ParallelCompile: use the native extensions

scripts/run_code_generation_hashes.json

@@ -170,17 +170,17 @@
   "OpenGL dispatch table:src/libANGLE/renderer/angle_format.py":
     "b18ca0fe4835114a4a2f54977b19e798",
   "OpenGL dispatch table:src/libANGLE/renderer/gl/DispatchTableGL_autogen.cpp":
-    "6556e48f03112aaf0e4d0fa2949727b4",
+    "2ed6500b69d5db151348ddccf6d9b2b8",
   "OpenGL dispatch table:src/libANGLE/renderer/gl/DispatchTableGL_autogen.h":
-    "caea949141a8c6b8692f1c021fb4fa42",
+    "4211cf4afa7d4a0f8a95de39db45a55e",
   "OpenGL dispatch table:src/libANGLE/renderer/gl/generate_gl_dispatch_table.py":
-    "86a66ba63f6dceac553d8af6c132b6fb",
+    "7571edb9e610891ed0c95dc496120cff",
   "OpenGL dispatch table:src/libANGLE/renderer/gl/gl_bindings_data.json":
-    "71079f089335ce1f67835d67a6d49d1a",
+    "6cd07767ecf12d24d2d3572194c3323b",
   "OpenGL dispatch table:src/libANGLE/renderer/gl/null_functions.cpp":
-    "774c21cf434656bb40735cc67fb4fa40",
+    "f10d2f520f2ceae32ad8e65322fdb0a5",
   "OpenGL dispatch table:src/libANGLE/renderer/gl/null_functions.h":
-    "594f92ec5ffaaf21409579009f579cd7",
+    "638724bfa3c1e926e93fb97a84b7617f",
   "Vulkan format:src/libANGLE/renderer/angle_format.py":
     "b18ca0fe4835114a4a2f54977b19e798",
   "Vulkan format:src/libANGLE/renderer/angle_format_map.json":

src/libANGLE/Context.cpp

@@ -7954,6 +7954,7 @@ void Context::maxShaderCompilerThreads(GLuint count)
         mThreadPool = angle::WorkerThreadPool::Create(count > 0);
     }
     mThreadPool->setMaxThreads(count);
+    mImplementation->setMaxShaderCompilerThreads(count);
 }
 
 bool Context::isGLES1() const

src/libANGLE/Shader.cpp

@@ -20,7 +20,6 @@
 #include "libANGLE/Constants.h"
 #include "libANGLE/Context.h"
 #include "libANGLE/ResourceManager.h"
-#include "libANGLE/WorkerThread.h"
 #include "libANGLE/renderer/GLImplFactory.h"
 #include "libANGLE/renderer/ShaderImpl.h"
 
@@ -108,47 +107,10 @@ class ScopedExit final : angle::NonCopyable
     std::function<void()> mExit;
 };
 
-using CompileImplFunctor = std::function<void(const std::string &, std::string &)>;
-class CompileTask : public angle::Closure
+struct Shader::CompilingState
 {
-  public:
-    CompileTask(ShHandle handle,
-                std::string &&sourcePath,
-                std::string &&source,
-                ShCompileOptions options,
-                CompileImplFunctor &&functor)
-        : mHandle(handle),
-          mSourcePath(sourcePath),
-          mSource(source),
-          mOptions(options),
-          mCompileImplFunctor(functor),
-          mResult(false)
-    {}
-    void operator()() override
-    {
-        std::vector<const char *> srcStrings;
-        if (!mSourcePath.empty())
-        {
-            srcStrings.push_back(mSourcePath.c_str());
-        }
-        srcStrings.push_back(mSource.c_str());
-        mResult = sh::Compile(mHandle, &srcStrings[0], srcStrings.size(), mOptions);
-        if (mResult)
-        {
-            mCompileImplFunctor(sh::GetObjectCode(mHandle), mInfoLog);
-        }
-    }
-    bool getResult() { return mResult; }
-    const std::string &getInfoLog() { return mInfoLog; }
-
-  private:
-    ShHandle mHandle;
-    std::string mSourcePath;
-    std::string mSource;
-    ShCompileOptions mOptions;
-    CompileImplFunctor mCompileImplFunctor;
-    bool mResult;
-    std::string mInfoLog;
+    std::shared_ptr<rx::WaitableCompileEvent> compileEvent;
+    ShCompilerInstance shCompilerInstance;
 };
 
 ShaderState::ShaderState(ShaderType shaderType)
@@ -359,17 +321,10 @@ void Shader::compile(const Context *context)
     mState.mCompileStatus = CompileStatus::COMPILE_REQUESTED;
     mBoundCompiler.set(context, context->getCompiler());
 
-    // Cache the compile source and options for compilation. Must be done now, since the source
-    // can change before the link call or another call that resolves the compile.
-
-    std::stringstream sourceStream;
-    std::string sourcePath;
-    ShCompileOptions options =
-        mImplementation->prepareSourceAndReturnOptions(context, &sourceStream, &sourcePath);
-    options |= (SH_OBJECT_CODE | SH_VARIABLES | SH_EMULATE_GL_DRAW_ID);
-    auto source = sourceStream.str();
+    ShCompileOptions options = (SH_OBJECT_CODE | SH_VARIABLES | SH_EMULATE_GL_DRAW_ID);
 
-    // Add default options to WebGL shaders to prevent unexpected behavior during compilation.
+    // Add default options to WebGL shaders to prevent unexpected behavior during
+    // compilation.
     if (context->getExtensions().webglCompatibility)
     {
         options |= SH_INIT_GL_POSITION;
@@ -379,9 +334,9 @@ void Shader::compile(const Context *context)
         options |= SH_INIT_SHARED_VARIABLES;
     }
 
-    // Some targets (eg D3D11 Feature Level 9_3 and below) do not support non-constant loop indexes
-    // in fragment shaders. Shader compilation will fail. To provide a better error message we can
-    // instruct the compiler to pre-validate.
+    // Some targets (eg D3D11 Feature Level 9_3 and below) do not support non-constant loop
+    // indexes in fragment shaders. Shader compilation will fail. To provide a better error
+    // message we can instruct the compiler to pre-validate.
     if (mRendererLimitations.shadersRequireIndexedLoopValidation)
     {
         options |= SH_VALIDATE_LOOP_INDEXING;
@@ -390,27 +345,15 @@ void Shader::compile(const Context *context)
     mCurrentMaxComputeWorkGroupInvocations = context->getCaps().maxComputeWorkGroupInvocations;
 
     ASSERT(mBoundCompiler.get());
-    mShCompilerInstance     = mBoundCompiler->getInstance(mState.mShaderType);
-    ShHandle compilerHandle = mShCompilerInstance.getHandle();
+    ShCompilerInstance compilerInstance = mBoundCompiler->getInstance(mState.mShaderType);
+    ShHandle compilerHandle             = compilerInstance.getHandle();
     ASSERT(compilerHandle);
-    mCompilerResourcesString = mShCompilerInstance.getBuiltinResourcesString();
+    mCompilerResourcesString = compilerInstance.getBuiltinResourcesString();
 
-    mWorkerPool   = context->getWorkerThreadPool();
-    std::function<void(const std::string &, std::string &)> compileImplFunctor;
-    if (mWorkerPool->isAsync())
-    {
-        compileImplFunctor = [this](const std::string &source, std::string &infoLog) {
-            mImplementation->compileAsync(source, infoLog);
-        };
-    }
-    else
-    {
-        compileImplFunctor = [](const std::string &source, std::string &infoLog) {};
-    }
-    mCompileTask =
-        std::make_shared<CompileTask>(compilerHandle, std::move(sourcePath), std::move(source),
-                                      options, std::move(compileImplFunctor));
-    mCompileEvent = mWorkerPool->postWorkerTask(mCompileTask);
+    mCompilingState.reset(new CompilingState());
+    mCompilingState->shCompilerInstance = std::move(compilerInstance);
+    mCompilingState->compileEvent =
+        mImplementation->compile(context, &(mCompilingState->shCompilerInstance), options);
 }
 
 void Shader::resolveCompile()
@@ -420,22 +363,20 @@ void Shader::resolveCompile()
         return;
     }
 
-    ASSERT(mCompileEvent.get());
-    ASSERT(mCompileTask.get());
-
-    mCompileEvent->wait();
+    ASSERT(mCompilingState.get());
 
-    mCompileEvent.reset();
-    mWorkerPool.reset();
+    mCompilingState->compileEvent->wait();
 
-    bool compiled = mCompileTask->getResult();
-    mInfoLog += mCompileTask->getInfoLog();
-    mCompileTask.reset();
+    mInfoLog += mCompilingState->compileEvent->getInfoLog();
 
-    ScopedExit exit([this]() { mBoundCompiler->putInstance(std::move(mShCompilerInstance)); });
+    ScopedExit exit([this]() {
+        mBoundCompiler->putInstance(std::move(mCompilingState->shCompilerInstance));
+        mCompilingState->compileEvent.reset();
+        mCompilingState.reset();
+    });
 
-    ShHandle compilerHandle = mShCompilerInstance.getHandle();
-    if (!compiled)
+    ShHandle compilerHandle = mCompilingState->shCompilerInstance.getHandle();
+    if (!mCompilingState->compileEvent->getResult())
     {
         mInfoLog += sh::GetInfoLog(compilerHandle);
         WARN() << std::endl << mInfoLog;
@@ -552,7 +493,7 @@ void Shader::resolveCompile()
 
     ASSERT(!mState.mTranslatedSource.empty());
 
-    bool success          = mImplementation->postTranslateCompile(&mShCompilerInstance, &mInfoLog);
+    bool success          = mCompilingState->compileEvent->postTranslate(&mInfoLog);
     mState.mCompileStatus = success ? CompileStatus::COMPILED : CompileStatus::NOT_COMPILED;
 }
 
@@ -594,7 +535,7 @@ bool Shader::isCompiled()
 
 bool Shader::isCompleted()
 {
-    return (!mState.compilePending() || mCompileEvent->isReady());
+    return (!mState.compilePending() || mCompilingState->compileEvent->isReady());
 }
 
 int Shader::getShaderVersion()

src/libANGLE/Shader.h

@@ -31,6 +31,7 @@ namespace rx
 class GLImplFactory;
 class ShaderImpl;
 class ShaderSh;
+class WaitableCompileEvent;
 }  // namespace rx
 
 namespace angle
@@ -189,6 +190,8 @@ class Shader final : angle::NonCopyable, public LabeledObject
     const std::string &getCompilerResourcesString() const;
 
   private:
+    struct CompilingState;
+
     ~Shader() override;
     static void GetSourceImpl(const std::string &source,
                               GLsizei bufSize,
@@ -208,10 +211,7 @@ class Shader final : angle::NonCopyable, public LabeledObject
 
     // We keep a reference to the translator in order to defer compiles while preserving settings.
     BindingPointer<Compiler> mBoundCompiler;
-    ShCompilerInstance mShCompilerInstance;
-    std::shared_ptr<CompileTask> mCompileTask;
-    std::shared_ptr<angle::WorkerThreadPool> mWorkerPool;
-    std::shared_ptr<angle::WaitableEvent> mCompileEvent;
+    std::unique_ptr<CompilingState> mCompilingState;
     std::string mCompilerResourcesString;
 
     ShaderProgramManager *mResourceManager;

src/libANGLE/WorkerThread.cpp

@@ -24,6 +24,13 @@ namespace angle
 WaitableEvent::WaitableEvent()  = default;
 WaitableEvent::~WaitableEvent() = default;
 
+void WaitableEventDone::wait() {}
+
+bool WaitableEventDone::isReady()
+{
+    return true;
+}
+
 WorkerThreadPool::WorkerThreadPool()  = default;
 WorkerThreadPool::~WorkerThreadPool() = default;
 
@@ -219,4 +226,17 @@ std::shared_ptr<WorkerThreadPool> WorkerThreadPool::Create(bool multithreaded)
     return pool;
 }
 
+// static
+std::shared_ptr<WaitableEvent> WorkerThreadPool::PostWorkerTask(
+    std::shared_ptr<WorkerThreadPool> pool,
+    std::shared_ptr<Closure> task)
+{
+    std::shared_ptr<WaitableEvent> event = pool->postWorkerTask(task);
+    if (event.get())
+    {
+        event->setWorkerThreadPool(pool);
+    }
+    return event;
+}
+
 }  // namespace angle

src/libANGLE/WorkerThread.h

@@ -21,6 +21,8 @@
 namespace angle
 {
 
+class WorkerThreadPool;
+
 // A callback function with no return value and no arguments.
 class Closure
 {
@@ -41,6 +43,7 @@ class WaitableEvent : angle::NonCopyable
 
     // Peeks whether the event is ready. If ready, wait() will not block.
     virtual bool isReady() = 0;
+    void setWorkerThreadPool(std::shared_ptr<WorkerThreadPool> pool) { mPool = pool; }
 
     template <size_t Count>
     static void WaitMany(std::array<std::shared_ptr<WaitableEvent>, Count> *waitables)
@@ -51,6 +54,17 @@ class WaitableEvent : angle::NonCopyable
             (*waitables)[index]->wait();
         }
     }
+
+  private:
+    std::shared_ptr<WorkerThreadPool> mPool;
+};
+
+// A dummy waitable event.
+class WaitableEventDone final : public WaitableEvent
+{
+  public:
+    void wait() override;
+    bool isReady() override;
 };
 
 // Request WorkerThreads from the WorkerThreadPool. Each pool can keep worker threads around so
@@ -62,14 +76,17 @@ class WorkerThreadPool : angle::NonCopyable
     virtual ~WorkerThreadPool();
 
     static std::shared_ptr<WorkerThreadPool> Create(bool multithreaded);
-
-    // Returns an event to wait on for the task to finish.
-    // If the pool fails to create the task, returns null.
-    virtual std::shared_ptr<WaitableEvent> postWorkerTask(std::shared_ptr<Closure> task) = 0;
+    static std::shared_ptr<WaitableEvent> PostWorkerTask(std::shared_ptr<WorkerThreadPool> pool,
+                                                         std::shared_ptr<Closure> task);
 
     virtual void setMaxThreads(size_t maxThreads) = 0;
 
     virtual bool isAsync() = 0;
+
+  private:
+    // Returns an event to wait on for the task to finish.
+    // If the pool fails to create the task, returns null.
+    virtual std::shared_ptr<WaitableEvent> postWorkerTask(std::shared_ptr<Closure> task) = 0;
 };
 
 }  // namespace angle

src/libANGLE/WorkerThread_unittest.cpp

@@ -35,8 +35,10 @@ TEST(WorkerPoolTest, SimpleTask)
             {std::make_shared<TestTask>(), std::make_shared<TestTask>(),
              std::make_shared<TestTask>(), std::make_shared<TestTask>()}};
         std::array<std::shared_ptr<WaitableEvent>, 4> waitables = {
-            {pool->postWorkerTask(tasks[0]), pool->postWorkerTask(tasks[1]),
-             pool->postWorkerTask(tasks[2]), pool->postWorkerTask(tasks[3])}};
+            {WorkerThreadPool::PostWorkerTask(pool, tasks[0]),
+             WorkerThreadPool::PostWorkerTask(pool, tasks[1]),
+             WorkerThreadPool::PostWorkerTask(pool, tasks[2]),
+             WorkerThreadPool::PostWorkerTask(pool, tasks[3])}};
 
         WaitableEvent::WaitMany(&waitables);
 

src/libANGLE/renderer/ContextImpl.cpp

@@ -124,4 +124,5 @@ void ContextImpl::handleError(GLenum errorCode,
     errorStream << "Internal error: " << gl::FmtHex(errorCode) << ": " << message;
     mErrors->handleError(errorCode, errorStream.str().c_str(), file, function, line);
 }
+
 }  // namespace rx

src/libANGLE/renderer/ContextImpl.h

@@ -140,6 +140,9 @@ class ContextImpl : public GLImplFactory
     virtual void pushDebugGroup(GLenum source, GLuint id, const std::string &message) = 0;
     virtual void popDebugGroup()                                                      = 0;
 
+    // KHR_parallel_shader_compile
+    virtual void setMaxShaderCompilerThreads(GLuint count) {}
+
     // State sync with dirty bits.
     virtual angle::Result syncState(const gl::Context *context,
                                     const gl::State::DirtyBits &dirtyBits,

src/libANGLE/renderer/ShaderImpl.cpp

+//
+// Copyright 2019 The ANGLE Project Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+//
+
+// ShaderImpl.cpp: Implementation methods of ShaderImpl
+
+#include "libANGLE/renderer/ShaderImpl.h"
+
+#include "libANGLE/Context.h"
+
+namespace rx
+{
+
+WaitableCompileEvent::WaitableCompileEvent(std::shared_ptr<angle::WaitableEvent> waitableEvent)
+    : mWaitableEvent(waitableEvent)
+{}
+
+WaitableCompileEvent::~WaitableCompileEvent()
+{
+    mWaitableEvent.reset();
+}
+
+void WaitableCompileEvent::wait()
+{
+    mWaitableEvent->wait();
+}
+
+bool WaitableCompileEvent::isReady()
+{
+    return mWaitableEvent->isReady();
+}
+
+const std::string &WaitableCompileEvent::getInfoLog()
+{
+    return mInfoLog;
+}
+
+class TranslateTask : public angle::Closure
+{
+  public:
+    TranslateTask(ShHandle handle, ShCompileOptions options, const std::string &source)
+        : mHandle(handle), mOptions(options), mSource(source), mResult(false)
+    {}
+
+    void operator()() override
+    {
+        const char *source = mSource.c_str();
+        mResult            = sh::Compile(mHandle, &source, 1, mOptions);
+    }
+
+    bool getResult() { return mResult; }
+
+    ShHandle getHandle() { return mHandle; }
+
+  private:
+    ShHandle mHandle;
+    ShCompileOptions mOptions;
+    std::string mSource;
+    bool mResult;
+};
+
+class WaitableCompileEventImpl final : public WaitableCompileEvent
+{
+  public:
+    WaitableCompileEventImpl(std::shared_ptr<angle::WaitableEvent> waitableEvent,
+                             std::shared_ptr<TranslateTask> translateTask)
+        : WaitableCompileEvent(waitableEvent), mTranslateTask(translateTask)
+    {}
+
+    bool getResult() override { return mTranslateTask->getResult(); }
+
+    bool postTranslate(std::string *infoLog) override { return true; }
+
+  private:
+    std::shared_ptr<TranslateTask> mTranslateTask;
+};
+
+std::shared_ptr<WaitableCompileEvent> ShaderImpl::compileImpl(
+    const gl::Context *context,
+    gl::ShCompilerInstance *compilerInstance,
+    const std::string &source,
+    ShCompileOptions compileOptions)
+{
+    auto workerThreadPool = context->getWorkerThreadPool();
+    auto translateTask =
+        std::make_shared<TranslateTask>(compilerInstance->getHandle(), compileOptions, source);
+
+    return std::make_shared<WaitableCompileEventImpl>(
+        angle::WorkerThreadPool::PostWorkerTask(workerThreadPool, translateTask), translateTask);
+}
+
+}  // namespace rx

src/libANGLE/renderer/ShaderImpl.h

@@ -9,12 +9,42 @@
 #ifndef LIBANGLE_RENDERER_SHADERIMPL_H_
 #define LIBANGLE_RENDERER_SHADERIMPL_H_
 
+#include <functional>
+
 #include "common/angleutils.h"
 #include "libANGLE/Shader.h"
+#include "libANGLE/WorkerThread.h"
+
+namespace gl
+{
+class ShCompilerInstance;
+}  // namespace gl
 
 namespace rx
 {
 
+using UpdateShaderStateFunctor = std::function<void(bool compiled, ShHandle handle)>;
+class WaitableCompileEvent : public angle::WaitableEvent
+{
+  public:
+    WaitableCompileEvent(std::shared_ptr<angle::WaitableEvent> waitableEvent);
+    ~WaitableCompileEvent() override;
+
+    void wait() override;
+
+    bool isReady() override;
+
+    virtual bool getResult() = 0;
+
+    virtual bool postTranslate(std::string *infoLog) = 0;
+
+    const std::string &getInfoLog();
+
+  protected:
+    std::shared_ptr<angle::WaitableEvent> mWaitableEvent;
+    std::string mInfoLog;
+};
+
 class ShaderImpl : angle::NonCopyable
 {
   public:
@@ -23,22 +53,20 @@ class ShaderImpl : angle::NonCopyable
 
     virtual void destroy() {}
 
-    // Returns additional sh::Compile options.
-    virtual ShCompileOptions prepareSourceAndReturnOptions(const gl::Context *context,
-                                                           std::stringstream *sourceStream,
-                                                           std::string *sourcePath) = 0;
-
-    // Uses the GL driver to compile the shader source in a worker thread.
-    virtual void compileAsync(const std::string &source, std::string &infoLog) {}
-
-    // Returns success for compiling on the driver. Returns success.
-    virtual bool postTranslateCompile(gl::ShCompilerInstance *compiler, std::string *infoLog) = 0;
+    virtual std::shared_ptr<WaitableCompileEvent> compile(const gl::Context *context,
+                                                          gl::ShCompilerInstance *compilerInstance,
+                                                          ShCompileOptions options) = 0;
 
     virtual std::string getDebugInfo() const = 0;
 
     const gl::ShaderState &getData() const { return mData; }
 
   protected:
+    std::shared_ptr<WaitableCompileEvent> compileImpl(const gl::Context *context,
+                                                      gl::ShCompilerInstance *compilerInstance,
+                                                      const std::string &source,
+                                                      ShCompileOptions compileOptions);
+
     const gl::ShaderState &mData;
 };
 

src/libANGLE/renderer/d3d/ProgramD3D.cpp

@@ -955,7 +955,7 @@ class ProgramD3D::LoadBinaryLinkEvent final : public LinkEvent
                         gl::BinaryInputStream *stream,
                         gl::InfoLog &infoLog)
         : mTask(std::make_shared<ProgramD3D::LoadBinaryTask>(context, program, stream, infoLog)),
-          mWaitableEvent(workerPool->postWorkerTask(mTask))
+          mWaitableEvent(angle::WorkerThreadPool::PostWorkerTask(workerPool, mTask))
     {}
 
     angle::Result wait(const gl::Context *context) override
@@ -1771,14 +1771,15 @@ class ProgramD3D::GraphicsProgramLinkEvent final : public LinkEvent
                              const ShaderD3D *vertexShader,
                              const ShaderD3D *fragmentShader)
         : mInfoLog(infoLog),
-          mWorkerPool(workerPool),
           mVertexTask(vertexTask),
           mPixelTask(pixelTask),
           mGeometryTask(geometryTask),
-          mWaitEvents(
-              {{std::shared_ptr<WaitableEvent>(workerPool->postWorkerTask(mVertexTask)),
-                std::shared_ptr<WaitableEvent>(workerPool->postWorkerTask(mPixelTask)),
-                std::shared_ptr<WaitableEvent>(workerPool->postWorkerTask(mGeometryTask))}}),
+          mWaitEvents({{std::shared_ptr<WaitableEvent>(
+                            angle::WorkerThreadPool::PostWorkerTask(workerPool, mVertexTask)),
+                        std::shared_ptr<WaitableEvent>(
+                            angle::WorkerThreadPool::PostWorkerTask(workerPool, mPixelTask)),
+                        std::shared_ptr<WaitableEvent>(
+                            angle::WorkerThreadPool::PostWorkerTask(workerPool, mGeometryTask))}}),
           mUseGS(useGS),
           mVertexShader(vertexShader),
           mFragmentShader(fragmentShader)
@@ -1863,7 +1864,6 @@ class ProgramD3D::GraphicsProgramLinkEvent final : public LinkEvent
     }
 
     gl::InfoLog &mInfoLog;
-    std::shared_ptr<WorkerThreadPool> mWorkerPool;
     std::shared_ptr<ProgramD3D::GetVertexExecutableTask> mVertexTask;
     std::shared_ptr<ProgramD3D::GetPixelExecutableTask> mPixelTask;
     std::shared_ptr<ProgramD3D::GetGeometryExecutableTask> mGeometryTask;

src/libANGLE/renderer/d3d/ShaderD3D.h

@@ -38,11 +38,10 @@ class ShaderD3D : public ShaderImpl
               const gl::Extensions &extensions);
     ~ShaderD3D() override;
 
-    // ShaderImpl implementation
-    ShCompileOptions prepareSourceAndReturnOptions(const gl::Context *context,
-                                                   std::stringstream *sourceStream,
-                                                   std::string *sourcePath) override;
-    bool postTranslateCompile(gl::ShCompilerInstance *compiler, std::string *infoLog) override;
+    std::shared_ptr<WaitableCompileEvent> compile(const gl::Context *context,
+                                                  gl::ShCompilerInstance *compilerInstance,
+                                                  ShCompileOptions options) override;
+
     std::string getDebugInfo() const override;
 
     // D3D-specific methods

src/libANGLE/renderer/gl/ContextGL.cpp

@@ -608,4 +608,10 @@ angle::Result ContextGL::memoryBarrierByRegion(const gl::Context *context, GLbit
 {
     return mRenderer->memoryBarrierByRegion(barriers);
 }
+
+void ContextGL::setMaxShaderCompilerThreads(GLuint count)
+{
+    mRenderer->setMaxShaderCompilerThreads(count);
+}
+
 }  // namespace rx

src/libANGLE/renderer/gl/ContextGL.h

@@ -213,6 +213,8 @@ class ContextGL : public ContextImpl
     angle::Result memoryBarrier(const gl::Context *context, GLbitfield barriers) override;
     angle::Result memoryBarrierByRegion(const gl::Context *context, GLbitfield barriers) override;
 
+    void setMaxShaderCompilerThreads(GLuint count) override;
+
   private:
     angle::Result setDrawArraysState(const gl::Context *context,
                                      GLint first,

src/libANGLE/renderer/gl/DispatchTableGL_autogen.cpp

@@ -1128,6 +1128,11 @@ void DispatchTableGL::initProcsDesktopGL(const gl::Version &version,
         ASSIGN("glIsQueryARB", isQuery);
     }
 
+    if (extensions.count("GL_ARB_parallel_shader_compile") != 0)
+    {
+        ASSIGN("glMaxShaderCompilerThreadsARB", maxShaderCompilerThreadsARB);
+    }
+
     if (extensions.count("GL_ARB_point_parameters") != 0)
     {
         ASSIGN("glPointParameterfARB", pointParameterf);
@@ -2699,6 +2704,11 @@ void DispatchTableGL::initProcsSharedExtensions(const std::set<std::string> &ext
         ASSIGN("glValidateProgramPipelineEXT", validateProgramPipeline);
     }
 
+    if (extensions.count("GL_KHR_parallel_shader_compile") != 0)
+    {
+        ASSIGN("glMaxShaderCompilerThreadsKHR", maxShaderCompilerThreadsKHR);
+    }
+
     if (extensions.count("GL_NV_fence") != 0)
     {
         ASSIGN("glDeleteFencesNV", deleteFencesNV);
@@ -3849,6 +3859,11 @@ void DispatchTableGL::initProcsDesktopGLNULL(const gl::Version &version,
         isQuery           = &glIsQueryNULL;
     }
 
+    if (extensions.count("GL_ARB_parallel_shader_compile") != 0)
+    {
+        maxShaderCompilerThreadsARB = &glMaxShaderCompilerThreadsARBNULL;
+    }
+
     if (extensions.count("GL_ARB_point_parameters") != 0)
     {
         pointParameterf  = &glPointParameterfNULL;
@@ -5419,6 +5434,11 @@ void DispatchTableGL::initProcsSharedExtensionsNULL(const std::set<std::string>
         validateProgramPipeline   = &glValidateProgramPipelineNULL;
     }
 
+    if (extensions.count("GL_KHR_parallel_shader_compile") != 0)
+    {
+        maxShaderCompilerThreadsKHR = &glMaxShaderCompilerThreadsKHRNULL;
+    }
+
     if (extensions.count("GL_NV_fence") != 0)
     {
         deleteFencesNV = &glDeleteFencesNVNULL;

src/libANGLE/renderer/gl/DispatchTableGL_autogen.h

@@ -730,6 +730,9 @@ class DispatchTableGL : angle::NonCopyable
     PFNGLBLENDBARRIERPROC blendBarrier = nullptr;
     PFNGLPRIMITIVEBOUNDINGBOXPROC primitiveBoundingBox = nullptr;
 
+    // GL_ARB_parallel_shader_compile
+    PFNGLMAXSHADERCOMPILERTHREADSARBPROC maxShaderCompilerThreadsARB = nullptr;
+
     // GL_EXT_debug_marker
     PFNGLINSERTEVENTMARKEREXTPROC insertEventMarkerEXT = nullptr;
     PFNGLPOPGROUPMARKEREXTPROC popGroupMarkerEXT = nullptr;
@@ -738,6 +741,9 @@ class DispatchTableGL : angle::NonCopyable
     // GL_EXT_discard_framebuffer
     PFNGLDISCARDFRAMEBUFFEREXTPROC discardFramebufferEXT = nullptr;
 
+    // GL_KHR_parallel_shader_compile
+    PFNGLMAXSHADERCOMPILERTHREADSKHRPROC maxShaderCompilerThreadsKHR = nullptr;
+
     // GL_NV_internalformat_sample_query
     PFNGLGETINTERNALFORMATSAMPLEIVNVPROC getInternalformatSampleivNV = nullptr;
 

src/libANGLE/renderer/gl/ProgramGL.cpp

@@ -148,16 +148,51 @@ class ProgramGL::LinkTask final : public angle::Closure
 };
 
 using PostLinkImplFunctor = std::function<angle::Result(bool, const std::string &)>;
+
+// The event for a parallelized linking using the native driver extension.
+class ProgramGL::LinkEventNativeParallel final : public LinkEvent
+{
+  public:
+    LinkEventNativeParallel(PostLinkImplFunctor &&functor,
+                            const FunctionsGL *functions,
+                            GLuint programID)
+        : mPostLinkImplFunctor(functor), mFunctions(functions), mProgramID(programID)
+    {}
+
+    angle::Result wait(const gl::Context *context) override
+    {
+        GLint linkStatus = GL_FALSE;
+        mFunctions->getProgramiv(mProgramID, GL_LINK_STATUS, &linkStatus);
+        if (linkStatus == GL_TRUE)
+        {
+            return mPostLinkImplFunctor(false, std::string());
+        }
+        return angle::Result::Incomplete;
+    }
+
+    bool isLinking() override
+    {
+        GLint completionStatus = GL_FALSE;
+        mFunctions->getProgramiv(mProgramID, GL_COMPLETION_STATUS, &completionStatus);
+        return completionStatus == GL_FALSE;
+    }
+
+  private:
+    PostLinkImplFunctor mPostLinkImplFunctor;
+    const FunctionsGL *mFunctions;
+    GLuint mProgramID;
+};
+
+// The event for a parallelized linking using the worker thread pool.
 class ProgramGL::LinkEventGL final : public LinkEvent
 {
   public:
     LinkEventGL(std::shared_ptr<angle::WorkerThreadPool> workerPool,
                 std::shared_ptr<ProgramGL::LinkTask> linkTask,
                 PostLinkImplFunctor &&functor)
-        : mWorkerPool(workerPool),
-          mLinkTask(linkTask),
-          mWaitableEvent(
-              std::shared_ptr<angle::WaitableEvent>(workerPool->postWorkerTask(mLinkTask))),
+        : mLinkTask(linkTask),
+          mWaitableEvent(std::shared_ptr<angle::WaitableEvent>(
+              angle::WorkerThreadPool::PostWorkerTask(workerPool, mLinkTask))),
           mPostLinkImplFunctor(functor)
     {}
 
@@ -170,7 +205,6 @@ class ProgramGL::LinkEventGL final : public LinkEvent
     bool isLinking() override { return !mWaitableEvent->isReady(); }
 
   private:
-    std::shared_ptr<angle::WorkerThreadPool> mWorkerPool;
     std::shared_ptr<ProgramGL::LinkTask> mLinkTask;
     std::shared_ptr<angle::WaitableEvent> mWaitableEvent;
     PostLinkImplFunctor mPostLinkImplFunctor;
@@ -430,7 +464,13 @@ std::unique_ptr<LinkEvent> ProgramGL::link(const gl::Context *context,
         return angle::Result::Continue;
     };
 
-    if (workerPool->isAsync() && (!mWorkarounds.dontRelinkProgramsInParallel || !mLinkedInParallel))
+    if (mRenderer->hasNativeParallelCompile())
+    {
+        mFunctions->linkProgram(mProgramID);
+        return std::make_unique<LinkEventNativeParallel>(postLinkImplTask, mFunctions, mProgramID);
+    }
+    else if (workerPool->isAsync() &&
+             (!mWorkarounds.dontRelinkProgramsInParallel || !mLinkedInParallel))
     {
         mLinkedInParallel = true;
         return std::make_unique<LinkEventGL>(workerPool, linkTask, postLinkImplTask);

src/libANGLE/renderer/gl/ProgramGL.h

@@ -117,6 +117,7 @@ class ProgramGL : public ProgramImpl
 
   private:
     class LinkTask;
+    class LinkEventNativeParallel;
     class LinkEventGL;
 
     void preLink();

src/libANGLE/renderer/gl/RendererGL.cpp

@@ -55,6 +55,19 @@ std::vector<GLuint> GatherPaths(const std::vector<gl::Path *> &paths)
     return ret;
 }
 
+void SetMaxShaderCompilerThreads(const rx::FunctionsGL *functions, GLuint count)
+{
+    if (functions->maxShaderCompilerThreadsKHR != nullptr)
+    {
+        functions->maxShaderCompilerThreadsKHR(count);
+    }
+    else
+    {
+        ASSERT(functions->maxShaderCompilerThreadsARB != nullptr);
+        functions->maxShaderCompilerThreadsARB(count);
+    }
+}
+
 }  // namespace
 
 static void INTERNAL_GL_APIENTRY LogGLDebugMessage(GLenum source,
@@ -175,7 +188,8 @@ RendererGL::RendererGL(std::unique_ptr<FunctionsGL> functions, const egl::Attrib
       mMultiviewClearer(nullptr),
       mUseDebugOutput(false),
       mCapsInitialized(false),
-      mMultiviewImplementationType(MultiviewImplementationTypeGL::UNSPECIFIED)
+      mMultiviewImplementationType(MultiviewImplementationTypeGL::UNSPECIFIED),
+      mNativeParallelCompileEnabled(false)
 {
     ASSERT(mFunctions);
     nativegl_gl::GenerateWorkarounds(mFunctions.get(), &mWorkarounds);
@@ -215,6 +229,12 @@ RendererGL::RendererGL(std::unique_ptr<FunctionsGL> functions, const egl::Attrib
             mFunctions->vertexAttrib4f(i, 0.0f, 0.0f, 0.0f, 1.0f);
         }
     }
+
+    if (hasNativeParallelCompile() && !mNativeParallelCompileEnabled)
+    {
+        SetMaxShaderCompilerThreads(mFunctions.get(), 0xffffffff);
+        mNativeParallelCompileEnabled = true;
+    }
 }
 
 RendererGL::~RendererGL()
@@ -617,6 +637,20 @@ unsigned int RendererGL::getMaxWorkerContexts()
     return std::min(16u, std::thread::hardware_concurrency());
 }
 
+bool RendererGL::hasNativeParallelCompile()
+{
+    return mFunctions->maxShaderCompilerThreadsKHR != nullptr ||
+           mFunctions->maxShaderCompilerThreadsARB != nullptr;
+}
+
+void RendererGL::setMaxShaderCompilerThreads(GLuint count)
+{
+    if (hasNativeParallelCompile())
+    {
+        SetMaxShaderCompilerThreads(mFunctions.get(), count);
+    }
+}
+
 ScopedWorkerContextGL::ScopedWorkerContextGL(RendererGL *renderer, std::string *infoLog)
     : mRenderer(renderer)
 {

src/libANGLE/renderer/gl/RendererGL.h

@@ -180,6 +180,10 @@ class RendererGL : angle::NonCopyable
 
     bool bindWorkerContext(std::string *infoLog);
     void unbindWorkerContext();
+    // Checks if the driver has the KHR_parallel_shader_compile or ARB_parallel_shader_compile
+    // extension.
+    bool hasNativeParallelCompile();
+    void setMaxShaderCompilerThreads(GLuint count);
 
     static unsigned int getMaxWorkerContexts();
 
@@ -218,6 +222,8 @@ class RendererGL : angle::NonCopyable
     std::list<std::unique_ptr<WorkerContext>> mWorkerContextPool;
     // Protect the concurrent accesses to worker contexts.
     std::mutex mWorkerMutex;
+
+    bool mNativeParallelCompileEnabled;
 };
 
 }  // namespace rx

src/libANGLE/renderer/gl/ShaderGL.h

@@ -28,23 +28,24 @@ class ShaderGL : public ShaderImpl
 
     void destroy() override;
 
-    // ShaderImpl implementation
-    ShCompileOptions prepareSourceAndReturnOptions(const gl::Context *context,
-                                                   std::stringstream *sourceStream,
-                                                   std::string *sourcePath) override;
-    void compileAsync(const std::string &source, std::string &infoLog) override;
-    bool postTranslateCompile(gl::ShCompilerInstance *compiler, std::string *infoLog) override;
+    std::shared_ptr<WaitableCompileEvent> compile(const gl::Context *context,
+                                                  gl::ShCompilerInstance *compilerInstance,
+                                                  ShCompileOptions options) override;
+
     std::string getDebugInfo() const override;
 
     GLuint getShaderID() const;
 
   private:
     void compileAndCheckShader(const char *source);
+    void compileShader(const char *source);
+    void checkShader();
+    bool peekCompletion();
+    bool compileAndCheckShaderInWorker(const char *source);
 
     GLuint mShaderID;
     MultiviewImplementationTypeGL mMultiviewImplementationType;
     std::shared_ptr<RendererGL> mRenderer;
-    bool mFallbackToMainThread;
     GLint mCompileStatus;
     std::string mInfoLog;
 };

src/libANGLE/renderer/gl/functionsgl_enums.h

@@ -1371,5 +1371,6 @@
 #define GL_TEXTURE_TARGET 0x1006
 #define GL_UNKNOWN_CONTEXT_RESET 0x8255
 #define GL_ZERO_TO_ONE 0x935F
+#define GL_COMPLETION_STATUS 0x91B1
 
 #endif  // LIBANGLE_RENDERER_GL_FUNCTIONSGLENUMS_H_

src/libANGLE/renderer/gl/functionsgl_typedefs.h

@@ -1857,6 +1857,13 @@ typedef void(INTERNAL_GL_APIENTRY *PFNGLINSERTEVENTMARKEREXTPROC)(GLsizei length
 typedef void(INTERNAL_GL_APIENTRY *PFNGLPUSHGROUPMARKEREXTPROC)(GLsizei length,
                                                                 const GLchar *marker);
 typedef void(INTERNAL_GL_APIENTRY *PFNGLPOPGROUPMARKEREXTPROC)(void);
+
+// KHR_parallel_shader_compile
+typedef void(INTERNAL_GL_APIENTRY *PFNGLMAXSHADERCOMPILERTHREADSKHRPROC)(GLuint count);
+
+// ARB_parallel_shader_compile
+typedef void(INTERNAL_GL_APIENTRY *PFNGLMAXSHADERCOMPILERTHREADSARBPROC)(GLuint count);
+
 }  // namespace rx
 
 #endif  // LIBANGLE_RENDERER_GL_FUNCTIONSGLTYPEDEFS_H_

src/libANGLE/renderer/gl/generate_gl_dispatch_table.py

@@ -297,7 +297,7 @@ def main():
             command_name = command.attrib['name']
             if 'gl' in support and 'gles2' in support:
                 # Special case for KHR extensions, since in GLES they are suffixed.
-                if '_KHR_' in extension_name:
+                if '_KHR_' in extension_name and not command_name.endswith('KHR'):
                     safe_append(gl_extension_commands, command_name, extension_name)
                     safe_append(gles2_extension_commands, command_name, extension_name)
                 else:

src/libANGLE/renderer/gl/gl_bindings_data.json

@@ -790,5 +790,15 @@
         "InsertEventMarkerEXT",
         "PushGroupMarkerEXT",
         "PopGroupMarkerEXT"
+    ],
+
+    "GL_ARB_parallel_shader_compile":
+    [
+        "MaxShaderCompilerThreadsARB"
+    ],
+
+    "GL_KHR_parallel_shader_compile":
+    [
+        "MaxShaderCompilerThreadsKHR"
     ]
 }

src/libANGLE/renderer/gl/glx/DisplayGLX.cpp

@@ -24,6 +24,17 @@
 #include "libANGLE/renderer/gl/glx/WindowSurfaceGLX.h"
 #include "libANGLE/renderer/gl/renderergl_utils.h"
 
+namespace
+{
+
+bool HasParallelShaderCompileExtension(const rx::FunctionsGL *functions)
+{
+    return functions->maxShaderCompilerThreadsKHR != nullptr ||
+           functions->maxShaderCompilerThreadsARB != nullptr;
+}
+
+}  // anonymous namespace
+
 namespace rx
 {
 
@@ -289,14 +300,23 @@ egl::Error DisplayGLX::initialize(egl::Display *display)
 
     if (mSharedContext)
     {
-        for (unsigned int i = 0; i < RendererGL::getMaxWorkerContexts(); ++i)
+        if (HasParallelShaderCompileExtension(functionsGL.get()))
+        {
+            mGLX.destroyContext(mSharedContext);
+            mSharedContext = nullptr;
+        }
+        else
         {
-            glx::Pbuffer workerPbuffer = mGLX.createPbuffer(mContextConfig, dummyPbufferAttribs);
-            if (!workerPbuffer)
+            for (unsigned int i = 0; i < RendererGL::getMaxWorkerContexts(); ++i)
             {
-                return egl::EglNotInitialized() << "Could not create the worker pbuffers.";
+                glx::Pbuffer workerPbuffer =
+                    mGLX.createPbuffer(mContextConfig, dummyPbufferAttribs);
+                if (!workerPbuffer)
+                {
+                    return egl::EglNotInitialized() << "Could not create the worker pbuffers.";
+                }
+                mWorkerPbufferPool.push_back(workerPbuffer);
             }
-            mWorkerPbufferPool.push_back(workerPbuffer);
         }
     }
 

src/libANGLE/renderer/gl/null_functions.cpp

@@ -1672,6 +1672,10 @@ void *INTERNAL_GL_APIENTRY glMapNamedBufferRangeNULL(GLuint buffer,
 
 void INTERNAL_GL_APIENTRY glMatrixLoadfEXTNULL(GLenum mode, const GLfloat *m) {}
 
+void INTERNAL_GL_APIENTRY glMaxShaderCompilerThreadsARBNULL(GLuint count) {}
+
+void INTERNAL_GL_APIENTRY glMaxShaderCompilerThreadsKHRNULL(GLuint count) {}
+
 void INTERNAL_GL_APIENTRY glMemoryBarrierNULL(GLbitfield barriers) {}
 
 void INTERNAL_GL_APIENTRY glMemoryBarrierByRegionNULL(GLbitfield barriers) {}

src/libANGLE/renderer/gl/null_functions.h

@@ -1036,6 +1036,8 @@ void *INTERNAL_GL_APIENTRY glMapNamedBufferRangeNULL(GLuint buffer,
                                                      GLsizeiptr length,
                                                      GLbitfield access);
 void INTERNAL_GL_APIENTRY glMatrixLoadfEXTNULL(GLenum mode, const GLfloat *m);
+void INTERNAL_GL_APIENTRY glMaxShaderCompilerThreadsARBNULL(GLuint count);
+void INTERNAL_GL_APIENTRY glMaxShaderCompilerThreadsKHRNULL(GLuint count);
 void INTERNAL_GL_APIENTRY glMemoryBarrierNULL(GLbitfield barriers);
 void INTERNAL_GL_APIENTRY glMemoryBarrierByRegionNULL(GLbitfield barriers);
 void INTERNAL_GL_APIENTRY glMinSampleShadingNULL(GLfloat value);

src/libANGLE/renderer/null/ShaderNULL.cpp

@@ -10,6 +10,7 @@
 #include "libANGLE/renderer/null/ShaderNULL.h"
 
 #include "common/debug.h"
+#include "libANGLE/Context.h"
 
 namespace rx
 {
@@ -18,17 +19,11 @@ ShaderNULL::ShaderNULL(const gl::ShaderState &data) : ShaderImpl(data) {}
 
 ShaderNULL::~ShaderNULL() {}
 
-ShCompileOptions ShaderNULL::prepareSourceAndReturnOptions(const gl::Context *context,
-                                                           std::stringstream *sourceStream,
-                                                           std::string *sourcePath)
+std::shared_ptr<WaitableCompileEvent> ShaderNULL::compile(const gl::Context *context,
+                                                          gl::ShCompilerInstance *compilerInstance,
+                                                          ShCompileOptions options)
 {
-    *sourceStream << mData.getSource();
-    return 0;
-}
-
-bool ShaderNULL::postTranslateCompile(gl::ShCompilerInstance *compiler, std::string *infoLog)
-{
-    return true;
+    return compileImpl(context, compilerInstance, mData.getSource(), options);
 }
 
 std::string ShaderNULL::getDebugInfo() const

src/libANGLE/renderer/null/ShaderNULL.h

@@ -21,12 +21,9 @@ class ShaderNULL : public ShaderImpl
     ShaderNULL(const gl::ShaderState &data);
     ~ShaderNULL() override;
 
-    // Returns additional sh::Compile options.
-    ShCompileOptions prepareSourceAndReturnOptions(const gl::Context *context,
-                                                   std::stringstream *sourceStream,
-                                                   std::string *sourcePath) override;
-    // Returns success for compiling on the driver. Returns success.
-    bool postTranslateCompile(gl::ShCompilerInstance *compiler, std::string *infoLog) override;
+    std::shared_ptr<WaitableCompileEvent> compile(const gl::Context *context,
+                                                  gl::ShCompilerInstance *compilerInstance,
+                                                  ShCompileOptions options) override;
 
     std::string getDebugInfo() const override;
 };

src/libANGLE/renderer/vulkan/ShaderVk.cpp

@@ -21,12 +21,10 @@ ShaderVk::ShaderVk(const gl::ShaderState &data) : ShaderImpl(data) {}
 
 ShaderVk::~ShaderVk() {}
 
-ShCompileOptions ShaderVk::prepareSourceAndReturnOptions(const gl::Context *context,
-                                                         std::stringstream *sourceStream,
-                                                         std::string *sourcePath)
+std::shared_ptr<WaitableCompileEvent> ShaderVk::compile(const gl::Context *context,
+                                                        gl::ShCompilerInstance *compilerInstance,
+                                                        ShCompileOptions options)
 {
-    *sourceStream << mData.getSource();
-
     ShCompileOptions compileOptions = SH_INITIALIZE_UNINITIALIZED_LOCALS;
 
     ContextVk *contextVk = vk::GetImpl(context);
@@ -36,13 +34,7 @@ ShCompileOptions ShaderVk::prepareSourceAndReturnOptions(const gl::Context *cont
         compileOptions |= SH_CLAMP_POINT_SIZE;
     }
 
-    return compileOptions;
-}
-
-bool ShaderVk::postTranslateCompile(gl::ShCompilerInstance *compiler, std::string *infoLog)
-{
-    // No work to do here.
-    return true;
+    return compileImpl(context, compilerInstance, mData.getSource(), compileOptions | options);
 }
 
 std::string ShaderVk::getDebugInfo() const

src/libANGLE/renderer/vulkan/ShaderVk.h

@@ -21,12 +21,9 @@ class ShaderVk : public ShaderImpl
     ShaderVk(const gl::ShaderState &data);
     ~ShaderVk() override;
 
-    // Returns additional sh::Compile options.
-    ShCompileOptions prepareSourceAndReturnOptions(const gl::Context *context,
-                                                   std::stringstream *sourceStream,
-                                                   std::string *sourcePath) override;
-    // Returns success for compiling on the driver. Returns success.
-    bool postTranslateCompile(gl::ShCompilerInstance *compiler, std::string *infoLog) override;
+    std::shared_ptr<WaitableCompileEvent> compile(const gl::Context *context,
+                                                  gl::ShCompilerInstance *compilerInstance,
+                                                  ShCompileOptions options) override;
 
     std::string getDebugInfo() const override;
 };

src/libGLESv2.gni

@@ -306,6 +306,7 @@ libangle_sources = [
   "src/libANGLE/renderer/RenderbufferImpl.h",
   "src/libANGLE/renderer/RenderTargetCache.h",
   "src/libANGLE/renderer/SamplerImpl.h",
+  "src/libANGLE/renderer/ShaderImpl.cpp",
   "src/libANGLE/renderer/ShaderImpl.h",
   "src/libANGLE/renderer/StreamProducerImpl.h",
   "src/libANGLE/renderer/SurfaceImpl.cpp",