[Mesa-dev] [PATCH 41/70] i965: Convert performance monitor over to begin/end batch

[Chris Wilson]

We need to wrap all batch buffer access inside begin/end. In order to
make that transformation for brw_performance_monitor.c requires a small
amount of code motion, but otherwise straightforward.

To simply the number of places where we try and access the batch, we
move the start/stop based on oa_users into the critical sections.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 src/mesa/drivers/dri/i965/brw_context.c            |  3 ++
 .../drivers/dri/i965/brw_performance_monitor.c     | 38 ++++++++++------------
 2 files changed, 20 insertions(+), 21 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_context.c b/src/mesa/drivers/dri/i965/brw_context.c
index e8d43db..26041e6 100644
--- a/src/mesa/drivers/dri/i965/brw_context.c
+++ b/src/mesa/drivers/dri/i965/brw_context.c
@@ -1575,6 +1575,9 @@ void brw_batch_finish_hook(brw_batch *batch)
     */
    brw_emit_query_end(brw);
 
+   /* We may also need to snapshot and disable OA counters. */
+   brw_perf_monitor_finish_batch(brw);
+
    brw->cache.bo_used_by_gpu = true;
 }
 
diff --git a/src/mesa/drivers/dri/i965/brw_performance_monitor.c b/src/mesa/drivers/dri/i965/brw_performance_monitor.c
index dfc003f..5545ac7 100644
--- a/src/mesa/drivers/dri/i965/brw_performance_monitor.c
+++ b/src/mesa/drivers/dri/i965/brw_performance_monitor.c
@@ -703,12 +703,6 @@ emit_mi_report_perf_count(struct brw_context *brw,
 {
    assert(offset_in_bytes % 64 == 0);
 
-   /* Make sure the commands to take a snapshot fits in a single batch. */
-   intel_batchbuffer_require_space(&brw->batch,
-                                   MI_REPORT_PERF_COUNT_BATCH_DWORDS * 4,
-                                   RENDER_RING);
-   int batch_used = USED_BATCH(&brw->batch);
-
    /* Reports apparently don't always get written unless we flush first. */
    brw_emit_mi_flush(brw);
 
@@ -749,9 +743,6 @@ emit_mi_report_perf_count(struct brw_context *brw,
 
    /* Reports apparently don't always get written unless we flush after. */
    brw_emit_mi_flush(brw);
-
-   (void) batch_used;
-   assert(USED_BATCH(&brw->batch) - batch_used <= MI_REPORT_PERF_COUNT_BATCH_DWORDS * 4);
 }
 
 /**
@@ -1101,6 +1092,13 @@ brw_begin_perf_monitor(struct gl_context *ctx,
    DBG("Begin(%d)\n", m->Name);
 
    reinitialize_perf_monitor(brw, monitor);
+
+   /* Ensure the OACONTROL enable and snapshot land in the same batch. */
+   if (brw_batch_begin(&brw->batch,
+                       (MI_REPORT_PERF_COUNT_BATCH_DWORDS + 3) * 4,
+                       RENDER_RING) < 0)
+      return false;
+
    if (monitor_needs_oa(brw, m)) {
       /* If the global OA bookend BO doesn't exist, allocate it.  This should
        * only happen once, but we delay until BeginPerfMonitor time to avoid
@@ -1126,12 +1124,8 @@ brw_begin_perf_monitor(struct gl_context *ctx,
          calloc(brw->perfmon.entries_per_oa_snapshot, sizeof(uint32_t));
 
       /* If the OA counters aren't already on, enable them. */
-      if (brw->perfmon.oa_users == 0) {
-         /* Ensure the OACONTROL enable and snapshot land in the same batch. */
-         int space = (MI_REPORT_PERF_COUNT_BATCH_DWORDS + 3) * 4;
-         intel_batchbuffer_require_space(&brw->batch, space, RENDER_RING);
+      if (brw->perfmon.oa_users++ == 0)
          start_oa_counters(brw);
-      }
 
       /* Take a starting OA counter snapshot. */
       emit_mi_report_perf_count(brw, monitor->oa_bo, 0, REPORT_ID);
@@ -1142,8 +1136,6 @@ brw_begin_perf_monitor(struct gl_context *ctx,
 
       /* Add the monitor to the unresolved list. */
       add_to_unresolved_monitor_list(brw, monitor);
-
-      ++brw->perfmon.oa_users;
    }
 
    if (monitor_needs_statistics_registers(brw, m)) {
@@ -1155,7 +1147,7 @@ brw_begin_perf_monitor(struct gl_context *ctx,
       snapshot_statistics_registers(brw, monitor, 0);
    }
 
-   return true;
+   return brw_batch_end(&brw->batch) == 0;
 }
 
 /**
@@ -1170,14 +1162,16 @@ brw_end_perf_monitor(struct gl_context *ctx,
 
    DBG("End(%d)\n", m->Name);
 
+   if (brw_batch_begin(&brw->batch, 80, RENDER_RING) < 0)
+      return;
+
    if (monitor_needs_oa(brw, m)) {
+      assert(brw->perfmon.oa_users > 0);
+
       /* Take an ending OA counter snapshot. */
       emit_mi_report_perf_count(brw, monitor->oa_bo,
                                 SECOND_SNAPSHOT_OFFSET_IN_BYTES, REPORT_ID);
-
-      --brw->perfmon.oa_users;
-
-      if (brw->perfmon.oa_users == 0)
+      if (--brw->perfmon.oa_users == 0)
          stop_oa_counters(brw);
 
       if (monitor->oa_head_end == brw->perfmon.bookend_snapshots) {
@@ -1208,6 +1202,8 @@ brw_end_perf_monitor(struct gl_context *ctx,
       snapshot_statistics_registers(brw, monitor,
                                     SECOND_SNAPSHOT_OFFSET_IN_BYTES);
    }
+
+   brw_batch_end(&brw->batch);
 }
 
 /**
-- 
2.5.0