Profile layer: Add frame serialization config option

Author: solidpixel

layer_gpu_profile/README_LAYER.md

@@ -140,6 +140,14 @@ to profile using the `sample_mode` config option:
 * `workload`: Sample every workload in each frame of interest.
 * `frame`: Sample at the end of each frame of interest.
 
+By default per-frame samples are isolated from other frames by inserting a
+`vkDeviceWaitIdle()` before and after the frame to ensure that workload
+in the sampled region does not overlap neighboring frames. Setting the
+`frame_serialization` config option to `false` will allow frames to overlap
+without serialization, but can add noise to the returned counter values. This
+option has no effect for per-workload sampling, which must always use
+serialization.
+
 ## Layer counters
 
 The current layer uses a hard-coded set of performance counters defined in the

layer_gpu_profile/layer_config.json

@@ -4,5 +4,6 @@
     "sample_mode": "frame",
     "periodic_min_frame": 1,
     "periodic_frame": 600,
-    "frame_list": []
+    "frame_list": [],
+    "frame_serialization": false
 }

layer_gpu_profile/source/layer_config.cpp

@@ -90,6 +90,9 @@ void LayerConfig::parseSamplingOptions(const json& config)
         rawSampleMode = "disabled";
     }
 
+    // Decode frame serialization mode
+    frameSerialization = config.at("frame_serialization");
+
     LAYER_LOG("Layer sampling configuration");
     LAYER_LOG("============================");
     LAYER_LOG(" - Frame selection mode: %s", rawFrameMode.c_str());
@@ -107,6 +110,11 @@ void LayerConfig::parseSamplingOptions(const json& config)
     }
 
     LAYER_LOG(" - Counter sampling mode: %s", rawSampleMode.c_str());
+
+    if (samplingMode == COUNTER_SAMPLING_FRAMES)
+    {
+        LAYER_LOG(" - Frame serialization: %u", frameSerialization);
+    }
 }
 
 /* See header for documentation. */
@@ -172,20 +180,29 @@ bool LayerConfig::isFrameOfInterest(
 }
 
 /* See header for documentation. */
-bool LayerConfig::isSamplingWorkloads() const {
+bool LayerConfig::isSamplingWorkloads() const
+{
     return frameMode != FRAME_SELECTION_DISABLED &&
            samplingMode == COUNTER_SAMPLING_WORKLOADS;
 }
 
 /* See header for documentation. */
-bool LayerConfig::isSamplingFrames() const {
+bool LayerConfig::isSamplingFrames() const
+{
     return frameMode != FRAME_SELECTION_DISABLED &&
            samplingMode == COUNTER_SAMPLING_FRAMES;
 }
 
 /* See header for documentation. */
-bool LayerConfig::isSamplingAny() const {
+bool LayerConfig::isSamplingAny() const
+{
     return frameMode != FRAME_SELECTION_DISABLED &&
            samplingMode != COUNTER_SAMPLING_DISABLED;
 }
 
+/* See header for documentation. */
+bool LayerConfig::isSerializingFrames() const
+{
+    return isSamplingWorkloads() ||
+           (isSamplingFrames() && frameSerialization);
+};

layer_gpu_profile/source/layer_config.hpp

@@ -79,6 +79,13 @@ class LayerConfig
      */
     bool isSamplingAny() const;
 
+    /**
+     * @brief Test if we are serializing frames.
+     *
+     * @return @c true if serializing, @c false otherwise.
+     */
+    bool isSerializingFrames() const;
+
 private:
     /**
      * @brief Supported frame selection modes.
@@ -119,6 +126,10 @@ class LayerConfig
      */
     CounterSamplingMode samplingMode {COUNTER_SAMPLING_DISABLED};
 
+    /**
+     * @brief The frame sample serialization mode.
+     */
+    bool frameSerialization {true};
 
     /**
      * @brief The sampling period in frames, or 0 if disabled.

layer_gpu_profile/source/layer_device_functions_queue.cpp

@@ -140,10 +140,16 @@ static void processFrameBoundaryPostSubmit(
     Device& layer,
     bool frameSample
 ) {
+    const auto& config = layer.instance->config;
+
     // If we are measuring performance ensure the previous frame has finished
     // and then take an initial sample to reset the counters
-    layer.driver.vkDeviceWaitIdle(layer.device);
-    workaroundDelay();
+    if (config.isSerializingFrames())
+    {
+        layer.driver.vkDeviceWaitIdle(layer.device);
+        workaroundDelay();
+    }
+
     auto ec = layer.lgcSampler->sample_now();
     if (ec)
     {