WebSVN – Kolibri OS – Blame – /programs/develop/libraries/libs-dev/libimg/xcf/composite_mmx.inc

Rev	Author	Line No.	Line
1921	dunkaist	1	proc blend_rgb
		2
		3	xchg al, bh
		4	mov ah, bh
		5	neg ax
		6	add ax, 0xffff
		7	mul ah
		8	neg ah
		9	add ah, 0xff
		10	xchg ah, bh
		11
		12	mov al, 0xff
		13	cmp ah, bh
		14	je @f
		15	not al
		16	div bh
		17	@@:
		18
		19	mov ah, al
		20	movd mm1, eax
		21	; pxor mm0, mm0 ; already xor'ed in composite function
		22	punpcklbw mm1, mm1
		23	punpcklbw mm1, mm0
		24	; punpcklbw mm3, mm0
		25
		26	movq mm7, mm1
		27	psrlw mm7, 7
		28	paddw mm1, mm7
		29
		30	psubw mm3, mm2
		31	pmullw mm3, mm1
		32	psllw mm2, 8
		33	paddw mm3, mm2
		34	pinsrw mm3, ebx, 3
		35	psrlw mm3, 8
		36	packuswb mm3, mm0
		37	movd eax, mm3
		38
		39	ret
		40	endp
		41
		42
		43	proc blend_gray
		44
		45	xchg al, bh
		46	mov ah, bh
		47	neg ax
		48	add ax, 0xffff
		49	mul ah
		50	neg ah
		51	add ah, 0xff
		52	xchg ah, bh
		53
		54	mov al, 0xff
		55	cmp ah, bh
		56	je @f
		57	not al
		58	div bh
		59	@@:
		60
		61	mov ah, al
		62	movd mm1, eax
		63	; pxor mm0, mm0 ; already xor'ed in composite function
		64	punpcklbw mm1, mm1
		65	punpcklbw mm1, mm0
		66	; punpcklbw mm3, mm0
		67
		68	movq mm7, mm1
		69	psrlw mm7, 7
		70	paddw mm1, mm7
		71
		72	psubw mm3, mm2
		73	pmullw mm3, mm1
		74	psllw mm2, 8
		75	paddw mm3, mm2
		76	pinsrw mm3, ebx, 1
		77	psrlw mm3, 8
		78	packuswb mm3, mm0
		79	movd eax, mm3
		80
		81	ret
		82	endp
		83
		84
		85	proc merge_32 _copy_width, _copy_height, _img_total_bpl, _bottom_total_bpl
		86	.rgb_line:
		87	mov ecx, [_copy_width]
		88	.rgb_pixel:
		89	mov ebx, [edi]
		90	lodsd
		91
		92	movd mm2, ebx
		93	movd mm3, eax
		94	shr eax, 24
		95	shr ebx, 16
		96	cmp al, bh
		97	jna @f
		98	mov al, bh
		99	@@: pxor mm0, mm0
		100	call edx
		101	call blend_rgb
		102	stosd
		103	dec ecx
		104	jnz .rgb_pixel
		105	add esi, [_img_total_bpl]
		106	add edi, [_bottom_total_bpl]
		107	dec [_copy_height]
		108	jnz .rgb_line
		109	emms
		110	ret
		111	endp
		112
		113
		114	proc merge_8a _copy_width, _copy_height, _img_total_bpl, _bottom_total_bpl
		115	.gray_line:
		116	mov ecx, [_copy_width]
		117	.gray_pixel:
		118	mov bx, word[edi]
		119	lodsw
		120	movd mm2, ebx
		121	movd mm3, eax
		122	shr eax, 8
		123	cmp al, bh
		124	jna @f
		125	mov al, bh
		126	@@: pxor mm0, mm0
		127	call edx
		128	call blend_gray
		129	stosw
		130	dec ecx
		131	jnz .gray_pixel
		132	add esi, [_img_total_bpl]
		133	add edi, [_bottom_total_bpl]
		134	dec [_copy_height]
		135	jnz .gray_line
		136	emms
		137	ret
		138	endp
		139
		140
		141	proc composite_rgb_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		142
		143	.line: mov ecx, [_copy_width]
		144	.pixel: mov ebx, [edi]
		145	lodsd
		146	movd mm2, ebx
		147	movd mm3, eax
		148
		149	shr eax, 24
		150	shr ebx, 16
		151
		152	xchg al, bh
		153	mov ah, bh
		154	neg ax
		155	add ax, 0xffff
		156	mul ah
		157	neg ah
		158	add ah, 0xff
		159	xchg ah, bh
		160
		161	mov al, 0xff
		162	cmp ah, bh
		163	je @f
		164	not al
		165	div bh
		166	@@:
		167
		168	mov ah, al
		169	movd mm1, eax
		170	pxor mm0, mm0
		171	punpcklbw mm1, mm1
		172	punpcklbw mm1, mm0
		173	punpcklbw mm2, mm0
		174	punpcklbw mm3, mm0
		175
		176	psubsw mm3, mm2
		177	pmullw mm3, mm1
		178	psllw mm2, 8
		179	paddw mm3, mm2
		180	pinsrw mm3, ebx, 3
		181	psrlw mm3, 8
		182	packuswb mm3, mm0
		183	movd eax, mm3
		184	stosd
		185
		186	dec ecx
		187	jnz .pixel
		188	add esi, [_img_total_bpl]
		189	add edi, [_bottom_total_bpl]
		190	dec [_copy_height]
		191	jnz .line
		192
		193	ret
		194	endp
		195
		196
		197	proc composite_gray_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		198
		199	.line: mov ecx, [_copy_width]
		200	.pixel: mov bx, [edi]
		201	lodsw
		202	movd mm2, ebx
		203	movd mm3, eax
		204
		205	shr eax, 8
		206
		207	xchg al, bh
		208	mov ah, bh
		209	neg ax
		210	add ax, 0xffff
		211	mul ah
		212	neg ah
		213	add ah, 0xff
		214	xchg ah, bh
		215
		216	mov al, 0xff
		217	cmp ah, bh
		218	je @f
		219	not al
		220	div bh
		221	@@:
		222
		223	mov ah, al
		224	movd mm1, eax
		225	pxor mm0, mm0
		226	punpcklbw mm1, mm1
		227	punpcklbw mm1, mm0
		228	punpcklbw mm2, mm0
		229	punpcklbw mm3, mm0
		230
		231	psubw mm3, mm2
		232	pmullw mm3, mm1
		233	psllw mm2, 8
		234	paddw mm3, mm2
		235	pinsrw mm3, ebx, 1
		236	psrlw mm3, 8
		237	packuswb mm3, mm0
		238	movd eax, mm3
		239	stosw
		240
		241	dec ecx
		242	jnz .pixel
		243	add esi, [_img_total_bpl]
		244	add edi, [_bottom_total_bpl]
		245	dec [_copy_height]
		246	jnz .line
		247
		248	ret
		249	endp
		250
		251
		252	proc composite_indexed_00 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		253
		254	.line: mov ecx, [_copy_width]
		255	.pixel: mov bx, [edi]
		256	lodsw
		257
		258	or ah, 0x7f
		259	test ah, 0x80
		260	jnz @f
		261	mov ax, bx
		262	@@: stosw
		263
		264	dec ecx
		265	jnz .pixel
		266	add esi, [_img_total_bpl]
		267	add edi, [_bottom_total_bpl]
		268	dec [_copy_height]
		269	jnz .line
		270	ret
		271	endp
		272
		273
		274	proc composite_rgb_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		275	pushad
		276
		277	pxor mm4, mm4
		278	movd mm4, [random_b]
		279	movd mm1, [random_a]
		280	movd mm2, [random_c]
		281
		282	.line: mov ecx, [_copy_width]
		283	.pixel: mov ebx, [edi]
		284	lodsd
		285
		286	movq mm0, mm4
		287	pmuludq mm0, mm1
		288	paddq mm0, mm2
		289	movd edx, mm0
		290	movd mm4, edx
		291	pxor mm0, mm0
		292
		293	rol eax, 8
		294	test al, al
		295	jz @f
		296	shr edx, 17
		297	cmp dl, al
		298	ja @f
		299	ror eax, 8
		300	or eax, 0xff000000
		301	jmp .done
		302	@@: mov eax, ebx
		303	.done: stosd
		304	dec ecx
		305	jnz .pixel
		306	add esi, [_img_total_bpl]
		307	add edi, [_bottom_total_bpl]
		308	dec [_copy_height]
		309	jnz .line
		310
		311	.quit: popad
		312	ret
		313	endp
		314
		315
		316	proc composite_gray_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		317	pushad
		318
		319	pxor mm4, mm4
		320	movd mm4, [random_b]
		321	movd mm1, [random_a]
		322	movd mm2, [random_c]
		323
		324	.line: mov ecx, [_copy_width]
		325	.pixel: mov ebx, [edi]
		326	lodsw
		327
		328	movq mm0, mm4
		329	pmuludq mm0, mm1
		330	paddq mm0, mm2
		331	movd edx, mm0
		332	movd mm4, edx
		333	pxor mm0, mm0
		334
		335	test ah, ah
		336	jz @f
		337	shr edx, 17
		338	cmp dl, ah
		339	ja @f
		340	or ax, 0xff00
		341	jmp .done
		342	@@: mov eax, ebx
		343	.done: stosw
		344	dec ecx
		345	jnz .pixel
		346	add esi, [_img_total_bpl]
		347	add edi, [_bottom_total_bpl]
		348	dec [_copy_height]
		349	jnz .line
		350
		351	.quit: popad
		352	ret
		353	endp
		354
		355
		356	;proc composite_indexed_01 _copy_width, _copy_height, _bottom_total_bpl, _img_total_bpl
		357	; pushad
		358	;
		359	; pxor mm4, mm4
		360	; movd mm4, [random_b]
		361	; movd mm1, [random_a]
		362	; movd mm2, [random_c]
		363	;
		364	;.line: mov ecx, [_copy_width]
		365	;.pixel: mov ebx, [edi]
		366	; lodsw
		367	;
		368	; movq mm0, mm4
		369	; pmuludq mm0, mm1
		370	; paddq mm0, mm2
		371	; movd edx, mm0
		372	; movd mm4, edx
		373	; pxor mm0, mm0
		374	;
		375	; test ah, ah
		376	; jz @f
		377	; shr edx, 17
		378	; cmp dl, ah
		379	; ja @f
		380	; or ax, 0xff00
		381	; jmp .done
		382	;@@: mov eax, ebx
		383	;.done: stosw
		384	; dec ecx
		385	; jnz .pixel
		386	; add esi, [_img_total_bpl]
		387	; add edi, [_bottom_total_bpl]
		388	; dec [_copy_height]
		389	; jnz .line
		390	;
		391	;.quit: popad
		392	; ret
		393	;endp
		394
		395
		396	proc composite_rgb_03 ; Multiply
		397
		398	punpcklbw mm2, mm0
		399	punpcklbw mm3, mm0
		400	pmullw mm3, mm2
		401	psrlw mm3, 8
		402
		403	ret
		404	endp
		405
		406
		407	proc composite_rgb_04 ; Screen
		408
		409	punpcklbw mm2, mm0
		410	punpcklbw mm3, mm0
		411	movq mm4, [mmx_00ff]
		412	movq mm5, mm4
		413	psubw mm4, mm2
		414	psubw mm5, mm3
		415	pmullw mm4, mm5
		416	psrlw mm4, 8
		417	movq mm3, [mmx_00ff]
		418	psubw mm3, mm4
		419
		420	ret
		421	endp
		422
		423
		424	proc composite_rgb_05 ; Overlay
		425
		426	punpcklbw mm2, mm0
		427	punpcklbw mm3, mm0
		428	movq mm4, [mmx_00ff]
		429	psubw mm4, mm2
		430	pmullw mm3, mm4
		431	psrlw mm3, 7
		432	paddw mm3, mm2
		433	pmullw mm3, mm2
		434	psrlw mm3, 8
		435
		436	ret
		437	endp
		438
		439
		440	proc composite_rgb_06 ; Difference
		441
		442	movq mm4, mm3
		443	pminub mm4, mm2
		444	pmaxub mm3, mm2
		445	psubusb mm3, mm4
		446	punpcklbw mm2, mm0
		447	punpcklbw mm3, mm0
		448
		449	ret
		450	endp
		451
		452
		453	proc composite_rgb_07 ; Addition
		454
		455	paddusb mm3, mm2
		456	punpcklbw mm2, mm0
		457	punpcklbw mm3, mm0
		458
		459	ret
		460	endp
		461
		462
		463	proc composite_rgb_08 ; Subtract
		464
		465	movq mm4, mm2
		466	psubusb mm4, mm3
		467	movq mm3, mm4
		468	punpcklbw mm2, mm0
		469	punpcklbw mm3, mm0
		470
		471	ret
		472	endp
		473
		474
		475	proc composite_rgb_09 ; Darken Only
		476
		477	pminub mm3, mm2
		478	punpcklbw mm2, mm0
		479	punpcklbw mm3, mm0
		480
		481	ret
		482	endp
		483
		484
		485	proc composite_rgb_10 ; Lighten Only
		486
		487	pmaxub mm3, mm2
		488	punpcklbw mm2, mm0
		489	punpcklbw mm3, mm0
		490
		491	ret
		492	endp
		493
		494
		495	proc composite_rgb_11 ; Hue (H of HSV)
		496	push eax ebx ecx edx
		497
		498	movd eax, mm3
		499	movd ebx, mm2
		500
		501	call pixel_rgb2hsv
		502	xchg eax, ebx
		503	call pixel_rgb2hsv
		504	xchg eax, ebx
		505
		506	test ah, ah
		507	jnz @f
		508	ror eax, 8
		509	ror ebx, 8
		510	mov ah, bh
		511	rol eax, 8
		512	rol ebx, 8
		513	@@:
		514	mov ax, bx
		515
		516	call pixel_hsv2rgb
		517
		518
		519	movd mm3, eax
		520
		521	punpcklbw mm2, mm0
		522	punpcklbw mm3, mm0
		523
		524	.quit:
		525	pop edx ecx ebx eax
		526	ret
		527	endp
		528
		529
		530	proc composite_rgb_12 ; Saturation (S of HSV)
		531	push eax ebx ecx edx
		532
		533	movd eax, mm3
		534	movd ebx, mm2
		535
		536	call pixel_rgb2hsv
		537	xchg eax, ebx
		538	call pixel_rgb2hsv
		539	xchg eax, ebx
		540
		541	ror eax, 8
		542	ror ebx, 8
		543	mov ah, bh
		544	rol eax, 8
		545	rol ebx, 8
		546	mov al, bl
		547
		548	call pixel_hsv2rgb
		549
		550
		551	movd mm3, eax
		552
		553	punpcklbw mm2, mm0
		554	punpcklbw mm3, mm0
		555
		556	.quit:
		557	pop edx ecx ebx eax
		558	ret
		559	endp
		560
		561
		562	proc composite_rgb_13 ; Color (H and S of HSL)
		563	push eax ebx ecx edx
		564
		565	movd eax, mm3
		566	movd ebx, mm2
		567
		568	call pixel_rgb2hsl
		569	xchg eax, ebx
		570	call pixel_rgb2hsl
		571	xchg eax, ebx
		572
		573	mov al, bl
		574
		575	call pixel_hsl2rgb
		576
		577
		578	movd mm3, eax
		579
		580	punpcklbw mm2, mm0
		581	punpcklbw mm3, mm0
		582
		583	.quit:
		584	pop edx ecx ebx eax
		585	ret
		586	endp
		587
		588
		589	proc composite_rgb_14 ; Value (V of HSV)
		590	push eax ebx ecx edx
		591
		592	movd eax, mm3
		593	movd ebx, mm2
		594
		595	call pixel_rgb2hsv
		596	xchg eax, ebx
		597	call pixel_rgb2hsv
		598	xchg eax, ebx
		599
		600	ror eax, 8
		601	ror ebx, 8
		602	mov ax, bx
		603	rol eax, 8
		604	rol ebx, 8
		605
		606	call pixel_hsv2rgb
		607
		608
		609	movd mm3, eax
		610
		611	punpcklbw mm2, mm0
		612	punpcklbw mm3, mm0
		613
		614	.quit:
		615	pop edx ecx ebx eax
		616	ret
		617	endp
		618
		619
		620	proc composite_rgb_15 ; Divide
		621	push eax ebx ecx
		622
		623	movd eax, mm3
		624	movd ebx, mm2
		625
		626	rol eax, 8
		627	rol ebx, 8
		628
		629	xchg eax, ebx
		630
		631	mov ecx, 3
		632
		633	.color: rol eax, 8
		634	rol ebx, 8
		635	shl ax, 8
		636	test bl, bl
		637	jz .clamp1
		638	cmp ah, bl
		639	jae .clamp2
		640	div bl
		641	jmp .done
		642	.clamp1:mov al, 0xff
		643	test ah, ah
		644	jnz @f
		645	not al
		646	@@: jmp .done
		647	.clamp2:mov al, 0xff
		648	jmp .done
		649	.done: mov ah, al
		650	loop .color
		651
		652	ror eax, 8
		653	movd mm3, eax
		654
		655	punpcklbw mm2, mm0
		656	punpcklbw mm3, mm0
		657
		658	pop ecx ebx eax
		659	ret
		660	endp
		661
		662
		663	proc composite_rgb_16 ; Dodge
		664	push eax ebx ecx
		665
		666	movd eax, mm3
		667	movd ebx, mm2
		668
		669	rol eax, 8
		670	rol ebx, 8
		671
		672	xchg eax, ebx
		673
		674	mov ecx, 3
		675
		676	.color: rol eax, 8
		677	rol ebx, 8
		678	shl ax, 8
		679	neg bl
		680	add bl, 0xff
		681	test bl, bl
		682	jz .clamp1
		683	cmp ah, bl
		684	jae .clamp2
		685	div bl
		686	jmp .done
		687	.clamp1:mov al, 0xff
		688	test ah, ah
		689	jnz @f
		690	not al
		691	@@: jmp .done
		692	.clamp2:mov al, 0xff
		693	jmp .done
		694	.done: mov ah, al
		695	loop .color
		696
		697	ror eax, 8
		698	movd mm3, eax
		699
		700	punpcklbw mm2, mm0
		701	punpcklbw mm3, mm0
		702
		703	pop ecx ebx eax
		704	ret
		705	endp
		706
		707
		708	proc composite_rgb_17 ; Burn
		709	push eax ebx ecx
		710
		711	movd eax, mm3
		712	movd ebx, mm2
		713
		714	rol eax, 8
		715	rol ebx, 8
		716
		717	xchg eax, ebx
		718
		719	mov ecx, 3
		720
		721	.color: rol eax, 8
		722	rol ebx, 8
		723	shl ax, 8
		724	neg ah
		725	add ah, 0xff
		726	test bl, bl
		727	jz .clamp1
		728	cmp ah, bl
		729	jae .clamp2
		730	div bl
		731	jmp .done
		732	.clamp1:mov al, 0xff
		733	test ah, ah
		734	jnz @f
		735	not al
		736	@@: jmp .done
		737	.clamp2:mov al, 0xff
		738	jmp .done
		739	.done: mov ah, al
		740	neg ah
		741	add ah, 0xff
		742	loop .color
		743
		744	ror eax, 8
		745	movd mm3, eax
		746
		747	punpcklbw mm2, mm0
		748	punpcklbw mm3, mm0
		749
		750	pop ecx ebx eax
		751	ret
		752	endp
		753
		754
		755	proc composite_rgb_18 ; Hard Light
		756	push eax ebx ecx
		757
		758	movd eax, mm3
		759	movd ebx, mm2
		760
		761	rol eax, 8
		762	rol ebx, 8
		763
		764	mov ecx, 3
		765
		766	.color: rol eax, 8
		767	rol ebx, 8
		768	cmp al, 127
		769	jna .part1
		770	mov ah, 0xff
		771	sub ah, bl
		772	neg al
		773	add al, 0xff
		774	mul ah
		775	shl ax, 1
		776	neg ah
		777	add ah, 0xff
		778	jmp .done
		779	.part1:
		780	mul bl
		781	shl ax, 1
		782	.done: loop .color
		783
		784	ror eax, 8
		785	movd mm3, eax
		786
		787	punpcklbw mm2, mm0
		788	punpcklbw mm3, mm0
		789
		790	pop ecx ebx eax
		791	ret
		792	endp
		793
		794
		795	proc composite_rgb_20 ; Grain Extract
		796
		797	movq mm4, [mmx_0080]
		798	punpcklbw mm2, mm0
		799	punpcklbw mm3, mm0
		800	movq mm5, mm2
		801	psubw mm3, mm4
		802	psubsw mm5, mm3
		803	packuswb mm5, mm0
		804	punpcklbw mm5, mm0
		805	movq mm3, mm5
		806
		807	ret
		808	endp
		809
		810
		811	proc composite_rgb_21 ; Grain Merge
		812
		813	movq mm4, [mmx_0080]
		814	punpcklbw mm2, mm0
		815	punpcklbw mm3, mm0
		816	movq mm5, mm2
		817	psubw mm5, mm4
		818	paddsw mm3, mm5
		819	packuswb mm3, mm0
		820	punpcklbw mm3, mm0
		821
		822	ret
		823	endp
		824
		825
		826	mmx_0080 dq 0x0080008000800080
		827	mmx_00ff dq 0x00ff00ff00ff00ff
		828	mmx_0100 dq 0x0100010001000100

Subversion Repositories Kolibri OS

(root)/programs/develop/libraries/libs-dev/libimg/xcf/composite_mmx.inc – Rev 2387