diff --git a/src/gl/compatibility/gl_20.cpp b/src/gl/compatibility/gl_20.cpp
index b06ce762bc..d0de0589c4 100644
--- a/src/gl/compatibility/gl_20.cpp
+++ b/src/gl/compatibility/gl_20.cpp
@@ -123,88 +123,6 @@ void gl_SetTextureMode(int type)
 	}
 }
 
-//===========================================================================
-//
-// FGLTex::WarpBuffer
-//
-//===========================================================================
-
-BYTE *gl_WarpBuffer(BYTE *buffer, int Width, int Height, int warp, float Speed)
-{
-	if (Width > 256 || Height > 256) return buffer;
-
-	DWORD *in = (DWORD*)buffer;
-	DWORD *out = (DWORD*)new BYTE[4 * Width*Height];
-
-	static DWORD linebuffer[256];	// anything larger will bring down performance so it is excluded above.
-	DWORD timebase = DWORD(r_FrameTime*Speed * 23 / 28);
-	int xsize = Width;
-	int ysize = Height;
-	int xmask = xsize - 1;
-	int ymask = ysize - 1;
-	int ds_xbits;
-	int i;
-
-	if (warp == 1)
-	{
-		for (ds_xbits = -1, i = Width; i; i >>= 1, ds_xbits++);
-
-		// pending consolidation with the software renderer's code.
-
-		/*
-		for (x = xsize - 1; x >= 0; x--)
-		{
-			int yt, yf = (finesine[(timebase + (x + 17) * 128)&FINEMASK] >> 13) & ymask;
-			const DWORD *source = in + x;
-			DWORD *dest = out + x;
-			for (yt = ysize; yt; yt--, yf = (yf + 1)&ymask, dest += xsize)
-			{
-				*dest = *(source + (yf << ds_xbits));
-			}
-		}
-		timebase = DWORD(r_FrameTime*Speed * 32 / 28);
-		int y;
-		for (y = ysize - 1; y >= 0; y--)
-		{
-			int xt, xf = (finesine[(timebase + y * 128)&FINEMASK] >> 13) & xmask;
-			DWORD *source = out + (y << ds_xbits);
-			DWORD *dest = linebuffer;
-			for (xt = xsize; xt; xt--, xf = (xf + 1)&xmask)
-			{
-				*dest++ = *(source + xf);
-			}
-			memcpy(out + y*xsize, linebuffer, xsize * sizeof(DWORD));
-		}
-		*/
-	}
-	else
-	{
-		int ybits;
-		for (ybits = -1, i = ysize; i; i >>= 1, ybits++);
-
-		/*
-		DWORD timebase = (r_FrameTime * Speed * 40 / 28);
-		for (x = xsize - 1; x >= 0; x--)
-		{
-			for (int y = ysize - 1; y >= 0; y--)
-			{
-				int xt = (x + 128
-					+ ((finesine[(y * 128 + timebase * 5 + 900) & 8191] * 2) >> FRACBITS)
-					+ ((finesine[(x * 256 + timebase * 4 + 300) & 8191] * 2) >> FRACBITS)) & xmask;
-				int yt = (y + 128
-					+ ((finesine[(y * 128 + timebase * 3 + 700) & 8191] * 2) >> FRACBITS)
-					+ ((finesine[(x * 256 + timebase * 4 + 1200) & 8191] * 2) >> FRACBITS)) & ymask;
-				const DWORD *source = in + (xt << ybits) + yt;
-				DWORD *dest = out + (x << ybits) + y;
-				*dest = *source;
-			}
-		}
-		*/
-	}
-	delete[] buffer;
-	return (BYTE*)out;
-}
-
 //==========================================================================
 //
 //
diff --git a/src/gl/textures/gl_material.cpp b/src/gl/textures/gl_material.cpp
index f3743bd3a2..a900a1e457 100644
--- a/src/gl/textures/gl_material.cpp
+++ b/src/gl/textures/gl_material.cpp
@@ -47,6 +47,7 @@
 #include "templates.h"
 #include "sc_man.h"
 #include "colormatcher.h"
+#include "textures/warpbuffer.h"
 
 //#include "gl/gl_intern.h"
 
@@ -72,7 +73,6 @@ EXTERN_CVAR(Bool, gl_texture_usehires)
 // The GL texture maintenance class
 //
 //===========================================================================
-BYTE *gl_WarpBuffer(BYTE *buffer, int Width, int Height, int warp, float Speed);
 
 //===========================================================================
 //
@@ -310,11 +310,18 @@ const FHardwareTexture *FGLTexture::Bind(int texunit, int clampmode, int transla
 			if (!tex->bHasCanvas)
 			{
 				buffer = CreateTexBuffer(translation, w, h, hirescheck, true, alphatrans);
-				if (tex->bWarped && gl.glslversion == 0)
+				if (tex->bWarped && gl.glslversion == 0 && w*h <= 256*256)	// do not software-warp larger textures, especially on the old systems that still need this fallback.
 				{
-					// need to warp
-					buffer = gl_WarpBuffer(buffer, w, h, tex->bWarped, static_cast<FWarpTexture*>(tex)->GetSpeed());
-					static_cast<FWarpTexture*>(tex)->GenTime = r_FrameTime;
+					// need to do software warping
+					FWarpTexture *wt = static_cast<FWarpTexture*>(tex);
+					unsigned char *warpbuffer = new unsigned char[w*h*4];
+					if (tex->bWarped != 2)
+						WarpBufferType1((DWORD*)warpbuffer, (const DWORD*)buffer, w, h, wt->WidthOffsetMultiplier, wt->HeightOffsetMultiplier, r_FrameTime, wt->Speed);
+					else
+						WarpBufferType2((DWORD*)warpbuffer, (const DWORD*)buffer, w, h, wt->WidthOffsetMultiplier, wt->HeightOffsetMultiplier, r_FrameTime, wt->Speed);
+					delete[] buffer;
+					buffer = warpbuffer;
+					wt->GenTime = r_FrameTime;
 				}
 				tex->ProcessData(buffer, w, h, false);
 			}
diff --git a/src/textures/textures.h b/src/textures/textures.h
index 778b955d46..4aa30605fa 100644
--- a/src/textures/textures.h
+++ b/src/textures/textures.h
@@ -555,12 +555,12 @@ public:
 	FTexture *GetRedirect(bool wantwarped);
 
 	DWORD GenTime;
+	float Speed;
+	int WidthOffsetMultiplier, HeightOffsetMultiplier;  // [mxd]
 protected:
 	FTexture *SourcePic;
 	BYTE *Pixels;
 	Span **Spans;
-	float Speed;
-	int WidthOffsetMultiplier, HeightOffsetMultiplier;  // [mxd]
 
 	virtual void MakeTexture (DWORD time);
 	int NextPo2 (int v); // [mxd]