fix turbo mode

src/driver/amdxdna/aie2_ctx.c

@@ -540,6 +540,7 @@ int aie2_hwctx_init(struct amdxdna_hwctx *hwctx)
 	struct drm_gpu_scheduler *sched;
 	struct amdxdna_hwctx_priv *priv;
 	struct amdxdna_gem_obj *heap;
+	unsigned int wq_flags;
 	int i, ret;
 
 	priv = kzalloc(sizeof(*hwctx->priv), GFP_KERNEL);
@@ -587,12 +588,21 @@ int aie2_hwctx_init(struct amdxdna_hwctx *hwctx)
 
 	sched = &priv->sched;
 	mutex_init(&priv->io_lock);
-	ret = drm_sched_init(sched, &sched_ops, NULL, DRM_SCHED_PRIORITY_COUNT,
+
+	wq_flags = __WQ_ORDERED;
+	if (!aie2_pm_is_turbo(xdna->dev_handle))
+		wq_flags |= WQ_UNBOUND;
+	priv->submit_wq = alloc_workqueue(hwctx->name, wq_flags, 1);
+	if (!priv->submit_wq) {
+		XDNA_ERR(xdna, "Failed to alloc submit wq");
+		goto free_cmd_bufs;
+	}
+	ret = drm_sched_init(sched, &sched_ops, priv->submit_wq, DRM_SCHED_PRIORITY_COUNT,
 			     HWCTX_MAX_CMDS, 0, MAX_SCHEDULE_TIMEOUT,
 			     NULL, NULL, hwctx->name, xdna->ddev.dev);
 	if (ret) {
 		XDNA_ERR(xdna, "Failed to init DRM scheduler. ret %d", ret);
-		goto free_cmd_bufs;
+		goto free_wq;
 	}
 
 	ret = drm_sched_entity_init(&priv->entity, DRM_SCHED_PRIORITY_NORMAL,
@@ -645,6 +655,8 @@ int aie2_hwctx_init(struct amdxdna_hwctx *hwctx)
 	drm_sched_entity_destroy(&priv->entity);
 free_sched:
 	drm_sched_fini(&priv->sched);
+free_wq:
+	destroy_workqueue(priv->submit_wq);
 free_cmd_bufs:
 	for (i = 0; i < ARRAY_SIZE(priv->cmd_buf); i++) {
 		if (!priv->cmd_buf[i])
@@ -681,6 +693,7 @@ void aie2_hwctx_fini(struct amdxdna_hwctx *hwctx)
 	aie2_hwctx_wait_for_idle(hwctx);
 	drm_sched_entity_destroy(&hwctx->priv->entity);
 	drm_sched_fini(&hwctx->priv->sched);
+	destroy_workqueue(hwctx->priv->submit_wq);
 
 	for (idx = 0; idx < HWCTX_MAX_CMDS; idx++) {
 		job = hwctx->priv->pending[idx];

src/driver/amdxdna/aie2_pci.h

@@ -199,6 +199,7 @@ struct amdxdna_hwctx_priv {
 	u32				num_pending;
 
 	struct amdxdna_gem_obj		*cmd_buf[HWCTX_MAX_CMDS];
+	struct workqueue_struct		*submit_wq;
 };
 
 struct async_events;

src/driver/tools/npu_perf_analyze.sh

@@ -98,6 +98,8 @@ fi
 echo "${event2_ts_num} events for: '${event2}'"
 
 # Caculate time difference between two events
+diffs_event1=()
+diffs_event2=()
 diffs=()
 i1=0
 i2=0
@@ -118,6 +120,8 @@ while [ 1 ]; do
 
 
 	(( i1-- ))
+	diffs_event1+=( $((event1_ts[i1])) )
+	diffs_event2+=( $((event2_ts[i2])) )
 	diffs+=( $((event2_ts[i2] - event1_ts[i1])) )
 	(( i1++ ))
 	(( i2++ ))
@@ -162,6 +166,6 @@ done
 
 # Output result
 total_events=$(( range_end - range_start ))
-echo Average over ${total_events} events: $(( total / total_events ))us
-echo Largest:  ${largest}us@${largest_idx}
-echo Smallest: ${smallest}us@${smallest_idx}
+echo Average over ${total_events} events: $(( total / total_events ))ns
+echo Largest:  ${largest}ns@${largest_idx}: event1=${diffs_event1[largest_idx]}, event2=${diffs_event2[largest_idx]}
+echo Smallest: ${smallest}ns@${smallest_idx}: event1=${diffs_event1[smallest_idx]}, event2=${diffs_event2[smallest_idx]}

src/driver/tools/npu_perf_trace.sh

@@ -127,7 +127,7 @@ eval $command
 tmp_file=/tmp/perf.out
 # convert timestamp from second to microsecond to avoid floating numbers
 #perf script | awk '{ $4=$4*1000000; print }' > ${tmp_file}
-perf script --reltime > ${tmp_file}
+perf script --reltime --ns > ${tmp_file}
 # replace IOCTL cmd number to name
 sed "$ioctl_sed_expr" "${tmp_file}" > perf.converted.out
 rm -rf ${tmp_file}