i965: Emit CS stall before MEDIA_VFE_STATE.

This fixes hangs on GFXBench 5's Aztec Ruins benchmark.

Unfortunately, it regresses OglCSCloth performance by about 10%. There
are some ideas for fixing that.

The Vulkan driver already emits this stall.

Reviewed-by: Matt Turner <mattst88@gmail.com>

diff --git a/src/mesa/drivers/dri/i965/genX_state_upload.c b/src/mesa/drivers/dri/i965/genX_state_upload.c
index 533ef8b..04a4925 100644 (file)
--- a/src/mesa/drivers/dri/i965/genX_state_upload.c
+++ b/src/mesa/drivers/dri/i965/genX_state_upload.c
@@ -4169,6 +4169,18 @@ genX(upload_cs_state)(struct brw_context *brw)
    uint32_t *bind = brw_state_batch(brw, prog_data->binding_table.size_bytes,
                                     32, &stage_state->bind_bo_offset);
 
+   /* The MEDIA_VFE_STATE documentation for Gen8+ says:
+    *
+    * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
+    *  the only bits that are changed are scoreboard related: Scoreboard
+    *  Enable, Scoreboard Type, Scoreboard Mask, Scoreboard * Delta. For
+    *  these scoreboard related states, a MEDIA_STATE_FLUSH is sufficient."
+    *
+    * Earlier generations say "MI_FLUSH" instead of "stalling PIPE_CONTROL",
+    * but MI_FLUSH isn't really a thing, so we assume they meant PIPE_CONTROL.
+    */
+   brw_emit_pipe_control_flush(brw, PIPE_CONTROL_CS_STALL);
+
    brw_batch_emit(brw, GENX(MEDIA_VFE_STATE), vfe) {
       if (prog_data->total_scratch) {
          uint32_t bo_offset;