+ assert(USED_BATCH(brw->batch) - batch_used <= MI_REPORT_PERF_COUNT_BATCH_DWORDS * 4);
+}
+
+/**
+ * Add a monitor to the global list of "unresolved monitors."
+ *
+ * Monitors are "unresolved" if they refer to OA counter snapshots in
+ * bookend_bo. Results (even partial ones) must be gathered for all
+ * unresolved monitors before it's safe to discard bookend_bo.
+ */
+static void
+add_to_unresolved_monitor_list(struct brw_context *brw,
+ struct brw_perf_monitor_object *monitor)
+{
+ if (brw->perfmon.unresolved_elements >=
+ brw->perfmon.unresolved_array_size) {
+ brw->perfmon.unresolved_array_size *= 2;
+ brw->perfmon.unresolved = reralloc(brw, brw->perfmon.unresolved,
+ struct brw_perf_monitor_object *,
+ brw->perfmon.unresolved_array_size);
+ }
+
+ brw->perfmon.unresolved[brw->perfmon.unresolved_elements++] = monitor;
+}
+
+/**
+ * If possible, throw away the contents of bookend BO.
+ *
+ * When all monitoring stops, and no monitors need data from bookend_bo to
+ * compute results, we can discard it and start writing snapshots at the
+ * beginning again. This helps reduce the amount of buffer wraparound.
+ */
+static void
+clean_bookend_bo(struct brw_context *brw)
+{
+ if (brw->perfmon.unresolved_elements == 0) {
+ DBG("***Resetting bookend snapshots to 0\n");
+ brw->perfmon.bookend_snapshots = 0;
+ }
+}
+
+/**
+ * Remove a monitor from the global list of "unresolved monitors."
+ *
+ * This can happen when:
+ * - We finish computing a completed monitor's results.
+ * - We discard unwanted monitor results.
+ * - A monitor's results can be computed without relying on bookend_bo.
+ */
+static void
+drop_from_unresolved_monitor_list(struct brw_context *brw,
+ struct brw_perf_monitor_object *monitor)
+{
+ for (int i = 0; i < brw->perfmon.unresolved_elements; i++) {
+ if (brw->perfmon.unresolved[i] == monitor) {
+ int last_elt = --brw->perfmon.unresolved_elements;
+
+ if (i == last_elt) {
+ brw->perfmon.unresolved[i] = NULL;
+ } else {
+ brw->perfmon.unresolved[i] = brw->perfmon.unresolved[last_elt];
+ }
+
+ clean_bookend_bo(brw);
+ return;
+ }
+ }
+}
+
+/**
+ * Given pointers to starting and ending OA snapshots, add the deltas for each
+ * counter to the results.
+ */
+static void
+add_deltas(struct brw_context *brw,
+ struct brw_perf_monitor_object *monitor,
+ uint32_t *start, uint32_t *end)
+{
+ /* Look for expected report ID values to ensure data is present. */
+ assert(start[0] == REPORT_ID);
+ assert(end[0] == REPORT_ID);
+
+ /* Subtract each counter's ending and starting values, then add the
+ * difference to the counter's value so far.
+ */
+ for (int i = 3; i < brw->perfmon.entries_per_oa_snapshot; i++) {
+ /* When debugging, it's useful to note when the ending value is less than
+ * the starting value; aggregating counters should always increase in
+ * value (or remain unchanged). This happens periodically due to
+ * wraparound, but can also indicate serious problems.
+ */
+#ifdef DEBUG
+ if (end[i] < start[i]) {
+ int counter = brw->perfmon.oa_snapshot_layout[i];
+ if (counter >= 0) {
+ DBG("WARNING: \"%s\" ending value was less than the starting "
+ "value: %u < %u (end - start = %u)\n",
+ brw->ctx.PerfMonitor.Groups[0].Counters[counter].Name,
+ end[i], start[i], end[i] - start[i]);
+ }
+ }
+#endif
+ monitor->oa_results[i] += end[i] - start[i];
+ }
+}
+
+/**
+ * Gather OA counter results (partial or full) from a series of snapshots.
+ *
+ * Monitoring can start or stop at any time, likely at some point mid-batch.
+ * We write snapshots for both events, storing them in monitor->oa_bo.
+ *
+ * Ideally, we would simply subtract those two snapshots to obtain the final
+ * counter results. Unfortunately, our hardware doesn't preserve their values
+ * across context switches or GPU sleep states. In order to support multiple
+ * concurrent OA clients, as well as reliable data across power management,
+ * we have to take snapshots at the start and end of batches as well.
+ *
+ * This results in a three-part sequence of (start, end) intervals:
+ * - The "head" is from the BeginPerfMonitor snapshot to the end of the first
+ * batchbuffer.
+ * - The "middle" is a series of (batch start, batch end) snapshots which
+ * bookend any batchbuffers between the ones which start/end monitoring.
+ * - The "tail" is from the start of the last batch where monitoring was
+ * active to the EndPerfMonitor snapshot.
+ *
+ * Due to wrapping in the bookend BO, we may have to accumulate partial results.
+ * If so, we handle the "head" and any "middle" results so far. When monitoring
+ * eventually ends, we handle additional "middle" batches and the "tail."
+ */
+static void
+gather_oa_results(struct brw_context *brw,
+ struct brw_perf_monitor_object *monitor,
+ uint32_t *bookend_buffer)
+{
+ struct gl_perf_monitor_object *m = &monitor->base;
+ assert(monitor->oa_bo != NULL);
+
+ drm_intel_bo_map(monitor->oa_bo, false);
+ uint32_t *monitor_buffer = monitor->oa_bo->virtual;
+
+ /* If monitoring was entirely contained within a single batch, then the
+ * bookend BO is irrelevant. Just subtract monitor->bo's two snapshots.
+ */
+ if (monitor->oa_middle_start == -1) {
+ add_deltas(brw, monitor,
+ monitor_buffer,
+ monitor_buffer + (SECOND_SNAPSHOT_OFFSET_IN_BYTES /
+ sizeof(uint32_t)));
+ drm_intel_bo_unmap(monitor->oa_bo);
+ return;
+ }
+
+ const ptrdiff_t snapshot_size = brw->perfmon.entries_per_oa_snapshot;
+
+ /* First, add the contributions from the "head" interval:
+ * (snapshot taken at BeginPerfMonitor time,
+ * snapshot taken at the end of the first batch after monitoring began)
+ */
+ if (monitor->oa_head_end != -1) {
+ assert(monitor->oa_head_end < brw->perfmon.bookend_snapshots);
+ add_deltas(brw, monitor,
+ monitor_buffer,
+ bookend_buffer + snapshot_size * monitor->oa_head_end);
+
+ /* Make sure we don't count the "head" again in the future. */
+ monitor->oa_head_end = -1;
+ }
+
+ /* Next, count the contributions from the "middle" batches. These are
+ * (batch begin, batch end) deltas while monitoring was active.
+ */
+ int last_snapshot;
+ if (m->Ended)
+ last_snapshot = monitor->oa_tail_start;
+ else
+ last_snapshot = brw->perfmon.bookend_snapshots;
+
+ for (int s = monitor->oa_middle_start; s < last_snapshot; s += 2) {
+ add_deltas(brw, monitor,
+ bookend_buffer + snapshot_size * s,
+ bookend_buffer + snapshot_size * (s + 1));
+ }
+
+ /* Finally, if the monitor has ended, we need to count the contributions of
+ * the "tail" interval:
+ * (start of the batch where monitoring ended, EndPerfMonitor snapshot)
+ */
+ if (m->Ended) {
+ assert(monitor->oa_tail_start != -1);
+ add_deltas(brw, monitor,
+ bookend_buffer + snapshot_size * monitor->oa_tail_start,
+ monitor_buffer + (SECOND_SNAPSHOT_OFFSET_IN_BYTES /
+ sizeof(uint32_t)));
+ }
+
+ drm_intel_bo_unmap(monitor->oa_bo);
+
+ /* If the monitor has ended, then we've gathered all the results, and
+ * can free the monitor's OA BO.
+ */
+ if (m->Ended) {
+ drm_intel_bo_unreference(monitor->oa_bo);
+ monitor->oa_bo = NULL;
+
+ /* The monitor's OA result is now resolved. */
+ DBG("Marking %d resolved - results gathered\n", m->Name);
+ drop_from_unresolved_monitor_list(brw, monitor);
+ }
+}
+
+/**
+ * Handle running out of space in the bookend BO.
+ *
+ * When we run out of space in the bookend BO, we need to gather up partial
+ * results for every unresolved monitor. This allows us to free the snapshot
+ * data in bookend_bo, freeing up the space for reuse. We call this "wrapping."
+ *
+ * This will completely compute the result for any unresolved monitors that
+ * have ended.
+ */
+static void
+wrap_bookend_bo(struct brw_context *brw)
+{
+ DBG("****Wrap bookend BO****\n");
+ /* Note that wrapping will only occur at the start of a batch, since that's
+ * where we reserve space. So the current batch won't reference bookend_bo
+ * or any monitor BOs. This means we don't need to worry about
+ * synchronization.
+ *
+ * Also, EndPerfMonitor guarantees that only monitors which span multiple
+ * batches exist in the unresolved monitor list.
+ */
+ assert(brw->perfmon.oa_users > 0);
+
+ drm_intel_bo_map(brw->perfmon.bookend_bo, false);
+ uint32_t *bookend_buffer = brw->perfmon.bookend_bo->virtual;
+ for (int i = 0; i < brw->perfmon.unresolved_elements; i++) {
+ struct brw_perf_monitor_object *monitor = brw->perfmon.unresolved[i];
+ struct gl_perf_monitor_object *m = &monitor->base;
+
+ gather_oa_results(brw, monitor, bookend_buffer);
+
+ if (m->Ended) {
+ /* gather_oa_results() dropped the monitor from the unresolved list,
+ * throwing our indices off by one.
+ */
+ --i;
+ } else {
+ /* When we create the new bookend_bo, snapshot #0 will be the
+ * beginning of another "middle" BO.
+ */
+ monitor->oa_middle_start = 0;
+ assert(monitor->oa_head_end == -1);
+ assert(monitor->oa_tail_start == -1);
+ }
+ }
+ drm_intel_bo_unmap(brw->perfmon.bookend_bo);
+
+ brw->perfmon.bookend_snapshots = 0;
+}
+
+/* This is fairly arbitrary; the trade off is memory usage vs. extra overhead
+ * from wrapping. On Gen7, 32768 should be enough for 128 snapshots before
+ * wrapping (since each is 256 bytes).
+ */
+#define BOOKEND_BO_SIZE_BYTES 32768
+
+/**
+ * Check whether bookend_bo has space for a given number of snapshots.
+ */
+static bool
+has_space_for_bookend_snapshots(struct brw_context *brw, int snapshots)
+{
+ int snapshot_bytes = brw->perfmon.entries_per_oa_snapshot * sizeof(uint32_t);
+
+ /* There are brw->perfmon.bookend_snapshots - 1 existing snapshots. */
+ int total_snapshots = (brw->perfmon.bookend_snapshots - 1) + snapshots;
+
+ return total_snapshots * snapshot_bytes < BOOKEND_BO_SIZE_BYTES;
+}
+
+/**
+ * Write an OA counter snapshot to bookend_bo.
+ */
+static void
+emit_bookend_snapshot(struct brw_context *brw)
+{
+ int snapshot_bytes = brw->perfmon.entries_per_oa_snapshot * sizeof(uint32_t);
+ int offset_in_bytes = brw->perfmon.bookend_snapshots * snapshot_bytes;
+
+ emit_mi_report_perf_count(brw, brw->perfmon.bookend_bo, offset_in_bytes,
+ REPORT_ID);
+ ++brw->perfmon.bookend_snapshots;
projects / mesa.git / blobdiff