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
2733	dunkaist	586	movdqu xmm4, xword[xcf._.xmm_00ff]
		587	movdqa xmm5, xmm4
		588	psubw xmm5, xmm3
		589	movdqa xmm3, xmm4
2388	dunkaist	590	psubw xmm4, xmm2
2733	dunkaist	591	pmullw xmm4, xmm5
		592	psrlw xmm4, 8
		593	psubw xmm3, xmm4
2388	dunkaist	594	ret
		595	endp
		596
		597
		598	proc xcf._.composite_rgb_05 ; Overlay
		599
		600	punpcklbw xmm2, xmm0
		601	punpcklbw xmm3, xmm0
		602	movdqu xmm4, xword[xcf._.xmm_00ff]
		603	psubw xmm4, xmm2
		604	pmullw xmm3, xmm4
		605	psrlw xmm3, 7
		606	paddw xmm3, xmm2
		607	pmullw xmm3, xmm2
		608	psrlw xmm3, 8
		609
		610	ret
		611	endp
		612
		613
		614	proc xcf._.composite_rgb_06 ; Difference
		615
		616	movdqa xmm4, xmm3
		617	pminub xmm4, xmm2
		618	pmaxub xmm3, xmm2
		619	psubusb xmm3, xmm4
		620	punpcklbw xmm2, xmm0
		621	punpcklbw xmm3, xmm0
		622
		623	ret
		624	endp
		625
		626
		627	proc xcf._.composite_rgb_07 ; Addition
		628
		629	paddusb xmm3, xmm2
		630	punpcklbw xmm2, xmm0
		631	punpcklbw xmm3, xmm0
		632
		633	ret
		634	endp
		635
		636
		637	proc xcf._.composite_rgb_08 ; Subtract
		638
		639	movdqa xmm4, xmm2
		640	psubusb xmm4, xmm3
		641	movq xmm3, xmm4
		642	punpcklbw xmm2, xmm0
		643	punpcklbw xmm3, xmm0
		644
		645	ret
		646	endp
		647
		648
		649	proc xcf._.composite_rgb_09 ; Darken Only
		650
		651	pminub xmm3, xmm2
		652	punpcklbw xmm2, xmm0
		653	punpcklbw xmm3, xmm0
		654
		655	ret
		656	endp
		657
		658
		659	proc xcf._.composite_rgb_10 ; Lighten Only
		660
		661	pmaxub xmm3, xmm2
		662	punpcklbw xmm2, xmm0
		663	punpcklbw xmm3, xmm0
		664
		665	ret
		666	endp
		667
		668
		669	proc xcf._.composite_rgb_11 ; Hue (H of HSV)
		670	push eax ebx ecx edx
		671
		672	movd eax, xmm3
		673	movd ebx, xmm2
		674
		675	call xcf._.rgb2hsv
		676	xchg eax, ebx
		677	call xcf._.rgb2hsv
		678	xchg eax, ebx
		679
		680	test ah, ah
		681	jnz @f
		682	ror eax, 8
		683	ror ebx, 8
		684	mov ah, bh
		685	rol eax, 8
		686	rol ebx, 8
		687	@@:
		688	mov ax, bx
		689
		690	call xcf._.hsv2rgb
		691
		692	push eax
		693
		694	movq xmm1, xmm3
		695	psrldq xmm1, 4
		696	movd eax, xmm1
		697	movq xmm1, xmm2
		698	psrldq xmm1, 4
		699	movd ebx, xmm1
		700
		701	call xcf._.rgb2hsv
		702	xchg eax, ebx
		703	call xcf._.rgb2hsv
		704	xchg eax, ebx
		705
		706	test ah, ah
		707	jnz @f
		708	ror eax, 8
		709	ror ebx, 8
		710	mov ah, bh
		711	rol eax, 8
		712	rol ebx, 8
		713	@@:
		714	mov ax, bx
		715
		716	call xcf._.hsv2rgb
		717
		718	movd xmm3, eax
		719	pslldq xmm3, 4
		720	pop eax
		721	movd xmm1, eax
		722	paddq xmm3, xmm1
		723
		724	punpcklbw xmm2, xmm0
		725	punpcklbw xmm3, xmm0
		726
		727	.quit:
		728	pop edx ecx ebx eax
		729	ret
		730	endp
		731
		732
		733	proc xcf._.composite_rgb_12 ; Saturation (S of HSV)
		734	push eax ebx ecx edx
		735
		736	movd eax, xmm3
		737	movd ebx, xmm2
		738
		739	call xcf._.rgb2hsv
		740	xchg eax, ebx
		741	call xcf._.rgb2hsv
		742	xchg eax, ebx
		743
		744	ror eax, 8
		745	ror ebx, 8
		746	mov ah, bh
		747	rol eax, 8
		748	rol ebx, 8
		749	mov al, bl
		750
		751	call xcf._.hsv2rgb
		752
		753	push eax
		754	movq xmm1, xmm3
		755	psrldq xmm1, 4
		756	movd eax, xmm1
		757	movq xmm1, xmm2
		758	psrldq xmm1, 4
		759	movd ebx, xmm1
		760
		761	call xcf._.rgb2hsv
		762	xchg eax, ebx
		763	call xcf._.rgb2hsv
		764	xchg eax, ebx
		765
		766	ror eax, 8
		767	ror ebx, 8
		768	mov ah, bh
		769	rol eax, 8
		770	rol ebx, 8
		771	mov al, bl
		772
		773	call xcf._.hsv2rgb
		774
		775
		776	movd xmm3, eax
		777	pslldq xmm3, 4
		778	pop eax
		779	movd xmm1, eax
		780	paddq xmm3, xmm1
		781
		782	punpcklbw xmm2, xmm0
		783	punpcklbw xmm3, xmm0
		784
		785	.quit:
		786	pop edx ecx ebx eax
		787	ret
		788	endp
		789
		790
		791	proc xcf._.composite_rgb_13 ; Color (H and S of HSL)
		792	push eax ebx ecx edx
		793
		794	movd eax, xmm3
		795	movd ebx, xmm2
		796
		797	call xcf._.rgb2hsl
		798	xchg eax, ebx
		799	call xcf._.rgb2hsl
		800	xchg eax, ebx
		801
		802	mov al, bl
		803
		804	call xcf._.hsl2rgb
		805
		806	push eax
		807	movq xmm1, xmm3
		808	psrldq xmm1, 4
		809	movd eax, xmm1
		810	movq xmm1, xmm2
		811	psrldq xmm1, 4
		812	movd ebx, xmm1
		813
		814	call xcf._.rgb2hsl
		815	xchg eax, ebx
		816	call xcf._.rgb2hsl
		817	xchg eax, ebx
		818
		819	mov al, bl
		820
		821	call xcf._.hsl2rgb
		822
		823	movd xmm3, eax
		824	pslldq xmm3, 4
		825	pop eax
		826	movd xmm1, eax
		827	paddq xmm3, xmm1
		828
		829	punpcklbw xmm2, xmm0
		830	punpcklbw xmm3, xmm0
		831
		832	.quit:
		833	pop edx ecx ebx eax
		834	ret
		835	endp
		836
		837
		838	proc xcf._.composite_rgb_14 ; Value (V of HSV)
		839	push eax ebx ecx edx
		840
		841	movd eax, xmm3
		842	movd ebx, xmm2
		843
		844	call xcf._.rgb2hsv
		845	xchg eax, ebx
		846	call xcf._.rgb2hsv
		847	xchg eax, ebx
		848
		849	ror eax, 8
		850	ror ebx, 8
		851	mov ax, bx
		852	rol eax, 8
		853	rol ebx, 8
		854
		855	call xcf._.hsv2rgb
		856
		857	push eax
		858	movq xmm1, xmm3
		859	psrldq xmm1, 4
		860	movd eax, xmm1
		861	movq xmm1, xmm2
		862	psrldq xmm1, 4
		863	movd ebx, xmm1
		864
		865	call xcf._.rgb2hsv
		866	xchg eax, ebx
		867	call xcf._.rgb2hsv
		868	xchg eax, ebx
		869
		870	ror eax, 8
		871	ror ebx, 8
		872	mov ax, bx
		873	rol eax, 8
		874	rol ebx, 8
		875
		876	call xcf._.hsv2rgb
		877
		878	movd xmm3, eax
		879	pslldq xmm3, 4
		880	pop eax
		881	movd xmm1, eax
		882	paddq xmm3, xmm1
		883
		884	punpcklbw xmm2, xmm0
		885	punpcklbw xmm3, xmm0
		886
		887	.quit:
		888	pop edx ecx ebx eax
		889	ret
		890	endp
		891
		892
		893	proc xcf._.composite_rgb_15 ; Divide
		894	push eax ebx ecx
		895
		896	movd eax, xmm3
		897	movd ebx, xmm2
		898
		899	rol eax, 8
		900	rol ebx, 8
		901
		902	xchg eax, ebx
		903
		904	mov ecx, 3
		905
		906	.color:
		907	rol eax, 8
		908	rol ebx, 8
		909	shl ax, 8
		910	test bl, bl
		911	jz .clamp1
		912	cmp ah, bl
		913	jae .clamp2
		914	div bl
		915	jmp .done
		916	.clamp1:
		917	mov al, 0xff
		918	test ah, ah
		919	jnz @f
		920	not al
		921	@@:
		922	jmp .done
		923	.clamp2:
		924	mov al, 0xff
		925	jmp .done
		926	.done:
		927	mov ah, al
		928	loop .color
		929
		930	ror eax, 8
		931
		932
		933	push eax
		934	movq xmm1, xmm3
		935	psrldq xmm1, 4
		936	movd eax, xmm1
		937	movq xmm1, xmm2
		938	psrldq xmm1, 4
		939	movd ebx, xmm1
		940
		941
		942	rol eax, 8
		943	rol ebx, 8
		944
		945	xchg eax, ebx
		946
		947	mov ecx, 3
		948
		949	.color2:
		950	rol eax, 8
		951	rol ebx, 8
		952	shl ax, 8
		953	test bl, bl
		954	jz .clamp12
		955	cmp ah, bl
		956	jae .clamp22
		957	div bl
		958	jmp .done2
		959	.clamp12:
		960	mov al, 0xff
		961	test ah, ah
		962	jnz @f
		963	not al
		964	@@:
		965	jmp .done2
		966	.clamp22:
		967	mov al, 0xff
		968	jmp .done2
		969	.done2:
		970	mov ah, al
		971	loop .color2
		972
		973	ror eax, 8
		974
		975
		976	movd xmm3, eax
		977	pslldq xmm3, 4
		978	pop eax
		979	movd xmm1, eax
		980	paddq xmm3, xmm1
		981
		982	punpcklbw xmm2, xmm0
		983	punpcklbw xmm3, xmm0
		984
		985	pop ecx ebx eax
		986	ret
		987	endp
		988
		989
		990	proc xcf._.composite_rgb_16 ; Dodge
		991	push eax ebx ecx
		992
		993	movd eax, xmm3
		994	movd ebx, xmm2
		995
		996	rol eax, 8
		997	rol ebx, 8
		998
		999	xchg eax, ebx
		1000
		1001	mov ecx, 3
		1002
		1003	.color:
		1004	rol eax, 8
		1005	rol ebx, 8
		1006	shl ax, 8
		1007	neg bl
		1008	add bl, 0xff
		1009	test bl, bl
		1010	jz .clamp1
		1011	cmp ah, bl
		1012	jae .clamp2
		1013	div bl
		1014	jmp .done
		1015	.clamp1:
		1016	mov al, 0xff
		1017	test ah, ah
		1018	jnz @f
		1019	not al
		1020	@@:
		1021	jmp .done
		1022	.clamp2:
		1023	mov al, 0xff
		1024	jmp .done
		1025	.done:
		1026	mov ah, al
		1027	loop .color
		1028
		1029	ror eax, 8
		1030
		1031
		1032	push eax
		1033	movq xmm1, xmm3
		1034	psrldq xmm1, 4
		1035	movd eax, xmm1
		1036	movq xmm1, xmm2
		1037	psrldq xmm1, 4
		1038	movd ebx, xmm1
		1039
		1040
		1041	rol eax, 8
		1042	rol ebx, 8
		1043
		1044	xchg eax, ebx
		1045
		1046	mov ecx, 3
		1047
		1048	.color2:
		1049	rol eax, 8
		1050	rol ebx, 8
		1051	shl ax, 8
		1052	neg bl
		1053	add bl, 0xff
		1054	test bl, bl
		1055	jz .clamp12
		1056	cmp ah, bl
		1057	jae .clamp22
		1058	div bl
		1059	jmp .done2
		1060	.clamp12:
		1061	mov al, 0xff
		1062	test ah, ah
		1063	jnz @f
		1064	not al
		1065	@@:
		1066	jmp .done2
		1067	.clamp22:
		1068	mov al, 0xff
		1069	jmp .done2
		1070	.done2:
		1071	mov ah, al
		1072	loop .color2
		1073
		1074	ror eax, 8
		1075
		1076
		1077	movd xmm3, eax
		1078	pslldq xmm3, 4
		1079	pop eax
		1080	movd xmm1, eax
		1081	paddq xmm3, xmm1
		1082
		1083	punpcklbw xmm2, xmm0
		1084	punpcklbw xmm3, xmm0
		1085
		1086	pop ecx ebx eax
		1087	ret
		1088	endp
		1089
		1090
		1091	proc xcf._.composite_rgb_17 ; Burn
		1092	push eax ebx ecx
		1093
		1094	movd eax, xmm3
		1095	movd ebx, xmm2
		1096
		1097	rol eax, 8
		1098	rol ebx, 8
		1099
		1100	xchg eax, ebx
		1101
		1102	mov ecx, 3
		1103
		1104	.color:
		1105	rol eax, 8
		1106	rol ebx, 8
		1107	shl ax, 8
		1108	neg ah
		1109	add ah, 0xff
		1110	test bl, bl
		1111	jz .clamp1
		1112	cmp ah, bl
		1113	jae .clamp2
		1114	div bl
		1115	jmp .done
		1116	.clamp1:
		1117	mov al, 0xff
		1118	test ah, ah
		1119	jnz @f
		1120	not al
		1121	@@:
		1122	jmp .done
		1123	.clamp2:
		1124	mov al, 0xff
		1125	jmp .done
		1126	.done:
		1127	mov ah, al
		1128	neg ah
		1129	add ah, 0xff
		1130	loop .color
		1131
		1132	ror eax, 8
		1133
		1134
		1135	push eax
		1136	movq xmm1, xmm3
		1137	psrldq xmm1, 4
		1138	movd eax, xmm1
		1139	movq xmm1, xmm2
		1140	psrldq xmm1, 4
		1141	movd ebx, xmm1
		1142
		1143
		1144	rol eax, 8
		1145	rol ebx, 8
		1146
		1147	xchg eax, ebx
		1148
		1149	mov ecx, 3
		1150
		1151	.color2:
		1152	rol eax, 8
		1153	rol ebx, 8
		1154	shl ax, 8
		1155	neg ah
		1156	add ah, 0xff
		1157	test bl, bl
		1158	jz .clamp12
		1159	cmp ah, bl
		1160	jae .clamp22
		1161	div bl
		1162	jmp .done2
		1163	.clamp12:
		1164	mov al, 0xff
		1165	test ah, ah
		1166	jnz @f
		1167	not al
		1168	@@:
		1169	jmp .done2
		1170	.clamp22:
		1171	mov al, 0xff
		1172	jmp .done2
		1173	.done2:
		1174	mov ah, al
		1175	neg ah
		1176	add ah, 0xff
		1177	loop .color2
		1178
		1179	ror eax, 8
		1180
		1181
		1182	movd xmm3, eax
		1183	pslldq xmm3, 4
		1184	pop eax
		1185	movd xmm1, eax
		1186	paddq xmm3, xmm1
		1187
		1188	punpcklbw xmm2, xmm0
		1189	punpcklbw xmm3, xmm0
		1190
		1191	pop ecx ebx eax
		1192	ret
		1193	endp
		1194
		1195
		1196	proc xcf._.composite_rgb_18 ; Hard Light
		1197	push eax ebx ecx
		1198
		1199	movd eax, xmm3
		1200	movd ebx, xmm2
		1201
		1202	rol eax, 8
		1203	rol ebx, 8
		1204
		1205	mov ecx, 3
		1206
		1207	.color:
		1208	rol eax, 8
		1209	rol ebx, 8
		1210	cmp al, 127
		1211	jna .part1
		1212	mov ah, 0xff
		1213	sub ah, bl
		1214	neg al
		1215	add al, 0xff
		1216	mul ah
		1217	shl ax, 1
		1218	neg ah
		1219	add ah, 0xff
		1220	jmp .done
		1221	.part1:
		1222	mul bl
		1223	shl ax, 1
		1224	.done:
		1225	loop .color
		1226
		1227	ror eax, 8
		1228
		1229
		1230	push eax
		1231	movq xmm1, xmm3
		1232	psrldq xmm1, 4
		1233	movd eax, xmm1
		1234	movq xmm1, xmm2
		1235	psrldq xmm1, 4
		1236	movd ebx, xmm1
		1237
		1238
		1239	rol eax, 8
		1240	rol ebx, 8
		1241
		1242	mov ecx, 3
		1243
		1244	.color2:
		1245	rol eax, 8
		1246	rol ebx, 8
		1247	cmp al, 127
		1248	jna .part12
		1249	mov ah, 0xff
		1250	sub ah, bl
		1251	neg al
		1252	add al, 0xff
		1253	mul ah
		1254	shl ax, 1
		1255	neg ah
		1256	add ah, 0xff
		1257	jmp .done2
		1258	.part12:
		1259	mul bl
		1260	shl ax, 1
		1261	.done2:
		1262	loop .color2
		1263
		1264	ror eax, 8
		1265
		1266
		1267	movd xmm3, eax
		1268	pslldq xmm3, 4
		1269	pop eax
		1270	movd xmm1, eax
		1271	paddq xmm3, xmm1
		1272
		1273	punpcklbw xmm2, xmm0
		1274	punpcklbw xmm3, xmm0
		1275
		1276	pop ecx ebx eax
		1277	ret
		1278	endp
		1279
		1280
		1281	proc xcf._.composite_rgb_20 ; Grain Extract
		1282
		1283	punpcklbw xmm2, xmm0
		1284	punpcklbw xmm3, xmm0
		1285	movdqu xmm4, xmm2
		1286	psubw xmm3, xword[xcf._.xmm_0080]
		1287	psubw xmm4, xmm3
		1288	movdqa xmm3, xmm4
		1289	packuswb xmm3, xmm0
		1290	punpcklbw xmm3, xmm0
		1291	ret
		1292	endp
		1293
		1294
		1295	proc xcf._.composite_rgb_21 ; Grain Merge
		1296
		1297	punpcklbw xmm2, xmm0
		1298	punpcklbw xmm3, xmm0
		1299	paddw xmm3, xmm2
		1300	psubusw xmm3, xword[xcf._.xmm_0080]
		1301	packuswb xmm3, xmm0
		1302	punpcklbw xmm3, xmm0
		1303	ret
		1304	endp
		1305
		1306
		1307	; starting numbers for pseudo-random number generator
		1308	xcf._.random_a dd 1103515245
		1309	xcf._.random_b dd 777
		1310	xcf._.random_c dd 12345
		1311
		1312	xcf._.xmm_8080 dq 0x8080808080808080, 0x8080808080808080
		1313	xcf._.xmm_0080 dq 0x0080008000800080, 0x0080008000800080
		1314	xcf._.xmm_00ff dq 0x00ff00ff00ff00ff, 0x00ff00ff00ff00ff
		1315	xcf._.xmm_0100 dq 0x0100010001000100, 0x0100010001000100
		1316	xcf._.xmm_000000ff dq 0x000000ff000000ff, 0x0000000000000000

Subversion Repositories Kolibri OS

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