diff --git a/src/radeon_exa_funcs.c b/src/radeon_exa_funcs.c
index f937c4e..25fbf5c 100644
--- a/src/radeon_exa_funcs.c
+++ b/src/radeon_exa_funcs.c
@@ -631,13 +631,21 @@ RADEONDownloadFromScreenCP(PixmapPtr pSrc, int x, int y, int w, int h,
 				    | (info->gartLocation + info->dri->bufStart
 				       + scratch->idx * scratch->total) >> 10;
 	drm_radeon_indirect_t indirect;
+	/* if the xfer is small, do the blit twice to make sure data hits
+	 * gart
+	 */
+	int small_image = (w < 32 || h < 32);
 	ACCEL_PREAMBLE();
 
 	RADEON_SWITCH_TO_2D();
 
 	/* Kick the first blit as early as possible */
 	RADEONBlitChunk(pScrn, NULL, NULL, datatype, src_pitch_offset,
-                        scratch_pitch_offset, x, y, 0, 0, w, hpass, 0, 0);
+			scratch_pitch_offset, x, y, 0, 0, w, hpass, 0, 0);
+	if (small_image) {
+	    RADEONBlitChunk(pScrn, NULL, NULL, datatype, src_pitch_offset,
+			    scratch_pitch_offset, x, y, 0, 0, w, hpass, 0, 0);
+	}
 	FLUSH_RING();
 
 #if X_BYTE_ORDER == X_BIG_ENDIAN
@@ -664,8 +672,13 @@ RADEONDownloadFromScreenCP(PixmapPtr pSrc, int x, int y, int w, int h,
 	    if (hpass) {
 		scratch_off = scratch->total/2 - scratch_off;
 		RADEONBlitChunk(pScrn, NULL, NULL, datatype, src_pitch_offset,
-                                scratch_pitch_offset + (scratch_off >> 10),
+				scratch_pitch_offset + (scratch_off >> 10),
 				x, y, 0, 0, w, hpass, 0, 0);
+		if (small_image) {
+		    RADEONBlitChunk(pScrn, NULL, NULL, datatype, src_pitch_offset,
+				    scratch_pitch_offset + (scratch_off >> 10),
+				    x, y, 0, 0, w, hpass, 0, 0);
+		}
 	    }
 
 	    /*