Mesa (main): anv: fix performance bug in INTEL_MEASURE

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 34a130fedf0abe1596891e85df48fb2b208385ed
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=34a130fedf0abe1596891e85df48fb2b208385ed

Author: Mark Janes <markjanes at swizzler.org>
Date:   Mon May 16 12:52:42 2022 -0700

anv: fix performance bug in INTEL_MEASURE

Re-allocating the buffer object for snapshots carries a heavy penalty
at run-time.  When resetting a command buffer, the buffer object that
is allocated for snapshots may be re-used directly on subsequent
renders.

Stale snapshot data will persist in the buffer object.  To verify that
rendering is complete, zero the final timestamp value and check that
it has been written before gathering data.

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16571>

---

 src/intel/vulkan/anv_measure.c | 26 +++++++++++---------------
 1 file changed, 11 insertions(+), 15 deletions(-)

diff --git a/src/intel/vulkan/anv_measure.c b/src/intel/vulkan/anv_measure.c
index 2b2e9b672f8..7b82926a5dd 100644
--- a/src/intel/vulkan/anv_measure.c
+++ b/src/intel/vulkan/anv_measure.c
@@ -332,16 +332,6 @@ anv_measure_reset(struct anv_cmd_buffer *cmd_buffer)
    measure->base.frame = 0;
    measure->base.event_count = 0;
    list_inithead(&measure->base.link);
-
-   anv_device_release_bo(device, measure->bo);
-   ASSERTED VkResult result =
-      anv_device_alloc_bo(device, "measure data",
-                          config->batch_size * sizeof(uint64_t),
-                          ANV_BO_ALLOC_MAPPED,
-                          0,
-                          (struct anv_bo**)&measure->bo);
-   measure->base.timestamps = measure->bo->map;
-   assert(result == VK_SUCCESS);
 }
 
 void
@@ -403,19 +393,25 @@ _anv_measure_submit(struct anv_cmd_buffer *cmd_buffer)
    if (measure == NULL)
       return;
 
-   if (measure->base.index == 0)
+   struct intel_measure_batch *base = &measure->base;
+   if (base->index == 0)
       /* no snapshots were started */
       return;
 
    /* finalize snapshots and enqueue them */
    static unsigned cmd_buffer_count = 0;
-   measure->base.batch_count = p_atomic_inc_return(&cmd_buffer_count);
+   base->batch_count = p_atomic_inc_return(&cmd_buffer_count);
 
-   if (measure->base.index %2 == 1) {
-      anv_measure_end_snapshot(cmd_buffer, measure->base.event_count);
-      measure->base.event_count = 0;
+   if (base->index %2 == 1) {
+      anv_measure_end_snapshot(cmd_buffer, base->event_count);
+      base->event_count = 0;
    }
 
+   /* Mark the final timestamp as 'not completed'.  This marker will be used
+    * to verify that rendering is complete.
+    */
+   base->timestamps[base->index - 1] = 0;
+
    /* add to the list of submitted snapshots */
    pthread_mutex_lock(&measure_device->mutex);
    list_addtail(&measure->base.link, &measure_device->queued_snapshots);