Speed up HLSL preprocessing and prepared SPIR-V hot paths

cmake/common.cmake

@@ -1259,10 +1259,10 @@ struct DeviceConfigCaps
 	if(NOT NBL_EMBED_BUILTIN_RESOURCES)
 		list(APPEND REQUIRED_OPTIONS
 			-no-nbl-builtins
-			-I "${NBL_ROOT_PATH}/include"
-			-I "${NBL_ROOT_PATH}/3rdparty/dxc/dxc/external/SPIRV-Headers/include"
-			-I "${NBL_ROOT_PATH}/3rdparty/boost/superproject/libs/preprocessor/include"
-			-I "${NBL_ROOT_PATH_BINARY}/src/nbl/device/include"
+			-isystem "${NBL_ROOT_PATH}/include"
+			-isystem "${NBL_ROOT_PATH}/3rdparty/dxc/dxc/external/SPIRV-Headers/include"
+			-isystem "${NBL_ROOT_PATH}/3rdparty/boost/superproject/libs/preprocessor/include"
+			-isystem "${NBL_ROOT_PATH_BINARY}/src/nbl/device/include"
 		)
 	endif()
 
@@ -1307,7 +1307,8 @@ struct DeviceConfigCaps
 		)
 
 		target_sources(${IMPL_TARGET} PUBLIC ${INCLUDE_FILE})
-		set_source_files_properties(${INCLUDE_FILE} PROPERTIES 
+		set_source_files_properties(${INCLUDE_FILE} PROPERTIES
+			GENERATED TRUE
 			HEADER_FILE_ONLY ON
 			VS_TOOL_OVERRIDE None
 		)

include/nbl/asset/utils/ISPIRVEntryPointTrimmer.h

@@ -7,10 +7,12 @@
 
 #include "nbl/system/ILogger.h"
 
+#include <mutex>
+
 namespace nbl::asset
 {
 
-class ISPIRVEntryPointTrimmer final : public core::IReferenceCounted
+class NBL_API2 ISPIRVEntryPointTrimmer final : public core::IReferenceCounted
 {
     public:
         ISPIRVEntryPointTrimmer();
@@ -46,6 +48,8 @@ class ISPIRVEntryPointTrimmer final : public core::IReferenceCounted
         };
 
         Result trim(const ICPUBuffer* spirvBuffer, const core::set<EntryPoint>& entryPoints, system::logger_opt_ptr logger = nullptr) const;
+        bool ensureValidated(const ICPUBuffer* spirvBuffer, system::logger_opt_ptr logger = nullptr) const;
+        void markValidated(const ICPUBuffer* spirvBuffer) const;
 
         inline core::smart_refctd_ptr<const IShader> trim(const IShader* shader, const core::set<EntryPoint>& entryPoints, system::logger_opt_ptr logger = nullptr) const
         {
@@ -72,6 +76,8 @@ class ISPIRVEntryPointTrimmer final : public core::IReferenceCounted
 
     private:
         core::smart_refctd_ptr<ISPIRVOptimizer> m_optimizer;
+        mutable std::mutex m_validationCacheMutex;
+        mutable core::unordered_set<core::blake3_hash_t> m_validatedSpirvHashes;
 };
 
 }

include/nbl/asset/utils/IShaderCompiler.h

@@ -17,6 +17,8 @@
 #include "nbl/builtin/hlsl/enums.hlsl"
 
 #include <functional>
+#include <mutex>
+#include <unordered_set>
 
 namespace nbl::asset
 {
@@ -26,6 +28,25 @@ class NBL_API2 IShaderCompiler : public core::IReferenceCounted
 	public:
 		IShaderCompiler(core::smart_refctd_ptr<system::ISystem>&& system);
 
+		enum class IncludeRootOrigin : uint8_t
+		{
+			User,
+			Builtin,
+			Generated
+		};
+
+		enum class HeaderClass : uint8_t
+		{
+			User,
+			System
+		};
+
+		struct IncludeClassification
+		{
+			IncludeRootOrigin origin = IncludeRootOrigin::User;
+			HeaderClass headerClass = HeaderClass::User;
+		};
+
 		class NBL_API2 IIncludeLoader : public core::IReferenceCounted
 		{
 			public:
@@ -34,12 +55,13 @@ class NBL_API2 IShaderCompiler : public core::IReferenceCounted
 					system::path absolutePath = {};
 					std::string contents = {};
 					core::blake3_hash_t hash = {}; // TODO: we're not yet using IFile::getPrecomputedHash(), so for builtins we can maybe use that in the future
+					IncludeClassification classification = {};
 					// Could be used in the future for early rejection of cache hit
 					//nbl::system::IFileBase::time_point_t lastWriteTime = {};
 
 					explicit inline operator bool() const {return !absolutePath.empty();}
 				};
-				virtual found_t getInclude(const system::path& searchPath, const std::string& includeName) const = 0;
+				virtual found_t getInclude(const system::path& searchPath, const std::string& includeName, bool needHash = true) const = 0;
 		};
 
 		class NBL_API2 IIncludeGenerator : public core::IReferenceCounted
@@ -65,7 +87,7 @@ class NBL_API2 IShaderCompiler : public core::IReferenceCounted
 			public:
 				CFileSystemIncludeLoader(core::smart_refctd_ptr<system::ISystem>&& system);
 
-				IIncludeLoader::found_t getInclude(const system::path& searchPath, const std::string& includeName) const override;
+				IIncludeLoader::found_t getInclude(const system::path& searchPath, const std::string& includeName, bool needHash = true) const override;
 
 			protected:
 				core::smart_refctd_ptr<system::ISystem> m_system;
@@ -74,37 +96,77 @@ class NBL_API2 IShaderCompiler : public core::IReferenceCounted
 		class NBL_API2 CIncludeFinder : public core::IReferenceCounted
 		{
 			public:
+				struct SSessionCache
+				{
+					enum class LookupResult : uint8_t
+					{
+						Miss,
+						Missing,
+						Found
+					};
+
+					explicit SSessionCache(const bool threadSafe = false) : threadSafe(threadSafe) {}
+
+					void clear();
+					LookupResult lookup(const std::string& key, IIncludeLoader::found_t& result) const;
+					void store(const std::string& key, IIncludeLoader::found_t result);
+
+					bool threadSafe = false;
+
+					mutable std::mutex mutex;
+					core::unordered_map<std::string, IIncludeLoader::found_t> found;
+					core::unordered_set<std::string> missing;
+				};
+
 				CIncludeFinder(core::smart_refctd_ptr<system::ISystem>&& system);
 
 				// ! includes within <>
 				// @param requestingSourceDir: the directory where the incude was requested
 				// @param includeName: the string within <> of the include preprocessing directive
-				IIncludeLoader::found_t getIncludeStandard(const system::path& requestingSourceDir, const std::string& includeName) const;
+				IIncludeLoader::found_t getIncludeStandard(const system::path& requestingSourceDir, const std::string& includeName, bool needHash = true, SSessionCache* sessionCache = nullptr) const;
 
 				// ! includes within ""
 				// @param requestingSourceDir: the directory where the incude was requested
 				// @param includeName: the string within "" of the include preprocessing directive
-				IIncludeLoader::found_t getIncludeRelative(const system::path& requestingSourceDir, const std::string& includeName) const;
+				IIncludeLoader::found_t getIncludeRelative(const system::path& requestingSourceDir, const std::string& includeName, bool needHash = true, SSessionCache* sessionCache = nullptr) const;
 
 				inline core::smart_refctd_ptr<CFileSystemIncludeLoader> getDefaultFileSystemLoader() const { return m_defaultFileSystemLoader; }
 
-				void addSearchPath(const std::string& searchPath, const core::smart_refctd_ptr<IIncludeLoader>& loader);
+				void addSearchPath(const std::string& searchPath, const core::smart_refctd_ptr<IIncludeLoader>& loader, IncludeClassification classification = {});
+
+				void addGenerator(const core::smart_refctd_ptr<IIncludeGenerator>& generator, IncludeClassification classification = {IncludeRootOrigin::Generated,HeaderClass::System});
 
-				void addGenerator(const core::smart_refctd_ptr<IIncludeGenerator>& generator);
+				bool isKnownGlobalInclude(std::string_view includeName) const;
+				IIncludeLoader::found_t classifyFound(IIncludeLoader::found_t found) const;
 
 			protected:
-				IIncludeLoader::found_t trySearchPaths(const std::string& includeName) const;
+				IIncludeLoader::found_t trySearchPaths(const std::string& includeName, bool needHash) const;
 
 				IIncludeLoader::found_t tryIncludeGenerators(const std::string& includeName) const;
+				void registerHeaderRoot(std::string rootPath, IncludeClassification classification);
 
 				struct LoaderSearchPath
 				{
 					core::smart_refctd_ptr<IIncludeLoader> loader = nullptr;
 					std::string searchPath = {};
+					IncludeClassification classification = {};
+				};
+
+				struct GeneratorEntry
+				{
+					core::smart_refctd_ptr<IIncludeGenerator> generator = nullptr;
+					IncludeClassification classification = {IncludeRootOrigin::Generated,HeaderClass::System};
+				};
+
+				struct HeaderRoot
+				{
+					std::string path = {};
+					IncludeClassification classification = {};
 				};
 
 				std::vector<LoaderSearchPath> m_loaders;
-				std::vector<core::smart_refctd_ptr<IIncludeGenerator>> m_generators;
+				std::vector<GeneratorEntry> m_generators;
+				std::vector<HeaderRoot> m_headerRoots;
 				core::smart_refctd_ptr<CFileSystemIncludeLoader> m_defaultFileSystemLoader;
 		};
 
@@ -134,9 +196,11 @@ class NBL_API2 IShaderCompiler : public core::IReferenceCounted
 			std::string_view sourceIdentifier = "";
 			system::logger_opt_ptr logger = nullptr;
 			const CIncludeFinder* includeFinder = nullptr;
+			mutable CIncludeFinder::SSessionCache* includeSessionCache = nullptr;
 			std::span<const SMacroDefinition> extraDefines = {};
 			E_SPIRV_VERSION targetSpirvVersion = E_SPIRV_VERSION::ESV_1_6;
 			bool depfile = false;
+			bool preserveComments = false;
 			system::path depfilePath = {};
 			std::function<void(std::string_view)> onPartialOutputOnFailure = {};
 		};

include/nbl/builtin/hlsl/bxdf/reflection/beckmann.hlsl

@@ -4,9 +4,7 @@
 #ifndef _NBL_BUILTIN_HLSL_BXDF_REFLECTION_BECKMANN_INCLUDED_
 #define _NBL_BUILTIN_HLSL_BXDF_REFLECTION_BECKMANN_INCLUDED_
 
-#include "nbl/builtin/hlsl/bxdf/common.hlsl"
 #include "nbl/builtin/hlsl/bxdf/bxdf_traits.hlsl"
-#include "nbl/builtin/hlsl/sampling/cos_weighted_spheres.hlsl"
 #include "nbl/builtin/hlsl/bxdf/ndf/beckmann.hlsl"
 #include "nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl"
 

include/nbl/builtin/hlsl/bxdf/reflection/ggx.hlsl

@@ -4,9 +4,7 @@
 #ifndef _NBL_BUILTIN_HLSL_BXDF_REFLECTION_GGX_INCLUDED_
 #define _NBL_BUILTIN_HLSL_BXDF_REFLECTION_GGX_INCLUDED_
 
-#include "nbl/builtin/hlsl/bxdf/common.hlsl"
 #include "nbl/builtin/hlsl/bxdf/bxdf_traits.hlsl"
-#include "nbl/builtin/hlsl/sampling/cos_weighted_spheres.hlsl"
 #include "nbl/builtin/hlsl/bxdf/ndf/ggx.hlsl"
 #include "nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl"
 

include/nbl/builtin/hlsl/bxdf/transmission/beckmann.hlsl

@@ -4,10 +4,8 @@
 #ifndef _NBL_BUILTIN_HLSL_BXDF_TRANSMISSION_BECKMANN_INCLUDED_
 #define _NBL_BUILTIN_HLSL_BXDF_TRANSMISSION_BECKMANN_INCLUDED_
 
-#include "nbl/builtin/hlsl/bxdf/common.hlsl"
 #include "nbl/builtin/hlsl/bxdf/bxdf_traits.hlsl"
-#include "nbl/builtin/hlsl/sampling/cos_weighted_spheres.hlsl"
-#include "nbl/builtin/hlsl/bxdf/reflection.hlsl"
+#include "nbl/builtin/hlsl/bxdf/ndf/beckmann.hlsl"
 #include "nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl"
 
 namespace nbl

include/nbl/builtin/hlsl/bxdf/transmission/ggx.hlsl

@@ -4,10 +4,7 @@
 #ifndef _NBL_BUILTIN_HLSL_BXDF_TRANSMISSION_GGX_INCLUDED_
 #define _NBL_BUILTIN_HLSL_BXDF_TRANSMISSION_GGX_INCLUDED_
 
-#include "nbl/builtin/hlsl/bxdf/common.hlsl"
 #include "nbl/builtin/hlsl/bxdf/bxdf_traits.hlsl"
-#include "nbl/builtin/hlsl/sampling/cos_weighted_spheres.hlsl"
-#include "nbl/builtin/hlsl/bxdf/reflection.hlsl"
 #include "nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl"
 
 namespace nbl

include/nbl/builtin/hlsl/emulated/float64_t.hlsl

@@ -2,7 +2,6 @@
 #define _NBL_BUILTIN_HLSL_EMULATED_FLOAT64_T_HLSL_INCLUDED_
 
 #include <nbl/builtin/hlsl/emulated/float64_t_impl.hlsl>
-#include <nbl/builtin/hlsl/concepts/core.hlsl>
 
 namespace nbl
 {

include/nbl/builtin/hlsl/member_test_macros.hlsl

@@ -5,7 +5,10 @@
 #define _NBL_BUILTIN_HLSL_MEMBER_TEST_MACROS_INCLUDED_
 
 #include <nbl/builtin/hlsl/utility.hlsl>
-#include <boost/preprocessor.hpp>
+#include <boost/preprocessor/arithmetic/inc.hpp>
+#include <boost/preprocessor/comparison/not_equal.hpp>
+#include <boost/preprocessor/punctuation/comma_if.hpp>
+#include <boost/preprocessor/repetition/repeat.hpp>
 
 #ifdef __HLSL_VERSION
 
@@ -123,4 +126,4 @@ GENERATE_METHOD_TESTER(set)
 
 
 #endif
-#endif
+#endif

include/nbl/builtin/hlsl/path_tracing/concepts.hlsl

@@ -5,7 +5,6 @@
 #define _NBL_BUILTIN_HLSL_PATH_TRACING_CONCEPTS_INCLUDED_
 
 #include <nbl/builtin/hlsl/concepts.hlsl>
-#include <nbl/builtin/hlsl/bxdf/common.hlsl>
 
 namespace nbl
 {
@@ -15,6 +14,17 @@ namespace path_tracing
 {
 namespace concepts
 {
+namespace impl
+{
+template<typename Vector3>
+struct DummyRayInteraction
+{
+    using vector3_type = Vector3;
+
+    vector3_type getN() NBL_CONST_MEMBER_FUNC;
+    bool isMaterialBSDF() NBL_CONST_MEMBER_FUNC;
+};
+}
 
 #define NBL_CONCEPT_NAME RandGenerator
 #define NBL_CONCEPT_TPLT_PRM_KINDS (typename)
@@ -38,7 +48,7 @@ NBL_CONCEPT_END(
 #define NBL_CONCEPT_TPLT_PRM_NAMES (T)
 #define NBL_CONCEPT_PARAM_0 (ray, T)
 #define NBL_CONCEPT_PARAM_1 (v, typename T::vector3_type)
-#define NBL_CONCEPT_PARAM_2 (interaction, bxdf::surface_interactions::SIsotropic<bxdf::ray_dir_info::SBasic<float>, typename T::spectral_type>)
+#define NBL_CONCEPT_PARAM_2 (interaction, impl::DummyRayInteraction<typename T::vector3_type>)
 #define NBL_CONCEPT_PARAM_3 (scalar, typename T::scalar_type)
 #define NBL_CONCEPT_PARAM_4 (color, typename T::spectral_type)
 NBL_CONCEPT_BEGIN(5)
@@ -52,7 +62,7 @@ NBL_CONCEPT_END(
     ((NBL_CONCEPT_REQ_TYPE)(T::vector3_type))
     ((NBL_CONCEPT_REQ_TYPE)(T::spectral_type))
     ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((ray.init(v/*origin*/, v/*direction*/)), ::nbl::hlsl::is_same_v, void))
-    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((ray.template setInteraction<bxdf::surface_interactions::SIsotropic<bxdf::ray_dir_info::SBasic<float>, typename T::spectral_type> >(interaction)), ::nbl::hlsl::is_same_v, void))
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((ray.setInteraction(interaction)), ::nbl::hlsl::is_same_v, void))
     ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((ray.initPayload()), ::nbl::hlsl::is_same_v, void))
     ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((ray.shouldDoMIS()), ::nbl::hlsl::is_same_v, bool))
     ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((ray.foundEmissiveMIS(scalar)), ::nbl::hlsl::is_same_v, typename T::scalar_type))
@@ -124,6 +134,7 @@ NBL_CONCEPT_END(
     ((NBL_CONCEPT_REQ_TYPE)(T::scene_type))
     ((NBL_CONCEPT_REQ_TYPE)(T::ray_type))
     ((NBL_CONCEPT_REQ_TYPE)(T::object_handle_type))
+    ((NBL_CONCEPT_REQ_TYPE)(T::anisotropic_interaction_type))
     ((NBL_CONCEPT_REQ_TYPE)(T::closest_hit_type))
     ((NBL_CONCEPT_REQ_TYPE_ALIAS_CONCEPT)(IntersectorClosestHit, typename T::closest_hit_type))
     ((NBL_CONCEPT_REQ_TYPE_ALIAS_CONCEPT)(Ray, typename T::ray_type))
@@ -136,6 +147,26 @@ NBL_CONCEPT_END(
 #undef intersect
 #include <nbl/builtin/hlsl/concepts/__end.hlsl>
 
+#define NBL_CONCEPT_NAME UnidirectionalInteractionContract
+#define NBL_CONCEPT_TPLT_PRM_KINDS (typename)(typename)(typename)
+#define NBL_CONCEPT_TPLT_PRM_NAMES (RayT)(IntersectorT)(MaterialSystemT)
+#define NBL_CONCEPT_PARAM_0 (ray, RayT)
+#define NBL_CONCEPT_PARAM_1 (hit, typename IntersectorT::closest_hit_type)
+#define NBL_CONCEPT_PARAM_2 (interaction, typename MaterialSystemT::anisotropic_interaction_type)
+NBL_CONCEPT_BEGIN(3)
+#define ray NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_0
+#define hit NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_1
+#define interaction NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_2
+NBL_CONCEPT_END(
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((hit.getInteraction()), ::nbl::hlsl::is_same_v, typename IntersectorT::anisotropic_interaction_type))
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((hit.getInteraction()), ::nbl::hlsl::is_same_v, typename MaterialSystemT::anisotropic_interaction_type))
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((ray.setInteraction(interaction)), ::nbl::hlsl::is_same_v, void))
+);
+#undef interaction
+#undef hit
+#undef ray
+#include <nbl/builtin/hlsl/concepts/__end.hlsl>
+
 #define NBL_CONCEPT_NAME BxdfNode
 #define NBL_CONCEPT_TPLT_PRM_KINDS (typename)
 #define NBL_CONCEPT_TPLT_PRM_NAMES (T)

include/nbl/builtin/hlsl/path_tracing/unidirectional.hlsl

@@ -5,11 +5,8 @@
 #define _NBL_BUILTIN_HLSL_PATH_TRACING_UNIDIRECTIONAL_INCLUDED_
 
 #include <nbl/builtin/hlsl/colorspace/EOTF.hlsl>
-#include <nbl/builtin/hlsl/colorspace/encodeCIEXYZ.hlsl>
 #include <nbl/builtin/hlsl/math/functions.hlsl>
 #include <nbl/builtin/hlsl/sampling/basic.hlsl>
-#include <nbl/builtin/hlsl/bxdf/bxdf_traits.hlsl>
-#include <nbl/builtin/hlsl/vector_utils/vector_traits.hlsl>
 #include <nbl/builtin/hlsl/path_tracing/concepts.hlsl>
 
 namespace nbl
@@ -22,6 +19,7 @@ namespace path_tracing
 template<class RandGen, class Ray, class Intersector, class MaterialSystem, /* class PathGuider, */ class NextEventEstimator, class Accumulator, class Scene
 NBL_PRIMARY_REQUIRES(concepts::RandGenerator<RandGen> && concepts::Ray<Ray> &&
     concepts::Intersector<Intersector> && concepts::MaterialSystem<MaterialSystem> &&
+    concepts::UnidirectionalInteractionContract<Ray, Intersector, MaterialSystem> &&
     concepts::NextEventEstimator<NextEventEstimator> && concepts::Accumulator<Accumulator> &&
     concepts::Scene<Scene>)
 struct Unidirectional

include/nbl/builtin/hlsl/tgmath.hlsl

@@ -7,12 +7,8 @@
 #include <nbl/builtin/hlsl/spirv_intrinsics/glsl.std.450.hlsl>
 #include <nbl/builtin/hlsl/tgmath/impl.hlsl>
 #include <nbl/builtin/hlsl/type_traits.hlsl>
-#include <nbl/builtin/hlsl/vector_utils/vector_traits.hlsl>
 #include <nbl/builtin/hlsl/cpp_compat.hlsl>
 #include <nbl/builtin/hlsl/ieee754.hlsl>
-#include <nbl/builtin/hlsl/spirv_intrinsics/core.hlsl>
-#include <nbl/builtin/hlsl/concepts/core.hlsl>
-#include <nbl/builtin/hlsl/concepts/vector.hlsl>
 #include <nbl/builtin/hlsl/cpp_compat/intrinsics.hlsl>
 
 // C++ headers