diff --git a/src/d3d/d3d.cpp b/src/d3d/d3d.cpp
index 0344585..a5e65a6 100644
--- a/src/d3d/d3d.cpp
+++ b/src/d3d/d3d.cpp
@@ -527,7 +527,9 @@ rasterSetFormat(Raster *raster)
 		natras->format = formatInfoRW[(raster->format >> 8) & 0xF].d3dformat;
 		raster->depth = formatInfoRW[(raster->format >> 8) & 0xF].depth;
 	}
+	natras->bpp = raster->depth/8;
 	natras->hasAlpha = formatInfoRW[(raster->format >> 8) & 0xF].hasAlpha;
+	raster->stride = raster->width*natras->bpp;
 }
 
 static Raster*
@@ -684,8 +686,11 @@ imageFindRasterFormat(Image *img, int32 type,
 
 	assert((type&0xF) == Raster::TEXTURE);
 
-	for(width = 1; width < img->width; width <<= 1);
-	for(height = 1; height < img->height; height <<= 1);
+//	for(width = 1; width < img->width; width <<= 1);
+//	for(height = 1; height < img->height; height <<= 1);
+	// Perhaps non-power-of-2 textures are acceptable?
+	width = img->width;
+	height = img->height;
 
 	depth = img->depth;
 
@@ -734,6 +739,8 @@ rasterFromImage(Raster *raster, Image *image)
 	if((raster->type&0xF) != Raster::TEXTURE)
 		return 0;
 
+	void (*conv)(uint8 *out, uint8 *in) = nil;
+
 	// Unpalettize image if necessary but don't change original
 	Image *truecolimg = nil;
 	if(image->depth <= 8 && !isP8supported){
@@ -748,22 +755,39 @@ rasterFromImage(Raster *raster, Image *image)
 	D3dRaster *natras = PLUGINOFFSET(D3dRaster, raster, nativeRasterOffset);
 	switch(image->depth){
 	case 32:
-		if(raster->format != Raster::C8888 &&
-		   raster->format != Raster::C888)
+		if(raster->format == Raster::C8888)
+			conv = conv_BGRA8888_from_RGBA8888;
+		else if(raster->format == Raster::C888)
+			conv = conv_BGR888_from_RGB888;
+		else
 			goto err;
 		break;
 	case 24:
-		if(raster->format != Raster::C888) goto err;
+		if(raster->format == Raster::C8888)
+			conv = conv_BGRA8888_from_RGB888;
+		else if(raster->format == Raster::C888)
+			conv = conv_BGR888_from_RGB888;
+		else
+			goto err;
 		break;
 	case 16:
-		if(raster->format != Raster::C1555) goto err;
+		if(raster->format == Raster::C1555)
+			conv = conv_ARGB1555_from_ARGB1555;
+		else
+			goto err;
 		break;
 	case 8:
-		if(raster->format != (Raster::PAL8 | Raster::C8888)) goto err;
+		if(raster->format == (Raster::PAL8 | Raster::C8888))
+			conv = conv_8_from_8;
+		else
+			goto err;
 		break;
 	case 4:
-		if(raster->format != (Raster::PAL4 | Raster::C8888)) goto err;
-		break;
+		if(raster->format == (Raster::PAL4 | Raster::C8888) ||
+		   raster->format == (Raster::PAL8 | Raster::C8888))
+			conv = conv_8_from_8;
+		else
+			goto err;
 	default:
 	err:
 		RWERROR((ERR_INVRASTER));
@@ -780,48 +804,30 @@ rasterFromImage(Raster *raster, Image *image)
 		in = image->palette;
 		out = (uint8*)natras->palette;
 		for(int32 i = 0; i < pallength; i++){
-			out[0] = in[0];
-			out[1] = in[1];
-			out[2] = in[2];
-			out[3] = in[3];
+			conv_RGBA8888_from_RGBA8888(out, in);
 			in += 4;
 			out += 4;
 		}
 	}
 
-	int32 inc = image->bpp;
-	in = image->pixels;
-	out = raster->lock(0, Raster::LOCKWRITE|Raster::LOCKNOFETCH);
-	if(pallength)
-		memcpy(out, in, raster->width*raster->height);
-	else
-		// TODO: stride
-		for(int32 y = 0; y < image->height; y++)
-			for(int32 x = 0; x < image->width; x++)
-				switch(raster->format & 0xF00){
-				case Raster::C8888:
-					out[0] = in[2];
-					out[1] = in[1];
-					out[2] = in[0];
-					out[3] = in[3];
-					in += inc;
-					out += 4;
-					break;
-				case Raster::C888:
-					out[0] = in[2];
-					out[1] = in[1];
-					out[2] = in[0];
-					out[3] = 0xFF;
-					in += inc;
-					out += 4;
-					break;
-				case Raster::C1555:
-					out[0] = in[0];
-					out[1] = in[1];
-					in += inc;
-					out += 2;
-					break;
-				}
+	uint8 *pixels = raster->lock(0, Raster::LOCKWRITE|Raster::LOCKNOFETCH);
+	assert(pixels);
+	uint8 *imgpixels = image->pixels;
+
+	int x, y;
+	assert(image->width == raster->width);
+	assert(image->height == raster->height);
+	for(y = 0; y < image->height; y++){
+		uint8 *imgrow = imgpixels;
+		uint8 *rasrow = pixels;
+		for(x = 0; x < image->width; x++){
+			conv(rasrow, imgrow);
+			imgrow += image->bpp;
+			rasrow += natras->bpp;
+		}
+		imgpixels += image->stride;
+		pixels += raster->stride;
+	}
 	raster->unlock(0);
 
 	if(truecolimg)
@@ -860,18 +866,24 @@ rasterToImage(Raster *raster)
 		raster->unlock(0);
 		return image;
 	}
+
+	void (*conv)(uint8 *out, uint8 *in) = nil;
 	switch(raster->format & 0xF00){
 	case Raster::C1555:
 		depth = 16;
+		conv = conv_ARGB1555_from_ARGB1555;
 		break;
 	case Raster::C8888:
 		depth = 32;
+		conv = conv_RGBA8888_from_BGRA8888;
 		break;
 	case Raster::C888:
 		depth = 24;
+		conv = conv_RGB888_from_BGR888;
 		break;
 	case Raster::C555:
 		depth = 16;
+		conv = conv_ARGB1555_from_RGB555;
 		break;
 
 	default:
@@ -898,52 +910,28 @@ rasterToImage(Raster *raster)
 		out = image->palette;
 		in = (uint8*)natras->palette;
 		for(int32 i = 0; i < pallength; i++){
-			out[0] = in[0];
-			out[1] = in[1];
-			out[2] = in[2];
-			out[3] = in[3];
+			conv_RGBA8888_from_RGBA8888(out, in);
 			in += 4;
 			out += 4;
 		}
 	}
 
-	uint8 *dst = image->pixels;
-	in = raster->lock(0, Raster::LOCKREAD);
-	if(pallength)
-		memcpy(dst, in, raster->width*raster->height);
-	else{
-		for(int32 y = 0; y < image->height; y++){
-			out = dst;
-			for(int32 x = 0; x < image->width; x++){
-				switch(raster->format & 0xF00){
-				case Raster::C8888:
-					out[0] = in[2];
-					out[1] = in[1];
-					out[2] = in[0];
-					out[3] = in[3];
-					in += 4;
-					break;
-				case Raster::C888:
-					out[0] = in[2];
-					out[1] = in[1];
-					out[2] = in[0];
-					in += 4;
-					break;
-				case Raster::C1555:
-					out[0] = in[0];
-					out[1] = in[1];
-					in += 2;
-					break;
-				case Raster::C555:
-					out[0] = in[0];
-					out[1] = in[1] | 0x80;
-					in += 2;
-					break;
-				}
-				out += image->bpp;
-			}
-			dst += image->stride;
+	uint8 *imgpixels = image->pixels;
+	uint8 *pixels = raster->lock(0, Raster::LOCKREAD);
+
+	int x, y;
+	assert(image->width == raster->width);
+	assert(image->height == raster->height);
+	for(y = 0; y < image->height; y++){
+		uint8 *imgrow = imgpixels;
+		uint8 *rasrow = pixels;
+		for(x = 0; x < image->width; x++){
+			conv(imgrow, rasrow);
+			imgrow += image->bpp;
+			rasrow += natras->bpp;
 		}
+		imgpixels += image->stride;
+		pixels += raster->stride;
 	}
 	raster->unlock(0);
 
diff --git a/src/d3d/rwd3d.h b/src/d3d/rwd3d.h
index 178a644..00cab50 100644
--- a/src/d3d/rwd3d.h
+++ b/src/d3d/rwd3d.h
@@ -162,6 +162,7 @@ struct D3dRaster
 	void *texture;
 	void *palette;
 	uint32 format;
+	uint32 bpp;	// bytes per pixel
 	bool32 hasAlpha;
 	bool32 customFormat;
 };
diff --git a/src/gl/gl3raster.cpp b/src/gl/gl3raster.cpp
index 85d12e5..f5800bf 100644
--- a/src/gl/gl3raster.cpp
+++ b/src/gl/gl3raster.cpp
@@ -315,8 +315,11 @@ imageFindRasterFormat(Image *img, int32 type,
 
 	assert((type&0xF) == Raster::TEXTURE);
 
-	for(width = 1; width < img->width; width <<= 1);
-	for(height = 1; height < img->height; height <<= 1);
+//	for(width = 1; width < img->width; width <<= 1);
+//	for(height = 1; height < img->height; height <<= 1);
+	// Perhaps non-power-of-2 textures are acceptable?
+	width = img->width;
+	height = img->height;
 
 	depth = img->depth;
 
@@ -440,8 +443,11 @@ rasterFromImage(Raster *raster, Image *image)
 		imgpixels -= image->stride;
 		pixels += raster->stride;
 	}
-
 	raster->unlock(0);
+
+	if(truecolimg)
+		truecolimg->destroy();
+
 	return 1;
 }
 
diff --git a/src/raster.cpp b/src/raster.cpp
index eb3ac19..183ec6b 100644
--- a/src/raster.cpp
+++ b/src/raster.cpp
@@ -230,6 +230,15 @@ conv_RGBA8888_from_RGBA8888(uint8 *out, uint8 *in)
 	out[3] = in[3];
 }
 
+void
+conv_BGRA8888_from_RGBA8888(uint8 *out, uint8 *in)
+{
+	out[2] = in[0];
+	out[1] = in[1];
+	out[0] = in[2];
+	out[3] = in[3];
+}
+
 void
 conv_RGBA8888_from_RGB888(uint8 *out, uint8 *in)
 {
@@ -239,6 +248,15 @@ conv_RGBA8888_from_RGB888(uint8 *out, uint8 *in)
 	out[3] = 0xFF;
 }
 
+void
+conv_BGRA8888_from_RGB888(uint8 *out, uint8 *in)
+{
+	out[2] = in[0];
+	out[1] = in[1];
+	out[0] = in[2];
+	out[3] = 0xFF;
+}
+
 void
 conv_RGB888_from_RGB888(uint8 *out, uint8 *in)
 {
@@ -247,6 +265,28 @@ conv_RGB888_from_RGB888(uint8 *out, uint8 *in)
 	out[2] = in[2];
 }
 
+void
+conv_BGR888_from_RGB888(uint8 *out, uint8 *in)
+{
+	out[2] = in[0];
+	out[1] = in[1];
+	out[0] = in[2];
+}
+
+void
+conv_ARGB1555_from_ARGB1555(uint8 *out, uint8 *in)
+{
+	out[0] = in[0];
+	out[1] = in[1];
+}
+
+void
+conv_ARGB1555_from_RGB555(uint8 *out, uint8 *in)
+{
+	out[0] = in[0];
+	out[1] = in[1] | 0x80;
+}
+
 void
 conv_RGBA5551_from_ARGB1555(uint8 *out, uint8 *in)
 {
diff --git a/src/rwbase.h b/src/rwbase.h
index d9cec4b..2c67ab8 100644
--- a/src/rwbase.h
+++ b/src/rwbase.h
@@ -74,6 +74,16 @@ typedef uint32 uint;
 
 #define nelem(A) (sizeof(A) / sizeof A[0])
 
+#ifdef __GNUC__
+#define RWALIGN(n) __attribute__ ((aligned (n)))
+#else
+#ifdef _MSC_VER
+#define RWALIGN(n) __declspec(align(n))
+#else
+#define RWALIGN(n)	// unknown compiler...ignore
+#endif
+#endif
+
 // Lists
 
 struct LLLink
@@ -286,7 +296,7 @@ inline V3d rotate(const V3d &v, const Quat &q) { return mult(mult(q, makeQuat(0.
 Quat lerp(const Quat &q, const Quat &p, float32 r);
 Quat slerp(const Quat &q, const Quat &p, float32 a);
 
-struct RawMatrix
+struct RWALIGN(16) RawMatrix
 {
 	V3d right;
 	float32 rightw;
@@ -302,7 +312,7 @@ struct RawMatrix
 	static void setIdentity(RawMatrix *dst);
 };
 
-struct Matrix
+struct RWALIGN(16) Matrix
 {
 	enum Type {
 		TYPENORMAL	= 1,
diff --git a/src/rwobjects.h b/src/rwobjects.h
index 7476054..fd0a09f 100644
--- a/src/rwobjects.h
+++ b/src/rwobjects.h
@@ -306,11 +306,20 @@ struct Raster
 };
 
 void conv_RGBA8888_from_RGBA8888(uint8 *out, uint8 *in);
+void conv_BGRA8888_from_RGBA8888(uint8 *out, uint8 *in);
 void conv_RGBA8888_from_RGB888(uint8 *out, uint8 *in);
+void conv_BGRA8888_from_RGB888(uint8 *out, uint8 *in);
 void conv_RGB888_from_RGB888(uint8 *out, uint8 *in);
+void conv_BGR888_from_RGB888(uint8 *out, uint8 *in);
+void conv_ARGB1555_from_ARGB1555(uint8 *out, uint8 *in);
+void conv_ARGB1555_from_RGB555(uint8 *out, uint8 *in);
 void conv_RGBA5551_from_ARGB1555(uint8 *out, uint8 *in);
 void conv_RGBA8888_from_ARGB1555(uint8 *out, uint8 *in);
 void conv_ABGR1555_from_ARGB1555(uint8 *out, uint8 *in);
+inline void conv_8_from_8(uint8 *out, uint8 *in) { *out = *in; }
+// some swaps are the same, so these are just more descriptive names
+inline void conv_RGBA8888_from_BGRA8888(uint8 *out, uint8 *in) { conv_BGRA8888_from_RGBA8888(out, in); }
+inline void conv_RGB888_from_BGR888(uint8 *out, uint8 *in) { conv_BGR888_from_RGB888(out, in); }
 inline void conv_ARGB1555_from_ABGR1555(uint8 *out, uint8 *in) { conv_ABGR1555_from_ARGB1555(out, in); }
 
 void expandPal4(uint8 *dst, uint32 dststride, uint8 *src, uint32 srcstride, int32 w, int32 h);