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
2881	leency	63
		64
		65	.dz12 equ [edi-8]
		66	.dbx12 equ dword[edi-12]
		67	.dby12 equ [edi-16]
		68	.dex12 equ dword[edi-20]
		69	.dey12 equ [edi-24]
		70	.dtx12 equ dword[edi-28]
		71	.dty12 equ [edi-32]
		72
		73
		74	.dz13 equ [ebp-52-4*2]
		75	.dbx13 equ dword[ebp-52-4*3]
		76	.dby13 equ [ebp-52-4*4]
		77	.dex13 equ dword[ebp-52-4*5]
		78	.dey13 equ [ebp-52-4*6]
		79	.dtx13 equ dword[ebp-52-4*7]
		80	.dty13 equ [ebp-52-4*8]
		81
		82
		83
		84	.dz23 equ [ebp-(52+4*10)]
		85	.dbx23 equ dword[ebp-(52+4*11)]
		86	.dby23 equ [ebp-(52+4*12)]
		87	.dex23 equ dword[ebp-(52+4*13)]
		88	.dey23 equ [ebp-(52+4*14)]
		89	.dtx23 equ dword[ebp-(52+4*15)]
		90	.dty23 equ [ebp-(52+4*16)]
		91
		92
		93
		94
1245	hidnplayr	95	.dz12 equ [ebp-28]
		96	.dbx12 equ dword[ebp-32]
		97	.dby12 equ [ebp-36]
		98	.dex12 equ dword[ebp-40]
		99	.dey12 equ [ebp-44]
		100	.dtx12 equ dword[ebp-48]
		101	.dty12 equ [ebp-52]
		102
		103
		104	.dz13 equ [ebp-52-4*2]
		105	.dbx13 equ dword[ebp-52-4*3]
		106	.dby13 equ [ebp-52-4*4]
		107	.dex13 equ dword[ebp-52-4*5]
		108	.dey13 equ [ebp-52-4*6]
		109	.dtx13 equ dword[ebp-52-4*7]
		110	.dty13 equ [ebp-52-4*8]
		111
		112
		113
		114	.dz23 equ [ebp-(52+4*10)]
		115	.dbx23 equ dword[ebp-(52+4*11)]
		116	.dby23 equ [ebp-(52+4*12)]
		117	.dex23 equ dword[ebp-(52+4*13)]
		118	.dey23 equ [ebp-(52+4*14)]
		119	.dtx23 equ dword[ebp-(52+4*15)]
		120	.dty23 equ [ebp-(52+4*16)]
		121
		122
2881	leency	123
		124
1819	yogev_ezra	125
		126
1245	hidnplayr	127	.cz1 equ [ebp-(52+4*18)]
		128	.cx2 equ dword[ebp-(52+4*19)]
		129	.cz2 equ [ebp-(52+4*20)]
		130	.cbx1 equ dword[ebp-(52+4*21)]
		131	.cby1 equ [ebp-(52+4*22)]
		132	.cbx2 equ dword[ebp-(52+4*23)]
		133	.cby2 equ [ebp-(52+4*24)]
		134	.cex1 equ dword[ebp-(52+4*25)]
		135	.cey1 equ [ebp-(52+4*26)]
		136	.cex2 equ dword[ebp-(52+4*27)]
		137	.cey2 equ [ebp-(52+4*28)]
		138
		139
		140	.cty1 equ [ebp-(52+4*30)]
		141	.ctx2 equ dword[ebp-(52+4*31)]
		142	.cty2 equ [ebp-(52+4*32)]
		143
		144
1819	yogev_ezra	145
		146
		147	.cz1 equ [ebp-(52+4*18)]
		148	.cbx1 equ dword[ebp-(52+4*19)]
		149	.cby1 equ [ebp-(52+4*20)]
		150	.cex1 equ dword[ebp-(52+4*21)]
		151	.cey1 equ [ebp-(52+4*22)]
		152	.ctx1 equ dword[ebp-(52+4*23)]
		153	.cty1 equ [ebp-(52+4*24)]
		154
		155
		156	.cz2 equ [ebp-(52+4*26)]
		157	.cbx2 equ dword[ebp-(52+4*27)]
		158	.cby2 equ [ebp-(52+4*28)]
		159	.cex2 equ dword[ebp-(52+4*29)]
		160	.cey2 equ [ebp-(52+4*30)]
		161	.ctx2 equ dword[ebp-(52+4*31)]
		162	.cty2 equ [ebp-(52+4*32)]
		163
		164
		165	cld
1245	hidnplayr	166	mov ebp,esp
		167	push edx ; store bump map
		168	push esi ; store e. map
		169	; sub esp,120
		170	.sort3: ; sort triangle coordinates...
		171	cmp ax,bx
		172	jle .sort1
		173	xchg eax,ebx
		174	mov edx,dword[.b_x1]
		175	xchg edx,dword[.b_x2]
		176	mov dword[.b_x1],edx
		177	mov edx,dword[.e_x1]
		178	xchg edx,dword[.e_x2]
		179	mov dword[.e_x1],edx
		180	mov edx,dword[.t_x1]
		181	xchg edx,dword[.t_x2]
		182	mov dword[.t_x1],edx
		183	mov dx,.z1
		184	xchg dx,.z2
		185	mov .z1,dx
		186	.sort1:
		187	cmp bx,cx
		188	jle .sort2
		189	xchg ebx,ecx
		190	mov edx,dword[.b_x2]
		191	xchg edx,dword[.b_x3]
		192	mov dword[.b_x2],edx
		193	mov edx,dword[.e_x2]
		194	xchg edx,dword[.e_x3]
		195	mov dword[.e_x2],edx
		196	mov edx,dword[.t_x2]
		197	xchg edx,dword[.t_x3]
		198	mov dword[.t_x2],edx
		199	mov dx,.z2
		200	xchg dx,.z3
		201	mov .z2,dx
		202	jmp .sort3
		203	.sort2:
		204	push eax ; store triangle coords in variables
		205	push ebx
		206	push ecx
		207	mov edx,80008000h ; eax,ebx,ecx are ANDd together into edx which means that
		208	and edx,ebx ; if all of them are negative a sign flag is raised
		209	and edx,ecx
		210	and edx,eax
		211	test edx,80008000h ; Check both X&Y at once
		212	jne .loop23_done
		213	; mov edx,eax ; eax,ebx,ecx are ORd together into edx which means that
		214	; or edx,ebx ; if any one of them is negative a sign flag is raised
		215	; or edx,ecx
		216	; test edx,80000000h ; Check only X
		217	; jne .loop23_done
		218
		219
		220	; jg .loop23_done
		221	; cmp .x2,SIZE_X ; This can be optimized with effort
		222	; jg .loop23_done
		223	; cmp .x3,SIZE_X
		224	; jg .loop23_done ; {
		225
		226
		227
		228	sub bx,.y1
		229	jnz .bt_dx12_make
		230	if 0 ;Ext >= SSE2
2881	leency	231	pxor xmm0,xmm0
		232	movups .dty12,xmm0
		233	movups .dey12,xmm0
		234	sub esp,16
		235	else
		236	mov ecx,8
1245	hidnplayr	237	xor edx,edx
		238	@@:
		239	push edx ;dword 0
		240	loop @b
		241	end if
2881	leency	242	jmp .bt_dx12_done
1245	hidnplayr	243	.bt_dx12_make:
		244	movsx ebx,bx
1776	yogev_ezra	245
1245	hidnplayr	246
		247
		248	sub esp,32
1819	yogev_ezra	249	; mov eax,256
1245	hidnplayr	250	cvtsi2ss xmm4,[i255d]
		251	cvtsi2ss xmm3,ebx ;rcps
2881	leency	252	if 0 ;Ext >= SSE2
		253	mov edi,ebp
		254	sub edi,512
		255	or edi,0x0000000f
		256	end if
		257	divss xmm3,xmm4
1245	hidnplayr	258	shufps xmm3,xmm3,0
		259
		260
		261	movd mm1,[.b_x2]
		262	movd mm2,[.e_x1]
		263	movd mm3,[.e_x2]
		264
		265
		266	punpcklwd mm0,mm4
		267	punpcklwd mm1,mm4
		268	punpcklwd mm2,mm4
		269	punpcklwd mm3,mm4
		270
		271
		272	psubd mm3,mm2
		273
		274
		275	movlhps xmm1,xmm1
		276	cvtpi2ps xmm1,mm3
		277
		278
		279
		280
		281	;xmm1--> \| dbx \| dby \| dex \| dey \|
		282	;1 movups .dey12,xmm1
1819	yogev_ezra	283	cvtps2pi mm0,xmm1 ;mm0,xmm1 ; mm0 -> 2 delta dwords
1776	yogev_ezra	284	movhlps xmm1,xmm1
1245	hidnplayr	285	cvtps2pi mm1,xmm1 ;mm1,xmm1
1776	yogev_ezra	286	movq .dey12,mm0
1245	hidnplayr	287	movq .dby12,mm1
		288	;-------------
		289	; pxor mm0,mm0
1776	yogev_ezra	290	; pxor mm1,mm1
		291	;/ pinsrw mm0,.z1,1
		292	;/ pinsrw mm0,.x1,0
		293	;/ pinsrw mm1,.z2,1
		294	;/ pinsrw mm1,.x2,0
		295	mov ax,.z2
		296	sub ax,.z1
		297	cwde
		298
1819	yogev_ezra	299
		300	sub dx,.x1
		301	movsx edx,dx
		302
		303
1776	yogev_ezra	304
		305
		306	;/ punpcklwd mm1,mm4
		307
		308
		309	; cvtpi2ps xmm2,mm0
		310	; subps xmm1,xmm2
		311
		312
		313
		314
		315	movd mm3,[.t_x2]
		316
		317
		318	punpcklwd mm3,mm4
		319	psubd mm3,mm2
		320
		321
		322	cvtsi2ss xmm1,eax
		323	movlhps xmm1,xmm1
		324	cvtsi2ss xmm1,edx
1819	yogev_ezra	325	; movss xmm1,xmm4
		326	shufps xmm1,xmm1,00101111b
		327	cvtpi2ps xmm1,mm3
1776	yogev_ezra	328
		329
1819	yogev_ezra	330
1776	yogev_ezra	331
		332	; xmm1--> \| dx \| dz \| dtx \| dty \|
1819	yogev_ezra	333	;1 movlps .dty12,xmm1
		334	;1 movhps .dz12,xmm1
		335	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords \| dtx \| dty \|
1776	yogev_ezra	336	movhlps xmm1,xmm1
		337	cvtps2pi mm1,xmm1
1819	yogev_ezra	338	movq .dty12,mm0
1776	yogev_ezra	339	movq .dz12,mm1
1819	yogev_ezra	340	;----
1776	yogev_ezra	341	; mov ax,.z2
1245	hidnplayr	342	; sub ax,.z1
		343	; cwde
		344	; mov bx,.x2
		345	; sub bx,.x1
		346	; movsx ebx,bx
		347	; movd mm1,eax
		348	; psllq mm1,32
		349	; movd mm1,ebx
		350
1776	yogev_ezra	351
1245	hidnplayr	352	;; push eax
		353	;; movq mm1,[esp]
		354	;; add esp,8
		355	;;; mov ax,.z1
		356	;;; mov bx,.z2
		357	;;; shl eax,16
		358	;;; shl ebx,16
		359	;;; mov ax,.x1
		360	;;; mov bx,.x2
		361	; movd mm2,[.t_x1]
		362	; movd mm3,[.t_x2]
		363	;; movd mm0,eax
		364	;; movd mm1,ebx
		365
		366
		367	;; punpcklwd mm0,mm4
		368	;; punpcklwd mm1,mm4
		369	; punpcklwd mm2,mm4
		370	; punpcklwd mm3,mm4
		371
		372
		373	; psubd mm3,mm2
		374
		375
		376
		377	; movlhps xmm1,xmm1
		378	; cvtpi2ps xmm1,mm3
		379
		380
		381
		382
		383	; xmm1--> \| dx \| dz \| dtx \| dty \|
		384	; cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords \| dtx \| dty \|
		385	; movhlps xmm1,xmm1
		386	; cvtps2pi mm1,xmm1 ; mm1 --> 2 delta dwords \| dx \| dz \|
		387	; movq .dty12,mm0
		388	; movq .dz12,mm1
		389	else
		390	mov ax,.x2
1819	yogev_ezra	391	sub ax,.x1
		392	cwde
		393	shl eax,ROUND
		394	cdq
		395	idiv ebx
		396	; mov .dx12,eax
		397	push eax
		398
		399
1776	yogev_ezra	400	sub ax,.z1
		401	cwde
		402	shl eax,CATMULL_SHIFT
		403	cdq
		404	idiv ebx
		405	push eax
		406
1245	hidnplayr	407
		408	sub ax,word[.b_x1]
		409	cwde
		410	shl eax,ROUND
		411	cdq
		412	idiv ebx
		413	; mov .dbx12,eax
		414	push eax
		415
		416
		417	sub ax,word[.b_y1]
		418	cwde
		419	shl eax,ROUND
		420	cdq
		421	idiv ebx
		422	; mov .dby12,eax
		423	push eax
		424
		425
		426	sub ax,word[.e_x1]
		427	cwde
		428	shl eax,ROUND
		429	cdq
		430	idiv ebx
		431	; mov .dex12,eax
		432	push eax
		433
		434
		435	sub ax,word[.e_y1]
		436	cwde
		437	shl eax,ROUND
		438	cdq
		439	idiv ebx
		440	; mov .dey12,eax
		441	push eax
		442
		443
		444	sub ax,word[.t_x1]
		445	cwde
		446	shl eax,ROUND
		447	cdq
		448	idiv ebx
		449	; mov .dtx12,eax
		450	push eax
		451
		452
		453	sub ax,word[.t_y1]
		454	cwde
		455	shl eax,ROUND
		456	cdq
		457	idiv ebx
		458	; mov .dty12,eax
		459	push eax
		460	end if
1776	yogev_ezra	461	.bt_dx12_done:
1245	hidnplayr	462
		463
		464	sub bx,.y1
		465	jnz .bt_dx13_make
		466	mov ecx,8
		467	xor edx,edx
		468	@@:
		469	push edx ;dword 0
		470	loop @b
		471	jmp .bt_dx13_done
		472	.bt_dx13_make:
		473	movsx ebx,bx
		474
		475
		476
		477
1819	yogev_ezra	478	; mov eax,256
		479	cvtsi2ss xmm4,[i255d]
1245	hidnplayr	480	cvtsi2ss xmm3,ebx ;rcps
		481	divss xmm3,xmm4
		482	shufps xmm3,xmm3,0
		483
		484
		485	movd mm1,[.b_x3]
		486	movd mm2,[.e_x1]
		487	movd mm3,[.e_x3]
		488
		489
		490	punpcklwd mm0,mm4
		491	punpcklwd mm1,mm4
		492	punpcklwd mm2,mm4
		493	punpcklwd mm3,mm4
		494
		495
		496	psubd mm3,mm2
		497
		498
		499	movlhps xmm1,xmm1
		500	cvtpi2ps xmm1,mm3
		501
		502
		503
		504
		505	;xmm1--> \| dbx \| dby \| dex \| dey \|
		506	;1 movups .dey13,xmm1
1819	yogev_ezra	507
		508
		509	movhlps xmm1,xmm1
1245	hidnplayr	510	cvtps2pi mm1,xmm1 ;mm1,xmm1
1819	yogev_ezra	511	movq .dey13,mm0
1245	hidnplayr	512	movq .dby13,mm1
		513
1819	yogev_ezra	514
1776	yogev_ezra	515	sub ax,.z1
		516	cwde
		517
		518
1819	yogev_ezra	519	sub dx,.x1
		520	movsx edx,dx
		521
		522
1776	yogev_ezra	523	movd mm3,[.t_x3]
		524
		525
		526	punpcklwd mm3,mm4
		527	psubd mm3,mm2
		528
		529
		530	movlhps xmm1,xmm1
		531	cvtsi2ss xmm1,edx
1819	yogev_ezra	532	shufps xmm1,xmm1,00101111b
		533	cvtpi2ps xmm1,mm3
1776	yogev_ezra	534
		535
1819	yogev_ezra	536
1776	yogev_ezra	537
		538	; xmm1--> \| dx \| dz \| dtx \| dty \|
1819	yogev_ezra	539	;1 movlps .dty13,xmm1
		540	;1 movhps .dz13,xmm1
		541
		542
1776	yogev_ezra	543	movhlps xmm1,xmm1
		544	cvtps2pi mm1,xmm1
1819	yogev_ezra	545	movq .dty13,mm0
1776	yogev_ezra	546	movq .dz13,mm1
1819	yogev_ezra	547
		548
1245	hidnplayr	549
1819	yogev_ezra	550
		551	sub ax,.x1
		552	cwde
		553	shl eax,ROUND
		554	cdq
		555	idiv ebx
		556	; mov .dx13,eax
		557	push eax
		558
		559
1776	yogev_ezra	560	sub ax,.z1
		561	cwde
		562	shl eax,CATMULL_SHIFT
		563	cdq
		564	idiv ebx
		565	; mov .dz13,eax
		566	push eax
		567
		568
		569
1245	hidnplayr	570	sub ax,word[.b_x1]
		571	cwde
		572	shl eax,ROUND
		573	cdq
		574	idiv ebx
		575	; mov .dbx13,eax
		576	push eax
		577
		578
		579	sub ax,word[.b_y1]
		580	cwde
		581	shl eax,ROUND
		582	cdq
		583	idiv ebx
		584	; mov .dby13,eax
		585	push eax
		586
		587
		588	sub ax,word[.e_x1]
		589	cwde
		590	shl eax,ROUND
		591	cdq
		592	idiv ebx
		593	; mov .dex13,eax
		594	push eax
		595
		596
		597	sub ax,word[.e_y1]
		598	cwde
		599	shl eax,ROUND
		600	cdq
		601	idiv ebx
		602	; mov .dey13,eax
		603	push eax
		604
		605
		606	sub ax,word[.t_x1]
		607	cwde
		608	shl eax,ROUND
		609	cdq
		610	idiv ebx
		611	; mov .dtx13,eax
		612	push eax
		613
		614
		615	sub ax,word[.t_y1]
		616	cwde
		617	shl eax,ROUND
		618	cdq
		619	idiv ebx
		620	; mov .dty13,eax
		621	push eax
		622	end if
1776	yogev_ezra	623	.bt_dx13_done:
1245	hidnplayr	624
		625
		626	sub bx,.y2
		627	jnz .bt_dx23_make
		628	mov ecx,8
		629	xor edx,edx
		630	@@:
		631	push edx ;dword 0
		632	loop @b
		633	jmp .bt_dx23_done
		634	.bt_dx23_make:
		635	movsx ebx,bx
		636
		637
		638
		639
1819	yogev_ezra	640	; mov eax,256
		641	cvtsi2ss xmm4,[i255d]
1245	hidnplayr	642	cvtsi2ss xmm3,ebx ;rcps
		643	divss xmm3,xmm4
		644	shufps xmm3,xmm3,0
		645
		646
		647	movd mm1,[.b_x3]
		648	movd mm2,[.e_x2]
		649	movd mm3,[.e_x3]
		650
		651
		652	punpcklwd mm0,mm4
		653	punpcklwd mm1,mm4
		654	punpcklwd mm2,mm4
		655	punpcklwd mm3,mm4
		656
		657
		658	psubd mm3,mm2
		659
		660
		661	movlhps xmm1,xmm1
		662	cvtpi2ps xmm1,mm3
		663
		664
		665
		666
		667	;xmm1--> \| dbx \| dby \| dex \| dey \|
		668	;1 movups .dey23,xmm1
1819	yogev_ezra	669
		670
		671	movhlps xmm1,xmm1
1245	hidnplayr	672	cvtps2pi mm1,xmm1 ;mm1,xmm1
1819	yogev_ezra	673	movq .dey23,mm0
1245	hidnplayr	674	movq .dby23,mm1
		675
1819	yogev_ezra	676
1776	yogev_ezra	677	sub ax,.z2
		678	cwde
		679
		680
1819	yogev_ezra	681	sub dx,.x2
		682	movsx edx,dx
		683
		684
1776	yogev_ezra	685	movd mm3,[.t_x3]
		686
		687
		688	punpcklwd mm3,mm4
		689	psubd mm3,mm2
		690
		691
		692	movlhps xmm1,xmm1
		693	cvtsi2ss xmm1,edx
1819	yogev_ezra	694	shufps xmm1,xmm1,00101111b
		695	cvtpi2ps xmm1,mm3
1776	yogev_ezra	696
		697
1819	yogev_ezra	698
1776	yogev_ezra	699
		700	; xmm1--> \| dx \| dz \| dtx \| dty \|
1819	yogev_ezra	701	; movlps .dty23,xmm1
		702	; movhps .dz23,xmm1
		703	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords \| dtx \| dty \|
1776	yogev_ezra	704	movhlps xmm1,xmm1
		705	cvtps2pi mm1,xmm1 ; mm1 --> 2 delta dwords \| dx \| dz \|
1819	yogev_ezra	706	movq .dty23,mm0
1776	yogev_ezra	707	movq .dz23,mm1
1819	yogev_ezra	708
		709
		710
1245	hidnplayr	711	mov ax,.x3
1819	yogev_ezra	712	sub ax,.x2
		713	cwde
		714	shl eax,ROUND
		715	cdq
		716	idiv ebx
		717	; mov .dx23,eax
		718	push eax
		719
1776	yogev_ezra	720
		721	sub ax,.z2
		722	cwde
		723	shl eax,CATMULL_SHIFT
		724	cdq
		725	idiv ebx
		726	; mov .dz23,eax
		727	push eax
		728
		729
1245	hidnplayr	730	sub ax,word[.b_x2]
		731	cwde
		732	shl eax,ROUND
		733	cdq
		734	idiv ebx
		735	; mov .dbx23,eax
		736	push eax
		737
		738
		739	sub ax,word[.b_y2]
		740	cwde
		741	shl eax,ROUND
		742	cdq
		743	idiv ebx
		744	; mov .dby23,eax
		745	push eax
		746
		747
		748	sub ax,word[.e_x2]
		749	cwde
		750	shl eax,ROUND
		751	cdq
		752	idiv ebx
		753	; mov .dex23,eax
		754	push eax
		755
		756
		757	sub ax,word[.e_y2]
		758	cwde
		759	shl eax,ROUND
		760	cdq
		761	idiv ebx
		762	; mov .dey23,eax
		763	push eax
		764
		765
1776	yogev_ezra	766
1245	hidnplayr	767	sub ax,word[.t_x2]
		768	cwde
		769	shl eax,ROUND
		770	cdq
		771	idiv ebx
		772	; mov .dtx23,eax
		773	push eax
		774
		775
		776	sub ax,word[.t_y2]
		777	cwde
		778	shl eax,ROUND
		779	cdq
		780	idiv ebx
		781	; mov .dty23,eax
		782	push eax
		783	end if
1776	yogev_ezra	784	; sub esp,40
1245	hidnplayr	785	.bt_dx23_done:
		786	sub esp,64
		787
1819	yogev_ezra	788
1245	hidnplayr	789	shl eax,ROUND
		790	mov .cx1,eax
		791	mov .cx2,eax
		792	; push eax
		793	; push eax
		794
		795
		796	shl ebx,ROUND
		797	mov .cbx1,ebx
		798	mov .cbx2,ebx
		799	; push ebx
		800	; push ebx
		801
		802
		803	shl ecx,ROUND
		804	mov .cby1,ecx
		805	mov .cby2,ecx
		806	; push ecx
		807	; push ecx
		808
		809
		810	shl edx,ROUND
		811	mov .cex1,edx
		812	mov .cex2,edx
		813	; push edx
		814	; push edx
		815
		816
		817	shl eax,ROUND
		818	mov .cey1,eax
		819	mov .cey2,eax
		820	; push eax
		821	; push eax
		822
		823
		824	shl ebx,CATMULL_SHIFT
		825	mov .cz1,ebx
		826	mov .cz2,ebx
		827	; push ebx
		828	; push ebx
		829
		830
		831	movsx ecx,word[.t_x1]
		832	shl ecx,ROUND
		833	mov .ctx1,ecx
		834	mov .ctx2,ecx
		835	;push ecx
		836	;push ecx
		837
		838
		839	shl edx,ROUND
		840	mov .cty1,edx
		841	mov .cty2,edx
		842	; push edx
		843	; push edx
		844
		845
2881	leency	846	movups xmm0,.cby1
		847	movups xmm1,.cty1
		848	movups xmm2,.cby2
		849	movups xmm3,.cty2
		850	movups xmm4,.dby13
		851	movups xmm5,.dty13
		852	movups xmm6,.dby12
		853	movups xmm7,.dty12
		854	.scby1 equ [edi]
		855	.scty1 equ [edi+16]
		856	.scby2 equ [edi+32]
		857	.scty2 equ [edi+48]
		858	.sdby13 equ [edi+64]
		859	.sdty13 equ [edi+80]
		860	.sdby12 equ [edi+96]
		861	.sdty12 equ [edi+128]
		862	push edi
		863	mov edi,sse_repository
		864	movaps .scby1,xmm0
		865	movaps .scty1,xmm1
		866	movaps .scby2,xmm2
		867	movaps .scty2,xmm3
		868	movaps .sdby13,xmm4
		869	movaps .sdty13,xmm5
		870	movaps .sdby12,xmm6
		871	movaps .sdty12,xmm7
		872	pop edi
		873
		874
		875	movsx ecx,.y1
2192	leency	876	cmp cx,.y2
		877	jge .loop12_done
		878	.loop12:
		879	;if Ext >= SSE2
		880	; fxsave [sse_repository]
		881	;end if
		882	call .call_line
		883	if Ext >= SSE2
		884	; fxrstor [sse_repository]
		885	movups xmm0,.cby1
		886	movups xmm1,.cty1
		887	movups xmm2,.cby2
		888	movups xmm3,.cty2
		889	; movups xmm4,.dby13
2881	leency	890	; movups xmm5,.dty13
		891	; movups xmm6,.dby12
		892	; movups xmm7,.dty12
		893	; paddd xmm0,xmm4
		894	; paddd xmm1,xmm5
		895	; paddd xmm2,xmm6
		896	; paddd xmm3,xmm7
		897	push edi
		898	mov edi,sse_repository
		899	paddd xmm0,.sdby13
		900	paddd xmm1,.sdty13
		901	paddd xmm2,.sdby12
		902	paddd xmm3,.sdty12
		903	pop edi
		904	movups .cby1,xmm0
2192	leency	905	movups .cty1,xmm1
		906	movups .cby2,xmm2
		907	movups .cty2,xmm3
		908	end if
		909
1245	hidnplayr	910
2192	leency	911	movq mm0,.cby2
1245	hidnplayr	912	movq mm1,.cby1
		913	movq mm2,.cey2
		914	movq mm3,.cey1
		915	movq mm4,.cty1
		916	movq mm5,.cty2
		917	movq mm6,.cz1
		918	movq mm7,.cz2
		919	paddd mm0,.dby12
		920	paddd mm1,.dby13
		921	paddd mm2,.dey12
		922	paddd mm3,.dey13
		923	paddd mm4,.dty13
		924	paddd mm5,.dty12
		925	paddd mm6,.dz13
		926	paddd mm7,.dz12
		927	movq .cby2,mm0
		928	movq .cby1,mm1
		929	movq .cey1,mm3
		930	movq .cey2,mm2
		931	movq .cty1,mm4
		932	movq .cty2,mm5
		933	movq .cz1,mm6
		934	movq .cz2,mm7
		935	end if
1819	yogev_ezra	936	if Ext = NON
		937	mov edx,.dbx13
1245	hidnplayr	938	add .cbx1,edx
		939	mov eax,.dbx12
		940	add .cbx2,eax
		941	mov ebx,.dby13
		942	add .cby1,ebx
		943	mov edx,.dby12
		944	add .cby2,edx
		945
		946
		947	add .cex1,eax
		948	mov ebx,.dex12
		949	add .cex2,ebx
		950	mov edx,.dey13
		951	add .cey1,edx
		952	mov eax,.dey12
		953	add .cey2,eax
		954
		955
		956	add .ctx1,eax
		957	mov ebx,.dtx12
		958	add .ctx2,ebx
		959	mov edx,.dty13
		960	add .cty1,edx
		961	mov eax,.dty12
		962	add .cty2,eax
		963
		964
		965	add .cx1,eax
		966	mov ebx,.dx12
		967	add .cx2,ebx
		968	mov ebx,.dz13
		969	add .cz1,ebx
		970	mov edx,.dz12
		971	add .cz2,edx
		972	end if
		973	inc ecx
		974	cmp cx,.y2
		975	jl .loop12
		976	.loop12_done:
		977
		978
		979	cmp cx,.y3
		980	jge .loop23_done
		981
		982
1819	yogev_ezra	983
1245	hidnplayr	984	shl eax,CATMULL_SHIFT
		985	mov .cz2,eax
		986
		987
		988	shl ebx,ROUND
		989	mov .cx2,ebx
		990
		991
		992	shl edx,ROUND
		993	mov .cbx2,edx
		994
		995
		996	shl eax,ROUND
		997	mov .cby2,eax
		998
		999
		1000	shl ebx,ROUND
		1001	mov .cex2,ebx
		1002
		1003
		1004	shl edx,ROUND
		1005	mov .cey2,edx
		1006
		1007
		1008	shl eax,ROUND
		1009	mov .ctx2,eax
		1010
		1011
		1012	shl ebx,ROUND
		1013	mov .cty2,ebx
		1014	if Ext >= SSE2
2881	leency	1015	movups xmm2,.cby2
		1016	movups xmm3,.cty2
		1017	; movups xmm4,.dby13
		1018	; movups xmm5,.dty13
		1019	movups xmm6,.dby23
		1020	movups xmm7,.dty23
		1021	; .scby1 equ [edi]
		1022	; .scty1 equ [edi+16]
		1023	; .scby2 equ [edi+32]
		1024	; .scty2 equ [edi+48]
		1025	; .sdby13 equ [edi+64]
		1026	; .sdty13 equ [edi+80]
		1027	.sdby23 equ [edi+160]
		1028	.sdty23 equ [edi+192]
		1029	push edi
		1030	mov edi,sse_repository
		1031	; movaps .scby1,xmm0
		1032	; movaps .scty1,xmm1
		1033	movaps .scby2,xmm2
		1034	movaps .scty2,xmm3
		1035	; movaps .sdby13,xmm4
		1036	; movaps .sdty13,xmm5
		1037	movaps .sdby23,xmm6
		1038	movaps .sdty23,xmm7
		1039	pop edi
		1040
		1041
		1042
		1043
2192	leency	1044	;if Ext >= SSE2
		1045	; fxsave [sse_repository]
		1046	;end if
		1047	call .call_line
		1048
		1049
		1050
2881	leency	1051
2192	leency	1052	movups xmm1,.cty1
		1053	movups xmm2,.cby2
		1054	movups xmm3,.cty2
		1055
2881	leency	1056
		1057
		1058	mov edi,sse_repository
		1059	paddd xmm0,.sdby13
		1060	paddd xmm1,.sdty13
		1061	paddd xmm2,.sdby23
		1062	paddd xmm3,.sdty23
		1063	pop edi
		1064	movups .cby1,xmm0
2192	leency	1065	movups .cty1,xmm1
		1066	movups .cby2,xmm2
		1067	movups .cty2,xmm3
		1068
2881	leency	1069
		1070
		1071
		1072
		1073	; movups xmm0,.cby1
		1074	; movups xmm1,.cty1
		1075	; movups xmm2,.cby2
		1076	; movups xmm3,.cty2
		1077	; movups xmm4,.dby13
		1078	; movups xmm5,.dty13
		1079	; movups xmm6,.dby23
		1080	; movups xmm7,.dty23
		1081	; paddd xmm0,xmm4
		1082	; paddd xmm1,xmm5
		1083	; paddd xmm2,xmm6
		1084	; paddd xmm3,xmm7
		1085	; movups .cby1,xmm0
		1086	; movups .cty1,xmm1
		1087	; movups .cby2,xmm2
		1088	; movups .cty2,xmm3
		1089	;
1819	yogev_ezra	1090	end if
2192	leency	1091	if (Ext = MMX) \| (Ext = SSE)
		1092	movq mm0,.cby2
1245	hidnplayr	1093	movq mm1,.cby1
		1094	movq mm2,.cey2
		1095	movq mm3,.cey1
		1096	movq mm4,.cty1
		1097	movq mm5,.cty2
		1098	movq mm6,.cz1
		1099	movq mm7,.cz2
		1100	paddd mm0,.dby23
		1101	paddd mm1,.dby13
		1102	paddd mm2,.dey23
		1103	paddd mm3,.dey13
		1104	paddd mm4,.dty13
		1105	paddd mm5,.dty23
		1106	paddd mm6,.dz13
		1107	paddd mm7,.dz23
		1108	movq .cby2,mm0
		1109	movq .cby1,mm1
		1110	movq .cey2,mm2
		1111	movq .cey1,mm3
		1112	movq .cty1,mm4
		1113	movq .cty2,mm5
		1114	movq .cz1,mm6
		1115	movq .cz2,mm7
		1116	end if
1819	yogev_ezra	1117	If Ext = NON
		1118	mov edx,.dbx13
1245	hidnplayr	1119	add .cbx1,edx
		1120	mov eax,.dbx23
		1121	add .cbx2,eax
		1122	mov ebx,.dby13
		1123	add .cby1,ebx
		1124	mov edx,.dby23
		1125	add .cby2,edx
		1126
		1127
		1128	add .cex1,eax
		1129	mov ebx,.dex23
		1130	add .cex2,ebx
		1131	mov edx,.dey13
		1132	add .cey1,edx
		1133	mov eax,.dey23
		1134	add .cey2,eax
		1135
		1136
		1137	add .cx1,eax
		1138	mov ebx,.dx23
		1139	add .cx2,ebx
		1140	mov ebx,.dz13
		1141	add .cz1,ebx
		1142	mov edx,.dz23
		1143	add .cz2,edx
		1144
		1145
		1146	add .ctx1,eax
		1147	mov ebx,.dtx23
		1148	add .ctx2,ebx
		1149	mov edx,.dty13
		1150	add .cty1,edx
		1151	mov eax,.dty23
		1152	add .cty2,eax
		1153	end if
		1154	inc ecx
		1155	cmp cx,.y3
		1156	jl .loop23
		1157	.loop23_done:
		1158
		1159
		1160	ret 50
		1161
		1162
		1163
		1164
		1165	; xmm0= cby1,cbx1,cz1,cx1
2881	leency	1166	; xmm1= cty1,ctx1,cey1,cex1
		1167	if Ext >= SSE2
		1168	sub esp,8
		1169	shufps xmm1,xmm1,10110001b
		1170	shufps xmm3,xmm3,10110001b
		1171	movlps [esp],xmm1
		1172	else
		1173	push dword .cty1
1245	hidnplayr	1174	push .ctx1
		1175	end if
2881	leency	1176	push dword .cz1
1245	hidnplayr	1177	if Ext>=SSE2
2881	leency	1178	sub esp,8
		1179	movlps [esp],xmm3
		1180	else
		1181	push dword .cty2
1819	yogev_ezra	1182	push .ctx2
		1183	end if
2881	leency	1184	push dword .cz2
1245	hidnplayr	1185	if Ext>=SSE2
2881	leency	1186	sub esp,32
		1187	movhps [esp+24],xmm3
		1188	shufps xmm2,xmm2,10110001b
		1189	movlps [esp+16],xmm2
		1190	movhps [esp+8],xmm1
		1191	shufps xmm0,xmm0,10110001b
		1192	movlps [esp],xmm0 ;================================
		1193
		1194
		1195	push dword .cey2
1245	hidnplayr	1196	push .cex2
		1197	push dword .cby2
1819	yogev_ezra	1198	push .cbx2
		1199	push dword .cey1
1245	hidnplayr	1200	push .cex1
		1201	push dword .cby1
		1202	push .cbx1
		1203	end if
2881	leency	1204
2192	leency	1205
2881	leency	1206	push .z_buff
		1207	push .t_emap
		1208	push .t_bmap
		1209
		1210
1245	hidnplayr	1211
		1212
		1213	sar eax,ROUND
		1214	mov ebx,.cx2
		1215	sar ebx,ROUND
		1216
		1217
		1218
		1219
		1220	;end if
1819	yogev_ezra	1221	ret
1245	hidnplayr	1222	bump_tex_line_z:
		1223	;--------------in: eax - x1
		1224	;-------------- ebx - x2
		1225	;-------------- edi - pointer to screen buffer
		1226	;stack - another parameters :
		1227	.y equ dword [ebp+4]
		1228	.bmap equ dword [ebp+8] ; bump map pointer
2881	leency	1229	.emap equ dword [ebp+12] ; env map pointer
		1230	.z_buff equ dword [ebp+16] ; z buffer
		1231	.tex_map equ dword [ebp+20] ; texture pointer
		1232
1245	hidnplayr	1233
2881	leency	1234	.by1 equ [ebp+28] ; \|
		1235	.ex1 equ [ebp+32] ; \|
		1236	.ey1 equ [ebp+36] ; \|
		1237	.bx2 equ [ebp+40] ; \|
		1238	.by2 equ [ebp+44] ; \|> b. map and e. map coords
		1239	.ex2 equ [ebp+48] ; \|> shifted shl ROUND
		1240	.ey2 equ [ebp+52] ; ---
		1241	.z2 equ [ebp+56]
		1242	.tx2 equ [ebp+60]
		1243	.ty2 equ [ebp+64]
		1244	.z1 equ [ebp+68]
		1245	.tx1 equ [ebp+72]
		1246	.ty1 equ [ebp+76]
		1247
1245	hidnplayr	1248
		1249
2881	leency	1250
1819	yogev_ezra	1251	.x2 equ [ebp-8]
		1252	.dbx equ [ebp-12]
		1253	.dby equ [ebp-16]
		1254	.dex equ [ebp-20]
		1255	.dey equ [ebp-24]
		1256	.dz equ [ebp-28]
		1257	.dtx equ [ebp-32]
		1258	.dty equ [ebp-36]
		1259
1245	hidnplayr	1260
1819	yogev_ezra	1261	.cby equ [ebp-44]
		1262	.cex equ [ebp-48]
		1263	.cey equ [ebp-52]
		1264	.cz equ [ebp-56]
		1265	.czbuff equ [ebp-60]
		1266	.ctx equ [ebp-64]
		1267	.cty equ [ebp-68]
		1268	.c_scr equ [ebp-72]
		1269
1245	hidnplayr	1270
		1271	.temp2 equ ebp-88
		1272	.temp3 equ ebp-76
		1273	.temp4 equ ebp-84
		1274	.temp5 equ ebp-92
		1275
		1276
		1277
		1278
		1279	or ecx,ecx
		1280	jl .bl_end
		1281	cmp ecx,SIZE_Y
		1282	jge .bl_end
		1283
		1284
		1285	jl .bl_ok
		1286	je .bl_end
		1287
		1288
2881	leency	1289
1245	hidnplayr	1290	mov edx,.bx1
		1291	xchg edx,.bx2
		1292	mov .bx1,edx
		1293	mov edx,.by1
		1294	xchg edx,.by2
		1295	mov .by1,edx
		1296
		1297
		1298	xchg edx,.ex2
		1299	mov .ex1,edx
		1300	mov edx,.ey1
		1301	xchg edx,.ey2
		1302	mov .ey1,edx
		1303
		1304
		1305	xchg edx,.tx2
		1306	mov .tx1,edx
		1307	mov edx,.ty1
		1308	xchg edx,.ty2
		1309	mov .ty1,edx
		1310	end if
		1311	if Ext = MMX
1819	yogev_ezra	1312	movq mm0,.bx1
		1313	movq mm1,.bx2
		1314	movq mm2,.ex1
		1315	movq mm3,.ex2
		1316	movq mm4,.tx1
		1317	movq mm5,.tx2
		1318	movq .bx2,mm0
		1319	movq .bx1,mm1
		1320	movq .ex1,mm3
		1321	movq .ex2,mm2
		1322	movq .tx1,mm5
		1323	movq .tx2,mm4
		1324	end if
		1325	if Ext>=SSE
		1326	movups xmm0,.bx1
		1327	movups xmm1,.bx2
		1328	movups .bx1,xmm1
		1329	movups .bx2,xmm0
		1330	movq mm0,.tx1
		1331	movq mm1,.tx2
		1332	movq .tx1,mm1
		1333	movq .tx2,mm0
		1334	end if
		1335	;if Ext>=SSE2
2881	leency	1336	; movaps xmm4,xmm0
		1337	; movaps xmm0,xmm2
		1338	; movaps xmm2,xmm4
		1339	; movaps xmm5,xmm1
		1340	; movaps xmm1,xmm3
		1341	; movaps xmm3,xmm5
		1342	;else
		1343
1245	hidnplayr	1344
2881	leency	1345	mov edx,.z1
1245	hidnplayr	1346	xchg edx,.z2
		1347	mov .z1,edx
		1348	;end if
2881	leency	1349	.bl_ok:
1245	hidnplayr	1350	;if Ext >= SSE2
2881	leency	1351	; shufps xmm0,xmm0,11100001b
		1352	; shufps xmm2,xmm2,11100001b
		1353	; movlps .bx1,xmm0
		1354	; movlps .bx2,xmm2
		1355
		1356
		1357
		1358	; shufps xmm2,xmm2,00011011b
		1359	; movd eax,xmm0
		1360	; movd ebx,xmm2
		1361	; shufps xmm0,xmm0,11000110b
		1362	; shufps xmm2,xmm2,11000110b
		1363	; movd .z1,xmm0
		1364	; movd .z2,xmm2
		1365	; shufps xmm1,xmm1,10110001b
		1366	; shufps xmm3,xmm3,10110001b
		1367	; movlps .ex1,xmm1
		1368	; movlps .ex2,xmm2
		1369	; movhps .tx1,xmm1
		1370	; movhps .tx2,xmm2
		1371
		1372
		1373	; mov edx,.z1
		1374	; xchg edx,.z2
		1375	; mov .z1,edx
		1376
		1377
		1378
		1379
		1380
1245	hidnplayr	1381	push ebx ;store x1, x2
		1382	cmp dword .x1,SIZE_X
1819	yogev_ezra	1383	jge .bl_end
1245	hidnplayr	1384	cmp dword .x2,0
1819	yogev_ezra	1385	jle .bl_end
1245	hidnplayr	1386
		1387
		1388	sub ebx,.x1
		1389
		1390
		1391
		1392
		1393	cvtsi2ss xmm3,ebx ;rcps
		1394	shufps xmm3,xmm3,0
		1395	; float using SSE variant ::-->
1819	yogev_ezra	1396	; movups xmm0,.bx1 ; new
		1397	; movups xmm1,.bx2 ; new
		1398
1245	hidnplayr	1399
1819	yogev_ezra	1400	movlhps xmm0,xmm0
1245	hidnplayr	1401	cvtpi2ps xmm0,.ex1 ;mm2
1819	yogev_ezra	1402	cvtpi2ps xmm1,.bx2 ;mm1
		1403	movlhps xmm1,xmm1
1245	hidnplayr	1404	cvtpi2ps xmm1,.ex2 ;mm3
1819	yogev_ezra	1405	subps xmm1,xmm0
1245	hidnplayr	1406
		1407
		1408
		1409
		1410	; movups .dey,xmm1 ; new
1819	yogev_ezra	1411	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
1245	hidnplayr	1412	movhlps xmm1,xmm1
		1413	cvtps2pi mm1,xmm1
		1414	movq .dey,mm0
1819	yogev_ezra	1415	movq .dby,mm1
		1416
1245	hidnplayr	1417
		1418	movd mm3,.z2
		1419
		1420
1819	yogev_ezra	1421	movlhps xmm0,xmm0
1245	hidnplayr	1422	cvtpi2ps xmm0,mm2
		1423	cvtpi2ps xmm1,.tx2 ;mm1
1819	yogev_ezra	1424	movlhps xmm1,xmm1
1245	hidnplayr	1425	cvtpi2ps xmm1,mm3
		1426	; movups xmm0,,z1 ; new
1819	yogev_ezra	1427	; movups xmm1,.z2 ; new
		1428	subps xmm1,xmm0
1245	hidnplayr	1429
		1430
		1431
		1432
1819	yogev_ezra	1433
		1434
1245	hidnplayr	1435	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		1436	movhlps xmm1,xmm1
		1437	cvtps2pi mm1,xmm1
		1438	movd .dz,mm0
		1439	movq .dty,mm1
1819	yogev_ezra	1440
1245	hidnplayr	1441
		1442
		1443
		1444	sub eax,.bx1
		1445	cdq
		1446	idiv ebx
		1447	push eax
		1448
		1449
		1450	sub eax,.by1
		1451	cdq
		1452	idiv ebx
		1453	push eax
		1454
		1455
		1456	sub eax,.ex1
		1457	cdq
		1458	idiv ebx
		1459	push eax
		1460
		1461
		1462	sub eax,.ey1
		1463	cdq
		1464	idiv ebx
		1465	push eax
		1466
		1467
		1468
		1469	sub eax,.z1
		1470	cdq
		1471	idiv ebx
		1472	push eax
		1473
		1474
		1475	sub eax,.tx1
		1476	cdq
		1477	idiv ebx
		1478	push eax
		1479
		1480
		1481	sub eax,.ty1
		1482	cdq
		1483	idiv ebx
		1484	push eax
		1485
		1486
		1487	cmp dword .x1,0 ; set correctly begin variable
1819	yogev_ezra	1488	jge @f ; CLIPPING ON FUNCTION
1245	hidnplayr	1489	; cutting triangle exceedes screen
		1490	mov ebx,.x1
		1491	neg ebx
		1492
1819	yogev_ezra	1493
		1494
		1495
		1496	; shufps xmm0,xmm0,0
		1497	; movups xmm1,.dey
		1498	; mulps xmm1,xmm0
		1499	; shufps xmm1,xmm1,00011011b
		1500	; movups xmm2,.bx1
		1501	; addps xmm2,xmm1
		1502	; movups .bx1,xmm2
		1503
		1504
		1505	imul ebx ; eax = .dz * abs(.x1)
1245	hidnplayr	1506	add .z1,eax
		1507	mov dword .x1,0
1819	yogev_ezra	1508
1245	hidnplayr	1509
		1510	imul ebx
		1511	add .bx1,eax
		1512
		1513
		1514	imul ebx
		1515	add .by1,eax
		1516
		1517
		1518	imul ebx
		1519	add .ex1,eax
		1520
		1521
		1522	imul ebx
		1523	add .ey1,eax
		1524
		1525
		1526	imul ebx
		1527	add .tx1,eax
		1528
		1529
		1530	imul ebx
		1531	add .ty1,eax
		1532
		1533
		1534	cmp dword .x2,SIZE_X
1819	yogev_ezra	1535	jl @f
1245	hidnplayr	1536	mov dword .x2,SIZE_X
1819	yogev_ezra	1537	@@:
1245	hidnplayr	1538	mov eax,SIZE_X ;calc memory begin in buffers
		1539	mul .y
		1540	add eax,.x1
		1541	lea esi,[4*eax]
		1542	add esi,.z_buff ; z-buffer filled with dd variables
		1543	lea eax,[eax*3]
		1544	add edi,eax
		1545
		1546
		1547
		1548	sub ecx,.x1
		1549	; init current variables
		1550	push dword .bx1 ; current b, e and t shifted shl ROUND .cbx
1819	yogev_ezra	1551	push dword .by1 ; .cby
		1552	push dword .ex1 ; .cex
		1553	push dword .ey1 ; .cey
		1554
1245	hidnplayr	1555
1819	yogev_ezra	1556	push esi ; .czbuff
1245	hidnplayr	1557
		1558
1819	yogev_ezra	1559	push dword .ty1 ; .cty
		1560	push edi ; .c_scr
1245	hidnplayr	1561	if Ext = SSE2
2984	leency	1562	mov eax,TEXTURE_SIZE
		1563	movd xmm1,eax
		1564	shufps xmm1,xmm1,0
		1565	push dword TEX_X
		1566	push dword -TEX_X
		1567	push dword 1
		1568	push dword -1
		1569	movups xmm2,[esp]
		1570	movd xmm3,.bmap
		1571	shufps xmm3,xmm3,0
		1572	end if
		1573
1979	yogev_ezra	1574
1245	hidnplayr	1575	movq mm7,.cty
1819	yogev_ezra	1576	movq mm6,.cby
		1577	movq mm5,.cey
		1578	; movq mm4,.dtyq
1245	hidnplayr	1579	; movq mm3,.dbyq
		1580	end if
		1581
		1582
		1583	; if TEX = SHIFTING ;bump drawing only in shifting mode
		1584	mov esi,.czbuff ; .czbuff current address in buffer
		1585	mov ebx,.cz ; .cz - cur z position
		1586	cmp ebx,dword[esi]
		1587	jge .skip
		1588	if Ext=NON
		1589	mov eax,.cby
		1590	shr eax,ROUND
		1591	mov esi,.cbx
		1592	shr esi,ROUND
		1593	else
		1594	movq mm1,mm6
		1595	psrld mm1,ROUND
		1596	movd eax,mm1
		1597	psrlq mm1,32
		1598	movd esi,mm1
		1599	end if
		1600
		1601
		1602	add esi,eax ;- ; esi - current bump map index
		1603
		1604
2984	leency	1605
		1606
		1607	shufps xmm0,xmm0,0
		1608	paddd xmm0,xmm2
		1609	pand xmm0,xmm1
		1610	paddd xmm0,xmm3
		1611
		1612
		1613	movzx eax,byte[ebx]
		1614	;
1979	yogev_ezra	1615	; shufps xmm0,xmm0,11100001b
		1616	psrldq xmm0,4
2984	leency	1617	movd ebx,xmm0
		1618	movzx ebx,byte[ebx]
		1619	sub eax,ebx
		1620	;
1979	yogev_ezra	1621	; shufps xmm0,xmm0,11111110b
		1622	psrldq xmm0,4
2984	leency	1623	movd ebx,xmm0
		1624	movzx edx, byte [ebx]
		1625	;
1979	yogev_ezra	1626	; shufps xmm0,xmm0,11111111b
		1627	psrldq xmm0,4
2984	leency	1628	movd ebx,xmm0
		1629	movzx ebx, byte [ebx]
		1630	sub edx,ebx
		1631	;
1979	yogev_ezra	1632	else
2984	leency	1633	; mov ebx,esi
		1634	; dec ebx
		1635	lea ebx,[esi-1]
		1636	and ebx,TEXTURE_SIZE
1245	hidnplayr	1637	add ebx,.bmap
		1638	movzx eax,byte [ebx]
		1639
		1640
2984	leency	1641	; inc ebx
		1642	lea ebx,[esi+1]
		1643	and ebx,TEXTURE_SIZE
1245	hidnplayr	1644	add ebx,.bmap
		1645	movzx ebx,byte [ebx]
		1646	sub eax,ebx
		1647
		1648
2984	leency	1649	; sub ebx,TEX_X
		1650	lea ebx,[esi-TEX_X]
		1651	and ebx,TEXTURE_SIZE
1245	hidnplayr	1652	add ebx,.bmap
		1653	movzx edx,byte [ebx]
		1654
		1655
2984	leency	1656	; add ebx,TEX_X
		1657	lea ebx,[esi+TEX_X]
		1658	and ebx,TEXTURE_SIZE
1245	hidnplayr	1659	add ebx,.bmap
		1660	movzx ebx,byte [ebx]
		1661	sub edx,ebx
		1662	end if
2984	leency	1663
1245	hidnplayr	1664
		1665	; edx - vertical sub modificated y coord
		1666	if Ext=NON
		1667	mov ebx,.cex ;.cex - current env map X
		1668	shr ebx,ROUND
		1669	add eax,ebx
		1670
		1671
		1672
		1673	shr ebx,ROUND
		1674	add edx,ebx
		1675
		1676
		1677	movq mm1,mm5 ; mm5 - copy of cur env coords
		1678	psrld mm1,ROUND
		1679	movd ebx,mm1
		1680	psrlq mm1,32
		1681	add eax,ebx
		1682	movd ebx,mm1
		1683	add edx,ebx
		1684	; movq qword[.temp1],mm3
		1685	; add eax,dword [.temp1]
		1686	; add edx,dword [.temp1+4]
		1687	end if
		1688
		1689
		1690	jl .black
		1691	cmp eax,TEX_X
		1692	jg .black
		1693	or edx,edx
		1694	jl .black
		1695	cmp edx,TEX_Y
		1696	jg .black
		1697
		1698
		1699	add edx,eax ; proponuje nie stawiac czarnego pixela tylko
		1700	lea esi,[edx*3] ; niezaburzony.
		1701	add esi,.emap ;
		1702	lodsd
		1703
		1704
		1705	mov edx,.cty
		1706	shr edx,ROUND ; sar
		1707
		1708
		1709	shr edi,ROUND ; sar
		1710	else
		1711	movq mm1,mm7
		1712	psrld mm1,ROUND
		1713	movd edx,mm1
		1714	psrlq mm1,32
		1715	movd edi,mm1
		1716
		1717
		1718
		1719
		1720	add edi,edx
		1721	and edi,TEXTURE_SIZE
		1722	lea esi,[edi*3]
		1723	add esi,.tex_map
		1724
		1725
		1726	mov edx,eax
		1727	lodsd
		1728	push ax
		1729	mul dl
		1730	mov dl,ah
		1731	pop ax
		1732	shr ax,8
		1733	mul dh
		1734	mov al,dl
		1735	mov edi,.c_scr
		1736	stosw
		1737	shr edx,16
		1738	shr eax,16
		1739	mul dl
		1740	shr ax,8
		1741	stosb
		1742	else
		1743	movd mm0,eax
		1744	pxor mm1,mm1
		1745	punpcklbw mm0,mm1
		1746	movd mm2,[esi]
		1747	punpcklbw mm2,mm1
		1748	pmullw mm0,mm2
		1749	psrlw mm0,8
		1750	packuswb mm0,mm1
		1751	mov edi,.c_scr
		1752	movd [edi],mm0
		1753
		1754
		1755
		1756
		1757	@@:
		1758	.black:
		1759	xor eax,eax
		1760	mov edi,.c_scr
		1761	stosd
		1762	.actual_zbuff:
		1763	mov eax,.cz
		1764	mov edi,.czbuff
		1765	stosd
		1766
		1767
		1768	add dword .czbuff,4
1819	yogev_ezra	1769	add dword .c_scr,3
		1770
1245	hidnplayr	1771
		1772	mov eax,.dbx
		1773	add .cbx,eax
		1774	mov ebx,.dby
		1775	add .cby,ebx
		1776
		1777
		1778	add .cex,edx
		1779	mov eax,.dey
		1780	add .cey,eax
		1781
		1782
		1783	add .ctx,ebx
		1784	mov edx,.dty
		1785	add .cty,edx
		1786
		1787
		1788	paddd mm7,.dty
1819	yogev_ezra	1789	paddd mm6,.dby
		1790	paddd mm5,.dey
		1791	end if
1245	hidnplayr	1792	mov eax,.dz
		1793	add .cz,eax
		1794
		1795
		1796	jnz .draw
		1797
		1798
		1799	mov esp,ebp
		1800	ret 76
		1801	;Ext = MMX
		1802
		1803
		1804	; movq mm5, qword[.temp1] ;-
		1805	; paddd mm5, qword[.temp5] ; .temp5 == low dword = TEX_X, high dword = -TEX_X
		1806	; pand mm5, qword[.temp3] ; .temp3 == low = high dword = TEX_SIZE
		1807	; paddd mm5, qword[.temp4] ; .temp4 == low = high dword = offset .bmap
		1808	; movd ebx,mm5
		1809	; psrlq mm5,32
		1810	; end if
		1811

Subversion Repositories Kolibri OS

(root)/programs/demos/3DS/BUMP_TEX.INC – Rev 2984