fix in some vertex functions

git-svn-id: svn://kolibrios.org@5175 a494cfbc-eb01-0410-851d-a64ba20cac60

programs/develop/libraries/TinyGL/asm_fork/clip.asm

@@ -169,29 +169,29 @@ proc interpolate uses eax ebx ecx, q:dword,p0:dword,p1:dword,t:dword
 	fld dword[t]
 
 	; интерполяция по координатам
-	fld dword[ecx+offs_vert_pc]
+	fld dword[ecx+offs_vert_pc+offs_X]
-	fsub dword[ebx+offs_vert_pc]
+	fsub dword[ebx+offs_vert_pc+offs_X]
 	fmul st0,st1
-	fadd dword[ebx+offs_vert_pc]
+	fadd dword[ebx+offs_vert_pc+offs_X]
-	fstp dword[eax+offs_vert_pc]
+	fstp dword[eax+offs_vert_pc+offs_X]
 
-	fld dword[ecx+offs_vert_pc+4]
+	fld dword[ecx+offs_vert_pc+offs_Y]
-	fsub dword[ebx+offs_vert_pc+4]
+	fsub dword[ebx+offs_vert_pc+offs_Y]
 	fmul st0,st1
-	fadd dword[ebx+offs_vert_pc+4]
+	fadd dword[ebx+offs_vert_pc+offs_Y]
-	fstp dword[eax+offs_vert_pc+4]
+	fstp dword[eax+offs_vert_pc+offs_Y]
 
-	fld dword[ecx+offs_vert_pc+8]
+	fld dword[ecx+offs_vert_pc+offs_Z]
-	fsub dword[ebx+offs_vert_pc+8]
+	fsub dword[ebx+offs_vert_pc+offs_Z]
 	fmul st0,st1
-	fadd dword[ebx+offs_vert_pc+8]
+	fadd dword[ebx+offs_vert_pc+offs_Z]
-	fstp dword[eax+offs_vert_pc+8]
+	fstp dword[eax+offs_vert_pc+offs_Z]
 
-	fld dword[ecx+offs_vert_pc+12]
+	fld dword[ecx+offs_vert_pc+offs_W]
-	fsub dword[ebx+offs_vert_pc+12]
+	fsub dword[ebx+offs_vert_pc+offs_W]
 	fmul st0,st1
-	fadd dword[ebx+offs_vert_pc+12]
+	fadd dword[ebx+offs_vert_pc+offs_W]
-	fstp dword[eax+offs_vert_pc+12]
+	fstp dword[eax+offs_vert_pc+offs_W]
 
 	; интерполяция по цвету
 	fld dword[ecx+offs_vert_color]
@@ -531,17 +531,17 @@ endl
 	mov edx,[a]
 	mov ebx,[b]
 	mov ecx,[c]
-	fld dword[ebx]
+	fld dword[ebx+offs_X]
-	fsub dword[edx]
+	fsub dword[edx+offs_X]
 	fstp dword[d_X] ;d_X = (b.X - a.X)
-	fld dword[ebx+4]
+	fld dword[ebx+offs_Y]
-	fsub dword[edx+4]
+	fsub dword[edx+offs_Y]
 	fstp dword[d_Y] ;d_Y = (b.Y - a.Y)
-	fld dword[ebx+8]
+	fld dword[ebx+offs_Z]
-	fsub dword[edx+8]
+	fsub dword[edx+offs_Z]
 	fstp dword[d_Z] ;d_Z = (b.Z - a.Z)
-	fld dword[ebx+12]
+	fld dword[ebx+offs_W]
-	fsub dword[edx+12]
+	fsub dword[edx+offs_W]
 	fst dword[d_W] ;d_W = (b.W - a.W)
 if sign eq 0
 	fadd dword[d#dir]
@@ -550,15 +550,13 @@ else
 end if
 
 	fldz
-	fcomp st1
+	fcompp
 	fstsw ax
 	sahf
 	ja @f
-		fincstp
 		fst dword[t] ;t=0
 		jmp .e_zero
 	@@: ;else
-		fincstp
 		fld dword[edx+offs#dir]
 if sign eq 0		
 		fchs
@@ -628,115 +626,189 @@ endp
 align 4
 clip_proc dd clip_xmin,clip_xmax, clip_ymin,clip_ymax, clip_zmin,clip_zmax
 
-;static inline void updateTmp(GLContext *c, GLVertex *q,GLVertex *p0,GLVertex *p1,float t)
+align 4
-;{
+proc updateTmp uses edx, context:dword, q:dword, p0:dword, p1:dword, t:dword
-;  if (c->current_shade_model == GL_SMOOTH) {
+	mov eax,[context]
+	cmp dword[edx+offs_cont_current_shade_model],GL_SMOOTH ;if (context.current_shade_model == GL_SMOOTH)
+	jne .els_0
 ;    q->color.v[0]=p0->color.v[0] + (p1->color.v[0]-p0->color.v[0])*t;
 ;    q->color.v[1]=p0->color.v[1] + (p1->color.v[1]-p0->color.v[1])*t;
 ;    q->color.v[2]=p0->color.v[2] + (p1->color.v[2]-p0->color.v[2])*t;
-;  } else {
+		jmp @f
+	.els_0:
 ;    q->color.v[0]=p0->color.v[0];
 ;    q->color.v[1]=p0->color.v[1];
 ;    q->color.v[2]=p0->color.v[2];
-;  }
+	@@:
 
-;  if (c->texture_2d_enabled) {
+	cmp dword[edx+offs_cont_texture_2d_enabled],0 ;if (context.texture_2d_enabled)
+	je @f
 ;    q->tex_coord.X=p0->tex_coord.X + (p1->tex_coord.X-p0->tex_coord.X)*t;
 ;    q->tex_coord.Y=p0->tex_coord.Y + (p1->tex_coord.Y-p0->tex_coord.Y)*t;
-;  }
+	@@:
 
 ;  q->clip_code=gl_clipcode(q->pc.X,q->pc.Y,q->pc.Z,q->pc.W);
 ;  if (q->clip_code==0){
 ;    gl_transform_to_viewport(c,q);
 ;    RGBFtoRGBI(q->color.v[0],q->color.v[1],q->color.v[2],q->zp.r,q->zp.g,q->zp.b);
 ;  }
-;}
+	ret
+endp
 
-;static void gl_draw_triangle_clip(GLContext *c, GLVertex *p0,GLVertex *p1,GLVertex *p2,int clip_bit);
+align 4
+proc gl_draw_triangle, context:dword, p0:dword, p1:dword, p2:dword
+locals
+	cc rd 3
+	front dd ?
+	norm dd ? ;float
+endl
+pushad
+	mov ebx,[p0]
+	mov ecx,[p1]
+	mov edx,[p2]
+	
+	mov edi,[ebx+offs_vert_clip_code]
+	mov dword[cc],edi
+	mov eax,[ecx+offs_vert_clip_code]
+	mov dword[cc+4],eax
+	or edi,eax
+	mov eax,[edx+offs_vert_clip_code]
+	mov dword[cc+8],eax
+	or edi,eax
 
-;void gl_draw_triangle(GLContext *c, GLVertex *p0,GLVertex *p1,GLVertex *p2)
+	; we handle the non clipped case here to go faster
-;{
+	cmp edi,0
-;  int co,c_and,cc[3],front;
+	jne .els_0
-;  float norm;
+		mov edi,dword[edx+offs_vert_zp+offs_zbup_x]
-;  
+		mov edi,dword[ebx+offs_vert_zp+offs_zbup_x]
-;  cc[0]=p0->clip_code;
+		mov dword[norm],edi
-;  cc[1]=p1->clip_code;
+		fild dword[norm]
-;  cc[2]=p2->clip_code;
+		mov esi,dword[ecx+offs_vert_zp+offs_zbup_y]
-;  
+		mov esi,dword[ebx+offs_vert_zp+offs_zbup_y]
-;  co=cc[0] | cc[1] | cc[2];
+		mov dword[norm],edi
-;
+		fimul dword[norm]
-;  /* we handle the non clipped case here to go faster */
+		mov edi,dword[ecx+offs_vert_zp+offs_zbup_x]
-;  if (co==0) {
+		sub edi,dword[ebx+offs_vert_zp+offs_zbup_x]
-;    
+		mov dword[norm],edi
-;      norm=(float)(p1->zp.x-p0->zp.x)*(float)(p2->zp.y-p0->zp.y)-
+		fild dword[norm]
-;        (float)(p2->zp.x-p0->zp.x)*(float)(p1->zp.y-p0->zp.y);
+		mov edi,dword[edx+offs_vert_zp+offs_zbup_y]
-;      
+		mov edi,dword[ebx+offs_vert_zp+offs_zbup_y]
-;      if (norm == 0) return;
+		mov dword[norm],edi
-;
+		fimul dword[norm]
-;      front = norm < 0.0;
+		fsub st0,st1
-;      front = front ^ c->current_front_face;
-;  
-;      /* back face culling */
-;      if (c->cull_face_enabled) {
-;        /* most used case first */
-;        if (c->current_cull_face == GL_BACK) {
-;          if (front == 0) return;
-;          c->draw_triangle_front(c,p0,p1,p2);
-;        } else if (c->current_cull_face == GL_FRONT) {
-;          if (front != 0) return;
-;          c->draw_triangle_back(c,p0,p1,p2);
-;        } else {
-;          return;
-;        }
-;      } else {
-;        /* no culling */
-;        if (front) {
-;          c->draw_triangle_front(c,p0,p1,p2);
-;        } else {
-;          c->draw_triangle_back(c,p0,p1,p2);
-;        }
-;      }
-;  } else {
-;    c_and=cc[0] & cc[1] & cc[2];
-;    if (c_and==0) {
-;      gl_draw_triangle_clip(c,p0,p1,p2,0);
-;    }
-;  }
-;}
 
-;static void gl_draw_triangle_clip(GLContext *c, GLVertex *p0,GLVertex *p1,GLVertex *p2,int clip_bit)
+		;st0 = (p1.zp.x-p0.zp.x)*(p2.zp.y-p0.zp.y) - (p2.zp.x-p0.zp.x)*(p1.zp.y-p0.zp.y)
-;{
-;  int co,c_and,co1,cc[3],edge_flag_tmp,clip_mask;
-;  GLVertex tmp1,tmp2,*q[3];
-;  float tt;
 
-;  cc[0]=p0->clip_code;
+		mov dword[front],0
-;  cc[1]=p1->clip_code;
+		fldz
-;  cc[2]=p2->clip_code;
+		fcom st1
+		fstsw ax
+		sahf
+		je .end_f
+		jb @f
+			inc dword[front] ;front = 0.0 > norm
+		@@:
+		mov edi,[context]
+		mov eax,dword[edi+offs_cont_current_front_face]
+		xor dword[front],eax ;front ^= context.current_front_face
 
-;  co=cc[0] | cc[1] | cc[2];
+		; back face culling
-;  if (co == 0) {
+		cmp dword[edi+offs_cont_cull_face_enabled],0
-;    gl_draw_triangle(c,p0,p1,p2);
+		je .els_1
-;  } else {
+			; most used case first
-;    c_and=cc[0] & cc[1] & cc[2];
+			cmp dword[edi+offs_cont_current_cull_face],GL_BACK
-;    /* the triangle is completely outside */
+			jne @f
-;    if (c_and!=0) return;
+				cmp dword[front],0
+				je .end_f
+					stdcall dword[edi+offs_cont_draw_triangle_front], edi,ebx,ecx,edx
+				jmp .end_f
+			@@:
+			cmp dword[edi+offs_cont_current_cull_face],GL_FRONT
+			jne .end_f
+				cmp dword[front],0
+				jne .end_f
+					stdcall dword[edi+offs_cont_draw_triangle_back], edi,ebx,ecx,edx
+			jmp .end_f
+		.els_1:
+			; no culling
+			cmp dword[front],0
+			je @f
+				stdcall dword[edi+offs_cont_draw_triangle_front], edi,ebx,ecx,edx
+			@@:
+				stdcall dword[edi+offs_cont_draw_triangle_back], edi,ebx,ecx,edx
+		jmp .end_f
+	.els_0:
+		and eax,[cc]
+		and eax,[cc+4]
+		cmp eax,0
+		jne @f
+			stdcall gl_draw_triangle_clip, [context],ebx,ecx,edx,0
+	.end_f:
+	finit
+	@@:
+popad
+	ret
+endp
 
-;    /* find the next direction to clip */
+align 4
+proc gl_draw_triangle_clip, context:dword, p0:dword, p1:dword, p2:dword, clip_bit:dword
+locals
+	co dd ?
+	co1 dd ?
+	cc rd 3
+	edge_flag_tmp dd ?
+	clip_mask dd ?
+	tmp1 GLVertex ?
+	tmp2 GLVertex ?
+	q rd 3 ;GLVertex*
+	tt dd ? ;float
+endl
+pushad
+
+	mov ebx,[p0]
+	mov ecx,[p1]
+	mov edx,[p2]
+	
+	mov edi,[ebx+offs_vert_clip_code]
+	mov dword[cc],edi
+	mov eax,[ecx+offs_vert_clip_code]
+	mov dword[cc+4],eax
+	or edi,eax
+	mov eax,[edx+offs_vert_clip_code]
+	mov dword[cc+8],eax
+	or edi,eax
+	mov dword[co],edi
+
+	cmp edi,0
+	jne .els_0
+		stdcall gl_draw_triangle, [context],ebx,ecx,edx
+		jmp .end_f
+	.els_0:
+		and eax,[cc]
+		and eax,[cc+4]
+
+		; the triangle is completely outside
+		cmp eax,0
+		jne .end_f
+
+	; find the next direction to clip
 ;    while (clip_bit < 6 && (co & (1 << clip_bit)) == 0)  {
 ;      clip_bit++;
 ;    }
 
-;    /* this test can be true only in case of rounding errors */
+	; this test can be true only in case of rounding errors
-;    if (clip_bit == 6) {
+	cmp dword[clip_bit],6
-;#if 0
+if 0
+	jne @f
 ;      printf("Error:\n");
 ;      printf("%f %f %f %f\n",p0->pc.X,p0->pc.Y,p0->pc.Z,p0->pc.W);
 ;      printf("%f %f %f %f\n",p1->pc.X,p1->pc.Y,p1->pc.Z,p1->pc.W);
 ;      printf("%f %f %f %f\n",p2->pc.X,p2->pc.Y,p2->pc.Z,p2->pc.W);
-;#endif
+		jmp .end_f
-;      return;
+	@@:
-;    }
+else
+	je .end_f
+end if
 
 ;    clip_mask = 1 << clip_bit;
 ;    co1=(cc[0] ^ cc[1] ^ cc[2]) & clip_mask;
@@ -781,7 +853,10 @@ clip_proc dd clip_xmin,clip_xmax, clip_ymin,clip_ymax, clip_zmin,clip_zmax
 ;      gl_draw_triangle_clip(c,q[0],&tmp1,&tmp2,clip_bit+1);
 ;    }
 ;  }
-;}
+	.end_f:
+popad
+	ret
+endp
 
 align 4
 proc gl_draw_triangle_select uses eax, context:dword, p0:dword,p1:dword,p2:dword
@@ -795,14 +870,16 @@ proc gl_draw_triangle_select uses eax, context:dword, p0:dword,p1:dword,p2:dword
 	ret
 endp
 
-;#ifdef PROFILE
+if PROFILE eq 1
-;int count_triangles,count_triangles_textured,count_pixels;
+	count_triangles dd ?
-;#endif
+	count_triangles_textured dd ?
+	count_pixels dd ?
+end if
 
 align 4
-proc gl_draw_triangle_fill uses eax edx, context:dword, p0:dword,p1:dword,p2:dword
+proc gl_draw_triangle_fill, context:dword, p0:dword,p1:dword,p2:dword
-;#ifdef PROFILE
+pushad
-;  {
+if PROFILE eq 1
 ;    int norm;
 ;    assert(p0->zp.x >= 0 && p0->zp.x < c->zb->xsize);
 ;    assert(p0->zp.y >= 0 && p0->zp.y < c->zb->ysize);
@@ -814,32 +891,40 @@ proc gl_draw_triangle_fill uses eax edx, context:dword, p0:dword,p1:dword,p2:dwo
 ;    norm=(p1->zp.x-p0->zp.x)*(p2->zp.y-p0->zp.y)-
 ;      (p2->zp.x-p0->zp.x)*(p1->zp.y-p0->zp.y);
 ;    count_pixels+=abs(norm)/2;
-;    count_triangles++;
+	inc dword[count_triangles]
-;  }
+end if
-;#endif
 
+	mov ebx,[p1]
+	add ebx,offs_vert_zp
+	mov ecx,[p2]
+	add ecx,offs_vert_zp
 	mov edx,[context]
 	cmp dword[edx+offs_cont_texture_2d_enabled],0
 	je .els_i
 		;if (context.texture_2d_enabled)
-;#ifdef PROFILE
+if PROFILE eq 1
-;    count_triangles_textured++;
+	inc dword[count_triangles_textured]
-;#endif
+end if
 		mov eax,dword[edx+offs_cont_current_texture]
 		mov eax,[eax] ;переход по указателю
 		;так как offs_text_images+offs_imag_pixmap = 0 то context.current_texture.images[0].pixmap = [eax]
 		stdcall ZB_setTexture, dword[edx+offs_cont_zb],dword[eax]
-;    ZB_fillTriangleMappingPerspective, dword[edx+offs_cont_zb],&p0->zp,&p1->zp,&p2->zp);
+		mov eax,[p0]
+		add eax,offs_vert_zp
+		stdcall ZB_fillTriangleMappingPerspective, dword[edx+offs_cont_zb],eax,ebx,ecx
 		jmp .end_f
 	.els_i:
-	cmp dword[edx+offs_cont_current_shade_model],GL_SMOOTH
+		mov eax,[p0]
-	jne .els
+		add eax,offs_vert_zp
-		;else if (context.current_shade_model == GL_SMOOTH)
+		cmp dword[edx+offs_cont_current_shade_model],GL_SMOOTH
-;    ZB_fillTriangleSmooth, dword[edx+offs_cont_zb],&p0->zp,&p1->zp,&p2->zp);
+		jne .els
-		jmp .end_f
+			;else if (context.current_shade_model == GL_SMOOTH)
-	.els:
+			stdcall ZB_fillTriangleSmooth, dword[edx+offs_cont_zb],eax,ebx,ecx
-;    ZB_fillTriangleFlat, dword[edx+offs_cont_zb],&p0->zp,&p1->zp,&p2->zp);
+			jmp .end_f
+		.els:
+			stdcall ZB_fillTriangleFlat, dword[edx+offs_cont_zb],eax,ebx,ecx
 	.end_f:
+popad
 	ret
 endp
 

programs/develop/libraries/TinyGL/asm_fork/matrix.asm

@@ -410,16 +410,18 @@ proc glopTranslate uses eax ebx ecx, context:dword, p:dword
 		fmul st0,st3     ;m[1] * y
 		fld dword[ebx+8] ;m[2]
 		fmul st0,st3     ;m[2] * z
-		fadd st0,st1
+		faddp
-		fadd st0,st2
+		faddp
 		fadd dword[ebx+12] ;m[3]
 		fstp dword[ebx+12] ;m[3] = m[0] * x + m[1] * y + m[2] * z + m[3]
-		ffree st0
-		fincstp
-		ffree st0
-		fincstp
 		add ebx,16
 	loop @b
+	ffree st0
+	fincstp
+	ffree st0
+	fincstp
+	ffree st0
+	fincstp
 
 if DEBUG ;glopTranslate
 	mov ebx,[eax+offs_cont_matrix_mode]

programs/develop/libraries/TinyGL/asm_fork/tinygl.asm

@@ -87,9 +87,9 @@ m_5 db '(5)',13,10,0
 buf_param rb 80
 
 align 4
-proc str_n_cat uses eax ecx edi esi, str1:dword, str2:dword, n:dword
+proc str_n_cat uses eax ecx edi esi, str1:dword, str2:dword, n_len:dword
 	mov esi,dword[str2]
-	mov ecx,dword[n]
+	mov ecx,dword[n_len]
 	mov edi,dword[str1]
 	stdcall str_len,edi
 	add edi,eax

programs/develop/libraries/TinyGL/asm_fork/vertex.asm

@@ -236,7 +236,6 @@ pushad
 	je .els_0
 		; eye coordinates needed for lighting
 		mov ebx,dword[eax+offs_cont_matrix_stack_ptr]
-		;;;mov edx,[v]
 		finit
 		fld dword[edx+offs_vert_coord+offs_X]
 		fld dword[edx+offs_vert_coord+offs_Y]
@@ -251,13 +250,9 @@ pushad
 			fld dword[ebx+8]   ;st0 = m[2]
 			fmul st0,st3       ;st0 *= v.coord.Z
 			fadd dword[ebx+12] ;st0 += m[3]
-			fadd st0,st1       ;st0 += v.coord.Z * m[2] 
+			faddp              ;st0 += v.coord.Z * m[2] 
-			fadd st0,st2       ;st0 += v.coord.Y * m[1] 
+			faddp              ;st0 += v.coord.Y * m[1] 
 			fstp dword[edx+offs_vert_ec] ;v.ec.X = v.coord.X * m[0] + v.coord.Y * m[1] + v.coord.Z * m[2] + m[3]
-			ffree st0
-			fincstp
-			ffree st0
-			fincstp
 			add ebx,16 ;следущая строка матрицы
 			add edx,4  ;следущая координата вектора
 		loop .cycle_0
@@ -279,13 +274,9 @@ pushad
 			fld dword[ebx+8]   ;st0 = m[2]
 			fmul st0,st3       ;st0 *= v.ec.Z
 			fadd dword[ebx+12] ;st0 += m[3]
-			fadd st0,st1       ;st0 += v.ec.Z * m[2]
+			faddp              ;st0 += v.ec.Z * m[2]
-			fadd st0,st2       ;st0 += v.ec.Y * m[1]
+			faddp              ;st0 += v.ec.Y * m[1]
 			fstp dword[edx+offs_vert_pc] ;v.pc.X = v.ec.X * m[0] + v.ec.Y * m[1] + v.ec.Z * m[2] + m[3]
-			ffree st0
-			fincstp
-			ffree st0
-			fincstp
 			add ebx,16 ;следущая строка матрицы
 			add edx,4  ;следущая координата вектора
 		loop .cycle_1
@@ -308,13 +299,9 @@ pushad
 			fmul st0,st3       ;st0 *= n.Y
 			fld dword[ebx+8]   ;st0 = m[2]
 			fmul st0,st3       ;st0 *= n.Z
-			fadd st0,st1       ;st0 += n.Z * m[2]
+			faddp              ;st0 += n.Z * m[2]
-			fadd st0,st2       ;st0 += n.Y * m[1]
+			faddp              ;st0 += n.Y * m[1]
 			fstp dword[edx+offs_vert_normal] ;v.normal.X = n.X * m[0] + n.Y * m[1] + n.Z * m[2]
-			ffree st0
-			fincstp
-			ffree st0
-			fincstp
 			add ebx,16 ;следущая строка матрицы
 			add edx,4  ;следущая координата вектора
 		loop .cycle_2
@@ -329,7 +316,6 @@ pushad
 		mov ebx,eax
 		add ebx,offs_cont_matrix_model_projection
 
-		;;;mov edx,[v]
 		finit
 		fld dword[edx+offs_vert_coord+offs_X]
 		fld dword[edx+offs_vert_coord+offs_Y]
@@ -345,13 +331,10 @@ pushad
 		fld dword[ebx+8]   ;st0 = m[2]
 		fmul st0,st3       ;st0 *= v.coord.Z
 		fadd dword[ebx+12] ;st0 += m[3]
-		fadd st0,st1       ;st0 += v.coord.Z * m[2]
+		faddp              ;st0 += v.coord.Z * m[2]
-		fadd st0,st2       ;st0 += v.coord.Y * m[1]
+		faddp              ;st0 += v.coord.Y * m[1]
 		fstp dword[esi]    ;v.pc.X = v.coord.X * m[0] + v.coord.Y * m[1] + v.coord.Z * m[2] + m[3]
-		ffree st0
+
-		fincstp
-		ffree st0
-		fincstp
 		fld dword[ebx+16]  ;st0 = m[4]
 		fmul st0,st3       ;st0 *= v.coord.X
 		fld dword[ebx+20]  ;st0 = m[5]
@@ -359,13 +342,10 @@ pushad
 		fld dword[ebx+24]  ;st0 = m[6]
 		fmul st0,st3       ;st0 *= v.coord.Z
 		fadd dword[ebx+28] ;st0 += m[7]
-		fadd st0,st1       ;st0 += v.coord.Z * m[6]
+		faddp              ;st0 += v.coord.Z * m[6]
-		fadd st0,st2       ;st0 += v.coord.Y * m[5]
+		faddp              ;st0 += v.coord.Y * m[5]
 		fstp dword[esi+4]  ;v.pc.X = v.coord.X * m[4] + v.coord.Y * m[5] + v.coord.Z * m[6] + m[7]
-		ffree st0
+
-		fincstp
-		ffree st0
-		fincstp
 		fld dword[ebx+32]  ;st0 = m[8]
 		fmul st0,st3       ;st0 *= v.coord.X
 		fld dword[ebx+36]  ;st0 = m[9]
@@ -373,8 +353,8 @@ pushad
 		fld dword[ebx+40]  ;st0 = m[10]
 		fmul st0,st3       ;st0 *= v.coord.Z
 		fadd dword[ebx+44] ;st0 += m[11]
-		fadd st0,st1       ;st0 += v.coord.Z * m[10]
+		faddp              ;st0 += v.coord.Z * m[10]
-		fadd st0,st2       ;st0 += v.coord.Y * m[9]
+		faddp              ;st0 += v.coord.Y * m[9]
 		fstp dword[esi+8]  ;v.pc.X = v.coord.X * m[8] + v.coord.Y * m[9] + v.coord.Z * m[10] + m[11]
 
 		cmp dword[eax+offs_cont_matrix_model_projection_no_w_transform],0
@@ -384,10 +364,6 @@ pushad
 			mov dword[esi+12],ebx ;v.pc.W = m[15]
 			jmp .end_els
 		.els_1:
-			ffree st0
-			fincstp
-			ffree st0
-			fincstp
 			fld dword[ebx+48]  ;st0 = m[12]
 			fmul st0,st3       ;st0 *= v.coord.X
 			fld dword[ebx+52]  ;st0 = m[13]
@@ -395,10 +371,17 @@ pushad
 			fld dword[ebx+56]  ;st0 = m[14]
 			fmul st0,st3       ;st0 *= v.coord.Z
 			fadd dword[ebx+60] ;st0 += m[15]
-			fadd st0,st1       ;st0 += v.coord.Z * m[14] 
+			faddp              ;st0 += v.coord.Z * m[14] 
-			fadd st0,st2       ;st0 += v.coord.Y * m[13] 
+			faddp              ;st0 += v.coord.Y * m[13] 
 			fstp dword[esi+12] ;v.pc.W = v.coord.X * m[12] + v.coord.Y * m[13] + v.coord.Z * m[14] + m[15]
 	.end_els:
+	ffree st0
+	fincstp
+	ffree st0
+	fincstp
+	ffree st0
+	fincstp
+
 if DEBUG ;gl_vertex_transform
 	stdcall dbg_print,f_vt,txt_nl
 	mov edx,[v]
@@ -509,8 +492,7 @@ pushad
 			mov eax,[edx+offs_cont_vertex]
 			push eax
 			add eax,sizeof.GLVertex
-			push eax
+			stdcall gl_draw_line, edx, eax
-			stdcall gl_draw_line, edx
 			xor eax,eax
 			mov dword[n],eax
 		jmp .end_f
@@ -534,8 +516,7 @@ pushad
 			mov eax,[edx+offs_cont_vertex]
 			push eax
 			add eax,sizeof.GLVertex
-			push eax
+			stdcall gl_draw_line, edx, eax
-			stdcall gl_draw_line, edx
 			mov edi,[edx+offs_cont_vertex]
 			mov esi,edi
 			add esi,sizeof.GLVertex
@@ -548,7 +529,12 @@ pushad
 	jne @f
 		cmp dword[n],3
 		jne .end_f
-;               gl_draw_triangle(c, &c->vertex[0], &c->vertex[1], &c->vertex[2]);
+			mov eax,[edx+offs_cont_vertex]
+			push eax
+			add eax,sizeof.GLVertex
+			push eax
+			add eax,sizeof.GLVertex
+			stdcall gl_draw_triangle, edx, eax
 			xor eax,eax
 			mov dword[n],eax
 		jmp .end_f

programs/develop/libraries/TinyGL/asm_fork/zgl.inc

@@ -2,19 +2,13 @@ include 'opengl_const.inc'
 include 'zbuffer.inc'
 include 'zmath.inc'
 
-;#define fputc(...) /*nothing*/
-;#define fprintf(...) /*nothing*/
-;#define vfprintf(...) /*nothing*/
-;#undef stderr
-;#define stderr ((FILE*)-1)
-
 
 ;enum { OP_ ## a , ... }
-s1 equ 0
+sum1 equ 0
 macro ADD_OP a,b,c
 {
-	OP_#a equ s1
+	OP_#a equ sum1
-	s1 equ (s1+1)
+	sum1 equ (sum1+1)
 }
 include 'opinfo.inc'
 

programs/develop/libraries/TinyGL/asm_fork/ztriangle.asm

@@ -7,13 +7,13 @@ if TGL_FEATURE_RENDER_BITS eq 24
 	mov ecx,[p2]
 	mov eax,[ecx+offs_zbup_r]
 	shr eax,8
-	mov [colorR],eax ;colorR=p2.r>>8
+	mov [colorR],al ;colorR=p2.r>>8
 	mov eax,[ecx+offs_zbup_g]
 	shr eax,8
-	mov [colorG],eax ;colorG=p2.g>>8
+	mov [colorG],al ;colorG=p2.g>>8
 	mov eax,[ecx+offs_zbup_b]
 	shr eax,8
-	mov [colorB],eax ;colorB=p2.b>>8
+	mov [colorB],al ;colorB=p2.b>>8
 else
 ;  color=RGB_TO_PIXEL(p2->r,p2->g,p2->b);
 end if
@@ -21,19 +21,27 @@ end if
 
 macro PUT_PIXEL _a
 {
-;    zz=z >> ZB_POINT_Z_FRAC_BITS;
+local .end_0
-;    if (zz >= pz[_a]) {
+	mov eax,[z]
+	shr eax, ZB_POINT_Z_FRAC_BITS
+	mov [zz],eax
+	mov ebx,[pz]
+	cmp ax,word[ebx+2*_a] ;if (zz >= pz[_a])
+	jl .end_0
+		;edi = pp
 if TGL_FEATURE_RENDER_BITS eq 24
-;      pp[3 * _a]=colorR;
+		mov cl,[colorR]
-;      pp[3 * _a + 1]=colorG;
+		mov ch,[colorG]
-;      pp[3 * _a + 2]=colorB;
+		mov word[edi+3*_a],cx
-;      pz[_a]=zz;
+		mov cl,[colorB]
+		mov byte[edi+3*_a +2],cl
 else
 ;      pp[_a]=color;
-;      pz[_a]=zz;
 end if
-;    }
+		mov word[ebx+2*_a],ax ;пишем в буфер глубины новое значение
-;    z+=dzdx;
+	.end_0:
+	mov eax,[dzdx]
+	add [z],eax
 }
 
 align 4
@@ -47,8 +55,6 @@ else
 	color dd ? ;int
 end if
 include 'ztriangle.inc'
-	ret
-endp
 
 ;
 ; Smooth filled triangle.
@@ -69,7 +75,9 @@ end if
 
 macro PUT_PIXEL _a
 {
-;    zz=z >> ZB_POINT_Z_FRAC_BITS;
+	mov eax,[z]
+	shr eax,ZB_POINT_Z_FRAC_BITS
+	mov [zz],eax
 if TGL_FEATURE_RENDER_BITS eq 24
 ;    if (zz >= pz[_a]) {
 ;      pp[3 * _a]=or1 >> 8;
@@ -77,7 +85,8 @@ if TGL_FEATURE_RENDER_BITS eq 24
 ;      pp[3 * _a + 2]=ob1 >> 8;
 ;      pz[_a]=zz;
 ;    }
-;    z+=dzdx;
+	mov eax,[dzdx]
+	add [z],eax
 ;    og1+=dgdx;
 ;    or1+=drdx;
 ;    ob1+=dbdx;
@@ -94,20 +103,26 @@ else
 ;      pp[_a] = RGB_TO_PIXEL(or1, og1, ob1);
 ;      pz[_a]=zz;
 ;    }
-;    z+=dzdx;
+	mov eax,[dzdx]
+	add [z],eax
 ;    og1+=dgdx;
 ;    or1+=drdx;
 ;    ob1+=dbdx;
 end if
 }
 
-macro DRAW_LINE
+DRAW_LINE_M equ 1
+
+macro DRAW_LINE code
 {
 if TGL_FEATURE_RENDER_BITS eq 16
+if code eq 0
 ;  register unsigned short *pz;
 ;  register PIXEL *pp;
 ;  register unsigned int tmp,z,zz,rgb,drgbdx;
 ;  register int n;
+end if
+if code eq 1
 ;  n=(x2 >> 16) - x1;
 ;  pp=pp1+x1;
 ;  pz=pz1+x1;
@@ -132,12 +147,134 @@ if TGL_FEATURE_RENDER_BITS eq 16
 ;    n-=1;
 ;  }
 end if
+end if
+}
+
+align 4
+proc ZB_fillTriangleSmooth, zb:dword, p0:dword, p1:dword, p2:dword
+locals
+if TGL_FEATURE_RENDER_BITS eq 16
+	_drgbdx dd ? ;int
+end if
+include 'ztriangle.inc'
+
+align 4
+proc ZB_setTexture uses eax ebx, zb:dword, texture:dword
+	mov eax,[zb]
+	mov ebx,[texture]
+	mov dword[eax+offs_zbuf_current_texture],ebx
+	ret
+endp
+
+INTERP_Z equ 1
+INTERP_ST equ 1
+
+macro DRAW_INIT
+{
+;  texture=zb->current_texture;
+}
+
+macro PUT_PIXEL _a
+{
+;   zz=z >> ZB_POINT_Z_FRAC_BITS;
 if TGL_FEATURE_RENDER_BITS eq 24
-;  register unsigned short *pz;
+;   unsigned char *ptr;
-;  register PIXEL *pp;
+;     if (zz >= pz[_a]) {
-;  register unsigned int s,t,z,zz;
+;       ptr = texture + (((t & 0x3FC00000) | s) >> 14) * 3;
-;  register int n,dsdx,dtdx;
+;       pp[3 * _a]= ptr[0];
-;  float sz,tz,fz,zinv;
+;       pp[3 * _a + 1]= ptr[1];
+;       pp[3 * _a + 2]= ptr[2];
+;       pz[_a]=zz;
+;    }
+else
+;     if (zz >= pz[_a]) {
+;       pp[_a]=texture[((t & 0x3FC00000) | s) >> 14];
+;       pz[_a]=zz;
+;    }
+end if
+	mov eax,[dzdx]
+	add [z],eax
+	mov eax,[dsdx]
+	add [s],eax
+	mov eax,[dtdx]
+	add [t],eax
+}
+
+align 4
+proc ZB_fillTriangleMapping, zb:dword, p0:dword, p1:dword, p2:dword
+locals
+	texture dd ? ;PIXEL*
+include 'ztriangle.inc'
+
+;
+; Texture mapping with perspective correction.
+; We use the gradient method to make less divisions.
+; TODO: pipeline the division
+;
+if 1
+
+INTERP_Z equ 1
+INTERP_STZ equ 1
+
+NB_INTERP equ 8
+
+macro DRAW_INIT
+{
+;  texture=zb->current_texture;
+;  fdzdx=(float)dzdx;
+;  fndzdx=NB_INTERP * fdzdx;
+;  ndszdx=NB_INTERP * dszdx;
+;  ndtzdx=NB_INTERP * dtzdx;
+}
+
+macro PUT_PIXEL _a
+{
+;   zz=z >> ZB_POINT_Z_FRAC_BITS;
+if TGL_FEATURE_RENDER_BITS eq 24
+;   unsigned char *ptr;
+;     if (zz >= pz[_a]) {
+;       ptr = texture + (((t & 0x3FC00000) | (s & 0x003FC000)) >> 14) * 3;
+;       pp[3 * _a]= ptr[0];
+;       pp[3 * _a + 1]= ptr[1];
+;       pp[3 * _a + 2]= ptr[2];
+;       pz[_a]=zz;
+;    }
+else
+;     if (zz >= pz[_a]) {
+;       pp[_a]=*(PIXEL *)((char *)texture+
+;               (((t & 0x3FC00000) | (s & 0x003FC000)) >> (17 - PSZSH)));
+;       pz[_a]=zz;
+;    }
+end if
+	mov eax,[dzdx]
+	add [z],eax
+	mov eax,[dsdx]
+	add [s],eax
+	mov eax,[dtdx]
+	add [t],eax
+}
+
+DRAW_LINE_M equ 1
+
+macro DRAW_LINE code
+{
+if TGL_FEATURE_RENDER_BITS eq 24
+if code eq 0
+	pz dd ? ;uint *
+	;edi = pp dd ?
+	s dd ? ;uint
+	t dd ? ;uint
+	z dd ? ;uint
+	zz dd ? ;uint
+	n dd ? ;int
+	dsdx dd ? ;int
+	dtdx dd ? ;int
+	s_z dd ? ;float
+	t_z dd ? ;float
+	fz dd ? ;float
+	zinv dd ? ;float
+end if
+if code eq 1
 ;  n=(x2>>16)-x1;
 ;  fz=(float)z1;
 ;  zinv=1.0 / fz;
@@ -188,118 +325,18 @@ if TGL_FEATURE_RENDER_BITS eq 24
 ;    n-=1;
 ;  }
 end if
-}
-
-align 4
-proc ZB_fillTriangleSmooth, zb:dword, p0:dword, p1:dword, p2:dword
-locals
-if TGL_FEATURE_RENDER_BITS eq 16
-	_drgbdx dd ? ;int
 end if
-include 'ztriangle.inc'
-	ret
-endp
-
-align 4
-proc ZB_setTexture uses eax ebx, zb:dword, texture:dword
-	mov eax,[zb]
-	mov ebx,[texture]
-	mov dword[eax+offs_zbuf_current_texture],ebx
-	ret
-endp
-
-INTERP_Z equ 1
-INTERP_ST equ 1
-
-macro DRAW_INIT
-{
-;  texture=zb->current_texture;
-}
-
-macro PUT_PIXEL _a
-{
-;   zz=z >> ZB_POINT_Z_FRAC_BITS;
-if TGL_FEATURE_RENDER_BITS eq 24
-;   unsigned char *ptr;
-;     if (zz >= pz[_a]) {
-;       ptr = texture + (((t & 0x3FC00000) | s) >> 14) * 3;
-;       pp[3 * _a]= ptr[0];
-;       pp[3 * _a + 1]= ptr[1];
-;       pp[3 * _a + 2]= ptr[2];
-;       pz[_a]=zz;
-;    }
-else
-;     if (zz >= pz[_a]) {
-;       pp[_a]=texture[((t & 0x3FC00000) | s) >> 14];
-;       pz[_a]=zz;
-;    }
-end if
-;    z+=dzdx;
-;    s+=dsdx;
-;    t+=dtdx;
-}
-
-align 4
-proc ZB_fillTriangleMapping, zb:dword, p0:dword, p1:dword, p2:dword
-locals
-	texture dd ? ;PIXEL*
-include 'ztriangle.inc'
-	ret
-endp
-
-;
-; Texture mapping with perspective correction.
-; We use the gradient method to make less divisions.
-; TODO: pipeline the division
-;
-if 1
-
-INTERP_Z equ 1
-INTERP_STZ equ 1
-
-NB_INTERP equ 8
-
-macro DRAW_INIT
-{
-;  texture=zb->current_texture;
-;  fdzdx=(float)dzdx;
-;  fndzdx=NB_INTERP * fdzdx;
-;  ndszdx=NB_INTERP * dszdx;
-;  ndtzdx=NB_INTERP * dtzdx;
-}
-
-macro PUT_PIXEL _a
-{
-;   zz=z >> ZB_POINT_Z_FRAC_BITS;
-if TGL_FEATURE_RENDER_BITS eq 24
-;   unsigned char *ptr;
-;     if (zz >= pz[_a]) {
-;       ptr = texture + (((t & 0x3FC00000) | (s & 0x003FC000)) >> 14) * 3;
-;       pp[3 * _a]= ptr[0];
-;       pp[3 * _a + 1]= ptr[1];
-;       pp[3 * _a + 2]= ptr[2];
-;       pz[_a]=zz;
-;    }
-else
-;     if (zz >= pz[_a]) {
-;       pp[_a]=*(PIXEL *)((char *)texture+
-;               (((t & 0x3FC00000) | (s & 0x003FC000)) >> (17 - PSZSH)));
-;       pz[_a]=zz;
-;    }
-end if
-;    z+=dzdx;
-;    s+=dsdx;
-;    t+=dtdx;
 }
 
 align 4
 proc ZB_fillTriangleMappingPerspective, zb:dword, p0:dword, p1:dword, p2:dword
 locals
-;    PIXEL *texture;
+	texture dd ? ;PIXEL *
-;    float fdzdx,fndzdx,ndszdx,ndtzdx;
+	fdzdx dd ? ;float
+	fndzdx dd ?
+	ndszdx dd ?
+	ndtzdx dd ?
 include 'ztriangle.inc'
-	ret
-endp
 
 end if
 
@@ -328,9 +365,12 @@ macro PUT_PIXEL _a
 ;       pp[_a]=texture[((t & 0x3FC00000) | s) >> 14];
 ;       pz[_a]=zz;
 ;    }
-;    z+=dzdx;
+	mov eax,[dzdx]
-;    sz+=dszdx;
+	add [z],eax
-;    tz+=dtzdx;
+	mov eax,[dszdx]
+	add [sz],eax
+	mov eax,[dtzdx]
+	add [tz],eax
 }
 
 align 4
@@ -338,7 +378,5 @@ proc ZB_fillTriangleMappingPerspective, zb:dword, p0:dword, p1:dword, p2:dword
 locals
 	texture dd ? ;PIXEL*
 include 'ztriangle.inc'
-	ret
-endp
 
 end if