diff --git a/lib/intel_batchbuffer.c b/lib/intel_batchbuffer.c
index 449d46de..14446df6 100644
--- a/lib/intel_batchbuffer.c
+++ b/lib/intel_batchbuffer.c
@@ -72,6 +72,8 @@ intel_batchbuffer_free(struct intel_batchbuffer *batch)
 	free(batch);
 }
 
+#define CMD_POLY_STIPPLE_OFFSET       0x7906
+
 void
 intel_batchbuffer_flush(struct intel_batchbuffer *batch)
 {
@@ -82,6 +84,13 @@ intel_batchbuffer_flush(struct intel_batchbuffer *batch)
 	if (used == 0)
 		return;
 
+	if (IS_GEN5(batch->devid)) {
+		BEGIN_BATCH(2);
+		OUT_BATCH(CMD_POLY_STIPPLE_OFFSET << 16);
+		OUT_BATCH(0);
+		ADVANCE_BATCH();
+	}
+
 	/* Round batchbuffer usage to 2 DWORDs. */
 	if ((used & 4) == 0) {
 		*(uint32_t *) (batch->ptr) = 0; /* noop */
diff --git a/tests/gem_stress.c b/tests/gem_stress.c
index 5aa6def8..e8d91880 100644
--- a/tests/gem_stress.c
+++ b/tests/gem_stress.c
@@ -166,13 +166,6 @@ static void keep_gpu_busy(void)
 	OUT_BATCH(src_pitch);
 	OUT_RELOC(busy_bo, I915_GEM_DOMAIN_RENDER, 0, 0);
 	ADVANCE_BATCH();
-
-	if (IS_GEN5(devid)) {
-		BEGIN_BATCH(2);
-		OUT_BATCH(CMD_POLY_STIPPLE_OFFSET << 16);
-		OUT_BATCH(0);
-		ADVANCE_BATCH();
-	}
 }
 
 static unsigned int copyfunc_seq = 0;
@@ -288,13 +281,6 @@ static void blitter_copyfunc(struct scratch_buf *src, unsigned src_x, unsigned s
 	OUT_RELOC(src->bo, I915_GEM_DOMAIN_RENDER, 0, 0);
 	ADVANCE_BATCH();
 
-	if (IS_GEN5(devid)) {
-		BEGIN_BATCH(2);
-		OUT_BATCH(CMD_POLY_STIPPLE_OFFSET << 16);
-		OUT_BATCH(0);
-		ADVANCE_BATCH();
-	}
-
 	if (!(keep_gpu_busy_counter & 1) && !fence_storm)
 		keep_gpu_busy();