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

Rev	Author	Line No.	Line
1245	hidnplayr	1	;SIZE_X equ 350
		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	CATMULL_SHIFT equ 8
		9	;TEXTURE_SIZE EQU (TEX_X * TEX_Y)-1
		10	;Ext = NON
		11	;MMX = 1
		12	;NON = 0
		13	;------- Big thanks to Majuma (www.majuma.xt.pl) for absolutely great---
		14	;------- DOS 13h mode demos --------------------------------------------
		15	;------- Procedure draws bump triangle using Catmull Z-buffer algorithm-
		16	;------- (Z coordinate interpolation)-----------------------------------
		17	bump_triangle_z:
		18	;------------------in - eax - x1 shl 16 + y1 -----------
		19	;---------------------- ebx - x2 shl 16 + y2 -----------
		20	;---------------------- ecx - x3 shl 16 + y3 -----------
		21	;---------------------- edx - pointer to bump map ------
		22	;---------------------- esi - pointer to environment map
		23	;---------------------- edi - pointer to screen buffer--
		24	;---------------------- stack : bump coordinates--------
		25	;---------------------- environment coordinates-
		26	;---------------------- Z position coordinates--
		27	;---------------------- pointer io Z buffer-----
		28	;-- Z-buffer - filled with coordinates as dword --------
		29	;-- (Z coor. as word) shl CATMULL_SHIFT ----------------
		30	.b_x1 equ ebp+4 ; procedure don't save registers !!!
		31	.b_y1 equ ebp+6 ; each coordinate as word
		32	.b_x2 equ ebp+8
		33	.b_y2 equ ebp+10
		34	.b_x3 equ ebp+12
		35	.b_y3 equ ebp+14
		36	.e_x1 equ ebp+16
		37	.e_y1 equ ebp+18
		38	.e_x2 equ ebp+20
		39	.e_y2 equ ebp+22
		40	.e_x3 equ ebp+24
		41	.e_y3 equ ebp+26
		42	.z1 equ word[ebp+28]
		43	.z2 equ word[ebp+30]
		44	.z3 equ word[ebp+32]
		45	.z_buff equ dword[ebp+34] ; pointer to Z-buffer
		46
		47
		48	.t_bmap equ dword[ebp-4] ; pointer to bump map
		49	.t_emap equ dword[ebp-8] ; pointer to e. map
		50	.x1 equ word[ebp-10]
		51	.y1 equ word[ebp-12]
		52	.x2 equ word[ebp-14]
		53	.y2 equ word[ebp-16]
		54	.x3 equ word[ebp-18]
		55	.y3 equ word[ebp-20]
		56
		57	.dx12 equ dword[ebp-24]
		58	.dz12 equ [ebp-28]
		59	.dbx12 equ dword[ebp-32]
		60	.dby12 equ [ebp-36]
		61	.dex12 equ dword[ebp-40]
		62	.dey12 equ [ebp-44]
		63
		64	.dx13 equ dword[ebp-48]
		65	.dz13 equ [ebp-52]
		66	.dbx13 equ dword[ebp-56]
		67	.dby13 equ [ebp-60]
		68	.dex13 equ dword[ebp-64]
		69	.dey13 equ [ebp-68]
		70
		71	.dx23 equ dword[ebp-72]
		72	.dz23 equ [ebp-76]
		73	.dbx23 equ dword[ebp-80]
		74	.dby23 equ [ebp-84]
		75	.dex23 equ dword[ebp-88]
		76	.dey23 equ [ebp-92]
		77
		78	.cx1 equ dword[ebp-96] ; current variables
		79	.cz1 equ [ebp-100]
		80	.cx2 equ dword[ebp-104]
		81	.cz2 equ [ebp-108]
		82	.cbx1 equ dword[ebp-112]
		83	.cby1 equ [ebp-116]
		84	.cbx2 equ dword[ebp-120]
		85	.cby2 equ [ebp-124]
		86	.cex1 equ dword[ebp-128]
		87	.cey1 equ [ebp-132]
		88	.cex2 equ dword[ebp-136]
		89	.cey2 equ [ebp-140]
		90
		91	mov ebp,esp
		92	push edx ; store bump map
		93	push esi ; store e. map
		94	; sub esp,120
		95	.sort3: ; sort triangle coordinates...
		96	cmp ax,bx
		97	jle .sort1
		98	xchg eax,ebx
		99	mov edx,dword[.b_x1]
		100	xchg edx,dword[.b_x2]
		101	mov dword[.b_x1],edx
		102	mov edx,dword[.e_x1]
		103	xchg edx,dword[.e_x2]
		104	mov dword[.e_x1],edx
		105	mov dx,.z1
		106	xchg dx,.z2
		107	mov .z1,dx
		108	.sort1:
		109	cmp bx,cx
		110	jle .sort2
		111	xchg ebx,ecx
		112	mov edx,dword[.b_x2]
		113	xchg edx,dword[.b_x3]
		114	mov dword[.b_x2],edx
		115	mov edx,dword[.e_x2]
		116	xchg edx,dword[.e_x3]
		117	mov dword[.e_x2],edx
		118	mov dx,.z2
		119	xchg dx,.z3
		120	mov .z2,dx
		121	jmp .sort3
		122	.sort2:
		123	push eax ; store triangle coords in variables
		124	push ebx
		125	push ecx
		126
		127	mov edx,80008000h ; eax,ebx,ecx are ANDd together into edx which means that
		128	and edx,ebx ; if all of them are negative a sign flag is raised
		129	and edx,ecx
		130	and edx,eax
		131	test edx,80008000h ; Check both X&Y at once
		132	jne .loop23_done
		133	; mov edx,eax ; eax,ebx,ecx are ORd together into edx which means that
		134	; or edx,ebx ; if any one of them is negative a sign flag is raised
		135	; or edx,ecx
		136	; test edx,80000000h ; Check only X
		137	; jne .loop23_done
		138
		139	; cmp .x1,SIZE_X ; {
		140	; jg .loop23_done
		141	; cmp .x2,SIZE_X ; This can be optimized with effort
		142	; jg .loop23_done
		143	; cmp .x3,SIZE_X
		144	; jg .loop23_done ; {
		145
		146
		147	mov bx,.y2 ; calc delta 12
		148	sub bx,.y1
		149	jnz .bt_dx12_make
		150	mov ecx,6
		151	xor edx,edx
		152	@@:
		153	push edx ;dword 0
		154	loop @b
		155	jmp .bt_dx12_done
		156	.bt_dx12_make:
		157	mov ax,.x2
		158	sub ax,.x1
		159	cwde
		160	movsx ebx,bx
		161	shl eax,ROUND
		162	cdq
		163	idiv ebx
		164	; mov .dx12,eax
		165	push eax
		166
		167	mov ax,.z2
		168	sub ax,.z1
		169	cwde
		170	shl eax,CATMULL_SHIFT
		171	cdq
		172	idiv ebx
		173	push eax
		174	if Ext>=SSE
		175
		176	sub esp,16
		177	cvtsi2ss xmm3,ebx ;rcps
		178	; mov eax,255
		179	cvtsi2ss xmm4,[i255d]
		180	divss xmm3,xmm4
		181	rcpss xmm3,xmm3
		182	; mulss xmm3,xmm4
		183	shufps xmm3,xmm3,0
		184
		185	movd mm0,[.b_x1]
		186	movd mm1,[.b_x2]
		187	movd mm2,[.e_x1]
		188	movd mm3,[.e_x2]
		189	pxor mm4,mm4
		190	punpcklwd mm0,mm4
		191	punpcklwd mm1,mm4
		192	punpcklwd mm2,mm4
		193	punpcklwd mm3,mm4
		194
		195	psubd mm1,mm0
		196	psubd mm3,mm2
		197
		198	; cvtpi2ps xmm0,mm0
		199	; movlhps xmm0,xmm0
		200	; cvtpi2ps xmm0,mm2
		201	cvtpi2ps xmm1,mm1
		202	movlhps xmm1,xmm1
		203	cvtpi2ps xmm1,mm3
		204	; subps xmm1,xmm0
		205
		206
		207	; cvtpi2ps xmm0,mm3
		208	; divps xmm1,xmm3
		209	mulps xmm1,xmm3
		210	shufps xmm1,xmm1,10110001b
		211	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		212	movhlps xmm1,xmm1
		213	cvtps2pi mm1,xmm1
		214	movq .dey12,mm0
		215	movq .dby12,mm1
		216
		217
		218	else
		219
		220	mov ax,word[.b_x2]
		221	sub ax,word[.b_x1]
		222	cwde
		223	shl eax,ROUND
		224	cdq
		225	idiv ebx
		226	; mov .dbx12,eax
		227	push eax
		228
		229	mov ax,word[.b_y2]
		230	sub ax,word[.b_y1]
		231	cwde
		232	shl eax,ROUND
		233	cdq
		234	idiv ebx
		235	; mov .dby12,eax
		236	push eax
		237
		238	mov ax,word[.e_x2]
		239	sub ax,word[.e_x1]
		240	cwde
		241	shl eax,ROUND
		242	cdq
		243	idiv ebx
		244	; mov .dex12,eax
		245	push eax
		246
		247	mov ax,word[.e_y2]
		248	sub ax,word[.e_y1]
		249	cwde
		250	shl eax,ROUND
		251	cdq
		252	idiv ebx
		253	; mov .dey12,eax
		254	push eax
		255
		256	end if
		257
		258	.bt_dx12_done:
		259
		260	mov bx,.y3 ; calc delta13
		261	sub bx,.y1
		262	jnz .bt_dx13_make
		263	mov ecx,6
		264	xor edx,edx
		265	@@:
		266	push edx ;dword 0
		267	loop @b
		268	jmp .bt_dx13_done
		269	.bt_dx13_make:
		270	mov ax,.x3
		271	sub ax,.x1
		272	cwde
		273	movsx ebx,bx
		274	shl eax,ROUND
		275	cdq
		276	idiv ebx
		277	; mov .dx13,eax
		278	push eax
		279
		280	mov ax,.z3
		281	sub ax,.z1
		282	cwde
		283	shl eax,CATMULL_SHIFT
		284	cdq
		285	idiv ebx
		286	; mov .dz13,eax
		287	push eax
		288
		289	if Ext>=SSE
		290
		291	sub esp,16
		292	cvtsi2ss xmm3,ebx ;rcps
		293	; mov eax,255
		294	cvtsi2ss xmm4,[i255d]
		295	divss xmm3,xmm4
		296	rcpss xmm3,xmm3
		297	; mulss xmm3,xmm4
		298	shufps xmm3,xmm3,0
		299
		300	movd mm0,[.b_x1]
		301	movd mm1,[.b_x3]
		302	movd mm2,[.e_x1]
		303	movd mm3,[.e_x3]
		304	pxor mm4,mm4
		305	punpcklwd mm0,mm4
		306	punpcklwd mm1,mm4
		307	punpcklwd mm2,mm4
		308	punpcklwd mm3,mm4
		309
		310	psubd mm1,mm0
		311	psubd mm3,mm2
		312
		313	; cvtpi2ps xmm0,mm0
		314	; movlhps xmm0,xmm0
		315	; cvtpi2ps xmm0,mm2
		316	cvtpi2ps xmm1,mm1
		317	movlhps xmm1,xmm1
		318	cvtpi2ps xmm1,mm3
		319	; subps xmm1,xmm0
		320
		321
		322	; cvtpi2ps xmm0,mm3
		323	; divps xmm1,xmm3
		324	mulps xmm1,xmm3
		325	shufps xmm1,xmm1,10110001b
		326	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		327	movhlps xmm1,xmm1
		328	cvtps2pi mm1,xmm1
		329	movq .dey13,mm0
		330	movq .dby13,mm1
		331
		332
		333	else
		334
		335	mov ax,word[.b_x3]
		336	sub ax,word[.b_x1]
		337	cwde
		338	shl eax,ROUND
		339	cdq
		340	idiv ebx
		341	; mov .dbx13,eax
		342	push eax
		343
		344	mov ax,word[.b_y3]
		345	sub ax,word[.b_y1]
		346	cwde
		347	shl eax,ROUND
		348	cdq
		349	idiv ebx
		350	; mov .dby13,eax
		351	push eax
		352
		353	mov ax,word[.e_x3]
		354	sub ax,word[.e_x1]
		355	cwde
		356	shl eax,ROUND
		357	cdq
		358	idiv ebx
		359	; mov .dex13,eax
		360	push eax
		361
		362	mov ax,word[.e_y3]
		363	sub ax,word[.e_y1]
		364	cwde
		365	shl eax,ROUND
		366	cdq
		367	idiv ebx
		368	; mov .dey13,eax
		369	push eax
		370	end if
		371
		372	.bt_dx13_done:
		373
		374	mov bx,.y3 ; calc delta23
		375	sub bx,.y2
		376	jnz .bt_dx23_make
		377	mov ecx,6
		378	xor edx,edx
		379	@@:
		380	push edx ;dword 0
		381	loop @b
		382	jmp .bt_dx23_done
		383	.bt_dx23_make:
		384	mov ax,.x3
		385	sub ax,.x2
		386	cwde
		387	movsx ebx,bx
		388	shl eax,ROUND
		389	cdq
		390	idiv ebx
		391	; mov .dx23,eax
		392	push eax
		393
		394	mov ax,.z3
		395	sub ax,.z2
		396	cwde
		397	shl eax,CATMULL_SHIFT
		398	cdq
		399	idiv ebx
		400	; mov .dz23,eax
		401	push eax
		402	; sub esp,40
		403	if Ext>=SSE
		404
		405	sub esp,16
		406	cvtsi2ss xmm3,ebx ;rcps
		407	; mov eax,255
		408	cvtsi2ss xmm4,[i255d]
		409	divss xmm3,xmm4
		410	rcpss xmm3,xmm3
		411	; mulss xmm3,xmm4
		412	shufps xmm3,xmm3,0
		413
		414	movd mm0,[.b_x2]
		415	movd mm1,[.b_x3]
		416	movd mm2,[.e_x2]
		417	movd mm3,[.e_x3]
		418	pxor mm4,mm4
		419	punpcklwd mm0,mm4
		420	punpcklwd mm1,mm4
		421	punpcklwd mm2,mm4
		422	punpcklwd mm3,mm4
		423
		424	psubd mm1,mm0
		425	psubd mm3,mm2
		426
		427	; cvtpi2ps xmm0,mm0
		428	; movlhps xmm0,xmm0
		429	; cvtpi2ps xmm0,mm2
		430	cvtpi2ps xmm1,mm1
		431	movlhps xmm1,xmm1
		432	cvtpi2ps xmm1,mm3
		433	; subps xmm1,xmm0
		434
		435
		436	; cvtpi2ps xmm0,mm3
		437	; divps xmm1,xmm3
		438	mulps xmm1,xmm3
		439	shufps xmm1,xmm1,10110001b
		440	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		441	movhlps xmm1,xmm1
		442	cvtps2pi mm1,xmm1
		443	movq .dey23,mm0
		444	movq .dby23,mm1
		445
		446	else
		447
		448	mov ax,word[.b_x3]
		449	sub ax,word[.b_x2]
		450	cwde
		451	shl eax,ROUND
		452	cdq
		453	idiv ebx
		454	; mov .dbx23,eax
		455	push eax
		456
		457	mov ax,word[.b_y3]
		458	sub ax,word[.b_y2]
		459	cwde
		460	shl eax,ROUND
		461	cdq
		462	idiv ebx
		463	; mov .dby23,eax
		464	push eax
		465
		466	mov ax,word[.e_x3]
		467	sub ax,word[.e_x2]
		468	cwde
		469	shl eax,ROUND
		470	cdq
		471	idiv ebx
		472	; mov .dex23,eax
		473	push eax
		474
		475	mov ax,word[.e_y3]
		476	sub ax,word[.e_y2]
		477	cwde
		478	shl eax,ROUND
		479	cdq
		480	idiv ebx
		481	; mov .dey23,eax
		482	push eax
		483
		484	end if
		485
		486	.bt_dx23_done:
		487	sub esp,48
		488
		489	movsx eax,.x1
		490	shl eax,ROUND
		491	mov .cx1,eax
		492	mov .cx2,eax
		493	; push eax
		494	; push eax
		495
		496	movsx eax,word[.b_x1]
		497	shl eax,ROUND
		498	mov .cbx1,eax
		499	mov .cbx2,eax
		500	; push eax
		501	; push eax
		502
		503	movsx eax,word[.b_y1]
		504	shl eax,ROUND
		505	mov .cby1,eax
		506	mov .cby2,eax
		507	; push eax
		508	; push eax
		509
		510	movsx eax,word[.e_x1]
		511	shl eax,ROUND
		512	mov .cex1,eax
		513	mov .cex2,eax
		514	; push eax
		515	; push eax
		516
		517	movsx eax,word[.e_y1]
		518	shl eax,ROUND
		519	mov .cey1,eax
		520	mov .cey2,eax
		521	; push eax
		522	; push eax
		523
		524	movsx eax,.z1
		525	shl eax,CATMULL_SHIFT
		526	mov .cz1,eax
		527	mov .cz2,eax
		528	; push eax
		529	; push eax
		530
		531	movsx ecx,.y1
		532	cmp cx,.y2
		533	jge .loop12_done
		534	.loop12:
		535	call .call_bump_line
		536
		537	if Ext >= MMX
		538	movq mm0,.cby2
		539	movq mm1,.cby1
		540	movq mm2,.cey2
		541	movq mm3,.cey1
		542	movq mm4,.cz1
		543	movq mm5,.cz2
		544	paddd mm0,.dby12
		545	paddd mm1,.dby13
		546	paddd mm2,.dey12
		547	paddd mm3,.dey13
		548	paddd mm4,.dz13
		549	paddd mm5,.dz12
		550	movq .cby2,mm0
		551	movq .cby1,mm1
		552	movq .cey1,mm3
		553	movq .cey2,mm2
		554	movq .cz1,mm4
		555	movq .cz2,mm5
		556	else
		557	mov edx,.dbx13
		558	add .cbx1,edx
		559	mov eax,.dbx12
		560	add .cbx2,eax
		561	mov ebx,.dby13
		562	add .cby1,ebx
		563	mov edx,.dby12
		564	add .cby2,edx
		565
		566	mov eax,.dex13
		567	add .cex1,eax
		568	mov ebx,.dex12
		569	add .cex2,ebx
		570	mov edx,.dey13
		571	add .cey1,edx
		572	mov eax,.dey12
		573	add .cey2,eax
		574
		575	mov eax,.dx13
		576	add .cx1,eax
		577	mov ebx,.dx12
		578	add .cx2,ebx
		579
		580	mov ebx,.dz13
		581	add .cz1,ebx
		582	mov edx,.dz12
		583	add .cz2,edx
		584	end if
		585	inc ecx
		586	cmp cx,.y2
		587	jl .loop12
		588	.loop12_done:
		589
		590	movsx ecx,.y2
		591	cmp cx,.y3
		592	jge .loop23_done
		593
		594	movsx eax,.z2
		595	shl eax,CATMULL_SHIFT
		596	mov .cz2,eax
		597
		598	movsx eax,.x2
		599	shl eax,ROUND
		600	mov .cx2,eax
		601
		602	movzx eax,word[.b_x2]
		603	shl eax,ROUND
		604	mov .cbx2,eax
		605
		606	movzx eax,word[.b_y2]
		607	shl eax,ROUND
		608	mov .cby2,eax
		609
		610	movzx eax,word[.e_x2]
		611	shl eax,ROUND
		612	mov .cex2,eax
		613
		614	movzx eax,word[.e_y2]
		615	shl eax,ROUND
		616	mov .cey2,eax
		617
		618	.loop23:
		619	call .call_bump_line
		620
		621	if Ext >= MMX
		622	movq mm0,.cby2
		623	movq mm1,.cby1
		624	movq mm2,.cey2
		625	movq mm3,.cey1
		626	movq mm4,.cz1
		627	movq mm5,.cz2
		628	paddd mm0,.dby23
		629	paddd mm1,.dby13
		630	paddd mm2,.dey23
		631	paddd mm3,.dey13
		632	paddd mm4,.dz13
		633	paddd mm5,.dz23
		634	movq .cby2,mm0
		635	movq .cby1,mm1
		636	movq .cey1,mm3
		637	movq .cey2,mm2
		638	movq .cz1,mm4
		639	movq .cz2,mm5
		640	else
		641	mov eax,.dx13
		642	add .cx1,eax
		643	mov ebx,.dx23
		644	add .cx2,ebx
		645
		646	mov edx,.dbx13
		647	add .cbx1,edx
		648	mov eax,.dbx23
		649	add .cbx2,eax
		650	mov ebx,.dby13
		651	add .cby1,ebx
		652	mov edx,.dby23
		653	add .cby2,edx
		654
		655	mov eax,.dex13
		656	add .cex1,eax
		657	mov ebx,.dex23
		658	add .cex2,ebx
		659	mov edx,.dey13
		660	add .cey1,edx
		661	mov eax,.dey23
		662	add .cey2,eax
		663
		664	mov ebx,.dz13
		665	add .cz1,ebx
		666	mov edx,.dz23
		667	add .cz2,edx
		668	end if
		669	inc ecx
		670	cmp cx,.y3
		671	jl .loop23
		672	.loop23_done:
		673
		674	mov esp,ebp
		675	ret 34
		676
		677	.call_bump_line:
		678
		679	; push ebp
		680	; push ecx
		681	pushad
		682
		683	push dword .cz1
		684	push dword .cz2
		685	push .z_buff
		686	push .t_emap
		687	push .t_bmap
		688	push dword .cey2
		689	push .cex2
		690	push dword .cey1
		691	push .cex1
		692	push dword .cby2
		693	push .cbx2
		694	push dword .cby1
		695	push .cbx1
		696	push ecx
		697
		698	mov eax,.cx1
		699	sar eax,ROUND
		700	mov ebx,.cx2
		701	sar ebx,ROUND
		702
		703	call bump_line_z
		704
		705	popad
		706	ret
		707	bump_line_z:
		708	;--------------in: eax - x1
		709	;-------------- ebx - x2
		710	;-------------- edi - pointer to screen buffer
		711	;stack - another parameters :
		712	.y equ dword [ebp+4]
		713	.bx1q equ [ebp+8]
		714	.bx2q equ [ebp+16]
		715	.ex1q equ [ebp+24]
		716	.ex2q equ [ebp+32]
		717	.bx1 equ dword [ebp+8] ; ---
		718	.by1 equ dword [ebp+12] ; \|
		719	.bx2 equ dword [ebp+16] ; \|
		720	.by2 equ dword [ebp+20] ; \|> bump and env coords
		721	.ex1 equ dword [ebp+24] ; \|> shifted shl ROUND
		722	.ey1 equ dword [ebp+28] ; \|
		723	.ex2 equ dword [ebp+32] ; \|
		724	.ey2 equ dword [ebp+36] ; ---
		725	;.bx1q equ qword [ebp+8] ; - new
		726	;.ex1q equ qword [ebp+24] ; - new
		727	.bmap equ dword [ebp+40]
		728	.emap equ dword [ebp+44]
		729	.z_buff equ dword [ebp+48]
		730	.z2 equ dword [ebp+52] ; -- \|> z coords shifted
		731	.z1 equ dword [ebp+56] ; -- shl CATMULL_SHIFT
		732
		733	.x1 equ dword [ebp-4]
		734	.x2 equ dword [ebp-8]
		735	.dbx equ dword [ebp-12]
		736	.dby equ dword [ebp-16]
		737	.dbyq equ qword [ebp-16] ; - new
		738	.dex equ dword [ebp-20]
		739	.dey equ dword [ebp-24]
		740	.deyq equ qword [ebp-24] ; - new
		741	.dz equ dword [ebp-28]
		742	.cbx equ dword [ebp-32]
		743	.cby equ dword [ebp-36]
		744	.cbyq equ qword [ebp-36] ; - new
		745	.cex equ dword [ebp-40]
		746	.cey equ dword [ebp-44]
		747	.ceyq equ qword [ebp-44] ; - new
		748	.cz equ dword [ebp-48]
		749	.czbuff equ dword [ebp-52]
		750	.temp1 equ ebp-60
		751	.temp2 equ ebp-68
		752	.temp3 equ ebp-76
		753	.temp4 equ ebp-84
		754	.temp5 equ ebp-92
		755
		756	mov ebp,esp
		757
		758	mov ecx,.y
		759	or ecx,ecx
		760	jl .bl_end
		761	cmp ecx,SIZE_Y
		762	jge .bl_end
		763
		764	cmp eax,ebx
		765	jl .bl_ok
		766	je .bl_end
		767
		768	xchg eax,ebx
		769	if Ext=NON
		770	mov edx,.bx1
		771	xchg edx,.bx2
		772	mov .bx1,edx
		773	mov edx,.by1
		774	xchg edx,.by2
		775	mov .by1,edx
		776
		777	mov edx,.ex1
		778	xchg edx,.ex2
		779	mov .ex1,edx
		780	mov edx,.ey1
		781	xchg edx,.ey2
		782	mov .ey1,edx
		783	else
		784
		785	movq mm0,.bx1q
		786	movq mm1,.ex1q
		787	movq mm2,.bx2q
		788	movq mm3,.ex2q
		789	movq .bx2q,mm0
		790	movq .ex2q,mm1
		791	movq .bx1q,mm2
		792	movq .ex1q,mm3
		793
		794	end if
		795
		796	mov edx,.z1
		797	xchg edx,.z2
		798	mov .z1,edx
		799
		800	.bl_ok:
		801
		802	push eax
		803	push ebx ;store x1, x2
		804
		805	cmp .x1,SIZE_X
		806	jge .bl_end
		807	cmp .x2,0
		808	jle .bl_end
		809
		810	mov ebx,.x2
		811	sub ebx,.x1
		812
		813	if Ext >= SSE
		814
		815	sub esp,16
		816	cvtsi2ss xmm3,ebx ;rcps
		817	shufps xmm3,xmm3,0
		818
		819	cvtpi2ps xmm0,.bx1q ;mm0
		820	movlhps xmm0,xmm0
		821	cvtpi2ps xmm0,.ex1q ;mm2
		822	cvtpi2ps xmm1,.bx2q ;mm1
		823	movlhps xmm1,xmm1
		824	cvtpi2ps xmm1,.ex2q ;mm3
		825	subps xmm1,xmm0
		826
		827	divps xmm1,xmm3
		828
		829	shufps xmm1,xmm1,10110001b
		830	cvtps2pi mm0,xmm1 ; mm0 -> 2 delta dwords
		831	movhlps xmm1,xmm1
		832	cvtps2pi mm1,xmm1
		833	movq .deyq,mm0
		834	movq .dbyq,mm1
		835
		836	else
		837
		838	mov eax,.bx2 ; calc .dbx
		839	sub eax,.bx1
		840	cdq
		841	idiv ebx
		842	push eax
		843
		844	mov eax,.by2 ; calc .dby
		845	sub eax,.by1
		846	cdq
		847	idiv ebx
		848	push eax
		849
		850	mov eax,.ex2 ; calc .dex
		851	sub eax,.ex1
		852	cdq
		853	idiv ebx
		854	push eax
		855
		856	mov eax,.ey2 ; calc .dey
		857	sub eax,.ey1
		858	cdq
		859	idiv ebx
		860	push eax
		861
		862	end if
		863
		864	mov eax,.z2 ; calc .dz
		865	sub eax,.z1
		866	cdq
		867	idiv ebx
		868	push eax
		869
		870	cmp .x1,0 ; set correctly begin variable
		871	jge @f ; CLIPPING ON FUNCTION
		872	; cutting triangle exceedes screen
		873	mov ebx,.x1
		874	neg ebx
		875	imul ebx ; eax = .dz * abs(.x1)
		876	add .z1,eax
		877	mov .x1,0
		878
		879	mov eax,.dbx
		880	imul ebx
		881	add .bx1,eax
		882
		883	mov eax,.dby
		884	imul ebx
		885	add .by1,eax
		886
		887	mov eax,.dex
		888	imul ebx
		889	add .ex1,eax
		890
		891	mov eax,.dey
		892	imul ebx
		893	add .ey1,eax
		894	@@:
		895	cmp .x2,SIZE_X
		896	jl @f
		897	mov .x2,SIZE_X
		898	@@:
		899	mov eax,SIZE_X ;calc memory begin in buffers
		900	mov ebx,.y
		901	mul ebx
		902	mov ebx,.x1
		903	add eax,ebx
		904	mov ebx,eax
		905	lea eax,[eax*3]
		906	add edi,eax
		907	mov esi,.z_buff ; z-buffer filled with dd variables
		908	shl ebx,2
		909	add esi,ebx
		910
		911	mov ecx,.x2
		912	sub ecx,.x1
		913	; init current variables
		914	push .bx1
		915	push .by1
		916	push .ex1
		917	push .ey1
		918
		919	push .z1 ; current z shl CATMULL_SHIFT
		920	push esi
		921	; It's my first attempt at MMX :), have mercy - Macgub
		922
		923	;; if Ext = MMX
		924	; mov dword[.temp1],esi
		925	; mov dword[.temp1+4],esi
		926	;; movq mm0,.cbyq ; mm0 - current bump coords
		927	;; movq mm1,.ceyq ; mm1 - current env coords
		928	;; movq mm2,.dbyq ; mm2 - delta bump
		929	;; movq mm3,.deyq ; mm3 - delta env
		930	; movd mm6,.z1 ; mm6 - cur z
		931	; movq mm7,qword.[temp1] ; mm7 = lo = hi dword = current z buff
		932	;; mov dword [.temp2],1
		933	;; mov dword [.temp2+4],-1
		934	;; mov dword [.temp3],TEXTURE_SIZE
		935	;; mov dword [.temp3+4],TEXTURE_SIZE
		936	;; mov esi,.bmap
		937	;; mov dword [.temp4],esi
		938	;; mov dword [.temp4+4],esi
		939	;; mov dword [.temp5],TEX_X
		940	;; mov dword [.temp5+4],- TEX_X
		941	; mov dword [.temp1],TEX_SHIFT
		942	; mov dword [.temp1+4],0
		943	;; end if
		944
		945	.draw:
		946	; if TEX = SHIFTING ;bump drawing only in shifting mode
		947	mov esi,.czbuff ; .czbuff current address in buffer
		948	mov ebx,.cz ; .cz - cur z position
		949	cmp ebx,dword[esi]
		950	jge .skip
		951	;; if Ext=NON
		952	mov eax,.cby
		953	sar eax,ROUND
		954	mov esi,.cbx
		955	sar esi,ROUND
		956	;; else
		957	;; movq mm4,mm0 ; mm4 - copies of cur bump coords
		958	;; psrad mm4,ROUND ; mm4 = lo dword = y b coord, hi dword = x b coord
		959	;; movd eax,mm4 ; -
		960	;; psrlq mm4,32 ; -
		961	;; movd esi,mm4 ; -
		962	;;
		963	;;; punpckldq mm5,mm4 ;
		964	;;; psllq mm5,TEX_SHIFT
		965	;;; paddq mm4,mm5 ; mm4 - lo dword index to b. map
		966	;;
		967	;; ; packqd mm4,mm5
		968	; movq mm5,mm4 ; mm5 ~~ current bump map index?
		969
		970	;; end if
		971	shl eax,TEX_SHIFT ;-
		972	add esi,eax ;- ; esi - current bump map index
		973	;; if Ext = NON
		974	mov ebx,esi
		975	dec ebx
		976	and ebx,TEXTURE_SIZE
		977	add ebx,.bmap
		978	movzx eax,byte [ebx]
		979
		980	mov ebx,esi
		981	inc ebx
		982	and ebx,TEXTURE_SIZE
		983	add ebx,.bmap
		984	movzx ebx,byte [ebx]
		985	;; else ;-------------------------------------------
		986	;; mov dword [.temp1],esi ;-
		987	;; mov dword [.temp1+4],esi ;-
		988	;; movq mm5, qword[.temp1] ;-
		989	;; paddd mm5, qword[.temp2] ; .temp2 == low dword = 1, high dword = -1
		990	;; pand mm5, qword[.temp3] ; .temp3 == low = high dword = TEX_SIZE
		991	;; paddd mm5, qword[.temp4] ; .temp4 == low = high dword = .bmap
		992	;; movd ebx,mm5
		993	;; psrlq mm5,32
		994	;; movd eax,mm5
		995	;; movzx ebx,byte[ebx]
		996	;; movzx eax,byte[eax]
		997	;; end if
		998	sub eax,ebx
		999	;; if Ext=NON
		1000	mov ebx,esi
		1001	sub ebx,TEX_X
		1002	and ebx,TEXTURE_SIZE
		1003	add ebx,.bmap
		1004	movzx edx,byte [ebx]
		1005
		1006	mov ebx,esi
		1007	add ebx,TEX_X
		1008	and ebx,TEXTURE_SIZE
		1009	add ebx,.bmap
		1010	movzx ebx,byte [ebx]
		1011	;; else
		1012	;; movq mm5, qword[.temp1] ;-
		1013	;; paddd mm5, qword[.temp5] ; .temp5 == low dword = TEX_X, high dword = -TEX_X
		1014	;; pand mm5, qword[.temp3] ; .temp3 == low = high dword = TEX_SIZE
		1015	;; paddd mm5, qword[.temp4] ; .temp4 == low = high dword = offset .bmap
		1016	;; movd ebx,mm5
		1017	;; psrlq mm5,32
		1018	;; movd edx,mm5
		1019	;; movzx ebx,byte[ebx]
		1020	;; movzx edx,byte[edx]
		1021	;; end if
		1022	sub edx,ebx
		1023	; eax - horizontal sub
		1024	; edx - vertical sub
		1025	;; if Ext=NON
		1026	mov ebx,.cex ;.cex - current env map X
		1027	sar ebx,ROUND
		1028	add eax,ebx ; eax - modified x coord
		1029
		1030	mov ebx,.cey ;.cey - current env map y
		1031	sar ebx,ROUND
		1032	add edx,ebx ; edx - modified y coord
		1033	;; else
		1034	;; movq mm5,mm1 ; mm5 - copy of cur env coords
		1035	;; psrad mm5,ROUND
		1036	;; movq qword[.temp1],mm5
		1037	;; add eax,dword [.temp1]
		1038	;; add edx,dword [.temp1+4]
		1039	;; ; movd ebx,mm5
		1040	;; ; add eax,ebx
		1041	;; ; psrlq mm5,32
		1042	;; ; movd ebx,mm5
		1043	; add edx,ebx
		1044	;; end if
		1045
		1046	or eax,eax
		1047	jl .black
		1048	cmp eax,TEX_X
		1049	jg .black
		1050	or edx,edx
		1051	jl .black
		1052	cmp edx,TEX_Y
		1053	jg .black
		1054
		1055	shl edx,TEX_SHIFT
		1056	add edx,eax
		1057	lea edx,[edx*3]
		1058	add edx,.emap
		1059	mov eax,dword[edx]
		1060	jmp .put_pixel
		1061	.black:
		1062	xor eax,eax
		1063	.put_pixel:
		1064	stosd
		1065	dec edi
		1066	mov ebx,.cz
		1067	mov esi,.czbuff
		1068	mov dword[esi],ebx
		1069	jmp .no_skip
		1070	.skip:
		1071	add edi,3
		1072	.no_skip:
		1073	add .czbuff,4
		1074	;; if Ext = NON
		1075	mov eax,.dbx
		1076	add .cbx,eax
		1077	mov eax,.dby
		1078	add .cby,eax
		1079	mov eax,.dex
		1080	add .cex,eax
		1081	mov eax,.dey
		1082	add .cey,eax
		1083	;; else
		1084	;; paddd mm0,mm2
		1085	;; paddd mm1,mm3
		1086	;; end if
		1087	mov eax,.dz
		1088	add .cz,eax
		1089
		1090	dec ecx
		1091	jnz .draw
		1092	; end if
		1093	.bl_end:
		1094	mov esp,ebp
		1095	ret 56

Subversion Repositories Kolibri OS

(root)/programs/demos/3DS/BUMP_CAT.INC – Rev 1245