diff --git a/programs/develop/libraries/TinyGL/asm_fork/examples/test_glu1.asm b/programs/develop/libraries/TinyGL/asm_fork/examples/test_glu1.asm
index 96252f65b6..ee8f2f367d 100644
--- a/programs/develop/libraries/TinyGL/asm_fork/examples/test_glu1.asm
+++ b/programs/develop/libraries/TinyGL/asm_fork/examples/test_glu1.asm
@@ -24,8 +24,9 @@ start:
 
 	mcall 40,0x27
 
-stdcall [kosglMakeCurrent], 10,10,300,225,ctx1
+stdcall [kosglMakeCurrent], 10,10,400,325,ctx1
 stdcall [glEnable], GL_DEPTH_TEST
+stdcall [glEnable], GL_NORMALIZE ;делам нормали одинаковой величины во избежание артефактов
 stdcall [gluNewQuadric]
 mov [qObj],eax
 
@@ -55,7 +56,7 @@ draw_window:
 	mcall 12,1
 
 	mov edx,0x33ffffff ;0x73ffffff
-	mcall 0,(50 shl 16)+330,(30 shl 16)+275,,,caption
+	mcall 0,(50 shl 16)+430,(30 shl 16)+375,,,caption
 	stdcall [kosglSwapBuffers]
 
 	mcall 12,2
@@ -150,15 +151,15 @@ stdcall [glPushMatrix]
 
 	stdcall [glRotatef], [angle_z],0.0,0.0,1.0
 	stdcall [glRotatef], [angle_y],0.0,1.0,0.0
-	stdcall [gluSphere], [qObj], 1.0, 16,16
+	stdcall [gluSphere], [qObj], 1.0, 32,32
 
 	stdcall [glColor3f], 1.0, 0.0, 0.0
 	stdcall [glTranslatef], -1.6,0.0,0.0
-	stdcall [gluSphere], [qObj], 0.55, 8,8
+	stdcall [gluSphere], [qObj], 0.55, 16,16
 
 	stdcall [glColor3f], 0.0, 0.0, 1.0
 	stdcall [glTranslatef], 3.2,0.0,0.0
-	stdcall [gluSphere], [qObj], 0.55, 8,8
+	stdcall [gluSphere], [qObj], 0.55, 16,16
 stdcall [glPopMatrix]
 ret
 
diff --git a/programs/develop/libraries/TinyGL/asm_fork/init.asm b/programs/develop/libraries/TinyGL/asm_fork/init.asm
index f661f8e5da..6c62259de3 100644
--- a/programs/develop/libraries/TinyGL/asm_fork/init.asm
+++ b/programs/develop/libraries/TinyGL/asm_fork/init.asm
@@ -39,12 +39,13 @@ proc glInit uses eax ebx ecx edx, zbuffer1:dword
 	mov dword[edx+offs_cont_vertex],eax
 
 	; viewport
-	mov dword[edx+offs_cont_viewport+offs_vpor_xmin],0
-	mov dword[edx+offs_cont_viewport+offs_vpor_ymin],0
+	xor eax,eax
+	mov dword[edx+offs_cont_viewport+offs_vpor_xmin],eax
+	mov dword[edx+offs_cont_viewport+offs_vpor_ymin],eax
 	mov eax,[ecx+offs_zbuf_xsize]
-	mov dword[edx+offs_cont_viewport+offs_vpor_xsize], eax
+	mov dword[edx+offs_cont_viewport+offs_vpor_xsize],eax
 	mov eax,[ecx+offs_zbuf_ysize]
-	mov dword[edx+offs_cont_viewport+offs_vpor_ysize], eax
+	mov dword[edx+offs_cont_viewport+offs_vpor_ysize],eax
 	mov dword[edx+offs_cont_viewport+offs_vpor_updated],1
 
 	; shared state
@@ -71,9 +72,7 @@ proc glInit uses eax ebx ecx edx, zbuffer1:dword
 		gl_V3_New eax+offs_ligh_norm_spot_direction, 0.0,0.0,-1.0
 		mov dword[eax+offs_ligh_spot_exponent],0.0
 		mov dword[eax+offs_ligh_spot_cutoff],180.0
-		mov dword[eax+offs_ligh_attenuation],1.0
-		mov dword[eax+offs_ligh_attenuation+4],0.0
-		mov dword[eax+offs_ligh_attenuation+8],0.0
+		gl_V3_New eax+offs_ligh_attenuation, 1.0,0.0,0.0
 		mov dword[eax+offs_ligh_enabled],0
 		add eax,sizeof.GLLight
 		dec ecx
@@ -107,25 +106,14 @@ proc glInit uses eax ebx ecx edx, zbuffer1:dword
 	stdcall glInitTextures,edx
 
 	; default state
-	mov dword[edx+offs_cont_current_color],1.0
-	mov dword[edx+offs_cont_current_color+4],1.0
-	mov dword[edx+offs_cont_current_color+8],1.0
-	mov dword[edx+offs_cont_current_color+12],1.0
-	mov dword[edx+offs_cont_longcurrent_color],65535
-	mov dword[edx+offs_cont_longcurrent_color+4],65535
-	mov dword[edx+offs_cont_longcurrent_color+8],65535
+	gl_V4_New edx+offs_cont_current_color, 1.0,1.0,1.0,1.0
+	gl_V3_New edx+offs_cont_longcurrent_color, 65535,65535,65535
 
-	mov dword[edx+offs_cont_current_normal],1.0
-	mov dword[edx+offs_cont_current_normal+4],0.0
-	mov dword[edx+offs_cont_current_normal+8],0.0
-	mov dword[edx+offs_cont_current_normal+12],0.0
+	gl_V4_New edx+offs_cont_current_normal, 1.0,0.0,0.0,0.0
 
 	mov dword[edx+offs_cont_current_edge_flag],1
   
-	mov dword[edx+offs_cont_current_tex_coord],0.0
-	mov dword[edx+offs_cont_current_tex_coord+4],0.0
-	mov dword[edx+offs_cont_current_tex_coord+8],0.0
-	mov dword[edx+offs_cont_current_tex_coord+12],1.0
+	gl_V4_New edx+offs_cont_current_tex_coord, 0.0,0.0,0.0,1.0
 
 	mov dword[edx+offs_cont_polygon_mode_front],GL_FILL
 	mov dword[edx+offs_cont_polygon_mode_back],GL_FILL
@@ -136,10 +124,7 @@ proc glInit uses eax ebx ecx edx, zbuffer1:dword
 	mov dword[edx+offs_cont_cull_face_enabled],0
   
 	; clear
-	mov dword[edx+offs_cont_clear_color],0.0
-	mov dword[edx+offs_cont_clear_color+4],0.0
-	mov dword[edx+offs_cont_clear_color+8],0.0
-	mov dword[edx+offs_cont_clear_color+12],0.0
+	gl_V4_New edx+offs_cont_clear_color, 0.0,0.0,0.0,0.0
 	mov dword[edx+offs_cont_clear_depth],0
 
 	; selection
@@ -150,9 +135,8 @@ proc glInit uses eax ebx ecx edx, zbuffer1:dword
 	; matrix
 	mov dword[edx+offs_cont_matrix_mode],0
 
-	mov dword[edx+offs_cont_matrix_stack_depth_max],MAX_MODELVIEW_STACK_DEPTH
-	mov dword[edx+offs_cont_matrix_stack_depth_max+4],MAX_PROJECTION_STACK_DEPTH
-	mov dword[edx+offs_cont_matrix_stack_depth_max+8],MAX_TEXTURE_STACK_DEPTH
+	gl_V3_New edx+offs_cont_matrix_stack_depth_max,\
+		MAX_MODELVIEW_STACK_DEPTH,MAX_PROJECTION_STACK_DEPTH,MAX_TEXTURE_STACK_DEPTH
 
 	mov ecx,3 ;for(i=0;i<3;i++)
 	mov ebx,edx
diff --git a/programs/develop/libraries/TinyGL/asm_fork/light.asm b/programs/develop/libraries/TinyGL/asm_fork/light.asm
index 03c483c62c..a50c5cafdc 100644
--- a/programs/develop/libraries/TinyGL/asm_fork/light.asm
+++ b/programs/develop/libraries/TinyGL/asm_fork/light.asm
@@ -1,3 +1,5 @@
+align 4
+sp128f dd 128.0
 
 align 4
 proc glopMaterial uses eax ebx ecx edi esi, context:dword, p:dword
@@ -14,7 +16,8 @@ proc glopMaterial uses eax ebx ecx edi esi, context:dword, p:dword
 		stdcall glopMaterial,eax,edi
 		mov ecx,GL_BACK
 	@@:
-	mov edi,[eax+offs_cont_materials]
+	mov edi,eax
+	add edi,offs_cont_materials
 	cmp ecx,GL_FRONT ;if (mode == GL_FRONT) m=&context.materials[0]
 	je @f
 		add edi,sizeof.GLMaterial ;else m=&context.materials[1]
@@ -54,7 +57,7 @@ proc glopMaterial uses eax ebx ecx edi esi, context:dword, p:dword
 		add edi,offs_mate_shininess
 		movsd
 		mov dword[edi],SPECULAR_BUFFER_RESOLUTION
-		fdiv dword[an180f]
+		fdiv dword[sp128f]
 		fimul dword[edi]
 		fistp dword[edi] ;m.shininess_i = (v[0]/128.0f)*SPECULAR_BUFFER_RESOLUTION
 		jmp .end_f
@@ -174,10 +177,12 @@ pushad
 		add edi,offs_ligh_spot_direction
 		mov ecx,3
 		rep movsd ;l.spot_direction=v[0,1,2]
-		mov esi,ebx
-		add esi,12
-		mov edi,edx
-		add edi,offs_ligh_norm_spot_direction
+		;mov esi,ebx
+		;add esi,12
+		sub esi,12
+		;mov edi,edx
+		;add edi,offs_ligh_norm_spot_direction
+		add edi,offs_ligh_norm_spot_direction-(offs_ligh_spot_direction+12)
 		mov ecx,3
 		rep movsd ;l.norm_spot_direction=v[0,1,2]
 		add edx,offs_ligh_norm_spot_direction
@@ -380,7 +385,8 @@ pushad
 ; esi -> GLVertex *v
 	mov esi,[v]
 	mov edx,[context]
-	mov ecx,[edx+offs_cont_materials] ;ecx(m) = &context.materials[0]
+	mov ecx,edx
+	add ecx,offs_cont_materials ;ecx(m) = &context.materials[0]
 	mov eax,[edx+offs_cont_light_model_two_side]
 	mov [twoside],eax
 
@@ -437,19 +443,26 @@ end if
 		fmul dword[ebx+offs_ligh_ambient+8]
 		fstp dword[lB] ;lB=l.ambient.v[2] * m.ambient.v[2]
 
-		cmp dword[ebx+offs_ligh_position+12],0 ;if (l.position.v[3] == 0)
+		fld dword[ebx+offs_ligh_position+offs_W]
+		ftst ;if (l.position.v[3] == 0)
+		fstsw ax
+		sahf
 		jne .els_0
 			; light at infinity
+			ffree st0 ;l.position.v[3]
+			fincstp
 			mov eax,[ebx+offs_ligh_position]
 			mov [d],eax ;d.X=l.position.v[0]
-			mov eax,[ebx+offs_ligh_position+4]
-			mov [d+4],eax ;d.Y=l.position.v[1]
-			mov eax,[ebx+offs_ligh_position+8]
-			mov [d+8],eax ;d.Z=l.position.v[2]
+			mov eax,[ebx+offs_ligh_position+offs_Y]
+			mov [d+offs_Y],eax ;d.Y=l.position.v[1]
+			mov eax,[ebx+offs_ligh_position+offs_Z]
+			mov [d+offs_Z],eax ;d.Z=l.position.v[2]
 			mov dword[att],1.0
 			jmp .els_0_end
 		.els_0:
 			; distance attenuation
+			ffree st0 ;l.position.v[3]
+			fincstp
 			fld dword[ebx+offs_ligh_position]
 			fsub dword[esi+offs_vert_ec]
 			fstp dword[d] ;d.X=l.position.v[0]-v.ec.v[0]
@@ -468,7 +481,7 @@ end if
 			fmul st0,st0
 			faddp
 			fsqrt
-			fst dword[dist] ;dist=sqrt(d.X^2+d^2+d^2)
+			fst dword[dist] ;dist=sqrt(d.X^2+d.Y^2+d.Z^2)
 			fcom dword[fl_1e_3]
 			fstsw ax
 			sahf
@@ -519,7 +532,7 @@ end if
 		ftst ;if (dot>0)
 		fstsw ax
 		sahf
-		jle .if0_end
+		jbe .if0_end
 			; diffuse light
 			fld dword[ecx+offs_mate_diffuse]
 			fmul dword[ebx+offs_ligh_diffuse]
@@ -654,7 +667,7 @@ end if
 			ftst ;if (dot_spec > 0)
 			fstsw ax
 			sahf
-			jae .if0_end
+			jbe .if0_end
 				fld dword[s]
 				fmul st0,st0
 				fld dword[s+offs_Y]
@@ -678,17 +691,16 @@ end if
 				; dot_spec= pow(dot_spec,m.shininess)
 				stdcall specbuf_get_buffer, edx, dword[ecx+offs_mate_shininess_i], dword[ecx+offs_mate_shininess]
 				mov edi,eax ;edi = specbuf
-				mov dword[idx],SPECULAR_BUFFER_SIZE
-				fild dword[idx]
+				mov dword[idx],SPECULAR_BUFFER_SIZE ;idx = SPECULAR_BUFFER_SIZE
 				fld1
 				fcomp
 				fstsw ax
+				fild dword[idx]
 				sahf
-				jae @f
-					fmul st0,st1
+				jae @f ;if(dot_spec < 1.0)
+					fmul st0,st1 ;idx *= dot_spec
 				@@:
-				fistp dword[idx] ;if (dot_spec < 1.0) idx = (int)(dot_spec*SPECULAR_BUFFER_SIZE)
-					;else idx = SPECULAR_BUFFER_SIZE;
+				fistp dword[idx]
 				ffree st0 ;dot_spec
 				fincstp
 				shl dword[idx],2
diff --git a/programs/develop/libraries/TinyGL/asm_fork/matrix.asm b/programs/develop/libraries/TinyGL/asm_fork/matrix.asm
index 4d29df8ed3..06a0e3fd2e 100644
--- a/programs/develop/libraries/TinyGL/asm_fork/matrix.asm
+++ b/programs/develop/libraries/TinyGL/asm_fork/matrix.asm
@@ -1,5 +1,9 @@
 
 if DEBUG
+align 4
+txt_gl_scal db 'glopScale',0
+txt_gl_tran db 'glopTranslate',0
+
 align 4
 proc gl_print_matrix uses eax ebx ecx edi, m:dword, rows:dword
 	mov ecx,[rows]
@@ -380,6 +384,7 @@ proc glopScale uses eax ebx ecx, context:dword, p:dword
 	loop @b
 
 if DEBUG ;glopScale
+	stdcall dbg_print,txt_gl_scal,txt_nl
 	mov ebx,[eax+offs_cont_matrix_mode]
 	shl ebx,2
 	add ebx,eax
@@ -408,9 +413,9 @@ proc glopTranslate uses eax ebx ecx, context:dword, p:dword
 		fmul st0,st3     ;m[0] * x
 		fld dword[ebx+4] ;m[1]
 		fmul st0,st3     ;m[1] * y
-		fld dword[ebx+8] ;m[2]
-		fmul st0,st3     ;m[2] * z
 		faddp
+		fld dword[ebx+8] ;m[2]
+		fmul st0,st2     ;m[2] * z
 		faddp
 		fadd dword[ebx+12] ;m[3]
 		fstp dword[ebx+12] ;m[3] = m[0] * x + m[1] * y + m[2] * z + m[3]
@@ -424,6 +429,7 @@ proc glopTranslate uses eax ebx ecx, context:dword, p:dword
 	fincstp
 
 if DEBUG ;glopTranslate
+	stdcall dbg_print,txt_gl_tran,txt_nl
 	mov ebx,[eax+offs_cont_matrix_mode]
 	shl ebx,2
 	add ebx,eax
diff --git a/programs/develop/libraries/TinyGL/asm_fork/vertex.asm b/programs/develop/libraries/TinyGL/asm_fork/vertex.asm
index 6610f49bba..25aea3a32c 100644
--- a/programs/develop/libraries/TinyGL/asm_fork/vertex.asm
+++ b/programs/develop/libraries/TinyGL/asm_fork/vertex.asm
@@ -108,13 +108,6 @@ endl
 	mov dword[zsize],(1 shl ZB_POINT_Z_FRAC_BITS) / 2
 	fiadd dword[zsize]
 	fstp dword[eax+offs_vpor_trans+offs_Z]
-if DEBUG ;gl_eval_viewport
-	stdcall dbg_print,f_ev,txt_nl
-	add eax,offs_vpor_scale
-	stdcall gl_print_matrix,eax,1
-	add eax,8
-	stdcall gl_print_matrix,eax,1
-end if
 	ret
 endp
 
@@ -260,11 +253,11 @@ pushad
 			fmul st0,st3       ;st0 *= v.coord.X
 			fld dword[ebx+4]   ;st0 = m[1]
 			fmul st0,st3       ;st0 *= v.coord.Y
+			faddp              ;st0 = v.coord.X * m[0] + v.coord.Y * m[1]
 			fld dword[ebx+8]   ;st0 = m[2]
-			fmul st0,st3       ;st0 *= v.coord.Z
+			fmul st0,st2       ;st0 *= v.coord.Z
 			fadd dword[ebx+12] ;st0 += m[3]
-			faddp              ;st0 += v.coord.Z * m[2] 
-			faddp              ;st0 += v.coord.Y * m[1] 
+			faddp              ;st0 += v.ec.X
 			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]
 			add ebx,16 ;следущая строка матрицы
 			add edx,4  ;следущая координата вектора
@@ -290,11 +283,11 @@ pushad
 			fmul st0,st3       ;st0 *= v.ec.X
 			fld dword[ebx+4]   ;st0 = m[1]
 			fmul st0,st3       ;st0 *= v.ec.Y
+			faddp              ;st0 = v.ec.X * m[0] + v.ec.Y * m[1]
 			fld dword[ebx+8]   ;st0 = m[2]
-			fmul st0,st3       ;st0 *= v.ec.Z
+			fmul st0,st2       ;st0 *= v.ec.Z
 			fadd dword[ebx+12] ;st0 += m[3]
-			faddp              ;st0 += v.ec.Z * m[2]
-			faddp              ;st0 += v.ec.Y * m[1]
+			faddp              ;st0 = v.pc.X
 			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]
 			add ebx,16 ;следущая строка матрицы
 			add edx,4  ;следущая координата вектора
@@ -317,24 +310,24 @@ pushad
 		fld dword[edi+offs_Z]
 
 		mov ecx,3
+		add edx,offs_vert_normal
 		.cycle_2:
 			fld dword[ebx]   ;st0 = m[0]
 			fmul st0,st3     ;st0 *= n.X
 			fld dword[ebx+4] ;st0 = m[1]
 			fmul st0,st3     ;st0 *= n.Y
+			faddp            ;st0 = n.X * m[0] + n.Y * m[1]
 			fld dword[ebx+8] ;st0 = m[2]
-			fmul st0,st3     ;st0 *= n.Z
-			faddp            ;st0 += n.Z * m[2]
-			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]
+			fmul st0,st2     ;st0 *= n.Z
+			faddp            ;st0 = v.normal.X
+			fstp dword[edx]  ;v.normal.X = n.X * m[0] + n.Y * m[1] + n.Z * m[2]
 			add ebx,16 ;следущая строка матрицы
 			add edx,4  ;следущая координата вектора
 		loop .cycle_2
 
 		cmp dword[eax+offs_cont_normalize_enabled],0
 		je .end_els
-			mov edx,[v]
-			add edx,offs_vert_normal
+			sub edx,12
 			stdcall gl_V3_Norm,edx
 		jmp .end_els
 	.els_0:
@@ -355,34 +348,34 @@ pushad
 		fmul st0,st3       ;st0 *= v.coord.X
 		fld dword[ebx+4]   ;st0 = m[1]
 		fmul st0,st3       ;st0 *= v.coord.Y
+		faddp              ;st0 = v.coord.X * m[0] + v.coord.Y * m[1]
 		fld dword[ebx+8]   ;st0 = m[2]
-		fmul st0,st3       ;st0 *= v.coord.Z
+		fmul st0,st2       ;st0 *= v.coord.Z
 		fadd dword[ebx+12] ;st0 += m[3]
-		faddp              ;st0 += v.coord.Z * m[2]
-		faddp              ;st0 += v.coord.Y * m[1]
+		faddp              ;st0 = v.pc.X
 		fstp dword[esi]    ;v.pc.X = v.coord.X * m[0] + v.coord.Y * m[1] + v.coord.Z * m[2] + m[3]
 
 		fld dword[ebx+16]  ;st0 = m[4]
 		fmul st0,st3       ;st0 *= v.coord.X
 		fld dword[ebx+20]  ;st0 = m[5]
 		fmul st0,st3       ;st0 *= v.coord.Y
+		faddp              ;st0 = v.coord.X * m[4] + v.coord.Y * m[5]
 		fld dword[ebx+24]  ;st0 = m[6]
-		fmul st0,st3       ;st0 *= v.coord.Z
+		fmul st0,st2       ;st0 *= v.coord.Z
 		fadd dword[ebx+28] ;st0 += m[7]
-		faddp              ;st0 += v.coord.Z * m[6]
-		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]
+		faddp              ;st0 = v.pc.Y
+		fstp dword[esi+4]  ;v.pc.Y = v.coord.X * m[4] + v.coord.Y * m[5] + v.coord.Z * m[6] + m[7]
 
 		fld dword[ebx+32]  ;st0 = m[8]
 		fmul st0,st3       ;st0 *= v.coord.X
 		fld dword[ebx+36]  ;st0 = m[9]
 		fmul st0,st3       ;st0 *= v.coord.Y
+		faddp              ;st0 = v.coord.X * m[8] + v.coord.Y * m[9]
 		fld dword[ebx+40]  ;st0 = m[10]
-		fmul st0,st3       ;st0 *= v.coord.Z
+		fmul st0,st2       ;st0 *= v.coord.Z
 		fadd dword[ebx+44] ;st0 += m[11]
-		faddp              ;st0 += v.coord.Z * m[10]
-		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]
+		faddp              ;st0 = v.pc.Z
+		fstp dword[esi+8]  ;v.pc.Z = 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
 		je .els_1
@@ -395,11 +388,11 @@ pushad
 			fmul st0,st3       ;st0 *= v.coord.X
 			fld dword[ebx+52]  ;st0 = m[13]
 			fmul st0,st3       ;st0 *= v.coord.Y
+			faddp              ;st0 = v.coord.X * m[12] + v.coord.Y * m[13]
 			fld dword[ebx+56]  ;st0 = m[14]
-			fmul st0,st3       ;st0 *= v.coord.Z
+			fmul st0,st2       ;st0 *= v.coord.Z
 			fadd dword[ebx+60] ;st0 += m[15]
-			faddp              ;st0 += v.coord.Z * m[14] 
-			faddp              ;st0 += v.coord.Y * m[13] 
+			faddp              ;st0 = v.pc.W
 			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
@@ -492,11 +485,20 @@ pushad
 	je @f
 		cmp dword[edx+offs_cont_apply_texture_matrix],0
 		je .els_1
-;           gl_M4_MulV4(&v->tex_coord, c->matrix_stack_ptr[2], &c->current_tex_coord);
+			mov eax,edx
+			add eax,offs_cont_current_tex_coord
+			push eax ;&context.current_tex_coord
+			mov eax,ebx
+			add eax,offs_vert_tex_coord
+			stdcall gl_M4_MulV4, eax, dword[edx+offs_cont_matrix_stack_ptr+8]
 			jmp @f
 		.els_1:
-			mov eax,[edx+offs_cont_current_tex_coord]
-			mov [ebx+offs_vert_tex_coord],eax
+			mov esi,edx
+			add esi,offs_cont_current_tex_coord
+			mov edi,ebx
+			add edi,offs_vert_tex_coord
+			mov ecx,4
+			rep movsd
 	@@:
 
 	; precompute the mapping to the viewport
@@ -572,20 +574,33 @@ pushad
 	@@:
 	cmp dword[edx+offs_cont_begin_type],GL_TRIANGLE_STRIP
 	jne @f
-;       if (c->vertex_cnt >= 3) {
-;           if (n == 3)
-;               n = 0;
-;            /* needed to respect triangle orientation */
-;            switch(c->vertex_cnt & 1) {
-;            case 0:
-;               gl_draw_triangle(c,&c->vertex[2],&c->vertex[1],&c->vertex[0]);
-;               break;
-;            default:
-;            case 1:
-;               gl_draw_triangle(c,&c->vertex[0],&c->vertex[1],&c->vertex[2]);
-;               break;
-;            }
-;       }
+		cmp dword[edx+offs_cont_vertex_cnt],3 ;if (context.vertex_cnt >= 3)
+		jl .end_f
+			cmp dword[n],3
+			jne .ts3
+				xor eax,eax
+				mov dword[n],eax
+			.ts3:
+			; needed to respect triangle orientation
+			mov eax,[edx+offs_cont_vertex]
+			bt dword[edx+offs_cont_vertex_cnt],0
+			jc .case_1
+				mov [esp-12],eax
+				add eax,sizeof.GLVertex
+				mov [esp-8],eax
+				add eax,sizeof.GLVertex
+				mov [esp-4],eax
+				sub esp,12
+				stdcall gl_draw_triangle, edx ;&v[2],&v[1],&v[0]
+				jmp .end_f
+			.case_1:
+				mov [esp-4],eax
+				add eax,sizeof.GLVertex
+				mov [esp-8],eax
+				add eax,sizeof.GLVertex
+				mov [esp-12],eax
+				sub esp,12
+				stdcall gl_draw_triangle, edx ;&v[0],&v[1],&v[2]
 		jmp .end_f
 	@@:
 	cmp dword[edx+offs_cont_begin_type],GL_TRIANGLE_FAN
diff --git a/programs/develop/libraries/TinyGL/asm_fork/zmath.asm b/programs/develop/libraries/TinyGL/asm_fork/zmath.asm
index 565bcd5f6f..a681cb28f1 100644
--- a/programs/develop/libraries/TinyGL/asm_fork/zmath.asm
+++ b/programs/develop/libraries/TinyGL/asm_fork/zmath.asm
@@ -605,8 +605,7 @@ proc gl_V3_Norm uses ebx, a:dword
 	fmul st0,st0
 	faddp
 	fsqrt ;st0 = sqrt(a.X^2 +a.Y^2 +a.Z^2)
-	fldz
-	fcomp
+	ftst
 	fstsw ax
 	sahf
 	je .r1 ;if (sqrt(...)==0) return 1