WebSVN – Kolibri OS – Blame – /programs/demos/view3ds/two_tex.inc

Rev	Author	Line No.	Line
9237	leency	1
		2	;SIZE_Y equ 350
		3	;ROUND equ 8
		4	;TEX_X equ 512
		5	;TEX_Y equ 512
		6	;TEXTURE_SIZE EQU (512*512)-1
		7
9512	IgorA	8
9237	leency	9
		10
9512	IgorA	11	TEXTURE_SIZE EQU (TEX_X * TEX_Y)-1
		12
		13
		14	;SSE2 = 3
		15	;SSE = 2
		16	;MMX = 1
9237	leency	17	;NON = 0
		18	;Ext = NON
9512	IgorA	19	;use32
9237	leency	20	;------- Big thanks to Majuma (www.majuma.xt.pl) for absolutely great---
		21	;------- DOS 13h mode demos --------------------------------------------
		22	;------- Procedure draws triangle with two overlapped textures, I use --
		23	;--------Catmull Z-buffer algorithm- (Z coordinate interpolation)-------
		24	;--------I calc texture pixel by this way: col1*col2/256 ---------------
		25	two_tex_triangle_z:
		26	;------------------in - eax - x1 shl 16 + y1 -----------
		27	;---------------------- ebx - x2 shl 16 + y2 -----------
		28	;---------------------- ecx - x3 shl 16 + y3 -----------
		29	;---------------------- edx - pointer to b. texture-----
		30	;---------------------- esi - pointer to e. texture-----
		31	;---------------------- edi - pointer to screen buffer--
		32	;---------------------- stack : b. tex coordinates------
		33	;---------------------- e. tex coordinates------
		34	;---------------------- Z position coordinates--
		35	;---------------------- pointer io Z buffer-----
		36	;-- Z-buffer - filled with coordinates as dword --------
		37	;-- (Z coor. as word) shl CATMULL_SHIFT ----------------
		38	.e_x1 equ ebp+4 ; procedure don't save registers !!!
9512	IgorA	39	.e_y1 equ ebp+6 ; each coordinate as word
		40	.b_x1 equ ebp+8
		41	.b_y1 equ ebp+10
		42	.e_x2 equ ebp+12
		43	.e_y2 equ ebp+14
		44	.b_x2 equ ebp+16
		45	.b_y2 equ ebp+18 ; b - first texture
		46	.e_x3 equ ebp+20
		47	.e_y3 equ ebp+22 ; e - second texture
		48	.b_x3 equ ebp+24
		49	.b_y3 equ ebp+26
		50	.z1 equ word[ebp+28]
		51	.z2 equ word[ebp+30]
		52	.z3 equ word[ebp+32]
		53	.z_buff equ dword[ebp+34] ; pointer to Z-buffer
		54
9237	leency	55
		56
9512	IgorA	57	.t_emap equ dword[ebp-8] ; pointer to e. texture
		58	.x1 equ word[ebp-10]
		59	.y1 equ word[ebp-12]
		60	.x2 equ word[ebp-14]
		61	.y2 equ word[ebp-16]
		62	.x3 equ word[ebp-18]
		63	.y3 equ word[ebp-20]
		64
9237	leency	65
		66	.dbx12 equ [ebp-28]
9512	IgorA	67	.dby12 equ dword[ebp-32]
9237	leency	68	.dby12q equ [ebp-32]
9512	IgorA	69	.dex12 equ dword[ebp-36]
9237	leency	70	.dey12 equ dword[ebp-40]
		71	.dey12q equ [ebp-40]
9512	IgorA	72	.dz12 equ [ebp-44]
		73
9237	leency	74
		75	.dbx13 equ [ebp-52]
9512	IgorA	76	.dby13 equ dword[ebp-56]
9237	leency	77	.dby13q equ [ebp-56]
9512	IgorA	78	.dex13 equ dword[ebp-60]
9237	leency	79	.dey13 equ dword[ebp-64]
		80	.dey13q equ [ebp-64]
9512	IgorA	81	.dz13 equ [ebp-68]
		82
9237	leency	83
		84	.dbx23 equ [ebp-76]
9512	IgorA	85	.dby23 equ dword[ebp-80]
9237	leency	86	.dby23q equ [ebp-80]
9512	IgorA	87	.dex23 equ dword[ebp-84]
9237	leency	88	.dey23 equ dword[ebp-88]
		89	.dey23q equ [ebp-88]
9512	IgorA	90	.dz23 equ [ebp-92]
		91
9237	leency	92
		93	.cx2 equ dword[ebp-100]
		94	.cbx1 equ dword[ebp-104]
		95	.cby1 equ [ebp-108]
		96	.cex1 equ dword[ebp-112]
		97	.cey1 equ [ebp-116]
		98	.cbx2 equ dword[ebp-120]
		99	.cby2 equ [ebp-124]
		100	.cex2 equ dword[ebp-128]
		101	.cey2 equ [ebp-132]
		102
		103
		104	.cz2 equ dword[ebp-140]
		105
		106
		107	emms
		108	else
		109	cld
		110	end if
		111	mov ebp,esp
		112	push edx esi ; store bump map
9512	IgorA	113	; push esi ; store e. map
9237	leency	114	; sub esp,120
		115	.sort3: ; sort triangle coordinates...
9512	IgorA	116	cmp ax,bx
9237	leency	117	jle .sort1
		118	xchg eax,ebx
		119	mov edx,dword[.b_x1]
		120	xchg edx,dword[.b_x2]
		121	mov dword[.b_x1],edx
		122	mov edx,dword[.e_x1]
		123	xchg edx,dword[.e_x2]
		124	mov dword[.e_x1],edx
		125	mov dx,.z1
		126	xchg dx,.z2
		127	mov .z1,dx
		128	.sort1:
		129	cmp bx,cx
9512	IgorA	130	jle .sort2
		131	xchg ebx,ecx
		132	mov edx,dword[.b_x2]
		133	xchg edx,dword[.b_x3]
		134	mov dword[.b_x2],edx
		135	mov edx,dword[.e_x2]
		136	xchg edx,dword[.e_x3]
		137	mov dword[.e_x2],edx
		138	mov dx,.z2
9237	leency	139	xchg dx,.z3
		140	mov .z2,dx
		141	jmp .sort3
9512	IgorA	142	.sort2:
9237	leency	143	push eax ebx ecx ; store triangle coords in variables
9512	IgorA	144
9237	leency	145
9512	IgorA	146	and edx,ebx ; if all of them are negative a sign flag is raised
		147	and edx,ecx
		148	and edx,eax
		149	test edx,80008000h ; Check both X&Y at once
		150	jne .loop23_done
		151	; mov edx,eax ; eax,ebx,ecx are ORd together into edx which means that
9237	leency	152	; or edx,ebx ; if any one of them is negative a sign flag is raised
		153	; or edx,ecx
		154	; test edx,80000000h ; Check only X
		155	; jne .loop23_done
		156
		157
		158	; jg .loop23_done
		159	; cmp .x2,SIZE_X ; This can be optimized with effort
		160	; jg .loop23_done
		161	; cmp .x3,SIZE_X
		162	; jg .loop23_done ; {
		163	; sub esp,18*4
9512	IgorA	164	; pxor xmm7,xmm7
		165	mov bx,.y2 ; calc delta 12
		166	sub bx,.y1
		167	jnz .bt_dx12_make
		168	mov ecx,6
		169	xor edx,edx
		170	@@:
9237	leency	171	push edx ;dword 0
9512	IgorA	172	loop @b
		173	jmp .bt_dx12_done
		174	.bt_dx12_make:
9237	leency	175
9512	IgorA	176
		177	mov eax,1 shl 15
		178	cdq
9237	leency	179	idiv ebx
9512	IgorA	180	mov ebx,eax
		181
9237	leency	182
9512	IgorA	183	sub ax,.x1
		184	cwde
		185	imul ebx
		186	sar eax,15 - ROUND
		187	push eax
		188	; mov .dx12,eax
		189
9237	leency	190
9512	IgorA	191	sub esp,4*4
		192	movd xmm0,ebx
		193	pshuflw xmm0,xmm0,0
		194	movq xmm1,[.e_x1]
		195	movq xmm2,[.e_x2]
		196	psubw xmm2,xmm1
		197	movdqa xmm3,xmm2
		198	pmullw xmm2,xmm0
		199	pmulhw xmm3,xmm0
		200	punpcklwd xmm2,xmm3
		201	psrad xmm2,15 - ROUND
		202	pshufd xmm2,xmm2,10110001b
		203	movdqu .dey12q,xmm2
		204	else
		205
9237	leency	206
9512	IgorA	207	sub ax,word[.b_x1]
		208	cwde
		209	imul ebx
		210	sar eax,15 - ROUND
		211	; mov .dbx23,eax
		212	push eax
		213
9237	leency	214
9512	IgorA	215	sub ax,word[.b_y1]
		216	cwde
		217	imul ebx
		218	sar eax,15 - ROUND
		219	; mov .dbx23,eax
		220	push eax
		221
9237	leency	222
9512	IgorA	223	sub ax,word[.e_x1]
		224	cwde
9237	leency	225	imul ebx
9512	IgorA	226	sar eax,15 - ROUND
		227	; mov .dbx23,eax
		228	push eax
		229
9237	leency	230
9512	IgorA	231	sub ax,word[.e_y1]
		232	cwde
9237	leency	233	imul ebx
9512	IgorA	234	sar eax,15 - ROUND
		235	; mov .dbx23,eax
		236	push eax
		237	end if
		238
9237	leency	239
		240
9512	IgorA	241	sub ax,.z1
		242	cwde
		243	imul ebx
		244	sar eax,15 - ROUND
		245	push eax
		246	; mov .dz12,eax
		247
9237	leency	248
		249
		250
		251
9512	IgorA	252	sub bx,.y1
		253	jnz .bt_dx13_make
		254	mov ecx,6
		255	xor edx,edx
		256	@@:
9237	leency	257	push edx ;dword 0
9512	IgorA	258	loop @b
		259	; movq .dbx13,xmm7
		260	; movdqu .dz13,xmm7
		261	jmp .bt_dx13_done
		262	.bt_dx13_make:
9237	leency	263	; sub esp,6*4
9512	IgorA	264
		265
		266	mov eax,1 shl 15
		267	cdq
9237	leency	268	idiv ebx
9512	IgorA	269	; push eax
		270	mov ebx,eax
		271
9237	leency	272
		273
9512	IgorA	274	sub ax,.x1
		275	cwde
		276	imul ebx
		277	sar eax,15 - ROUND
		278	push eax
		279	; mov .dx13,eax
		280	if 1
		281	sub esp,4*4
		282	movd xmm0,ebx
		283	pshuflw xmm0,xmm0,0
		284	movq xmm1,[.e_x1]
		285	movq xmm2,[.e_x3]
		286	psubw xmm2,xmm1
		287	movdqa xmm3,xmm2
		288	pmullw xmm2,xmm0
		289	pmulhw xmm3,xmm0
		290	punpcklwd xmm2,xmm3
		291	psrad xmm2,15 - ROUND
		292	pshufd xmm2,xmm2,10110001b
		293	movdqu .dey13q,xmm2
		294	; punpcklwd xmm4,xmm5
		295	; psrad xmm4,15 - ROUND
		296	; movq .tex_dx12,xmm4
		297	else
9237	leency	298	mov ax,word[.b_x3]
9512	IgorA	299	sub ax,word[.b_x1]
		300	cwde
9237	leency	301	; shl eax,ROUND
9512	IgorA	302	; cdq
		303	imul ebx
		304	sar eax,15 - ROUND
		305	; mov .dbx23,eax
		306	push eax
		307
9237	leency	308
9512	IgorA	309	sub ax,word[.b_y1]
		310	cwde
9237	leency	311	imul ebx
9512	IgorA	312	sar eax,15 - ROUND
		313	; mov .dbx23,eax
		314	push eax
		315
9237	leency	316
9512	IgorA	317	sub ax,word[.e_x1]
		318	cwde
9237	leency	319	imul ebx
9512	IgorA	320	sar eax,15 - ROUND
		321	; mov .dbx23,eax
		322	push eax
		323
9237	leency	324
9512	IgorA	325	sub ax,word[.e_y1]
		326	cwde
9237	leency	327	imul ebx
9512	IgorA	328	sar eax,15 - ROUND
		329	; mov .dbx23,eax
		330	push eax
		331	end if
		332	mov ax,.z3
		333	sub ax,.z1
		334	cwde
		335	imul ebx
		336	sar eax,15 - ROUND
		337	push eax
		338	; mov .dz13,eax
		339
9237	leency	340
		341
		342
9512	IgorA	343
		344	sub bx,.y2
		345	jnz .bt_dx23_make
		346	mov ecx,6
		347	xor edx,edx
		348	@@:
9237	leency	349	push edx ;dword 0
9512	IgorA	350	loop @b
		351	; movq .dbx23,xmm7
		352	; movdqu .dz23,xmm7
		353
		354
		355	.bt_dx23_make:
9237	leency	356
		357
		358
9512	IgorA	359	mov eax,1 shl 15
		360	cdq
		361	idiv ebx
		362	; push eax
		363	mov ebx,eax
		364
9237	leency	365
		366
9512	IgorA	367	sub ax,.x2
		368	cwde
		369	imul ebx
		370	sar eax,15 - ROUND
		371	push eax
		372	; mov .dx23,eax
		373
9237	leency	374
9512	IgorA	375	sub esp,4*4
		376	movd xmm0,ebx
		377	pshuflw xmm0,xmm0,0
		378	movq xmm1,[.e_x2]
		379	movq xmm2,[.e_x3]
		380	psubw xmm2,xmm1
		381	movdqa xmm3,xmm2
		382	pmullw xmm2,xmm0
		383	pmulhw xmm3,xmm0
		384	punpcklwd xmm2,xmm3
		385	psrad xmm2,15 - ROUND
		386	pshufd xmm2,xmm2,10110001b
		387	movdqu .dey23q,xmm2
		388	else
9237	leency	389
		390
9512	IgorA	391	sub ax,word[.b_x2]
		392	cwde
9237	leency	393	; shl eax,ROUND
9512	IgorA	394	; cdq
		395	imul ebx
		396	sar eax,15 - ROUND
		397	; mov .dbx23,eax
9237	leency	398	push eax
9512	IgorA	399
9237	leency	400
9512	IgorA	401	sub ax,word[.b_y2]
		402	cwde
9237	leency	403	imul ebx
9512	IgorA	404	sar eax,15 - ROUND
		405	; mov .dbx23,eax
		406	push eax
		407
9237	leency	408
9512	IgorA	409	sub ax,word[.e_x2]
		410	cwde
9237	leency	411	imul ebx
9512	IgorA	412	sar eax,15 - ROUND
		413	; mov .dbx23,eax
		414	push eax
		415
9237	leency	416
9512	IgorA	417	sub ax,word[.e_y2]
		418	cwde
9237	leency	419	imul ebx
9512	IgorA	420	sar eax,15 - ROUND
		421	; mov .dbx23,eax
		422	push eax
		423	end if
9237	leency	424
9512	IgorA	425
		426
		427	sub ax,.z2
		428	cwde
		429	imul ebx
		430	sar eax,15 - ROUND
		431	push eax
		432	; mov .dz23,eax
		433
		434
		435
		436
		437
		438
		439
		440
9237	leency	441
9512	IgorA	442
		443	shl eax,ROUND
		444	; mov .cx1,eax
9237	leency	445	; mov .cx2,eax
		446	push eax eax
9512	IgorA	447	; push eax
9237	leency	448	sub esp,32
9512	IgorA	449	; push eax eax
		450	;push eax
		451
9237	leency	452
9512	IgorA	453	shl eax,CATMULL_SHIFT
		454	; mov .cz1,eax
		455	; mov .cz2,eax
		456	push eax eax
		457	;push eax
		458
		459
		460	shl eax,ROUND
		461	mov .cbx1,eax
		462	mov .cbx2,eax
		463	; push eax eax
9237	leency	464	; push eax
		465
		466
9512	IgorA	467	shl eax,ROUND
		468	mov .cby1,eax
		469	mov .cby2,eax
		470	; push eax eax
9237	leency	471	; push eax
		472
		473
9512	IgorA	474	shl eax,ROUND
		475	mov .cex1,eax
		476	mov .cex2,eax
		477	; push eax eax
9237	leency	478	;push eax
		479
		480
9512	IgorA	481	shl eax,ROUND
		482	mov .cey1,eax
		483	mov .cey2,eax
		484
9237	leency	485
9512	IgorA	486	cmp cx,.y2
		487	jge .loop12_done
		488	.loop12:
9237	leency	489	call .call_line
9512	IgorA	490
9237	leency	491
9512	IgorA	492	add .cx1,eax
		493	mov ebx,.dx12
		494	add .cx2,ebx
		495	if Ext>= SSE2
9237	leency	496	movups xmm0,.cey1
9512	IgorA	497	movups xmm1,.cey2
		498	movups xmm2,.dey12q
		499	movups xmm3,.dey13q
		500	paddd xmm0,xmm3
		501	paddd xmm1,xmm2
		502	movups .cey1,xmm0
		503	movups .cey2,xmm1
		504	else if (Ext = MMX) \| (Ext=SSE)
9237	leency	505	movq mm0,.cby2 ; with this optimization object
9512	IgorA	506	movq mm1,.cby1 ; looks bit annoying
		507	movq mm2,.cey2
		508	movq mm3,.cey1
		509	paddd mm0,.dby12q
		510	paddd mm1,.dby13q
		511	paddd mm2,.dey12q
		512	paddd mm3,.dey13q
		513	movq .cby2,mm0
		514	movq .cby1,mm1
		515	movq .cey1,mm3
		516	movq .cey2,mm2
		517	else
9237	leency	518	mov edx,.dbx13
9512	IgorA	519	add .cbx1,edx
		520	mov eax,.dbx12
		521	add .cbx2,eax
		522	mov ebx,.dby13
		523	add .cby1,ebx
		524	mov edx,.dby12
		525	add .cby2,edx
		526
9237	leency	527
9512	IgorA	528	add .cex1,eax
		529	mov ebx,.dex12
		530	add .cex2,ebx
		531	mov edx,.dey13
		532	add .cey1,edx
		533	mov eax,.dey12
		534	add .cey2,eax
		535
9237	leency	536
		537	mov ebx,.dz13
9512	IgorA	538	add .cz1,ebx
		539	mov edx,.dz12
		540	add .cz2,edx
		541
9237	leency	542
9512	IgorA	543	cmp cx,.y2
		544	jl .loop12
		545	.loop12_done:
9237	leency	546
		547
9512	IgorA	548	cmp cx,.y3
		549	jge .loop23_done
		550
9237	leency	551
9512	IgorA	552	shl eax,CATMULL_SHIFT
		553	mov .cz2,eax
		554
9237	leency	555
9512	IgorA	556	shl eax,ROUND
		557	mov .cx2,eax
		558
9237	leency	559
9512	IgorA	560	shl eax,ROUND
		561	mov .cbx2,eax
		562
9237	leency	563
9512	IgorA	564	shl eax,ROUND
		565	mov .cby2,eax
		566
9237	leency	567
9512	IgorA	568	shl eax,ROUND
		569	mov .cex2,eax
		570
9237	leency	571
9512	IgorA	572	shl eax,ROUND
		573	mov .cey2,eax
		574
9237	leency	575
		576	call .call_line
9512	IgorA	577	;if Ext = NON
9237	leency	578	mov eax,.dx13
9512	IgorA	579	add .cx1,eax
		580	mov ebx,.dx23
		581	add .cx2,ebx
		582	if Ext>= SSE2
9237	leency	583	movups xmm0,.cey1
9512	IgorA	584	movups xmm1,.cey2
		585	movups xmm2,.dey23q
		586	movups xmm3,.dey13q
		587	paddd xmm0,xmm3
		588	paddd xmm1,xmm2
		589	movups .cey1,xmm0
		590	movups .cey2,xmm1
		591	else if (Ext = MMX) \| ( Ext = SSE)
9237	leency	592	movq mm0,.cby2 ; with this mmx optimization object looks bit
9512	IgorA	593	movq mm1,.cby1 ; annoying
		594	movq mm2,.cey2
		595	movq mm3,.cey1
		596	paddd mm0,.dby23q
		597	paddd mm1,.dby13q
		598	paddd mm2,.dey23q
		599	paddd mm3,.dey13q
		600	movq .cby2,mm0
		601	movq .cby1,mm1
		602	movq .cey2,mm2
		603	movq .cey1,mm3
		604
9237	leency	605
		606	mov edx,.dbx13
9512	IgorA	607	add .cbx1,edx
		608	mov eax,.dbx23
		609	add .cbx2,eax
		610	mov ebx,.dby13
		611	add .cby1,ebx
		612	mov edx,.dby23
		613	add .cby2,edx
		614
9237	leency	615
9512	IgorA	616	add .cex1,eax
		617	mov ebx,.dex23
		618	add .cex2,ebx
		619	mov edx,.dey13
		620	add .cey1,edx
		621	mov eax,.dey23
		622	add .cey2,eax
		623	end if
9237	leency	624
		625
9512	IgorA	626	add .cz1,ebx
		627	mov edx,.dz23
		628	add .cz2,edx
		629	;else
9237	leency	630	; movq mm0,.db13q
		631	; movq mm1,.cbx1q
		632
		633
9512	IgorA	634	cmp cx,.y3
		635	jl .loop23
		636	.loop23_done:
9237	leency	637
		638
9512	IgorA	639	ret 34
9237	leency	640
		641
		642
		643
		644
		645
9512	IgorA	646	push .cz2
		647	push .z_buff
		648	push .t_bmap
		649	push .t_emap
		650	push dword .cey2
		651	push .cex2
		652	push dword .cey1
		653	push .cex1
		654	push dword .cby2
		655	push .cbx2
		656	push dword .cby1
		657	push .cbx1
		658	push ecx
		659
9237	leency	660
9512	IgorA	661	sar eax,ROUND
		662	mov ebx,.cx2
		663	sar ebx,ROUND
		664
9237	leency	665
9512	IgorA	666
9237	leency	667
		668	ret
		669	two_tex_line_z:
		670	;--------------in: eax - x1
		671	;-------------- ebx - x2
		672	;-------------- edi - pointer to screen buffer
		673	;stack - another parameters :
		674	.y equ dword [ebp+4]
9512	IgorA	675	.bx1 equ [ebp+8] ; ---
		676	.by1 equ [ebp+12] ; \|
		677	.bx2 equ [ebp+16] ; \|
		678	.by2 equ [ebp+20] ; \|> b. texture and e. texture coords
		679	.ex1 equ [ebp+24] ; \|> shifted shl ROUND
		680	.ey1 equ [ebp+28] ; \|
		681	.ex2 equ [ebp+32] ; \|
		682	.ey2 equ [ebp+36] ; ---
		683	.emap equ [ebp+40] ; b texture offset
		684	.bmap equ [ebp+44] ; e texture offset
		685	.z_buff equ dword [ebp+48]
9237	leency	686	.z2 equ dword [ebp+52] ; -- \|> z coords shifted
9512	IgorA	687	.z1 equ dword [ebp+56] ; -- shl CATMULL_SHIFT
		688
9237	leency	689
9512	IgorA	690	.x2 equ dword [ebp-8]
		691	.dbx equ [ebp-12]
		692	.dex equ [ebp-16]
		693	.dby equ [ebp-20]
		694	.dey equ [ebp-24]
		695	.dz equ dword [ebp-28]
		696	.cbx equ [ebp-32]
		697	.cex equ [ebp-36]
		698	.cby equ [ebp-40]
		699	.cey equ [ebp-44]
		700	.cz equ dword [ebp-48]
		701	.czbuff equ dword [ebp-52]
9237	leency	702
		703
9512	IgorA	704
9237	leency	705
9512	IgorA	706	or ecx,ecx
		707	jl .bl_end
		708	mov dx,word[size_y_var]
		709	dec dx
		710	cmp cx,dx ;word[size_y_var] ;SIZE_Y
		711	jge .bl_end
		712
9237	leency	713
9512	IgorA	714	jl @f
		715	je .bl_end
		716
9237	leency	717
9512	IgorA	718	if Ext=NON
9237	leency	719	mov edx,.bx1
9512	IgorA	720	xchg edx,.bx2
		721	mov .bx1,edx
		722	mov edx,.by1
		723	xchg edx,.by2
		724	mov .by1,edx
		725
9237	leency	726
9512	IgorA	727	xchg edx,.ex2
		728	mov .ex1,edx
		729	mov edx,.ey1
		730	xchg edx,.ey2
		731	mov .ey1,edx
		732	else
9237	leency	733	movq mm0,.bx1
9512	IgorA	734	movq mm1,.ex1
		735	movq mm2,.bx2
		736	movq mm3,.ex2
		737	movq .bx2,mm0
		738	movq .ex2,mm1
		739	movq .bx1,mm2
		740	movq .ex1,mm3
		741	end if
9237	leency	742	mov edx,.z1
9512	IgorA	743	xchg edx,.z2
		744	mov .z1,edx
		745	@@:
9237	leency	746	push eax ebx ;store x1, x2
9512	IgorA	747	mov ebx,.x1
		748	movzx edx,word[size_x_var]
		749	dec edx
		750	cmp ebx,edx
		751	; cmp bx,word[size_x_var] ;SIZE_X
9237	leency	752	jg .bl_end
9512	IgorA	753	cmp .x2,0
		754	jle .bl_end
		755
9237	leency	756
9512	IgorA	757	sub ebx,.x1
		758
9237	leency	759
		760
		761
9512	IgorA	762	cvtsi2ss xmm3,ebx ;rcps
		763	shufps xmm3,xmm3,0
		764
9237	leency	765
		766	; movq mm1,.bx2q
		767	; movq mm2,.ex1q
		768	; movq mm3,.ex2q
		769	; psubd mm1,mm0
		770	; psubd mm3,mm2
		771	; cvtpi2ps xmm1,mm1
		772	; movlhps xmm1,xmm1
		773	; cvtpi2ps xmm1,mm3
		774
		775
		776	movlhps xmm0,xmm0
9512	IgorA	777	cvtpi2ps xmm0,.ex1 ;mm2 ; ex1; ey1
9237	leency	778	cvtpi2ps xmm1,.bx2 ;mm1 ; bx2; by2
		779	movlhps xmm1,xmm1
9512	IgorA	780	cvtpi2ps xmm1,.ex2 ;mm3 ; ex2; ey2
9237	leency	781	subps xmm1,xmm0
9512	IgorA	782	; hi lo
		783	divps xmm1,xmm3 ; xmm1 -> dby; dbx; dey; dex
		784
9237	leency	785
9512	IgorA	786	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		787	movhlps xmm1,xmm1
		788	cvtps2pi mm1,xmm1
9237	leency	789	movq .dex,mm0 ; hi - lo -> dbx, dex
9512	IgorA	790	movq .dey,mm1 ; hi - lo -> dby, dey
		791
9237	leency	792
		793
		794
		795
		796
		797
		798
9512	IgorA	799
		800	sub eax,.bx1
		801	cdq
		802	idiv ebx
		803	push eax
		804
		805
		806	sub eax,.ex1
		807	cdq
		808	idiv ebx
		809	push eax
		810
		811
		812	sub eax,.by1
		813	cdq
		814	idiv ebx
		815	push eax
		816
		817
		818	sub eax,.ey1
		819	cdq
		820	idiv ebx
		821	push eax
		822
		823
9237	leency	824
		825
9512	IgorA	826	sub eax,.z1
		827	cdq
		828	idiv ebx
		829	push eax
		830
9237	leency	831
9512	IgorA	832	jge @f ; CLIPPING ON FUNCTION
		833	; cutting triangle exceedes screen
		834	mov ebx,.x1
		835	neg ebx
		836	imul ebx ; eax = .dz * abs(.x1)
		837	add .z1,eax
		838	mov .x1,0
		839
9237	leency	840
9512	IgorA	841	imul ebx
		842	add .bx1,eax
		843
9237	leency	844
9512	IgorA	845	imul ebx
		846	add .by1,eax
		847
9237	leency	848
9512	IgorA	849	imul ebx
		850	add .ex1,eax
		851
9237	leency	852
9512	IgorA	853	imul ebx
		854	add .ey1,eax
		855	@@:
9237	leency	856	movzx eax,word[size_x_var] ;SIZE_X ;word[size_x_var]
9512	IgorA	857	mov ebx,.x2
		858	cmp eax,ebx
		859	jg @f
		860	mov .x2,eax
		861	@@:
9237	leency	862	; movd mm0,eax
		863	; movd mm1,.x2
		864	; pminsw mm0,mm1
		865	; movd .x2,mm0
		866	; cmp .x2,SIZE_X ;eax \|
		867	; jl @f \|> this dont work idk cause
		868	; mov .x2,SIZE_X ;eax \|
		869	@@:
		870	; movzx eax,word[size_x_var] ;calc memory begin in buffers
		871	mov ebx,.y
9512	IgorA	872	mul ebx
		873	mov ebx,.x1
		874	add eax,ebx
		875	mov ebx,eax
		876	lea eax,[eax*3]
		877	add edi,eax ; edi - screen
		878	mov esi,.z_buff ; z-buffer filled with dd variables
		879	shl ebx,2
		880	add esi,ebx ; esi - Z buffer
		881
9237	leency	882
9512	IgorA	883	sub ecx,.x1
		884	; init current variables
		885	push dword .bx1 ;.by1 .ex1 .ey1 .z1 esi
		886	push dword .ex1
		887	push dword .by1
		888	push dword .ey1
		889
9237	leency	890
9512	IgorA	891	push esi
		892
9237	leency	893
		894	pxor mm0,mm0
		895	movq mm3,.cex ; hi - lo -> cbx; cex
		896	movq mm4,.cey ; hi - lo -> cby; cey
		897	; movq mm5,mm3
		898	; movq mm6,mm4
		899	; psrad mm5,ROUND
		900	; psrad mm6,ROUND
		901	; movq .ceyq,mm5
		902	; movq .cbyq,mm6
		903	mov edx,.czbuff
		904	else
		905	cld
		906	end if
		907	.draw:
		908	; if TEX = SHIFTING ;bump drawing only in shifting mode
		909	if Ext=NON
		910	mov esi,.czbuff ; .czbuff current address in buffer
9512	IgorA	911	mov ebx,.cz ; .cz - cur z position
		912	cmp ebx,dword[esi]
		913	else
9237	leency	914	mov ebx,.cz
9512	IgorA	915	cmp ebx,dword[edx]
		916	end if
9237	leency	917	jge .skip
9512	IgorA	918
9237	leency	919
		920	mov eax,.cby
9512	IgorA	921	mov esi,.cbx
		922	sar eax,ROUND
		923	sar esi,ROUND
		924	shl eax,TEX_SHIFT ;-
		925	add esi,eax
		926	lea esi,[esi*3] ;- ; esi - current b. texture addres
		927	add esi,.bmap
		928
9237	leency	929
9512	IgorA	930	mov eax,.cey ;.cey - current env map y
		931	sar ebx,ROUND
		932	sar eax,ROUND
		933
9237	leency	934
9512	IgorA	935	add ebx,eax
		936	lea ebx,[ebx*3]
		937	add ebx,.emap
		938
9237	leency	939
		940
		941	movq mm5,mm4 ;.cey
9512	IgorA	942	psrad mm5,ROUND
		943	pslld mm5,TEX_SHIFT
		944	movq mm6,mm3 ;.cex
		945	psrad mm6,ROUND
		946	paddd mm5,mm6
		947	movq mm6,mm5
		948	paddd mm5,mm5
		949	paddd mm5,mm6
		950	paddd mm5,.emap
		951	movd esi,mm5
		952	psrlq mm5,32
		953	movd ebx,mm5
		954	end if
9237	leency	955	if Ext>=MMX
		956	movd mm1,[esi]
9512	IgorA	957	movd mm2,[ebx]
		958	punpcklbw mm1,mm0
		959	punpcklbw mm2,mm0
		960	pmullw mm1,mm2
		961	psrlw mm1,8
		962	packuswb mm1,mm0
		963	movd [edi],mm1
		964	mov ebx,.cz
		965	mov dword[edx],ebx
		966	else
9237	leency	967	cld ; esi - tex e.
9512	IgorA	968	lodsb ; ebx - tex b.
		969	mov dl,[ebx]
		970	mul dl
		971	shr ax,8
		972	stosb
		973	inc ebx
		974	lodsb
		975	mov dl,[ebx]
		976	mul dl
		977	shr ax,8
		978	stosb
		979	inc ebx
		980	lodsb
		981	mov dl,[ebx]
		982	mul dl
		983	shr ax,8
		984	stosb
		985	mov ebx,.cz
		986	mov esi,.czbuff
		987	mov dword[esi],ebx
		988	jmp .no_skip
		989	end if
9237	leency	990	.skip:
		991	add edi,3
9512	IgorA	992
9237	leency	993
		994	.no_skip:
		995	add .czbuff,4
9512	IgorA	996	mov eax,.dbx
		997	add .cbx,eax
		998	mov eax,.dby
		999	add .cby,eax
		1000	mov eax,.dex
		1001	add .cex,eax
		1002	mov eax,.dey
		1003	add .cey,eax
		1004	else
9237	leency	1005	add edx,4
9512	IgorA	1006	paddd mm3,.dex
		1007	paddd mm4,.dey
		1008	; movq mm5,mm3
9237	leency	1009	; movq mm6,mm4
		1010	; psrad mm5,ROUND
		1011	; psrad mm6,ROUND
		1012	; movq .cex,mm3
		1013	; movq .cey,mm4
		1014	end if
		1015	mov eax,.dz
9512	IgorA	1016	add .cz,eax
		1017	if Ext = NON
9237	leency	1018	dec ecx
9512	IgorA	1019	jnz .draw
		1020	else
9237	leency	1021	loop .draw
9512	IgorA	1022	end if
9237	leency	1023
		1024
		1025	mov esp,ebp
9512	IgorA	1026	ret 56
9237	leency	1027	;Ext = SSE2
9512	IgorA	1028	;the_zero:
		1029	;size_y_var:
		1030	;size_x_var:
		1031

Subversion Repositories Kolibri OS

(root)/programs/demos/view3ds/two_tex.inc @ 8666 – Rev 9512