Improve code-correctness script (#8659)

- Sort failures alphabetically
- Add way to suppress error when necessary
- Add way to run test on a subset of files for testing
- Fix case where number of files is less than number of workers

.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

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)

build/common/check-headers.sh

@@ -56,10 +56,16 @@ popd >/dev/null
 rm -rf out/check-headers
 mkdir -p out/check-headers
 
+TMP_FILE=out/check-headers/temp.cpp
 echo "Checking that public headers compile independently..."
 for include in "${includes[@]}"; do
+    rm -f ${TMP_FILE}
     echo "Checking ${include}"
-    echo "#include <${include}>" >> out/check-headers/temp.cpp
-    clang -std=c++17 -I "${FILAMENT_HEADERS}" out/check-headers/temp.cpp -c -o /dev/null
+    if [[ "${include}" == "utils/Systrace.h" ]]; then
+        # A necessary define before we can include utils/Systrace.h
+        echo "#define SYSTRACE_TAG SYSTRACE_TAG_DISABLED" >> ${TMP_FILE}
+    fi
+    echo "#include <${include}>" >> ${TMP_FILE}
+    clang -std=c++17 -I "${FILAMENT_HEADERS}" ${TMP_FILE} -c -o /dev/null
 done
 echo "Done!"

filament/backend/CMakeLists.txt

@@ -194,6 +194,8 @@ if (FILAMENT_SUPPORTS_VULKAN)
             src/vulkan/VulkanDriver.cpp
             src/vulkan/VulkanDriver.h
             src/vulkan/VulkanDriverFactory.h
+            src/vulkan/VulkanExternalImageManager.cpp
+            src/vulkan/VulkanExternalImageManager.h
             src/vulkan/VulkanFboCache.cpp
             src/vulkan/VulkanFboCache.h
             src/vulkan/VulkanHandles.cpp
@@ -259,6 +261,7 @@ 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/include/backend/platforms/VulkanPlatform.h

@@ -294,6 +294,16 @@ public:
     VkQueue getProtectedGraphicsQueue() const noexcept;
 
     struct ExternalImageMetadata {
+        /**
+         * The Filament texture format.
+         */
+        TextureFormat filamentFormat;
+
+        /**
+         * The Filament texture usage.
+         */
+        TextureUsage filamentUsage;
+
         /**
          * The width of the external image
          */
@@ -304,11 +314,6 @@ public:
          */
         uint32_t height;
 
-        /**
-         * The layerCount of the external image
-         */
-        uint32_t layerCount;
-
         /**
          * The layer count of the external image
          */
@@ -324,11 +329,6 @@ public:
          */
         VkFormat format;
 
-        /**
-         * An external buffer can be protected. This tells you if it is.
-         */
-        bool isProtected;
-
         /**
          * The type of external format (opaque int) if used.
          */
@@ -348,20 +348,61 @@ public:
          * Heap information
          */
         uint32_t memoryTypeBits;
+
+        /**
+         * Ycbcr conversion components
+         */
+        VkComponentMapping ycbcrConversionComponents;
+
+        /**
+         * Ycbcr model
+         */
+        VkSamplerYcbcrModelConversion ycbcrModel;
+
+        /**
+         * Ycbcr range
+         */
+        VkSamplerYcbcrRange ycbcrRange;
+
+        /**
+         * Ycbcr x chroma offset
+         */
+        VkChromaLocation xChromaOffset;
+
+        /**
+         * Ycbcr y chroma offset
+         */
+        VkChromaLocation yChromaOffset;
     };
-    virtual ExternalImageMetadata getExternalImageMetadata(ExternalImageHandleRef externalImage);
 
-    using ImageData = std::pair<VkImage, VkDeviceMemory>;
-    virtual ImageData createExternalImageData(ExternalImageHandleRef externalImage,
-            const ExternalImageMetadata& metadata, uint32_t memoryTypeIndex,
-            VkImageUsageFlags usage);
 
-    virtual VkSampler createExternalSampler(SamplerYcbcrConversion chroma,
-            SamplerParams sampler, uint32_t internalFormat);
+    // Note that the image metadata might change per-frame, hence we need a method for extracting
+    // it.
+    virtual ExternalImageMetadata extractExternalImageMetadata(ExternalImageHandleRef image) const {
+        return {};
+    }
 
-    virtual VkImageView createExternalImageView(SamplerYcbcrConversion chroma,
-            uint32_t internalFormat, VkImage image, VkImageSubresourceRange range,
-            VkImageViewType viewType, VkComponentMapping swizzle);
+    struct ImageData {
+        struct Bundle {
+            VkImage image = VK_NULL_HANDLE;
+            VkDeviceMemory memory = VK_NULL_HANDLE;
+
+            inline bool valid() const noexcept {
+                return image != VK_NULL_HANDLE;
+            }
+        };
+        // It's possible for the external image to also have a known VK format. We need to create an
+        // image for that in case we are not looking to use an external "sampler" with this image.
+        Bundle internal;
+
+        // If we get a externalFormat in the metadata, then we should create an image with
+        // VK_FORMAT_UNDEFINED
+        Bundle external;
+    };
+
+    virtual ImageData createVkImageFromExternal(ExternalImageHandleRef image) const {
+        return {};
+    }
 
 protected:
     virtual ExtensionSet getSwapchainInstanceExtensions() const;
@@ -374,20 +415,6 @@ private:
     // Platform dependent helper methods
     static ExtensionSet getSwapchainInstanceExtensionsImpl();
 
-    static ExternalImageMetadata getExternalImageMetadataImpl(ExternalImageHandleRef externalImage,
-            VkDevice device);
-
-    static ImageData createExternalImageDataImpl(ExternalImageHandleRef externalImage,
-            VkDevice device, const ExternalImageMetadata& metadata, uint32_t memoryTypeIndex,
-            VkImageUsageFlags usage);
-    static VkSampler createExternalSamplerImpl(VkDevice device,
-            SamplerYcbcrConversion chroma, SamplerParams sampler,
-            uint32_t internalFormat);
-    static VkImageView createExternalImageViewImpl(VkDevice device,
-            SamplerYcbcrConversion chroma, uint32_t internalFormat, VkImage image,
-            VkImageSubresourceRange range, VkImageViewType viewType,
-            VkComponentMapping swizzle);
-
     // Platform dependent helper methods
     static SurfaceBundle createVkSurfaceKHRImpl(void* nativeWindow, VkInstance instance,
             uint64_t flags) noexcept;

filament/backend/include/backend/platforms/VulkanPlatformAndroid.h

@@ -26,7 +26,7 @@ namespace filament::backend {
 
 class VulkanPlatformAndroid : public VulkanPlatform {
 public:
-    Platform::ExternalImageHandle UTILS_PUBLIC createExternalImage(AHardwareBuffer const* buffer,
+    ExternalImageHandle UTILS_PUBLIC createExternalImage(AHardwareBuffer const* buffer,
             bool sRGB) noexcept;
 
     struct UTILS_PUBLIC ExternalImageDescAndroid {
@@ -39,31 +39,26 @@ public:
     ExternalImageDescAndroid UTILS_PUBLIC getExternalImageDesc(
             ExternalImageHandleRef externalImage) const noexcept;
 
+    virtual ExternalImageMetadata extractExternalImageMetadata(
+            ExternalImageHandleRef image) const override;
+
+    virtual ImageData createVkImageFromExternal(ExternalImageHandleRef image) const override;
+
 protected:
+    virtual ExtensionSet getSwapchainInstanceExtensions() const override;
+
+    using SurfaceBundle = VulkanPlatform::SurfaceBundle;
+    virtual SurfaceBundle createVkSurfaceKHR(void* nativeWindow, VkInstance instance,
+            uint64_t flags) const noexcept override;
+
+private:
     struct ExternalImageVulkanAndroid : public Platform::ExternalImage {
         AHardwareBuffer* aHardwareBuffer = nullptr;
         bool sRGB = false;
-        unsigned int width;  // Texture width
-        unsigned int height; // Texture height
-        TextureFormat format;// Texture format
-        TextureUsage usage;  // Texture usage flags
 
     protected:
         ~ExternalImageVulkanAndroid() override;
     };
-
-    virtual ExternalImageMetadata getExternalImageMetadata(ExternalImageHandleRef externalImage);
-
-    using ImageData = VulkanPlatform::ImageData;
-    virtual ImageData createExternalImageData(ExternalImageHandleRef externalImage,
-            const ExternalImageMetadata& metadata, uint32_t memoryTypeIndex,
-            VkImageUsageFlags usage);
-
-    virtual ExtensionSet getSwapchainInstanceExtensions() const;
-
-    using SurfaceBundle = VulkanPlatform::SurfaceBundle;
-    virtual SurfaceBundle createVkSurfaceKHR(void* nativeWindow, VkInstance instance,
-            uint64_t flags) const noexcept;
 };
 
 }// namespace filament::backend

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/metal/MetalDriver.mm

@@ -1408,8 +1408,8 @@ void MetalDriver::setRenderPrimitiveBuffer(Handle<HwRenderPrimitive> rph, Primit
     auto primitive = handle_cast<MetalRenderPrimitive>(rph);
     auto vertexBuffer = handle_cast<MetalVertexBuffer>(vbh);
     auto indexBuffer = handle_cast<MetalIndexBuffer>(ibh);
-    MetalVertexBufferInfo const* const vbi = handle_cast<MetalVertexBufferInfo>(vertexBuffer->vbih);
-    primitive->setBuffers(vbi, vertexBuffer, indexBuffer);
+    primitive->vertexBuffer = vertexBuffer;
+    primitive->indexBuffer = indexBuffer;
     primitive->type = pt;
 }
 

filament/backend/src/metal/MetalHandles.h

@@ -194,12 +194,8 @@ struct MetalIndexBuffer : public HwIndexBuffer {
 };
 
 struct MetalRenderPrimitive : public HwRenderPrimitive {
-    MetalRenderPrimitive();
-    void setBuffers(MetalVertexBufferInfo const* const vbi,
-            MetalVertexBuffer* vertexBuffer, MetalIndexBuffer* indexBuffer);
     // The pointers to MetalVertexBuffer and MetalIndexBuffer are "weak".
     // The MetalVertexBuffer and MetalIndexBuffer must outlive the MetalRenderPrimitive.
-
     MetalVertexBuffer* vertexBuffer = nullptr;
     MetalIndexBuffer* indexBuffer = nullptr;
 };
@@ -380,7 +376,6 @@ public:
 
     math::uint2 getAttachmentSize() noexcept;
 
-    bool isDefaultRenderTarget() const { return defaultRenderTarget; }
     uint8_t getSamples() const { return samples; }
 
     Attachment getDrawColorAttachment(size_t index);

filament/backend/src/metal/MetalHandles.mm

@@ -536,15 +536,6 @@ MetalIndexBuffer::MetalIndexBuffer(MetalContext& context, BufferUsage usage, uin
         uint32_t indexCount) : HwIndexBuffer(elementSize, indexCount),
         buffer(context, BufferObjectBinding::VERTEX, usage, elementSize * indexCount, true) { }
 
-MetalRenderPrimitive::MetalRenderPrimitive() {
-}
-
-void MetalRenderPrimitive::setBuffers(MetalVertexBufferInfo const* const vbi,
-        MetalVertexBuffer* vertexBuffer, MetalIndexBuffer* indexBuffer) {
-    this->vertexBuffer = vertexBuffer;
-    this->indexBuffer = indexBuffer;
-}
-
 MetalProgram::MetalProgram(MetalContext& context, Program&& program) noexcept
     : HwProgram(program.getName()), mContext(context) {
     mToken = context.shaderCompiler->createProgram(program.getName(), std::move(program));

filament/backend/src/opengl/GLUtils.cpp

@@ -16,9 +16,15 @@
 
 #include "GLUtils.h"
 
+#include "private/backend/Driver.h"
+
+#include <utils/compiler.h>
+#include <utils/ostream.h>
 #include <utils/trap.h>
 
-#include "private/backend/Driver.h"
+#include <string_view>
+
+#include <stddef.h>
 
 namespace filament::backend {
 
@@ -28,38 +34,31 @@ using namespace utils;
 namespace GLUtils {
 
 UTILS_NOINLINE
-const char* getGLError(GLenum error) noexcept {
-    const char* string = "unknown";
+std::string_view getGLErrorString(GLenum error) noexcept {
     switch (error) {
         case GL_NO_ERROR:
-            string = "GL_NO_ERROR";
-            break;
+            return "GL_NO_ERROR";
         case GL_INVALID_ENUM:
-            string = "GL_INVALID_ENUM";
-            break;
+            return "GL_INVALID_ENUM";
         case GL_INVALID_VALUE:
-            string = "GL_INVALID_VALUE";
-            break;
+            return "GL_INVALID_VALUE";
         case GL_INVALID_OPERATION:
-            string = "GL_INVALID_OPERATION";
-            break;
+            return "GL_INVALID_OPERATION";
         case GL_INVALID_FRAMEBUFFER_OPERATION:
-            string = "GL_INVALID_FRAMEBUFFER_OPERATION";
-            break;
+            return "GL_INVALID_FRAMEBUFFER_OPERATION";
         case GL_OUT_OF_MEMORY:
-            string = "GL_OUT_OF_MEMORY";
-            break;
+            return "GL_OUT_OF_MEMORY";
         default:
             break;
     }
-    return string;
+    return "unknown";
 }
 
 UTILS_NOINLINE
 GLenum checkGLError(io::ostream& out, const char* function, size_t line) noexcept {
     GLenum const error = glGetError();
-    if (error != GL_NO_ERROR) {
-        const char* string = getGLError(error);
+    if (UTILS_VERY_UNLIKELY(error != GL_NO_ERROR)) {
+        auto const string = getGLErrorString(error);
         out << "OpenGL error " << io::hex << error << " (" << string << ") in \""
             << function << "\" at line " << io::dec << line << io::endl;
     }
@@ -69,46 +68,39 @@ GLenum checkGLError(io::ostream& out, const char* function, size_t line) noexcep
 UTILS_NOINLINE
 void assertGLError(io::ostream& out, const char* function, size_t line) noexcept {
     GLenum const err = checkGLError(out, function, line);
-    if (err != GL_NO_ERROR) {
+    if (UTILS_VERY_UNLIKELY(err != GL_NO_ERROR)) {
         debug_trap();
     }
 }
 
 UTILS_NOINLINE
-const char* getFramebufferStatus(GLenum status) noexcept {
-    const char* string = "unknown";
+std::string_view getFramebufferStatusString(GLenum status) noexcept {
     switch (status) {
         case GL_FRAMEBUFFER_COMPLETE:
-            string = "GL_FRAMEBUFFER_COMPLETE";
-            break;
+            return "GL_FRAMEBUFFER_COMPLETE";
         case GL_FRAMEBUFFER_INCOMPLETE_ATTACHMENT:
-            string = "GL_FRAMEBUFFER_INCOMPLETE_ATTACHMENT";
-            break;
+            return "GL_FRAMEBUFFER_INCOMPLETE_ATTACHMENT";
         case GL_FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT:
-            string = "GL_FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT";
-            break;
+            return "GL_FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT";
         case GL_FRAMEBUFFER_UNSUPPORTED:
-            string = "GL_FRAMEBUFFER_UNSUPPORTED";
-            break;
+            return "GL_FRAMEBUFFER_UNSUPPORTED";
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
         case GL_FRAMEBUFFER_UNDEFINED:
-            string = "GL_FRAMEBUFFER_UNDEFINED";
-            break;
+            return "GL_FRAMEBUFFER_UNDEFINED";
         case GL_FRAMEBUFFER_INCOMPLETE_MULTISAMPLE:
-            string = "GL_FRAMEBUFFER_INCOMPLETE_MULTISAMPLE";
-            break;
+            return "GL_FRAMEBUFFER_INCOMPLETE_MULTISAMPLE";
 #endif
         default:
             break;
     }
-    return string;
+    return "unknown";
 }
 
 UTILS_NOINLINE
 GLenum checkFramebufferStatus(io::ostream& out, GLenum target, const char* function, size_t line) noexcept {
     GLenum const status = glCheckFramebufferStatus(target);
-    if (status != GL_FRAMEBUFFER_COMPLETE) {
-        const char* string = getFramebufferStatus(status);
+    if (UTILS_VERY_UNLIKELY(status != GL_FRAMEBUFFER_COMPLETE)) {
+        auto const string = getFramebufferStatusString(status);
         out << "OpenGL framebuffer error " << io::hex << status << " (" << string << ") in \""
             << function << "\" at line " << io::dec << line << io::endl;
     }
@@ -118,7 +110,7 @@ GLenum checkFramebufferStatus(io::ostream& out, GLenum target, const char* funct
 UTILS_NOINLINE
 void assertFramebufferStatus(io::ostream& out, GLenum target, const char* function, size_t line) noexcept {
     GLenum const status = checkFramebufferStatus(out, target, function, line);
-    if (status != GL_FRAMEBUFFER_COMPLETE) {
+    if (UTILS_VERY_UNLIKELY(status != GL_FRAMEBUFFER_COMPLETE)) {
         debug_trap();
     }
 }

filament/backend/src/opengl/GLUtils.h

@@ -17,29 +17,26 @@
 #ifndef TNT_FILAMENT_BACKEND_OPENGL_GLUTILS_H
 #define TNT_FILAMENT_BACKEND_OPENGL_GLUTILS_H
 
-#include <utils/compiler.h>
 #include <utils/debug.h>
-#include <utils/Log.h>
+#include <utils/ostream.h>
 
 #include <backend/DriverEnums.h>
 
 #include <string_view>
 #include <unordered_set>
 
-#include <stdint.h>
 #include <stddef.h>
-#include <string.h>
+#include <stdint.h>
 
 #include "gl_headers.h"
 
-namespace filament::backend {
-namespace GLUtils {
+namespace filament::backend::GLUtils {
 
-const char* getGLError(GLenum error) noexcept;
+std::string_view getGLErrorString(GLenum error) noexcept;
 GLenum checkGLError(utils::io::ostream& out, const char* function, size_t line) noexcept;
 void assertGLError(utils::io::ostream& out, const char* function, size_t line) noexcept;
 
-const char* getFramebufferStatus(GLenum err) noexcept;
+std::string_view getFramebufferStatusString(GLenum err) noexcept;
 GLenum checkFramebufferStatus(utils::io::ostream& out, GLenum target, const char* function, size_t line) noexcept;
 void assertFramebufferStatus(utils::io::ostream& out, GLenum target, const char* function, size_t line) noexcept;
 
@@ -53,7 +50,7 @@ void assertFramebufferStatus(utils::io::ostream& out, GLenum target, const char*
 #   define CHECK_GL_FRAMEBUFFER_STATUS(out, target) { GLUtils::checkFramebufferStatus(out, target, __func__, __LINE__); }
 #endif
 
-constexpr GLuint getComponentCount(ElementType type) noexcept {
+constexpr GLuint getComponentCount(ElementType const type) noexcept {
     using ElementType = ElementType;
     switch (type) {
         case ElementType::BYTE:
@@ -87,27 +84,29 @@ constexpr GLuint getComponentCount(ElementType type) noexcept {
         case ElementType::USHORT4:
             return 4;
     }
+    // should never happen
+    return 1;
 }
 
 // ------------------------------------------------------------------------------------------------
 // Our enums to GLenum conversions
 // ------------------------------------------------------------------------------------------------
 
-constexpr GLbitfield getAttachmentBitfield(TargetBufferFlags flags) noexcept {
+constexpr GLbitfield getAttachmentBitfield(TargetBufferFlags const flags) noexcept {
     GLbitfield mask = 0;
     if (any(flags & TargetBufferFlags::COLOR_ALL)) {
-        mask |= (GLbitfield)GL_COLOR_BUFFER_BIT;
+        mask |= GLbitfield(GL_COLOR_BUFFER_BIT);
     }
     if (any(flags & TargetBufferFlags::DEPTH)) {
-        mask |= (GLbitfield)GL_DEPTH_BUFFER_BIT;
+        mask |= GLbitfield(GL_DEPTH_BUFFER_BIT);
     }
     if (any(flags & TargetBufferFlags::STENCIL)) {
-        mask |= (GLbitfield)GL_STENCIL_BUFFER_BIT;
+        mask |= GLbitfield(GL_STENCIL_BUFFER_BIT);
     }
     return mask;
 }
 
-constexpr GLenum getBufferUsage(BufferUsage usage) noexcept {
+constexpr GLenum getBufferUsage(BufferUsage const usage) noexcept {
     switch (usage) {
         case BufferUsage::STATIC:
             return GL_STATIC_DRAW;
@@ -116,7 +115,7 @@ constexpr GLenum getBufferUsage(BufferUsage usage) noexcept {
     }
 }
 
-constexpr GLenum getBufferBindingType(BufferObjectBinding bindingType) noexcept {
+constexpr GLenum getBufferBindingType(BufferObjectBinding const bindingType) noexcept {
     switch (bindingType) {
         case BufferObjectBinding::VERTEX:
             return GL_ARRAY_BUFFER;
@@ -135,13 +134,15 @@ constexpr GLenum getBufferBindingType(BufferObjectBinding bindingType) noexcept
             return 0x90D2; // just to return something
 #endif
     }
+    // should never happen
+    return GL_ARRAY_BUFFER;
 }
 
-constexpr GLboolean getNormalization(bool normalized) noexcept {
+constexpr GLboolean getNormalization(bool const normalized) noexcept {
     return GLboolean(normalized ? GL_TRUE : GL_FALSE);
 }
 
-constexpr GLenum getComponentType(ElementType type) noexcept {
+constexpr GLenum getComponentType(ElementType const type) noexcept {
     using ElementType = ElementType;
     switch (type) {
         case ElementType::BYTE:
@@ -184,9 +185,11 @@ constexpr GLenum getComponentType(ElementType type) noexcept {
             return GL_HALF_FLOAT_OES;
 #endif
     }
+    // should never happen
+    return GL_INT;
 }
 
-constexpr GLenum getTextureTargetNotExternal(SamplerType target) noexcept {
+constexpr GLenum getTextureTargetNotExternal(SamplerType const target) noexcept {
     switch (target) {
         case SamplerType::SAMPLER_2D:
             return GL_TEXTURE_2D;
@@ -202,14 +205,16 @@ constexpr GLenum getTextureTargetNotExternal(SamplerType target) noexcept {
             // we should never be here
             return GL_TEXTURE_2D;
     }
+    // should never happen
+    return GL_TEXTURE_2D;
 }
 
-constexpr GLenum getCubemapTarget(uint16_t layer) noexcept {
+constexpr GLenum getCubemapTarget(uint16_t const layer) noexcept {
     assert_invariant(layer <= 5);
     return GL_TEXTURE_CUBE_MAP_POSITIVE_X + layer;
 }
 
-constexpr GLenum getWrapMode(SamplerWrapMode mode) noexcept {
+constexpr GLenum getWrapMode(SamplerWrapMode const mode) noexcept {
     using SamplerWrapMode = SamplerWrapMode;
     switch (mode) {
         case SamplerWrapMode::REPEAT:
@@ -219,6 +224,8 @@ constexpr GLenum getWrapMode(SamplerWrapMode mode) noexcept {
         case SamplerWrapMode::MIRRORED_REPEAT:
             return GL_MIRRORED_REPEAT;
     }
+    // should never happen
+    return GL_CLAMP_TO_EDGE;
 }
 
 constexpr GLenum getTextureFilter(SamplerMinFilter filter) noexcept {
@@ -234,6 +241,8 @@ constexpr GLenum getTextureFilter(SamplerMinFilter filter) noexcept {
             return GL_NEAREST_MIPMAP_NEAREST
                    - GLenum(SamplerMinFilter::NEAREST_MIPMAP_NEAREST) + GLenum(filter);
     }
+    // should never happen
+    return GL_NEAREST;
 }
 
 constexpr GLenum getTextureFilter(SamplerMagFilter filter) noexcept {
@@ -241,7 +250,7 @@ constexpr GLenum getTextureFilter(SamplerMagFilter filter) noexcept {
 }
 
 
-constexpr GLenum getBlendEquationMode(BlendEquation mode) noexcept {
+constexpr GLenum getBlendEquationMode(BlendEquation const mode) noexcept {
     using BlendEquation = BlendEquation;
     switch (mode) {
         case BlendEquation::ADD:               return GL_FUNC_ADD;
@@ -250,9 +259,11 @@ constexpr GLenum getBlendEquationMode(BlendEquation mode) noexcept {
         case BlendEquation::MIN:               return GL_MIN;
         case BlendEquation::MAX:               return GL_MAX;
     }
+    // should never happen
+    return GL_FUNC_ADD;
 }
 
-constexpr GLenum getBlendFunctionMode(BlendFunction mode) noexcept {
+constexpr GLenum getBlendFunctionMode(BlendFunction const mode) noexcept {
     using BlendFunction = BlendFunction;
     switch (mode) {
         case BlendFunction::ZERO:                  return GL_ZERO;
@@ -267,9 +278,11 @@ constexpr GLenum getBlendFunctionMode(BlendFunction mode) noexcept {
         case BlendFunction::ONE_MINUS_DST_ALPHA:   return GL_ONE_MINUS_DST_ALPHA;
         case BlendFunction::SRC_ALPHA_SATURATE:    return GL_SRC_ALPHA_SATURATE;
     }
+    // should never happen
+    return GL_ONE;
 }
 
-constexpr GLenum getCompareFunc(SamplerCompareFunc func) noexcept {
+constexpr GLenum getCompareFunc(SamplerCompareFunc const func) noexcept {
     switch (func) {
         case SamplerCompareFunc::LE:    return GL_LEQUAL;
         case SamplerCompareFunc::GE:    return GL_GEQUAL;
@@ -280,28 +293,30 @@ constexpr GLenum getCompareFunc(SamplerCompareFunc func) noexcept {
         case SamplerCompareFunc::A:     return GL_ALWAYS;
         case SamplerCompareFunc::N:     return GL_NEVER;
     }
+    // should never happen
+    return GL_LEQUAL;
 }
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-constexpr GLenum getTextureCompareMode(SamplerCompareMode mode) noexcept {
+constexpr GLenum getTextureCompareMode(SamplerCompareMode const mode) noexcept {
     return mode == SamplerCompareMode::NONE ?
            GL_NONE : GL_COMPARE_REF_TO_TEXTURE;
 }
 
-constexpr GLenum getTextureCompareFunc(SamplerCompareFunc func) noexcept {
+constexpr GLenum getTextureCompareFunc(SamplerCompareFunc const func) noexcept {
     return getCompareFunc(func);
 }
 #endif
 
-constexpr GLenum getDepthFunc(SamplerCompareFunc func) noexcept {
+constexpr GLenum getDepthFunc(SamplerCompareFunc const func) noexcept {
     return getCompareFunc(func);
 }
 
-constexpr GLenum getStencilFunc(SamplerCompareFunc func) noexcept {
+constexpr GLenum getStencilFunc(SamplerCompareFunc const func) noexcept {
     return getCompareFunc(func);
 }
 
-constexpr GLenum getStencilOp(StencilOperation op) noexcept {
+constexpr GLenum getStencilOp(StencilOperation const op) noexcept {
     switch (op) {
         case StencilOperation::KEEP:        return GL_KEEP;
         case StencilOperation::ZERO:        return GL_ZERO;
@@ -312,9 +327,11 @@ constexpr GLenum getStencilOp(StencilOperation op) noexcept {
         case StencilOperation::DECR_WRAP:   return GL_DECR_WRAP;
         case StencilOperation::INVERT:      return GL_INVERT;
     }
+    // should never happen
+    return GL_KEEP;
 }
 
-constexpr GLenum getFormat(PixelDataFormat format) noexcept {
+constexpr GLenum getFormat(PixelDataFormat const format) noexcept {
     using PixelDataFormat = PixelDataFormat;
     switch (format) {
         case PixelDataFormat::RGB:              return GL_RGB;
@@ -336,9 +353,11 @@ constexpr GLenum getFormat(PixelDataFormat format) noexcept {
         default: return GL_NONE;
 #endif
     }
+    // should never happen
+    return GL_RGBA;
 }
 
-constexpr GLenum getType(PixelDataType type) noexcept {
+constexpr GLenum getType(PixelDataType const type) noexcept {
     using PixelDataType = PixelDataType;
     switch (type) {
         case PixelDataType::UBYTE:                return GL_UNSIGNED_BYTE;
@@ -360,10 +379,12 @@ constexpr GLenum getType(PixelDataType type) noexcept {
         default: return GL_NONE;
 #endif
     }
+    // should never happen
+    return GL_UNSIGNED_INT;
 }
 
 #if !defined(__EMSCRIPTEN__)  && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
-constexpr GLenum getSwizzleChannel(TextureSwizzle c) noexcept {
+constexpr GLenum getSwizzleChannel(TextureSwizzle const c) noexcept {
     using TextureSwizzle = TextureSwizzle;
     switch (c) {
         case TextureSwizzle::SUBSTITUTE_ZERO:   return GL_ZERO;
@@ -373,10 +394,12 @@ constexpr GLenum getSwizzleChannel(TextureSwizzle c) noexcept {
         case TextureSwizzle::CHANNEL_2:         return GL_BLUE;
         case TextureSwizzle::CHANNEL_3:         return GL_ALPHA;
     }
+    // should never happen
+    return GL_RED;
 }
 #endif
 
-constexpr GLenum getCullingMode(CullingMode mode) noexcept {
+constexpr GLenum getCullingMode(CullingMode const mode) noexcept {
     switch (mode) {
         case CullingMode::NONE:
             // should never happen
@@ -388,11 +411,13 @@ constexpr GLenum getCullingMode(CullingMode mode) noexcept {
         case CullingMode::FRONT_AND_BACK:
             return GL_FRONT_AND_BACK;
     }
+    // should never happen
+    return GL_FRONT_AND_BACK;
 }
 
 // ES2 supported internal formats for texturing and how they  map to a format/type
 constexpr std::pair<GLenum, GLenum> textureFormatToFormatAndType(
-        TextureFormat format) noexcept {
+        TextureFormat const format) noexcept {
     switch (format) {
         case TextureFormat::R8:         return { 0x1909 /*GL_LUMINANCE*/, GL_UNSIGNED_BYTE };
         case TextureFormat::RGB8:       return { GL_RGB,                  GL_UNSIGNED_BYTE };
@@ -413,7 +438,7 @@ constexpr std::pair<GLenum, GLenum> textureFormatToFormatAndType(
 // clang loses it on this one, and generates a huge jump table when
 // inlined. So we don't  mark it as inline (only constexpr) which solves the problem,
 // strangely, when not inlined, clang simply generates an array lookup.
-constexpr /* inline */ GLenum getInternalFormat(TextureFormat format) noexcept {
+constexpr /* inline */ GLenum getInternalFormat(TextureFormat const format) noexcept {
     switch (format) {
 
         /* Formats supported by our ES2 implementations */
@@ -661,7 +686,7 @@ public:
 
 unordered_string_set split(const char* extensions) noexcept;
 
-} // namespace GLUtils
-} // namespace filament::backend
+} // namespace filament::backend::GLUtils
+
 
 #endif // TNT_FILAMENT_BACKEND_OPENGL_GLUTILS_H

filament/backend/src/opengl/OpenGLDriver.cpp

@@ -2098,6 +2098,7 @@ void OpenGLDriver::setAcquiredImage(Handle<HwStream> sh, void* hwbuffer, const m
 
     glstream->user_thread.pending = mPlatform.transformAcquiredImage({
             hwbuffer, cb, userData, handler });
+    glstream->user_thread.transform = transform;
 
     if (glstream->user_thread.pending.image != nullptr) {
         // If there's no pending image, do nothing. Note that GL_OES_EGL_image does not let you pass

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/opengl/platforms/CocoaExternalImage.mm

@@ -17,9 +17,11 @@
 #define COREVIDEO_SILENCE_GL_DEPRECATION
 
 #include "CocoaExternalImage.h"
-#include <utils/Panic.h>
 #include "../GLUtils.h"
 
+#include <utils/Panic.h>
+#include <utils/Log.h>
+
 namespace filament::backend {
 
 static const char *s_vertex = R"SHADER(#version 410 core

filament/backend/src/opengl/platforms/CocoaTouchExternalImage.mm

@@ -28,6 +28,7 @@
 #include <utils/compiler.h>
 #include <utils/Panic.h>
 #include <utils/debug.h>
+#include <utils/Log.h>
 
 namespace filament::backend {
 

filament/backend/src/vulkan/VulkanCommands.cpp

@@ -295,6 +295,8 @@ VulkanCommandBuffer& CommandBufferPool::getRecording() {
 }
 
 void CommandBufferPool::gc() {
+    FVK_SYSTRACE_CONTEXT();
+    FVK_SYSTRACE_START("CommandBufferPool::gc");
     ActiveBuffers reclaimed;
     mSubmitted.forEachSetBit([this,&reclaimed] (size_t index) {
         auto& buffer = mBuffers[index];
@@ -304,6 +306,7 @@ void CommandBufferPool::gc() {
         }
     });
     mSubmitted &= ~reclaimed;
+    FVK_SYSTRACE_END();
 }
 
 void CommandBufferPool::update() {
@@ -333,7 +336,9 @@ void CommandBufferPool::wait() {
     mSubmitted.forEachSetBit([this, &count, &fences] (size_t index) {
         fences[count++] = mBuffers[index]->getVkFence();
     });
-    vkWaitForFences(mDevice, count, fences, VK_TRUE, UINT64_MAX);
+    if (count) {
+        vkWaitForFences(mDevice, count, fences, VK_TRUE, UINT64_MAX);
+    }
     update();
 }
 

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/vulkan/VulkanDescriptorSetCache.cpp

@@ -31,8 +31,8 @@ namespace filament::backend {
 
 namespace {
 
-using DescriptorCount = VulkanDescriptorSetLayout::Count;
 using DescriptorSetLayoutArray = VulkanDescriptorSetCache::DescriptorSetLayoutArray;
+using DescriptorCount = VulkanDescriptorSetCache::DescriptorCount;
 
 // We create a pool for each layout as defined by the number of descriptors of each type. For
 // example, a layout of
@@ -203,11 +203,10 @@ public:
     DescriptorInfinitePool(VkDevice device)
         : mDevice(device) {}
 
-    VkDescriptorSet obtainSet(fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout) {
-        auto const vklayout = layout->getVkLayout();
+    VkDescriptorSet obtainSet(DescriptorCount const& count, VkDescriptorSetLayout vklayout) {
         DescriptorPool* sameTypePool = nullptr;
         for (auto& pool: mPools) {
-            if (!pool->canAllocate(layout->count)) {
+            if (!pool->canAllocate(count)) {
                 continue;
             }
             if (auto set = pool->obtainSet(vklayout); set != VK_NULL_HANDLE) {
@@ -225,8 +224,7 @@ public:
         }
 
         // We need to increase the set of pools by one.
-        mPools.push_back(std::make_unique<DescriptorPool>(mDevice,
-                DescriptorCount::fromLayoutBitmask(layout->bitmask), capacity));
+        mPools.push_back(std::make_unique<DescriptorPool>(mDevice, count, capacity));
         auto& pool = mPools.back();
         auto ret = pool->obtainSet(vklayout);
         assert_invariant(ret != VK_NULL_HANDLE && "failed to obtain a set?");
@@ -276,39 +274,36 @@ void VulkanDescriptorSetCache::unbind(uint8_t setIndex) {
 }
 
 void VulkanDescriptorSetCache::commit(VulkanCommandBuffer* commands,
-        VkPipelineLayout pipelineLayout, fvkutils::DescriptorSetMask const& setMask) {
+        VkPipelineLayout pipelineLayout, fvkutils::DescriptorSetMask const& useExternalSamplers,
+        fvkutils::DescriptorSetMask const& setMask) {
     // setMask indicates the set of descriptor sets the driver wants to bind, curMask is the
     // actual set of sets that *needs* to be bound.
     fvkutils::DescriptorSetMask curMask = setMask;
 
-    auto& updateSets = mStashedSets;
-    bool const pipelineLayoutIsSame = mLastBoundInfo.pipelineLayout == pipelineLayout;
-
-    if (pipelineLayoutIsSame) {
-        auto& lastBoundSets = mLastBoundInfo.boundSets;
-        setMask.forEachSetBit([&](size_t index) {
-            if (!updateSets[index] || updateSets[index] == lastBoundSets[index]) {
-                curMask.unset(index);
-            }
-        });
-        if (curMask.none() &&
-                mLastBoundInfo.setMask == setMask && mLastBoundInfo.boundSets == updateSets) {
-            return;
+    auto const& updateSets = mStashedSets;
+    curMask.forEachSetBit([&](size_t index) {
+        if (!updateSets[index]) {
+            curMask.unset(index);
         }
-    } else {
-        setMask.forEachSetBit([&](size_t index) {
-            if (!updateSets[index]) {
+    });
+
+    if (mLastBoundInfo.pipelineLayout == pipelineLayout) {
+        auto& lastBoundSets = mLastBoundInfo.boundSets;
+        curMask.forEachSetBit([&](size_t index) {
+            if (updateSets[index] == lastBoundSets[index] && !useExternalSamplers[index]) {
                 curMask.unset(index);
             }
         });
     }
 
-    curMask.forEachSetBit([&updateSets, commands, pipelineLayout](size_t index) {
+    curMask.forEachSetBit([&](size_t index) {
         // This code actually binds the descriptor sets.
         auto set = updateSets[index];
         VkCommandBuffer const cmdbuffer = commands->buffer();
+        VkDescriptorSet vkset = useExternalSamplers[index] ? set->getExternalSamplerVkSet() :
+            set->getVkSet();
         vkCmdBindDescriptorSets(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, index,
-                1, &set->getVkSet(), set->uniqueDynamicUboCount, set->getOffsets()->data());
+                1, &vkset, set->uniqueDynamicUboCount, set->getOffsets()->data());
         commands->acquire(set);
     });
 
@@ -334,7 +329,7 @@ void VulkanDescriptorSetCache::updateBuffer(fvkmemory::resource_ptr<VulkanDescri
     if (set->dynamicUboMask.test(binding)) {
         type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC;
     }
-    VkWriteDescriptorSet const descriptorWrite = {
+    VkWriteDescriptorSet descriptorWrite = {
         .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
         .dstSet = set->getVkSet(),
         .dstBinding = binding,
@@ -343,12 +338,17 @@ void VulkanDescriptorSetCache::updateBuffer(fvkmemory::resource_ptr<VulkanDescri
         .pBufferInfo = &info,
     };
     vkUpdateDescriptorSets(mDevice, 1, &descriptorWrite, 0, nullptr);
+
+    if (auto externalSamplerSet = set->getExternalSamplerVkSet();
+            externalSamplerSet != VK_NULL_HANDLE) {
+        descriptorWrite.dstSet = externalSamplerSet;
+        vkUpdateDescriptorSets(mDevice, 1, &descriptorWrite, 0, nullptr);
+    }
     set->acquire(bufferObject);
 }
 
-void VulkanDescriptorSetCache::updateSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
-        uint8_t binding, fvkmemory::resource_ptr<VulkanTexture> texture,
-        VkSampler sampler) noexcept {
+void VulkanDescriptorSetCache::updateSamplerImpl(VkDescriptorSet vkset, uint8_t binding,
+        fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept {
     VkImageSubresourceRange range = texture->getPrimaryViewRange();
     VkImageViewType const expectedType = texture->getViewType();
     if (any(texture->usage & TextureUsage::DEPTH_ATTACHMENT) &&
@@ -364,16 +364,29 @@ void VulkanDescriptorSetCache::updateSampler(fvkmemory::resource_ptr<VulkanDescr
         .imageLayout = fvkutils::getVkLayout(texture->getDefaultLayout()),
     };
 
-    VkWriteDescriptorSet const descriptorWrite = {
+    VkWriteDescriptorSet descriptorWrite = {
         .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
         .pNext = nullptr,
-        .dstSet = set->getVkSet(),
+        .dstSet = vkset,
         .dstBinding = binding,
         .descriptorCount = 1,
         .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
         .pImageInfo = &info,
     };
     vkUpdateDescriptorSets(mDevice, 1, &descriptorWrite, 0, nullptr);
+}
+
+void VulkanDescriptorSetCache::updateSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
+        uint8_t binding, fvkmemory::resource_ptr<VulkanTexture> texture,
+        VkSampler sampler) noexcept {
+    updateSamplerImpl(set->getVkSet(), binding, texture, sampler);
+    set->acquire(texture);
+}
+
+void VulkanDescriptorSetCache::updateSamplerForExternalSamplerSet(
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t binding,
+        fvkmemory::resource_ptr<VulkanTexture> texture) noexcept {
+    updateSamplerImpl(set->getExternalSamplerVkSet(), binding, texture, VK_NULL_HANDLE);
     set->acquire(texture);
 }
 
@@ -383,32 +396,32 @@ void VulkanDescriptorSetCache::updateInputAttachment(
     // TOOD: fill this in.
 }
 
+
 fvkmemory::resource_ptr<VulkanDescriptorSet> VulkanDescriptorSetCache::createSet(
         Handle<HwDescriptorSet> handle, fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout) {
-    auto const vkSet = getVkSet(layout);
+    auto const vkSet = mDescriptorPool->obtainSet(layout->count, layout->getVkLayout());
     auto const& count = layout->count;
     auto const vklayout = layout->getVkLayout();
-    auto set = fvkmemory::resource_ptr<VulkanDescriptorSet>::make(mResourceManager, handle,
-            layout->bitmask.dynamicUbo, layout->count.dynamicUbo,
-            [vkSet, count, vklayout, this](VulkanDescriptorSet*) {
-                // Note that mDescriptorPool could be gone due to terminate (when the backend shuts
-                // down).
-                if (mDescriptorPool) {
-                    mDescriptorPool->recycle(count, vklayout, vkSet);
-                }
-            });
-    set->setVkSet(vkSet);
+    auto set = fvkmemory::resource_ptr<VulkanDescriptorSet>::make(
+            mResourceManager, handle, layout->bitmask.dynamicUbo, layout->count.dynamicUbo,
+            [vkSet, count, vklayout, this](
+                    VulkanDescriptorSet*) { this->manualRecycle(count, vklayout, vkSet); },
+            vkSet);
     return set;
 }
 
-VkDescriptorSet VulkanDescriptorSetCache::getVkSet(
-        fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout) {
-    return mDescriptorPool->obtainSet(layout);
+VkDescriptorSet VulkanDescriptorSetCache::getVkSet(DescriptorCount const& count,
+        VkDescriptorSetLayout vklayout) {
+    return mDescriptorPool->obtainSet(count, vklayout);
 }
 
-void VulkanDescriptorSetCache::manualRecyle(VulkanDescriptorSetLayout::Count const& count,
+void VulkanDescriptorSetCache::manualRecycle(VulkanDescriptorSetLayout::Count const& count,
         VkDescriptorSetLayout vklayout, VkDescriptorSet vkSet) {
-    mDescriptorPool->recycle(count, vklayout, vkSet);
+    // Note that mDescriptorPool could be gone due to terminate (when the backend shuts
+    // down).
+    if (mDescriptorPool) {
+        mDescriptorPool->recycle(count, vklayout, vkSet);
+    }
 }
 
 void VulkanDescriptorSetCache::gc() { mStashedSets = {}; }

filament/backend/src/vulkan/VulkanDescriptorSetCache.h

@@ -43,6 +43,7 @@ public:
     using DescriptorSetLayoutArray = VulkanDescriptorSetLayout::DescriptorSetLayoutArray;
     using DescriptorSetArray =
             std::array<fvkmemory::resource_ptr<VulkanDescriptorSet>, UNIQUE_DESCRIPTOR_SET_COUNT>;
+    using DescriptorCount = VulkanDescriptorSetLayout::Count;
 
     VulkanDescriptorSetCache(VkDevice device, fvkmemory::ResourceManager* resourceManager);
     ~VulkanDescriptorSetCache();
@@ -56,6 +57,10 @@ public:
     void updateSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t binding,
             fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept;
 
+    void updateSamplerForExternalSamplerSet(fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t binding,
+            fvkmemory::resource_ptr<VulkanTexture> texture) noexcept;
+
+
     void updateInputAttachment(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
             VulkanAttachment const& attachment) noexcept;
 
@@ -65,17 +70,17 @@ public:
     void unbind(uint8_t setIndex);
 
     void commit(VulkanCommandBuffer* commands, VkPipelineLayout pipelineLayout,
+            fvkutils::DescriptorSetMask const& useExternalSamplerMask,
             fvkutils::DescriptorSetMask const& setMask);
 
     fvkmemory::resource_ptr<VulkanDescriptorSet> createSet(Handle<HwDescriptorSet> handle,
             fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout);
 
-    // This method is only meant to be used with external samplers (or internally within this
-    // class).
-    VkDescriptorSet getVkSet(fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout);
+    // This method is meant to be used with external samplers
+    VkDescriptorSet getVkSet(DescriptorCount const& count, VkDescriptorSetLayout vklayout);
 
-    // This method is only meant to be used with external samplers.
-    void manualRecyle(VulkanDescriptorSetLayout::Count const& count, VkDescriptorSetLayout vklayout,
+    // This method is meant to be used with external samplers
+    void manualRecycle(VulkanDescriptorSetLayout::Count const& count, VkDescriptorSetLayout vklayout,
             VkDescriptorSet vkSet);
 
     DescriptorSetArray const& getBoundSets() const { return mStashedSets; }
@@ -83,6 +88,9 @@ public:
     void gc();
 
 private:
+    void updateSamplerImpl(VkDescriptorSet set, uint8_t binding,
+            fvkmemory::resource_ptr<VulkanTexture> texture, VkSampler sampler) noexcept;
+
     class DescriptorInfinitePool;
 
     VkDevice mDevice;

filament/backend/src/vulkan/VulkanDescriptorSetLayoutCache.cpp

@@ -127,6 +127,7 @@ void VulkanDescriptorSetLayoutCache::terminate() noexcept {
 
 VkDescriptorSetLayout VulkanDescriptorSetLayoutCache::getVkLayout(
         VulkanDescriptorSetLayout::Bitmask const& bitmasks,
+        fvkutils::SamplerBitmask externalSamplers,
         utils::FixedCapacityVector<VkSampler> immutableSamplers) {
     LayoutKey key = {
         .bitmask = bitmasks,
@@ -141,7 +142,7 @@ VkDescriptorSetLayout VulkanDescriptorSetLayoutCache::getVkLayout(
     count += appendBindings(&toBind[count], VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC,
             bitmasks.dynamicUbo);
     count += appendBindings(&toBind[count], VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, bitmasks.ubo);
-    count += appendSamplerBindings(&toBind[count], bitmasks.sampler, bitmasks.externalSampler,
+    count += appendSamplerBindings(&toBind[count], bitmasks.sampler, externalSamplers,
             immutableSamplers);
     count += appendBindings(&toBind[count], VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT,
             bitmasks.inputAttachment);
@@ -160,9 +161,9 @@ VkDescriptorSetLayout VulkanDescriptorSetLayoutCache::getVkLayout(
 
 fvkmemory::resource_ptr<VulkanDescriptorSetLayout> VulkanDescriptorSetLayoutCache::createLayout(
         Handle<HwDescriptorSetLayout> handle, backend::DescriptorSetLayout&& info) {
+    BitmaskGroup maskGroup = VulkanDescriptorSetLayout::Bitmask::fromLayoutDescription(info);
     auto layout = fvkmemory::resource_ptr<VulkanDescriptorSetLayout>::make(mResourceManager, handle,
-            info);
-    layout->setVkLayout(getVkLayout(layout->bitmask));
+            std::move(info), getVkLayout(maskGroup, maskGroup.externalSampler));
     return layout;
 }
 

filament/backend/src/vulkan/VulkanDescriptorSetLayoutCache.h

@@ -46,6 +46,7 @@ public:
 
     // This method is meant to be used with external samplers
     VkDescriptorSetLayout getVkLayout(VulkanDescriptorSetLayout::Bitmask const& bitmasks,
+            fvkutils::SamplerBitmask externalSamplers,
             utils::FixedCapacityVector<VkSampler> immutableSamplers = {});
 
 private:

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -30,7 +30,6 @@
 #include "vulkan/memory/ResourcePointer.h"
 #include "vulkan/utils/Conversion.h"
 #include "vulkan/utils/Definitions.h"
-#include "vulkan/vulkan_core.h"
 
 #include <backend/DriverEnums.h>
 #include <backend/platforms/VulkanPlatform.h>
@@ -198,11 +197,9 @@ Dispatcher VulkanDriver::getDispatcher() const noexcept {
 }
 
 VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& context,
-        Platform::DriverConfig const& driverConfig) noexcept
+        Platform::DriverConfig const& driverConfig)
     : mPlatform(platform),
-      mResourceManager(
-              driverConfig.handleArenaSize,
-              driverConfig.disableHandleUseAfterFreeCheck,
+      mResourceManager(driverConfig.handleArenaSize, driverConfig.disableHandleUseAfterFreeCheck,
               driverConfig.disableHeapHandleTags),
       mAllocator(createAllocator(mPlatform->getInstance(), mPlatform->getPhysicalDevice(),
               mPlatform->getDevice())),
@@ -221,6 +218,8 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
       mDescriptorSetLayoutCache(mPlatform->getDevice(), &mResourceManager),
       mDescriptorSetCache(mPlatform->getDevice(), &mResourceManager),
       mQueryManager(mPlatform->getDevice()),
+      mExternalImageManager(platform, &mSamplerCache, &mYcbcrConversionCache, &mDescriptorSetCache,
+              &mDescriptorSetLayoutCache),
       mIsSRGBSwapChainSupported(mPlatform->getCustomization().isSRGBSwapChainSupported),
       mStereoscopicType(driverConfig.stereoscopicType) {
 
@@ -251,7 +250,7 @@ VulkanDriver::~VulkanDriver() noexcept = default;
 
 UTILS_NOINLINE
 Driver* VulkanDriver::create(VulkanPlatform* platform, VulkanContext const& context,
-         Platform::DriverConfig const& driverConfig) noexcept {
+         Platform::DriverConfig const& driverConfig) {
 #if 0
     // this is useful for development, but too verbose even for debug builds
     // For reference on a 64-bits machine in Release mode:
@@ -313,7 +312,7 @@ void VulkanDriver::terminate() {
 
     mCurrentSwapChain = {};
     mDefaultRenderTarget = {};
-    mBoundPipeline = {};
+    mPipelineState = {};
 
     mQueryManager.terminate();
 
@@ -325,6 +324,10 @@ void VulkanDriver::terminate() {
 
     mCommands.terminate();
 
+    // Must come before samplerCache, ycbcrConversionCache, descriptorSetCache,
+    // descriptorSetLayoutCache
+    mExternalImageManager.terminate();
+
     mStagePool.terminate();
     mPipelineCache.terminate();
     mFramebufferCache.terminate();
@@ -381,6 +384,10 @@ void VulkanDriver::beginFrame(int64_t monotonic_clock_ns,
         int64_t refreshIntervalNs, uint32_t frameId) {
     FVK_PROFILE_MARKER(PROFILE_NAME_BEGINFRAME);
     // Do nothing.
+
+    if (mAppState.hasExternalSamplers()) {
+        mExternalImageManager.onBeginFrame();
+    }
 }
 
 void VulkanDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch, CallbackHandler* handler,
@@ -421,12 +428,17 @@ void VulkanDriver::updateDescriptorSetTexture(
     auto set = resource_ptr<VulkanDescriptorSet>::cast(&mResourceManager, dsh);
     auto texture = resource_ptr<VulkanTexture>::cast(&mResourceManager, th);
 
-    // TODO: YcbcrConversion?
-    VulkanSamplerCache::Params cacheParams = {
-        .sampler = params,
-    };
-    VkSampler const vksampler = mSamplerCache.getSampler(cacheParams);
-    mDescriptorSetCache.updateSampler(set, binding, texture, vksampler);
+    if (mExternalImageManager.isExternallySampledTexture(texture)) {
+        mExternalImageManager.bindExternallySampledTexture(set, binding, texture, params);
+        mAppState.hasBoundExternalImages = true;
+    } else {
+        VulkanSamplerCache::Params cacheParams = {
+            .sampler = params,
+        };
+        VkSampler const vksampler = mSamplerCache.getSampler(cacheParams);
+        mDescriptorSetCache.updateSampler(set, binding, texture, vksampler);
+        mExternalImageManager.clearTextureBinding(set, binding);
+    }
 }
 
 void VulkanDriver::flush(int) {
@@ -446,6 +458,7 @@ void VulkanDriver::finish(int dummy) {
 void VulkanDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph,
         Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh,
         PrimitiveType pt) {
+    FVK_SYSTRACE_SCOPE();
     auto vb = resource_ptr<VulkanVertexBuffer>::cast(&mResourceManager, vbh);
     auto ib = resource_ptr<VulkanIndexBuffer>::cast(&mResourceManager, ibh);
     auto ptr = resource_ptr<VulkanRenderPrimitive>::make(&mResourceManager, rph, pt, vb, ib);
@@ -456,12 +469,14 @@ void VulkanDriver::destroyRenderPrimitive(Handle<HwRenderPrimitive> rph) {
     if (!rph) {
         return;
     }
+    FVK_SYSTRACE_SCOPE();
     auto ptr = resource_ptr<VulkanRenderPrimitive>::cast(&mResourceManager, rph);
     ptr.dec();
 }
 
 void VulkanDriver::createVertexBufferInfoR(Handle<HwVertexBufferInfo> vbih, uint8_t bufferCount,
         uint8_t attributeCount, AttributeArray attributes) {
+    FVK_SYSTRACE_SCOPE();
     auto vbi = resource_ptr<VulkanVertexBufferInfo>::make(&mResourceManager, vbih, bufferCount,
             attributeCount, attributes);
     vbi.inc();
@@ -471,12 +486,14 @@ void VulkanDriver::destroyVertexBufferInfo(Handle<HwVertexBufferInfo> vbih) {
     if (!vbih) {
         return;
     }
+    FVK_SYSTRACE_SCOPE();
     auto vbi = resource_ptr<VulkanVertexBufferInfo>::cast(&mResourceManager, vbih);
     vbi.dec();
 }
 
 void VulkanDriver::createVertexBufferR(Handle<HwVertexBuffer> vbh, uint32_t vertexCount,
         Handle<HwVertexBufferInfo> vbih) {
+    FVK_SYSTRACE_SCOPE();
     auto vbi = resource_ptr<VulkanVertexBufferInfo>::cast(&mResourceManager, vbih);
     auto vb = resource_ptr<VulkanVertexBuffer>::make(&mResourceManager, vbh, mContext, mStagePool,
             vertexCount, vbi);
@@ -487,12 +504,14 @@ void VulkanDriver::destroyVertexBuffer(Handle<HwVertexBuffer> vbh) {
     if (!vbh) {
         return;
     }
+    FVK_SYSTRACE_SCOPE();
     auto vb = resource_ptr<VulkanVertexBuffer>::cast(&mResourceManager, vbh);
     vb.dec();
 }
 
 void VulkanDriver::createIndexBufferR(Handle<HwIndexBuffer> ibh, ElementType elementType,
         uint32_t indexCount, BufferUsage usage) {
+    FVK_SYSTRACE_SCOPE();
     auto elementSize = (uint8_t) getElementTypeSize(elementType);
     auto ib = resource_ptr<VulkanIndexBuffer>::make(&mResourceManager, ibh, mAllocator, mStagePool,
             elementSize, indexCount);
@@ -503,12 +522,14 @@ void VulkanDriver::destroyIndexBuffer(Handle<HwIndexBuffer> ibh) {
     if (!ibh) {
         return;
     }
+    FVK_SYSTRACE_SCOPE();
     auto ib = resource_ptr<VulkanIndexBuffer>::cast(&mResourceManager, ibh);
     ib.dec();
 }
 
 void VulkanDriver::createBufferObjectR(Handle<HwBufferObject> boh, uint32_t byteCount,
         BufferObjectBinding bindingType, BufferUsage usage) {
+    FVK_SYSTRACE_SCOPE();
     auto bo = resource_ptr<VulkanBufferObject>::make(&mResourceManager, boh, mAllocator, mStagePool,
             byteCount, bindingType);
     bo.inc();
@@ -518,6 +539,7 @@ void VulkanDriver::destroyBufferObject(Handle<HwBufferObject> boh) {
     if (!boh) {
         return;
     }
+    FVK_SYSTRACE_SCOPE();
     auto bo = resource_ptr<VulkanBufferObject>::cast(&mResourceManager, boh);
     bo.dec();
 }
@@ -563,41 +585,48 @@ void VulkanDriver::createTextureExternalImage2R(Handle<HwTexture> th, backend::S
         backend::TextureFormat format, uint32_t width, uint32_t height, backend::TextureUsage usage,
         Platform::ExternalImageHandleRef externalImage) {
     FVK_SYSTRACE_SCOPE();
-    auto const& metadata = mPlatform->getExternalImageMetadata(externalImage);
-    if (metadata.isProtected) {
-        usage |= backend::TextureUsage::PROTECTED;
-    }
-
-    VkImageUsageFlags vkUsage = metadata.usage;
-    if (any(usage & TextureUsage::BLIT_SRC)) {
-        vkUsage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
-    }
-
-    if (any(usage & (TextureUsage::BLIT_DST | TextureUsage::UPLOADABLE))) {
-        vkUsage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT;
-    }
+    auto metadata = mPlatform->extractExternalImageMetadata(externalImage);
 
     assert_invariant(width == metadata.width);
     assert_invariant(height == metadata.height);
-    assert_invariant(fvkutils::getVkFormat(format) == metadata.format);
 
-    VkMemoryPropertyFlags const requiredMemoryFlags = any(usage & TextureUsage::UPLOADABLE)
-                                                              ? VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
-                                                              : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
-    uint32_t const memoryTypeIndex =
-            mContext.selectMemoryType(metadata.memoryTypeBits, requiredMemoryFlags);
-    FILAMENT_CHECK_POSTCONDITION(memoryTypeIndex != VK_MAX_MEMORY_TYPES)
-            << "failed to find a valid memory type for external image memory.";
+    // We do not check the format since AHB could return both a known format and an external format.
+    // In which case, we choose one or the other, but this choice is not known to the client.
+    // Therefore the following lines are commented out.
+    // assert_invariant(format == metadata.filamentFormat);
+    // assert_invariant(fvkutils::getVkFormat(format) == metadata.format);
 
-    VkImage vkimg;
-    VkDeviceMemory deviceMemory;
-    std::tie(vkimg, deviceMemory) =
-            mPlatform->createExternalImageData(externalImage, metadata, memoryTypeIndex, vkUsage);
+    auto imgData = mPlatform->createVkImageFromExternal(externalImage);
 
+    assert_invariant(imgData.internal.valid() || imgData.external.valid());
+
+    VkFormat vkformat = metadata.format;
+    VkImage vkimage = VK_NULL_HANDLE;
+    VkDeviceMemory memory = VK_NULL_HANDLE;
+    if (imgData.internal.valid()) {
+        metadata.externalFormat = 0;
+        vkimage = imgData.internal.image;
+        memory = imgData.internal.memory;
+    } else { // imgData.external.valid()
+        vkformat = VK_FORMAT_UNDEFINED;
+        vkimage = imgData.external.image;
+        memory = imgData.external.memory;
+    }
+
+    VkSamplerYcbcrConversion const conversion =
+            mExternalImageManager.getVkSamplerYcbcrConversion(metadata);
     auto texture = resource_ptr<VulkanTexture>::make(&mResourceManager, th, mContext,
-            mPlatform->getDevice(), mAllocator, &mResourceManager, &mCommands, vkimg, deviceMemory,
-            metadata.format, VK_NULL_HANDLE, metadata.samples, metadata.width, metadata.height,
-            metadata.layerCount, usage, mStagePool);
+            mPlatform->getDevice(), mAllocator, &mResourceManager, &mCommands, vkimage, memory,
+            vkformat, conversion, metadata.samples, metadata.width, metadata.height,
+            metadata.layers, usage, mStagePool);
+    auto& commands = mCommands.get();
+    // Unlike uploaded textures or swapchains, we need to explicit transition this
+    // texture into the read layout.
+    texture->transitionLayout(&commands, texture->getPrimaryViewRange(), VulkanLayout::READ_ONLY);
+
+    if (imgData.external.valid()) {
+        mExternalImageManager.addExternallySampledTexture(texture, externalImage);
+    }
 
     texture.inc();
 }
@@ -630,6 +659,8 @@ void VulkanDriver::destroyTexture(Handle<HwTexture> th) {
     }
     auto texture = resource_ptr<VulkanTexture>::cast(&mResourceManager, th);
     texture.dec();
+
+    mExternalImageManager.removeExternallySampledTexture(texture);
 }
 
 void VulkanDriver::createProgramR(Handle<HwProgram> ph, Program&& program) {
@@ -746,6 +777,7 @@ void VulkanDriver::createFenceR(Handle<HwFence> fh, int) {
 }
 
 void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow, uint64_t flags) {
+    FVK_SYSTRACE_SCOPE();
     // Running gc() to guard against an edge case where the old swapchains need to have been
     // destroyed before the new swapchain can be created. Otherwise, we would fail
     // vkCreateSwapchainKHR with VK_ERROR_NATIVE_WINDOW_IN_USE_KHR.
@@ -787,6 +819,7 @@ void VulkanDriver::createTimerQueryR(Handle<HwTimerQuery> tqh, int) {
 
 void VulkanDriver::createDescriptorSetLayoutR(Handle<HwDescriptorSetLayout> dslh,
         backend::DescriptorSetLayout&& info) {
+    FVK_SYSTRACE_SCOPE();
     auto layout = mDescriptorSetLayoutCache.createLayout(dslh, std::move(info));
     layout.inc();
 }
@@ -798,6 +831,10 @@ void VulkanDriver::createDescriptorSetR(Handle<HwDescriptorSet> dsh,
             fvkmemory::resource_ptr<VulkanDescriptorSetLayout>::cast(&mResourceManager, dslh);
     auto set = mDescriptorSetCache.createSet(dsh, layout);
     set.inc();
+
+    if (layout->hasExternalSamplers()) {
+        mAppState.hasExternalSamplerLayouts = true;
+    }
 }
 
 Handle<HwVertexBufferInfo> VulkanDriver::createVertexBufferInfoS() noexcept {
@@ -922,6 +959,10 @@ void VulkanDriver::destroyDescriptorSetLayout(Handle<HwDescriptorSetLayout> dslh
 void VulkanDriver::destroyDescriptorSet(Handle<HwDescriptorSet> dsh) {
     auto set = resource_ptr<VulkanDescriptorSet>::cast(&mResourceManager, dsh);
     set.dec();
+
+    if (mAppState.hasExternalSamplers() && set->getExternalSamplerVkSet() != VK_NULL_HANDLE) {
+        mExternalImageManager.removeDescriptorSet(set);
+    }
 }
 
 Handle<HwStream> VulkanDriver::createStreamNative(void* nativeStream) {
@@ -1482,7 +1523,6 @@ void VulkanDriver::endRenderPass(int) {
     // pipeline barrier between framebuffer writes and shader reads.
     rt->emitBarriersEndRenderPass(*mCurrentRenderPass.commandBuffer);
 
-    mRenderPassFboInfo = {};
     mCurrentRenderPass.renderTarget = {};
     mCurrentRenderPass.renderPass = VK_NULL_HANDLE;
 
@@ -1542,10 +1582,10 @@ void VulkanDriver::commit(Handle<HwSwapChain> sch) {
 
 void VulkanDriver::setPushConstant(backend::ShaderStage stage, uint8_t index,
         backend::PushConstantVariant value) {
-    assert_invariant(mBoundPipeline.program && "Expect a program when writing to push constants");
+    assert_invariant(mPipelineState.program && "Expect a program when writing to push constants");
     assert_invariant(mCurrentRenderPass.commandBuffer && "Should be called within a renderpass");
-    mBoundPipeline.program->writePushConstant(mCurrentRenderPass.commandBuffer->buffer(),
-            mBoundPipeline.pipelineLayout, stage, index, value);
+    mPipelineState.program->writePushConstant(mCurrentRenderPass.commandBuffer->buffer(),
+            mPipelineState.pipelineLayout, stage, index, value);
 }
 
 void VulkanDriver::insertEventMarker(char const* string) {
@@ -1724,16 +1764,69 @@ void VulkanDriver::blitDEPRECATED(TargetBufferFlags buffers,
 }
 
 void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
+    // This resets all of the pipeline states; the most relevant (needing reset) is .bindInDraw.
+    mPipelineState = {};
+
+    auto& setLayouts = pipelineState.pipelineLayout.setLayout;
+    DescriptorSetLayoutHandleList layoutHandles;
+    uint8_t layoutCount = 0;
+    std::transform(setLayouts.begin(), setLayouts.end(), layoutHandles.begin(),
+            [&](auto const& handle) -> resource_ptr<VulkanDescriptorSetLayout> {
+                if (!handle) {
+                    return {};
+                }
+                layoutCount++;
+                return resource_ptr<VulkanDescriptorSetLayout>::cast(&mResourceManager, handle);
+            });
+
+    constexpr uint8_t descriptorSetMaskTable[4] = {0x1, 0x3, 0x7, 0xF};
+    fvkutils::DescriptorSetMask const descriptorSetMask =
+            fvkutils::DescriptorSetMask(descriptorSetMaskTable[layoutCount]);
+
+    if (mAppState.hasExternalSamplers()) {
+        auto const haveExternalSamplers = [&](auto layoutHandle) {
+            if (!layoutHandle) {
+                return false;
+            }
+            return layoutHandle->hasExternalSamplers();
+        };
+        if (std::any_of(layoutHandles.begin(), layoutHandles.end(), haveExternalSamplers)) {
+            BindInDrawBundle bundle = {
+                .pipelineState = pipelineState,
+                .dsLayoutHandles = layoutHandles,
+                .descriptorSetMask = descriptorSetMask,
+            };
+            mPipelineState.bindInDraw = { true, bundle };
+            return;
+        }
+    }
+
+    // The normal, non-external sampler path
+    using VkDescriptorSetLayoutArray = VulkanPipelineLayoutCache::DescriptorSetLayoutArray;
+    VkDescriptorSetLayoutArray vkLayouts;
+    std::transform(layoutHandles.begin(), layoutHandles.end(), vkLayouts.begin(),
+            [](auto const& layout) -> VkDescriptorSetLayout {
+                if (!layout) {
+                    return VK_NULL_HANDLE;
+                }
+                return layout->getVkLayout();
+            });
+    auto program = resource_ptr<VulkanProgram>::cast(&mResourceManager, pipelineState.program);
+    auto pipelineLayout = mPipelineLayoutCache.getLayout(vkLayouts, program);
+    bindPipelineImpl(pipelineState, pipelineLayout, descriptorSetMask);
+}
+
+void VulkanDriver::bindPipelineImpl(PipelineState const& pipelineState,
+        VkPipelineLayout pipelineLayout, fvkutils::DescriptorSetMask descriptorSetMask) {
     FVK_SYSTRACE_SCOPE();
     auto commands = mCurrentRenderPass.commandBuffer;
     auto vbi = resource_ptr<VulkanVertexBufferInfo>::cast(&mResourceManager,
             pipelineState.vertexBufferInfo);
 
-    Handle<HwProgram> programHandle = pipelineState.program;
     RasterState const& rasterState = pipelineState.rasterState;
     PolygonOffset const& depthOffset = pipelineState.polygonOffset;
 
-    auto program = resource_ptr<VulkanProgram>::cast(&mResourceManager, programHandle);
+    auto program = resource_ptr<VulkanProgram>::cast(&mResourceManager, pipelineState.program);
     commands->acquire(program);
 
     // Update the VK raster state.
@@ -1775,28 +1868,11 @@ void VulkanDriver::bindPipeline(PipelineState const& pipelineState) {
     mPipelineCache.bindPrimitiveTopology(topology);
     mPipelineCache.bindVertexArray(attribDesc, bufferDesc, vbi->getAttributeCount());
 
-    auto& setLayouts = pipelineState.pipelineLayout.setLayout;
-    VulkanDescriptorSetLayout::DescriptorSetLayoutArray layoutList;
-    uint8_t layoutCount = 0;
-    std::transform(setLayouts.begin(), setLayouts.end(), layoutList.begin(),
-            [&](Handle<HwDescriptorSetLayout> handle) -> VkDescriptorSetLayout {
-                if (!handle) {
-                    return VK_NULL_HANDLE;
-                }
-                auto layout =
-                        resource_ptr<VulkanDescriptorSetLayout>::cast(&mResourceManager, handle);
-                layoutCount++;
-                return layout->getVkLayout();
-            });
-    auto pipelineLayout = mPipelineLayoutCache.getLayout(layoutList, program);
-
-    constexpr uint8_t descriptorSetMaskTable[4] = {0x1, 0x3, 0x7, 0xF};
-
-    mBoundPipeline = {
-        .program = program,
-        .pipelineLayout = pipelineLayout,
-        .descriptorSetMask = fvkutils::DescriptorSetMask(descriptorSetMaskTable[layoutCount]),
-    };
+    // Note that we cannot reinit mPipeline because the .bindInDraw metadata that needs to carry
+    // over even on bind.
+    mPipelineState.program = program;
+    mPipelineState.pipelineLayout = pipelineLayout;
+    mPipelineState.descriptorSetMask = descriptorSetMask;
 
     mPipelineCache.bindLayout(pipelineLayout);
     mPipelineCache.bindPipeline(mCurrentRenderPass.commandBuffer);
@@ -1841,15 +1917,41 @@ void VulkanDriver::bindDescriptorSet(
 void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
     FVK_SYSTRACE_SCOPE();
     VkCommandBuffer cmdbuffer = mCurrentRenderPass.commandBuffer->buffer();
+    auto const& [doBindInDraw, bundle] = mPipelineState.bindInDraw;
 
-    mDescriptorSetCache.commit(mCurrentRenderPass.commandBuffer,
-            mBoundPipeline.pipelineLayout,
-            mBoundPipeline.descriptorSetMask);
+    fvkutils::DescriptorSetMask setsWithExternalSamplers = {};
+    if (doBindInDraw) {
+        auto& layoutHandles = bundle.dsLayoutHandles;
+        setsWithExternalSamplers = mExternalImageManager.prepareBindSets(layoutHandles,
+                mDescriptorSetCache.getBoundSets());
+
+        VulkanDescriptorSetLayout::DescriptorSetLayoutArray vklayouts;
+        for (size_t i = 0; i < layoutHandles.size(); i++) {
+            if (!layoutHandles[i]) {
+                vklayouts[i] = VK_NULL_HANDLE;
+                continue;
+            }
+            if (setsWithExternalSamplers[i]) {
+                vklayouts[i] = layoutHandles[i]->getExternalSamplerVkLayout();
+            } else {
+                vklayouts[i] = layoutHandles[i]->getVkLayout();
+            }
+        }
+        auto program =
+                resource_ptr<VulkanProgram>::cast(&mResourceManager, bundle.pipelineState.program);
+        VkPipelineLayout const pipelineLayout = mPipelineLayoutCache.getLayout(vklayouts, program);
+        if (pipelineLayout != mPipelineState.pipelineLayout) {
+            bindPipelineImpl(bundle.pipelineState, pipelineLayout, bundle.descriptorSetMask);
+        }
+        mPipelineState.bindInDraw.first = false;
+    }
+    mDescriptorSetCache.commit(mCurrentRenderPass.commandBuffer, mPipelineState.pipelineLayout,
+            setsWithExternalSamplers, mPipelineState.descriptorSetMask);
 
     // Finally, make the actual draw call. TODO: support subranges
-    const uint32_t firstIndex = indexOffset;
-    const int32_t vertexOffset = 0;
-    const uint32_t firstInstId = 0;
+    uint32_t const firstIndex = indexOffset;
+    constexpr int32_t vertexOffset = 0;
+    constexpr uint32_t firstInstId = 0;
 
     vkCmdDrawIndexed(cmdbuffer, indexCount, instanceCount, firstIndex, vertexOffset, firstInstId);
 }

filament/backend/src/vulkan/VulkanDriver.h

@@ -30,6 +30,7 @@
 #include "VulkanYcbcrConversionCache.h"
 #include "vulkan/VulkanDescriptorSetCache.h"
 #include "vulkan/VulkanDescriptorSetLayoutCache.h"
+#include "vulkan/VulkanExternalImageManager.h"
 #include "vulkan/VulkanPipelineLayoutCache.h"
 #include "vulkan/memory/ResourceManager.h"
 #include "vulkan/memory/ResourcePointer.h"
@@ -54,7 +55,7 @@ constexpr uint8_t MAX_RENDERTARGET_ATTACHMENT_TEXTURES =
 class VulkanDriver final : public DriverBase {
 public:
     static Driver* create(VulkanPlatform* platform, VulkanContext const& context,
-            Platform::DriverConfig const& driverConfig) noexcept;
+            Platform::DriverConfig const& driverConfig);
 
 #if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS)
     // Encapsulates the VK_EXT_debug_utils extension.  In particular, we use
@@ -89,8 +90,8 @@ private:
     void debugCommandBegin(CommandStream* cmds, bool synchronous,
             const char* methodName) noexcept override;
 
-    inline VulkanDriver(VulkanPlatform* platform, VulkanContext const& context,
-            Platform::DriverConfig const& driverConfig) noexcept;
+    VulkanDriver(VulkanPlatform* platform, VulkanContext const& context,
+            Platform::DriverConfig const& driverConfig);
 
     ~VulkanDriver() noexcept override;
 
@@ -119,6 +120,8 @@ private:
 
 private:
     void collectGarbage();
+    void bindPipelineImpl(PipelineState const& pipelineState, VkPipelineLayout pipelineLayout,
+            fvkutils::DescriptorSetMask descriptorSetMask);
 
     VulkanPlatform* mPlatform = nullptr;
     fvkmemory::ResourceManager mResourceManager;
@@ -143,21 +146,39 @@ private:
     VulkanDescriptorSetLayoutCache mDescriptorSetLayoutCache;
     VulkanDescriptorSetCache mDescriptorSetCache;
     VulkanQueryManager mQueryManager;
+    VulkanExternalImageManager mExternalImageManager;
 
     // This is necessary for us to write to push constants after binding a pipeline.
-    struct {
-        resource_ptr<VulkanProgram> program;
-        VkPipelineLayout pipelineLayout;
-        fvkutils::DescriptorSetMask descriptorSetMask;
-    } mBoundPipeline = {};
+    using DescriptorSetLayoutHandleList = std::array<resource_ptr<VulkanDescriptorSetLayout>,
+            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+
+    struct BindInDrawBundle {
+        PipelineState pipelineState = {};
+        DescriptorSetLayoutHandleList dsLayoutHandles = {};
+        fvkutils::DescriptorSetMask descriptorSetMask = {};
+        resource_ptr<VulkanProgram> program = {};
+    };
 
-    // We need to store information about a render pass to enable better barriers at the end of a
-    // renderpass.
     struct {
-        using AttachmentArray =
-                fvkutils::StaticVector<VulkanAttachment, MAX_RENDERTARGET_ATTACHMENT_TEXTURES>;
-        AttachmentArray attachments;
-    } mRenderPassFboInfo = {};
+        // For push constant
+        resource_ptr<VulkanProgram> program = {};
+        // For push commiting dynamic ubos in draw()
+        VkPipelineLayout pipelineLayout = VK_NULL_HANDLE;
+        fvkutils::DescriptorSetMask descriptorSetMask = {};
+
+        std::pair<bool, BindInDrawBundle> bindInDraw = {false, {}};
+    } mPipelineState = {};
+
+    struct {
+        // This tracks whether the app has seen external samplers bound to a the descriptor set.
+        // This will force bindPipeline to take a slow path.
+        bool hasExternalSamplerLayouts = false;
+        bool hasBoundExternalImages = false;
+
+        bool hasExternalSamplers() const noexcept {
+            return hasExternalSamplerLayouts && hasBoundExternalImages;
+        }
+    } mAppState;
 
     bool const mIsSRGBSwapChainSupported;
     backend::StereoscopicType const mStereoscopicType;

filament/backend/src/vulkan/VulkanExternalImageManager.cpp

@@ -0,0 +1,294 @@
+/*
+ * 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 "VulkanExternalImageManager.h"
+
+#include "VulkanDescriptorSetCache.h"
+#include "VulkanDescriptorSetLayoutCache.h"
+#include "VulkanSamplerCache.h"
+#include "VulkanYcbcrConversionCache.h"
+#include "vulkan/memory/ResourcePointer.h"
+#include "vulkan/utils/Conversion.h"
+
+#include <backend/platforms/VulkanPlatform.h>
+
+#include <algorithm>
+
+namespace filament::backend {
+
+namespace {
+
+using Bitmask = fvkutils::UniformBufferBitmask;
+static_assert(sizeof(Bitmask) * 8 == fvkutils::MAX_DESCRIPTOR_SET_BITMASK_BITS);
+
+template<typename T>
+void erasep(std::vector<T>& v, std::function<bool(T const&)> f) {
+    auto newEnd = std::remove_if(v.begin(), v.end(), f);
+    v.erase(newEnd, v.end());
+}
+
+using ImageData = VulkanExternalImageManager::VulkanExternalImageManager::ImageData;
+ImageData& findImage(std::vector<ImageData>& images,
+        fvkmemory::resource_ptr<VulkanTexture> texture) {
+    auto itr = std::find_if(images.begin(), images.end(), [&](ImageData const& data) {
+        return data.image == texture;
+    });
+    assert_invariant(itr != images.end());
+    return *itr;
+}
+
+void copySet(VkDevice device, VkDescriptorSet srcSet, VkDescriptorSet dstSet, Bitmask bindings) {
+    // TODO: fix the size for better memory management
+    std::vector<VkCopyDescriptorSet> copies;
+    bindings.forEachSetBit([&](size_t index) {
+        copies.push_back({
+            .sType = VK_STRUCTURE_TYPE_COPY_DESCRIPTOR_SET,
+            .srcSet = srcSet,
+            .srcBinding = (uint32_t) index,
+            .dstSet = dstSet,
+            .dstBinding = (uint32_t) index,
+            .descriptorCount = 1,
+        });
+    });
+    vkUpdateDescriptorSets(device, 0, nullptr, copies.size(), copies.data());
+}
+
+Bitmask foldBitsInHalf(Bitmask bitset) {
+    Bitmask outBitset;
+    bitset.forEachSetBit([&](size_t index) {
+        constexpr size_t BITMASK_LOWER_BITS_LEN = sizeof(outBitset) * 4;
+        outBitset.set(index % BITMASK_LOWER_BITS_LEN);
+    });
+    return outBitset;
+}
+
+}// namespace
+
+VulkanExternalImageManager::VulkanExternalImageManager(VulkanPlatform* platform,
+        VulkanSamplerCache* samplerCache, VulkanYcbcrConversionCache* ycbcrConversionCache,
+        VulkanDescriptorSetCache* setCache, VulkanDescriptorSetLayoutCache* layoutCache)
+    : mPlatform(platform),
+      mSamplerCache(samplerCache),
+      mYcbcrConversionCache(ycbcrConversionCache),
+      mDescriptorSetCache(setCache),
+      mDescriptorSetLayoutCache(layoutCache) {
+}
+
+VulkanExternalImageManager::~VulkanExternalImageManager() = default;
+
+void VulkanExternalImageManager::terminate() {
+    mSetBindings.clear();
+    mImages.clear();
+}
+
+void VulkanExternalImageManager::onBeginFrame() {
+    std::for_each(mImages.begin(), mImages.end(), [](ImageData& image) {
+        image.hasBeenValidated = false;
+    });
+    std::for_each(mSetBindings.begin(), mSetBindings.end(), [](SetBindingInfo& info) {
+        info.bound = false;
+    });
+}
+
+fvkutils::DescriptorSetMask VulkanExternalImageManager::prepareBindSets(LayoutArray const& layouts,
+        SetArray const& sets) {
+    fvkutils::DescriptorSetMask shouldUseExternalSampler{};
+    for (uint8_t i = 0; i < sets.size(); i++) {
+        auto set = sets[i];
+        auto layout = layouts[i];
+        if (!set || !layout) {
+            continue;
+        }
+        if (hasExternalSampler(set)) {
+            updateSetAndLayout(set, layout);
+            shouldUseExternalSampler.set(i);
+        }
+    }
+    return shouldUseExternalSampler;
+}
+
+bool VulkanExternalImageManager::hasExternalSampler(
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set) {
+    auto itr = std::find_if(mSetBindings.begin(), mSetBindings.end(),
+            [&](SetBindingInfo const& info) { return info.set == set; });
+    return itr != mSetBindings.end();
+}
+
+void VulkanExternalImageManager::updateSetAndLayout(
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set,
+        fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout) {
+    utils::FixedCapacityVector<
+            std::tuple<uint8_t, VkSampler, fvkmemory::resource_ptr<VulkanTexture>>>
+            samplerAndBindings;
+    samplerAndBindings.reserve(MAX_SAMPLER_COUNT);
+
+    fvkutils::SamplerBitmask actualExternalSamplers;
+    for (auto& bindingInfo : mSetBindings) {
+        if (bindingInfo.set != set || bindingInfo.bound) {
+            continue;
+        }
+        auto& imageData = findImage(mImages, bindingInfo.image);
+        updateImage(&imageData);
+
+        auto samplerParams = bindingInfo.samplerParams;
+        // according to spec, these must match chromaFilter
+        // https://registry.khronos.org/vulkan/specs/latest/man/html/VkSamplerCreateInfo.html#VUID-VkSamplerCreateInfo-minFilter-01645
+        samplerParams.filterMag = SamplerMagFilter::NEAREST;
+        samplerParams.filterMin = SamplerMinFilter::NEAREST;
+
+        auto sampler = mSamplerCache->getSampler({
+            .sampler = samplerParams,
+            .conversion = imageData.conversion,
+        });
+        actualExternalSamplers.set(bindingInfo.binding);
+        samplerAndBindings.push_back({ bindingInfo.binding, sampler, bindingInfo.image });
+        bindingInfo.bound = true;
+    }
+
+    if (samplerAndBindings.empty()) {
+        return;
+    }
+
+    // Sort by binding number
+    std::sort(samplerAndBindings.begin(), samplerAndBindings.end(), [](auto const& a, auto const& b) {
+        return std::get<0>(a) < std::get<0>(b);
+    });
+
+    utils::FixedCapacityVector<VkSampler> outSamplers;
+    outSamplers.reserve(MAX_SAMPLER_COUNT);
+    std::for_each(samplerAndBindings.begin(), samplerAndBindings.end(),
+            [&](auto const& b) { outSamplers.push_back(std::get<1>(b)); });
+
+    VkDescriptorSetLayout const oldLayout = layout->getExternalSamplerVkLayout();
+    VkDescriptorSetLayout const newLayout = mDescriptorSetLayoutCache->getVkLayout(layout->bitmask,
+            actualExternalSamplers, outSamplers);
+
+    // Need to copy the set
+    VkDescriptorSet const oldSet = set->getExternalSamplerVkSet();
+    if (oldLayout != newLayout || oldSet == VK_NULL_HANDLE) {
+        // Build a new descriptor set from the new layout
+        VkDescriptorSet const newSet = mDescriptorSetCache->getVkSet(layout->count, newLayout);
+        auto const ubo = layout->bitmask.ubo | layout->bitmask.dynamicUbo;
+        auto const samplers = layout->bitmask.sampler & (~actualExternalSamplers);
+
+        // Each bitmask denotes a binding index, and separated into two stages - vertex and buffer
+        // We fold the two stages into just the lower half of the bits to denote a combined set of
+        // bindings.
+        Bitmask const copyBindings = foldBitsInHalf(ubo | samplers);
+        VkDescriptorSet const srcSet = oldSet != VK_NULL_HANDLE ? oldSet : set->getVkSet();
+        copySet(mPlatform->getDevice(), srcSet, newSet, copyBindings);
+
+        set->setExternalSamplerVkSet(newSet, [&](VulkanDescriptorSet*) {
+            mDescriptorSetCache->manualRecycle(layout->count, newLayout, newSet);
+        });
+        if (oldLayout != newLayout) {
+            layout->setExternalSamplerVkLayout(newLayout);
+        }
+    }
+
+    // Update the external samplers in the set
+    for (auto& [binding, sampler, image]: samplerAndBindings) {
+        mDescriptorSetCache->updateSamplerForExternalSamplerSet(set, binding, image);
+    }
+}
+
+VkSamplerYcbcrConversion VulkanExternalImageManager::getVkSamplerYcbcrConversion(
+        VulkanPlatform::ExternalImageMetadata const& metadata) {
+    // This external image does not require external sampler (YUV conversion).
+    if (metadata.externalFormat == 0 && !fvkutils::isVKYcbcrConversionFormat(metadata.format)) {
+        return VK_NULL_HANDLE;
+    }
+    VulkanYcbcrConversionCache::Params ycbcrParams = {
+        .conversion = {
+            .ycbcrModel = fvkutils::getYcbcrModelConversionFilament(metadata.ycbcrModel),
+            .r = fvkutils::getSwizzleFilament(metadata.ycbcrConversionComponents.r, 0),
+            .g = fvkutils::getSwizzleFilament(metadata.ycbcrConversionComponents.g, 1),
+            .b = fvkutils::getSwizzleFilament(metadata.ycbcrConversionComponents.b, 2),
+            .a = fvkutils::getSwizzleFilament(metadata.ycbcrConversionComponents.a, 3),
+            .ycbcrRange = fvkutils::getYcbcrRangeFilament(metadata.ycbcrRange),
+            .xChromaOffset = fvkutils::getChromaLocationFilament(metadata.xChromaOffset),
+            .yChromaOffset = fvkutils::getChromaLocationFilament(metadata.yChromaOffset),
+
+            // Unclear where to get the chromaFilter, we just assume it's nearest.
+            .chromaFilter = SamplerMagFilter::NEAREST,
+        },
+        .format = metadata.format,
+        .externalFormat = metadata.externalFormat,
+    };
+    return mYcbcrConversionCache->getConversion(ycbcrParams);
+}
+
+void VulkanExternalImageManager::updateImage(ImageData* image) {
+    if (image->hasBeenValidated) {
+        return;
+    }
+    image->hasBeenValidated = true;
+
+    auto metadata = mPlatform->extractExternalImageMetadata(image->platformHandle);
+    auto vkYcbcr = getVkSamplerYcbcrConversion(metadata);
+    if (vkYcbcr == image->conversion) {
+        return;
+    }
+
+    image->image->setYcbcrConversion(vkYcbcr);
+    image->conversion = vkYcbcr;
+    return;
+}
+
+void VulkanExternalImageManager::removeDescriptorSet(
+        fvkmemory::resource_ptr<VulkanDescriptorSet> inSet) {
+    erasep<SetBindingInfo>(mSetBindings,
+            [&](auto const& bindingInfo) { return (bindingInfo.set == inSet); });
+}
+
+void VulkanExternalImageManager::bindExternallySampledTexture(
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t bindingPoint,
+        fvkmemory::resource_ptr<VulkanTexture> image, SamplerParams samplerParams) {
+    // Should we do duplicate validation here?
+    auto& imageData = findImage(mImages, image);
+    mSetBindings.push_back({ bindingPoint, imageData.image, set, samplerParams });
+}
+
+void VulkanExternalImageManager::addExternallySampledTexture(
+       fvkmemory::resource_ptr<VulkanTexture> image,
+        Platform::ExternalImageHandleRef platformHandleRef) {
+    mImages.push_back({ image, platformHandleRef, false });
+}
+
+void VulkanExternalImageManager::removeExternallySampledTexture(
+        fvkmemory::resource_ptr<VulkanTexture> image) {
+    erasep<SetBindingInfo>(mSetBindings,
+            [&](auto const& bindingInfo) { return (bindingInfo.image == image); });
+    erasep<ImageData>(mImages, [&](auto const& imageData) {
+        return imageData.image == image;
+    });
+}
+
+bool VulkanExternalImageManager::isExternallySampledTexture(
+        fvkmemory::resource_ptr<VulkanTexture> image) const {
+    return std::find_if(mImages.begin(), mImages.end(), [&](auto const& imageData) {
+        return imageData.image == image;
+    }) != mImages.end();
+}
+
+void VulkanExternalImageManager::clearTextureBinding(
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set, uint8_t bindingPoint) {
+    erasep<SetBindingInfo>(mSetBindings, [&](auto const& bindingInfo) {
+        return (bindingInfo.set == set && bindingInfo.binding == bindingPoint);
+    });
+}
+
+} // namesapce filament::backend

filament/backend/src/vulkan/VulkanExternalImageManager.h

@@ -0,0 +1,121 @@
+/*
+ * 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.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_CACHING_VULKANEXTERNALIMAGEMANAGER_H
+#define TNT_FILAMENT_BACKEND_CACHING_VULKANEXTERNALIMAGEMANAGER_H
+
+#include "VulkanHandles.h"
+
+#include <backend/DriverEnums.h>
+
+#include <array>
+#include <vector>
+
+namespace filament::backend {
+
+class VulkanYcbcrConversionCache;
+class VulkanSamplerCache;
+class VulkanDescriptorSetLayoutCache;
+class VulkanDescriptorSetCache;
+
+// Manages the logic of external images and their quirks wrt Vulikan.
+class VulkanExternalImageManager {
+public:
+
+    VulkanExternalImageManager(
+            VulkanPlatform* platform,
+            VulkanSamplerCache* samplerCache,
+            VulkanYcbcrConversionCache* ycbcrConversionCache,
+            VulkanDescriptorSetCache* setCache,
+            VulkanDescriptorSetLayoutCache* layoutCache);
+
+    ~VulkanExternalImageManager();
+
+    void terminate();
+
+    void onBeginFrame();
+
+    using SetArray = std::array<fvkmemory::resource_ptr<VulkanDescriptorSet>,
+            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+
+    using LayoutArray = std::array<fvkmemory::resource_ptr<VulkanDescriptorSetLayout>,
+            VulkanDescriptorSetLayout::UNIQUE_DESCRIPTOR_SET_COUNT>;
+
+    using VkLayoutArray = VulkanDescriptorSetLayout::DescriptorSetLayoutArray;
+
+    // Returns  bitmask to indicate whether or not to use the external sampler version of each
+    // descriptor set.
+    fvkutils::DescriptorSetMask prepareBindSets(LayoutArray const& layouts, SetArray const& sets);
+
+    void removeDescriptorSet(fvkmemory::resource_ptr<VulkanDescriptorSet> set);
+
+    void bindExternallySampledTexture(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
+            uint8_t bindingPoint, fvkmemory::resource_ptr<VulkanTexture> image,
+            SamplerParams samplerParams);
+
+    void clearTextureBinding(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
+                             uint8_t bindingPoint);
+
+    void addExternallySampledTexture(fvkmemory::resource_ptr<VulkanTexture> external,
+            Platform::ExternalImageHandleRef platformHandleRef);
+
+    void removeExternallySampledTexture(fvkmemory::resource_ptr<VulkanTexture> image);
+
+    bool isExternallySampledTexture(fvkmemory::resource_ptr<VulkanTexture> image) const;
+
+    VkSamplerYcbcrConversion getVkSamplerYcbcrConversion(
+            VulkanPlatform::ExternalImageMetadata const& metadata);
+
+    struct ImageData {
+        fvkmemory::resource_ptr<VulkanTexture> image;
+        Platform::ExternalImageHandle platformHandle;
+        bool hasBeenValidated = false; // indicates whether the image has been validated *this frame*
+        VkSamplerYcbcrConversion conversion = VK_NULL_HANDLE;
+    };
+
+private:
+    bool hasExternalSampler(fvkmemory::resource_ptr<VulkanDescriptorSet> set);
+
+    void updateSetAndLayout(fvkmemory::resource_ptr<VulkanDescriptorSet> set,
+            fvkmemory::resource_ptr<VulkanDescriptorSetLayout> layout);
+
+    void updateImage(ImageData* imageData);
+
+    VulkanPlatform* mPlatform;
+    VulkanSamplerCache* mSamplerCache;
+    VulkanYcbcrConversionCache* mYcbcrConversionCache;
+    VulkanDescriptorSetCache* mDescriptorSetCache;
+    VulkanDescriptorSetLayoutCache* mDescriptorSetLayoutCache;
+
+    using SetAndLayout = std::pair<fvkmemory::resource_ptr<VulkanDescriptorSet>,
+            fvkmemory::resource_ptr<VulkanDescriptorSetLayout>>;
+
+    struct SetBindingInfo {
+        uint8_t binding = 0;
+        fvkmemory::resource_ptr<VulkanTexture> image;
+        fvkmemory::resource_ptr<VulkanDescriptorSet> set;
+        SamplerParams samplerParams;
+        bool bound = false;
+    };
+
+    // Use vectors instead of hash maps because we only expect small number of entries.
+    std::vector<SetBindingInfo> mSetBindings;
+    std::vector<ImageData> mImages;
+};
+
+} // filament::backend
+
+#endif // TNT_FILAMENT_BACKEND_CACHING_VULKANEXTERNALIMAGEMANAGER_H

filament/backend/src/vulkan/VulkanFboCache.cpp

@@ -330,12 +330,20 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey const& config) noexcept
     mRenderPassCache[config] = {renderPass, mCurrentTime};
 
 #if FVK_ENABLED(FVK_DEBUG_FBO_CACHE)
-    FVK_LOGD << "Created render pass " << renderPass << " with "
-        << "samples = " << int(config.samples) << ", "
-        << "depth = " << (hasDepth ? 1 : 0) << ", "
-        << "colorAttachmentCount[0] = " << subpasses[0].colorAttachmentCount
-        << utils::io::endl;
-    #endif
+    FVK_LOGD << "Created render pass " << renderPass << " with ";
+    for (int i = 0; i < MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT; ++i) {
+        FVK_LOGD << (int) config.colorFormat[i] << " ";
+    }
+    FVK_LOGD << ", "
+             << "depth = " << config.depthFormat << ", "
+             << "initialDepthLayout = " << (int) config.initialDepthLayout << ", "
+             << "samples = " << int(config.samples) << ", "
+             << "needsResolveMask = " << int(config.needsResolveMask) << ", "
+             << "usesLazilyAllocatedMemory = " << int(config.usesLazilyAllocatedMemory) << ", "
+             << "viewCount = " << int(config.viewCount) << ", "
+             << "colorAttachmentCount[0] = " << subpasses[0].colorAttachmentCount
+             << utils::io::endl;
+#endif
 
     return renderPass;
 }

filament/backend/src/vulkan/VulkanHandles.cpp

@@ -149,11 +149,18 @@ void VulkanDescriptorSet::acquire(fvkmemory::resource_ptr<VulkanBufferObject> ob
     mResources.push_back(obj);
 }
 
-VulkanDescriptorSetLayout::VulkanDescriptorSetLayout(DescriptorSetLayout const& layout)
+VulkanDescriptorSetLayout::VulkanDescriptorSetLayout(DescriptorSetLayout&& layout,
+        VkDescriptorSetLayout vkLayout)
     : bitmask(fromBackendLayout(layout)),
-      count(Count::fromLayoutBitmask(bitmask)) {}
+      count(Count::fromLayoutBitmask(bitmask)),
+      mVkLayout(vkLayout) {}
 
-PushConstantDescription::PushConstantDescription(backend::Program const& program) noexcept {
+VulkanDescriptorSetLayout::Bitmask VulkanDescriptorSetLayout::Bitmask::fromLayoutDescription(
+        DescriptorSetLayout const& layout) {
+    return fromBackendLayout(layout);
+}
+
+PushConstantDescription::PushConstantDescription(backend::Program const& program) {
     mRangeCount = 0;
     for (auto stage : { ShaderStage::VERTEX, ShaderStage::FRAGMENT, ShaderStage::COMPUTE }) {
         auto const& constants = program.getPushConstants(stage);

filament/backend/src/vulkan/VulkanHandles.h

@@ -71,7 +71,7 @@ struct VulkanDescriptorSetLayout : public HwDescriptorSetLayout, fvkmemory::Reso
         fvkutils::SamplerBitmask sampler;           // 8 bytes
         fvkutils::InputAttachmentBitmask inputAttachment; // 8 bytes
 
-        // This is a subset of the bitmask.sampler field.
+        // This is a subset of the sampler field.
         fvkutils::SamplerBitmask externalSampler; // 8 bytes
 
         bool operator==(Bitmask const& right) const {
@@ -79,6 +79,8 @@ struct VulkanDescriptorSetLayout : public HwDescriptorSetLayout, fvkmemory::Reso
                    inputAttachment == right.inputAttachment &&
                    externalSampler == right.externalSampler;
         }
+
+        static Bitmask fromLayoutDescription(DescriptorSetLayout const& layout);
     };
     static_assert(sizeof(Bitmask) == 40);
 
@@ -120,16 +122,20 @@ struct VulkanDescriptorSetLayout : public HwDescriptorSetLayout, fvkmemory::Reso
         }
     };
 
-    VulkanDescriptorSetLayout(DescriptorSetLayout const& layout);
+    VulkanDescriptorSetLayout(DescriptorSetLayout&& layout, VkDescriptorSetLayout vkLayout);
 
     // Note that we don't destroy the vklayout. This is done by the layout cache.
     ~VulkanDescriptorSetLayout() = default;
 
-    VkDescriptorSetLayout const& getVkLayout() const noexcept { return mVkLayout; }
+    VkDescriptorSetLayout getVkLayout() const noexcept { return mVkLayout; }
 
-    // It is possible to have the layout switch out due to AHardwarebuffer (external image) format
-    // changes.
-    void setVkLayout(VkDescriptorSetLayout vklayout) noexcept { mVkLayout = vklayout; }
+    VkDescriptorSetLayout getExternalSamplerVkLayout() const noexcept {
+        return mExternalSamplerVkLayout;
+    }
+
+    void setExternalSamplerVkLayout(VkDescriptorSetLayout vklayout) noexcept {
+        mExternalSamplerVkLayout = vklayout;
+    }
 
     bool hasExternalSamplers() const noexcept { return bitmask.externalSampler.count() > 0; }
 
@@ -137,7 +143,11 @@ struct VulkanDescriptorSetLayout : public HwDescriptorSetLayout, fvkmemory::Reso
     Count const count;
 
 private:
-    VkDescriptorSetLayout mVkLayout = VK_NULL_HANDLE;
+    // This is the layout without any immutable samplers.
+    VkDescriptorSetLayout const mVkLayout = VK_NULL_HANDLE;
+
+    // This is the layout with immutable samplers, and can be updated.
+    VkDescriptorSetLayout mExternalSamplerVkLayout = VK_NULL_HANDLE;
 };
 
 struct VulkanDescriptorSet : public HwDescriptorSet, fvkmemory::Resource {
@@ -149,23 +159,37 @@ public:
     VulkanDescriptorSet(
             fvkutils::UniformBufferBitmask const& dynamicUboMask,
             uint8_t uniqueDynamicUboCount,
-            OnRecycle&& onRecycleFn)
+            OnRecycle&& onRecycleFn, VkDescriptorSet vkSet)
         : dynamicUboMask(dynamicUboMask),
           uniqueDynamicUboCount(uniqueDynamicUboCount),
+          mVkSet(vkSet),
           mOnRecycleFn(std::move(onRecycleFn)) {}
 
+    // NOLINTNEXTLINE(bugprone-exception-escape)
     ~VulkanDescriptorSet() {
         if (mOnRecycleFn) {
             mOnRecycleFn(this);
         }
+        if (mOnRecycleExternalSamplerFn) {
+            mOnRecycleExternalSamplerFn(this);
+        }
     }
 
-    VkDescriptorSet const& getVkSet() const noexcept {
+    VkDescriptorSet getVkSet() const noexcept {
         return mVkSet;
     }
 
-    // Note that the only case where you'd set it more than once is with external images/samplers.
-    void setVkSet(VkDescriptorSet vkset) noexcept { mVkSet = vkset; }
+    VkDescriptorSet getExternalSamplerVkSet() const noexcept {
+        return mExternalSamplerVkSet;
+    }
+
+    void setExternalSamplerVkSet(VkDescriptorSet vkset, OnRecycle onRecycle) {
+        mExternalSamplerVkSet = vkset;
+        if (mOnRecycleExternalSamplerFn) {
+            mOnRecycleExternalSamplerFn(this);
+        }
+        mOnRecycleExternalSamplerFn = onRecycle;
+    }
 
     void setOffsets(backend::DescriptorSetOffsetArray&& offsets) noexcept {
         mOffsets = std::move(offsets);
@@ -182,17 +206,20 @@ public:
     uint8_t const uniqueDynamicUboCount;
 
 private:
-    VkDescriptorSet mVkSet = VK_NULL_HANDLE;
+    VkDescriptorSet const mVkSet;
+    VkDescriptorSet mExternalSamplerVkSet = VK_NULL_HANDLE;
+
     backend::DescriptorSetOffsetArray mOffsets;
     std::vector<fvkmemory::resource_ptr<fvkmemory::Resource>> mResources;
     OnRecycle mOnRecycleFn;
+    OnRecycle mOnRecycleExternalSamplerFn;
 };
 
 using PushConstantNameArray = utils::FixedCapacityVector<char const*>;
 using PushConstantNameByStage = std::array<PushConstantNameArray, Program::SHADER_TYPE_COUNT>;
 
 struct PushConstantDescription {
-    explicit PushConstantDescription(backend::Program const& program) noexcept;
+    explicit PushConstantDescription(backend::Program const& program);
 
     VkPushConstantRange const* getVkRanges() const noexcept { return mRanges; }
     uint32_t getVkRangeCount() const noexcept { return mRangeCount; }

filament/backend/src/vulkan/VulkanPipelineCache.cpp

@@ -66,8 +66,12 @@ void VulkanPipelineCache::bindPipeline(VulkanCommandBuffer* commands) {
     // If an error occurred, allow higher levels to handle it gracefully.
     assert_invariant(cacheEntry != nullptr && "Failed to create/find pipeline");
 
-    mBoundPipeline = mPipelineRequirements;
-    vkCmdBindPipeline(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, cacheEntry->handle);
+
+    static PipelineEqual equal;
+    if (!equal(mBoundPipeline, mPipelineRequirements)) {
+        mBoundPipeline = mPipelineRequirements;
+        vkCmdBindPipeline(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, cacheEntry->handle);
+    }
 }
 
 VulkanPipelineCache::PipelineCacheEntry* VulkanPipelineCache::createPipeline() noexcept {

filament/backend/src/vulkan/VulkanReadPixels.cpp

@@ -335,7 +335,7 @@ void VulkanReadPixels::run(fvkmemory::resource_ptr<VulkanRenderTarget> srcTarget
     mTaskHandler->post(std::move(waitFenceFunc), std::move(cleanPbdFunc));
 }
 
-void VulkanReadPixels::runUntilComplete() noexcept {
+void VulkanReadPixels::runUntilComplete() {
     if (!mTaskHandler) {
         return;
     }

filament/backend/src/vulkan/VulkanReadPixels.h

@@ -79,7 +79,7 @@ public:
             OnReadCompleteFunction const& readCompleteFunc);
 
     // This method will block until all of the in-flight requests are complete.
-    void runUntilComplete() noexcept;
+    void runUntilComplete();
 
 private:
     VkDevice mDevice = VK_NULL_HANDLE;

filament/backend/src/vulkan/VulkanSamplerCache.cpp

@@ -28,7 +28,7 @@ namespace filament::backend {
 VulkanSamplerCache::VulkanSamplerCache(VkDevice device)
     : mDevice(device) {}
 
-VkSampler VulkanSamplerCache::getSampler(Params params) noexcept {
+VkSampler VulkanSamplerCache::getSampler(Params params) {
     auto iter = mCache.find(params);
     if (UTILS_LIKELY(iter != mCache.end())) {
         return iter->second;

filament/backend/src/vulkan/VulkanSamplerCache.h

@@ -38,7 +38,7 @@ public:
     static_assert(sizeof(Params) == 16);
 
     explicit VulkanSamplerCache(VkDevice device);
-    VkSampler getSampler(Params params) noexcept;
+    VkSampler getSampler(Params params);
     void terminate() noexcept;
 private:
     VkDevice mDevice;

filament/backend/src/vulkan/VulkanStagePool.cpp

@@ -24,7 +24,7 @@
 
 #include <utils/Panic.h>
 
-static constexpr uint32_t TIME_BEFORE_EVICTION = FVK_MAX_COMMAND_BUFFERS;
+static constexpr uint32_t TIME_BEFORE_EVICTION = 3;
 
 namespace filament::backend {
 
@@ -39,7 +39,7 @@ VulkanStage const* VulkanStagePool::acquireStage(uint32_t numBytes) {
         auto stage = iter->second;
         mFreeStages.erase(iter);
         stage->lastAccessed = mCurrentFrame;
-        mUsedStages.insert(stage);
+        mUsedStages.push_back(stage);
         return stage;
     }
     // We were not able to find a sufficiently large stage, so create a new one.
@@ -51,7 +51,7 @@ VulkanStage const* VulkanStagePool::acquireStage(uint32_t numBytes) {
     });
 
     // Create the VkBuffer.
-    mUsedStages.insert(stage);
+    mUsedStages.push_back(stage);
     VkBufferCreateInfo bufferInfo {
         .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
         .size = numBytes,
@@ -77,7 +77,7 @@ VulkanStageImage const* VulkanStagePool::acquireImage(PixelDataFormat format, Pi
         if (image->format == vkformat && image->width == width && image->height == height) {
             mFreeImages.erase(image);
             image->lastAccessed = mCurrentFrame;
-            mUsedImages.insert(image);
+            mUsedImages.push_back(image);
             return image;
         }
     }
@@ -89,7 +89,7 @@ VulkanStageImage const* VulkanStagePool::acquireImage(PixelDataFormat format, Pi
         .lastAccessed = mCurrentFrame,
     });
 
-    mUsedImages.insert(image);
+    mUsedImages.push_back(image);
 
     const VkImageCreateInfo imageInfo = {
         .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
@@ -161,7 +161,7 @@ void VulkanStagePool::gc() noexcept {
             stage->lastAccessed = mCurrentFrame;
             mFreeStages.insert(std::make_pair(stage->capacity, stage));
         } else {
-            mUsedStages.insert(stage);
+            mUsedStages.push_back(stage);
         }
     }
 
@@ -185,7 +185,7 @@ void VulkanStagePool::gc() noexcept {
             image->lastAccessed = mCurrentFrame;
             mFreeImages.insert(image);
         } else {
-            mUsedImages.insert(image);
+            mUsedImages.push_back(image);
         }
     }
     FVK_SYSTRACE_END();

filament/backend/src/vulkan/VulkanStagePool.h

@@ -22,6 +22,7 @@
 
 #include <map>
 #include <unordered_set>
+#include <vector>
 
 namespace filament::backend {
 
@@ -73,10 +74,10 @@ private:
     std::multimap<uint32_t, VulkanStage const*> mFreeStages;
 
     // Simple unordered set for stashing a list of in-use stages that can be reclaimed later.
-    std::unordered_set<VulkanStage const*> mUsedStages;
+    std::vector<VulkanStage const*> mUsedStages;
 
     std::unordered_set<VulkanStageImage const*> mFreeImages;
-    std::unordered_set<VulkanStageImage const*> mUsedImages;
+    std::vector<VulkanStageImage const*> mUsedImages;
 
     // Store the current "time" (really just a frame count) and LRU eviction parameters.
     uint64_t mCurrentFrame = 0;

filament/backend/src/vulkan/VulkanTexture.cpp

@@ -223,8 +223,7 @@ VkImageUsageFlags getUsage(VulkanContext const& context, uint8_t samples,
 VulkanTextureState::VulkanTextureState(VulkanStagePool& stagePool, VulkanCommands* commands,
         VmaAllocator allocator, VkDevice device, VkImage image, VkDeviceMemory deviceMemory,
         VkFormat format, VkImageViewType viewType, uint8_t levels, uint8_t layerCount,
-        VkSamplerYcbcrConversion ycbcrConversion, bool isExternalFormat, VkImageUsageFlags usage,
-        bool isProtected)
+        VkSamplerYcbcrConversion ycbcrConversion, VkImageUsageFlags usage, bool isProtected)
     : mStagePool(stagePool),
       mCommands(commands),
       mAllocator(allocator),
@@ -234,17 +233,17 @@ VulkanTextureState::VulkanTextureState(VulkanStagePool& stagePool, VulkanCommand
       mVkFormat(format),
       mViewType(viewType),
       mFullViewRange{ fvkutils::getImageAspect(format), 0, levels, 0, layerCount },
-      mYcbcr{ ycbcrConversion, isExternalFormat },
+      mYcbcr{ ycbcrConversion },
       mDefaultLayout(getDefaultLayoutImpl(usage)),
       mUsage(usage),
       mIsProtected(isProtected) {}
 
 VulkanTextureState::~VulkanTextureState() {
+    clearCachedImageViews();
     if (mTextureImageMemory != VK_NULL_HANDLE) {
         vkDestroyImage(mDevice, mTextureImage, VKALLOC);
         vkFreeMemory(mDevice, mTextureImageMemory, VKALLOC);
     }
-    clearCachedImageViews();
 }
 
 void VulkanTextureState::clearCachedImageViews() noexcept {
@@ -272,7 +271,7 @@ VkImageView VulkanTextureState::getImageView(VkImageSubresourceRange range, VkIm
         .flags = 0,
         .image = mTextureImage,
         .viewType = viewType,
-        .format = mYcbcr.isExternalFormat ? VK_FORMAT_UNDEFINED : mVkFormat,
+        .format = mYcbcr.conversion != VK_NULL_HANDLE ? VK_FORMAT_UNDEFINED : mVkFormat,
         .components = swizzle,
         .subresourceRange = range,
     };
@@ -294,7 +293,6 @@ VulkanTexture::VulkanTexture(VulkanContext const& context, VkDevice device, VmaA
               commands, allocator, device, image, memory, format,
               fvkutils::getViewType(SamplerType::SAMPLER_2D),
               /*mipLevels=*/1, getLayerCountFromDepth(depth), conversion,
-              /*isExternalFormat=*/false,
               getUsage(context, samples, VK_NULL_HANDLE, format, tusage),
               any(usage & TextureUsage::PROTECTED))) {
     mPrimaryViewRange = mState->mFullViewRange;
@@ -424,8 +422,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
     mState = fvkmemory::resource_ptr<VulkanTextureState>::construct(resourceManager, stagePool,
             commands, allocator, device, textureImage, textureImageMemory, vkFormat,
             fvkutils::getViewType(target), levels, getLayerCount(target, depth),
-            VK_NULL_HANDLE /* ycbcrConversion */, false /*isExternalFormat*/, imageInfo.usage,
-            isProtected);
+            VK_NULL_HANDLE /* ycbcrConversion */, imageInfo.usage, isProtected);
 
     // Spec out the "primary" VkImageView that shaders use to sample from the image.
     mPrimaryViewRange = mState->mFullViewRange;
@@ -764,15 +761,13 @@ void VulkanTexture::setLayout(VkImageSubresourceRange const& range, VulkanLayout
     }
 }
 
-void VulkanTexture::setYcbcrConversion(VkSamplerYcbcrConversion conversion, bool isExternalFormat) {
+void VulkanTexture::setYcbcrConversion(VkSamplerYcbcrConversion conversion) {
     // Note that this comparison is valid because we only ever create VkSamplerYcbcrConversion from
     // a cache.  So for each set of parameters, there is exactly one conversion (similar to
     // samplers).
     VulkanTextureState::Ycbcr ycbcr = {
         .conversion = conversion,
-        .isExternalFormat = isExternalFormat,
     };
-
     if (mState->mYcbcr != ycbcr) {
         mState->mYcbcr = ycbcr;
         mState->clearCachedImageViews();

filament/backend/src/vulkan/VulkanTexture.h

@@ -40,8 +40,7 @@ struct VulkanTextureState : public fvkmemory::Resource {
     VulkanTextureState(VulkanStagePool& stagePool, VulkanCommands* commands, VmaAllocator allocator,
             VkDevice device, VkImage image, VkDeviceMemory deviceMemory, VkFormat format,
             VkImageViewType viewType, uint8_t levels, uint8_t layerCount,
-            VkSamplerYcbcrConversion ycbcrConversion, bool isExternalFormat,
-            VkImageUsageFlags usage, bool isProtected);
+            VkSamplerYcbcrConversion ycbcrConversion, VkImageUsageFlags usage, bool isProtected);
 
     ~VulkanTextureState();
 
@@ -86,10 +85,9 @@ private:
     // conversion matrix per-frame.
     struct Ycbcr {
         VkSamplerYcbcrConversion conversion;
-        bool isExternalFormat;
 
         bool operator==(Ycbcr const& other) const {
-            return conversion == other.conversion && isExternalFormat == other.isExternalFormat;
+            return conversion == other.conversion;
         }
 
         bool operator!=(Ycbcr const& other) const {
@@ -209,7 +207,7 @@ struct VulkanTexture : public HwTexture, fvkmemory::Resource {
     // This is used in the case of external images and external samplers. AHB might update the
     // conversion per-frame. This implies that we need to invalidate the view cache when that
     // happens.
-    void setYcbcrConversion(VkSamplerYcbcrConversion conversion, bool isExternal);
+    void setYcbcrConversion(VkSamplerYcbcrConversion conversion);
 
 #if FVK_ENABLED(FVK_DEBUG_TEXTURE)
     void print() const;

filament/backend/src/vulkan/VulkanYcbcrConversionCache.cpp

@@ -33,7 +33,7 @@ VulkanYcbcrConversionCache::VulkanYcbcrConversionCache(VkDevice device)
     : mDevice(device) {}
 
 VkSamplerYcbcrConversion VulkanYcbcrConversionCache::getConversion(
-        VulkanYcbcrConversionCache::Params params) noexcept {
+        VulkanYcbcrConversionCache::Params params) {
     auto iter = mCache.find(params);
     if (UTILS_LIKELY(iter != mCache.end())) {
         return iter->second;
@@ -43,7 +43,7 @@ VkSamplerYcbcrConversion VulkanYcbcrConversionCache::getConversion(
     TextureSwizzle const swizzleArray[] = { chroma.r, chroma.g, chroma.b, chroma.a };
     VkSamplerYcbcrConversionCreateInfo conversionInfo = {
         .sType = VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_CREATE_INFO,
-        .format = fvkutils::getVkFormat(params.format),
+        .format = params.format,
         .ycbcrModel = fvkutils::getYcbcrModelConversion(chroma.ycbcrModel),
         .ycbcrRange = fvkutils::getYcbcrRange(chroma.ycbcrRange),
         .components = fvkutils::getSwizzleMap(swizzleArray),

filament/backend/src/vulkan/VulkanYcbcrConversionCache.h

@@ -30,15 +30,14 @@ namespace filament::backend {
 class VulkanYcbcrConversionCache {
 public:
     struct Params {
-        SamplerYcbcrConversion conversion = {};
-        TextureFormat format = {};
-        uint16_t padding = 0;
-        uint64_t externalFormat = 0;
+        SamplerYcbcrConversion conversion = {}; // 4
+        VkFormat format;                        // 4
+        uint64_t externalFormat = 0;            // 8
     };
     static_assert(sizeof(Params) == 16);
 
     explicit VulkanYcbcrConversionCache(VkDevice device);
-    VkSamplerYcbcrConversion getConversion(Params params) noexcept;
+    VkSamplerYcbcrConversion getConversion(Params params);
     void terminate() noexcept;
 
 private:
@@ -48,7 +47,8 @@ private:
         bool operator()(Params lhs, Params rhs) const noexcept {
             SamplerYcbcrConversion::EqualTo equal;
             return equal(lhs.conversion, rhs.conversion) &&
-                   lhs.externalFormat == rhs.externalFormat;
+                   lhs.externalFormat == rhs.externalFormat &&
+                   lhs.format == rhs.format;
         }
     };
     using ConversionHashFn = utils::hash::MurmurHashFn<Params>;

filament/backend/src/vulkan/memory/ResourceManager.cpp

@@ -31,6 +31,8 @@ ResourceManager::ResourceManager(size_t arenaSize, bool disableUseAfterFreeCheck
     : mHandleAllocatorImpl("Handles", arenaSize, disableUseAfterFreeCheck, disablePoolHandleTags) {}
 
 void ResourceManager::gc() noexcept {
+    FVK_SYSTRACE_CONTEXT();
+    FVK_SYSTRACE_START("ResourceManager::gc");
     auto destroyAll = [this](GcList& list) {
         for (auto const& [type, id]: list) {
             destroyWithType(type, id);
@@ -49,6 +51,7 @@ void ResourceManager::gc() noexcept {
     GcList gcs;
     std::swap(gcs, mGcList);
     destroyAll(gcs);
+    FVK_SYSTRACE_END();
 }
 
 void ResourceManager::terminate() noexcept {

filament/backend/src/vulkan/memory/ResourcePointer.h

@@ -132,6 +132,10 @@ public:
         return id() == other.id() && type() == other.type();
     }
 
+    inline bool operator!=(resource_ptr<D> const& other) const {
+        return !((*this) == other);
+    }
+
     inline explicit operator bool() const {
         return bool(mRef);
     }

filament/backend/src/vulkan/platform/VulkanPlatform.cpp

@@ -212,6 +212,8 @@ ExtensionSet getDeviceExtensions(VkPhysicalDevice device) {
         VK_KHR_SAMPLER_YCBCR_CONVERSION_EXTENSION_NAME,
         VK_ANDROID_EXTERNAL_MEMORY_ANDROID_HARDWARE_BUFFER_EXTENSION_NAME,
         VK_EXT_QUEUE_FAMILY_FOREIGN_EXTENSION_NAME,
+        // This is needed for external images.  See VulkanPlatformAndroid
+        VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME,
 #endif
         // MoltenVk is the only non-conformant implementation we're interested in.
 #if defined(__APPLE__)
@@ -993,30 +995,6 @@ VkQueue VulkanPlatform::getProtectedGraphicsQueue() const noexcept {
     return mImpl->mProtectedGraphicsQueue;
 }
 
-VulkanPlatform::ExternalImageMetadata VulkanPlatform::getExternalImageMetadata(
-        ExternalImageHandleRef externalImage) {
-    return getExternalImageMetadataImpl(externalImage, mImpl->mDevice);
-}
-
-VulkanPlatform::ImageData VulkanPlatform::createExternalImageData(
-        ExternalImageHandleRef externalImage, const ExternalImageMetadata& metadata,
-        uint32_t memoryTypeIndex, VkImageUsageFlags usage) {
-    return createExternalImageDataImpl(externalImage, mImpl->mDevice, metadata, memoryTypeIndex,
-            usage);
-}
-
-VkSampler VulkanPlatform::createExternalSampler(SamplerYcbcrConversion chroma,
-        SamplerParams sampler, uint32_t internalFormat) {
-    return createExternalSamplerImpl(mImpl->mDevice, chroma, sampler, internalFormat);
-}
-
-VkImageView VulkanPlatform::createExternalImageView(SamplerYcbcrConversion chroma,
-        uint32_t internalFormat, VkImage image, VkImageSubresourceRange range,
-        VkImageViewType viewType, VkComponentMapping swizzle) {
-    return createExternalImageViewImpl(mImpl->mDevice, chroma, internalFormat, image, range,
-            viewType, swizzle);
-}
-
 ExtensionSet VulkanPlatform::getSwapchainInstanceExtensions() const {
     return getSwapchainInstanceExtensionsImpl();
 }

filament/backend/src/vulkan/platform/VulkanPlatformAndroid.cpp

@@ -15,14 +15,14 @@
  */
 #include <backend/platforms/VulkanPlatformAndroid.h>
 
+#include "vulkan/VulkanConstants.h"
+#include "vulkan/VulkanContext.h"
+
 #include <backend/DriverEnums.h>
 #include <private/backend/BackendUtilsAndroid.h>
 
-#include "vulkan/VulkanConstants.h"
-
 #include <utils/Panic.h>
 #include "vulkan/utils/Image.h"
-#include "vulkan/utils/Conversion.h"
 
 #include <bluevk/BlueVK.h>
 
@@ -39,6 +39,14 @@ namespace {
 
 VkFormat transformVkFormat(VkFormat format, bool sRGB) {
     if (!sRGB) {
+        switch (format) {
+            case VK_FORMAT_R8G8B8A8_SRGB:
+               return VK_FORMAT_R8G8B8A8_UNORM;
+            case VK_FORMAT_R8G8B8_SRGB:
+                return VK_FORMAT_R8G8B8_UNORM;
+            default:
+                break;
+        }
         return format;
     }
 
@@ -57,7 +65,7 @@ VkFormat transformVkFormat(VkFormat format, bool sRGB) {
 }
 
 bool isProtectedFromUsage(uint64_t usage) {
-    return (usage & AHARDWAREBUFFER_USAGE_PROTECTED_CONTENT) ? true : false;
+    return usage & AHARDWAREBUFFER_USAGE_PROTECTED_CONTENT;
 }
 
 std::pair<VkFormat, VkImageUsageFlags> getVKFormatAndUsage(const AHardwareBuffer_Desc& desc,
@@ -120,7 +128,9 @@ std::pair<VkFormat, VkImageUsageFlags> getVKFormatAndUsage(const AHardwareBuffer
     usage = 0;
     if (desc.usage & AHARDWAREBUFFER_USAGE_GPU_SAMPLED_IMAGE) {
         usage |= VK_IMAGE_USAGE_SAMPLED_BIT;
-        usage |= VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT;
+
+        // We shouldn't be using external samplers as input attachments
+        // usage |= VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT;
     }
     if (desc.usage & AHARDWAREBUFFER_USAGE_GPU_FRAMEBUFFER) {
         if (isDepthFormat) {
@@ -136,87 +146,37 @@ std::pair<VkFormat, VkImageUsageFlags> getVKFormatAndUsage(const AHardwareBuffer
     return { format, usage };
 }
 
-VulkanPlatform::ImageData allocateExternalImage(AHardwareBuffer* buffer, VkDevice device,
-        VulkanPlatform::ExternalImageMetadata const& metadata, uint32_t memoryTypeIndex,
-        VkImageUsageFlags usage) {
-    VulkanPlatform::ImageData data;
-
-    // if external format we need to specifiy it in the allocation
-    const bool useExternalFormat = metadata.format == VK_FORMAT_UNDEFINED;
-
-    const VkExternalFormatANDROID externalFormat = {
-        .sType = VK_STRUCTURE_TYPE_EXTERNAL_FORMAT_ANDROID,
-        .pNext = nullptr,
-        // pass down the format (external means we don't have it VK defined)
-        .externalFormat = metadata.externalFormat,
+std::pair<TextureFormat, TextureUsage> getFilamentFormatAndUsage(const AHardwareBuffer_Desc& desc,
+        bool sRGB) {
+    auto const format = mapToFilamentFormat(desc.format, sRGB);
+    return {
+        format,
+        mapToFilamentUsage(desc.usage, format),
     };
-    const VkExternalMemoryImageCreateInfo externalCreateInfo = {
-        .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
-        .pNext = useExternalFormat ? &externalFormat : nullptr,
-        .handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID,
-    };
-
-    VkImageCreateInfo imageInfo{ .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO };
-    imageInfo.pNext = &externalCreateInfo;
-    imageInfo.format = metadata.format;
-    imageInfo.imageType = VK_IMAGE_TYPE_2D;
-    imageInfo.extent = {
-        metadata.width,
-        metadata.height,
-        1u,
-    };
-    imageInfo.mipLevels = 1;
-    imageInfo.arrayLayers = metadata.layers;
-    imageInfo.samples = metadata.samples;
-    imageInfo.usage = usage;
-
-    VkResult result = vkCreateImage(device, &imageInfo, VKALLOC, &data.first);
-    FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
-            << "vkCreateImage failed with error=" << static_cast<int32_t>(result);
-
-    // Allocate the memory
-    VkImportAndroidHardwareBufferInfoANDROID androidHardwareBufferInfo = {
-        .sType = VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID,
-        .pNext = nullptr,
-        .buffer = buffer,
-    };
-    VkMemoryDedicatedAllocateInfo memoryDedicatedAllocateInfo = {
-        .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
-        .pNext = &androidHardwareBufferInfo,
-        .image = data.first,
-        .buffer = VK_NULL_HANDLE,
-    };
-    VkMemoryAllocateInfo allocInfo = {
-        .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
-        .pNext = &memoryDedicatedAllocateInfo,
-        .allocationSize = metadata.allocationSize,
-        .memoryTypeIndex = memoryTypeIndex,
-    };
-    result = vkAllocateMemory(device, &allocInfo, VKALLOC, &data.second);
-    FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
-            << "vkAllocateMemory failed with error=" << static_cast<int32_t>(result);
-
-    return data;
 }
 
 }// namespace
 
-VulkanPlatformAndroid::ExternalImageVulkanAndroid::~ExternalImageVulkanAndroid() = default;
+VulkanPlatformAndroid::ExternalImageVulkanAndroid::~ExternalImageVulkanAndroid() {
+    if (__builtin_available(android 26, *)) {
+        if (aHardwareBuffer) {
+            AHardwareBuffer_release(aHardwareBuffer);
+        }
+    }
+}
 
 Platform::ExternalImageHandle VulkanPlatformAndroid::createExternalImage(
         AHardwareBuffer const* buffer, bool sRGB) noexcept {
     if (__builtin_available(android 26, *)) {
+        auto bufferImpl = const_cast<AHardwareBuffer*>(buffer);
+        AHardwareBuffer_acquire(bufferImpl);
+
         AHardwareBuffer_Desc hardwareBufferDescription = {};
         AHardwareBuffer_describe(buffer, &hardwareBufferDescription);
 
         auto* const p = new (std::nothrow) ExternalImageVulkanAndroid;
         p->aHardwareBuffer = const_cast<AHardwareBuffer*>(buffer);
         p->sRGB = sRGB;
-        p->height = hardwareBufferDescription.height;
-        p->width = hardwareBufferDescription.width;
-        TextureFormat textureFormat = mapToFilamentFormat(hardwareBufferDescription.format, sRGB);
-        p->format = textureFormat;
-        p->usage = mapToFilamentUsage(hardwareBufferDescription.usage, textureFormat);
         return Platform::ExternalImageHandle{ p };
     }
 
@@ -225,23 +185,20 @@ Platform::ExternalImageHandle VulkanPlatformAndroid::createExternalImage(
 
 VulkanPlatformAndroid::ExternalImageDescAndroid VulkanPlatformAndroid::getExternalImageDesc(
         ExternalImageHandleRef externalImage) const noexcept {
-    auto const* fvkExternalImage =
-            static_cast<ExternalImageVulkanAndroid const*>(externalImage.get());
-
+    auto metadata = extractExternalImageMetadata(externalImage);
     return {
-        .width = fvkExternalImage->width,
-        .height = fvkExternalImage->height,
-        .format = fvkExternalImage->format,
-        .usage = fvkExternalImage->usage,
+        .width = metadata.width,
+        .height = metadata.height,
+        .format = metadata.filamentFormat,
+        .usage = metadata.filamentUsage,
     };
 }
 
-VulkanPlatform::ExternalImageMetadata VulkanPlatformAndroid::getExternalImageMetadata(
-        ExternalImageHandleRef externalImage) {
-    auto const* fvkExternalImage =
-            static_cast<ExternalImageVulkanAndroid const*>(externalImage.get());
+VulkanPlatform::ExternalImageMetadata VulkanPlatformAndroid::extractExternalImageMetadata(
+        ExternalImageHandleRef image) const {
+    auto const* fvkExternalImage = static_cast<ExternalImageVulkanAndroid const*>(image.get());
 
-    ExternalImageMetadata metadata;
+    ExternalImageMetadata metadata = {};
     AHardwareBuffer* buffer = fvkExternalImage->aHardwareBuffer;
     if (__builtin_available(android 26, *)) {
         AHardwareBuffer_Desc bufferDesc;
@@ -249,16 +206,29 @@ VulkanPlatform::ExternalImageMetadata VulkanPlatformAndroid::getExternalImageMet
         metadata.width = bufferDesc.width;
         metadata.height = bufferDesc.height;
         metadata.layers = bufferDesc.layers;
-        metadata.isProtected = isProtectedFromUsage(bufferDesc.usage);
         std::tie(metadata.format, metadata.usage) =
                 getVKFormatAndUsage(bufferDesc, fvkExternalImage->sRGB);
-    }
+        std::tie(metadata.filamentFormat, metadata.filamentUsage) =
+                getFilamentFormatAndUsage(bufferDesc, fvkExternalImage->sRGB);
 
+        if (isProtectedFromUsage(bufferDesc.usage)) {
+            metadata.filamentUsage |= TextureUsage::PROTECTED;
+        }
+
+        // TODO: The following seems unnecessary. we should be able to discern directly from the
+        // bufferDesc.
+        if (any(metadata.filamentUsage & TextureUsage::BLIT_SRC)) {
+            metadata.usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+        }
+
+        if (any(metadata.filamentUsage & (TextureUsage::BLIT_DST | TextureUsage::UPLOADABLE))) {
+            metadata.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT;
+        }
+    }
     metadata.samples = VK_SAMPLE_COUNT_1_BIT;
 
     VkAndroidHardwareBufferFormatPropertiesANDROID formatInfo = {
         .sType = VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_FORMAT_PROPERTIES_ANDROID,
-        .pNext = nullptr,
     };
     VkAndroidHardwareBufferPropertiesANDROID properties = {
         .sType = VK_STRUCTURE_TYPE_ANDROID_HARDWARE_BUFFER_PROPERTIES_ANDROID,
@@ -268,138 +238,155 @@ VulkanPlatform::ExternalImageMetadata VulkanPlatformAndroid::getExternalImageMet
     FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
             << "vkGetAndroidHardwareBufferProperties failed with error="
             << static_cast<int32_t>(result);
-
     VkFormat bufferPropertiesFormat = transformVkFormat(formatInfo.format, fvkExternalImage->sRGB);
     FILAMENT_CHECK_POSTCONDITION(metadata.format == bufferPropertiesFormat)
             << "mismatched image format( " << metadata.format << ") and queried format("
             << bufferPropertiesFormat << ") for external image (AHB)";
-    metadata.externalFormat = formatInfo.externalFormat;
+
+    bool const requiresConversion =
+        metadata.format == VK_FORMAT_UNDEFINED ||
+        fvkutils::isVKYcbcrConversionFormat(metadata.format);
+    if (requiresConversion) {
+      metadata.format = VK_FORMAT_UNDEFINED;
+      metadata.usage = VK_IMAGE_USAGE_SAMPLED_BIT;
+      metadata.externalFormat = formatInfo.externalFormat;
+    } else {
+      metadata.externalFormat = 0;
+    }
+
     metadata.allocationSize = properties.allocationSize;
     metadata.memoryTypeBits = properties.memoryTypeBits;
+
+    metadata.ycbcrConversionComponents = formatInfo.samplerYcbcrConversionComponents;
+    metadata.ycbcrModel = formatInfo.suggestedYcbcrModel;
+    metadata.ycbcrRange = formatInfo.suggestedYcbcrRange;
+    metadata.xChromaOffset = formatInfo.suggestedXChromaOffset;
+    metadata.yChromaOffset = formatInfo.suggestedYChromaOffset;
+
     return metadata;
 }
 
-VulkanPlatformAndroid::ImageData VulkanPlatformAndroid::createExternalImageData(
-        ExternalImageHandleRef externalImage, const ExternalImageMetadata& metadata,
-        uint32_t memoryTypeIndex, VkImageUsageFlags usage) {
+VulkanPlatform::ImageData VulkanPlatformAndroid::createVkImageFromExternal(
+        ExternalImageHandleRef externalImage) const {
+    auto metadata = extractExternalImageMetadata(externalImage);
+
     auto const* fvkExternalImage =
             static_cast<ExternalImageVulkanAndroid const*>(externalImage.get());
-    ImageData data = allocateExternalImage(fvkExternalImage->aHardwareBuffer, getDevice(), metadata,
-            memoryTypeIndex, usage);
-    VkResult result = vkBindImageMemory(getDevice(), data.first, data.second, 0);
-    FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
-        << "vkBindImageMemory error=" << static_cast<int32_t>(result);
-    return data;
-}
+    AHardwareBuffer* buffer = fvkExternalImage->aHardwareBuffer;
 
-VkImageView VulkanPlatform::createExternalImageViewImpl(VkDevice device, SamplerYcbcrConversion chroma,
-            uint32_t internalFormat, VkImage image, VkImageSubresourceRange range,
-            VkImageViewType viewType, VkComponentMapping swizzle){
-    VkExternalFormatANDROID externalFormat = {
-        .sType = VK_STRUCTURE_TYPE_EXTERNAL_FORMAT_ANDROID,
-        .externalFormat = internalFormat,
+    VkDevice const device = getDevice();
+    VkPhysicalDevice const physicalDevice = getPhysicalDevice();
+    auto buildImage = [&](ExternalImageMetadata const& metadata) {
+        bool const isExternal = metadata.externalFormat != 0;
+        VkExternalFormatANDROID const externalFormat = {
+            .sType = VK_STRUCTURE_TYPE_EXTERNAL_FORMAT_ANDROID,
+            .pNext = nullptr,
+            .externalFormat = metadata.externalFormat,
+        };
+        VkExternalMemoryImageCreateInfo externalCreateInfo = {
+            .sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
+            .pNext = isExternal ? &externalFormat : nullptr,
+            .handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID,
+        };
+
+        VkFormat formats[2] = {};
+        VkImageFormatListCreateInfo imageFormatListInfo = {
+            .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO,
+            .pNext = nullptr,
+            .viewFormatCount = 2,
+            .pViewFormats = formats,
+        };
+
+        if (fvkExternalImage->sRGB) {
+            formats[0] = metadata.format;
+            formats[1] = transformVkFormat(metadata.format, /*sRGB=*/false);
+            imageFormatListInfo.pNext = externalCreateInfo.pNext;
+            externalCreateInfo.pNext = &imageFormatListInfo;
+        }
+
+        VkImageCreateInfo const imageInfo = {
+            .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+            .pNext = &externalCreateInfo,
+            .flags = fvkExternalImage->sRGB ? VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT : 0u,
+            .imageType = VK_IMAGE_TYPE_2D,
+            // For non external images, use the same format as the AHB, which isn't in SRGB
+            // Fix VUID-VkMemoryAllocateInfo-pNext-02387
+            .format = transformVkFormat(metadata.format, /*sRGB=*/false),
+            .extent = {
+                metadata.width,
+                metadata.height,
+                1u,
+            },
+            .mipLevels = 1,
+            .arrayLayers = metadata.layers,
+            .samples = metadata.samples,
+            .usage = metadata.usage,
+        };
+        VkImage image;
+        VkResult result = vkCreateImage(device, &imageInfo, VKALLOC, &image);
+        FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
+            << "vkCreateImage failed with error=" << static_cast<int32_t>(result);
+        return image;
     };
 
-    TextureSwizzle const swizzleArray[] = {chroma.r, chroma.g, chroma.b, chroma.a};
-    VkSamplerYcbcrConversionCreateInfo conversionInfo = {
-        .sType = VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_CREATE_INFO,
-        .pNext = &externalFormat,
-        .format = VK_FORMAT_UNDEFINED,
-        .ycbcrModel = fvkutils::getYcbcrModelConversion(chroma.ycbcrModel),
-        .ycbcrRange = fvkutils::getYcbcrRange(chroma.ycbcrRange),
-        .components = fvkutils::getSwizzleMap(swizzleArray),
-        .xChromaOffset = fvkutils::getChromaLocation(chroma.xChromaOffset),
-        .yChromaOffset = fvkutils::getChromaLocation(chroma.yChromaOffset),
-        .chromaFilter = fvkutils::getFilter(chroma.chromaFilter),
-    };
-    VkSamplerYcbcrConversion conversion = VK_NULL_HANDLE;
-    VkResult result = vkCreateSamplerYcbcrConversion(device, &conversionInfo,
-                                                     nullptr, &conversion);
-    FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
-        << "Unable to create Ycbcr Conversion."
-        << " error=" << static_cast<int32_t>(result);
+    auto allocMem = [&](VkImage image, ExternalImageMetadata const& metadata) {
+        bool const isExternal = metadata.externalFormat != 0;
+        // Allocate the memory
+        VkImportAndroidHardwareBufferInfoANDROID const androidHardwareBufferInfo = {
+            .sType = VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID,
+            .buffer = buffer,
+        };
+        VkMemoryDedicatedAllocateInfo const memoryDedicatedAllocateInfo = {
+            .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
+            .pNext = &androidHardwareBufferInfo,
+            .image = image,
+            .buffer = VK_NULL_HANDLE,
+        };
+        VkPhysicalDeviceMemoryProperties memoryProperties;
+        vkGetPhysicalDeviceMemoryProperties(physicalDevice, &memoryProperties);
+        VkMemoryPropertyFlags requiredMemoryFlags =
+                !isExternal && any(metadata.filamentUsage & TextureUsage::UPLOADABLE)
+                        ? VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
+                        : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
 
-    VkSamplerYcbcrConversionInfo samplerYcbcrConversionInfo = {
-        .sType = VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_INFO,
-        .conversion = conversion,
+        if (any(metadata.filamentUsage & TextureUsage::PROTECTED)) {
+            requiredMemoryFlags |= VK_MEMORY_PROPERTY_PROTECTED_BIT;
+        }
+
+        uint32_t const memoryTypeIndex = VulkanContext::selectMemoryType(memoryProperties,
+                metadata.memoryTypeBits, requiredMemoryFlags);
+
+        VkMemoryAllocateInfo const allocInfo = {
+            .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+            .pNext = &memoryDedicatedAllocateInfo,
+            .allocationSize = metadata.allocationSize,
+            .memoryTypeIndex = memoryTypeIndex,
+        };
+        VkDeviceMemory memory;
+        VkResult result = vkAllocateMemory(device, &allocInfo, VKALLOC, &memory);
+        FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
+                << "vkAllocateMemory failed with error=" << static_cast<int32_t>(result);
+        result = vkBindImageMemory(getDevice(), image, memory, 0);
+        FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
+                << "vkBindImageMemory error=" << static_cast<int32_t>(result);
+        return memory;
     };
 
-    VkImageViewCreateInfo viewInfo = {
-        .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
-        .pNext = &samplerYcbcrConversionInfo,
-        .flags = 0,
-        .image = image,
-        .viewType = viewType,
-        .format = VK_FORMAT_UNDEFINED,
-        .components = swizzle,
-        .subresourceRange = range,
-    };
-    VkImageView imageView;
-    result = vkCreateImageView(device, &viewInfo, VKALLOC, &imageView);
-    FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
-        << "Unable to create VkImageView."
-        << " error=" << static_cast<int32_t>(result);
+    VulkanPlatform::ImageData::Bundle internal = {}, external = {};
+    auto img = buildImage(metadata);
+    auto mem = allocMem(img, metadata);
 
-    return imageView;
-}
+    // Note that we're always choosing a non-externally sampled format if it exists.
+    if (metadata.externalFormat == 0) {
+        internal = { img, mem };
+    } else {
+        external = { img, mem };
+    }
 
-VkSampler VulkanPlatform::createExternalSamplerImpl(
-        VkDevice device, SamplerYcbcrConversion chroma, SamplerParams params,
-        uint32_t internalFormat) {
-    VkExternalFormatANDROID externalFormat = {
-        .sType = VK_STRUCTURE_TYPE_EXTERNAL_FORMAT_ANDROID,
-        .externalFormat = internalFormat,
+    return {
+        .internal = internal,
+        .external = external,
     };
-
-    TextureSwizzle const swizzleArray[] = {chroma.r, chroma.g, chroma.b, chroma.a};
-    VkSamplerYcbcrConversionCreateInfo conversionInfo = {
-        .sType = VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_CREATE_INFO,
-        .pNext = &externalFormat,
-        .format = VK_FORMAT_UNDEFINED,
-        .ycbcrModel = fvkutils::getYcbcrModelConversion(chroma.ycbcrModel),
-        .ycbcrRange = fvkutils::getYcbcrRange(chroma.ycbcrRange),
-        .components = fvkutils::getSwizzleMap(swizzleArray),
-        .xChromaOffset = fvkutils::getChromaLocation(chroma.xChromaOffset),
-        .yChromaOffset = fvkutils::getChromaLocation(chroma.yChromaOffset),
-        .chromaFilter = fvkutils::getFilter(chroma.chromaFilter),
-    };
-    VkSamplerYcbcrConversion conversion = VK_NULL_HANDLE;
-    VkResult result = vkCreateSamplerYcbcrConversion(device, &conversionInfo,
-                                                     nullptr, &conversion);
-    FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
-        << "Unable to create Ycbcr Conversion."
-        << " error=" << static_cast<int32_t>(result);
-
-    VkSamplerYcbcrConversionInfo samplerYcbcrConversionInfo = {
-        .sType = VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_INFO,
-        .pNext = nullptr,
-        .conversion = conversion,
-    };
-
-    VkSamplerCreateInfo samplerInfo = {
-        .sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO,
-        .pNext = &samplerYcbcrConversionInfo,
-        .magFilter = fvkutils::getFilter(params.filterMag),
-        .minFilter = fvkutils::getFilter(params.filterMin),
-        .mipmapMode = fvkutils::getMipmapMode(params.filterMin),
-        .addressModeU = fvkutils::getWrapMode(params.wrapS),
-        .addressModeV = fvkutils::getWrapMode(params.wrapT),
-        .addressModeW = fvkutils::getWrapMode(params.wrapR),
-        .anisotropyEnable = params.anisotropyLog2 == 0 ? VK_FALSE : VK_TRUE,
-        .maxAnisotropy = (float)(1u << params.anisotropyLog2),
-        .compareEnable = fvkutils::getCompareEnable(params.compareMode),
-        .compareOp = fvkutils::getCompareOp(params.compareFunc),
-        .minLod = 0.0f,
-        .maxLod = fvkutils::getMaxLod(params.filterMin),
-        .borderColor = VK_BORDER_COLOR_INT_OPAQUE_BLACK,
-        .unnormalizedCoordinates = VK_FALSE,
-    };
-    VkSampler sampler;
-    result = vkCreateSampler(device, &samplerInfo, VKALLOC, &sampler);
-    FILAMENT_CHECK_POSTCONDITION(result == VK_SUCCESS)
-        << "Unable to create sampler."
-        << " error=" << static_cast<int32_t>(result);
-    return sampler;
 }
 
 VulkanPlatform::ExtensionSet VulkanPlatformAndroid::getSwapchainInstanceExtensions() const {
@@ -413,7 +400,7 @@ VulkanPlatform::SurfaceBundle VulkanPlatformAndroid::createVkSurfaceKHR(void* na
     VkSurfaceKHR surface;
     VkExtent2D extent;
 
-    VkAndroidSurfaceCreateInfoKHR const createInfo{
+    VkAndroidSurfaceCreateInfoKHR const createInfo = {
         .sType = VK_STRUCTURE_TYPE_ANDROID_SURFACE_CREATE_INFO_KHR,
         .window = (ANativeWindow*) nativeWindow,
     };
@@ -427,20 +414,9 @@ VulkanPlatform::SurfaceBundle VulkanPlatformAndroid::createVkSurfaceKHR(void* na
 // Deprecated platform dependent helper methods
 VulkanPlatform::ExtensionSet VulkanPlatform::getSwapchainInstanceExtensionsImpl() { return {}; }
 
-VulkanPlatform::ExternalImageMetadata VulkanPlatform::getExternalImageMetadataImpl(
-        ExternalImageHandleRef externalImage, VkDevice device) {
-    return ExternalImageMetadata{};
-}
-
-VulkanPlatform::ImageData VulkanPlatform::createExternalImageDataImpl(
-        ExternalImageHandleRef externalImage, VkDevice device,
-        const ExternalImageMetadata& metadata, uint32_t memoryTypeIndex, VkImageUsageFlags usage) {
-    return ImageData{};
-}
-
 VulkanPlatform::SurfaceBundle VulkanPlatform::createVkSurfaceKHRImpl(void* nativeWindow,
         VkInstance instance, uint64_t flags) noexcept {
     return SurfaceBundle{};
 }
 
-}// namespace filament::backend
+} // namespace filament::backend

filament/backend/src/vulkan/platform/VulkanPlatformApple.mm

@@ -43,30 +43,6 @@ VulkanPlatform::ExtensionSet VulkanPlatform::getSwapchainInstanceExtensionsImpl(
     return ret;
 }
 
-VulkanPlatform::ExternalImageMetadata VulkanPlatform::getExternalImageMetadataImpl(
-        ExternalImageHandleRef externalImage, VkDevice device) {
-    return {};
-}
-
-VulkanPlatform::ImageData VulkanPlatform::createExternalImageDataImpl(
-        ExternalImageHandleRef externalImage, VkDevice device,
-        const ExternalImageMetadata& metadata, uint32_t memoryTypeIndex, VkImageUsageFlags usage) {
-    return {};
-}
-
-VkSampler VulkanPlatform::createExternalSamplerImpl(VkDevice device,
-    SamplerYcbcrConversion chroma,
-    SamplerParams sampler,
-    uint32_t internalFormat) {
-    return VK_NULL_HANDLE;
-}
-
-VkImageView VulkanPlatform::createExternalImageViewImpl(VkDevice device,
-        SamplerYcbcrConversion chroma, uint32_t internalFormat, VkImage image,
-        VkImageSubresourceRange range, VkImageViewType viewType, VkComponentMapping swizzle) {
-    return VK_NULL_HANDLE;
-}
-
 VulkanPlatform::SurfaceBundle VulkanPlatform::createVkSurfaceKHRImpl(void* nativeWindow,
         VkInstance instance, uint64_t flags) noexcept {
     VkSurfaceKHR surface;

filament/backend/src/vulkan/platform/VulkanPlatformLinuxWindows.cpp

@@ -84,30 +84,6 @@ using namespace bluevk;
 
 namespace filament::backend {
 
-VulkanPlatform::ExternalImageMetadata VulkanPlatform::getExternalImageMetadataImpl(
-        ExternalImageHandleRef externalImage, VkDevice device) {
-    return {};
-}
-
-VulkanPlatform::ImageData VulkanPlatform::createExternalImageDataImpl(
-        ExternalImageHandleRef externalImage, VkDevice device,
-        const ExternalImageMetadata& metadata, uint32_t memoryTypeIndex, VkImageUsageFlags usage) {
-    return {};
-}
-
-VkSampler VulkanPlatform::createExternalSamplerImpl(VkDevice device,
-    SamplerYcbcrConversion chroma,
-    SamplerParams sampler,
-    uint32_t internalFormat) {
-    return VK_NULL_HANDLE;
-}
-
-VkImageView VulkanPlatform::createExternalImageViewImpl(VkDevice device,
-        SamplerYcbcrConversion chroma, uint32_t internalFormat, VkImage image,
-        VkImageSubresourceRange range, VkImageViewType viewType, VkComponentMapping swizzle) {
-    return VK_NULL_HANDLE;
-}
-
 VulkanPlatform::ExtensionSet VulkanPlatform::getSwapchainInstanceExtensionsImpl() {
     VulkanPlatform::ExtensionSet const ret = {
 #if defined(__linux__) && defined(FILAMENT_SUPPORTS_WAYLAND)

filament/backend/src/vulkan/utils/Conversion.cpp

@@ -51,6 +51,8 @@ VkFormat getVkFormat(ElementType type, bool normalized, bool integer) {
                 return VK_FORMAT_UNDEFINED;
         }
     }
+
+    // Non-normalized case
     switch (type) {
         // Single Component Types
         case ElementType::BYTE: return integer ? VK_FORMAT_R8_SINT : VK_FORMAT_R8_SSCALED;

filament/backend/src/vulkan/utils/Image.cpp

@@ -184,6 +184,46 @@ bool isVkStencilFormat(VkFormat format) {
     return (getImageAspect(format) & VK_IMAGE_ASPECT_STENCIL_BIT) != 0;
 }
 
+bool isVKYcbcrConversionFormat(VkFormat format) {
+    switch (format) {
+        case VK_FORMAT_G8B8G8R8_422_UNORM:
+        case VK_FORMAT_B8G8R8G8_422_UNORM:
+        case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
+        case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
+        case VK_FORMAT_G8_B8_R8_3PLANE_422_UNORM:
+        case VK_FORMAT_G8_B8R8_2PLANE_422_UNORM:
+        case VK_FORMAT_G8_B8_R8_3PLANE_444_UNORM:
+        case VK_FORMAT_G10X6B10X6G10X6R10X6_422_UNORM_4PACK16:
+        case VK_FORMAT_B10X6G10X6R10X6G10X6_422_UNORM_4PACK16:
+        case VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_420_UNORM_3PACK16:
+        case VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16:
+        case VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_422_UNORM_3PACK16:
+        case VK_FORMAT_G10X6_B10X6R10X6_2PLANE_422_UNORM_3PACK16:
+        case VK_FORMAT_G10X6_B10X6_R10X6_3PLANE_444_UNORM_3PACK16:
+        case VK_FORMAT_G12X4B12X4G12X4R12X4_422_UNORM_4PACK16:
+        case VK_FORMAT_B12X4G12X4R12X4G12X4_422_UNORM_4PACK16:
+        case VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_420_UNORM_3PACK16:
+        case VK_FORMAT_G12X4_B12X4R12X4_2PLANE_420_UNORM_3PACK16:
+        case VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_422_UNORM_3PACK16:
+        case VK_FORMAT_G12X4_B12X4R12X4_2PLANE_422_UNORM_3PACK16:
+        case VK_FORMAT_G12X4_B12X4_R12X4_3PLANE_444_UNORM_3PACK16:
+        case VK_FORMAT_G16B16G16R16_422_UNORM:
+        case VK_FORMAT_B16G16R16G16_422_UNORM:
+        case VK_FORMAT_G16_B16_R16_3PLANE_420_UNORM:
+        case VK_FORMAT_G16_B16R16_2PLANE_420_UNORM:
+        case VK_FORMAT_G16_B16_R16_3PLANE_422_UNORM:
+        case VK_FORMAT_G16_B16R16_2PLANE_422_UNORM:
+        case VK_FORMAT_G16_B16_R16_3PLANE_444_UNORM:
+        case VK_FORMAT_G8_B8R8_2PLANE_444_UNORM:
+        case VK_FORMAT_G10X6_B10X6R10X6_2PLANE_444_UNORM_3PACK16:
+        case VK_FORMAT_G12X4_B12X4R12X4_2PLANE_444_UNORM_3PACK16:
+        case VK_FORMAT_G16_B16R16_2PLANE_444_UNORM:
+            return true;
+        default:
+            return false;
+    }
+}
+
 static uint32_t mostSignificantBit(uint32_t x) { return 1ul << (31ul - utils::clz(x)); }
 
 uint8_t reduceSampleCount(uint8_t sampleCount, VkSampleCountFlags mask) {

filament/backend/src/vulkan/utils/Image.h

@@ -98,6 +98,8 @@ bool isVkDepthFormat(VkFormat format);
 
 bool isVkStencilFormat(VkFormat format);
 
+bool isVKYcbcrConversionFormat(VkFormat format);
+
 VkImageAspectFlags getImageAspect(VkFormat format);
 
 uint8_t reduceSampleCount(uint8_t sampleCount, VkSampleCountFlags mask);

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

@@ -17,7 +17,6 @@
 #include "webgpu/WebGPUDriver.h"
 
 #include "WebGPUSwapChain.h"
-#include "webgpu/WebGPUConstants.h"
 #include <backend/platforms/WebGPUPlatform.h>
 
 #include "CommandStreamDispatcher.h"
@@ -184,8 +183,6 @@ void printAdapterDetails(wgpu::Adapter const& adapter) {
 }
 #endif
 
-
-
 #if FWGPU_ENABLED(FWGPU_PRINT_SYSTEM)
 void printDeviceDetails(wgpu::Device const& device) {
     wgpu::SupportedFeatures supportedFeatures{};
@@ -237,6 +234,7 @@ WebGPUDriver::WebGPUDriver(WebGPUPlatform& platform, const Platform::DriverConfi
 #if FWGPU_ENABLED(FWGPU_PRINT_SYSTEM)
     printDeviceDetails(mDevice);
 #endif
+    mQueue = mDevice.GetQueue();
 }
 
 WebGPUDriver::~WebGPUDriver() noexcept = default;
@@ -319,12 +317,18 @@ void WebGPUDriver::destroyIndexBuffer(Handle<HwIndexBuffer> ibh) {
 }
 
 void WebGPUDriver::destroyBufferObject(Handle<HwBufferObject> boh) {
+    if (boh) {
+        destructHandle<WGPUBufferObject>(boh);
+    }
 }
 
 void WebGPUDriver::destroyTexture(Handle<HwTexture> th) {
 }
 
 void WebGPUDriver::destroyProgram(Handle<HwProgram> ph) {
+    if (ph) {
+        destructHandle<WGPUProgram>(ph);
+    }
 }
 
 void WebGPUDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
@@ -350,6 +354,9 @@ void WebGPUDriver::destroyDescriptorSetLayout(Handle<HwDescriptorSetLayout> tqh)
 }
 
 void WebGPUDriver::destroyDescriptorSet(Handle<HwDescriptorSet> tqh) {
+    if (tqh) {
+        destructHandle<WebGPUDescriptorSet>(tqh);
+    }
 }
 
 Handle<HwSwapChain> WebGPUDriver::createSwapChainS() noexcept {
@@ -369,7 +376,7 @@ 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 {
@@ -401,7 +408,7 @@ Handle<HwVertexBuffer> WebGPUDriver::createVertexBufferS() noexcept {
 }
 
 Handle<HwDescriptorSet> WebGPUDriver::createDescriptorSetS() noexcept {
-    return Handle<HwDescriptorSet>((Handle<HwDescriptorSet>::HandleId) mNextFakeHandle++);
+    return allocHandle<WebGPUDescriptorSet>();
 }
 
 Handle<HwRenderPrimitive> WebGPUDriver::createRenderPrimitiveS() noexcept {
@@ -441,7 +448,6 @@ void WebGPUDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
     assert_invariant(!mSwapChain);
     wgpu::Surface surface = mPlatform.createSurface(nativeWindow, flags);
 
-    mQueue = mDevice.GetQueue();
     wgpu::Extent2D surfaceSize = mPlatform.getSurfaceExtent(mNativeWindow);
     mSwapChain = constructHandle<WebGPUSwapChain>(sch, std::move(surface), surfaceSize, mAdapter,
             mDevice, flags);
@@ -464,16 +470,27 @@ void WebGPUDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch, uint32_t wi
         uint32_t height, uint64_t flags) {}
 
 void WebGPUDriver::createVertexBufferInfoR(Handle<HwVertexBufferInfo> vbih, uint8_t bufferCount,
-        uint8_t attributeCount, AttributeArray attributes) {}
+        uint8_t attributeCount, AttributeArray attributes) {
+    constructHandle<WGPUVertexBufferInfo>(vbih, bufferCount, attributeCount, attributes);
+}
 
 void WebGPUDriver::createVertexBufferR(Handle<HwVertexBuffer> vbh, uint32_t vertexCount,
-        Handle<HwVertexBufferInfo> vbih) {}
+        Handle<HwVertexBufferInfo> vbih) {
+    auto* vertexBufferInfo = handleCast<WGPUVertexBufferInfo>(vbih);
+    constructHandle<WGPUVertexBuffer>(vbh, mDevice, vertexCount, vertexBufferInfo->bufferCount,
+            vbih);
+}
 
 void WebGPUDriver::createIndexBufferR(Handle<HwIndexBuffer> ibh, ElementType elementType,
-        uint32_t indexCount, BufferUsage usage) {}
+        uint32_t indexCount, BufferUsage usage) {
+    auto elementSize = static_cast<uint8_t>(getElementTypeSize(elementType));
+    constructHandle<WGPUIndexBuffer>(ibh, mDevice, elementSize, indexCount);
+}
 
 void WebGPUDriver::createBufferObjectR(Handle<HwBufferObject> boh, uint32_t byteCount,
-        BufferObjectBinding bindingType, BufferUsage usage) {}
+        BufferObjectBinding bindingType, BufferUsage usage) {
+    constructHandle<WGPUBufferObject>(boh, mDevice, bindingType, byteCount);
+}
 
 void WebGPUDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint8_t levels,
         TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
@@ -503,9 +520,20 @@ void WebGPUDriver::importTextureR(Handle<HwTexture> th, intptr_t id, SamplerType
         uint32_t depth, TextureUsage usage) {}
 
 void WebGPUDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph, Handle<HwVertexBuffer> vbh,
-        Handle<HwIndexBuffer> ibh, PrimitiveType pt) {}
+        Handle<HwIndexBuffer> ibh, PrimitiveType pt) {
+    assert_invariant(mDevice);
 
-void WebGPUDriver::createProgramR(Handle<HwProgram> ph, Program&& program) {}
+    auto* renderPrimitive = constructHandle<WGPURenderPrimitive>(rph);
+    auto* vertexBuffer = handleCast<WGPUVertexBuffer>(vbh);
+    auto* indexBuffer = handleCast<WGPUIndexBuffer>(ibh);
+    renderPrimitive->vertexBuffer = vertexBuffer;
+    renderPrimitive->indexBuffer = indexBuffer;
+    renderPrimitive->type = pt;
+}
+
+void WebGPUDriver::createProgramR(Handle<HwProgram> ph, Program&& program) {
+    constructHandle<WGPUProgram>(ph, mDevice, program);
+}
 
 void WebGPUDriver::createDefaultRenderTargetR(Handle<HwRenderTarget> rth, int) {
     assert_invariant(!mDefaultRenderTarget);
@@ -527,7 +555,10 @@ void WebGPUDriver::createDescriptorSetLayoutR(Handle<HwDescriptorSetLayout> dslh
 }
 
 void WebGPUDriver::createDescriptorSetR(Handle<HwDescriptorSet> dsh,
-        Handle<HwDescriptorSetLayout> dslh) {}
+        Handle<HwDescriptorSetLayout> dslh) {
+    auto layout = handleCast<WebGPUDescriptorSetLayout>(dslh);
+    constructHandle<WebGPUDescriptorSet>(dsh, layout->getLayout(), layout->getLayoutSize());
+}
 
 Handle<HwStream> WebGPUDriver::createStreamNative(void* nativeStream) {
     return {};
@@ -567,7 +598,7 @@ bool WebGPUDriver::isTextureFormatSupported(TextureFormat format) {
 }
 
 bool WebGPUDriver::isTextureSwizzleSupported() {
-    return true;
+    return false;
 }
 
 bool WebGPUDriver::isTextureFormatMipmappable(TextureFormat format) {
@@ -656,20 +687,21 @@ size_t WebGPUDriver::getMaxArrayTextureLayers() {
 
 void WebGPUDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor&& p,
         uint32_t byteOffset) {
-    scheduleDestroy(std::move(p));
+    updateGPUBuffer(handleCast<WGPUIndexBuffer>(ibh), std::move(p), byteOffset);
 }
 
 void WebGPUDriver::updateBufferObject(Handle<HwBufferObject> ibh, BufferDescriptor&& p,
         uint32_t byteOffset) {
-    scheduleDestroy(std::move(p));
+    updateGPUBuffer(handleCast<WGPUBufferObject>(ibh), std::move(p), byteOffset);
 }
 
-void WebGPUDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> ibh, BufferDescriptor&& p,
-        uint32_t byteOffset) {
-    scheduleDestroy(std::move(p));
+void WebGPUDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> ibh,
+        BufferDescriptor&& p, uint32_t byteOffset) {
+    updateGPUBuffer(handleCast<WGPUBufferObject>(ibh), std::move(p), byteOffset);
 }
 
 void WebGPUDriver::resetBufferObject(Handle<HwBufferObject> boh) {
+    // Is there something that needs to be done here? Vulkan has left it unimplemented.
 }
 
 void WebGPUDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t index,
@@ -677,7 +709,8 @@ void WebGPUDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t in
     auto* vertexBuffer = handleCast<WGPUVertexBuffer>(vbh);
     auto* bufferObject = handleCast<WGPUBufferObject>(boh);
     assert_invariant(index < vertexBuffer->buffers.size());
-    vertexBuffer->setBuffer(bufferObject, index);
+    assert_invariant(bufferObject->buffer.GetUsage() & wgpu::BufferUsage::Vertex);
+    vertexBuffer->buffers[index] = bufferObject->buffer;
 }
 
 void WebGPUDriver::update3DImage(Handle<HwTexture> th,
@@ -715,6 +748,7 @@ void WebGPUDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     };
     mCommandEncoder = mDevice.CreateCommandEncoder(&commandEncoderDescriptor);
     assert_invariant(mCommandEncoder);
+
     // TODO: Remove this code once WebGPU pipeline is implemented
     static float red = 1.0f;
     if (red - 0.01 > 0) {
@@ -828,6 +862,19 @@ void WebGPUDriver::bindPipeline(PipelineState const& pipelineState) {
 }
 
 void WebGPUDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
+    auto* renderPrimitive = handleCast<WGPURenderPrimitive>(rph);
+
+    // This *must* match the WGPUVertexBufferInfo that was bound in bindPipeline(). But we want
+    // to allow to call this before bindPipeline(), so the validation can only happen in draw()
+    auto vbi = handleCast<WGPUVertexBufferInfo>(renderPrimitive->vertexBuffer->vbih);
+    assert_invariant(
+            vbi->getVertexBufferLayoutSize() == renderPrimitive->vertexBuffer->buffers.size());
+    for (uint32_t i = 0; i < vbi->getVertexBufferLayoutSize(); i++) {
+        mRenderPassEncoder.SetVertexBuffer(i, renderPrimitive->vertexBuffer->buffers[i]);
+    }
+
+    mRenderPassEncoder.SetIndexBuffer(renderPrimitive->indexBuffer->buffer,
+            renderPrimitive->indexBuffer->indexFormat);
 }
 
 void WebGPUDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
@@ -853,25 +900,48 @@ void WebGPUDriver::endTimerQuery(Handle<HwTimerQuery> tqh) {
 void WebGPUDriver::resetState(int) {
 }
 
-void WebGPUDriver::updateDescriptorSetBuffer(
-        Handle<HwDescriptorSet> dsh,
-        backend::descriptor_binding_t binding,
-        Handle<HwBufferObject> boh,
-        uint32_t offset,
+void WebGPUDriver::updateDescriptorSetBuffer(Handle<HwDescriptorSet> dsh,
+        backend::descriptor_binding_t binding, Handle<HwBufferObject> boh, uint32_t offset,
         uint32_t size) {
+    auto bindGroup = handleCast<WebGPUDescriptorSet>(dsh);
+    auto buffer = handleCast<WGPUBufferObject>(boh);
+    if (!bindGroup->getIsLocked()) {
+        // TODO making assumptions that size and offset mean the same thing here.
+        wgpu::BindGroupEntry entry{ .binding = static_cast<uint32_t>(binding * 2),
+            .buffer = buffer->buffer,
+            .offset = offset,
+            .size = size };
+        bindGroup->addEntry(entry.binding, std::move(entry));
+    }
 }
 
-void WebGPUDriver::updateDescriptorSetTexture(
-        Handle<HwDescriptorSet> dsh,
-        backend::descriptor_binding_t binding,
-        Handle<HwTexture> th,
-        SamplerParams params) {
+void WebGPUDriver::updateDescriptorSetTexture(Handle<HwDescriptorSet> dsh,
+        backend::descriptor_binding_t binding, Handle<HwTexture> th, SamplerParams params) {
+    /*
+    auto bindGroup = handleCast<WebGPUDescriptorSet>(dsh);
+    auto texture = handleCast<WGPUTexture>(th);
+
+    // TODO very high odds badd assumptions are in here about handling HwTexture. Revisit with more
+    // understanding. Right now assuming there is a wgpu::TextureView filled in
+    if (!bindGroup->getIsLocked()) {
+        // TODO making assumptions that size and offset mean the same thing here.
+        wgpu::BindGroupEntry tEntry{ .binding = static_cast<uint32_t>(binding * 2),
+            .textureView = texture->texView };
+        bindGroup->addEntry(tEntry.binding, std::move(tEntry));
+
+        wgpu::BindGroupEntry sEntry{ .binding = static_cast<uint32_t>(binding * 2 + 1),
+            .sampler = texture->sampler };
+        bindGroup->addEntry(sEntry.binding, std::move(sEntry));
+    }
+    //TODO Just the setup, this function stilll needs the rest of logic implemented
+     */
 }
 
-void WebGPUDriver::bindDescriptorSet(
-        Handle<HwDescriptorSet> dsh,
-        backend::descriptor_set_t set,
+void WebGPUDriver::bindDescriptorSet(Handle<HwDescriptorSet> dsh, backend::descriptor_set_t set,
         backend::DescriptorSetOffsetArray&& offsets) {
+    auto bindGroup = handleCast<WebGPUDescriptorSet>(dsh);
+    // TODO: presume we need this, use it. Probably Encoder::SetBindGroup
+    auto wbg = bindGroup->lockAndReturn(mDevice);
 }
 
 void WebGPUDriver::setDebugTag(HandleBase::HandleId handleId, utils::CString tag) {

filament/backend/src/webgpu/WebGPUDriver.h

@@ -18,6 +18,7 @@
 #define TNT_FILAMENT_BACKEND_WEBGPUDRIVER_H
 
 #include "WebGPUHandles.h"
+#include "webgpu/WebGPUConstants.h"
 #include <backend/platforms/WebGPUPlatform.h>
 
 #include "DriverBase.h"
@@ -34,7 +35,7 @@
 #include <memory>
 
 #ifndef FILAMENT_WEBGPU_HANDLE_ARENA_SIZE_IN_MB
-#    define FILAMENT_WEBGPU_HANDLE_ARENA_SIZE_IN_MB 8
+#define FILAMENT_WEBGPU_HANDLE_ARENA_SIZE_IN_MB 8
 #endif
 
 namespace filament::backend {
@@ -56,6 +57,27 @@ private:
     [[nodiscard]] ShaderModel getShaderModel() const noexcept final;
     [[nodiscard]] ShaderLanguage getShaderLanguage() const noexcept final;
 
+    template<typename GPUBufferObject>
+    void updateGPUBuffer(GPUBufferObject* gpuBufferObject, BufferDescriptor&& bufferDescriptor,
+            uint32_t byteOffset) {
+        FILAMENT_CHECK_PRECONDITION(bufferDescriptor.buffer)
+                << "copyIntoBuffer called with a null buffer";
+        FILAMENT_CHECK_PRECONDITION(
+                bufferDescriptor.size + byteOffset <= gpuBufferObject->buffer.GetSize())
+                << "Attempting to copy " << bufferDescriptor.size << " bytes into a buffer of size "
+                << gpuBufferObject->buffer.GetSize() << " at offset " << byteOffset;
+
+        // TODO: All buffer objects are created with CopyDst usage.
+        // This may have some performance implications. That should be investigated later.
+        assert_invariant(gpuBufferObject->buffer.GetUsage() & wgpu::BufferUsage::CopyDst);
+
+        // WriteBuffer is an async call. But cpu buffer data is already written to the staging
+        // buffer on return from the WriteBuffer.
+        mQueue.WriteBuffer(gpuBufferObject->buffer, byteOffset, bufferDescriptor.buffer,
+                bufferDescriptor.size);
+        scheduleDestroy(std::move(bufferDescriptor));
+    }
+
     // the platform (e.g. OS) specific aspects of the WebGPU backend are strictly only
     // handled in the WebGPUPlatform
     WebGPUPlatform& mPlatform;
@@ -99,8 +121,8 @@ private:
         return mHandleAllocator.allocate<D>();
     }
 
-    template<typename D, typename B, typename ... ARGS>
-    D* constructHandle(Handle<B>& handle, ARGS&& ... args) noexcept {
+    template<typename D, typename B, typename... ARGS>
+    D* constructHandle(Handle<B>& handle, ARGS&&... args) noexcept {
         return mHandleAllocator.construct<D>(handle, std::forward<ARGS>(args)...);
     }
 
@@ -112,7 +134,7 @@ private:
     template<typename D, typename B>
     void destructHandle(Handle<B>& handle) noexcept {
         auto* p = mHandleAllocator.handle_cast<D*>(handle);
-        return mHandleAllocator.deallocate(handle, p);
+        mHandleAllocator.deallocate(handle, p);
     }
 };
 

filament/backend/src/webgpu/WebGPUHandles.cpp

@@ -1,5 +1,5 @@
 /*
-* Copyright (C) 2025 The Android Open Source Project
+ * 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.
@@ -19,46 +19,159 @@
 #include <utility>
 
 namespace {
-
-wgpu::Buffer createIndexBuffer(wgpu::Device const& device, uint8_t elementSize, uint32_t indexCount) {
-    wgpu::BufferDescriptor descriptor{ .label = "index_buffer",
-        .usage = wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Index,
-        .size = elementSize * indexCount,
-        .mappedAtCreation = false };
-    return device.CreateBuffer(&descriptor);
-}
-} // namespace
-
-namespace filament::backend {
-
-
-WGPUIndexBuffer::WGPUIndexBuffer(wgpu::Device const& device, uint8_t elementSize,
-        uint32_t indexCount)
-    : buffer(createIndexBuffer(device, elementSize, indexCount)) {}
-
-
-WGPUVertexBuffer::WGPUVertexBuffer(wgpu::Device const &device, uint32_t vextexCount, uint32_t bufferCount,
-                                   Handle<WGPUVertexBufferInfo> vbih)
-        : HwVertexBuffer(vextexCount),
-          vbih(vbih),
-          buffers(bufferCount) {
-    wgpu::BufferDescriptor descriptor {
-            .usage = wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Vertex,
-            .size = vextexCount * bufferCount,
-            .mappedAtCreation = false };
-
-    for (uint32_t i = 0; i < bufferCount; ++i) {
-        descriptor.label = ("vertex_buffer_" + std::to_string(i)).c_str();
-        buffers[i] = device.CreateBuffer(&descriptor);
+constexpr wgpu::BufferUsage getBufferObjectUsage(
+        filament::backend::BufferObjectBinding bindingType) noexcept {
+    switch (bindingType) {
+        case filament::backend::BufferObjectBinding::VERTEX:
+            return wgpu::BufferUsage::Vertex;
+        case filament::backend::BufferObjectBinding::UNIFORM:
+            return wgpu::BufferUsage::Uniform;
+        case filament::backend::BufferObjectBinding::SHADER_STORAGE:
+            return wgpu::BufferUsage::Storage;
     }
 }
 
-// TODO: Empty function is a place holder for verxtex buffer updates and should be
-// updated for that purpose.
-void WGPUVertexBuffer::setBuffer(WGPUBufferObject* bufferObject, uint32_t index) {}
+wgpu::Buffer createBuffer(wgpu::Device const& device, wgpu::BufferUsage usage, uint32_t size,
+        char const* label) {
+    wgpu::BufferDescriptor descriptor{ .label = label,
+        .usage = usage,
+        .size = size,
+        .mappedAtCreation = false };
+    return device.CreateBuffer(&descriptor);
+}
 
-WGPUBufferObject::WGPUBufferObject(BufferObjectBinding bindingType, uint32_t byteCount)
+wgpu::VertexFormat getVertexFormat(filament::backend::ElementType type, bool normalized, bool integer) {
+    using ElementType = filament::backend::ElementType;
+    using VertexFormat = wgpu::VertexFormat;
+    if (normalized) {
+        switch (type) {
+            // Single Component Types
+            case ElementType::BYTE: return VertexFormat::Snorm8;
+            case ElementType::UBYTE: return VertexFormat::Unorm8;
+            case ElementType::SHORT: return VertexFormat::Snorm16;
+            case ElementType::USHORT: return VertexFormat::Unorm16;
+            // Two Component Types
+            case ElementType::BYTE2: return VertexFormat::Snorm8x2;
+            case ElementType::UBYTE2: return VertexFormat::Unorm8x2;
+            case ElementType::SHORT2: return VertexFormat::Snorm16x2;
+            case ElementType::USHORT2: return VertexFormat::Unorm16x2;
+            // Three Component Types
+            // There is no vertex format type for 3 byte data in webgpu. Use
+            // 4 byte signed normalized type and ignore the last byte.
+            // TODO: This is to be verified.
+            case ElementType::BYTE3: return VertexFormat::Snorm8x4;    // NOT MINSPEC
+            case ElementType::UBYTE3: return VertexFormat::Unorm8x4;   // NOT MINSPEC
+            case ElementType::SHORT3: return VertexFormat::Snorm16x4;  // NOT MINSPEC
+            case ElementType::USHORT3: return VertexFormat::Unorm16x4; // NOT MINSPEC
+            // Four Component Types
+            case ElementType::BYTE4: return VertexFormat::Snorm8x4;
+            case ElementType::UBYTE4: return VertexFormat::Unorm8x4;
+            case ElementType::SHORT4: return VertexFormat::Snorm16x4;
+            case ElementType::USHORT4: return VertexFormat::Unorm8x4;
+            default:
+                FILAMENT_CHECK_POSTCONDITION(false) << "Normalized format does not exist.";
+                return VertexFormat::Float32x3;
+        }
+    }
+    switch (type) {
+        // Single Component Types
+        // There is no direct alternative for SSCALED in webgpu. Convert them to Float32 directly.
+        // This will result in increased memory on the cpu side.
+        // TODO: Is Float16 acceptable instead with some potential accuracy errors?
+        case ElementType::BYTE: return integer ? VertexFormat::Sint8 : VertexFormat::Float32;
+        case ElementType::UBYTE: return integer ? VertexFormat::Uint8 : VertexFormat::Float32;
+        case ElementType::SHORT: return integer ? VertexFormat::Sint16 : VertexFormat::Float32;
+        case ElementType::USHORT: return integer ? VertexFormat::Uint16 : VertexFormat::Float32;
+        case ElementType::HALF: return VertexFormat::Float16;
+        case ElementType::INT: return VertexFormat::Sint32;
+        case ElementType::UINT: return VertexFormat::Uint32;
+        case ElementType::FLOAT: return VertexFormat::Float32;
+        // Two Component Types
+        case ElementType::BYTE2: return integer ? VertexFormat::Sint8x2 : VertexFormat::Float32x2;
+        case ElementType::UBYTE2: return integer ? VertexFormat::Uint8x2 : VertexFormat::Float32x2;
+        case ElementType::SHORT2: return integer ? VertexFormat::Sint16x2 : VertexFormat::Float32x2;
+        case ElementType::USHORT2: return integer ? VertexFormat::Uint16x2 : VertexFormat::Float32x2;
+        case ElementType::HALF2: return VertexFormat::Float16x2;
+        case ElementType::FLOAT2: return VertexFormat::Float32x2;
+        // Three Component Types
+        case ElementType::BYTE3: return VertexFormat::Sint8x4;    // NOT MINSPEC
+        case ElementType::UBYTE3: return VertexFormat::Uint8x4;   // NOT MINSPEC
+        case ElementType::SHORT3: return VertexFormat::Sint16x4;  // NOT MINSPEC
+        case ElementType::USHORT3: return VertexFormat::Uint16x4; // NOT MINSPEC
+        case ElementType::HALF3: return VertexFormat::Float16x4;  // NOT MINSPEC
+        case ElementType::FLOAT3: return VertexFormat::Float32x3;
+        // Four Component Types
+        case ElementType::BYTE4: return integer ? VertexFormat::Sint8x4 : VertexFormat::Float32x4;
+        case ElementType::UBYTE4: return integer ? VertexFormat::Uint8x4 : VertexFormat::Float32x4;
+        case ElementType::SHORT4: return integer ? VertexFormat::Sint16x4 : VertexFormat::Float32x4;
+        case ElementType::USHORT4: return integer ? VertexFormat::Uint16x4 : VertexFormat::Float32x4;
+        case ElementType::HALF4: return VertexFormat::Float16x4;
+        case ElementType::FLOAT4: return VertexFormat::Float32x4;
+    }
+}
+
+}// namespace
+
+namespace filament::backend {
+
+WGPUVertexBufferInfo::WGPUVertexBufferInfo(uint8_t bufferCount, uint8_t attributeCount,
+        AttributeArray const& attributes)
+    : HwVertexBufferInfo(bufferCount, attributeCount),
+      mVertexBufferLayout(bufferCount),
+      mAttributes(bufferCount) {
+    assert_invariant(attributeCount > 0);
+    assert_invariant(bufferCount > 0);
+    for (uint32_t attribIndex = 0; attribIndex < attributes.size(); attribIndex++) {
+        Attribute const& attrib = attributes[attribIndex];
+        // Ignore the attributes which are not bind to vertex buffers.
+        if (attrib.buffer == Attribute::BUFFER_UNUSED) {
+            continue;
+        }
+
+        assert_invariant(attrib.buffer < bufferCount);
+        bool const isInteger = attrib.flags & Attribute::FLAG_INTEGER_TARGET;
+        bool const isNormalized = attrib.flags & Attribute::FLAG_NORMALIZED;
+        wgpu::VertexFormat vertexFormat = getVertexFormat(attrib.type, isNormalized, isInteger);
+
+        // Attributes are sequential per buffer
+        mAttributes[attrib.buffer].push_back({
+            .format = vertexFormat,
+            .offset = attrib.offset,
+            .shaderLocation = static_cast<uint32_t>(mAttributes[attrib.buffer].size()),
+        });
+
+        mVertexBufferLayout[attrib.buffer].stepMode = wgpu::VertexStepMode::Vertex;
+        if (mVertexBufferLayout[attrib.buffer].arrayStride == 0) {
+            mVertexBufferLayout[attrib.buffer].arrayStride = attrib.stride;
+        } else {
+            assert_invariant(mVertexBufferLayout[attrib.buffer].arrayStride == attrib.stride);
+        }
+    }
+
+    for (uint32_t bufferIndex = 0; bufferIndex < bufferCount; bufferIndex++) {
+        mVertexBufferLayout[bufferIndex].attributeCount = mAttributes[bufferIndex].size();
+        mVertexBufferLayout[bufferIndex].attributes = mAttributes[bufferIndex].data();
+    }
+}
+
+WGPUIndexBuffer::WGPUIndexBuffer(wgpu::Device const& device, uint8_t elementSize,
+        uint32_t indexCount)
+    : buffer(createBuffer(device, wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Index,
+              elementSize * indexCount, "index_buffer")),
+      indexFormat(elementSize == 2 ? wgpu::IndexFormat::Uint16 : wgpu::IndexFormat::Uint32) {}
+
+
+WGPUVertexBuffer::WGPUVertexBuffer(wgpu::Device const& device, uint32_t vertexCount,
+        uint32_t bufferCount, Handle<HwVertexBufferInfo> vbih)
+    : HwVertexBuffer(vertexCount),
+      vbih(vbih),
+      buffers(bufferCount) {}
+
+WGPUBufferObject::WGPUBufferObject(wgpu::Device const& device, BufferObjectBinding bindingType,
+        uint32_t byteCount)
     : HwBufferObject(byteCount),
+      buffer(createBuffer(device, wgpu::BufferUsage::CopyDst | getBufferObjectUsage(bindingType),
+              byteCount, "buffer_object")),
       bufferObjectBinding(bindingType) {}
 
 wgpu::ShaderStage WebGPUDescriptorSetLayout::filamentStageToWGPUStage(ShaderStageFlags fFlags) {
@@ -146,7 +259,39 @@ WebGPUDescriptorSetLayout::WebGPUDescriptorSetLayout(DescriptorSetLayout const&
         .entries = wEntries.data()
     };
     // TODO Do we need to defer this until we have more info on textures and samplers??
+    mLayoutSize = wEntries.size();
     mLayout = device.CreateBindGroupLayout(&layoutDescriptor);
 }
+
 WebGPUDescriptorSetLayout::~WebGPUDescriptorSetLayout() {}
+
+WebGPUDescriptorSet::WebGPUDescriptorSet(const wgpu::BindGroupLayout& layout, uint layoutSize)
+    : mLayout(layout),
+      entries(layoutSize, wgpu::BindGroupEntry{}) {
+    // Establish the size of entries based on the layout. This should be reliable and efficient.
+}
+WebGPUDescriptorSet::~WebGPUDescriptorSet() {}
+
+wgpu::BindGroup WebGPUDescriptorSet::lockAndReturn(const wgpu::Device& device) {
+    if (mBindGroup) {
+        return mBindGroup;
+    }
+    // TODO label? Should we just copy layout label?
+    wgpu::BindGroupDescriptor desc{ .layout = mLayout,
+        .entryCount = entries.size(),
+        .entries = entries.data() };
+    mBindGroup = device.CreateBindGroup(&desc);
+    return mBindGroup;
+}
+
+void WebGPUDescriptorSet::addEntry(uint index, wgpu::BindGroupEntry&& entry) {
+    if (mBindGroup) {
+        // We will keep getting hits from future updates, but shouldn't do anything
+        // Filament guarantees this won't change after things have locked.
+        return;
+    }
+    // TODO: Putting some level of trust that Filament is not going to reuse indexes or go past the
+    // layout index for efficiency. Add guards if wrong.
+    entries[index] = std::move(entry);
+}
 }// namespace filament::backend

filament/backend/src/webgpu/WebGPUHandles.h

@@ -28,27 +28,59 @@
 #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.
-struct WGPUVertexBufferInfo : public HwVertexBufferInfo {
+
+// VertexBufferInfo contains layout info for Vertex Buffer based on WebGPU structs. In WebGPU each
+// VertexBufferLayout is associated with a single vertex buffer. So number of mVertexBufferLayout
+// is equal to bufferCount. Each VertexBufferLayout can contain multiple VertexAttribute. Bind index
+// of vertex buffer is implicitly calculated by the position of VertexBufferLayout in an array.
+class WGPUVertexBufferInfo : public HwVertexBufferInfo {
+public:
     WGPUVertexBufferInfo(uint8_t bufferCount, uint8_t attributeCount,
-            AttributeArray const& attributes)
-        : HwVertexBufferInfo(bufferCount, attributeCount),
-          attributes(attributes) {}
-    AttributeArray attributes;
+            AttributeArray const& attributes);
+    inline  wgpu::VertexBufferLayout const* getVertexBufferLayout() const {
+        return mVertexBufferLayout.data();
+    }
+
+    inline uint32_t getVertexBufferLayoutSize() const {
+        return mVertexBufferLayout.size();
+    }
+
+    inline wgpu::VertexAttribute const* getVertexAttributeForIndex(uint32_t index) const {
+        assert_invariant(index < mAttributes.size());
+        return mAttributes[index].data();
+    }
+
+    inline uint32_t getVertexAttributeSize(uint32_t index) const {
+        assert_invariant(index < mAttributes.size());
+        return mAttributes[index].size();
+    }
+
+private:
+    // TODO: can we do better in terms on heap management.
+    std::vector<wgpu::VertexBufferLayout> mVertexBufferLayout {};
+    std::vector<std::vector<wgpu::VertexAttribute>> mAttributes {};
 };
 
 struct WGPUVertexBuffer : public HwVertexBuffer {
-    WGPUVertexBuffer(wgpu::Device const &device, uint32_t vextexCount, uint32_t bufferCount,
-                     Handle<WGPUVertexBufferInfo> vbih);
+    WGPUVertexBuffer(wgpu::Device const &device, uint32_t vertexCount, uint32_t bufferCount,
+                     Handle<HwVertexBufferInfo> vbih);
 
-    void setBuffer(WGPUBufferObject *bufferObject, uint32_t index);
-
-    Handle<WGPUVertexBufferInfo> vbih;
+    Handle<HwVertexBufferInfo> vbih;
     utils::FixedCapacityVector<wgpu::Buffer> buffers;
 };
 
@@ -57,29 +89,49 @@ struct WGPUIndexBuffer : public HwIndexBuffer {
                     uint32_t indexCount);
 
     wgpu::Buffer buffer;
+    wgpu::IndexFormat indexFormat;
 };
 
-// TODO: Currently WGPUVertexBufferInfo is not used by WebGPU for useful task.
-// Update the struct when used by WebGPU driver.
 struct WGPUBufferObject : HwBufferObject {
-    WGPUBufferObject(BufferObjectBinding bindingType, uint32_t byteCount);
+    WGPUBufferObject(wgpu::Device const &device, BufferObjectBinding bindingType, uint32_t byteCount);
 
-    wgpu::Buffer buffer;
+    wgpu::Buffer buffer = nullptr;
     const BufferObjectBinding bufferObjectBinding;
 };
-class WebGPUDescriptorSetLayout : public HwDescriptorSetLayout {
+class WebGPUDescriptorSetLayout final : public HwDescriptorSetLayout {
 public:
     WebGPUDescriptorSetLayout(DescriptorSetLayout const& layout, wgpu::Device const& device);
     ~WebGPUDescriptorSetLayout();
+    [[nodiscard]] const wgpu::BindGroupLayout& getLayout() const { return mLayout; }
+    [[nodiscard]] uint getLayoutSize() const { return mLayoutSize; }
 
 private:
     // TODO: If this is useful elsewhere, remove it from this class
     // Convert Filament Shader Stage Flags bitmask to webgpu equivilant
     static wgpu::ShaderStage filamentStageToWGPUStage(ShaderStageFlags fFlags);
-
+    uint mLayoutSize;
     wgpu::BindGroupLayout mLayout;
 };
 
+class WebGPUDescriptorSet final : public HwDescriptorSet {
+public:
+    WebGPUDescriptorSet(const wgpu::BindGroupLayout& layout, uint layoutSize);
+    ~WebGPUDescriptorSet();
+
+    wgpu::BindGroup lockAndReturn(wgpu::Device const& device);
+    void addEntry(uint index, wgpu::BindGroupEntry&& entry);
+    [[nodiscard]] bool getIsLocked() const { return mBindGroup != nullptr; }
+
+private:
+    // TODO: Consider storing what we used to make the layout. However we need to essentially
+    // Recreate some of the info (Sampler in slot X with the actual sampler) so letting Dawn confirm
+    // there isn't a mismatch may be easiest.
+    // Also storing the wgpu ObjectBase takes care of ownership challenges in theory
+    wgpu::BindGroupLayout mLayout;
+    std::vector<wgpu::BindGroupEntry> entries;
+    wgpu::BindGroup mBindGroup;
+};
+
 // TODO: Currently WGPUTexture is not used by WebGPU for useful task.
 // Update the struct when used by WebGPU driver.
 struct WGPUTexture : public HwTexture {
@@ -90,10 +142,15 @@ struct WGPUTexture : public HwTexture {
     WGPUTexture(WGPUTexture const* src, uint8_t baseLevel, uint8_t levelCount) noexcept;
 
     wgpu::Texture texture = nullptr;
+    // TODO: Adding this but not yet setting it up. Filament "Textures" are combined image samplers,
+    // rep both.
+    wgpu::Sampler sampler = nullptr;
+    //TODO: Not sure all the ways HwTexture is used. Overloading like this might be entirely wrong.
+    wgpu::TextureView texView = nullptr;
 };
 
 struct WGPURenderPrimitive : public HwRenderPrimitive {
-    WGPURenderPrimitive();
+    WGPURenderPrimitive() {}
 
     void setBuffers(WGPUVertexBufferInfo const* const vbi,
             WGPUVertexBuffer* vertexBuffer, WGPUIndexBuffer* indexBuffer);

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();
 }