WebSVN – Kolibri OS – Blame – /programs/demos/3DS/BUMP_TEX.INC

Rev	Author	Line No.	Line
1245	hidnplayr	1
		2	;TEXTURE_SIZE EQU (TEX_X * TEX_Y)-1
		3	;ROUND equ 8
		4	;Ext = NON
		5	;MMX = 1
		6	;NON = 0
		7	;------- Big thanks to Majuma (www.majuma.xt.pl) for absolutely great---
		8	;------- DOS 13h mode demos --------------------------------------------
		9	;------- Procedure draws bump triangle with texture, I use -------------
		10	;--------Catmull Z-buffer algorithm- (Z coordinate interpolation)-------
		11	;--------I calc texture pixel by this way: col1*col2/256 ---------------
		12	bump_tex_triangle_z:
		13	;------------------in - eax - x1 shl 16 + y1 -----------
		14	;---------------------- ebx - x2 shl 16 + y2 -----------
		15	;---------------------- ecx - x3 shl 16 + y3 -----------
		16	;---------------------- edx - pointer to bump map-------
		17	;---------------------- esi - pointer to env map--------
		18	;---------------------- edi - pointer to screen buffer--
		19	;---------------------- stack : bump coordinates--------
		20	;---------------------- environment coordinates-
		21	;---------------------- Z position coordinates--
		22	;---------------------- pointer to Z buffer-----
		23	;---------------------- pointer to texture------
		24	;---------------------- texture coordinates-----
		25	;-- Z-buffer - filled with coordinates as dword --------
		26	;-- (Z coor. as word) shl CATMULL_SHIFT ----------------
		27	.b_x1 equ ebp+4 ; procedure don't save registers !!!
		28	.b_y1 equ ebp+6 ; each coordinate as word
		29	.b_x2 equ ebp+8
		30	.b_y2 equ ebp+10 ; b - bump map coords
		31	.b_x3 equ ebp+12 ; e - env map coords
		32	.b_y3 equ ebp+14
		33	.e_x1 equ ebp+16
		34	.e_y1 equ ebp+18
		35	.e_x2 equ ebp+20
		36	.e_y2 equ ebp+22
		37	.e_x3 equ ebp+24
		38	.e_y3 equ ebp+26
		39	.z1 equ word[ebp+28]
		40	.z2 equ word[ebp+30]
		41	.z3 equ word[ebp+32]
		42	.z_buff equ dword[ebp+34] ; pointer to Z-buffer
		43	.tex_ptr equ dword[ebp+38] ; ptr to texture
		44	.t_x1 equ ebp+42 ; texture coords
		45	.t_y1 equ ebp+44
		46	.t_x2 equ ebp+46
		47	.t_y2 equ ebp+48
		48	.t_x3 equ ebp+50
		49	.t_y3 equ ebp+52
		50
		51
		52
		53
		54	.t_emap equ dword[ebp-8] ; pointer to env map
		55	.x1 equ word[ebp-10]
		56	.y1 equ word[ebp-12]
		57	.x2 equ word[ebp-14]
		58	.y2 equ word[ebp-16]
		59	.x3 equ word[ebp-18]
		60	.y3 equ word[ebp-20]
		61
		62
		63	.dz12 equ [ebp-28]
		64	.dbx12 equ dword[ebp-32]
		65	.dby12 equ [ebp-36]
		66	.dex12 equ dword[ebp-40]
		67	.dey12 equ [ebp-44]
		68	.dtx12 equ dword[ebp-48]
		69	.dty12 equ [ebp-52]
		70
		71
		72	.dz13 equ [ebp-52-4*2]
		73	.dbx13 equ dword[ebp-52-4*3]
		74	.dby13 equ [ebp-52-4*4]
		75	.dex13 equ dword[ebp-52-4*5]
		76	.dey13 equ [ebp-52-4*6]
		77	.dtx13 equ dword[ebp-52-4*7]
		78	.dty13 equ [ebp-52-4*8]
		79
		80
		81
		82	.dz23 equ [ebp-(52+4*10)]
		83	.dbx23 equ dword[ebp-(52+4*11)]
		84	.dby23 equ [ebp-(52+4*12)]
		85	.dex23 equ dword[ebp-(52+4*13)]
		86	.dey23 equ [ebp-(52+4*14)]
		87	.dtx23 equ dword[ebp-(52+4*15)]
		88	.dty23 equ [ebp-(52+4*16)]
		89
		90
		91	.cz1 equ [ebp-(52+4*18)]
		92	.cx2 equ dword[ebp-(52+4*19)]
		93	.cz2 equ [ebp-(52+4*20)]
		94	.cbx1 equ dword[ebp-(52+4*21)]
		95	.cby1 equ [ebp-(52+4*22)]
		96	.cbx2 equ dword[ebp-(52+4*23)]
		97	.cby2 equ [ebp-(52+4*24)]
		98	.cex1 equ dword[ebp-(52+4*25)]
		99	.cey1 equ [ebp-(52+4*26)]
		100	.cex2 equ dword[ebp-(52+4*27)]
		101	.cey2 equ [ebp-(52+4*28)]
		102
		103
		104	.cty1 equ [ebp-(52+4*30)]
		105	.ctx2 equ dword[ebp-(52+4*31)]
		106	.cty2 equ [ebp-(52+4*32)]
		107
		108
		109	mov ebp,esp
		110	push edx ; store bump map
		111	push esi ; store e. map
		112	; sub esp,120
		113	.sort3: ; sort triangle coordinates...
		114	cmp ax,bx
		115	jle .sort1
		116	xchg eax,ebx
		117	mov edx,dword[.b_x1]
		118	xchg edx,dword[.b_x2]
		119	mov dword[.b_x1],edx
		120	mov edx,dword[.e_x1]
		121	xchg edx,dword[.e_x2]
		122	mov dword[.e_x1],edx
		123	mov edx,dword[.t_x1]
		124	xchg edx,dword[.t_x2]
		125	mov dword[.t_x1],edx
		126	mov dx,.z1
		127	xchg dx,.z2
		128	mov .z1,dx
		129	.sort1:
		130	cmp bx,cx
		131	jle .sort2
		132	xchg ebx,ecx
		133	mov edx,dword[.b_x2]
		134	xchg edx,dword[.b_x3]
		135	mov dword[.b_x2],edx
		136	mov edx,dword[.e_x2]
		137	xchg edx,dword[.e_x3]
		138	mov dword[.e_x2],edx
		139	mov edx,dword[.t_x2]
		140	xchg edx,dword[.t_x3]
		141	mov dword[.t_x2],edx
		142	mov dx,.z2
		143	xchg dx,.z3
		144	mov .z2,dx
		145	jmp .sort3
		146	.sort2:
		147	push eax ; store triangle coords in variables
		148	push ebx
		149	push ecx
		150	mov edx,80008000h ; eax,ebx,ecx are ANDd together into edx which means that
		151	and edx,ebx ; if all of them are negative a sign flag is raised
		152	and edx,ecx
		153	and edx,eax
		154	test edx,80008000h ; Check both X&Y at once
		155	jne .loop23_done
		156	; mov edx,eax ; eax,ebx,ecx are ORd together into edx which means that
		157	; or edx,ebx ; if any one of them is negative a sign flag is raised
		158	; or edx,ecx
		159	; test edx,80000000h ; Check only X
		160	; jne .loop23_done
		161
		162
		163	; jg .loop23_done
		164	; cmp .x2,SIZE_X ; This can be optimized with effort
		165	; jg .loop23_done
		166	; cmp .x3,SIZE_X
		167	; jg .loop23_done ; {
		168
		169
		170
		171	sub bx,.y1
		172	jnz .bt_dx12_make
		173	mov ecx,8
		174	xor edx,edx
		175	@@:
		176	push edx ;dword 0
		177	loop @b
		178	jmp .bt_dx12_done
		179	.bt_dx12_make:
		180	movsx ebx,bx
1776	yogev_ezra	181
1245	hidnplayr	182
		183	sub ax,.x1
		184	cwde
		185	shl eax,ROUND
		186	cdq
		187	idiv ebx
		188	; mov .dx12,eax
		189	push eax
		190
		191
		192
		193
1776	yogev_ezra	194	; mov eax,256
1245	hidnplayr	195	cvtsi2ss xmm4,[i255d]
		196	cvtsi2ss xmm3,ebx ;rcps
		197	divss xmm3,xmm4
		198	shufps xmm3,xmm3,0
		199
		200
		201	movd mm1,[.b_x2]
		202	movd mm2,[.e_x1]
		203	movd mm3,[.e_x2]
		204
		205
		206	punpcklwd mm0,mm4
		207	punpcklwd mm1,mm4
		208	punpcklwd mm2,mm4
		209	punpcklwd mm3,mm4
		210
		211
		212	psubd mm3,mm2
		213
		214
		215	movlhps xmm1,xmm1
		216	cvtpi2ps xmm1,mm3
		217
		218
		219
		220
		221	;xmm1--> \| dbx \| dby \| dex \| dey \|
		222	cvtps2pi mm0,xmm1 ;mm0,xmm1 ; mm0 -> 2 delta dwords
1776	yogev_ezra	223	movhlps xmm1,xmm1
1245	hidnplayr	224	cvtps2pi mm1,xmm1 ;mm1,xmm1
1776	yogev_ezra	225	movq .dey12,mm0
1245	hidnplayr	226	movq .dby12,mm1
		227	;-------------
		228	; pxor mm0,mm0
1776	yogev_ezra	229	; pxor mm1,mm1
		230	;/ pinsrw mm0,.z1,1
		231	;/ pinsrw mm0,.x1,0
		232	;/ pinsrw mm1,.z2,1
		233	;/ pinsrw mm1,.x2,0
		234	mov ax,.z2
		235	sub ax,.z1
		236	cwde
		237	;/ movd mm1,eax
		238
		239
		240	;/ punpcklwd mm1,mm4
		241
		242
		243	; cvtpi2ps xmm2,mm0
		244	; subps xmm1,xmm2
		245
		246
		247
		248
		249	movd mm3,[.t_x2]
		250
		251
		252	punpcklwd mm3,mm4
		253	psubd mm3,mm2
		254
		255
		256	cvtsi2ss xmm1,eax
		257	movlhps xmm1,xmm1
		258	cvtpi2ps xmm1,mm3
		259
		260
		261
		262
		263	; xmm1--> \| xxx \| dz \| dtx \| dty \|
		264	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords \| dtx \| dty \|
		265	movhlps xmm1,xmm1
		266	cvtss2si eax,xmm1 ; eax --> 1 delta dword \| dz \|
		267	movq .dty12,mm0
		268	mov .dz12,eax
		269	;----
		270	; mov ax,.z2
1245	hidnplayr	271	; sub ax,.z1
		272	; cwde
		273	; mov bx,.x2
		274	; sub bx,.x1
		275	; movsx ebx,bx
		276	; movd mm1,eax
		277	; psllq mm1,32
		278	; movd mm1,ebx
		279
1776	yogev_ezra	280
1245	hidnplayr	281	;; push eax
		282	;; movq mm1,[esp]
		283	;; add esp,8
		284	;;; mov ax,.z1
		285	;;; mov bx,.z2
		286	;;; shl eax,16
		287	;;; shl ebx,16
		288	;;; mov ax,.x1
		289	;;; mov bx,.x2
		290	; movd mm2,[.t_x1]
		291	; movd mm3,[.t_x2]
		292	;; movd mm0,eax
		293	;; movd mm1,ebx
		294
		295
		296	;; punpcklwd mm0,mm4
		297	;; punpcklwd mm1,mm4
		298	; punpcklwd mm2,mm4
		299	; punpcklwd mm3,mm4
		300
		301
		302	; psubd mm3,mm2
		303
		304
		305
		306	; movlhps xmm1,xmm1
		307	; cvtpi2ps xmm1,mm3
		308
		309
		310
		311
		312	; xmm1--> \| dx \| dz \| dtx \| dty \|
		313	; cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords \| dtx \| dty \|
		314	; movhlps xmm1,xmm1
		315	; cvtps2pi mm1,xmm1 ; mm1 --> 2 delta dwords \| dx \| dz \|
		316	; movq .dty12,mm0
		317	; movq .dz12,mm1
		318	else
		319	mov ax,.z2
1776	yogev_ezra	320	sub ax,.z1
		321	cwde
		322	shl eax,CATMULL_SHIFT
		323	cdq
		324	idiv ebx
		325	push eax
		326
1245	hidnplayr	327
		328	sub ax,word[.b_x1]
		329	cwde
		330	shl eax,ROUND
		331	cdq
		332	idiv ebx
		333	; mov .dbx12,eax
		334	push eax
		335
		336
		337	sub ax,word[.b_y1]
		338	cwde
		339	shl eax,ROUND
		340	cdq
		341	idiv ebx
		342	; mov .dby12,eax
		343	push eax
		344
		345
		346	sub ax,word[.e_x1]
		347	cwde
		348	shl eax,ROUND
		349	cdq
		350	idiv ebx
		351	; mov .dex12,eax
		352	push eax
		353
		354
		355	sub ax,word[.e_y1]
		356	cwde
		357	shl eax,ROUND
		358	cdq
		359	idiv ebx
		360	; mov .dey12,eax
		361	push eax
		362
		363
1776	yogev_ezra	364
1245	hidnplayr	365
		366	sub ax,word[.t_x1]
		367	cwde
		368	shl eax,ROUND
		369	cdq
		370	idiv ebx
		371	; mov .dtx12,eax
		372	push eax
		373
		374
		375	sub ax,word[.t_y1]
		376	cwde
		377	shl eax,ROUND
		378	cdq
		379	idiv ebx
		380	; mov .dty12,eax
		381	push eax
		382	end if
1776	yogev_ezra	383	.bt_dx12_done:
1245	hidnplayr	384
		385
		386	sub bx,.y1
		387	jnz .bt_dx13_make
		388	mov ecx,8
		389	xor edx,edx
		390	@@:
		391	push edx ;dword 0
		392	loop @b
		393	jmp .bt_dx13_done
		394	.bt_dx13_make:
		395	mov ax,.x3
		396	sub ax,.x1
		397	cwde
		398	movsx ebx,bx
		399	shl eax,ROUND
		400	cdq
		401	idiv ebx
		402	; mov .dx13,eax
		403	push eax
		404
		405
		406
		407
1776	yogev_ezra	408	; mov eax,255
1245	hidnplayr	409	cvtsi2ss xmm4,[i255d]
		410	cvtsi2ss xmm3,ebx ;rcps
		411	divss xmm3,xmm4
		412	shufps xmm3,xmm3,0
		413
		414
		415	movd mm1,[.b_x3]
		416	movd mm2,[.e_x1]
		417	movd mm3,[.e_x3]
		418
		419
		420	punpcklwd mm0,mm4
		421	punpcklwd mm1,mm4
		422	punpcklwd mm2,mm4
		423	punpcklwd mm3,mm4
		424
		425
		426	psubd mm3,mm2
		427
		428
		429	movlhps xmm1,xmm1
		430	cvtpi2ps xmm1,mm3
		431
		432
		433
		434
		435	;xmm1--> \| dbx \| dby \| dex \| dey \|
		436	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		437	movhlps xmm1,xmm1
		438	cvtps2pi mm1,xmm1
		439	movq .dey13,mm0
		440	movq .dby13,mm1
		441	;----------
1776	yogev_ezra	442	mov ax,.z3
		443	sub ax,.z1
		444	cwde
		445
		446
		447	movd mm3,[.t_x3]
		448
		449
		450	punpcklwd mm3,mm4
		451	psubd mm3,mm2
		452
		453
		454	movlhps xmm1,xmm1
		455	cvtpi2ps xmm1,mm3
		456
		457
		458
		459
		460	; xmm1--> \| xxx \| dz \| dtx \| dty \|
		461	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords \| dtx \| dty \|
		462	movhlps xmm1,xmm1
		463	cvtss2si eax,xmm1 ; eax --> 1 delta dword \| dz \|
		464	movq .dty13,mm0
		465	mov .dz13,eax
		466	else
1245	hidnplayr	467	mov ax,.z3
1776	yogev_ezra	468	sub ax,.z1
		469	cwde
		470	shl eax,CATMULL_SHIFT
		471	cdq
		472	idiv ebx
		473	; mov .dz13,eax
		474	push eax
		475
		476
		477
1245	hidnplayr	478	sub ax,word[.b_x1]
		479	cwde
		480	shl eax,ROUND
		481	cdq
		482	idiv ebx
		483	; mov .dbx13,eax
		484	push eax
		485
		486
		487	sub ax,word[.b_y1]
		488	cwde
		489	shl eax,ROUND
		490	cdq
		491	idiv ebx
		492	; mov .dby13,eax
		493	push eax
		494
		495
		496	sub ax,word[.e_x1]
		497	cwde
		498	shl eax,ROUND
		499	cdq
		500	idiv ebx
		501	; mov .dex13,eax
		502	push eax
		503
		504
		505	sub ax,word[.e_y1]
		506	cwde
		507	shl eax,ROUND
		508	cdq
		509	idiv ebx
		510	; mov .dey13,eax
		511	push eax
		512
		513
		514	sub ax,word[.t_x1]
		515	cwde
		516	shl eax,ROUND
		517	cdq
		518	idiv ebx
		519	; mov .dtx13,eax
		520	push eax
		521
		522
		523	sub ax,word[.t_y1]
		524	cwde
		525	shl eax,ROUND
		526	cdq
		527	idiv ebx
		528	; mov .dty13,eax
		529	push eax
		530	end if
1776	yogev_ezra	531	.bt_dx13_done:
1245	hidnplayr	532
		533
		534	sub bx,.y2
		535	jnz .bt_dx23_make
		536	mov ecx,8
		537	xor edx,edx
		538	@@:
		539	push edx ;dword 0
		540	loop @b
		541	jmp .bt_dx23_done
		542	.bt_dx23_make:
		543	mov ax,.x3
		544	sub ax,.x2
		545	cwde
		546	movsx ebx,bx
		547	shl eax,ROUND
		548	cdq
		549	idiv ebx
		550	; mov .dx23,eax
		551	push eax
		552
		553
		554
		555
1776	yogev_ezra	556	; mov eax,255
1245	hidnplayr	557	cvtsi2ss xmm4,[i255d]
		558	cvtsi2ss xmm3,ebx ;rcps
		559	divss xmm3,xmm4
		560	shufps xmm3,xmm3,0
		561
		562
		563	movd mm1,[.b_x3]
		564	movd mm2,[.e_x2]
		565	movd mm3,[.e_x3]
		566
		567
		568	punpcklwd mm0,mm4
		569	punpcklwd mm1,mm4
		570	punpcklwd mm2,mm4
		571	punpcklwd mm3,mm4
		572
		573
		574	psubd mm3,mm2
		575
		576
		577	movlhps xmm1,xmm1
		578	cvtpi2ps xmm1,mm3
		579
		580
		581
		582
		583	;xmm1--> \| dbx \| dby \| dex \| dey \|
		584	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		585	movhlps xmm1,xmm1
		586	cvtps2pi mm1,xmm1
		587	movq .dey23,mm0
		588	movq .dby23,mm1
		589	;----------
1776	yogev_ezra	590	mov ax,.z3
		591	sub ax,.z2
		592	cwde
		593
		594
		595	movd mm3,[.t_x3]
		596
		597
		598	punpcklwd mm3,mm4
		599	psubd mm3,mm2
		600
		601
		602	movlhps xmm1,xmm1
		603	cvtpi2ps xmm1,mm3
		604
		605
		606
		607
		608	; xmm1--> \| xxx \| dz \| dtx \| dty \|
		609	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords \| dtx \| dty \|
		610	movhlps xmm1,xmm1
		611	cvtss2si eax,xmm1 ; eax --> 1 delta dword \| dz \|
		612	movq .dty23,mm0
		613	mov .dz23,eax
		614	else
1245	hidnplayr	615
1776	yogev_ezra	616
		617	sub ax,.z2
		618	cwde
		619	shl eax,CATMULL_SHIFT
		620	cdq
		621	idiv ebx
		622	; mov .dz23,eax
		623	push eax
		624
		625
1245	hidnplayr	626	sub ax,word[.b_x2]
		627	cwde
		628	shl eax,ROUND
		629	cdq
		630	idiv ebx
		631	; mov .dbx23,eax
		632	push eax
		633
		634
		635	sub ax,word[.b_y2]
		636	cwde
		637	shl eax,ROUND
		638	cdq
		639	idiv ebx
		640	; mov .dby23,eax
		641	push eax
		642
		643
		644	sub ax,word[.e_x2]
		645	cwde
		646	shl eax,ROUND
		647	cdq
		648	idiv ebx
		649	; mov .dex23,eax
		650	push eax
		651
		652
		653	sub ax,word[.e_y2]
		654	cwde
		655	shl eax,ROUND
		656	cdq
		657	idiv ebx
		658	; mov .dey23,eax
		659	push eax
		660
		661
1776	yogev_ezra	662
1245	hidnplayr	663	sub ax,word[.t_x2]
		664	cwde
		665	shl eax,ROUND
		666	cdq
		667	idiv ebx
		668	; mov .dtx23,eax
		669	push eax
		670
		671
		672	sub ax,word[.t_y2]
		673	cwde
		674	shl eax,ROUND
		675	cdq
		676	idiv ebx
		677	; mov .dty23,eax
		678	push eax
		679	end if
1776	yogev_ezra	680	; sub esp,40
1245	hidnplayr	681	.bt_dx23_done:
		682	sub esp,64
		683	movsx eax,.x1
		684	shl eax,ROUND
		685	mov .cx1,eax
		686	mov .cx2,eax
		687	; push eax
		688	; push eax
		689
		690
		691	shl ebx,ROUND
		692	mov .cbx1,ebx
		693	mov .cbx2,ebx
		694	; push ebx
		695	; push ebx
		696
		697
		698	shl ecx,ROUND
		699	mov .cby1,ecx
		700	mov .cby2,ecx
		701	; push ecx
		702	; push ecx
		703
		704
		705	shl edx,ROUND
		706	mov .cex1,edx
		707	mov .cex2,edx
		708	; push edx
		709	; push edx
		710
		711
		712	shl eax,ROUND
		713	mov .cey1,eax
		714	mov .cey2,eax
		715	; push eax
		716	; push eax
		717
		718
		719	shl ebx,CATMULL_SHIFT
		720	mov .cz1,ebx
		721	mov .cz2,ebx
		722	; push ebx
		723	; push ebx
		724
		725
		726	movsx ecx,word[.t_x1]
		727	shl ecx,ROUND
		728	mov .ctx1,ecx
		729	mov .ctx2,ecx
		730	;push ecx
		731	;push ecx
		732
		733
		734	shl edx,ROUND
		735	mov .cty1,edx
		736	mov .cty2,edx
		737	; push edx
		738	; push edx
		739
		740
		741
		742	cmp cx,.y2
		743	jge .loop12_done
		744	.loop12:
		745	call .call_line
		746
		747
		748	movq mm0,.cby2
		749	movq mm1,.cby1
		750	movq mm2,.cey2
		751	movq mm3,.cey1
		752	movq mm4,.cty1
		753	movq mm5,.cty2
		754	movq mm6,.cz1
		755	movq mm7,.cz2
		756	paddd mm0,.dby12
		757	paddd mm1,.dby13
		758	paddd mm2,.dey12
		759	paddd mm3,.dey13
		760	paddd mm4,.dty13
		761	paddd mm5,.dty12
		762	paddd mm6,.dz13
		763	paddd mm7,.dz12
		764	movq .cby2,mm0
		765	movq .cby1,mm1
		766	movq .cey1,mm3
		767	movq .cey2,mm2
		768	movq .cty1,mm4
		769	movq .cty2,mm5
		770	movq .cz1,mm6
		771	movq .cz2,mm7
		772	else
		773	mov edx,.dbx13
		774	add .cbx1,edx
		775	mov eax,.dbx12
		776	add .cbx2,eax
		777	mov ebx,.dby13
		778	add .cby1,ebx
		779	mov edx,.dby12
		780	add .cby2,edx
		781
		782
		783	add .cex1,eax
		784	mov ebx,.dex12
		785	add .cex2,ebx
		786	mov edx,.dey13
		787	add .cey1,edx
		788	mov eax,.dey12
		789	add .cey2,eax
		790
		791
		792	add .ctx1,eax
		793	mov ebx,.dtx12
		794	add .ctx2,ebx
		795	mov edx,.dty13
		796	add .cty1,edx
		797	mov eax,.dty12
		798	add .cty2,eax
		799
		800
		801	add .cx1,eax
		802	mov ebx,.dx12
		803	add .cx2,ebx
		804	mov ebx,.dz13
		805	add .cz1,ebx
		806	mov edx,.dz12
		807	add .cz2,edx
		808	end if
		809	inc ecx
		810	cmp cx,.y2
		811	jl .loop12
		812	.loop12_done:
		813
		814
		815	cmp cx,.y3
		816	jge .loop23_done
		817
		818
		819	shl eax,CATMULL_SHIFT
		820	mov .cz2,eax
		821
		822
		823	shl ebx,ROUND
		824	mov .cx2,ebx
		825
		826
		827	shl edx,ROUND
		828	mov .cbx2,edx
		829
		830
		831	shl eax,ROUND
		832	mov .cby2,eax
		833
		834
		835	shl ebx,ROUND
		836	mov .cex2,ebx
		837
		838
		839	shl edx,ROUND
		840	mov .cey2,edx
		841
		842
		843	shl eax,ROUND
		844	mov .ctx2,eax
		845
		846
		847	shl ebx,ROUND
		848	mov .cty2,ebx
		849
		850
		851	call .call_line
		852
		853
		854	movq mm0,.cby2
		855	movq mm1,.cby1
		856	movq mm2,.cey2
		857	movq mm3,.cey1
		858	movq mm4,.cty1
		859	movq mm5,.cty2
		860	movq mm6,.cz1
		861	movq mm7,.cz2
		862	paddd mm0,.dby23
		863	paddd mm1,.dby13
		864	paddd mm2,.dey23
		865	paddd mm3,.dey13
		866	paddd mm4,.dty13
		867	paddd mm5,.dty23
		868	paddd mm6,.dz13
		869	paddd mm7,.dz23
		870	movq .cby2,mm0
		871	movq .cby1,mm1
		872	movq .cey2,mm2
		873	movq .cey1,mm3
		874	movq .cty1,mm4
		875	movq .cty2,mm5
		876	movq .cz1,mm6
		877	movq .cz2,mm7
		878	else
		879	mov edx,.dbx13
		880	add .cbx1,edx
		881	mov eax,.dbx23
		882	add .cbx2,eax
		883	mov ebx,.dby13
		884	add .cby1,ebx
		885	mov edx,.dby23
		886	add .cby2,edx
		887
		888
		889	add .cex1,eax
		890	mov ebx,.dex23
		891	add .cex2,ebx
		892	mov edx,.dey13
		893	add .cey1,edx
		894	mov eax,.dey23
		895	add .cey2,eax
		896
		897
		898	add .cx1,eax
		899	mov ebx,.dx23
		900	add .cx2,ebx
		901	mov ebx,.dz13
		902	add .cz1,ebx
		903	mov edx,.dz23
		904	add .cz2,edx
		905
		906
		907	add .ctx1,eax
		908	mov ebx,.dtx23
		909	add .ctx2,ebx
		910	mov edx,.dty13
		911	add .cty1,edx
		912	mov eax,.dty23
		913	add .cty2,eax
		914
		915
		916	inc ecx
		917	cmp cx,.y3
		918	jl .loop23
		919	.loop23_done:
		920
		921
		922	ret 50
		923
		924
		925
		926
		927	push .tex_ptr
		928	push dword .cty2
		929	push .ctx2
		930	push dword .cty1
		931	push .ctx1
		932	push dword .cz1
		933	push dword .cz2
		934	push .z_buff
		935	push .t_emap
		936	push .t_bmap
		937	push dword .cey2
		938	push .cex2
		939	push dword .cey1
		940	push .cex1
		941	push dword .cby2
		942	push .cbx2
		943	push dword .cby1
		944	push .cbx1
		945	push ecx
		946
		947
		948	sar eax,ROUND
		949	mov ebx,.cx2
		950	sar ebx,ROUND
		951
		952
		953
		954
		955	ret
		956	bump_tex_line_z:
		957	;--------------in: eax - x1
		958	;-------------- ebx - x2
		959	;-------------- edi - pointer to screen buffer
		960	;stack - another parameters :
		961	.y equ dword [ebp+4]
		962	.bx1q equ [ebp+8]
		963	.bx2q equ [ebp+16]
		964	.ex1q equ [ebp+24]
		965	.ex2q equ [ebp+32]
		966	.tx1q equ [ebp+60]
		967	.tx2q equ [ebp+68]
		968	;.bx1q equ [ebp+8]
		969	;.bx2q equ [ebp+16]
		970	;.ex1q equ [ebp+24]
		971	;.exyq equ [ebp+32]
		972	.bx1 equ dword [ebp+8] ; ---
		973	.by1 equ dword [ebp+12] ; \|
		974	.bx2 equ dword [ebp+16] ; \|
		975	.by2 equ dword [ebp+20] ; \|> b. map and e. map coords
		976	.ex1 equ dword [ebp+24] ; \|> shifted shl ROUND
		977	.ey1 equ dword [ebp+28] ; \|
		978	.ex2 equ dword [ebp+32] ; \|
		979	.ey2 equ dword [ebp+36] ; ---
		980	.bmap equ dword [ebp+40] ; bump map offset
		981	.emap equ dword [ebp+44] ; env map offset
		982	.z_buff equ dword [ebp+48]
		983	.z2 equ dword [ebp+52] ; -- \|> z coords shifted
		984	.z1 equ dword [ebp+56] ; -- shl CATMULL_SHIFT
		985
		986
		987	.ty1 equ dword [ebp+64] ; \|> shifted shl ROUND
		988	.tx2 equ dword [ebp+68] ; \|
		989	.ty2 equ dword [ebp+72] ; ---
		990	.tex_map equ dword [ebp+76] ; texture offset ( pointer )
		991
		992
		993
		994	.x2 equ dword [ebp-8]
		995	.dbx equ dword [ebp-12]
		996	.dby equ dword [ebp-16]
		997	.dbyq equ qword [ebp-16] ; - new
		998	.dex equ dword [ebp-20]
		999	.dey equ dword [ebp-24]
		1000	.deyq equ qword [ebp-24] ; - new
		1001	.dz equ dword [ebp-28]
		1002	.dtx equ dword [ebp-32]
		1003	.dty equ dword [ebp-36]
		1004	.dtyq equ qword [ebp-36]
		1005
		1006
		1007	.cby equ dword [ebp-44]
		1008	.cbyq equ qword [ebp-44] ; - new
		1009	.cex equ dword [ebp-48]
		1010	.cey equ dword [ebp-52]
		1011	.ceyq equ qword [ebp-52] ; - new
		1012	.cz equ dword [ebp-56]
		1013	.czbuff equ dword [ebp-60]
		1014	.ctx equ dword [ebp-64]
		1015	.cty equ dword [ebp-68]
		1016	.ctyq equ qword [ebp-68]
		1017	.c_scr equ dword [ebp-72]
		1018
		1019
		1020	.temp2 equ ebp-88
		1021	.temp3 equ ebp-76
		1022	.temp4 equ ebp-84
		1023	.temp5 equ ebp-92
		1024
		1025
		1026
		1027
		1028
		1029
		1030	or ecx,ecx
		1031	jl .bl_end
		1032	cmp ecx,SIZE_Y
		1033	jge .bl_end
		1034
		1035
		1036	jl .bl_ok
		1037	je .bl_end
		1038
		1039
		1040	if Ext=NON
		1041	mov edx,.bx1
		1042	xchg edx,.bx2
		1043	mov .bx1,edx
		1044	mov edx,.by1
		1045	xchg edx,.by2
		1046	mov .by1,edx
		1047
		1048
		1049	xchg edx,.ex2
		1050	mov .ex1,edx
		1051	mov edx,.ey1
		1052	xchg edx,.ey2
		1053	mov .ey1,edx
		1054
		1055
		1056	xchg edx,.tx2
		1057	mov .tx1,edx
		1058	mov edx,.ty1
		1059	xchg edx,.ty2
		1060	mov .ty1,edx
		1061	else
		1062	movq mm0,.bx1q
		1063	movq mm1,.bx2q
		1064	movq mm2,.ex1q
		1065	movq mm3,.ex2q
		1066	movq mm4,.tx1q
		1067	movq mm5,.tx2q
		1068	movq .bx2q,mm0
		1069	movq .bx1q,mm1
		1070	movq .ex1q,mm3
		1071	movq .ex2q,mm2
		1072	movq .tx1q,mm5
		1073	movq .tx2q,mm4
		1074	end if
		1075
		1076
		1077	xchg edx,.z2
		1078	mov .z1,edx
		1079	.bl_ok:
		1080	push eax
		1081	push ebx ;store x1, x2
		1082	cmp .x1,SIZE_X
		1083	jge .bl_end
		1084	cmp .x2,0
		1085	jle .bl_end
		1086
		1087
		1088	sub ebx,.x1
		1089
		1090
		1091
		1092
		1093	cvtsi2ss xmm3,ebx ;rcps
		1094	shufps xmm3,xmm3,0
		1095
		1096
		1097	movlhps xmm0,xmm0
		1098	cvtpi2ps xmm0,.ex1q ;mm2
		1099	cvtpi2ps xmm1,.bx2q ;mm1
		1100	movlhps xmm1,xmm1
		1101	cvtpi2ps xmm1,.ex2q ;mm3
		1102	subps xmm1,xmm0
		1103
		1104
		1105
		1106
		1107	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		1108	movhlps xmm1,xmm1
		1109	cvtps2pi mm1,xmm1
		1110	movq .deyq,mm0
		1111	movq .dbyq,mm1
		1112
		1113
		1114	movd mm3,.z2
		1115
		1116
		1117	movlhps xmm0,xmm0
		1118	cvtpi2ps xmm0,mm2
		1119	cvtpi2ps xmm1,.tx2q ;mm1
		1120	movlhps xmm1,xmm1
		1121	cvtpi2ps xmm1,mm3
		1122	subps xmm1,xmm0
		1123
		1124
		1125
		1126
		1127	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		1128	movhlps xmm1,xmm1
		1129	cvtps2pi mm1,xmm1
		1130	movd .dz,mm0
		1131	movq .dtyq,mm1
		1132
		1133
		1134
		1135
		1136	sub eax,.bx1
		1137	cdq
		1138	idiv ebx
		1139	push eax
		1140
		1141
		1142	sub eax,.by1
		1143	cdq
		1144	idiv ebx
		1145	push eax
		1146
		1147
		1148	sub eax,.ex1
		1149	cdq
		1150	idiv ebx
		1151	push eax
		1152
		1153
		1154	sub eax,.ey1
		1155	cdq
		1156	idiv ebx
		1157	push eax
		1158
		1159
		1160
		1161	sub eax,.z1
		1162	cdq
		1163	idiv ebx
		1164	push eax
		1165
		1166
		1167	sub eax,.tx1
		1168	cdq
		1169	idiv ebx
		1170	push eax
		1171
		1172
		1173	sub eax,.ty1
		1174	cdq
		1175	idiv ebx
		1176	push eax
		1177
		1178
		1179	cmp .x1,0 ; set correctly begin variable
		1180	jge @f ; CLIPPING ON FUNCTION
		1181	; cutting triangle exceedes screen
		1182	mov ebx,.x1
		1183	neg ebx
		1184	imul ebx ; eax = .dz * abs(.x1)
		1185	add .z1,eax
		1186	mov .x1,0
		1187
		1188
		1189	imul ebx
		1190	add .bx1,eax
		1191
		1192
		1193	imul ebx
		1194	add .by1,eax
		1195
		1196
		1197	imul ebx
		1198	add .ex1,eax
		1199
		1200
		1201	imul ebx
		1202	add .ey1,eax
		1203
		1204
		1205	imul ebx
		1206	add .tx1,eax
		1207
		1208
		1209	imul ebx
		1210	add .ty1,eax
		1211
		1212
		1213	cmp .x2,SIZE_X
		1214	jl @f
		1215	mov .x2,SIZE_X
		1216	@@:
		1217	mov eax,SIZE_X ;calc memory begin in buffers
		1218	mul .y
		1219	add eax,.x1
		1220	lea esi,[4*eax]
		1221	add esi,.z_buff ; z-buffer filled with dd variables
		1222	lea eax,[eax*3]
		1223	add edi,eax
		1224
		1225
		1226
		1227	sub ecx,.x1
		1228	; init current variables
		1229	push .bx1 ; current b, e and t shifted shl ROUND .cbx
		1230	push .by1 ; .cby
		1231	push .ex1 ; .cex
		1232	push .ey1 ; .cey
		1233
		1234
		1235	push esi ; .czbuff
		1236
		1237
		1238	push .ty1 ; .cty
		1239	push edi ; .c_scr
		1240	if Ext>=MMX
		1241	movq mm7,.ctyq
		1242	movq mm6,.cbyq
		1243	movq mm5,.ceyq
		1244	; movq mm4,.dtyq
		1245	; movq mm3,.dbyq
		1246	end if
		1247
		1248
		1249	; if TEX = SHIFTING ;bump drawing only in shifting mode
		1250	mov esi,.czbuff ; .czbuff current address in buffer
		1251	mov ebx,.cz ; .cz - cur z position
		1252	cmp ebx,dword[esi]
		1253	jge .skip
		1254	if Ext=NON
		1255	mov eax,.cby
		1256	shr eax,ROUND
		1257	mov esi,.cbx
		1258	shr esi,ROUND
		1259	else
		1260	movq mm1,mm6
		1261	psrld mm1,ROUND
		1262	movd eax,mm1
		1263	psrlq mm1,32
		1264	movd esi,mm1
		1265	end if
		1266
		1267
		1268
		1269
		1270	add esi,eax ;- ; esi - current bump map index
		1271
		1272
		1273	dec ebx
		1274	and ebx,TEXTURE_SIZE
		1275	add ebx,.bmap
		1276	movzx eax,byte [ebx]
		1277
		1278
		1279	inc ebx
		1280	and ebx,TEXTURE_SIZE
		1281	add ebx,.bmap
		1282	movzx ebx,byte [ebx]
		1283	sub eax,ebx
		1284
		1285
		1286	sub ebx,TEX_X
		1287	and ebx,TEXTURE_SIZE
		1288	add ebx,.bmap
		1289	movzx edx,byte [ebx]
		1290
		1291
		1292	add ebx,TEX_X
		1293	and ebx,TEXTURE_SIZE
		1294	add ebx,.bmap
		1295	movzx ebx,byte [ebx]
		1296	sub edx,ebx
		1297
		1298
		1299	; edx - vertical sub modificated y coord
		1300	if Ext=NON
		1301	mov ebx,.cex ;.cex - current env map X
		1302	shr ebx,ROUND
		1303	add eax,ebx
		1304
		1305
		1306
		1307	shr ebx,ROUND
		1308	add edx,ebx
		1309
		1310
		1311	movq mm1,mm5 ; mm5 - copy of cur env coords
		1312	psrld mm1,ROUND
		1313	movd ebx,mm1
		1314	psrlq mm1,32
		1315	add eax,ebx
		1316	movd ebx,mm1
		1317	add edx,ebx
		1318	; movq qword[.temp1],mm3
		1319	; add eax,dword [.temp1]
		1320	; add edx,dword [.temp1+4]
		1321	end if
		1322
		1323
		1324	jl .black
		1325	cmp eax,TEX_X
		1326	jg .black
		1327	or edx,edx
		1328	jl .black
		1329	cmp edx,TEX_Y
		1330	jg .black
		1331
		1332
		1333	add edx,eax ; proponuje nie stawiac czarnego pixela tylko
		1334	lea esi,[edx*3] ; niezaburzony.
		1335	add esi,.emap ;
		1336	lodsd
		1337
		1338
		1339	mov edx,.cty
		1340	shr edx,ROUND ; sar
		1341
		1342
		1343	shr edi,ROUND ; sar
		1344	else
		1345	movq mm1,mm7
		1346	psrld mm1,ROUND
		1347	movd edx,mm1
		1348	psrlq mm1,32
		1349	movd edi,mm1
		1350
		1351
		1352
		1353
		1354	add edi,edx
		1355	and edi,TEXTURE_SIZE
		1356	lea esi,[edi*3]
		1357	add esi,.tex_map
		1358
		1359
		1360	mov edx,eax
		1361	lodsd
		1362	push ax
		1363	mul dl
		1364	mov dl,ah
		1365	pop ax
		1366	shr ax,8
		1367	mul dh
		1368	mov al,dl
		1369	mov edi,.c_scr
		1370	stosw
		1371	shr edx,16
		1372	shr eax,16
		1373	mul dl
		1374	shr ax,8
		1375	stosb
		1376	else
		1377	movd mm0,eax
		1378	pxor mm1,mm1
		1379	punpcklbw mm0,mm1
		1380	movd mm2,[esi]
		1381	punpcklbw mm2,mm1
		1382	pmullw mm0,mm2
		1383	psrlw mm0,8
		1384	packuswb mm0,mm1
		1385	mov edi,.c_scr
		1386	movd [edi],mm0
		1387
		1388
		1389
		1390
		1391	@@:
		1392	.black:
		1393	xor eax,eax
		1394	mov edi,.c_scr
		1395	stosd
		1396	.actual_zbuff:
		1397	mov eax,.cz
		1398	mov edi,.czbuff
		1399	stosd
		1400
		1401
		1402	add .czbuff,4
		1403	add .c_scr,3
		1404
		1405
		1406	mov eax,.dbx
		1407	add .cbx,eax
		1408	mov ebx,.dby
		1409	add .cby,ebx
		1410
		1411
		1412	add .cex,edx
		1413	mov eax,.dey
		1414	add .cey,eax
		1415
		1416
		1417	add .ctx,ebx
		1418	mov edx,.dty
		1419	add .cty,edx
		1420
		1421
		1422	paddd mm7,.dtyq
		1423	paddd mm6,.dbyq
		1424	paddd mm5,.deyq
		1425	end if
		1426	mov eax,.dz
		1427	add .cz,eax
		1428
		1429
		1430	jnz .draw
		1431
		1432
		1433	mov esp,ebp
		1434	ret 76
		1435	;Ext = MMX
		1436
		1437
		1438	; movq mm5, qword[.temp1] ;-
		1439	; paddd mm5, qword[.temp5] ; .temp5 == low dword = TEX_X, high dword = -TEX_X
		1440	; pand mm5, qword[.temp3] ; .temp3 == low = high dword = TEX_SIZE
		1441	; paddd mm5, qword[.temp4] ; .temp4 == low = high dword = offset .bmap
		1442	; movd ebx,mm5
		1443	; psrlq mm5,32
		1444	; end if
		1445

Subversion Repositories Kolibri OS

(root)/programs/demos/3DS/BUMP_TEX.INC – Rev 1776