Vulkan: Ignore glFlush to reduce vkQueueSubmits in Asphalt 9

src/libANGLE/renderer/vulkan/ContextVk.cpp

@@ -396,7 +396,8 @@ ContextVk::ContextVk(const gl::State &state, gl::ErrorSet *errorSet, RendererVk
       mRenderPassCommands(nullptr),
       mGpuEventsEnabled(false),
       mSyncObjectPendingFlush(false),
-      mDeferredFlushCount(0),
+      mEGLSyncObjectPendingFlush(false),
+      mHasDeferredFlush(false),
       mGpuClockSync{std::numeric_limits<double>::max(), std::numeric_limits<double>::max()},
       mGpuEventTimestampOrigin(0),
       mPerfCounters{},
@@ -700,14 +701,29 @@ angle::Result ContextVk::initialize()
 
 angle::Result ContextVk::flush(const gl::Context *context)
 {
-    // If we are in middle of renderpass and it is not a shared context, then we will defer the
-    // glFlush call here until the renderpass ends. If sync object has been used, we must respect
-    // glFlush call, otherwise we a wait for sync object without GL_SYNC_FLUSH_COMMANDS_BIT may
-    // never come back.
-    if (mRenderer->getFeatures().deferFlushUntilEndRenderPass.enabled && !context->isShared() &&
-        !mSyncObjectPendingFlush && hasStartedRenderPass())
+    // If a sync object has been used or this is a shared context, then we need to flush the
+    // commands and end the render pass to make sure the sync object (and any preceding commands)
+    // lands in the correct place within the command stream.
+    // EGL sync objects can span across context share groups, so don't defer flushes if there's one
+    // pending a flush.
+    if (mSyncObjectPendingFlush && context->isShared() && !mEGLSyncObjectPendingFlush)
     {
-        mDeferredFlushCount++;
+        // Flush the commands to create a sync point in the command stream.
+        ANGLE_TRY(flushCommandsAndEndRenderPass());
+        // Move the resources to the share group, so they are released during the next vkQueueSubmit
+        // performed by any context in the share group. Note that this relies heavily on the global
+        // mutex to guarantee that no two contexts are modifying the lists at the same time.
+        getShareGroupVk()->acquireResourceUseList(std::move(mResourceUseList));
+        mHasDeferredFlush = true;
+        return angle::Result::Continue;
+    }
+
+    // EGL sync objects can span across context share groups, so don't defer flushes if there's one
+    // pending a flush.
+    if (!mEGLSyncObjectPendingFlush &&
+        mRenderer->getFeatures().deferFlushUntilEndRenderPass.enabled && hasStartedRenderPass())
+    {
+        mHasDeferredFlush = true;
         return angle::Result::Continue;
     }
 
@@ -1569,9 +1585,11 @@ angle::Result ContextVk::submitFrame(const vk::Semaphore *signalSemaphore)
         dumpCommandStreamDiagnostics();
     }
 
-    ANGLE_TRY(mRenderer->submitFrame(
-        this, mContextPriority, std::move(mWaitSemaphores), std::move(mWaitSemaphoreStageMasks),
-        signalSemaphore, std::move(mResourceUseList), std::move(mCurrentGarbage), &mCommandPool));
+    getShareGroupVk()->acquireResourceUseList(std::move(mResourceUseList));
+    ANGLE_TRY(mRenderer->submitFrame(this, mContextPriority, std::move(mWaitSemaphores),
+                                     std::move(mWaitSemaphoreStageMasks), signalSemaphore,
+                                     getShareGroupVk()->releaseResourceUseLists(),
+                                     std::move(mCurrentGarbage), &mCommandPool));
 
     onRenderPassFinished();
     mComputeDirtyBits |= mNewComputeCommandBufferDirtyBits;
@@ -4256,9 +4274,9 @@ angle::Result ContextVk::flushImpl(const vk::Semaphore *signalSemaphore)
 {
     ANGLE_TRACE_EVENT0("gpu.angle", "ContextVk::flushImpl");
 
-    // We must set this to zero before calling flushCommandsAndEndRenderPass to prevent it from
+    // We must set this to false before calling flushCommandsAndEndRenderPass to prevent it from
     // calling back to flushImpl.
-    mDeferredFlushCount     = 0;
+    mHasDeferredFlush       = false;
     mSyncObjectPendingFlush = false;
 
     ANGLE_TRY(flushCommandsAndEndRenderPass());
@@ -4651,7 +4669,6 @@ angle::Result ContextVk::flushCommandsAndEndRenderPass()
 
     if (!mRenderPassCommands->started())
     {
-        ASSERT(!mDeferredFlushCount);
         onRenderPassFinished();
         return angle::Result::Continue;
     }
@@ -4707,7 +4724,7 @@ angle::Result ContextVk::flushCommandsAndEndRenderPass()
         ANGLE_TRY(flushOutsideRenderPassCommands());
     }
 
-    if (mDeferredFlushCount > 0)
+    if (mHasDeferredFlush)
     {
         // If we have deferred glFlush call in the middle of renderpass, flush them now.
         ANGLE_TRY(flushImpl(nullptr));

src/libANGLE/renderer/vulkan/ContextVk.h

@@ -569,6 +569,7 @@ class ContextVk : public ContextImpl, public vk::Context, public MultisampleText
     vk::PerfCounters &getPerfCounters() { return mPerfCounters; }
 
     void onSyncHelperInitialize() { mSyncObjectPendingFlush = true; }
+    void onEGLSyncHelperInitialize() { mEGLSyncObjectPendingFlush = true; }
 
     // When UtilsVk issues a draw call on the currently running render pass, the pipelines and
     // descriptor sets it binds need to be undone.
@@ -1000,7 +1001,8 @@ class ContextVk : public ContextImpl, public vk::Context, public MultisampleText
     // Track SyncHelper object been added into secondary command buffer that has not been flushed to
     // vulkan.
     bool mSyncObjectPendingFlush;
-    uint32_t mDeferredFlushCount;
+    bool mEGLSyncObjectPendingFlush;
+    bool mHasDeferredFlush;
 
     // Semaphores that must be waited on in the next submission.
     std::vector<VkSemaphore> mWaitSemaphores;

src/libANGLE/renderer/vulkan/DisplayVk.h

@@ -12,6 +12,7 @@
 
 #include "common/MemoryBuffer.h"
 #include "libANGLE/renderer/DisplayImpl.h"
+#include "libANGLE/renderer/vulkan/ResourceVk.h"
 #include "libANGLE/renderer/vulkan/vk_cache_utils.h"
 #include "libANGLE/renderer/vulkan/vk_utils.h"
 
@@ -34,6 +35,15 @@ class ShareGroupVk : public ShareGroupImpl
     DescriptorSetLayoutCache &getDescriptorSetLayoutCache() { return mDescriptorSetLayoutCache; }
     ShareContextSet *getShareContextSet() { return &mShareContextSet; }
 
+    std::vector<vk::ResourceUseList> &&releaseResourceUseLists()
+    {
+        return std::move(mResourceUseLists);
+    }
+    void acquireResourceUseList(vk::ResourceUseList &&resourceUseList)
+    {
+        mResourceUseLists.emplace_back(std::move(resourceUseList));
+    }
+
   private:
     // ANGLE uses a PipelineLayout cache to store compatible pipeline layouts.
     PipelineLayoutCache mPipelineLayoutCache;
@@ -43,6 +53,10 @@ class ShareGroupVk : public ShareGroupImpl
 
     // The list of contexts within the share group
     ShareContextSet mShareContextSet;
+
+    // List of resources currently used that need to be freed when any ContextVk in this
+    // ShareGroupVk submits the next command.
+    std::vector<vk::ResourceUseList> mResourceUseLists;
 };
 
 class DisplayVk : public DisplayImpl, public vk::Context

src/libANGLE/renderer/vulkan/FenceNVVk.cpp

@@ -29,7 +29,7 @@ void FenceNVVk::onDestroy(const gl::Context *context)
 angle::Result FenceNVVk::set(const gl::Context *context, GLenum condition)
 {
     ASSERT(condition == GL_ALL_COMPLETED_NV);
-    return mFenceSync.initialize(vk::GetImpl(context));
+    return mFenceSync.initialize(vk::GetImpl(context), false);
 }
 
 angle::Result FenceNVVk::test(const gl::Context *context, GLboolean *outFinished)

src/libANGLE/renderer/vulkan/RendererVk.cpp

@@ -2511,7 +2511,7 @@ angle::Result RendererVk::submitFrame(vk::Context *context,
                                       std::vector<VkSemaphore> &&waitSemaphores,
                                       std::vector<VkPipelineStageFlags> &&waitSemaphoreStageMasks,
                                       const vk::Semaphore *signalSemaphore,
-                                      vk::ResourceUseList &&resourceUseList,
+                                      std::vector<vk::ResourceUseList> &&resourceUseLists,
                                       vk::GarbageList &&currentGarbage,
                                       vk::CommandPool *commandPool)
 {
@@ -2538,7 +2538,10 @@ angle::Result RendererVk::submitFrame(vk::Context *context,
 
     waitSemaphores.clear();
     waitSemaphoreStageMasks.clear();
-    resourceUseList.releaseResourceUsesAndUpdateSerials(submitQueueSerial);
+    for (vk::ResourceUseList &it : resourceUseLists)
+    {
+        it.releaseResourceUsesAndUpdateSerials(submitQueueSerial);
+    }
 
     return angle::Result::Continue;
 }

src/libANGLE/renderer/vulkan/RendererVk.h

@@ -332,7 +332,7 @@ class RendererVk : angle::NonCopyable
                               std::vector<VkSemaphore> &&waitSemaphores,
                               std::vector<VkPipelineStageFlags> &&waitSemaphoreStageMasks,
                               const vk::Semaphore *signalSemaphore,
-                              vk::ResourceUseList &&resourceUseList,
+                              std::vector<vk::ResourceUseList> &&resourceUseLists,
                               vk::GarbageList &&currentGarbage,
                               vk::CommandPool *commandPool);
 

src/libANGLE/renderer/vulkan/SyncVk.cpp

@@ -83,7 +83,7 @@ void SyncHelper::releaseToRenderer(RendererVk *renderer)
     renderer->collectGarbageAndReinit(&mUse, &mEvent);
 }
 
-angle::Result SyncHelper::initialize(ContextVk *contextVk)
+angle::Result SyncHelper::initialize(ContextVk *contextVk, bool isEglSyncObject)
 {
     ASSERT(!mEvent.valid());
 
@@ -108,7 +108,14 @@ angle::Result SyncHelper::initialize(ContextVk *contextVk)
     commandBuffer->setEvent(mEvent.getHandle(), VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT);
     retain(&contextVk->getResourceUseList());
 
-    contextVk->onSyncHelperInitialize();
+    if (isEglSyncObject)
+    {
+        contextVk->onEGLSyncHelperInitialize();
+    }
+    else
+    {
+        contextVk->onSyncHelperInitialize();
+    }
 
     return angle::Result::Continue;
 }
@@ -137,19 +144,13 @@ angle::Result SyncHelper::clientWait(Context *context,
         return angle::Result::Continue;
     }
 
-    if (flushCommands && contextVk)
+    // We defer (ignore) flushes, so it's possible that the glFence's signal operation is pending
+    // submission.
+    if ((flushCommands && contextVk) || usedInRecordedCommands())
     {
         ANGLE_TRY(contextVk->flushImpl(nullptr));
     }
 
-    // Undefined behaviour. Early exit.
-    if (usedInRecordedCommands())
-    {
-        WARN() << "Waiting on a sync that is not flushed";
-        *outResult = VK_TIMEOUT;
-        return angle::Result::Continue;
-    }
-
     ASSERT(mUse.getSerial().valid());
 
     VkResult status = VK_SUCCESS;
@@ -401,7 +402,7 @@ angle::Result SyncVk::set(const gl::Context *context, GLenum condition, GLbitfie
     ASSERT(condition == GL_SYNC_GPU_COMMANDS_COMPLETE);
     ASSERT(flags == 0);
 
-    return mSyncHelper.initialize(vk::GetImpl(context));
+    return mSyncHelper.initialize(vk::GetImpl(context), false);
 }
 
 angle::Result SyncVk::clientWait(const gl::Context *context,
@@ -484,7 +485,7 @@ egl::Error EGLSyncVk::initialize(const egl::Display *display,
         case EGL_SYNC_FENCE_KHR:
             ASSERT(mAttribs.isEmpty());
             mSyncHelper = new vk::SyncHelper();
-            if (mSyncHelper->initialize(vk::GetImpl(context)) == angle::Result::Stop)
+            if (mSyncHelper->initialize(vk::GetImpl(context), true) == angle::Result::Stop)
             {
                 return egl::Error(EGL_BAD_ALLOC, "eglCreateSyncKHR failed to create sync object");
             }

src/libANGLE/renderer/vulkan/SyncVk.h

@@ -38,7 +38,7 @@ class SyncHelper : public vk::Resource
 
     virtual void releaseToRenderer(RendererVk *renderer);
 
-    virtual angle::Result initialize(ContextVk *contextVk);
+    virtual angle::Result initialize(ContextVk *contextVk, bool isEglSyncObject);
     virtual angle::Result clientWait(Context *context,
                                      ContextVk *contextVk,
                                      bool flushCommands,

src/tests/egl_tests/EGLContextSharingTest.cpp

@@ -13,6 +13,7 @@
 #include "test_utils/gl_raii.h"
 #include "util/EGLWindow.h"
 
+#include <atomic>
 #include <condition_variable>
 #include <mutex>
 #include <thread>
@@ -360,6 +361,8 @@ TEST_P(EGLContextSharingTest, SamplerLifetime)
 TEST_P(EGLContextSharingTest, DeleteReaderOfSharedTexture)
 {
     ANGLE_SKIP_TEST_IF(!platformSupportsMultithreading());
+    // GL Fences require GLES 3.0+
+    ANGLE_SKIP_TEST_IF(getClientMajorVersion() < 3);
 
     // Initialize contexts
     EGLWindow *window = getEGLWindow();
@@ -423,6 +426,8 @@ TEST_P(EGLContextSharingTest, DeleteReaderOfSharedTexture)
     // Synchronization tools to ensure the two threads are interleaved as designed by this test.
     std::mutex mutex;
     std::condition_variable condVar;
+    std::atomic<GLsync> deletingThreadSyncObj;
+    std::atomic<GLsync> continuingThreadSyncObj;
 
     enum class Step
     {
@@ -498,16 +503,25 @@ TEST_P(EGLContextSharingTest, DeleteReaderOfSharedTexture)
         // Draw using the shared texture.
         drawQuad(program[0].get(), essl1_shaders::PositionAttrib(), 0.5f);
 
+        deletingThreadSyncObj = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+        ASSERT_GL_NO_ERROR();
+        // Force the fence to be created
+        glFlush();
+
         // Wait for the other thread to also draw using the shared texture.
         nextStep(Step::Thread0Draw);
         ASSERT_TRUE(waitForStep(Step::Thread1Draw));
 
+        ASSERT_TRUE(continuingThreadSyncObj != nullptr);
+        glWaitSync(continuingThreadSyncObj, 0, GL_TIMEOUT_IGNORED);
+        ASSERT_GL_NO_ERROR();
+        glDeleteSync(continuingThreadSyncObj);
+        ASSERT_GL_NO_ERROR();
+        continuingThreadSyncObj = nullptr;
+
         // Delete this thread's framebuffer (reader of the shared texture).
         fbo[0].reset();
 
-        // Flush to make sure the graph nodes associated with this context are deleted.
-        glFlush();
-
         // Wait for the other thread to use the shared texture again before unbinding the
         // context (so no implicit flush happens).
         nextStep(Step::Thread0Delete);
@@ -527,9 +541,21 @@ TEST_P(EGLContextSharingTest, DeleteReaderOfSharedTexture)
         // Wait for first thread to draw using the shared texture.
         ASSERT_TRUE(waitForStep(Step::Thread0Draw));
 
+        ASSERT_TRUE(deletingThreadSyncObj != nullptr);
+        glWaitSync(deletingThreadSyncObj, 0, GL_TIMEOUT_IGNORED);
+        ASSERT_GL_NO_ERROR();
+        glDeleteSync(deletingThreadSyncObj);
+        ASSERT_GL_NO_ERROR();
+        deletingThreadSyncObj = nullptr;
+
         // Draw using the shared texture.
         drawQuad(program[0].get(), essl1_shaders::PositionAttrib(), 0.5f);
 
+        continuingThreadSyncObj = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+        ASSERT_GL_NO_ERROR();
+        // Force the fence to be created
+        glFlush();
+
         // Wait for the other thread to delete its framebuffer.
         nextStep(Step::Thread1Draw);
         ASSERT_TRUE(waitForStep(Step::Thread0Delete));

src/tests/egl_tests/EGLSyncTest.cpp

@@ -214,17 +214,13 @@ TEST_P(EGLSyncTest, BasicOperations)
 
     glFlush();
 
-    EGLint value           = 0;
-    unsigned int loopCount = 0;
-
-    // Use 'loopCount' to make sure the test doesn't get stuck in an infinite loop
-    while (value != EGL_SIGNALED_KHR && loopCount <= 1000000)
-    {
-        loopCount++;
-        EXPECT_EGL_TRUE(eglGetSyncAttribKHR(display, sync, EGL_SYNC_STATUS_KHR, &value));
-    }
+    glClear(GL_COLOR_BUFFER_BIT);
 
-    ASSERT_EQ(value, EGL_SIGNALED_KHR);
+    // Don't wait forever to make sure the test terminates
+    constexpr GLuint64 kTimeout = 1'000'000'000;  // 1 second
+    EGLint value                = 0;
+    ASSERT_EQ(EGL_CONDITION_SATISFIED_KHR,
+              eglClientWaitSyncKHR(display, sync, EGL_SYNC_FLUSH_COMMANDS_BIT_KHR, kTimeout));
 
     for (size_t i = 0; i < 20; i++)
     {

src/tests/gl_tests/FenceSyncTests.cpp

@@ -230,6 +230,9 @@ TEST_P(FenceSyncTest, BasicQueries)
 // Test that basic usage works and doesn't generate errors or crash
 TEST_P(FenceSyncTest, BasicOperations)
 {
+    // TODO(http://anglebug.com/5425): glClientWaitSync() returns GL_TIMEOUT_EXPIRED
+    ANGLE_SKIP_TEST_IF(IsAMD() && IsD3D11());
+
     glClearColor(1.0f, 0.0f, 1.0f, 1.0f);
 
     GLsync sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
@@ -238,28 +241,17 @@ TEST_P(FenceSyncTest, BasicOperations)
     glWaitSync(sync, 0, GL_TIMEOUT_IGNORED);
     EXPECT_GL_NO_ERROR();
 
-    GLsizei length         = 0;
-    GLint value            = 0;
-    unsigned int loopCount = 0;
-
     glFlush();
 
-    // Use 'loopCount' to make sure the test doesn't get stuck in an infinite loop
-    while (value != GL_SIGNALED && loopCount <= 1000000)
-    {
-        loopCount++;
-
-        glGetSynciv(sync, GL_SYNC_STATUS, 1, &length, &value);
-        ASSERT_GL_NO_ERROR();
-    }
-
-    ASSERT_GLENUM_EQ(GL_SIGNALED, value);
-
+    // Don't wait forever to make sure the test terminates
+    constexpr GLuint64 kTimeout = 1'000'000'000;  // 1 second
+    GLint value                 = 0;
     for (size_t i = 0; i < 20; i++)
     {
         glClear(GL_COLOR_BUFFER_BIT);
-        glClientWaitSync(sync, GL_SYNC_FLUSH_COMMANDS_BIT, GL_TIMEOUT_IGNORED);
+        value = glClientWaitSync(sync, GL_SYNC_FLUSH_COMMANDS_BIT, kTimeout);
         EXPECT_GL_NO_ERROR();
+        ASSERT_TRUE(value == GL_CONDITION_SATISFIED || value == GL_ALREADY_SIGNALED);
     }
 }