Metal: add RelaxedtoHalfPass for improved ALU performance (#5504)

RELEASE_NOTES.md

@@ -5,6 +5,8 @@ A new header is inserted each time a *tag* is created.
 
 ## v1.22.1 (currently main branch)
 
+- Metal: Shaders now use `half` floating-point arithmetic when possible for improved performance. [⚠️ **Recompile Materials**]
+
 ## v1.22.0
 
 - engine: Changed UBOs layout [⚠️ **Material breakage**].

libs/filamat/src/GLSLPostProcessor.cpp

@@ -441,6 +441,13 @@ std::shared_ptr<spvtools::Optimizer> GLSLPostProcessor::createOptimizer(
         registerSizePasses(*optimizer, config);
     } else if (optimization == MaterialBuilder::Optimization::PERFORMANCE) {
         registerPerformancePasses(*optimizer, config);
+        // Metal doesn't support relaxed precision, but does have support for float16 math operations.
+        if (config.targetApi == MaterialBuilder::TargetApi::METAL) {
+            optimizer->RegisterPass(CreateConvertRelaxedToHalfPass());
+            optimizer->RegisterPass(CreateSimplificationPass());
+            optimizer->RegisterPass(CreateRedundancyEliminationPass());
+            optimizer->RegisterPass(CreateAggressiveDCEPass());
+        }
     }
 
     return optimizer;

libs/filamat/src/GLSLPostProcessor.h

@@ -51,6 +51,7 @@ public:
 
     struct Config {
         filament::Variant variant;
+        MaterialBuilder::TargetApi targetApi;
         filament::backend::ShaderType shaderType;
         filament::backend::ShaderModel shaderModel;
         filament::MaterialDomain domain;

libs/filamat/src/MaterialBuilder.cpp

@@ -727,6 +727,7 @@ bool MaterialBuilder::generateShaders(JobSystem& jobSystem, const std::vector<Va
 #ifndef FILAMAT_LITE
                 GLSLPostProcessor::Config config{
                         .variant = v.variant,
+                        .targetApi = targetApi,
                         .shaderType = v.stage,
                         .shaderModel = shaderModel,
                         .domain = mMaterialDomain,