Merge branch 'main' into exv/mi-sc

# Conflicts:
#	NEW_RELEASE_NOTES.md
#	filament/src/LocalProgramCache.cpp
#	filament/src/PostProcessManager.cpp
#	filament/src/PostProcessManager.h
#	filament/src/details/Material.cpp

CMakeLists.txt

@@ -691,6 +691,11 @@ if (FILAMENT_SUPPORTS_WEBGPU)
     set(MATC_API_FLAGS ${MATC_API_FLAGS} -a webgpu --variant-filter=stereo)
 endif()
 
+# Disable ESSL 1.0 code generation.
+if (NOT FILAMENT_ENABLE_FEATURE_LEVEL_0)
+    set(MATC_API_FLAGS ${MATC_API_FLAGS} -1)
+endif()
+
 # Enable debug info (preserves names in SPIR-V)
 if (FILAMENT_ENABLE_MATDBG)
     set(MATC_OPT_FLAGS ${MATC_OPT_FLAGS} -d)

android/filament-android/src/main/cpp/MaterialInstance.cpp

@@ -20,6 +20,8 @@
 #include <filament/Texture.h>
 #include <filament/TextureSampler.h>
 
+#include "common/CallbackUtils.h"
+
 #include <math/mat3.h>
 #include <math/mat4.h>
 #include <math/vec2.h>
@@ -256,6 +258,69 @@ Java_com_google_android_filament_MaterialInstance_nSetFloatParameterArray(JNIEnv
     env->ReleaseStringUTFChars(name_, name);
 }
 
+extern "C"
+JNIEXPORT jboolean JNICALL
+Java_com_google_android_filament_MaterialInstance_nGetConstantBool(JNIEnv *env, jclass,
+        jlong nativeMaterialInstance, jstring name_) {
+    MaterialInstance* instance = (MaterialInstance*) nativeMaterialInstance;
+    const char *name = env->GetStringUTFChars(name_, 0);
+    jboolean result = instance->getConstant<bool>(name);
+    env->ReleaseStringUTFChars(name_, name);
+    return result;
+}
+
+extern "C"
+JNIEXPORT jfloat JNICALL
+Java_com_google_android_filament_MaterialInstance_nGetConstantFloat(JNIEnv *env, jclass,
+        jlong nativeMaterialInstance, jstring name_) {
+    MaterialInstance* instance = (MaterialInstance*) nativeMaterialInstance;
+    const char *name = env->GetStringUTFChars(name_, 0);
+    jfloat result = instance->getConstant<float>(name);
+    env->ReleaseStringUTFChars(name_, name);
+    return result;
+}
+
+extern "C"
+JNIEXPORT jint JNICALL
+Java_com_google_android_filament_MaterialInstance_nGetConstantInt(JNIEnv *env, jclass,
+        jlong nativeMaterialInstance, jstring name_) {
+    MaterialInstance* instance = (MaterialInstance*) nativeMaterialInstance;
+    const char *name = env->GetStringUTFChars(name_, 0);
+    jint result = instance->getConstant<int32_t>(name);
+    env->ReleaseStringUTFChars(name_, name);
+    return result;
+}
+
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_MaterialInstance_nSetConstantBool(JNIEnv *env, jclass,
+        jlong nativeMaterialInstance, jstring name_, jboolean x) {
+    MaterialInstance* instance = (MaterialInstance*) nativeMaterialInstance;
+    const char *name = env->GetStringUTFChars(name_, 0);
+    instance->setConstant<bool>(name, x);
+    env->ReleaseStringUTFChars(name_, name);
+}
+
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_MaterialInstance_nSetConstantFloat(JNIEnv *env, jclass,
+        jlong nativeMaterialInstance, jstring name_, jfloat x) {
+    MaterialInstance* instance = (MaterialInstance*) nativeMaterialInstance;
+    const char *name = env->GetStringUTFChars(name_, 0);
+    instance->setConstant<float>(name, x);
+    env->ReleaseStringUTFChars(name_, name);
+}
+
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_MaterialInstance_nSetConstantInt(JNIEnv *env, jclass,
+        jlong nativeMaterialInstance, jstring name_, jint x) {
+    MaterialInstance* instance = (MaterialInstance*) nativeMaterialInstance;
+    const char *name = env->GetStringUTFChars(name_, 0);
+    instance->setConstant<int32_t>(name, x);
+    env->ReleaseStringUTFChars(name_, name);
+}
+
 // defined in TextureSampler.cpp
 namespace filament::JniUtils {
     TextureSampler from_long(jlong params) noexcept;
@@ -592,3 +657,17 @@ Java_com_google_android_filament_MaterialInstance_nGetTransparencyMode(JNIEnv*,
     MaterialInstance* instance = (MaterialInstance*) nativeMaterialInstance;
     return (jint) instance->getTransparencyMode();
 }
+
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_MaterialInstance_nCompile(JNIEnv *env, jclass clazz,
+        jlong nativeMaterialInstance, jint priority, jint variants, jobject handler, jobject runnable) {
+    MaterialInstance* materialInstance = (MaterialInstance*) nativeMaterialInstance;
+    JniCallback* jniCallback = JniCallback::make(env, handler, runnable);
+    materialInstance->compile(
+            (MaterialInstance::CompilerPriorityQueue) priority,
+            (UserVariantFilterBit) variants,
+            jniCallback->getHandler(), [jniCallback](MaterialInstance*){
+                JniCallback::postToJavaAndDestroy(jniCallback);
+            });
+}

android/filament-android/src/main/java/com/google/android/filament/Material.java

@@ -523,11 +523,19 @@ public class Material {
      * for stereoscopic rendering. If an application is not planning to render in stereo, this bit
      * should be turned off to avoid unnecessary material compilations.
      *</p>
+     *<p>
+     * Note that it is possible to override specialization constants on a per-MaterialInstance basis
+     * (see {@link MaterialInstance#setConstant}). In that case, the programs compiled by a call to
+     * Material::compile() may not be reusable by that MaterialInstance. It's better to call
+     * MaterialInstance::compile() in cases where you intend to override specialization constants.
+     *</p>
      * @param priority      Which priority queue to use, LOW or HIGH.
      * @param variants      Variants to include to the compile command.
      * @param handler       An {@link java.util.concurrent.Executor Executor}. On Android this can also be a {@link android.os.Handler Handler}.
      * @param callback      callback called on the main thread when the compilation is done on
      *                      by backend.
+     *
+     * @see MaterialInstance#compile
      */
     public void compile(@NonNull CompilerPriorityQueue priority,
                         int variants,

android/filament-android/src/main/java/com/google/android/filament/MaterialInstance.java

@@ -18,6 +18,7 @@ package com.google.android.filament;
 
 import androidx.annotation.IntRange;
 import androidx.annotation.NonNull;
+import androidx.annotation.Nullable;
 import androidx.annotation.Size;
 
 import com.google.android.filament.proguard.UsedByNative;
@@ -142,6 +143,28 @@ public class MaterialInstance {
         return mMaterial;
     }
 
+    /**
+     * Asynchronously ensures that a subset of this MaterialInstance's variants are compiled.
+     *
+     * <p>This function behaves identically to {@link Material#compile}, but takes into account
+     * the specific constants overridden by {@link #setConstant}.</p>
+     *
+     * @param priority      Priority of the compile command.
+     * @param variants      Variants to include to the compile command.
+     * @param handler       An {@link java.util.concurrent.Executor Executor}. On Android this can also be a {@link android.os.Handler Handler}.
+     * @param callback      callback called on the main thread when the compilation is done on
+     *                      by backend.
+     *
+     * @see Material#compile
+     * @see #setConstant
+     */
+    public void compile(@NonNull Material.CompilerPriorityQueue priority,
+                        int variants,
+                        @Nullable Object handler,
+                        @Nullable Runnable callback) {
+        nCompile(getNativeObject(), priority.ordinal(), variants, handler, callback);
+    }
+
     /** @return the name associated with this instance */
     @NonNull
     public String getName() {
@@ -420,6 +443,69 @@ public class MaterialInstance {
         nSetParameterFloat4(getNativeObject(), name, color[0], color[1], color[2], color[3]);
     }
 
+    /**
+     * Overrides a specialization constant of this material instance.
+     *
+     * @param name  The name of the constant as defined in the material.
+     * @param value The value of the constant.
+     * @see Material.Builder#constant
+     */
+    public void setConstant(@NonNull String name, boolean value) {
+        nSetConstantBool(getNativeObject(), name, value);
+    }
+
+    /**
+     * Overrides a specialization constant of this material instance.
+     *
+     * @param name  The name of the constant as defined in the material.
+     * @param value The value of the constant.
+     * @see Material.Builder#constant
+     */
+    public void setConstant(@NonNull String name, float value) {
+        nSetConstantFloat(getNativeObject(), name, value);
+    }
+
+    /**
+     * Overrides a specialization constant of this material instance.
+     *
+     * @param name  The name of the constant as defined in the material.
+     * @param value The value of the constant.
+     * @see Material.Builder#constant
+     */
+    public void setConstant(@NonNull String name, int value) {
+        nSetConstantInt(getNativeObject(), name, value);
+    }
+
+    /**
+     * Gets the value of a specialization constant by name.
+     *
+     * @param name  The name of the constant as defined in the material.
+     * @return The value of the constant.
+     */
+    public boolean getConstantBoolean(@NonNull String name) {
+        return nGetConstantBool(getNativeObject(), name);
+    }
+
+    /**
+     * Gets the value of a specialization constant by name.
+     *
+     * @param name  The name of the constant as defined in the material.
+     * @return The value of the constant.
+     */
+    public float getConstantFloat(@NonNull String name) {
+        return nGetConstantFloat(getNativeObject(), name);
+    }
+
+    /**
+     * Gets the value of a specialization constant by name.
+     *
+     * @param name  The name of the constant as defined in the material.
+     * @return The value of the constant.
+     */
+    public int getConstantInt(@NonNull String name) {
+        return nGetConstantInt(getNativeObject(), name);
+    }
+
     /**
      * Set-up a custom scissor rectangle; by default it is disabled.
      *
@@ -955,6 +1041,17 @@ public class MaterialInstance {
             @NonNull String name, int element, @NonNull @Size(min = 1) float[] v,
             @IntRange(from = 0) int offset, @IntRange(from = 1) int count);
 
+    private static native boolean nGetConstantBool(long nativeMaterialInstance, @NonNull String name);
+    private static native float nGetConstantFloat(long nativeMaterialInstance, @NonNull String name);
+    private static native int nGetConstantInt(long nativeMaterialInstance, @NonNull String name);
+
+    private static native void nSetConstantBool(long nativeMaterialInstance,
+            @NonNull String name, boolean x);
+    private static native void nSetConstantFloat(long nativeMaterialInstance,
+            @NonNull String name, float x);
+    private static native void nSetConstantInt(long nativeMaterialInstance,
+            @NonNull String name, int x);
+
     private static native void nSetParameterTexture(long nativeMaterialInstance,
             @NonNull String name, long nativeTexture, long sampler);
 
@@ -1018,4 +1115,5 @@ public class MaterialInstance {
     private static native int nGetDepthFunc(long nativeMaterialInstance);
     private static native void nSetTransparencyMode(long nativeMaterialInstance, int mode);
     private static native int nGetTransparencyMode(long nativeMaterialInstance);
+    private static native void nCompile(long nativeMaterialInstance, int priority, int variants, Object handler, Runnable callback);
 }

build.sh

@@ -3,7 +3,7 @@ set -e
 
 # Host tools required by Android, WASM, and iOS builds
 MOBILE_HOST_TOOLS="matc resgen cmgen filamesh uberz"
-WEB_HOST_TOOLS="${MOBILE_HOST_TOOLS} mipgen filamesh glslminifier"
+WEB_HOST_TOOLS="${MOBILE_HOST_TOOLS} mipgen filamesh"
 
 function print_help {
     local self_name=$(basename "$0")

build/common/get-mesa.sh

@@ -134,23 +134,7 @@ pushd .
 cd ${MESA_DIR}
 
 # Need >= 24 to have llvmpipe for swrast.  llvmpipe is needed for GL >= 4.1.
-git checkout -f mesa-${MESA_VERSION}
-
-# Apply custom patch to fix a double-free in OSMesa
-git apply << 'EOF'
-diff --git a/src/mesa/program/program.c b/src/mesa/program/program.c
-index 74bd6a6c33b..a70814e53a1 100644
---- a/src/mesa/program/program.c
-+++ b/src/mesa/program/program.c
-@@ -130,6 +130,7 @@ _mesa_free_program_data(struct gl_context *ctx)
-       ctx->ATIFragmentShader.Current->RefCount--;
-       if (ctx->ATIFragmentShader.Current->RefCount <= 0) {
-          free(ctx->ATIFragmentShader.Current);
-+         ctx->ATIFragmentShader.Current = NULL;
-       }
-    }
- 
-EOF
+git checkout mesa-${MESA_VERSION}
 
 mkdir -p out
 

build/common/upload-release-assets/package-lock.json

@@ -13,20 +13,12 @@
       }
     },
     "node_modules/@actions/core": {
-      "version": "1.11.1",
-      "resolved": "https://registry.npmjs.org/@actions/core/-/core-1.11.1.tgz",
-      "integrity": "sha512-hXJCSrkwfA46Vd9Z3q4cpEpHB1rL5NG04+/rbqW9d3+CSvtB1tYe8UTpAlixa1vj0m/ULglfEK2UKxMGxCxv5A==",
+      "version": "1.10.0",
+      "resolved": "https://registry.npmjs.org/@actions/core/-/core-1.10.0.tgz",
+      "integrity": "sha512-2aZDDa3zrrZbP5ZYg159sNoLRb61nQ7awl5pSvIq5Qpj81vwDzdMRKzkWJGJuwVvWpvZKx7vspJALyvaaIQyug==",
       "dependencies": {
-        "@actions/exec": "^1.1.1",
-        "@actions/http-client": "^2.0.1"
-      }
-    },
-    "node_modules/@actions/exec": {
-      "version": "1.1.1",
-      "resolved": "https://registry.npmjs.org/@actions/exec/-/exec-1.1.1.tgz",
-      "integrity": "sha512-+sCcHHbVdk93a0XT19ECtO/gIXoxvdsgQLzb2fE2/5sIZmWQuluYyjPQtrtTHdU1YzTZ7bAPN4sITq2xi1679w==",
-      "dependencies": {
-        "@actions/io": "^1.0.1"
+        "@actions/http-client": "^2.0.1",
+        "uuid": "^8.3.2"
       }
     },
     "node_modules/@actions/glob": {
@@ -46,11 +38,6 @@
         "tunnel": "^0.0.6"
       }
     },
-    "node_modules/@actions/io": {
-      "version": "1.1.3",
-      "resolved": "https://registry.npmjs.org/@actions/io/-/io-1.1.3.tgz",
-      "integrity": "sha512-wi9JjgKLYS7U/z8PPbco+PvTb/nRWjeoFlJ1Qer83k/3C5PHQi28hiVdeE2kHXmIL99mQFawx8qt/JPjZilJ8Q=="
-    },
     "node_modules/@octokit/auth-token": {
       "version": "5.1.2",
       "resolved": "https://registry.npmjs.org/@octokit/auth-token/-/auth-token-5.1.2.tgz",
@@ -270,24 +257,24 @@
       "resolved": "https://registry.npmjs.org/universal-user-agent/-/universal-user-agent-7.0.2.tgz",
       "integrity": "sha512-0JCqzSKnStlRRQfCdowvqy3cy0Dvtlb8xecj/H8JFZuCze4rwjPZQOgvFvn0Ws/usCHQFGpyr+pB9adaGwXn4Q==",
       "license": "ISC"
+    },
+    "node_modules/uuid": {
+      "version": "8.3.2",
+      "resolved": "https://registry.npmjs.org/uuid/-/uuid-8.3.2.tgz",
+      "integrity": "sha512-+NYs2QeMWy+GWFOEm9xnn6HCDp0l7QBD7ml8zLUmJ+93Q5NF0NocErnwkTkXVFNiX3/fpC6afS8Dhb/gz7R7eg==",
+      "bin": {
+        "uuid": "dist/bin/uuid"
+      }
     }
   },
   "dependencies": {
     "@actions/core": {
-      "version": "1.11.1",
-      "resolved": "https://registry.npmjs.org/@actions/core/-/core-1.11.1.tgz",
-      "integrity": "sha512-hXJCSrkwfA46Vd9Z3q4cpEpHB1rL5NG04+/rbqW9d3+CSvtB1tYe8UTpAlixa1vj0m/ULglfEK2UKxMGxCxv5A==",
+      "version": "1.10.0",
+      "resolved": "https://registry.npmjs.org/@actions/core/-/core-1.10.0.tgz",
+      "integrity": "sha512-2aZDDa3zrrZbP5ZYg159sNoLRb61nQ7awl5pSvIq5Qpj81vwDzdMRKzkWJGJuwVvWpvZKx7vspJALyvaaIQyug==",
       "requires": {
-        "@actions/exec": "^1.1.1",
-        "@actions/http-client": "^2.0.1"
-      }
-    },
-    "@actions/exec": {
-      "version": "1.1.1",
-      "resolved": "https://registry.npmjs.org/@actions/exec/-/exec-1.1.1.tgz",
-      "integrity": "sha512-+sCcHHbVdk93a0XT19ECtO/gIXoxvdsgQLzb2fE2/5sIZmWQuluYyjPQtrtTHdU1YzTZ7bAPN4sITq2xi1679w==",
-      "requires": {
-        "@actions/io": "^1.0.1"
+        "@actions/http-client": "^2.0.1",
+        "uuid": "^8.3.2"
       }
     },
     "@actions/glob": {
@@ -307,11 +294,6 @@
         "tunnel": "^0.0.6"
       }
     },
-    "@actions/io": {
-      "version": "1.1.3",
-      "resolved": "https://registry.npmjs.org/@actions/io/-/io-1.1.3.tgz",
-      "integrity": "sha512-wi9JjgKLYS7U/z8PPbco+PvTb/nRWjeoFlJ1Qer83k/3C5PHQi28hiVdeE2kHXmIL99mQFawx8qt/JPjZilJ8Q=="
-    },
     "@octokit/auth-token": {
       "version": "5.1.2",
       "resolved": "https://registry.npmjs.org/@octokit/auth-token/-/auth-token-5.1.2.tgz",
@@ -462,6 +444,11 @@
       "version": "7.0.2",
       "resolved": "https://registry.npmjs.org/universal-user-agent/-/universal-user-agent-7.0.2.tgz",
       "integrity": "sha512-0JCqzSKnStlRRQfCdowvqy3cy0Dvtlb8xecj/H8JFZuCze4rwjPZQOgvFvn0Ws/usCHQFGpyr+pB9adaGwXn4Q=="
+    },
+    "uuid": {
+      "version": "8.3.2",
+      "resolved": "https://registry.npmjs.org/uuid/-/uuid-8.3.2.tgz",
+      "integrity": "sha512-+NYs2QeMWy+GWFOEm9xnn6HCDp0l7QBD7ml8zLUmJ+93Q5NF0NocErnwkTkXVFNiX3/fpC6afS8Dhb/gz7R7eg=="
     }
   }
 }

filament/backend/include/backend/platforms/PlatformOSMesa.h

@@ -75,8 +75,7 @@ private:
         OSMesaContext context;
         std::unique_ptr<uint8_t[]> buffer;
     };
-    using ContextMap = std::unordered_map<std::thread::id, ContextInfo>;
-    ContextMap mAdditionalContexts;
+    std::unordered_map<std::thread::id, ContextInfo> mAdditionalContexts;
     mutable std::shared_mutex mAdditionalContextsLock;
 };
 

filament/backend/src/opengl/OpenGLDriver.cpp

@@ -403,9 +403,6 @@ void OpenGLDriver::terminate() {
     if (getJobWorker()) {
         getJobWorker()->terminate();
     }
-    // wait for the GPU again because JobWorker might have queued more work.
-    glFinish();
-
     if constexpr (UTILS_HAS_THREADING) {
         stopServiceThread();
     }

filament/backend/src/opengl/platforms/PlatformOSMesa.cpp

@@ -20,7 +20,6 @@
 #include <utils/Panic.h>
 #include <utils/ThreadUtils.h>
 
-#include <algorithm>
 #include <dlfcn.h>
 #include <memory>
 
@@ -204,7 +203,6 @@ void PlatformOSMesa::createContext(bool shared) {
 void PlatformOSMesa::releaseContext() noexcept {
     std::thread::id currentThreadId = utils::ThreadUtils::getThreadId();
     OSMesaContext context = nullptr;
-    std::unique_ptr<uint8_t[]> buffer;
 
     {
         std::lock_guard<std::shared_mutex> lock(mAdditionalContextsLock);
@@ -214,15 +212,16 @@ void PlatformOSMesa::releaseContext() noexcept {
             return;
         }
         context = it->second.context;
-        buffer = std::move(it->second.buffer);
-        mAdditionalContexts.erase(it);
     }
 
     OSMesaAPI* api = (OSMesaAPI*) mOsMesaApi;
-    if (api) {
-        // Passing NULL as the context is the standard way to unbind in OSMesa.
-        api->fOSMesaMakeCurrent(NULL, NULL, 0, 0, 0);
-        api->fOSMesaDestroyContext(context);
+    // Passing NULL as the context is the standard way to unbind in OSMesa.
+    api->fOSMesaMakeCurrent(NULL, NULL, 0, 0, 0);
+    api->fOSMesaDestroyContext(context);
+
+    {
+        std::lock_guard<std::shared_mutex> lock(mAdditionalContextsLock);
+        mAdditionalContexts.erase(currentThreadId);
     }
 }
 

filament/backend/src/webgpu/WebGPUProgram.cpp

@@ -43,6 +43,101 @@ namespace filament::backend {
 
 namespace {
 
+/**
+ * Given the source code of a [WGSL] WebGPU shader in text (string view to it) and the constants
+ * to be overwritten now (at runtime), returns updated [WGSL] shader code text, where the constants
+ * have been replaced.
+ * @param shaderLabel Something to call this shader for troubleshooting, error messaging, etc.
+ * @param shaderSource The original WGSL WebGPU shader code as text to be processed. This is a
+ *                     view to that text and this function does not change it; it is immutable.
+ *                     Instead, the processed version of this source is returned by the function.
+ * @param specConstants The constants to replace in the shader code, indexed by constant id.
+ * @return Processed version of the WGSL WebGPU shader code provided, where the constants have
+ * been replaced with the values provided by the specConstants parameter.
+ */
+[[nodiscard]] std::string replaceSpecConstants(std::string_view shaderLabel,
+        std::string_view shaderSource,
+        utils::FixedCapacityVector<Program::SpecializationConstant> const& specConstants) {
+    // this function is not expected to be called at all if no spec constants are to be replaced
+    assert_invariant(!specConstants.empty());
+    static constexpr std::string_view specConstantPrefix = "FILAMENT_SPEC_CONST_";
+    static constexpr size_t specConstantPrefixSize = specConstantPrefix.size();
+    const char* const sourceData = shaderSource.data();
+    std::stringstream processedShaderSource{};
+    size_t pos = 0;
+    while (pos < shaderSource.size()) {
+        const size_t posOfNextSpecConstant = shaderSource.find(specConstantPrefix, pos);
+        if (posOfNextSpecConstant == std::string::npos) {
+            // no more spec constants, so just stream the rest of the source code string
+            processedShaderSource << std::string_view(sourceData + pos, shaderSource.size() - pos);
+            break;
+        }
+        const size_t posOfId = posOfNextSpecConstant + specConstantPrefixSize;
+        const size_t posAfterId = shaderSource.find('_', posOfId);
+        FILAMENT_CHECK_POSTCONDITION(posAfterId != std::string::npos)
+                << "malformed " << shaderLabel << ". Found spec constant prefix '"
+                << specConstantPrefix << "' without an id or '_' after it.";
+        const std::string_view idStr =
+                std::string_view(sourceData + posOfId, posAfterId - posOfId);
+        const size_t posEndOfStatement = shaderSource.find(';', posAfterId);
+        FILAMENT_CHECK_POSTCONDITION(posEndOfStatement != std::string::npos)
+                << "malformed " << shaderLabel << ". Found spec constant assignment with id "
+                << idStr << " without a terminating ';' character?";
+        // this is a view into part of the statement, from after the id to the ';'
+        const std::string_view statementSegment =
+                std::string_view(sourceData + posAfterId, posEndOfStatement - posAfterId);
+        size_t posOfEqual = statementSegment.find('=');
+        if (posOfEqual == std::string::npos) {
+            // Not an assignment, so we don't need to replace it.
+            processedShaderSource << std::string_view(sourceData + pos,
+                    posEndOfStatement + 1 - pos);
+            pos = posEndOfStatement + 1;
+            continue;
+        }
+        posOfEqual += posAfterId; // position in original source overall, not just the segment
+        int constantId = 0;
+        char* endPtr;
+        errno = 0; // Clear errno before the call
+
+        long tempConstantId = std::strtol(idStr.data(), &endPtr, 10);
+
+        // Check for conversion errors
+        if (endPtr == idStr.data() || *endPtr != '\0' || errno == ERANGE) {
+            // Parsing failed, or conversion was out of range for 'long'
+            // or the string contained non-numeric characters after the number.
+            PANIC_POSTCONDITION("Invalid spec constant id '%s' in %s (not a valid integer or out "
+                                "of range for 'long'?)",
+                    idStr.data(), shaderLabel.data());
+        } else {
+            // Check if the parsed long value fits into an int
+            if (tempConstantId > std::numeric_limits<int>::max() ||
+                    tempConstantId < std::numeric_limits<int>::min()) {
+                PANIC_POSTCONDITION(
+                        "Invalid spec constant id '%s' in %s (value out of range for 'int')",
+                        idStr.data(), shaderLabel.data());
+            } else {
+                constantId = static_cast<int>(tempConstantId);
+            }
+        }
+        const std::variant<int32_t, float, bool> newValue = specConstants[constantId];
+        // stream up to the equal sign...
+        processedShaderSource << std::string_view(sourceData + pos, posOfEqual + 1 - pos);
+        // stream the new value...
+        if (auto* v = std::get_if<int32_t>(&newValue)) {
+            processedShaderSource << " " << *v << "i";
+        } else if (auto* f = std::get_if<float>(&newValue)) {
+            processedShaderSource << " " << *f << "f";
+        } else if (auto* b = std::get_if<bool>(&newValue)) {
+            processedShaderSource << " " << ((*b) ? "true" : "false");
+        }
+        // end the statement...
+        processedShaderSource << ";";
+        // and skip to after the end of the statement in the original source and continue...
+        pos = posEndOfStatement + 1;
+    }
+    return processedShaderSource.str();
+}
+
 /**
  * Creates a WebGPU shader module for a given "program" "stage", accounting for override constants.
  * Effectively, this function is responsible for preprocessing the shader source and compiling it.
@@ -50,12 +145,14 @@ namespace {
  * a shader module
  * @param program The "program" to compile/create the shader, which includes the shader source
  * @param stage The stage (e.g. vertex, fragment, etc.) to create the shader module
+ * @param specConstants Override constants to apply when creating/compiling the shader module.
  * @return the proper WebGPU shader module compiled/created from the input parameters. This might
  * wrap a null handle if the shader is not present (if the shader source is empty), such as
  * a missing fragment or compute shader.
  */
 [[nodiscard]] wgpu::ShaderModule createShaderModule(wgpu::Device const& device,
-        Program const& program, const ShaderStage stage) {
+        Program const& program, const ShaderStage stage,
+        utils::FixedCapacityVector<Program::SpecializationConstant> const& specConstants) {
     const char* const programName = program.getName().c_str_safe();
     std::array<utils::FixedCapacityVector<uint8_t>, Program::SHADER_TYPE_COUNT> const&
             shaderSource = program.getShadersSource();
@@ -67,9 +164,13 @@ namespace {
     std::stringstream labelStream;
     labelStream << programName << " " << filamentShaderStageToString(stage) << " shader";
     const auto label = labelStream.str();
-
+    const std::string processedShaderSource =
+            specConstants.empty()
+                    ? reinterpret_cast<const char*>(sourceBytes.data())
+                    : replaceSpecConstants(label, reinterpret_cast<const char*>(sourceBytes.data()),
+                              specConstants);
     wgpu::ShaderSourceWGSL wgslDescriptor{};
-    wgslDescriptor.code = wgpu::StringView(reinterpret_cast<const char*>(sourceBytes.data()));
+    wgslDescriptor.code = wgpu::StringView(processedShaderSource);
     const wgpu::ShaderModuleDescriptor descriptor{
         .nextInChain = &wgslDescriptor,
         .label = label.data()
@@ -150,11 +251,14 @@ namespace {
 }// namespace
 
 WebGPUProgram::WebGPUProgram(wgpu::Device const& device, Program const& program)
-        : HwProgram{ program.getName() } {
+    : HwProgram{ program.getName() } {
+    utils::FixedCapacityVector<Program::SpecializationConstant> const& specConstants =
+            program.getSpecializationConstants();
     // TODO: Consider creating/compiling these shaders in parallel.
-    vertexShaderModule = createShaderModule(device, program, ShaderStage::VERTEX);
-    fragmentShaderModule = createShaderModule(device, program, ShaderStage::FRAGMENT);
-    computeShaderModule = createShaderModule(device, program, ShaderStage::COMPUTE);
+    vertexShaderModule = createShaderModule(device, program, ShaderStage::VERTEX, specConstants);
+    fragmentShaderModule =
+            createShaderModule(device, program, ShaderStage::FRAGMENT, specConstants);
+    computeShaderModule = createShaderModule(device, program, ShaderStage::COMPUTE, specConstants);
 }
 
 }// namespace filament::backend

filament/backend/src/webgpu/WebGPUVertexBufferInfo.cpp

@@ -225,73 +225,42 @@ struct AttributeInfo final {
  * associated with a vertex buffer, and the slot each one belongs
  */
 void createBufferLayoutsBindingSlotsAndAttributeInfos(AttributeArray const& attributes,
-        const uint32_t deviceMaxVertexBuffers,
+        const uint8_t attributeCount, const uint32_t deviceMaxVertexBuffers,
         std::vector<WebGPUVertexBufferInfo::WebGPUSlotBindingInfo>& outWebGPUSlotBindingInfos,
         std::vector<wgpu::VertexBufferLayout>& outVertexBufferLayouts,
-        std::array<AttributeInfo, MAX_VERTEX_ATTRIBUTE_COUNT>& outAttributeInfos,
-        uint8_t& outActualAttributeCount) {
+        std::array<AttributeInfo, MAX_VERTEX_ATTRIBUTE_COUNT>& outAttributeInfos) {
     uint8_t currentWebGPUSlotIndex = 0;
     uint8_t currentAttributeIndex = 0;
     outVertexBufferLayouts.reserve(MAX_VERTEX_BUFFER_COUNT);
     outWebGPUSlotBindingInfos.reserve(outVertexBufferLayouts.capacity());
-
-    // Find a valid attribute to use as a dummy for unused slots (typically POSITION at index 0)
-    Attribute dummyAttribute{};
-    bool hasDummy = false;
-    for (uint32_t i = 0; i < attributes.size(); ++i) {
-        if (attributes[i].buffer != Attribute::BUFFER_UNUSED) {
-            dummyAttribute = attributes[i];
-            hasDummy = true;
-            break;
-        }
-    }
-
-    /*
-     * WebGPU Strict Validation Workaround:
-     *
-     * 1. Completeness: WebGPU requires ALL shader-declared attributes to exist in the Pipeline's
-     *    VertexState, even if unused by a specific mesh variant (e.g., BONE_WEIGHTS).
-     * 2. Stride Boundary: Offset + Format_Size <= arrayStride.
-     *
-     * If we omit an unused attribute, WebGPU crashes (violates 1).
-     * If we pad it as Float32x4 (16 bytes) into Slot 0 (e.g., Position, stride 12), WebGPU crashes
-     * (violates 2):
-     *
-     *   [ Slot 0 Stride: 12 bytes ]
-     *   |-------POSITION--------|
-     *   |------ BONE_WEIGHTS (Float32x4 = 16b) -----| <-- OVERFLOWS STRIDE!
-     *
-     * FIX: We alias unused attributes into Slot 0 as a 4-byte format (Float32 or Uint32).
-     *
-     *   [ Slot 0 Stride: 12 bytes ]
-     *   |-------POSITION--------|
-     *   |-BONE-| <-- (Float32 = 4b) Safely fits inside the 12-byte stride!
-     *
-     * WebGPU safely promotes the 4-byte Float32 into a WGSL vec4<f32> as (x, 0.0, 0.0, 1.0).
-     */
     for (uint32_t attributeIndex = 0; attributeIndex < attributes.size(); ++attributeIndex) {
-        auto attribute = attributes[attributeIndex];
-        wgpu::VertexFormat vertexFormat;
-
+        const auto& attribute = attributes[attributeIndex];
         if (attribute.buffer == Attribute::BUFFER_UNUSED) {
-            if (!hasDummy) {
-                continue;
-            }
-            // HACK: Re-use the dummy buffer for disabled attributes to satisfy WebGPU validation.
-            // Filament's shaders expect vec4 or uvec4. We provide a dummy format.
-            const bool isInteger = attribute.flags & Attribute::FLAG_INTEGER_TARGET;
-            vertexFormat = isInteger ? wgpu::VertexFormat::Uint8x4 : wgpu::VertexFormat::Unorm8x4;
-            attribute = dummyAttribute;
-        } else {
-            const bool isInteger = attribute.flags & Attribute::FLAG_INTEGER_TARGET;
-            const bool isNormalized = attribute.flags & Attribute::FLAG_NORMALIZED;
-            vertexFormat = getVertexFormat(attribute.type, isNormalized, isInteger);
+            // We ignore "unused" attributes, ones not associated with a buffer. If a shader
+            // references such an attribute we have a bug one way or another. Either the API/CPU
+            // will produce an error (best case scenario), where an attribute is referenced in the
+            // shader but not provided by the backend API (CPU-side), or the shader would be getting
+            // junk/undefined data from the GPU, since we do not have a valid buffer of data to
+            // provide to the shader/GPU.
+            continue;
         }
-
+        const bool isInteger = attribute.flags & Attribute::FLAG_INTEGER_TARGET;
+        const bool isNormalized = attribute.flags & Attribute::FLAG_NORMALIZED;
+        const wgpu::VertexFormat vertexFormat =
+                getVertexFormat(attribute.type, isNormalized, isInteger);
         uint8_t existingSlot = INVALID_SLOT_INDEX;
         for (uint32_t slot = 0; slot < currentWebGPUSlotIndex; slot++) {
             WebGPUVertexBufferInfo::WebGPUSlotBindingInfo const& info =
                     outWebGPUSlotBindingInfos[slot];
+            // We consider attributes to be in the same slot/layout only if they belong to the
+            // same buffer and are interleaved; they cannot belong to separate partitions in the
+            // same buffer, for example.
+            // For the attributes to be interleaved, the stride must match (among other things).
+            // The attribute offset being less than the slot's/layout's buffer offset indicates
+            // that it is in a separate partition before this slot/layout, thus not part of it.
+            // The difference from the attribute's offset and this slot's/layout's buffer offset
+            // being greater than the stride indicates it is in a separate partition after this
+            // slot/layout, thus not part of it.
             if (info.sourceBufferIndex == attribute.buffer && info.stride == attribute.stride &&
                     attribute.offset >= info.bufferOffset &&
                     ((attribute.offset - info.bufferOffset) < attribute.stride)) {
@@ -300,18 +269,26 @@ void createBufferLayoutsBindingSlotsAndAttributeInfos(AttributeArray const& attr
             }
         }
         if (existingSlot == INVALID_SLOT_INDEX) {
+            // New combination, use a new WebGPU slot
             FILAMENT_CHECK_PRECONDITION(currentWebGPUSlotIndex < MAX_VERTEX_BUFFER_COUNT &&
                                         currentWebGPUSlotIndex < deviceMaxVertexBuffers)
                     << "Number of vertex buffer layouts must not exceed MAX_VERTEX_BUFFER_COUNT ("
                     << MAX_VERTEX_BUFFER_COUNT << ") or the device limit ("
                     << deviceMaxVertexBuffers << ")";
             existingSlot = currentWebGPUSlotIndex++;
-            outWebGPUSlotBindingInfos.push_back({ .sourceBufferIndex = attribute.buffer,
+            outWebGPUSlotBindingInfos.push_back({
+                .sourceBufferIndex = attribute.buffer,
                 .bufferOffset = attribute.offset,
                 .stride = attribute.stride });
-            outVertexBufferLayouts.push_back(
-                    { .stepMode = wgpu::VertexStepMode::Vertex, .arrayStride = attribute.stride });
+            // we need to have a vertex buffer layout for each slot
+            outVertexBufferLayouts.push_back({
+                .stepMode = wgpu::VertexStepMode::Vertex,
+                .arrayStride = attribute.stride
+                // we do not know attributeCount or attributes at this time. We get those
+                // in a subsequent pass over the attributeInfos we collect in this loop.
+            });
         }
+        // we don't use a designated initializer here because certain compilers can't handle it
         outAttributeInfos[currentAttributeIndex++] = AttributeInfo(
                 existingSlot,
                 {
@@ -320,8 +297,9 @@ void createBufferLayoutsBindingSlotsAndAttributeInfos(AttributeArray const& attr
                             attribute.offset - outWebGPUSlotBindingInfos[existingSlot].bufferOffset,
                     .shaderLocation = attributeIndex });
     }
-
-    outActualAttributeCount = currentAttributeIndex;
+    FILAMENT_CHECK_POSTCONDITION(currentAttributeIndex == attributeCount)
+            << "Using " << currentAttributeIndex << " attributes, but " << attributeCount
+            << " where indicated.";
     outVertexBufferLayouts.shrink_to_fit();
     outWebGPUSlotBindingInfos.shrink_to_fit();
 }
@@ -338,21 +316,21 @@ WebGPUVertexBufferInfo::WebGPUVertexBufferInfo(const uint8_t bufferCount,
             << ") exceeds Filament's MAX_VERTEX_ATTRIBUTE_COUNT limit ("
             << MAX_VERTEX_ATTRIBUTE_COUNT << ") and/or the device's limit ("
             << deviceLimits.maxVertexAttributes << ")";
-    mVertexAttributes.reserve(MAX_VERTEX_ATTRIBUTE_COUNT); // Reserve max as we might add dummies
+    mVertexAttributes.reserve(attributeCount);
     if (attributeCount == 0) {
         mVertexBufferLayouts.resize(0);
         mWebGPUSlotBindingInfos.resize(0);
         return; // should not be possible, but being defensive. nothing to do otherwise
     }
     std::array<AttributeInfo, MAX_VERTEX_ATTRIBUTE_COUNT> attributeInfos{};
-    uint8_t actualAttributeCount = 0;
-    createBufferLayoutsBindingSlotsAndAttributeInfos(attributes, deviceLimits.maxVertexBuffers,
-            mWebGPUSlotBindingInfos, mVertexBufferLayouts, attributeInfos, actualAttributeCount);
+    createBufferLayoutsBindingSlotsAndAttributeInfos(attributes, attributeCount,
+            deviceLimits.maxVertexBuffers, mWebGPUSlotBindingInfos, mVertexBufferLayouts,
+            attributeInfos);
     // sort attribute infos by increasing slot (by increasing offset within each slot).
     // We do this to ensure that attributes for the same slot/layout are contiguous in
     // the vector, so the vertex buffer layout associated with these contiguous attributes
     // can directly reference them in the mVertexAttributes vector below.
-    std::sort(attributeInfos.data(), attributeInfos.data() + actualAttributeCount,
+    std::sort(attributeInfos.data(), attributeInfos.data() + attributeCount,
             [](AttributeInfo const& first, AttributeInfo const& second) {
                 if (first.slot < second.slot) {
                     return true;
@@ -366,7 +344,7 @@ WebGPUVertexBufferInfo::WebGPUVertexBufferInfo(const uint8_t bufferCount,
             });
     // populate mVertexAttributes and update mVertexBufferLayouts to reference the correct
     // attributes in it (which will be contiguous in memory as ensured by the sorting above)...
-    for (uint32_t attributeIndex = 0; attributeIndex < actualAttributeCount; ++attributeIndex) {
+    for (uint32_t attributeIndex = 0; attributeIndex < attributeCount; ++attributeIndex) {
         AttributeInfo const& info = attributeInfos[attributeIndex];
         mVertexAttributes.push_back(info.attribute);
         if (mVertexBufferLayouts[info.slot].attributeCount == 0) {

filament/backend/test/BackendTest.cpp

@@ -83,8 +83,6 @@ BackendTest::~BackendTest() {
     driver->terminate();
     delete driver;
     recordFailedImages();
-
-    delete mPlatform;
 }
 
 void BackendTest::initializeDriver() {

filament/include/filament/Material.h

@@ -139,10 +139,8 @@ public:
         Builder& package(const void* UTILS_NONNULL payload, size_t size);
 
         template<typename T>
-        using is_supported_constant_parameter_t = std::enable_if_t<
-                std::is_same_v<int32_t, T> ||
-                std::is_same_v<float, T> ||
-                std::is_same_v<bool, T>>;
+        using is_supported_constant_parameter_t =
+                MaterialInstance::is_supported_constant_parameter_t<T>;
 
         /**
          * Specialize a constant parameter specified in the material definition with a concrete
@@ -243,11 +241,18 @@ public:
      * for stereoscopic rendering. If an application is not planning to render in stereo, this bit
      * should be turned off to avoid unnecessary material compilations.
      *
+     * Note that it is possible to override specialization constants on a per-MaterialInstance basis
+     * (@see MaterialInstance::setConstant). In that case, the programs compiled by a call to
+     * Material::compile() may not be reusable by that MaterialInstance. It's better to call
+     * MaterialInstance::compile() in cases where you intend to override specialization constants.
+     *
      * @param priority      Which priority queue to use, LOW or HIGH.
      * @param variants      Variants to include to the compile command.
      * @param handler       Handler to dispatch the callback or nullptr for the default handler
      * @param callback      callback called on the main thread when the compilation is done on
      *                      by backend.
+     *
+     * @see Material::compile
      */
     void compile(CompilerPriorityQueue priority,
             UserVariantFilterMask variants,

filament/include/filament/MaterialInstance.h

@@ -94,6 +94,12 @@ public:
             std::is_same_v<math::mat3f, T>
     >;
 
+    template<typename T>
+    using is_supported_constant_parameter_t = std::enable_if_t<
+            std::is_same_v<int32_t, T> ||
+            std::is_same_v<float, T> ||
+            std::is_same_v<bool, T>>;
+
     /**
      * Creates a new MaterialInstance using another MaterialInstance as a template for initialization.
      * The new MaterialInstance is an instance of the same Material of the template instance and
@@ -276,6 +282,91 @@ public:
         return getParameter<T>(name, strlen(name));
     }
 
+    /**
+     * Overrides a specialization constant of this material instance.
+     *
+     * @tparam T    The type of the constant. Must be int32_t, float, or bool.
+     * @param name  The name of the constant as defined in the material. Cannot be nullptr.
+     * @param nameLength Length in `char` of the name parameter.
+     * @param value The value of the constant.
+     *
+     * @see Material::Builder::constant
+     */
+    template<typename T, typename = is_supported_constant_parameter_t<T>>
+    void setConstant(const char* UTILS_NONNULL name, size_t nameLength, T value);
+
+    /** inline helper to provide the name as a null-terminated string literal */
+    template<typename T, typename = is_supported_constant_parameter_t<T>>
+    void setConstant(StringLiteral const name, T value) {
+        setConstant<T>(name.data, name.size, value);
+    }
+
+    /** inline helper to provide the name as a null-terminated C string */
+    template<typename T, typename = is_supported_constant_parameter_t<T>>
+    void setConstant(const char* UTILS_NONNULL name, T value) {
+        setConstant<T>(name, strlen(name), value);
+    }
+
+    /**
+     * Gets the value of a specialization constant by name.
+     *
+     * @tparam T    The type of the constant. Must be int32_t, float, or bool.
+     * @param name  The name of the constant as defined in the material. Cannot be nullptr.
+     * @param nameLength Length in `char` of the name parameter.
+     * @return The value of the constant.
+     */
+    template<typename T, typename = is_supported_constant_parameter_t<T>>
+    T getConstant(const char* UTILS_NONNULL name, size_t nameLength) const;
+
+    /** inline helper to provide the name as a null-terminated C string */
+    template<typename T, typename = is_supported_constant_parameter_t<T>>
+    T getConstant(StringLiteral const name) const {
+        return getConstant<T>(name.data, name.size);
+    }
+
+    /** inline helper to provide the name as a null-terminated C string */
+    template<typename T, typename = is_supported_constant_parameter_t<T>>
+    T getConstant(const char* UTILS_NONNULL name) const {
+        return getConstant<T>(name, strlen(name));
+    }
+
+    using CompilerPriorityQueue = backend::CompilerPriorityQueue;
+
+    /**
+     * Asynchronously ensures that a subset of this MaterialInstance's variants are compiled.
+     *
+     * This function behaves identically to Material::compile(), but takes into account the
+     * specific constants overridden by setConstant().
+     *
+     * @param priority      Which priority queue to use, LOW or HIGH.
+     * @param variants      Variants to include to the compile command.
+     * @param handler       Handler to dispatch the callback or nullptr for the default handler
+     * @param callback      callback called on the main thread when the compilation is done on
+     *                      by backend.
+     *
+     * @see Material::compile
+     * @see setConstant
+     */
+    void compile(CompilerPriorityQueue priority,
+            UserVariantFilterMask variants,
+            backend::CallbackHandler* UTILS_NULLABLE handler = nullptr,
+            utils::Invocable<void(MaterialInstance* UTILS_NONNULL)>&& callback = {}) noexcept;
+
+    inline void compile(CompilerPriorityQueue priority,
+            UserVariantFilterBit variants,
+            backend::CallbackHandler* UTILS_NULLABLE handler = nullptr,
+            utils::Invocable<void(MaterialInstance* UTILS_NONNULL)>&& callback = {}) noexcept {
+        compile(priority, UserVariantFilterMask(variants), handler,
+                std::forward<utils::Invocable<void(MaterialInstance* UTILS_NONNULL)>>(callback));
+    }
+
+    inline void compile(CompilerPriorityQueue priority,
+            backend::CallbackHandler* UTILS_NULLABLE handler = nullptr,
+            utils::Invocable<void(MaterialInstance* UTILS_NONNULL)>&& callback = {}) noexcept {
+        compile(priority, UserVariantFilterBit::ALL, handler,
+                std::forward<utils::Invocable<void(MaterialInstance* UTILS_NONNULL)>>(callback));
+    }
+
     /**
      * Set-up a custom scissor rectangle; by default it is disabled.
      *

filament/src/LocalProgramCache.cpp

@@ -182,15 +182,11 @@ Program::SpecializationConstant LocalProgramCache::getConstantImpl(
         std::string_view name) const noexcept {
     assert_invariant(mMaterial != nullptr);
 
-    MaterialDefinition const& definition = mMaterial->getDefinition();
-    auto it = definition.specializationConstantsNameToIndex.find(name);
-    if (it != definition.specializationConstantsNameToIndex.cend()) {
-        return getConstantImpl(it->second + CONFIG_MAX_RESERVED_SPEC_CONSTANTS);
-    }
+    auto const& constants = mMaterial->getDefinition().specializationConstantsNameToIndex;
+    auto it = constants.find(name);
+    FILAMENT_CHECK_PRECONDITION(it != constants.end()) << "Constant " << name << " does not exist";
 
-    CString name_cstring(name);
-    PANIC_PRECONDITION("No such constant exists: %s", name_cstring.c_str());
-    return {};
+    return getConstantImpl(it->second + CONFIG_MAX_RESERVED_SPEC_CONSTANTS);
 }
 
 void LocalProgramCache::setConstants(
@@ -240,6 +236,13 @@ void LocalProgramCache::setConstants(
     }
 }
 
+void LocalProgramCache::setConstants(
+        FixedCapacityVector<Program::SpecializationConstant> constants) noexcept {
+    assert_invariant(mMaterial != nullptr);
+
+    setConstantsImpl(std::move(constants));
+}
+
 void LocalProgramCache::setConstantsImpl(
         FixedCapacityVector<Program::SpecializationConstant> constants) noexcept {
     FEngine& engine = mMaterial->getEngine();

filament/src/LocalProgramCache.h

@@ -119,7 +119,14 @@ public:
             std::pair<std::string_view, backend::Program::SpecializationConstant>>
                     constants) noexcept;
 
+    // Set constants list directly.
+    void setConstants(utils::FixedCapacityVector<backend::Program::SpecializationConstant>
+                    constants) noexcept;
+
 private:
+    // Apply any pending specialization constants. Invalidates programs as necessary.
+    void flushConstants() const;
+
     backend::Handle<backend::HwProgram> prepareProgramSlow(backend::DriverApi& driver,
             Variant const variant,
             backend::CompilerPriorityQueue const priorityQueue) const noexcept;

filament/src/MaterialDefinition.cpp

@@ -574,15 +574,13 @@ void MaterialDefinition::processPushConstants() {
             [&](MaterialPushConstant const& constant) {
                 snprintf(buf, sizeof(buf), "%s.%s", structVarName.c_str(), constant.name.c_str());
 
-                CString const cs(buf, strlen(buf));
-
                 switch (constant.stage) {
                     case ShaderStage::VERTEX:
-                        vertexConstants.push_back({cs, constant.type});
+                        vertexConstants.push_back({CString(buf), constant.type});
                         vertexCount++;
                         break;
                     case ShaderStage::FRAGMENT:
-                        fragmentConstants.push_back({cs, constant.type});
+                        fragmentConstants.push_back({CString(buf), constant.type});
                         fragmentCount++;
                         break;
                     case ShaderStage::COMPUTE:

filament/src/MaterialInstance.cpp

@@ -233,6 +233,26 @@ template UTILS_PUBLIC mat3f MaterialInstance::getParameter<mat3f>      (const ch
 
 // ------------------------------------------------------------------------------------------------
 
+template<typename T, typename>
+void MaterialInstance::setConstant(const char* name, size_t nameLength, T value) {
+    downcast(this)->setConstantImpl(std::string_view{name, nameLength}, value);
+}
+
+template UTILS_PUBLIC void MaterialInstance::setConstant<int32_t>(const char* name, size_t nameLength, int32_t value);
+template UTILS_PUBLIC void MaterialInstance::setConstant<float>(const char* name, size_t nameLength, float value);
+template UTILS_PUBLIC void MaterialInstance::setConstant<bool>(const char* name, size_t nameLength, bool value);
+
+template<typename T, typename>
+T MaterialInstance::getConstant(const char* name, size_t nameLength) const {
+    return downcast(this)->getConstantImpl<T>(std::string_view{name, nameLength});
+}
+
+template UTILS_PUBLIC int32_t MaterialInstance::getConstant<int32_t>(const char* name, size_t nameLength) const;
+template UTILS_PUBLIC float MaterialInstance::getConstant<float>(const char* name, size_t nameLength) const;
+template UTILS_PUBLIC bool MaterialInstance::getConstant<bool>(const char* name, size_t nameLength) const;
+
+// ------------------------------------------------------------------------------------------------
+
 Material const* MaterialInstance::getMaterial() const noexcept {
     return downcast(this)->getMaterial();
 }
@@ -403,4 +423,10 @@ void MaterialInstance::commit(Engine& engine) const {
     downcast(this)->commit(downcast(engine));
 }
 
+void MaterialInstance::compile(CompilerPriorityQueue const priority,
+        UserVariantFilterMask const variants, CallbackHandler* handler,
+        utils::Invocable<void(MaterialInstance*)>&& callback) noexcept {
+    downcast(this)->compile(priority, variants, handler, std::move(callback));
+}
+
 } // namespace filament

filament/src/PostProcessManager.cpp

@@ -186,12 +186,10 @@ void PostProcessManager::PostProcessMaterial::loadMaterial(FEngine& engine) cons
 }
 
 UTILS_NOINLINE
-FMaterial* PostProcessManager::PostProcessMaterial::getMaterial(FEngine& engine,
-        DriverApi& driver, Variant::type_t const variant) const noexcept {
+FMaterial* PostProcessManager::PostProcessMaterial::getMaterial(FEngine& engine) const noexcept {
     if (UTILS_UNLIKELY(mSize)) {
         loadMaterial(engine);
     }
-    mMaterial->prepareProgram(driver, Variant{ variant }, CompilerPriorityQueue::CRITICAL);
     return mMaterial;
 }
 
@@ -249,6 +247,20 @@ void PostProcessManager::bindPerRenderableDescriptorSet(DriverApi& driver) const
             { { 0, 0 }, driver });
 }
 
+FMaterialInstance* PostProcessManager::getMaterialInstance(backend::DriverApi& driver,
+        FMaterial const* ma, Variant::type_t variant) const {
+    FMaterialInstance* mi = mMaterialInstanceManager.getMaterialInstance(ma);
+    mi->prepareProgram(driver, Variant{ variant }, backend::CompilerPriorityQueue::CRITICAL);
+    return mi;
+}
+
+FMaterialInstance* PostProcessManager::getMaterialInstanceWithTag(backend::DriverApi& driver,
+        FMaterial const* ma, uint32_t tag, Variant::type_t variant) const {
+    FMaterialInstance* mi = mMaterialInstanceManager.getMaterialInstance(ma, tag);
+    mi->prepareProgram(driver, Variant { variant }, backend::CompilerPriorityQueue::CRITICAL);
+    return mi;
+}
+
 UboManager* PostProcessManager::getUboManager() const noexcept {
     return mEngine.getUboManager();
 }
@@ -468,9 +480,10 @@ void PostProcessManager::unbindAllDescriptorSets(DriverApi& driver) noexcept {
 
 UTILS_NOINLINE
 PipelineState PostProcessManager::getPipelineState(
-        FMaterial const* const ma, Variant::type_t const variant) const noexcept {
+        FMaterialInstance const* const mi, Variant::type_t const variant) const noexcept {
+    FMaterial const* const ma = mi->getMaterial();
     return {
-            .program = ma->getProgram(Variant{ variant }),
+            .program = mi->getProgram(Variant{ variant }),
             .vertexBufferInfo = mFullScreenQuadVbih,
             .pipelineLayout = {
                     .setLayout = {
@@ -478,7 +491,7 @@ PipelineState PostProcessManager::getPipelineState(
                             mPerRenderableDslh,
                             ma->getDescriptorSetLayout().getHandle()
                     }},
-            .rasterState = ma->getRasterState()
+            .rasterState = mi->getRasterState()
     };
 }
 
@@ -518,8 +531,7 @@ void PostProcessManager::commitAndRenderFullScreenQuad(DriverApi& driver,
         PostProcessVariant const variant) const noexcept {
     mi->commit(driver, getUboManager());
     mi->use(driver);
-    FMaterial const* const ma = mi->getMaterial();
-    PipelineState const pipeline = getPipelineState(ma, variant);
+    PipelineState const pipeline = getPipelineState(mi, variant);
 
     assert_invariant(
             ((out.params.readOnlyDepthStencil & RenderPassParams::READONLY_DEPTH)
@@ -636,12 +648,12 @@ PostProcessManager::StructurePassOutput PostProcessManager::structure(FrameGraph
 
                 auto in = resources.getTexture(data.depth);
                 auto& material = getPostProcessMaterial("mipmapDepth");
-                FMaterial const* const ma = material.getMaterial(mEngine, driver);
-                FMaterialInstance* const mi = getMaterialInstance(ma);
+                FMaterial const* const ma = material.getMaterial(mEngine);
+                FMaterialInstance* const mi = getMaterialInstance(driver, ma);
                 // Only the depth texture is changing in the material instance (no UBO updates),
                 // we do not move getMaterialInstance() inside the loop.
 
-                auto pipeline = getPipelineState(ma);
+                auto pipeline = getPipelineState(mi);
 
                 // The first mip already exists, so we process n-1 lods
                 for (size_t level = 0; level < levelCount - 1; level++) {
@@ -1001,14 +1013,13 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::screenSpaceAmbientOcclusion(
 #endif
                 auto& material = getPostProcessMaterial(materialName);
 
-                FMaterial* ma = material.getMaterial(mEngine, driver);
+                FMaterial* ma = material.getMaterial(mEngine);
                 ma->getPrograms().setConstants({
                     { "useVisibilityBitmasks", options.gtao.useVisibilityBitmasks },
                     { "linearThickness", options.gtao.linearThickness },
                 });
 
-                ma = material.getMaterial(mEngine, driver);
-                FMaterialInstance* const mi = getMaterialInstance(ma);
+                FMaterialInstance* const mi = getMaterialInstance(driver, ma);
 
                 // Set AO type specific material parameters
                 switch (aoType) {
@@ -1084,7 +1095,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::screenSpaceAmbientOcclusion(
                 mi->commit(driver, getUboManager());
                 mi->use(driver);
 
-                auto pipeline = getPipelineState(ma);
+                auto pipeline = getPipelineState(mi);
                 pipeline.rasterState.depthFunc = RasterState::DepthFunc::L;
                 assert_invariant(ssao.params.readOnlyDepthStencil & RenderPassParams::READONLY_DEPTH);
                 renderFullScreenQuad(ssao, pipeline, driver);
@@ -1194,8 +1205,8 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::bilateralBlurPass(FrameGraph
                 auto& material = config.bentNormals ?
                         getPostProcessMaterial("bilateralBlurBentNormals") :
                         getPostProcessMaterial("bilateralBlur");
-                FMaterial const* const ma = material.getMaterial(mEngine, driver);
-                FMaterialInstance* const mi = getMaterialInstance(ma);
+                FMaterial const* const ma = material.getMaterial(mEngine);
+                FMaterialInstance* const mi = getMaterialInstance(driver, ma);
                 mi->setParameter("ssao", ssao, { /* only reads level 0 */ });
                 mi->setParameter("axis", axis / float2{desc.width, desc.height});
                 mi->setParameter("kernel", kGaussianSamples, kGaussianCount);
@@ -1205,7 +1216,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::bilateralBlurPass(FrameGraph
                 mi->commit(driver, getUboManager());
                 mi->use(driver);
 
-                auto pipeline = getPipelineState(ma);
+                auto pipeline = getPipelineState(mi);
                 pipeline.rasterState.depthFunc = RasterState::DepthFunc::L;
                 renderFullScreenQuad(blurred, pipeline, driver);
                 unbindAllDescriptorSets(driver);
@@ -1364,7 +1375,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::gaussianBlurPass(FrameGraph&
                             "separableGaussianBlur4L"sv : "separableGaussianBlur4"sv;   break;
                 }
                 auto const& separableGaussianBlur = getPostProcessMaterial(materialName);
-                auto ma = separableGaussianBlur.getMaterial(mEngine, driver);
+                auto ma = separableGaussianBlur.getMaterial(mEngine);
 
                 const size_t kernelStorageSize = ma->reflect("kernel")->size;
                 float2 kernel[64];
@@ -1887,9 +1898,10 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::dof(FrameGraph& fg,
                 auto inOutCoc   = resources.getTexture(data.inOutCoc);
 
                 auto const& material = getPostProcessMaterial("dofMipmap");
-                FMaterial const* const ma = material.getMaterial(mEngine, driver);
+                FMaterial const* const ma = material.getMaterial(mEngine);
+                FMaterialInstance* const mi = getMaterialInstance(driver, ma);
 
-                auto const pipeline = getPipelineState(ma, variant);
+                auto const pipeline = getPipelineState(mi, variant);
 
                 for (size_t level = 0 ; level < mipmapCount - 1u ; level++) {
                     const float w = FTexture::valueForLevel(level, desc.width);
@@ -1897,7 +1909,8 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::dof(FrameGraph& fg,
                     auto const& out = resources.getRenderPassInfo(data.rp[level]);
                     auto inColor = driver.createTextureView(inOutColor, level, 1);
                     auto inCoc = driver.createTextureView(inOutCoc, level, 1);
-                    FMaterialInstance* const mi = getMaterialInstance(ma);
+                    // FIXME: is this necessary?
+                    FMaterialInstance* const mi = getMaterialInstance(driver, ma);
 
                     mi->setParameter("color", inColor, SamplerParams{
                             .filterMin = SamplerMinFilter::NEAREST_MIPMAP_NEAREST });
@@ -2426,9 +2439,9 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloom(FrameGraph& fg,
                 auto const& outDesc = resources.getDescriptor(data.out);
 
                 auto const& material = getPostProcessMaterial("bloomUpsample");
-                FMaterial const* const ma = material.getMaterial(mEngine, driver);
+                FMaterial const* const ma = material.getMaterial(mEngine);
 
-                auto pipeline = getPipelineState(ma);
+                auto pipeline = getPipelineState(getMaterialInstance(driver, ma));
                 pipeline.rasterState.blendFunctionSrcRGB = BlendFunction::ONE;
                 pipeline.rasterState.blendFunctionDstRGB = BlendFunction::ONE;
 
@@ -2436,7 +2449,7 @@ PostProcessManager::BloomPassOutput PostProcessManager::bloom(FrameGraph& fg,
                     // Note that we wouldn't want to use the same instance for each pass since that
                     // would imply using the same UBOs, which implies synchronization across the
                     // passes.
-                    FMaterialInstance* mi = getMaterialInstance(ma);
+                    FMaterialInstance* mi = getMaterialInstance(driver, ma);
                     auto hwDstRT = resources.getRenderPassInfo(data.outRT[i - 1]);
                     hwDstRT.params.flags.discardStart = TargetBufferFlags::NONE; // b/c we'll blend
                     hwDstRT.params.flags.discardEnd = TargetBufferFlags::NONE;
@@ -2569,11 +2582,12 @@ void PostProcessManager::colorGradingSubpass(DriverApi& driver,
             PostProcessVariant::TRANSLUCENT : PostProcessVariant::OPAQUE;
 
     auto const& material = getPostProcessMaterial("colorGradingAsSubpass");
-    FMaterial const* const ma = material.getMaterial(mEngine, driver, variant);
+    FMaterial const* const ma = material.getMaterial(mEngine);
     // the UBO has been set and committed in colorGradingPrepareSubpass()
-    FMaterialInstance const* mi = mMaterialInstanceManager.getMaterialInstance(ma, colorGradingConfig.translucent);
+    FMaterialInstance const* mi =
+            getMaterialInstanceWithTag(driver, ma, colorGradingConfig.translucent, variant);
     mi->use(driver);
-    auto const pipeline = getPipelineState(ma, variant);
+    auto const pipeline = getPipelineState(mi, variant);
     driver.nextSubpass();
     driver.scissor(mi->getScissor());
     driver.draw(pipeline, mFullScreenQuadRph, 0, 3, 1);
@@ -2581,8 +2595,8 @@ void PostProcessManager::colorGradingSubpass(DriverApi& driver,
 
 void PostProcessManager::customResolvePrepareSubpass(DriverApi& driver, CustomResolveOp const op) noexcept {
     auto const& material = getPostProcessMaterial("customResolveAsSubpass");
-    auto const ma = material.getMaterial(mEngine, driver, PostProcessVariant::OPAQUE);
-    auto* const mi = mMaterialInstanceManager.getMaterialInstance(ma, 0);
+    auto const ma = material.getMaterial(mEngine);
+    auto* const mi = getMaterialInstance(driver, ma, 0);
     mi->setParameter("direction", op == CustomResolveOp::COMPRESS ? 1.0f : -1.0f),
     mi->commit(driver, getUboManager());
 }
@@ -2592,12 +2606,12 @@ void PostProcessManager::customResolveSubpass(DriverApi& driver) noexcept {
     bindPerRenderableDescriptorSet(driver);
 
     auto const& material = getPostProcessMaterial("customResolveAsSubpass");
-    FMaterial const* const ma = material.getMaterial(mEngine, driver);
+    FMaterial const* const ma = material.getMaterial(mEngine);
     // the UBO has been set and committed in customResolvePrepareSubpass()
-    FMaterialInstance const* mi = mMaterialInstanceManager.getMaterialInstance(ma, 0);
+    FMaterialInstance const* mi = getMaterialInstance(driver, ma, 0);
     mi->use(driver);
 
-    auto const pipeline = getPipelineState(ma);
+    auto const pipeline = getPipelineState(mi);
     driver.nextSubpass();
     driver.scissor(mi->getScissor());
     driver.draw(pipeline, mFullScreenQuadRph, 0, 3, 1);
@@ -2633,8 +2647,8 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::customResolveUncompressPass(
 void PostProcessManager::clearAncillaryBuffersPrepare(DriverApi& driver,
         Variant::type_t variant) noexcept {
     auto const& material = getPostProcessMaterial("clearDepth");
-    auto const ma = material.getMaterial(mEngine, driver, variant);
-    auto const mi = mMaterialInstanceManager.getMaterialInstance(ma, 0);
+    auto const ma = material.getMaterial(mEngine);
+    auto const mi = getMaterialInstanceWithTag(driver, ma, 0, variant);
     mi->commit(driver, getUboManager());
 }
 
@@ -2649,13 +2663,13 @@ void PostProcessManager::clearAncillaryBuffers(DriverApi& driver,
     bindPerRenderableDescriptorSet(driver);
 
     auto const& material = getPostProcessMaterial("clearDepth");
-    FMaterial const* const ma = material.getMaterial(mEngine, driver, variant);
+    FMaterial const* const ma = material.getMaterial(mEngine);
 
     // the UBO has been set and committed in clearAncillaryBuffersPrepare()
-    FMaterialInstance const* const mi = mMaterialInstanceManager.getMaterialInstance(ma, 0);
+    FMaterialInstance const* const mi = getMaterialInstanceWithTag(driver, ma, 0, variant);
     mi->use(driver);
 
-    auto pipeline = getPipelineState(ma, variant);
+    auto pipeline = getPipelineState(mi, variant);
     pipeline.rasterState.depthFunc = RasterState::DepthFunc::A;
 
     driver.scissor(mi->getScissor());
@@ -2665,7 +2679,7 @@ void PostProcessManager::clearAncillaryBuffers(DriverApi& driver,
 void PostProcessManager::fogPrepare(DriverApi& driver) noexcept {
     // ensures the material is loaded and material instance created
     auto const& material = getPostProcessMaterial("fog");
-    FMaterial const* const ma = material.getMaterial(mEngine, driver, PostProcessVariant::OPAQUE);
+    FMaterial const* const ma = material.getMaterial(mEngine);
     FMaterialInstance const* mi = ma->getDefaultInstance();
     mi->commit(driver, getUboManager());
 }
@@ -2676,11 +2690,11 @@ void PostProcessManager::fog(DriverApi& driver) noexcept {
     bindPerRenderableDescriptorSet(driver);
 
     auto const& material = getPostProcessMaterial("fog");
-    FMaterial const* const ma = material.getMaterial(mEngine, driver);
+    FMaterial const* const ma = material.getMaterial(mEngine);
     FMaterialInstance const* mi = ma->getDefaultInstance();
     mi->use(driver);
 
-    auto pipeline = getPipelineState(ma, Variant::NO_VARIANT);
+    auto pipeline = getPipelineState(mi, Variant::NO_VARIANT);
     driver.scissor(mi->getScissor());
     driver.draw(pipeline, mFullScreenQuadRph, 0, 3, 1);
 }
@@ -2937,7 +2951,7 @@ void PostProcessManager::TaaJitterCamera(
 void PostProcessManager::configureTemporalAntiAliasingMaterial(backend::DriverApi& driver,
         TemporalAntiAliasingOptions const& taaOptions) noexcept {
 
-    FMaterial* const ma = getPostProcessMaterial("taa").getMaterial(mEngine, driver);
+    FMaterial* const ma = getPostProcessMaterial("taa").getMaterial(mEngine);
     ma->getPrograms().setConstants({
         { "upscaling", taaOptions.upscaling > 1.0f },
         { "historyReprojection", taaOptions.historyReprojection },
@@ -2956,7 +2970,7 @@ FMaterialInstance* PostProcessManager::configureColorGradingMaterial(backend::Dr
         PostProcessMaterial const& material, FColorGrading const* colorGrading,
         ColorGradingConfig const& colorGradingConfig, VignetteOptions const& vignetteOptions,
         uint32_t const width, uint32_t const height) noexcept {
-    FMaterial* ma = material.getMaterial(mEngine, driver);
+    FMaterial* ma = material.getMaterial(mEngine);
     ma->getPrograms().setConstants({
         { "isOneDimensional", colorGrading->isOneDimensional() },
         { "isLDR", colorGrading->isLDR() },
@@ -2965,8 +2979,9 @@ FMaterialInstance* PostProcessManager::configureColorGradingMaterial(backend::Dr
     PostProcessVariant const variant = colorGradingConfig.translucent
                                                ? PostProcessVariant::TRANSLUCENT
                                                : PostProcessVariant::OPAQUE;
-    ma = material.getMaterial(mEngine, driver, variant);
-    FMaterialInstance* mi = mMaterialInstanceManager.getMaterialInstance(ma, colorGradingConfig.translucent);
+    ma = material.getMaterial(mEngine);
+    FMaterialInstance* mi =
+            getMaterialInstanceWithTag(driver, ma, colorGradingConfig.translucent, variant);
 
     const SamplerParams params = SamplerParams{
             .filterMag = SamplerMagFilter::LINEAR,
@@ -3078,9 +3093,9 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::taa(FrameGraph& fg,
                 PostProcessVariant const variant = colorGradingConfig.translucent ?
                         PostProcessVariant::TRANSLUCENT : PostProcessVariant::OPAQUE;
 
-                FMaterial const* const ma = material.getMaterial(mEngine, driver, variant);
+                FMaterial const* const ma = material.getMaterial(mEngine);
 
-                FMaterialInstance* mi = getMaterialInstance(ma);
+                FMaterialInstance* mi = getMaterialInstance(driver, ma);
                 mi->setParameter("color",  color, SamplerParams{});  // nearest
                 mi->setParameter("depth",  depth, SamplerParams{});  // nearest
                 mi->setParameter("history", history, SamplerParams{
@@ -3120,7 +3135,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::taa(FrameGraph& fg,
                 if (colorGradingConfig.asSubpass) {
                     out.params.subpassMask = 1;
                 }
-                auto const pipeline = getPipelineState(ma, variant);
+                auto const pipeline = getPipelineState(mi, variant);
 
                 driver.beginRenderPass(out.target, out.params);
                 driver.draw(pipeline, mFullScreenQuadRph, 0, 3, 1);
@@ -3202,7 +3217,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::rcas(
                 mi->commit(driver, getUboManager());
                 mi->use(driver);
 
-                auto pipeline = getPipelineState(material.getMaterial(mEngine, driver), variant);
+                auto pipeline = getPipelineState(mi, variant);
                 if (mode == RcasMode::BLENDED) {
                     pipeline.rasterState.blendFunctionSrcRGB = BlendFunction::ONE;
                     pipeline.rasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
@@ -3295,7 +3310,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::upscaleBilinear(FrameGraph&
 
                 auto out = resources.getRenderPassInfo();
 
-                auto pipeline = getPipelineState(material.getMaterial(mEngine, driver));
+                auto pipeline = getPipelineState(mi);
                 if (blended) {
                     pipeline.rasterState.blendFunctionSrcRGB = BlendFunction::ONE;
                     pipeline.rasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
@@ -3509,15 +3524,15 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::upscaleFSR1(FrameGraph& fg,
                 auto out = resources.getRenderPassInfo();
 
                 if (UTILS_UNLIKELY(twoPassesEASU)) {
-                    auto pipeline0 = getPipelineState(splitEasuMaterial->getMaterial(mEngine, driver));
-                    auto pipeline1 = getPipelineState(easuMaterial->getMaterial(mEngine, driver));
+                    auto pipeline0 = getPipelineState(getMaterialInstance(mEngine, driver, *splitEasuMaterial));
+                    auto pipeline1 = getPipelineState(getMaterialInstance(mEngine, driver, *easuMaterial));
                     pipeline1.rasterState.depthFunc = SamplerCompareFunc::NE;
                     driver.beginRenderPass(out.target, out.params);
                     driver.draw(pipeline0, mFullScreenQuadRph, 0, 3, 1);
                     driver.draw(pipeline1, mFullScreenQuadRph, 0, 3, 1);
                     driver.endRenderPass();
                 } else {
-                    auto pipeline = getPipelineState(easuMaterial->getMaterial(mEngine, driver));
+                    auto pipeline = getPipelineState(getMaterialInstance(mEngine, driver, *easuMaterial));
                     renderFullScreenQuad(out, pipeline, driver);
                 }
                 unbindAllDescriptorSets(driver);
@@ -3569,8 +3584,8 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::blit(FrameGraph& fg, bool co
 
                 PostProcessMaterial const& material =
                         getPostProcessMaterial(layer ? "blitArray" : "blitLow");
-                FMaterial const* const ma = material.getMaterial(mEngine, driver);
-                auto* mi = getMaterialInstance(ma);
+                FMaterial const* const ma = material.getMaterial(mEngine);
+                auto* mi = getMaterialInstance(driver, ma);
                 mi->setParameter("color", color, SamplerParams{
                         .filterMag = filterMag,
                         .filterMin = filterMin
@@ -3588,7 +3603,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::blit(FrameGraph& fg, bool co
                 mi->commit(driver, getUboManager());
                 mi->use(driver);
 
-                auto pipeline = getPipelineState(ma);
+                auto pipeline = getPipelineState(mi);
                 if (translucent) {
                     pipeline.rasterState.blendFunctionSrcRGB = BlendFunction::ONE;
                     pipeline.rasterState.blendFunctionSrcAlpha = BlendFunction::ONE;
@@ -3886,10 +3901,10 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::debugCombineArrayTexture(Fra
                 // set uniforms
 
                 PostProcessMaterial const& material = getPostProcessMaterial("blitArray");
-                FMaterial const* const ma = material.getMaterial(mEngine, driver);
+                FMaterial const* const ma = material.getMaterial(mEngine);
                 // It should be ok to not move this getMaterialInstance to inside the loop, since
                 // this is a pass meant for debug.
-                auto* mi = getMaterialInstance(ma);
+                auto* mi = getMaterialInstance(driver, ma);
                 mi->setParameter("color", color, SamplerParams{
                         .filterMag = filterMag,
                         .filterMin = filterMin
@@ -3903,7 +3918,7 @@ FrameGraphId<FrameGraphTexture> PostProcessManager::debugCombineArrayTexture(Fra
                 mi->commit(driver, getUboManager());
                 mi->use(driver);
 
-                auto pipeline = getPipelineState(ma);
+                auto pipeline = getPipelineState(mi);
                 if (translucent) {
                     pipeline.rasterState.blendFunctionSrcRGB = BlendFunction::ONE;
                     pipeline.rasterState.blendFunctionSrcAlpha = BlendFunction::ONE;

filament/src/PostProcessManager.h

@@ -360,13 +360,7 @@ public:
 
         void terminate(FEngine& engine) noexcept;
 
-        FMaterial* getMaterial(FEngine& engine, backend::DriverApi& driver,
-                Variant::type_t variant) const noexcept;
-
-        FMaterial* getMaterial(FEngine& engine, backend::DriverApi& driver,
-                PostProcessVariant variant = PostProcessVariant::OPAQUE) const noexcept {
-                return getMaterial(engine, driver, Variant::type_t(variant));
-        }
+        FMaterial* getMaterial(FEngine& engine) const noexcept;
 
     private:
         void loadMaterial(FEngine& engine) const noexcept;
@@ -393,11 +387,12 @@ public:
 
     void bindPerRenderableDescriptorSet(backend::DriverApi& driver) const noexcept;
 
-    backend::PipelineState getPipelineState(FMaterial const* ma, Variant::type_t variant) const noexcept;
+    backend::PipelineState getPipelineState(FMaterialInstance const* mi,
+            Variant::type_t variant) const noexcept;
 
-    backend::PipelineState getPipelineState(FMaterial const* ma,
-                    PostProcessVariant variant = PostProcessVariant::OPAQUE) const noexcept {
-            return getPipelineState(ma, Variant::type_t(variant));
+    backend::PipelineState getPipelineState(FMaterialInstance const* mi,
+            PostProcessVariant variant = PostProcessVariant::OPAQUE) const noexcept {
+        return getPipelineState(mi, Variant::type_t(variant));
     }
 
     void renderFullScreenQuad(FrameGraphResources::RenderPassInfo const& out,
@@ -427,27 +422,47 @@ public:
 
     void resetForRender();
 
-
     MaterialInstanceManager& getMaterialInstanceManager() noexcept {
             return mMaterialInstanceManager;
     }
 
-    // Helper to get a MaterialInstance from a FMaterial
-    // This currently just call FMaterial::getDefaultInstance().
-    FMaterialInstance* getMaterialInstance(FMaterial const* ma) {
-        return mMaterialInstanceManager.getMaterialInstance(ma);
-    }
-
-    // Helper to get a MaterialInstance from a PostProcessMaterial.
-    FMaterialInstance* getMaterialInstance(FEngine& engine, backend::DriverApi& driver,
-            PostProcessMaterial const& material, PostProcessVariant variant = PostProcessVariant::OPAQUE) {
-        FMaterial const* ma = material.getMaterial(engine, driver, variant);
-        return getMaterialInstance(ma);
-    }
-
     static void unbindAllDescriptorSets(backend::DriverApi& driver) noexcept;
 
 private:
+
+    // Helpers to get MaterialInstances.
+    //
+    // These funcions additionally ensure that the necessary shader programs are compiled via
+    // prepareProgram().
+    FMaterialInstance* getMaterialInstance(backend::DriverApi& driver, FMaterial const* ma,
+            Variant::type_t variant) const;
+
+    FMaterialInstance* getMaterialInstance(backend::DriverApi& driver, FMaterial const* ma,
+            PostProcessVariant variant = PostProcessVariant::OPAQUE) const {
+        return getMaterialInstance(driver, ma, Variant::type_t(variant));
+    }
+
+    FMaterialInstance* getMaterialInstance(FEngine& engine, backend::DriverApi& driver,
+            PostProcessMaterial const& material,
+            PostProcessVariant variant = PostProcessVariant::OPAQUE) const {
+        return getMaterialInstance(driver, material.getMaterial(engine), Variant::type_t(variant));
+    }
+
+    FMaterialInstance* getMaterialInstanceWithTag(backend::DriverApi& driver, FMaterial const* ma,
+            uint32_t tag, Variant::type_t variant) const;
+
+    FMaterialInstance* getMaterialInstanceWithTag(backend::DriverApi& driver, FMaterial const* ma,
+            uint32_t tag, PostProcessVariant variant = PostProcessVariant::OPAQUE) const {
+        return getMaterialInstanceWithTag(driver, ma, tag, Variant::type_t(variant));
+    }
+
+    FMaterialInstance* getMaterialInstanceWithTag(FEngine& engine, backend::DriverApi& driver,
+            PostProcessMaterial const& material, uint32_t tag,
+            PostProcessVariant variant = PostProcessVariant::OPAQUE) const {
+        return getMaterialInstanceWithTag(driver, material.getMaterial(engine), tag,
+                Variant::type_t(variant));
+    }
+
     UboManager* getUboManager() const noexcept;
 
     backend::RenderPrimitiveHandle mFullScreenQuadRph;

filament/src/RenderPass.cpp

@@ -243,8 +243,8 @@ void RenderPass::appendCommands(FEngine const& engine, backend::DriverApi& drive
     // This must be done from the main thread.
     for (Command const* first = curr, *last = curr + commandCount ; first != last ; ++first) {
         if (UTILS_LIKELY((first->key & CUSTOM_MASK) == uint64_t(CustomCommand::PASS))) {
-            auto ma = first->info.mi->getMaterial();
-            ma->prepareProgram(driver, first->info.materialVariant, CompilerPriorityQueue::CRITICAL);
+            first->info.mi->prepareProgram(driver, first->info.materialVariant,
+                    CompilerPriorityQueue::CRITICAL);
         }
     }
 }
@@ -426,16 +426,18 @@ void RenderPass::setupColorCommand(Command& cmdDraw, Variant variant,
     bool const isBlendingCommand = !hasScreenSpaceRefraction &&
             (blendingMode != BlendingMode::OPAQUE && blendingMode != BlendingMode::MASKED);
 
+    RasterState rasterState = mi->getRasterState();
+
     uint64_t keyDraw = cmdDraw.key;
     keyDraw &= ~(PASS_MASK | BLENDING_MASK | MATERIAL_MASK);
     keyDraw |= uint64_t(hasScreenSpaceRefraction ? Pass::REFRACT : Pass::COLOR);
     keyDraw |= uint64_t(CustomCommand::PASS);
     keyDraw |= mi->getSortingKey(); // already all set-up for direct or'ing
     keyDraw |= makeField(variant.key, MATERIAL_VARIANT_KEY_MASK, MATERIAL_VARIANT_KEY_SHIFT);
-    keyDraw |= makeField(ma->getRasterState().alphaToCoverage, BLENDING_MASK, BLENDING_SHIFT);
+    keyDraw |= makeField(rasterState.alphaToCoverage, BLENDING_MASK, BLENDING_SHIFT);
 
     cmdDraw.key = isBlendingCommand ? keyBlending : keyDraw;
-    cmdDraw.info.rasterState = ma->getRasterState();
+    cmdDraw.info.rasterState = rasterState;
 
     // for SSR pass, the blending mode of opaques (including MASKED) must be off
     // see Material.cpp.
@@ -446,10 +448,6 @@ void RenderPass::setupColorCommand(Command& cmdDraw, Variant variant,
             BlendFunction::ZERO : cmdDraw.info.rasterState.blendFunctionDstAlpha;
 
     cmdDraw.info.rasterState.inverseFrontFaces = inverseFrontFaces;
-    cmdDraw.info.rasterState.culling = mi->getCullingMode();
-    cmdDraw.info.rasterState.colorWrite = mi->isColorWriteEnabled();
-    cmdDraw.info.rasterState.depthWrite = mi->isDepthWriteEnabled();
-    cmdDraw.info.rasterState.depthFunc = mi->getDepthFunc();
     cmdDraw.info.rasterState.depthClamp = hasDepthClamp;
     cmdDraw.info.materialVariant = variant;
     // we keep "RasterState::colorWrite" to the value set by material (could be disabled)
@@ -1090,8 +1088,7 @@ void RenderPass::Executor::execute(FEngine const& engine, DriverApi& driver,
                     mi->use(driver, info.materialVariant);
                 }
 
-                assert_invariant(ma);
-                pipeline.program = ma->getProgram(info.materialVariant);
+                pipeline.program = mi->getProgram(info.materialVariant);
 
                 if (UTILS_UNLIKELY(memcmp(&pipeline, &currentPipeline, sizeof(PipelineState)) != 0)) {
                     currentPipeline = pipeline;

filament/src/ShadowMapManager.cpp

@@ -636,7 +636,7 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::gaussianBlurSeparatedPass(
 
             // get the material
             auto const& separableGaussianBlur = ppm.getPostProcessMaterial("gaussian");
-            auto const ma = separableGaussianBlur.getMaterial(engine, driver);
+            auto const ma = separableGaussianBlur.getMaterial(engine);
 
             // Generates half of a Gaussian kernel (center + one side)
             auto generateGaussianWeights = [](std::array<float, 32>& weights,
@@ -685,7 +685,7 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::gaussianBlurSeparatedPass(
             for (auto const mi : miList) {
                 mi->use(driver);
                 driver.scissor(mi->getScissor());
-                driver.draw(ppm.getPipelineState(ma), engine.getFullScreenRenderPrimitive(), 0, 3, 1);
+                driver.draw(ppm.getPipelineState(mi), engine.getFullScreenRenderPrimitive(), 0, 3, 1);
             }
             driver.endRenderPass();
 
@@ -733,12 +733,12 @@ FrameGraphId<FrameGraphTexture> ShadowMapManager::vsmMipmapPass(
                 ppm.bindPerRenderableDescriptorSet(driver);
 
                 auto& material = ppm.getPostProcessMaterial("vsmMipmap");
-                FMaterial const* const ma = material.getMaterial(engine, driver);
+                FMaterial const* const ma = material.getMaterial(engine);
+                auto const mi = ppm.getMaterialInstanceManager().getMaterialInstance(ma);
 
-                auto const pipeline = ppm.getPipelineState(ma);
+                auto const pipeline = ppm.getPipelineState(mi);
                 backend::Viewport const scissor = { 0, 0, dim, dim };
 
-                auto const mi = ppm.getMaterialInstanceManager().getMaterialInstance(ma);
                 mi->setParameter("color", in, SamplerParams{
                         .filterMag = SamplerMagFilter::NEAREST,
                         .filterMin = SamplerMinFilter::NEAREST_MIPMAP_NEAREST

filament/src/details/Engine.cpp

@@ -1830,7 +1830,7 @@ void FEngine::compile(
         CallbackHandler* handler,
         Invocable<void(Material*)>&& callback) {
     auto const variants = getMaterialCompileVariants(view, shadowReceiver, skinning);
-    material->compile(priority, variants, handler, std::move(callback));
+    const_cast<FMaterial*>(material)->compile(priority, variants, handler, std::move(callback));
 }
 
 // ------------------------------------------------------------------------------------------------

filament/src/details/Material.cpp

@@ -170,8 +170,6 @@ FMaterial::FMaterial(FEngine& engine, const Builder& builder, MaterialDefinition
 
     DriverApi& driver = engine.getDriverApi();
 
-    mIsStereoSupported = driver.isStereoSupported();
-    mIsParallelShaderCompileSupported = driver.isParallelShaderCompileSupported();
     mDepthPrecacheDisabled =
             driver.isWorkaroundNeeded(Workaround::DISABLE_DEPTH_PRECACHE_FOR_DEFAULT_MATERIAL);
     mDefaultMaterial = engine.getDefaultMaterial();
@@ -240,44 +238,33 @@ filament::DescriptorSetLayout const& FMaterial::getPerViewDescriptorSetLayout(
 void FMaterial::compile(CompilerPriorityQueue const priority,
         UserVariantFilterMask variantSpec,
         CallbackHandler* handler,
-        Invocable<void(Material*)>&& callback) const noexcept {
-    DriverApi& driver = mEngine.getDriverApi();
-
-    // Turn off the STE variant if stereo is not supported.
-    if (!mIsStereoSupported) {
-        variantSpec &= ~UserVariantFilterMask(UserVariantFilterBit::STE);
+        Invocable<void(Material*)>&& callback) noexcept {
+    FMaterialInstance* mi = getDefaultInstance();
+    if (callback) {
+        mi->compile(priority, variantSpec, handler,
+                [this, callback = std::move(callback)](MaterialInstance*) {
+                    callback(this);
+                });
+    } else {
+        mi->compile(priority, variantSpec, handler, {});
     }
-
-    UserVariantFilterMask const variantFilter = ~variantSpec & UserVariantFilterMask(UserVariantFilterBit::ALL);
-    ShaderModel const shaderModel = mEngine.getShaderModel();
-    bool const isStereoSupported = mEngine.getDriverApi().isStereoSupported();
-
-    if (UTILS_LIKELY(mIsParallelShaderCompileSupported)) {
-        for (auto const variant: mDefinition.getVariants()) {
-            if (!variantFilter || variant == Variant::filterUserVariant(variant, variantFilter)) {
-                if (mDefinition.hasVariant(variant, shaderModel, isStereoSupported)) {
-                    prepareProgram(driver, variant, priority);
-                }
-            }
-        }
-    }
-
-    compileAllPrograms(priority, handler, std::move(callback));
 }
 
 void FMaterial::compile(CompilerPriorityQueue const priority,
         FixedCapacityVector<Variant> const& variants,
         CallbackHandler* handler,
-        Invocable<void(Material*)>&& callback) const noexcept {
+        Invocable<void(Material*)>&& callback) noexcept {
     DriverApi& driver = mEngine.getDriverApi();
+    FMaterialInstance* mi = getDefaultInstance();
 
     ShaderModel const shaderModel = mEngine.getShaderModel();
     bool const isStereoSupported = driver.isStereoSupported();
+    bool const isParallelShaderCompileSupported = driver.isParallelShaderCompileSupported();
 
-    if (UTILS_LIKELY(mIsParallelShaderCompileSupported)) {
+    if (UTILS_LIKELY(isParallelShaderCompileSupported)) {
         for (auto const variant : variants) {
             if (mDefinition.hasVariant(variant, shaderModel, isStereoSupported)) {
-                prepareProgram(driver, variant, priority);
+                mi->prepareProgram(driver, variant, priority);
             }
         }
     }
@@ -287,7 +274,7 @@ void FMaterial::compile(CompilerPriorityQueue const priority,
 
 void FMaterial::compileAllPrograms(CompilerPriorityQueue const priority,
         CallbackHandler* handler,
-        Invocable<void(Material*)>&& callback) const noexcept {
+        Invocable<void(Material*)>&& callback) noexcept {
     DriverApi& driver = mEngine.getDriverApi();
 
     if (callback) {

filament/src/details/Material.h

@@ -106,12 +106,12 @@ public:
     void compile(CompilerPriorityQueue priority,
             UserVariantFilterMask variantSpec,
             backend::CallbackHandler* handler,
-            utils::Invocable<void(Material*)>&& callback) const noexcept;
+            utils::Invocable<void(Material*)>&& callback) noexcept;
 
     void compile(CompilerPriorityQueue priority,
         utils::FixedCapacityVector<Variant> const& variants,
         backend::CallbackHandler* handler,
-        utils::Invocable<void(Material*)>&& callback) const noexcept;
+        utils::Invocable<void(Material*)>&& callback) noexcept;
 
     // Creates an instance of this material, specifying the batching mode.
     FMaterialInstance* createInstance(const char* name) const noexcept;
@@ -130,40 +130,6 @@ public:
 
     FEngine& getEngine() const noexcept  { return mEngine; }
 
-    // prepareProgram creates the program for the material's given variant at the backend level.
-    // Must be called outside of backend render pass.
-    // Must be called before getProgram() below.
-    backend::Handle<backend::HwProgram> prepareProgram(backend::DriverApi& driver,
-            Variant const variant,
-            backend::CompilerPriorityQueue const priorityQueue) const noexcept {
-        return mPrograms.prepareProgram(driver, variant, priorityQueue);
-    }
-
-    // getProgram returns the backend program for the material's given variant.
-    // Must be called after prepareProgram().
-    [[nodiscard]]
-    backend::Handle<backend::HwProgram> getProgram(Variant variant) const noexcept {
-
-        if (UTILS_UNLIKELY(mEngine.features.material.enable_fog_as_postprocess)) {
-            // if the fog as post-process feature is enabled, we need to proceed "as-if" the material
-            // didn't have the FOG variant bit.
-            if (getMaterialDomain() == MaterialDomain::SURFACE) {
-                BlendingMode const blendingMode = getBlendingMode();
-                bool const hasScreenSpaceRefraction = getRefractionMode() == RefractionMode::SCREEN_SPACE;
-                bool const isBlendingCommand = !hasScreenSpaceRefraction &&
-                        (blendingMode != BlendingMode::OPAQUE && blendingMode != BlendingMode::MASKED);
-                if (!isBlendingCommand) {
-                    variant.setFog(false);
-                }
-            }
-        }
-
-#if FILAMENT_ENABLE_MATDBG
-        updateActiveProgramsForMatdbg(variant);
-#endif
-        return mPrograms.getProgram(variant);
-    }
-
     // MaterialInstance::use() binds descriptor sets before drawing. For shared variants,
     // however, the material instance will call useShared() to bind the default material's sets
     // instead.
@@ -310,6 +276,9 @@ public:
     }
 
     /** @}*/
+
+    // Called by getProgram() to update active program list for matdbg UI.
+    void updateActiveProgramsForMatdbg(Variant const variant) const noexcept;
 #endif
 
 private:
@@ -318,15 +287,13 @@ private:
 
     void compileAllPrograms(CompilerPriorityQueue priority,
         backend::CallbackHandler* handler,
-        utils::Invocable<void(Material*)>&& callback) const noexcept;
+        utils::Invocable<void(Material*)>&& callback) noexcept;
 
     MaterialDefinition const& mDefinition;
 
     bool mIsDefaultMaterial = false;
 
     bool mUseUboBatching = false;
-    bool mIsStereoSupported = false;
-    bool mIsParallelShaderCompileSupported = false;
     bool mDepthPrecacheDisabled = false;
 
     FMaterial const* mDefaultMaterial = nullptr;
@@ -341,8 +308,6 @@ private:
     mutable utils::Mutex mPendingEditsLock;
     std::unique_ptr<MaterialParser> mPendingEdits;
     std::unique_ptr<MaterialParser> mEditedMaterialParser;
-    // Called by getProgram() to update active program list for matdbg UI.
-    void updateActiveProgramsForMatdbg(Variant const variant) const noexcept;
     void setPendingEdits(std::unique_ptr<MaterialParser> pendingEdits) noexcept;
     bool hasPendingEdits() const noexcept;
     void latchPendingEdits() noexcept;

filament/src/details/MaterialInstance.cpp

@@ -17,6 +17,7 @@
 #include <filament/MaterialInstance.h>
 
 #include "RenderPass.h"
+#include "MaterialParser.h"
 
 #include "ds/DescriptorSetLayout.h"
 
@@ -131,6 +132,7 @@ FMaterialInstance::FMaterialInstance(FEngine& engine,
           mTextureParameters(other->mTextureParameters),
           mDescriptorSet(other->mDescriptorSet.duplicate(
                 "MaterialInstance", mMaterial->getDescriptorSetLayout())),
+          mPrograms(other->mPrograms),
           mPolygonOffset(other->mPolygonOffset),
           mStencilState(other->mStencilState),
           mMaskThreshold(other->mMaskThreshold),
@@ -207,6 +209,10 @@ void FMaterialInstance::terminate(FEngine& engine) {
     if (ubHandle){
         driver.destroyBufferObject(*ubHandle);
     }
+
+    if (mPrograms.isInitialized()) {
+        mPrograms.terminate(engine);
+    }
 }
 
 void FMaterialInstance::commit(FEngine& engine) const {
@@ -264,6 +270,39 @@ void FMaterialInstance::commit(FEngine::DriverApi& driver, UboManager* uboManage
 
 // ------------------------------------------------------------------------------------------------
 
+template<typename T>
+void FMaterialInstance::setConstantImpl(std::string_view name, T value) {
+    auto const& constants = mMaterial->getDefinition().specializationConstantsNameToIndex;
+    auto it = constants.find(name);
+    FILAMENT_CHECK_PRECONDITION(it != constants.end()) << "Constant " << name << " does not exist";
+
+    if (UTILS_UNLIKELY(mPendingSpecializationConstants.empty())) {
+        mPendingSpecializationConstants =
+                FixedCapacityVector<backend::Program::SpecializationConstant>(
+                        getPrograms().getSpecializationConstants());
+    }
+
+    uint32_t id = it->second + CONFIG_MAX_RESERVED_SPEC_CONSTANTS;
+    mPendingSpecializationConstants[id] = value;
+}
+
+template<typename T>
+T FMaterialInstance::getConstantImpl(std::string_view name) const {
+    auto const& constants = mMaterial->getDefinition().specializationConstantsNameToIndex;
+    auto it = constants.find(name);
+    FILAMENT_CHECK_PRECONDITION(it != constants.end()) << "Constant " << name << " does not exist";
+
+    uint32_t id = it->second + CONFIG_MAX_RESERVED_SPEC_CONSTANTS;
+
+    if (UTILS_UNLIKELY(!mPendingSpecializationConstants.empty())) {
+        return std::get<T>(mPendingSpecializationConstants[id]);
+    }
+
+    return getPrograms().getConstant<T>(id);
+}
+
+// ------------------------------------------------------------------------------------------------
+
 void FMaterialInstance::setParameter(std::string_view const name,
         Handle<HwTexture> texture, SamplerParams const params) {
     auto const binding = mMaterial->getSamplerBinding(name);
@@ -383,6 +422,15 @@ void FMaterialInstance::setTransparencyMode(TransparencyMode const mode) noexcep
     mTransparencyMode = mode;
 }
 
+RasterState FMaterialInstance::getRasterState() const noexcept {
+    RasterState rs = mMaterial->getRasterState();
+    rs.culling = mCulling;
+    rs.depthWrite = mDepthWrite;
+    rs.depthFunc = mDepthFunc;
+    rs.colorWrite = mColorWrite;
+    return rs;
+}
+
 void FMaterialInstance::setDepthCulling(bool const enable) noexcept {
     mDepthFunc = enable ? RasterState::DepthFunc::GE : RasterState::DepthFunc::A;
 }
@@ -403,6 +451,51 @@ const char* FMaterialInstance::getName() const noexcept {
 
 // ------------------------------------------------------------------------------------------------
 
+void FMaterialInstance::compile(CompilerPriorityQueue const priority,
+        UserVariantFilterMask variantSpec, CallbackHandler* handler,
+        Invocable<void(MaterialInstance*)>&& callback) noexcept {
+    FEngine& engine = mMaterial->getEngine();
+    DriverApi& driver = engine.getDriverApi();
+    MaterialDefinition const& definition = mMaterial->getDefinition();
+
+    bool const isStereoSupported = driver.isStereoSupported();
+
+    // Turn off the STE variant if stereo is not supported.
+    if (UTILS_LIKELY(!isStereoSupported)) {
+        variantSpec &= ~UserVariantFilterMask(UserVariantFilterBit::STE);
+    }
+
+    UserVariantFilterMask const variantFilter =
+            ~variantSpec & UserVariantFilterMask(UserVariantFilterBit::ALL);
+    ShaderModel const shaderModel = engine.getShaderModel();
+
+    if (UTILS_LIKELY(driver.isParallelShaderCompileSupported())) {
+        for (auto const variant: definition.getVariants()) {
+            if (!variantFilter || variant == Variant::filterUserVariant(variant, variantFilter)) {
+                if (definition.hasVariant(variant, shaderModel, isStereoSupported)) {
+                    prepareProgram(driver, variant, priority);
+                }
+            }
+        }
+    }
+
+    if (callback) {
+        struct Callback {
+            Invocable<void(MaterialInstance*)> f;
+            MaterialInstance* m;
+            static void func(void* user) {
+                auto* const c = static_cast<Callback*>(user);
+                c->f(c->m);
+                delete c;
+            }
+        };
+        auto* const user = new (std::nothrow) Callback{ std::move(callback), this };
+        driver.compilePrograms(priority, handler, &Callback::func, user);
+    } else {
+        driver.compilePrograms(priority, nullptr, nullptr, nullptr);
+    }
+}
+
 void FMaterialInstance::use(FEngine::DriverApi& driver, Variant variant) const {
     assert_invariant(mDescriptorSet.getHandle());
     assert_invariant(!isUsingUboBatching() || BufferAllocator::isValid(getAllocationId()));
@@ -507,4 +600,36 @@ void FMaterialInstance::fixMissingSamplers() const {
     }
 }
 
+LocalProgramCache const& FMaterialInstance::getPrograms() const noexcept {
+    return mPrograms.isInitialized() ? mPrograms : mMaterial->getPrograms();
+}
+
+void FMaterialInstance::flushSpecializationConstants() const noexcept {
+    if (mPendingSpecializationConstants.empty()) {
+        return;
+    }
+
+    if (!mPrograms.isInitialized()) {
+        mPrograms.initializeForMaterialInstance(mMaterial->getEngine(), *mMaterial);
+    }
+    mPrograms.setConstants(std::move(mPendingSpecializationConstants));
+    mPendingSpecializationConstants.clear();
+}
+
+#if FILAMENT_ENABLE_MATDBG
+
+void FMaterialInstance::updateActiveProgramsForMatdbg(Variant const variant) const noexcept {
+    mMaterial->updateActiveProgramsForMatdbg(variant);
+}
+
+#endif // FILAMENT_ENABLE_MATDBG
+
+template void FMaterialInstance::setConstantImpl<int32_t>(std::string_view name, int32_t value);
+template void FMaterialInstance::setConstantImpl<float>(std::string_view name, float value);
+template void FMaterialInstance::setConstantImpl<bool>(std::string_view name, bool value);
+
+template int32_t FMaterialInstance::getConstantImpl<int32_t>(std::string_view name) const;
+template float FMaterialInstance::getConstantImpl<float>(std::string_view name) const;
+template bool FMaterialInstance::getConstantImpl<bool>(std::string_view name) const;
+
 } // namespace filament

filament/src/details/MaterialInstance.h

@@ -18,7 +18,7 @@
 #define TNT_FILAMENT_DETAILS_MATERIALINSTANCE_H
 
 #include "downcast.h"
-
+#include "LocalProgramCache.h"
 #include "UniformBuffer.h"
 
 #include "ds/DescriptorSet.h"
@@ -26,8 +26,6 @@
 #include "details/BufferAllocator.h"
 #include "details/Engine.h"
 
-#include "private/backend/DriverApi.h"
-
 #include <filament/MaterialInstance.h>
 
 #include <private/filament/Variant.h>
@@ -67,7 +65,7 @@ public:
     ~FMaterialInstance() noexcept;
 
     void terminate(FEngine& engine);
-    
+
     void commit(FEngine& engine) const;
 
     void commit(FEngine::DriverApi& driver, UboManager* uboManager) const;
@@ -86,6 +84,32 @@ public:
 
     UniformBuffer const& getUniformBuffer() const noexcept { return mUniforms; }
 
+    void compile(backend::CompilerPriorityQueue priority, UserVariantFilterMask variantSpec,
+            backend::CallbackHandler* handler,
+            utils::Invocable<void(MaterialInstance*)>&& callback) noexcept;
+
+    // prepareProgram creates the program for the material's given variant at the backend level.
+    // Must be called outside of backend render pass.
+    // Must be called before getProgram() below.
+    backend::Handle<backend::HwProgram> prepareProgram(backend::DriverApi& driver,
+            Variant const variant,
+            backend::CompilerPriorityQueue const priorityQueue) const noexcept {
+        flushSpecializationConstants();
+        return getPrograms().prepareProgram(driver, variant, priorityQueue);
+    }
+
+    // getProgram returns the backend program for the material's given variant.
+    // Must be called after prepareProgram().
+    //
+    // See also Material::getProgram().
+    [[nodiscard]]
+    backend::Handle<backend::HwProgram> getProgram(Variant const variant) const noexcept {
+#if FILAMENT_ENABLE_MATDBG
+        updateActiveProgramsForMatdbg(variant);
+#endif
+        return getPrograms().getProgram(variant);
+    }
+
     void setScissor(uint32_t const left, uint32_t const bottom, uint32_t const width, uint32_t const height) noexcept {
         constexpr uint32_t maxvalu = std::numeric_limits<int32_t>::max();
         mScissorRect = { int32_t(left), int32_t(bottom),
@@ -107,6 +131,8 @@ public:
 
     bool hasScissor() const noexcept { return mHasScissor; }
 
+    backend::RasterState getRasterState() const noexcept;
+
     backend::CullingMode getCullingMode() const noexcept { return mCulling; }
 
     backend::CullingMode getShadowCullingMode() const noexcept { return mShadowCulling; }
@@ -254,11 +280,15 @@ public:
             backend::Handle<backend::HwTexture> texture, backend::SamplerParams params);
 
     using MaterialInstance::setParameter;
+    using MaterialInstance::setConstant;
 
 private:
     friend class FMaterial;
     friend class MaterialInstance;
 
+    // Cannot inline since it inspects the FMaterial class.
+    LocalProgramCache const& getPrograms() const noexcept;
+
     template<size_t Size>
     void setParameterUntypedImpl(std::string_view name, const void* value);
 
@@ -277,6 +307,19 @@ private:
     template<typename T>
     T getParameterImpl(std::string_view name) const;
 
+    template<typename T>
+    void setConstantImpl(std::string_view name, T value);
+
+    template<typename T>
+    T getConstantImpl(std::string_view name) const;
+
+    void flushSpecializationConstants() const noexcept;
+
+#if FILAMENT_ENABLE_MATDBG
+    // Called by getProgram() to update active program list for matdbg UI.
+    void updateActiveProgramsForMatdbg(Variant const variant) const noexcept;
+#endif
+
     // keep these grouped, they're accessed together in the render-loop
     FMaterial const* mMaterial = nullptr;
 
@@ -291,6 +334,11 @@ private:
     mutable DescriptorSet mDescriptorSet;
     UniformBuffer mUniforms;
 
+    // HACK: Mutable so that prepareProgram() can update specialization constants.
+    mutable LocalProgramCache mPrograms;
+    mutable utils::FixedCapacityVector<backend::Program::SpecializationConstant>
+            mPendingSpecializationConstants;
+
     backend::PolygonOffset mPolygonOffset{};
     backend::StencilState mStencilState{};
 

filament/test/filament_test_material.cpp

@@ -19,6 +19,7 @@
 
 #include <filament/Engine.h>
 #include <filament/Material.h>
+#include <filament/MaterialInstance.h>
 
 #include "filament_test_resources.h"
 
@@ -167,3 +168,52 @@ TEST(Material, MaterialSettingInvalidApiLevelReturnsAnInvalidPackage) {
 
     Engine::destroy(engine);
 }
+
+TEST(MaterialInstanceTest, SetConstant) {
+    Engine* engine = Engine::create(Engine::Backend::NOOP);
+
+    std::string shaderCode(R"(
+        void material(inout MaterialInputs material) {
+            prepareMaterial(material);
+            material.baseColor = vec4(1.0);
+        }
+    )");
+
+    filamat::MaterialBuilder builder;
+    builder.init();
+    builder.name("MaterialInstanceTest");
+    builder.material(shaderCode.c_str());
+    builder.constant("myFloat", filamat::MaterialBuilder::ConstantType::FLOAT, 1.0f);
+    builder.constant("myInt", filamat::MaterialBuilder::ConstantType::INT, 2);
+    builder.constant("myBool", filamat::MaterialBuilder::ConstantType::BOOL, false);
+
+    filamat::Package result = builder.build(engine->getJobSystem());
+    ASSERT_TRUE(result.isValid());
+
+    Material* material = Material::Builder()
+            .package(result.getData(), result.getSize())
+            .build(*engine);
+    ASSERT_NE(material, nullptr);
+
+    MaterialInstance* instance = material->createInstance();
+    ASSERT_NE(instance, nullptr);
+
+    // Verify default values
+    EXPECT_EQ(instance->getConstant<float>("myFloat"), 1.0f);
+    EXPECT_EQ(instance->getConstant<int32_t>("myInt"), 2);
+    EXPECT_EQ(instance->getConstant<bool>("myBool"), false);
+
+    // Set new values
+    instance->setConstant("myFloat", 3.0f);
+    instance->setConstant("myInt", 4);
+    instance->setConstant("myBool", true);
+
+    // Verify new values
+    EXPECT_EQ(instance->getConstant<float>("myFloat"), 3.0f);
+    EXPECT_EQ(instance->getConstant<int32_t>("myInt"), 4);
+    EXPECT_EQ(instance->getConstant<bool>("myBool"), true);
+
+    engine->destroy(instance);
+    engine->destroy(material);
+    Engine::destroy(engine);
+}

libs/filamat/src/shaders/CodeGenerator.cpp

@@ -881,6 +881,14 @@ utils::io::sstream& CodeGenerator::generateSpecializationConstant(utils::io::sst
 
     static const char* types[] = { "int", "float", "bool" };
 
+    // Spec constants aren't fully supported in Tint,
+    // workaround until https://issues.chromium.org/issues/42250586 is resolved
+    if (mTargetApi == TargetApi::WEBGPU) {
+        std::string const variableName = "FILAMENT_SPEC_CONST_" + std::to_string(id) + "_" + name;
+        out << " const " << types[value.index()] << " " << variableName << " = " << constantString << ";\n";
+        out << types[value.index()] << " " << name << " =  " << variableName << ";\n";
+        return out;
+    }
     if (mTargetLanguage == MaterialBuilderBase::TargetLanguage::SPIRV) {
         out << "layout (constant_id = " << id << ") const "
             << types[value.index()] << " " << name << " = " << constantString << ";\n";

test/backend/test.sh

@@ -14,7 +14,6 @@ if [ ! -d "${PROJECT_ROOT_DIR}/mesa/out" ]; then
 fi
 
 BACKEND_TEST_TARGET=''
-ASAN_FLAG=''
 
 # Set environment variables to use Mesa drivers.
 os_name=$(uname -s)
@@ -28,18 +27,15 @@ if [[ "$os_name" == "Linux" ]]; then
         export VK_ICD_FILENAMES="${PROJECT_ROOT_DIR}/mesa/out/share/vulkan/icd.d/lvp_icd.x86_64.json"
     fi
     BACKEND_TEST_TARGET=backend_test_linux
-    ASAN_FLAG="-b"
 elif [[ "$os_name" == "Darwin" ]]; then
     export DYLD_LIBRARY_PATH="${PROJECT_ROOT_DIR}/mesa/out/lib"
     export VK_ICD_FILENAMES="${PROJECT_ROOT_DIR}/mesa/out/share/vulkan/icd.d/lvp_icd.aarch64.json"
     BACKEND_TEST_TARGET=backend_test_mac
-    # asan is too slow for macOs build of the backend test
-    ASAN_FLAG=""
 fi
 
-# Build backend test
-echo "Building ${BACKEND_TEST_TARGET}..."
-"${PROJECT_ROOT_DIR}/build.sh" ${ASAN_FLAG} -W -y release -p desktop -X "${PROJECT_ROOT_DIR}/mesa" debug ${BACKEND_TEST_TARGET}
+# Build backend_test_mac
+echo "Building backend_test_mac..."
+"${PROJECT_ROOT_DIR}/build.sh" -W -p desktop -X "${PROJECT_ROOT_DIR}/mesa" debug ${BACKEND_TEST_TARGET}
 
 set +e
 
@@ -55,20 +51,10 @@ do
     fi
 done
 
-# Mesa OSMesa is known to leak part of the context.
-LSAN_CMD_PREFIX=""
-if [[ "${ASAN_FLAG}" == "-b" ]]; then
-    echo "leak:PlatformOSMesa.cpp" > leak_skip.txt
-    export LSAN_OPTIONS=suppressions=leak_skip.txt
-fi
-
 FINAL_RESULT=0
 for BACKEND in ${BACKENDS[@]}; do
     echo "----- ${BACKEND} backend test -----"
-
-    ${PROJECT_ROOT_DIR}/out/cmake-debug/filament/backend/${BACKEND_TEST_TARGET} \
-                           -a ${BACKEND} --ci --headless_only ${GTEST_FILTER_ARG}
-
+    ${PROJECT_ROOT_DIR}/out/cmake-debug/filament/backend/${BACKEND_TEST_TARGET} -a ${BACKEND} --ci --headless_only ${GTEST_FILTER_ARG}
     RESULT=$(echo $?)
     if [ ${RESULT} -gt 0 ]; then
         echo "----- Error: backend ${BACKEND} test failed with result ${RESULT} -----"

third_party/abseil/absl/base/config.h

@@ -302,19 +302,7 @@ static_assert(ABSL_INTERNAL_INLINE_NAMESPACE_STR[0] != 'h' ||
 #if (defined(__clang__) && !defined(_WIN32)) ||           \
     (defined(__CUDACC__) && __CUDACC_VER_MAJOR__ >= 9) || \
     (defined(__GNUC__) && !defined(__clang__) && !defined(__CUDACC__))
-
-// Disable int128 instrinsic on arm when asan is enabled
-#if defined(__arm__) || defined(__aarch64__)
-    #if defined(__has_feature) || defined(__SANITIZE_ADDRESS__)
-        #if __has_feature(address_sanitizer) || defined(__SANITIZE_ADDRESS__)
-            #define ABSL_ASAN_ON_ARM_DISABLE_INT128 1
-        #endif
-    #endif
-#endif
-#ifndef ABSL_ASAN_ON_ARM_DISABLE_INT128
-    #define ABSL_HAVE_INTRINSIC_INT128 1
-#endif
-
+#define ABSL_HAVE_INTRINSIC_INT128 1
 #elif defined(__CUDACC__)
 // __CUDACC_VER__ is a full version number before CUDA 9, and is defined to a
 // string explaining that it has been removed starting with CUDA 9. We use

third_party/abseil/tnt/README.md

@@ -32,13 +32,3 @@ This folder previously last updated as follows:
     rsync -r abseil_new/ abseil/ --delete --exclude tnt
     rm -rf master.zip abseil_new
     git add abseil ; git status
-
-## Custom Changes
-
-A custom patch has been applied to fix compilation on ARM architectures when AddressSanitizer (ASAN) is enabled (specifically regarding `ABSL_HAVE_INTRINSIC_INT128`).
-
-If you update the `abseil` folder, you may need to re-apply the patch. You can do so by running the following command from the repository root:
-
-```shell
-git apply third_party/abseil/tnt/asan_arm_int128.diff
-```

third_party/dawn/tnt/CMakeLists.txt

@@ -52,6 +52,8 @@ if (ANDROID AND FILAMENT_BUILD_FILAMAT)
         tint_lang_spirv_intrinsic
         tint_lang_spirv_ir
         tint_lang_spirv_reader
+        tint_lang_spirv_reader_ast_lower
+        tint_lang_spirv_reader_ast_parser
         tint_lang_spirv_reader_common
         tint_lang_spirv_reader_lower
         tint_lang_spirv_reader_parser
@@ -75,7 +77,6 @@ if (ANDROID AND FILAMENT_BUILD_FILAMAT)
         tint_utils_rtti
         tint_utils_strconv
         tint_utils_symbol
-        tint_utils_system
         tint_utils_text
         tint_utils_text_generator
     )

web/examples/materials/nonlit.mat

@@ -4,8 +4,7 @@ material {
         color
     ],
     shadingModel : unlit,
-    culling : none,
-    featureLevel: 0
+    culling : none
 }
 
 fragment {

web/filament-js/CMakeLists.txt

@@ -31,7 +31,6 @@ set(COPTS "${COPTS} -DEMSCRIPTEN_HAS_UNBOUND_TYPE_NAMES=0")
 set(LOPTS "${LOPTS} --bind")
 set(LOPTS "${LOPTS} -s ALLOW_MEMORY_GROWTH=1")
 set(LOPTS "${LOPTS} -s EXPORTED_FUNCTIONS=['_malloc','_free']")
-set(LOPTS "${LOPTS} -s EXPORTED_RUNTIME_METHODS=['HEAP8','HEAP16','HEAP32','HEAPU8','HEAPU16','HEAPU32','HEAPF32','HEAPF64']")
 set(LOPTS "${LOPTS} -s MODULARIZE=1")
 set(LOPTS "${LOPTS} -s EXPORT_NAME=Filament")
 set(LOPTS "${LOPTS} -s USE_WEBGL2=1")