freedreno: dummy-draw workaround for a320

Fixes gpu lockups in supertuxkart.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

src/gallium/drivers/freedreno/a3xx/fd3_emit.c

@@ -231,6 +231,7 @@ emit_cache_flush(struct fd_ringbuffer *ring)
 	OUT_PKT3(ring, CP_EVENT_WRITE, 1);
 	OUT_RING(ring, CACHE_FLUSH);
 
+	/* probably only really needed on a320: */
 	OUT_PKT3(ring, CP_DRAW_INDX, 3);
 	OUT_RING(ring, 0x00000000);
 	OUT_RING(ring, DRAW(1, DI_SRC_SEL_AUTO_INDEX,
@@ -523,7 +524,7 @@ fd3_emit_restore(struct fd_context *ctx)
 	OUT_RING(ring, A3XX_RB_MSAA_CONTROL_DISABLE |
 			A3XX_RB_MSAA_CONTROL_SAMPLES(MSAA_ONE) |
 			A3XX_RB_MSAA_CONTROL_SAMPLE_MASK(0xffff));
-	OUT_RING(ring, 0x00000000);        /* UNKNOWN_20C3 */
+	OUT_RING(ring, 0x00000000);        /* RB_ALPHA_REF */
 
 	OUT_PKT0(ring, REG_A3XX_GRAS_CL_GB_CLIP_ADJ, 1);
 	OUT_RING(ring, A3XX_GRAS_CL_GB_CLIP_ADJ_HORZ(0) |

src/gallium/drivers/freedreno/freedreno_draw.h

@@ -59,6 +59,21 @@ fd_draw(struct fd_context *ctx, enum pc_di_primtype primtype,
 	 */
 	emit_marker(ring, 7);
 
+	if (ctx->screen->gpu_id == 320) {
+		/* dummy-draw workaround: */
+		OUT_PKT3(ring, CP_DRAW_INDX, 3);
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, DRAW(1, DI_SRC_SEL_AUTO_INDEX,
+				INDEX_SIZE_IGN, IGNORE_VISIBILITY));
+		OUT_RING(ring, 0);             /* NumIndices */
+
+		/* ugg, hard-code register offset to avoid pulling in the
+		 * a3xx register headers into something #included from a2xx
+		 */
+		OUT_PKT0(ring, 0x2206, 1);     /* A3XX_HLSQ_CONST_VSPRESV_RANGE_REG */
+		OUT_RING(ring, 0);
+	}
+
 	OUT_PKT3(ring, CP_DRAW_INDX, idx_bo ? 5 : 3);
 	OUT_RING(ring, 0x00000000);        /* viz query info. */
 	OUT_RING(ring, DRAW(primtype, src_sel,