WebSVN – Kolibri OS – Blame – /programs/develop/libraries/libs-dev/libimg/xcf/composite_sse.asm

Rev	Author	Line No.	Line
2388	dunkaist	1	proc xcf._.blend_rgb
		2
		3	push eax ebx
		4
		5	xchg al, bh
		6	mov ah, bh
		7	neg ax
		8	add ax, 0xffff
		9	mul ah
		10	neg ah
		11	add ah, 0xff
		12	xchg ah, bh
		13
		14	mov al, 0xff
		15	cmp ah, bh
		16	je @f
		17	not al
		18	div bh
		19	@@:
		20	mov ah, al
		21	movd xmm1, eax
		22
		23	pop ebx eax
		24	push ebx
		25
		26	shr eax, 8
		27	shr ebx, 8
		28
		29	xchg al, bh
		30	mov ah, bh
		31	neg ax
		32	add ax, 0xffff
		33	mul ah
		34	neg ah
		35	add ah, 0xff
		36	xchg ah, bh
		37
		38	mov al, 0xff
		39	cmp ah, bh
		40	je @f
		41	not al
		42	div bh
		43	@@:
		44	mov ah, al
		45	movd ebx, xmm1
		46	ror ebx, 16
		47	mov bx, ax
		48	rol ebx, 16
		49	movd xmm1, ebx
		50
		51	pop ebx
		52
		53	; movdqu xmm1, xword[xcf._.xmm_000000ff]
		54	; movdqa xmm4, xmm1
		55	; movdqa xmm5, xmm1
		56	; movdqa xmm6, xmm2
		57	; psrldq xmm6, 3
		58	; pand xmm6, xmm1
		59	; psubw xmm4, xmm6
		60	; movdqa xmm6, xmm3
		61	; psrldq xmm6, 3
		62	; pand xmm6, xmm1
		63	; psubw xmm5, xmm6
		64	; pmullw xmm4, xmm5
		65	; psrlw xmm4, 8
		66	; psubw xmm1, xmm4
		67	; movdqa xmm4, xmm1
		68	; movdqa xmm1, xmm6
		69	; divps xmm1, xmm4
		70	; packuswb xmm1, xmm0
		71	; packuswb xmm1, xmm0
		72	; punpcklbw xmm1, xmm1
		73
		74	punpcklbw xmm1, xmm1
		75	punpcklbw xmm1, xmm0
		76
		77	movdqa xmm7, xmm1
		78	psrlw xmm7, 7
		79	paddw xmm1, xmm7
		80
		81	psubw xmm3, xmm2
		82	pmullw xmm3, xmm1
		83	psllw xmm2, 8
		84	paddw xmm3, xmm2
		85	pinsrw xmm3, ebx, 3
		86	shr ebx, 8
		87	pinsrw xmm3, ebx, 7
		88	psrlw xmm3, 8
		89	packuswb xmm3, xmm0
		90
		91	ret
		92	endp
		93
		94
		95	proc xcf._.blend_gray
		96
		97	xchg al, bh
		98	mov ah, bh
		99	neg ax
		100	add ax, 0xffff
		101	mul ah
		102	neg ah
		103	add ah, 0xff
		104	xchg ah, bh
		105
		106	mov al, 0xff
		107	cmp ah, bh
		108	je @f
		109	not al
		110	div bh
		111	@@:
		112
		113	mov ah, al
		114
		115	movd xmm1, eax
		116	punpcklbw xmm1, xmm1
		117	punpcklbw xmm1, xmm0
		118
		119	movq xmm7, xmm1
		120	psrlw xmm7, 7
		121	paddw xmm1, xmm7
		122
		123	psubw xmm3, xmm2
		124	pmullw xmm3, xmm1
		125	psllw xmm2, 8
		126	paddw xmm3, xmm2
		127	pinsrw xmm3, ebx, 1
		128	psrlw xmm3, 8
		129	packuswb xmm3, xmm0
		130
		131	ret
		132	endp
		133
		134
		135	proc xcf._.merge_32 _copy_width, _copy_height, _img_total_bpl, _bottom_total_bpl
		136
		137	pxor xmm0, xmm0
		138
		139	.line:
		140	mov ecx, [_copy_width]
		141	bt ecx, 0
		142	jnc .even
		143	.odd:
		144	movd xmm2, [edi]
		145	movd xmm3, [esi]
		146	add esi, 4
		147
		148	movdqa xmm4, xmm2
		149	pminub xmm4, xmm3
		150	pextrw eax, xmm4, 3
		151	pextrw ebx, xmm4, 1
		152	mov al, bh
		153
		154	push eax
		155	pextrw eax, xmm2, 3
		156	pextrw ebx, xmm2, 1
		157	mov bl, ah
		158	shl ebx, 8
		159	pop eax
		160
		161	call edx
		162	call xcf._.blend_rgb
		163	movd [edi], xmm3
		164	add edi, 4
		165
		166	cmp ecx, 1
		167	je .done
		168
		169	.even:
		170	sub ecx, 2
		171	.pixel:
		172	movq xmm2, [edi]
		173	movq xmm3, [esi]
		174	add esi, 8
		175
		176	movdqa xmm4, xmm2
		177	pminub xmm4, xmm3
		178	pextrw eax, xmm4, 3
		179	pextrw ebx, xmm4, 1
		180	mov al, bh
		181
		182	push eax
		183	pextrw eax, xmm2, 3
		184	pextrw ebx, xmm2, 1
		185	mov bl, ah
		186	shl ebx, 8
		187	pop eax
		188
		189	call edx
		190	call xcf._.blend_rgb
		191	movq [edi], xmm3
		192	add edi, 8
		193	sub ecx, 2
		194	jns .pixel
		195	add esi, [_img_total_bpl]
		196	add edi, [_bottom_total_bpl]
		197	dec [_copy_height]
		198	jnz .line
		199	.done:
		200	ret
		201	endp
		202
		203
		204	proc xcf._.merge_8a _copy_width, _copy_height, _img_total_bpl, _bottom_total_bpl
		205	.gray_line:
		206	mov ecx, [_copy_width]
		207	.gray_pixel:
		208	mov bx, word[edi]
		209	lodsw
		210	movd xmm2, ebx
		211	movd xmm3, eax
		212	shr eax, 8
		213	cmp al, bh
		214	jna @f
		215	mov al, bh
		216	@@:
		217	pxor xmm0, xmm0
		218	call edx
		219	call xcf._.blend_gray
		220	movd eax, xmm3
		221	stosw
		222	dec ecx
		223	jnz .gray_pixel
		224	add esi, [_img_total_bpl]
		225	add edi, [_bottom_total_bpl]
		226	dec [_copy_height]
		227	jnz .gray_line
		228	ret
		229	endp
		230
		231
		232	proc xcf._.composite_rgb_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		233
		234	pxor xmm0, xmm0
		235
		236	.line:
		237	mov ecx, [_copy_width]
		238	bt ecx, 0
		239	jnc .even
		240	.odd:
		241	movlpd xmm2, [edi]
		242	movlpd xmm3, [esi]
		243	add esi, 4
		244
		245	pextrw eax, xmm3, 3
		246	pextrw ebx, xmm3, 1
		247	mov al, bh
		248
		249	push eax
		250	pextrw eax, xmm2, 3
		251	pextrw ebx, xmm2, 1
		252	mov bl, ah
		253	shl ebx, 8
		254	pop eax
		255
		256	xchg al, bh
		257	mov ah, bh
		258	neg al
		259	neg ah
		260	dec al
		261	dec ah
		262	mul ah
		263	neg ah
		264	dec ah
		265	xchg ah, bh
		266
		267	mov al, 0xff
		268	cmp ah, bh
		269	je @f
		270	inc al
		271	div bh
		272	@@:
		273	mov ah, al
		274	movd xmm1, eax
		275
		276	punpcklbw xmm1, xmm1
		277	punpcklbw xmm1, xmm0
		278	punpcklbw xmm2, xmm0
		279	punpcklbw xmm3, xmm0
		280
		281	psubsw xmm3, xmm2
		282	pmullw xmm3, xmm1
		283	psllw xmm2, 8
		284	paddw xmm3, xmm2
		285	pinsrw xmm3, ebx, 3
		286	shr ebx, 8
		287	pinsrw xmm3, ebx, 7
		288	psrlw xmm3, 8
		289	packuswb xmm3, xmm0
		290
		291	movd [edi], xmm3
		292	add edi, 4
		293
		294	cmp ecx, 1
		295	je .done
		296
		297	.even:
		298	sub ecx, 2
		299	.pixel:
		300	movlpd xmm2, [edi]
		301	movlpd xmm3, [esi]
		302	add esi, 8
		303
		304	pextrw eax, xmm3, 3
		305	pextrw ebx, xmm3, 1
		306	mov al, bh
		307
		308	push eax
		309	pextrw eax, xmm2, 3
		310	pextrw ebx, xmm2, 1
		311	mov bl, ah
		312	shl ebx, 8
		313	pop eax
		314
		315
		316	push eax ebx
		317
		318	xchg al, bh
		319	mov ah, bh
		320	neg al
		321	neg ah
		322	dec al
		323	dec ah
		324	mul ah
		325	neg ah
		326	dec ah
		327	xchg ah, bh
		328
		329	mov al, 0xff
		330	cmp ah, bh
		331	je @f
		332	inc al
		333	div bh
		334	@@:
		335	mov ah, al
		336	movd xmm1, eax
		337
		338	pop ebx eax
		339	push ebx
		340
		341	shr eax, 8
		342	shr ebx, 8
		343
		344	xchg al, bh
		345	mov ah, bh
		346	neg ax
		347	add ax, 0xffff
		348	mul ah
		349	neg ah
		350	add ah, 0xff
		351	xchg ah, bh
		352
		353	mov al, 0xff
		354	cmp ah, bh
		355	je @f
		356	not al
		357	div bh
		358	@@:
		359	mov ah, al
		360	movd ebx, xmm1
		361	ror ebx, 16
		362	mov bx, ax
		363	rol ebx, 16
		364	movd xmm1, ebx
		365
		366	pop ebx
		367
		368	punpcklbw xmm1, xmm1
		369	punpcklbw xmm1, xmm0
		370	punpcklbw xmm2, xmm0
		371	punpcklbw xmm3, xmm0
		372
		373	psubsw xmm3, xmm2
		374	pmullw xmm3, xmm1
		375	psllw xmm2, 8
		376	paddw xmm3, xmm2
		377	pinsrw xmm3, ebx, 3
		378	shr ebx, 8
		379	pinsrw xmm3, ebx, 7
		380	psrlw xmm3, 8
		381	packuswb xmm3, xmm0
		382
		383	movq [edi], xmm3
		384	add edi, 8
		385	sub ecx, 2
		386	jns .pixel
		387	add esi, [_img_total_bpl]
		388	add edi, [_bottom_total_bpl]
		389	dec [_copy_height]
		390	jnz .line
		391	.done:
		392	ret
		393	endp
		394
		395
		396	proc xcf._.composite_gray_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		397
		398	.line:
		399	mov ecx, [_copy_width]
		400	.pixel:
		401	mov bx, [edi]
		402	lodsw
		403	movd xmm2, ebx
		404	movd xmm3, eax
		405
		406	shr eax, 8
		407
		408	xchg al, bh
		409	mov ah, bh
		410	neg ax
		411	add ax, 0xffff
		412	mul ah
		413	neg ah
		414	add ah, 0xff
		415	xchg ah, bh
		416
		417	mov al, 0xff
		418	cmp ah, bh
		419	je @f
		420	not al
		421	div bh
		422	@@:
		423
		424	mov ah, al
		425
		426	movd xmm1, eax
		427	pxor xmm0, xmm0
		428	punpcklbw xmm1, xmm1
		429	punpcklbw xmm1, xmm0
		430	punpcklbw xmm2, xmm0
		431	punpcklbw xmm3, xmm0
		432
		433	psubw xmm3, xmm2
		434	pmullw xmm3, xmm1
		435	psllw xmm2, 8
		436	paddw xmm3, xmm2
		437	pinsrw xmm3, ebx, 1
		438	psrlw xmm3, 8
		439	packuswb xmm3, xmm0
		440	movd eax, xmm3
		441	stosw
		442
		443	dec ecx
		444	jnz .pixel
		445	add esi, [_img_total_bpl]
		446	add edi, [_bottom_total_bpl]
		447	dec [_copy_height]
		448	jnz .line
		449
		450	ret
		451	endp
		452
		453
		454	proc xcf._.composite_indexed_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		455
		456	.line:
		457	mov ecx, [_copy_width]
		458	.pixel:
		459	mov bx, [edi]
		460	lodsw
		461
		462	or ah, 0x7f
		463	test ah, 0x80
		464	jnz @f
		465	mov ax, bx
		466	@@:
		467	stosw
		468
		469	dec ecx
		470	jnz .pixel
		471	add esi, [_img_total_bpl]
		472	add edi, [_bottom_total_bpl]
		473	dec [_copy_height]
		474	jnz .line
		475	ret
		476	endp
		477
		478
		479	proc xcf._.composite_rgb_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		480	pushad
		481
		482	pxor xmm4, xmm4
		483	movd xmm4, [xcf._.random_b]
		484	movd xmm1, [xcf._.random_a]
		485	movd xmm2, [xcf._.random_c]
		486
		487	.line:
		488	mov ecx, [_copy_width]
		489	.pixel:
		490	mov ebx, [edi]
		491	lodsd
		492
		493	movq xmm0, xmm4
		494	pmuludq xmm0, xmm1
		495	paddq xmm0, xmm2
		496	movd edx, xmm0
		497	movd xmm4, edx
		498	pxor xmm0, xmm0
		499
		500	rol eax, 8
		501	test al, al
		502	jz @f
		503	shr edx, 17
		504	cmp dl, al
		505	ja @f
		506	ror eax, 8
		507	or eax, 0xff000000
		508	jmp .done
		509	@@:
		510	mov eax, ebx
		511	.done:
		512	stosd
		513	dec ecx
		514	jnz .pixel
		515	add esi, [_img_total_bpl]
		516	add edi, [_bottom_total_bpl]
		517	dec [_copy_height]
		518	jnz .line
		519
		520	.quit:
		521	popad
		522	ret
		523	endp
		524
		525
		526	proc xcf._.composite_gray_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		527	pushad
		528
		529	pxor xmm4, xmm4
		530	movd xmm4, [xcf._.random_b]
		531	movd xmm1, [xcf._.random_a]
		532	movd xmm2, [xcf._.random_c]
		533
		534	.line:
		535	mov ecx, [_copy_width]
		536	.pixel:
		537	mov ebx, [edi]
		538	lodsw
		539
		540	movq xmm0, xmm4
		541	pmuludq xmm0, xmm1
		542	paddq xmm0, xmm2
		543	movd edx, xmm0
		544	movd xmm4, edx
		545	pxor xmm0, xmm0
		546
		547	test ah, ah
		548	jz @f
		549	shr edx, 17
		550	cmp dl, ah
		551	ja @f
		552	or ax, 0xff00
		553	jmp .done
		554	@@:
		555	mov eax, ebx
		556	.done:
		557	stosw
		558	dec ecx
		559	jnz .pixel
		560	add esi, [_img_total_bpl]
		561	add edi, [_bottom_total_bpl]
		562	dec [_copy_height]
		563	jnz .line
		564
		565	.quit:
		566	popad
		567	ret
		568	endp
		569
		570
		571	proc xcf._.composite_rgb_03 ; Multiply
		572
		573	punpcklbw xmm2, xmm0
		574	punpcklbw xmm3, xmm0
		575	pmullw xmm3, xmm2
		576	psrlw xmm3, 8
		577
		578	ret
		579	endp
		580
		581
		582	proc xcf._.composite_rgb_04 ; Screen
		583
		584	punpcklbw xmm2, xmm0
		585	punpcklbw xmm3, xmm0
		586	movdqu xmm5, xword[xcf._.xmm_00ff]
		587	movdqa xmm4, xmm5
		588	psubw xmm4, xmm2
		589	psubw xmm3, xmm5
		590	pmullw xmm3, xmm4
		591	psrlw xmm3, 8
		592	paddw xmm3, xmm5
		593	ret
		594	endp
		595
		596
		597	proc xcf._.composite_rgb_05 ; Overlay
		598
		599	punpcklbw xmm2, xmm0
		600	punpcklbw xmm3, xmm0
		601	movdqu xmm4, xword[xcf._.xmm_00ff]
		602	psubw xmm4, xmm2
		603	pmullw xmm3, xmm4
		604	psrlw xmm3, 7
		605	paddw xmm3, xmm2
		606	pmullw xmm3, xmm2
		607	psrlw xmm3, 8
		608
		609	ret
		610	endp
		611
		612
		613	proc xcf._.composite_rgb_06 ; Difference
		614
		615	movdqa xmm4, xmm3
		616	pminub xmm4, xmm2
		617	pmaxub xmm3, xmm2
		618	psubusb xmm3, xmm4
		619	punpcklbw xmm2, xmm0
		620	punpcklbw xmm3, xmm0
		621
		622	ret
		623	endp
		624
		625
		626	proc xcf._.composite_rgb_07 ; Addition
		627
		628	paddusb xmm3, xmm2
		629	punpcklbw xmm2, xmm0
		630	punpcklbw xmm3, xmm0
		631
		632	ret
		633	endp
		634
		635
		636	proc xcf._.composite_rgb_08 ; Subtract
		637
		638	movdqa xmm4, xmm2
		639	psubusb xmm4, xmm3
		640	movq xmm3, xmm4
		641	punpcklbw xmm2, xmm0
		642	punpcklbw xmm3, xmm0
		643
		644	ret
		645	endp
		646
		647
		648	proc xcf._.composite_rgb_09 ; Darken Only
		649
		650	pminub xmm3, xmm2
		651	punpcklbw xmm2, xmm0
		652	punpcklbw xmm3, xmm0
		653
		654	ret
		655	endp
		656
		657
		658	proc xcf._.composite_rgb_10 ; Lighten Only
		659
		660	pmaxub xmm3, xmm2
		661	punpcklbw xmm2, xmm0
		662	punpcklbw xmm3, xmm0
		663
		664	ret
		665	endp
		666
		667
		668	proc xcf._.composite_rgb_11 ; Hue (H of HSV)
		669	push eax ebx ecx edx
		670
		671	movd eax, xmm3
		672	movd ebx, xmm2
		673
		674	call xcf._.rgb2hsv
		675	xchg eax, ebx
		676	call xcf._.rgb2hsv
		677	xchg eax, ebx
		678
		679	test ah, ah
		680	jnz @f
		681	ror eax, 8
		682	ror ebx, 8
		683	mov ah, bh
		684	rol eax, 8
		685	rol ebx, 8
		686	@@:
		687	mov ax, bx
		688
		689	call xcf._.hsv2rgb
		690
		691	push eax
		692
		693	movq xmm1, xmm3
		694	psrldq xmm1, 4
		695	movd eax, xmm1
		696	movq xmm1, xmm2
		697	psrldq xmm1, 4
		698	movd ebx, xmm1
		699
		700	call xcf._.rgb2hsv
		701	xchg eax, ebx
		702	call xcf._.rgb2hsv
		703	xchg eax, ebx
		704
		705	test ah, ah
		706	jnz @f
		707	ror eax, 8
		708	ror ebx, 8
		709	mov ah, bh
		710	rol eax, 8
		711	rol ebx, 8
		712	@@:
		713	mov ax, bx
		714
		715	call xcf._.hsv2rgb
		716
		717	movd xmm3, eax
		718	pslldq xmm3, 4
		719	pop eax
		720	movd xmm1, eax
		721	paddq xmm3, xmm1
		722
		723	punpcklbw xmm2, xmm0
		724	punpcklbw xmm3, xmm0
		725
		726	.quit:
		727	pop edx ecx ebx eax
		728	ret
		729	endp
		730
		731
		732	proc xcf._.composite_rgb_12 ; Saturation (S of HSV)
		733	push eax ebx ecx edx
		734
		735	movd eax, xmm3
		736	movd ebx, xmm2
		737
		738	call xcf._.rgb2hsv
		739	xchg eax, ebx
		740	call xcf._.rgb2hsv
		741	xchg eax, ebx
		742
		743	ror eax, 8
		744	ror ebx, 8
		745	mov ah, bh
		746	rol eax, 8
		747	rol ebx, 8
		748	mov al, bl
		749
		750	call xcf._.hsv2rgb
		751
		752	push eax
		753	movq xmm1, xmm3
		754	psrldq xmm1, 4
		755	movd eax, xmm1
		756	movq xmm1, xmm2
		757	psrldq xmm1, 4
		758	movd ebx, xmm1
		759
		760	call xcf._.rgb2hsv
		761	xchg eax, ebx
		762	call xcf._.rgb2hsv
		763	xchg eax, ebx
		764
		765	ror eax, 8
		766	ror ebx, 8
		767	mov ah, bh
		768	rol eax, 8
		769	rol ebx, 8
		770	mov al, bl
		771
		772	call xcf._.hsv2rgb
		773
		774
		775	movd xmm3, eax
		776	pslldq xmm3, 4
		777	pop eax
		778	movd xmm1, eax
		779	paddq xmm3, xmm1
		780
		781	punpcklbw xmm2, xmm0
		782	punpcklbw xmm3, xmm0
		783
		784	.quit:
		785	pop edx ecx ebx eax
		786	ret
		787	endp
		788
		789
		790	proc xcf._.composite_rgb_13 ; Color (H and S of HSL)
		791	push eax ebx ecx edx
		792
		793	movd eax, xmm3
		794	movd ebx, xmm2
		795
		796	call xcf._.rgb2hsl
		797	xchg eax, ebx
		798	call xcf._.rgb2hsl
		799	xchg eax, ebx
		800
		801	mov al, bl
		802
		803	call xcf._.hsl2rgb
		804
		805	push eax
		806	movq xmm1, xmm3
		807	psrldq xmm1, 4
		808	movd eax, xmm1
		809	movq xmm1, xmm2
		810	psrldq xmm1, 4
		811	movd ebx, xmm1
		812
		813	call xcf._.rgb2hsl
		814	xchg eax, ebx
		815	call xcf._.rgb2hsl
		816	xchg eax, ebx
		817
		818	mov al, bl
		819
		820	call xcf._.hsl2rgb
		821
		822	movd xmm3, eax
		823	pslldq xmm3, 4
		824	pop eax
		825	movd xmm1, eax
		826	paddq xmm3, xmm1
		827
		828	punpcklbw xmm2, xmm0
		829	punpcklbw xmm3, xmm0
		830
		831	.quit:
		832	pop edx ecx ebx eax
		833	ret
		834	endp
		835
		836
		837	proc xcf._.composite_rgb_14 ; Value (V of HSV)
		838	push eax ebx ecx edx
		839
		840	movd eax, xmm3
		841	movd ebx, xmm2
		842
		843	call xcf._.rgb2hsv
		844	xchg eax, ebx
		845	call xcf._.rgb2hsv
		846	xchg eax, ebx
		847
		848	ror eax, 8
		849	ror ebx, 8
		850	mov ax, bx
		851	rol eax, 8
		852	rol ebx, 8
		853
		854	call xcf._.hsv2rgb
		855
		856	push eax
		857	movq xmm1, xmm3
		858	psrldq xmm1, 4
		859	movd eax, xmm1
		860	movq xmm1, xmm2
		861	psrldq xmm1, 4
		862	movd ebx, xmm1
		863
		864	call xcf._.rgb2hsv
		865	xchg eax, ebx
		866	call xcf._.rgb2hsv
		867	xchg eax, ebx
		868
		869	ror eax, 8
		870	ror ebx, 8
		871	mov ax, bx
		872	rol eax, 8
		873	rol ebx, 8
		874
		875	call xcf._.hsv2rgb
		876
		877	movd xmm3, eax
		878	pslldq xmm3, 4
		879	pop eax
		880	movd xmm1, eax
		881	paddq xmm3, xmm1
		882
		883	punpcklbw xmm2, xmm0
		884	punpcklbw xmm3, xmm0
		885
		886	.quit:
		887	pop edx ecx ebx eax
		888	ret
		889	endp
		890
		891
		892	proc xcf._.composite_rgb_15 ; Divide
		893	push eax ebx ecx
		894
		895	movd eax, xmm3
		896	movd ebx, xmm2
		897
		898	rol eax, 8
		899	rol ebx, 8
		900
		901	xchg eax, ebx
		902
		903	mov ecx, 3
		904
		905	.color:
		906	rol eax, 8
		907	rol ebx, 8
		908	shl ax, 8
		909	test bl, bl
		910	jz .clamp1
		911	cmp ah, bl
		912	jae .clamp2
		913	div bl
		914	jmp .done
		915	.clamp1:
		916	mov al, 0xff
		917	test ah, ah
		918	jnz @f
		919	not al
		920	@@:
		921	jmp .done
		922	.clamp2:
		923	mov al, 0xff
		924	jmp .done
		925	.done:
		926	mov ah, al
		927	loop .color
		928
		929	ror eax, 8
		930
		931
		932	push eax
		933	movq xmm1, xmm3
		934	psrldq xmm1, 4
		935	movd eax, xmm1
		936	movq xmm1, xmm2
		937	psrldq xmm1, 4
		938	movd ebx, xmm1
		939
		940
		941	rol eax, 8
		942	rol ebx, 8
		943
		944	xchg eax, ebx
		945
		946	mov ecx, 3
		947
		948	.color2:
		949	rol eax, 8
		950	rol ebx, 8
		951	shl ax, 8
		952	test bl, bl
		953	jz .clamp12
		954	cmp ah, bl
		955	jae .clamp22
		956	div bl
		957	jmp .done2
		958	.clamp12:
		959	mov al, 0xff
		960	test ah, ah
		961	jnz @f
		962	not al
		963	@@:
		964	jmp .done2
		965	.clamp22:
		966	mov al, 0xff
		967	jmp .done2
		968	.done2:
		969	mov ah, al
		970	loop .color2
		971
		972	ror eax, 8
		973
		974
		975	movd xmm3, eax
		976	pslldq xmm3, 4
		977	pop eax
		978	movd xmm1, eax
		979	paddq xmm3, xmm1
		980
		981	punpcklbw xmm2, xmm0
		982	punpcklbw xmm3, xmm0
		983
		984	pop ecx ebx eax
		985	ret
		986	endp
		987
		988
		989	proc xcf._.composite_rgb_16 ; Dodge
		990	push eax ebx ecx
		991
		992	movd eax, xmm3
		993	movd ebx, xmm2
		994
		995	rol eax, 8
		996	rol ebx, 8
		997
		998	xchg eax, ebx
		999
		1000	mov ecx, 3
		1001
		1002	.color:
		1003	rol eax, 8
		1004	rol ebx, 8
		1005	shl ax, 8
		1006	neg bl
		1007	add bl, 0xff
		1008	test bl, bl
		1009	jz .clamp1
		1010	cmp ah, bl
		1011	jae .clamp2
		1012	div bl
		1013	jmp .done
		1014	.clamp1:
		1015	mov al, 0xff
		1016	test ah, ah
		1017	jnz @f
		1018	not al
		1019	@@:
		1020	jmp .done
		1021	.clamp2:
		1022	mov al, 0xff
		1023	jmp .done
		1024	.done:
		1025	mov ah, al
		1026	loop .color
		1027
		1028	ror eax, 8
		1029
		1030
		1031	push eax
		1032	movq xmm1, xmm3
		1033	psrldq xmm1, 4
		1034	movd eax, xmm1
		1035	movq xmm1, xmm2
		1036	psrldq xmm1, 4
		1037	movd ebx, xmm1
		1038
		1039
		1040	rol eax, 8
		1041	rol ebx, 8
		1042
		1043	xchg eax, ebx
		1044
		1045	mov ecx, 3
		1046
		1047	.color2:
		1048	rol eax, 8
		1049	rol ebx, 8
		1050	shl ax, 8
		1051	neg bl
		1052	add bl, 0xff
		1053	test bl, bl
		1054	jz .clamp12
		1055	cmp ah, bl
		1056	jae .clamp22
		1057	div bl
		1058	jmp .done2
		1059	.clamp12:
		1060	mov al, 0xff
		1061	test ah, ah
		1062	jnz @f
		1063	not al
		1064	@@:
		1065	jmp .done2
		1066	.clamp22:
		1067	mov al, 0xff
		1068	jmp .done2
		1069	.done2:
		1070	mov ah, al
		1071	loop .color2
		1072
		1073	ror eax, 8
		1074
		1075
		1076	movd xmm3, eax
		1077	pslldq xmm3, 4
		1078	pop eax
		1079	movd xmm1, eax
		1080	paddq xmm3, xmm1
		1081
		1082	punpcklbw xmm2, xmm0
		1083	punpcklbw xmm3, xmm0
		1084
		1085	pop ecx ebx eax
		1086	ret
		1087	endp
		1088
		1089
		1090	proc xcf._.composite_rgb_17 ; Burn
		1091	push eax ebx ecx
		1092
		1093	movd eax, xmm3
		1094	movd ebx, xmm2
		1095
		1096	rol eax, 8
		1097	rol ebx, 8
		1098
		1099	xchg eax, ebx
		1100
		1101	mov ecx, 3
		1102
		1103	.color:
		1104	rol eax, 8
		1105	rol ebx, 8
		1106	shl ax, 8
		1107	neg ah
		1108	add ah, 0xff
		1109	test bl, bl
		1110	jz .clamp1
		1111	cmp ah, bl
		1112	jae .clamp2
		1113	div bl
		1114	jmp .done
		1115	.clamp1:
		1116	mov al, 0xff
		1117	test ah, ah
		1118	jnz @f
		1119	not al
		1120	@@:
		1121	jmp .done
		1122	.clamp2:
		1123	mov al, 0xff
		1124	jmp .done
		1125	.done:
		1126	mov ah, al
		1127	neg ah
		1128	add ah, 0xff
		1129	loop .color
		1130
		1131	ror eax, 8
		1132
		1133
		1134	push eax
		1135	movq xmm1, xmm3
		1136	psrldq xmm1, 4
		1137	movd eax, xmm1
		1138	movq xmm1, xmm2
		1139	psrldq xmm1, 4
		1140	movd ebx, xmm1
		1141
		1142
		1143	rol eax, 8
		1144	rol ebx, 8
		1145
		1146	xchg eax, ebx
		1147
		1148	mov ecx, 3
		1149
		1150	.color2:
		1151	rol eax, 8
		1152	rol ebx, 8
		1153	shl ax, 8
		1154	neg ah
		1155	add ah, 0xff
		1156	test bl, bl
		1157	jz .clamp12
		1158	cmp ah, bl
		1159	jae .clamp22
		1160	div bl
		1161	jmp .done2
		1162	.clamp12:
		1163	mov al, 0xff
		1164	test ah, ah
		1165	jnz @f
		1166	not al
		1167	@@:
		1168	jmp .done2
		1169	.clamp22:
		1170	mov al, 0xff
		1171	jmp .done2
		1172	.done2:
		1173	mov ah, al
		1174	neg ah
		1175	add ah, 0xff
		1176	loop .color2
		1177
		1178	ror eax, 8
		1179
		1180
		1181	movd xmm3, eax
		1182	pslldq xmm3, 4
		1183	pop eax
		1184	movd xmm1, eax
		1185	paddq xmm3, xmm1
		1186
		1187	punpcklbw xmm2, xmm0
		1188	punpcklbw xmm3, xmm0
		1189
		1190	pop ecx ebx eax
		1191	ret
		1192	endp
		1193
		1194
		1195	proc xcf._.composite_rgb_18 ; Hard Light
		1196	push eax ebx ecx
		1197
		1198	movd eax, xmm3
		1199	movd ebx, xmm2
		1200
		1201	rol eax, 8
		1202	rol ebx, 8
		1203
		1204	mov ecx, 3
		1205
		1206	.color:
		1207	rol eax, 8
		1208	rol ebx, 8
		1209	cmp al, 127
		1210	jna .part1
		1211	mov ah, 0xff
		1212	sub ah, bl
		1213	neg al
		1214	add al, 0xff
		1215	mul ah
		1216	shl ax, 1
		1217	neg ah
		1218	add ah, 0xff
		1219	jmp .done
		1220	.part1:
		1221	mul bl
		1222	shl ax, 1
		1223	.done:
		1224	loop .color
		1225
		1226	ror eax, 8
		1227
		1228
		1229	push eax
		1230	movq xmm1, xmm3
		1231	psrldq xmm1, 4
		1232	movd eax, xmm1
		1233	movq xmm1, xmm2
		1234	psrldq xmm1, 4
		1235	movd ebx, xmm1
		1236
		1237
		1238	rol eax, 8
		1239	rol ebx, 8
		1240
		1241	mov ecx, 3
		1242
		1243	.color2:
		1244	rol eax, 8
		1245	rol ebx, 8
		1246	cmp al, 127
		1247	jna .part12
		1248	mov ah, 0xff
		1249	sub ah, bl
		1250	neg al
		1251	add al, 0xff
		1252	mul ah
		1253	shl ax, 1
		1254	neg ah
		1255	add ah, 0xff
		1256	jmp .done2
		1257	.part12:
		1258	mul bl
		1259	shl ax, 1
		1260	.done2:
		1261	loop .color2
		1262
		1263	ror eax, 8
		1264
		1265
		1266	movd xmm3, eax
		1267	pslldq xmm3, 4
		1268	pop eax
		1269	movd xmm1, eax
		1270	paddq xmm3, xmm1
		1271
		1272	punpcklbw xmm2, xmm0
		1273	punpcklbw xmm3, xmm0
		1274
		1275	pop ecx ebx eax
		1276	ret
		1277	endp
		1278
		1279
		1280	proc xcf._.composite_rgb_20 ; Grain Extract
		1281
		1282	punpcklbw xmm2, xmm0
		1283	punpcklbw xmm3, xmm0
		1284	movdqu xmm4, xmm2
		1285	psubw xmm3, xword[xcf._.xmm_0080]
		1286	psubw xmm4, xmm3
		1287	movdqa xmm3, xmm4
		1288	packuswb xmm3, xmm0
		1289	punpcklbw xmm3, xmm0
		1290	ret
		1291	endp
		1292
		1293
		1294	proc xcf._.composite_rgb_21 ; Grain Merge
		1295
		1296	punpcklbw xmm2, xmm0
		1297	punpcklbw xmm3, xmm0
		1298	paddw xmm3, xmm2
		1299	psubusw xmm3, xword[xcf._.xmm_0080]
		1300	packuswb xmm3, xmm0
		1301	punpcklbw xmm3, xmm0
		1302	ret
		1303	endp
		1304
		1305
		1306	; starting numbers for pseudo-random number generator
		1307	xcf._.random_a dd 1103515245
		1308	xcf._.random_b dd 777
		1309	xcf._.random_c dd 12345
		1310
		1311	xcf._.xmm_8080 dq 0x8080808080808080, 0x8080808080808080
		1312	xcf._.xmm_0080 dq 0x0080008000800080, 0x0080008000800080
		1313	xcf._.xmm_00ff dq 0x00ff00ff00ff00ff, 0x00ff00ff00ff00ff
		1314	xcf._.xmm_0100 dq 0x0100010001000100, 0x0100010001000100
		1315	xcf._.xmm_000000ff dq 0x000000ff000000ff, 0x0000000000000000

Subversion Repositories Kolibri OS

(root)/programs/develop/libraries/libs-dev/libimg/xcf/composite_sse.asm @ 2733 – Rev 2388