intel_gpu: properly reset metrics

Reset the appropriate fields in the context struct when PAPI_cleanup() is called. This corresponds to intel_gpu_update_control_state() being called with the value 'count' equal to zero. Also reset the internal metric counts when PAPI_start() is called, so that the previous values do not persist after the next PAPI_start(). These changes have been tested on the Intel Ponte Vecchio architecture.
icl-utk-edu · Oct 3, 2024 · 1942be9 · 1942be9
1 parent 437cee0
commit 1942be9
Show file tree

Hide file tree

Showing 2 changed files with 52 additions and 62 deletions.
diff --git a/src/components/intel_gpu/internal/src/GPUMetricHandler.cpp b/src/components/intel_gpu/internal/src/GPUMetricHandler.cpp
@@ -1230,18 +1230,6 @@ int GPUMetricHandler::EnableTimeBasedStream(uint32_t timePeriod, uint32_t numRep
 	} else {
 		DebugPrintError("EnableTimeBasedStream: failed on device [%p], status 0x%x\n", 
 						m_device, status);
-		if (m_metricStreamer) {
-			status = zetMetricStreamerCloseFunc(m_metricStreamer);
-			m_metricStreamer = nullptr;
-		}
-		if (m_event) {
-			status = zeEventDestroyFunc(m_event);
-			m_event = nullptr;
-		}
-		if (m_eventPool) {
-			status = zeEventPoolDestroyFunc(m_eventPool);
-			m_eventPool = nullptr;
-		}
 		status = zetContextActivateMetricGroupsFunc(m_context, m_device, 0, nullptr);
 		m_status = COLLECTION_INIT;
 		ret = 1;
@@ -1331,22 +1319,6 @@ int GPUMetricHandler::EnableEventBasedQuery()
 		ret  = 0;
 	} else {
 		DebugPrintError("EnableEventBasedQuery: failed with status 0x%x, abort.\n", status);
-		if (m_tracer) {
-			status = zetTracerExpDestroyFunc(m_tracer);
-			m_tracer = nullptr;
-		}
-		if (m_event) {
-			status = zeEventDestroyFunc(m_event);
-			m_event = nullptr;
-		}
-		if (m_eventPool) {
-			status = zeEventPoolDestroyFunc(m_eventPool);
-			m_eventPool = nullptr;
-		}
-		if (m_queryPool) {
-			status = zetMetricQueryPoolDestroyFunc(m_queryPool);
-			m_queryPool = nullptr;
-		}
 		status = zetContextActivateMetricGroupsFunc(m_context, m_device, 0, nullptr);
 		m_status = COLLECTION_INIT;
 		ret  = retError;
@@ -1375,30 +1347,6 @@ GPUMetricHandler::DisableMetricGroup()
 	}
 	m_status = COLLECTION_DISABLED;
 
-	if (m_groupType == ZET_METRIC_GROUP_SAMPLING_TYPE_FLAG_TIME_BASED) {
-		if (m_metricStreamer) {
-			zetMetricStreamerCloseFunc(m_metricStreamer);
-			m_metricStreamer = nullptr;
-		}
-	}
-	if (m_groupType == ZET_METRIC_GROUP_SAMPLING_TYPE_FLAG_EVENT_BASED) {
-		if (m_tracer) {
-			zetTracerExpDestroyFunc(m_tracer);
-			m_tracer = nullptr;
-		}
-		if (m_queryPool) {
-			zetMetricQueryPoolDestroyFunc(m_queryPool);
-			m_queryPool = nullptr;
-		}
-	}
-	if (m_event) {
-		zeEventDestroyFunc(m_event);
-		m_event = nullptr;
-	}
-	if (m_eventPool) {
-		zeEventPoolDestroyFunc(m_eventPool);
-		m_eventPool = nullptr;
-	}
 	zetContextActivateMetricGroupsFunc(m_context,  m_device, 0, nullptr);
 	m_lock.unlock();
 	return;

diff --git a/src/components/intel_gpu/linux_intel_gpu_metrics.c b/src/components/intel_gpu/linux_intel_gpu_metrics.c
@@ -150,6 +150,22 @@ addMetricToDevice(uint32_t code, int rootDev) {
 	return i;
 }
 
+/*!
+ * @brief  Reset metric counts to zero.
+ */
+void
+metricReset( MetricContext *mContext )
+{
+    for (uint32_t i=0; i<num_avail_devices; i++) {
+        uint32_t dev_idx = mContext->active_devices[i];
+        if  (dev_idx < num_active_devices) {
+            DeviceContext *dev = &active_devices[dev_idx];
+            GPUSetMetricControl(dev->handle, METRIC_RESET);
+        }
+    }
+}
+
+
 
 /************************* PAPI Functions **********************************/
 
@@ -277,12 +293,40 @@ intel_gpu_update_control_state( hwd_control_state_t *ctl,
 	(void)ctl;
 
 	// use local maintained context,
-	if (!count ||!native)  {
+	MetricContext *mContext = (MetricContext *)ctx;
+
+    /* This check accounts for calls to PAPI_cleanup_eventset(). */
+	if ( !count )  {
+	    metricReset(mContext);
+        /* Free devices. */
+        for (uint32_t i=0; i<mContext->num_devices; i++) {
+            uint32_t dev_idx = mContext->active_devices[i];
+            if (dev_idx >= num_active_devices) {
+                return PAPI_ENOMEM;
+            }
+            DeviceContext *dev = &active_devices[dev_idx];
+            DEVICE_HANDLE handle = dev->handle;
+            if( !handle ) {
+                GPUFreeDevice(handle);
+            }
+            mContext->active_devices[i] = 0;
+        }
+        mContext->num_devices = 0;
+        /* Free metric slots. */
+        for (uint32_t midx=0; midx<mContext->num_metrics; midx++) {
+            mContext->metric_idx[midx] = 0;
+            mContext->metric_values[midx] = 0;
+            mContext->dev_ctx_idx[midx] = 0;
+            mContext->subdev_idx[midx] = 0;
+        }
+        mContext->num_metrics = 0;
+        return PAPI_OK;
+	}
+
+	if ( !native ) {
 		return PAPI_OK;
 	}
 
-	MetricContext *mContext = (MetricContext *)ctx;
-
 #if defined(_DEBUG)
 	for (int i=0; i<count; i++) {
 		 GPUDEBUG("\t i=%d, ni_event 0x%x,  ni_papi_code 0x%x, ni_position %d, ni_owners %d   \n", 
@@ -354,6 +398,9 @@ intel_gpu_start( hwd_context_t * ctx, hwd_control_state_t * ctl )
 	MetricContext *mContext = (MetricContext *)ctx;
 
 	int	 ret	= PAPI_OK;
+
+    metricReset(mContext);
+
 	if (mContext->num_metrics == 0) {
 		GPUDEBUG("intel_gpu_start : No metric selected, abort.\n");
 		return PAPI_EINVAL;
@@ -521,13 +568,8 @@ intel_gpu_reset( hwd_context_t *ctx, hwd_control_state_t *ctl)
 	GPUDEBUG("Entering intel_gpu_reset\n");
 	MetricContext *mContext = (MetricContext *)ctx;
 
-	for (uint32_t i=0; i<num_avail_devices; i++) {
-		uint32_t dev_idx = mContext->active_devices[i];
-		if  (dev_idx < num_active_devices) {
-			DeviceContext *dev = &active_devices[dev_idx];
-			GPUSetMetricControl(dev->handle, METRIC_RESET);
-		}
-	}
+	metricReset(mContext);
+
 	return PAPI_OK;
 }