fixes in draw triangle functions

git-svn-id: svn://kolibrios.org@5187 a494cfbc-eb01-0410-851d-a64ba20cac60

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

@@ -43,6 +43,9 @@ endl
 	fadd dword[eax+offs_cont_viewport+offs_vpor_trans+offs_Z]
 	fistp dword[ebx+offs_vert_zp+offs_zbup_z] ;v.zp.z = st0, st0 = st1
 
+	ffree st0
+	fincstp
+
 	; color
 	bt dword[eax+offs_cont_lighting_enabled],0
 	jnc @f
@@ -695,13 +698,13 @@ pushad
 		mov edi,dword[ebx+offs_vert_zp+offs_zbup_y]
 		mov dword[norm],edi
 		fimul dword[norm]
-		fsub st0,st1
+		fsubp
 
 		;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)
 
 		mov dword[front],0
 		fldz
-		fcom st1
+		fcompp
 		fstsw ax
 		sahf
 		je .end_f
@@ -741,11 +744,9 @@ pushad
 		and eax,[cc]
 		and eax,[cc+4]
 		cmp eax,0
-		jne @f
+		jne .end_f
 			stdcall gl_draw_triangle_clip, [context],ebx,ecx,edx,0
 	.end_f:
-	finit
-	@@:
 popad
 	ret
 endp
@@ -986,14 +987,14 @@ proc gl_draw_triangle_point uses eax ebx edx, context:dword, p0:dword,p1:dword,p
 		add ebx,offs_vert_zp
 		stdcall ZB_plot,dword[edx+offs_cont_zb],ebx
 	@@:
-	add eax,[p1]
+	mov eax,[p1]
 	cmp dword[eax+offs_vert_edge_flag],0
 	je @f
 		mov ebx,eax
 		add ebx,offs_vert_zp
 		stdcall ZB_plot,dword[edx+offs_cont_zb],ebx
 	@@:
-	add eax,[p2]
+	mov eax,[p2]
 	cmp dword[eax+offs_vert_edge_flag],0
 	je @f
 		mov ebx,eax

programs/develop/libraries/TinyGL/asm_fork/examples/test3.asm

@@ -0,0 +1,174 @@
+use32
+	org 0x0
+	db 'MENUET01'
+	dd 0x1
+	dd start
+	dd i_end
+	dd mem,stacktop
+	dd 0,cur_dir_path
+
+include '../../../../../../programs/proc32.inc'
+include '../../../../../../programs/macros.inc'
+include '../../../../../../programs/develop/libraries/box_lib/load_lib.mac'
+include '../../../../../../programs/dll.inc'
+include '../opengl_const.inc'
+
+@use_library
+
+align 4
+start:
+	load_library name_tgl, cur_dir_path, library_path, system_path, \
+		err_message_found_lib, head_f_l, import_lib_tinygl, err_message_import, head_f_i
+	cmp eax,-1
+	jz button.exit
+
+	mcall 40,0x27
+
+stdcall [kosglMakeCurrent], 10,10,300,225,ctx1
+;;;stdcall [glEnable], GL_DEPTH_TEST
+stdcall [glClearColor], 0.2,0.0,0.2,0.0
+stdcall [glShadeModel],GL_FLAT
+
+call draw_3d
+
+align 4
+red_win:
+	call draw_window
+
+align 4
+still:
+	mcall 10
+	cmp al,1
+	jz red_win
+	cmp al,2
+	jz key
+	cmp al,3
+	jz button
+	jmp still
+
+align 4
+draw_window:
+	pushad
+	mcall 12,1
+
+	mov edx,0x33ffffff ;0x73ffffff
+	mcall 0,(50 shl 16)+330,(30 shl 16)+275,,,caption
+	stdcall [kosglSwapBuffers]
+
+	mcall 12,2
+	popad
+	ret
+
+align 4
+key:
+	mcall 2
+
+	cmp ah,27 ;Esc
+	je button.exit
+
+	;178 ;Up
+	;177 ;Down
+	cmp ah,176 ;Left
+	jne @f
+		fld dword[angle_z]
+		fadd dword[delt_size]
+		fstp dword[angle_z]
+		call draw_3d
+		stdcall [kosglSwapBuffers]
+	@@:
+	cmp ah,179 ;Right
+	jne @f
+		fld dword[angle_z]
+		fsub dword[delt_size]
+		fstp dword[angle_z]
+		call draw_3d
+		stdcall [kosglSwapBuffers]
+	@@:
+
+	jmp still
+
+align 4
+button:
+	mcall 17
+	cmp ah,1
+	jne still
+.exit:
+	mcall -1
+
+
+align 4
+caption db 'Test tinygl library, [Esc] - exit, [<-] and [->] - rotate',0
+align 4
+ctx1 db 28 dup (0) ;TinyGLContext or KOSGLContext
+;sizeof.TinyGLContext = 28
+
+align 4
+draw_3d:
+stdcall [glClear], GL_COLOR_BUFFER_BIT + GL_DEPTH_BUFFER_BIT ;очистим буфер цвета и глубины
+
+stdcall [glPushMatrix]
+	stdcall [glRotatef], [angle_z],0.0,0.0,1.0
+
+stdcall [glColor3f],0.0, 0.0, 1.0
+	stdcall [glBegin],GL_LINE_LOOP
+	stdcall [glVertex3f], 0.0,    0.5, 0.3
+	stdcall [glVertex3f], 0.354,  0.354, 0.3
+	stdcall [glVertex3f], 0.5,    0.0, 0.3
+	stdcall [glEnd]
+;stdcall [glColor3f],1.0, 1.0, 0.0
+	stdcall [glBegin],GL_LINE_LOOP
+	stdcall [glVertex3f], 0.354, -0.354, 0.3
+	stdcall [glVertex3f], 0.0,   -0.5, 0.3
+	stdcall [glVertex3f], -0.354,-0.354, 0.3
+	stdcall [glEnd]
+
+	stdcall [glColor3f],1.0, 0.0, 0.0
+	stdcall [glBegin],GL_TRIANGLES
+	stdcall [glVertex3f], 0.0,    0.5, 0.1
+	stdcall [glVertex3f], 0.354,  0.354, 0.1
+	stdcall [glVertex3f], 0.5,    0.0, 0.1
+stdcall [glColor3f],1.0, 1.0, 0.0
+	stdcall [glVertex3f], 0.354, -0.354, 0.1
+	stdcall [glVertex3f], 0.0,   -0.5, 0.1
+	stdcall [glVertex3f], -0.354,-0.354, 0.1
+	stdcall [glEnd]
+
+stdcall [glPopMatrix]
+ret
+
+angle_z dd 15.0
+delt_size dd 3.0
+
+;--------------------------------------------------
+align 4
+import_lib_tinygl:
+
+macro E_LIB n
+{
+	n dd sz_#n
+}
+include '../export.inc'
+	dd 0,0
+macro E_LIB n
+{
+	sz_#n db `n,0
+}
+include '../export.inc'
+
+;--------------------------------------------------
+system_path db '/sys/lib/'
+name_tgl db 'tinygl.obj',0
+err_message_found_lib db 'Sorry I cannot load library tinygl.obj',0
+head_f_i:
+head_f_l db 'System error',0
+err_message_import db 'Error on load import library tinygl.obj',0
+;--------------------------------------------------
+
+i_end:
+	rb 1024
+stacktop:
+cur_dir_path:
+	rb 4096
+library_path:
+	rb 4096
+mem:

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

@@ -83,6 +83,10 @@ m_2 db '(2)',13,10,0
 m_3 db '(3)',13,10,0
 m_4 db '(4)',13,10,0
 m_5 db '(5)',13,10,0
+m_6 db '(6)',13,10,0
+m_7 db '(7)',13,10,0
+m_8 db '(8)',13,10,0
+m_9 db '(9)',13,10,0
 
 buf_param rb 80
 
@@ -161,6 +165,11 @@ f_alloc_l db 'alloc_list',0
 f_is_l db 'glIsList',0
 f_gen_l db 'glGenLists',0
 f_end_l db 'glEndList',0
+;f_fill_trf db 'ZB_fillTriangleFlat',0
+;f_fill_trrgb db 'ZB_fillTriangleSmooth',0
+f_fill_tr db 'ZB_fillTriangle...',0
+f_fill_tr_nl db ' lines',0
+f_fill_tr_nll db ' len',0
 
 align 4
 proc dbg_print, fun:dword, mes:dword

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

@@ -14,7 +14,7 @@ if TGL_FEATURE_RENDER_BITS eq 24
 	mov eax,[ecx+offs_zbup_b]
 	shr eax,8
 	mov [colorB],al ;colorB=p2.b>>8
-else
+;else
 ;  color=RGB_TO_PIXEL(p2->r,p2->g,p2->b);
 end if
 }
@@ -29,16 +29,16 @@ local .end_0
 	cmp ax,word[ebx+2*_a] ;if (zz >= pz[_a])
 	jl .end_0
 		;edi = pp
+		mov word[ebx+2*_a],ax ;пишем в буфер глубины новое значение
 if TGL_FEATURE_RENDER_BITS eq 24
 		mov cl,[colorR]
 		mov ch,[colorG]
 		mov word[edi+3*_a],cx
 		mov cl,[colorB]
 		mov byte[edi+3*_a +2],cl
-else
+;else
 ;      pp[_a]=color;
 end if
-		mov word[ebx+2*_a],ax ;пишем в буфер глубины новое значение
 	.end_0:
 	mov eax,[dzdx]
 	add [z],eax
@@ -75,55 +75,67 @@ end if
 
 macro PUT_PIXEL _a
 {
+local .end_0
 	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
+		mov word[ebx+2*_a],ax ;пишем в буфер глубины новое значение
 if TGL_FEATURE_RENDER_BITS eq 24
-;    if (zz >= pz[_a]) {
+		mov eax,[or1]
-;      pp[3 * _a]=or1 >> 8;
+		mov cl,ah
-;      pp[3 * _a + 1]=og1 >> 8;
+		mov eax,[og1]
-;      pp[3 * _a + 2]=ob1 >> 8;
+		mov ch,ah
-;      pz[_a]=zz;
+		mov word[edi+3*_a],cx
-;    }
+		mov eax,[ob1]
-	mov eax,[dzdx]
+		mov byte[edi+3*_a +2],ah
-	add [z],eax
+end if
-;    og1+=dgdx;
+if TGL_FEATURE_RENDER_BITS eq 16
-;    or1+=drdx;
-;    ob1+=dbdx;
-elseif TGL_FEATURE_RENDER_BITS eq 16
-;    if (zz >= pz[_a]) {
 ;      tmp=rgb & 0xF81F07E0;
 ;      pp[_a]=tmp | (tmp >> 16);
-;      pz[_a]=zz;
+;else
-;    }
-;    z+=dzdx;
-;    rgb=(rgb+drgbdx) & ( ~ 0x00200800);
-else
-;    if (zz >= pz[_a]) {
 ;      pp[_a] = RGB_TO_PIXEL(or1, og1, ob1);
-;      pz[_a]=zz;
+end if
-;    }
+	.end_0:
 	mov eax,[dzdx]
 	add [z],eax
-;    og1+=dgdx;
+if TGL_FEATURE_RENDER_BITS eq 16
-;    or1+=drdx;
+;    rgb=(rgb+drgbdx) & ( ~ 0x00200800);
-;    ob1+=dbdx;
+end if
+if TGL_FEATURE_RENDER_BITS <> 16
+	mov eax,[dgdx]
+	add [og1],eax
+	mov eax,[drdx]
+	add [or1],eax
+	mov eax,[dbdx]
+	add [ob1],eax
 end if
 }
 
-DRAW_LINE_M equ 1
+;;;DRAW_LINE_M equ 1
 
 macro DRAW_LINE code
 {
+local .cycle_0
+local .cycle_1
 if TGL_FEATURE_RENDER_BITS eq 16
 if code eq 0
-;  register unsigned short *pz;
+	pz dd ? ;uint*
-;  register PIXEL *pp;
+	tmp dd ? ;uint
-;  register unsigned int tmp,z,zz,rgb,drgbdx;
+	z dd ? ;uint
-;  register int n;
+	zz dd ? ;uint
+	rgb dd ? ;uint
+	drgbdx dd ? ;uint
+	n dd ? ;int
 end if
 if code eq 1
-;  n=(x2 >> 16) - x1;
+	mov eax,[x2]
+	shr eax,16
+	sub eax,[x1]
+	mov [n],eax ;n = (x2 >> 16) - x1
 ;  pp=pp1+x1;
 ;  pz=pz1+x1;
 ;  z=z1;
@@ -131,21 +143,27 @@ if code eq 1
 ;  rgb|=(g1 >> 5) & 0x000007FF;
 ;  rgb|=(b1 << 5) & 0x001FF000;
 ;  drgbdx=_drgbdx;
-;  while (n>=3) {
+align 4
-;    PUT_PIXEL(0);
+	.cycle_0: ;while (n>=3)
-;    PUT_PIXEL(1);
+	cmp dword[n],3
-;    PUT_PIXEL(2);
+	jl .cycle_1
-;    PUT_PIXEL(3);
+		PUT_PIXEL 0
-;    pz+=4;
+		PUT_PIXEL 1
-;    pp+=4;
+		PUT_PIXEL 2
-;    n-=4;
+		PUT_PIXEL 3
-;  }
+		add dword[pz],8
-;  while (n>=0) {
+		add edi,4
-;    PUT_PIXEL(0);
+		sub [n],4
-;    pz+=1;
+	jmp .cycle_0
-;    pp+=1;
+	.cycle_1: ;while (n>=0)
-;    n-=1;
+	cmp dword[n],0
-;  }
+	jl .cycle_1_end
+		PUT_PIXEL 0
+		add dword[pz],2
+		inc edi
+		dec dword[n]
+		jmp .cycle_1
+	.cycle_1_end:
 end if
 end if
 }
@@ -171,27 +189,31 @@ INTERP_ST equ 1
 
 macro DRAW_INIT
 {
-;  texture=zb->current_texture;
+	mov eax,[zb]
+	mov eax,[eax+offs_zbuf_current_texture]
+	mov [texture],eax
 }
 
 macro PUT_PIXEL _a
 {
-;   zz=z >> ZB_POINT_Z_FRAC_BITS;
+local .end_0
+	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
+;       pz[_a]=zz;
 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
+	.end_0:
 	mov eax,[dzdx]
 	add [z],eax
 	mov eax,[dsdx]
@@ -220,8 +242,11 @@ NB_INTERP equ 8
 
 macro DRAW_INIT
 {
-;  texture=zb->current_texture;
+	mov eax,[zb]
-;  fdzdx=(float)dzdx;
+	mov eax,[eax+offs_zbuf_current_texture]
+	mov [texture],eax
+	fild dword[dzdx]
+	fstp dword[fdzdx]
 ;  fndzdx=NB_INTERP * fdzdx;
 ;  ndszdx=NB_INTERP * dszdx;
 ;  ndtzdx=NB_INTERP * dtzdx;
@@ -229,23 +254,25 @@ macro DRAW_INIT
 
 macro PUT_PIXEL _a
 {
-;   zz=z >> ZB_POINT_Z_FRAC_BITS;
+local .end_0
+	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
+;       pz[_a]=zz;
 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)));
+;           (((t & 0x3FC00000) | (s & 0x003FC000)) >> (17 - PSZSH)));
-;       pz[_a]=zz;
-;    }
 end if
+	.end_0:
 	mov eax,[dzdx]
 	add [z],eax
 	mov eax,[dsdx]
@@ -261,12 +288,11 @@ 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
+	n1 dd ? ;int - длинна горизонтальной линии в пикселях
 	dsdx dd ? ;int
 	dtdx dd ? ;int
 	s_z dd ? ;float
@@ -275,7 +301,7 @@ if code eq 0
 	zinv dd ? ;float
 end if
 if code eq 1
-;  n=(x2>>16)-x1;
+;  n1=(x2>>16)-x1;
 ;  fz=(float)z1;
 ;  zinv=1.0 / fz;
 ;  pp=(pp1 + x1 * PSZB);
@@ -283,7 +309,7 @@ if code eq 1
 ;  z=z1;
 ;  sz=sz1;
 ;  tz=tz1;
-;  while (n>=(NB_INTERP-1)) {
+;  while (n1>=(NB_INTERP-1)) {
 ;    {
 ;      float ss,tt;
 ;      ss=(sz * zinv);
@@ -305,7 +331,7 @@ if code eq 1
 ;    PUT_PIXEL(7);
 ;    pz+=NB_INTERP;
 ;    pp=(pp + NB_INTERP * PSZB);
-;    n-=NB_INTERP;
+;    n1-=NB_INTERP;
 ;    sz+=ndszdx;
 ;    tz+=ndtzdx;
 ;  }
@@ -318,11 +344,11 @@ if code eq 1
 ;      dsdx= (int)( (dszdx - ss*fdzdx)*zinv );
 ;      dtdx= (int)( (dtzdx - tt*fdzdx)*zinv );
 ;    }
-;  while (n>=0) {
+;  while (n1>=0) {
-;    PUT_PIXEL(0);
+;;;		PUT_PIXEL 0
 ;    pz+=1;
 ;    pp=(PIXEL *)((char *)pp + PSZB);
-;    n-=1;
+		dec dword[n1]
 ;  }
 end if
 end if
@@ -333,9 +359,9 @@ proc ZB_fillTriangleMappingPerspective, zb:dword, p0:dword, p1:dword, p2:dword
 locals
 	texture dd ? ;PIXEL *
 	fdzdx dd ? ;float
-	fndzdx dd ?
+	fndzdx dd ? ;float
-	ndszdx dd ?
+	ndszdx dd ? ;float
-	ndtzdx dd ?
+	ndtzdx dd ? ;float
 include 'ztriangle.inc'
 
 end if
@@ -350,27 +376,37 @@ INTERP_STZ equ 1
 
 macro DRAW_INIT
 {
-;  texture=zb->current_texture;
+	mov eax,[zb]
+	mov eax,[eax+offs_zbuf_current_texture]
+	mov [texture],eax
 }
 
 macro PUT_PIXEL _a
 {
-;   float zinv;
+local .end_0
 ;   int s,t;
-;   zz=z >> ZB_POINT_Z_FRAC_BITS;
+	mov eax,[z]
-;     if (zz >= pz[_a]) {
+	shr eax,ZB_POINT_Z_FRAC_BITS
-;       zinv= 1.0 / (float) z;
+	mov [zz],eax
-;       s= (int) (sz * zinv);
+	mov ebx,[pz]
-;       t= (int) (tz * zinv);
+	cmp ax,word[ebx+2*_a] ;if (zz >= pz[_a])
-;       pp[_a]=texture[((t & 0x3FC00000) | s) >> 14];
+	jl .end_0
 ;       pz[_a]=zz;
-;    }
+		fild dword[z]
+		fld dword[s_z]
+		fdiv st0,st1
+		;fistp dword[...s...] ;s = (int) (s_z / (float) z)
+;       t= (int) (t_z / (float) z);
+;       pp[_a]=texture[((t & 0x3FC00000) | s) >> 14];
+	.end_0:
 	mov eax,[dzdx]
 	add [z],eax
-	mov eax,[dszdx]
+	fld dword[dszdx]
-	add [sz],eax
+	fadd dword[s_z]
-	mov eax,[dtzdx]
+	fstp dword[s_z]
-	add [tz],eax
+	fld dword,[dtzdx]
+	fadd dword[t_z]
+	fstp dword[t_z]
 }
 
 align 4

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

@@ -19,27 +19,27 @@
 	update_left dd ?
 	update_right dd ?
 
-	nb_lines dd ?
+	nb_lines dd ? ;число горизонтальных линий в половине треугольника
 	dx1 dd ?
 	dy1 dd ?
-	dx2 dd ?
+	;dx2 dd ?
 	dy2 dd ?
 
-	error dd ?
+	error dd ? ;int
-	derror dd ?
+	derror dd ? ;int
-	x1 dd ?
+	x1 dd ? ;int
-	dxdy_min dd ?
+	dxdy_min dd ? ;int
-	dxdy_max dd ?
+	dxdy_max dd ? ;int
 	; warning: x2 is multiplied by 2^16
-	x2 dd ?
+	x2 dd ? ;int
-	dx2dy2 dd ?
+	dx2dy2 dd ? ;int
 
 if INTERP_Z eq 1
-	z1 dd ?
+	z1 dd ? ;int
-	dzdx dd ?
+	dzdx dd ? ;int
-	dzdy dd ?
+	dzdy dd ? ;int
-	dzdl_min dd ?
+	dzdl_min dd ? ;int
-	dzdl_max dd ?
+	dzdl_max dd ? ;int
 end if
 if INTERP_RGB eq 1
 	r1 dd ? ;int
@@ -59,16 +59,16 @@ if INTERP_RGB eq 1
 	dbdl_max dd ?
 end if
 if INTERP_ST eq 1
-	s1 dd ?
+	s1 dd ? ;int
-	dsdx dd ?
+	dsdx dd ? ;int
-	dsdy dd ?
+	dsdy dd ? ;int
-	dsdl_min dd ?
+	dsdl_min dd ? ;int
-	dsdl_max dd ?
+	dsdl_max dd ? ;int
-	t1 dd ?
+	t1 dd ? ;int
-	dtdx dd ?
+	dtdx dd ? ;int
-	dtdy dd ?
+	dtdy dd ? ;int
-	dtdl_min dd ?
+	dtdl_min dd ? ;int
-	dtdl_max dd ?
+	dtdl_max dd ? ;int
 end if
 if INTERP_STZ eq 1
 	sz1 dd ? ;float
@@ -83,11 +83,11 @@ if INTERP_STZ eq 1
 	dtzdl_max dd ? ;float
 end if
 
+
 if DRAW_LINE_M eq 1
 	DRAW_LINE 0 ;переменные делаются в макросе
 else
-	;edi = pp dd ?
+	n dd ? ;int - длинна горизонтальной линии в пикселях
-	n dd ? ;int
 if INTERP_Z eq 1
 	pz dd ? ;unsigned short *
 	z dd ? ;uint
@@ -109,41 +109,36 @@ end if
 end if
 
 endl
+pushad
+
+if DEBUG ;(1)
+stdcall dbg_print,f_fill_tr,m_1
+end if
 
 	; we sort the vertex with increasing y
 	mov ebx,[p0]
 	mov ecx,[p1]
-	mov eax,[ebx+offs_zbup_y]
-	cmp [ecx+offs_zbup_y],eax
-	jge @f
-		;if (p1.y < p0.y)
-		mov [p0],ecx
-		mov [p1],ebx
-		xchg ebx,ecx
-		mov eax,[ebx+offs_zbup_y] ;обновляем p0.y для следующего сравнения
-	@@:
 	mov edx,[p2]
-	cmp [edx+offs_zbup_y],eax
+	mov eax,[edx+offs_zbup_y]
-	jge @f
+	cmp [ecx+offs_zbup_y],eax ;(2-1)
-		;if (p2.y < p0.y)
+	jle @f
-		mov [p0],edx
+		xchg edx,ecx
-		mov [p1],ebx
-		mov [p2],ecx
-		mov ebx,[p0]
-		mov ecx,[p1]
-		mov edx,[p2]
-		jmp .end_e0
 	@@:
 	mov eax,[ecx+offs_zbup_y]
-	cmp [edx+offs_zbup_y],eax
+	cmp [ebx+offs_zbup_y],eax ;(1-0)
-	jge .end_e0
+	jle @f
-		;else if (p2.y < p1.y)
+		xchg ecx,ebx
-		mov [p1],edx
+	@@:
-		mov [p2],ecx
+	mov eax,[edx+offs_zbup_y]
-	.end_e0:
+	cmp [ecx+offs_zbup_y],eax ;(2-1)
+	jle @f
+		xchg edx,ecx
+	@@:
+	mov [p0],ebx
+	mov [p1],ecx
+	mov [p2],edx
 
 	; we compute dXdx and dXdy for all interpolated values
-
 	mov eax,[ecx+offs_zbup_x]
 	sub eax,[ebx+offs_zbup_x]
 	mov [fdx1],eax ;p1.x - p0.x
@@ -158,23 +153,26 @@ endl
 	sub eax,[ebx+offs_zbup_y]
 	mov [fdy2],eax ;p2.y - p0.y
 
-	fild dword[fdx2]
-	fst dword[fdx2]
-	fild dword[fdy1]
-	fst dword[fdy1]
-	fmulp
 	fild dword[fdx1]
 	fst dword[fdx1]
 	fild dword[fdy2]
 	fst dword[fdy2]
 	fmulp
-	fsubp ;st0 = st0-st1
+	fild dword[fdx2]
+	fst dword[fdx2]
+	fild dword[fdy1]
+	fst dword[fdy1]
+	fmulp
+	fsubp ;st0 = st1-st0
 	fst dword[fz] ;fz = fdx1 * fdy2 - fdx2 * fdy1
 	fldz
-	fcompp ;if (fz == 0)
+	fcompp ;if (fz == 0) return
 	fstsw ax
 	sahf
 	je .end_f
+if DEBUG ;(2)
+stdcall dbg_print,txt_sp,m_2
+end if
 	fld1
 	fdiv dword[fz] ;fz = 1.0 / fz
 	fst dword[fz]  ;st0 = fz
@@ -201,26 +199,24 @@ if INTERP_Z eq 1
 	mov eax,[edx+offs_zbup_z]
 	sub eax,[ebx+offs_zbup_z]
 	mov [d2],eax
-	fild dword[d1]
+	fild dword[d1] ;d1 = p1.z - p0.z
-	fst dword[d1] ;d1 = p1.z - p0.z
+	fild dword[d2] ;d2 = p2.z - p0.z
-	fild dword[d2]
-	fst dword[d2] ;d2 = p2.z - p0.z
 
-	fld dword[fdy1]
-	fmul st0,st1
 	fld dword[fdy2]
-	fmul st0,st3
+	fmul st0,st2
-	fsub st0,st1
+	fld dword[fdy1]
+	fmul st0,st2
+	fsubp
 	fistp dword[dzdx] ;dzdx = (int) (fdy2*d1 - fdy1*d2)
-	ffree st0
-	fincstp
-	fld dword[fdx2]
-	fmul st0,st2
 	fld dword[fdx1]
-	fmul st0,st2
+	fmul st0,st1
-	fsub st0,st1
+	fld dword[fdx2]
+	fmul st0,st3
+	fsubp
 	fistp dword[dzdy] ;dzdy = (int) (fdx1*d2 - fdx2*d1)
-	ffree st0
+	ffree st0 ;free d2
+	fincstp
+	ffree st0 ;free d1
 	fincstp
 end if
 
@@ -231,26 +227,24 @@ if INTERP_RGB eq 1
 	mov eax,[edx+offs_zbup_r]
 	sub eax,[ebx+offs_zbup_r]
 	mov [d2],eax
-	fild dword[d1]
+	fild dword[d1] ;d1 = p1.r - p0.r
-	fst dword[d1] ;d1 = p1.r - p0.r
+	fild dword[d2] ;d2 = p2.r - p0.r
-	fild dword[d2]
-	fst dword[d2] ;d2 = p2.r - p0.r
 
-	fld dword[fdy1]
-	fmul st0,st1
 	fld dword[fdy2]
-	fmul st0,st3
+	fmul st0,st2
-	fsub st0,st1
+	fld dword[fdy1]
+	fmul st0,st2
+	fsubp
 	fistp dword[drdx] ;drdx = (int) (fdy2*d1 - fdy1*d2)
-	ffree st0
-	fincstp
-	fld dword[fdx2]
-	fmul st0,st2
 	fld dword[fdx1]
-	fmul st0,st2
+	fmul st0,st1
-	fsub st0,st1
+	fld dword[fdx2]
+	fmul st0,st3
+	fsubp
 	fistp dword[drdy] ;drdy = (int) (fdx1*d2 - fdx2*d1)
-	ffree st0
+	ffree st0 ;free d2
+	fincstp
+	ffree st0 ;free d1
 	fincstp
 
 	mov eax,[ecx+offs_zbup_g]
@@ -259,26 +253,24 @@ if INTERP_RGB eq 1
 	mov eax,[edx+offs_zbup_g]
 	sub eax,[ebx+offs_zbup_g]
 	mov [d2],eax
-	fild dword[d1]
+	fild dword[d1] ;d1 = p1.g - p0.g
-	fst dword[d1] ;d1 = p1.g - p0.g
+	fild dword[d2] ;d2 = p2.g - p0.g
-	fild dword[d2]
-	fst dword[d2] ;d2 = p2.g - p0.g
 
-	fld dword[fdy1]
-	fmul st0,st1
 	fld dword[fdy2]
-	fmul st0,st3
+	fmul st0,st2
-	fsub st0,st1
+	fld dword[fdy1]
+	fmul st0,st2
+	fsubp
 	fistp dword[dgdx] ;dgdx = (int) (fdy2*d1 - fdy1*d2)
-	ffree st0
-	fincstp
-	fld dword[fdx2]
-	fmul st0,st2
 	fld dword[fdx1]
-	fmul st0,st2
+	fmul st0,st1
-	fsub st0,st1
+	fld dword[fdx2]
+	fmul st0,st3
+	fsubp
 	fistp dword[dgdy] ;dgdy = (int) (fdx1*d2 - fdx2*d1)
-	ffree st0
+	ffree st0 ;free d2
+	fincstp
+	ffree st0 ;free d1
 	fincstp
 
 	mov eax,[ecx+offs_zbup_b]
@@ -287,26 +279,24 @@ if INTERP_RGB eq 1
 	mov eax,[edx+offs_zbup_b]
 	sub eax,[ebx+offs_zbup_b]
 	mov [d2],eax
-	fild dword[d1]
+	fild dword[d1] ;d1 = p1.b - p0.b
-	fst dword[d1] ;d1 = p1.b - p0.b
+	fild dword[d2] ;d2 = p2.b - p0.b
-	fild dword[d2]
-	fst dword[d2] ;d2 = p2.b - p0.b
 
-	fld dword[fdy1]
-	fmul st0,st1
 	fld dword[fdy2]
-	fmul st0,st3
+	fmul st0,st2
-	fsub st0,st1
+	fld dword[fdy1]
+	fmul st0,st2
+	fsubp
 	fistp dword[dbdx] ;dbdx = (int) (fdy2*d1 - fdy1*d2)
-	ffree st0
-	fincstp
-	fld dword[fdx2]
-	fmul st0,st2
 	fld dword[fdx1]
-	fmul st0,st2
+	fmul st0,st1
-	fsub st0,st1
+	fld dword[fdx2]
+	fmul st0,st3
+	fsubp
 	fistp dword[dbdy] ;dbdy = (int) (fdx1*d2 - fdx2*d1)
-	ffree st0
+	ffree st0 ;free d2
+	fincstp
+	ffree st0 ;free d1
 	fincstp
 end if
 
@@ -317,19 +307,15 @@ if INTERP_ST eq 1
 	mov eax,[edx+offs_zbup_s]
 	sub eax,[ebx+offs_zbup_s]
 	mov [d2],eax
-	fild dword[d1]
+	fild dword[d1] ;d1 = p1.s - p0.s
-	fst dword[d1] ;d1 = p1.s - p0.s
+	fild dword[d2] ;d2 = p2.s - p0.s
-	fild dword[d2]
-	fst dword[d2] ;d2 = p2.s - p0.s
 
-	fld dword[fdy1]
-	fmul st0,st1
 	fld dword[fdy2]
-	fmul st0,st3
+	fmul st0,st2
-	fsub st0,st1
+	fld dword[fdy1]
+	fmul st0,st2
+	fsubp
 	fistp dword[dsdx] ;dsdx = (int) (fdy2*d1 - fdy1*d2)
-	ffree st0
-	fincstp
 	fld dword[fdx2]
 	fmul st0,st2
 	fld dword[fdx1]
@@ -338,6 +324,10 @@ if INTERP_ST eq 1
 	fistp dword[dsdy] ;dsdy = (int) (fdx1*d2 - fdx2*d1)
 	ffree st0
 	fincstp
+	ffree st0 ;free d2
+	fincstp
+	ffree st0 ;free d1
+	fincstp
 
 	mov eax,[ecx+offs_zbup_t]
 	sub eax,[ebx+offs_zbup_t]
@@ -345,10 +335,8 @@ if INTERP_ST eq 1
 	mov eax,[edx+offs_zbup_t]
 	sub eax,[ebx+offs_zbup_t]
 	mov [d2],eax
-	fild dword[d1]
+	fild dword[d1] ;d1 = p1.t - p0.t
-	fst dword[d1] ;d1 = p1.t - p0.t
+	fild dword[d2] ;d2 = p2.t - p0.t
-	fild dword[d2]
-	fst dword[d2] ;d2 = p2.t - p0.t
 
 	fld dword[fdy1]
 	fmul st0,st1
@@ -366,6 +354,10 @@ if INTERP_ST eq 1
 	fistp dword[dtdy] ;dtdy = (int) (fdx1*d2 - fdx2*d1)
 	ffree st0
 	fincstp
+	ffree st0 ;free d2
+	fincstp
+	ffree st0 ;free d1
+	fincstp
 end if
 
 if INTERP_STZ eq 1
@@ -374,68 +366,60 @@ if INTERP_STZ eq 1
 	fmul st0,st1
 	fstp dword[ebx+offs_zbup_sz] ;p0.sz = (float) p0.s * p0.z
 	fild dword[ebx+offs_zbup_t]
-	fmul st0,st1
+	fmulp
 	fstp dword[ebx+offs_zbup_tz] ;p0.tz = (float) p0.t * p0.z
-	ffree st0
-	fincstp
 
 	fild dword[ecx+offs_zbup_z]
 	fild dword[ecx+offs_zbup_s]
 	fmul st0,st1
 	fstp dword[ecx+offs_zbup_sz] ;p1.sz = (float) p1.s * p1.z
 	fild dword[ecx+offs_zbup_t]
-	fmul st0,st1
+	fmulp
 	fstp dword[ecx+offs_zbup_tz] ;p1.tz = (float) p1.t * p1.z
-	ffree st0
-	fincstp
 
 	fild dword[edx+offs_zbup_z]
 	fild dword[edx+offs_zbup_s]
 	fmul st0,st1
 	fstp dword[edx+offs_zbup_sz] ;p2.sz = (float) p2.s * p2.z
 	fild dword[edx+offs_zbup_t]
-	fmul st0,st1
+	fmulp
 	fstp dword[edx+offs_zbup_tz] ;p2.tz = (float) p2.t * p2.z
-	ffree st0
-	fincstp
 
 	fld dword[ecx+offs_zbup_sz]
-	fsub dword[ebx+offs_zbup_sz]
+	fsub dword[ebx+offs_zbup_sz] ;d1 = p1.sz - p0.sz
-	fst dword[d1] ;d1 = p1.sz - p0.sz
 	fld dword[edx+offs_zbup_sz]
-	fsub dword[ebx+offs_zbup_sz]
+	fsub dword[ebx+offs_zbup_sz] ;d2 = p2.sz - p0.sz
-	fst dword[d2] ;d2 = p2.sz - p0.sz
 
-	fld dword[fdy1]
-	fmul st0,st1
 	fld dword[fdy2]
-	fmul st0,st3
+	fmul st0,st2
-	fsub st0,st1
+	fld dword[fdy1]
-	fistp dword[dszdx] ;dszdx = (int) (fdy2*d1 - fdy1*d2)
+	fmul st0,st2
-	ffree st0
+	fsubp
-	fincstp
+	fstp dword[dszdx] ;dszdx = (fdy2*d1 - fdy1*d2)
 	fld dword[fdx2]
 	fmul st0,st2
 	fld dword[fdx1]
 	fmul st0,st2
 	fsub st0,st1
-	fistp dword[dszdy] ;dszdy = (int) (fdx1*d2 - fdx2*d1)
+	fstp dword[dszdy] ;dszdy = (fdx1*d2 - fdx2*d1)
 	ffree st0
 	fincstp
+	ffree st0 ;free d2
+	fincstp
+	ffree st0 ;free d1
+	fincstp
 
 	fld dword[ecx+offs_zbup_tz]
-	fsub dword[ebx+offs_zbup_tz]
+	fsub dword[ebx+offs_zbup_tz] ;d1 = p1.tz - p0.tz
-	fst dword[d1] ;d1 = p1.tz - p0.tz
 	fld dword[edx+offs_zbup_tz]
-	fsub dword[ebx+offs_zbup_tz]
+	fsub dword[ebx+offs_zbup_tz] ;d2 = p2.tz - p0.tz
-	fst dword[d2] ;d2 = p2.tz - p0.tz
 
 	fld dword[fdy1]
 	fmul st0,st1
 	fld dword[fdy2]
 	fmul st0,st3
 	fsub st0,st1
-	fistp dword[dtzdx] ;dtzdx = (int) (fdy2*d1 - fdy1*d2)
+	fstp dword[dtzdx] ;dtzdx = (fdy2*d1 - fdy1*d2)
 	ffree st0
 	fincstp
 	fld dword[fdx2]
@@ -443,13 +427,16 @@ if INTERP_STZ eq 1
 	fld dword[fdx1]
 	fmul st0,st2
 	fsub st0,st1
-	fistp dword[dtzdy] ;dtzdy = (int) (fdx1*d2 - fdx2*d1)
+	fstp dword[dtzdy] ;dtzdy = (fdx1*d2 - fdx2*d1)
 	ffree st0
 	fincstp
+	ffree st0 ;free d2
+	fincstp
+	ffree st0 ;free d1
+	fincstp
 end if
 
 	; screen coordinates
-
 	mov eax,[zb]
 	mov edx,[eax+offs_zbuf_linesize]
 	imul edx,[ebx+offs_zbup_y]
@@ -465,10 +452,13 @@ end if
 
 	mov dword[part],0
 	.cycle_0:
+if DEBUG ;(3)
+stdcall dbg_print,txt_sp,m_3
+end if
 		mov ebx,[p0]
 		mov ecx,[p1]
 		mov edx,[p2]
-		cmp dword[part],0
+		cmp dword[part],0 ;if (part == 0)
 		jne .els_0
 			mov dword[update_left],1
 			mov dword[update_right],1
@@ -493,18 +483,20 @@ end if
 			jmp .end_0
 		.els_0:
 			; second part
-			mov dword[update_left],0
-			mov dword[update_right],1
 			fldz
 			fld dword[fz]
 			fcompp ;if (fz > 0)
 			fstsw ax
 			sahf
 			jbe .els_2
+				mov dword[update_left],0
+				mov dword[update_right],1
 				mov [pr1],ecx
 				mov [pr2],edx
 				jmp .end_2
 			.els_2:
+				mov dword[update_left],1
+				mov dword[update_right],0
 				mov [l1],ecx
 				mov [l2],edx
 			.end_2:
@@ -515,37 +507,78 @@ end if
 		.end_0:
 
 	; compute the values for the left edge
-
 	cmp dword[update_left],0 ;if (update_left)
 	je .end_upd_l
 		mov ebx,[l1]
 		mov ecx,[l2]
+		mov edx,[ecx+offs_zbup_y]
+		sub edx,[ebx+offs_zbup_y]
+		mov [dy1],edx ;dy1 = l2.y - l1.y
 		mov eax,[ecx+offs_zbup_x]
 		sub eax,[ebx+offs_zbup_x]
 		mov [dx1],eax ;dx1 = l2.x - l1.x
-		mov eax,[ecx+offs_zbup_y]
+		cmp edx,0 ;if (dy1 > 0) 
-		sub eax,[ebx+offs_zbup_y]
-		mov [dy1],eax ;dy1 = l2.y - l1.y
-		cmp eax,0 ;if (dy1 > 0) 
 		jle .els_3
-			mov eax,[dx1]
-			shl eax,16
 			xor edx,edx
+			cmp eax,0
+			jl .otr_dx1
+			shl eax,16
 			div dword[dy1] ;eax = (dx1 << 16) / dy1
 			jmp .end_3
+			.otr_dx1:
+			neg eax
+			inc eax
+			shl eax,16
+			div dword[dy1] ;eax = (-dx1 << 16) / dy1
+			neg eax
+			inc eax
+			jmp .end_3
 		.els_3:
 			xor eax,eax
 		.end_3:
 		mov edx,[ebx+offs_zbup_x]
-		mov [x1],edx
+		mov [x1],edx ;x1 = l1.x
-		mov dword[error],0
+		mov dword[error],0 ;error = 0
 		mov dword[derror],eax
-		and dword[derror],0xffff
+		and dword[derror],0xffff ;derror = eax & 0x0000ffff
 		shr eax,16
-		mov [dxdy_min],eax
+		mov [dxdy_min],eax ;dxdy_min = eax >> 16
 		inc eax
 		mov [dxdy_max],eax
 
+if DEBUG ;(4) update_left [dx1], [dy1], [dxdy_min], [dxdy_max]
+push ecx edi
+	mov ecx,80
+	lea edi,[buf_param]
+	mov eax,[dx1]
+	stdcall convert_int_to_str,ecx
+	stdcall str_n_cat,edi,txt_zp_sp,2
+	stdcall str_len,edi
+	add edi,eax
+	sub ecx,eax
+
+	mov eax,[dy1]
+	stdcall convert_int_to_str,ecx
+	stdcall str_n_cat,edi,txt_zp_sp,2
+	stdcall str_len,edi
+	add edi,eax
+	sub ecx,eax
+
+	mov eax,[dxdy_min]
+	stdcall convert_int_to_str,ecx
+	stdcall str_n_cat,edi,txt_zp_sp,2
+	stdcall str_len,edi
+	add edi,eax
+	sub ecx,eax
+
+	mov eax,[dxdy_max]
+	stdcall convert_int_to_str,ecx
+
+	stdcall str_n_cat,edi,txt_nl,2
+	stdcall dbg_print,m_4,buf_param
+pop edi ecx
+end if
+
 if INTERP_Z eq 1
 	mov eax,[l1]
 	mov eax,[eax+offs_zbup_z]
@@ -585,7 +618,6 @@ if INTERP_RGB eq 1
 	mov [dbdl_min],eax ;dbdl_min = (dbdy +dbdx*dxdy_min)
 	add eax,[dbdx]
 	mov [dbdl_max],eax ;dbdl_max = dbdl_min +dbdx
-
 end if
 if INTERP_ST eq 1
 	mov ebx,[l1]
@@ -635,29 +667,85 @@ end if
 	je .end_upd_r
 		mov ebx,[pr1]
 		mov ecx,[pr2]
+		mov edx,[ebx+offs_zbup_x]
 		mov eax,[ecx+offs_zbup_x]
-		sub eax,[ebx+offs_zbup_x]
+		sub eax,edx
-		mov [dx2],eax ;dx2 = pr2.x - pr1.x
+		;mov [dx2],eax ;dx2 = pr2.x - pr1.x
-		mov eax,[ecx+offs_zbup_y]
+		shl edx,16
-		sub eax,[ebx+offs_zbup_y]
+		mov [x2],edx ; x2 = pr1.x << 16
-		mov [dy2],eax ;dy2 = pr2.y - pr1.y
+		mov edx,[ecx+offs_zbup_y]
-		cmp eax,0 ;if (dy2 > 0) 
+		sub edx,[ebx+offs_zbup_y]
+		mov [dy2],edx ;dy2 = pr2.y - pr1.y
+		cmp edx,0 ;if (dy2 > 0) 
 		jle .els_4
-			mov eax,[dx2]
-			shl eax,16
 			xor edx,edx
+			cmp eax,0
+			jl .otr_dx2
+			shl eax,16
 			div dword[dy2] ;eax = (dx2 << 16) / dy2
 			jmp .end_4
+			.otr_dx2:
+			neg eax
+			inc eax ;dx2 *= -1
+			shl eax,16
+			div dword[dy2] ;eax = (-dx2 << 16) / dy2
+			neg eax
+			inc eax
+			jmp .end_4
 		.els_4:
 			xor eax,eax
 		.end_4:
 		mov [dx2dy2],eax
-		mov eax,[ebx+offs_zbup_x]
+
-		shl eax,16
+if DEBUG ;(5) update_right [dx2dy2], [dy2], [pr1.x], [pr2.x]
-		mov [x2],eax ; x2 = pr1.x << 16
+push ecx edi
+	mov eax,[dx2dy2]
+	shr eax,16
+	mov ecx,80
+	lea edi,[buf_param]
+	stdcall convert_int_to_str,ecx
+	stdcall str_n_cat,edi,txt_zp_sp,2
+	stdcall str_len,edi
+	add edi,eax
+	sub ecx,eax
+
+	mov eax,[dy2]
+	stdcall convert_int_to_str,ecx
+	stdcall str_n_cat,edi,txt_zp_sp,2
+	stdcall str_len,edi
+	add edi,eax
+	sub ecx,eax
+
+	mov ebx,[pr1]
+	mov eax,[ebx+offs_zbup_x]
+	stdcall convert_int_to_str,ecx
+	stdcall str_n_cat,edi,txt_zp_sp,2
+	stdcall str_len,edi
+	add edi,eax
+	sub ecx,eax
+
+	mov ebx,[pr2]
+	mov eax,[ebx+offs_zbup_x]
+	stdcall convert_int_to_str,ecx
+
+	stdcall str_n_cat,edi,txt_nl,2
+	stdcall dbg_print,m_5,buf_param
+pop edi ecx
+end if
 	.end_upd_r:
 
 	; we draw all the scan line of the part
+if DEBUG ;[nb_lines]
+push ecx edi
+	mov eax,[nb_lines]
+	mov ecx,80
+	lea edi,[buf_param]
+	stdcall convert_int_to_str,ecx
+
+	stdcall str_n_cat,edi,txt_nl,2
+	stdcall dbg_print,f_fill_tr_nl,buf_param
+pop edi ecx
+end if
 
 	.beg_w_lin:
 	cmp dword[nb_lines],0 ;while (nb_lines>0)
@@ -674,6 +762,34 @@ else
 		mov [n],eax ;n = (x2 >> 16) - x1
 		imul edi,PSZB
 		add edi,[pp1] ;pp = pp1 + x1 * PSZB
+
+if DEBUG ;[n], [x1], [x2]>>16
+push ecx edi
+	mov eax,[n]
+	mov ecx,80
+	lea edi,[buf_param]
+	stdcall convert_int_to_str,ecx
+	stdcall str_n_cat,edi,txt_zp_sp,2
+	stdcall str_len,edi
+	add edi,eax
+	sub ecx,eax
+
+	mov eax,[x1]
+	stdcall convert_int_to_str,ecx
+	stdcall str_n_cat,edi,txt_zp_sp,2
+	stdcall str_len,edi
+	add edi,eax
+	sub ecx,eax
+
+	mov eax,[x2]
+	shr eax,16
+	stdcall convert_int_to_str,ecx
+
+	stdcall str_n_cat,edi,txt_nl,2
+	stdcall dbg_print,f_fill_tr_nll,buf_param
+pop edi ecx
+end if
+
 if INTERP_Z eq 1
 		mov eax,[x1]
 		shl eax,1
@@ -702,7 +818,11 @@ if INTERP_STZ eq 1
 		mov eax,[tz1]
 		mov [t_z],eax
 end if
+
+align 4
 		.cycle_1: ;while (n>=3)
+		cmp dword[n],3
+		jl .cycle_2
 			PUT_PIXEL 0
 			PUT_PIXEL 1
 			PUT_PIXEL 2
@@ -712,23 +832,24 @@ if INTERP_Z eq 1
 end if
 			add edi,4*PSZB
 			sub dword[n],4
-			cmp dword[n],3
+		jmp .cycle_1
-		jge .cycle_1
 		.cycle_2: ;while (n>=0)
+		cmp dword[n],0
+		jl .cycle_2_end
 			PUT_PIXEL 0
 if INTERP_Z eq 1
 			add dword[pz],2 ;=sizeof(uint)
 end if
 			add edi,PSZB
 			dec dword[n]
-			cmp dword[n],0
+		jmp .cycle_2
-		jge .cycle_2
+		.cycle_2_end:
-end if
+end if ;проверка от макроса DRAW_LINE
 
 	; left edge
 	mov eax,[derror]
 	add [error],eax
-	cmp eax,0 ;if (error > 0)
+	cmp dword[error],0 ;if (error > 0)
 	jle .els_er
 		sub dword[error],0x10000
 		mov eax,[dxdy_max]
@@ -808,6 +929,7 @@ end if
 		cmp dword[part],2
 		jl .cycle_0
 	.end_f:
+popad
 	ret
 endp