[PATCH] etnaviv: fix bogus flush requests in transfer handling

Lucas Stach l.stach at pengutronix.de
Fri Sep 29 13:40:11 UTC 2017


The logic to decide if we need to flush the GPU command stream was broken
and hard to reason about. Fix and clarify this.

Fixes the data sync subtests from piglit arb_vertex_buffer_object.

Signed-off-by: Lucas Stach <l.stach at pengutronix.de>
---
 src/gallium/drivers/etnaviv/etnaviv_transfer.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/src/gallium/drivers/etnaviv/etnaviv_transfer.c b/src/gallium/drivers/etnaviv/etnaviv_transfer.c
index ee5cda5e8ec2..08ec1987d8f7 100644
--- a/src/gallium/drivers/etnaviv/etnaviv_transfer.c
+++ b/src/gallium/drivers/etnaviv/etnaviv_transfer.c
@@ -243,11 +243,16 @@ etna_transfer_map(struct pipe_context *pctx, struct pipe_resource *prsc,
 
    struct etna_resource_level *res_level = &rsc->levels[level];
 
-   /* Always sync if we have the temporary resource.  The PIPE_TRANSFER_READ
-    * case could be optimised if we knew whether the resource has outstanding
-    * rendering. */
-   if ((usage & PIPE_TRANSFER_READ || trans->rsc) &&
-       rsc->status & ETNA_PENDING_WRITE)
+   /*
+    * Always flush if we have the temporary resource and have a copy to this
+    * outstanding. Otherwise infer flush requirement from resource access and
+    * current GPU usage (reads must wait for GPU writes, writes must have
+    * exclusive access to the buffer).
+    */
+   if ((trans->rsc && (etna_resource(trans->rsc)->status & ETNA_PENDING_WRITE)) ||
+       (!trans->rsc &&
+        (((usage & PIPE_TRANSFER_READ) && (rsc->status & ETNA_PENDING_WRITE)) ||
+        ((usage & PIPE_TRANSFER_WRITE) && rsc->status))))
       pctx->flush(pctx, NULL, 0);
 
    /* XXX we don't handle PIPE_TRANSFER_FLUSH_EXPLICIT; this flag can be ignored
-- 
2.11.0



More information about the etnaviv mailing list