freedreno/a5xx: cargo-cult end-batch sequence more faithfully

Fixes some issues at least with GMEM bypass mode, where we'd sometimes
end up with some FS quads not hitting memory.

Signed-off-by: Rob Clark <robdclark@gmail.com>

4 files changed, 39 insertions, 4 deletions

diff --git a/src/gallium/drivers/freedreno/a5xx/fd5_context.h b/src/gallium/drivers/freedreno/a5xx/fd5_context.h
index 30a11d0e14..846c4b9224 100644
--- a/src/gallium/drivers/freedreno/a5xx/fd5_context.h
+++ b/src/gallium/drivers/freedreno/a5xx/fd5_context.h
@@ -49,7 +49,11 @@ struct fd5_context {
 	 */
 	struct fd_bo *vsc_size_mem;
 
-	/* TODO not sure what this is for.. */
+	/* TODO not sure what this is for.. probably similar to
+	 * CACHE_FLUSH_TS on kernel side, where value gets written
+	 * to this address synchronized w/ 3d (ie. a way to
+	 * synchronize when the CP is running far ahead)
+	 */
 	struct fd_bo *blit_mem;
 
 	struct u_upload_mgr *border_color_uploader;
diff --git a/src/gallium/drivers/freedreno/a5xx/fd5_gmem.c b/src/gallium/drivers/freedreno/a5xx/fd5_gmem.c
index 775d5b9cb2..21104c2f1d 100644
--- a/src/gallium/drivers/freedreno/a5xx/fd5_gmem.c
+++ b/src/gallium/drivers/freedreno/a5xx/fd5_gmem.c
@@ -96,7 +96,7 @@ emit_mrt(struct fd_ringbuffer *ring, unsigned nr_bufs,
 		OUT_RING(ring, A5XX_RB_MRT_BUF_INFO_COLOR_FORMAT(format) |
 				A5XX_RB_MRT_BUF_INFO_COLOR_TILE_MODE(tile_mode) |
 				A5XX_RB_MRT_BUF_INFO_COLOR_SWAP(swap) |
-				0x800 | /* XXX 0x1000 for RECTLIST clear, 0x0 for BLIT.. */
+				COND(gmem, 0x800) | /* XXX 0x1000 for RECTLIST clear, 0x0 for BLIT.. */
 				COND(srgb, A5XX_RB_MRT_BUF_INFO_COLOR_SRGB));
 		OUT_RING(ring, A5XX_RB_MRT_PITCH(stride));
 		OUT_RING(ring, A5XX_RB_MRT_ARRAY_PITCH(size));
@@ -467,8 +467,16 @@ fd5_emit_tile_gmem2mem(struct fd_batch *batch, struct fd_tile *tile)
 static void
 fd5_emit_tile_fini(struct fd_batch *batch)
 {
-	fd5_cache_flush(batch, batch->gmem);
-	fd5_set_render_mode(batch->ctx, batch->gmem, BYPASS);
+	struct fd_ringbuffer *ring = batch->gmem;
+
+	OUT_PKT7(ring, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
+	OUT_RING(ring, 0x0);
+
+	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
+	OUT_RING(ring, UNK_26);
+
+	fd5_cache_flush(batch, ring);
+	fd5_set_render_mode(batch->ctx, ring, BYPASS);
 }
 
 static void
@@ -545,6 +553,24 @@ fd5_emit_sysmem_prep(struct fd_batch *batch)
 			A5XX_GRAS_SC_DEST_MSAA_CNTL_MSAA_DISABLE);
 }
 
+static void
+fd5_emit_sysmem_fini(struct fd_batch *batch)
+{
+	struct fd5_context *fd5_ctx = fd5_context(batch->ctx);
+	struct fd_ringbuffer *ring = batch->gmem;
+
+	OUT_PKT7(ring, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
+	OUT_RING(ring, 0x0);
+
+	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
+	OUT_RING(ring, UNK_26);
+
+	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
+	OUT_RING(ring, UNK_1D);
+	OUT_RELOCW(ring, fd5_ctx->blit_mem, 0, 0, 0);  /* ADDR_LO/HI */
+	OUT_RING(ring, 0x00000000);
+}
+
 void
 fd5_gmem_init(struct pipe_context *pctx)
 {
@@ -557,4 +583,5 @@ fd5_gmem_init(struct pipe_context *pctx)
 	ctx->emit_tile_gmem2mem = fd5_emit_tile_gmem2mem;
 	ctx->emit_tile_fini = fd5_emit_tile_fini;
 	ctx->emit_sysmem_prep = fd5_emit_sysmem_prep;
+	ctx->emit_sysmem_fini = fd5_emit_sysmem_fini;
 }
diff --git a/src/gallium/drivers/freedreno/freedreno_context.h b/src/gallium/drivers/freedreno/freedreno_context.h
index e83b208ab8..995e7d4c43 100644
--- a/src/gallium/drivers/freedreno/freedreno_context.h
+++ b/src/gallium/drivers/freedreno/freedreno_context.h
@@ -261,6 +261,7 @@ struct fd_context {
 
 	/* optional, for GMEM bypass: */
 	void (*emit_sysmem_prep)(struct fd_batch *batch);
+	void (*emit_sysmem_fini)(struct fd_batch *batch);
 
 	/* draw: */
 	bool (*draw_vbo)(struct fd_context *ctx, const struct pipe_draw_info *info);
diff --git a/src/gallium/drivers/freedreno/freedreno_gmem.c b/src/gallium/drivers/freedreno/freedreno_gmem.c
index c7ac0a23a2..d9f707d9c9 100644
--- a/src/gallium/drivers/freedreno/freedreno_gmem.c
+++ b/src/gallium/drivers/freedreno/freedreno_gmem.c
@@ -358,6 +358,9 @@ render_sysmem(struct fd_batch *batch)
 	/* emit IB to drawcmds: */
 	ctx->emit_ib(batch->gmem, batch->draw);
 	fd_reset_wfi(batch);
+
+	if (ctx->emit_sysmem_fini)
+		ctx->emit_sysmem_fini(batch);
 }
 
 static void