support rasterization kill

src/Cafe/HW/Latte/Core/LatteShader.cpp

@@ -501,9 +501,28 @@ void LatteSHRC_UpdateVSBaseHash(uint8* vertexShaderPtr, uint32 vertexShaderSize,
 	uint64 vsHash2 = 0;
 	_calculateShaderProgramHash(vsProgramCode, vertexShaderSize, &hashCacheVS, &vsHash1, &vsHash2);
 	uint64 vsHash = vsHash1 + vsHash2 + _activeFetchShader->key + _activePSImportTable.key + (usesGeometryShader ? 0x1111ULL : 0ULL);
-	if (g_renderer->GetType() == RendererAPI::Metal && usesGeometryShader)
+	if (g_renderer->GetType() == RendererAPI::Metal)
+	{
+	    if (usesGeometryShader)
 	        vsHash += _activeFetchShader->mtlShaderHashObject;
 
+		// Rasterization
+		bool rasterizationEnabled = !LatteGPUState.contextNew.PA_CL_CLIP_CNTL.get_DX_RASTERIZATION_KILL();
+
+		// HACK
+		if (!LatteGPUState.contextNew.PA_CL_VTE_CNTL.get_VPORT_X_OFFSET_ENA())
+			rasterizationEnabled = true;
+
+		const auto& polygonControlReg = LatteGPUState.contextNew.PA_SU_SC_MODE_CNTL;
+		uint32 cullFront = polygonControlReg.get_CULL_FRONT();
+		uint32 cullBack = polygonControlReg.get_CULL_BACK();
+		if (cullFront && cullBack)
+		    rasterizationEnabled = false;
+
+		if (rasterizationEnabled)
+		    vsHash += 51ULL;
+	}
+
 	uint32 tmp = LatteGPUState.contextNew.PA_CL_VTE_CNTL.getRawValue() ^ 0x43F;
 	vsHash += tmp;
 

src/Cafe/HW/Latte/LegacyShaderDecompiler/LatteDecompilerEmitMSL.cpp

@@ -20,6 +20,8 @@
 
 #define _CRLF	"\r\n"
 
+static bool rasterizationEnabled;
+
 void LatteDecompiler_emitAttributeDecodeMSL(LatteDecompilerShader* shaderContext, StringBuf* src, LatteParsedFetchShaderAttribute_t* attrib);
 
 /*
@@ -3108,6 +3110,9 @@ static void _emitExportGPRReadCode(LatteDecompilerShaderContext* shaderContext,
 
 static void _emitExportCode(LatteDecompilerShaderContext* shaderContext, LatteDecompilerCFInstruction* cfInstruction)
 {
+    if (!rasterizationEnabled)
+        return;
+
 	StringBuf* src = shaderContext->shaderSource;
 	src->add("// export" _CRLF);
 	if(shaderContext->shaderType == LatteConst::ShaderType::Vertex )
@@ -3332,6 +3337,9 @@ static void _emitCFRingWriteCode(LatteDecompilerShaderContext* shaderContext, La
 		return;
 	}
 
+	if (!rasterizationEnabled)
+	    return;
+
 	if (shaderContext->shaderType == LatteConst::ShaderType::Vertex)
 	{
 		if (cfInstruction->memWriteElemSize != 3)
@@ -3861,6 +3869,23 @@ void LatteDecompiler_emitMSLShader(LatteDecompilerShaderContext* shaderContext,
 {
     bool isRectVertexShader = (static_cast<LattePrimitiveMode>(shaderContext->contextRegisters[mmVGT_PRIMITIVE_TYPE]) == LattePrimitiveMode::RECTS);
 
+    // Rasterization
+    rasterizationEnabled = true;
+    if (shader->shaderType == LatteConst::ShaderType::Vertex)
+    {
+        rasterizationEnabled = !shaderContext->contextRegistersNew->PA_CL_CLIP_CNTL.get_DX_RASTERIZATION_KILL();
+
+    	// HACK
+    	if (!shaderContext->contextRegistersNew->PA_CL_VTE_CNTL.get_VPORT_X_OFFSET_ENA())
+    		rasterizationEnabled = true;
+
+    	const auto& polygonControlReg = shaderContext->contextRegistersNew->PA_SU_SC_MODE_CNTL;
+    	uint32 cullFront = polygonControlReg.get_CULL_FRONT();
+    	uint32 cullBack = polygonControlReg.get_CULL_BACK();
+    	if (cullFront && cullBack)
+    	    rasterizationEnabled = false;
+    }
+
 	StringBuf* src = new StringBuf(1024*1024*12); // reserve 12MB for generated source (we resize-to-fit at the end)
 	shaderContext->shaderSource = src;
 
@@ -3874,7 +3899,7 @@ void LatteDecompiler_emitMSLShader(LatteDecompilerShaderContext* shaderContext,
     src->add("#include <metal_stdlib>" _CRLF);
     src->add("using namespace metal;" _CRLF);
 	// header part (definitions for inputs and outputs)
-	LatteDecompiler::emitHeader(shaderContext, isRectVertexShader);
+	LatteDecompiler::emitHeader(shaderContext, isRectVertexShader, rasterizationEnabled);
 	// helper functions
 	LatteDecompiler_emitHelperFunctions(shaderContext, src);
 	const char* functionType = "";
@@ -4010,7 +4035,10 @@ void LatteDecompiler_emitMSLShader(LatteDecompilerShaderContext* shaderContext,
 		else
 		{
     	    functionType = "vertex";
+            if (rasterizationEnabled)
        	        outputTypeName = "VertexOut";
+            else
+                outputTypeName = "void";
 		}
 		break;
 	case LatteConst::ShaderType::Geometry:
@@ -4048,6 +4076,7 @@ void LatteDecompiler_emitMSLShader(LatteDecompilerShaderContext* shaderContext,
 	}
 	else
 	{
+	    if (rasterizationEnabled)
 	        src->addFmt("{} out;" _CRLF, outputTypeName);
 	}
 	// variable definition
@@ -4285,9 +4314,9 @@ void LatteDecompiler_emitMSLShader(LatteDecompilerShaderContext* shaderContext,
 	//if(shader->shaderType == LatteConst::ShaderType::Geometry)
 	//	src->add("EndPrimitive();" _CRLF);
 	// vertex shader should write renderstate point size at the end if required but not modified by shader
-	if (shaderContext->analyzer.outputPointSize && shaderContext->analyzer.writesPointSize == false)
+	if (shaderContext->analyzer.outputPointSize && !shaderContext->analyzer.writesPointSize)
 	{
-		if (shader->shaderType == LatteConst::ShaderType::Vertex && !shaderContext->options->usesGeometryShader)
+		if (shader->shaderType == LatteConst::ShaderType::Vertex && !shaderContext->options->usesGeometryShader && rasterizationEnabled)
 			src->add("out.pointSize = supportBuffer.pointSize;" _CRLF);
 	}
 
@@ -4325,13 +4354,15 @@ void LatteDecompiler_emitMSLShader(LatteDecompilerShaderContext* shaderContext,
         }
 	}
 
-    // TODO: this should be handled outside of the shader, because clipping currently wouldn't work (or would it?)
+    if (rasterizationEnabled)
-	if ((shader->shaderType == LatteConst::ShaderType::Vertex && !shaderContext->options->usesGeometryShader) || shader->shaderType == LatteConst::ShaderType::Geometry)
+    {
+    	if (shader->shaderType == LatteConst::ShaderType::Vertex && !shaderContext->options->usesGeometryShader)
     		src->add("out.position.z = (out.position.z + out.position.w) / 2.0;" _CRLF);
 
     	// Return
     	if (!(shaderContext->options->usesGeometryShader || isRectVertexShader) || shader->shaderType == LatteConst::ShaderType::Pixel)
             src->add("return out;" _CRLF);
+    }
 
 	// end of shader main
 	src->add("}" _CRLF);

src/Cafe/HW/Latte/LegacyShaderDecompiler/LatteDecompilerEmitMSLHeader.hpp

@@ -262,7 +262,7 @@ namespace LatteDecompiler
 		src->add("};" _CRLF _CRLF);
 	}
 
-	static void _emitInputsAndOutputs(LatteDecompilerShaderContext* decompilerContext, bool isRectVertexShader)
+	static void _emitInputsAndOutputs(LatteDecompilerShaderContext* decompilerContext, bool isRectVertexShader, bool rasterizationEnabled)
 	{
 		auto src = decompilerContext->shaderSource;
 
@@ -300,7 +300,7 @@ namespace LatteDecompiler
 
 		if (!decompilerContext->options->usesGeometryShader)
 		{
-    		if (decompilerContext->shaderType == LatteConst::ShaderType::Vertex)
+    		if (decompilerContext->shaderType == LatteConst::ShaderType::Vertex && rasterizationEnabled)
     			_emitVSOutputs(decompilerContext, isRectVertexShader);
 		}
 		else
@@ -351,7 +351,7 @@ namespace LatteDecompiler
 		}
 	}
 
-	static void emitHeader(LatteDecompilerShaderContext* decompilerContext, bool isRectVertexShader)
+	static void emitHeader(LatteDecompilerShaderContext* decompilerContext, bool isRectVertexShader, bool rasterizationEnabled)
 	{
 	    auto src = decompilerContext->shaderSource;
 
@@ -410,7 +410,7 @@ namespace LatteDecompiler
 		// uniform buffers
 		_emitUniformBuffers(decompilerContext);
 		// inputs and outputs
-		_emitInputsAndOutputs(decompilerContext, isRectVertexShader);
+		_emitInputsAndOutputs(decompilerContext, isRectVertexShader, rasterizationEnabled);
 
 		if (dump_shaders_enabled)
 			decompilerContext->shaderSource->add("// end of shader inputs/outputs" _CRLF);

src/Cafe/HW/Latte/Renderer/Metal/MetalPipelineCache.cpp

@@ -9,6 +9,7 @@
 
 #include "Cafe/HW/Latte/Core/FetchShader.h"
 #include "Cafe/HW/Latte/ISA/RegDefines.h"
+#include "Cemu/Logging/CemuLogging.h"
 #include "config/ActiveSettings.h"
 
 static void rectsEmulationGS_outputSingleVertex(std::string& gsSrc, const LatteDecompilerShader* vertexShader, LatteShaderPSInputTable* psInputTable, sint32 vIdx, const LatteContextRegister& latteRegister)
@@ -188,8 +189,31 @@ extern std::atomic_int g_compiled_shaders_total;
 extern std::atomic_int g_compiled_shaders_async;
 
 template<typename T>
-void SetFragmentState(T* desc, CachedFBOMtl* lastUsedFBO, CachedFBOMtl* activeFBO, const LatteContextRegister& lcr)
+void SetFragmentState(T* desc, CachedFBOMtl* lastUsedFBO, CachedFBOMtl* activeFBO, const LatteDecompilerShader* pixelShader, const LatteContextRegister& lcr)
 {
+	// Rasterization
+	bool rasterizationEnabled = !LatteGPUState.contextNew.PA_CL_CLIP_CNTL.get_DX_RASTERIZATION_KILL();
+
+	// HACK
+	// TODO: include this in the hash?
+	if (!LatteGPUState.contextNew.PA_CL_VTE_CNTL.get_VPORT_X_OFFSET_ENA())
+		rasterizationEnabled = true;
+
+	// Culling both front and back faces effectively disables rasterization
+	const auto& polygonControlReg = LatteGPUState.contextNew.PA_SU_SC_MODE_CNTL;
+	uint32 cullFront = polygonControlReg.get_CULL_FRONT();
+	uint32 cullBack = polygonControlReg.get_CULL_BACK();
+	if (cullFront && cullBack)
+	    rasterizationEnabled = false;
+
+	desc->setRasterizationEnabled(rasterizationEnabled);
+
+	if (rasterizationEnabled)
+	{
+	    auto pixelShaderMtl = static_cast<RendererShaderMtl*>(pixelShader->shader);
+		desc->setFragmentFunction(pixelShaderMtl->GetFunction());
+	}
+
     // Color attachments
 	const Latte::LATTE_CB_COLOR_CONTROL& colorControlReg = lcr.CB_COLOR_CONTROL;
 	uint32 blendEnableMask = colorControlReg.get_BLEND_MASK();
@@ -284,7 +308,7 @@ MetalPipelineCache::~MetalPipelineCache()
     m_binaryArchive->serializeToURL(m_binaryArchiveURL, &error);
     if (error)
     {
-        debug_printf("failed to serialize binary archive: %s\n", error->localizedDescription()->utf8String());
+        cemuLog_log(LogType::Force, "failed to serialize binary archive: {}", error->localizedDescription()->utf8String());
         error->release();
     }
     m_binaryArchive->release();
@@ -362,17 +386,15 @@ MTL::RenderPipelineState* MetalPipelineCache::GetRenderPipelineState(const Latte
 		layout->setStride(bufferStride);
 	}
 
-	auto mtlVertexShader = static_cast<RendererShaderMtl*>(vertexShader->shader);
+	auto vertexShaderMtl = static_cast<RendererShaderMtl*>(vertexShader->shader);
-	auto mtlPixelShader = static_cast<RendererShaderMtl*>(pixelShader->shader);
 
 	// Render pipeline state
 	MTL::RenderPipelineDescriptor* desc = MTL::RenderPipelineDescriptor::alloc()->init();
-	desc->setVertexFunction(mtlVertexShader->GetFunction());
+	desc->setVertexFunction(vertexShaderMtl->GetFunction());
-	desc->setFragmentFunction(mtlPixelShader->GetFunction());
 	// TODO: don't always set the vertex descriptor?
 	desc->setVertexDescriptor(vertexDescriptor);
 
-	SetFragmentState(desc, lastUsedFBO, activeFBO, lcr);
+	SetFragmentState(desc, lastUsedFBO, activeFBO, pixelShader, lcr);
 
 	TryLoadBinaryArchive();
 
@@ -391,9 +413,6 @@ MTL::RenderPipelineState* MetalPipelineCache::GetRenderPipelineState(const Latte
 #endif
 	pipeline = m_mtlr->GetDevice()->newRenderPipelineState(desc, MTL::PipelineOptionFailOnBinaryArchiveMiss, nullptr, &error);
 
-	//static uint32 oldPipelineCount = 0;
-	//static uint32 newPipelineCount = 0;
-
 	// Pipeline wasn't found in the binary archive, we need to compile it
 	if (error)
 	{
@@ -407,7 +426,7 @@ MTL::RenderPipelineState* MetalPipelineCache::GetRenderPipelineState(const Latte
 	    pipeline = m_mtlr->GetDevice()->newRenderPipelineState(desc, &error);
 		if (error)
 		{
-		    debug_printf("error creating render pipeline state: %s\n", error->localizedDescription()->utf8String());
+		    cemuLog_log(LogType::Force, "error creating render pipeline state: {}", error->localizedDescription()->utf8String());
 			error->release();
 		}
 		else
@@ -419,19 +438,12 @@ MTL::RenderPipelineState* MetalPipelineCache::GetRenderPipelineState(const Latte
                 m_binaryArchive->addRenderPipelineFunctions(desc, &error);
                 if (error)
                 {
-                    debug_printf("error saving render pipeline functions: %s\n", error->localizedDescription()->utf8String());
+                    cemuLog_log(LogType::Force, "error saving render pipeline functions: {}", error->localizedDescription()->utf8String());
                     error->release();
                 }
 			}
 		}
-
-		//newPipelineCount++;
 	}
-	//else
-    //{
-    //    oldPipelineCount++;
-    //}
-    //debug_printf("%u pipelines were found in the binary archive, %u new were created\n", oldPipelineCount, newPipelineCount);
 	desc->release();
 	vertexDescriptor->release();
 
@@ -452,26 +464,24 @@ MTL::RenderPipelineState* MetalPipelineCache::GetMeshPipelineState(const LatteFe
     if (pipeline)
         return pipeline;
 
-	auto mtlObjectShader = static_cast<RendererShaderMtl*>(vertexShader->shader);
+	auto objectShaderMtl = static_cast<RendererShaderMtl*>(vertexShader->shader);
-	RendererShaderMtl* mtlMeshShader;
+	RendererShaderMtl* meshShaderMtl;
 	if (geometryShader)
 	{
-        mtlMeshShader = static_cast<RendererShaderMtl*>(geometryShader->shader);
+        meshShaderMtl = static_cast<RendererShaderMtl*>(geometryShader->shader);
 	}
     else
     {
         // If there is no geometry shader, it means that we are emulating rects
-        mtlMeshShader = rectsEmulationGS_generate(m_mtlr, vertexShader, lcr);
+        meshShaderMtl = rectsEmulationGS_generate(m_mtlr, vertexShader, lcr);
     }
-	auto mtlPixelShader = static_cast<RendererShaderMtl*>(pixelShader->shader);
 
 	// Render pipeline state
 	MTL::MeshRenderPipelineDescriptor* desc = MTL::MeshRenderPipelineDescriptor::alloc()->init();
-	desc->setObjectFunction(mtlObjectShader->GetFunction());
+	desc->setObjectFunction(objectShaderMtl->GetFunction());
-	desc->setMeshFunction(mtlMeshShader->GetFunction());
+	desc->setMeshFunction(meshShaderMtl->GetFunction());
-	desc->setFragmentFunction(mtlPixelShader->GetFunction());
 
-	SetFragmentState(desc, lastUsedFBO, activeFBO, lcr);
+	SetFragmentState(desc, lastUsedFBO, activeFBO, pixelShader, lcr);
 
 	TryLoadBinaryArchive();
 
@@ -486,7 +496,7 @@ MTL::RenderPipelineState* MetalPipelineCache::GetMeshPipelineState(const LatteFe
 	desc->release();
 	if (error)
 	{
-    	debug_printf("error creating render pipeline state: %s\n", error->localizedDescription()->utf8String());
+    	cemuLog_log(LogType::Force, "error creating mesh render pipeline state: {}", error->localizedDescription()->utf8String());
         error->release();
 	}
 
@@ -594,7 +604,7 @@ void MetalPipelineCache::TryLoadBinaryArchive()
         m_binaryArchive = m_mtlr->GetDevice()->newBinaryArchive(desc, &error);
         if (error)
         {
-            debug_printf("failed to create binary archive: %s\n", error->localizedDescription()->utf8String());
+            cemuLog_log(LogType::Force, "failed to create binary archive: {}", error->localizedDescription()->utf8String());
             error->release();
         }
     }

src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp

@@ -879,7 +879,7 @@ void MetalRenderer::draw_beginSequence()
 	LatteRenderTarget_updateScissorBox();
 
 	// check for conditions which would turn the drawcalls into no-ops
-	bool rasterizerEnable = LatteGPUState.contextNew.PA_CL_CLIP_CNTL.get_DX_RASTERIZATION_KILL() == false;
+	bool rasterizerEnable = !LatteGPUState.contextNew.PA_CL_CLIP_CNTL.get_DX_RASTERIZATION_KILL();
 
 	// GX2SetSpecialState(0, true) enables DX_RASTERIZATION_KILL, but still expects depth writes to happen? -> Research which stages are disabled by DX_RASTERIZATION_KILL exactly
 	// for now we use a workaround:
@@ -888,18 +888,6 @@ void MetalRenderer::draw_beginSequence()
 
 	if (!rasterizerEnable && !streamoutEnable)
 		m_state.m_skipDrawSequence = true;
-
-	// Both faces are culled
-	// TODO: can we really skip the draw?
-	const auto& polygonControlReg = LatteGPUState.contextNew.PA_SU_SC_MODE_CNTL;
-	uint32 cullFront = polygonControlReg.get_CULL_FRONT();
-	uint32 cullBack = polygonControlReg.get_CULL_BACK();
-	if (cullFront && cullBack)
-	    m_state.m_skipDrawSequence = true;
-
-	// TODO: is this even needed?
-	if (!m_state.m_activeFBO)
-	    m_state.m_skipDrawSequence = true;
 }
 
 void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32 instanceCount, uint32 count, MPTR indexDataMPTR, Latte::LATTE_VGT_DMA_INDEX_TYPE::E_INDEX_TYPE indexType, bool isFirst)
@@ -1065,10 +1053,9 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
 
 	// Cull mode
 
-	// Handled in draw_beginSequence
+	// Cull front and back is handled by disabling rasterization
-	if (cullFront && cullBack)
+	if (!(cullFront && cullBack))
-	    cemu_assert_suspicious();
+	{
-
         MTL::CullMode cullMode;
        	if (cullFront)
       		cullMode = MTL::CullModeFront;
@@ -1082,6 +1069,7 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
        	    renderCommandEncoder->setCullMode(cullMode);
       		encoderState.m_cullMode = cullMode;
        	}
+	}
 
 	// Front face
 	MTL::Winding frontFaceWinding;
@@ -1164,12 +1152,8 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
 	else
         renderPipelineState = m_pipelineCache->GetRenderPipelineState(fetchShader, vertexShader, pixelShader, m_state.m_lastUsedFBO, m_state.m_activeFBO, LatteGPUState.contextNew);
 
-    // HACK
     if (!renderPipelineState)
-    {
-        printf("invalid render pipeline state, skipping draw\n");
         return;
-    }
 
 	if (renderPipelineState != encoderState.m_renderPipelineState)
    	{