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	;TEX_SHIFT EQU 9
		8
		9
		10	;TEXTURE_SIZE EQU (TEX_X * TEX_Y)-1
		11	;Ext = SSE
		12	;SSE = 3
		13	;MMX = 1
		14	;NON = 0
		15	;use32
		16	;------- Big thanks to Majuma (www.majuma.xt.pl) for absolutely great---
		17	;------- DOS 13h mode demos --------------------------------------------
		18	;------- Procedure draws triangle with two overlapped textures, I use --
		19	;--------Catmull Z-buffer algorithm- (Z coordinate interpolation)-------
		20	;--------I calc texture pixel by this way: col1*col2/256 ---------------
		21	two_tex_triangle_z:
		22	;------------------in - eax - x1 shl 16 + y1 -----------
		23	;---------------------- ebx - x2 shl 16 + y2 -----------
		24	;---------------------- ecx - x3 shl 16 + y3 -----------
		25	;---------------------- edx - pointer to b. texture-----
		26	;---------------------- esi - pointer to e. texture-----
		27	;---------------------- edi - pointer to screen buffer--
		28	;---------------------- stack : b. tex coordinates------
		29	;---------------------- e. tex coordinates------
		30	;---------------------- Z position coordinates--
		31	;---------------------- pointer io Z buffer-----
		32	;-- Z-buffer - filled with coordinates as dword --------
		33	;-- (Z coor. as word) shl CATMULL_SHIFT ----------------
		34	.b_x1 equ ebp+4 ; procedure don't save registers !!!
		35	.b_y1 equ ebp+6 ; each coordinate as word
		36	.b_x2 equ ebp+8
		37	.b_y2 equ ebp+10 ; b - first texture
		38	.b_x3 equ ebp+12
		39	.b_y3 equ ebp+14 ; e - second texture
		40	.e_x1 equ ebp+16
		41	.e_y1 equ ebp+18
		42	.e_x2 equ ebp+20
		43	.e_y2 equ ebp+22
		44	.e_x3 equ ebp+24
		45	.e_y3 equ ebp+26
		46	.z1 equ word[ebp+28]
		47	.z2 equ word[ebp+30]
		48	.z3 equ word[ebp+32]
		49	.z_buff equ dword[ebp+34] ; pointer to Z-buffer
		50
		51
		52
		53	.t_emap equ dword[ebp-8] ; pointer to e. texture
		54	.x1 equ word[ebp-10]
		55	.y1 equ word[ebp-12]
		56	.x2 equ word[ebp-14]
		57	.y2 equ word[ebp-16]
		58	.x3 equ word[ebp-18]
		59	.y3 equ word[ebp-20]
		60
		61
		62	.dbx12 equ dword[ebp-28]
		63	.dby12 equ dword[ebp-32]
		64	.dby12q equ [ebp-32]
		65	.dex12 equ dword[ebp-36]
		66	.dey12 equ dword[ebp-40]
		67	.dey12q equ [ebp-40]
		68	.dz12 equ dword[ebp-44]
		69
		70
		71	.dbx13 equ dword[ebp-52]
		72	.dby13 equ dword[ebp-56]
		73	.dby13q equ [ebp-56]
		74	.dex13 equ dword[ebp-60]
		75	.dey13 equ dword[ebp-64]
		76	.dey13q equ [ebp-64]
		77	.dz13 equ dword[ebp-68]
		78
		79
		80	.dbx23 equ dword[ebp-76]
		81	.dby23 equ dword[ebp-80]
		82	.dby23q equ [ebp-80]
		83	.dex23 equ dword[ebp-84]
		84	.dey23 equ dword[ebp-88]
		85	.dey23q equ [ebp-88]
		86	.dz23 equ dword[ebp-92]
		87
		88
		89	.cx2 equ dword[ebp-100]
		90	.cbx1 equ dword[ebp-104]
		91	.cby1 equ [ebp-108]
		92	.cex1 equ dword[ebp-112]
		93	.cey1 equ [ebp-116]
		94	.cbx2 equ dword[ebp-120]
		95	.cby2 equ [ebp-124]
		96	.cex2 equ dword[ebp-128]
		97	.cey2 equ [ebp-132]
		98
		99
		100	.cz2 equ dword[ebp-140]
		101
		102
		103	emms
		104	else
		105	cld
		106	end if
		107	mov ebp,esp
		108	push edx esi ; store bump map
		109	; push esi ; store e. map
		110	; sub esp,120
		111	.sort3: ; sort triangle coordinates...
		112	cmp ax,bx
		113	jle .sort1
		114	xchg eax,ebx
		115	mov edx,dword[.b_x1]
		116	xchg edx,dword[.b_x2]
		117	mov dword[.b_x1],edx
		118	mov edx,dword[.e_x1]
		119	xchg edx,dword[.e_x2]
		120	mov dword[.e_x1],edx
		121	mov dx,.z1
		122	xchg dx,.z2
		123	mov .z1,dx
		124	.sort1:
		125	cmp bx,cx
		126	jle .sort2
		127	xchg ebx,ecx
		128	mov edx,dword[.b_x2]
		129	xchg edx,dword[.b_x3]
		130	mov dword[.b_x2],edx
		131	mov edx,dword[.e_x2]
		132	xchg edx,dword[.e_x3]
		133	mov dword[.e_x2],edx
		134	mov dx,.z2
		135	xchg dx,.z3
		136	mov .z2,dx
		137	jmp .sort3
		138	.sort2:
		139	push eax ebx ecx ; store triangle coords in variables
		140	; push ebx
		141	; push ecx
		142
		143
		144	and edx,ebx ; if all of them are negative a sign flag is raised
		145	and edx,ecx
		146	and edx,eax
		147	test edx,80008000h ; Check both X&Y at once
		148	jne .loop23_done
		149	; mov edx,eax ; eax,ebx,ecx are ORd together into edx which means that
		150	; or edx,ebx ; if any one of them is negative a sign flag is raised
		151	; or edx,ecx
		152	; test edx,80000000h ; Check only X
		153	; jne .loop23_done
		154
		155
		156	; jg .loop23_done
		157	; cmp .x2,SIZE_X ; This can be optimized with effort
		158	; jg .loop23_done
		159	; cmp .x3,SIZE_X
		160	; jg .loop23_done ; {
		161
		162
		163
		164	sub bx,.y1
		165	jnz .bt_dx12_make
		166	mov ecx,6
		167	xor edx,edx
		168	@@:
		169	push edx ;dword 0
		170	loop @b
		171	jmp .bt_dx12_done
		172	.bt_dx12_make:
		173	mov ax,.x2
		174	sub ax,.x1
		175	cwde
		176	movsx ebx,bx
		177	shl eax,ROUND
		178	cdq
		179	idiv ebx
		180	; mov .dx12,eax
		181	push eax
		182
		183
		184
		185
		186	cvtsi2ss xmm3,ebx ;rcps
		187	; mov eax,255
		188	cvtsi2ss xmm4,[i255d] ;eax
		189	divss xmm3,xmm4
		190	rcpss xmm3,xmm3
		191	; mulss xmm3,xmm4
		192	shufps xmm3,xmm3,0
		193
		194
		195	movd mm1,[.b_x2]
		196	movd mm2,[.e_x1]
		197	movd mm3,[.e_x2]
		198	; psubsw mm3,mm2
		199	; psubsw mm1,mm0
		200	pxor mm4,mm4
		201	punpcklwd mm0,mm4
		202	punpcklwd mm1,mm4
		203	punpcklwd mm2,mm4
		204	punpcklwd mm3,mm4
		205	; pslld mm0,ROUND
		206	; pslld mm1,ROUND
		207	; pslld mm2,ROUND
		208	; pslld mm3,ROUND
		209	cvtpi2ps xmm0,mm0
		210	movlhps xmm0,xmm0
		211	cvtpi2ps xmm0,mm2
		212	cvtpi2ps xmm1,mm1
		213	movlhps xmm1,xmm1
		214	cvtpi2ps xmm1,mm3
		215	subps xmm1,xmm0
		216
		217
		218	; movq mm5,mm1
		219	; movq mm6,mm1
		220	; pcmpeqb mm5,mm4
		221	; psubd mm1,mm0
		222	; psubd mm3,mm2
		223
		224
		225	; movq mm1,[.e_x1] ; ex1 ey1 ex2 ey2
		226	; pxor
		227	; punpcklhd mm0,mm1 ; lwd ;
		228	; psubw mm1,mm0 ; mm1, mm0
		229	; pxor mm2,mm2
		230	; pmovmaskb eax,mm1
		231	; and eax,10101010b
		232	; pcmpgtw mm2,mm1
		233	; punpcklwd mm1,mm2
		234	; psllw mm0,ROUND
		235	; psllw mm1,ROUND
		236	; movq mm2,mm0
		237	; psrlq mm0,32
		238
		239
		240	; movlhps xmm0,xmm0
		241	; cvtpi2ps xmm0,mm3
		242	; divps xmm1,xmm3
		243	mulps xmm1,xmm3
		244	shufps xmm1,xmm1,10110001b
		245	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		246	movhlps xmm1,xmm1
		247	cvtps2pi mm1,xmm1
		248	movq .dey12q,mm0
		249	movq .dby12q,mm1
		250
		251
		252	; psrlq mm0,32
		253	; movd .dey12,mm0
		254	; movhlps xmm1,xmm1
		255	; cvtps2pi mm0,xmm1
		256	; movd .dbx12,mm0
		257	; psrlq mm0,32
		258	; movd .dby12,mm0
		259
		260
		261	mov ax,word[.b_x2]
		262	sub ax,word[.b_x1]
		263	cwde
		264	shl eax,ROUND
		265	cdq
		266	idiv ebx
		267	; mov .dbx12,eax
		268	push eax
		269
		270
		271	sub ax,word[.b_y1]
		272	cwde
		273	shl eax,ROUND
		274	cdq
		275	idiv ebx
		276	; mov .dby12,eax
		277	push eax
		278
		279
		280	; mov ebx,.dby12
		281	; int3
		282
		283
		284	sub ax,word[.e_x1]
		285	cwde
		286	shl eax,ROUND
		287	cdq
		288	idiv ebx
		289	; mov .dex12,eax
		290	push eax
		291
		292
		293	sub ax,word[.e_y1]
		294	cwde
		295	shl eax,ROUND
		296	cdq
		297	idiv ebx
		298	; mov .dey12,eax
		299	push eax
		300
		301
		302	mov ax,.z2
		303	sub ax,.z1
		304	cwde
		305	shl eax,CATMULL_SHIFT
		306	cdq
		307	idiv ebx
		308	push eax
		309	.bt_dx12_done:
		310
		311
		312	sub bx,.y1
		313	jnz .bt_dx13_make
		314	mov ecx,6
		315	xor edx,edx
		316	@@:
		317	push edx ;dword 0
		318	loop @b
		319	jmp .bt_dx13_done
		320	.bt_dx13_make:
		321	mov ax,.x3
		322	sub ax,.x1
		323	cwde
		324	movsx ebx,bx
		325	shl eax,ROUND
		326	cdq
		327	idiv ebx
		328	; mov .dx13,eax
		329	push eax
		330
		331
		332
		333
		334	; mov eax,255
		335	cvtsi2ss xmm4,[i255d]
		336	divss xmm3,xmm4
		337	rcpss xmm3,xmm3
		338	; mulss xmm3,xmm4
		339	shufps xmm3,xmm3,0
		340	sub esp,16
		341
		342
		343	movd mm1,[.b_x3]
		344	movd mm2,[.e_x1]
		345	movd mm3,[.e_x3]
		346
		347
		348	punpcklwd mm0,mm4
		349	punpcklwd mm1,mm4
		350	punpcklwd mm2,mm4
		351	punpcklwd mm3,mm4
		352
		353
		354	movlhps xmm0,xmm0
		355	cvtpi2ps xmm0,mm2
		356	cvtpi2ps xmm1,mm1
		357	movlhps xmm1,xmm1
		358	cvtpi2ps xmm1,mm3
		359	subps xmm1,xmm0
		360
		361
		362	mulps xmm1,xmm3
		363	shufps xmm1,xmm1,10110001b
		364	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		365	movhlps xmm1,xmm1
		366	cvtps2pi mm1,xmm1
		367	movq .dey13q,mm0
		368	movq .dby13q,mm1
		369
		370
		371
		372
		373	sub ax,word[.b_x1]
		374	cwde
		375	shl eax,ROUND
		376	cdq
		377	idiv ebx
		378	; mov .dbx13,eax
		379	push eax
		380
		381
		382	sub ax,word[.b_y1]
		383	cwde
		384	shl eax,ROUND
		385	cdq
		386	idiv ebx
		387	; mov .dby13,eax
		388	push eax
		389
		390
		391	sub ax,word[.e_x1]
		392	cwde
		393	shl eax,ROUND
		394	cdq
		395	idiv ebx
		396	; mov .dex13,eax
		397	push eax
		398
		399
		400	sub ax,word[.e_y1]
		401	cwde
		402	shl eax,ROUND
		403	cdq
		404	idiv ebx
		405	; mov .dey13,eax
		406	push eax
		407
		408
		409
		410
		411	sub ax,.z1
		412	cwde
		413	shl eax,CATMULL_SHIFT
		414	cdq
		415	idiv ebx
		416	; mov .dz13,eax
		417	push eax
		418	.bt_dx13_done:
		419
		420
		421	sub bx,.y2
		422	jnz .bt_dx23_make
		423	mov ecx,6
		424	xor edx,edx
		425	@@:
		426	push edx ;dword 0
		427	loop @b
		428	jmp .bt_dx23_done
		429	.bt_dx23_make:
		430	mov ax,.x3
		431	sub ax,.x2
		432	cwde
		433	movsx ebx,bx
		434	shl eax,ROUND
		435	cdq
		436	idiv ebx
		437	; mov .dx23,eax
		438	push eax
		439
		440
		441
		442
		443	; mov eax,255
		444	cvtsi2ss xmm4,[i255d] ;eax
		445	divss xmm3,xmm4
		446	shufps xmm3,xmm3,0
		447	sub esp,16
		448
		449
		450	movd mm1,[.b_x3]
		451	movd mm2,[.e_x2]
		452	movd mm3,[.e_x3]
		453
		454
		455	punpcklwd mm0,mm4
		456	punpcklwd mm1,mm4
		457	punpcklwd mm2,mm4
		458	punpcklwd mm3,mm4
		459
		460
		461	movlhps xmm0,xmm0
		462	cvtpi2ps xmm0,mm2
		463	cvtpi2ps xmm1,mm1
		464	movlhps xmm1,xmm1
		465	cvtpi2ps xmm1,mm3
		466	subps xmm1,xmm0
		467
		468
		469	shufps xmm1,xmm1,10110001b
		470	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		471	movhlps xmm1,xmm1
		472	cvtps2pi mm1,xmm1
		473	movq .dey23q,mm0
		474	movq .dby23q,mm1
		475
		476
		477
		478
		479	sub ax,word[.b_x2]
		480	cwde
		481	shl eax,ROUND
		482	cdq
		483	idiv ebx
		484	; mov .dbx23,eax
		485	push eax
		486
		487
		488	sub ax,word[.b_y2]
		489	cwde
		490	shl eax,ROUND
		491	cdq
		492	idiv ebx
		493	; mov .dby23,eax
		494	push eax
		495
		496
		497	sub ax,word[.e_x2]
		498	cwde
		499	shl eax,ROUND
		500	cdq
		501	idiv ebx
		502	; mov .dex23,eax
		503	push eax
		504
		505
		506	sub ax,word[.e_y2]
		507	cwde
		508	shl eax,ROUND
		509	cdq
		510	idiv ebx
		511	; mov .dey23,eax
		512	push eax
		513	end if
		514	mov ax,.z3
		515	sub ax,.z2
		516	cwde
		517	shl eax,CATMULL_SHIFT
		518	cdq
		519	idiv ebx
		520	; mov .dz23,eax
		521	push eax
		522	; sub esp,40
		523	.bt_dx23_done:
		524	movsx eax,.x1
		525	shl eax,ROUND
		526	; mov .cx1,eax
		527	; mov .cx2,eax
		528	push eax eax
		529	; push eax
		530
		531
		532	shl eax,ROUND
		533	mov .cbx1,eax
		534	mov .cbx2,eax
		535	; push eax eax
		536	; push eax
		537
		538
		539	shl eax,ROUND
		540	mov .cby1,eax
		541	mov .cby2,eax
		542	; push eax eax
		543	; push eax
		544
		545
		546	shl eax,ROUND
		547	mov .cex1,eax
		548	mov .cex2,eax
		549	; push eax eax
		550	;push eax
		551
		552
		553	shl eax,ROUND
		554	mov .cey1,eax
		555	mov .cey2,eax
		556	sub esp,32
		557	; push eax eax
		558	;push eax
		559
		560
		561	shl eax,CATMULL_SHIFT
		562	; mov .cz1,eax
		563	; mov .cz2,eax
		564	push eax eax
		565	;push eax
		566
		567
		568	cmp cx,.y2
		569	jge .loop12_done
		570	.loop12:
		571	call .call_line
		572
		573
		574	add .cx1,eax
		575	mov ebx,.dx12
		576	add .cx2,ebx
		577	if Ext>= SSE2
		578	movups xmm0,.cey1
		579	movups xmm1,.cey2
		580	movups xmm2,.dey12q
		581	movups xmm3,.dey13q
		582	paddd xmm0,xmm3
		583	paddd xmm1,xmm2
		584	movups .cey1,xmm0
		585	movups .cey2,xmm1
		586	else if (Ext = MMX) \| (Ext=SSE)
		587	movq mm0,.cby2 ; with this optimization object
		588	movq mm1,.cby1 ; looks bit annoying
		589	movq mm2,.cey2
		590	movq mm3,.cey1
		591	paddd mm0,.dby12q
		592	paddd mm1,.dby13q
		593	paddd mm2,.dey12q
		594	paddd mm3,.dey13q
		595	movq .cby2,mm0
		596	movq .cby1,mm1
		597	movq .cey1,mm3
		598	movq .cey2,mm2
		599	else
		600	mov edx,.dbx13
		601	add .cbx1,edx
		602	mov eax,.dbx12
		603	add .cbx2,eax
		604	mov ebx,.dby13
		605	add .cby1,ebx
		606	mov edx,.dby12
		607	add .cby2,edx
		608
		609
		610	add .cex1,eax
		611	mov ebx,.dex12
		612	add .cex2,ebx
		613	mov edx,.dey13
		614	add .cey1,edx
		615	mov eax,.dey12
		616	add .cey2,eax
		617
		618
		619	mov ebx,.dz13
		620	add .cz1,ebx
		621	mov edx,.dz12
		622	add .cz2,edx
		623
		624
		625	cmp cx,.y2
		626	jl .loop12
		627	.loop12_done:
		628
		629
		630	cmp cx,.y3
		631	jge .loop23_done
		632
		633
		634	shl eax,CATMULL_SHIFT
		635	mov .cz2,eax
		636
		637
		638	shl eax,ROUND
		639	mov .cx2,eax
		640
		641
		642	shl eax,ROUND
		643	mov .cbx2,eax
		644
		645
		646	shl eax,ROUND
		647	mov .cby2,eax
		648
		649
		650	shl eax,ROUND
		651	mov .cex2,eax
		652
		653
		654	shl eax,ROUND
		655	mov .cey2,eax
		656
		657
		658	call .call_line
		659	;if Ext = NON
		660	mov eax,.dx13
		661	add .cx1,eax
		662	mov ebx,.dx23
		663	add .cx2,ebx
		664	if Ext>= SSE2
		665	movups xmm0,.cey1
		666	movups xmm1,.cey2
		667	movups xmm2,.dey23q
		668	movups xmm3,.dey13q
		669	paddd xmm0,xmm3
		670	paddd xmm1,xmm2
		671	movups .cey1,xmm0
		672	movups .cey2,xmm1
		673	else if (Ext = MMX) \| ( Ext = SSE)
		674	movq mm0,.cby2 ; with this mmx optimization object looks bit
		675	movq mm1,.cby1 ; annoying
		676	movq mm2,.cey2
		677	movq mm3,.cey1
		678	paddd mm0,.dby23q
		679	paddd mm1,.dby13q
		680	paddd mm2,.dey23q
		681	paddd mm3,.dey13q
		682	movq .cby2,mm0
		683	movq .cby1,mm1
		684	movq .cey2,mm2
		685	movq .cey1,mm3
		686
		687
		688	mov edx,.dbx13
		689	add .cbx1,edx
		690	mov eax,.dbx23
		691	add .cbx2,eax
		692	mov ebx,.dby13
		693	add .cby1,ebx
		694	mov edx,.dby23
		695	add .cby2,edx
		696
		697
		698	add .cex1,eax
		699	mov ebx,.dex23
		700	add .cex2,ebx
		701	mov edx,.dey13
		702	add .cey1,edx
		703	mov eax,.dey23
		704	add .cey2,eax
		705	end if
		706
		707
		708	add .cz1,ebx
		709	mov edx,.dz23
		710	add .cz2,edx
		711	;else
		712	; movq mm0,.db13q
		713	; movq mm1,.cbx1q
		714
		715
		716	cmp cx,.y3
		717	jl .loop23
		718	.loop23_done:
		719
		720
		721	ret 34
		722
		723
		724
		725
		726
		727
		728	push .cz2
		729	push .z_buff
		730	push .t_bmap
		731	push .t_emap
		732	push dword .cey2
		733	push .cex2
		734	push dword .cey1
		735	push .cex1
		736	push dword .cby2
		737	push .cbx2
		738	push dword .cby1
		739	push .cbx1
		740	push ecx
		741
		742
		743	sar eax,ROUND
		744	mov ebx,.cx2
		745	sar ebx,ROUND
		746
		747
		748
		749
		750	ret
		751	two_tex_line_z:
		752	;--------------in: eax - x1
		753	;-------------- ebx - x2
		754	;-------------- edi - pointer to screen buffer
		755	;stack - another parameters :
		756	.y equ dword [ebp+4]
		757	.bx1 equ [ebp+8] ; ---
		758	.by1 equ [ebp+12] ; \|
		759	.bx2 equ [ebp+16] ; \|
		760	.by2 equ [ebp+20] ; \|> b. texture and e. texture coords
		761	.ex1 equ [ebp+24] ; \|> shifted shl ROUND
		762	.ey1 equ [ebp+28] ; \|
		763	.ex2 equ [ebp+32] ; \|
		764	.ey2 equ [ebp+36] ; ---
		765	.emap equ [ebp+40] ; b texture offset
		766	.bmap equ [ebp+44] ; e texture offset
		767	.z_buff equ dword [ebp+48]
		768	.z2 equ dword [ebp+52] ; -- \|> z coords shifted
		769	.z1 equ dword [ebp+56] ; -- shl CATMULL_SHIFT
		770
		771
		772	.x2 equ dword [ebp-8]
		773	.dbx equ [ebp-12]
		774	.dex equ [ebp-16]
		775	.dby equ [ebp-20]
		776	.dey equ [ebp-24]
		777	.dz equ dword [ebp-28]
		778	.cbx equ [ebp-32]
		779	.cex equ [ebp-36]
		780	.cby equ [ebp-40]
		781	.cey equ [ebp-44]
		782	.cz equ dword [ebp-48]
		783	.czbuff equ dword [ebp-52]
		784
		785
		786
		787
		788	or ecx,ecx
		789	jl .bl_end
		790	mov dx,word[size_y_var]
		791	dec dx
		792	cmp cx,dx ;word[size_y_var] ;SIZE_Y
		793	jge .bl_end
		794
		795
		796	jl @f
		797	je .bl_end
		798
		799
		800	if Ext=NON
		801	mov edx,.bx1
		802	xchg edx,.bx2
		803	mov .bx1,edx
		804	mov edx,.by1
		805	xchg edx,.by2
		806	mov .by1,edx
		807
		808
		809	xchg edx,.ex2
		810	mov .ex1,edx
		811	mov edx,.ey1
		812	xchg edx,.ey2
		813	mov .ey1,edx
		814	else
		815	movq mm0,.bx1
		816	movq mm1,.ex1
		817	movq mm2,.bx2
		818	movq mm3,.ex2
		819	movq .bx2,mm0
		820	movq .ex2,mm1
		821	movq .bx1,mm2
		822	movq .ex1,mm3
		823	end if
		824	mov edx,.z1
		825	xchg edx,.z2
		826	mov .z1,edx
		827	@@:
		828	push eax ebx ;store x1, x2
		829	mov ebx,.x1
		830	movzx edx,word[size_x_var]
		831	dec edx
		832	cmp ebx,edx
		833	; cmp bx,word[size_x_var] ;SIZE_X
		834	jg .bl_end
		835	cmp .x2,0
		836	jle .bl_end
		837
		838
		839	sub ebx,.x1
		840
		841
		842
		843
		844	cvtsi2ss xmm3,ebx ;rcps
		845	shufps xmm3,xmm3,0
		846
		847
		848	; movq mm1,.bx2q
		849	; movq mm2,.ex1q
		850	; movq mm3,.ex2q
		851	; psubd mm1,mm0
		852	; psubd mm3,mm2
		853	; cvtpi2ps xmm1,mm1
		854	; movlhps xmm1,xmm1
		855	; cvtpi2ps xmm1,mm3
		856
		857
		858	movlhps xmm0,xmm0
		859	cvtpi2ps xmm0,.ex1 ;mm2 ; ex1; ey1
		860	cvtpi2ps xmm1,.bx2 ;mm1 ; bx2; by2
		861	movlhps xmm1,xmm1
		862	cvtpi2ps xmm1,.ex2 ;mm3 ; ex2; ey2
		863	subps xmm1,xmm0
		864	; hi lo
		865	divps xmm1,xmm3 ; xmm1 -> dby; dbx; dey; dex
		866
		867
		868	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		869	movhlps xmm1,xmm1
		870	cvtps2pi mm1,xmm1
		871	movq .dex,mm0 ; hi - lo -> dbx, dex
		872	movq .dey,mm1 ; hi - lo -> dby, dey
		873
		874
		875
		876
		877	sub eax,.bx1
		878	cdq
		879	idiv ebx
		880	push eax
		881
		882
		883	sub eax,.ex1
		884	cdq
		885	idiv ebx
		886	push eax
		887
		888
		889	sub eax,.by1
		890	cdq
		891	idiv ebx
		892	push eax
		893
		894
		895	sub eax,.ey1
		896	cdq
		897	idiv ebx
		898	push eax
		899
		900
		901
		902
		903	sub eax,.z1
		904	cdq
		905	idiv ebx
		906	push eax
		907
		908
		909	jge @f ; CLIPPING ON FUNCTION
		910	; cutting triangle exceedes screen
		911	mov ebx,.x1
		912	neg ebx
		913	imul ebx ; eax = .dz * abs(.x1)
		914	add .z1,eax
		915	mov .x1,0
		916
		917
		918	imul ebx
		919	add .bx1,eax
		920
		921
		922	imul ebx
		923	add .by1,eax
		924
		925
		926	imul ebx
		927	add .ex1,eax
		928
		929
		930	imul ebx
		931	add .ey1,eax
		932	@@:
		933	movzx eax,word[size_x_var] ;SIZE_X ;word[size_x_var]
		934	mov ebx,.x2
		935	cmp eax,ebx
		936	jg @f
		937	mov .x2,eax
		938	@@:
		939	; movd mm0,eax
		940	; movd mm1,.x2
		941	; pminsw mm0,mm1
		942	; movd .x2,mm0
		943	; cmp .x2,SIZE_X ;eax \|
		944	; jl @f \|> this dont work idk cause
		945	; mov .x2,SIZE_X ;eax \|
		946	@@:
		947	; movzx eax,word[size_x_var] ;calc memory begin in buffers
		948	mov ebx,.y
		949	mul ebx
		950	mov ebx,.x1
		951	add eax,ebx
		952	mov ebx,eax
		953	lea eax,[eax*3]
		954	add edi,eax ; edi - screen
		955	mov esi,.z_buff ; z-buffer filled with dd variables
		956	shl ebx,2
		957	add esi,ebx ; esi - Z buffer
		958
		959
		960	sub ecx,.x1
		961	; init current variables
		962	push dword .bx1 ;.by1 .ex1 .ey1 .z1 esi
		963	push dword .ex1
		964	push dword .by1
		965	push dword .ey1
		966
		967
		968	push esi
		969
		970
		971	pxor mm0,mm0
		972	movq mm3,.cex ; hi - lo -> cbx; cex
		973	movq mm4,.cey ; hi - lo -> cby; cey
		974	; movq mm5,mm3
		975	; movq mm6,mm4
		976	; psrad mm5,ROUND
		977	; psrad mm6,ROUND
		978	; movq .ceyq,mm5
		979	; movq .cbyq,mm6
		980	mov edx,.czbuff
		981	else
		982	cld
		983	end if
		984	.draw:
		985	; if TEX = SHIFTING ;bump drawing only in shifting mode
		986	if Ext=NON
		987	mov esi,.czbuff ; .czbuff current address in buffer
		988	mov ebx,.cz ; .cz - cur z position
		989	cmp ebx,dword[esi]
		990	else
		991	mov ebx,.cz
		992	cmp ebx,dword[edx]
		993	end if
		994	jge .skip
		995
		996
		997	mov eax,.cby
		998	mov esi,.cbx
		999	sar eax,ROUND
		1000	sar esi,ROUND
		1001	shl eax,TEX_SHIFT ;-
		1002	add esi,eax
		1003	lea esi,[esi*3] ;- ; esi - current b. texture addres
		1004	add esi,.bmap
		1005
		1006
		1007	mov eax,.cey ;.cey - current env map y
		1008	sar ebx,ROUND
		1009	sar eax,ROUND
		1010
		1011
		1012	add ebx,eax
		1013	lea ebx,[ebx*3]
		1014	add ebx,.emap
		1015
		1016
		1017
		1018	movq mm5,mm4 ;.cey
		1019	psrad mm5,ROUND
		1020	pslld mm5,TEX_SHIFT
		1021	movq mm6,mm3 ;.cex
		1022	psrad mm6,ROUND
		1023	paddd mm5,mm6
		1024	movq mm6,mm5
		1025	paddd mm5,mm5
		1026	paddd mm5,mm6
		1027	paddd mm5,.emap
		1028	movd esi,mm5
		1029	psrlq mm5,32
		1030	movd ebx,mm5
		1031	end if
		1032	if Ext>=MMX
		1033	movd mm1,[esi]
		1034	movd mm2,[ebx]
		1035	punpcklbw mm1,mm0
		1036	punpcklbw mm2,mm0
		1037	pmullw mm1,mm2
		1038	psrlw mm1,8
		1039	packuswb mm1,mm0
		1040	movd [edi],mm1
		1041	mov ebx,.cz
		1042	mov dword[edx],ebx
		1043	else
		1044	cld ; esi - tex e.
		1045	lodsb ; ebx - tex b.
		1046	mov dl,[ebx]
		1047	mul dl
		1048	shr ax,8
		1049	stosb
		1050	inc ebx
		1051	lodsb
		1052	mov dl,[ebx]
		1053	mul dl
		1054	shr ax,8
		1055	stosb
		1056	inc ebx
		1057	lodsb
		1058	mov dl,[ebx]
		1059	mul dl
		1060	shr ax,8
		1061	stosb
		1062	mov ebx,.cz
		1063	mov esi,.czbuff
		1064	mov dword[esi],ebx
		1065	jmp .no_skip
		1066	end if
		1067	.skip:
		1068	add edi,3
		1069
		1070
		1071	.no_skip:
		1072	add .czbuff,4
		1073	mov eax,.dbx
		1074	add .cbx,eax
		1075	mov eax,.dby
		1076	add .cby,eax
		1077	mov eax,.dex
		1078	add .cex,eax
		1079	mov eax,.dey
		1080	add .cey,eax
		1081	else
		1082	add edx,4
		1083	paddd mm3,.dex
		1084	paddd mm4,.dey
		1085	; movq mm5,mm3
		1086	; movq mm6,mm4
		1087	; psrad mm5,ROUND
		1088	; psrad mm6,ROUND
		1089	; movq .cex,mm3
		1090	; movq .cey,mm4
		1091	end if
		1092	mov eax,.dz
		1093	add .cz,eax
		1094	if Ext = NON
		1095	dec ecx
		1096	jnz .draw
		1097	else
		1098	loop .draw
		1099	end if
		1100
		1101
		1102	mov esp,ebp
		1103	ret 56
		1104

Subversion Repositories Kolibri OS

(root)/programs/demos/view3ds/two_tex.inc – Rev 9237