diff --git a/r_context.c b/r_context.c
index f7072be2..7c22764e 100644
--- a/r_context.c
+++ b/r_context.c
@@ -376,6 +376,7 @@ void R_AllowFog(qboolean allowed)
 
 void GL_SetRenderMode(int mode)
 {
+	vid.rendermode = mode;
 	/// TODO: table shading/blending???
 	/// maybe, setup block drawing function pointers here
 }
diff --git a/r_draw.c b/r_draw.c
index aa9c8ba7..9aeb1933 100644
--- a/r_draw.c
+++ b/r_draw.c
@@ -108,12 +108,12 @@ void R_DrawStretchPicImplementation (int x, int y, int w, int h, int s1, int t1,
 	{
 		sv = (skip + v)*(t2-t1)/h + t1;
 		source = pic->pixels[0] + sv*pic->width + s1;
-		if (w == s2 - s1)
-			memcpy (dest, source, w * 2);
-		else
 		{
 			f = 0;
 			fstep = s2*0x10000/w;
+			if( w == s2 - s1 )
+				fstep = 0x10000;
+#if 0
 			for (u=0 ; u<w ; u+=4)
 			{
 				dest[u] = source[f>>16];
@@ -125,6 +125,37 @@ void R_DrawStretchPicImplementation (int x, int y, int w, int h, int s1, int t1,
 				dest[u+3] = source[f>>16];
 				f += fstep;
 			}
+#else
+			for (u=0 ; u<w ; u++)
+			{
+				pixel_t src = source[f>>16];
+				int alpha = vid.alpha;
+
+				if( pic->transparent )
+				{
+					alpha &= src >> 16 - 3;
+					src = src << 3;
+				}
+
+				if( vid.color != COLOR_WHITE )
+					src = vid.modmap[src & 0xff00|(vid.color>>8)] << 8 | (src & vid.color & 0xff) | ((src & 0xff) >> 1);
+
+				if( vid.rendermode == kRenderTransAdd)
+				{
+					pixel_t screen = dest[u];
+					dest[u] = vid.addmap[src & 0xff00|(screen>>8)] << 8 | (screen & 0xff) | ((src & 0xff) >> 0);
+				}
+				else if( alpha < 7) // && (vid.rendermode == kRenderTransAlpha || vid.rendermode == kRenderTransTexture ) )
+				{
+					pixel_t screen = dest[u];
+					dest[u] = vid.alphamap[(vid.alpha << 16)|(src & 0xff00)|(screen>>8)] << 8 | (screen & 0xff) | ((src & 0xff) >> 1);
+
+				}
+				else
+					dest[u] = src;
+				f += fstep;
+			}
+#endif
 		}
 	}
 }
@@ -160,7 +191,7 @@ void R_DrawTileClear( int texnum, int x, int y, int w, int h )
 	image_t	*pic;
 	pixel_t *psrc, *pdest;
 
-	//GL_SetRenderMode( kRenderNormal );
+	GL_SetRenderMode( kRenderNormal );
 	_TriColor4f( 1.0f, 1.0f, 1.0f, 1.0f );
 	GL_Bind( XASH_TEXTURE0, texnum );
 
@@ -239,6 +270,7 @@ R_Set2DMode
 */
 void R_Set2DMode( qboolean enable )
 {
+	vid.color = COLOR_WHITE;
 	if( enable )
 	{
 //		if( glState.in2DMode )
diff --git a/r_glblit.c b/r_glblit.c
index 776feb02..f92cd824 100644
--- a/r_glblit.c
+++ b/r_glblit.c
@@ -38,9 +38,103 @@ int tex;
 unsigned short *buffer;
 
 #define LOAD(x) p##x = gEngfuncs.GL_GetProcAddress(#x)
-void R_InitBlit()
+
+void R_BuildScreenMap()
 {
 	int i;
+#ifdef SEPARATE_BLIT
+	for( i = 0; i < 256; i++ )
+	{
+		unsigned int r,g,b;
+
+		// 332 to 565
+		r = ((i >> (8 - 3) )<< 2 ) & MASK(5);
+		g = ((i >> (8 - 3 - 3)) << 3) & MASK(6);
+		b = ((i >> (8 - 3 - 3 - 2)) << 3) & MASK(5);
+		vid.screen_major[i] = r << (6 + 5) | (g << 5) | b;
+
+
+		// restore minor GBRGBRGB
+		r = MOVE_BIT(i, 5, 1) | MOVE_BIT(i, 2, 0);
+		g = MOVE_BIT(i, 7, 2) | MOVE_BIT(i, 4, 1) | MOVE_BIT(i, 1, 0);
+		b = MOVE_BIT(i, 6, 2) | MOVE_BIT(i, 3, 1) | MOVE_BIT(i, 0, 0);
+		vid.screen_minor[i] = r << (6 + 5) | (g << 5) | b;
+
+	}
+#else
+	for( i = 0; i < 256; i++ )
+	{
+		unsigned int r,g,b , major, j;
+
+		// 332 to 565
+		r = ((i >> (8 - 3) )<< 2 ) & MASK(5);
+		g = ((i >> (8 - 3 - 3)) << 3) & MASK(6);
+		b = ((i >> (8 - 3 - 3 - 2)) << 3) & MASK(5);
+		major = r << (6 + 5) | (g << 5) | b;
+
+
+		for( j = 0; j < 256; j++ )
+		{
+			// restore minor GBRGBRGB
+			r = MOVE_BIT(i, 5, 1) | MOVE_BIT(i, 2, 0);
+			g = MOVE_BIT(i, 7, 2) | MOVE_BIT(i, 4, 1) | MOVE_BIT(i, 1, 0);
+			b = MOVE_BIT(i, 6, 2) | MOVE_BIT(i, 3, 1) | MOVE_BIT(i, 0, 0);
+			vid.screen[(i<<8)|j] = r << (6 + 5) | (g << 5) | b | major;
+
+		}
+
+	}
+#endif
+}
+
+void R_BuildBlendMaps()
+{
+	unsigned int r1, g1, b1;
+	unsigned int r2, g2, b2;
+
+	for( r1 = 0; r1 < BIT(3); r1++ )
+		for( g1 = 0; g1 < BIT(3); g1++ )
+			for( b1 = 0; b1 < BIT(2); b1++ )
+				for( r2 = 0; r2 < BIT(3); r2++ )
+					for( g2 = 0; g2 < BIT(3); g2++ )
+						for( b2 = 0; b2 < BIT(2); b2++ )
+						{
+							unsigned int r, g, b;
+							unsigned short index1 = r1 << (2 + 3) | g1 << 2 | b1;
+							unsigned short index2 = (r2 << (2 + 3) | g2 << 2 | b2) << 8;
+							unsigned int a;
+
+							r = r1 + r2;
+							g = g1 + g2;
+							b = b1 + b2;
+							if( r > MASK(2) )
+								r = MASK(2);
+							if( g > MASK(2) )
+								g = MASK(2);
+							if( b > MASK(1) )
+								b = MASK(1);
+							ASSERT(!vid.addmap[index2|index1]);
+
+							vid.addmap[index2|index1] =  r << (2 + 3) | g << 2 | b;
+							r = r1 * r2 / MASK(2);
+							g = g1 * g2 / MASK(2);
+							b = b1 * b2 / MASK(1);
+							vid.modmap[index2|index1] =  r << (2 + 3) | g << 2 | b;
+
+							for( a = 0; a < 8; a++ )
+							{
+								r = r1 * a / 7 + r2 * (7 - a) / 7;
+								g = g1 * a / 7 + g2 * (7 - a) / 7;
+								b = b1 * a / 7 + b2 * (7 - a) / 7;
+								vid.alphamap[a << 16|index2|index1] =  r << (2 + 3) | g << 2 | b;
+							}
+
+						}
+}
+
+void R_InitBlit()
+{
+
 	LOAD(glBegin);
 	LOAD(glEnd);
 	LOAD(glTexCoord2f);
@@ -58,7 +152,7 @@ void R_InitBlit()
 	LOAD(glGetError);
 	LOAD(glGenTextures);
 	LOAD(glTexParameteri);
-
+#ifdef GLDEBUG
 	if( gpGlobals->developer )
 	{
 		gEngfuncs.Con_Reportf( "Installing GL_DebugOutput...\n");
@@ -70,25 +164,14 @@ void R_InitBlit()
 
 	// enable all the low priority messages
 	pglDebugMessageControlARB( GL_DONT_CARE, GL_DONT_CARE, GL_DEBUG_SEVERITY_LOW_ARB, 0, NULL, true );
+#endif
 	pglGenTextures( 1, &tex );
+
+
 	buffer = Mem_Malloc( r_temppool, 1920*1080*2 );
 
-	for( i = 0; i < 256; i++ )
-	{
-		unsigned int r,g,b;
-
-		// 332 to 565
-		r = ((i >> (8 - 3) )<< 2 ) & MASK(5);
-		g = ((i >> (8 - 3 - 3)) << 3) & MASK(6);
-		b = ((i >> (8 - 3 - 3 - 2)) << 3) & MASK(5);
-		vid.screen_major[i] = r << (6 + 5) | (g << 5) | b;
-
-		// restore minor GBRGBRGB
-		r = MOVE_BIT(i, 5, 1) | MOVE_BIT(i, 2, 0);
-		g = MOVE_BIT(i, 7, 2) | MOVE_BIT(i, 4, 1) | MOVE_BIT(i, 1, 0);
-		b = MOVE_BIT(i, 6, 2) | MOVE_BIT(i, 3, 1) | MOVE_BIT(i, 0, 0);
-		vid.screen_minor[i] = r << (6 + 5) | (g << 5) | b;
-	}
+	R_BuildScreenMap();
+	R_BuildBlendMaps();
 }
 
 void R_BlitScreen()
@@ -99,11 +182,18 @@ void R_BlitScreen()
 
 	for( i = 0; i < vid.width * vid.height;i++)
 	{
+#ifdef SEPARATE_BLIT
+		// need only 1024 bytes table, but slower
+		// wtf?? maybe some prefetch???
 		byte major = buf[(i<<1)+1];
 		byte minor = buf[(i<<1)];
 
-		buffer[i] = vid.screen_major[major]|vid.screen_minor[minor];
+		buffer[i] = vid.screen_major[major] |vid.screen_minor[minor];
+#else
+		buffer[i] = vid.screen[vid.buffer[i]];
+#endif
 	}
+
 	pglBindTexture(GL_TEXTURE_2D, tex);
 	pglViewport( 0, 0, gpGlobals->width, gpGlobals->height );
 	pglMatrixMode( GL_PROJECTION );
@@ -119,6 +209,7 @@ void R_BlitScreen()
 	pglTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 	pglTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 	//gEngfuncs.Con_Printf("%d\n",pglGetError());
+
 	pglTexImage2D( GL_TEXTURE_2D, 0, GL_RGB, vid.width, vid.height, 0, GL_RGB, GL_UNSIGNED_SHORT_5_6_5, buffer );
 	//gEngfuncs.Con_Printf("%d\n",pglGetError());
 	pglBegin( GL_QUADS );
@@ -136,4 +227,5 @@ void R_BlitScreen()
 	pglEnd();
 	pglDisable( GL_TEXTURE_2D );
 	gEngfuncs.GL_SwapBuffers();
+	memset( vid.buffer, 0, vid.width * vid.height * 2 );
 }
diff --git a/r_image.c b/r_image.c
index 778ce912..07900cbf 100644
--- a/r_image.c
+++ b/r_image.c
@@ -220,7 +220,7 @@ static void GL_SetTextureFormat( image_t *tex, pixformat_t format, int channelMa
 	qboolean	haveAlpha = ( channelMask & IMAGE_HAS_ALPHA );
 
 	Assert( tex != NULL );
-	tex->transparent = !!( channelMask & IMAGE_HAS_ALPHA );
+	//tex->transparent = !!( channelMask & IMAGE_HAS_ALPHA );
 }
 
 /*
@@ -509,16 +509,26 @@ static qboolean GL_UploadTexture( image_t *tex, rgbdata_t *pic )
 	if( !pic->buffer )
 		return true;
 
+	/// TODO: generate mipmaps
+
+	if( tex->flags & TF_HAS_ALPHA )
+		tex->transparent = true;
+
 	tex->pixels[0] = Mem_Calloc( r_temppool, tex->width * tex->height * sizeof(pixel_t) + 64 );
 
 	for( i = 0; i < tex->width * tex->height; i++ )
 	{
 		unsigned int r, g, b, major, minor;
-
+#if 0
+		r = pic->buffer[i * 4 + 0] * MASK(5-1) / 255;
+		g = pic->buffer[i * 4 + 1] * MASK(6-1) / 255;
+		b = pic->buffer[i * 4 + 2] * MASK(5-1) / 255;
+#else
+		// seems to look better
 		r = pic->buffer[i * 4 + 0] * BIT(5) / 256;
 		g = pic->buffer[i * 4 + 1] * BIT(6) / 256;
 		b = pic->buffer[i * 4 + 2] * BIT(5) / 256;
-
+#endif
 		// 565 to 332
 		major = (((r >> 2) & MASK(3)) << 5) |( (( (g >> 3) & MASK(3)) << 2 )  )| (((b >> 3) & MASK(2)));
 
@@ -526,9 +536,14 @@ static qboolean GL_UploadTexture( image_t *tex, rgbdata_t *pic )
 		minor = MOVE_BIT(r,1,5) | MOVE_BIT(r,0,2) | MOVE_BIT(g,2,7) | MOVE_BIT(g,1,4) | MOVE_BIT(g,0,1) | MOVE_BIT(b,2,6)| MOVE_BIT(b,1,3)|MOVE_BIT(b,0,0);
 
 		tex->pixels[0][i] = major << 8 | (minor & 0xFF);
+		if( tex->transparent )
+		{
+			unsigned int alpha = (pic->buffer[i * 4 + 3] * 8 / 256) << (16 - 3);
+			tex->pixels[0][i] = (tex->pixels[0][i] >> 3) | alpha;
+		}
 	}
 
-	/// TODO: generate mipmaps
+
 #if 0
 
 
diff --git a/r_local.h b/r_local.h
index 7058201f..1548f63e 100644
--- a/r_local.h
+++ b/r_local.h
@@ -124,13 +124,27 @@ typedef struct vrect_s
 	struct vrect_s  *pnext;
 } vrect_t;
 
+#define COLOR_WHITE 0xFFFF
+#define SEPARATE_BLIT
 typedef struct
 {
 	pixel_t                 *buffer;                // invisible buffer
 	pixel_t                 *colormap;              // 256 * VID_GRADES size
-	pixel_t                 *alphamap;              // 256 * 256 translucency map
+	//pixel_t                 *alphamap;              // 256 * 256 translucency map
+#ifdef SEPARATE_BLIT
 	pixel_t					screen_minor[256];
 	pixel_t					screen_major[256];
+#else
+	pixel_t					screen[256*256];
+#endif
+	byte					addmap[256*256];
+	byte					modmap[256*256];
+	pixel_t                 alphamap[8*256*256];
+	pixel_t					color;
+	byte alpha;
+
+	// maybe compute colormask for minor byte?
+	int						rendermode;
 	int                             rowbytes;               // may be > width if displayed in a window
 									// can be negative for stupid dibs
 	int						width;
diff --git a/r_triapi.c b/r_triapi.c
index 1ff6f817..1ee810b4 100644
--- a/r_triapi.c
+++ b/r_triapi.c
@@ -39,7 +39,7 @@ set rendermode
 */
 void TriRenderMode( int mode )
 {
-	ds.renderMode = mode;
+	ds.renderMode = vid.rendermode = mode;
 #if 0
 	switch( mode )
 	{
@@ -130,9 +130,29 @@ _TriColor4f
 
 =============
 */
-void _TriColor4f( float r, float g, float b, float a )
+void _TriColor4f( float rr, float gg, float bb, float aa )
 {
 	//pglColor4f( r, g, b, a );
+	unsigned short r,g,b;
+
+	if( rr == 1 && gg == 1 && bb == 1 )
+	{
+		vid.color = COLOR_WHITE;
+		return;
+	}
+
+	r = rr * 31, g = gg * 63, b = bb * 31;
+
+	vid.alpha = aa * 7;
+	if( vid.alpha > 7 )
+		vid.alpha = 7;
+
+	unsigned int major = (((r >> 2) & MASK(3)) << 5) |( (( (g >> 3) & MASK(3)) << 2 )  )| (((b >> 3) & MASK(2)));
+
+	// save minor GBRGBRGB
+	unsigned int minor = MOVE_BIT(r,1,5) | MOVE_BIT(r,0,2) | MOVE_BIT(g,2,7) | MOVE_BIT(g,1,4) | MOVE_BIT(g,0,1) | MOVE_BIT(b,2,6)| MOVE_BIT(b,1,3)|MOVE_BIT(b,0,0);
+
+	vid.color =  major << 8 | (minor & 0xFF);
 }
 
 /*
@@ -148,7 +168,7 @@ void TriColor4ub( byte r, byte g, byte b, byte a )
 	ds.triRGBA[2] = b * (1.0f / 255.0f);
 	ds.triRGBA[3] = a * (1.0f / 255.0f);
 
-	_TriColor4f( ds.triRGBA[0], ds.triRGBA[1], ds.triRGBA[2], 1.0f );
+	_TriColor4f( ds.triRGBA[0], ds.triRGBA[1], ds.triRGBA[2], ds.triRGBA[3] ); //1.0f );
 }
 
 /*