diff --git a/src/glamor_copyarea.c b/src/glamor_copyarea.c
index 4e6f953..257fb5c 100644
--- a/src/glamor_copyarea.c
+++ b/src/glamor_copyarea.c
@@ -163,16 +163,37 @@ glamor_copy_n_to_n_textured(DrawablePtr src,
 	src_pixmap_priv = glamor_get_pixmap_private(src_pixmap);
 	dst_pixmap_priv = glamor_get_pixmap_private(dst_pixmap);
 
+	glamor_get_drawable_deltas(src, src_pixmap, &src_x_off,
+					   &src_y_off);
+	glamor_get_drawable_deltas(dst, dst_pixmap, &dst_x_off,
+				   &dst_y_off);
+
 	if (!src_pixmap_priv->base.gl_fbo) {
 #ifndef GLAMOR_PIXMAP_DYNAMIC_UPLOAD
 		glamor_delayed_fallback(dst->pScreen, "src has no fbo.\n");
 		return FALSE;
 #else
-		src_status = glamor_upload_pixmap_to_texture(src_pixmap);
-		if (src_status != GLAMOR_UPLOAD_DONE)
-			return FALSE;
-
-		src_pixmap_priv = glamor_get_pixmap_private(src_pixmap);
+		/* Optimize when the source doesn't have an FBO, just upload
+		   the data directly to the dest FBO */
+		int src_stride = src_pixmap->devKind;
+		int bpp = src_pixmap->drawable.bitsPerPixel;
+		CARD8 *src_data = NULL;
+		for (i = 0; i < nbox; i++) {
+			int x = box[i].x1 + dst_x_off;
+			int y = box[i].y1 + dst_y_off;
+			int w = box[i].x2 - box[i].x1;
+			int h = box[i].y2 - box[i].y1;
+			src_data = src_pixmap->devPrivate.ptr + (box[i].y1 + dy +
+								src_y_off) * src_stride +
+								(box[i].x1 + dx + src_x_off) * (bpp / 8);
+			if (!glamor_upload_sub_pixmap_to_texture(dst_pixmap,
+					x, y, w, h,
+					src_stride, src_data, 0)) {
+				ErrorF("Failed to upload the sub pixmap to dst\n");
+				return FALSE;
+			}
+		}
+		return TRUE;
 #endif
 	}
 
@@ -180,8 +201,6 @@ glamor_copy_n_to_n_textured(DrawablePtr src,
 	pixmap_priv_get_dest_scale(dst_pixmap_priv, &dst_xscale, &dst_yscale);
 	pixmap_priv_get_scale(src_pixmap_priv, &src_xscale, &src_yscale);
 
-	glamor_get_drawable_deltas(dst, dst_pixmap, &dst_x_off,
-				   &dst_y_off);
 
 	dispatch = glamor_get_dispatch(glamor_priv);
 
@@ -192,11 +211,11 @@ glamor_copy_n_to_n_textured(DrawablePtr src,
 					vertices);
 	dispatch->glEnableVertexAttribArray(GLAMOR_VERTEX_POS);
 
-	glamor_get_drawable_deltas(src, src_pixmap, &src_x_off,
-				   &src_y_off);
+
 	dx += src_x_off;
 	dy += src_y_off;
 
+
 	dispatch->glActiveTexture(GL_TEXTURE0);
 	dispatch->glBindTexture(GL_TEXTURE_2D,
 				src_pixmap_priv->base.fbo->tex);
@@ -335,7 +354,7 @@ __glamor_copy_n_to_n(DrawablePtr src,
 			      * 4 >
 			      src_pixmap->drawable.width *
 			      src_pixmap->drawable.height)
-		             || !(glamor_check_fbo_size(glamor_priv,
+				|| !(glamor_check_fbo_size(glamor_priv,
 					src_pixmap->drawable.width,
 					src_pixmap->drawable.height))))) {
 
diff --git a/src/glamor_pixmap.c b/src/glamor_pixmap.c
index 84694ec..68a9214 100644
--- a/src/glamor_pixmap.c
+++ b/src/glamor_pixmap.c
@@ -387,7 +387,7 @@ static void
 __glamor_upload_pixmap_to_texture(PixmapPtr pixmap, unsigned int *tex,
 				  GLenum format,
 				  GLenum type,
-				  int x, int y, int w, int h,
+				  int x, int y, int w, int h, int stride,
 				  void *bits, int pbo)
 {
 	glamor_screen_private *glamor_priv =
@@ -395,6 +395,12 @@ __glamor_upload_pixmap_to_texture(PixmapPtr pixmap, unsigned int *tex,
 	glamor_gl_dispatch *dispatch;
 	int non_sub = 0;
 	unsigned int iformat = 0;
+	int bpp = depth_for_type(type);
+	int alignment = bpp / 8;
+	if (stride == 0) {
+		alignment = 4;
+	}
+
 
 	dispatch = glamor_get_dispatch(glamor_priv);
 	if (*tex == 0) {
@@ -412,11 +418,13 @@ __glamor_upload_pixmap_to_texture(PixmapPtr pixmap, unsigned int *tex,
 				  GL_NEAREST);
 	dispatch->glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER,
 				  GL_NEAREST);
-	dispatch->glPixelStorei(GL_UNPACK_ALIGNMENT, 4);
+	dispatch->glPixelStorei(GL_UNPACK_ALIGNMENT, alignment);
 
 	if (bits == NULL)
 		dispatch->glBindBuffer(GL_PIXEL_UNPACK_BUFFER,
 				       pbo);
+
+	dispatch->glPixelStorei(GL_UNPACK_ROW_LENGTH, stride / alignment);
 	if (non_sub)
 		dispatch->glTexImage2D(GL_TEXTURE_2D,
 				       0, iformat, w, h, 0,
@@ -428,6 +436,8 @@ __glamor_upload_pixmap_to_texture(PixmapPtr pixmap, unsigned int *tex,
 					  format, type,
 					  bits);
 
+	dispatch->glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
+
 	if (bits == NULL)
 		dispatch->glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 	glamor_put_dispatch(glamor_priv);
@@ -511,7 +521,7 @@ ready_to_upload:
 		assert(y + fbo_y_off + h <= pixmap_priv->base.fbo->height);
 		__glamor_upload_pixmap_to_texture(pixmap, &pixmap_priv->base.fbo->tex,
 						  format, type,
-						  x + fbo_x_off, y + fbo_y_off, w, h,
+						  x + fbo_x_off, y + fbo_y_off, w, h, stride,
 						  bits, pbo);
 		return TRUE;
 	}
@@ -542,7 +552,7 @@ ready_to_upload:
 	glamor_set_destination_pixmap_priv_nc(pixmap_priv);
 	__glamor_upload_pixmap_to_texture(pixmap, &tex,
 					  format, type,
-					  0, 0, w, h,
+					  0, 0, w, h, stride,
 					  bits, pbo);
 	dispatch->glActiveTexture(GL_TEXTURE0);
 	dispatch->glBindTexture(GL_TEXTURE_2D, tex);
@@ -807,7 +817,7 @@ glamor_upload_pixmap_to_texture(PixmapPtr pixmap)
 	if (glamor_upload_sub_pixmap_to_texture(pixmap, 0, 0,
 						pixmap->drawable.width,
 						pixmap->drawable.height,
-						pixmap->devKind,
+						0,
 						data, pbo))
 		ret = GLAMOR_UPLOAD_DONE;
 	else
diff --git a/src/glamor_putimage.c b/src/glamor_putimage.c
index 99f7ac6..c389685 100644
--- a/src/glamor_putimage.c
+++ b/src/glamor_putimage.c
@@ -300,13 +300,13 @@ _glamor_put_image(DrawablePtr drawable, GCPtr gc, int depth, int x, int y,
 		}
 
 		glamor_upload_sub_pixmap_to_texture(temp_pixmap, 0, 0, w, h,
-						    pixmap->devKind, bits, 0);
-
+						   0, bits, 0);
 		glamor_copy_area(&temp_pixmap->drawable, drawable, gc, 0, 0, w, h, x, y);
 		glamor_destroy_pixmap(temp_pixmap);
-	} else
+	} else {
 		glamor_upload_sub_pixmap_to_texture(pixmap, x + drawable->x + x_off, y + drawable->y + y_off,
-						    w, h, PixmapBytePad(w, depth), bits, 0);
+								    w, h, 0, bits, 0);
+	}
 	ret = TRUE;
 	goto done;
 
diff --git a/src/glamor_setspans.c b/src/glamor_setspans.c
index 3d447b6..32fd8cf 100644
--- a/src/glamor_setspans.c
+++ b/src/glamor_setspans.c
@@ -69,7 +69,7 @@ _glamor_set_spans(DrawablePtr drawable, GCPtr gc, char *src,
 			if (x1 >= x2)
 				continue;
 			glamor_upload_sub_pixmap_to_texture(dest_pixmap, x1 + x_off,  y1 + y_off, x2 - x1, 1,
-							    PixmapBytePad(widths[i], drawable->depth),
+							    0,
 							    drawpixels_src, 0);
 		}
 		drawpixels_src += PixmapBytePad(widths[i], drawable->depth);
diff --git a/src/glamor_utils.h b/src/glamor_utils.h
index d307838..cf0f1eb 100644
--- a/src/glamor_utils.h
+++ b/src/glamor_utils.h
@@ -880,6 +880,26 @@ format_for_depth(int depth)
 	}
 }
 
+static inline int
+depth_for_type(GLenum type)
+{
+	switch (type) {
+	case GL_UNSIGNED_BYTE:
+		return 8;
+	case GL_UNSIGNED_INT_8_8_8_8:
+	case GL_UNSIGNED_INT_8_8_8_8_REV:
+	case GL_UNSIGNED_INT_2_10_10_10_REV:
+		return 32;
+	case GL_UNSIGNED_SHORT_5_6_5:
+	case GL_UNSIGNED_SHORT_5_6_5_REV:
+	case GL_UNSIGNED_SHORT_1_5_5_5_REV:
+	case GL_UNSIGNED_SHORT_4_4_4_4_REV:
+		return 16;
+	default:
+		return 0;
+	}
+}
+
 static inline void
 gl_iformat_for_depth(int depth, GLenum * format)
 {