anv/batch_chain: Don't start a new BO just for BATCH_BUFFER_START
authorJason Ekstrand <jason.ekstrand@intel.com>
Tue, 2 Oct 2018 22:19:32 +0000 (17:19 -0500)
committerJason Ekstrand <jason.ekstrand@intel.com>
Wed, 3 Oct 2018 14:03:12 +0000 (09:03 -0500)
Previously, we just went ahead and emitted MI_BATCH_BUFFER_START as
normal.  If we are near enough to the end, this can cause us to start a
new BO just for the MI_BATCH_BUFFER_START which messes up chaining.  We
always reserve enough space at the end for an MI_BATCH_BUFFER_START so
we can just increment cmd_buffer->batch.end prior to emitting the
command.

Fixes: a0b133286a3 "anv/batch_chain: Simplify secondary batch return..."
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=107926
Tested-by: Alex Smith <asmith@feralinteractive.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
src/intel/vulkan/anv_batch_chain.c

index 3e13553ac18cd6e39be1a0755fe0d77cb56e7f00..e08e07ad7bd7062416c9c54256e3f92c8f09e755 100644 (file)
@@ -894,8 +894,17 @@ anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer *cmd_buffer)
           * It doesn't matter where it points now so long as has a valid
           * relocation.  We'll adjust it later as part of the chaining
           * process.
+          *
+          * We set the end of the batch a little short so we would be sure we
+          * have room for the chaining command.  Since we're about to emit the
+          * chaining command, let's set it back where it should go.
           */
+         cmd_buffer->batch.end += GEN8_MI_BATCH_BUFFER_START_length * 4;
+         assert(cmd_buffer->batch.start == batch_bo->bo.map);
+         assert(cmd_buffer->batch.end == batch_bo->bo.map + batch_bo->bo.size);
+
          emit_batch_buffer_start(cmd_buffer, &batch_bo->bo, 0);
+         assert(cmd_buffer->batch.start == batch_bo->bo.map);
       } else {
          cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN;
       }