Update Gitignore to ignore generated files.

- Re-enable renderdiff test.
- Cache the mesa directory so that we're not spending time pulling
  and compiling it.
- Move python prereqs into the script and use venv.

.github/workflows/presubmit.yml

@@ -108,16 +108,27 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.x'
-      - name: Install python prereqs
-        run: pip install mako setuptools pyyaml
-      - name: Run script
+      - name: Cache Mesa and deps
+        id: mesa-cache
+        uses: actions/cache@v4 # Use a specific version
+        with:
+          path: |
+            $HOME/Library/Caches/Homebrew
+            mesa
+          key: ${{ runner.os }}-mesa-deps-${{ vars.MESA_VERSION }}
+      - name: Get Mesa
+        id: mesa-prereq
+        env:
+          MESA_VERSION: ${{ vars.MESA_VERSION }}
         run: |
-          echo "disabled test"
-#          bash test/renderdiff/test.sh
-#      - uses: actions/upload-artifact@v4
-#        with:
-#          name: presubmit-renderdiff-result
-#          path: ./out/renderdiff_tests
+          bash test/utils/get_mesa.sh
+      - name: Run Test
+        run: |
+          bash test/renderdiff/test.sh
+      - uses: actions/upload-artifact@v4
+        with:
+          name: presubmit-renderdiff-result
+          path: ./out/renderdiff_tests
 
   validate-wgsl-webgpu:
     name: validate-wgsl-webgpu

.gitignore

@@ -18,3 +18,18 @@ test*.json
 results
 /compile_commands.json
 /.cache
+build/.cmake/
+build/CMakeCache.txt
+build/CMakeFiles/
+build/Makefile
+build/SPIRV-Tools*
+build/cmake_install.cmake
+build/compile_commands.json
+build/filament/
+build/include/
+build/libs/
+build/mac/ninja
+build/samples/
+build/shaders/
+build/third_party/
+build/tools/

CMakeLists.txt

@@ -801,6 +801,7 @@ add_subdirectory(${EXTERNAL}/draco/tnt)
 add_subdirectory(${EXTERNAL}/jsmn/tnt)
 add_subdirectory(${EXTERNAL}/stb/tnt)
 add_subdirectory(${EXTERNAL}/getopt)
+add_subdirectory(${EXTERNAL}/perfetto/tnt)
 
 # Note that this has to be placed after mikktspace in order for combine_static_libs to work.
 add_subdirectory(${LIBRARIES}/geometry)

android/filamat-android/CMakeLists.txt

@@ -26,6 +26,10 @@ add_library(utils STATIC IMPORTED)
 set_target_properties(utils PROPERTIES IMPORTED_LOCATION
         ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libutils.a)
 
+add_library(perfetto STATIC IMPORTED)
+set_target_properties(perfetto PROPERTIES IMPORTED_LOCATION
+        ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libperfetto.a)
+
 add_library(filabridge STATIC IMPORTED)
 set_target_properties(filabridge PROPERTIES IMPORTED_LOCATION
         ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libfilabridge.a)
@@ -40,6 +44,7 @@ set_target_properties(shaders PROPERTIES IMPORTED_LOCATION
 
 set(FILAMAT_INCLUDE_DIRS
         ../../libs/utils/include
+        ../../third_party/perfetto
         )
 
 include_directories(${FILAMENT_DIR}/include)
@@ -55,6 +60,7 @@ target_link_libraries(filamat-jni
         filabridge
         shaders
         utils
+        perfetto
         log
         smol-v
         $<$<STREQUAL:${FILAMENT_SUPPORTS_WEBGPU},ON>:tint>

android/filament-android/CMakeLists.txt

@@ -21,6 +21,10 @@ add_library(utils STATIC IMPORTED)
 set_target_properties(utils PROPERTIES IMPORTED_LOCATION
         ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libutils.a)
 
+add_library(perfetto STATIC IMPORTED)
+set_target_properties(perfetto PROPERTIES IMPORTED_LOCATION
+        ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libperfetto.a)
+
 add_library(ibl-lite STATIC IMPORTED)
 set_target_properties(ibl-lite PROPERTIES IMPORTED_LOCATION
         ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libibl-lite.a)
@@ -123,6 +127,7 @@ target_link_libraries(filament-jni
     PRIVATE android
     PRIVATE jnigraphics
     PRIVATE utils
+    PRIVATE perfetto
 
     # libgeometry is PUBLIC because gltfio uses it.
     PUBLIC geometry
@@ -141,6 +146,7 @@ target_include_directories(filament-jni PRIVATE
         ${FILAMENT_DIR}/include
         ../../filament/backend/include
         ../../third_party/robin-map
+        ../../third_party/perfetto
         ../../libs/utils/include)
 
 # Force a relink when the version script is changed:

android/gltfio-android/CMakeLists.txt

@@ -35,6 +35,10 @@ add_library(utils STATIC IMPORTED)
 set_target_properties(utils PROPERTIES IMPORTED_LOCATION
         ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libutils.a)
 
+add_library(perfetto STATIC IMPORTED)
+set_target_properties(perfetto PROPERTIES IMPORTED_LOCATION
+        ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libperfetto.a)
+
 add_library(uberzlib STATIC IMPORTED)
 set_target_properties(uberzlib PROPERTIES IMPORTED_LOCATION
         ${FILAMENT_DIR}/lib/${ANDROID_ABI}/libuberzlib.a)
@@ -121,6 +125,7 @@ set(GLTFIO_INCLUDE_DIRS
         ../../third_party/meshoptimizer/src
         ../../third_party/robin-map
         ../../third_party/stb
+        ../../third_party/perfetto
         ../../libs/utils/include
         ../../libs/ktxreader/include
 )
@@ -129,7 +134,7 @@ add_library(gltfio-jni SHARED ${GLTFIO_SRCS})
 target_include_directories(gltfio-jni PRIVATE ${GLTFIO_INCLUDE_DIRS})
 set_target_properties(gltfio-jni PROPERTIES LINK_DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/libgltfio-jni.symbols)
 set_target_properties(gltfio-jni PROPERTIES LINK_DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/libgltfio-jni.map)
-target_link_libraries(gltfio-jni filament-jni utils uberzlib log stb ktxreader basis_transcoder zstd uberarchive)
+target_link_libraries(gltfio-jni filament-jni utils perfetto uberzlib log stb ktxreader basis_transcoder zstd uberarchive)
 target_link_libraries(gltfio-jni dracodec meshoptimizer)
 target_compile_definitions(gltfio-jni PUBLIC GLTFIO_DRACO_SUPPORTED=1)
 target_include_directories(gltfio-jni PRIVATE ${DRACO_DIR}/src)

filament/backend/CMakeLists.txt

@@ -259,6 +259,7 @@ if (FILAMENT_SUPPORTS_WEBGPU)
             src/webgpu/WebGPUHandles.h
             src/webgpu/WebGPUSwapChain.cpp
             src/webgpu/WebGPUSwapChain.h
+            src/webgpu/WGPUProgram.cpp
     )
     if (WIN32)
         list(APPEND SRCS src/webgpu/platform/WebGPUPlatformWindows.cpp)

filament/backend/src/CommandStream.cpp

@@ -84,7 +84,7 @@ void CommandStream::execute(void* buffer) {
 
     Profiler profiler;
 
-    if (SYSTRACE_TAG) {
+    if constexpr (SYSTRACE_TAG) {
         if (UTILS_UNLIKELY(mUsePerformanceCounter)) {
             // we want to remove all this when tracing is completely disabled
             profiler.resetEvents(Profiler::EV_CPU_CYCLES  | Profiler::EV_BPU_MISSES);
@@ -100,7 +100,7 @@ void CommandStream::execute(void* buffer) {
         }
     });
 
-    if (SYSTRACE_TAG) {
+    if constexpr (SYSTRACE_TAG) {
         if (UTILS_UNLIKELY(mUsePerformanceCounter)) {
             // we want to remove all this when tracing is completely disabled
             profiler.stop();

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -85,6 +85,7 @@ OpenGLProgram::~OpenGLProgram() noexcept {
         delete lazyInitializationData;
 
         ShaderCompilerService::terminate(mToken);
+        assert_invariant(!mToken);
     }
 
     delete [] mUniformsRecords;

filament/backend/src/opengl/ShaderCompilerService.cpp

@@ -17,11 +17,15 @@
 #include "ShaderCompilerService.h"
 
 #include "BlobCacheKey.h"
+#include "CallbackManager.h"
+#include "CompilerThreadPool.h"
 #include "OpenGLBlobCache.h"
 #include "OpenGLDriver.h"
 
+#include <iterator>
 #include <private/backend/BackendUtils.h>
 
+#include <backend/DriverEnums.h>
 #include <backend/Program.h>
 
 #include <utils/compiler.h>
@@ -34,9 +38,9 @@
 #include <utils/Panic.h>
 #include <utils/Systrace.h>
 
+#include <algorithm>
 #include <array>
 #include <cctype>
-#include <chrono>
 #include <mutex>
 #include <memory>
 #include <string>
@@ -47,6 +51,7 @@
 
 #include <stddef.h>
 #include <stdint.h>
+#include <string.h>
 
 namespace filament::backend {
 
@@ -54,17 +59,17 @@ using namespace utils;
 
 // ------------------------------------------------------------------------------------------------
 
-static inline std::string to_string(bool b) noexcept { return b ? "true" : "false"; }
-static inline std::string to_string(int i) noexcept { return std::to_string(i); }
-static inline std::string to_string(float f) noexcept { return "float(" + std::to_string(f) + ")"; }
+static std::string to_string(bool const b) { return b ? "true" : "false"; }
+static std::string to_string(int const i) { return std::to_string(i); }
+static std::string to_string(float const f) { return "float(" + std::to_string(f) + ")"; }
 
 static void logCompilationError(io::ostream& out, ShaderStage shaderType, const char* name,
         GLuint shaderId, CString const& sourceCode) noexcept;
 static void logProgramLinkError(io::ostream& out, char const* name, GLuint program) noexcept;
 
-static void process_GOOGLE_cpp_style_line_directive(OpenGLContext& context, char* source,
+static void process_GOOGLE_cpp_style_line_directive(OpenGLContext const& context, char* source,
         size_t len) noexcept;
-static void process_OVR_multiview2(OpenGLContext& context, int32_t eyeCount, char* source,
+static void process_OVR_multiview2(OpenGLContext const& context, int32_t eyeCount, char* source,
         size_t len) noexcept;
 static std::string_view process_ARB_shading_language_packing(OpenGLContext& context) noexcept;
 static std::array<std::string_view, 3> splitShaderSource(std::string_view source) noexcept;
@@ -72,20 +77,15 @@ static std::array<std::string_view, 3> splitShaderSource(std::string_view source
 // ------------------------------------------------------------------------------------------------
 
 struct ShaderCompilerService::OpenGLProgramToken : ProgramToken {
-    struct ProgramData {
-        GLuint program{};
-        shaders_t shaders{};
-    };
-
     ~OpenGLProgramToken() override;
 
-    OpenGLProgramToken(ShaderCompilerService& compiler, utils::CString const& name) noexcept
-            : compiler(compiler), name(name) {
+    OpenGLProgramToken(ShaderCompilerService& compiler, CString const& name) noexcept
+            : compiler(compiler), name(name), handle(compiler.issueCallbackHandle()) {
     }
 
     ShaderCompilerService& compiler;
-    utils::CString const& name;
-    utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> attributes;
+    CString const& name;
+    FixedCapacityVector<std::pair<CString, uint8_t>> attributes;
     shaders_source_t shaderSourceCode;
     void* user = nullptr;
     struct {
@@ -93,48 +93,34 @@ struct ShaderCompilerService::OpenGLProgramToken : ProgramToken {
         GLuint program = 0;
     } gl; // 12 bytes
 
-
-    // Sets the programData, typically from the compiler thread, and signal the main thread.
-    // This is similar to std::promise::set_value.
-    void set(ProgramData const& data) noexcept {
+    // Used in THREAD_POOL mode. The job from ThreadPool should call this when the token is ready to
+    // be used. It sends a signal to the engine thread being blocked upon the `wait` call, so that
+    // the engine thread resumes its processing with the token.
+    void signal() noexcept {
         std::unique_lock const l(lock);
-        programData = data;
         signaled = true;
         cond.notify_one();
     }
 
-    // Get the programBinary, wait if necessary.
-    // This is similar to std::future::get
-    ProgramData const& get() const noexcept {
-        std::unique_lock l(lock);
-        cond.wait(l, [this](){ return signaled; });
-        return programData;
-    }
-
+    // Used in THREAD_POOL mode. The engine thread should call this before accessing token's fields.
+    // This may block until the token is ready to be used.
     void wait() const noexcept {
         std::unique_lock l(lock);
-        cond.wait(l, [this](){ return signaled; });
-    }
-
-    // Checks if the programBinary is ready.
-    // This is similar to std::future::wait_for(0s)
-    bool isReady() const noexcept {
-        std::unique_lock l(lock);
-        using namespace std::chrono_literals;
-        return cond.wait_for(l, 0s, [this](){ return signaled; });
+        cond.wait(l, [this] { return signaled; });
     }
 
     CallbackManager::Handle handle{};
     BlobCacheKey key;
-    mutable utils::Mutex lock;
-    mutable utils::Condition cond;
-    ProgramData programData;
-    bool signaled = false;
 
-    bool canceled = false; // not part of the signaling
+    // Used for the `THREAD_POOL` mode.
+    mutable Mutex lock;
+    mutable Condition cond;
+    bool signaled = false;
 };
 
-ShaderCompilerService::OpenGLProgramToken::~OpenGLProgramToken() = default;
+ShaderCompilerService::OpenGLProgramToken::~OpenGLProgramToken() {
+    compiler.submitCallbackHandle(handle);
+}
 
 /* static */ void ShaderCompilerService::setUserData(const program_token_t& token,
         void* user) noexcept {
@@ -189,9 +175,9 @@ void ShaderCompilerService::init() noexcept {
     if (mMode == Mode::THREAD_POOL) {
         // - on Adreno there is a single compiler object. We can't use a pool > 1
         //   also glProgramBinary blocks if other threads are compiling.
-        // - on Mali shader compilation can be multi-threaded, but program linking happens on
+        // - on Mali shader compilation can be multithreaded, but program linking happens on
         //   a single service thread, so we don't bother using more than one thread either.
-        // - on PowerVR shader compilation and linking can be multi-threaded.
+        // - on PowerVR shader compilation and linking can be multithreaded.
         //   How many threads should we use?
         // - on macOS (M1 MacBook Pro/Ventura) there is global lock around all GL APIs when using
         //   a shared context, so parallel shader compilation yields no benefit.
@@ -220,7 +206,7 @@ void ShaderCompilerService::init() noexcept {
 
         mShaderCompilerThreadCount = poolSize;
         mCompilerThreadPool.init(mShaderCompilerThreadCount,
-                [&platform = mDriver.mPlatform, priority]() {
+                [&platform = mDriver.mPlatform, priority] {
                     // give the thread a name
                     JobSystem::setThreadName("CompilerThreadPool");
                     // run at a slightly lower priority than other filament threads
@@ -228,7 +214,7 @@ void ShaderCompilerService::init() noexcept {
                     // create a gl context current to this thread
                     platform.createContext(true);
                 },
-                [&platform = mDriver.mPlatform]() {
+                [&platform = mDriver.mPlatform] {
                     // release context and thread state
                     platform.releaseContext();
                 });
@@ -249,119 +235,65 @@ void ShaderCompilerService::terminate() noexcept {
 }
 
 ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
-        utils::CString const& name, Program&& program) {
+        CString const& name, Program&& program) {
     auto& gl = mDriver.getContext();
 
+    // Create a token. A callback condition (handle) is internally created upon token creation.
     auto token = std::make_shared<OpenGLProgramToken>(*this, name);
     if (UTILS_UNLIKELY(gl.isES2())) {
         token->attributes = std::move(program.getAttributes());
     }
 
+    // Try retrieving the cached program blob if available.
     token->gl.program = mBlobCache.retrieve(&token->key, mDriver.mPlatform, program);
     if (token->gl.program) {
         return token;
     }
 
-    token->handle = mCallbackManager.get();
-
+    // Initiate program compilation.
     CompilerPriorityQueue const priorityQueue = program.getPriorityQueue();
     switch (mMode) {
         case Mode::THREAD_POOL: {
-            // queue a compile job
             mCompilerThreadPool.queue(priorityQueue, token,
                     [this, &gl, program = std::move(program), token]() mutable {
-                        // compile the shaders
-                        shaders_t shaders{};
-                        compileShaders(gl,
-                                std::move(program.getShadersSource()),
-                                program.getSpecializationConstants(),
-                                program.isMultiview(),
-                                shaders,
-                                token->shaderSourceCode);
-
-                        // link the program
-                        GLuint const glProgram = linkProgram(gl, shaders, token->attributes);
-
-                        OpenGLProgramToken::ProgramData programData;
-                        programData.shaders = shaders;
-
-                        // We need to query the link status here to guarantee that the
-                        // program is compiled and linked now (we don't want this to be
-                        // deferred to later). We don't care about the result at this point.
-                        GLint status = GL_FALSE;
-                        glGetProgramiv(glProgram, GL_LINK_STATUS, &status);
-                        programData.program = glProgram;
-
-                        // we don't need to check for success here, it'll be done on the
-                        // main thread side.
-                        token->set(programData);
-
-                        mCallbackManager.put(token->handle);
-
-                        // caching must be the last thing we do
-                        if (token->key && status == GL_TRUE) {
-                            // Attempt to cache. This calls glGetProgramBinary.
-                            mBlobCache.insert(mDriver.mPlatform, token->key, glProgram);
-                        }
+                        compileShaders(gl, std::move(program.getShadersSource()),
+                                program.getSpecializationConstants(), program.isMultiview(), token);
+                        linkProgram(gl, token);
+                        // Now `token->gl.program` must be populated, so we signal the completion
+                        // of the linking. We don't need to check the result of the program here
+                        // because it'll be done in the engine thread.
+                        token->signal();
+                        // We try caching the program blob after sending the signal. This allows us
+                        // to unblock the engine thread as soon as the token is ready while
+                        // performing an expensive caching operation still in the pool.
+                        tryCachingProgram(mBlobCache, mDriver.mPlatform, token);
                     });
             break;
         }
 
         case Mode::SYNCHRONOUS:
         case Mode::ASYNCHRONOUS: {
-            // this cannot fail because we check compilation status after linking the program
-            // shaders[] is filled with id of shader stages present.
-            compileShaders(gl,
-                    std::move(program.getShadersSource()),
-                    program.getSpecializationConstants(),
-                    program.isMultiview(),
-                    token->gl.shaders,
-                    token->shaderSourceCode);
+            compileShaders(gl, std::move(program.getShadersSource()),
+                    program.getSpecializationConstants(), program.isMultiview(), token);
 
             runAtNextTick(priorityQueue, token, [this, token](Job const&) {
                 assert_invariant(mMode != Mode::THREAD_POOL);
                 if (mMode == Mode::ASYNCHRONOUS) {
-                    // don't attempt to link this program if all shaders are not done compiling
-                    GLint status;
+                    // Check link completion if link was initiated.
                     if (token->gl.program) {
-                        glGetProgramiv(token->gl.program, GL_COMPLETION_STATUS, &status);
-                        if (status == GL_FALSE) {
-                            return false;
-                        }
-                    } else {
-                        for (auto shader: token->gl.shaders) {
-                            if (shader) {
-                                glGetShaderiv(shader, GL_COMPLETION_STATUS, &status);
-                                if (status == GL_FALSE) {
-                                    return false;
-                                }
-                            }
-                        }
+                        return isLinkCompleted(token);
                     }
-                }
-
-                if (!token->gl.program) {
-                    // link the program, this also cannot fail because status is checked later.
-                    token->gl.program = linkProgram(mDriver.getContext(),
-                            token->gl.shaders, token->attributes);
-                    if (mMode == Mode::ASYNCHRONOUS) {
-                        // wait until the link finishes...
+                    // Link hasn't been initiated, then check compile completion.
+                    if (!isCompileCompleted(token)) {
                         return false;
                     }
                 }
-
-                assert_invariant(token->gl.program);
-
-                mCallbackManager.put(token->handle);
-
-                if (token->key) {
-                    // TODO: technically we don't have to cache right now. Is it advantageous to
-                    //       do this later, maybe depending on CPU usage?
-                    // attempt to cache if we don't have a thread pool (otherwise it's done
-                    // by the pool).
-                    mBlobCache.insert(mDriver.mPlatform, token->key, token->gl.program);
+                if (!token->gl.program) {
+                    linkProgram(mDriver.getContext(), token);
+                    if (mMode == Mode::ASYNCHRONOUS) {
+                        return false;// Wait until the link finishes.
+                    }
                 }
-
                 return true;
             });
             break;
@@ -375,7 +307,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
     return token;
 }
 
-GLuint ShaderCompilerService::getProgram(ShaderCompilerService::program_token_t& token) {
+GLuint ShaderCompilerService::getProgram(program_token_t& token) {
     GLuint const program = initialize(token);
     assert_invariant(token == nullptr);
 #if !FILAMENT_ENABLE_MATDBG
@@ -384,43 +316,29 @@ GLuint ShaderCompilerService::getProgram(ShaderCompilerService::program_token_t&
     return program;
 }
 
+/*
+ * Cancel program compilation. This function is responsible for cleaning up the ongoing
+ * compilation & link process. If the process is already completed by calling `initialize(token)`,
+ * this function is not called.
+ */
 /* static */ void ShaderCompilerService::terminate(program_token_t& token) {
-    assert_invariant(token);
 
-    token->canceled = true;
-
-    bool const isTickOpCanceled = token->compiler.cancelTickOp(token);
+    assert_invariant(token);// This function should be called when the token is still alive.
 
     if (token->compiler.mMode == Mode::THREAD_POOL) {
-        auto job = token->compiler.mCompilerThreadPool.dequeue(token);
+        auto const job = token->compiler.mCompilerThreadPool.dequeue(token);
         if (!job) {
-            // The job is being executed right now. We need to wait for it to finish to avoid a
-            // race.
+            // It's likely that the job was already completed. But it may be still being
+            // executed at this moment. Just try waiting for it to avoid a race.
             token->wait();
-        } else {
-            // The job has not been executed, but we still need to inform the callback manager in
-            // order for future callbacks to be successfully called.
-            token->compiler.mCallbackManager.put(token->handle);
-        }
-    } else if (isTickOpCanceled) {
-        // Since the tick op was canceled, we need to .put the token here.
-        token->compiler.mCallbackManager.put(token->handle);
-    }
-
-    for (GLuint& shader: token->gl.shaders) {
-        if (shader) {
-            if (token->gl.program) {
-                glDetachShader(token->gl.program, shader);
-            }
-            glDeleteShader(shader);
-            shader = 0;
         }
     }
-    if (token->gl.program) {
-        glDeleteProgram(token->gl.program);
-    }
 
-    token.reset();
+    cleanupProgramAndShaders(token);
+
+    // Cleanup the token.
+    token->compiler.cancelTickOp(token);
+    token = nullptr;// This will submit a callback condition (handle) to the callback manager.
 }
 
 void ShaderCompilerService::tick() {
@@ -430,8 +348,16 @@ void ShaderCompilerService::tick() {
     }
 }
 
+CallbackManager::Handle ShaderCompilerService::issueCallbackHandle() const noexcept {
+    return mCallbackManager.get();
+}
+
+void ShaderCompilerService::submitCallbackHandle(CallbackManager::Handle handle) noexcept {
+    mCallbackManager.put(handle);
+}
+
 void ShaderCompilerService::notifyWhenAllProgramsAreReady(
-        CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
+        CallbackHandler* handler, CallbackHandler::Callback const callback, void* user) {
     if (callback) {
         mCallbackManager.setCallback(handler, callback, user);
     }
@@ -439,117 +365,137 @@ void ShaderCompilerService::notifyWhenAllProgramsAreReady(
 
 // ------------------------------------------------------------------------------------------------
 
-/* static */ void ShaderCompilerService::getProgramFromCompilerPool(
-        program_token_t& token) noexcept {
-    OpenGLProgramToken::ProgramData const& programData{ token->get() };
-    if (!token->canceled) {
-        token->gl.shaders = programData.shaders;
-        token->gl.program = programData.program;
-    }
-}
+GLuint ShaderCompilerService::initialize(program_token_t& token) {
 
-GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
     SYSTRACE_CALL();
-    if (!token->gl.program) {
-        switch (mMode) {
-            case Mode::THREAD_POOL: {
-                // we need this program right now, remove it from the queue
-                auto job = mCompilerThreadPool.dequeue(token);
-                if (job) {
-                    // if we were able to remove it, we execute the job now, otherwise it means
-                    // it's being executed right now.
-                    job();
-                }
 
-                if (!token->canceled) {
-                    token->compiler.cancelTickOp(token);
-                }
+    assert_invariant(token);// This function should be called when the token is still alive.
 
-                // Block until we get the program from the pool. Generally this wouldn't block
-                // because we just compiled the program above, when executing job.
-                ShaderCompilerService::getProgramFromCompilerPool(token);
-                break;
-            }
-
-            case Mode::ASYNCHRONOUS: {
-                // we force the program link -- which might stall, either here or below in
-                // checkProgramStatus(), but we don't have a choice, we need to use the program now.
-                token->compiler.cancelTickOp(token);
-
-                token->gl.program =
-                        linkProgram(mDriver.getContext(), token->gl.shaders, token->attributes);
-
-                assert_invariant(token->gl.program);
-
-                mCallbackManager.put(token->handle);
-
-                if (token->key) {
-                    mBlobCache.insert(mDriver.mPlatform, token->key, token->gl.program);
-                }
-                break;
-            }
-
-            case Mode::SYNCHRONOUS: {
-                // if we don't have a program yet, block until we get it.
-                tick();
-                break;
-            }
-
-            case Mode::UNDEFINED: {
-                assert_invariant(false);
-            }
-        }
-    }
-
-    // by this point we must have a GL program
+    ensureTokenIsReady(token);
     assert_invariant(token->gl.program);
 
-    GLuint program = 0;
+    // Check status of program linking. If it failed, errors will be logged.
+    bool const linked = checkLinkStatusAndCleanupShaders(token);
 
-    // check status of program linking and shader compilation, logs error and free all resources
-    // in case of error.
-    bool const success = checkProgramStatus(token);
-
-    // Unless we have matdbg, we panic if a program is invalid. Otherwise, we'd get a UB.
-    // The compilation error has been logged to log.e by this point.
-    FILAMENT_CHECK_POSTCONDITION(FILAMENT_ENABLE_MATDBG || success)
+    // We panic if it failed to create the program.
+    FILAMENT_CHECK_POSTCONDITION(linked)
             << "OpenGL program " << token->name.c_str_safe() << " failed to link or compile";
 
-    if (UTILS_LIKELY(success)) {
-        program = token->gl.program;
-        // no need to keep the shaders around
-        UTILS_NOUNROLL
-        for (GLuint& shader: token->gl.shaders) {
-            if (shader) {
-                glDetachShader(program, shader);
-                glDeleteShader(shader);
-                shader = 0;
-            }
-        }
+    // The program is successfully created. Try caching the program blob. In the THREAD_POOL mode,
+    // caching is performed in the pool.
+    if (mMode != Mode::THREAD_POOL) {
+        tryCachingProgram(mBlobCache, mDriver.mPlatform, token);
     }
 
-    // and destroy all temporary init data
-    token = nullptr;
+    GLuint const program = token->gl.program;
+
+    // Cleanup the token.
+    token->compiler.cancelTickOp(token);
+    token = nullptr;// This will submit a callback condition (handle) to the callback manager.
 
     return program;
 }
 
+void ShaderCompilerService::ensureTokenIsReady(program_token_t const& token) {
+    if (token->gl.program) {
+        return;// It's ready.
+    }
+
+    switch (mMode) {
+        case Mode::THREAD_POOL: {
+            // We need this program right now, make sure the job is finished.
+            if (auto job = mCompilerThreadPool.dequeue(token)) {
+                job();// The job hasn't started yet, so execute it now.
+            }
+
+            // This may block if the job was already taken by a thread ahead of the `dequeue`
+            // above and currently being executed. Otherwise, the job must have already been
+            // completed by this point from either the code above or the other thread.
+            token->wait();
+            break;
+        }
+
+        case Mode::ASYNCHRONOUS: {
+            // Technically the shader compilation may not have finished yet. To deal with the case,
+            // ideally, we should wait here until the compilation is finished. However, for now, we
+            // just log warnings here instead of repeatedly checking compile status. If this turns
+            // out to be a real issue later, we would need to consider doing the canonical way.
+            if (!isCompileCompleted(token)) {
+                slog.w << "Shader compilation for OpenGL program " << token->name.c_str_safe()
+                       << " is not completed yet. The following program link may not succeed.";
+            }
+
+            linkProgram(mDriver.getContext(), token);
+            break;
+        }
+
+        case Mode::SYNCHRONOUS: {
+            // We must not have called the TickOp yet until now. Call now to have
+            // `token->gl.program` ready to use.
+            tick();
+            break;
+        }
+
+        case Mode::UNDEFINED: {
+            assert_invariant(false);
+        }
+    }
+}
+
+// ------------------------------------------------------------------------------------------------
+
+void ShaderCompilerService::runAtNextTick(CompilerPriorityQueue priority,
+        program_token_t const& token, Job job) noexcept {
+    // insert items in order of priority and at the end of the range
+    auto& ops = mRunAtNextTickOps;
+    auto const pos = std::lower_bound(ops.begin(), ops.end(), priority,
+            [](ContainerType const& lhs, CompilerPriorityQueue const priorityQueue) {
+                return std::get<0>(lhs) < priorityQueue;
+            });
+    ops.emplace(pos, priority, token, std::move(job));
+
+    SYSTRACE_CONTEXT();
+    SYSTRACE_VALUE32("ShaderCompilerService Jobs", mRunAtNextTickOps.size());
+}
+
+bool ShaderCompilerService::cancelTickOp(program_token_t const& token) noexcept {
+    // We do a linear search here, but this is rare, and we know the list is pretty small.
+    auto& ops = mRunAtNextTickOps;
+    auto const pos = std::find_if(ops.begin(), ops.end(), [&](const auto& item) {
+        return std::get<1>(item) == token;
+    });
+    if (pos != ops.end()) {
+        ops.erase(pos);
+        return true;
+    }
+    SYSTRACE_CONTEXT();
+    SYSTRACE_VALUE32("ShaderCompilerService Jobs", ops.size());
+    return false;
+}
+
+void ShaderCompilerService::executeTickOps() noexcept {
+    auto& ops = mRunAtNextTickOps;
+    auto it = ops.begin();
+    while (it != ops.end()) {
+        Job const& job = std::get<2>(*it);
+        bool const remove = job.fn(job);
+        if (remove) {
+            it = ops.erase(it);
+        } else {
+            ++it;
+        }
+    }
+    SYSTRACE_CONTEXT();
+    SYSTRACE_VALUE32("ShaderCompilerService Jobs", ops.size());
+}
 
-/*
- * Compile shaders in the ShaderSource. This cannot fail because compilation failures are not
- * checked until after the program is linked.
- * This always returns the GL shader IDs or zero a shader stage is not present.
- */
 /* static */ void ShaderCompilerService::compileShaders(OpenGLContext& context,
         Program::ShaderSource shadersSource,
-        utils::FixedCapacityVector<Program::SpecializationConstant> const& specializationConstants,
-        bool multiview,
-        shaders_t& outShaders,
-        UTILS_UNUSED_IN_RELEASE shaders_source_t& outShaderSourceCode) noexcept {
-
+        FixedCapacityVector<Program::SpecializationConstant> const& specializationConstants,
+        bool multiview, program_token_t const& token) noexcept {
     SYSTRACE_CALL();
 
-    auto appendSpecConstantString = +[](std::string& s, Program::SpecializationConstant const& sc) {
+    auto const appendSpecConstantString = +[](std::string& s, Program::SpecializationConstant const& sc) {
         s += "#define SPIRV_CROSS_CONSTANT_ID_" + std::to_string(sc.id) + ' ';
         s += std::visit([](auto&& arg) { return to_string(arg); }, sc.value);
         s += '\n';
@@ -558,7 +504,7 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
 
     std::string specializationConstantString;
     int32_t numViews = 2;
-    for (auto const& sc : specializationConstants) {
+    for (auto const& sc: specializationConstants) {
         appendSpecConstantString(specializationConstantString, sc);
         if (sc.id == 8) {
             // This constant must match
@@ -596,7 +542,7 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
         if (UTILS_LIKELY(!shadersSource[i].empty())) {
             Program::ShaderBlob& shader = shadersSource[i];
             char* shader_src = reinterpret_cast<char*>(shader.data());
-            size_t shader_len = shader.size();
+            size_t const shader_len = shader.size();
 
             // remove GOOGLE_cpp_style_line_directive
             process_GOOGLE_cpp_style_line_directive(context, shader_src, shader_len);
@@ -619,178 +565,191 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
             }
 
             std::array<std::string_view, 5> sources = {
-                version,
-                prolog,
-                specializationConstantString,
-                packingFunctions,
-                { body.data(), body.size() - 1 }  // null-terminated
+                version, prolog, specializationConstantString, packingFunctions,
+                { body.data(), body.size() - 1 }// null-terminated
             };
 
             // Some of the sources may be zero-length. Remove them as to avoid passing lengths of
             // zero to glShaderSource(). glShaderSource should work with lengths of zero, but some
             // drivers instead interpret zero as a sentinel for a null-terminated string.
-            auto partitionPoint = std::stable_partition(
-                    sources.begin(), sources.end(), [](std::string_view s) { return !s.empty(); });
-            size_t count = std::distance(sources.begin(), partitionPoint);
+            auto const partitionPoint = std::stable_partition(sources.begin(), sources.end(),
+                    [](std::string_view s) { return !s.empty(); });
+            size_t const count = std::distance(sources.begin(), partitionPoint);
 
             std::array<const char*, 5> shaderStrings;
             std::array<GLint, 5> lengths;
-            for (size_t i = 0; i < count; i++) {
-                shaderStrings[i] = sources[i].data();
-                lengths[i] = sources[i].size();
+            for (size_t j = 0; j < count; j++) {
+                shaderStrings[j] = sources[j].data();
+                lengths[j] = GLint(sources[j].size());
             }
 
             GLuint const shaderId = glCreateShader(glShaderType);
-            glShaderSource(shaderId, count, shaderStrings.data(), lengths.data());
-
+            glShaderSource(shaderId, GLsizei(count), shaderStrings.data(), lengths.data());
             glCompileShader(shaderId);
-
 #ifndef NDEBUG
             // for debugging we return the original shader source (without the modifications we
             // made here), otherwise the line numbers wouldn't match.
-            outShaderSourceCode[i] = { shader_src, shader_len };
+            token->shaderSourceCode[i] = { shader_src, shader_len };
 #endif
-
-            outShaders[i] = shaderId;
+            token->gl.shaders[i] = shaderId;
         }
     }
 }
 
-/*
- * Create a program from the given shader IDs and links it. This cannot fail because errors
- * are checked later. This always returns a valid GL program ID (which doesn't mean the
- * program itself is valid).
- */
-/* static */ GLuint ShaderCompilerService::linkProgram(OpenGLContext& context,
-        shaders_t const& shaders,
-        utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& attributes) noexcept {
+/* static */ bool ShaderCompilerService::isCompileCompleted(program_token_t const& token) noexcept {
+    GLenum param = GL_COMPLETION_STATUS;
+    if (UTILS_UNLIKELY(token->compiler.mMode != Mode::ASYNCHRONOUS)) {
+        param = GL_COMPILE_STATUS;
+    }
 
+    for (auto shader: token->gl.shaders) {
+        if (!shader) {
+            continue;
+        }
+        GLint status;
+        glGetShaderiv(shader, param, &status);
+        if (status == GL_FALSE) {
+            return false;
+        }
+    }
+    return true;
+}
+
+/* static */ void ShaderCompilerService::checkCompileStatus(program_token_t const& token) noexcept {
     SYSTRACE_CALL();
 
+    UTILS_NOUNROLL
+    for (size_t i = 0; i < Program::SHADER_TYPE_COUNT; i++) {
+        const GLuint shader = token->gl.shaders[i];
+        if (!shader) {
+            continue;// We're not using this shader stage.
+        }
+        // GL_COMPILE_STATUS may block until the compilation is completed.
+        GLint status;
+        glGetShaderiv(shader, GL_COMPILE_STATUS, &status);
+        if (UTILS_LIKELY(status == GL_TRUE)) {
+            continue;// Succeeded in compilation.
+        }
+        // Something went wrong. Log the error message.
+        const ShaderStage type = static_cast<ShaderStage>(i);
+        logCompilationError(slog.e, type, token->name.c_str_safe(), shader,
+                token->shaderSourceCode[i]);
+    }
+}
+
+/* static */ void ShaderCompilerService::linkProgram(OpenGLContext const& context,
+        program_token_t const& token) noexcept {
+    SYSTRACE_CALL();
+
+    // Shader compilation should be completed by now. Check the status and log errors on failure.
+    checkCompileStatus(token);
+
+    // Link program
     GLuint const program = glCreateProgram();
-    for (auto shader : shaders) {
+    for (auto const shader: token->gl.shaders) {
         if (shader) {
             glAttachShader(program, shader);
         }
     }
-
     if (UTILS_UNLIKELY(context.isES2())) {
-        for (auto const& [ name, loc ] : attributes) {
+        for (auto const& [name, loc]: token->attributes) {
             glBindAttribLocation(program, loc, name.c_str());
         }
     }
-
     glLinkProgram(program);
-
-    return program;
+    token->gl.program = program;
 }
 
-// ------------------------------------------------------------------------------------------------
-
-void ShaderCompilerService::runAtNextTick(CompilerPriorityQueue priority,
-        const program_token_t& token, Job job) noexcept {
-    // insert items in order of priority and at the end of the range
-    auto& ops = mRunAtNextTickOps;
-    auto const pos = std::lower_bound(ops.begin(), ops.end(), priority,
-            [](ContainerType const& lhs, CompilerPriorityQueue priorityQueue) {
-                return std::get<0>(lhs) < priorityQueue;
-            });
-    ops.emplace(pos, priority, token, std::move(job));
-
-    SYSTRACE_CONTEXT();
-    SYSTRACE_VALUE32("ShaderCompilerService Jobs", mRunAtNextTickOps.size());
-}
-
-bool ShaderCompilerService::cancelTickOp(program_token_t token) noexcept {
-    // We do a linear search here, but this is rare, and we know the list is pretty small.
-    auto& ops = mRunAtNextTickOps;
-    auto pos = std::find_if(ops.begin(), ops.end(), [&](const auto& item) {
-        return std::get<1>(item) == token;
-    });
-    if (pos != ops.end()) {
-        ops.erase(pos);
-        return true;
-    }
-    SYSTRACE_CONTEXT();
-    SYSTRACE_VALUE32("ShaderCompilerService Jobs", ops.size());
-    return false;
-}
-
-void ShaderCompilerService::executeTickOps() noexcept {
-    auto& ops = mRunAtNextTickOps;
-    auto it = ops.begin();
-    while (it != ops.end()) {
-        Job const& job = std::get<2>(*it);
-        bool const remove = job.fn(job);
-        if (remove) {
-            it = ops.erase(it);
-        } else {
-            ++it;
-        }
-    }
-    SYSTRACE_CONTEXT();
-    SYSTRACE_VALUE32("ShaderCompilerService Jobs", ops.size());
-}
-
-// ------------------------------------------------------------------------------------------------
-
-/*
- * Checks a program link status and logs errors and frees resources on failure.
- * Returns true on success.
- */
-/* static */ bool ShaderCompilerService::checkProgramStatus(program_token_t const& token) noexcept {
-
-    SYSTRACE_CALL();
-
+/* static */ bool ShaderCompilerService::isLinkCompleted(program_token_t const& token) noexcept {
     assert_invariant(token->gl.program);
 
-    GLint status;
-    glGetProgramiv(token->gl.program, GL_LINK_STATUS, &status);
-    if (UTILS_LIKELY(status == GL_TRUE)) {
-        return true;
+    GLenum param = GL_COMPLETION_STATUS;
+    if (UTILS_UNLIKELY(token->compiler.mMode != Mode::ASYNCHRONOUS)) {
+        param = GL_LINK_STATUS;
     }
 
-    // only if the link fails, we check the compilation status
+    GLint status;
+    glGetProgramiv(token->gl.program, param, &status);
+    return (status == GL_TRUE);
+}
+
+/* static */ bool ShaderCompilerService::checkLinkStatusAndCleanupShaders(
+        program_token_t const& token) noexcept {
+    SYSTRACE_CALL();
+    assert_invariant(token->gl.program);
+
+    bool linked = true;
+    GLint status;
+    // GL_LINK_STATUS may block until the link is completed.
+    glGetProgramiv(token->gl.program, GL_LINK_STATUS, &status);
+    if (UTILS_UNLIKELY(status != GL_TRUE)) {
+        // Something went wrong. Log the error message.
+        logProgramLinkError(slog.e, token->name.c_str_safe(), token->gl.program);
+        linked = false;
+    }
+    // No need to keep the shaders around regardless of the result of the program linking.
     UTILS_NOUNROLL
-    for (size_t i = 0; i < Program::SHADER_TYPE_COUNT; i++) {
-        const ShaderStage type = static_cast<ShaderStage>(i);
-        const GLuint shader = token->gl.shaders[i];
+    for (GLuint& shader: token->gl.shaders) {
         if (shader) {
-            glGetShaderiv(shader, GL_COMPILE_STATUS, &status);
-            if (status != GL_TRUE) {
-                logCompilationError(slog.e, type,
-                        token->name.c_str_safe(), shader, token->shaderSourceCode[i]);
-            }
             glDetachShader(token->gl.program, shader);
             glDeleteShader(shader);
-            token->gl.shaders[i] = 0;
+            shader = 0;
         }
     }
-    // log the link error as well
-    logProgramLinkError(slog.e, token->name.c_str_safe(), token->gl.program);
-    glDeleteProgram(token->gl.program);
-    token->gl.program = 0;
-    return false;
+    return linked;
+}
+
+/* static */ void ShaderCompilerService::tryCachingProgram(OpenGLBlobCache& cache,
+        OpenGLPlatform& platform, program_token_t const& token) noexcept {
+    if (!token->key || !token->gl.program) {
+        return; // Invalid params
+    }
+    GLint status = GL_FALSE;
+    glGetProgramiv(token->gl.program, GL_LINK_STATUS, &status);
+    if (status == GL_FALSE) {
+        return;// Link failure
+    }
+
+    cache.insert(platform, token->key, token->gl.program);
+}
+
+/* static */ void ShaderCompilerService::cleanupProgramAndShaders(
+        program_token_t const& token) noexcept {
+    for (GLuint& shader: token->gl.shaders) {
+        if (!shader) {
+            continue;
+        }
+        if (token->gl.program) {
+            glDetachShader(token->gl.program, shader);
+        }
+        glDeleteShader(shader);
+        shader = 0;
+    }
+    if (token->gl.program) {
+        glDeleteProgram(token->gl.program);
+        token->gl.program = 0;
+    }
 }
 
 // ------------------------------------------------------------------------------------------------
 
 UTILS_NOINLINE
 /* static */ void logCompilationError(io::ostream& out, ShaderStage shaderType, const char* name,
-        GLuint shaderId, UTILS_UNUSED_IN_RELEASE CString const& sourceCode) noexcept {
+        GLuint const shaderId, UTILS_UNUSED_IN_RELEASE CString const& sourceCode) noexcept {
 
-    auto to_string = [](ShaderStage type) -> const char* {
-        switch (type) {
-            case ShaderStage::VERTEX:
-                return "vertex";
-            case ShaderStage::FRAGMENT:
-                return "fragment";
-            case ShaderStage::COMPUTE:
-                return "compute";
-        }
-    };
+    { // scope for the temporary string storage
+        auto to_string = [](ShaderStage type) -> const char* {
+            switch (type) {
+                case ShaderStage::VERTEX:
+                    return "vertex";
+                case ShaderStage::FRAGMENT:
+                    return "fragment";
+                case ShaderStage::COMPUTE:
+                    return "compute";
+            }
+            return "unknown";
+        };
 
-    {// scope for the temporary string storage
         GLint length = 0;
         glGetShaderiv(shaderId, GL_INFO_LOG_LENGTH, &length);
 
@@ -837,8 +796,8 @@ UTILS_NOINLINE
 
 // If usages of the Google-style line directive are present, remove them, as some
 // drivers don't allow the quotation marks. This source modification happens in-place.
-/* static */ void process_GOOGLE_cpp_style_line_directive(OpenGLContext& context, char* source,
-        size_t len) noexcept {
+/* static */ void process_GOOGLE_cpp_style_line_directive(OpenGLContext const& context,
+        char* source, size_t len) noexcept {
     if (!context.ext.GOOGLE_cpp_style_line_directive) {
         if (UTILS_UNLIKELY(requestsGoogleLineDirectivesExtension({ source, len }))) {
             removeGoogleLineDirectives(source, len);// length is unaffected
@@ -850,13 +809,13 @@ UTILS_NOINLINE
 // necessary for OpenGL because OpenGL relies on the number specified in shader files to determine
 // the number of views, which is assumed as a single digit, for multiview.
 // This source modification happens in-place.
-/* static */ void process_OVR_multiview2(OpenGLContext& context, int32_t eyeCount, char* source,
-        size_t len) noexcept {
+/* static */ void process_OVR_multiview2(OpenGLContext const& context, int32_t const eyeCount,
+    char* source, size_t const len) noexcept {
     // We don't use regular expression in favor of performance.
     if (context.ext.OVR_multiview2) {
         const std::string_view shader{ source, len };
-        const std::string_view layout = "layout";
-        const std::string_view num_views = "num_views";
+        constexpr std::string_view layout = "layout";
+        constexpr std::string_view num_views = "num_views";
         size_t found = 0;
         while (true) {
             found = shader.find(layout, found);
@@ -1011,20 +970,20 @@ mediump vec4 unpackSnorm4x8(highp uint v) {
 // - extensions
 // - everything else
 /* static */ std::array<std::string_view, 3> splitShaderSource(std::string_view source) noexcept {
-    auto version_start = source.find("#version");
+    auto const version_start = source.find("#version");
     assert_invariant(version_start != std::string_view::npos);
 
-    auto version_eol = source.find('\n', version_start) + 1;
+    auto const version_eol = source.find('\n', version_start) + 1;
     assert_invariant(version_eol != std::string_view::npos);
 
-    auto prolog_start = version_eol;
+    auto const prolog_start = version_eol;
     auto prolog_eol = source.rfind("\n#extension");// last #extension line
     if (prolog_eol == std::string_view::npos) {
         prolog_eol = prolog_start;
     } else {
         prolog_eol = source.find('\n', prolog_eol + 1) + 1;
     }
-    auto body_start = prolog_eol;
+    auto const body_start = prolog_eol;
 
     std::string_view const version = source.substr(version_start, version_eol - version_start);
     std::string_view const prolog = source.substr(prolog_start, prolog_eol - prolog_start);

filament/backend/src/opengl/ShaderCompilerService.h

@@ -24,23 +24,23 @@
 #include "OpenGLBlobCache.h"
 
 #include <backend/CallbackHandler.h>
+#include <backend/DriverEnums.h>
 #include <backend/Program.h>
 
 #include <utils/CString.h>
 #include <utils/FixedCapacityVector.h>
-#include <utils/Invocable.h>
 #include <utils/JobSystem.h>
 
-#include <atomic>
-#include <condition_variable>
-#include <deque>
+#include <array>
 #include <functional>
 #include <memory>
 #include <mutex>
-#include <thread>
+#include <tuple>
 #include <utility>
 #include <vector>
 
+#include <stdint.h>
+
 namespace filament::backend {
 
 class OpenGLDriver;
@@ -84,6 +84,7 @@ public:
     void tick();
 
     // Destroys a valid token and all associated resources. Used to "cancel" a program compilation.
+    // This function is not called if `initialize(token)` is already invoked.
     static void terminate(program_token_t& token);
 
     // stores a user data pointer in the token
@@ -92,6 +93,12 @@ public:
     // retrieves the user data pointer stored in the token
     static void* getUserData(const program_token_t& token) noexcept;
 
+    // Issue one callback handle.
+    CallbackManager::Handle issueCallbackHandle() const noexcept;
+
+    // Return a callback handle to the callback manager.
+    void submitCallbackHandle(CallbackManager::Handle handle) noexcept;
+
     // call the callback when all active programs are ready
     void notifyWhenAllProgramsAreReady(
             CallbackHandler* handler, CallbackHandler::Callback callback, void* user);
@@ -99,7 +106,7 @@ public:
 private:
     struct Job {
         template<typename FUNC>
-        Job(FUNC&& fn) : fn(std::forward<FUNC>(fn)) {}
+        Job(FUNC&& fn) : fn(std::forward<FUNC>(fn)) {} // NOLINT(*-explicit-constructor)
         Job(std::function<bool(Job const& job)> fn,
                 CallbackHandler* handler, void* user, CallbackHandler::Callback callback)
                 : fn(std::move(fn)), handler(handler), user(user), callback(callback) {
@@ -128,26 +135,49 @@ private:
     using ContainerType = std::tuple<CompilerPriorityQueue, program_token_t, Job>;
     std::vector<ContainerType> mRunAtNextTickOps;
 
-    GLuint initialize(ShaderCompilerService::program_token_t& token) noexcept;
+    GLuint initialize(program_token_t& token);
+    void ensureTokenIsReady(program_token_t const& token);
 
-    static void getProgramFromCompilerPool(program_token_t& token) noexcept;
-
-    static void compileShaders(
-            OpenGLContext& context,
-            Program::ShaderSource shadersSource,
-            utils::FixedCapacityVector<Program::SpecializationConstant> const& specializationConstants,
-            bool multiview, shaders_t& outShaders, shaders_source_t& outShaderSourceCode) noexcept;
-
-    static GLuint linkProgram(OpenGLContext& context, shaders_t const& shaders,
-            utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> const& attributes) noexcept;
-
-    static bool checkProgramStatus(program_token_t const& token) noexcept;
-
-    void runAtNextTick(CompilerPriorityQueue priority,
-            const program_token_t& token, Job job) noexcept;
+    void runAtNextTick(CompilerPriorityQueue priority, program_token_t const& token,
+            Job job) noexcept;
     void executeTickOps() noexcept;
-    bool cancelTickOp(program_token_t token) noexcept;
-    // order of insertion is important
+    bool cancelTickOp(program_token_t const& token) noexcept;
+
+    // Compile shaders with the given `shaderSource`. `gl.shaders` is always populated with valid
+    // shader IDs after this method. But this doesn't necessarily mean the shaders are successfully
+    // compiled. Errors can be checked by calling `checkCompileStatus` later.
+    static void compileShaders(OpenGLContext& context, Program::ShaderSource shadersSource,
+            utils::FixedCapacityVector<Program::SpecializationConstant> const&
+                    specializationConstants,
+            bool multiview, program_token_t const& token) noexcept;
+
+    // Check if the shader compilation is completed. You may want to call this when the extension
+    // `KHR_parallel_shader_compile` is enabled.
+    static bool isCompileCompleted(program_token_t const& token) noexcept;
+
+    // Check compilation status of the shaders and log errors on failure.
+    static void checkCompileStatus(program_token_t const& token) noexcept;
+
+    // Create a program by linking the compiled shaders. `gl.program` is always populated with a
+    // valid program ID after this method. But this doesn't necessarily mean the program is
+    // successfully linked. Errors can be checked by calling `checkLinkStatusAndCleanupShaders`
+    // later.
+    static void linkProgram(OpenGLContext const& context, program_token_t const& token) noexcept;
+
+    // Check if the program link is completed. You may want to call this when the extension
+    // `KHR_parallel_shader_compile` is enabled.
+    static bool isLinkCompleted(program_token_t const& token) noexcept;
+
+    // Check link status of the program and log errors on failure. Return the result of the link.
+    // Also cleanup shaders regardless of the result.
+    static bool checkLinkStatusAndCleanupShaders(program_token_t const& token) noexcept;
+
+    // Try caching the program if we haven't done it yet. Cache it only when the program is valid.
+    static void tryCachingProgram(OpenGLBlobCache& cache, OpenGLPlatform& platform,
+            program_token_t const& token) noexcept;
+
+    // Cleanup GL resources.
+    static void cleanupProgramAndShaders(program_token_t const& token) noexcept;
 };
 
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanConstants.h

@@ -152,12 +152,12 @@ static_assert(FVK_ENABLED(FVK_DEBUG_VALIDATION));
 #elif FVK_ENABLED(FVK_DEBUG_SYSTRACE)
 
     #include <utils/Systrace.h>
-    
+
     #define FVK_SYSTRACE_CONTEXT()      SYSTRACE_CONTEXT()
     #define FVK_SYSTRACE_START(marker)  SYSTRACE_NAME_BEGIN(marker)
     #define FVK_SYSTRACE_END()          SYSTRACE_NAME_END()
-    #define FVK_SYSTRACE_SCOPE()        SYSTRACE_NAME(__func__)
-    #define FVK_PROFILE_MARKER(marker)  FVK_SYSTRACE_SCOPE()
+    #define FVK_SYSTRACE_SCOPE()        SYSTRACE_CALL()
+    #define FVK_PROFILE_MARKER(marker)  SYSTRACE_CALL()
 
 #else
     #define FVK_SYSTRACE_CONTEXT()

filament/backend/src/webgpu/WGPUProgram.cpp

@@ -0,0 +1,150 @@
+/*
+ * Copyright (C) 2025 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "WebGPUHandles.h"
+
+#include "WebGPUConstants.h"
+
+#include "DriverBase.h"
+#include <backend/DriverEnums.h>
+#include <backend/Program.h>
+
+#include <utils/Panic.h>
+#include <utils/ostream.h>
+
+#include <webgpu/webgpu_cpp.h>
+
+#include <sstream>
+#include <string_view>
+#include <vector>
+
+namespace filament::backend {
+
+namespace {
+
+[[nodiscard]] constexpr std::string_view toString(ShaderStage stage) {
+    switch (stage) {
+        case ShaderStage::VERTEX:
+            return "vertex";
+        case ShaderStage::FRAGMENT:
+            return "fragment";
+        case ShaderStage::COMPUTE:
+            return "compute";
+    }
+}
+
+[[nodiscard]] wgpu::ShaderModule createShaderModule(wgpu::Device& device, const char* programName,
+        std::array<utils::FixedCapacityVector<uint8_t>, Program::SHADER_TYPE_COUNT> const&
+                shaderSource,
+        ShaderStage stage) {
+    utils::FixedCapacityVector<uint8_t> const& sourceBytes =
+            shaderSource[static_cast<size_t>(stage)];
+    if (sourceBytes.empty()) {
+        return nullptr;// nothing to compile, the shader was not provided
+    }
+    wgpu::ShaderModuleWGSLDescriptor wgslDescriptor{};
+    wgslDescriptor.code = wgpu::StringView(reinterpret_cast<const char*>(sourceBytes.data()));
+    std::stringstream labelStream;
+    labelStream << programName << " " << toString(stage) << " shader";
+    auto label = labelStream.str();
+    wgpu::ShaderModuleDescriptor descriptor{
+        .nextInChain = &wgslDescriptor,
+        .label = label.data()
+    };
+    wgpu::ShaderModule module = device.CreateShaderModule(&descriptor);
+    FILAMENT_CHECK_POSTCONDITION(module != nullptr) << "Failed to create " << descriptor.label;
+    module.GetCompilationInfo(wgpu::CallbackMode::AllowSpontaneous,
+            [&descriptor](auto const& status, wgpu::CompilationInfo const* info) {
+                switch (status) {
+                    case wgpu::CompilationInfoRequestStatus::CallbackCancelled:
+                        FWGPU_LOGW << "Shader compilation info callback cancelled for "
+                                   << descriptor.label << "?" << utils::io::endl;
+                        return;
+                    case wgpu::CompilationInfoRequestStatus::Success:
+                        break;
+                }
+                if (info != nullptr) {
+                    std::stringstream errorStream;
+                    int errorCount = 0;
+                    for (size_t msgIndex = 0; msgIndex < info->messageCount; msgIndex++) {
+                        wgpu::CompilationMessage const& message = info->messages[msgIndex];
+                        switch (message.type) {
+                            case wgpu::CompilationMessageType::Info:
+                                FWGPU_LOGI << descriptor.label << ": " << message.message
+                                           << " line#:" << message.lineNum
+                                           << " linePos:" << message.linePos
+                                           << " offset:" << message.offset
+                                           << " length:" << message.length << utils::io::endl;
+                                break;
+                            case wgpu::CompilationMessageType::Warning:
+                                FWGPU_LOGW << "Warning compiling " << descriptor.label << ": "
+                                           << message.message << " line#:" << message.lineNum
+                                           << " linePos:" << message.linePos
+                                           << " offset:" << message.offset
+                                           << " length:" << message.length << utils::io::endl;
+                                break;
+                            case wgpu::CompilationMessageType::Error:
+                                errorCount++;
+                                errorStream << "Error " << errorCount << " : "
+                                            << std::string_view(message.message)
+                                            << " line#:" << message.lineNum
+                                            << " linePos:" << message.linePos
+                                            << " offset:" << message.offset
+                                            << " length:" << message.length << "\n";
+                                break;
+                        }
+                    }
+                    FILAMENT_CHECK_POSTCONDITION(errorCount < 1)
+                            << errorCount << " error(s) compiling " << descriptor.label << ":\n"
+                            << errorStream.str();
+                }
+                FWGPU_LOGD << descriptor.label << " compiled successfully" << utils::io::endl;
+            });
+    return module;
+}
+
+std::vector<wgpu::ConstantEntry> convertConstants(
+        utils::FixedCapacityVector<filament::backend::Program::SpecializationConstant> const&
+                constantsInfo) {
+    std::vector<wgpu::ConstantEntry> constants(constantsInfo.size());
+    for (size_t i = 0; i < constantsInfo.size(); i++) {
+        filament::backend::Program::SpecializationConstant const& specConstant = constantsInfo[i];
+        wgpu::ConstantEntry& constantEntry = constants[i];
+        constantEntry.key = wgpu::StringView(std::to_string(specConstant.id));
+        if (auto* v = std::get_if<int32_t>(&specConstant.value)) {
+            constantEntry.value = static_cast<double>(*v);
+        } else if (auto* f = std::get_if<float>(&specConstant.value)) {
+            constantEntry.value = static_cast<double>(*f);
+        } else if (auto* b = std::get_if<bool>(&specConstant.value)) {
+            constantEntry.value = *b ? 0.0 : 1.0;
+        }
+    }
+    return constants;
+}
+
+}// namespace
+
+WGPUProgram::WGPUProgram(wgpu::Device& device, Program& program)
+    : HwProgram(program.getName()),
+      vertexShaderModule(createShaderModule(device, name.c_str_safe(), program.getShadersSource(),
+              ShaderStage::VERTEX)),
+      fragmentShaderModule(createShaderModule(device, name.c_str_safe(), program.getShadersSource(),
+              ShaderStage::FRAGMENT)),
+      computeShaderModule(createShaderModule(device, name.c_str_safe(), program.getShadersSource(),
+              ShaderStage::COMPUTE)),
+      constants(convertConstants(program.getSpecializationConstants())) {}
+
+}// namespace filament::backend

filament/backend/src/webgpu/WebGPUDriver.cpp

@@ -325,6 +325,9 @@ void WebGPUDriver::destroyTexture(Handle<HwTexture> th) {
 }
 
 void WebGPUDriver::destroyProgram(Handle<HwProgram> ph) {
+    if (ph) {
+        destructHandle<WGPUProgram>(ph);
+    }
 }
 
 void WebGPUDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
@@ -369,7 +372,7 @@ Handle<HwTexture> WebGPUDriver::importTextureS() noexcept {
 }
 
 Handle<HwProgram> WebGPUDriver::createProgramS() noexcept {
-    return Handle<HwProgram>((Handle<HwProgram>::HandleId) mNextFakeHandle++);
+    return allocHandle<WGPUProgram>();
 }
 
 Handle<HwFence> WebGPUDriver::createFenceS() noexcept {
@@ -505,7 +508,9 @@ void WebGPUDriver::importTextureR(Handle<HwTexture> th, intptr_t id, SamplerType
 void WebGPUDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph, Handle<HwVertexBuffer> vbh,
         Handle<HwIndexBuffer> ibh, PrimitiveType pt) {}
 
-void WebGPUDriver::createProgramR(Handle<HwProgram> ph, Program&& program) {}
+void WebGPUDriver::createProgramR(Handle<HwProgram> ph, Program&& program) {
+    constructHandle<WGPUProgram>(ph, mDevice, program);
+}
 
 void WebGPUDriver::createDefaultRenderTargetR(Handle<HwRenderTarget> rth, int) {
     assert_invariant(!mDefaultRenderTarget);

filament/backend/src/webgpu/WebGPUHandles.h

@@ -28,9 +28,20 @@
 #include <webgpu/webgpu_cpp.h>
 
 #include <cstdint>
+#include <vector>
 
 namespace filament::backend {
 
+class WGPUProgram final : public HwProgram {
+public:
+    WGPUProgram(wgpu::Device&, Program&);
+
+    wgpu::ShaderModule vertexShaderModule = nullptr;
+    wgpu::ShaderModule fragmentShaderModule = nullptr;
+    wgpu::ShaderModule computeShaderModule = nullptr;
+    std::vector<wgpu::ConstantEntry> constants;
+};
+
 struct WGPUBufferObject;
 // TODO: Currently WGPUVertexBufferInfo is not used by WebGPU for useful task.
 // Update the struct when used by WebGPU driver.

filament/backend/test/BackendTest.cpp

@@ -55,14 +55,17 @@ void BackendTest::init(Backend backend, OperatingSystem operatingSystem, bool is
 BackendTest::BackendTest() : commandBufferQueue(CONFIG_MIN_COMMAND_BUFFERS_SIZE,
         CONFIG_COMMAND_BUFFERS_SIZE, /*mPaused=*/false) {
     initializeDriver();
+    mImageExpectations.emplace(getDriverApi());
 }
 
 BackendTest::~BackendTest() {
+    // Ensure all graphics commands and callbacks are finished.
+    flushAndWait();
+    mImageExpectations->evaluate();
     // Note: Don't terminate the driver for OpenGL, as it wipes away the context and removes the buffer from the screen.
     if (sBackend == Backend::OPENGL) {
         return;
     }
-    flushAndWait();
     driver->terminate();
     delete driver;
 }
@@ -156,51 +159,6 @@ void BackendTest::renderTriangle(
     api.endRenderPass();
 }
 
-void BackendTest::readPixelsAndAssertHash(const char* testName, size_t width, size_t height,
-        Handle<HwRenderTarget> rt, uint32_t expectedHash, bool exportScreenshot) {
-    void* buffer = calloc(1, width * height * 4);
-
-    struct Capture {
-        uint32_t expectedHash;
-        char* name;
-        bool exportScreenshot;
-        size_t width, height;
-    };
-    auto* c = new Capture();
-    c->expectedHash = expectedHash;
-    c->name = strdup(testName);
-    c->exportScreenshot = exportScreenshot;
-    c->width = width;
-    c->height = height;
-
-    PixelBufferDescriptor pbd(buffer, width * height * 4, PixelDataFormat::RGBA, PixelDataType::UBYTE,
-            1, 0, 0, width, [](void* buffer, size_t size, void* user) {
-                auto* c = (Capture*)user;
-
-                // Export a screenshot, if requested.
-                if (c->exportScreenshot) {
-#ifndef FILAMENT_IOS
-                    LinearImage image(c->width, c->height, 4);
-                    image = toLinearWithAlpha<uint8_t>(c->width, c->height, c->width * 4,
-                            (uint8_t*) buffer);
-                    const std::string png = std::string(c->name) + ".png";
-                    std::ofstream outputStream(png.c_str(), std::ios::binary | std::ios::trunc);
-                    ImageEncoder::encode(outputStream, ImageEncoder::Format::PNG, image, "",
-                            png);
-#endif
-                }
-
-                // Hash the contents of the buffer and check that they match.
-                uint32_t hash = utils::hash::murmur3((const uint32_t*) buffer, size / 4, 0);
-                ASSERT_EQ(hash, c->expectedHash) << c->name << " failed: hashes do not match." << std::endl;
-
-                free(buffer);
-                free(c->name);
-                free(c);
-            }, (void*)c);
-    getDriverApi().readPixels(rt, 0, 0, width, height, std::move(pbd));
-}
-
 bool BackendTest::matchesEnvironment(Backend backend) {
     return sBackend == backend;
 }

filament/backend/test/BackendTest.h

@@ -25,6 +25,7 @@
 #include "private/backend/DriverApi.h"
 
 #include "PlatformRunner.h"
+#include "ImageExpectations.h"
 
 namespace test {
 
@@ -65,17 +66,14 @@ protected:
             filament::backend::Handle<filament::backend::HwProgram> program,
             const filament::backend::RenderPassParams& params);
 
-    void readPixelsAndAssertHash(const char* testName, size_t width, size_t height,
-            filament::backend::Handle<filament::backend::HwRenderTarget> rt, uint32_t expectedHash,
-            bool exportScreenshot = false);
-
     filament::backend::DriverApi& getDriverApi() { return *commandStream; }
     filament::backend::Driver& getDriver() { return *driver; }
 
+    ImageExpectations& getExpectations() { return *mImageExpectations; }
+
     static bool matchesEnvironment(Backend backend);
     static bool matchesEnvironment(OperatingSystem operatingSystem);
     static bool matchesEnvironment(OperatingSystem operatingSystem, Backend backend);
-
 private:
 
     filament::backend::Driver* driver = nullptr;
@@ -83,6 +81,10 @@ private:
     std::unique_ptr<filament::backend::DriverApi> commandStream;
 
     filament::backend::Handle<filament::backend::HwBufferObject> uniform;
+
+    // This isn't truly optional, it just needs to delay construction until after the driver has
+    // been initialized
+    std::optional<ImageExpectations> mImageExpectations;
 };
 
 } // namespace test

filament/backend/test/ImageExpectations.cpp

@@ -97,16 +97,21 @@ void ImageExpectation::evaluate() {
 
 void ImageExpectation::compareImage() const {
     bool bytesFilled = mResult.bytesFilled();
+    // If this fails, it likely means that BackendTest::flushAndWait needs to be called before
+    // ImageExpectations is evaluated or destroyed.
     EXPECT_THAT(bytesFilled, testing::IsTrue())
                         << "Render target wasn't copied to the buffer for " << mFileName;
     if (bytesFilled) {
-        LoadedPng loadedImage(mParams.expectedFilePath());
         // Rather than directly compare the two images compare their hashes because comparing very
         // large arrays generates way too much debug output to be useful.
         uint32_t actualHash = mResult.hash();
+#ifndef FILAMENT_IOS
+        LoadedPng loadedImage(mParams.expectedFilePath());
         uint32_t loadedImageHash = loadedImage.hash();
-        EXPECT_THAT(actualHash, testing::Eq(loadedImageHash));
-        EXPECT_THAT(actualHash, testing::Eq(mParams.expectedHash()));
+        EXPECT_THAT(actualHash, testing::Eq(loadedImageHash)) << mParams.expectedFileName();
+#endif
+        // For builds that can't load PNGs (currently iOS only) use the expected hash.
+        EXPECT_THAT(actualHash, testing::Eq(mParams.expectedHash())) << mParams.expectedFileName();
         // TODO: Add better debug output, such as generating a diff image.
     }
 }
@@ -121,12 +126,13 @@ ImageExpectations::~ImageExpectations() {
 
 void ImageExpectations::addExpectation(const char* fileName, int lineNumber,
         filament::backend::RenderTargetHandle renderTarget, ScreenshotParams params) {
-    mExpectations.emplace_back(fileName, lineNumber, mApi, std::move(params), renderTarget);
+    mExpectations.emplace_back(std::make_unique<ImageExpectation>(fileName, lineNumber, mApi,
+            std::move(params), renderTarget));
 }
 
 void ImageExpectations::evaluate() {
     for (auto& expectation: mExpectations) {
-        expectation.evaluate();
+        expectation->evaluate();
     }
     mExpectations.clear();
 }
@@ -134,16 +140,13 @@ void ImageExpectations::evaluate() {
 RenderTargetDump::RenderTargetDump(filament::backend::DriverApi& api,
         filament::backend::RenderTargetHandle renderTarget, const ScreenshotParams& params)
         : mInternal(std::make_unique<RenderTargetDump::Internal>(params)) {
-#ifdef FILAMENT_IOS
-    bytesFilled_ = true;
-    bytes_.resize(size);
-    std::fill(bytes_.begin(), bytes_.end(), 0);
-#else
     const size_t size = mInternal->params.width() * mInternal->params.height() * 4;
     mInternal->bytes.resize(size);
 
     auto cb = [](void* buffer, size_t size, void* user) {
         auto* internal = static_cast<RenderTargetDump::Internal*>(user);
+        internal->bytesFilled = true;
+#ifndef FILAMENT_IOS
         image::LinearImage image(internal->params.width(), internal->params.width(), 4);
         image = image::toLinearWithAlpha<uint8_t>(internal->params.width(),
                 internal->params.height(),
@@ -152,14 +155,13 @@ RenderTargetDump::RenderTargetDump(filament::backend::DriverApi& api,
         std::ofstream pngStream(filePath, std::ios::binary | std::ios::trunc);
         image::ImageEncoder::encode(pngStream, image::ImageEncoder::Format::PNG, image, "",
                 filePath);
-        internal->bytesFilled = true;
+#endif
     };
     filament::backend::PixelBufferDescriptor pb(mInternal->bytes.data(), size,
             filament::backend::PixelDataFormat::RGBA, filament::backend::PixelDataType::UBYTE, cb,
             (void*)mInternal.get());
     api.readPixels(renderTarget, 0, 0, mInternal->params.width(), mInternal->params.height(),
             std::move(pb));
-#endif
 }
 
 RenderTargetDump::~RenderTargetDump() {
@@ -183,9 +185,9 @@ bool RenderTargetDump::bytesFilled() const {
 
 RenderTargetDump::Internal::Internal(const ScreenshotParams& params) : params(params) {}
 
-LoadedPng::LoadedPng(std::string filePath) {
+LoadedPng::LoadedPng(std::string filePath) : mFilePath(std::move(filePath)) {
 #ifndef FILAMENT_IOS
-    std::ifstream pngStream(filePath, std::ios::binary);
+    std::ifstream pngStream(mFilePath, std::ios::binary);
     image::LinearImage loadedImage = image::ImageDecoder::decode(pngStream, filePath,
             image::ImageDecoder::ColorSpace::LINEAR);
     size_t valuesInImage = loadedImage.getWidth() * loadedImage.getHeight() *
@@ -201,7 +203,8 @@ LoadedPng::LoadedPng(std::string filePath) {
 }
 
 uint32_t LoadedPng::hash() const {
-    EXPECT_THAT(mBytes, testing::Not(testing::IsEmpty()));
+    EXPECT_THAT(mBytes, testing::Not(testing::IsEmpty()))
+            << "Failed to load expected test result: " << mFilePath;
     if (mBytes.empty()) {
         return 0;
     }

filament/backend/test/ImageExpectations.h

@@ -106,6 +106,7 @@ public:
     uint32_t hash() const;
 
 private:
+    std::string mFilePath;
     std::vector<unsigned char> mBytes;
 };
 
@@ -141,7 +142,8 @@ public:
 
 private:
     filament::backend::DriverApi& mApi;
-    std::vector<ImageExpectation> mExpectations;
+    // Store expectations in unique pointers because they are self referential.
+    std::vector<std::unique_ptr<ImageExpectation>> mExpectations;
 };
 
 #endif //TNT_IMAGE_EXPECTATIONS_H

filament/backend/test/test_Blit.cpp

@@ -197,16 +197,10 @@ TEST_F(BlitTest, ColorMagnify) {
     }
 
     {
-        ImageExpectations expectations(api);
-
-        {
-            RenderFrame frame(api);
-            EXPECT_IMAGE(dstRenderTargets[0], expectations,
-                    ScreenshotParams(kDstTexWidth, kDstTexHeight, "ColorMagnify", 0x410bdd31));
-            api.commit(swapChain);
-        }
-
-        flushAndWait();
+        RenderFrame frame(api);
+        EXPECT_IMAGE(dstRenderTargets[0], getExpectations(),
+                ScreenshotParams(kDstTexWidth, kDstTexHeight, "ColorMagnify", 0x410bdd31));
+        api.commit(swapChain);
     }
 }
 
@@ -260,14 +254,8 @@ TEST_F(BlitTest, ColorMinify) {
             { 0, 0, kSrcTexWidth >> srcLevel, kSrcTexHeight >> srcLevel },
             SamplerMagFilter::LINEAR);
 
-    {
-        ImageExpectations expectations(api);
-
-        EXPECT_IMAGE(dstRenderTargets[0], expectations,
-                ScreenshotParams(kDstTexWidth, kDstTexHeight, "ColorMinify", 0xf3d9c53f));
-
-        flushAndWait();
-    }
+    EXPECT_IMAGE(dstRenderTargets[0], getExpectations(),
+            ScreenshotParams(kDstTexWidth, kDstTexHeight, "ColorMinify", 0xf3d9c53f));
 }
 
 TEST_F(BlitTest, ColorResolve) {
@@ -351,14 +339,8 @@ TEST_F(BlitTest, ColorResolve) {
             srcRenderTarget, { 0, 0, kSrcTexWidth, kSrcTexHeight },
             SamplerMagFilter::NEAREST);
 
-    {
-        ImageExpectations expectations(api);
-
-        EXPECT_IMAGE(dstRenderTarget, expectations,
-                ScreenshotParams(kDstTexWidth, kDstTexHeight, "ColorResolve", 0xebfac2ef));
-
-        flushAndWait();
-    }
+    EXPECT_IMAGE(dstRenderTarget, getExpectations(),
+            ScreenshotParams(kDstTexWidth, kDstTexHeight, "ColorResolve", 0xebfac2ef));
 }
 
 TEST_F(BlitTest, Blit2DTextureArray) {
@@ -423,17 +405,11 @@ TEST_F(BlitTest, Blit2DTextureArray) {
     }
 
     {
-        ImageExpectations expectations(api);
-
-        {
-            RenderFrame frame(api);
-            EXPECT_IMAGE(dstRenderTarget, expectations,
-                    ScreenshotParams(kDstTexWidth, kDstTexHeight, "Blit2DTextureArray",
-                            0x8de7d55b));
-            api.commit(swapChain);
-        }
-
-        flushAndWait();
+        RenderFrame frame(api);
+        EXPECT_IMAGE(dstRenderTarget, getExpectations(),
+                ScreenshotParams(kDstTexWidth, kDstTexHeight, "Blit2DTextureArray",
+                        0x8de7d55b));
+        api.commit(swapChain);
     }
 }
 
@@ -503,18 +479,12 @@ TEST_F(BlitTest, BlitRegion) {
     }
 
     {
-        ImageExpectations expectations(api);
-
-        {
-            RenderFrame frame(api);
-            // TODO: for some reason, this test has very, very slight (as in one pixel) differences
-            // between OpenGL and Metal. So disable golden checking for now.
-            // EXPECT_IMAGE(dstRenderTarget, expectations, ScreenshotParams(kDstTexWidth,
-            //         kDstTexHeight, "BlitRegion", 0x74fa34ed));
-            api.commit(swapChain);
-        }
-
-        flushAndWait();
+        RenderFrame frame(api);
+        // TODO: for some reason, this test has very, very slight (as in one pixel) differences
+        // between OpenGL and Metal. So disable golden checking for now.
+        // EXPECT_IMAGE(dstRenderTarget, expectations, ScreenshotParams(kDstTexWidth,
+        //         kDstTexHeight, "BlitRegion", 0x74fa34ed));
+        api.commit(swapChain);
     }
 }
 
@@ -567,25 +537,19 @@ TEST_F(BlitTest, BlitRegionToSwapChain) {
     };
 
     {
-        ImageExpectations expectations(api);
-        {
-            {
-                RenderFrame frame(api);
+        RenderFrame frame(api);
 
-                api.blitDEPRECATED(TargetBufferFlags::COLOR0, dstRenderTarget,
-                        dstRect, srcRenderTargets[srcLevel],
-                        srcRect, SamplerMagFilter::LINEAR);
+        api.blitDEPRECATED(TargetBufferFlags::COLOR0, dstRenderTarget,
+                dstRect, srcRenderTargets[srcLevel],
+                srcRect, SamplerMagFilter::LINEAR);
 
-                api.commit(swapChain);
-            }
-
-            // TODO: for some reason, this test has been disabled. It needs to be tested on all
-            // machines.
-            // EXPECT_IMAGE(dstRenderTarget, expectations,
-            //         ScreenshotParams(kDstTexWidth, kDstTexHeight, "BlitRegionToSwapChain", 0x0));
-        }
-        flushAndWait();
+        api.commit(swapChain);
     }
+
+    // TODO: for some reason, this test has been disabled. It needs to be tested on all
+    // machines.
+    // EXPECT_IMAGE(dstRenderTarget, expectations,
+    //         ScreenshotParams(kDstTexWidth, kDstTexHeight, "BlitRegionToSwapChain", 0x0));
 }
 
 } // namespace test

filament/backend/test/test_BufferUpdates.cpp

@@ -16,6 +16,7 @@
 
 #include "BackendTest.h"
 
+#include "ImageExpectations.h"
 #include "Lifetimes.h"
 #include "Shader.h"
 #include "SharedShaders.h"
@@ -220,9 +221,9 @@ TEST_F(BufferUpdatesTest, BufferObjectUpdateWithOffset) {
     renderTriangle({{ DescriptorSetLayoutHandle{}, shader.getDescriptorSetLayout() }},
             renderTarget, swapChain, shader.getProgram(), params);
 
-    static const uint32_t expectedHash = 91322442;
-    readPixelsAndAssertHash(
-            "BufferObjectUpdateWithOffset", 512, 512, renderTarget, expectedHash, true);
+
+    EXPECT_IMAGE(renderTarget, getExpectations(),
+            ScreenshotParams(512, 512, "BufferObjectUpdateWithOffset", 91322442));
 
     api.flush();
     api.commit(swapChain);

filament/backend/test/test_FeedbackLoops.cpp

@@ -71,8 +71,6 @@ void main() {
     fragColor = textureLod(test_tex, uv, params.sourceLevel);
 })";
 
-static uint32_t sPixelHashResult = 0;
-
 // Selecting a NPOT texture size seems to exacerbate the bug seen with Intel GPU's.
 // Note that Filament uses a higher precision format (R11F_G11F_B10F) but this does not seem
 // necessary to trigger the bug.
@@ -96,25 +94,6 @@ struct MaterialParams {
     float unused;
 };
 
-static void dumpScreenshot(DriverApi& dapi, Handle<HwRenderTarget> rt) {
-    const size_t size = kTexWidth * kTexHeight * 4;
-    void* buffer = calloc(1, size);
-    auto cb = [](void* buffer, size_t size, void* user) {
-        int w = kTexWidth, h = kTexHeight;
-        const uint32_t* texels = (uint32_t*) buffer;
-        sPixelHashResult = utils::hash::murmur3(texels, size / 4, 0);
-#ifndef FILAMENT_IOS
-        LinearImage image(w, h, 4);
-        image = toLinearWithAlpha<uint8_t>(w, h, w * 4, (uint8_t*) buffer);
-        std::ofstream pngstrm("feedback.png", std::ios::binary | std::ios::trunc);
-        ImageEncoder::encode(pngstrm, ImageEncoder::Format::PNG, image, "", "feedback.png");
-#endif
-        free(buffer);
-    };
-    PixelBufferDescriptor pb(buffer, size, PixelDataFormat::RGBA, PixelDataType::UBYTE, cb);
-    dapi.readPixels(rt, 0, 0, kTexWidth, kTexHeight, std::move(pb));
-}
-
 // TODO: This test needs work to get Metal and OpenGL to agree on results.
 // The problems are caused by both uploading and rendering into the same texture, since the OpenGL
 // backend's readPixels does not work correctly with textures that have image data uploaded.
@@ -259,7 +238,8 @@ TEST_F(BackendTest, FeedbackLoops) {
             // NOTE: Calling glReadPixels on any miplevel other than the base level
             // seems to be un-reliable on some GPU's.
             if (frame == kNumFrames - 1) {
-                dumpScreenshot(api, renderTargets[0]);
+                EXPECT_IMAGE(renderTargets[0], getExpectations(),
+                        ScreenshotParams(kTexWidth, kTexHeight, "FeedbackLoops", 0x70695aa1));
             }
 
             api.flush();
@@ -270,10 +250,6 @@ TEST_F(BackendTest, FeedbackLoops) {
             getDriver().purge();
         }
     }
-
-    const uint32_t expected = 0x70695aa1;
-    printf("Computed hash is 0x%8.8x, Expected 0x%8.8x\n", sPixelHashResult, expected);
-    EXPECT_TRUE(sPixelHashResult == expected);
 }
 
 } // namespace test

filament/backend/test/test_LoadImage.cpp

@@ -246,42 +246,42 @@ TEST_F(LoadImageTest, UpdateImage2D) {
     std::vector<TestCase> testCases;
 
     // Test basic upload.
-    testCases.emplace_back("RGBA, UBYTE -> RGBA8", PixelDataFormat::RGBA, PixelDataType::UBYTE, TextureFormat::RGBA8);
+    testCases.emplace_back("RGBA UBYTE to RGBA8", PixelDataFormat::RGBA, PixelDataType::UBYTE, TextureFormat::RGBA8);
 
     // Test format conversion.
     // TODO: Vulkan crashes with `Texture at colorAttachment[0] has usage (0x01) which doesn't specify MTLTextureUsageRenderTarget (0x04)'
-    testCases.emplace_back("RGBA, FLOAT -> RGBA16F", PixelDataFormat::RGBA, PixelDataType::FLOAT, TextureFormat::RGBA16F);
+    testCases.emplace_back("RGBA FLOAT to RGBA16F", PixelDataFormat::RGBA, PixelDataType::FLOAT, TextureFormat::RGBA16F);
 
     // Test texture formats not all backends support natively.
     // TODO: Vulkan crashes with "VK_FORMAT_R32G32B32_SFLOAT is not supported"
-    testCases.emplace_back("RGB, FLOAT -> RGB32F", PixelDataFormat::RGB, PixelDataType::FLOAT, TextureFormat::RGB32F);
-    testCases.emplace_back("RGB, FLOAT -> RGB16F", PixelDataFormat::RGB, PixelDataType::FLOAT, TextureFormat::RGB16F);
+    testCases.emplace_back("RGB FLOAT to RGB32F", PixelDataFormat::RGB, PixelDataType::FLOAT, TextureFormat::RGB32F);
+    testCases.emplace_back("RGB FLOAT to RGB16F", PixelDataFormat::RGB, PixelDataType::FLOAT, TextureFormat::RGB16F);
 
     // Test packed format uploads.
     // TODO: Vulkan crashes with "Texture at colorAttachment[0] has usage (0x01) which doesn't specify MTLTextureUsageRenderTarget (0x04)"
-    testCases.emplace_back("RGBA, UINT_2_10_10_10_REV -> RGB10_A2", PixelDataFormat::RGBA, PixelDataType::UINT_2_10_10_10_REV, TextureFormat::RGB10_A2);
-    testCases.emplace_back("RGB, UINT_10F_11F_11F_REV -> R11F_G11F_B10F", PixelDataFormat::RGB, PixelDataType::UINT_10F_11F_11F_REV, TextureFormat::R11F_G11F_B10F);
-    testCases.emplace_back("RGB, HALF -> R11F_G11F_B10F", PixelDataFormat::RGB, PixelDataType::HALF, TextureFormat::R11F_G11F_B10F);
+    testCases.emplace_back("RGBA UINT_2_10_10_10_REV to RGB10_A2", PixelDataFormat::RGBA, PixelDataType::UINT_2_10_10_10_REV, TextureFormat::RGB10_A2);
+    testCases.emplace_back("RGB UINT_10F_11F_11F_REV to R11F_G11F_B10F", PixelDataFormat::RGB, PixelDataType::UINT_10F_11F_11F_REV, TextureFormat::R11F_G11F_B10F);
+    testCases.emplace_back("RGB HALF to R11F_G11F_B10F", PixelDataFormat::RGB, PixelDataType::HALF, TextureFormat::R11F_G11F_B10F);
 
     // Test integer format uploads.
     // TODO: These cases fail on OpenGL and Vulkan.
     // TODO: These cases now also fail on Metal, but at some point previously worked.
-     testCases.emplace_back("RGB_INTEGER, UBYTE -> RGB8UI", PixelDataFormat::RGB_INTEGER, PixelDataType::UBYTE, TextureFormat::RGB8UI);
-     testCases.emplace_back("RGB_INTEGER, USHORT -> RGB16UI", PixelDataFormat::RGB_INTEGER, PixelDataType::USHORT, TextureFormat::RGB16UI);
-     testCases.emplace_back("RGB_INTEGER, INT -> RGB32I", PixelDataFormat::RGB_INTEGER, PixelDataType::INT, TextureFormat::RGB32I);
+     testCases.emplace_back("RGB_INTEGER UBYTE to RGB8UI", PixelDataFormat::RGB_INTEGER, PixelDataType::UBYTE, TextureFormat::RGB8UI);
+     testCases.emplace_back("RGB_INTEGER USHORT to RGB16UI", PixelDataFormat::RGB_INTEGER, PixelDataType::USHORT, TextureFormat::RGB16UI);
+     testCases.emplace_back("RGB_INTEGER INT to RGB32I", PixelDataFormat::RGB_INTEGER, PixelDataType::INT, TextureFormat::RGB32I);
 
     // Test uploads with buffer padding.
     // TODO: Vulkan crashes with "Assertion failed: (offset + size <= allocationSize)"
-    testCases.emplace_back("RGBA, UBYTE -> RGBA8 (with buffer padding)", PixelDataFormat::RGBA, PixelDataType::UBYTE, TextureFormat::RGBA8, 64u);
-    testCases.emplace_back("RGBA, FLOAT -> RGBA16F (with buffer padding)", PixelDataFormat::RGBA, PixelDataType::FLOAT, TextureFormat::RGBA16F, 64u);
-    testCases.emplace_back("RGB, FLOAT -> RGB32F (with buffer padding)", PixelDataFormat::RGB, PixelDataType::FLOAT, TextureFormat::RGB32F, 64u);
+    testCases.emplace_back("RGBA UBYTE to RGBA8 (with buffer padding)", PixelDataFormat::RGBA, PixelDataType::UBYTE, TextureFormat::RGBA8, 64u);
+    testCases.emplace_back("RGBA FLOAT to RGBA16F (with buffer padding)", PixelDataFormat::RGBA, PixelDataType::FLOAT, TextureFormat::RGBA16F, 64u);
+    testCases.emplace_back("RGB FLOAT to RGB32F (with buffer padding)", PixelDataFormat::RGB, PixelDataType::FLOAT, TextureFormat::RGB32F, 64u);
 
     // Upload subregions separately.
     // TODO: Vulkan crashes with "Offsets not yet supported"
-    testCases.emplace_back("RGBA, UBYTE -> RGBA8 (subregions)", PixelDataFormat::RGBA, PixelDataType::UBYTE, TextureFormat::RGBA8, 0u, true);
-    testCases.emplace_back("RGBA, FLOAT -> RGBA16F (subregions)", PixelDataFormat::RGBA, PixelDataType::FLOAT, TextureFormat::RGBA16F, 0u, true);
-    testCases.emplace_back("RGBA, UBYTE -> RGBA8 (subregions, buffer padding)", PixelDataFormat::RGBA, PixelDataType::UBYTE, TextureFormat::RGBA8, 64u, true);
-    testCases.emplace_back("RGB, FLOAT -> RGB32F (subregions, buffer padding)", PixelDataFormat::RGB, PixelDataType::FLOAT, TextureFormat::RGB32F, 64u, true);
+    testCases.emplace_back("RGBA UBYTE to RGBA8 (subregions)", PixelDataFormat::RGBA, PixelDataType::UBYTE, TextureFormat::RGBA8, 0u, true);
+    testCases.emplace_back("RGBA FLOAT to RGBA16F (subregions)", PixelDataFormat::RGBA, PixelDataType::FLOAT, TextureFormat::RGBA16F, 0u, true);
+    testCases.emplace_back("RGBA UBYTE to RGBA8 (subregions and buffer padding)", PixelDataFormat::RGBA, PixelDataType::UBYTE, TextureFormat::RGBA8, 64u, true);
+    testCases.emplace_back("RGB FLOAT to RGB32F (subregions and buffer padding)", PixelDataFormat::RGB, PixelDataType::FLOAT, TextureFormat::RGB32F, 64u, true);
 
     auto& api = getDriverApi();
 
@@ -339,16 +339,14 @@ TEST_F(LoadImageTest, UpdateImage2D) {
         renderTriangle({{ DescriptorSetLayoutHandle{}, shader.getDescriptorSetLayout() }},
                 defaultRenderTarget, swapChain, shader.getProgram());
 
-        readPixelsAndAssertHash(t.name, 512, 512, defaultRenderTarget, expectedHash);
+        EXPECT_IMAGE(defaultRenderTarget, getExpectations(),
+                ScreenshotParams(512, 512, t.name, expectedHash));
 
         api.commit(swapChain);
         api.endFrame(0);
     }
 
-    api.finish();
     api.stopCapture();
-
-    flushAndWait();
 }
 
 TEST_F(LoadImageTest, UpdateImageSRGB) {
@@ -372,7 +370,7 @@ TEST_F(LoadImageTest, UpdateImageSRGB) {
             getSamplerTypeName(textureFormat), fragmentTemplate);
     Shader shader(api, cleanup, ShaderConfig{
         .vertexShader = mVertexShader, .fragmentShader = fragment, .uniforms = {{
-            "text_tex", DescriptorType::SAMPLER, samplerInfo
+            "test_tex", DescriptorType::SAMPLER, samplerInfo
     }}});
 
     // Create a texture.
@@ -416,15 +414,12 @@ TEST_F(LoadImageTest, UpdateImageSRGB) {
     renderTriangle({{ DescriptorSetLayoutHandle{}, shader.getDescriptorSetLayout() }},
             defaultRenderTarget, swapChain, shader.getProgram());
 
-    static const uint32_t expectedHash = 359858623;
-    readPixelsAndAssertHash("UpdateImageSRGB", 512, 512, defaultRenderTarget, expectedHash);
+    EXPECT_IMAGE(defaultRenderTarget, getExpectations(),
+            ScreenshotParams(512, 512, "UpdateImageSRGB", 359858623));
 
-    api.flush();
     api.commit(swapChain);
     api.endFrame(0);
 
-    // This ensures all driver commands have finished before exiting the test.
-    api.finish();
     api.stopCapture();
 }
 
@@ -478,15 +473,12 @@ TEST_F(LoadImageTest, UpdateImageMipLevel) {
     renderTriangle({{ DescriptorSetLayoutHandle{}, shader.getDescriptorSetLayout() }},
             defaultRenderTarget, swapChain, shader.getProgram());
 
-    static const uint32_t expectedHash = 3644679986;
-    readPixelsAndAssertHash("UpdateImageMipLevel", 512, 512, defaultRenderTarget, expectedHash);
+    EXPECT_IMAGE(defaultRenderTarget, getExpectations(),
+            ScreenshotParams(512, 512, "UpdateImageMipLevel", 3644679986));
 
-    api.flush();
     api.commit(swapChain);
     api.endFrame(0);
 
-    // This ensures all driver commands have finished before exiting the test.
-    api.finish();
     api.stopCapture();
 }
 
@@ -538,29 +530,24 @@ TEST_F(LoadImageTest, UpdateImage3D) {
 
     api.update3DImage(texture, 0, 0, 0, 0, 512, 512, 4, std::move(descriptor));
 
-    api.beginFrame(0, 0, 0);
+    {
+        RenderFrame frame(api);
 
-    // Update samplers.
-    DescriptorSetHandle  descriptorSet = shader.createDescriptorSet(api);
-    api.updateDescriptorSetTexture(descriptorSet, 0, texture, {
-            .filterMag = SamplerMagFilter::LINEAR,
-            .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST
-    });
+        // Update samplers.
+        DescriptorSetHandle descriptorSet = shader.createDescriptorSet(api);
+        api.updateDescriptorSetTexture(descriptorSet, 0, texture,
+                { .filterMag = SamplerMagFilter::LINEAR,
+                    .filterMin = SamplerMinFilter::LINEAR_MIPMAP_NEAREST });
 
-    api.bindDescriptorSet(descriptorSet, 1, {});
+        api.bindDescriptorSet(descriptorSet, 1, {});
 
-    renderTriangle({{ DescriptorSetLayoutHandle{}, shader.getDescriptorSetLayout() }},
-            defaultRenderTarget, swapChain, shader.getProgram());
+        renderTriangle({ { DescriptorSetLayoutHandle{}, shader.getDescriptorSetLayout() } },
+                defaultRenderTarget, swapChain, shader.getProgram());
 
-    static const uint32_t expectedHash = 3644679986;
-    readPixelsAndAssertHash("UpdateImage3D", 512, 512, defaultRenderTarget, expectedHash);
+        EXPECT_IMAGE(defaultRenderTarget, getExpectations(),
+                ScreenshotParams(512, 512, "UpdateImage3D", 3644679986));
+    }
 
-    api.flush();
-    api.commit(swapChain);
-    api.endFrame(0);
-
-    // This ensures all driver commands have finished before exiting the test.
-    api.finish();
     api.stopCapture();
 }
 

filament/backend/test/test_PushConstants.cpp

@@ -16,6 +16,7 @@
 
 #include "BackendTest.h"
 
+#include "ImageExpectations.h"
 #include "Lifetimes.h"
 #include "ShaderGenerator.h"
 #include "Skip.h"
@@ -151,19 +152,14 @@ TEST_F(BackendTest, PushConstants) {
 
         api.endRenderPass();
 
-        readPixelsAndAssertHash("pushConstants", 512, 512, renderTarget, 1957275826, true);
+        EXPECT_IMAGE(renderTarget, getExpectations(),
+                ScreenshotParams(512, 512, "pushConstants", 1957275826));
 
         api.commit(swapChain);
         api.endFrame(0);
     }
 
     api.stopCapture(0);
-
-    // Wait for the ReadPixels result to come back.
-    api.finish();
-
-    executeCommands();
-    getDriver().purge();
 }
 
 } // namespace test

filament/backend/test/test_RenderExternalImage.cpp

@@ -16,6 +16,7 @@
 
 #include "BackendTest.h"
 
+#include "ImageExpectations.h"
 #include "Lifetimes.h"
 #include "Shader.h"
 #include "SharedShaders.h"
@@ -200,17 +201,16 @@ TEST_F(BackendTest, RenderExternalImage) {
     api.draw(state, triangle.getRenderPrimitive(), 0, 3, 1);
     api.endRenderPass();
 
-    readPixelsAndAssertHash("RenderExternalImage", 512, 512, defaultRenderTarget, 267229901, true);
-
     api.flush();
     api.commit(swapChain);
     api.endFrame(0);
+    EXPECT_IMAGE(defaultRenderTarget, getExpectations(),
+            ScreenshotParams(512, 512, "RenderExternalImage", 267229901));
 
     api.stopCapture(0);
-
     api.finish();
+    flushAndWait();
 
-    executeCommands();
 }
 
 } // namespace test

filament/backend/test/test_Scissor.cpp

@@ -16,6 +16,7 @@
 
 #include "BackendTest.h"
 
+#include "ImageExpectations.h"
 #include "Lifetimes.h"
 #include "Shader.h"
 #include "SharedShaders.h"
@@ -130,20 +131,14 @@ TEST_F(BackendTest, ScissorViewportRegion) {
         api.draw(ps, triangle.getRenderPrimitive(), 0, 3, 1);
         api.endRenderPass();
 
-        readPixelsAndAssertHash("scissor", kSrcTexWidth >> 1, kSrcTexHeight >> 1, fullRenderTarget,
-                0xAB3D1C53, true);
+        EXPECT_IMAGE(fullRenderTarget, getExpectations(),
+                ScreenshotParams(kSrcTexWidth >> 1, kSrcTexHeight >> 1, "scissor", 0xAB3D1C53));
 
         api.commit(swapChain);
         api.endFrame(0);
 
         api.stopCapture(0);
     }
-
-    // Wait for the ReadPixels result to come back.
-    api.finish();
-
-    executeCommands();
-    getDriver().purge();
 }
 
 // Verify that a negative Viewport origin works with scissor.
@@ -226,20 +221,14 @@ TEST_F(BackendTest, ScissorViewportEdgeCases) {
         api.draw(ps, triangle.getRenderPrimitive(), 0, 3, 1);
         api.endRenderPass();
 
-        readPixelsAndAssertHash(
-                "ScissorViewportEdgeCases", 512, 512, renderTarget, 0x6BF00F31, true);
+        EXPECT_IMAGE(renderTarget, getExpectations(),
+                ScreenshotParams(512, 512, "ScissorViewportEdgeCases", 0x6BF00F31));
 
         api.commit(swapChain);
         api.endFrame(0);
 
         api.stopCapture(0);
     }
-
-    // Wait for the ReadPixels result to come back.
-    api.finish();
-
-    executeCommands();
-    getDriver().purge();
 }
 
 } // namespace test

filament/backend/test/test_StencilBuffer.cpp

@@ -16,6 +16,7 @@
 
 #include "BackendTest.h"
 
+#include "ImageExpectations.h"
 #include "Lifetimes.h"
 #include "Shader.h"
 #include "SharedShaders.h"
@@ -129,7 +130,8 @@ TEST_F(BasicStencilBufferTest, StencilBuffer) {
 
     RunTest(renderTarget);
 
-    readPixelsAndAssertHash("StencilBuffer", 512, 512, renderTarget, 0x3B1AEF0F, true);
+    EXPECT_IMAGE(renderTarget, getExpectations(),
+            ScreenshotParams(512, 512, "StencilBuffer", 0x3B1AEF0F));
 
     flushAndWait();
     getDriver().purge();
@@ -151,7 +153,8 @@ TEST_F(BasicStencilBufferTest, DepthAndStencilBuffer) {
 
     RunTest(renderTarget);
 
-    readPixelsAndAssertHash("DepthAndStencilBuffer", 512, 512, renderTarget, 0x3B1AEF0F, true);
+    EXPECT_IMAGE(renderTarget, getExpectations(),
+            ScreenshotParams(512, 512, "DepthAndStencilBuffer", 0x3B1AEF0F));
 
     flushAndWait();
     getDriver().purge();
@@ -233,7 +236,8 @@ TEST_F(BasicStencilBufferTest, StencilBufferMSAA) {
     api.stopCapture(0);
     api.endFrame(0);
 
-    readPixelsAndAssertHash("StencilBufferAutoResolve", 512, 512, renderTarget1, 0x6CEFAC8F, true);
+    EXPECT_IMAGE(renderTarget1, getExpectations(),
+            ScreenshotParams(512, 512, "StencilBufferAutoResolve", 0x6CEFAC8F));
 
     flushAndWait();
     getDriver().purge();

libs/gltfio/src/AssetLoader.cpp

@@ -50,6 +50,8 @@
 #include <utils/Log.h>
 #include <utils/Panic.h>
 #include <utils/NameComponentManager.h>
+
+#define SYSTRACE_TAG SYSTRACE_TAG_GLTFIO
 #include <utils/Systrace.h>
 
 #include <tsl/robin_map.h>

libs/gltfio/src/ResourceLoader.cpp

@@ -37,7 +37,10 @@
 #include <utils/compiler.h>
 #include <utils/JobSystem.h>
 #include <utils/Log.h>
+
+#define SYSTRACE_TAG SYSTRACE_TAG_GLTFIO
 #include <utils/Systrace.h>
+
 #include <utils/Path.h>
 
 #include <cgltf.h>

libs/gltfio/src/Utility.cpp

@@ -21,6 +21,8 @@
 #include "GltfEnums.h"
 
 #include <utils/Log.h>
+
+#define SYSTRACE_TAG SYSTRACE_TAG_GLTFIO
 #include <utils/Systrace.h>
 
 #define CGLTF_IMPLEMENTATION

libs/utils/CMakeLists.txt

@@ -112,6 +112,7 @@ if (ANDROID)
     target_link_libraries(${TARGET} PUBLIC log)
     target_link_libraries(${TARGET} PRIVATE dl)
     target_link_libraries(${TARGET} PUBLIC android)
+    target_link_libraries(${TARGET} PUBLIC perfetto)
 endif()
 
 if (WIN32)

libs/utils/include/utils/Systrace.h

@@ -17,24 +17,23 @@
 #ifndef TNT_UTILS_SYSTRACE_H
 #define TNT_UTILS_SYSTRACE_H
 
-
-#define SYSTRACE_TAG_NEVER          (0)
-#define SYSTRACE_TAG_ALWAYS         (1<<0)
-#define SYSTRACE_TAG_FILAMENT       (1<<1)  // don't change, used in makefiles
-#define SYSTRACE_TAG_JOBSYSTEM      (1<<2)
+#define SYSTRACE_TAG_DISABLED       (0)
+#define SYSTRACE_TAG_FILAMENT       (2) // don't change used in makefiles
+#define SYSTRACE_TAG_JOBSYSTEM      (3)
+#define SYSTRACE_TAG_GLTFIO         (4)
 
 /*
- * The SYSTRACE_ macros use SYSTRACE_TAG as a the TAG, which should be defined
- * before this file is included. If not, the SYSTRACE_TAG_ALWAYS tag will be used.
+ * The SYSTRACE_ macros use SYSTRACE_TAG as a category, which must be defined
+ * before this file is included.
  */
 
 #ifndef SYSTRACE_TAG
-#define SYSTRACE_TAG (SYSTRACE_TAG_ALWAYS)
+#   error SYSTRACE_TAG must be set to SYSTRACE_TAG_{DISABLED|FILAMENT|JOBSYSTEM}
 #endif
 
 // Systrace on Apple platforms is fragile and adds overhead, should only be enabled in dev builds.
 #ifndef FILAMENT_APPLE_SYSTRACE
-#define FILAMENT_APPLE_SYSTRACE 0
+#   define FILAMENT_APPLE_SYSTRACE 0
 #endif
 
 #if defined(__ANDROID__)
@@ -44,7 +43,6 @@
 #else
 
 #define SYSTRACE_ENABLE()
-#define SYSTRACE_DISABLE()
 #define SYSTRACE_CONTEXT()
 #define SYSTRACE_NAME(name)
 #define SYSTRACE_FRAME_ID(frame)
@@ -56,6 +54,6 @@
 #define SYSTRACE_VALUE32(name, val)
 #define SYSTRACE_VALUE64(name, val)
 
-#endif // ANDROID
+#endif
 
 #endif // TNT_UTILS_SYSTRACE_H

libs/utils/include/utils/android/Systrace.h

@@ -17,226 +17,75 @@
 #ifndef TNT_UTILS_ANDROID_SYSTRACE_H
 #define TNT_UTILS_ANDROID_SYSTRACE_H
 
-#include <atomic>
+#include <perfetto/perfetto.h>
 
 #include <stdint.h>
-#include <stdio.h>
-#include <unistd.h>
 
-#include <utils/compiler.h>
+PERFETTO_DEFINE_CATEGORIES_IN_NAMESPACE(systrace,
+        perfetto::Category("filament"),
+        perfetto::Category("jobsystem"),
+        perfetto::Category("gltfio"));
 
-// enable tracing
-#define SYSTRACE_ENABLE() ::utils::details::Systrace::enable(SYSTRACE_TAG)
+PERFETTO_USE_CATEGORIES_FROM_NAMESPACE(systrace);
 
-// disable tracing
-#define SYSTRACE_DISABLE() ::utils::details::Systrace::disable(SYSTRACE_TAG)
+#if SYSTRACE_TAG == SYSTRACE_TAG_FILAMENT
+#   define UTILS_PERFETTO_CATEGORY "filament"
+#elif SYSTRACE_TAG == SYSTRACE_TAG_JOBSYSTEM
+#   define UTILS_PERFETTO_CATEGORY "jobsystem"
+#elif SYSTRACE_TAG == SYSTRACE_TAG_GLTFIO
+#   define UTILS_PERFETTO_CATEGORY "gltfio"
+#endif
 
+#if SYSTRACE_TAG == SYSTRACE_TAG_DISABLED
 
-/**
- * Creates a Systrace context in the current scope. needed for calling all other systrace
- * commands below.
- */
-#define SYSTRACE_CONTEXT() ::utils::details::Systrace ___trctx(SYSTRACE_TAG)
+#define SYSTRACE_ENABLE()
+#define SYSTRACE_CONTEXT()
+#define SYSTRACE_NAME(name)
+#define SYSTRACE_FRAME_ID(frame)
+#define SYSTRACE_NAME_BEGIN(name)
+#define SYSTRACE_NAME_END()
+#define SYSTRACE_CALL()
+#define SYSTRACE_ASYNC_BEGIN(name, cookie)
+#define SYSTRACE_ASYNC_END(name, cookie)
+#define SYSTRACE_VALUE32(name, val)
+#define SYSTRACE_VALUE64(name, val)
 
+#else
 
-// SYSTRACE_NAME traces the beginning and end of the current scope.  To trace
-// the correct start and end times this macro should be declared first in the
-// scope body.
-// It also automatically creates a Systrace context
-#define SYSTRACE_NAME(name) ::utils::details::ScopedTrace ___tracer(SYSTRACE_TAG, name)
+#define SYSTRACE_ENABLE()
+#define SYSTRACE_CONTEXT()
 
-// Denotes that a new frame has started processing.
-#define SYSTRACE_FRAME_ID(frame) \
-    { /* scope for frame id trace */ \
-        char buf[64]; \
-        snprintf(buf, 64, "frame %u", frame); \
-        SYSTRACE_NAME(buf); \
-    }
+#define SYSTRACE_CALL() \
+    auto constexpr FILAMENT_SYSTRACE_FUNCTION = perfetto::StaticString(__FUNCTION__); \
+    TRACE_EVENT(UTILS_PERFETTO_CATEGORY, FILAMENT_SYSTRACE_FUNCTION)
 
-// SYSTRACE_CALL is an SYSTRACE_NAME that uses the current function name.
-#define SYSTRACE_CALL() SYSTRACE_NAME(__FUNCTION__)
+#define SYSTRACE_NAME(name) TRACE_EVENT(UTILS_PERFETTO_CATEGORY, nullptr,   \
+        [&](perfetto::EventContext ctx) {                                   \
+            ctx.event()->set_name(name);                                    \
+        })
 
-#define SYSTRACE_NAME_BEGIN(name) \
-        ___trctx.traceBegin(SYSTRACE_TAG, name)
+#define SYSTRACE_NAME_BEGIN(name) TRACE_EVENT_BEGIN(UTILS_PERFETTO_CATEGORY, nullptr,   \
+        [&](perfetto::EventContext ctx) {                                               \
+            ctx.event()->set_name(name);                                                \
+        })
 
-#define SYSTRACE_NAME_END() \
-        ___trctx.traceEnd(SYSTRACE_TAG)
+#define SYSTRACE_NAME_END() TRACE_EVENT_END(UTILS_PERFETTO_CATEGORY)
 
-
-/**
- * Trace the beginning of an asynchronous event. Unlike ATRACE_BEGIN/ATRACE_END
- * contexts, asynchronous events do not need to be nested. The name describes
- * the event, and the cookie provides a unique identifier for distinguishing
- * simultaneous events. The name and cookie used to begin an event must be
- * used to end it.
- */
 #define SYSTRACE_ASYNC_BEGIN(name, cookie) \
-        ___trctx.asyncBegin(SYSTRACE_TAG, name, cookie)
+        TRACE_EVENT_BEGIN(UTILS_PERFETTO_CATEGORY, name, perfetto::Track(cookie))
 
-/**
- * Trace the end of an asynchronous event.
- * This should have a corresponding SYSTRACE_ASYNC_BEGIN.
- */
 #define SYSTRACE_ASYNC_END(name, cookie) \
-        ___trctx.asyncEnd(SYSTRACE_TAG, name, cookie)
+        TRACE_EVENT_END(UTILS_PERFETTO_CATEGORY, perfetto::Track(cookie))
+
+#define SYSTRACE_FRAME_ID(frame) \
+    TRACE_EVENT_INSTANT(UTILS_PERFETTO_CATEGORY, "frame", "id", frame)
 
-/**
- * Traces an integer counter value.  name is used to identify the counter.
- * This can be used to track how a value changes over time.
- */
 #define SYSTRACE_VALUE32(name, val) \
-        ___trctx.value(SYSTRACE_TAG, name, int32_t(val))
+    TRACE_COUNTER(UTILS_PERFETTO_CATEGORY, name, val)
 
 #define SYSTRACE_VALUE64(name, val) \
-        ___trctx.value(SYSTRACE_TAG, name, int64_t(val))
+    TRACE_COUNTER(UTILS_PERFETTO_CATEGORY, name, val)
 
-// ------------------------------------------------------------------------------------------------
-// No user serviceable code below...
-// ------------------------------------------------------------------------------------------------
-
-namespace utils {
-namespace details {
-
-class UTILS_PUBLIC Systrace {
-   public:
-
-    enum tags {
-        NEVER       = SYSTRACE_TAG_NEVER,
-        ALWAYS      = SYSTRACE_TAG_ALWAYS,
-        FILAMENT    = SYSTRACE_TAG_FILAMENT,
-        JOBSYSTEM   = SYSTRACE_TAG_JOBSYSTEM
-        // we could define more TAGS here, as we need them.
-    };
-
-    explicit Systrace(uint32_t tag) noexcept {
-        if (tag) init(tag);
-    }
-
-    static void enable(uint32_t tags) noexcept;
-    static void disable(uint32_t tags) noexcept;
-
-
-    inline void traceBegin(uint32_t tag, const char* name) noexcept {
-        if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
-            beginSection(this, name);
-        }
-    }
-
-    inline void traceEnd(uint32_t tag) noexcept {
-        if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
-            endSection(this);
-        }
-    }
-
-    inline void asyncBegin(uint32_t tag, const char* name, int32_t cookie) noexcept {
-        if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
-            beginAsyncSection(this, name, cookie);
-        }
-    }
-
-    inline void asyncEnd(uint32_t tag, const char* name, int32_t cookie) noexcept {
-        if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
-            endAsyncSection(this, name, cookie);
-        }
-    }
-
-    inline void value(uint32_t tag, const char* name, int32_t value) noexcept {
-        if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
-            setCounter(this, name, value);
-        }
-    }
-
-    inline void value(uint32_t tag, const char* name, int64_t value) noexcept {
-        if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
-            setCounter(this, name, value);
-        }
-    }
-
-   private:
-    friend class ScopedTrace;
-
-    // whether tracing is supported at all by the platform
-
-    using ATrace_isEnabled_t          = bool (*)();
-    using ATrace_beginSection_t       = void (*)(const char* sectionName);
-    using ATrace_endSection_t         = void (*)();
-    using ATrace_beginAsyncSection_t  = void (*)(const char* sectionName, int32_t cookie);
-    using ATrace_endAsyncSection_t    = void (*)(const char* sectionName, int32_t cookie);
-    using ATrace_setCounter_t         = void (*)(const char* counterName, int64_t counterValue);
-
-    struct GlobalState {
-        bool isTracingAvailable;
-        std::atomic<uint32_t> isTracingEnabled;
-        int markerFd;
-
-        ATrace_isEnabled_t ATrace_isEnabled;
-        ATrace_beginSection_t ATrace_beginSection;
-        ATrace_endSection_t ATrace_endSection;
-        ATrace_beginAsyncSection_t ATrace_beginAsyncSection;
-        ATrace_endAsyncSection_t ATrace_endAsyncSection;
-        ATrace_setCounter_t ATrace_setCounter;
-
-        void (*beginSection)(Systrace* that, const char* name);
-        void (*endSection)(Systrace* that);
-        void (*beginAsyncSection)(Systrace* that, const char* name, int32_t cookie);
-        void (*endAsyncSection)(Systrace* that, const char* name, int32_t cookie);
-        void (*setCounter)(Systrace* that, const char* name, int64_t value);
-    };
-
-    static GlobalState sGlobalState;
-
-
-    // per-instance versions for better performance
-    ATrace_isEnabled_t ATrace_isEnabled;
-    ATrace_beginSection_t ATrace_beginSection;
-    ATrace_endSection_t ATrace_endSection;
-    ATrace_beginAsyncSection_t ATrace_beginAsyncSection;
-    ATrace_endAsyncSection_t ATrace_endAsyncSection;
-    ATrace_setCounter_t ATrace_setCounter;
-
-    void (*beginSection)(Systrace* that, const char* name);
-    void (*endSection)(Systrace* that);
-    void (*beginAsyncSection)(Systrace* that, const char* name, int32_t cookie);
-    void (*endAsyncSection)(Systrace* that, const char* name, int32_t cookie);
-    void (*setCounter)(Systrace* that, const char* name, int64_t value);
-
-    void init(uint32_t tag) noexcept;
-
-    // cached values for faster access, no need to be initialized
-    bool mIsTracingEnabled;
-    int mMarkerFd = -1;
-    pid_t mPid;
-
-    static void setup() noexcept;
-    static void init_once() noexcept;
-    static bool isTracingEnabled(uint32_t tag) noexcept;
-
-    static void begin_body(int fd, int pid, const char* name) noexcept;
-    static void end_body(int fd, int pid) noexcept;
-    static void async_begin_body(int fd, int pid, const char* name, int32_t cookie) noexcept;
-    static void async_end_body(int fd, int pid, const char* name, int32_t cookie) noexcept;
-    static void int64_body(int fd, int pid, const char* name, int64_t value) noexcept;
-};
-
-// ------------------------------------------------------------------------------------------------
-
-class UTILS_PUBLIC ScopedTrace {
-public:
-    // we don't inline this because it's relatively heavy due to a global check
-    ScopedTrace(uint32_t tag, const char* name) noexcept: mTrace(tag), mTag(tag) {
-        mTrace.traceBegin(tag, name);
-    }
-
-    inline ~ScopedTrace() noexcept {
-        mTrace.traceEnd(mTag);
-    }
-
-private:
-    Systrace mTrace;
-    const uint32_t mTag;
-};
-
-} // namespace details
-} // namespace utils
+#endif // SYSTRACE_TAG == SYSTRACE_TAG_DISABLED
 
 #endif // TNT_UTILS_ANDROID_SYSTRACE_H

libs/utils/include/utils/darwin/Systrace.h

@@ -29,13 +29,25 @@
 #include <utils/compiler.h>
 #include <stack>
 
+#if SYSTRACE_TAG == SYSTRACE_TAG_DISABLED
+
+#define SYSTRACE_ENABLE()
+#define SYSTRACE_CONTEXT()
+#define SYSTRACE_NAME(name)
+#define SYSTRACE_FRAME_ID(frame)
+#define SYSTRACE_NAME_BEGIN(name)
+#define SYSTRACE_NAME_END()
+#define SYSTRACE_CALL()
+#define SYSTRACE_ASYNC_BEGIN(name, cookie)
+#define SYSTRACE_ASYNC_END(name, cookie)
+#define SYSTRACE_VALUE32(name, val)
+#define SYSTRACE_VALUE64(name, val)
+
+#else
+
 // enable tracing
 #define SYSTRACE_ENABLE() ::utils::details::Systrace::enable(SYSTRACE_TAG)
 
-// disable tracing
-#define SYSTRACE_DISABLE() ::utils::details::Systrace::disable(SYSTRACE_TAG)
-
-
 /**
  * Creates a Systrace context in the current scope. needed for calling all other systrace
  * commands below.
@@ -93,6 +105,8 @@ extern thread_local std::stack<const char*> ___tracerSections;
 #define SYSTRACE_VALUE64(name, val) \
         ___tracer.value(SYSTRACE_TAG, name, int64_t(val))
 
+#endif // SYSTRACE_TAG == SYSTRACE_TAG_DISABLED
+
 // ------------------------------------------------------------------------------------------------
 // No user serviceable code below...
 // ------------------------------------------------------------------------------------------------
@@ -118,50 +132,40 @@ namespace utils {
 namespace details {
 
 class Systrace {
-   public:
-
-    enum tags {
-        NEVER       = SYSTRACE_TAG_NEVER,
-        ALWAYS      = SYSTRACE_TAG_ALWAYS,
-        FILAMENT    = SYSTRACE_TAG_FILAMENT,
-        JOBSYSTEM   = SYSTRACE_TAG_JOBSYSTEM
-        // we could define more TAGS here, as we need them.
-    };
-
+public:
     explicit Systrace(uint32_t tag) noexcept {
         if (tag) init(tag);
     }
 
-    static void enable(uint32_t tags) noexcept;
-    static void disable(uint32_t tags) noexcept;
+    static void enable(uint32_t tag) noexcept;
 
-    inline void traceBegin(uint32_t tag, const char* name) noexcept {
+    void traceBegin(uint32_t tag, const char* name) noexcept {
         if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
             APPLE_SIGNPOST_EMIT(sGlobalState.systraceLog, OS_SIGNPOST_INTERVAL_BEGIN,
                     OS_SIGNPOST_ID_EXCLUSIVE, name, name)
         }
     }
 
-    inline void traceEnd(uint32_t tag, const char* name) noexcept {
+    void traceEnd(uint32_t tag, const char* name) noexcept {
         if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
             APPLE_SIGNPOST_EMIT(sGlobalState.systraceLog, OS_SIGNPOST_INTERVAL_END,
                     OS_SIGNPOST_ID_EXCLUSIVE, name, "")
         }
     }
 
-    inline void asyncBegin(uint32_t tag, const char* name, int32_t cookie) noexcept {
+    void asyncBegin(uint32_t tag, const char* name, int32_t cookie) noexcept {
         if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
             // TODO
         }
     }
 
-    inline void asyncEnd(uint32_t tag, const char* name, int32_t cookie) noexcept {
+    void asyncEnd(uint32_t tag, const char* name, int32_t cookie) noexcept {
         if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
             // TODO
         }
     }
 
-    inline void value(uint32_t tag, const char* name, int32_t value) noexcept {
+    void value(uint32_t tag, const char* name, int32_t value) noexcept {
         if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
             char buf[64];
             snprintf(buf, 64, "%s - %d", name, value);
@@ -170,7 +174,7 @@ class Systrace {
         }
     }
 
-    inline void value(uint32_t tag, const char* name, int64_t value) noexcept {
+    void value(uint32_t tag, const char* name, int64_t value) noexcept {
         if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
             char buf[64];
             snprintf(buf, 64, "%s - %lld", name, value);
@@ -179,16 +183,16 @@ class Systrace {
         }
     }
 
-    inline void frameId(uint32_t tag, uint32_t frame) noexcept {
+    void frameId(uint32_t tag, uint32_t frame) noexcept {
         if (tag && UTILS_UNLIKELY(mIsTracingEnabled)) {
-            char buf[64]; \
-            snprintf(buf, 64, "frame %u", frame); \
+            char buf[64];
+            snprintf(buf, 64, "frame %u", frame);
             APPLE_SIGNPOST_EMIT(sGlobalState.frameIdLog, OS_SIGNPOST_EVENT,
                     OS_SIGNPOST_ID_EXCLUSIVE, "frame", buf)
         }
     }
 
-   private:
+private:
     friend class ScopedTrace;
 
     struct GlobalState {
@@ -213,25 +217,25 @@ class Systrace {
 // ------------------------------------------------------------------------------------------------
 
 class ScopedTrace {
-   public:
+public:
     // we don't inline this because it's relatively heavy due to a global check
     ScopedTrace(uint32_t tag, const char* name) noexcept : mTrace(tag), mName(name), mTag(tag) {
         mTrace.traceBegin(tag, name);
     }
 
-    inline ~ScopedTrace() noexcept {
+    ~ScopedTrace() noexcept {
         mTrace.traceEnd(mTag, mName);
     }
 
-    inline void value(uint32_t tag, const char* name, int32_t v) noexcept {
+    void value(uint32_t tag, const char* name, int32_t v) noexcept {
         mTrace.value(tag, name, v);
     }
 
-    inline void value(uint32_t tag, const char* name, int64_t v) noexcept {
+    void value(uint32_t tag, const char* name, int64_t v) noexcept {
         mTrace.value(tag, name, v);
     }
 
-   private:
+private:
     Systrace mTrace;
     const char* mName;
     const uint32_t mTag;

libs/utils/src/JobSystem.cpp

@@ -15,10 +15,8 @@
  */
 
 // Note: The overhead of SYSTRACE_TAG_JOBSYSTEM is not negligible especially with parallel_for().
-#ifndef SYSTRACE_TAG
+#define SYSTRACE_TAG SYSTRACE_TAG_DISABLED
 //#define SYSTRACE_TAG SYSTRACE_TAG_JOBSYSTEM
-#define SYSTRACE_TAG SYSTRACE_TAG_NEVER
-#endif
 
 // when SYSTRACE_TAG_JOBSYSTEM is used, enables even heavier systraces
 #define HEAVY_SYSTRACE  0

libs/utils/src/android/Systrace.cpp

@@ -14,206 +14,26 @@
  * limitations under the License.
  */
 
-#include <utils/Systrace.h>
-#include <utils/Log.h>
+#include <utils/compiler.h>
+#include <utils/android/Systrace.h>
 
-#include <cinttypes>
+#include <perfetto/perfetto.h>
 
-#include <string.h>
+PERFETTO_TRACK_EVENT_STATIC_STORAGE_IN_NAMESPACE(systrace);
 
-#include <errno.h>
-#include <fcntl.h>
-#include <pthread.h>
-#include <dlfcn.h>
+namespace {
 
-namespace utils {
-namespace details {
-
-static pthread_once_t atrace_once_control = PTHREAD_ONCE_INIT;
-
-template <typename T>
-static void loadSymbol(T*& pfn, const char *symbol) noexcept {
-    pfn = (T*)dlsym(RTLD_DEFAULT, symbol);
-}
-
-Systrace::GlobalState Systrace::sGlobalState = {};
-
-void Systrace::init_once() noexcept {
-    GlobalState& s = sGlobalState;
-
-    s.markerFd = -1;
-
-    // API 23
-    loadSymbol(s.ATrace_isEnabled, "ATrace_isEnabled");
-    loadSymbol(s.ATrace_beginSection, "ATrace_beginSection");
-    loadSymbol(s.ATrace_endSection, "ATrace_endSection");
-    // API 29
-    loadSymbol(s.ATrace_beginAsyncSection, "ATrace_beginAsyncSection");
-    loadSymbol(s.ATrace_endAsyncSection, "ATrace_endAsyncSection");
-    loadSymbol(s.ATrace_setCounter, "ATrace_setCounter");
-
-
-    const bool hasBasicAtrace = s.ATrace_isEnabled &&
-                                s.ATrace_beginSection &&
-                                s.ATrace_endSection;
-
-    const bool hasFullATrace = hasBasicAtrace &&
-                               s.ATrace_beginAsyncSection &&
-                               s.ATrace_endAsyncSection &&
-                               s.ATrace_setCounter;
-
-    if (!hasFullATrace) {
-        s.markerFd = open("/sys/kernel/debug/tracing/trace_marker", O_WRONLY | O_CLOEXEC);
+class SystraceStaticInitialization {
+public:
+    SystraceStaticInitialization() {
+        perfetto::TracingInitArgs args;
+        args.backends |= perfetto::kSystemBackend;
+        perfetto::Tracing::Initialize(args);
+        systrace::TrackEvent::Register();
     }
+};
 
-    if (hasBasicAtrace && !hasFullATrace) {
-        // no-op if we don't have all these
-        s.ATrace_beginAsyncSection = [](const char* sectionName, int32_t cookie){};
-        s.ATrace_endAsyncSection = [](const char* sectionName, int32_t cookie){};
-        s.ATrace_setCounter = [](const char* sectionName, int64_t counterValue){};
-    }
+UTILS_UNUSED SystraceStaticInitialization sSystraceStaticInitialization{};
 
-    const bool hasLegacySystrace = s.markerFd != -1;
-
-    if (hasLegacySystrace && !hasFullATrace) {
-        // use legacy
-        s.beginSection = [](Systrace* that, const char* name) {
-            begin_body(that->mMarkerFd, that->mPid, name);
-        };
-        s.endSection = [](Systrace* that) {
-            end_body(that->mMarkerFd, that->mPid);
-        };
-        s.beginAsyncSection = [](Systrace* that, const char* name, int32_t cookie) {
-            async_begin_body(that->mMarkerFd, that->mPid, name, cookie);
-        };
-        s.endAsyncSection = [](Systrace* that, const char* name, int32_t cookie) {
-            async_end_body(that->mMarkerFd, that->mPid, name, cookie);
-        };
-        s.setCounter = [](Systrace* that, const char* name, int64_t value) {
-            int64_body(that->mMarkerFd, that->mPid, name, value);
-        };
-    } else if (hasBasicAtrace) {
-        // we have at least basic ATrace
-        s.beginSection = [](Systrace* that, const char* name) {
-            that->ATrace_beginSection(name);
-        };
-        s.endSection = [](Systrace* that) {
-            that->ATrace_endSection();
-        };
-        s.beginAsyncSection = [](Systrace* that, const char* name, int32_t cookie) {
-            that->ATrace_beginAsyncSection(name, cookie);
-        };
-        s.endAsyncSection = [](Systrace* that, const char* name, int32_t cookie) {
-            that->ATrace_endAsyncSection(name, cookie);
-        };
-        s.setCounter = [](Systrace* that, const char* name, int64_t value) {
-            that->ATrace_setCounter(name, value);
-        };
-    }
-
-    s.isTracingAvailable = hasLegacySystrace || hasFullATrace || hasBasicAtrace;
 }
 
-void Systrace::setup() noexcept {
-    pthread_once(&atrace_once_control, init_once);
-}
-
-void Systrace::enable(uint32_t tags) noexcept {
-    setup();
-    if (UTILS_LIKELY(sGlobalState.isTracingAvailable)) {
-        sGlobalState.isTracingEnabled.fetch_or(tags, std::memory_order_relaxed);
-    }
-}
-
-void Systrace::disable(uint32_t tags) noexcept {
-    sGlobalState.isTracingEnabled.fetch_and(~tags, std::memory_order_relaxed);
-}
-
-// unfortunately, this generates quite a bit of code because reading a global is not
-// trivial. For this reason, we do not inline this method.
-bool Systrace::isTracingEnabled(uint32_t tag) noexcept {
-    if (tag) {
-        setup();
-        return bool((sGlobalState.isTracingEnabled.load(std::memory_order_relaxed) | SYSTRACE_TAG_ALWAYS) & tag);
-    }
-    return false;
-}
-
-// ------------------------------------------------------------------------------------------------
-
-void Systrace::init(uint32_t tag) noexcept {
-    // must be called first
-    mIsTracingEnabled = isTracingEnabled(tag);
-
-    // cache static variables for better efficiency
-    GlobalState& s = sGlobalState;
-    ATrace_isEnabled         = s.ATrace_isEnabled;
-    ATrace_beginSection      = s.ATrace_beginSection;
-    ATrace_endSection        = s.ATrace_endSection;
-    ATrace_beginAsyncSection = s.ATrace_beginAsyncSection;
-    ATrace_endAsyncSection   = s.ATrace_endAsyncSection;
-    ATrace_setCounter        = s.ATrace_setCounter;
-
-    beginSection = s.beginSection;
-    endSection = s.endSection;
-    beginAsyncSection = s.beginAsyncSection;
-    endAsyncSection = s.endAsyncSection;
-    setCounter = s.setCounter;
-
-    mMarkerFd = s.markerFd;
-
-    mPid = getpid();
-}
-
-// ------------------------------------------------------------------------------------------------
-
-/**
- * Maximum size of a message that can be logged to the trace buffer.
- * Note this message includes a tag, the pid, and the string given as the name.
- * Names should be kept short to get the most use of the trace buffer.
- */
-#define ATRACE_MESSAGE_LENGTH 512
-
-#define WRITE_MSG(format_begin, format_end, pid, name, value) {                 \
-    char buf[ATRACE_MESSAGE_LENGTH];                                            \
-    int len = snprintf(buf, sizeof(buf), format_begin "%s" format_end, pid,     \
-        name, value);                                                           \
-    if (len >= (int) sizeof(buf)) {                                             \
-        /* Given the sizeof(buf), and all of the current format buffers,        \
-         * it is impossible for name_len to be < 0 if len >= sizeof(buf). */    \
-        int name_len = strlen(name) - (len - sizeof(buf)) - 1;                  \
-        /* Truncate the name to make the message fit. */                        \
-        len = snprintf(buf, sizeof(buf), format_begin "%.*s" format_end, pid,   \
-            name_len, name, value);                                             \
-    }                                                                           \
-    write(fd, buf, len);                                                        \
-}
-
-void Systrace::begin_body(int fd, int pid, const char* name) noexcept {
-    char buf[ATRACE_MESSAGE_LENGTH];
-    ssize_t len = snprintf(buf, sizeof(buf), "B|%d|%s", pid, name);
-    if (len >= sizeof(buf)) {
-        len = sizeof(buf) - 1;
-    }
-    write(fd, buf, size_t(len));
-}
-
-void Systrace::end_body(int fd, int pid) noexcept {
-    const char END_TAG = 'E';
-    write(fd, &END_TAG, 1);
-}
-
-void Systrace::async_begin_body(int fd, int pid, const char* name, int32_t cookie) noexcept {
-    WRITE_MSG("S|%d|", "|%" PRId32, pid, name, cookie);
-}
-
-void Systrace::async_end_body(int fd, int pid, const char* name, int32_t cookie) noexcept {
-    WRITE_MSG("F|%d|", "|%" PRId32, pid, name, cookie);
-}
-
-void Systrace::int64_body(int fd, int pid, const char* name, int64_t value) noexcept {
-    WRITE_MSG("C|%d|", "|%" PRId64, pid, name, value);
-}
-
-} // namespace details
-} // namespace utils

libs/utils/src/darwin/Systrace.cpp

@@ -14,11 +14,17 @@
  * limitations under the License.
  */
 
-#include <utils/Systrace.h>
-#include <utils/Log.h>
+#include <utils/darwin/Systrace.h>
+
+#ifndef FILAMENT_APPLE_SYSTRACE
+#   define FILAMENT_APPLE_SYSTRACE 0
+#endif
 
 #if FILAMENT_APPLE_SYSTRACE
 
+#include <atomic>
+#include <stack>
+#include <stdint.h>
 #include <pthread.h>
 
 static pthread_once_t atrace_once_control = PTHREAD_ONCE_INIT;
@@ -41,21 +47,24 @@ void Systrace::setup() noexcept {
     pthread_once(&atrace_once_control, init_once);
 }
 
-void Systrace::enable(uint32_t tags) noexcept {
+void Systrace::enable(uint32_t tag) noexcept {
     setup();
-    sGlobalState.isTracingEnabled.fetch_or(tags, std::memory_order_relaxed);
+    uint32_t const mask = 1 << tag;
+    sGlobalState.isTracingEnabled.fetch_or(mask, std::memory_order_relaxed);
 }
 
-void Systrace::disable(uint32_t tags) noexcept {
-    sGlobalState.isTracingEnabled.fetch_and(~tags, std::memory_order_relaxed);
+void Systrace::disable(uint32_t tag) noexcept {
+    uint32_t const mask = 1 << tag;
+    sGlobalState.isTracingEnabled.fetch_and(~mask, std::memory_order_relaxed);
 }
 
-// unfortunately, this generates quite a bit of code because reading a global is not
+// Unfortunately, this generates quite a bit of code because reading a global is not
 // trivial. For this reason, we do not inline this method.
 bool Systrace::isTracingEnabled(uint32_t tag) noexcept {
     if (tag) {
         setup();
-        return bool((sGlobalState.isTracingEnabled.load(std::memory_order_relaxed) | SYSTRACE_TAG_ALWAYS) & tag);
+        uint32_t const mask = 1 << tag;
+        return bool(sGlobalState.isTracingEnabled.load(std::memory_order_relaxed) & mask);
     }
     return false;
 }

test/renderdiff/test.sh

@@ -41,10 +41,10 @@ function prepare_mesa() {
 #  - Run the python script that runs the test
 #  - Zip up the result
 
-set -e && set -x && prepare_mesa && \
+set -ex && prepare_mesa && \
     mkdir -p ${OUTPUT_DIR} && \
-    ./build.sh -X ${MESA_DIR} -p desktop debug gltf_viewer && \
-    python3 ${RENDERDIFF_TEST_DIR}/src/run.py \
+        CXX=`which clang++` CC=`which clang` ./build.sh -X ${MESA_DIR} -p desktop debug gltf_viewer && \
+        python3 ${RENDERDIFF_TEST_DIR}/src/run.py \
             --gltf_viewer="$(pwd)/out/cmake-debug/samples/gltf_viewer" \
             --test=${RENDERDIFF_TEST_DIR}/tests/presubmit.json \
             --output_dir=${OUTPUT_DIR} \

test/utils/get_mesa.sh

@@ -14,13 +14,34 @@
 
 #!/usr/bin/bash
 
-set -xe
+set -x
+if [[ "$GITHUB_WORKFLOW" ]]; then
+    set -e
+fi
 
-# GITHUB_CLANG_VERSION is set in build/linux/ci-common.sh
 os_name=$(uname -s)
 LLVM_VERSION=16
-MESA_DIR=$(pwd)/mesa
+MESA_VERSION=${MESA_VERSION-24.2.1}
+ORIG_DIR=$(pwd)
+MESA_DIR=${MESA_DIR-${ORIG_DIR}/mesa}
 
+if [[ "$os_name" == "Linux" ]]; then
+    sudo apt install python3-venv
+fi
+
+# Install python deps
+python3 -m venv ${ORIG_DIR}/venv
+source ${ORIG_DIR}/venv/bin/activate
+
+NEEDED_PYTHON_DEPS=("mako" "setuptools" "pyyaml")
+for cmd in "${NEEDED_PYTHON_DEPS[@]}"; do
+    if ! python3 -m pip show "${cmd}" >/dev/null 2>&1; then
+        python3 -m pip install ${cmd}
+    fi
+done
+deactivate
+
+# Install system deps
 if [[ "$os_name" == "Linux" ]]; then
     if [[ "$GITHUB_WORKFLOW" ]]; then
         # We only want to do this if it is a CI machine.
@@ -44,12 +65,16 @@ if [[ "$os_name" == "Linux" ]]; then
         set -e
         CURRENT_CLANG_VERSION=$(clang --version | head -n 1 | awk '{ print $4 }' | awk 'BEGIN { FS="\\." } { print $1 }')
         GITHUB_CLANG_VERSION=${GITHUB_CLANG_VERSION:-${CURRENT_CLANG_VERSION}}
+        GITHUB_CLANG_VERSION=${GITHUB_CLANG_VERSION:-${LLVM_VERSION}}
         sudo apt-get -y install clang-${GITHUB_CLANG_VERSION} \
              libc++-${GITHUB_CLANG_VERSION}-dev \
              libc++abi-${GITHUB_CLANG_VERSION}-dev \
              llvm-${LLVM_VERSION} \
              llvm-${LLVM_VERSION}-{dev,tools,runtime}
-     fi # [[ "$GITHUB_WORKFLOW" ]]
+        ! command -v clang > /dev/null 2>&1 && \
+            sudo ln -s /usr/bin/clang-${GITHUB_CLANG_VERSION} /usr/bin/clang && \
+            sudo ln -s /usr/bin/clang++-${GITHUB_CLANG_VERSION} /usr/bin/clang++
+    fi # [[ "$GITHUB_WORKFLOW" ]]
 elif [[ "$os_name" == "Darwin" ]]; then
     if [[ ! "$GITHUB_WORKFLOW" ]]; then
         if [ ! command -v brew > /dev/null 2>&1 ]; then
@@ -57,14 +82,7 @@ elif [[ "$os_name" == "Darwin" ]]; then
             exit 1
         fi
     fi
-    brew install autoconf automake libx11 libxext libxrandr llvm@${LLVM_VERSION} ninja meson pkg-config libxshmfence
-
-    NEEDED_PYTHON_DEPS=("mako" "setuptools")
-    for cmd in "${NEEDED_PYTHON_DEPS[@]}"; do
-        if ! pip3 show "${cmd}" >/dev/null 2>&1; then
-            sudo pip3 install ${cmd}
-        fi
-    done
+    HOMEBREW_NO_INSTALLED_DEPENDENTS_CHECK=true brew install autoconf automake libx11 libxext libxrandr llvm@${LLVM_VERSION} ninja meson pkg-config libxshmfence
 fi # [[ "$os_name" == x ]]
 
 LOCAL_LDFLAGS=${LDFLAGS}
@@ -88,10 +106,8 @@ fi
 if [[ "$CHECKOUT_MESA" = "true" ]]; then
     rm -rf ${MESA_DIR}
 
-    #git clone https://gitlab.freedesktop.org/mesa/mesa.git
-    git clone git://anongit.freedesktop.org/mesa/mesa
-    # Due to gitlab mesa outage.    
-    if [[ "${MESA_DIR}" != "$(pwd)/mesa" ]]; then
+    git clone https://gitlab.freedesktop.org/mesa/mesa.git
+    if [[ "${MESA_DIR}" != "${ORIG_DIR}/mesa" ]]; then
         mv mesa ${MESA_DIR}
     fi
 fi
@@ -100,10 +116,12 @@ pushd .
 cd ${MESA_DIR}
 
 # Need >= 24 to have llvmpipe for swrast.  llvmpipe is needed for GL >= 4.1.
-git checkout mesa-24.2.1
+git checkout mesa-${MESA_VERSION}
 
 mkdir -p out
 
+source ${ORIG_DIR}/venv/bin/activate
+
 if  [[ "$os_name" == "Darwin" ]]; then
     LOCAL_LDFLAGS="-L/opt/homebrew/opt/llvm@${LLVM_VERSION}/lib"
     LOCAL_CPPFLAGS="-I/opt/homebrew/opt/llvm@${LLVM_VERSION}/include -I/opt/homebrew/include"
@@ -118,8 +136,16 @@ fi
 # -Dvulkan-drivers=swrast   => builds VK software rasterizer
 # -Dgallium-drivers=llvmpipe is needed for GL >= 4.1 pipe-screen (see src/gallium/auxiliary/target-helpers/inline_sw_helper.h)
 CXX=${LOCAL_CXX} CC=${LOCAL_CC} PATH=${LOCAL_PATH} LDFLAGS=${LOCAL_LDFLAGS} CPPFLAGS=${LOCAL_CPPFLAGS} \
-   meson setup --wipe builddir/ -Dprefix="$(pwd)/out" -Dglx=xlib -Dosmesa=true -Dgallium-drivers=llvmpipe,swrast -Dvulkan-drivers=swrast
+   meson setup --wipe builddir/ -Dprefix="${MESA_DIR}/out" -Dglx=xlib -Dosmesa=true -Dgallium-drivers=llvmpipe,swrast -Dvulkan-drivers=swrast
 CXX=${LOCAL_CXX} CC=${LOCAL_CC} PATH=${LOCAL_PATH} LDFLAGS=${LOCAL_LDFLAGS} CPPFLAGS=${LOCAL_CPPFLAGS} \
    meson install -C builddir/
 
+# Disable python venv
+deactivate
+
 popd
+
+if [[ "$GITHUB_WORKFLOW" ]]; then
+    set +e
+fi
+set +x

third_party/perfetto/tnt/CMakeLists.txt

@@ -0,0 +1,30 @@
+cmake_minimum_required(VERSION 3.19)
+project(perfetto)
+
+set(OUR_DIR ${CMAKE_CURRENT_SOURCE_DIR}/..)
+
+set(TARGET              perfetto)
+set(SRC_DIR             ${OUR_DIR}/${TARGET})
+set(PUBLIC_HDR_DIR      ${OUR_DIR})
+
+set(PUBLIC_HDRS
+    ${PUBLIC_HDR_DIR}/${TARGET}/perfetto.h
+)
+
+add_library(${TARGET} STATIC ${PUBLIC_HDRS} ${SRC_DIR}/perfetto.cc)
+
+if (WIN32)
+    # The perfetto library contains many symbols, so it needs the big object
+    # format.
+    target_compile_options(perfetto PRIVATE "/bigobj")
+    # Disable legacy features in windows.h.
+    add_definitions(-DWIN32_LEAN_AND_MEAN -DNOMINMAX)
+    # On Windows we should link to WinSock2.
+    target_link_libraries(${TARGET} ws2_32)
+endif (WIN32)
+
+# specify where the public headers of this library are
+target_include_directories(${TARGET} PUBLIC ${PUBLIC_HDR_DIR})
+set_target_properties(${TARGET} PROPERTIES FOLDER ThirdParty)
+
+install(TARGETS ${TARGET} ARCHIVE DESTINATION lib/${DIST_DIR})

third_party/perfetto/tnt/README.txt

@@ -0,0 +1,8 @@
+Useful links:
+https://perfetto.dev/
+https://perfetto.dev/docs/instrumentation/tracing-sdk
+
+Perfetto was fetched as:
+git clone https://github.com/google/perfetto.git -b v50.1
+
+Only the sdk/ directory is needed.