i965: perf: flush batchbuffers at the beginning of queries

As Chris commented, it makes more sense to have batch buffer flushes
before the query. Usually applications like frame_retrace do a series
of queries and in that case, with flushes at the end of the queries,
we might still have the first query contained in 2 different batchs.
More generally it would be quite usual to have the query contained in
2 batch buffers because we never now what's the fill rate of the
current batch buffer.

If we move the flushing at the beginning of the queries, it's pretty
much guaranteed that queries will be contained in a single batch
buffer (unless the amount of commands is huge, but then it's only fair
to include reloading request times in the measurements).

Fixes: adafe4b733c02 ("i965: perf: minimize the chances to spread queries across batchbuffers")
Reported-by: Chris Wilson <[email protected]>
Signed-off-by: Lionel Landwerlin <[email protected]>
Cc: "17.2 17.1" <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>

1 files changed, 8 insertions, 8 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_performance_query.c b/src/mesa/drivers/dri/i965/brw_performance_query.c
index d7902de836c..d8680b48793 100644
--- a/src/mesa/drivers/dri/i965/brw_performance_query.c
+++ b/src/mesa/drivers/dri/i965/brw_performance_query.c
@@ -1212,6 +1212,14 @@ brw_begin_perf_query(struct gl_context *ctx,
       obj->oa.begin_report_id = brw->perfquery.next_query_start_report_id;
       brw->perfquery.next_query_start_report_id += 2;
 
+      /* We flush the batchbuffer here to minimize the chances that MI_RPC
+       * delimiting commands end up in different batchbuffers. If that's the
+       * case, the measurement will include the time it takes for the kernel
+       * scheduler to load a new request into the hardware. This is manifested in
+       * tools like frameretrace by spikes in the "GPU Core Clocks" counter.
+       */
+      intel_batchbuffer_flush(brw);
+
       /* Take a starting OA counter snapshot. */
       brw->vtbl.emit_mi_report_perf_count(brw, obj->oa.bo, 0,
                                           obj->oa.begin_report_id);
@@ -1298,14 +1306,6 @@ brw_end_perf_query(struct gl_context *ctx,
                                              obj->oa.begin_report_id + 1);
       }
 
-      /* We flush the batchbuffer here to minimize the chances that MI_RPC
-       * delimiting commands end up in different batchbuffers. If that's the
-       * case, the measurement will include the time it takes for the kernel
-       * scheduler to load a new request into the hardware. This is manifested
-       * in tools like frameretrace by spikes in the "GPU Core Clocks"
-       * counter.
-       */
-      intel_batchbuffer_flush(brw);
       --brw->perfquery.n_active_oa_queries;
 
       /* NB: even though the query has now ended, it can't be accumulated