WebSVN – Kolibri OS – Blame – /programs/demos/view3ds/3r_phg.inc

Rev	Author	Line No.	Line
9237	leency	1	; Real Phong's shading implemented if flat assembler
		2	; by Maciej Guba.
		3	; http://macgub.co.pl
		4
		5	ROUND2 equ 10
		6	real_phong_tri_z:
		7	;----procedure render Phongs shaded triangle with z coord
		8	;----interpolation ( Catmull alghoritm )-----------------
		9	;----I normalize normal vector in every pixel -----------
		10	;------------------in - eax - x1 shl 16 + y1 ------------
		11	;---------------------- ebx - x2 shl 16 + y2 ------------
		12	;---------------------- ecx - x3 shl 16 + y3 ------------
		13	;---------------------- esi - pointer to Z-buffer filled-
		14	;---------------------- with dd float variables--------
		15	;---------------------- edi - pointer to screen buffer---
		16	;---------------------- xmm0 - 1st normal vector --------
		17	;---------------------- xmm1 - 2cond normal vector ------
		18	;---------------------- xmm2 - 3rd normal vector --------
		19	;---------------------- xmm3 - normalized light vector --
		20	;---------------------- xmm4 - lo -> hi z1, z2, z3 coords
		21	;---------------------- as dwords floats ---------------
		22	;---------------------- xmm5 - lo -> hi y_min, y_max, ---
		23	;---------------------- x_min, x_max as dword integers -
		24	;---------------------- stack - no parameters -----------
		25	;--------------------------------------------------------
		26	;----------------- procedure don't save registers !! ----
		27
		28
		29
		30
		31	push ebp
		32	mov ebp,esp
		33	sub esp,512
		34	sub ebp,16
		35	and ebp,0xfffffff0
		36
		37	.1_nv equ [ebp-16]
		38	.2_nv equ [ebp-32]
		39	.3_nv equ [ebp-48]
		40	.l_v equ [ebp-64]
		41	.z3 equ [ebp-72]
		42	.z2 equ [ebp-76]
		43	.z1 equ [ebp-80]
		44	.x1 equ [ebp-82]
		45	.y1 equ [ebp-84]
		46	.x2 equ [ebp-86]
		47	.y2 equ [ebp-88]
		48	.x3 equ [ebp-90]
		49	.y3 equ [ebp-92]
		50	.Zbuf equ [ebp-96]
		51	.x_max equ [ebp-100]
		52	.x_min equ [ebp-104]
		53	.y_max equ [ebp-108]
		54	.y_min equ [ebp-112]
		55	.screen equ [ebp-116]
		56	.dx12 equ [ebp-120]
		57	.dx13 equ [ebp-124]
		58	.dx23 equ [ebp-128]
		59	.dn12 equ [ebp-144]
		60	.dn13 equ [ebp-160]
		61	.dn23 equ [ebp-176]
		62	.dz12 equ [ebp-180]
		63	.dz13 equ [ebp-184]
		64	.dz23 equ [ebp-188]
		65
		66	.cnv1 equ [ebp-208] ; cur normal vectors
		67	.cnv2 equ [ebp-224]
		68	.cz2 equ [ebp-228]
		69	.cz1 equ [ebp-232]
		70
		71
		72
		73
		74
		75	.sort3: ; sort triangle coordinates...
		76	cmp ax,bx
		77	jle .sort1
		78	xchg eax,ebx
		79	shufps xmm4,xmm4,11100001b
		80	movaps xmm6,xmm0
		81	movaps xmm0,xmm1
		82	movaps xmm1,xmm6
		83
		84
		85	.sort1:
		86	cmp bx,cx
		87	jle .sort2
		88	xchg ebx,ecx
		89	shufps xmm4,xmm4,11011000b
		90	movaps xmm6,xmm1
		91	movaps xmm1,xmm2
		92	movaps xmm2,xmm6
		93
		94	jmp .sort3
		95
		96	.sort2:
		97
		98	movaps .z1,xmm4
		99	mov .y1,eax
		100	mov .y2,ebx
		101	mov .y3,ecx
		102
		103	movdqa .y_min,xmm5
		104	if 1 ; check if at last only fragment
		105	packssdw xmm5,xmm5 ; of triangle is in visable area
		106	pshuflw xmm5,xmm5,11011000b
		107	movdqu xmm7,.y3
		108	movdqa xmm6,xmm5
		109	pshufd xmm5,xmm5,0 ; xmm5 lo-hi -> broadcasted y_min, x_min
		110	pshufd xmm6,xmm6,01010101b ;xmm6 -> brd y_max x_max
		111	movdqa xmm4,xmm7
		112	pcmpgtw xmm7,xmm5
		113	pcmpgtw xmm4,xmm6
		114	pxor xmm7,xmm4
		115	pmovmskb eax,xmm7
		116	and eax,0x00aaaaaa
		117	or eax,eax
		118	jz .rpt_loop2_end
		119	end if
		120	movaps .1_nv,xmm0
		121	movaps .2_nv,xmm1
		122	movaps .3_nv,xmm2
		123	movaps .l_v,xmm3
		124	mov .Zbuf,esi
		125	mov .screen,edi
		126
		127
		128
		129	mov bx,.y2 ; calc deltas
		130	sub bx,.y1
		131	jnz .rpt_dx12_make
		132
		133	xorps xmm7,xmm7
		134	mov dword .dx12,0
		135	mov dword .dz12,0
		136	movaps .dn12,xmm7
		137	jmp .rpt_dx12_done
		138
		139	.rpt_dx12_make:
		140	mov ax,.x2
		141	sub ax,.x1
		142	cwde
		143	movsx ebx,bx
		144	shl eax,ROUND2
		145	cdq
		146	idiv ebx
		147	mov .dx12,eax
		148
		149	cvtsi2ss xmm6,ebx
		150	movss xmm5,.z2
		151	subss xmm5,.z1
		152	divss xmm5,xmm6
		153	movss .dz12,xmm5
		154
		155	movaps xmm0,.2_nv
		156	subps xmm0,.1_nv
		157	shufps xmm6,xmm6,0
		158	divps xmm0,xmm6
		159	movaps .dn12,xmm0
		160
		161
		162	.rpt_dx12_done:
		163
		164	mov bx,.y3 ; calc deltas
		165	sub bx,.y1
		166	jnz .rpt_dx13_make
		167
		168	xorps xmm7,xmm7
		169	mov dword .dx13,0
		170	mov dword .dz13,0
		171	movaps .dn13,xmm7
		172	jmp .rpt_dx13_done
		173
		174	.rpt_dx13_make:
		175	mov ax,.x3
		176	sub ax,.x1
		177	cwde
		178	movsx ebx,bx
		179	shl eax,ROUND2
		180	cdq
		181	idiv ebx
		182	mov .dx13,eax
		183
		184	cvtsi2ss xmm6,ebx
		185	movss xmm5,.z3
		186	subss xmm5,.z1
		187	divss xmm5,xmm6
		188	movss .dz13,xmm5
		189
		190	movaps xmm0,.3_nv
		191	subps xmm0,.1_nv
		192	shufps xmm6,xmm6,0
		193	divps xmm0,xmm6
		194	movaps .dn13,xmm0
		195
		196	.rpt_dx13_done:
		197
		198	mov bx,.y3 ; calc deltas
		199	sub bx,.y2
		200	jnz .rpt_dx23_make
		201
		202	xorps xmm7,xmm7
		203	mov dword .dx23,0
		204	mov dword .dz23,0
		205	movaps .dn23,xmm7
		206	jmp .rpt_dx23_done
		207
		208	.rpt_dx23_make:
		209	mov ax,.x3
		210	sub ax,.x2
		211	cwde
		212	movsx ebx,bx
		213	shl eax,ROUND2
		214	cdq
		215	idiv ebx
		216	mov .dx23,eax
		217
		218	cvtsi2ss xmm6,ebx
		219	movss xmm5,.z3
		220	subss xmm5,.z2
		221	divss xmm5,xmm6
		222	movss .dz23,xmm5
		223
		224	movaps xmm0,.3_nv
		225	subps xmm0,.2_nv
		226	shufps xmm6,xmm6,0
		227	divps xmm0,xmm6
		228	movaps .dn23,xmm0
		229
		230	.rpt_dx23_done:
		231
		232
		233	movsx eax,word .x1
		234	shl eax,ROUND2
		235	mov ebx,eax
		236	mov edx,.z1
		237	mov .cz1,edx
		238	mov .cz2,edx
		239	movaps xmm0,.1_nv
		240	movaps .cnv1,xmm0
		241	movaps .cnv2,xmm0
		242
		243
		244	movsx ecx,word .y1
		245	cmp cx,.y2
		246
		247	jge .rpt_loop1_end
		248
		249	.rpt_loop1:
		250	pushad
		251
		252	movaps xmm2,.y_min
		253	movaps xmm0,.cnv1
		254	movaps xmm1,.cnv2
		255	movlps xmm3,.cz1
		256	movaps xmm4,.l_v
		257	sar ebx,ROUND2
		258	sar eax,ROUND2
		259	mov edi,.screen
		260	mov esi,.Zbuf
		261
		262	call real_phong_line_z
		263
		264	popad
		265	movaps xmm0,.cnv1
		266	movaps xmm1,.cnv2
		267	movss xmm2,.cz1
		268	movss xmm3,.cz2
		269	addps xmm0,.dn13
		270	addps xmm1,.dn12
		271	addss xmm2,.dz13
		272	addss xmm3,.dz12
		273	add eax,.dx13
		274	add ebx,.dx12
		275
		276	movaps .cnv1,xmm0
		277	movaps .cnv2,xmm1
		278	movss .cz1,xmm2
		279	movss .cz2,xmm3
		280
		281	add ecx,1
		282	cmp cx,.y2
		283	jl .rpt_loop1
		284
		285
		286
		287
		288
		289	.rpt_loop1_end:
		290	movsx ecx,word .y2
		291	cmp cx,.y3
		292	jge .rpt_loop2_end
		293
		294	movsx ebx,word .x2 ; eax - cur x1
		295	shl ebx,ROUND2 ; ebx - cur x2
		296	push dword .z2
		297	pop dword .cz2
		298	movaps xmm0,.2_nv
		299	movaps .cnv2,xmm0
		300
		301
		302	.rpt_loop2:
		303	pushad
		304
		305	movaps xmm2,.y_min
		306	movaps xmm0,.cnv1
		307	movaps xmm1,.cnv2
		308	movlps xmm3,.cz1
		309	movaps xmm4,.l_v
		310	sar ebx,ROUND2
		311	sar eax,ROUND2
		312	mov edi,.screen
		313	mov esi,.Zbuf
		314
		315	call real_phong_line_z
		316
		317	popad
		318	movaps xmm0,.cnv1
		319	movaps xmm1,.cnv2
		320	movss xmm2,.cz1
		321	movss xmm3,.cz2
		322	addps xmm0,.dn13
		323	addps xmm1,.dn23
		324	addss xmm2,.dz13
		325	addss xmm3,.dz23
		326	add eax,.dx13
		327	add ebx,.dx23
		328
		329	movaps .cnv1,xmm0
		330	movaps .cnv2,xmm1
		331	movss .cz1,xmm2
		332	movss .cz2,xmm3
		333
		334	add ecx,1
		335	cmp cx,.y3
		336	jl .rpt_loop2
		337
		338	.rpt_loop2_end:
		339
		340	add esp,512
		341	pop ebp
		342
		343	ret
		344	real_phong_line_z:
		345	; in:
		346	; xmm0 - normal vector 1
		347	; xmm1 - normal vect 2
		348	; xmm3 - lo -> hi z1, z2 coords as dwords floats
		349	; xmm2 - lo -> hi y_min, y_max, x_min, x_max
		350	; as dword integers
		351	; xmm4 - normalized light vector
		352	; eax - x1
		353	; ebx - x2
		354	; ecx - y
		355	; edi - screen buffer
		356	; esi - z buffer filled with dd floats
		357
		358	push ebp
		359	mov ebp,esp
		360	sub esp,160
		361	sub ebp,16
		362	and ebp,0xfffffff0
		363
		364	.n1 equ [ebp-16]
		365	.n2 equ [ebp-32]
		366	.lv equ [ebp-48]
		367	.lx1 equ [ebp-52]
		368	.lx2 equ [ebp-56]
		369	.z2 equ [ebp-60]
		370	.z1 equ [ebp-64]
		371	.screen equ [ebp-68]
		372	.zbuff equ [ebp-72]
		373	.x_max equ [ebp-74]
		374	.x_min equ [ebp-76]
		375	.y_max equ [ebp-78]
		376	.y_min equ [ebp-80]
		377	.dn equ [ebp-96]
		378	.dz equ [ebp-100]
		379	.y equ [ebp-104]
		380	.cnv equ [ebp-128]
		381
		382	mov .y,ecx
		383	packssdw xmm2,xmm2
		384	movq .y_min,xmm2
		385	cmp cx,.y_min
		386	jl .end_rp_line
		387	cmp cx,.y_max
		388	jge .end_rp_line ;
		389
		390	cmp eax,ebx
		391	je .end_rp_line
		392	jl @f
		393	xchg eax,ebx
		394	movaps xmm7,xmm0
		395	movaps xmm0,xmm1
		396	movaps xmm1,xmm7
		397	shufps xmm3,xmm3,11100001b
		398	@@:
		399
		400	cmp ax,.x_max
		401	jge .end_rp_line
		402	cmp bx,.x_min
		403	jle .end_rp_line
		404	movaps .lv,xmm4
		405	movaps .n1,xmm0
		406	movaps .n2,xmm1
		407	mov .lx1,eax
		408	mov .lx2,ebx
		409	movlps .z1,xmm3
		410
		411	sub ebx,eax
		412	cvtsi2ss xmm7,ebx
		413	shufps xmm7,xmm7,0
		414	subps xmm1,xmm0
		415	divps xmm1,xmm7
		416	movaps .dn,xmm1
		417	psrldq xmm3,4
		418	subss xmm3,.z1
		419	divss xmm3,xmm7
		420	movss .dz,xmm3
		421
		422
		423
		424	mov ebx,.lx1
		425	cmp bx,.x_min ; clipping on function4
		426	jge @f
		427	movzx eax,word .x_min
		428	sub eax,ebx
		429	cvtsi2ss xmm7,eax
		430	shufps xmm7,xmm7,0
		431	mulss xmm3,xmm7
		432	mulps xmm1,xmm7
		433	addss xmm3,.z1
		434	addps xmm1,.n1
		435	movsx eax,word .x_min
		436	movss .z1,xmm3
		437	movaps .n1,xmm1
		438	mov dword .lx1,eax
		439
		440	@@:
		441	movzx eax,word .x_max
		442	cmp .lx2,eax
		443	jl @f
		444	mov .lx2,eax
		445	@@:
		446	movzx eax,word[size_x_var]
		447	mul dword .y
		448	; mov edx,.x1
		449	add eax,.lx1
		450	shl eax,2
		451	add edi,eax
		452	add esi,eax
		453
		454	mov ecx,.lx2
		455	sub ecx,.lx1
		456	movaps xmm0,.n1
		457	movss xmm2,.z1
9512	IgorA	458
9237	leency	459	.ddraw:
		460	movss xmm7,xmm2
		461	cmpnltss xmm7,dword[esi]
		462	movd eax,xmm7
		463	or eax,eax
		464	jnz .skip
		465	movss [esi],xmm2
		466	movaps xmm7,xmm0
		467	mulps xmm7,xmm7 ; normalize
		468	haddps xmm7,xmm7
		469	haddps xmm7,xmm7
		470	rsqrtps xmm7,xmm7
		471	mulps xmm7,xmm0
		472	movaps .cnv,xmm7
		473
		474	mov edx,lights_aligned ; lights - global variable
		475	xorps xmm1,xmm1 ; instead global can be used .lv - light vect.
		476	@@:
		477	movaps xmm6,[edx+16]
		478	movaps xmm5,[edx]
		479	movaps xmm3,[edx+48]
		480	andps xmm5,[zero_hgst_dd] ; global
		481
		482	mulps xmm5,.cnv ;.lv ; last dword should be zeroed
		483	haddps xmm5,xmm5
		484	haddps xmm5,xmm5
		485	; mulps xmm5,[env_const2]
		486	; maxps xmm5,[dot_min]
		487	; minps xmm5,[dot_max]
		488	movaps xmm7,xmm5
		489	; mulps xmm7,[env_const2]
		490	; mulps xmm7,[env_const2]
		491	; maxps xmm7,[dot_min]
		492	; minps xmm7,[dot_max]
		493
		494	mulps xmm7,xmm7
		495	mulps xmm7,xmm7
		496	mulps xmm5,xmm6
		497	mulps xmm7,xmm7
		498	mulps xmm7,xmm3
		499
		500	addps xmm5,xmm7
		501	minps xmm5,[mask_255f] ; global
		502	maxps xmm1,xmm5
		503	; movq xmm3,[edx+20] ; minimal color
		504	; punpcklwd xmm3,[minimum0]
		505	; cvtdq2ps xmm3,xmm3
		506	; maxps xmm1,xmm3
		507	add edx,64
		508	cmp edx,lights_aligned_end ; global
		509	jnz @b
		510
		511	cvtps2dq xmm1,xmm1
		512	packssdw xmm1,xmm1
		513	packuswb xmm1,xmm1
		514	movd [edi],xmm1
		515	.skip:
		516	add edi,4
		517	add esi,4
		518	addps xmm0,.dn
		519	addss xmm2,.dz
		520	sub ecx,1
		521	jnz .ddraw
		522
		523	.end_rp_line:
		524	add esp,160
		525	pop ebp
		526
		527	ret

Subversion Repositories Kolibri OS

(root)/programs/demos/view3ds/3r_phg.inc – Rev 9512