sf: reuse luma sampling buffer when available

Reuse the luma sampling buffer when available, saving a frequent buffer
reallocation. Depending on driver refcounting behavior, this could have
resulted in the buffer free on the SF main thread (and typically did on
referenced device, costing about ~500us @10hz).

Test: verify luma sampling work reduces from ~2ms -> 1.5ms
Test: RegionSamplingTest in libgui
Fixes: 131416627

Change-Id: I8e6d57ae25bd37ceec828c82796f0f4f8f45636e

2 files changed, 18 insertions, 5 deletions

diff --git a/services/surfaceflinger/RegionSamplingThread.cpp b/services/surfaceflinger/RegionSamplingThread.cpp
index 252ff0d611..0d142675db 100644
--- a/services/surfaceflinger/RegionSamplingThread.cpp
+++ b/services/surfaceflinger/RegionSamplingThread.cpp
@@ -377,10 +377,15 @@ void RegionSamplingThread::captureSample() {
         mFlinger.traverseLayersInDisplay(device, filterVisitor);
     };
 
-    const uint32_t usage = GRALLOC_USAGE_SW_READ_OFTEN | GRALLOC_USAGE_HW_RENDER;
-    sp<GraphicBuffer> buffer =
-            new GraphicBuffer(sampledArea.getWidth(), sampledArea.getHeight(),
-                              PIXEL_FORMAT_RGBA_8888, 1, usage, "RegionSamplingThread");
+    sp<GraphicBuffer> buffer = nullptr;
+    if (mCachedBuffer && mCachedBuffer->getWidth() == sampledArea.getWidth() &&
+        mCachedBuffer->getHeight() == sampledArea.getHeight()) {
+        buffer = mCachedBuffer;
+    } else {
+        const uint32_t usage = GRALLOC_USAGE_SW_READ_OFTEN | GRALLOC_USAGE_HW_RENDER;
+        buffer = new GraphicBuffer(sampledArea.getWidth(), sampledArea.getHeight(),
+                                   PIXEL_FORMAT_RGBA_8888, 1, usage, "RegionSamplingThread");
+    }
 
     // When calling into SF, we post a message into the SF message queue (so the
     // screen capture runs on the main thread). This message blocks until the
@@ -415,6 +420,12 @@ void RegionSamplingThread::captureSample() {
     for (size_t d = 0; d < activeDescriptors.size(); ++d) {
         activeDescriptors[d].listener->onSampleCollected(lumas[d]);
     }
+
+    // Extend the lifetime of mCachedBuffer from the previous frame to here to ensure that:
+    // 1) The region sampling thread is the last owner of the buffer, and the freeing of the buffer
+    // happens in this thread, as opposed to the main thread.
+    // 2) The listener(s) receive their notifications prior to freeing the buffer.
+    mCachedBuffer = buffer;
     ATRACE_INT(lumaSamplingStepTag, static_cast<int>(samplingStep::noWorkNeeded));
 }
 
diff --git a/services/surfaceflinger/RegionSamplingThread.h b/services/surfaceflinger/RegionSamplingThread.h
index 979642912b..72b20420ef 100644
--- a/services/surfaceflinger/RegionSamplingThread.h
+++ b/services/surfaceflinger/RegionSamplingThread.h
@@ -24,13 +24,13 @@
 
 #include <android-base/thread_annotations.h>
 #include <binder/IBinder.h>
+#include <ui/GraphicBuffer.h>
 #include <ui/Rect.h>
 #include <utils/StrongPointer.h>
 #include "Scheduler/IdleTimer.h"
 
 namespace android {
 
-class GraphicBuffer;
 class IRegionSamplingListener;
 class Layer;
 class Scheduler;
@@ -121,6 +121,8 @@ private:
     std::unordered_map<wp<IBinder>, Descriptor, WpHash> mDescriptors GUARDED_BY(mMutex);
     std::chrono::nanoseconds lastSampleTime GUARDED_BY(mMutex);
     bool mDiscardedFrames GUARDED_BY(mMutex) = false;
+
+    sp<GraphicBuffer> mCachedBuffer GUARDED_BY(mMutex) = nullptr;
 };
 
 } // namespace android