* if (t0 == 0xff) && (t2 == 0xff) */
michael@0	842	move s3, s0
michael@0	843	lhu t2, 0(a0) /* t2 = dst */
michael@0	844	lhu t3, 2(a0) /* t3 = dst */
michael@0	845
michael@0	846	CONVERT_2x0565_TO_2x8888 t2, t3, s2, s3, t7, t8, s4, s5, s6, s7
michael@0	847	OVER_2x8888_2x8_2x8888 a1, a1, t0, t1, s2, s3, \
michael@0	848	t2, t3, t4, t9, s4, s5, s6, s7, s8
michael@0	849	CONVERT_2x8888_TO_2x0565 t2, t3, s2, s3, t6, t7, t8, s4, s5
michael@0	850	22:
michael@0	851	sh s2, 0(a0)
michael@0	852	sh s3, 2(a0)
michael@0	853	222:
michael@0	854	addiu a3, a3, -2
michael@0	855	addiu t0, a3, -1
michael@0	856	bgtz t0, 21b
michael@0	857	addiu a0, a0, 4
michael@0	858	3:
michael@0	859	blez a3, 4f
michael@0	860	nop
michael@0	861	/* a1 = src */
michael@0	862	lbu t0, 0(a2) /* t0 = mask */
michael@0	863	beqz t0, 4f /* if (t0 == 0) */
michael@0	864	nop
michael@0	865	lhu t1, 0(a0) /* t1 = dst */
michael@0	866	CONVERT_1x0565_TO_1x8888 t1, t2, t3, t7
michael@0	867	beq t0, t5, 31f /* if (t0 == 0xff) */
michael@0	868	move t3, a1
michael@0	869
michael@0	870	MIPS_UN8x4_MUL_UN8 a1, t0, t3, t4, t7, t8, t9
michael@0	871	31:
michael@0	872	not t6, t3
michael@0	873	srl t6, t6, 24
michael@0	874	MIPS_UN8x4_MUL_UN8 t2, t6, t2, t4, t7, t8, t9
michael@0	875	addu_s.qb t1, t2, t3
michael@0	876	CONVERT_1x8888_TO_1x0565 t1, t2, t3, t7
michael@0	877	sh t2, 0(a0)
michael@0	878	4:
michael@0	879	RESTORE_REGS_FROM_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	880	j ra
michael@0	881	nop
michael@0	882
michael@0	883	END(pixman_composite_over_n_8_0565_asm_mips)
michael@0	884
michael@0	885	LEAF_MIPS_DSPR2(pixman_composite_over_8888_n_8888_asm_mips)
michael@0	886	/*
michael@0	887	* a0 - dst (a8r8g8b8)
michael@0	888	* a1 - src (a8r8g8b8)
michael@0	889	* a2 - mask (32bit constant)
michael@0	890	* a3 - w
michael@0	891	*/
michael@0	892
michael@0	893	SAVE_REGS_ON_STACK 0, s0
michael@0	894	li t4, 0x00ff00ff
michael@0	895	beqz a3, 3f
michael@0	896	nop
michael@0	897	addiu t1, a3, -1
michael@0	898	srl a2, a2, 24
michael@0	899	beqz t1, 2f
michael@0	900	nop
michael@0	901
michael@0	902	1:
michael@0	903	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	904	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	905	/* a2 = mask (32bit constant) */
michael@0	906	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	907	lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
michael@0	908	addiu a1, a1, 8
michael@0	909
michael@0	910	OVER_2x8888_2x8_2x8888 t0, t1, a2, a2, t2, t3, \
michael@0	911	t5, t6, t4, t7, t8, t9, t0, t1, s0
michael@0	912
michael@0	913	sw t5, 0(a0)
michael@0	914	sw t6, 4(a0)
michael@0	915	addiu a3, a3, -2
michael@0	916	addiu t1, a3, -1
michael@0	917	bgtz t1, 1b
michael@0	918	addiu a0, a0, 8
michael@0	919	2:
michael@0	920	beqz a3, 3f
michael@0	921	nop
michael@0	922	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	923	/* a2 = mask (32bit constant) */
michael@0	924	lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
michael@0	925
michael@0	926	OVER_8888_8_8888 t0, a2, t1, t3, t4, t5, t6, t7, t8
michael@0	927
michael@0	928	sw t3, 0(a0)
michael@0	929	3:
michael@0	930	RESTORE_REGS_FROM_STACK 0, s0
michael@0	931	j ra
michael@0	932	nop
michael@0	933
michael@0	934	END(pixman_composite_over_8888_n_8888_asm_mips)
michael@0	935
michael@0	936	LEAF_MIPS_DSPR2(pixman_composite_over_8888_n_0565_asm_mips)
michael@0	937	/*
michael@0	938	* a0 - dst (r5g6b5)
michael@0	939	* a1 - src (a8r8g8b8)
michael@0	940	* a2 - mask (32bit constant)
michael@0	941	* a3 - w
michael@0	942	*/
michael@0	943
michael@0	944	SAVE_REGS_ON_STACK 0, s0, s1, s2, s3
michael@0	945	li t6, 0x00ff00ff
michael@0	946	li t7, 0xf800f800
michael@0	947	li t8, 0x07e007e0
michael@0	948	li t9, 0x001F001F
michael@0	949	beqz a3, 3f
michael@0	950	nop
michael@0	951	srl a2, a2, 24
michael@0	952	addiu t1, a3, -1
michael@0	953	beqz t1, 2f
michael@0	954	nop
michael@0	955	1:
michael@0	956	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	957	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	958	/* a2 = mask (32bit constant) */
michael@0	959	lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
michael@0	960	lhu t3, 2(a0) /* t2 = destination (r5g6b5) */
michael@0	961	addiu a1, a1, 8
michael@0	962
michael@0	963	CONVERT_2x0565_TO_2x8888 t2, t3, t4, t5, t8, t9, s0, s1, t2, t3
michael@0	964	OVER_2x8888_2x8_2x8888 t0, t1, a2, a2, t4, t5, \
michael@0	965	t2, t3, t6, t0, t1, s0, s1, s2, s3
michael@0	966	CONVERT_2x8888_TO_2x0565 t2, t3, t4, t5, t7, t8, t9, s0, s1
michael@0	967
michael@0	968	sh t4, 0(a0)
michael@0	969	sh t5, 2(a0)
michael@0	970	addiu a3, a3, -2
michael@0	971	addiu t1, a3, -1
michael@0	972	bgtz t1, 1b
michael@0	973	addiu a0, a0, 4
michael@0	974	2:
michael@0	975	beqz a3, 3f
michael@0	976	nop
michael@0	977	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	978	/* a2 = mask (32bit constant) */
michael@0	979	lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
michael@0	980
michael@0	981	CONVERT_1x0565_TO_1x8888 t1, t2, t4, t5
michael@0	982	OVER_8888_8_8888 t0, a2, t2, t1, t6, t3, t4, t5, t7
michael@0	983	CONVERT_1x8888_TO_1x0565 t1, t3, t4, t5
michael@0	984
michael@0	985	sh t3, 0(a0)
michael@0	986	3:
michael@0	987	RESTORE_REGS_FROM_STACK 0, s0, s1, s2, s3
michael@0	988	j ra
michael@0	989	nop
michael@0	990
michael@0	991	END(pixman_composite_over_8888_n_0565_asm_mips)
michael@0	992
michael@0	993	LEAF_MIPS_DSPR2(pixman_composite_over_0565_n_0565_asm_mips)
michael@0	994	/*
michael@0	995	* a0 - dst (r5g6b5)
michael@0	996	* a1 - src (r5g6b5)
michael@0	997	* a2 - mask (32bit constant)
michael@0	998	* a3 - w
michael@0	999	*/
michael@0	1000
michael@0	1001	SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5
michael@0	1002	li t6, 0x00ff00ff
michael@0	1003	li t7, 0xf800f800
michael@0	1004	li t8, 0x07e007e0
michael@0	1005	li t9, 0x001F001F
michael@0	1006	beqz a3, 3f
michael@0	1007	nop
michael@0	1008	srl a2, a2, 24
michael@0	1009	addiu t1, a3, -1
michael@0	1010	beqz t1, 2f
michael@0	1011	nop
michael@0	1012	1:
michael@0	1013	lhu t0, 0(a1) /* t0 = source (r5g6b5) */
michael@0	1014	lhu t1, 2(a1) /* t1 = source (r5g6b5) */
michael@0	1015	/* a2 = mask (32bit constant) */
michael@0	1016	lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
michael@0	1017	lhu t3, 2(a0) /* t3 = destination (r5g6b5) */
michael@0	1018	addiu a1, a1, 4
michael@0	1019
michael@0	1020	CONVERT_2x0565_TO_2x8888 t0, t1, t4, t5, t8, t9, s0, s1, s2, s3
michael@0	1021	CONVERT_2x0565_TO_2x8888 t2, t3, s0, s1, t8, t9, s2, s3, s4, s5
michael@0	1022	OVER_2x8888_2x8_2x8888 t4, t5, a2, a2, s0, s1, \
michael@0	1023	t0, t1, t6, s2, s3, s4, s5, t4, t5
michael@0	1024	CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t7, t8, t9, s2, s3
michael@0	1025
michael@0	1026	sh s0, 0(a0)
michael@0	1027	sh s1, 2(a0)
michael@0	1028	addiu a3, a3, -2
michael@0	1029	addiu t1, a3, -1
michael@0	1030	bgtz t1, 1b
michael@0	1031	addiu a0, a0, 4
michael@0	1032	2:
michael@0	1033	beqz a3, 3f
michael@0	1034	nop
michael@0	1035	lhu t0, 0(a1) /* t0 = source (r5g6b5) */
michael@0	1036	/* a2 = mask (32bit constant) */
michael@0	1037	lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
michael@0	1038
michael@0	1039	CONVERT_1x0565_TO_1x8888 t0, t2, t4, t5
michael@0	1040	CONVERT_1x0565_TO_1x8888 t1, t3, t4, t5
michael@0	1041	OVER_8888_8_8888 t2, a2, t3, t0, t6, t1, t4, t5, t7
michael@0	1042	CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5
michael@0	1043
michael@0	1044	sh t3, 0(a0)
michael@0	1045	3:
michael@0	1046	RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5
michael@0	1047	j ra
michael@0	1048	nop
michael@0	1049
michael@0	1050	END(pixman_composite_over_0565_n_0565_asm_mips)
michael@0	1051
michael@0	1052	LEAF_MIPS_DSPR2(pixman_composite_over_8888_8_8888_asm_mips)
michael@0	1053	/*
michael@0	1054	* a0 - dst (a8r8g8b8)
michael@0	1055	* a1 - src (a8r8g8b8)
michael@0	1056	* a2 - mask (a8)
michael@0	1057	* a3 - w
michael@0	1058	*/
michael@0	1059
michael@0	1060	SAVE_REGS_ON_STACK 0, s0, s1
michael@0	1061	li t4, 0x00ff00ff
michael@0	1062	beqz a3, 3f
michael@0	1063	nop
michael@0	1064	addiu t1, a3, -1
michael@0	1065	beqz t1, 2f
michael@0	1066	nop
michael@0	1067	1:
michael@0	1068	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1069	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	1070	lbu t2, 0(a2) /* t2 = mask (a8) */
michael@0	1071	lbu t3, 1(a2) /* t3 = mask (a8) */
michael@0	1072	lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */
michael@0	1073	lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */
michael@0	1074	addiu a1, a1, 8
michael@0	1075	addiu a2, a2, 2
michael@0	1076
michael@0	1077	OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, t5, t6, \
michael@0	1078	t7, t8, t4, t9, s0, s1, t0, t1, t2
michael@0	1079
michael@0	1080	sw t7, 0(a0)
michael@0	1081	sw t8, 4(a0)
michael@0	1082	addiu a3, a3, -2
michael@0	1083	addiu t1, a3, -1
michael@0	1084	bgtz t1, 1b
michael@0	1085	addiu a0, a0, 8
michael@0	1086	2:
michael@0	1087	beqz a3, 3f
michael@0	1088	nop
michael@0	1089	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1090	lbu t1, 0(a2) /* t1 = mask (a8) */
michael@0	1091	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	1092
michael@0	1093	OVER_8888_8_8888 t0, t1, t2, t3, t4, t5, t6, t7, t8
michael@0	1094
michael@0	1095	sw t3, 0(a0)
michael@0	1096	3:
michael@0	1097	RESTORE_REGS_FROM_STACK 0, s0, s1
michael@0	1098	j ra
michael@0	1099	nop
michael@0	1100
michael@0	1101	END(pixman_composite_over_8888_8_8888_asm_mips)
michael@0	1102
michael@0	1103	LEAF_MIPS_DSPR2(pixman_composite_over_8888_8_0565_asm_mips)
michael@0	1104	/*
michael@0	1105	* a0 - dst (r5g6b5)
michael@0	1106	* a1 - src (a8r8g8b8)
michael@0	1107	* a2 - mask (a8)
michael@0	1108	* a3 - w
michael@0	1109	*/
michael@0	1110
michael@0	1111	SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5
michael@0	1112	li t6, 0x00ff00ff
michael@0	1113	li t7, 0xf800f800
michael@0	1114	li t8, 0x07e007e0
michael@0	1115	li t9, 0x001F001F
michael@0	1116	beqz a3, 3f
michael@0	1117	nop
michael@0	1118	addiu t1, a3, -1
michael@0	1119	beqz t1, 2f
michael@0	1120	nop
michael@0	1121	1:
michael@0	1122	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1123	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	1124	lbu t2, 0(a2) /* t2 = mask (a8) */
michael@0	1125	lbu t3, 1(a2) /* t3 = mask (a8) */
michael@0	1126	lhu t4, 0(a0) /* t4 = destination (r5g6b5) */
michael@0	1127	lhu t5, 2(a0) /* t5 = destination (r5g6b5) */
michael@0	1128	addiu a1, a1, 8
michael@0	1129	addiu a2, a2, 2
michael@0	1130
michael@0	1131	CONVERT_2x0565_TO_2x8888 t4, t5, s0, s1, t8, t9, s2, s3, s4, s5
michael@0	1132	OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, s0, s1, \
michael@0	1133	t4, t5, t6, s2, s3, s4, s5, t0, t1
michael@0	1134	CONVERT_2x8888_TO_2x0565 t4, t5, s0, s1, t7, t8, t9, s2, s3
michael@0	1135
michael@0	1136	sh s0, 0(a0)
michael@0	1137	sh s1, 2(a0)
michael@0	1138	addiu a3, a3, -2
michael@0	1139	addiu t1, a3, -1
michael@0	1140	bgtz t1, 1b
michael@0	1141	addiu a0, a0, 4
michael@0	1142	2:
michael@0	1143	beqz a3, 3f
michael@0	1144	nop
michael@0	1145	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1146	lbu t1, 0(a2) /* t1 = mask (a8) */
michael@0	1147	lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
michael@0	1148
michael@0	1149	CONVERT_1x0565_TO_1x8888 t2, t3, t4, t5
michael@0	1150	OVER_8888_8_8888 t0, t1, t3, t2, t6, t4, t5, t7, t8
michael@0	1151	CONVERT_1x8888_TO_1x0565 t2, t3, t4, t5
michael@0	1152
michael@0	1153	sh t3, 0(a0)
michael@0	1154	3:
michael@0	1155	RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5
michael@0	1156	j ra
michael@0	1157	nop
michael@0	1158
michael@0	1159	END(pixman_composite_over_8888_8_0565_asm_mips)
michael@0	1160
michael@0	1161	LEAF_MIPS_DSPR2(pixman_composite_over_0565_8_0565_asm_mips)
michael@0	1162	/*
michael@0	1163	* a0 - dst (r5g6b5)
michael@0	1164	* a1 - src (r5g6b5)
michael@0	1165	* a2 - mask (a8)
michael@0	1166	* a3 - w
michael@0	1167	*/
michael@0	1168
michael@0	1169	SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5
michael@0	1170	li t4, 0xf800f800
michael@0	1171	li t5, 0x07e007e0
michael@0	1172	li t6, 0x001F001F
michael@0	1173	li t7, 0x00ff00ff
michael@0	1174	beqz a3, 3f
michael@0	1175	nop
michael@0	1176	addiu t1, a3, -1
michael@0	1177	beqz t1, 2f
michael@0	1178	nop
michael@0	1179	1:
michael@0	1180	lhu t0, 0(a1) /* t0 = source (r5g6b5) */
michael@0	1181	lhu t1, 2(a1) /* t1 = source (r5g6b5) */
michael@0	1182	lbu t2, 0(a2) /* t2 = mask (a8) */
michael@0	1183	lbu t3, 1(a2) /* t3 = mask (a8) */
michael@0	1184	lhu t8, 0(a0) /* t8 = destination (r5g6b5) */
michael@0	1185	lhu t9, 2(a0) /* t9 = destination (r5g6b5) */
michael@0	1186	addiu a1, a1, 4
michael@0	1187	addiu a2, a2, 2
michael@0	1188
michael@0	1189	CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5
michael@0	1190	CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, t0, t1
michael@0	1191	OVER_2x8888_2x8_2x8888 s0, s1, t2, t3, s2, s3, \
michael@0	1192	t0, t1, t7, s4, s5, t8, t9, s0, s1
michael@0	1193	CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3
michael@0	1194
michael@0	1195	sh s0, 0(a0)
michael@0	1196	sh s1, 2(a0)
michael@0	1197	addiu a3, a3, -2
michael@0	1198	addiu t1, a3, -1
michael@0	1199	bgtz t1, 1b
michael@0	1200	addiu a0, a0, 4
michael@0	1201	2:
michael@0	1202	beqz a3, 3f
michael@0	1203	nop
michael@0	1204	lhu t0, 0(a1) /* t0 = source (r5g6b5) */
michael@0	1205	lbu t1, 0(a2) /* t1 = mask (a8) */
michael@0	1206	lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
michael@0	1207
michael@0	1208	CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5
michael@0	1209	CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6
michael@0	1210	OVER_8888_8_8888 t3, t1, t4, t0, t7, t2, t5, t6, t8
michael@0	1211	CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5
michael@0	1212
michael@0	1213	sh t3, 0(a0)
michael@0	1214	3:
michael@0	1215	RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5
michael@0	1216	j ra
michael@0	1217	nop
michael@0	1218
michael@0	1219	END(pixman_composite_over_0565_8_0565_asm_mips)
michael@0	1220
michael@0	1221	LEAF_MIPS_DSPR2(pixman_composite_over_8888_8888_8888_asm_mips)
michael@0	1222	/*
michael@0	1223	* a0 - dst (a8r8g8b8)
michael@0	1224	* a1 - src (a8r8g8b8)
michael@0	1225	* a2 - mask (a8r8g8b8)
michael@0	1226	* a3 - w
michael@0	1227	*/
michael@0	1228
michael@0	1229	SAVE_REGS_ON_STACK 0, s0, s1, s2
michael@0	1230	li t4, 0x00ff00ff
michael@0	1231	beqz a3, 3f
michael@0	1232	nop
michael@0	1233	addiu t1, a3, -1
michael@0	1234	beqz t1, 2f
michael@0	1235	nop
michael@0	1236	1:
michael@0	1237	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1238	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	1239	lw t2, 0(a2) /* t2 = mask (a8r8g8b8) */
michael@0	1240	lw t3, 4(a2) /* t3 = mask (a8r8g8b8) */
michael@0	1241	lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */
michael@0	1242	lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */
michael@0	1243	addiu a1, a1, 8
michael@0	1244	addiu a2, a2, 8
michael@0	1245	srl t2, t2, 24
michael@0	1246	srl t3, t3, 24
michael@0	1247
michael@0	1248	OVER_2x8888_2x8_2x8888 t0, t1, t2, t3, t5, t6, t7, t8, t4, t9, s0, s1, s2, t0, t1
michael@0	1249
michael@0	1250	sw t7, 0(a0)
michael@0	1251	sw t8, 4(a0)
michael@0	1252	addiu a3, a3, -2
michael@0	1253	addiu t1, a3, -1
michael@0	1254	bgtz t1, 1b
michael@0	1255	addiu a0, a0, 8
michael@0	1256	2:
michael@0	1257	beqz a3, 3f
michael@0	1258	nop
michael@0	1259	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1260	lw t1, 0(a2) /* t1 = mask (a8r8g8b8) */
michael@0	1261	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	1262	srl t1, t1, 24
michael@0	1263
michael@0	1264	OVER_8888_8_8888 t0, t1, t2, t3, t4, t5, t6, t7, t8
michael@0	1265
michael@0	1266	sw t3, 0(a0)
michael@0	1267	3:
michael@0	1268	RESTORE_REGS_FROM_STACK 0, s0, s1, s2
michael@0	1269	j ra
michael@0	1270	nop
michael@0	1271
michael@0	1272	END(pixman_composite_over_8888_8888_8888_asm_mips)
michael@0	1273
michael@0	1274	LEAF_MIPS_DSPR2(pixman_composite_over_8888_8888_asm_mips)
michael@0	1275	/*
michael@0	1276	* a0 - dst (a8r8g8b8)
michael@0	1277	* a1 - src (a8r8g8b8)
michael@0	1278	* a2 - w
michael@0	1279	*/
michael@0	1280
michael@0	1281	SAVE_REGS_ON_STACK 0, s0, s1, s2
michael@0	1282	li t4, 0x00ff00ff
michael@0	1283	beqz a2, 3f
michael@0	1284	nop
michael@0	1285	addiu t1, a2, -1
michael@0	1286	beqz t1, 2f
michael@0	1287	nop
michael@0	1288	1:
michael@0	1289	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1290	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	1291	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	1292	lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
michael@0	1293	addiu a1, a1, 8
michael@0	1294
michael@0	1295	not t5, t0
michael@0	1296	srl t5, t5, 24
michael@0	1297	not t6, t1
michael@0	1298	srl t6, t6, 24
michael@0	1299
michael@0	1300	or t7, t5, t6
michael@0	1301	beqz t7, 11f
michael@0	1302	or t8, t0, t1
michael@0	1303	beqz t8, 12f
michael@0	1304
michael@0	1305	MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t5, t6, t7, t8, t4, t9, s0, s1, s2, t2, t3
michael@0	1306
michael@0	1307	addu_s.qb t0, t7, t0
michael@0	1308	addu_s.qb t1, t8, t1
michael@0	1309	11:
michael@0	1310	sw t0, 0(a0)
michael@0	1311	sw t1, 4(a0)
michael@0	1312	12:
michael@0	1313	addiu a2, a2, -2
michael@0	1314	addiu t1, a2, -1
michael@0	1315	bgtz t1, 1b
michael@0	1316	addiu a0, a0, 8
michael@0	1317	2:
michael@0	1318	beqz a2, 3f
michael@0	1319	nop
michael@0	1320
michael@0	1321	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1322	lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
michael@0	1323	addiu a1, a1, 4
michael@0	1324
michael@0	1325	not t2, t0
michael@0	1326	srl t2, t2, 24
michael@0	1327
michael@0	1328	beqz t2, 21f
michael@0	1329	nop
michael@0	1330	beqz t0, 3f
michael@0	1331
michael@0	1332	MIPS_UN8x4_MUL_UN8 t1, t2, t3, t4, t5, t6, t7
michael@0	1333
michael@0	1334	addu_s.qb t0, t3, t0
michael@0	1335	21:
michael@0	1336	sw t0, 0(a0)
michael@0	1337
michael@0	1338	3:
michael@0	1339	RESTORE_REGS_FROM_STACK 0, s0, s1, s2
michael@0	1340	j ra
michael@0	1341	nop
michael@0	1342
michael@0	1343	END(pixman_composite_over_8888_8888_asm_mips)
michael@0	1344
michael@0	1345	LEAF_MIPS_DSPR2(pixman_composite_over_n_0565_asm_mips)
michael@0	1346	/*
michael@0	1347	* a0 - dst (r5g6b5)
michael@0	1348	* a1 - src (32bit constant)
michael@0	1349	* a2 - w
michael@0	1350	*/
michael@0	1351
michael@0	1352	beqz a2, 5f
michael@0	1353	nop
michael@0	1354
michael@0	1355	not t0, a1
michael@0	1356	srl t0, t0, 24
michael@0	1357	bgtz t0, 1f
michael@0	1358	nop
michael@0	1359	CONVERT_1x8888_TO_1x0565 a1, t1, t2, t3
michael@0	1360	0:
michael@0	1361	sh t1, 0(a0)
michael@0	1362	addiu a2, a2, -1
michael@0	1363	bgtz a2, 0b
michael@0	1364	addiu a0, a0, 2
michael@0	1365	j ra
michael@0	1366	nop
michael@0	1367
michael@0	1368	1:
michael@0	1369	SAVE_REGS_ON_STACK 0, s0, s1, s2
michael@0	1370	li t4, 0x00ff00ff
michael@0	1371	li t5, 0xf800f800
michael@0	1372	li t6, 0x07e007e0
michael@0	1373	li t7, 0x001F001F
michael@0	1374	addiu t1, a2, -1
michael@0	1375	beqz t1, 3f
michael@0	1376	nop
michael@0	1377	2:
michael@0	1378	lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
michael@0	1379	lhu t2, 2(a0) /* t2 = destination (r5g6b5) */
michael@0	1380
michael@0	1381	CONVERT_2x0565_TO_2x8888 t1, t2, t3, t8, t6, t7, t9, s0, s1, s2
michael@0	1382	MIPS_2xUN8x4_MUL_2xUN8 t3, t8, t0, t0, t1, t2, t4, t9, s0, s1, s2, t3, t8
michael@0	1383	addu_s.qb t1, t1, a1
michael@0	1384	addu_s.qb t2, t2, a1
michael@0	1385	CONVERT_2x8888_TO_2x0565 t1, t2, t3, t8, t5, t6, t7, s0, s1
michael@0	1386
michael@0	1387	sh t3, 0(a0)
michael@0	1388	sh t8, 2(a0)
michael@0	1389
michael@0	1390	addiu a2, a2, -2
michael@0	1391	addiu t1, a2, -1
michael@0	1392	bgtz t1, 2b
michael@0	1393	addiu a0, a0, 4
michael@0	1394	3:
michael@0	1395	beqz a2, 4f
michael@0	1396	nop
michael@0	1397
michael@0	1398	lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
michael@0	1399
michael@0	1400	CONVERT_1x0565_TO_1x8888 t1, t2, s0, s1
michael@0	1401	MIPS_UN8x4_MUL_UN8 t2, t0, t1, t4, s0, s1, s2
michael@0	1402	addu_s.qb t1, t1, a1
michael@0	1403	CONVERT_1x8888_TO_1x0565 t1, t2, s0, s1
michael@0	1404
michael@0	1405	sh t2, 0(a0)
michael@0	1406
michael@0	1407	4:
michael@0	1408	RESTORE_REGS_FROM_STACK 0, s0, s1, s2
michael@0	1409	5:
michael@0	1410	j ra
michael@0	1411	nop
michael@0	1412
michael@0	1413	END(pixman_composite_over_n_0565_asm_mips)
michael@0	1414
michael@0	1415	LEAF_MIPS_DSPR2(pixman_composite_over_n_8888_asm_mips)
michael@0	1416	/*
michael@0	1417	* a0 - dst (a8r8g8b8)
michael@0	1418	* a1 - src (32bit constant)
michael@0	1419	* a2 - w
michael@0	1420	*/
michael@0	1421
michael@0	1422	beqz a2, 5f
michael@0	1423	nop
michael@0	1424
michael@0	1425	not t0, a1
michael@0	1426	srl t0, t0, 24
michael@0	1427	bgtz t0, 1f
michael@0	1428	nop
michael@0	1429	0:
michael@0	1430	sw a1, 0(a0)
michael@0	1431	addiu a2, a2, -1
michael@0	1432	bgtz a2, 0b
michael@0	1433	addiu a0, a0, 4
michael@0	1434	j ra
michael@0	1435	nop
michael@0	1436
michael@0	1437	1:
michael@0	1438	SAVE_REGS_ON_STACK 0, s0, s1, s2
michael@0	1439	li t4, 0x00ff00ff
michael@0	1440	addiu t1, a2, -1
michael@0	1441	beqz t1, 3f
michael@0	1442	nop
michael@0	1443	2:
michael@0	1444	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	1445	lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
michael@0	1446
michael@0	1447	MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t0, t0, t7, t8, t4, t9, s0, s1, s2, t2, t3
michael@0	1448
michael@0	1449	addu_s.qb t7, t7, a1
michael@0	1450	addu_s.qb t8, t8, a1
michael@0	1451
michael@0	1452	sw t7, 0(a0)
michael@0	1453	sw t8, 4(a0)
michael@0	1454
michael@0	1455	addiu a2, a2, -2
michael@0	1456	addiu t1, a2, -1
michael@0	1457	bgtz t1, 2b
michael@0	1458	addiu a0, a0, 8
michael@0	1459	3:
michael@0	1460	beqz a2, 4f
michael@0	1461	nop
michael@0	1462
michael@0	1463	lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
michael@0	1464
michael@0	1465	MIPS_UN8x4_MUL_UN8 t1, t0, t3, t4, t5, t6, t7
michael@0	1466
michael@0	1467	addu_s.qb t3, t3, a1
michael@0	1468
michael@0	1469	sw t3, 0(a0)
michael@0	1470
michael@0	1471	4:
michael@0	1472	RESTORE_REGS_FROM_STACK 0, s0, s1, s2
michael@0	1473	5:
michael@0	1474	j ra
michael@0	1475	nop
michael@0	1476
michael@0	1477	END(pixman_composite_over_n_8888_asm_mips)
michael@0	1478
michael@0	1479	LEAF_MIPS_DSPR2(pixman_composite_add_8_8_8_asm_mips)
michael@0	1480	/*
michael@0	1481	* a0 - dst (a8)
michael@0	1482	* a1 - src (a8)
michael@0	1483	* a2 - mask (a8)
michael@0	1484	* a3 - w
michael@0	1485	*/
michael@0	1486
michael@0	1487	SAVE_REGS_ON_STACK 0, v0, v1
michael@0	1488	li t9, 0x00ff00ff
michael@0	1489	beqz a3, 3f
michael@0	1490	nop
michael@0	1491
michael@0	1492	srl v0, a3, 2 /* v0 = how many multiples of 4 dst pixels */
michael@0	1493	beqz v0, 1f /* branch if less than 4 src pixels */
michael@0	1494	nop
michael@0	1495
michael@0	1496	0:
michael@0	1497	beqz v0, 1f
michael@0	1498	addiu v0, v0, -1
michael@0	1499	lbu t0, 0(a2)
michael@0	1500	lbu t1, 1(a2)
michael@0	1501	lbu t2, 2(a2)
michael@0	1502	lbu t3, 3(a2)
michael@0	1503	lbu t4, 0(a0)
michael@0	1504	lbu t5, 1(a0)
michael@0	1505	lbu t6, 2(a0)
michael@0	1506	lbu t7, 3(a0)
michael@0	1507
michael@0	1508	addiu a2, a2, 4
michael@0	1509
michael@0	1510	precr_sra.ph.w t1, t0, 0
michael@0	1511	precr_sra.ph.w t3, t2, 0
michael@0	1512	precr_sra.ph.w t5, t4, 0
michael@0	1513	precr_sra.ph.w t7, t6, 0
michael@0	1514
michael@0	1515	precr.qb.ph t0, t3, t1
michael@0	1516	precr.qb.ph t1, t7, t5
michael@0	1517
michael@0	1518	lbu t4, 0(a1)
michael@0	1519	lbu v1, 1(a1)
michael@0	1520	lbu t7, 2(a1)
michael@0	1521	lbu t8, 3(a1)
michael@0	1522
michael@0	1523	addiu a1, a1, 4
michael@0	1524
michael@0	1525	precr_sra.ph.w v1, t4, 0
michael@0	1526	precr_sra.ph.w t8, t7, 0
michael@0	1527
michael@0	1528	muleu_s.ph.qbl t2, t0, t8
michael@0	1529	muleu_s.ph.qbr t3, t0, v1
michael@0	1530	shra_r.ph t4, t2, 8
michael@0	1531	shra_r.ph t5, t3, 8
michael@0	1532	and t4, t4, t9
michael@0	1533	and t5, t5, t9
michael@0	1534	addq.ph t2, t2, t4
michael@0	1535	addq.ph t3, t3, t5
michael@0	1536	shra_r.ph t2, t2, 8
michael@0	1537	shra_r.ph t3, t3, 8
michael@0	1538	precr.qb.ph t0, t2, t3
michael@0	1539
michael@0	1540	addu_s.qb t2, t0, t1
michael@0	1541
michael@0	1542	sb t2, 0(a0)
michael@0	1543	srl t2, t2, 8
michael@0	1544	sb t2, 1(a0)
michael@0	1545	srl t2, t2, 8
michael@0	1546	sb t2, 2(a0)
michael@0	1547	srl t2, t2, 8
michael@0	1548	sb t2, 3(a0)
michael@0	1549	addiu a3, a3, -4
michael@0	1550	b 0b
michael@0	1551	addiu a0, a0, 4
michael@0	1552
michael@0	1553	1:
michael@0	1554	beqz a3, 3f
michael@0	1555	nop
michael@0	1556	2:
michael@0	1557	lbu t8, 0(a1)
michael@0	1558	lbu t0, 0(a2)
michael@0	1559	lbu t1, 0(a0)
michael@0	1560	addiu a1, a1, 1
michael@0	1561	addiu a2, a2, 1
michael@0	1562
michael@0	1563	mul t2, t0, t8
michael@0	1564	shra_r.ph t3, t2, 8
michael@0	1565	andi t3, t3, 0xff
michael@0	1566	addq.ph t2, t2, t3
michael@0	1567	shra_r.ph t2, t2, 8
michael@0	1568	andi t2, t2, 0xff
michael@0	1569
michael@0	1570	addu_s.qb t2, t2, t1
michael@0	1571	sb t2, 0(a0)
michael@0	1572	addiu a3, a3, -1
michael@0	1573	bnez a3, 2b
michael@0	1574	addiu a0, a0, 1
michael@0	1575
michael@0	1576	3:
michael@0	1577	RESTORE_REGS_FROM_STACK 0, v0, v1
michael@0	1578	j ra
michael@0	1579	nop
michael@0	1580
michael@0	1581	END(pixman_composite_add_8_8_8_asm_mips)
michael@0	1582
michael@0	1583	LEAF_MIPS_DSPR2(pixman_composite_add_n_8_8_asm_mips)
michael@0	1584	/*
michael@0	1585	* a0 - dst (a8)
michael@0	1586	* a1 - src (32bit constant)
michael@0	1587	* a2 - mask (a8)
michael@0	1588	* a3 - w
michael@0	1589	*/
michael@0	1590
michael@0	1591	SAVE_REGS_ON_STACK 0, v0
michael@0	1592	li t9, 0x00ff00ff
michael@0	1593	beqz a3, 3f
michael@0	1594	nop
michael@0	1595
michael@0	1596	srl v0, a3, 2 /* v0 = how many multiples of 4 dst pixels */
michael@0	1597	beqz v0, 1f /* branch if less than 4 src pixels */
michael@0	1598	nop
michael@0	1599
michael@0	1600	srl t8, a1, 24
michael@0	1601	replv.ph t8, t8
michael@0	1602
michael@0	1603	0:
michael@0	1604	beqz v0, 1f
michael@0	1605	addiu v0, v0, -1
michael@0	1606	lbu t0, 0(a2)
michael@0	1607	lbu t1, 1(a2)
michael@0	1608	lbu t2, 2(a2)
michael@0	1609	lbu t3, 3(a2)
michael@0	1610	lbu t4, 0(a0)
michael@0	1611	lbu t5, 1(a0)
michael@0	1612	lbu t6, 2(a0)
michael@0	1613	lbu t7, 3(a0)
michael@0	1614
michael@0	1615	addiu a2, a2, 4
michael@0	1616
michael@0	1617	precr_sra.ph.w t1, t0, 0
michael@0	1618	precr_sra.ph.w t3, t2, 0
michael@0	1619	precr_sra.ph.w t5, t4, 0
michael@0	1620	precr_sra.ph.w t7, t6, 0
michael@0	1621
michael@0	1622	precr.qb.ph t0, t3, t1
michael@0	1623	precr.qb.ph t1, t7, t5
michael@0	1624
michael@0	1625	muleu_s.ph.qbl t2, t0, t8
michael@0	1626	muleu_s.ph.qbr t3, t0, t8
michael@0	1627	shra_r.ph t4, t2, 8
michael@0	1628	shra_r.ph t5, t3, 8
michael@0	1629	and t4, t4, t9
michael@0	1630	and t5, t5, t9
michael@0	1631	addq.ph t2, t2, t4
michael@0	1632	addq.ph t3, t3, t5
michael@0	1633	shra_r.ph t2, t2, 8
michael@0	1634	shra_r.ph t3, t3, 8
michael@0	1635	precr.qb.ph t0, t2, t3
michael@0	1636
michael@0	1637	addu_s.qb t2, t0, t1
michael@0	1638
michael@0	1639	sb t2, 0(a0)
michael@0	1640	srl t2, t2, 8
michael@0	1641	sb t2, 1(a0)
michael@0	1642	srl t2, t2, 8
michael@0	1643	sb t2, 2(a0)
michael@0	1644	srl t2, t2, 8
michael@0	1645	sb t2, 3(a0)
michael@0	1646	addiu a3, a3, -4
michael@0	1647	b 0b
michael@0	1648	addiu a0, a0, 4
michael@0	1649
michael@0	1650	1:
michael@0	1651	beqz a3, 3f
michael@0	1652	nop
michael@0	1653	srl t8, a1, 24
michael@0	1654	2:
michael@0	1655	lbu t0, 0(a2)
michael@0	1656	lbu t1, 0(a0)
michael@0	1657	addiu a2, a2, 1
michael@0	1658
michael@0	1659	mul t2, t0, t8
michael@0	1660	shra_r.ph t3, t2, 8
michael@0	1661	andi t3, t3, 0xff
michael@0	1662	addq.ph t2, t2, t3
michael@0	1663	shra_r.ph t2, t2, 8
michael@0	1664	andi t2, t2, 0xff
michael@0	1665
michael@0	1666	addu_s.qb t2, t2, t1
michael@0	1667	sb t2, 0(a0)
michael@0	1668	addiu a3, a3, -1
michael@0	1669	bnez a3, 2b
michael@0	1670	addiu a0, a0, 1
michael@0	1671
michael@0	1672	3:
michael@0	1673	RESTORE_REGS_FROM_STACK 0, v0
michael@0	1674	j ra
michael@0	1675	nop
michael@0	1676
michael@0	1677	END(pixman_composite_add_n_8_8_asm_mips)
michael@0	1678
michael@0	1679	LEAF_MIPS_DSPR2(pixman_composite_add_n_8_8888_asm_mips)
michael@0	1680	/*
michael@0	1681	* a0 - dst (a8r8g8b8)
michael@0	1682	* a1 - src (32bit constant)
michael@0	1683	* a2 - mask (a8)
michael@0	1684	* a3 - w
michael@0	1685	*/
michael@0	1686
michael@0	1687	SAVE_REGS_ON_STACK 0, s0, s1, s2
michael@0	1688	li t4, 0x00ff00ff
michael@0	1689	beqz a3, 3f
michael@0	1690	nop
michael@0	1691	addiu t1, a3, -1
michael@0	1692	beqz t1, 2f
michael@0	1693	nop
michael@0	1694	1:
michael@0	1695	/* a1 = source (32bit constant) */
michael@0	1696	lbu t0, 0(a2) /* t0 = mask (a8) */
michael@0	1697	lbu t1, 1(a2) /* t1 = mask (a8) */
michael@0	1698	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	1699	lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
michael@0	1700	addiu a2, a2, 2
michael@0	1701
michael@0	1702	MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 a1, a1, \
michael@0	1703	t0, t1, \
michael@0	1704	t2, t3, \
michael@0	1705	t5, t6, \
michael@0	1706	t4, t7, t8, t9, s0, s1, s2
michael@0	1707
michael@0	1708	sw t5, 0(a0)
michael@0	1709	sw t6, 4(a0)
michael@0	1710	addiu a3, a3, -2
michael@0	1711	addiu t1, a3, -1
michael@0	1712	bgtz t1, 1b
michael@0	1713	addiu a0, a0, 8
michael@0	1714	2:
michael@0	1715	beqz a3, 3f
michael@0	1716	nop
michael@0	1717	/* a1 = source (32bit constant) */
michael@0	1718	lbu t0, 0(a2) /* t0 = mask (a8) */
michael@0	1719	lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
michael@0	1720
michael@0	1721	MIPS_UN8x4_MUL_UN8_ADD_UN8x4 a1, t0, t1, t2, t4, t3, t5, t6
michael@0	1722
michael@0	1723	sw t2, 0(a0)
michael@0	1724	3:
michael@0	1725	RESTORE_REGS_FROM_STACK 0, s0, s1, s2
michael@0	1726	j ra
michael@0	1727	nop
michael@0	1728
michael@0	1729	END(pixman_composite_add_n_8_8888_asm_mips)
michael@0	1730
michael@0	1731	LEAF_MIPS_DSPR2(pixman_composite_add_0565_8_0565_asm_mips)
michael@0	1732	/*
michael@0	1733	* a0 - dst (r5g6b5)
michael@0	1734	* a1 - src (r5g6b5)
michael@0	1735	* a2 - mask (a8)
michael@0	1736	* a3 - w
michael@0	1737	*/
michael@0	1738
michael@0	1739	SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	1740	li t4, 0xf800f800
michael@0	1741	li t5, 0x07e007e0
michael@0	1742	li t6, 0x001F001F
michael@0	1743	li t7, 0x00ff00ff
michael@0	1744	beqz a3, 3f
michael@0	1745	nop
michael@0	1746	addiu t1, a3, -1
michael@0	1747	beqz t1, 2f
michael@0	1748	nop
michael@0	1749	1:
michael@0	1750	lhu t0, 0(a1) /* t0 = source (r5g6b5) */
michael@0	1751	lhu t1, 2(a1) /* t1 = source (r5g6b5) */
michael@0	1752	lbu t2, 0(a2) /* t2 = mask (a8) */
michael@0	1753	lbu t3, 1(a2) /* t3 = mask (a8) */
michael@0	1754	lhu t8, 0(a0) /* t8 = destination (r5g6b5) */
michael@0	1755	lhu t9, 2(a0) /* t9 = destination (r5g6b5) */
michael@0	1756	addiu a1, a1, 4
michael@0	1757	addiu a2, a2, 2
michael@0	1758
michael@0	1759	CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5
michael@0	1760	CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, s6, s7
michael@0	1761	MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 s0, s1, \
michael@0	1762	t2, t3, \
michael@0	1763	s2, s3, \
michael@0	1764	t0, t1, \
michael@0	1765	t7, s4, s5, s6, s7, t8, t9
michael@0	1766	CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3
michael@0	1767
michael@0	1768	sh s0, 0(a0)
michael@0	1769	sh s1, 2(a0)
michael@0	1770	addiu a3, a3, -2
michael@0	1771	addiu t1, a3, -1
michael@0	1772	bgtz t1, 1b
michael@0	1773	addiu a0, a0, 4
michael@0	1774	2:
michael@0	1775	beqz a3, 3f
michael@0	1776	nop
michael@0	1777	lhu t0, 0(a1) /* t0 = source (r5g6b5) */
michael@0	1778	lbu t1, 0(a2) /* t1 = mask (a8) */
michael@0	1779	lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
michael@0	1780
michael@0	1781	CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5
michael@0	1782	CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6
michael@0	1783	MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t3, t1, t4, t0, t7, t2, t5, t6
michael@0	1784	CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5
michael@0	1785
michael@0	1786	sh t3, 0(a0)
michael@0	1787	3:
michael@0	1788	RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	1789	j ra
michael@0	1790	nop
michael@0	1791
michael@0	1792	END(pixman_composite_add_0565_8_0565_asm_mips)
michael@0	1793
michael@0	1794	LEAF_MIPS_DSPR2(pixman_composite_add_8888_8_8888_asm_mips)
michael@0	1795	/*
michael@0	1796	* a0 - dst (a8r8g8b8)
michael@0	1797	* a1 - src (a8r8g8b8)
michael@0	1798	* a2 - mask (a8)
michael@0	1799	* a3 - w
michael@0	1800	*/
michael@0	1801
michael@0	1802	SAVE_REGS_ON_STACK 0, s0, s1, s2
michael@0	1803	li t4, 0x00ff00ff
michael@0	1804	beqz a3, 3f
michael@0	1805	nop
michael@0	1806	addiu t1, a3, -1
michael@0	1807	beqz t1, 2f
michael@0	1808	nop
michael@0	1809	1:
michael@0	1810	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1811	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	1812	lbu t2, 0(a2) /* t2 = mask (a8) */
michael@0	1813	lbu t3, 1(a2) /* t3 = mask (a8) */
michael@0	1814	lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */
michael@0	1815	lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */
michael@0	1816	addiu a1, a1, 8
michael@0	1817	addiu a2, a2, 2
michael@0	1818
michael@0	1819	MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \
michael@0	1820	t2, t3, \
michael@0	1821	t5, t6, \
michael@0	1822	t7, t8, \
michael@0	1823	t4, t9, s0, s1, s2, t0, t1
michael@0	1824
michael@0	1825	sw t7, 0(a0)
michael@0	1826	sw t8, 4(a0)
michael@0	1827	addiu a3, a3, -2
michael@0	1828	addiu t1, a3, -1
michael@0	1829	bgtz t1, 1b
michael@0	1830	addiu a0, a0, 8
michael@0	1831	2:
michael@0	1832	beqz a3, 3f
michael@0	1833	nop
michael@0	1834	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1835	lbu t1, 0(a2) /* t1 = mask (a8) */
michael@0	1836	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	1837
michael@0	1838	MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t3, t4, t5, t6, t7
michael@0	1839
michael@0	1840	sw t3, 0(a0)
michael@0	1841	3:
michael@0	1842	RESTORE_REGS_FROM_STACK 0, s0, s1, s2
michael@0	1843	j ra
michael@0	1844	nop
michael@0	1845
michael@0	1846	END(pixman_composite_add_8888_8_8888_asm_mips)
michael@0	1847
michael@0	1848	LEAF_MIPS_DSPR2(pixman_composite_add_8888_n_8888_asm_mips)
michael@0	1849	/*
michael@0	1850	* a0 - dst (a8r8g8b8)
michael@0	1851	* a1 - src (a8r8g8b8)
michael@0	1852	* a2 - mask (32bit constant)
michael@0	1853	* a3 - w
michael@0	1854	*/
michael@0	1855
michael@0	1856	SAVE_REGS_ON_STACK 0, s0, s1, s2
michael@0	1857	li t4, 0x00ff00ff
michael@0	1858	beqz a3, 3f
michael@0	1859	nop
michael@0	1860	srl a2, a2, 24
michael@0	1861	addiu t1, a3, -1
michael@0	1862	beqz t1, 2f
michael@0	1863	nop
michael@0	1864	1:
michael@0	1865	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1866	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	1867	/* a2 = mask (32bit constant) */
michael@0	1868	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	1869	lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
michael@0	1870	addiu a1, a1, 8
michael@0	1871
michael@0	1872	MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \
michael@0	1873	a2, a2, \
michael@0	1874	t2, t3, \
michael@0	1875	t5, t6, \
michael@0	1876	t4, t7, t8, t9, s0, s1, s2
michael@0	1877
michael@0	1878	sw t5, 0(a0)
michael@0	1879	sw t6, 4(a0)
michael@0	1880	addiu a3, a3, -2
michael@0	1881	addiu t1, a3, -1
michael@0	1882	bgtz t1, 1b
michael@0	1883	addiu a0, a0, 8
michael@0	1884	2:
michael@0	1885	beqz a3, 3f
michael@0	1886	nop
michael@0	1887	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1888	/* a2 = mask (32bit constant) */
michael@0	1889	lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
michael@0	1890
michael@0	1891	MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, a2, t1, t3, t4, t5, t6, t7
michael@0	1892
michael@0	1893	sw t3, 0(a0)
michael@0	1894	3:
michael@0	1895	RESTORE_REGS_FROM_STACK 0, s0, s1, s2
michael@0	1896	j ra
michael@0	1897	nop
michael@0	1898
michael@0	1899	END(pixman_composite_add_8888_n_8888_asm_mips)
michael@0	1900
michael@0	1901	LEAF_MIPS_DSPR2(pixman_composite_add_8888_8888_8888_asm_mips)
michael@0	1902	/*
michael@0	1903	* a0 - dst (a8r8g8b8)
michael@0	1904	* a1 - src (a8r8g8b8)
michael@0	1905	* a2 - mask (a8r8g8b8)
michael@0	1906	* a3 - w
michael@0	1907	*/
michael@0	1908
michael@0	1909	SAVE_REGS_ON_STACK 0, s0, s1, s2
michael@0	1910	li t4, 0x00ff00ff
michael@0	1911	beqz a3, 3f
michael@0	1912	nop
michael@0	1913	addiu t1, a3, -1
michael@0	1914	beqz t1, 2f
michael@0	1915	nop
michael@0	1916	1:
michael@0	1917	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1918	lw t1, 4(a1) /* t1 = source (a8r8g8b8) */
michael@0	1919	lw t2, 0(a2) /* t2 = mask (a8r8g8b8) */
michael@0	1920	lw t3, 4(a2) /* t3 = mask (a8r8g8b8) */
michael@0	1921	lw t5, 0(a0) /* t5 = destination (a8r8g8b8) */
michael@0	1922	lw t6, 4(a0) /* t6 = destination (a8r8g8b8) */
michael@0	1923	addiu a1, a1, 8
michael@0	1924	addiu a2, a2, 8
michael@0	1925	srl t2, t2, 24
michael@0	1926	srl t3, t3, 24
michael@0	1927
michael@0	1928	MIPS_2xUN8x4_MUL_2xUN8_ADD_2xUN8x4 t0, t1, \
michael@0	1929	t2, t3, \
michael@0	1930	t5, t6, \
michael@0	1931	t7, t8, \
michael@0	1932	t4, t9, s0, s1, s2, t0, t1
michael@0	1933
michael@0	1934	sw t7, 0(a0)
michael@0	1935	sw t8, 4(a0)
michael@0	1936	addiu a3, a3, -2
michael@0	1937	addiu t1, a3, -1
michael@0	1938	bgtz t1, 1b
michael@0	1939	addiu a0, a0, 8
michael@0	1940	2:
michael@0	1941	beqz a3, 3f
michael@0	1942	nop
michael@0	1943	lw t0, 0(a1) /* t0 = source (a8r8g8b8) */
michael@0	1944	lw t1, 0(a2) /* t1 = mask (a8r8g8b8) */
michael@0	1945	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	1946	srl t1, t1, 24
michael@0	1947
michael@0	1948	MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t3, t4, t5, t6, t7
michael@0	1949
michael@0	1950	sw t3, 0(a0)
michael@0	1951	3:
michael@0	1952	RESTORE_REGS_FROM_STACK 0, s0, s1, s2
michael@0	1953	j ra
michael@0	1954	nop
michael@0	1955
michael@0	1956	END(pixman_composite_add_8888_8888_8888_asm_mips)
michael@0	1957
michael@0	1958	LEAF_MIPS_DSPR2(pixman_composite_add_8_8_asm_mips)
michael@0	1959	/*
michael@0	1960	* a0 - dst (a8)
michael@0	1961	* a1 - src (a8)
michael@0	1962	* a2 - w
michael@0	1963	*/
michael@0	1964
michael@0	1965	beqz a2, 3f
michael@0	1966	nop
michael@0	1967	srl t9, a2, 2 /* t9 = how many multiples of 4 dst pixels */
michael@0	1968	beqz t9, 1f /* branch if less than 4 src pixels */
michael@0	1969	nop
michael@0	1970
michael@0	1971	0:
michael@0	1972	beqz t9, 1f
michael@0	1973	addiu t9, t9, -1
michael@0	1974	lbu t0, 0(a1)
michael@0	1975	lbu t1, 1(a1)
michael@0	1976	lbu t2, 2(a1)
michael@0	1977	lbu t3, 3(a1)
michael@0	1978	lbu t4, 0(a0)
michael@0	1979	lbu t5, 1(a0)
michael@0	1980	lbu t6, 2(a0)
michael@0	1981	lbu t7, 3(a0)
michael@0	1982
michael@0	1983	addiu a1, a1, 4
michael@0	1984
michael@0	1985	precr_sra.ph.w t1, t0, 0
michael@0	1986	precr_sra.ph.w t3, t2, 0
michael@0	1987	precr_sra.ph.w t5, t4, 0
michael@0	1988	precr_sra.ph.w t7, t6, 0
michael@0	1989
michael@0	1990	precr.qb.ph t0, t3, t1
michael@0	1991	precr.qb.ph t1, t7, t5
michael@0	1992
michael@0	1993	addu_s.qb t2, t0, t1
michael@0	1994
michael@0	1995	sb t2, 0(a0)
michael@0	1996	srl t2, t2, 8
michael@0	1997	sb t2, 1(a0)
michael@0	1998	srl t2, t2, 8
michael@0	1999	sb t2, 2(a0)
michael@0	2000	srl t2, t2, 8
michael@0	2001	sb t2, 3(a0)
michael@0	2002	addiu a2, a2, -4
michael@0	2003	b 0b
michael@0	2004	addiu a0, a0, 4
michael@0	2005
michael@0	2006	1:
michael@0	2007	beqz a2, 3f
michael@0	2008	nop
michael@0	2009	2:
michael@0	2010	lbu t0, 0(a1)
michael@0	2011	lbu t1, 0(a0)
michael@0	2012	addiu a1, a1, 1
michael@0	2013
michael@0	2014	addu_s.qb t2, t0, t1
michael@0	2015	sb t2, 0(a0)
michael@0	2016	addiu a2, a2, -1
michael@0	2017	bnez a2, 2b
michael@0	2018	addiu a0, a0, 1
michael@0	2019
michael@0	2020	3:
michael@0	2021	j ra
michael@0	2022	nop
michael@0	2023
michael@0	2024	END(pixman_composite_add_8_8_asm_mips)
michael@0	2025
michael@0	2026	LEAF_MIPS_DSPR2(pixman_composite_add_8888_8888_asm_mips)
michael@0	2027	/*
michael@0	2028	* a0 - dst (a8r8g8b8)
michael@0	2029	* a1 - src (a8r8g8b8)
michael@0	2030	* a2 - w
michael@0	2031	*/
michael@0	2032
michael@0	2033	beqz a2, 4f
michael@0	2034	nop
michael@0	2035
michael@0	2036	srl t9, a2, 2 /* t1 = how many multiples of 4 src pixels */
michael@0	2037	beqz t9, 3f /* branch if less than 4 src pixels */
michael@0	2038	nop
michael@0	2039	1:
michael@0	2040	addiu t9, t9, -1
michael@0	2041	beqz t9, 2f
michael@0	2042	addiu a2, a2, -4
michael@0	2043
michael@0	2044	lw t0, 0(a1)
michael@0	2045	lw t1, 4(a1)
michael@0	2046	lw t2, 8(a1)
michael@0	2047	lw t3, 12(a1)
michael@0	2048	lw t4, 0(a0)
michael@0	2049	lw t5, 4(a0)
michael@0	2050	lw t6, 8(a0)
michael@0	2051	lw t7, 12(a0)
michael@0	2052	addiu a1, a1, 16
michael@0	2053
michael@0	2054	addu_s.qb t4, t4, t0
michael@0	2055	addu_s.qb t5, t5, t1
michael@0	2056	addu_s.qb t6, t6, t2
michael@0	2057	addu_s.qb t7, t7, t3
michael@0	2058
michael@0	2059	sw t4, 0(a0)
michael@0	2060	sw t5, 4(a0)
michael@0	2061	sw t6, 8(a0)
michael@0	2062	sw t7, 12(a0)
michael@0	2063	b 1b
michael@0	2064	addiu a0, a0, 16
michael@0	2065	2:
michael@0	2066	lw t0, 0(a1)
michael@0	2067	lw t1, 4(a1)
michael@0	2068	lw t2, 8(a1)
michael@0	2069	lw t3, 12(a1)
michael@0	2070	lw t4, 0(a0)
michael@0	2071	lw t5, 4(a0)
michael@0	2072	lw t6, 8(a0)
michael@0	2073	lw t7, 12(a0)
michael@0	2074	addiu a1, a1, 16
michael@0	2075
michael@0	2076	addu_s.qb t4, t4, t0
michael@0	2077	addu_s.qb t5, t5, t1
michael@0	2078	addu_s.qb t6, t6, t2
michael@0	2079	addu_s.qb t7, t7, t3
michael@0	2080
michael@0	2081	sw t4, 0(a0)
michael@0	2082	sw t5, 4(a0)
michael@0	2083	sw t6, 8(a0)
michael@0	2084	sw t7, 12(a0)
michael@0	2085
michael@0	2086	beqz a2, 4f
michael@0	2087	addiu a0, a0, 16
michael@0	2088	3:
michael@0	2089	lw t0, 0(a1)
michael@0	2090	lw t1, 0(a0)
michael@0	2091	addiu a1, a1, 4
michael@0	2092	addiu a2, a2, -1
michael@0	2093	addu_s.qb t1, t1, t0
michael@0	2094	sw t1, 0(a0)
michael@0	2095	bnez a2, 3b
michael@0	2096	addiu a0, a0, 4
michael@0	2097	4:
michael@0	2098	jr ra
michael@0	2099	nop
michael@0	2100
michael@0	2101	END(pixman_composite_add_8888_8888_asm_mips)
michael@0	2102
michael@0	2103	LEAF_MIPS_DSPR2(pixman_composite_out_reverse_8_0565_asm_mips)
michael@0	2104	/*
michael@0	2105	* a0 - dst (r5g6b5)
michael@0	2106	* a1 - src (a8)
michael@0	2107	* a2 - w
michael@0	2108	*/
michael@0	2109
michael@0	2110	beqz a2, 4f
michael@0	2111	nop
michael@0	2112
michael@0	2113	SAVE_REGS_ON_STACK 0, s0, s1, s2, s3
michael@0	2114	li t2, 0xf800f800
michael@0	2115	li t3, 0x07e007e0
michael@0	2116	li t4, 0x001F001F
michael@0	2117	li t5, 0x00ff00ff
michael@0	2118
michael@0	2119	addiu t1, a2, -1
michael@0	2120	beqz t1, 2f
michael@0	2121	nop
michael@0	2122	1:
michael@0	2123	lbu t0, 0(a1) /* t0 = source (a8) */
michael@0	2124	lbu t1, 1(a1) /* t1 = source (a8) */
michael@0	2125	lhu t6, 0(a0) /* t6 = destination (r5g6b5) */
michael@0	2126	lhu t7, 2(a0) /* t7 = destination (r5g6b5) */
michael@0	2127	addiu a1, a1, 2
michael@0	2128
michael@0	2129	not t0, t0
michael@0	2130	not t1, t1
michael@0	2131	andi t0, 0xff /* t0 = neg source1 */
michael@0	2132	andi t1, 0xff /* t1 = neg source2 */
michael@0	2133	CONVERT_2x0565_TO_2x8888 t6, t7, t8, t9, t3, t4, s0, s1, s2, s3
michael@0	2134	MIPS_2xUN8x4_MUL_2xUN8 t8, t9, t0, t1, t6, t7, t5, s0, s1, s2, s3, t8, t9
michael@0	2135	CONVERT_2x8888_TO_2x0565 t6, t7, t8, t9, t2, t3, t4, s0, s1
michael@0	2136
michael@0	2137	sh t8, 0(a0)
michael@0	2138	sh t9, 2(a0)
michael@0	2139	addiu a2, a2, -2
michael@0	2140	addiu t1, a2, -1
michael@0	2141	bgtz t1, 1b
michael@0	2142	addiu a0, a0, 4
michael@0	2143	2:
michael@0	2144	beqz a2, 3f
michael@0	2145	nop
michael@0	2146	lbu t0, 0(a1) /* t0 = source (a8) */
michael@0	2147	lhu t1, 0(a0) /* t1 = destination (r5g6b5) */
michael@0	2148
michael@0	2149	not t0, t0
michael@0	2150	andi t0, 0xff /* t0 = neg source */
michael@0	2151	CONVERT_1x0565_TO_1x8888 t1, t2, t3, t4
michael@0	2152	MIPS_UN8x4_MUL_UN8 t2, t0, t1, t5, t3, t4, t6
michael@0	2153	CONVERT_1x8888_TO_1x0565 t1, t2, t3, t4
michael@0	2154
michael@0	2155	sh t2, 0(a0)
michael@0	2156	3:
michael@0	2157	RESTORE_REGS_FROM_STACK 0, s0, s1, s2, s3
michael@0	2158	4:
michael@0	2159	j ra
michael@0	2160	nop
michael@0	2161
michael@0	2162	END(pixman_composite_out_reverse_8_0565_asm_mips)
michael@0	2163
michael@0	2164	LEAF_MIPS_DSPR2(pixman_composite_out_reverse_8_8888_asm_mips)
michael@0	2165	/*
michael@0	2166	* a0 - dst (a8r8g8b8)
michael@0	2167	* a1 - src (a8)
michael@0	2168	* a2 - w
michael@0	2169	*/
michael@0	2170
michael@0	2171	beqz a2, 3f
michael@0	2172	nop
michael@0	2173	li t4, 0x00ff00ff
michael@0	2174	addiu t1, a2, -1
michael@0	2175	beqz t1, 2f
michael@0	2176	nop
michael@0	2177	1:
michael@0	2178	lbu t0, 0(a1) /* t0 = source (a8) */
michael@0	2179	lbu t1, 1(a1) /* t1 = source (a8) */
michael@0	2180	lw t2, 0(a0) /* t2 = destination (a8r8g8b8) */
michael@0	2181	lw t3, 4(a0) /* t3 = destination (a8r8g8b8) */
michael@0	2182	addiu a1, a1, 2
michael@0	2183	not t0, t0
michael@0	2184	not t1, t1
michael@0	2185	andi t0, 0xff /* t0 = neg source */
michael@0	2186	andi t1, 0xff /* t1 = neg source */
michael@0	2187
michael@0	2188	MIPS_2xUN8x4_MUL_2xUN8 t2, t3, t0, t1, t5, t6, t4, t7, t8, t9, t2, t3, t0
michael@0	2189
michael@0	2190	sw t5, 0(a0)
michael@0	2191	sw t6, 4(a0)
michael@0	2192	addiu a2, a2, -2
michael@0	2193	addiu t1, a2, -1
michael@0	2194	bgtz t1, 1b
michael@0	2195	addiu a0, a0, 8
michael@0	2196	2:
michael@0	2197	beqz a2, 3f
michael@0	2198	nop
michael@0	2199	lbu t0, 0(a1) /* t0 = source (a8) */
michael@0	2200	lw t1, 0(a0) /* t1 = destination (a8r8g8b8) */
michael@0	2201	not t0, t0
michael@0	2202	andi t0, 0xff /* t0 = neg source */
michael@0	2203
michael@0	2204	MIPS_UN8x4_MUL_UN8 t1, t0, t2, t4, t3, t5, t6
michael@0	2205
michael@0	2206	sw t2, 0(a0)
michael@0	2207	3:
michael@0	2208	j ra
michael@0	2209	nop
michael@0	2210
michael@0	2211	END(pixman_composite_out_reverse_8_8888_asm_mips)
michael@0	2212
michael@0	2213	LEAF_MIPS_DSPR2(pixman_composite_over_reverse_n_8888_asm_mips)
michael@0	2214	/*
michael@0	2215	* a0 - dst (a8r8g8b8)
michael@0	2216	* a1 - src (32bit constant)
michael@0	2217	* a2 - w
michael@0	2218	*/
michael@0	2219
michael@0	2220	beqz a2, 5f
michael@0	2221	nop
michael@0	2222
michael@0	2223	SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2224	li t0, 0x00ff00ff
michael@0	2225	srl t9, a2, 2 /* t9 = how many multiples of 4 src pixels */
michael@0	2226	beqz t9, 2f /* branch if less than 4 src pixels */
michael@0	2227	nop
michael@0	2228	1:
michael@0	2229	beqz t9, 2f
michael@0	2230	addiu t9, t9, -1
michael@0	2231
michael@0	2232	lw t1, 0(a0)
michael@0	2233	lw t2, 4(a0)
michael@0	2234	lw t3, 8(a0)
michael@0	2235	lw t4, 12(a0)
michael@0	2236
michael@0	2237	addiu a2, a2, -4
michael@0	2238
michael@0	2239	not t5, t1
michael@0	2240	not t6, t2
michael@0	2241	not t7, t3
michael@0	2242	not t8, t4
michael@0	2243	srl t5, t5, 24
michael@0	2244	srl t6, t6, 24
michael@0	2245	srl t7, t7, 24
michael@0	2246	srl t8, t8, 24
michael@0	2247	replv.ph t5, t5
michael@0	2248	replv.ph t6, t6
michael@0	2249	replv.ph t7, t7
michael@0	2250	replv.ph t8, t8
michael@0	2251	muleu_s.ph.qbl s0, a1, t5
michael@0	2252	muleu_s.ph.qbr s1, a1, t5
michael@0	2253	muleu_s.ph.qbl s2, a1, t6
michael@0	2254	muleu_s.ph.qbr s3, a1, t6
michael@0	2255	muleu_s.ph.qbl s4, a1, t7
michael@0	2256	muleu_s.ph.qbr s5, a1, t7
michael@0	2257	muleu_s.ph.qbl s6, a1, t8
michael@0	2258	muleu_s.ph.qbr s7, a1, t8
michael@0	2259
michael@0	2260	shra_r.ph t5, s0, 8
michael@0	2261	shra_r.ph t6, s1, 8
michael@0	2262	shra_r.ph t7, s2, 8
michael@0	2263	shra_r.ph t8, s3, 8
michael@0	2264	and t5, t5, t0
michael@0	2265	and t6, t6, t0
michael@0	2266	and t7, t7, t0
michael@0	2267	and t8, t8, t0
michael@0	2268	addq.ph s0, s0, t5
michael@0	2269	addq.ph s1, s1, t6
michael@0	2270	addq.ph s2, s2, t7
michael@0	2271	addq.ph s3, s3, t8
michael@0	2272	shra_r.ph s0, s0, 8
michael@0	2273	shra_r.ph s1, s1, 8
michael@0	2274	shra_r.ph s2, s2, 8
michael@0	2275	shra_r.ph s3, s3, 8
michael@0	2276	shra_r.ph t5, s4, 8
michael@0	2277	shra_r.ph t6, s5, 8
michael@0	2278	shra_r.ph t7, s6, 8
michael@0	2279	shra_r.ph t8, s7, 8
michael@0	2280	and t5, t5, t0
michael@0	2281	and t6, t6, t0
michael@0	2282	and t7, t7, t0
michael@0	2283	and t8, t8, t0
michael@0	2284	addq.ph s4, s4, t5
michael@0	2285	addq.ph s5, s5, t6
michael@0	2286	addq.ph s6, s6, t7
michael@0	2287	addq.ph s7, s7, t8
michael@0	2288	shra_r.ph s4, s4, 8
michael@0	2289	shra_r.ph s5, s5, 8
michael@0	2290	shra_r.ph s6, s6, 8
michael@0	2291	shra_r.ph s7, s7, 8
michael@0	2292
michael@0	2293	precr.qb.ph t5, s0, s1
michael@0	2294	precr.qb.ph t6, s2, s3
michael@0	2295	precr.qb.ph t7, s4, s5
michael@0	2296	precr.qb.ph t8, s6, s7
michael@0	2297	addu_s.qb t5, t1, t5
michael@0	2298	addu_s.qb t6, t2, t6
michael@0	2299	addu_s.qb t7, t3, t7
michael@0	2300	addu_s.qb t8, t4, t8
michael@0	2301
michael@0	2302	sw t5, 0(a0)
michael@0	2303	sw t6, 4(a0)
michael@0	2304	sw t7, 8(a0)
michael@0	2305	sw t8, 12(a0)
michael@0	2306	b 1b
michael@0	2307	addiu a0, a0, 16
michael@0	2308
michael@0	2309	2:
michael@0	2310	beqz a2, 4f
michael@0	2311	nop
michael@0	2312	3:
michael@0	2313	lw t1, 0(a0)
michael@0	2314
michael@0	2315	not t2, t1
michael@0	2316	srl t2, t2, 24
michael@0	2317	replv.ph t2, t2
michael@0	2318
michael@0	2319	muleu_s.ph.qbl t4, a1, t2
michael@0	2320	muleu_s.ph.qbr t5, a1, t2
michael@0	2321	shra_r.ph t6, t4, 8
michael@0	2322	shra_r.ph t7, t5, 8
michael@0	2323
michael@0	2324	and t6,t6,t0
michael@0	2325	and t7,t7,t0
michael@0	2326
michael@0	2327	addq.ph t8, t4, t6
michael@0	2328	addq.ph t9, t5, t7
michael@0	2329
michael@0	2330	shra_r.ph t8, t8, 8
michael@0	2331	shra_r.ph t9, t9, 8
michael@0	2332
michael@0	2333	precr.qb.ph t9, t8, t9
michael@0	2334
michael@0	2335	addu_s.qb t9, t1, t9
michael@0	2336	sw t9, 0(a0)
michael@0	2337
michael@0	2338	addiu a2, a2, -1
michael@0	2339	bnez a2, 3b
michael@0	2340	addiu a0, a0, 4
michael@0	2341	4:
michael@0	2342	RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2343	5:
michael@0	2344	j ra
michael@0	2345	nop
michael@0	2346
michael@0	2347	END(pixman_composite_over_reverse_n_8888_asm_mips)
michael@0	2348
michael@0	2349	LEAF_MIPS_DSPR2(pixman_composite_in_n_8_asm_mips)
michael@0	2350	/*
michael@0	2351	* a0 - dst (a8)
michael@0	2352	* a1 - src (a8r8g8b8)
michael@0	2353	* a2 - w
michael@0	2354	*/
michael@0	2355
michael@0	2356	beqz a2, 5f
michael@0	2357	nop
michael@0	2358
michael@0	2359	SAVE_REGS_ON_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2360	move t7, a1
michael@0	2361	srl t5, t7, 24
michael@0	2362	replv.ph t5, t5
michael@0	2363	srl t9, a2, 2 /* t1 = how many multiples of 4 src pixels */
michael@0	2364	beqz t9, 2f /* branch if less than 4 src pixels */
michael@0	2365	nop
michael@0	2366
michael@0	2367	1:
michael@0	2368	addiu t9, t9, -1
michael@0	2369	addiu a2, a2, -4
michael@0	2370	lbu t0, 0(a0)
michael@0	2371	lbu t1, 1(a0)
michael@0	2372	lbu t2, 2(a0)
michael@0	2373	lbu t3, 3(a0)
michael@0	2374
michael@0	2375	muleu_s.ph.qbl s0, t0, t5
michael@0	2376	muleu_s.ph.qbr s1, t0, t5
michael@0	2377	muleu_s.ph.qbl s2, t1, t5
michael@0	2378	muleu_s.ph.qbr s3, t1, t5
michael@0	2379	muleu_s.ph.qbl s4, t2, t5
michael@0	2380	muleu_s.ph.qbr s5, t2, t5
michael@0	2381	muleu_s.ph.qbl s6, t3, t5
michael@0	2382	muleu_s.ph.qbr s7, t3, t5
michael@0	2383
michael@0	2384	shrl.ph t4, s0, 8
michael@0	2385	shrl.ph t6, s1, 8
michael@0	2386	shrl.ph t7, s2, 8
michael@0	2387	shrl.ph t8, s3, 8
michael@0	2388	addq.ph t0, s0, t4
michael@0	2389	addq.ph t1, s1, t6
michael@0	2390	addq.ph t2, s2, t7
michael@0	2391	addq.ph t3, s3, t8
michael@0	2392	shra_r.ph t0, t0, 8
michael@0	2393	shra_r.ph t1, t1, 8
michael@0	2394	shra_r.ph t2, t2, 8
michael@0	2395	shra_r.ph t3, t3, 8
michael@0	2396	shrl.ph t4, s4, 8
michael@0	2397	shrl.ph t6, s5, 8
michael@0	2398	shrl.ph t7, s6, 8
michael@0	2399	shrl.ph t8, s7, 8
michael@0	2400	addq.ph s0, s4, t4
michael@0	2401	addq.ph s1, s5, t6
michael@0	2402	addq.ph s2, s6, t7
michael@0	2403	addq.ph s3, s7, t8
michael@0	2404	shra_r.ph t4, s0, 8
michael@0	2405	shra_r.ph t6, s1, 8
michael@0	2406	shra_r.ph t7, s2, 8
michael@0	2407	shra_r.ph t8, s3, 8
michael@0	2408
michael@0	2409	precr.qb.ph s0, t0, t1
michael@0	2410	precr.qb.ph s1, t2, t3
michael@0	2411	precr.qb.ph s2, t4, t6
michael@0	2412	precr.qb.ph s3, t7, t8
michael@0	2413
michael@0	2414	sb s0, 0(a0)
michael@0	2415	sb s1, 1(a0)
michael@0	2416	sb s2, 2(a0)
michael@0	2417	sb s3, 3(a0)
michael@0	2418	bgtz t9, 1b
michael@0	2419	addiu a0, a0, 4
michael@0	2420	2:
michael@0	2421	beqz a2, 4f
michael@0	2422	nop
michael@0	2423	3:
michael@0	2424	lbu t1, 0(a0)
michael@0	2425
michael@0	2426	muleu_s.ph.qbl t4, t1, t5
michael@0	2427	muleu_s.ph.qbr t7, t1, t5
michael@0	2428	shrl.ph t6, t4, 8
michael@0	2429	shrl.ph t0, t7, 8
michael@0	2430	addq.ph t8, t4, t6
michael@0	2431	addq.ph t9, t7, t0
michael@0	2432	shra_r.ph t8, t8, 8
michael@0	2433	shra_r.ph t9, t9, 8
michael@0	2434	precr.qb.ph t2, t8, t9
michael@0	2435	sb t2, 0(a0)
michael@0	2436	addiu a2, a2, -1
michael@0	2437	bnez a2, 3b
michael@0	2438	addiu a0, a0, 1
michael@0	2439	4:
michael@0	2440	RESTORE_REGS_FROM_STACK 20, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2441	5:
michael@0	2442	j ra
michael@0	2443	nop
michael@0	2444
michael@0	2445	END(pixman_composite_in_n_8_asm_mips)
michael@0	2446
michael@0	2447	LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_8888_8_0565_OVER_asm_mips)
michael@0	2448	/*
michael@0	2449	* a0 - dst (r5g6b5)
michael@0	2450	* a1 - src (a8r8g8b8)
michael@0	2451	* a2 - mask (a8)
michael@0	2452	* a3 - w
michael@0	2453	* 16(sp) - vx
michael@0	2454	* 20(sp) - unit_x
michael@0	2455	*/
michael@0	2456	beqz a3, 4f
michael@0	2457	nop
michael@0	2458
michael@0	2459	SAVE_REGS_ON_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5
michael@0	2460	lw v0, 36(sp) /* v0 = vx */
michael@0	2461	lw v1, 40(sp) /* v1 = unit_x */
michael@0	2462	li t6, 0x00ff00ff
michael@0	2463	li t7, 0xf800f800
michael@0	2464	li t8, 0x07e007e0
michael@0	2465	li t9, 0x001F001F
michael@0	2466
michael@0	2467	addiu t1, a3, -1
michael@0	2468	beqz t1, 2f
michael@0	2469	nop
michael@0	2470	1:
michael@0	2471	sra t0, v0, 16 /* t0 = vx >> 16 */
michael@0	2472	sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */
michael@0	2473	addu t0, a1, t0
michael@0	2474	lw t0, 0(t0) /* t0 = source (a8r8g8b8) */
michael@0	2475	addu v0, v0, v1 /* v0 = vx + unit_x */
michael@0	2476	sra t1, v0, 16 /* t1 = vx >> 16 */
michael@0	2477	sll t1, t1, 2 /* t1 = t1 * 4 (a8r8g8b8) */
michael@0	2478	addu t1, a1, t1
michael@0	2479	lw t1, 0(t1) /* t1 = source (a8r8g8b8) */
michael@0	2480	addu v0, v0, v1 /* v0 = vx + unit_x */
michael@0	2481	lbu t2, 0(a2) /* t2 = mask (a8) */
michael@0	2482	lbu t3, 1(a2) /* t3 = mask (a8) */
michael@0	2483	lhu t4, 0(a0) /* t4 = destination (r5g6b5) */
michael@0	2484	lhu t5, 2(a0) /* t5 = destination (r5g6b5) */
michael@0	2485	addiu a2, a2, 2
michael@0	2486
michael@0	2487	CONVERT_2x0565_TO_2x8888 t4, t5, s0, s1, t8, t9, s2, s3, s4, s5
michael@0	2488	OVER_2x8888_2x8_2x8888 t0, t1, \
michael@0	2489	t2, t3, \
michael@0	2490	s0, s1, \
michael@0	2491	t4, t5, \
michael@0	2492	t6, s2, s3, s4, s5, t2, t3
michael@0	2493	CONVERT_2x8888_TO_2x0565 t4, t5, s0, s1, t7, t8, t9, s2, s3
michael@0	2494
michael@0	2495	sh s0, 0(a0)
michael@0	2496	sh s1, 2(a0)
michael@0	2497	addiu a3, a3, -2
michael@0	2498	addiu t1, a3, -1
michael@0	2499	bgtz t1, 1b
michael@0	2500	addiu a0, a0, 4
michael@0	2501	2:
michael@0	2502	beqz a3, 3f
michael@0	2503	nop
michael@0	2504	sra t0, v0, 16 /* t0 = vx >> 16 */
michael@0	2505	sll t0, t0, 2 /* t0 = t0 * 4 (a8r8g8b8) */
michael@0	2506	addu t0, a1, t0
michael@0	2507	lw t0, 0(t0) /* t0 = source (a8r8g8b8) */
michael@0	2508	lbu t1, 0(a2) /* t1 = mask (a8) */
michael@0	2509	lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
michael@0	2510
michael@0	2511	CONVERT_1x0565_TO_1x8888 t2, t3, t4, t5
michael@0	2512	OVER_8888_8_8888 t0, t1, t3, t2, t6, t4, t5, t7, t8
michael@0	2513	CONVERT_1x8888_TO_1x0565 t2, t3, t4, t5
michael@0	2514
michael@0	2515	sh t3, 0(a0)
michael@0	2516	3:
michael@0	2517	RESTORE_REGS_FROM_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5
michael@0	2518	4:
michael@0	2519	j ra
michael@0	2520	nop
michael@0	2521
michael@0	2522	END(pixman_scaled_nearest_scanline_8888_8_0565_OVER_asm_mips)
michael@0	2523
michael@0	2524	LEAF_MIPS_DSPR2(pixman_scaled_nearest_scanline_0565_8_0565_OVER_asm_mips)
michael@0	2525	/*
michael@0	2526	* a0 - dst (r5g6b5)
michael@0	2527	* a1 - src (r5g6b5)
michael@0	2528	* a2 - mask (a8)
michael@0	2529	* a3 - w
michael@0	2530	* 16(sp) - vx
michael@0	2531	* 20(sp) - unit_x
michael@0	2532	*/
michael@0	2533
michael@0	2534	beqz a3, 4f
michael@0	2535	nop
michael@0	2536	SAVE_REGS_ON_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5
michael@0	2537	lw v0, 36(sp) /* v0 = vx */
michael@0	2538	lw v1, 40(sp) /* v1 = unit_x */
michael@0	2539	li t4, 0xf800f800
michael@0	2540	li t5, 0x07e007e0
michael@0	2541	li t6, 0x001F001F
michael@0	2542	li t7, 0x00ff00ff
michael@0	2543
michael@0	2544	addiu t1, a3, -1
michael@0	2545	beqz t1, 2f
michael@0	2546	nop
michael@0	2547	1:
michael@0	2548	sra t0, v0, 16 /* t0 = vx >> 16 */
michael@0	2549	sll t0, t0, 1 /* t0 = t0 * 2 (r5g6b5) */
michael@0	2550	addu t0, a1, t0
michael@0	2551	lhu t0, 0(t0) /* t0 = source (r5g6b5) */
michael@0	2552	addu v0, v0, v1 /* v0 = vx + unit_x */
michael@0	2553	sra t1, v0, 16 /* t1 = vx >> 16 */
michael@0	2554	sll t1, t1, 1 /* t1 = t1 * 2 (r5g6b5) */
michael@0	2555	addu t1, a1, t1
michael@0	2556	lhu t1, 0(t1) /* t1 = source (r5g6b5) */
michael@0	2557	addu v0, v0, v1 /* v0 = vx + unit_x */
michael@0	2558	lbu t2, 0(a2) /* t2 = mask (a8) */
michael@0	2559	lbu t3, 1(a2) /* t3 = mask (a8) */
michael@0	2560	lhu t8, 0(a0) /* t8 = destination (r5g6b5) */
michael@0	2561	lhu t9, 2(a0) /* t9 = destination (r5g6b5) */
michael@0	2562	addiu a2, a2, 2
michael@0	2563
michael@0	2564	CONVERT_2x0565_TO_2x8888 t0, t1, s0, s1, t5, t6, s2, s3, s4, s5
michael@0	2565	CONVERT_2x0565_TO_2x8888 t8, t9, s2, s3, t5, t6, s4, s5, t0, t1
michael@0	2566	OVER_2x8888_2x8_2x8888 s0, s1, \
michael@0	2567	t2, t3, \
michael@0	2568	s2, s3, \
michael@0	2569	t0, t1, \
michael@0	2570	t7, t8, t9, s4, s5, s0, s1
michael@0	2571	CONVERT_2x8888_TO_2x0565 t0, t1, s0, s1, t4, t5, t6, s2, s3
michael@0	2572
michael@0	2573	sh s0, 0(a0)
michael@0	2574	sh s1, 2(a0)
michael@0	2575	addiu a3, a3, -2
michael@0	2576	addiu t1, a3, -1
michael@0	2577	bgtz t1, 1b
michael@0	2578	addiu a0, a0, 4
michael@0	2579	2:
michael@0	2580	beqz a3, 3f
michael@0	2581	nop
michael@0	2582	sra t0, v0, 16 /* t0 = vx >> 16 */
michael@0	2583	sll t0, t0, 1 /* t0 = t0 * 2 (r5g6b5) */
michael@0	2584	addu t0, a1, t0
michael@0	2585
michael@0	2586	lhu t0, 0(t0) /* t0 = source (r5g6b5) */
michael@0	2587	lbu t1, 0(a2) /* t1 = mask (a8) */
michael@0	2588	lhu t2, 0(a0) /* t2 = destination (r5g6b5) */
michael@0	2589
michael@0	2590	CONVERT_1x0565_TO_1x8888 t0, t3, t4, t5
michael@0	2591	CONVERT_1x0565_TO_1x8888 t2, t4, t5, t6
michael@0	2592	OVER_8888_8_8888 t3, t1, t4, t0, t7, t2, t5, t6, t8
michael@0	2593	CONVERT_1x8888_TO_1x0565 t0, t3, t4, t5
michael@0	2594
michael@0	2595	sh t3, 0(a0)
michael@0	2596	3:
michael@0	2597	RESTORE_REGS_FROM_STACK 20, v0, v1, s0, s1, s2, s3, s4, s5
michael@0	2598	4:
michael@0	2599	j ra
michael@0	2600	nop
michael@0	2601
michael@0	2602	END(pixman_scaled_nearest_scanline_0565_8_0565_OVER_asm_mips)
michael@0	2603
michael@0	2604	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_SRC_asm_mips)
michael@0	2605	/*
michael@0	2606	* a0 - *dst
michael@0	2607	* a1 - *src_top
michael@0	2608	* a2 - *src_bottom
michael@0	2609	* a3 - w
michael@0	2610	* 16(sp) - wt
michael@0	2611	* 20(sp) - wb
michael@0	2612	* 24(sp) - vx
michael@0	2613	* 28(sp) - unit_x
michael@0	2614	*/
michael@0	2615
michael@0	2616	beqz a3, 1f
michael@0	2617	nop
michael@0	2618
michael@0	2619	SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2620
michael@0	2621	lw s0, 36(sp) /* s0 = wt */
michael@0	2622	lw s1, 40(sp) /* s1 = wb */
michael@0	2623	lw s2, 44(sp) /* s2 = vx */
michael@0	2624	lw s3, 48(sp) /* s3 = unit_x */
michael@0	2625	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	2626
michael@0	2627	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2628	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2629	0:
michael@0	2630	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	2631	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	2632	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	2633
michael@0	2634	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	2635	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	2636	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	2637	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	2638
michael@0	2639	sra t9, s2, 16
michael@0	2640	sll t9, t9, 2
michael@0	2641	addiu t8, t9, 4
michael@0	2642	lwx t0, t9(a1) /* t0 = tl */
michael@0	2643	lwx t1, t8(a1) /* t1 = tr */
michael@0	2644	addiu a3, a3, -1
michael@0	2645	lwx t2, t9(a2) /* t2 = bl */
michael@0	2646	lwx t3, t8(a2) /* t3 = br */
michael@0	2647
michael@0	2648	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	2649
michael@0	2650	addu s2, s2, s3 /* vx += unit_x; */
michael@0	2651	sw t0, 0(a0)
michael@0	2652	bnez a3, 0b
michael@0	2653	addiu a0, a0, 4
michael@0	2654
michael@0	2655	RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2656	1:
michael@0	2657	j ra
michael@0	2658	nop
michael@0	2659
michael@0	2660	END(pixman_scaled_bilinear_scanline_8888_8888_SRC_asm_mips)
michael@0	2661
michael@0	2662	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_0565_SRC_asm_mips)
michael@0	2663	/*
michael@0	2664	* a0 - *dst
michael@0	2665	* a1 - *src_top
michael@0	2666	* a2 - *src_bottom
michael@0	2667	* a3 - w
michael@0	2668	* 16(sp) - wt
michael@0	2669	* 20(sp) - wb
michael@0	2670	* 24(sp) - vx
michael@0	2671	* 28(sp) - unit_x
michael@0	2672	*/
michael@0	2673
michael@0	2674	beqz a3, 1f
michael@0	2675	nop
michael@0	2676
michael@0	2677	SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2678
michael@0	2679	lw s0, 36(sp) /* s0 = wt */
michael@0	2680	lw s1, 40(sp) /* s1 = wb */
michael@0	2681	lw s2, 44(sp) /* s2 = vx */
michael@0	2682	lw s3, 48(sp) /* s3 = unit_x */
michael@0	2683	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	2684
michael@0	2685	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2686	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2687	0:
michael@0	2688	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	2689	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	2690	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	2691
michael@0	2692	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	2693	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	2694	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	2695	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	2696
michael@0	2697	sra t9, s2, 16
michael@0	2698	sll t9, t9, 2
michael@0	2699	addiu t8, t9, 4
michael@0	2700	lwx t0, t9(a1) /* t0 = tl */
michael@0	2701	lwx t1, t8(a1) /* t1 = tr */
michael@0	2702	addiu a3, a3, -1
michael@0	2703	lwx t2, t9(a2) /* t2 = bl */
michael@0	2704	lwx t3, t8(a2) /* t3 = br */
michael@0	2705
michael@0	2706	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	2707	CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
michael@0	2708
michael@0	2709	addu s2, s2, s3 /* vx += unit_x; */
michael@0	2710	sh t1, 0(a0)
michael@0	2711	bnez a3, 0b
michael@0	2712	addiu a0, a0, 2
michael@0	2713
michael@0	2714	RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2715	1:
michael@0	2716	j ra
michael@0	2717	nop
michael@0	2718
michael@0	2719	END(pixman_scaled_bilinear_scanline_8888_0565_SRC_asm_mips)
michael@0	2720
michael@0	2721	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8888_SRC_asm_mips)
michael@0	2722	/*
michael@0	2723	* a0 - *dst
michael@0	2724	* a1 - *src_top
michael@0	2725	* a2 - *src_bottom
michael@0	2726	* a3 - w
michael@0	2727	* 16(sp) - wt
michael@0	2728	* 20(sp) - wb
michael@0	2729	* 24(sp) - vx
michael@0	2730	* 28(sp) - unit_x
michael@0	2731	*/
michael@0	2732
michael@0	2733	beqz a3, 1f
michael@0	2734	nop
michael@0	2735
michael@0	2736	SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	2737
michael@0	2738	lw s0, 44(sp) /* s0 = wt */
michael@0	2739	lw s1, 48(sp) /* s1 = wb */
michael@0	2740	lw s2, 52(sp) /* s2 = vx */
michael@0	2741	lw s3, 56(sp) /* s3 = unit_x */
michael@0	2742	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	2743	li v1, 0x07e007e0
michael@0	2744	li s8, 0x001f001f
michael@0	2745
michael@0	2746	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2747	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2748	0:
michael@0	2749	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	2750	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	2751	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	2752
michael@0	2753	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	2754	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	2755	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	2756	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	2757
michael@0	2758	sra t9, s2, 16
michael@0	2759	sll t9, t9, 1
michael@0	2760	addiu t8, t9, 2
michael@0	2761	lhx t0, t9(a1) /* t0 = tl */
michael@0	2762	lhx t1, t8(a1) /* t1 = tr */
michael@0	2763	andi t1, t1, 0xffff
michael@0	2764	addiu a3, a3, -1
michael@0	2765	lhx t2, t9(a2) /* t2 = bl */
michael@0	2766	lhx t3, t8(a2) /* t3 = br */
michael@0	2767	andi t3, t3, 0xffff
michael@0	2768
michael@0	2769	CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7
michael@0	2770	CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7
michael@0	2771	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	2772
michael@0	2773	addu s2, s2, s3 /* vx += unit_x; */
michael@0	2774	sw t0, 0(a0)
michael@0	2775	bnez a3, 0b
michael@0	2776	addiu a0, a0, 4
michael@0	2777
michael@0	2778	RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	2779	1:
michael@0	2780	j ra
michael@0	2781	nop
michael@0	2782
michael@0	2783	END(pixman_scaled_bilinear_scanline_0565_8888_SRC_asm_mips)
michael@0	2784
michael@0	2785	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_0565_SRC_asm_mips)
michael@0	2786	/*
michael@0	2787	* a0 - *dst
michael@0	2788	* a1 - *src_top
michael@0	2789	* a2 - *src_bottom
michael@0	2790	* a3 - w
michael@0	2791	* 16(sp) - wt
michael@0	2792	* 20(sp) - wb
michael@0	2793	* 24(sp) - vx
michael@0	2794	* 28(sp) - unit_x
michael@0	2795	*/
michael@0	2796
michael@0	2797	beqz a3, 1f
michael@0	2798	nop
michael@0	2799
michael@0	2800	SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	2801
michael@0	2802	lw s0, 44(sp) /* s0 = wt */
michael@0	2803	lw s1, 48(sp) /* s1 = wb */
michael@0	2804	lw s2, 52(sp) /* s2 = vx */
michael@0	2805	lw s3, 56(sp) /* s3 = unit_x */
michael@0	2806	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	2807	li v1, 0x07e007e0
michael@0	2808	li s8, 0x001f001f
michael@0	2809
michael@0	2810	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2811	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2812	0:
michael@0	2813	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	2814	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	2815	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	2816
michael@0	2817	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	2818	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	2819	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	2820	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	2821
michael@0	2822	sra t9, s2, 16
michael@0	2823	sll t9, t9, 1
michael@0	2824	addiu t8, t9, 2
michael@0	2825	lhx t0, t9(a1) /* t0 = tl */
michael@0	2826	lhx t1, t8(a1) /* t1 = tr */
michael@0	2827	andi t1, t1, 0xffff
michael@0	2828	addiu a3, a3, -1
michael@0	2829	lhx t2, t9(a2) /* t2 = bl */
michael@0	2830	lhx t3, t8(a2) /* t3 = br */
michael@0	2831	andi t3, t3, 0xffff
michael@0	2832
michael@0	2833	CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7
michael@0	2834	CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7
michael@0	2835	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	2836	CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
michael@0	2837
michael@0	2838	addu s2, s2, s3 /* vx += unit_x; */
michael@0	2839	sh t1, 0(a0)
michael@0	2840	bnez a3, 0b
michael@0	2841	addiu a0, a0, 2
michael@0	2842
michael@0	2843	RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	2844	1:
michael@0	2845	j ra
michael@0	2846	nop
michael@0	2847
michael@0	2848	END(pixman_scaled_bilinear_scanline_0565_0565_SRC_asm_mips)
michael@0	2849
michael@0	2850	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_OVER_asm_mips)
michael@0	2851	/*
michael@0	2852	* a0 - *dst
michael@0	2853	* a1 - *src_top
michael@0	2854	* a2 - *src_bottom
michael@0	2855	* a3 - w
michael@0	2856	* 16(sp) - wt
michael@0	2857	* 20(sp) - wb
michael@0	2858	* 24(sp) - vx
michael@0	2859	* 28(sp) - unit_x
michael@0	2860	*/
michael@0	2861
michael@0	2862	beqz a3, 1f
michael@0	2863	nop
michael@0	2864
michael@0	2865	SAVE_REGS_ON_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	2866
michael@0	2867	lw s0, 40(sp) /* s0 = wt */
michael@0	2868	lw s1, 44(sp) /* s1 = wb */
michael@0	2869	lw s2, 48(sp) /* s2 = vx */
michael@0	2870	lw s3, 52(sp) /* s3 = unit_x */
michael@0	2871	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	2872	li s8, 0x00ff00ff
michael@0	2873
michael@0	2874	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2875	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2876	0:
michael@0	2877	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	2878	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	2879	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	2880
michael@0	2881	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	2882	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	2883	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	2884	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	2885
michael@0	2886	sra t9, s2, 16
michael@0	2887	sll t9, t9, 2
michael@0	2888	addiu t8, t9, 4
michael@0	2889	lwx t0, t9(a1) /* t0 = tl */
michael@0	2890	lwx t1, t8(a1) /* t1 = tr */
michael@0	2891	addiu a3, a3, -1
michael@0	2892	lwx t2, t9(a2) /* t2 = bl */
michael@0	2893	lwx t3, t8(a2) /* t3 = br */
michael@0	2894
michael@0	2895	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	2896	lw t1, 0(a0) /* t1 = dest */
michael@0	2897	OVER_8888_8888 t0, t1, t2, s8, t3, t4, t5, t6
michael@0	2898
michael@0	2899	addu s2, s2, s3 /* vx += unit_x; */
michael@0	2900	sw t2, 0(a0)
michael@0	2901	bnez a3, 0b
michael@0	2902	addiu a0, a0, 4
michael@0	2903
michael@0	2904	RESTORE_REGS_FROM_STACK 24, v0, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	2905	1:
michael@0	2906	j ra
michael@0	2907	nop
michael@0	2908
michael@0	2909	END(pixman_scaled_bilinear_scanline_8888_8888_OVER_asm_mips)
michael@0	2910
michael@0	2911	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8888_ADD_asm_mips)
michael@0	2912	/*
michael@0	2913	* a0 - *dst
michael@0	2914	* a1 - *src_top
michael@0	2915	* a2 - *src_bottom
michael@0	2916	* a3 - w
michael@0	2917	* 16(sp) - wt
michael@0	2918	* 20(sp) - wb
michael@0	2919	* 24(sp) - vx
michael@0	2920	* 28(sp) - unit_x
michael@0	2921	*/
michael@0	2922
michael@0	2923	beqz a3, 1f
michael@0	2924	nop
michael@0	2925
michael@0	2926	SAVE_REGS_ON_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2927
michael@0	2928	lw s0, 36(sp) /* s0 = wt */
michael@0	2929	lw s1, 40(sp) /* s1 = wb */
michael@0	2930	lw s2, 44(sp) /* s2 = vx */
michael@0	2931	lw s3, 48(sp) /* s3 = unit_x */
michael@0	2932	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	2933
michael@0	2934	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2935	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2936	0:
michael@0	2937	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	2938	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	2939	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	2940
michael@0	2941	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	2942	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	2943	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	2944	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	2945
michael@0	2946	sra t9, s2, 16
michael@0	2947	sll t9, t9, 2
michael@0	2948	addiu t8, t9, 4
michael@0	2949	lwx t0, t9(a1) /* t0 = tl */
michael@0	2950	lwx t1, t8(a1) /* t1 = tr */
michael@0	2951	addiu a3, a3, -1
michael@0	2952	lwx t2, t9(a2) /* t2 = bl */
michael@0	2953	lwx t3, t8(a2) /* t3 = br */
michael@0	2954
michael@0	2955	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	2956	lw t1, 0(a0)
michael@0	2957	addu_s.qb t2, t0, t1
michael@0	2958
michael@0	2959	addu s2, s2, s3 /* vx += unit_x; */
michael@0	2960	sw t2, 0(a0)
michael@0	2961	bnez a3, 0b
michael@0	2962	addiu a0, a0, 4
michael@0	2963
michael@0	2964	RESTORE_REGS_FROM_STACK 20, v0, s0, s1, s2, s3, s4, s5, s6, s7
michael@0	2965	1:
michael@0	2966	j ra
michael@0	2967	nop
michael@0	2968
michael@0	2969	END(pixman_scaled_bilinear_scanline_8888_8888_ADD_asm_mips)
michael@0	2970
michael@0	2971	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_SRC_asm_mips)
michael@0	2972	/*
michael@0	2973	* a0 - *dst
michael@0	2974	* a1 - *mask
michael@0	2975	* a2 - *src_top
michael@0	2976	* a3 - *src_bottom
michael@0	2977	* 16(sp) - wt
michael@0	2978	* 20(sp) - wb
michael@0	2979	* 24(sp) - vx
michael@0	2980	* 28(sp) - unit_x
michael@0	2981	* 32(sp) - w
michael@0	2982	*/
michael@0	2983
michael@0	2984	lw v1, 32(sp)
michael@0	2985	beqz v1, 1f
michael@0	2986	nop
michael@0	2987
michael@0	2988	SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	2989
michael@0	2990	lw s0, 44(sp) /* s0 = wt */
michael@0	2991	lw s1, 48(sp) /* s1 = wb */
michael@0	2992	lw s2, 52(sp) /* s2 = vx */
michael@0	2993	lw s3, 56(sp) /* s3 = unit_x */
michael@0	2994	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	2995	li s8, 0x00ff00ff
michael@0	2996
michael@0	2997	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2998	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	2999	0:
michael@0	3000	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	3001	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	3002	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	3003
michael@0	3004	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	3005	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	3006	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	3007	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	3008
michael@0	3009	sra t9, s2, 16
michael@0	3010	sll t9, t9, 2
michael@0	3011	addiu t8, t9, 4
michael@0	3012	lwx t0, t9(a2) /* t0 = tl */
michael@0	3013	lwx t1, t8(a2) /* t1 = tr */
michael@0	3014	addiu v1, v1, -1
michael@0	3015	lwx t2, t9(a3) /* t2 = bl */
michael@0	3016	lwx t3, t8(a3) /* t3 = br */
michael@0	3017
michael@0	3018	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	3019	lbu t1, 0(a1) /* t1 = mask */
michael@0	3020	addiu a1, a1, 1
michael@0	3021	MIPS_UN8x4_MUL_UN8 t0, t1, t0, s8, t2, t3, t4
michael@0	3022
michael@0	3023	addu s2, s2, s3 /* vx += unit_x; */
michael@0	3024	sw t0, 0(a0)
michael@0	3025	bnez v1, 0b
michael@0	3026	addiu a0, a0, 4
michael@0	3027
michael@0	3028	RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	3029	1:
michael@0	3030	j ra
michael@0	3031	nop
michael@0	3032
michael@0	3033	END(pixman_scaled_bilinear_scanline_8888_8_8888_SRC_asm_mips)
michael@0	3034
michael@0	3035	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_0565_SRC_asm_mips)
michael@0	3036	/*
michael@0	3037	* a0 - *dst
michael@0	3038	* a1 - *mask
michael@0	3039	* a2 - *src_top
michael@0	3040	* a3 - *src_bottom
michael@0	3041	* 16(sp) - wt
michael@0	3042	* 20(sp) - wb
michael@0	3043	* 24(sp) - vx
michael@0	3044	* 28(sp) - unit_x
michael@0	3045	* 32(sp) - w
michael@0	3046	*/
michael@0	3047
michael@0	3048	lw v1, 32(sp)
michael@0	3049	beqz v1, 1f
michael@0	3050	nop
michael@0	3051
michael@0	3052	SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	3053
michael@0	3054	lw s0, 44(sp) /* s0 = wt */
michael@0	3055	lw s1, 48(sp) /* s1 = wb */
michael@0	3056	lw s2, 52(sp) /* s2 = vx */
michael@0	3057	lw s3, 56(sp) /* s3 = unit_x */
michael@0	3058	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	3059	li s8, 0x00ff00ff
michael@0	3060
michael@0	3061	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3062	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3063	0:
michael@0	3064	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	3065	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	3066	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	3067
michael@0	3068	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	3069	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	3070	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	3071	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	3072
michael@0	3073	sra t9, s2, 16
michael@0	3074	sll t9, t9, 2
michael@0	3075	addiu t8, t9, 4
michael@0	3076	lwx t0, t9(a2) /* t0 = tl */
michael@0	3077	lwx t1, t8(a2) /* t1 = tr */
michael@0	3078	addiu v1, v1, -1
michael@0	3079	lwx t2, t9(a3) /* t2 = bl */
michael@0	3080	lwx t3, t8(a3) /* t3 = br */
michael@0	3081
michael@0	3082	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	3083	lbu t1, 0(a1) /* t1 = mask */
michael@0	3084	addiu a1, a1, 1
michael@0	3085	MIPS_UN8x4_MUL_UN8 t0, t1, t0, s8, t2, t3, t4
michael@0	3086	CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
michael@0	3087
michael@0	3088	addu s2, s2, s3 /* vx += unit_x; */
michael@0	3089	sh t1, 0(a0)
michael@0	3090	bnez v1, 0b
michael@0	3091	addiu a0, a0, 2
michael@0	3092
michael@0	3093	RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	3094	1:
michael@0	3095	j ra
michael@0	3096	nop
michael@0	3097
michael@0	3098	END(pixman_scaled_bilinear_scanline_8888_8_0565_SRC_asm_mips)
michael@0	3099
michael@0	3100	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8_x888_SRC_asm_mips)
michael@0	3101	/*
michael@0	3102	* a0 - *dst
michael@0	3103	* a1 - *mask
michael@0	3104	* a2 - *src_top
michael@0	3105	* a3 - *src_bottom
michael@0	3106	* 16(sp) - wt
michael@0	3107	* 20(sp) - wb
michael@0	3108	* 24(sp) - vx
michael@0	3109	* 28(sp) - unit_x
michael@0	3110	* 32(sp) - w
michael@0	3111	*/
michael@0	3112
michael@0	3113	lw t0, 32(sp)
michael@0	3114	beqz t0, 1f
michael@0	3115	nop
michael@0	3116
michael@0	3117	SAVE_REGS_ON_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra
michael@0	3118
michael@0	3119	lw s0, 48(sp) /* s0 = wt */
michael@0	3120	lw s1, 52(sp) /* s1 = wb */
michael@0	3121	lw s2, 56(sp) /* s2 = vx */
michael@0	3122	lw s3, 60(sp) /* s3 = unit_x */
michael@0	3123	lw ra, 64(sp) /* ra = w */
michael@0	3124	li v0, 0x00ff00ff
michael@0	3125	li v1, 0x07e007e0
michael@0	3126	li s8, 0x001f001f
michael@0	3127
michael@0	3128	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3129	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3130	0:
michael@0	3131	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	3132	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	3133	li t5, BILINEAR_INTERPOLATION_RANGE
michael@0	3134	subu t5, t5, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	3135
michael@0	3136	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	3137	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	3138	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	3139	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	3140
michael@0	3141	sra t9, s2, 16
michael@0	3142	sll t9, t9, 1
michael@0	3143	addiu t8, t9, 2
michael@0	3144	lhx t0, t9(a2) /* t0 = tl */
michael@0	3145	lhx t1, t8(a2) /* t1 = tr */
michael@0	3146	andi t1, t1, 0xffff
michael@0	3147	addiu ra, ra, -1
michael@0	3148	lhx t2, t9(a3) /* t2 = bl */
michael@0	3149	lhx t3, t8(a3) /* t3 = br */
michael@0	3150	andi t3, t3, 0xffff
michael@0	3151
michael@0	3152	CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7
michael@0	3153	CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7
michael@0	3154	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	3155	lbu t1, 0(a1) /* t1 = mask */
michael@0	3156	addiu a1, a1, 1
michael@0	3157	MIPS_UN8x4_MUL_UN8 t0, t1, t0, v0, t2, t3, t4
michael@0	3158
michael@0	3159	addu s2, s2, s3 /* vx += unit_x; */
michael@0	3160	sw t0, 0(a0)
michael@0	3161	bnez ra, 0b
michael@0	3162	addiu a0, a0, 4
michael@0	3163
michael@0	3164	RESTORE_REGS_FROM_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra
michael@0	3165	1:
michael@0	3166	j ra
michael@0	3167	nop
michael@0	3168
michael@0	3169	END(pixman_scaled_bilinear_scanline_0565_8_x888_SRC_asm_mips)
michael@0	3170
michael@0	3171	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_0565_8_0565_SRC_asm_mips)
michael@0	3172	/*
michael@0	3173	* a0 - *dst
michael@0	3174	* a1 - *mask
michael@0	3175	* a2 - *src_top
michael@0	3176	* a3 - *src_bottom
michael@0	3177	* 16(sp) - wt
michael@0	3178	* 20(sp) - wb
michael@0	3179	* 24(sp) - vx
michael@0	3180	* 28(sp) - unit_x
michael@0	3181	* 32(sp) - w
michael@0	3182	*/
michael@0	3183
michael@0	3184	lw t0, 32(sp)
michael@0	3185	beqz t0, 1f
michael@0	3186	nop
michael@0	3187
michael@0	3188	SAVE_REGS_ON_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra
michael@0	3189
michael@0	3190	lw s0, 48(sp) /* s0 = wt */
michael@0	3191	lw s1, 52(sp) /* s1 = wb */
michael@0	3192	lw s2, 56(sp) /* s2 = vx */
michael@0	3193	lw s3, 60(sp) /* s3 = unit_x */
michael@0	3194	lw ra, 64(sp) /* ra = w */
michael@0	3195	li v0, 0x00ff00ff
michael@0	3196	li v1, 0x07e007e0
michael@0	3197	li s8, 0x001f001f
michael@0	3198
michael@0	3199	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3200	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3201	0:
michael@0	3202	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	3203	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	3204	li t5, BILINEAR_INTERPOLATION_RANGE
michael@0	3205	subu t5, t5, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	3206
michael@0	3207	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	3208	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	3209	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	3210	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	3211
michael@0	3212	sra t9, s2, 16
michael@0	3213	sll t9, t9, 1
michael@0	3214	addiu t8, t9, 2
michael@0	3215	lhx t0, t9(a2) /* t0 = tl */
michael@0	3216	lhx t1, t8(a2) /* t1 = tr */
michael@0	3217	andi t1, t1, 0xffff
michael@0	3218	addiu ra, ra, -1
michael@0	3219	lhx t2, t9(a3) /* t2 = bl */
michael@0	3220	lhx t3, t8(a3) /* t3 = br */
michael@0	3221	andi t3, t3, 0xffff
michael@0	3222
michael@0	3223	CONVERT_2x0565_TO_2x8888 t0, t1, t0, t1, v1, s8, t4, t5, t6, t7
michael@0	3224	CONVERT_2x0565_TO_2x8888 t2, t3, t2, t3, v1, s8, t4, t5, t6, t7
michael@0	3225	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	3226	lbu t1, 0(a1) /* t1 = mask */
michael@0	3227	addiu a1, a1, 1
michael@0	3228	MIPS_UN8x4_MUL_UN8 t0, t1, t0, v0, t2, t3, t4
michael@0	3229	CONVERT_1x8888_TO_1x0565 t0, t1, t2, t3
michael@0	3230
michael@0	3231	addu s2, s2, s3 /* vx += unit_x; */
michael@0	3232	sh t1, 0(a0)
michael@0	3233	bnez ra, 0b
michael@0	3234	addiu a0, a0, 2
michael@0	3235
michael@0	3236	RESTORE_REGS_FROM_STACK 32, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8, ra
michael@0	3237	1:
michael@0	3238	j ra
michael@0	3239	nop
michael@0	3240
michael@0	3241	END(pixman_scaled_bilinear_scanline_0565_8_0565_SRC_asm_mips)
michael@0	3242
michael@0	3243	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_OVER_asm_mips)
michael@0	3244	/*
michael@0	3245	* a0 - dst (a8r8g8b8)
michael@0	3246	* a1 - mask (a8)
michael@0	3247	* a2 - src_top (a8r8g8b8)
michael@0	3248	* a3 - src_bottom (a8r8g8b8)
michael@0	3249	* 16(sp) - wt
michael@0	3250	* 20(sp) - wb
michael@0	3251	* 24(sp) - vx
michael@0	3252	* 28(sp) - unit_x
michael@0	3253	* 32(sp) - w
michael@0	3254	*/
michael@0	3255
michael@0	3256	SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	3257
michael@0	3258	lw v1, 60(sp) /* v1 = w(sp + 32 + 28 save regs stack offset)*/
michael@0	3259	beqz v1, 1f
michael@0	3260	nop
michael@0	3261
michael@0	3262	lw s0, 44(sp) /* s0 = wt */
michael@0	3263	lw s1, 48(sp) /* s1 = wb */
michael@0	3264	lw s2, 52(sp) /* s2 = vx */
michael@0	3265	lw s3, 56(sp) /* s3 = unit_x */
michael@0	3266	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	3267	li s8, 0x00ff00ff
michael@0	3268
michael@0	3269	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3270	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3271
michael@0	3272	0:
michael@0	3273	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	3274	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	3275	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	3276
michael@0	3277	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	3278	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	3279	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	3280	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	3281
michael@0	3282	sra t9, s2, 16
michael@0	3283	sll t9, t9, 2
michael@0	3284	addiu t8, t9, 4
michael@0	3285	lwx t0, t9(a2) /* t0 = tl */
michael@0	3286	lwx t1, t8(a2) /* t1 = tr */
michael@0	3287	addiu v1, v1, -1
michael@0	3288	lwx t2, t9(a3) /* t2 = bl */
michael@0	3289	lwx t3, t8(a3) /* t3 = br */
michael@0	3290
michael@0	3291	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, \
michael@0	3292	t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	3293	lbu t1, 0(a1) /* t1 = mask */
michael@0	3294	lw t2, 0(a0) /* t2 = dst */
michael@0	3295	addiu a1, a1, 1
michael@0	3296	OVER_8888_8_8888 t0, t1, t2, t0, s8, t3, t4, t5, t6
michael@0	3297
michael@0	3298	addu s2, s2, s3 /* vx += unit_x; */
michael@0	3299	sw t0, 0(a0)
michael@0	3300	bnez v1, 0b
michael@0	3301	addiu a0, a0, 4
michael@0	3302
michael@0	3303	1:
michael@0	3304	RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	3305	j ra
michael@0	3306	nop
michael@0	3307
michael@0	3308	END(pixman_scaled_bilinear_scanline_8888_8_8888_OVER_asm_mips)
michael@0	3309
michael@0	3310	LEAF_MIPS_DSPR2(pixman_scaled_bilinear_scanline_8888_8_8888_ADD_asm_mips)
michael@0	3311	/*
michael@0	3312	* a0 - *dst
michael@0	3313	* a1 - *mask
michael@0	3314	* a2 - *src_top
michael@0	3315	* a3 - *src_bottom
michael@0	3316	* 16(sp) - wt
michael@0	3317	* 20(sp) - wb
michael@0	3318	* 24(sp) - vx
michael@0	3319	* 28(sp) - unit_x
michael@0	3320	* 32(sp) - w
michael@0	3321	*/
michael@0	3322
michael@0	3323	lw v1, 32(sp)
michael@0	3324	beqz v1, 1f
michael@0	3325	nop
michael@0	3326
michael@0	3327	SAVE_REGS_ON_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	3328
michael@0	3329	lw s0, 44(sp) /* s0 = wt */
michael@0	3330	lw s1, 48(sp) /* s1 = wb */
michael@0	3331	lw s2, 52(sp) /* s2 = vx */
michael@0	3332	lw s3, 56(sp) /* s3 = unit_x */
michael@0	3333	li v0, BILINEAR_INTERPOLATION_RANGE
michael@0	3334	li s8, 0x00ff00ff
michael@0	3335
michael@0	3336	sll s0, s0, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3337	sll s1, s1, (2 * (8 - BILINEAR_INTERPOLATION_BITS))
michael@0	3338	0:
michael@0	3339	andi t4, s2, 0xffff /* t4 = (short)vx */
michael@0	3340	srl t4, t4, (16 - BILINEAR_INTERPOLATION_BITS) /* t4 = vx >> 8 */
michael@0	3341	subu t5, v0, t4 /* t5 = ( 256 - (vx>>8)) */
michael@0	3342
michael@0	3343	mul s4, s0, t5 /* s4 = wt*(256-(vx>>8)) */
michael@0	3344	mul s5, s0, t4 /* s5 = wt*(vx>>8) */
michael@0	3345	mul s6, s1, t5 /* s6 = wb*(256-(vx>>8)) */
michael@0	3346	mul s7, s1, t4 /* s7 = wb*(vx>>8) */
michael@0	3347
michael@0	3348	sra t9, s2, 16
michael@0	3349	sll t9, t9, 2
michael@0	3350	addiu t8, t9, 4
michael@0	3351	lwx t0, t9(a2) /* t0 = tl */
michael@0	3352	lwx t1, t8(a2) /* t1 = tr */
michael@0	3353	addiu v1, v1, -1
michael@0	3354	lwx t2, t9(a3) /* t2 = bl */
michael@0	3355	lwx t3, t8(a3) /* t3 = br */
michael@0	3356
michael@0	3357	BILINEAR_INTERPOLATE_SINGLE_PIXEL t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, s4, s5, s6, s7
michael@0	3358	lbu t1, 0(a1) /* t1 = mask */
michael@0	3359	lw t2, 0(a0) /* t2 = dst */
michael@0	3360	addiu a1, a1, 1
michael@0	3361	MIPS_UN8x4_MUL_UN8_ADD_UN8x4 t0, t1, t2, t0, s8, t3, t4, t5
michael@0	3362
michael@0	3363	addu s2, s2, s3 /* vx += unit_x; */
michael@0	3364	sw t0, 0(a0)
michael@0	3365	bnez v1, 0b
michael@0	3366	addiu a0, a0, 4
michael@0	3367
michael@0	3368	RESTORE_REGS_FROM_STACK 28, v0, v1, s0, s1, s2, s3, s4, s5, s6, s7, s8
michael@0	3369	1:
michael@0	3370	j ra
michael@0	3371	nop
michael@0	3372
michael@0	3373	END(pixman_scaled_bilinear_scanline_8888_8_8888_ADD_asm_mips)