From a26a373d15429b774448196f9cb6997e2eee5a25 Mon Sep 17 00:00:00 2001
From: Spoike <acceptthis@users.sourceforge.net>
Date: Fri, 14 Aug 2015 02:46:38 +0000
Subject: [PATCH] fixed picmip simpleitems issue. fixed double picmip bug. use
 pbos for videocapture, for extra speed. mwahaha. fix issues with 2d batching.
 including drawflags.

git-svn-id: https://svn.code.sf.net/p/fteqw/code/trunk@4962 fc73d0e0-1445-4013-8a0c-d673dee63da5
---
 engine/client/cl_ents.c      |   2 +-
 engine/client/cl_screen.c    | 185 ++++++++++++++++++++++++++---------
 engine/client/client.h       |   2 +-
 engine/client/image.c        | 119 +++++++++++-----------
 engine/client/m_mp3.c        | 151 ++++++++++++++++++++++++----
 engine/client/merged.h       |   4 +-
 engine/client/pr_menu.c      |   6 ++
 engine/client/r_2d.c         |   9 +-
 engine/client/renderer.c     |   2 +-
 engine/client/screen.h       |   3 +-
 engine/client/vid.h          |   2 +-
 engine/client/vid_headless.c |   3 +-
 engine/d3d/vid_d3d.c         |  11 ++-
 engine/d3d/vid_d3d11.c       |   6 +-
 engine/gl/gl_font.c          |  13 +--
 engine/gl/gl_screen.c        |  44 ++++++---
 engine/gl/glsupp.h           |   4 +
 engine/server/savegame.c     |   5 +-
 plugins/avplug/avencode.c    |  28 +++++-
 19 files changed, 424 insertions(+), 175 deletions(-)

diff --git a/engine/client/cl_ents.c b/engine/client/cl_ents.c
index 6665e6717..a81825b82 100644
--- a/engine/client/cl_ents.c
+++ b/engine/client/cl_ents.c
@@ -3686,7 +3686,7 @@ void CL_LinkPacketEntities (void)
 					Q_snprintfz(name, sizeof(name), "textures/bmodels/simple_%s_%i.tga", basename, ent->skinnum);
 				else
 					Q_snprintfz(name, sizeof(name), "textures/models/simple_%s_%i.tga", basename, ent->skinnum);
-				model->simpleskin[ent->skinnum] = R_RegisterShader(name, 0, va("{\nprogram defaultsprite\nsurfaceparm noshadows\nsurfaceparm nodlight\nsort seethrough\n{\nmap \"%s\"\nalphafunc ge128\n}\n}\n", name));
+				model->simpleskin[ent->skinnum] = R_RegisterShader(name, 0, va("{\nnomipmaps\nprogram defaultsprite\nsurfaceparm noshadows\nsurfaceparm nodlight\nsort seethrough\n{\nmap \"%s\"\nalphafunc ge128\n}\n}\n", name));
 			}
 			VectorCopy(le->angles, angles);
 
diff --git a/engine/client/cl_screen.c b/engine/client/cl_screen.c
index 8c8f99226..ab1305ffc 100644
--- a/engine/client/cl_screen.c
+++ b/engine/client/cl_screen.c
@@ -2031,13 +2031,96 @@ typedef struct _TargaHeader {
 
 
 #if defined(AVAIL_JPEGLIB) && !defined(NO_JPEG)
-qboolean screenshotJPEG(char *filename, enum fs_relative fsroot, int compression, qbyte *screendata, int screenwidth, int screenheight);
+qboolean screenshotJPEG(char *filename, enum fs_relative fsroot, int compression, qbyte *screendata, int screenwidth, int screenheight, enum uploadfmt fmt);
 #endif
 #ifdef AVAIL_PNGLIB
-int Image_WritePNG (char *filename, enum fs_relative fsroot, int compression, qbyte *pixels, int width, int height);
+int Image_WritePNG (char *filename, enum fs_relative fsroot, int compression, qbyte *pixels, int width, int height, enum uploadfmt fmt);
 #endif
 void WriteBMPFile(char *filename, enum fs_relative fsroot, qbyte *in, int width, int height);
 
+qboolean WriteTGA(char *filename, enum fs_relative fsroot, qbyte *rgb_buffer, int width, int height, enum uploadfmt fmt)
+{
+	size_t c, i;
+	vfsfile_t *vfs;
+	if (fmt != TF_BGRA32 && fmt != TF_RGB24 && fmt != TF_RGBA32 && fmt != TF_BGR24)
+		return false;
+	FS_CreatePath(filename, fsroot);
+	vfs = FS_OpenVFS(filename, "wb", fsroot);
+	if (vfs)
+	{
+		unsigned char header[18];
+		memset (header, 0, 18);
+		header[2] = 2;          // uncompressed type
+		header[12] = width&255;
+		header[13] = width>>8;
+		header[14] = height&255;
+		header[15] = height>>8;
+		header[16] = 24;        // pixel size
+
+		if (fmt == TF_BGRA32)
+		{
+#if 0
+			header[16] = 32;
+#else
+			qbyte tmp[3];
+			// compact+swap
+			c = width*height;
+			for (i=0 ; i<c ; i++)
+			{
+				tmp[0] = rgb_buffer[i*4+0];
+				tmp[1] = rgb_buffer[i*4+1];
+				tmp[2] = rgb_buffer[i*4+2];
+				rgb_buffer[i*3+0] = tmp[0];
+				rgb_buffer[i*3+1] = tmp[1];
+				rgb_buffer[i*3+2] = tmp[2];
+			}
+			c *= 3;
+#endif
+		}
+		else if (fmt == TF_BGR24)
+			c = width*height*3;
+		else if (fmt == TF_RGBA32)
+		{
+			int s = 3;
+			qbyte tmp[3];
+#if 0
+			s = 4;
+			header[16] = s*8;
+#endif
+			// compact+swap
+			c = width*height;
+			for (i=0 ; i<c ; i++)
+			{
+				tmp[0] = rgb_buffer[i*4+0];
+				tmp[1] = rgb_buffer[i*4+1];
+				tmp[2] = rgb_buffer[i*4+2];
+				rgb_buffer[i*s+0] = tmp[2];
+				rgb_buffer[i*s+1] = tmp[1];
+				rgb_buffer[i*s+2] = tmp[0];
+			}
+			c *= s;
+		}
+		else if (fmt == TF_RGB24)
+		{
+			qbyte temp;
+			// swap r+b in place
+			c = width*height*3;
+			for (i=0 ; i<c ; i+=3)
+			{
+				temp = rgb_buffer[i];
+				rgb_buffer[i] = rgb_buffer[i+2];
+				rgb_buffer[i+2] = temp;
+			}
+		}
+		else
+			c = 0;
+		VFS_WRITE(vfs, header, sizeof(header));
+		VFS_WRITE(vfs, rgb_buffer, c);
+		VFS_CLOSE(vfs);
+	}
+	return true;
+}
+
 /*
 Find closest color in the palette for named color
 */
@@ -2071,9 +2154,8 @@ int MipColor(int r, int g, int b)
 	return best;
 }
 
-qboolean SCR_ScreenShot (char *filename, enum fs_relative fsroot, void *rgb_buffer, int width, int height)
+qboolean SCR_ScreenShot (char *filename, enum fs_relative fsroot, void *rgb_buffer, int width, int height, enum uploadfmt fmt)
 {
-	int		i, c, temp;
 #if defined(AVAIL_PNGLIB) || defined(AVAIL_JPEGLIB)
 	extern cvar_t scr_sshot_compression;
 #endif
@@ -2088,70 +2170,66 @@ qboolean SCR_ScreenShot (char *filename, enum fs_relative fsroot, void *rgb_buff
 #ifdef AVAIL_PNGLIB
 	if (!Q_strcasecmp(ext, "png"))
 	{
-		return Image_WritePNG(filename, fsroot, scr_sshot_compression.value, rgb_buffer, width, height);
+		//png can do bgr+rgb
+		//rgba bgra will result in an extra alpha chan
+		return Image_WritePNG(filename, fsroot, scr_sshot_compression.value, rgb_buffer, width, height, fmt);
 	}
 	else
 #endif
 #ifdef AVAIL_JPEGLIB
 		if (!Q_strcasecmp(ext, "jpeg") || !Q_strcasecmp(ext, "jpg"))
 	{
-		return screenshotJPEG(filename, fsroot, scr_sshot_compression.value, rgb_buffer, width, height);
+		return screenshotJPEG(filename, fsroot, scr_sshot_compression.value, rgb_buffer, width, height, fmt);
 	}
 	else
 #endif
 	/*	if (!Q_strcasecmp(ext, "bmp"))
 	{
-		WriteBMPFile(pcxname, rgb_buffer, width, height);
+		return WriteBMPFile(pcxname, rgb_buffer, width, height);
 	}
 	else*/
 		if (!Q_strcasecmp(ext, "pcx"))
 	{
-		int y, x;
+		int y, x, s;
 		qbyte *src, *dest;
 		qbyte *newbuf = rgb_buffer;
-		// convert in-place to eight bit
-		for (y = 0; y < height; y++)
+		if (fmt == TF_RGB24 || fmt == TF_RGBA32)
 		{
-			src = newbuf + (width * 3 * y);
-			dest = newbuf + (width * y);
+			s = (fmt == TF_RGB24)?3:4;
+			// convert in-place to eight bit
+			for (y = 0; y < height; y++)
+			{
+				src = newbuf + (width * s * y);
+				dest = newbuf + (width * y);
 
-			for (x = 0; x < width; x++) {
-				*dest++ = MipColor(src[0], src[1], src[2]);
-				src += 3;
+				for (x = 0; x < width; x++) {
+					*dest++ = MipColor(src[0], src[1], src[2]);
+					src += s;
+				}
 			}
 		}
+		else if (fmt == TF_BGR24 || fmt == TF_BGRA32)
+		{
+			s = (fmt == TF_BGR24)?3:4;
+			// convert in-place to eight bit
+			for (y = 0; y < height; y++)
+			{
+				src = newbuf + (width * s * y);
+				dest = newbuf + (width * y);
+
+				for (x = 0; x < width; x++) {
+					*dest++ = MipColor(src[2], src[1], src[0]);
+					src += s;
+				}
+			}
+		}
+		else
+			return false;
 
 		WritePCXfile (filename, fsroot, newbuf, width, height, width, host_basepal, false);
 	}
 	else if (!Q_strcasecmp(ext, "tga"))	//tga
-	{
-		vfsfile_t *vfs;
-		FS_CreatePath(filename, fsroot);
-		vfs = FS_OpenVFS(filename, "wb", fsroot);
-		if (vfs)
-		{
-			unsigned char header[18];
-			memset (header, 0, 18);
-			header[2] = 2;          // uncompressed type
-			header[12] = width&255;
-			header[13] = width>>8;
-			header[14] = height&255;
-			header[15] = height>>8;
-			header[16] = 24;        // pixel size
-			VFS_WRITE(vfs, header, sizeof(header));
-
-			// swap rgb to bgr
-			c = width*height*3;
-			for (i=0 ; i<c ; i+=3)
-			{
-				temp = ((qbyte*)rgb_buffer)[i];
-				((qbyte*)rgb_buffer)[i] = ((qbyte*)rgb_buffer)[i+2];
-				((qbyte*)rgb_buffer)[i+2] = temp;
-			}
-			VFS_WRITE(vfs, rgb_buffer, c);
-			VFS_CLOSE(vfs);
-		}
-	}
+		return WriteTGA(filename, fsroot, rgb_buffer, width, height, fmt);
 	else	//extension / type not recognised.
 		return false;
 	return true;
@@ -2170,6 +2248,7 @@ void SCR_ScreenShot_f (void)
 	vfsfile_t *vfs;
 	void *rgbbuffer;
 	int width, height;
+	enum uploadmfmt fmt;
 
 	if (!VID_GetRGBInfo)
 	{
@@ -2213,10 +2292,10 @@ void SCR_ScreenShot_f (void)
 
 	FS_NativePath(pcxname, FS_GAMEONLY, sysname, sizeof(sysname));
 
-	rgbbuffer = VID_GetRGBInfo(0, &width, &height);
+	rgbbuffer = VID_GetRGBInfo(&width, &height, &fmt);
 	if (rgbbuffer)
 	{
-		if (SCR_ScreenShot(pcxname, FS_GAMEONLY, rgbbuffer, width, height))
+		if (SCR_ScreenShot(pcxname, FS_GAMEONLY, rgbbuffer, width, height, fmt))
 		{
 			Con_Printf ("Wrote %s\n", sysname);
 			BZ_Free(rgbbuffer);
@@ -2233,6 +2312,7 @@ void SCR_ScreenShot_Mega_f(void)
 	int height;
 	qbyte *rgbbuffer;
 	char filename[MAX_QPATH];
+	enum uploadfmt fmt;
 
 	//poke the various modes into redrawing the screen (without huds), to avoid any menus or console drawn over the top of the current backbuffer.
 	//FIXME: clear-to-black first
@@ -2286,10 +2366,10 @@ void SCR_ScreenShot_Mega_f(void)
 	//okay, we drew something, we're good to save a screeny.
 	if (okay)
 	{
-		rgbbuffer = VID_GetRGBInfo(0, &width, &height);
+		rgbbuffer = VID_GetRGBInfo(&width, &height, &fmt);
 		if (rgbbuffer)
 		{
-			if (SCR_ScreenShot(filename, FS_GAMEONLY, rgbbuffer, width, height))
+			if (SCR_ScreenShot(filename, FS_GAMEONLY, rgbbuffer, width, height, fmt))
 			{
 				char			sysname[1024];
 				FS_NativePath(filename, FS_GAMEONLY, sysname, sizeof(sysname));
@@ -2374,6 +2454,7 @@ qboolean SCR_RSShot (void)
 	float fracw, frach;
 	char st[80];
 	time_t now;
+	enum uploadfmt fmt;
 
 	if (!scr_allowsnap.ival)
 		return false;
@@ -2394,7 +2475,15 @@ qboolean SCR_RSShot (void)
 //
 // save the pcx file
 //
-	newbuf = VID_GetRGBInfo(0, &truewidth, &trueheight);
+	newbuf = VID_GetRGBInfo(&truewidth, &trueheight, &fmt);
+
+	if (fmt == TF_INVALID)
+		return false;
+	if (fmt != TF_RGB24)
+	{
+		BZ_Free(newbuf);
+		return false;
+	}
 
 	w = RSSHOT_WIDTH;
 	h = RSSHOT_HEIGHT;
diff --git a/engine/client/client.h b/engine/client/client.h
index 318c8d1d4..06ac80377 100644
--- a/engine/client/client.h
+++ b/engine/client/client.h
@@ -1572,7 +1572,7 @@ typedef struct
 typedef struct {
 	char *drivername;
 	void *(VARGS *capture_begin) (char *streamname, int videorate, int width, int height, int *sndkhz, int *sndchannels, int *sndbits);
-	void (VARGS *capture_video) (void *ctx, void *data, int frame, int width, int height);
+	void (VARGS *capture_video) (void *ctx, void *data, int frame, int width, int height, enum uploadfmt fmt);
 	void (VARGS *capture_audio) (void *ctx, void *data, int bytes);
 	void (VARGS *capture_end) (void *ctx);
 } media_encoder_funcs_t;
diff --git a/engine/client/image.c b/engine/client/image.c
index ba6c20dcf..fddfa7420 100644
--- a/engine/client/image.c
+++ b/engine/client/image.c
@@ -720,6 +720,7 @@ void (PNGAPI *qpng_set_expand_gray_1_2_4_to_8) PNGARG((png_structp png_ptr)) PST
 #else
 void (PNGAPI *qpng_set_gray_1_2_4_to_8) PNGARG((png_structp png_ptr)) PSTATIC(png_set_gray_1_2_4_to_8);
 #endif
+void (PNGAPI *qpng_set_bgr) PNGARG((png_structp png_ptr)) PSTATIC(png_set_bgr);
 void (PNGAPI *qpng_set_filler) PNGARG((png_structp png_ptr, png_uint_32 filler, int flags)) PSTATIC(png_set_filler);
 void (PNGAPI *qpng_set_palette_to_rgb) PNGARG((png_structp png_ptr)) PSTATIC(png_set_palette_to_rgb);
 png_uint_32 (PNGAPI *qpng_get_IHDR) PNGARG((png_structp png_ptr, png_infop info_ptr, png_uint_32 *width, png_uint_32 *height,
@@ -767,6 +768,7 @@ qboolean LibPNG_Init(void)
                 #else
 		{(void **) &qpng_set_gray_1_2_4_to_8,	"png_set_gray_1_2_4_to_8"},
                 #endif
+		{(void **) &qpng_set_bgr,						"png_set_bgr"},
 		{(void **) &qpng_set_filler,					"png_set_filler"},
 		{(void **) &qpng_set_palette_to_rgb,			"png_set_palette_to_rgb"},
 		{(void **) &qpng_get_IHDR,						"png_get_IHDR"},
@@ -979,7 +981,7 @@ error:
 
 
 #ifndef NPFTE
-int Image_WritePNG (char *filename, enum fs_relative fsroot, int compression, qbyte *pixels, int width, int height)
+int Image_WritePNG (char *filename, enum fs_relative fsroot, int compression, qbyte *pixels, int width, int height, enum uploadfmt fmt)
 {
 	char name[MAX_OSPATH];
 	int i;
@@ -988,6 +990,7 @@ int Image_WritePNG (char *filename, enum fs_relative fsroot, int compression, qb
 	png_infop info_ptr;
 	png_byte **row_pointers;
 	struct pngerr errctx;
+	int pxsize;
 
 	if (!FS_NativePath(filename, fsroot, name, sizeof(name)))
 		return false;
@@ -1036,14 +1039,25 @@ err:
 #endif
 	qpng_set_compression_level(png_ptr, Z_NO_COMPRESSION + (compression*(Z_BEST_COMPRESSION-Z_NO_COMPRESSION))/100);
 
-	qpng_set_IHDR(png_ptr, info_ptr, width, height, 8, PNG_COLOR_TYPE_RGB, PNG_INTERLACE_NONE, PNG_COMPRESSION_TYPE_DEFAULT, PNG_FILTER_TYPE_DEFAULT);
+	if (fmt == TF_BGR24 || fmt == TF_BGRA32)
+		qpng_set_bgr(png_ptr);
+	if (fmt == TF_RGBA32 || fmt == TF_BGRA32)
+	{
+		pxsize = 4;
+		qpng_set_IHDR(png_ptr, info_ptr, width, height, 8, PNG_COLOR_TYPE_RGBA, PNG_INTERLACE_NONE, PNG_COMPRESSION_TYPE_DEFAULT, PNG_FILTER_TYPE_DEFAULT);
+	}
+	else
+	{
+		pxsize = 3;
+		qpng_set_IHDR(png_ptr, info_ptr, width, height, 8, PNG_COLOR_TYPE_RGB, PNG_INTERLACE_NONE, PNG_COMPRESSION_TYPE_DEFAULT, PNG_FILTER_TYPE_DEFAULT);
+	}
 	qpng_write_info(png_ptr, info_ptr);
 
 	row_pointers = BZ_Malloc (sizeof(png_byte *) * height);
 	if (!row_pointers)
 		goto err;
 	for (i = 0; i < height; i++)
-		row_pointers[height - i - 1] = pixels + i * width * 3;
+		row_pointers[height - i - 1] = pixels + i * width * pxsize;
 	qpng_write_image(png_ptr, row_pointers);
 	qpng_write_end(png_ptr, info_ptr);
 	BZ_Free(row_pointers);
@@ -1463,6 +1477,19 @@ badjpeg:
 /*end read*/
 #ifndef NPFTE
 /*begin write*/
+
+
+#ifndef DYNAMIC_LIBJPEG
+#define qjpeg_std_error			jpeg_std_error
+#define qjpeg_destroy_compress	jpeg_destroy_compress
+#define qjpeg_CreateCompress	jpeg_CreateCompress
+#define qjpeg_set_defaults		jpeg_set_defaults
+#define qjpeg_set_quality		jpeg_set_quality
+#define qjpeg_start_compress	jpeg_start_compress
+#define qjpeg_write_scanlines	jpeg_write_scanlines
+#define qjpeg_finish_compress	jpeg_finish_compress
+#define qjpeg_destroy_compress	jpeg_destroy_compress
+#endif
 #define OUTPUT_BUF_SIZE 4096
 typedef struct  {
 	struct jpeg_error_mgr pub;
@@ -1535,7 +1562,7 @@ METHODDEF(void) jpeg_error_exit (j_common_ptr cinfo)
 {
   longjmp(((jpeg_error_mgr_wrapper *) cinfo->err)->setjmp_buffer, 1);
 }
-qboolean screenshotJPEG(char *filename, enum fs_relative fsroot, int compression, qbyte *screendata, int screenwidth, int screenheight)	//input is rgb NOT rgba
+qboolean screenshotJPEG(char *filename, enum fs_relative fsroot, int compression, qbyte *screendata, int screenwidth, int screenheight, enum uploadfmt fmt)
 {
 	qbyte	*buffer;
 	vfsfile_t	*outfile;
@@ -1543,6 +1570,9 @@ qboolean screenshotJPEG(char *filename, enum fs_relative fsroot, int compression
 	struct jpeg_compress_struct cinfo;
 	JSAMPROW row_pointer[1];
 
+	if (fmt != TF_RGB24)
+		return false;
+
 	if (!LIBJPEG_LOADED())
 		return false;
 
@@ -1556,29 +1586,17 @@ qboolean screenshotJPEG(char *filename, enum fs_relative fsroot, int compression
 		}
 	}
 
-	#ifdef DYNAMIC_LIBJPEG
-		cinfo.err = qjpeg_std_error(&jerr.pub);
-	#else
-		cinfo.err = jpeg_std_error(&jerr.pub);
-	#endif
+	cinfo.err = qjpeg_std_error(&jerr.pub);
 	jerr.pub.error_exit = jpeg_error_exit;
 	if (setjmp(jerr.setjmp_buffer))
 	{
-		#ifdef DYNAMIC_LIBJPEG
-			qjpeg_destroy_compress(&cinfo);
-		#else
-			jpeg_destroy_compress(&cinfo);
-		#endif
+		qjpeg_destroy_compress(&cinfo);
 		VFS_CLOSE(outfile);
 		FS_Remove(filename, FS_GAME);
 		Con_Printf("Failed to create jpeg\n");
 		return false;
 	}
-	#ifdef DYNAMIC_LIBJPEG
-		qjpeg_create_compress(&cinfo);
-	#else
-		jpeg_create_compress(&cinfo);
-	#endif
+	qjpeg_create_compress(&cinfo);
 
 	buffer = screendata;
 
@@ -1587,42 +1605,18 @@ qboolean screenshotJPEG(char *filename, enum fs_relative fsroot, int compression
 	cinfo.image_height = screenheight;
 	cinfo.input_components = 3;
 	cinfo.in_color_space = JCS_RGB;
-	#ifdef DYNAMIC_LIBJPEG
-		qjpeg_set_defaults(&cinfo);
-	#else
-		jpeg_set_defaults(&cinfo);
-	#endif
-	#ifdef DYNAMIC_LIBJPEG
-		qjpeg_set_quality (&cinfo, bound(0, compression, 100), true);
-	#else
-		jpeg_set_quality (&cinfo, bound(0, compression, 100), true);
-	#endif
-	#ifdef DYNAMIC_LIBJPEG
-		qjpeg_start_compress(&cinfo, true);
-	#else
-		jpeg_start_compress(&cinfo, true);
-	#endif
+	qjpeg_set_defaults(&cinfo);
+	qjpeg_set_quality (&cinfo, bound(0, compression, 100), true);
+	qjpeg_start_compress(&cinfo, true);
 
 	while (cinfo.next_scanline < cinfo.image_height)
 	{
 		*row_pointer = &buffer[(cinfo.image_height - cinfo.next_scanline - 1) * cinfo.image_width * 3];
-		#ifdef DYNAMIC_LIBJPEG
-			qjpeg_write_scanlines(&cinfo, row_pointer, 1);
-		#else
-			jpeg_write_scanlines(&cinfo, row_pointer, 1);
-		#endif
+		qjpeg_write_scanlines(&cinfo, row_pointer, 1);
 	}
-	#ifdef DYNAMIC_LIBJPEG
-		qjpeg_finish_compress(&cinfo);
-	#else
-		jpeg_finish_compress(&cinfo);
-	#endif
+	qjpeg_finish_compress(&cinfo);
 	VFS_CLOSE(outfile);
-	#ifdef DYNAMIC_LIBJPEG
-		qjpeg_destroy_compress(&cinfo);
-	#else
-		jpeg_destroy_compress(&cinfo);
-	#endif
+	qjpeg_destroy_compress(&cinfo);
 	return true;
 }
 #endif
@@ -3155,21 +3149,24 @@ static void Image_RoundDimensions(int *scaled_width, int *scaled_height, unsigne
 		}
 	}
 
-	if (flags & IF_NOMIPMAP)
+	if (!(flags & IF_NOPICMIP))
 	{
-		if (gl_picmip2d.ival > 0)
+		if (flags & IF_NOMIPMAP)
 		{
-			*scaled_width >>= gl_picmip2d.ival;
-			*scaled_height >>= gl_picmip2d.ival;
+			if (gl_picmip2d.ival > 0)
+			{
+				*scaled_width >>= gl_picmip2d.ival;
+				*scaled_height >>= gl_picmip2d.ival;
+			}
 		}
-	}
-	else
-	{
-		if (gl_picmip.ival > 0)
+		else
 		{
-			TRACE(("dbg: GL_RoundDimensions: %f\n", gl_picmip.value));
-			*scaled_width >>= gl_picmip.ival;
-			*scaled_height >>= gl_picmip.ival;
+			if (gl_picmip.ival > 0)
+			{
+				TRACE(("dbg: GL_RoundDimensions: %f\n", gl_picmip.value));
+				*scaled_width >>= gl_picmip.ival;
+				*scaled_height >>= gl_picmip.ival;
+			}
 		}
 	}
 
@@ -3500,6 +3497,7 @@ static qboolean Image_GenMip0(struct pendingtextureinfo *mips, unsigned int flag
 	case TF_MIP4_LUM8:
 		//8bit opaque data
 		Image_RoundDimensions(&mips->mip[0].width, &mips->mip[0].height, flags);
+		flags |= IF_NOPICMIP;
 		if (mips->mip[0].width == imgwidth && mips->mip[0].height == imgheight)
 		{
 			unsigned int pixels =
@@ -3537,6 +3535,7 @@ static qboolean Image_GenMip0(struct pendingtextureinfo *mips, unsigned int flag
 	case TF_MIP4_SOLID8:
 		//8bit opaque data
 		Image_RoundDimensions(&mips->mip[0].width, &mips->mip[0].height, flags);
+		flags |= IF_NOPICMIP;
 		if (mips->mip[0].width == imgwidth && mips->mip[0].height == imgheight && sh_config.texfmt[PTI_RGBX8])
 		{
 			unsigned int pixels =
diff --git a/engine/client/m_mp3.c b/engine/client/m_mp3.c
index 4c368eccb..c0661b01b 100644
--- a/engine/client/m_mp3.c
+++ b/engine/client/m_mp3.c
@@ -697,7 +697,6 @@ void Media_LoadTrackNames (char *listname);
 
 void M_Media_Draw (menu_t *menu)
 {
-	mpic_t	*p;
 	mediatrack_t *track;
 	int y;
 	int op, i;
@@ -2672,6 +2671,12 @@ int captureoldfbo;
 qboolean capturingfbo;
 texid_t	capturetexture;
 qboolean captureframeforce;
+#ifdef GLQUAKE
+//ring buffer
+int pbo_handles[4];
+enum uploadfmt pbo_format;
+#endif
+int pbo_oldest;
 
 qboolean capturepaused;
 extern cvar_t vid_conautoscale;
@@ -2751,13 +2756,13 @@ static void *QDECL capture_raw_begin (char *streamname, int videorate, int width
 	}
 	return ctx;
 }
-static void QDECL capture_raw_video (void *vctx, void *data, int frame, int width, int height)
+static void QDECL capture_raw_video (void *vctx, void *data, int frame, int width, int height, enum uploadfmt fmt)
 {
 	struct capture_raw_ctx *ctx = vctx;
 	char filename[MAX_OSPATH];
 	ctx->frames = frame+1;
 	Q_snprintfz(filename, sizeof(filename), "%s%8.8i.%s", ctx->videonameprefix, frame, ctx->videonameextension);
-	SCR_ScreenShot(filename, ctx->fsroot, data, width, height);
+	SCR_ScreenShot(filename, ctx->fsroot, data, width, height, fmt);
 }
 static void QDECL capture_raw_audio (void *vctx, void *data, int bytes)
 {
@@ -2949,19 +2954,39 @@ static void *QDECL capture_avi_begin (char *streamname, int videorate, int width
 	return ctx;
 }
 
-static void QDECL capture_avi_video(void *vctx, void *vdata, int frame, int width, int height)
+static void QDECL capture_avi_video(void *vctx, void *vdata, int frame, int width, int height, enum uploadfmt fmt)
 {
 	struct capture_avi_ctx *ctx = vctx;
 	qbyte *data = vdata;
 	int c, i;
 	qbyte temp;
-	// swap rgb to bgr
-	c = width*height*3;
-	for (i=0 ; i<c ; i+=3)
+
+	if (fmt == TF_BGRA32)
 	{
-		temp = data[i];
-		data[i] = data[i+2];
-		data[i+2] = temp;
+		// truncate bgra to bgr
+		c = width*height;
+		for (i=0 ; i<c ; i++)
+		{
+			data[i*3+0] = data[i*4+0];
+			data[i*3+1] = data[i*4+1];
+			data[i*3+2] = data[i*4+2];
+		}
+	}
+	else if (fmt == TF_RGB24)
+	{
+		// swap rgb to bgr
+		c = width*height*3;
+		for (i=0 ; i<c ; i+=3)
+		{
+			temp = data[i];
+			data[i] = data[i+2];
+			data[i+2] = temp;
+		}
+	}
+	else if (fmt != TF_BGR24)
+	{
+		Con_Printf("Unsupported image format\n");
+		return;
 	}
 	//write it
 	if (FAILED(qAVIStreamWrite(avi_video_stream(ctx), frame, 1, data, width*height * 3, ((frame%15) == 0)?AVIIF_KEYFRAME:0, NULL, NULL)))
@@ -2993,7 +3018,7 @@ static void *QDECL capture_null_begin (char *streamname, int videorate, int widt
 {
 	return (void*)~0;
 }
-static void QDECL capture_null_video(void *vctx, void *vdata, int frame, int width, int height)
+static void QDECL capture_null_video(void *vctx, void *vdata, int frame, int width, int height, enum uploadfmt fmt)
 {
 }
 static void QDECL capture_null_audio(void *vctx, void *data, int bytes)
@@ -3098,6 +3123,7 @@ void Media_RecordFrame (void)
 {
 	char *buffer;
 	int truewidth, trueheight;
+	enum uploadfmt fmt;
 
 	if (!currentcapture_funcs)
 		return;
@@ -3162,17 +3188,65 @@ void Media_RecordFrame (void)
 	if (R2D_Flush)
 		R2D_Flush();
 
-	//submit the current video frame. audio will be mixed to match.
-	buffer = VID_GetRGBInfo(0, &truewidth, &trueheight);
-	if (buffer)
+#ifdef GLQUAKE
+	if (pbo_format != TF_INVALID)
 	{
-		currentcapture_funcs->capture_video(currentcapture_ctx, buffer, captureframe, truewidth, trueheight);
-		capturelastvideotime += captureframeinterval;
-		captureframe++;
-		BZ_Free (buffer);
+		int imagesize = vid.fbpwidth * vid.fbpheight * 4;
+		while (pbo_oldest + countof(pbo_handles) <= captureframe)
+		{
+			qglBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, pbo_handles[pbo_oldest%countof(pbo_handles)]);
+			buffer = qglMapBufferARB(GL_PIXEL_PACK_BUFFER_ARB, GL_READ_ONLY_ARB);
+			if (buffer)
+			{
+				currentcapture_funcs->capture_video(currentcapture_ctx, buffer, pbo_oldest, vid.fbpwidth, vid.fbpheight, pbo_format);
+				qglUnmapBufferARB(GL_PIXEL_PACK_BUFFER_ARB);
+			}
+			pbo_oldest++;
+		}
+
+		if (!pbo_handles[captureframe%countof(pbo_handles)])
+		{
+			qglGenBuffersARB(1, &pbo_handles[captureframe%countof(pbo_handles)]);
+			qglBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, pbo_handles[captureframe%countof(pbo_handles)]);
+			qglBufferDataARB(GL_PIXEL_PACK_BUFFER_ARB, imagesize, NULL, GL_STATIC_READ_ARB);
+		}
+		qglBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, pbo_handles[captureframe%countof(pbo_handles)]);
+		switch(pbo_format)
+		{
+		case TF_BGR24:
+			qglReadPixels(0, 0, vid.fbpwidth, vid.fbpheight, GL_BGR_EXT, GL_UNSIGNED_BYTE, 0);
+			break;
+		case TF_BGRA32:
+			qglReadPixels(0, 0, vid.fbpwidth, vid.fbpheight, GL_BGRA_EXT, GL_UNSIGNED_INT_8_8_8_8_REV, 0);
+			break;
+		case TF_RGB24:
+			qglReadPixels(0, 0, vid.fbpwidth, vid.fbpheight, GL_RGB, GL_UNSIGNED_BYTE, 0);
+			break;
+		case TF_RGBA32:
+			qglReadPixels(0, 0, vid.fbpwidth, vid.fbpheight, GL_RGBA, GL_UNSIGNED_BYTE, 0);
+			break;
+		}
+		qglBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, 0);
 	}
 	else
-		Con_DPrintf("Unable to grab video image\n");
+#endif
+	{
+		pbo_oldest = captureframe+1;
+		//submit the current video frame. audio will be mixed to match.
+		buffer = VID_GetRGBInfo(&truewidth, &trueheight, &fmt);
+		if (buffer)
+		{
+			currentcapture_funcs->capture_video(currentcapture_ctx, buffer, captureframe, truewidth, trueheight, fmt);
+			BZ_Free (buffer);
+		}
+		else
+		{
+			Con_DPrintf("Unable to grab video image\n");
+			currentcapture_funcs->capture_video(currentcapture_ctx, NULL, captureframe, 0, 0, TF_INVALID);
+		}
+	}
+	captureframe++;
+	capturelastvideotime += captureframeinterval;
 
 	captureframeforce = false;
 
@@ -3358,6 +3432,35 @@ void Media_InitFakeSoundDevice (int speed, int channels, int samplebits)
 
 void Media_StopRecordFilm_f (void)
 {
+#ifdef GLQUAKE
+	if (pbo_format)
+	{
+		int i;
+		int imagesize = vid.fbpwidth * vid.fbpheight * 4;
+		while (pbo_oldest < captureframe)
+		{
+			qbyte *buffer;
+			qglBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, pbo_handles[pbo_oldest%countof(pbo_handles)]);
+			buffer = qglMapBufferARB(GL_PIXEL_PACK_BUFFER_ARB, GL_READ_ONLY_ARB);
+			if (buffer)
+			{
+				currentcapture_funcs->capture_video(currentcapture_ctx, buffer, pbo_oldest, vid.fbpwidth, vid.fbpheight, TF_BGR24);
+//				currentcapture_funcs->capture_video(currentcapture_ctx, buffer, pbo_oldest, vid.fbpwidth, vid.fbpheight, TF_BGRA32);
+				qglUnmapBufferARB(GL_PIXEL_PACK_BUFFER_ARB);
+			}
+			pbo_oldest++;
+		}
+
+		for (i = 0; i < countof(pbo_handles); i++)
+		{
+			if (pbo_handles[i])
+				qglDeleteBuffersARB(1, &pbo_handles[i]);
+			pbo_handles[i] = 0;
+		}
+	}
+#endif
+
+
 	if (capture_fakesounddevice)
 		S_ShutdownCard(capture_fakesounddevice);
 	capture_fakesounddevice = NULL;
@@ -3401,7 +3504,7 @@ static void Media_RecordFilm (char *recordingname, qboolean demo)
 
 	Con_ClearNotify();
 
-	captureframe = 0;
+	captureframe = pbo_oldest = 0;
 	for (i = 0; i < sizeof(pluginencodersfunc)/sizeof(pluginencodersfunc[0]); i++)
 	{
 		if (pluginencodersfunc[i])
@@ -3438,11 +3541,17 @@ static void Media_RecordFilm (char *recordingname, qboolean demo)
 	{
 		capturingfbo = true;
 		capturetexture = R2D_RT_Configure("$democapture", capturewidth.ival, captureheight.ival, TF_BGRA32);
-		captureoldfbo = GLBE_FBO_Update(&capturefbo, FBO_RB_DEPTH, &capturetexture, 1, r_nulltex, capturewidth.ival, captureheight.ival, 0);
+		captureoldfbo = GLBE_FBO_Update(&capturefbo, FBO_RB_DEPTH|(Sh_StencilShadowsActive()?FBO_RB_STENCIL:0), &capturetexture, 1, r_nulltex, capturewidth.ival, captureheight.ival, 0);
 		vid.fbpwidth = capturewidth.ival;
 		vid.fbpheight = captureheight.ival;
 		vid.framebuffer = capturetexture;
 	}
+
+	pbo_format = TF_INVALID;
+	if (qrenderer == QR_OPENGL && !gl_config.gles && gl_config.glversion >= 2.1)
+	{	//both tgas and vfw favour bgr24, so lets get the gl drivers to suffer instead of us.
+		pbo_format = TF_BGR24;
+	}
 #endif
 
 	recordingdemo = demo;
diff --git a/engine/client/merged.h b/engine/client/merged.h
index 23bb35428..8a780724e 100644
--- a/engine/client/merged.h
+++ b/engine/client/merged.h
@@ -101,7 +101,7 @@ extern void	(*R_RenderView)							(void);		// must set r_refdef first
 
 extern qboolean	(*VID_Init)							(rendererstate_t *info, unsigned char *palette);
 extern void	(*VID_DeInit)							(void);
-extern char *(*VID_GetRGBInfo)						(int prepad, int *truevidwidth, int *truevidheight);
+extern char *(*VID_GetRGBInfo)						(int *truevidwidth, int *truevidheight, enum uploadfmt *fmt);
 extern void	(*VID_SetWindowCaption)					(char *msg);
 
 extern void SCR_Init								(void);
@@ -413,7 +413,7 @@ typedef struct rendererinfo_s {
 	void (*VID_DestroyCursor)			(void *cursor);	//may be null
 
 	void	(*VID_SetWindowCaption)		(char *msg);
-	char	*(*VID_GetRGBInfo)			(int prepad, int *truevidwidth, int *truevidheight);
+	char	*(*VID_GetRGBInfo)			(int *truevidwidth, int *truevidheight, enum uploadfmt *fmt);
 
 	void	(*SCR_UpdateScreen)			(void);
 
diff --git a/engine/client/pr_menu.c b/engine/client/pr_menu.c
index 17c6a444a..813a2b231 100644
--- a/engine/client/pr_menu.c
+++ b/engine/client/pr_menu.c
@@ -55,6 +55,9 @@ void QCBUILTIN PF_CL_drawfill (pubprogfuncs_t *prinst, struct globalvars_s *pr_g
 void QCBUILTIN PF_CL_drawsetcliparea (pubprogfuncs_t *prinst, struct globalvars_s *pr_globals)
 {
 	srect_t srect;
+	if (R2D_Flush)
+		R2D_Flush();
+
 	csqc_dp_lastwas3d = false;
 
 	srect.x = G_FLOAT(OFS_PARM0) / (float)vid.fbvwidth;
@@ -71,6 +74,9 @@ void QCBUILTIN PF_CL_drawsetcliparea (pubprogfuncs_t *prinst, struct globalvars_
 //void	drawresetcliparea(void) = #459;
 void QCBUILTIN PF_CL_drawresetcliparea (pubprogfuncs_t *prinst, struct globalvars_s *pr_globals)
 {
+	if (R2D_Flush)
+		R2D_Flush();
+
 	csqc_dp_lastwas3d = false;
 
 	BE_Scissor(NULL);
diff --git a/engine/client/r_2d.c b/engine/client/r_2d.c
index 6d2f03949..9c87ea815 100644
--- a/engine/client/r_2d.c
+++ b/engine/client/r_2d.c
@@ -31,6 +31,7 @@ shader_t *shader_polyblend;
 shader_t *shader_menutint;
 
 #define DRAW_QUADS 128
+static int		draw_active_flags;
 static shader_t *draw_active_shader;
 static avec4_t	draw_active_colour;
 static mesh_t	draw_mesh;
@@ -576,7 +577,7 @@ void R2D_ImageAtlas(float x, float y, float w, float h, float s1, float t1, floa
 
 void R2D_ImageFlush(void)
 {
-	BE_DrawMesh_Single(draw_active_shader, &draw_mesh, NULL, r2d_be_flags);
+	BE_DrawMesh_Single(draw_active_shader, &draw_mesh, NULL, draw_active_flags);
 
 	R2D_Flush = NULL;
 	draw_active_shader = NULL;
@@ -597,12 +598,13 @@ void R2D_Image(float x, float y, float w, float h, float s1, float t1, float s2,
 			return;
 	}
 
-	if (draw_active_shader != pic || draw_mesh.numvertexes+4 > DRAW_QUADS)
+	if (draw_active_shader != pic || draw_active_flags != r2d_be_flags || draw_mesh.numvertexes+4 > DRAW_QUADS)
 	{
 		if (R2D_Flush)
 			R2D_Flush();
 
 		draw_active_shader = pic;
+		draw_active_flags = r2d_be_flags;
 		R2D_Flush = R2D_ImageFlush;
 
 		draw_mesh.numindexes = 0;
@@ -663,12 +665,13 @@ void R2D_FillBlock(float x, float y, float w, float h)
 		pic = shader_draw_fill_trans;
 	else
 		pic = shader_draw_fill;
-	if (draw_active_shader != pic || draw_mesh.numvertexes+4 > DRAW_QUADS)
+	if (draw_active_shader != pic || draw_active_flags != r2d_be_flags || draw_mesh.numvertexes+4 > DRAW_QUADS)
 	{
 		if (R2D_Flush)
 			R2D_Flush();
 
 		draw_active_shader = pic;
+		draw_active_flags = r2d_be_flags;
 		R2D_Flush = R2D_ImageFlush;
 
 		draw_mesh.numindexes = 0;
diff --git a/engine/client/renderer.c b/engine/client/renderer.c
index febcbb36c..f07bd3bb4 100644
--- a/engine/client/renderer.c
+++ b/engine/client/renderer.c
@@ -856,7 +856,7 @@ void	(*R_RenderView)				(void);		// must set r_refdef first
 
 qboolean (*VID_Init)				(rendererstate_t *info, unsigned char *palette);
 void	 (*VID_DeInit)				(void);
-char	*(*VID_GetRGBInfo)			(int prepad, int *truevidwidth, int *truevidheight);
+char	*(*VID_GetRGBInfo)			(int *truevidwidth, int *truevidheight, enum uploadfmt *fmt);
 void	(*VID_SetWindowCaption)		(char *msg);
 
 void	(*SCR_UpdateScreen)			(void);
diff --git a/engine/client/screen.h b/engine/client/screen.h
index 3d67d5ec3..5c86808e3 100644
--- a/engine/client/screen.h
+++ b/engine/client/screen.h
@@ -80,7 +80,8 @@ void SCR_ShowPic_Remove_f(void);
 //a header is better than none...
 void Draw_TextBox (int x, int y, int width, int lines);
 enum fs_relative;
-qboolean SCR_ScreenShot (char *filename, enum fs_relative fsroot, void *rgb_buffer, int width, int height);
+enum uploadfmt;
+qboolean SCR_ScreenShot (char *filename, enum fs_relative fsroot, void *rgb_buffer, int width, int height, enum uploadfmt fmt);
 
 void SCR_DrawTwoDimensional(int uimenu, qboolean nohud);
 
diff --git a/engine/client/vid.h b/engine/client/vid.h
index 039d58805..abea41160 100644
--- a/engine/client/vid.h
+++ b/engine/client/vid.h
@@ -110,6 +110,6 @@ int GLVID_SetMode (rendererstate_t *info, unsigned char *palette);
 qboolean GLVID_Is8bit(void);
 
 void GLVID_SwapBuffers(void);
-char *GLVID_GetRGBInfo(int prepadbytes, int *truewidth, int *trueheight);
+char *GLVID_GetRGBInfo(int *truewidth, int *trueheight, enum uploadfmt *fmt);
 void GLVID_SetCaption(char *caption);
 #endif
diff --git a/engine/client/vid_headless.c b/engine/client/vid_headless.c
index 5d1d3e3e1..f4c355724 100644
--- a/engine/client/vid_headless.c
+++ b/engine/client/vid_headless.c
@@ -163,8 +163,9 @@ static qboolean Headless_VID_ApplyGammaRamps		(unsigned short *ramps)
 static void	Headless_VID_SetWindowCaption		(char *msg)
 {
 }
-static char	*Headless_VID_GetRGBInfo			(int prepad, int *truevidwidth, int *truevidheight)
+static char	*Headless_VID_GetRGBInfo			(int *truevidwidth, int *truevidheight, enum uploadfmt *fmt)
 {
+	*fmt = TF_INVALID;
 	*truevidwidth = *truevidheight = 0;
 	return NULL;
 }
diff --git a/engine/d3d/vid_d3d.c b/engine/d3d/vid_d3d.c
index 891536113..8b58ba3f2 100644
--- a/engine/d3d/vid_d3d.c
+++ b/engine/d3d/vid_d3d.c
@@ -802,7 +802,7 @@ qboolean D3D9_VID_ApplyGammaRamps		(unsigned short *ramps)
 		IDirect3DDevice9_SetGammaRamp(pD3DDev9, 0, D3DSGR_NO_CALIBRATION, (D3DGAMMARAMP *)ramps);
 	return true;
 }
-static char	*(D3D9_VID_GetRGBInfo)			(int prepad, int *truevidwidth, int *truevidheight)
+static char	*(D3D9_VID_GetRGBInfo)			(int *truevidwidth, int *truevidheight, enum uploadfmt *fmt)
 {
 	IDirect3DSurface9 *backbuf, *surf;
 	D3DLOCKED_RECT rect;
@@ -828,14 +828,16 @@ static char	*(D3D9_VID_GetRGBInfo)			(int prepad, int *truevidwidth, int *truevi
 			if (!FAILED(IDirect3DDevice9_GetRenderTargetData(pD3DDev9, backbuf, surf)))
 			if (!FAILED(IDirect3DSurface9_LockRect(surf, &rect, NULL, D3DLOCK_NO_DIRTY_UPDATE|D3DLOCK_READONLY|D3DLOCK_NOSYSLOCK)))
 			{
-				ret = BZ_Malloc(prepad + desc.Width*desc.Height*3);
+				ret = BZ_Malloc(desc.Width*desc.Height*3);
 				if (ret)
 				{
+					*fmt = TF_RGB24;
+
 					// read surface rect and convert 32 bgra to 24 rgb and flip
-					c = prepad+desc.Width*desc.Height*3;
+					c = desc.Width*desc.Height*3;
 					p = (qbyte *)rect.pBits;
 
-					for (i=c-(3*desc.Width); i>=prepad; i-=(3*desc.Width))
+					for (i=c-(3*desc.Width); i>=0; i-=(3*desc.Width))
 					{
 						for (j=0; j<desc.Width; j++)
 						{
@@ -1239,7 +1241,6 @@ qboolean (D3D9_VID_Init)				(rendererstate_t *info, unsigned char *palette);
 void	 (D3D9_VID_DeInit)				(void);
 void	(D3D9_VID_SetPalette)			(unsigned char *palette);
 void	(D3D9_VID_ShiftPalette)			(unsigned char *palette);
-char	*(D3D9_VID_GetRGBInfo)			(int prepad, int *truevidwidth, int *truevidheight);
 void	(D3D9_VID_SetWindowCaption)		(char *msg);
 
 void	(D3D9_SCR_UpdateScreen)			(void);
diff --git a/engine/d3d/vid_d3d11.c b/engine/d3d/vid_d3d11.c
index 47123c2c1..0b4c15b61 100644
--- a/engine/d3d/vid_d3d11.c
+++ b/engine/d3d/vid_d3d11.c
@@ -1102,7 +1102,7 @@ static qboolean	D3D11_VID_ApplyGammaRamps(unsigned short *ramps)
 	}
 	return false;
 }
-static char	*D3D11_VID_GetRGBInfo(int prepad, int *truevidwidth, int *truevidheight)
+static char	*D3D11_VID_GetRGBInfo(int *truevidwidth, int *truevidheight, enum uploadfmt *fmt)
 {
 	//don't directly map the frontbuffer, as that can hold other things.
 	//create a texture, copy the (gpu)backbuffer to that (cpu)texture
@@ -1132,8 +1132,7 @@ static char	*D3D11_VID_GetRGBInfo(int prepad, int *truevidwidth, int *truevidhei
 	ID3D11Resource_Release(backbuffer);
 	if (!FAILED(ID3D11DeviceContext_Map(d3ddevctx, (ID3D11Resource*)texture, 0, D3D11_MAP_READ, 0, &lock)))
 	{
-		r = rgb = BZ_Malloc(prepad + 3 * vid.pixelwidth * vid.pixelheight);
-		r += prepad;
+		r = rgb = BZ_Malloc(3 * vid.pixelwidth * vid.pixelheight);
 		for (y = vid.pixelheight; y-- > 0; )
 		{
 			in = lock.pData;
@@ -1150,6 +1149,7 @@ static char	*D3D11_VID_GetRGBInfo(int prepad, int *truevidwidth, int *truevidhei
 	ID3D11Texture2D_Release(texture);
 	*truevidwidth = vid.pixelwidth;
 	*truevidheight = vid.pixelheight;
+	*fmt = TF_RGB24;
 	return r;
 }
 static void	(D3D11_VID_SetWindowCaption)		(char *msg)
diff --git a/engine/gl/gl_font.c b/engine/gl/gl_font.c
index 9a21fbdeb..ef0f346f8 100644
--- a/engine/gl/gl_font.c
+++ b/engine/gl/gl_font.c
@@ -28,6 +28,7 @@ int Font_LineBreaks(conchar_t *start, conchar_t *end, int maxpixelwidth, int max
 struct font_s *font_default;
 struct font_s *font_console;
 struct font_s *font_tiny;
+static int font_be_flags;
 extern unsigned int r2d_be_flags;
 
 //by adding 'extern' to one definition of a function in a translation unit, then the definition in that TU is NOT considered an inline definition. meaning non-inlined references in other TUs can link to it instead of their own if needed.
@@ -409,10 +410,10 @@ static void Font_Flush(void)
 		font_backmesh.numvertexes = font_foremesh.numvertexes;
 		font_backmesh.istrifan = font_foremesh.istrifan;
 
-		BE_DrawMesh_Single(fontplanes.backshader, &font_backmesh, NULL, r2d_be_flags);
+		BE_DrawMesh_Single(fontplanes.backshader, &font_backmesh, NULL, font_be_flags);
 	}
 	TEXASSIGN(fontplanes.shader->defaulttextures->base, font_texture);
-	BE_DrawMesh_Single(fontplanes.shader, &font_foremesh, NULL, r2d_be_flags);
+	BE_DrawMesh_Single(fontplanes.shader, &font_foremesh, NULL, font_be_flags);
 	font_foremesh.numindexes = 0;
 	font_foremesh.numvertexes = 0;
 }
@@ -1523,10 +1524,10 @@ void Font_Free(struct font_s *f)
 //maps a given virtual screen coord to a pixel coord, which matches the font's height/width values
 void Font_BeginString(struct font_s *font, float vx, float vy, int *px, int *py)
 {
-	if (R2D_Flush && (R2D_Flush != Font_Flush || curfont != font))
+	if (R2D_Flush && (R2D_Flush != Font_Flush || curfont != font || font_be_flags != r2d_be_flags))
 		R2D_Flush();
 	R2D_Flush = Font_Flush;
-
+	font_be_flags = r2d_be_flags;
 	curfont = font;
 	*px = (vx*(int)vid.rotpixelwidth) / (float)vid.width;
 	*py = (vy*(int)vid.rotpixelheight) / (float)vid.height;
@@ -1544,10 +1545,10 @@ void Font_Transform(float vx, float vy, int *px, int *py)
 }
 void Font_BeginScaledString(struct font_s *font, float vx, float vy, float szx, float szy, float *px, float *py)
 {
-	if (R2D_Flush && (R2D_Flush != Font_Flush || curfont != font))
+	if (R2D_Flush && (R2D_Flush != Font_Flush || curfont != font || font_be_flags != r2d_be_flags))
 		R2D_Flush();
 	R2D_Flush = Font_Flush;
-
+	font_be_flags = r2d_be_flags;
 	curfont = font;
 	*px = (vx*(float)vid.rotpixelwidth) / (float)vid.width;
 	*py = (vy*(float)vid.rotpixelheight) / (float)vid.height;
diff --git a/engine/gl/gl_screen.c b/engine/gl/gl_screen.c
index b67a554cf..dce08e0e5 100644
--- a/engine/gl/gl_screen.c
+++ b/engine/gl/gl_screen.c
@@ -243,7 +243,7 @@ void GLSCR_UpdateScreen (void)
 }
 
 
-char *GLVID_GetRGBInfo(int prepadbytes, int *truewidth, int *trueheight)
+char *GLVID_GetRGBInfo(int *truewidth, int *trueheight, enum uploadfmt *fmt)
 {	//returns a BZ_Malloced array
 	extern qboolean gammaworks;
 	int i, c;
@@ -259,14 +259,14 @@ char *GLVID_GetRGBInfo(int prepadbytes, int *truewidth, int *trueheight)
 		p = BZ_Malloc(vid.pixelwidth*vid.pixelheight*sizeof(float));
 		qglReadPixels (0, 0, vid.pixelwidth, vid.pixelheight, GL_DEPTH_COMPONENT, GL_FLOAT, p); 
 
-		ret = BZ_Malloc(prepadbytes + vid.pixelwidth*vid.pixelheight*3);
+		ret = BZ_Malloc(vid.pixelwidth*vid.pixelheight*3);
 
 		c = vid.pixelwidth*vid.pixelheight;
 		for (i = 1; i < c; i++)
 		{
-			ret[prepadbytes+i*3+0]=p[i]*p[i]*p[i]*255;
-			ret[prepadbytes+i*3+1]=p[i]*p[i]*p[i]*255;
-			ret[prepadbytes+i*3+2]=p[i]*p[i]*p[i]*255;
+			ret[i*3+0]=p[i]*p[i]*p[i]*255;
+			ret[i*3+1]=p[i]*p[i]*p[i]*255;
+			ret[i*3+2]=p[i]*p[i]*p[i]*255;
 		}
 		BZ_Free(p);
 	}
@@ -274,30 +274,46 @@ char *GLVID_GetRGBInfo(int prepadbytes, int *truewidth, int *trueheight)
 	{
 		qbyte *p;
 
-		// gles only guarantees GL_RGBA/GL_UNSIGNED_BYTE so downconvert and resize
-		ret = BZ_Malloc(prepadbytes + (*truewidth)*(*trueheight)*4);
-		qglReadPixels (0, 0, (*truewidth), (*trueheight), GL_RGBA, GL_UNSIGNED_BYTE, ret + prepadbytes); 
+		//gles:
+		//Only two format/type parameter pairs are accepted.
+		//GL_RGBA/GL_UNSIGNED_BYTE is always accepted, and the other acceptable pair can be discovered by querying GL_IMPLEMENTATION_COLOR_READ_FORMAT and GL_IMPLEMENTATION_COLOR_READ_TYPE.
+		//thus its simpler to only use GL_RGBA/GL_UNSIGNED_BYTE
+		//desktopgl:
+		//total line byte length must be aligned to GL_PACK_ALIGNMENT. by reading rgba instead of rgb, we can ensure the line is a multiple of 4 bytes.
 
+		ret = BZ_Malloc((*truewidth)*(*trueheight)*4);
+		qglReadPixels (0, 0, (*truewidth), (*trueheight), GL_RGBA, GL_UNSIGNED_BYTE, ret); 
+
+		*fmt = TF_RGB24;
 		c = (*truewidth)*(*trueheight);
-		p = ret + prepadbytes;
+		p = ret;
 		for (i = 1; i < c; i++)
 		{
 			p[i*3+0]=p[i*4+0];
 			p[i*3+1]=p[i*4+1];
 			p[i*3+2]=p[i*4+2];
 		}
-		ret = BZ_Realloc(ret, prepadbytes + (*truewidth)*(*trueheight)*3);
+		ret = BZ_Realloc(ret, (*truewidth)*(*trueheight)*3);
 	}
+#ifdef _DEBUG
+	else if (!gl_config.gles && gl_config.glversion >= 1.2)
+	{
+		*fmt = TF_BGRA32;
+		ret = BZ_Malloc((*truewidth)*(*trueheight)*4);
+		qglReadPixels (0, 0, (*truewidth), (*trueheight), GL_BGRA_EXT, GL_UNSIGNED_INT_8_8_8_8_REV, ret); 
+	}
+#endif
 	else
 	{
-		ret = BZ_Malloc(prepadbytes + (*truewidth)*(*trueheight)*3);
-		qglReadPixels (0, 0, (*truewidth), (*trueheight), GL_RGB, GL_UNSIGNED_BYTE, ret + prepadbytes); 
+		*fmt = TF_RGB24;
+		ret = BZ_Malloc((*truewidth)*(*trueheight)*3);
+		qglReadPixels (0, 0, (*truewidth), (*trueheight), GL_RGB, GL_UNSIGNED_BYTE, ret); 
 	}
 
 	if (gammaworks)
 	{
-		c = prepadbytes+(*truewidth)*(*trueheight)*3;
-		for (i=prepadbytes ; i<c ; i+=3)
+		c = (*truewidth)*(*trueheight)*3;
+		for (i=0 ; i<c ; i+=3)
 		{
 			extern qbyte		gammatable[256];
 			ret[i+0] = gammatable[ret[i+0]];
diff --git a/engine/gl/glsupp.h b/engine/gl/glsupp.h
index e98f54bf6..fce1010ff 100644
--- a/engine/gl/glsupp.h
+++ b/engine/gl/glsupp.h
@@ -612,6 +612,10 @@ typedef void		(APIENTRYP PFNGLGETSHADERSOURCEARBPROC)		(GLhandleARB obj, GLsizei
 #define GL_FRAMEBUFFER_SRGB							0x8DB9
 #endif
 
+#ifndef GL_ARB_pixel_buffer_object
+#define GL_PIXEL_PACK_BUFFER_ARB                        0x88EB
+#define GL_PIXEL_UNPACK_BUFFER_ARB                      0x88EC
+#endif
 
 #ifndef GL_ARB_vertex_buffer_object
 #define GL_BUFFER_SIZE_ARB                0x8764
diff --git a/engine/server/savegame.c b/engine/server/savegame.c
index 842ccb5ff..173a20c21 100644
--- a/engine/server/savegame.c
+++ b/engine/server/savegame.c
@@ -1221,10 +1221,11 @@ void SV_Savegame (char *savename)
 		//okay, we drew something, we're good to save a screeny.
 		if (okay)
 		{
-			rgbbuffer = VID_GetRGBInfo(0, &width, &height);
+			enum uploadfmt fmt;
+			rgbbuffer = VID_GetRGBInfo(&width, &height, &fmt);
 			if (rgbbuffer)
 			{
-				SCR_ScreenShot(savefilename, FS_GAMEONLY, rgbbuffer, width, height);
+				SCR_ScreenShot(savefilename, FS_GAMEONLY, rgbbuffer, width, height, fmt);
 				BZ_Free(rgbbuffer);
 	
 
diff --git a/plugins/avplug/avencode.c b/plugins/avplug/avencode.c
index 435beeb0d..ab5d28a79 100644
--- a/plugins/avplug/avencode.c
+++ b/plugins/avplug/avencode.c
@@ -115,20 +115,38 @@ void close_video(struct encctx *ctx)
 	}
 	av_free(ctx->video_outbuf);
 }
-static void AVEnc_Video (void *vctx, void *data, int frame, int width, int height)
+static void AVEnc_Video (void *vctx, void *data, int frame, int width, int height, enum uploadfmt qpfmt)
 {
 	struct encctx *ctx = vctx;
-	//weird maths to flip it.
-	const uint8_t *srcslices[2] = {(uint8_t*)data + (height-1)*width*3, NULL};
-	int srcstride[2] = {-width*3, 0};
+	const uint8_t *srcslices[2];
+	int srcstride[2];
 	int success;
 	AVPacket pkt;
+	int avpfmt;
+	int inbpp;
 
 	if (!ctx->video_st)
 		return;
 
+	switch(qpfmt)
+	{
+	case TF_BGRA32: avpfmt = AV_PIX_FMT_BGRA; inbpp = 4; break;
+	case TF_RGBA32: avpfmt = AV_PIX_FMT_RGBA; inbpp = 4; break;
+	case TF_BGR24: avpfmt = AV_PIX_FMT_BGR24; inbpp = 3; break;
+	case TF_RGB24: avpfmt = AV_PIX_FMT_RGB24; inbpp = 3; break;
+	default:
+		return;
+	}
+
+	//weird maths to flip it.
+	srcslices[0] = (uint8_t*)data + (height-1)*width*inbpp;
+	srcstride[0] = -width*inbpp;
+	srcslices[1] = NULL;
+	srcstride[1] = 0;
+
 	//convert RGB to whatever the codec needs (ie: yuv...).
-	ctx->scale_ctx = sws_getCachedContext(ctx->scale_ctx, width, height, AV_PIX_FMT_RGB24, ctx->picture->width, ctx->picture->height, ctx->video_st->codec->pix_fmt, SWS_POINT, 0, 0, 0);
+	//also rescales, but only if the user resizes the video while recording. which is a stupid thing to do.
+	ctx->scale_ctx = sws_getCachedContext(ctx->scale_ctx, width, height, qpfmt, ctx->picture->width, ctx->picture->height, ctx->video_st->codec->pix_fmt, SWS_POINT, 0, 0, 0);
 	sws_scale(ctx->scale_ctx, srcslices, srcstride, 0, height, ctx->picture->data, ctx->picture->linesize);
 
 	av_init_packet(&pkt);