anv/gpu_memcpy: CS Stall before a MI memcpy on gen7

This fixes a pile of hangs caused by the recent shuffling of resolves
and transitions.  The particularly problematic case is when you have at
least three attachments with load ops of CLEAR, LOAD, CLEAR.  In this
case, we execute the first CLEAR followed by a MI memcpy to copy the
clear values over for the LOAD followed by a second CLEAR.  The MI
commands cause the first CLEAR to hang which causes us to get stuck on
the 3DSTATE_MULTISAMPLE in the second CLEAR.

We also add guards for BLORP to fix the same issue.  These shouldn't
actually do anything right now because the only use of indirect clears
in BLORP today is for resolves which are already guarded by a render
cache flush and CS stall.  However, this will guard us against potential
issues in the future.

Acked-by: Kenneth Graunke <[email protected]>
Acked-by: Nanley Chery <[email protected]>

1 files changed, 22 insertions, 0 deletions

diff --git a/src/intel/vulkan/genX_gpu_memcpy.c b/src/intel/vulkan/genX_gpu_memcpy.c
index f3ada93333b..1d527fb1549 100644
--- a/src/intel/vulkan/genX_gpu_memcpy.c
+++ b/src/intel/vulkan/genX_gpu_memcpy.c
@@ -62,6 +62,28 @@ genX(cmd_buffer_mi_memcpy)(struct anv_cmd_buffer *cmd_buffer,
    assert(dst_offset % 4 == 0);
    assert(src_offset % 4 == 0);
 
+#if GEN_GEN == 7
+   /* On gen7, the combination of commands used here(MI_LOAD_REGISTER_MEM
+    * and MI_STORE_REGISTER_MEM) can cause GPU hangs if any rendering is
+    * in-flight when they are issued even if the memory touched is not
+    * currently active for rendering.  The weird bit is that it is not the
+    * MI_LOAD/STORE_REGISTER_MEM commands which hang but rather the in-flight
+    * rendering hangs such that the next stalling command after the
+    * MI_LOAD/STORE_REGISTER_MEM commands will catch the hang.
+    *
+    * It is unclear exactly why this hang occurs.  Both MI commands come with
+    * warnings about the 3D pipeline but that doesn't seem to fully explain
+    * it.  My (Jason's) best theory is that it has something to do with the
+    * fact that we're using a GPU state register as our temporary and that
+    * something with reading/writing it is causing problems.
+    *
+    * In order to work around this issue, we emit a PIPE_CONTROL with the
+    * command streamer stall bit set.
+    */
+   cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
+   genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+#endif
+
    for (uint32_t i = 0; i < size; i += 4) {
       const struct anv_address src_addr =
          (struct anv_address) { src, src_offset + i};