dawn: fixes to reconstructPosition, and line rendering

malytomas · malytomas · commit ba6906fcf361 · 2025-10-30T17:32:49.000+01:00
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -18,6 +18,13 @@ endif()
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 include(cage_all)
 
+option(cage_faster_debug "enable some optimizations to improve performance in debug builds" ON)
+option(cage_faster_release "enable more aggressive inlining in release builds" OFF)
+option(cage_assert_in_release_enabled "enable runtime asserts in non-debug builds" OFF)
+option(cage_validate_graphics "enable runtime validation in webgpu/dawn" ON)
+option(cage_profiling_enabled "enable compiling-in instructions for profiling" ON)
+option(cage_use_steam_sockets "include Game Networking Sockets library by Valve/Steam" OFF)
+
 if(CMAKE_CURRENT_BINARY_DIR STREQUAL CMAKE_BINARY_DIR)
 	cage_default_init() # standalone build
 else()
diff --git a/cmake/cage_build_configuration.cmake b/cmake/cage_build_configuration.cmake
@@ -55,7 +55,6 @@ macro(cage_build_configuration)
 		set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd26451") # Arithmetic overflow: Using operator ___ on a 4 byte value and then casting the result to a 8 byte value.
 
 		# optionally improve runtime performance in debug builds (basic inlining)
-		option(cage_faster_debug "enable some optimizations to improve performance in debug builds" ON)
 		if(cage_faster_debug)
 			string(REGEX REPLACE "/Ob[0-9]" "" CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG}")
 			string(REGEX REPLACE "/Ob[0-9]" "" CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}")
@@ -64,7 +63,6 @@ macro(cage_build_configuration)
 		endif()
 
 		# optionally improve runtime performance in release builds (more aggressive inlining)
-		option(cage_faster_release "enable more aggressive inlining in release builds" OFF)
 		if(cage_faster_release)
 			string(REGEX REPLACE "/Ob[0-9]" "" CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE}")
 			string(REGEX REPLACE "/Ob[0-9]" "" CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE}")
diff --git a/cmake/cage_fetch_cage_externals.cmake b/cmake/cage_fetch_cage_externals.cmake
@@ -1,5 +1,13 @@
 
+if(NOT DEFINED egac_fetch_cage_externals_done) # INTERNAL implies FORCE
+	set(egac_fetch_cage_externals_done OFF CACHE INTERNAL "")
+endif()
+
 function(cage_fetch_cage_externals)
+	if(egac_fetch_cage_externals_done)
+		return()
+	endif()
 	execute_process(COMMAND "bash" "${CMAKE_CURRENT_FUNCTION_LIST_DIR}/../submodules.sh" WORKING_DIRECTORY "${CMAKE_CURRENT_FUNCTION_LIST_DIR}/.." COMMAND_ECHO STDOUT COMMAND_ERROR_IS_FATAL ANY)
+	set(egac_fetch_cage_externals_done ON CACHE INTERNAL "" FORCE)
 endfunction(cage_fetch_cage_externals)
 
diff --git a/data/cage/shaders/engine/decal.glsl b/data/cage/shaders/engine/decal.glsl
@@ -13,16 +13,15 @@ void main()
 
 
 $include fragment.glsl
-
 $include ../functions/reconstructPosition.glsl
 
 void main()
 {
+	updateNormal();
+	Material material = loadMaterial();
 	vec3 prevPos = reconstructPosition(texDepth, uniProjection.vpInv, uniViewport.viewport);
 	if (length(varPosition - prevPos) > 0.05)
 		discard;
-	updateNormal();
-	Material material = loadMaterial();
 #ifdef CutOut
 	if (material.opacity < 0.5)
 		discard;
diff --git a/data/cage/shaders/functions/reconstructPosition.glsl b/data/cage/shaders/functions/reconstructPosition.glsl
@@ -1,8 +1,9 @@
 
 vec3 reconstructPosition(sampler2D texDepth, mat4 vpInv, vec4 viewport)
 {
-	float depth = texelFetch(texDepth, ivec2(gl_FragCoord.xy), 0).x * 2 - 1;
-	vec4 pos = vec4((gl_FragCoord.xy - viewport.xy) / viewport.zw * 2 - 1, depth, 1);
+	float depth = texelFetch(texDepth, ivec2(gl_FragCoord.xy), 0).x;
+	vec4 pos = vec4(gl_FragCoord.xy / viewport.zw * 2 - 1, depth, 1);
+	pos.y = -pos.y;
 	pos = vpInv * pos;
 	return pos.xyz / pos.w;
 }
diff --git a/externals/CMakeLists.txt b/externals/CMakeLists.txt
@@ -88,8 +88,6 @@ macro(cage_heavily_optimized)
 	endif()
 endmacro(cage_heavily_optimized)
 
-option(cage_use_steam_sockets "include Game Networking Sockets library by Valve/Steam" OFF)
-
 message(STATUS "----------------------------------------------------------------")
 message(STATUS "starting cage external subdirectories")
 message(STATUS "----------------------------------------------------------------")
diff --git a/externals/dawn/CMakeLists.txt b/externals/dawn/CMakeLists.txt
@@ -24,6 +24,7 @@ set(DAWN_ENABLE_WEBGPU_ON_WEBGPU OFF CACHE INTERNAL "Enable compilation of the W
 set(DAWN_ENABLE_DESKTOP_GL OFF CACHE INTERNAL "Enable compilation of the OpenGL backend" FORCE)
 set(DAWN_ENABLE_OPENGLES OFF CACHE INTERNAL "Enable compilation of the OpenGL ES backend" FORCE)
 set(DAWN_ENABLE_VULKAN ${vulkan} CACHE INTERNAL "Enable compilation of the Vulkan backend" FORCE)
+set(DAWN_ENABLE_SPIRV_VALIDATION ${cage_validate_graphics} CACHE INTERNAL "Enable validation of SPIR-V" FORCE)
 set(DAWN_FORCE_SYSTEM_COMPONENT_LOAD ON CACHE INTERNAL "Allow system component fallback" FORCE)
 set(DAWN_USE_WAYLAND OFF CACHE INTERNAL "Enable support for Wayland surface" FORCE)
 set(DAWN_USE_X11 OFF CACHE INTERNAL "Enable support for X11 surface" FORCE)
@@ -48,11 +49,15 @@ set(TINT_ENABLE_INSTALL OFF CACHE INTERNAL "Enable install step for Tint librari
 set(TINT_BUILD_CMD_TOOLS OFF CACHE INTERNAL "Build the Tint command line tools" FORCE)
 set(TINT_BUILD_IR_BINARY OFF CACHE INTERNAL "Build IR binary format support" FORCE)
 set(TINT_BUILD_TESTS OFF CACHE INTERNAL "Build tests" FORCE)
-set(TINT_ENABLE_IR_VALIDATION ON CACHE INTERNAL "Enable IR validation for backend codegen" FORCE)
+set(TINT_ENABLE_IR_VALIDATION ${cage_validate_graphics} CACHE INTERNAL "Enable IR validation for backend codegen" FORCE)
 
 unset(metal)
 unset(vulkan)
 
+if(NOT CMAKE_BUILD_TYPE)
+	set(CMAKE_BUILD_TYPE "RELEASE")
+endif()
+
 # replace constants header
 set(constants_path "${CMAKE_CURRENT_LIST_DIR}/dawn/src/dawn/common/Constants.h")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/patches")
diff --git a/sources/CMakeLists.txt b/sources/CMakeLists.txt
@@ -46,13 +46,11 @@ endif()
 target_compile_definitions(cage-core PUBLIC "$<$<CONFIG:DEBUG>:CAGE_DEBUG>")
 target_compile_definitions(cage-core PUBLIC "$<$<CONFIG:DEBUG>:CAGE_DEBUG_BOOL=true>")
 target_compile_definitions(cage-core PUBLIC "$<$<NOT:$<CONFIG:DEBUG>>:CAGE_DEBUG_BOOL=false>")
-option(cage_assert_in_release_enabled "enable runtime asserts in non-debug builds" OFF)
 if(cage_assert_in_release_enabled)
 	target_compile_definitions(cage-core PUBLIC "CAGE_ASSERT_ENABLED")
 else()
 	target_compile_definitions(cage-core PUBLIC "$<$<CONFIG:DEBUG>:CAGE_ASSERT_ENABLED>")
 endif()
-option(cage_profiling_enabled "enable compiling-in instructions for profiling" ON)
 if(cage_profiling_enabled)
 	target_compile_definitions(cage-core PUBLIC CAGE_PROFILING_ENABLED)
 	cage_embed_file(cage-core "profiling.htm" profiling_htm)
diff --git a/sources/asset-processor/shader.cpp b/sources/asset-processor/shader.cpp
@@ -518,6 +518,8 @@ void processShader()
 		{
 			const uint32 cdc = toUint32(defines["customDataCount"]);
 			CAGE_LOG(SeverityEnum::Info, "assetProcessor", Stringizer() + "custom data count: " + cdc);
+			if ((cdc % 4) != 0)
+				CAGE_THROW_ERROR(Exception, "customDataCount must be divisible by 4");
 			header.customDataCount = cdc;
 		}
 		header.variantsCount = variants.size();
diff --git a/sources/include/cage-engine/graphicsEncoder.h b/sources/include/cage-engine/graphicsEncoder.h
@@ -80,7 +80,7 @@ namespace cage
 	public:
 		void nextPass(const PassConfig &config);
 
-		//void scissors(Vec2i origin, Vec2i size);
+		void scissors(Vec2i origin, Vec2i size);
 		void draw(const DrawConfig &config);
 		void execute(wgpu::RenderBundle bundle);
 
diff --git a/sources/libengine/assets/shader.cpp b/sources/libengine/assets/shader.cpp
@@ -36,6 +36,8 @@ namespace cage
 
 			CAGE_ASSERT(des.available() == 0);
 
+			shr->customDataCount = header.customDataCount;
+
 			context->assetHolder = std::move(shr).cast<void>();
 		}
 	}
diff --git a/sources/libengine/graphics/encoder.cpp b/sources/libengine/graphics/encoder.cpp
@@ -35,6 +35,7 @@ namespace cage
 			wgpu::CompareFunction depthTest = wgpu::CompareFunction::Less;
 			bool depthWrite = true;
 			bool backFaceCulling = true;
+			uint32 primitiveTopology = 0;
 			BlendingEnum blending = BlendingEnum::None;
 
 			std::size_t hash = 0;
@@ -56,6 +57,7 @@ namespace cage
 					depthWrite = config.depthWrite;
 				}
 				backFaceCulling = config.backFaceCulling;
+				primitiveTopology = config.model->primitiveType;
 				blending = config.blending;
 
 				auto hashCombine = [&](std::unsigned_integral auto v) { hash ^= std::hash<std::decay_t<decltype(v)>>{}(v) + 0x9e3779b9 + (hash << 6) + (hash >> 2); };
@@ -112,6 +114,20 @@ namespace cage
 			return bs;
 		}
 
+		static wgpu::PrimitiveTopology convertTopology(const Model *model)
+		{
+			switch (model->primitiveType)
+			{
+				case 1:
+					return wgpu::PrimitiveTopology::PointList;
+				case 2:
+					return wgpu::PrimitiveTopology::LineList;
+				case 3:
+				default:
+					return wgpu::PrimitiveTopology::TriangleList;
+			}
+		}
+
 		wgpu::RenderPipeline createPipeline(const PassData &passData, const DrawConfig &config)
 		{
 			wgpu::DepthStencilState dss = {};
@@ -144,6 +160,7 @@ namespace cage
 			rpd.vertex.bufferCount = 1;
 			rpd.vertex.buffers = &config.model->getLayout();
 			rpd.primitive.cullMode = config.backFaceCulling ? wgpu::CullMode::Back : wgpu::CullMode::None;
+			rpd.primitive.topology = convertTopology(config.model);
 			if (passData.depthTarget)
 				rpd.depthStencil = &dss;
 			rpd.fragment = &fs;
@@ -371,13 +388,11 @@ namespace cage
 		impl->nextPass(config);
 	}
 
-	/*
 	void GraphicsEncoder::scissors(Vec2i origin, Vec2i size)
 	{
 		GraphicsEncoderImpl *impl = (GraphicsEncoderImpl *)this;
 		impl->scissors(origin, size);
 	}
-	*/
 
 	void GraphicsEncoder::draw(const DrawConfig &config)
 	{
diff --git a/sources/libengine/graphics/sceneRender.cpp b/sources/libengine/graphics/sceneRender.cpp
@@ -172,13 +172,6 @@ namespace cage
 			}
 		};
 
-		template<class T>
-		PointerRange<T> subRange(PointerRange<T> base, uint32 off, uint32 num)
-		{
-			CAGE_ASSERT(off + num <= base.size());
-			return { base.begin() + off, base.begin() + off + num };
-		}
-
 		enum class RenderModeEnum
 		{
 			Shadowmap,
@@ -382,7 +375,8 @@ namespace cage
 					return;
 				if (e && e->has<ShaderDataComponent>())
 				{
-					const auto sub = subRange(PointerRange<const Real>(e->value<ShaderDataComponent>().matrix.data).cast<const float>(), 0, customDataCount);
+					const auto wholeRange = PointerRange<const Real>(e->value<ShaderDataComponent>().matrix.data).cast<const float>();
+					const auto sub = wholeRange.subRange(0, customDataCount);
 					for (const auto it : sub)
 						uniCustomData.push_back(it);
 				}
@@ -412,14 +406,16 @@ namespace cage
 				return HashString("MaterialTexRegular");
 			}
 
-			CAGE_FORCE_INLINE static Holder<Shader> pickShaderVariant(MultiShader *multiShader, const Model *model, uint32 variant, const RenderModeEnum renderMode, const bool skeletalAnimation)
+			CAGE_FORCE_INLINE static Holder<Shader> pickShaderVariant(MultiShader *multiShader, const Model *model, uint32 textureVariant, const RenderModeEnum renderMode, const bool skeletalAnimation)
 			{
 				CAGE_ASSERT(multiShader);
+				uint32 variant = 0;
 				const auto &v = [&](uint32 c)
 				{
 					if (multiShader->checkKeyword(c))
 						variant += c;
 				};
+				v(textureVariant);
 				if (renderMode != RenderModeEnum::Color)
 					v(HashString("DepthOnly"));
 				if (any(model->flags & MeshRenderFlags::CutOut))
diff --git a/sources/libengine/graphics/shader.cpp b/sources/libengine/graphics/shader.cpp
@@ -106,7 +106,10 @@ namespace cage
 		MultiShaderImpl *impl = (MultiShaderImpl *)this;
 		auto it = impl->shaders.find(variant);
 		if (it == impl->shaders.end())
+		{
+			CAGE_LOG_THROW(Stringizer() + "shader name: " + label);
 			CAGE_THROW_ERROR(Exception, "unknown shader variant");
+		}
 		return it->second.share();
 	}
 

Original file line number	Diff line number	Diff line change
`@@ -1,8 +1,9 @@`
`1`	`1`
`2`	`2`	`vec3 reconstructPosition(sampler2D texDepth, mat4 vpInv, vec4 viewport)`
`3`	`3`	`{`
`4`		`- float depth = texelFetch(texDepth, ivec2(gl_FragCoord.xy), 0).x * 2 - 1;`
`5`		`- vec4 pos = vec4((gl_FragCoord.xy - viewport.xy) / viewport.zw * 2 - 1, depth, 1);`
	`4`	`+ float depth = texelFetch(texDepth, ivec2(gl_FragCoord.xy), 0).x;`
	`5`	`+ vec4 pos = vec4(gl_FragCoord.xy / viewport.zw * 2 - 1, depth, 1);`
	`6`	`+ pos.y = -pos.y;`
`6`	`7`	`pos = vpInv * pos;`
`7`	`8`	`return pos.xyz / pos.w;`
`8`	`9`	`}`
Original file line number	Diff line number	Diff line change
`@@ -518,6 +518,8 @@ void processShader()`
`518`	`518`	`{`
`519`	`519`	`const uint32 cdc = toUint32(defines["customDataCount"]);`
`520`	`520`	`CAGE_LOG(SeverityEnum::Info, "assetProcessor", Stringizer() + "custom data count: " + cdc);`
	`521`	`+ if ((cdc % 4) != 0)`
	`522`	`+ CAGE_THROW_ERROR(Exception, "customDataCount must be divisible by 4");`
`521`	`523`	`header.customDataCount = cdc;`
`522`	`524`	`}`
`523`	`525`	`header.variantsCount = variants.size();`