This is a test for pushing branch to github

Full commit hash is b6df2b9b35

DOCS_ALLOW_DIRECT_EDITS

.github/actions/get-commit-msg/action.yml

@@ -1,43 +1,53 @@
+# This action retrieves the latest commit message from a push or pull_request event
+# and makes it available as an output variable named 'msg'.
 name: 'Get commit message'
 outputs:
   msg:
     value: ${{ steps.action_output.outputs.msg }}
+  hash:
+    value: ${{ steps.action_output.outputs.hash }}
 runs:
   using: "composite"
   steps:
     - name: Find commit message (on push)
       if: github.event_name == 'push'
       shell: bash
+      env:
+        COMMIT_MESSAGE: ${{ github.event.head_commit.message }}
       run: |
         AUTHOR_NAME="${{ github.event.head_commit.author.name }}"
         AUTHOR_EMAIL="${{ github.event.head_commit.author.email }}"
         TSTAMP="${{ github.event.head_commit.timestamp }}"
-        echo "commit ${{ github.event.head_commit.id }}" >> /tmp/commit_msg.txt
+        HASH="${{ github.event.head_commit.id }}"
+        echo "commit $HASH" >> /tmp/commit_msg.txt
         echo "Author: ${AUTHOR_NAME}<${AUTHOR_EMAIL}>" >> /tmp/commit_msg.txt
         echo "Date: ${TSTAMP}" >> /tmp/commit_msg.txt
         echo "" >> /tmp/commit_msg.txt
-        echo "${{ github.event.head_commit.message }}" >> /tmp/commit_msg.txt
+        echo "$COMMIT_MESSAGE" >> /tmp/commit_msg.txt
+        echo "$HASH" > /tmp/commit_hash.txt
     - name: Find commit message (PR)
       shell: bash
       id: checkout_code
       if: github.event_name == 'pull_request'
       run: |
-        echo "+++++ head commit message +++++"
-        echo "$(git log -1 --no-merges)"
-        echo "+++++++++++++++++++++++++++++++"
-        echo "hash=$(git rev-parse HEAD)" >> "$GITHUB_OUTPUT"
-        git checkout ${{ github.event.pull_request.head.sha }}
-        echo "$(git log -1 --no-merges)" >> /tmp/commit_msg.txt
+        BEFORE_HASH=$(git rev-parse HEAD)
+        echo "hash=$BEFORE_HASH" >> "$GITHUB_OUTPUT"
+        # Next we will checkout the actual head (not the merge commits) of the PR
+        AFTER_HASH="${{ github.event.pull_request.head.sha }}"
+        git checkout $AFTER_HASH
+        COMMIT_MESSAGE=$(git log -1 --no-merges)
+        echo "$COMMIT_MESSAGE" > /tmp/commit_msg.txt
+        echo "$AFTER_HASH" > /tmp/commit_hash.txt
     - shell: bash
       id: action_output
       run: |
+        # Get the commit message
         DELIMITER="EOF_FILE_CONTENT_$(date +%s)" # Using timestamp to make it more unique
         echo "msg<<$DELIMITER" >> "$GITHUB_OUTPUT"
         cat /tmp/commit_msg.txt >> "$GITHUB_OUTPUT"
         echo "$DELIMITER" >> "$GITHUB_OUTPUT"
-        echo "----- got commit message ---"
-        cat /tmp/commit_msg.txt
-        echo "----------------------------"
+        # Get the commit hash
+        echo "hash=$(cat /tmp/commit_hash.txt)" >> "$GITHUB_OUTPUT"
     - name: Cleanup Find commit message (PR)
       shell: bash
       if: github.event_name == 'pull_request'

.github/workflows/postsubmit.yml

@@ -23,7 +23,7 @@ jobs:
           GH_TOKEN: ${{ secrets.FILAMENTBOT_TOKEN }}
         run: |
           GOLDEN_BRANCH=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | python3 test/renderdiff/src/commit_msg.py)
-          COMMIT_HASH=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | head -n 1 | sed "s/commit //g")
+          COMMIT_HASH="${{ steps.get_commit_msg.outputs.hash }}"
           if [[ "${GOLDEN_BRANCH}" != "main" ]]; then
               git config --global user.email "filament.bot@gmail.com"
               git config --global user.name "Filament Bot"
@@ -51,7 +51,7 @@ jobs:
           GH_TOKEN: ${{ secrets.FILAMENTBOT_TOKEN }}
         run: |
           bash docs_src/build/install_mdbook.sh && source ~/.bashrc
-          COMMIT_HASH=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | head -n 1 | sed "s/commit //g")
+          COMMIT_HASH="${{ steps.get_commit_msg.outputs.hash }}"
           git config --global user.email "filament.bot@gmail.com"
           git config --global user.name "Filament Bot"
           git config --global credential.helper cache

.github/workflows/presubmit.yml

@@ -108,8 +108,7 @@ jobs:
         uses: ./.github/actions/get-commit-msg
       - name: Check for manual edits to /docs
         run: |
-          COMMIT_ID=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | head -n 1 | sed "s/commit //g")
-          bash docs_src/build/presubmit_check.sh ${COMMIT_ID}
+          bash docs_src/build/presubmit_check.sh ${{ steps.get_commit_msg.outputs.hash }}
 
   test-renderdiff:
     name: test-renderdiff
@@ -121,7 +120,7 @@ jobs:
       - uses: ./.github/actions/mac-prereq
       - uses: ./.github/actions/get-gltf-assets
       - uses: ./.github/actions/get-mesa
-      - uses: ./.github/actions/get-vulkan-sdk      
+      - uses: ./.github/actions/get-vulkan-sdk
       - id: get_commit_msg
         uses: ./.github/actions/get-commit-msg
       - name: Prerequisites
@@ -132,12 +131,14 @@ jobs:
         shell: bash
       - name: Render and compare
         id: render_compare
+        env:
+          COMMIT_MESSAGE: ${{ steps.get_commit_msg.outputs.msg }}
         run: |
           ls ./gltf/Models
           TEST_DIR=test/renderdiff
           source ${TEST_DIR}/src/preamble.sh
           start_
-          GOLDEN_BRANCH=$(echo "${{ steps.get_commit_msg.outputs.msg }}" | python3 ${TEST_DIR}/src/commit_msg.py)
+          GOLDEN_BRANCH=$(echo "${COMMIT_MESSAGE}" | python3 ${TEST_DIR}/src/commit_msg.py)
           bash ${TEST_DIR}/generate.sh && \
               python3 ${TEST_DIR}/src/golden_manager.py \
                   --branch=${GOLDEN_BRANCH} \

docs/dup/intro.html

@@ -181,7 +181,7 @@ important for <code>matc</code> (material compiler).</p>
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.65.4'
+    implementation 'com.google.android.filament:filament-android:1.66.0'
 }
 </code></pre>
 <p>Here are all the libraries available in the group <code>com.google.android.filament</code>:</p>
@@ -196,7 +196,7 @@ dependencies {
 </div>
 <h3 id="ios"><a class="header" href="#ios">iOS</a></h3>
 <p>iOS projects can use CocoaPods to install the latest release:</p>
-<pre><code class="language-shell">pod 'Filament', '~&gt; 1.65.4'
+<pre><code class="language-shell">pod 'Filament', '~&gt; 1.66.0'
 </code></pre>
 <h2 id="documentation"><a class="header" href="#documentation">Documentation</a></h2>
 <ul>

docs/dup/renderdiff.html

@@ -159,16 +159,16 @@
                 <div id="content" class="content">
                     <main>
                         <h1 id="rendering-difference-test"><a class="header" href="#rendering-difference-test">Rendering Difference Test</a></h1>
-<p>We created a few scripts to run <code>gltf_viewer</code> and produce headless renderings.</p>
+<p>This tool is a collections of scripts to run <code>gltf_viewer</code> and produce headless renderings.</p>
 <p>This is mainly useful for continuous integration where GPUs are generally not available on cloud
 machines. To perform software rasterization, these scripts are centered around <a href="https://docs.mesa3d.org">Mesa</a>'s
 software rasterizers, but nothing bars us from using another rasterizer like <a href="https://github.com/google/swiftshader">SwiftShader</a>.
 Additionally, we should be able to use GPUs where available (though this is more of a future
 work).</p>
 <p>The script <code>render.py</code> contains the core logic for taking input parameters (such as the test
-description file) and then running gltf_viewer to produce the renderings.</p>
+description file) and then running <code>gltf_viewer</code> to produce the renderings.</p>
 <p>In the <code>test</code> directory is a list of test descriptions that are specified in json. Please see
-<code>sample.json</code> to parse the structure.</p>
+<code>sample.json</code> to glean the structure.</p>
 <h2 id="setting-up-python"><a class="header" href="#setting-up-python">Setting up python</a></h2>
 <p>The <code>renderdiff</code> project uses <code>python</code> extensively. To install the dependencies for producing
 renderings, do the following step</p>

docs_src/build/checks.py

@@ -77,11 +77,11 @@ def check_has_source_edits(commit_hash, printing=True):
 
 # Returns true in a given TAG is found in the commit msg
 def commit_msg_has_tag(commit_hash, tag, printing=True):
-  res, ret = execute(f'git log --pretty=%B {commit_hash}', cwd=ROOT_DIR)
+  res, ret = execute(f'git log -n1 --pretty=%B {commit_hash}', cwd=ROOT_DIR)
   for l in ret.split('\n'):
     if tag == l.strip():
       if printing:
-        print(f'Found tag={tag} in commit message')
+        print(f'Found tag={tag} in commit={commit_hash}')
       return True
   return False
 

filament/CMakeLists.txt

@@ -117,6 +117,7 @@ set(SRCS
         src/components/LightManager.cpp
         src/components/RenderableManager.cpp
         src/components/TransformManager.cpp
+        src/details/BufferAllocator.cpp
         src/details/BufferObject.cpp
         src/details/Camera.cpp
         src/details/ColorGrading.cpp
@@ -196,6 +197,7 @@ set(PRIVATE_HDRS
         src/components/LightManager.h
         src/components/RenderableManager.h
         src/components/TransformManager.h
+        src/details/BufferAllocator.h
         src/details/BufferObject.h
         src/details/Camera.h
         src/details/ColorGrading.h

filament/backend/src/opengl/OpenGLContext.cpp

@@ -760,6 +760,7 @@ void OpenGLContext::initExtensionsGLES(Extensions* ext, GLint major, GLint minor
     ext->EXT_texture_compression_rgtc = exts.has("GL_EXT_texture_compression_rgtc"sv);
     ext->EXT_texture_compression_bptc = exts.has("GL_EXT_texture_compression_bptc"sv);
     ext->EXT_texture_cube_map_array = exts.has("GL_EXT_texture_cube_map_array"sv) || exts.has("GL_OES_texture_cube_map_array"sv);
+    ext->EXT_texture_filter_anisotropic = exts.has("GL_EXT_texture_filter_anisotropic"sv);
     ext->GOOGLE_cpp_style_line_directive = exts.has("GL_GOOGLE_cpp_style_line_directive"sv);
     ext->KHR_debug = exts.has("GL_KHR_debug"sv);
     ext->KHR_parallel_shader_compile = exts.has("GL_KHR_parallel_shader_compile"sv);

filament/backend/src/webgpu/WebGPUBlitter.cpp

@@ -84,128 +84,12 @@ constexpr uint32_t UNIFORM_BINDING_INDEX{ 1 };
 constexpr uint32_t SAMPLER_BINDING_INDEX{ 2 };
 constexpr std::string_view VERTEX_SHADER_ENTRY_POINT{ "vertexShaderMain" };
 constexpr std::string_view FRAGMENT_SHADER_ENTRY_POINT{ "fragmentShaderMain" };
-// note that the placeholders below must start and end with this prefix and suffix and should not
-// otherwise be present in the template:
-constexpr std::string_view PLACEHOLDER_PREFIX{ "{{" };
-constexpr std::string_view PLACEHOLDER_SUFFIX{ "}}" };
-// texture_2d<f32> or texture_multisampled_2d<f32> or texture_3d<f32> or
-// texture_depth_2d or texture_depth_multisampled_2d
-constexpr std::string_view TEXTURE_TYPE_PLACEHOLDER{ "TEXTURE_TYPE" };
-// texture_2d<f32> or texture_depth_2d
-[[maybe_unused]] constexpr std::string_view TEXTURE_2D_TYPE_PLACEHOLDER { "TEXTURE_2D_TYPE" };
-// "" for no sampler, otherwise "@group(0) @binding(2) var sourceSampler: sampler;"
-constexpr std::string_view SAMPLER_DECLARATION_PLACEHOLDER{ "SAMPLER_DECLARATION" };
-constexpr std::string_view FRAGMENT_SHADER_SNIPPET_PLACEHOLDER{ "FRAGMENT_SHADER_SNIPPET" };
-// @location(0) vec4<f32> or @builtin(frag_depth) f32
-constexpr std::string_view FRAGMENT_RETURN_ATTRIBUTE_AND_TYPE_PLACEHOLDER{
-    "FRAGMENT_RETURN_ATTRIBUTE_AND_TYPE"
-};
-constexpr std::string_view SHADER_SOURCE_TEMPLATE{ R"(
-    struct BlitFragmentShaderArgs {
-        depthPlane:        u32,
-        scale:             vec2<f32>,
-        sourceOffset:      vec2<u32>,
-        destinationOffset: vec2<u32>,
-    };
-
-    @group(0) @binding(0) var sourceTexture: {{TEXTURE_TYPE}};
-    @group(0) @binding(1) var<uniform> fragmentShaderArgs: BlitFragmentShaderArgs;
-    {{SAMPLER_DECLARATION}}
-
-    fn getUnnormalizedSourceTextureCoordinates(position: vec2<f32>) -> vec2<f32> {
-        // These coordinates match the Vulkan vkCmdBlitImage spec:
-        // https://www.khronos.org/registry/vulkan/specs/1.2-extensions/man/html/vkCmdBlitImage.html
-        let uvOffset: vec2<f32> = position - vec2<f32>(f32(fragmentShaderArgs.destinationOffset.x),f32(fragmentShaderArgs.destinationOffset.y));
-        let uvScaled: vec2<f32> = uvOffset * fragmentShaderArgs.scale;
-        return uvScaled + vec2<f32>(f32(fragmentShaderArgs.sourceOffset.x), f32(fragmentShaderArgs.sourceOffset.y));
-    }
-
-    fn normalize2dSourceTextureCoordinates(
-            unnormalizedSourceTextureCoordinates: vec2<f32>,
-            sourceDimensions: vec2<u32>) -> vec2<f32> {
-        return unnormalizedSourceTextureCoordinates /
-            vec2<f32>(f32(sourceDimensions.x), f32(sourceDimensions.y));
-    }
-
-    fn getNormalized2dSourceTextureCoordinates(
-            position: vec2<f32>,
-            sourceTexture: {{TEXTURE_2D_TYPE}}) -> vec2<f32> {
-        let sourceDimensions: vec2<u32> = textureDimensions(sourceTexture);
-        return normalize2dSourceTextureCoordinates(
-            getUnnormalizedSourceTextureCoordinates(position),
-            sourceDimensions
-        );
-    }
-
-    fn normalize3dSourceTextureCoordinates(
-            unnormalizedSourceTextureCoordinates: vec2<f32>,
-            sourceDimensions: vec3<u32>) -> vec3<f32> {
-        let uvNormalized: vec2<f32> = normalize2dSourceTextureCoordinates(
-            unnormalizedSourceTextureCoordinates,
-            sourceDimensions.xy
-        );
-        return vec3<f32>(
-            uvNormalized,
-            (f32(fragmentShaderArgs.depthPlane) + 0.5) / f32(sourceDimensions.z)
-        );
-    }
-
-    fn getNormalized3dSourceTextureCoordinates(
-            position: vec2<f32>,
-            sourceTexture: texture_3d<f32>) -> vec3<f32> {
-        let sourceDimensions: vec3<u32> = textureDimensions(sourceTexture);
-        return normalize3dSourceTextureCoordinates(
-            getUnnormalizedSourceTextureCoordinates(position),
-            sourceDimensions
-        );
-    }
-
-    @vertex
-    fn vertexShaderMain(@builtin(vertex_index) vertexIndex: u32) -> @builtin(position) vec4<f32> {
-        let fullScreenTriangleVertices = array<vec2<f32>, 3>(
-            vec2<f32>(-1.0, -1.0),
-            vec2<f32>( 3.0, -1.0),
-            vec2<f32>(-1.0,  3.0)
-        );
-        return vec4<f32>(fullScreenTriangleVertices[vertexIndex].xy, 0.0, 1.0);
-    }
-
-    {{FRAGMENT_SHADER_SNIPPET}}
-)" };
-
-constexpr std::string_view FRAGMENT_SHADER_SNIPPET_MSAA_INPUT_TEMPLATE{ R"(
-    @fragment
-    fn fragmentShaderMain(@builtin(position) position: vec4<f32>) -> {{FRAGMENT_RETURN_ATTRIBUTE_AND_TYPE}} {
-        var color: vec4<f32> = vec4<f32>(0.0, 0.0, 0.0, 0.0);
-        let numberOfSamples: u32 = textureNumSamples(sourceTexture);
-        let coordinatesF: vec2<f32> = getUnnormalizedSourceTextureCoordinates(position.xy);
-        let coordinates: vec2<u32> = vec2<u32>(u32(coordinatesF.x), u32(coordinatesF.y));
-        for (var sampleIndex: u32 = 0; sampleIndex < numberOfSamples; sampleIndex++) {
-            color += textureLoad(sourceTexture, coordinates, sampleIndex);
-        }
-        color /= f32(numberOfSamples);
-        return color;
-    }
-)" };
-
-constexpr std::string_view FRAGMENT_SHADER_SNIPPET_3D_INPUT_TEMPLATE{ R"(
-    @fragment
-    fn fragmentShaderMain(@builtin(position) position: vec4<f32>) -> {{FRAGMENT_RETURN_ATTRIBUTE_AND_TYPE}} {
-        let coordinates: vec3<f32> = getNormalized3dSourceTextureCoordinates(position.xy, sourceTexture);
-        return textureSample(sourceTexture, sourceSampler, coordinates);
-    }
-)" };
-
-constexpr std::string_view FRAGMENT_SHADER_SNIPPET_2D_INPUT_TEMPLATE{ R"(
-    @fragment
-    fn fragmentShaderMain(@builtin(position) position: vec4<f32>) -> {{FRAGMENT_RETURN_ATTRIBUTE_AND_TYPE}} {
-        let coordinates: vec2<f32> = getNormalized2dSourceTextureCoordinates(position.xy, sourceTexture);
-        return textureSample(sourceTexture, sourceSampler, coordinates);
-    }
-)" };
 
 } // namespace
 
+// Include the wgsl template sources
+#include "WebGPUBlitter_wgsl.inc"
+
 // It can perform a direct memory copy if the formats are compatible and no scaling or format
 // conversion is needed. Otherwise, it uses a render pass with a custom shader to perform
 // the blit. This allows for scaling, format conversion, and resolving multisampled textures.
@@ -395,11 +279,25 @@ void WebGPUBlitter::blit(wgpu::Queue const& queue, wgpu::CommandEncoder const& c
     const bool multisampledSource{ args.source.texture.GetSampleCount() > 1 };
     const bool depthSource{ hasDepth(args.source.texture.GetFormat()) };
     const bool depthDestination{ hasDepth(args.destination.texture.GetFormat()) };
+
+    wgpu::TextureSampleType srcSampleType = {};
+    if (depthSource) {
+        srcSampleType = wgpu::TextureSampleType::Depth;
+    } else if (multisampledSource) {
+        srcSampleType = wgpu::TextureSampleType::UnfilterableFloat;
+    } else if (isIntFormat(args.source.texture.GetFormat())) {
+        srcSampleType = wgpu::TextureSampleType::Sint;
+    } else if (isUIntFormat(args.source.texture.GetFormat())) {
+        srcSampleType = wgpu::TextureSampleType::Uint;
+    } else {
+        srcSampleType = wgpu::TextureSampleType::Float;
+    }
+
     const PipelineLayoutKey pipelineLayoutKey{
         .sourceDimension = sourceDimension,
+        .sourceSampleType = srcSampleType,
         .filterType = args.filter,
         .multisampledSource = multisampledSource,
-        .depthSource = depthSource,
     };
     const wgpu::BindGroupDescriptor textureBindGroupDescriptor{
         .label = "blit_texture_bind_group",
@@ -454,10 +352,11 @@ void WebGPUBlitter::blit(wgpu::Queue const& queue, wgpu::CommandEncoder const& c
             << "Failed to create render pass encoder for blit.";
     const RenderPipelineKey renderPipelineKey{
         .sourceDimension = sourceDimension,
+        .sourceTextureFormat = args.source.texture.GetFormat(),
+        .sourceSampleType = srcSampleType,
         .destinationTextureFormat = args.destination.texture.GetFormat(),
         .sourceSampleCount = static_cast<uint8_t>(args.source.texture.GetSampleCount()),
         .filterType = args.filter,
-        .depthSource = depthSource,
     };
     renderPassEncoder.SetPipeline(getOrCreateRenderPipeline(renderPipelineKey));
     renderPassEncoder.SetBindGroup(TEXTURE_BIND_GROUP_INDEX, textureBindGroup);
@@ -557,9 +456,9 @@ wgpu::RenderPipeline WebGPUBlitter::createRenderPipeline(RenderPipelineKey const
     };
     const ShaderModuleKey shaderModuleKey{
         .sourceDimension = key.sourceDimension,
+        .sourceTextureFormat = key.sourceTextureFormat,
+        .destinationTextureFormat = key.destinationTextureFormat,
         .multisampledSource = key.sourceSampleCount > 1,
-        .depthSource = key.depthSource,
-        .depthDestination = hasDepth(key.destinationTextureFormat),
     };
     wgpu::ShaderModule const& shaderModule{ getOrCreateShaderModule(shaderModuleKey) };
     const wgpu::FragmentState fragmentState{
@@ -572,9 +471,9 @@ wgpu::RenderPipeline WebGPUBlitter::createRenderPipeline(RenderPipelineKey const
     };
     const PipelineLayoutKey pipelineLayoutKey{
         .sourceDimension = key.sourceDimension,
+        .sourceSampleType = key.sourceSampleType,
         .filterType = key.filterType,
         .multisampledSource = key.sourceSampleCount > 1,
-        .depthSource = key.depthSource,
     };
     const wgpu::RenderPipelineDescriptor pipelineDescriptor{
         .label = "render_pass_blit_pipeline",
@@ -663,13 +562,7 @@ wgpu::BindGroupLayout WebGPUBlitter::createTextureBindGroupLayout(PipelineLayout
             .binding = TEXTURE_BINDING_INDEX,
             .visibility = wgpu::ShaderStage::Fragment,
             .texture = {
-                .sampleType = key.depthSource
-                                      ? wgpu::TextureSampleType::Depth
-                                      : (key.multisampledSource
-                                                ? wgpu::TextureSampleType::UnfilterableFloat
-                                                : wgpu::TextureSampleType::Float), // only F32 scalar sample
-                                                                                   // type supported for now
-                                                                                   // (aside from depth)
+                .sampleType = key.sourceSampleType,
                 .viewDimension = key.sourceDimension,
                 .multisampled = key.multisampledSource,
             },
@@ -695,7 +588,7 @@ wgpu::BindGroupLayout WebGPUBlitter::createTextureBindGroupLayout(PipelineLayout
     };
     const wgpu::BindGroupLayoutDescriptor textureBindGroupLayoutDescriptor{
         .label = "render_pass_blit_texture_bind_group_layout",
-        // TODO, doesnt make any sense but gets rid of the error. Are the entries 0 based?
+        // No need for the last entry (sampler) if source is multisampled.
         .entryCount = key.multisampledSource ? (MAX_TEXTURE_BIND_GROUP_ENTRY_SIZE - 1)
                                              : (MAX_TEXTURE_BIND_GROUP_ENTRY_SIZE),
         .entries = bindGroupLayoutEntries,
@@ -720,63 +613,109 @@ wgpu::ShaderModule const& WebGPUBlitter::getOrCreateShaderModule(ShaderModuleKey
 // The shader source is generated from a template, with placeholders filled in based on the
 // blit configuration (e.g., texture type, sample count).
 wgpu::ShaderModule WebGPUBlitter::createShaderModule(ShaderModuleKey const& key) {
-    std::string_view textureType;
-    if (key.depthSource) {
-        if (key.multisampledSource) {
-            textureType = "texture_depth_multisampled_2d";
-        } else {
-            textureType = "texture_depth_2d";
-        }
+    bool const isDepthSrc = hasDepth(key.sourceTextureFormat);
+    bool const isDepthDst = hasDepth(key.destinationTextureFormat);
+
+    std::string_view vecDim;
+    if (key.sourceDimension == wgpu::TextureViewDimension::e3D) {
+        vecDim = "vec3";
     } else {
-        if (key.multisampledSource) {
-            textureType = "texture_multisampled_2d<f32>";
-        } else {
-            if (key.sourceDimension == wgpu::TextureViewDimension::e3D) {
-                textureType = "texture_3d<f32>";
-            } else {
-                textureType = "texture_2d<f32>";
-            }
-        }
+        vecDim = "vec2";
     }
-    const std::string_view texture2dType{ key.depthSource ? "texture_depth_2d"
-                                                          : "texture_2d<f32>" };
-    // we don't declare a sampler in the shader or pipeline for the multisampled case
-    const std::string_view samplerDeclaration{
-        key.multisampledSource ? "" : "@group(0) @binding(2) var sourceSampler: sampler;"
-    };
-    const std::string_view fragmentReturnAttributeAndType{
-        key.depthDestination ? "@builtin(frag_depth) f32" : "@location(0) vec4<f32>"
-    };
-    const std::unordered_map<std::string_view, std::string_view>
-            valueByPlaceholderNameForFragmentSnippet{
-                { FRAGMENT_RETURN_ATTRIBUTE_AND_TYPE_PLACEHOLDER, fragmentReturnAttributeAndType },
-            };
-    std::string fragmentShaderSnippet;
-    if (key.multisampledSource) {
-        fragmentShaderSnippet = webgpuutils::processPlaceholderTemplate(
-                FRAGMENT_SHADER_SNIPPET_MSAA_INPUT_TEMPLATE, PLACEHOLDER_PREFIX, PLACEHOLDER_SUFFIX,
-                valueByPlaceholderNameForFragmentSnippet);
+
+    std::string_view srcPrimType;
+    std::string_view sampleImpl;
+    if (isIntFormat(key.sourceTextureFormat)) {
+        srcPrimType = "<i32>";
+        sampleImpl = INT_TEXTURE_SAMPLE_IMPL_TEMPLATE;
+    } else if (isUIntFormat(key.sourceTextureFormat)) {
+        srcPrimType = "<u32>";
+        sampleImpl = INT_TEXTURE_SAMPLE_IMPL_TEMPLATE;
     } else {
-        if (key.sourceDimension == wgpu::TextureViewDimension::e3D) {
-            fragmentShaderSnippet = webgpuutils::processPlaceholderTemplate(
-                    FRAGMENT_SHADER_SNIPPET_3D_INPUT_TEMPLATE, PLACEHOLDER_PREFIX,
-                    PLACEHOLDER_SUFFIX, valueByPlaceholderNameForFragmentSnippet);
-        } else {
-            fragmentShaderSnippet = webgpuutils::processPlaceholderTemplate(
-                    FRAGMENT_SHADER_SNIPPET_2D_INPUT_TEMPLATE, PLACEHOLDER_PREFIX,
-                    PLACEHOLDER_SUFFIX, valueByPlaceholderNameForFragmentSnippet);
-        }
+        srcPrimType = "<f32>";
+        sampleImpl = FLOAT_TEXTURE_SAMPLE_IMPL_TEMPLATE;
     }
-    const std::unordered_map<std::string_view, std::string_view> valueByPlaceholderNameForModule{
-        { TEXTURE_TYPE_PLACEHOLDER, textureType },
-        { TEXTURE_2D_TYPE_PLACEHOLDER, texture2dType },
-        { SAMPLER_DECLARATION_PLACEHOLDER, samplerDeclaration },
-        { FRAGMENT_SHADER_SNIPPET_PLACEHOLDER, fragmentShaderSnippet },
+
+    std::string_view fragmentTemplate;
+    std::string_view srcTextureType;
+    if (isDepthSrc && key.multisampledSource) {
+        srcTextureType = "texture_depth_multisampled_2d";
+        fragmentTemplate = FRAGMENT_SHADER_SNIPPET_2D_INPUT_TEMPLATE;
+    } else if (isDepthSrc && !key.multisampledSource) {
+        srcTextureType = "texture_depth_2d";
+        fragmentTemplate = FRAGMENT_SHADER_SNIPPET_2D_INPUT_TEMPLATE;
+    } else if (key.multisampledSource) {
+        srcTextureType = "texture_multisampled_2d{{SRC_PRIM_TYPE}}";
+        fragmentTemplate = FRAGMENT_SHADER_SNIPPET_MSAA_INPUT_TEMPLATE;
+    } else if (key.sourceDimension == wgpu::TextureViewDimension::e3D) {
+        srcTextureType = "texture_3d{{SRC_PRIM_TYPE}}";
+        fragmentTemplate = FRAGMENT_SHADER_SNIPPET_3D_INPUT_TEMPLATE;
+    } else {
+        srcTextureType = "texture_2d{{SRC_PRIM_TYPE}}";
+        fragmentTemplate = FRAGMENT_SHADER_SNIPPET_2D_INPUT_TEMPLATE;
+    }
+
+    std::string_view dstPrimType;
+    if (isIntFormat(key.destinationTextureFormat)) {
+        dstPrimType = "<i32>";
+    } else if (isUIntFormat(key.destinationTextureFormat)) {
+        dstPrimType = "<u32>";
+    } else {
+        dstPrimType = "<f32>";
+    }
+
+    std::string_view retAttribute;
+    std::string_view retType;
+    if (isDepthDst) {
+        retAttribute = "@builtin(frag_depth)";
+        retType = "f32";
+    } else {
+        retAttribute = "@location(0)";
+        retType = "vec4{{DST_PRIM_TYPE}}";
+    }
+
+    using ReplacementMap = std::unordered_map<std::string_view, std::string_view>;
+    auto const replace = [&](std::string_view src, ReplacementMap const& replacements) {
+        return webgpuutils::processPlaceholderTemplate(
+                src, PLACEHOLDER_PREFIX, PLACEHOLDER_SUFFIX, replacements);
     };
-    const std::string shaderSource{ webgpuutils::processPlaceholderTemplate(SHADER_SOURCE_TEMPLATE,
-            PLACEHOLDER_PREFIX, PLACEHOLDER_SUFFIX, valueByPlaceholderNameForModule) };
+
+    // The dependency chain is shader module -> fragment source -> sampling function
+    // Hence we need to fill in the templates in reverse order.
+
+    // Fill in sampling implementation in the sampling func
+    std::string source = replace(TEXTURE_SAMPLE_IMPL_MAIN_TEMPLATE, {
+        { TEXTURE_SAMPLE_IMPL, sampleImpl },
+    });
+
+    // Fill in sampling func in the fragment
+    source = replace(fragmentTemplate, {
+        { TEXTURE_SAMPLE_IMPL_MAIN, source },
+    });
+
+    // Fill in fragment in the shader module, first pass
+    source = replace(SHADER_SOURCE_TEMPLATE, {
+        { FRAGMENT_SHADER_SNIPPET, source },
+    });
+
+    // Fill in fragment in the shader module, second pass
+    source = replace(source, {
+        { TEXTURE_TYPE, srcTextureType },
+        { RET_ATTRIBUTE, retAttribute },
+        { RET_TYPE, retType },
+    });
+
+    // Fill in fragment in the shader module, third pass
+    // TEXTURE_TYPE, RET_ATTRIBUTE, RET_TYPE have a dependency on SRC_PRIM_TYPE, DST_PRIM_TYPE,
+    // VECTOR_DIM so we need another pass to fill them in.
+    source = replace(source, {
+        { VECTOR_DIM, vecDim },
+        { SRC_PRIM_TYPE, srcPrimType },
+        { DST_PRIM_TYPE, dstPrimType },
+    });
+
     wgpu::ShaderModuleWGSLDescriptor wgslDescriptor{};
-    wgslDescriptor.code = shaderSource.data();
+    wgslDescriptor.code = source.data();
     const wgpu::ShaderModuleDescriptor shaderModuleDescriptor{
         .nextInChain = &wgslDescriptor,
         .label = "render_pass_blit_shaders",

filament/backend/src/webgpu/WebGPUBlitter.h

@@ -97,11 +97,12 @@ private:
 
     struct RenderPipelineKey {
         wgpu::TextureViewDimension sourceDimension;   // 4 bytes
+        wgpu::TextureFormat sourceTextureFormat;      // 4
+        wgpu::TextureSampleType sourceSampleType;     // 4
         wgpu::TextureFormat destinationTextureFormat; // 4
         uint8_t sourceSampleCount;                    // 1
         SamplerMagFilter filterType;                  // 1
-        bool depthSource;                             // 1
-        uint8_t padding = 0;                          // 1
+        uint8_t padding[2] = {};                      // 2
 
         bool operator==(const RenderPipelineKey& other) const;
         using Hasher = utils::hash::MurmurHashFn<RenderPipelineKey>;
@@ -109,10 +110,10 @@ private:
 
     struct PipelineLayoutKey {
         wgpu::TextureViewDimension sourceDimension; // 4 bytes
+        wgpu::TextureSampleType sourceSampleType;   // 4
         SamplerMagFilter filterType;                // 1
         bool multisampledSource;                    // 1
-        bool depthSource;                           // 1
-        uint8_t padding = 0;                        // 1
+        uint8_t padding[2] = {};                    // 2
 
         bool operator==(const PipelineLayoutKey& other) const;
         using Hasher = utils::hash::MurmurHashFn<PipelineLayoutKey>;
@@ -120,20 +121,20 @@ private:
 
     struct ShaderModuleKey {
         wgpu::TextureViewDimension sourceDimension;   // 4 bytes
+        wgpu::TextureFormat sourceTextureFormat;      // 4
+        wgpu::TextureFormat destinationTextureFormat; // 4
         bool multisampledSource;                      // 1
-        bool depthSource;                             // 1
-        bool depthDestination;                        // 1
-        uint8_t padding = 0;                          // 1
+        uint8_t padding0[3] = {};                     // 3
 
         bool operator==(const ShaderModuleKey& other) const;
         using Hasher = utils::hash::MurmurHashFn<ShaderModuleKey>;
     };
 
-    static_assert(sizeof(RenderPipelineKey) == 12,
+    static_assert(sizeof(RenderPipelineKey) == 20,
             "RenderPipelineKey must not have implicit padding.");
-    static_assert(sizeof(PipelineLayoutKey) == 8,
+    static_assert(sizeof(PipelineLayoutKey) == 12,
             "PipelineLayoutKey must not have implicit padding.");
-    static_assert(sizeof(ShaderModuleKey) == 8, "ShaderModuleKey must not have implicit padding.");
+    static_assert(sizeof(ShaderModuleKey) == 16, "ShaderModuleKey must not have implicit padding.");
 
     wgpu::Device mDevice;
     wgpu::Sampler mNearestSampler{ nullptr };

filament/backend/src/webgpu/WebGPUBlitter_wgsl.inc

@@ -0,0 +1,174 @@
+/*
+ * 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.
+ */
+
+
+/*
+ * NOTE: This file is meant to be included with #include in WebGPUBlitter.cpp
+ * It is separated out for clarity.
+ */
+
+namespace {
+// note that the placeholders below must start and end with this prefix and suffix and should not
+// otherwise be present in the template:
+constexpr std::string_view PLACEHOLDER_PREFIX{ "{{" };
+constexpr std::string_view PLACEHOLDER_SUFFIX{ "}}" };
+
+#define PLACEHOLDER_DEF(name) constexpr std::string_view name = #name
+
+PLACEHOLDER_DEF(FRAGMENT_SHADER_SNIPPET);
+PLACEHOLDER_DEF(TEXTURE_SAMPLE_IMPL);
+PLACEHOLDER_DEF(TEXTURE_SAMPLE_IMPL_MAIN);
+
+// texture_multisampled_2d<f32>
+// texture_2d<f32/i32/u32>
+// texture_3d<f32/i32/u32>
+// texture_depth_2d
+// texture_depth_multisampled_2d
+PLACEHOLDER_DEF(TEXTURE_TYPE);
+
+// vec2, vec3
+PLACEHOLDER_DEF(VECTOR_DIM);
+
+// vec_<f32>, vec_<u32>, vec_<i32>, f32
+PLACEHOLDER_DEF(RET_TYPE);
+
+// @builtin(frag_depth), @location(0)
+PLACEHOLDER_DEF(RET_ATTRIBUTE);
+
+// <f32>, <u32>, <i32>
+PLACEHOLDER_DEF(DST_PRIM_TYPE);
+PLACEHOLDER_DEF(SRC_PRIM_TYPE);
+
+#undef PLACEHOLDER_DEF
+
+constexpr std::string_view SHADER_SOURCE_TEMPLATE{ R"(
+    struct BlitFragmentShaderArgs {
+        depthPlane:        u32,
+        scale:             vec2<f32>,
+        sourceOffset:      vec2<u32>,
+        destinationOffset: vec2<u32>,
+    };
+
+    @group(0) @binding(0) var sourceTexture: {{TEXTURE_TYPE}};
+    @group(0) @binding(1) var<uniform> fragmentShaderArgs: BlitFragmentShaderArgs;
+
+    fn getUnnormalizedSourceTextureCoordinates(position: vec2<f32>) -> vec2<f32> {
+        // These coordinates match the Vulkan vkCmdBlitImage spec:
+        // https://www.khronos.org/registry/vulkan/specs/1.2-extensions/man/html/vkCmdBlitImage.html
+        let uvOffset: vec2<f32> = position - vec2<f32>(f32(fragmentShaderArgs.destinationOffset.x),f32(fragmentShaderArgs.destinationOffset.y));
+        let uvScaled: vec2<f32> = uvOffset * fragmentShaderArgs.scale;
+        return uvScaled + vec2<f32>(f32(fragmentShaderArgs.sourceOffset.x), f32(fragmentShaderArgs.sourceOffset.y));
+    }
+
+    fn normalize2dSourceTextureCoordinates(
+            unnormalizedSourceTextureCoordinates: vec2<f32>,
+            sourceDimensions: vec2<u32>) -> vec2<f32> {
+        return unnormalizedSourceTextureCoordinates /
+            vec2<f32>(f32(sourceDimensions.x), f32(sourceDimensions.y));
+    }
+
+    @vertex
+    fn vertexShaderMain(@builtin(vertex_index) vertexIndex: u32) -> @builtin(position) vec4<f32> {
+        let fullScreenTriangleVertices = array<vec2<f32>, 3>(
+            vec2<f32>(-1.0, -1.0),
+            vec2<f32>( 3.0, -1.0),
+            vec2<f32>(-1.0,  3.0)
+        );
+        return vec4<f32>(fullScreenTriangleVertices[vertexIndex].xy, 0.0, 1.0);
+    }
+
+    {{FRAGMENT_SHADER_SNIPPET}}
+)"
+};
+
+constexpr std::string_view FLOAT_TEXTURE_SAMPLE_IMPL_TEMPLATE{ R"(
+        return textureSample(sourceTexture, sourceSampler, coordinates);
+)" };
+
+constexpr std::string_view INT_TEXTURE_SAMPLE_IMPL_TEMPLATE { R"(
+        let texelCoords = {{VECTOR_DIM}}<u32>(coordinates * {{VECTOR_DIM}}<f32>(sourceDimensions));
+        return textureLoad(sourceTexture, texelCoords, 0);
+)"};
+
+constexpr std::string_view TEXTURE_SAMPLE_IMPL_MAIN_TEMPLATE { R"(
+    // Assumes that "@group(0) @binding(2) var sourceSampler: sampler;" has been declared in the main program;
+    fn sampleTextureImpl(sourceTexture: {{TEXTURE_TYPE}}, sourceDimensions: {{VECTOR_DIM}}<u32>,
+                  coordinates: {{VECTOR_DIM}}<f32>) -> vec4{{DST_PRIM_TYPE}} {
+        {{TEXTURE_SAMPLE_IMPL}}
+    }
+)"};
+
+constexpr std::string_view FRAGMENT_SHADER_SNIPPET_MSAA_INPUT_TEMPLATE{ R"(
+    @fragment
+    fn fragmentShaderMain(@builtin(position) position: vec4<f32>) -> {{RET_ATTRIBUTE}} {{RET_TYPE}} {
+        var color: vec4<f32> = vec4<f32>(0.0, 0.0, 0.0, 0.0);
+        let numberOfSamples: u32 = textureNumSamples(sourceTexture);
+        let coordinatesF: vec2<f32> = getUnnormalizedSourceTextureCoordinates(position.xy);
+        let coordinates: vec2<u32> = vec2<u32>(u32(coordinatesF.x), u32(coordinatesF.y));
+        for (var sampleIndex: u32 = 0; sampleIndex < numberOfSamples; sampleIndex++) {
+            color += textureLoad(sourceTexture, coordinates, sampleIndex);
+        }
+        color /= f32(numberOfSamples);
+        return color;
+    }
+)" };
+
+constexpr std::string_view FRAGMENT_SHADER_SNIPPET_3D_INPUT_TEMPLATE{ R"(
+    @group(0) @binding(2) var sourceSampler: sampler;
+
+    {{TEXTURE_SAMPLE_IMPL_MAIN}}
+
+    fn normalize3dSourceTextureCoordinates(
+            unnormalizedSourceTextureCoordinates: vec2<f32>,
+            sourceDimensions: vec3<u32>) -> vec3<f32> {
+        let uvNormalized: vec2<f32> = normalize2dSourceTextureCoordinates(
+            unnormalizedSourceTextureCoordinates,
+            sourceDimensions.xy
+        );
+        return vec3<f32>(
+            uvNormalized,
+            (f32(fragmentShaderArgs.depthPlane) + 0.5) / f32(sourceDimensions.z)
+        );
+    }
+
+    @fragment
+    fn fragmentShaderMain(@builtin(position) position: vec4<f32>) -> {{RET_ATTRIBUTE}} {{RET_TYPE}} {
+        let sourceDimensions: vec3<u32> = textureDimensions(sourceTexture);
+        let coordinates: vec3<f32> = normalize3dSourceTextureCoordinates(
+            getUnnormalizedSourceTextureCoordinates(position),
+            sourceDimensions
+        );
+        return sampleTextureImpl(sourceTexture, sourceDimensions, coordinates);
+    }
+)" };
+
+constexpr std::string_view FRAGMENT_SHADER_SNIPPET_2D_INPUT_TEMPLATE { R"(
+    @group(0) @binding(2) var sourceSampler: sampler;
+
+    {{TEXTURE_SAMPLE_IMPL_MAIN}}
+
+    @fragment
+    fn fragmentShaderMain(@builtin(position) position: vec4<f32>) -> {{RET_ATTRIBUTE}} {{RET_TYPE}} {
+        let sourceDimensions: vec2<u32> = textureDimensions(sourceTexture);
+        let coordinates: vec2<f32> = normalize2dSourceTextureCoordinates(
+            getUnnormalizedSourceTextureCoordinates(position.xy),
+            sourceDimensions
+        );
+        return sampleTextureImpl(sourceTexture, sourceDimensions, coordinates);
+    }
+)" };
+
+} // namespace

filament/backend/src/webgpu/WebGPUDriver.cpp

@@ -446,6 +446,7 @@ void WebGPUDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch, uint32_t wi
 void WebGPUDriver::createVertexBufferInfoR(Handle<HwVertexBufferInfo> vertexBufferInfoHandle,
         const uint8_t bufferCount, const uint8_t attributeCount, const AttributeArray attributes,
         utils::ImmutableCString&& tag) {
+    // Hello world! This is a test for pushing branch to Github
     FWGPU_SYSTRACE_SCOPE();
     constructHandle<WebGPUVertexBufferInfo>(vertexBufferInfoHandle, bufferCount, attributeCount,
             attributes, mDeviceLimits);
@@ -1497,7 +1498,7 @@ void WebGPUDriver::readPixels(Handle<HwRenderTarget> sourceRenderTargetHandle, c
 
     mReadPixelMapsCounter.startTask();
     userData->buffer.MapAsync(
-            wgpu::MapMode::Read, 0, bufferSize, wgpu::CallbackMode::AllowProcessEvents,
+            wgpu::MapMode::Read, 0, bufferSize, wgpu::CallbackMode::AllowSpontaneous,
             [](wgpu::MapAsyncStatus status, const char* message, UserData* userdata) {
                 std::unique_ptr<UserData> data(static_cast<UserData*>(userdata));
                 if (UTILS_LIKELY(status == wgpu::MapAsyncStatus::Success)) {

filament/backend/src/webgpu/WebGPUTextureHelpers.h

@@ -41,6 +41,45 @@ namespace filament::backend {
            textureFormat == wgpu::TextureFormat::Depth32FloatStencil8;
 }
 
+[[nodiscard]] constexpr bool isUIntFormat(const wgpu::TextureFormat format) {
+    // see https://www.w3.org/TR/webgpu/#texture-formats
+    //  and https://www.w3.org/TR/webgpu/#texture-format-caps
+    switch (format) {
+        case wgpu::TextureFormat::R8Uint:
+        case wgpu::TextureFormat::R16Uint:
+        case wgpu::TextureFormat::RG8Uint:
+        case wgpu::TextureFormat::R32Uint:
+        case wgpu::TextureFormat::RG16Uint:
+        case wgpu::TextureFormat::RGBA8Uint:
+        case wgpu::TextureFormat::RGB10A2Uint:
+        case wgpu::TextureFormat::RG32Uint:
+        case wgpu::TextureFormat::RGBA16Uint:
+        case wgpu::TextureFormat::RGBA32Uint:
+            return true;
+        default:
+            return false;
+    }
+}
+
+[[nodiscard]] constexpr bool isIntFormat(const wgpu::TextureFormat format) {
+    // see https://www.w3.org/TR/webgpu/#texture-formats
+    //  and https://www.w3.org/TR/webgpu/#texture-format-caps
+    switch (format) {
+        case wgpu::TextureFormat::R8Sint:
+        case wgpu::TextureFormat::R16Sint:
+        case wgpu::TextureFormat::RG8Sint:
+        case wgpu::TextureFormat::R32Sint:
+        case wgpu::TextureFormat::RG16Sint:
+        case wgpu::TextureFormat::RGBA8Sint:
+        case wgpu::TextureFormat::RG32Sint:
+        case wgpu::TextureFormat::RGBA16Sint:
+        case wgpu::TextureFormat::RGBA32Sint:
+            return true;
+        default:
+            return false;
+    }
+}
+
 [[nodiscard]] constexpr std::string_view toString(const PixelDataFormat format) {
     switch (format) {
         case PixelDataFormat::R:               return "R";

filament/backend/src/webgpu/utils/StringPlaceholderTemplateProcessor.cpp

@@ -52,15 +52,14 @@ std::string processPlaceholderTemplate(std::string_view const& stringTemplate,
                          "Malformed source with missing suffix to placeholder");
         const std::string_view placeholderName{ std::string_view(sourceData + positionOfPlaceholder,
                 positionAfterPlaceholder - positionOfPlaceholder) };
+
         if (const auto iter{ valueByPlaceholderName.find(placeholderName) };
-                iter == valueByPlaceholderName.end()) {
-            // wrapping placeholderName in a std::string to null terminate the C string....
-            PANIC_POSTCONDITION("Found placeholder '%s' in template, but this is not present in "
-                                "valueByPlaceholderName",
-                    std::string{ placeholderName }.data());
-        } else {
+                iter != valueByPlaceholderName.end()) {
             const std::string_view value{ iter->second };
             out << value;
+        } else {
+            // It's ok to not replace a template item.  We assume replacement can take multiple passes.
+            out << placeholderPrefix << placeholderName << placeholderSuffix;
         }
         // update the cursor for after the placeholder...
         positionCursorInTemplateString = positionAfterPlaceholder + placeholderSuffix.size();

filament/backend/test/Arguments.cpp

@@ -23,13 +23,17 @@
 
 namespace test {
 
-Backend parseArgumentsForBackend(int argc, char* argv[]) {
-    Backend backend = Backend::OPENGL;
+TestArguments parseArguments(int argc, char* argv[]) {
+    TestArguments arguments = {};
+    arguments.backend = Backend::OPENGL;
+
     // The first colon in OPTSTR turns on silent error reporting. This is important, as the
     // arguments may also contain gtest parameters we don't know about.
-    static constexpr const char* OPTSTR = ":a:";
+    static constexpr const char* OPTSTR = ":a:kc";
     static const struct option OPTIONS[] = {
             { "api", required_argument, nullptr, 'a' },
+            { "headless_only", no_argument, nullptr, 'k' },
+            { "ci", no_argument, nullptr, 'c' },
             { nullptr, 0, nullptr, 0 }  // termination of the option list
     };
 
@@ -41,13 +45,13 @@ Backend parseArgumentsForBackend(int argc, char* argv[]) {
         switch (opt) {
             case 'a':
                 if (arg == "opengl") {
-                    backend = Backend::OPENGL;
+                    arguments.backend = Backend::OPENGL;
                 } else if (arg == "vulkan") {
-                    backend = Backend::VULKAN;
+                    arguments.backend = Backend::VULKAN;
                 } else if (arg == "metal") {
-                    backend = Backend::METAL;
+                    arguments.backend = Backend::METAL;
                 } else if (arg == "webgpu") {
-                    backend = Backend::WEBGPU;
+                    arguments.backend = Backend::WEBGPU;
                 } else {
                     std::cerr << "Unrecognized target API. Must be 'opengl'|'vulkan'|'metal'|'webgpu'."
                               << std::endl
@@ -55,10 +59,16 @@ Backend parseArgumentsForBackend(int argc, char* argv[]) {
                               << std::endl;
                 }
                 break;
+            case 'k':
+                arguments.headlessOnly = true;
+                break;
+            case 'c':
+                arguments.isContinuousIntegration = true;
+                break;
         }
     }
 
-    return backend;
+    return arguments;
 }
 
 } // namespace test

filament/backend/test/PlatformRunner.h

@@ -41,6 +41,7 @@ enum class OperatingSystem: uint8_t {
     LINUX = 2,
     // Also represents iOS phones.
     APPLE = 3,
+    CONTINUOUS_INTEGRATION = 4,
     // TODO: When tests support windows add it here.
 };
 
@@ -73,11 +74,17 @@ void initTests(Backend backend, OperatingSystem operatingSystem, bool isMobile,
  */
 int runTests();
 
+struct TestArguments {
+    Backend backend;
+    bool headlessOnly = false;
+    bool isContinuousIntegration = false;
+};
+
 /**
  * A utility method that can be invoked by test runners to parse arguments.
  * Looks through the provided command-line arguments and finds any -a <backend> arguments.
  */
-Backend parseArgumentsForBackend(int argc, char* argv[]);
+TestArguments parseArguments(int argc, char* argv[]);
 
 } // namespace test
 

filament/backend/test/linux_runner.cpp

@@ -43,7 +43,10 @@ std::array<test::Backend, 2> const VALID_BACKENDS{
 }// namespace
 
 int main(int argc, char* argv[]) {
-    auto backend = test::parseArgumentsForBackend(argc, argv);
+    const auto arguments = test::parseArguments(argc, argv);
+    const auto backend = arguments.backend;
+
+    // Note that Linux is headless-only.
 
     if (!std::any_of(VALID_BACKENDS.begin(), VALID_BACKENDS.end(),
                 [backend](test::Backend validBackend) { return backend == validBackend; })) {
@@ -51,6 +54,8 @@ int main(int argc, char* argv[]) {
         return 1;
     }
 
-    test::initTests(backend, test::OperatingSystem::LINUX, false, argc, argv);
+    const auto operatingSystem = arguments.isContinuousIntegration ?
+            test::OperatingSystem::CONTINUOUS_INTEGRATION : test::OperatingSystem::LINUX;
+    test::initTests(backend, operatingSystem, false, argc, argv);
     return test::runTests();
 }

filament/backend/test/mac_runner.mm

@@ -32,29 +32,34 @@ test::NativeView getNativeView() {
 @interface AppDelegate : NSObject <NSApplicationDelegate>
 
 @property test::Backend backend;
+@property bool headlessOnly;
 
 @end
 
 @implementation AppDelegate
 
 - (void)applicationDidFinishLaunching:(NSNotification *)aNotification {
-    NSView* view = [self createView];
-
-    if (self.backend == test::Backend::OPENGL) {
-        nativeView.ptr = (void*) view;
+    if (self.headlessOnly) {
+        nativeView.ptr = nullptr;
+        nativeView.width = test::WINDOW_WIDTH;
+        nativeView.height = test::WINDOW_HEIGHT;
+    } else {
+        NSView* view = [self createView];
+        switch (self.backend) {
+            case test::Backend::OPENGL:
+                nativeView.ptr = (void*) view;
+                break;
+            case test::Backend::METAL:
+            case test::Backend::VULKAN:
+            case test::Backend::WEBGPU:
+            case test::Backend::NOOP:
+                nativeView.ptr = (void*) view.layer;
+                break;
+        }
+        CGSize drawableSize = ((CAMetalLayer*) view.layer).drawableSize;
+        nativeView.width = static_cast<size_t>(drawableSize.width);
+        nativeView.height = static_cast<size_t>(drawableSize.height);
     }
-    if (self.backend == test::Backend::METAL) {
-        nativeView.ptr = (void*) view.layer;
-    }
-    if (self.backend == test::Backend::VULKAN) {
-        nativeView.ptr = (void*) view.layer;
-    }
-    if (self.backend == test::Backend::WEBGPU) {
-        nativeView.ptr = (void*) view.layer;
-    }
-    CGSize drawableSize = ((CAMetalLayer*) view.layer).drawableSize;
-    nativeView.width = static_cast<size_t>(drawableSize.width);
-    nativeView.height = static_cast<size_t>(drawableSize.height);
 
     exit(test::runTests());
 }
@@ -100,11 +105,25 @@ test::NativeView getNativeView() {
 @end
 
 int main(int argc, char* argv[]) {
-    auto backend = test::parseArgumentsForBackend(argc, argv);
-    test::initTests(backend, test::OperatingSystem::APPLE, false, argc, argv);
-    AppDelegate* delegate = [AppDelegate new];
-    delegate.backend = backend;
+    const auto arguments = test::parseArguments(argc, argv);
+    const auto operatingSystem = arguments.isContinuousIntegration ?
+            test::OperatingSystem::CONTINUOUS_INTEGRATION : test::OperatingSystem::APPLE;
+    test::initTests(arguments.backend, operatingSystem, false, argc, argv);
+
     NSApplication* app = [NSApplication sharedApplication];
+    AppDelegate* delegate = [AppDelegate new];
+    delegate.backend = arguments.backend;
+    delegate.headlessOnly = arguments.headlessOnly;
     [app setDelegate:delegate];
+
+    if (arguments.headlessOnly) {
+        // In headless mode, we don't want to start the NSApplication event loop.
+        // Instead, we can manually "finish" launching the app, which will trigger the tests to run.
+        [app finishLaunching];
+        [delegate applicationDidFinishLaunching:nil];
+        // The line above calls exit(), so we should not reach here.
+        return 0;
+    }
+
     [app run];
 }

filament/backend/test/test_Blit.cpp

@@ -158,7 +158,7 @@ TEST_F(BlitTest, ColorMagnify) {
     constexpr int kNumLevels = 3;
 
     // Create a SwapChain and make it current. We don't really use it so the res doesn't matter.
-    auto swapChain = mCleanup.add(api.createSwapChainHeadless(256, 256, 0));
+    auto swapChain = mCleanup.add(createSwapChain());
     api.makeCurrent(swapChain, swapChain);
 
     // Create a source texture.
@@ -222,7 +222,7 @@ TEST_F(BlitTest, ColorMinify) {
     constexpr int kNumLevels = 3;
 
     // Create a SwapChain and make it current. We don't really use it so the res doesn't matter.
-    auto swapChain = mCleanup.add(api.createSwapChainHeadless(256, 256, 0));
+    auto swapChain = mCleanup.add(createSwapChain());
     api.makeCurrent(swapChain, swapChain);
 
     // Create a source texture.
@@ -371,7 +371,7 @@ TEST_F(BlitTest, Blit2DTextureArray) {
     constexpr int kDstTexLayer = 0;
 
     // Create a SwapChain and make it current. We don't really use it so the res doesn't matter.
-    auto swapChain = mCleanup.add(api.createSwapChainHeadless(256, 256, 0));
+    auto swapChain = mCleanup.add(createSwapChain());
     api.makeCurrent(swapChain, swapChain);
 
     // Create a source texture.
@@ -437,7 +437,7 @@ TEST_F(BlitTest, BlitRegion) {
     constexpr int kDstLevel = 0;
 
     // Create a SwapChain and make it current. We don't really use it so the res doesn't matter.
-    auto swapChain = mCleanup.add(api.createSwapChainHeadless(256, 256, 0));
+    auto swapChain = mCleanup.add(createSwapChain());
     api.makeCurrent(swapChain, swapChain);
 
     // Create a source texture.

filament/backend/test/test_Callbacks.cpp

@@ -103,7 +103,7 @@ TEST_F(BackendTest, FrameCompletedCallback) {
     Cleanup cleanup(api);
 
     // Create a SwapChain.
-    auto swapChain = cleanup.add(api.createSwapChainHeadless(256, 256, 0));
+    auto swapChain = cleanup.add(createSwapChain());
 
     int callbackCountA = 0;
     api.setFrameCompletedCallback(swapChain, nullptr,

filament/backend/test/test_MemoryMappedBuffer.cpp

@@ -22,6 +22,7 @@
 #include "SharedShaders.h"
 #include "SharedShadersConstants.h"
 #include "Skip.h"
+#include "Workarounds.h"
 
 #include <backend/BufferDescriptor.h>
 #include <backend/DriverEnums.h>
@@ -55,7 +56,7 @@ protected:
         auto colorTexture = cleanup.add(
                 api.createTexture(SamplerType::SAMPLER_2D, 1, TextureFormat::RGBA8, 1,
                         screenWidth(), screenHeight(), 1,
-                        TextureUsage::COLOR_ATTACHMENT | TextureUsage::SAMPLEABLE));
+                        TextureUsage::COLOR_ATTACHMENT | TextureUsage::SAMPLEABLE TEXTURE_USAGE_READ_PIXELS));
         auto renderTarget = cleanup.add(api.createRenderTarget(TargetBufferFlags::COLOR0,
                 screenWidth(), screenHeight(), 1, 1, { { colorTexture } }, {}, {}));
         return renderTarget;
@@ -90,19 +91,27 @@ protected:
         return { vbih, vbh, ibh };
     }
 
-    Shader getSimpleShader(Cleanup& cleanup, math::float4 const& color) {
+    std::pair<Shader, DescriptorSetHandle> getSimpleShader(Cleanup& cleanup,
+            math::float4 const& color) {
         auto& api = getDriverApi();
-        Shader const shader = SharedShaders::makeShader(getDriverApi(), cleanup, ShaderRequest{
-            .mVertexType = VertexShaderType::Simple,
-            .mFragmentType = FragmentShaderType::SolidColored,
-            .mUniformType = ShaderUniformType::Simple
-        });
+        Shader const shader = SharedShaders::makeShader(getDriverApi(), cleanup,
+                ShaderRequest{
+                    .mVertexType = VertexShaderType::Simple,
+                    .mFragmentType = FragmentShaderType::SolidColored,
+                    .mUniformType = ShaderUniformType::Simple,
+                });
 
+        auto descSet = shader.createDescriptorSet(api);
+        UniformBindingConfig uboBindingConfig = {
+            .descriptorSet = descSet,
+        };
         auto const ubuffer = cleanup.add(api.createBufferObject(sizeof(SimpleMaterialParams),
                 BufferObjectBinding::UNIFORM, BufferUsage::DYNAMIC_BIT));
-        shader.bindUniform<SimpleMaterialParams>(api, ubuffer);
+        // This will bind the UBO to descSet and also bind descSet.  But we also need to manually
+        // bind descset every begin/end frame.
+        shader.bindUniform<SimpleMaterialParams>(api, ubuffer, uboBindingConfig);
         shader.uploadUniform(api, ubuffer, SimpleMaterialParams{ .color = color });
-        return shader;
+        return { shader, descSet };
     }
 
     void copyData(const void* data, size_t const size, size_t const offset,
@@ -122,7 +131,9 @@ protected:
             int64_t const frame,
             SwapChainHandle swapChain,
             RenderTargetHandle renderTarget,
-            RenderPrimitiveHandle renderPrimitive, PipelineState const& state) {
+            RenderPrimitiveHandle renderPrimitive,
+            DescriptorSetHandle descSet,
+            PipelineState const& state) {
         auto& api = getDriverApi();
         RenderPassParams params = getClearColorRenderPass({0,0,0,1});
         params.viewport = getFullViewport();
@@ -130,6 +141,8 @@ protected:
         api.beginFrame(frame, 0, 0);
         api.beginRenderPass(renderTarget, params);
         api.bindPipeline(state);
+        // The binding index is always 0 for the simple shaders we're using.
+        api.bindDescriptorSet(descSet, 0, {});
         api.bindRenderPrimitive(renderPrimitive);
         api.draw2(0, count, 1);
         api.endRenderPass();
@@ -185,7 +198,7 @@ protected:
         flushAndWait();
         EXPECT_EQ(callbackExecuted, 1);
 
-        Shader const shader = getSimpleShader(cleanup, color);
+        auto [shader, descset] = getSimpleShader(cleanup, color);
 
         auto [vbih, vbh, ibh] = setupGeometryBuffer(cleanup, 3, stride, baseVertex, bufferObject);
 
@@ -197,7 +210,7 @@ protected:
         state.primitiveType = PrimitiveType::TRIANGLES;
         state.vertexBufferInfo = vbih;
 
-        render(3, 0, swapChain, renderTarget, renderPrimitive, state);
+        render(3, 0, swapChain, renderTarget, renderPrimitive, descset, state);
 
         EXPECT_IMAGE(renderTarget, ScreenshotParams(screenWidth(), screenHeight(), screenshotName, 0));
     }
@@ -288,7 +301,7 @@ TEST_F(MemoryMappedTest, MultipleCopies) {
     flushAndWait();
     EXPECT_EQ(callbacksExecuted, 3);
 
-    Shader const shader = getSimpleShader(cleanup, { 1, 1, 0, 1 });
+    auto [shader, descset] = getSimpleShader(cleanup, { 1, 1, 0, 1 });
 
     auto [vbih, vbh, ibh] = setupGeometryBuffer(cleanup, 9, sizeof(math::float2), 0, bufferObject);
 
@@ -300,7 +313,7 @@ TEST_F(MemoryMappedTest, MultipleCopies) {
     state.primitiveType = PrimitiveType::TRIANGLES;
     state.vertexBufferInfo = vbih;
 
-    render(9, 0, swapChain, renderTarget, renderPrimitive, state);
+    render(9, 0, swapChain, renderTarget, renderPrimitive, descset, state);
 
     EXPECT_IMAGE(renderTarget, ScreenshotParams(screenWidth(), screenHeight(), "MultipleCopies", 0));
 }
@@ -345,7 +358,7 @@ TEST_F(MemoryMappedTest, UpdatePartial) {
     }
 
 
-    Shader const shader = getSimpleShader(cleanup, { 1, 1, 0, 1 });
+    auto [shader, descset] = getSimpleShader(cleanup, { 1, 1, 0, 1 });
 
     auto [vbih, vbh, ibh] = setupGeometryBuffer(cleanup, 9, sizeof(math::float2), 0, bufferObject);
 
@@ -357,7 +370,7 @@ TEST_F(MemoryMappedTest, UpdatePartial) {
     state.primitiveType = PrimitiveType::TRIANGLES;
     state.vertexBufferInfo = vbih;
 
-    render(9, 0, swapChain, renderTarget, renderPrimitive, state);
+    render(9, 0, swapChain, renderTarget, renderPrimitive, descset, state);
 
     EXPECT_IMAGE(renderTarget, ScreenshotParams(screenWidth(), screenHeight(), "UpdatePartial_before", 0));
 
@@ -380,7 +393,7 @@ TEST_F(MemoryMappedTest, UpdatePartial) {
     api.makeCurrent(swapChain, swapChain);
 
     // Second render, after update
-    render(9, 1, swapChain, renderTarget, renderPrimitive, state);
+    render(9, 1, swapChain, renderTarget, renderPrimitive, descset, state);
 
     EXPECT_IMAGE(renderTarget, ScreenshotParams(screenWidth(), screenHeight(), "UpdatePartial_after", 0));
 }

filament/backend/test/test_ReadPixels.cpp

@@ -385,8 +385,7 @@ TEST_F(ReadPixelsTest, ReadPixelsPerformance) {
     Cleanup cleanup(api);
 
     // Create a platform-specific SwapChain and make it current.
-    auto swapChain = cleanup.add(
-            api.createSwapChainHeadless(renderTargetSize, renderTargetSize, 0));
+    auto swapChain = cleanup.add(createSwapChain());
     api.makeCurrent(swapChain, swapChain);
 
     Shader shader = SharedShaders::makeShader(api, cleanup, ShaderRequest{

filament/backend/test/test_Scissor.cpp

@@ -70,7 +70,7 @@ TEST_F(BackendTest, ScissorViewportRegion) {
     // executeCommands().
     {
         // Create a SwapChain and make it current. We don't really use it so the res doesn't matter.
-        auto swapChain = cleanup.add(api.createSwapChainHeadless(256, 256, 0));
+        auto swapChain = cleanup.add(createSwapChain());
         api.makeCurrent(swapChain, swapChain);
 
         Shader shader = SharedShaders::makeShader(api, cleanup,
@@ -149,7 +149,7 @@ TEST_F(BackendTest, ScissorViewportEdgeCases) {
     // executeCommands().
     {
         // Create a SwapChain and make it current. We don't really use it so the res doesn't matter.
-        auto swapChain = cleanup.add(api.createSwapChainHeadless(256, 256, 0));
+        auto swapChain = cleanup.add(createSwapChain());
         api.makeCurrent(swapChain, swapChain);
 
         Shader shader = SharedShaders::makeShader(api, cleanup, ShaderRequest{

filament/backend/test/test_Template.cpp

@@ -33,8 +33,7 @@ TEST_F(BackendTest, TestTemplate) {
 
     auto& api = getDriverApi();
     Cleanup cleanup(api);
-    auto swapChain =
-            cleanup.add(api.createSwapChainHeadless(kRenderTargetSize, kRenderTargetSize, 0));
+    auto swapChain = cleanup.add(createSwapChain());
     api.makeCurrent(swapChain, swapChain);
     RenderTargetHandle renderTarget = cleanup.add(api.createDefaultRenderTarget());
 

filament/src/details/BufferAllocator.cpp

@@ -0,0 +1,234 @@
+/*
+ * 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 "details/BufferAllocator.h"
+
+#include <utils/Panic.h>
+#include <utils/debug.h>
+
+namespace filament {
+namespace {
+
+static bool isValidId(BufferAllocator::AllocationId id) {
+    return id != BufferAllocator::UNALLOCATED && id != BufferAllocator::REALLOCATION_REQUIRED;
+}
+
+#ifndef NDEBUG
+constexpr static bool isPowerOfTwo(uint32_t n) {
+    return (n > 0) && ((n & (n - 1)) == 0);
+}
+#endif
+
+} // anonymous namespace
+
+BufferAllocator::BufferAllocator(allocation_size_t totalSize, allocation_size_t slotSize)
+    : mTotalSize(totalSize),
+      mSlotSize(slotSize) {
+    assert_invariant(mSlotSize > 0);
+    assert_invariant(isPowerOfTwo(mSlotSize));
+
+    reset(mTotalSize);
+}
+
+void BufferAllocator::reset(allocation_size_t newTotalSize) {
+    assert_invariant(newTotalSize % mSlotSize == 0);
+
+    mTotalSize = newTotalSize;
+
+    mSlotPool.clear();
+    mFreeList.clear();
+    mOffsetMap.clear();
+
+    // Initialize the pool with a single large free slot.
+    mSlotPool.emplace_back(InternalSlotNode{
+        .mSlot = {
+            .offset = 0,
+            .slotSize = mTotalSize,
+            .isAllocated = false,
+            .gpuUseCount = 0
+        }
+    });
+
+    InternalSlotNode* firstNode = &mSlotPool.front();
+
+    // Add this initial free slot to the free list and offset map.
+    auto freeListIter = mFreeList.emplace(newTotalSize, firstNode);
+    auto offsetMapIter = mOffsetMap.emplace(0, firstNode);
+
+    firstNode->mSlotPoolIterator = mSlotPool.begin();
+    firstNode->mFreeListIterator = freeListIter;
+    firstNode->mOffsetMapIterator = offsetMapIter.first;
+}
+
+std::pair<BufferAllocator::AllocationId, BufferAllocator::allocation_size_t>
+    BufferAllocator::allocate(allocation_size_t size) noexcept {
+    if (size == 0) {
+        return { UNALLOCATED, 0 };
+    }
+
+    const allocation_size_t alignedSize = alignUp(size);
+    auto bestFitIter = mFreeList.lower_bound(alignedSize);
+
+    if (bestFitIter == mFreeList.end()) {
+        return { REALLOCATION_REQUIRED, 0 };
+    }
+
+    InternalSlotNode* targetNode = bestFitIter->second;
+    const allocation_size_t originalSlotSize = targetNode->mSlot.slotSize;
+
+    mFreeList.erase(bestFitIter);
+    targetNode->mFreeListIterator = mFreeList.end();
+    targetNode->mSlot.isAllocated = true;
+
+    // Split the slot if it is larger than what we need.
+    if (originalSlotSize > alignedSize) {
+        targetNode->mSlot.slotSize = alignedSize;
+
+        allocation_size_t remainingSize = originalSlotSize - alignedSize;
+        allocation_size_t newSlotOffset = targetNode->mSlot.offset + alignedSize;
+        assert_invariant(remainingSize % mSlotSize == 0);
+        assert_invariant(newSlotOffset % mSlotSize == 0);
+
+        // Create a new node for the remaining free space.
+        auto insertPos = std::next(targetNode->mSlotPoolIterator);
+        auto newNodeIter = mSlotPool.emplace(insertPos, InternalSlotNode{
+            .mSlot = {
+                .offset = newSlotOffset,
+                .slotSize = remainingSize,
+                .isAllocated = false,
+                .gpuUseCount = 0
+            }
+        });
+        InternalSlotNode* newNode = &(*newNodeIter);
+
+        // Add the new free slot to our tracking maps.
+        auto freeListIter = mFreeList.emplace(remainingSize, newNode);
+        auto offsetMapIter = mOffsetMap.emplace(newSlotOffset, newNode);
+        newNode->mSlotPoolIterator = newNodeIter;
+        newNode->mFreeListIterator = freeListIter;
+        newNode->mOffsetMapIterator = offsetMapIter.first;
+    }
+
+    auto allocationId = calculateIdByOffset(targetNode->mSlot.offset);
+    return { allocationId, targetNode->mSlot.offset };
+}
+
+BufferAllocator::InternalSlotNode* BufferAllocator::getNodeById(
+        AllocationId id) const noexcept {
+    if (!isValidId(id)) {
+        return nullptr;
+    }
+
+    auto offset = getAllocationOffset(id);
+    auto iter = mOffsetMap.find(offset);
+
+    // We cannot find the corresponding node in the map.
+    if (iter == mOffsetMap.end()) {
+        return nullptr;
+    }
+    return iter->second;
+}
+
+void BufferAllocator::retire(AllocationId id) {
+    auto targetNode = getNodeById(id);
+    assert_invariant(targetNode != nullptr);
+
+    targetNode->mSlot.isAllocated = false;
+}
+
+void BufferAllocator::acquireGpu(AllocationId id) {
+    auto targetNode = getNodeById(id);
+    assert_invariant(targetNode != nullptr);
+
+    targetNode->mSlot.gpuUseCount++;
+}
+
+void BufferAllocator::releaseGpu(AllocationId id) {
+    auto targetNode = getNodeById(id);
+    assert_invariant(targetNode != nullptr);
+    assert_invariant(targetNode->mSlot.gpuUseCount > 0);
+
+    targetNode->mSlot.gpuUseCount--;
+}
+
+void BufferAllocator::releaseFreeSlots() {
+    auto curr = mSlotPool.begin();
+    while (curr != mSlotPool.end()) {
+        if (!curr->mSlot.isFree()) {
+            ++curr;
+            continue;
+        }
+
+        auto next = std::next(curr);
+        bool merged = false;
+        while (next != mSlotPool.end() && next->mSlot.isFree()) {
+            merged = true;
+            // Combine the size of free slots
+            curr->mSlot.slotSize += next->mSlot.slotSize;
+            assert_invariant(curr->mSlot.slotSize % mSlotSize == 0);
+
+            // Erase the merged slot from all maps
+            if (next->mFreeListIterator != mFreeList.end()) {
+                mFreeList.erase(next->mFreeListIterator);
+            }
+            mOffsetMap.erase(next->mOffsetMapIterator);
+            next = mSlotPool.erase(next);
+        }
+
+        // If we performed any merge, the current block's size has changed.
+        // We need to update its position in the mFreeList.
+        if (curr->mFreeListIterator != mFreeList.end()) {
+            // If it's already in the free list and we merged, we need to update it.
+            if (merged) {
+                mFreeList.erase(curr->mFreeListIterator);
+                curr->mFreeListIterator = mFreeList.emplace(curr->mSlot.slotSize, &(*curr));
+            }
+        } else {
+            // If it's not in the free list, it must be a newly freed block. Add it.
+            curr->mFreeListIterator = mFreeList.emplace(curr->mSlot.slotSize, &(*curr));
+        }
+
+        curr = next;
+    }
+}
+
+BufferAllocator::allocation_size_t BufferAllocator::getTotalSize() const noexcept {
+    return mTotalSize;
+}
+
+BufferAllocator::allocation_size_t
+    BufferAllocator::getAllocationOffset(AllocationId id) const {
+    assert_invariant(isValidId(id));
+
+    return (id - 1) * mSlotSize;
+}
+
+BufferAllocator::AllocationId BufferAllocator::calculateIdByOffset(
+        allocation_size_t offset) const {
+    assert_invariant(offset % mSlotSize == 0);
+
+    // The ID is 1-based since we use 0 for UNALLOCATED.
+    return (offset / mSlotSize) + 1;
+}
+
+BufferAllocator::allocation_size_t BufferAllocator::alignUp(
+        allocation_size_t size) const noexcept {
+    if (size == 0) return 0;
+
+    return (size + mSlotSize - 1) & ~(mSlotSize - 1);
+}
+
+} // namespace filament

filament/src/details/BufferAllocator.h

@@ -0,0 +1,118 @@
+/*
+ * 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_DETAILS_BUFFERALLOCATOR_H
+#define TNT_FILAMENT_DETAILS_BUFFERALLOCATOR_H
+
+#include <cstdint>
+#include <list>
+#include <map>
+#include <unordered_map>
+
+namespace filament {
+
+// This class is NOT thread-safe.
+//
+// It internally manages shared state (e.g., mSlotPool, mFreeList, mOffsetMap) without any
+// synchronization primitives. Concurrent access from multiple threads to the same
+// BufferAllocator instance will result in data races and undefined behavior.
+//
+// If an instance of this class is to be shared between threads, all calls to its member
+// functions MUST be protected by external synchronization (e.g., a std::mutex).
+class BufferAllocator {
+public:
+    using allocation_size_t = uint32_t;
+    using AllocationId = uint32_t;
+
+    static constexpr AllocationId UNALLOCATED = 0;
+    static constexpr AllocationId REALLOCATION_REQUIRED = ~0u;
+
+    struct Slot {
+        const allocation_size_t offset;       // 4 bytes
+        allocation_size_t slotSize;           // 4 bytes
+        bool isAllocated;                     // 1 byte
+        char padding[3];                      // 3 bytes
+        uint32_t gpuUseCount;                 // 4 bytes
+
+        [[nodiscard]] bool isFree() const noexcept {
+            return !isAllocated && gpuUseCount == 0;
+        }
+    };
+
+    // `slotSize` is derived from the GPU's uniform buffer offset alignment requirement,
+    // which can be up to 256 bytes.
+    explicit BufferAllocator(allocation_size_t totalSize,
+            allocation_size_t slotSize);
+
+    BufferAllocator(BufferAllocator const&) = delete;
+    BufferAllocator(BufferAllocator&&) = delete;
+
+    // Allocate a new slot and return its id and slot offset in the UBO.
+    // If the returned id is not valid, that means there's no large enough slot for allocation.
+    [[nodiscard]] std::pair<AllocationId, allocation_size_t> allocate(
+            allocation_size_t size) noexcept;
+
+    // Call it when MaterialInstance gives up the ownership of the allocation.
+    // We don't release the slot immediately in this function even if it is not being used,
+    // the release is centralized in releaseFreeSlots().
+    void retire(AllocationId id);
+
+    // Increments the GPU read-lock.
+    void acquireGpu(AllocationId id);
+
+    // Decrements the GPU read-lock.
+    // We don't release the slot immediately in this function even if it is not being used,
+    // the release is centralized in releaseFreeSlots().
+    void releaseGpu(AllocationId id);
+
+    // Traverse all slots and free all slots that are not being used by both CPU and GPU.
+    // Perform the merge at the same time.
+    void releaseFreeSlots();
+
+    // Resets the allocator to its initial state with a new total size.
+    // All existing allocations are cleared.
+    void reset(allocation_size_t newTotalSize);
+
+    // Size of the UBO in bytes.
+    [[nodiscard]] allocation_size_t getTotalSize() const noexcept;
+
+    // Query the allocation offset by AllocationId.
+    allocation_size_t getAllocationOffset(AllocationId id) const;
+
+private:
+    [[nodiscard]] AllocationId calculateIdByOffset(allocation_size_t offset) const;
+    [[nodiscard]] allocation_size_t alignUp(allocation_size_t size) const noexcept;
+
+    // Having an internal node type holding the base slot node and additional information.
+    struct InternalSlotNode {
+        Slot mSlot;
+        std::list<InternalSlotNode>::iterator mSlotPoolIterator;
+        std::multimap<allocation_size_t, InternalSlotNode*>::iterator mFreeListIterator;
+        std::unordered_map<allocation_size_t, InternalSlotNode*>::iterator mOffsetMapIterator;
+    };
+
+    [[nodiscard]] InternalSlotNode* getNodeById(AllocationId id) const noexcept;
+
+    allocation_size_t mTotalSize;
+    const allocation_size_t mSlotSize; // Size of a single slot in bytes.
+    std::list<InternalSlotNode> mSlotPool; // All slots, including both allocated and freed
+    std::multimap</*slot size*/allocation_size_t, InternalSlotNode*> mFreeList;
+    std::unordered_map</*slot offset*/allocation_size_t, InternalSlotNode*> mOffsetMap;
+};
+
+} // namespace filament
+
+#endif // TNT_FILAMENT_DETAILS_BUFFERALLOCATOR_H

filament/src/details/Engine.h

@@ -776,6 +776,7 @@ public:
         } backend;
         struct {
             bool check_crc32_after_loading = false;
+            bool enable_material_instance_uniform_batching = false;
         } material;
     } features;
 
@@ -825,6 +826,9 @@ public:
             { "material.check_crc32_after_loading",
               "Verify the checksum of package data when a material is loaded.",
               &features.material.check_crc32_after_loading, false },
+            { "material.enable_material_instance_uniform_batching",
+              "Make all MaterialInstances share a common large uniform buffer and use sub-allocations within it.",
+              &features.material.enable_material_instance_uniform_batching, false },
     }};
 
     utils::Slice<const FeatureFlag> getFeatureFlags() const noexcept {

filament/src/details/MaterialInstance.cpp

@@ -60,8 +60,13 @@ namespace filament {
 using namespace backend;
 
 FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
-                                     const char* name) noexcept
-        : mMaterial(material),
+        const char* name) noexcept
+        : FMaterialInstance(engine, material, name,
+                  engine.features.material.enable_material_instance_uniform_batching) {
+}
+
+FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
+                      const char* name, bool useUboBatching) noexcept : mMaterial(material),
           mDescriptorSet("MaterialInstance", material->getDescriptorSetLayout()),
           mCulling(CullingMode::BACK),
           mShadowCulling(CullingMode::BACK),
@@ -71,6 +76,7 @@ FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
           mHasScissor(false),
           mIsDoubleSided(false),
           mIsDefaultInstance(false),
+          mUseUboBatching(useUboBatching),
           mTransparencyMode(TransparencyMode::DEFAULT),
           mName(name ? CString(name) : material->getName()) {
 
@@ -80,12 +86,16 @@ FMaterialInstance::FMaterialInstance(FEngine& engine, FMaterial const* material,
     // expected by the per-material descriptor-set layout
     size_t const uboSize = std::max(size_t(16), material->getUniformInterfaceBlock().getSize());
     mUniforms = UniformBuffer(uboSize);
-    mUbHandle = driver.createBufferObject(mUniforms.getSize(), BufferObjectBinding::UNIFORM,
-            BufferUsage::STATIC, utils::ImmutableCString{ material->getName().c_str_safe() });
 
-    // set the UBO, always descriptor 0
-    mDescriptorSet.setBuffer(material->getDescriptorSetLayout(),
-        0, mUbHandle, 0, mUniforms.getSize());
+    if (mUseUboBatching) {
+        mUboData = BufferAllocator::UNALLOCATED;
+    } else {
+        mUboData = driver.createBufferObject(mUniforms.getSize(), BufferObjectBinding::UNIFORM,
+                BufferUsage::STATIC, ImmutableCString{ material->getName().c_str_safe() });
+        // set the UBO, always descriptor 0
+        mDescriptorSet.setBuffer(material->getDescriptorSetLayout(),
+                0, std::get<Handle<HwBufferObject>>(mUboData), 0, mUniforms.getSize());
+    }
 
     const RasterState& rasterState = material->getRasterState();
     // At the moment, only MaterialInstances have a stencil state, but in the future it should be
@@ -140,6 +150,7 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
           mHasScissor(false),
           mIsDoubleSided(other->mIsDoubleSided),
           mIsDefaultInstance(false),
+          mUseUboBatching(other->mUseUboBatching),
           mScissorRect(other->mScissorRect),
           mName(name ? CString(name) : other->mName) {
 
@@ -147,12 +158,16 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
     FMaterial const* const material = other->getMaterial();
 
     mUniforms.setUniforms(other->getUniformBuffer());
-    mUbHandle = driver.createBufferObject(mUniforms.getSize(), BufferObjectBinding::UNIFORM,
-            BufferUsage::DYNAMIC, ImmutableCString{ material->getName().c_str_safe() });
 
-    // set the UBO, always descriptor 0
-    mDescriptorSet.setBuffer(mMaterial->getDescriptorSetLayout(),
-            0, mUbHandle, 0, mUniforms.getSize());
+    if (mUseUboBatching) {
+        mUboData = BufferAllocator::UNALLOCATED;
+    } else {
+        mUboData = driver.createBufferObject(mUniforms.getSize(), BufferObjectBinding::UNIFORM,
+                BufferUsage::DYNAMIC, ImmutableCString{ material->getName().c_str_safe() });
+        // set the UBO, always descriptor 0
+        mDescriptorSet.setBuffer(material->getDescriptorSetLayout(),
+                0, std::get<Handle<HwBufferObject>>(mUboData), 0, mUniforms.getSize());
+    }
 
     if (material->hasDoubleSidedCapability()) {
         setDoubleSided(mIsDoubleSided);
@@ -173,7 +188,7 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
             material->getId(), material->generateMaterialInstanceId());
 
     // If the original descriptor set has been commited, the copy needs to commit as well.
-    if (other->mDescriptorSet.getHandle()) {
+    if (!mUseUboBatching && other->mDescriptorSet.getHandle()) {
         mDescriptorSet.commitSlow(mMaterial->getDescriptorSetLayout(), driver);
     }
 }
@@ -190,13 +205,16 @@ FMaterialInstance::~FMaterialInstance() noexcept = default;
 void FMaterialInstance::terminate(FEngine& engine) {
     FEngine::DriverApi& driver = engine.getDriverApi();
     mDescriptorSet.terminate(driver);
-    driver.destroyBufferObject(mUbHandle);
+    auto* ubHandle = std::get_if<Handle<HwBufferObject>>(&mUboData);
+    if (ubHandle){
+        driver.destroyBufferObject(*ubHandle);
+    }
 }
 
 void FMaterialInstance::commitStreamUniformAssociations(FEngine::DriverApi& driver) {
     mHasStreamUniformAssociations = false;
     if (!mTextureParameters.empty()) {
-        backend::BufferObjectStreamDescriptor descriptor;
+        BufferObjectStreamDescriptor descriptor;
         for (auto const& [binding, p]: mTextureParameters) {
             ssize_t offset = mMaterial->getUniformInterfaceBlock().getTransformFieldOffset(binding);
             if (offset >= 0) {
@@ -206,7 +224,12 @@ void FMaterialInstance::commitStreamUniformAssociations(FEngine::DriverApi& driv
             }
         }
         if (descriptor.mStreams.size() > 0) {
-            driver.registerBufferObjectStreams(mUbHandle, std::move(descriptor));
+            // UBO batching is incompatible with stream uniform associations because streams require a
+            // dedicated UBO handle, not a sub-allocation. Assert that any instance here uses its own UBO.
+            assert_invariant(!mUseUboBatching);
+            auto const* ubHandle = std::get_if<Handle<HwBufferObject>>(&mUboData);
+            assert_invariant(ubHandle);
+            driver.registerBufferObjectStreams(*ubHandle, std::move(descriptor));
         }
     }
 }
@@ -217,11 +240,17 @@ void FMaterialInstance::commit(FEngine& engine) const {
     }
 }
 
-void FMaterialInstance::commit(DriverApi& driver) const {
-    // update uniforms if needed
+void FMaterialInstance::commit(FEngine::DriverApi& driver) const {
     if (mUniforms.isDirty() || mHasStreamUniformAssociations) {
         mUniforms.clean();
-        driver.updateBufferObject(mUbHandle, mUniforms.toBufferDescriptor(driver), 0);
+        if (mUseUboBatching) {
+            // TODO: update the content by `copyToMemoryMappedBuffer`
+        }
+        else {
+            auto* ubHandle = std::get_if<Handle<HwBufferObject>>(&mUboData);
+            assert_invariant(ubHandle != nullptr);
+            driver.updateBufferObject(*ubHandle, mUniforms.toBufferDescriptor(driver), 0);
+        }
     }
     if (!mTextureParameters.empty()) {
         for (auto const& [binding, p]: mTextureParameters) {
@@ -411,6 +440,21 @@ void FMaterialInstance::use(FEngine::DriverApi& driver, Variant variant) const {
     mDescriptorSet.bind(driver, DescriptorSetBindingPoints::PER_MATERIAL);
 }
 
+void FMaterialInstance::assignUboAllocation(
+        const Handle<HwBufferObject>& ubHandle,
+        BufferAllocator::AllocationId id,
+        BufferAllocator::allocation_size_t offset) {
+    assert_invariant(mUseUboBatching);
+    mUboData = id;
+    mDescriptorSet.setBuffer(mMaterial->getDescriptorSetLayout(), 0, ubHandle, offset,
+            mUniforms.getSize());
+}
+
+BufferAllocator::AllocationId FMaterialInstance::getAllocationId() const noexcept {
+    auto const* allocationId = std::get_if<BufferAllocator::AllocationId>(&mUboData);
+    return allocationId ? *allocationId : BufferAllocator::UNALLOCATED;
+}
+
 void FMaterialInstance::fixMissingSamplers() const {
     // Here we check that all declared sampler parameters are set, this is required by
     // Vulkan and Metal; GL is more permissive. If a sampler parameter is not set, we will

filament/src/details/MaterialInstance.h

@@ -23,6 +23,7 @@
 
 #include "ds/DescriptorSet.h"
 
+#include "details/BufferAllocator.h"
 #include "details/Engine.h"
 
 #include "private/backend/DriverApi.h"
@@ -57,6 +58,9 @@ class FMaterialInstance : public MaterialInstance {
 public:
     FMaterialInstance(FEngine& engine, FMaterial const* material,
                       const char* name) noexcept;
+    // Use this constructor when you need to override the ubo batching flag for an individual MI.
+    FMaterialInstance(FEngine& engine, FMaterial const* material,
+                      const char* name, bool useUboBatching) noexcept;
     FMaterialInstance(FEngine& engine, FMaterialInstance const* other, const char* name);
     FMaterialInstance(const FMaterialInstance& rhs) = delete;
     FMaterialInstance& operator=(const FMaterialInstance& rhs) = delete;
@@ -75,6 +79,12 @@ public:
 
     void use(FEngine::DriverApi& driver, Variant variant = {}) const;
 
+    void assignUboAllocation(const backend::Handle<backend::HwBufferObject>& ubHandle,
+            BufferAllocator::AllocationId id,
+            BufferAllocator::allocation_size_t offset);
+
+    BufferAllocator::AllocationId getAllocationId() const noexcept;
+
     FMaterial const* getMaterial() const noexcept { return mMaterial; }
 
     uint64_t getSortingKey() const noexcept { return mMaterialSortingKey; }
@@ -234,6 +244,8 @@ public:
         return mIsDefaultInstance;
     }
 
+    bool isUsingUboBatching() const noexcept { return mUseUboBatching; }
+
     // Called by the engine to ensure that unset samplers are initialized with placedholders.
     void fixMissingSamplers() const;
 
@@ -274,7 +286,7 @@ private:
         backend::SamplerParams params;
     };
 
-    backend::Handle<backend::HwBufferObject> mUbHandle;
+    std::variant<BufferAllocator::AllocationId, backend::Handle<backend::HwBufferObject>> mUboData;
     tsl::robin_map<backend::descriptor_binding_t, TextureParameter> mTextureParameters;
     mutable DescriptorSet mDescriptorSet;
     UniformBuffer mUniforms;
@@ -296,6 +308,7 @@ private:
     bool mHasScissor : 1;
     bool mIsDoubleSided : 1;
     bool mIsDefaultInstance : 1;
+    bool mUseUboBatching : 1;
     TransparencyMode mTransparencyMode : 2;
 
     uint64_t mMaterialSortingKey = 0;

filament/test/CMakeLists.txt

@@ -43,6 +43,8 @@ list(APPEND RESGEN_SOURCE ${DUMMY_SRC})
 if (TNT_DEV)
     add_executable(test_${TARGET}
             filament_AtlasAllocator_test.cpp
+            test_BufferAllocator.cpp
+            test_BufferAllocatorStress.cpp
             test_CircularQueue.cpp
             filament_test_exposure.cpp
             filament_rendering_test.cpp

filament/test/test_BufferAllocator.cpp

@@ -0,0 +1,415 @@
+/*
+ * 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 <gtest/gtest.h>
+
+#include "../src/details/BufferAllocator.h"
+#include "utils/Panic.h"
+
+#include <utility>
+#include <vector>
+
+using namespace filament;
+
+namespace {
+
+class BufferAllocatorTest : public ::testing::Test {
+protected:
+    // We use a total size of 1024 and a slot size (alignment) of 64.
+    // This gives us 1024 / 64 = 16 total possible slots if aligned.
+    static constexpr BufferAllocator::allocation_size_t TOTAL_SIZE = 1024;
+    static constexpr BufferAllocator::allocation_size_t SLOT_SIZE = 64;
+
+    BufferAllocatorTest() : mAllocator(TOTAL_SIZE, SLOT_SIZE) {}
+
+    BufferAllocator mAllocator;
+};
+
+TEST_F(BufferAllocatorTest, ConstructorFailure) {
+    // The constructor requires slotSize to be a power of two.
+    constexpr BufferAllocator::allocation_size_t NON_POT_SLOT_SIZE = 60;
+    EXPECT_DEATH(BufferAllocator(TOTAL_SIZE, NON_POT_SLOT_SIZE), "failed assertion");
+}
+
+TEST_F(BufferAllocatorTest, InitialState) {
+    EXPECT_EQ(mAllocator.getTotalSize(), TOTAL_SIZE);
+    // Initially, there should be one large free block of the total size.
+    // Let's try to allocate the whole thing.
+    auto [id, offset] = mAllocator.allocate(TOTAL_SIZE);
+    EXPECT_EQ(id, 1); // The first ID should be 1.
+    EXPECT_EQ(offset, 0);
+}
+
+TEST_F(BufferAllocatorTest, SimpleAllocation) {
+    // Allocate 100 bytes, which should be aligned up to 128 (2 * 64).
+    auto [id, offset] = mAllocator.allocate(100);
+    EXPECT_EQ(id, 1);
+    EXPECT_EQ(offset, 0);
+    EXPECT_EQ(mAllocator.getAllocationOffset(id), offset);
+
+    // Try to allocate again. The next allocation should start after the first one.
+    auto [id2, offset2] = mAllocator.allocate(50); // Aligns to 64
+    EXPECT_EQ(id2, 3); // ID is (128 / 64) + 1 = 3
+    EXPECT_EQ(offset2, 128);
+    EXPECT_EQ(mAllocator.getAllocationOffset(id2), offset2);
+}
+
+TEST_F(BufferAllocatorTest, AllocateZeroSize) {
+    // Allocating zero bytes should return an unallocated ID and not affect the state.
+    auto [id, offset] = mAllocator.allocate(0);
+    EXPECT_EQ(id, BufferAllocator::UNALLOCATED);
+    EXPECT_EQ(offset, 0);
+
+    // The allocator should still be in its initial state, able to allocate the full size.
+    auto [fullId, fullOffset] = mAllocator.allocate(TOTAL_SIZE);
+    EXPECT_EQ(fullId, 1);
+    EXPECT_EQ(fullOffset, 0);
+}
+
+TEST_F(BufferAllocatorTest, AllocateAll) {
+    auto [id1, offset1] = mAllocator.allocate(512);
+    EXPECT_EQ(id1, 1);
+    EXPECT_EQ(offset1, 0);
+
+    auto [id2, offset2] = mAllocator.allocate(512);
+    EXPECT_EQ(id2, 9); // ID is (512 / 64) + 1 = 9
+    EXPECT_EQ(offset2, 512);
+
+    // The buffer is now full. The next allocation should fail.
+    auto [id3, offset3] = mAllocator.allocate(1);
+    EXPECT_EQ(id3, BufferAllocator::REALLOCATION_REQUIRED);
+}
+
+TEST_F(BufferAllocatorTest, AllocationFailure) {
+    auto [id1, offset1] = mAllocator.allocate(TOTAL_SIZE - 1); // Allocate most of the buffer.
+    EXPECT_EQ(id1, 1);
+    EXPECT_EQ(offset1, 0);
+
+    // Try to allocate more than the remaining space.
+    auto [id2, offset2] = mAllocator.allocate(100);
+    EXPECT_EQ(id2, BufferAllocator::REALLOCATION_REQUIRED);
+    EXPECT_EQ(offset2, 0);
+}
+
+TEST_F(BufferAllocatorTest, AllocationLifecycle) {
+    // 1. Allocate
+    auto [id, offset] = mAllocator.allocate(128);
+    EXPECT_EQ(id, 1);
+
+    // 2. Retire from CPU side
+    mAllocator.retire(id);
+
+    // 3. The slot is not free yet because the GPU might still be using it.
+    // Let's simulate the GPU acquiring and releasing it.
+    mAllocator.acquireGpu(id);
+    mAllocator.releaseGpu(id);
+
+    // Now the slot should be considered free, but it's not merged yet.
+    // Let's try to allocate something else to see where it goes.
+    auto [id2, offset2] = mAllocator.allocate(200); // Aligns to 256
+    EXPECT_EQ(id2, 3); // It should use the space after the first retired slot.
+    EXPECT_EQ(offset2, 128);
+
+    // Now, let's release the free slots.
+    mAllocator.releaseFreeSlots();
+
+    // The first slot (128 bytes) should now be available again.
+    // Let's try to allocate something that fits in it.
+    auto [id3, offset3] = mAllocator.allocate(100); // Aligns to 128
+    EXPECT_EQ(id3, 1); // It should reuse the first slot.
+    EXPECT_EQ(offset3, 0);
+}
+
+TEST_F(BufferAllocatorTest, RetireThenReleaseGpu) {
+    // 1. Allocate a block and a dummy block next to it.
+    auto [id1, offset1] = mAllocator.allocate(128);
+    auto [id2, offset2] = mAllocator.allocate(128);
+    EXPECT_EQ(id1, 1);
+    EXPECT_EQ(id2, 3);
+    EXPECT_EQ(offset1, 0);
+    EXPECT_EQ(offset2, 128);
+
+    // 2. Retire the first block (CPU is done), then acquire it for the GPU.
+    mAllocator.retire(id1);
+    mAllocator.acquireGpu(id1); // gpuUseCount = 1
+
+    // 3. The slot is now free from CPU but locked by GPU. releaseFreeSlots should do nothing to it.
+    mAllocator.releaseFreeSlots();
+
+    // 4. Try to allocate the same space. It should fail, and the allocation should go
+    //    to the next available free space.
+    auto [id3, offset3] = mAllocator.allocate(64);
+    EXPECT_EQ(id3, 5);
+    EXPECT_EQ(offset3, 256); // It should be allocated after id2.
+
+    // 5. Now, release the GPU lock.
+    mAllocator.releaseGpu(id1); // gpuUseCount = 0
+
+    // 6. Call releaseFreeSlots again. This time it should be freed.
+    mAllocator.releaseFreeSlots();
+
+    // 7. The original slot should now be available for allocation.
+    auto [id4, offset4] = mAllocator.allocate(64);
+    EXPECT_EQ(id4, 1);
+    EXPECT_EQ(offset4, 0); // Success! It reuses the first slot.
+}
+
+TEST_F(BufferAllocatorTest, MultipleGpuAcquires) {
+    // 1. Allocate a block.
+    auto [id, offset] = mAllocator.allocate(128);
+    EXPECT_EQ(id, 1);
+    EXPECT_EQ(offset, 0);
+
+    // 2. Retire from CPU, then acquire multiple times for GPU (e.g., used in 3 command buffers).
+    mAllocator.retire(id);
+    mAllocator.acquireGpu(id); // gpuUseCount = 1
+    mAllocator.acquireGpu(id); // gpuUseCount = 2
+    mAllocator.acquireGpu(id); // gpuUseCount = 3
+
+    // 3. Release GPU lock once. The slot should still be locked.
+    mAllocator.releaseGpu(id); // gpuUseCount = 2
+    mAllocator.releaseFreeSlots();
+    auto [failId1, failOffset1] = mAllocator.allocate(64);
+    EXPECT_NE(failOffset1, 0); // Should not be able to allocate at offset 0.
+    EXPECT_NE(failId1, 1);
+
+    // 4. Release GPU lock again. The slot should still be locked.
+    mAllocator.releaseGpu(id); // gpuUseCount = 1
+    mAllocator.releaseFreeSlots();
+    auto [failId2, failOffset2] = mAllocator.allocate(64);
+    EXPECT_NE(failOffset2, 0); // Still cannot allocate at offset 0.
+    EXPECT_NE(failId2, 1);
+
+    // 5. Final release. The lock count is now 0.
+    mAllocator.releaseGpu(id); // gpuUseCount = 0
+
+    // 6. Now it should be freed and available.
+    mAllocator.releaseFreeSlots();
+    auto [successId, successOffset] = mAllocator.allocate(64);
+    EXPECT_EQ(successOffset, 0);
+    EXPECT_EQ(successId, 1);
+}
+
+TEST_F(BufferAllocatorTest, GpuPanicOnUnderflow) {
+    // 1. Allocate a block and acquire it.
+    auto [id, _] = mAllocator.allocate(128);
+    mAllocator.acquireGpu(id);
+
+    // 2. Release it once, which is fine.
+    mAllocator.releaseGpu(id);
+
+    // 3. Releasing it again when the count is 0 should trigger a failed assertion.
+    EXPECT_DEATH(mAllocator.releaseGpu(id), "failed assertion");
+}
+
+TEST_F(BufferAllocatorTest, MergeFreeSlots) {
+    // Allocate three blocks
+    auto [id1, offset1] = mAllocator.allocate(128); // Slot 0-127
+    auto [id2, offset2] = mAllocator.allocate(128); // Slot 128-255
+    auto [id3, offset3] = mAllocator.allocate(128); // Slot 256-383
+
+    EXPECT_EQ(id1, 1);
+    EXPECT_EQ(id2, 3);
+    EXPECT_EQ(id3, 5);
+
+    EXPECT_EQ(offset1, 0);
+    EXPECT_EQ(offset2, 128);
+    EXPECT_EQ(offset3, 256);
+
+    // Retire the first and third blocks
+    mAllocator.retire(id1);
+    mAllocator.retire(id3);
+
+    // At this point, we have: [Free, Allocated, Free, Free (remaining)]
+    // releaseFreeSlots should not merge slot 1 and 3 because they are not adjacent.
+    // It should merge slot 3 and slot 4 instead.
+    mAllocator.releaseFreeSlots();
+
+    // At this point, we have: [Free, Allocated, Free (remaining)]
+    // Let's verify by trying to allocate 200. It should go into the merged slot.
+    auto [id4, offset4] = mAllocator.allocate(200); // Aligns to 256
+    EXPECT_EQ(id4, 5);
+    EXPECT_EQ(offset4, 256);
+
+    // Now, retire the middle block
+    mAllocator.retire(id2);
+
+    // Now we have: [Free, Free, Allocated, Free (remaining)]
+    // The first two blocks are now adjacent and free.
+    mAllocator.releaseFreeSlots();
+
+    // After merging, the first 256 bytes should be one large free block.
+    // Let's try to allocate something that requires this merged space.
+    auto [id5, offset5] = mAllocator.allocate(200); // Aligns to 256
+    EXPECT_EQ(id5, 1);
+    EXPECT_EQ(offset5, 0);
+}
+
+TEST_F(BufferAllocatorTest, MergeAllSlots) {
+    // Allocate the entire buffer in small chunks.
+    constexpr BufferAllocator::allocation_size_t CHUNK_SIZE = 128;
+    constexpr uint32_t NUM_CHUNKS = TOTAL_SIZE / CHUNK_SIZE;
+    std::vector<BufferAllocator::AllocationId> ids;
+    for (uint32_t i = 0; i < NUM_CHUNKS; ++i) {
+        auto [id, offset] = mAllocator.allocate(CHUNK_SIZE);
+        ASSERT_NE(id, BufferAllocator::REALLOCATION_REQUIRED);
+        ids.push_back(id);
+    }
+
+    // The buffer should be full.
+    auto [failId, failOffset] = mAllocator.allocate(1);
+    EXPECT_EQ(failId, BufferAllocator::REALLOCATION_REQUIRED);
+    EXPECT_EQ(failOffset, 0);
+
+    // Retire all chunks.
+    for (auto id : ids) {
+        mAllocator.retire(id);
+    }
+
+    // Release and merge.
+    mAllocator.releaseFreeSlots();
+
+    // Now, the allocator should be back to its initial state with one large free block.
+    // We should be able to allocate the entire buffer again.
+    auto [fullId, fullOffset] = mAllocator.allocate(TOTAL_SIZE);
+    EXPECT_EQ(fullId, 1);
+    EXPECT_EQ(fullOffset, 0);
+}
+
+TEST_F(BufferAllocatorTest, NoMergePossible) {
+    // Allocate blocks in an alternating pattern.
+    auto [id1, offset1] = mAllocator.allocate(128);
+    auto [id2, offset2] = mAllocator.allocate(128);
+    auto [id3, offset3] = mAllocator.allocate(128);
+    auto [id4, offset4] = mAllocator.allocate(128);
+    EXPECT_EQ(id1, 1);
+    EXPECT_EQ(id2, 3);
+    EXPECT_EQ(id3, 5);
+    EXPECT_EQ(id4, 7);
+    EXPECT_EQ(offset1, 0);
+    EXPECT_EQ(offset2, 128);
+    EXPECT_EQ(offset3, 256);
+    EXPECT_EQ(offset4, 384);
+
+    // Retire the first and third blocks, leaving allocated blocks in between.
+    mAllocator.retire(id1);
+    mAllocator.retire(id3);
+
+    // At this point, we have: [Free, Allocated, Free, Allocated, Free (remaining)]
+    mAllocator.releaseFreeSlots();
+
+    // Now, let's try to re-allocate the two 128-byte slots.
+    auto [id5, offset5] = mAllocator.allocate(128);
+    auto [id6, offset6] = mAllocator.allocate(128);
+    EXPECT_EQ(offset5, 0);
+    EXPECT_EQ(offset6, 256);
+}
+
+TEST_F(BufferAllocatorTest, Reset) {
+    auto [id1, offset1] = mAllocator.allocate(100);
+    auto [id2, offset2] = mAllocator.allocate(200);
+    EXPECT_EQ(id1, 1);
+    EXPECT_EQ(id2, 3);
+    EXPECT_EQ(offset1, 0);
+    EXPECT_EQ(offset2, 128);
+
+    // Reset the allocator to a new size.
+    mAllocator.reset(2048);
+    EXPECT_EQ(mAllocator.getTotalSize(), 2048);
+
+    // After reset, all previous allocations should be gone,
+    // and we should be able to allocate the entire new size.
+    auto [id3, offset3] = mAllocator.allocate(2048);
+    EXPECT_EQ(id3, 1);
+    EXPECT_EQ(offset3, 0);
+}
+
+TEST_F(BufferAllocatorTest, ResetWithInvalidSize) {
+    // Reset to a size which is not a power of two.
+    EXPECT_DEATH(mAllocator.reset(123), "failed assertion");
+}
+
+
+TEST_F(BufferAllocatorTest, ResetWithGpuLock) {
+    // 1. Allocate a block and acquire a GPU lock on it.
+    auto [id1, offset1] = mAllocator.allocate(128);
+    EXPECT_EQ(id1, 1);
+    mAllocator.acquireGpu(id1); // gpuUseCount = 1
+
+    // 2. Call reset. This should disregard the GPU lock and clear everything.
+    constexpr BufferAllocator::allocation_size_t NEW_TOTAL_SIZE = 4096;
+    mAllocator.reset(NEW_TOTAL_SIZE);
+
+    // 3. Verify the allocator is in a pristine state with the new size.
+    EXPECT_EQ(mAllocator.getTotalSize(), NEW_TOTAL_SIZE);
+
+    // 4. The strongest verification is to allocate the entire new size, which should
+    //    succeed, proving that the old GPU-locked block is gone.
+    auto [id2, offset2] = mAllocator.allocate(NEW_TOTAL_SIZE);
+    EXPECT_EQ(id2, 1);
+    EXPECT_EQ(offset2, 0);
+}
+
+TEST_F(BufferAllocatorTest, InvalidOperations) {
+    // These operations on invalid IDs should not crash and should be handled gracefully.
+    EXPECT_DEATH(mAllocator.retire(BufferAllocator::UNALLOCATED), "failed assertion");
+    EXPECT_DEATH(mAllocator.retire(999), "failed assertion"); // Non-existent ID
+
+    EXPECT_DEATH(mAllocator.acquireGpu(BufferAllocator::UNALLOCATED), "failed assertion");
+    EXPECT_DEATH(mAllocator.acquireGpu(999), "failed assertion");
+
+    EXPECT_DEATH(mAllocator.releaseGpu(BufferAllocator::UNALLOCATED), "failed assertion");
+    EXPECT_DEATH(mAllocator.releaseGpu(999), "failed assertion");
+
+    // Check that an invalid offset query panics in debug/testing builds.
+    EXPECT_DEATH(mAllocator.getAllocationOffset(BufferAllocator::UNALLOCATED), "failed assertion");
+    EXPECT_DEATH(mAllocator.getAllocationOffset(BufferAllocator::REALLOCATION_REQUIRED),
+            "failed assertion");
+}
+
+TEST_F(BufferAllocatorTest, ComplexScenario) {
+    std::vector<BufferAllocator::AllocationId> ids;
+    // 1. Allocate 4 blocks of 256 bytes
+    for (int i = 0; i < 4; ++i) {
+        auto [id, offset] = mAllocator.allocate(256);
+        EXPECT_EQ(offset, i * 256);
+        ids.push_back(id);
+    }
+
+    // Buffer should be full
+    auto [failId1, failOffset1] = mAllocator.allocate(1);
+    EXPECT_EQ(failId1, BufferAllocator::REALLOCATION_REQUIRED);
+
+    // 2. Retire the 2nd and 3rd blocks (ids[1] and ids[2])
+    mAllocator.retire(ids[1]);
+    mAllocator.retire(ids[2]);
+
+    // 3. Release free slots. This should merge the two retired blocks.
+    mAllocator.releaseFreeSlots();
+
+    // We now have a free block of 512 bytes at offset 256.
+    // Let's allocate 512 bytes. It should fit perfectly.
+    auto [id, offset] = mAllocator.allocate(512);
+    EXPECT_EQ(id, 5); // (256 / 64) + 1 = 5
+    EXPECT_EQ(offset, 256);
+
+    // 4. The buffer should be full again.
+    auto [failId2,failOffset2] = mAllocator.allocate(1);
+    EXPECT_EQ(failId2, BufferAllocator::REALLOCATION_REQUIRED);
+    EXPECT_EQ(failOffset2,0);
+}
+
+} // anonymous namespace

filament/test/test_BufferAllocatorStress.cpp

@@ -0,0 +1,122 @@
+/*
+ * 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 <gtest/gtest.h>
+
+#include "../src/details/BufferAllocator.h"
+#include "utils/Panic.h"
+
+#include <algorithm>
+#include <random>
+#include <vector>
+
+using namespace filament;
+
+namespace {
+
+class BufferAllocatorStressTest : public ::testing::Test {
+protected:
+    // We use a total size of 1024 * 64 and a slot size (alignment) of 64.
+    // This gives us 1024 total possible slots if aligned.
+    static constexpr BufferAllocator::allocation_size_t SLOT_SIZE = 64;
+    static constexpr BufferAllocator::allocation_size_t SLOT_COUNT = 4096;
+    static constexpr BufferAllocator::allocation_size_t TOTAL_SIZE = SLOT_COUNT * SLOT_SIZE;
+
+    BufferAllocatorStressTest() : mAllocator(TOTAL_SIZE, SLOT_SIZE) {
+    }
+
+    BufferAllocator mAllocator;
+};
+
+TEST_F(BufferAllocatorStressTest, StressTest) {
+    // Many operations.
+    constexpr int operationCount = 5000;
+
+    // 1. Prepare the random number distributions.
+    std::random_device rd;
+    std::mt19937 gen(rd());
+    std::uniform_int_distribution<> slotCountDistrib(1, SLOT_COUNT);
+    // Random the slot offset we allocate within a slot.
+    std::uniform_int_distribution<> slotOffsetDistrib(0, SLOT_SIZE - 1);
+    // Random the operation we perform: 0,1,2 -> allocate, 3 -> retire, 4 -> release slots
+    // We bias the operations to make more allocations than releases.
+    std::uniform_int_distribution operationDistrib(0, 4);
+
+    // 2. Randomly do some actions and expect to have no crash
+    std::vector<BufferAllocator::AllocationId> ids;
+    for (int i = 0; i < operationCount; i++) {
+        switch (operationDistrib(gen)) {
+            case 0: // Allocate
+            case 1: // Allocate
+            case 2: // Allocate
+            {
+                // Random the slot count + its offset.
+                const BufferAllocator::allocation_size_t size =
+                        slotCountDistrib(gen) * SLOT_SIZE - slotOffsetDistrib(gen);
+
+                if (auto [id, _] = mAllocator.allocate(size);
+                    id != BufferAllocator::REALLOCATION_REQUIRED && id !=
+                    BufferAllocator::UNALLOCATED) {
+                    ids.push_back(id);
+                }
+                break;
+            }
+            case 3: // Allocate
+            {
+                if (ids.empty()) {
+                    continue;
+                }
+
+                // Retire a random slot.
+                std::uniform_int_distribution<uint32_t> idDistrib(0, ids.size() - 1);
+                uint32_t indexToRetire = idDistrib(gen);
+
+                BufferAllocator::AllocationId idToRetire = ids[indexToRetire];
+                mAllocator.retire(idToRetire);
+
+                // Remove the retired id from the list.
+                std::swap(ids[indexToRetire], ids.back());
+                ids.pop_back();
+                break;
+            }
+            case 4: // Allocate
+            {
+                mAllocator.releaseFreeSlots();
+                break;
+            }
+            default:
+                break;
+        }
+    }
+
+    // 3. Retire all remaining allocations.
+    for (const auto& id: ids) {
+        mAllocator.retire(id);
+    }
+    ids.clear();
+
+    // 4. Release and merge everything.
+    mAllocator.releaseFreeSlots();
+
+    // 5. The allocator should now be in a pristine state.
+    //    A final allocation of the total size should succeed.
+    auto [finalId, finalOffset] = mAllocator.allocate(TOTAL_SIZE);
+    EXPECT_NE(finalId, BufferAllocator::REALLOCATION_REQUIRED);
+    EXPECT_NE(finalId, BufferAllocator::UNALLOCATED);
+    EXPECT_EQ(finalOffset, 0);
+}
+
+} // anonymous namespace

libs/filament-matp/src/MaterialParser.cpp

@@ -459,13 +459,6 @@ utils::Status MaterialParser::processTemplateSubstitutions(
 utils::Status MaterialParser::parse(filamat::MaterialBuilder& builder,
         const Config& config, ssize_t& size, std::unique_ptr<const char[]>& buffer) {
 
-    if (builder.getFeatureLevel() > config.getFeatureLevel()) {
-        utils::io::sstream errorMessage;
-        errorMessage << "Material feature level (" << +builder.getFeatureLevel()
-            << ") is higher than maximum allowed (" << +config.getFeatureLevel() << ")";
-        return utils::Status::invalidArgument(errorMessage.c_str());
-    }
-
     // Before attempting an expensive lex, let's find out if we were sent pure JSON.
     utils::Status parsedStatus;
     if (utils::Status validJson = isValidJsonStart(buffer.get(), size_t(size));
@@ -479,6 +472,13 @@ utils::Status MaterialParser::parse(filamat::MaterialBuilder& builder,
         return parsedStatus;
     }
 
+    if (builder.getFeatureLevel() > config.getFeatureLevel()) {
+        utils::io::sstream errorMessage;
+        errorMessage << "Material feature level (" << +builder.getFeatureLevel()
+            << ") is higher than maximum allowed (" << +config.getFeatureLevel() << ")";
+        return utils::Status::invalidArgument(errorMessage.c_str());
+    }
+
     switch (config.getReflectionTarget()) {
         case Config::Metadata::NONE:
             break;

samples/material_sandbox.cpp

@@ -191,6 +191,8 @@ static int handleCommandLineArgments(int argc, char* argv[], Config* config) {
 }
 
 static void cleanup(Engine* engine, View*, Scene*) {
+    g_meshSet.reset(nullptr);
+
     for (const auto& material : g_meshMaterialInstances) {
         engine->destroy(material.second);
     }
@@ -203,8 +205,6 @@ static void cleanup(Engine* engine, View*, Scene*) {
         engine->destroy(i);
     }
 
-    g_meshSet.reset(nullptr);
-
     engine->destroy(g_params.light);
     engine->destroy(g_params.spotLight);
     engine->destroy(g_colorGrading);

test/renderdiff/README.md

@@ -1,6 +1,6 @@
 # Rendering Difference Test
 
-We created a few scripts to run `gltf_viewer` and produce headless renderings.
+This tool is a collections of scripts to run `gltf_viewer` and produce headless renderings.
 
 This is mainly useful for continuous integration where GPUs are generally not available on cloud
 machines. To perform software rasterization, these scripts are centered around [Mesa]'s
@@ -9,10 +9,10 @@ Additionally, we should be able to use GPUs where available (though this is more
 work).
 
 The script `render.py` contains the core logic for taking input parameters (such as the test
-description file) and then running gltf_viewer to produce the renderings.
+description file) and then running `gltf_viewer` to produce the renderings.
 
 In the `test` directory is a list of test descriptions that are specified in json. Please see
-`sample.json` to parse the structure.
+`sample.json` to glean the structure.
 
 ## Setting up python
 

test/renderdiff/src/compare.py

@@ -229,7 +229,7 @@ if __name__ == '__main__':
   important_print(f'Successfully compared {success_count} / {len(results)} images')
 
   if tolerance_used_details:
-    pstr = 'Tolerance-based passes:'
+    pstr = 'Passed:'
     for detail in tolerance_used_details:
       pstr += '\n' + detail
     important_print(pstr)
@@ -237,7 +237,7 @@ if __name__ == '__main__':
   if failed_details:
     pstr = 'Failed:'
     for detail in failed_details:
-      pstr = '\n' + detail
+      pstr += '\n' + detail
     important_print(pstr)
   if len(failed) > 0:
     exit(1)

test/renderdiff/src/update_golden.py

@@ -74,7 +74,7 @@ def _get_deletes_updates(update_dir, golden_dir):
     for fpath in base.intersection(new):
       base_fpath = os.path.join(golden_dir, fpath)
       new_fpath = os.path.join(update_dir, fpath)
-      if (ext == 'tif' and not same_image(new_fpath, base_fpath)) or \
+      if (ext == 'tif' and not same_image(new_fpath, base_fpath)[0]) or \
          (ext == 'json' and _file_as_str(new_fpath) != _file_as_str(base_fpath)):
         update.append(fpath)
 

test/renderdiff/tests/gltf_models.txt

@@ -1,12 +1,13 @@
 AlphaBlendModeTest
 AttenuationTest
-Box
 BoomBoxWithAxes
+Box
 BoxInterleaved
 BoxTextured
 BoxTexturedNonPowerOfTwo
 Duck
 IridescenceSuzanne
+IridescentDishWithOlives
 Lantern
 MetalRoughSpheres
 NegativeScaleTest
@@ -17,4 +18,5 @@ Sponza
 Suzanne
 TextureCoordinateTest
 TextureSettingsTest
+TransmissionRoughnessTest
 TwoSidedPlane

test/renderdiff/tests/presubmit.json

@@ -5,9 +5,9 @@
     "presets": [
         {
             "name": "base",
-            "models": ["Box", "BoxTextured", "Duck", "lucy", "FlightHelmet"],
+            "models": ["Box", "BoxTextured", "FlightHelmet", "lucy"],
             "rendering": {
-                "viewer.cameraFocusDistance": 0,
+                "viewer.cameraFocalLength": 35.0,
                 "view.postProcessingEnabled": true,
                 "view.dithering": "NONE"
             },
@@ -17,27 +17,54 @@
             }
         },
         {
-            "name": "helmet_only",
+            "name": "bloom_models",
             "models": ["DamagedHelmet"],
             "rendering": {}
+        },
+        {
+            "name": "ssao_models",
+            "models": ["Duck", "lucy"],
+            "rendering": {}
+        },
+        {
+            "name": "transmission_models",
+            "models": ["TransmissionRoughnessTest", "IridescentDishWithOlives"],
+            "rendering": {}
         }
     ],
     "tests": [
         {
             "name": "Bloom",
-            "description": "Testing bloom",
-            "apply_presets": ["base", "helmet_only"],
+            "description": "Bloom",
+            "apply_presets": ["base", "bloom_models"],
             "rendering": {
                 "view.bloom.enabled": true
             }
         },
         {
             "name": "MSAA",
-            "description": "Testing multisampling anti-aliasing",
+            "description": "Multisampling anti-aliasing",
             "apply_presets": ["base"],
             "rendering": {
                 "view.msaa.enabled": true
             }
+        },
+        {
+            "name": "Transimssion",
+            "description": "transmission",
+            "apply_presets": ["base", "transmission_models"],
+            "rendering": {
+                "viewer.cameraFocalLength": 52.0,
+                "view.screenSpaceReflections.enabled": true
+            }
+        },
+        {
+            "name": "SSAO",
+            "description": "Screen space ambient occulsion",
+            "apply_presets": ["base", "ssao_models"],
+            "rendering": {
+                "view.ssao.enabled": true
+            }
         }
     ]
 }

third_party/perfetto/tnt/README.md

@@ -0,0 +1,12 @@
+To update the Perfetto library, run the following command from the tnt directory:
+```shell
+./update_perfetto.sh <version_tag>
+```
+
+For example:
+```shell
+./update_perfetto.sh v51.2
+```
+**Useful links:**
+https://perfetto.dev/
+https://perfetto.dev/docs/instrumentation/tracing-sdk

third_party/perfetto/tnt/README.txt

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

third_party/perfetto/tnt/update_perfetto.sh

@@ -0,0 +1,86 @@
+#!/bin/bash
+
+set -euo pipefail
+
+function usage() {
+  echo "Usage: $0 [-h] <version_tag>"
+  echo "  -h: Display this help message"
+  echo "  <version_tag>: The Perfetto version to download (e.g., v50.1)"
+  exit 1
+}
+
+if [[ $# -ne 1 || "$1" == "-h" ]]; then
+  usage
+fi
+
+DOWNLOAD_REF=$1
+DOWNLOAD_URL="https://github.com/google/perfetto/archive/refs/tags/${DOWNLOAD_REF}.zip"
+ZIP_FILE_NAME="${DOWNLOAD_REF}.zip"
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+THIRD_PARTY_DIR="$(dirname "$(dirname "$SCRIPT_DIR")")"
+
+if ! command -v curl &> /dev/null; then
+    echo "Error: curl is not installed." >&2
+    exit 1
+fi
+
+if ! command -v unzip &> /dev/null; then
+    echo "Error: unzip is not installed." >&2
+    exit 1
+fi
+
+if ! command -v rsync &> /dev/null; then
+    echo "Error: rsync is not installed." >&2
+    exit 1
+fi
+
+cd "${THIRD_PARTY_DIR}"
+
+echo "Downloading Perfetto ${DOWNLOAD_REF}..."
+if ! curl -f -L -o "${ZIP_FILE_NAME}" "${DOWNLOAD_URL}"; then
+    echo "Error: Failed to download Perfetto ${DOWNLOAD_REF}." >&2
+    echo "Please check the version number and your network connection." >&2
+    rm -f "${ZIP_FILE_NAME}"
+    exit 1
+fi
+
+echo "Unzipping..."
+
+TEMP_DIR="perfetto_temp_unzip"
+# Clean up temp dir from previous runs, if any.
+rm -rf "${TEMP_DIR}"
+mkdir "${TEMP_DIR}"
+
+if ! unzip -q "${ZIP_FILE_NAME}" -d "${TEMP_DIR}"; then
+    echo "Error: Failed to unzip the downloaded file." >&2
+    rm -f "${ZIP_FILE_NAME}"
+    rm -rf "${TEMP_DIR}"
+    exit 1
+fi
+
+# The archive contains a single directory. Find it.
+EXTRACTED_DIR=$(find "${TEMP_DIR}" -mindepth 1 -maxdepth 1 -type d)
+
+if [ -z "${EXTRACTED_DIR}" ] || [ ! -d "${EXTRACTED_DIR}" ]; then
+    echo "Error: Could not find the extracted directory inside the archive." >&2
+    rm -f "${ZIP_FILE_NAME}"
+    rm -rf "${TEMP_DIR}"
+    exit 1
+fi
+
+echo "Replacing existing perfetto contents with new version..."
+# If perfetto_new exists, remove it.
+rm -rf perfetto_new
+mv "${EXTRACTED_DIR}" perfetto_new
+
+# The perfetto directory should contain the content of the sdk directory from the archive, and our tnt directory.
+rsync -a --delete perfetto_new/sdk/ perfetto/perfetto/
+
+echo "Cleaning up..."
+rm -rf "${ZIP_FILE_NAME}" perfetto_new "${TEMP_DIR}"
+
+echo "Staging changes..."
+git add perfetto
+
+echo "Done. Please commit the changes."