Reduce the amount of thread switches needed for fetching textures

Combine multiple texture fetches into a single sync call on another
thread.

Change-Id: Id372d5a11290105b36b34a4fc56bff9929304c20
Reviewed-by: Jüri Valdmann <juri.valdmann@qt.io>
Reviewed-by: Alexandru Croitor <alexandru.croitor@qt.io>

2 files changed, 44 insertions, 7 deletions

diff --git a/src/core/delegated_frame_node.cpp b/src/core/delegated_frame_node.cpp
index d6ee87950..78241a431 100644
--- a/src/core/delegated_frame_node.cpp
+++ b/src/core/delegated_frame_node.cpp
@@ -1263,6 +1263,8 @@ void DelegatedFrameNode::fetchAndSyncMailboxes(QList<MailboxTexture *> &mailboxe
     QList<gl::TransferableFence> transferredFences;
     {
         QMutexLocker lock(&m_mutex);
+        QVector<MailboxTexture *> mailboxesToPull;
+        mailboxesToPull.reserve(mailboxesToFetch.size());
 
         gpu::SyncPointManager *syncPointManager = sync_point_manager();
         base::MessageLoop *gpuMessageLoop = gpu_message_loop();
@@ -1271,9 +1273,12 @@ void DelegatedFrameNode::fetchAndSyncMailboxes(QList<MailboxTexture *> &mailboxe
         for (MailboxTexture *mailboxTexture : qAsConst(mailboxesToFetch)) {
             gpu::SyncToken &syncToken = mailboxTexture->mailboxHolder().sync_token;
             const auto task = base::Bind(&DelegatedFrameNode::pullTexture, this, mailboxTexture);
-            if (syncPointManager->WaitOutOfOrderNonThreadSafe(syncToken, gpuMessageLoop->task_runner(), task))
-                continue;
-            gpuMessageLoop->task_runner()->PostTask(FROM_HERE, task);
+            if (!syncPointManager->WaitOutOfOrderNonThreadSafe(syncToken, gpuMessageLoop->task_runner(), std::move(task)))
+                mailboxesToPull.append(mailboxTexture);
+        }
+        if (!mailboxesToPull.isEmpty()) {
+            auto task = base::BindOnce(&DelegatedFrameNode::pullTextures, this, std::move(mailboxesToPull));
+            gpuMessageLoop->task_runner()->PostTask(FROM_HERE, std::move(task));
         }
 
         m_mailboxesFetchedWaitCond.wait(&m_mutex);
@@ -1349,6 +1354,25 @@ void DelegatedFrameNode::fetchAndSyncMailboxes(QList<MailboxTexture *> &mailboxe
 }
 
 
+void DelegatedFrameNode::pullTextures(DelegatedFrameNode *frameNode, const QVector<MailboxTexture *> textures)
+{
+#ifndef QT_NO_OPENGL
+    gpu::gles2::MailboxManager *mailboxManager = mailbox_manager();
+    for (MailboxTexture *texture : textures) {
+        gpu::SyncToken &syncToken = texture->mailboxHolder().sync_token;
+        if (syncToken.HasData())
+            mailboxManager->PullTextureUpdates(syncToken);
+        texture->fetchTexture(mailboxManager);
+        --frameNode->m_numPendingSyncPoints;
+    }
+
+    fenceAndUnlockQt(frameNode);
+#else
+    Q_UNUSED(frameNode)
+    Q_UNUSED(textures)
+#endif
+}
+
 void DelegatedFrameNode::pullTexture(DelegatedFrameNode *frameNode, MailboxTexture *texture)
 {
 #ifndef QT_NO_OPENGL
@@ -1357,6 +1381,18 @@ void DelegatedFrameNode::pullTexture(DelegatedFrameNode *frameNode, MailboxTextu
     if (syncToken.HasData())
         mailboxManager->PullTextureUpdates(syncToken);
     texture->fetchTexture(mailboxManager);
+    --frameNode->m_numPendingSyncPoints;
+
+    fenceAndUnlockQt(frameNode);
+#else
+    Q_UNUSED(frameNode)
+    Q_UNUSED(texture)
+#endif
+}
+
+void DelegatedFrameNode::fenceAndUnlockQt(DelegatedFrameNode *frameNode)
+{
+#ifndef QT_NO_OPENGL
     if (!!gl::GLContext::GetCurrent() && gl::GLFence::IsSupported()) {
         // Create a fence on the Chromium GPU-thread and context
         gl::GLFence *fence = gl::GLFence::Create();
@@ -1364,15 +1400,14 @@ void DelegatedFrameNode::pullTexture(DelegatedFrameNode *frameNode, MailboxTextu
         frameNode->m_textureFences.append(fence->Transfer());
         delete fence;
     }
-    if (--frameNode->m_numPendingSyncPoints == 0)
-        base::ThreadTaskRunnerHandle::Get()->PostTask(FROM_HERE, base::Bind(&DelegatedFrameNode::fenceAndUnlockQt, frameNode));
+    if (frameNode->m_numPendingSyncPoints == 0)
+        base::ThreadTaskRunnerHandle::Get()->PostTask(FROM_HERE, base::Bind(&DelegatedFrameNode::unlockQt, frameNode));
 #else
     Q_UNUSED(frameNode)
-    Q_UNUSED(texture)
 #endif
 }
 
-void DelegatedFrameNode::fenceAndUnlockQt(DelegatedFrameNode *frameNode)
+void DelegatedFrameNode::unlockQt(DelegatedFrameNode *frameNode)
 {
     QMutexLocker lock(&frameNode->m_mutex);
     // Signal preprocess() the textures are ready
diff --git a/src/core/delegated_frame_node.h b/src/core/delegated_frame_node.h
index c627cdf95..4bddf4a62 100644
--- a/src/core/delegated_frame_node.h
+++ b/src/core/delegated_frame_node.h
@@ -123,7 +123,9 @@ private:
     // Making those callbacks static bypasses base::Bind's ref-counting requirement
     // of the this pointer when the callback is a method.
     static void pullTexture(DelegatedFrameNode *frameNode, MailboxTexture *mailbox);
+    static void pullTextures(DelegatedFrameNode *frameNode, const QVector<MailboxTexture *> mailboxes);
     static void fenceAndUnlockQt(DelegatedFrameNode *frameNode);
+    static void unlockQt(DelegatedFrameNode *frameNode);
 
     ResourceHolder *findAndHoldResource(unsigned resourceId, QHash<unsigned, QSharedPointer<ResourceHolder> > &candidates);
     void holdResources(const cc::DrawQuad *quad, QHash<unsigned, QSharedPointer<ResourceHolder> > &candidates);