anv/blorp: Align vertex buffers to 64B

This fixes issues seen when adding support for full 48-bit addresses.
The 48-bit addresses themselves have nothing to do with it other than
that it caused the kernel to place buffers slightly differently so they
interacted differently with the caches.

Reviewed-by: Kenneth Graunke <[email protected]>
Cc: "13.0 17.0" <[email protected]>

1 files changed, 14 insertions, 1 deletions

diff --git a/src/intel/vulkan/genX_blorp_exec.c b/src/intel/vulkan/genX_blorp_exec.c
index 1d07fd7b5ea..7f22b677ea6 100644
--- a/src/intel/vulkan/genX_blorp_exec.c
+++ b/src/intel/vulkan/genX_blorp_exec.c
@@ -115,8 +115,21 @@ blorp_alloc_vertex_buffer(struct blorp_batch *batch, uint32_t size,
                           struct blorp_address *addr)
 {
    struct anv_cmd_buffer *cmd_buffer = batch->driver_batch;
+
+   /* From the Skylake PRM, 3DSTATE_VERTEX_BUFFERS:
+    *
+    *    "The VF cache needs to be invalidated before binding and then using
+    *    Vertex Buffers that overlap with any previously bound Vertex Buffer
+    *    (at a 64B granularity) since the last invalidation.  A VF cache
+    *    invalidate is performed by setting the "VF Cache Invalidation Enable"
+    *    bit in PIPE_CONTROL."
+    *
+    * This restriction first appears in the Skylake PRM but the internal docs
+    * also list it as being an issue on Broadwell.  In order to avoid this
+    * problem, we align all vertex buffer allocations to 64 bytes.
+    */
    struct anv_state vb_state =
-      anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 16);
+      anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 64);
 
    *addr = (struct blorp_address) {
       .buffer = &cmd_buffer->device->dynamic_state_block_pool.bo,