[Mesa-dev] [PATCH v2 12/15] i965: Make BLORP properly avoid batch wrapping.

[Chris Wilson]

Quoting Kenneth Graunke (2017-09-13 21:54:14)
> We need to set brw->no_batch_wrap to actually avoid flushing in the
> middle of our BLORP operation, and instead grow the batchbuffer.
> ---
>  src/mesa/drivers/dri/i965/genX_blorp_exec.c | 16 ++--------------
>  1 file changed, 2 insertions(+), 14 deletions(-)
> 
> diff --git a/src/mesa/drivers/dri/i965/genX_blorp_exec.c b/src/mesa/drivers/dri/i965/genX_blorp_exec.c
> index feb87923ccb..5bff7eaff59 100644
> --- a/src/mesa/drivers/dri/i965/genX_blorp_exec.c
> +++ b/src/mesa/drivers/dri/i965/genX_blorp_exec.c
> @@ -224,9 +224,7 @@ genX(blorp_exec)(struct blorp_batch *batch,
>  retry:
>     intel_batchbuffer_require_space(brw, estimated_max_batch_usage, RENDER_RING);
>     intel_batchbuffer_save_state(brw);
> -   struct brw_bo *saved_bo = brw->batch.bo;
> -   uint32_t saved_used = USED_BATCH(brw->batch);
> -   uint32_t saved_state_used = brw->batch.state_used;
> +   brw->no_batch_wrap = true;
>  
>  #if GEN_GEN == 6
>     /* Emit workaround flushes when we switch from drawing to blorping. */
> @@ -254,17 +252,7 @@ retry:
>  
>     blorp_exec(batch, params);
>  
> -   /* Make sure we didn't wrap the batch unintentionally, and make sure we
> -    * reserved enough space that a wrap will never happen.
> -    */
> -   assert(brw->batch.bo == saved_bo);
> -   assert((USED_BATCH(brw->batch) - saved_used) * 4 +
> -          (brw->batch.state_used - saved_state_used) <
> -          estimated_max_batch_usage);
> -   /* Shut up compiler warnings on release build */
> -   (void)saved_bo;
> -   (void)saved_used;
> -   (void)saved_state_used;
> +   brw->no_batch_wrap = false;

Hmm, did you add an assert(brw->no_batch_wrap) into do_flush_locked()?
Would be good to have that assertion back now that you should have fixed
all the early flushing...
-Chris