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

Rev	Author	Line No.	Line
1245	hidnplayr	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 !!!
2192	leency	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
1245	hidnplayr	51
		52
2192	leency	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
1245	hidnplayr	61
		62	.dbx12 equ dword[ebp-28]
		63	.dby12 equ dword[ebp-32]
		64	.dby12q equ [ebp-32]
2192	leency	65	.dex12 equ dword[ebp-36]
1245	hidnplayr	66	.dey12 equ dword[ebp-40]
		67	.dey12q equ [ebp-40]
2192	leency	68	.dz12 equ dword[ebp-44]
1245	hidnplayr	69
		70
		71	.dbx13 equ dword[ebp-52]
		72	.dby13 equ dword[ebp-56]
		73	.dby13q equ [ebp-56]
2192	leency	74	.dex13 equ dword[ebp-60]
1245	hidnplayr	75	.dey13 equ dword[ebp-64]
		76	.dey13q equ [ebp-64]
2192	leency	77	.dz13 equ dword[ebp-68]
1245	hidnplayr	78
		79
		80	.dbx23 equ dword[ebp-76]
		81	.dby23 equ dword[ebp-80]
		82	.dby23q equ [ebp-80]
2192	leency	83	.dex23 equ dword[ebp-84]
1245	hidnplayr	84	.dey23 equ dword[ebp-88]
		85	.dey23q equ [ebp-88]
2192	leency	86	.dz23 equ dword[ebp-92]
1245	hidnplayr	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]
2192	leency	93	.cey1 equ [ebp-116]
		94	.cbx2 equ dword[ebp-120]
		95	.cby2 equ [ebp-124]
		96	.cex2 equ dword[ebp-128]
1245	hidnplayr	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
2192	leency	109	; push esi ; store e. map
1245	hidnplayr	110	; sub esp,120
		111	.sort3: ; sort triangle coordinates...
2192	leency	112	cmp ax,bx
1245	hidnplayr	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
2192	leency	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
1245	hidnplayr	135	xchg dx,.z3
		136	mov .z2,dx
		137	jmp .sort3
2192	leency	138	.sort2:
1245	hidnplayr	139	push eax ebx ecx ; store triangle coords in variables
2192	leency	140	; push ebx
1245	hidnplayr	141	; push ecx
		142
		143
2192	leency	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
1245	hidnplayr	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
2192	leency	164	sub bx,.y1
		165	jnz .bt_dx12_make
		166	mov ecx,6
		167	xor edx,edx
		168	@@:
1245	hidnplayr	169	push edx ;dword 0
2192	leency	170	loop @b
		171	jmp .bt_dx12_done
		172	.bt_dx12_make:
1245	hidnplayr	173	mov ax,.x2
2192	leency	174	sub ax,.x1
		175	cwde
1245	hidnplayr	176	movsx ebx,bx
2192	leency	177	shl eax,ROUND
		178	cdq
1245	hidnplayr	179	idiv ebx
2192	leency	180	; mov .dx12,eax
1245	hidnplayr	181	push eax
2192	leency	182
1245	hidnplayr	183
		184
		185
2192	leency	186	cvtsi2ss xmm3,ebx ;rcps
		187	; mov eax,255
1245	hidnplayr	188	cvtsi2ss xmm4,[i255d] ;eax
		189	divss xmm3,xmm4
2192	leency	190	rcpss xmm3,xmm3
		191	; mulss xmm3,xmm4
1245	hidnplayr	192	shufps xmm3,xmm3,0
2192	leency	193
1245	hidnplayr	194
2192	leency	195	movd mm1,[.b_x2]
		196	movd mm2,[.e_x1]
		197	movd mm3,[.e_x2]
		198	; psubsw mm3,mm2
1245	hidnplayr	199	; psubsw mm1,mm0
		200	pxor mm4,mm4
2192	leency	201	punpcklwd mm0,mm4
1245	hidnplayr	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
2192	leency	211	cvtpi2ps xmm0,mm2
1245	hidnplayr	212	cvtpi2ps xmm1,mm1
		213	movlhps xmm1,xmm1
2192	leency	214	cvtpi2ps xmm1,mm3
1245	hidnplayr	215	subps xmm1,xmm0
2192	leency	216
1245	hidnplayr	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
2192	leency	244	shufps xmm1,xmm1,10110001b
		245	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		246	movhlps xmm1,xmm1
		247	cvtps2pi mm1,xmm1
1245	hidnplayr	248	movq .dey12q,mm0
2192	leency	249	movq .dby12q,mm1
		250
1245	hidnplayr	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]
2192	leency	262	sub ax,word[.b_x1]
		263	cwde
1245	hidnplayr	264	shl eax,ROUND
2192	leency	265	cdq
1245	hidnplayr	266	idiv ebx
2192	leency	267	; mov .dbx12,eax
1245	hidnplayr	268	push eax
2192	leency	269
1245	hidnplayr	270
2192	leency	271	sub ax,word[.b_y1]
		272	cwde
1245	hidnplayr	273	shl eax,ROUND
2192	leency	274	cdq
1245	hidnplayr	275	idiv ebx
2192	leency	276	; mov .dby12,eax
1245	hidnplayr	277	push eax
2192	leency	278
1245	hidnplayr	279
		280	; mov ebx,.dby12
		281	; int3
		282
		283
2192	leency	284	sub ax,word[.e_x1]
		285	cwde
1245	hidnplayr	286	shl eax,ROUND
2192	leency	287	cdq
1245	hidnplayr	288	idiv ebx
2192	leency	289	; mov .dex12,eax
1245	hidnplayr	290	push eax
2192	leency	291
1245	hidnplayr	292
2192	leency	293	sub ax,word[.e_y1]
		294	cwde
1245	hidnplayr	295	shl eax,ROUND
2192	leency	296	cdq
1245	hidnplayr	297	idiv ebx
2192	leency	298	; mov .dey12,eax
1245	hidnplayr	299	push eax
2192	leency	300
1245	hidnplayr	301
		302	mov ax,.z2
2192	leency	303	sub ax,.z1
		304	cwde
		305	shl eax,CATMULL_SHIFT
		306	cdq
		307	idiv ebx
		308	push eax
		309	.bt_dx12_done:
1245	hidnplayr	310
		311
2192	leency	312	sub bx,.y1
		313	jnz .bt_dx13_make
		314	mov ecx,6
		315	xor edx,edx
		316	@@:
1245	hidnplayr	317	push edx ;dword 0
2192	leency	318	loop @b
		319	jmp .bt_dx13_done
		320	.bt_dx13_make:
1245	hidnplayr	321	mov ax,.x3
2192	leency	322	sub ax,.x1
		323	cwde
1245	hidnplayr	324	movsx ebx,bx
2192	leency	325	shl eax,ROUND
		326	cdq
1245	hidnplayr	327	idiv ebx
2192	leency	328	; mov .dx13,eax
1245	hidnplayr	329	push eax
2192	leency	330
1245	hidnplayr	331
		332
		333
		334	; mov eax,255
		335	cvtsi2ss xmm4,[i255d]
		336	divss xmm3,xmm4
2192	leency	337	rcpss xmm3,xmm3
		338	; mulss xmm3,xmm4
1245	hidnplayr	339	shufps xmm3,xmm3,0
2192	leency	340	sub esp,16
		341
1245	hidnplayr	342
2192	leency	343	movd mm1,[.b_x3]
		344	movd mm2,[.e_x1]
		345	movd mm3,[.e_x3]
		346
1245	hidnplayr	347
2192	leency	348	punpcklwd mm0,mm4
1245	hidnplayr	349	punpcklwd mm1,mm4
		350	punpcklwd mm2,mm4
		351	punpcklwd mm3,mm4
		352
		353
		354	movlhps xmm0,xmm0
2192	leency	355	cvtpi2ps xmm0,mm2
1245	hidnplayr	356	cvtpi2ps xmm1,mm1
		357	movlhps xmm1,xmm1
2192	leency	358	cvtpi2ps xmm1,mm3
1245	hidnplayr	359	subps xmm1,xmm0
2192	leency	360
1245	hidnplayr	361
		362	mulps xmm1,xmm3
2192	leency	363	shufps xmm1,xmm1,10110001b
		364	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		365	movhlps xmm1,xmm1
		366	cvtps2pi mm1,xmm1
1245	hidnplayr	367	movq .dey13q,mm0
2192	leency	368	movq .dby13q,mm1
		369
1245	hidnplayr	370
		371
		372
2192	leency	373	sub ax,word[.b_x1]
		374	cwde
1245	hidnplayr	375	shl eax,ROUND
2192	leency	376	cdq
1245	hidnplayr	377	idiv ebx
2192	leency	378	; mov .dbx13,eax
1245	hidnplayr	379	push eax
2192	leency	380
1245	hidnplayr	381
2192	leency	382	sub ax,word[.b_y1]
		383	cwde
1245	hidnplayr	384	shl eax,ROUND
2192	leency	385	cdq
1245	hidnplayr	386	idiv ebx
2192	leency	387	; mov .dby13,eax
1245	hidnplayr	388	push eax
2192	leency	389
1245	hidnplayr	390
2192	leency	391	sub ax,word[.e_x1]
		392	cwde
1245	hidnplayr	393	shl eax,ROUND
2192	leency	394	cdq
1245	hidnplayr	395	idiv ebx
2192	leency	396	; mov .dex13,eax
1245	hidnplayr	397	push eax
2192	leency	398
1245	hidnplayr	399
2192	leency	400	sub ax,word[.e_y1]
		401	cwde
1245	hidnplayr	402	shl eax,ROUND
2192	leency	403	cdq
1245	hidnplayr	404	idiv ebx
2192	leency	405	; mov .dey13,eax
1245	hidnplayr	406	push eax
2192	leency	407
1245	hidnplayr	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
2192	leency	421	sub bx,.y2
		422	jnz .bt_dx23_make
		423	mov ecx,6
		424	xor edx,edx
		425	@@:
1245	hidnplayr	426	push edx ;dword 0
2192	leency	427	loop @b
		428	jmp .bt_dx23_done
		429	.bt_dx23_make:
1245	hidnplayr	430	mov ax,.x3
2192	leency	431	sub ax,.x2
		432	cwde
1245	hidnplayr	433	movsx ebx,bx
2192	leency	434	shl eax,ROUND
		435	cdq
1245	hidnplayr	436	idiv ebx
2192	leency	437	; mov .dx23,eax
1245	hidnplayr	438	push eax
2192	leency	439
1245	hidnplayr	440
		441
		442
		443	; mov eax,255
		444	cvtsi2ss xmm4,[i255d] ;eax
		445	divss xmm3,xmm4
2192	leency	446	shufps xmm3,xmm3,0
		447	sub esp,16
		448
1245	hidnplayr	449
2192	leency	450	movd mm1,[.b_x3]
		451	movd mm2,[.e_x2]
		452	movd mm3,[.e_x3]
		453
1245	hidnplayr	454
2192	leency	455	punpcklwd mm0,mm4
1245	hidnplayr	456	punpcklwd mm1,mm4
		457	punpcklwd mm2,mm4
		458	punpcklwd mm3,mm4
		459
		460
		461	movlhps xmm0,xmm0
2192	leency	462	cvtpi2ps xmm0,mm2
1245	hidnplayr	463	cvtpi2ps xmm1,mm1
		464	movlhps xmm1,xmm1
2192	leency	465	cvtpi2ps xmm1,mm3
1245	hidnplayr	466	subps xmm1,xmm0
2192	leency	467
1245	hidnplayr	468
2192	leency	469	shufps xmm1,xmm1,10110001b
		470	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		471	movhlps xmm1,xmm1
		472	cvtps2pi mm1,xmm1
1245	hidnplayr	473	movq .dey23q,mm0
2192	leency	474	movq .dby23q,mm1
		475
1245	hidnplayr	476
		477
		478
2192	leency	479	sub ax,word[.b_x2]
		480	cwde
1245	hidnplayr	481	shl eax,ROUND
2192	leency	482	cdq
1245	hidnplayr	483	idiv ebx
2192	leency	484	; mov .dbx23,eax
1245	hidnplayr	485	push eax
2192	leency	486
1245	hidnplayr	487
2192	leency	488	sub ax,word[.b_y2]
		489	cwde
1245	hidnplayr	490	shl eax,ROUND
2192	leency	491	cdq
1245	hidnplayr	492	idiv ebx
2192	leency	493	; mov .dby23,eax
1245	hidnplayr	494	push eax
2192	leency	495
1245	hidnplayr	496
2192	leency	497	sub ax,word[.e_x2]
		498	cwde
1245	hidnplayr	499	shl eax,ROUND
2192	leency	500	cdq
1245	hidnplayr	501	idiv ebx
2192	leency	502	; mov .dex23,eax
1245	hidnplayr	503	push eax
2192	leency	504
1245	hidnplayr	505
2192	leency	506	sub ax,word[.e_y2]
		507	cwde
1245	hidnplayr	508	shl eax,ROUND
2192	leency	509	cdq
1245	hidnplayr	510	idiv ebx
2192	leency	511	; mov .dey23,eax
1245	hidnplayr	512	push eax
2192	leency	513	end if
1245	hidnplayr	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
2192	leency	525	shl eax,ROUND
		526	; mov .cx1,eax
1245	hidnplayr	527	; mov .cx2,eax
		528	push eax eax
2192	leency	529	; push eax
1245	hidnplayr	530
		531
2192	leency	532	shl eax,ROUND
		533	mov .cbx1,eax
		534	mov .cbx2,eax
		535	; push eax eax
1245	hidnplayr	536	; push eax
		537
		538
2192	leency	539	shl eax,ROUND
		540	mov .cby1,eax
		541	mov .cby2,eax
		542	; push eax eax
1245	hidnplayr	543	; push eax
		544
		545
2192	leency	546	shl eax,ROUND
		547	mov .cex1,eax
		548	mov .cex2,eax
		549	; push eax eax
1245	hidnplayr	550	;push eax
		551
		552
2192	leency	553	shl eax,ROUND
		554	mov .cey1,eax
		555	mov .cey2,eax
		556	sub esp,32
		557	; push eax eax
1245	hidnplayr	558	;push eax
		559
		560
2192	leency	561	shl eax,CATMULL_SHIFT
		562	; mov .cz1,eax
1245	hidnplayr	563	; mov .cz2,eax
		564	push eax eax
		565	;push eax
		566
		567
2192	leency	568	cmp cx,.y2
		569	jge .loop12_done
		570	.loop12:
1245	hidnplayr	571	call .call_line
2192	leency	572
1245	hidnplayr	573
2192	leency	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
1245	hidnplayr	600	mov edx,.dbx13
2192	leency	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
1245	hidnplayr	609
2192	leency	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
1245	hidnplayr	618
		619	mov ebx,.dz13
2192	leency	620	add .cz1,ebx
		621	mov edx,.dz12
		622	add .cz2,edx
		623
1245	hidnplayr	624
2192	leency	625	cmp cx,.y2
		626	jl .loop12
		627	.loop12_done:
1245	hidnplayr	628
		629
2192	leency	630	cmp cx,.y3
		631	jge .loop23_done
		632
1245	hidnplayr	633
2192	leency	634	shl eax,CATMULL_SHIFT
		635	mov .cz2,eax
		636
1245	hidnplayr	637
2192	leency	638	shl eax,ROUND
		639	mov .cx2,eax
		640
1245	hidnplayr	641
2192	leency	642	shl eax,ROUND
		643	mov .cbx2,eax
		644
1245	hidnplayr	645
2192	leency	646	shl eax,ROUND
		647	mov .cby2,eax
		648
1245	hidnplayr	649
2192	leency	650	shl eax,ROUND
		651	mov .cex2,eax
		652
1245	hidnplayr	653
2192	leency	654	shl eax,ROUND
		655	mov .cey2,eax
		656
1245	hidnplayr	657
		658	call .call_line
2192	leency	659	;if Ext = NON
1245	hidnplayr	660	mov eax,.dx13
2192	leency	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
1245	hidnplayr	687
		688	mov edx,.dbx13
2192	leency	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
1245	hidnplayr	697
2192	leency	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
1245	hidnplayr	706
		707
2192	leency	708	add .cz1,ebx
		709	mov edx,.dz23
		710	add .cz2,edx
		711	;else
1245	hidnplayr	712	; movq mm0,.db13q
		713	; movq mm1,.cbx1q
		714
		715
2192	leency	716	cmp cx,.y3
		717	jl .loop23
		718	.loop23_done:
1245	hidnplayr	719
		720
2192	leency	721	ret 34
1245	hidnplayr	722
		723
		724
		725
		726
		727
2192	leency	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
1245	hidnplayr	742
2192	leency	743	sar eax,ROUND
		744	mov ebx,.cx2
		745	sar ebx,ROUND
		746
1245	hidnplayr	747
2192	leency	748
1245	hidnplayr	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]
2192	leency	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]
1245	hidnplayr	768	.z2 equ dword [ebp+52] ; -- \|> z coords shifted
2192	leency	769	.z1 equ dword [ebp+56] ; -- shl CATMULL_SHIFT
		770
1245	hidnplayr	771
2192	leency	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]
1245	hidnplayr	784
		785
2192	leency	786
1245	hidnplayr	787
2192	leency	788	or ecx,ecx
		789	jl .bl_end
		790	cmp ecx,SIZE_Y
		791	jge .bl_end
		792
1245	hidnplayr	793
2192	leency	794	jl @f
		795	je .bl_end
		796
1245	hidnplayr	797
2192	leency	798	if Ext=NON
1245	hidnplayr	799	mov edx,.bx1
2192	leency	800	xchg edx,.bx2
		801	mov .bx1,edx
		802	mov edx,.by1
		803	xchg edx,.by2
		804	mov .by1,edx
		805
1245	hidnplayr	806
2192	leency	807	xchg edx,.ex2
		808	mov .ex1,edx
		809	mov edx,.ey1
		810	xchg edx,.ey2
		811	mov .ey1,edx
		812	else
1245	hidnplayr	813	movq mm0,.bx1
2192	leency	814	movq mm1,.ex1
		815	movq mm2,.bx2
		816	movq mm3,.ex2
		817	movq .bx2,mm0
		818	movq .ex2,mm1
		819	movq .bx1,mm2
		820	movq .ex1,mm3
		821	end if
1245	hidnplayr	822	mov edx,.z1
2192	leency	823	xchg edx,.z2
		824	mov .z1,edx
		825	@@:
1245	hidnplayr	826	push eax ebx
2192	leency	827	; push ebx ;store x1, x2
1245	hidnplayr	828
		829
2192	leency	830	jge .bl_end
		831	cmp .x2,0
		832	jle .bl_end
		833
1245	hidnplayr	834
2192	leency	835	sub ebx,.x1
		836
1245	hidnplayr	837
1931	yogev_ezra	838
1245	hidnplayr	839
2192	leency	840	cvtsi2ss xmm3,ebx ;rcps
		841	shufps xmm3,xmm3,0
		842
1245	hidnplayr	843
		844	; movq mm1,.bx2q
		845	; movq mm2,.ex1q
		846	; movq mm3,.ex2q
		847	; psubd mm1,mm0
		848	; psubd mm3,mm2
		849	; cvtpi2ps xmm1,mm1
		850	; movlhps xmm1,xmm1
		851	; cvtpi2ps xmm1,mm3
		852
		853
1931	yogev_ezra	854	movlhps xmm0,xmm0
2192	leency	855	cvtpi2ps xmm0,.ex1 ;mm2 ; ex1; ey1
1931	yogev_ezra	856	cvtpi2ps xmm1,.bx2 ;mm1 ; bx2; by2
		857	movlhps xmm1,xmm1
2192	leency	858	cvtpi2ps xmm1,.ex2 ;mm3 ; ex2; ey2
1931	yogev_ezra	859	subps xmm1,xmm0
2192	leency	860	; hi lo
		861	divps xmm1,xmm3 ; xmm1 -> dby; dbx; dey; dex
		862
1245	hidnplayr	863
2192	leency	864	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		865	movhlps xmm1,xmm1
		866	cvtps2pi mm1,xmm1
1245	hidnplayr	867	movq .dex,mm0 ; hi - lo -> dbx, dex
2192	leency	868	movq .dey,mm1 ; hi - lo -> dby, dey
		869
1245	hidnplayr	870
		871
		872
2192	leency	873	sub eax,.bx1
		874	cdq
		875	idiv ebx
		876	push eax
		877
1245	hidnplayr	878
2192	leency	879	sub eax,.ex1
		880	cdq
		881	idiv ebx
		882	push eax
		883
1245	hidnplayr	884
2192	leency	885	sub eax,.by1
		886	cdq
		887	idiv ebx
		888	push eax
		889
1245	hidnplayr	890
2192	leency	891	sub eax,.ey1
		892	cdq
		893	idiv ebx
		894	push eax
		895
1245	hidnplayr	896
		897
		898
2192	leency	899	sub eax,.z1
		900	cdq
		901	idiv ebx
		902	push eax
		903
1245	hidnplayr	904
2192	leency	905	jge @f ; CLIPPING ON FUNCTION
		906	; cutting triangle exceedes screen
		907	mov ebx,.x1
		908	neg ebx
		909	imul ebx ; eax = .dz * abs(.x1)
		910	add .z1,eax
		911	mov .x1,0
		912
1245	hidnplayr	913
2192	leency	914	imul ebx
		915	add .bx1,eax
		916
1245	hidnplayr	917
2192	leency	918	imul ebx
		919	add .by1,eax
		920
1245	hidnplayr	921
2192	leency	922	imul ebx
		923	add .ex1,eax
		924
1245	hidnplayr	925
2192	leency	926	imul ebx
		927	add .ey1,eax
		928	@@:
1245	hidnplayr	929	cmp .x2,SIZE_X
2192	leency	930	jl @f
		931	mov .x2,SIZE_X
		932	@@:
1245	hidnplayr	933	mov eax,SIZE_X ;calc memory begin in buffers
2192	leency	934	mov ebx,.y
		935	mul ebx
		936	mov ebx,.x1
		937	add eax,ebx
		938	mov ebx,eax
		939	lea eax,[eax*3]
		940	add edi,eax ; edi - screen
		941	mov esi,.z_buff ; z-buffer filled with dd variables
		942	shl ebx,2
		943	add esi,ebx ; esi - Z buffer
		944
1245	hidnplayr	945
2192	leency	946	sub ecx,.x1
		947	; init current variables
		948	push dword .bx1 ;.by1 .ex1 .ey1 .z1 esi
		949	push dword .ex1
		950	push dword .by1
		951	push dword .ey1
		952
1245	hidnplayr	953
2192	leency	954	push esi
		955
1245	hidnplayr	956
		957	pxor mm0,mm0
		958	movq mm3,.cex ; hi - lo -> cbx; cex
1931	yogev_ezra	959	movq mm4,.cey ; hi - lo -> cby; cey
		960	; movq mm5,mm3
1245	hidnplayr	961	; movq mm6,mm4
		962	; psrad mm5,ROUND
		963	; psrad mm6,ROUND
		964	; movq .ceyq,mm5
		965	; movq .cbyq,mm6
		966	mov edx,.czbuff
		967	else
		968	cld
		969	end if
		970	.draw:
		971	; if TEX = SHIFTING ;bump drawing only in shifting mode
		972	if Ext=NON
		973	mov esi,.czbuff ; .czbuff current address in buffer
2192	leency	974	mov ebx,.cz ; .cz - cur z position
		975	cmp ebx,dword[esi]
		976	else
1245	hidnplayr	977	mov ebx,.cz
2192	leency	978	cmp ebx,dword[edx]
		979	end if
1245	hidnplayr	980	jge .skip
2192	leency	981
1245	hidnplayr	982
1931	yogev_ezra	983	mov eax,.cby
2192	leency	984	mov esi,.cbx
		985	sar eax,ROUND
		986	sar esi,ROUND
		987	shl eax,TEX_SHIFT ;-
		988	add esi,eax
		989	lea esi,[esi*3] ;- ; esi - current b. texture addres
		990	add esi,.bmap
		991
1245	hidnplayr	992
2192	leency	993	mov eax,.cey ;.cey - current env map y
		994	sar ebx,ROUND
		995	sar eax,ROUND
		996
1931	yogev_ezra	997
2192	leency	998	add ebx,eax
		999	lea ebx,[ebx*3]
		1000	add ebx,.emap
		1001
1931	yogev_ezra	1002
		1003
		1004	movq mm5,mm4 ;.cey
2192	leency	1005	psrad mm5,ROUND
		1006	pslld mm5,TEX_SHIFT
		1007	movq mm6,mm3 ;.cex
		1008	psrad mm6,ROUND
		1009	paddd mm5,mm6
		1010	movq mm6,mm5
		1011	paddd mm5,mm5
		1012	paddd mm5,mm6
		1013	paddd mm5,.emap
		1014	movd esi,mm5
		1015	psrlq mm5,32
		1016	movd ebx,mm5
		1017	end if
1931	yogev_ezra	1018	if Ext>=MMX
1245	hidnplayr	1019	movd mm1,[esi]
2192	leency	1020	movd mm2,[ebx]
		1021	punpcklbw mm1,mm0
		1022	punpcklbw mm2,mm0
		1023	pmullw mm1,mm2
		1024	psrlw mm1,8
		1025	packuswb mm1,mm0
		1026	movd [edi],mm1
		1027	mov ebx,.cz
		1028	mov dword[edx],ebx
		1029	else
1245	hidnplayr	1030	cld ; esi - tex e.
2192	leency	1031	lodsb ; ebx - tex b.
		1032	mov dl,[ebx]
		1033	mul dl
		1034	shr ax,8
		1035	stosb
		1036	inc ebx
		1037	lodsb
		1038	mov dl,[ebx]
		1039	mul dl
		1040	shr ax,8
		1041	stosb
		1042	inc ebx
		1043	lodsb
		1044	mov dl,[ebx]
		1045	mul dl
		1046	shr ax,8
		1047	stosb
		1048	mov ebx,.cz
		1049	mov esi,.czbuff
		1050	mov dword[esi],ebx
		1051	jmp .no_skip
		1052	end if
1245	hidnplayr	1053	.skip:
		1054	add edi,3
2192	leency	1055
1245	hidnplayr	1056
		1057	.no_skip:
		1058	add .czbuff,4
2192	leency	1059	mov eax,.dbx
		1060	add .cbx,eax
		1061	mov eax,.dby
		1062	add .cby,eax
		1063	mov eax,.dex
		1064	add .cex,eax
		1065	mov eax,.dey
		1066	add .cey,eax
		1067	else
1245	hidnplayr	1068	add edx,4
2192	leency	1069	paddd mm3,.dex
		1070	paddd mm4,.dey
		1071	; movq mm5,mm3
1245	hidnplayr	1072	; movq mm6,mm4
		1073	; psrad mm5,ROUND
		1074	; psrad mm6,ROUND
		1075	; movq .cex,mm3
1979	yogev_ezra	1076	; movq .cey,mm4
		1077	end if
1245	hidnplayr	1078	mov eax,.dz
2192	leency	1079	add .cz,eax
		1080	if Ext = NON
1245	hidnplayr	1081	dec ecx
2192	leency	1082	jnz .draw
		1083	else
1245	hidnplayr	1084	loop .draw
2192	leency	1085	end if
1245	hidnplayr	1086
		1087
		1088	mov esp,ebp
2192	leency	1089	ret 56
1245	hidnplayr	1090

Subversion Repositories Kolibri OS

(root)/programs/demos/3DS/TWO_TEX.INC @ 2564 – Rev 2192