fix: writing to depth from a fragment shader

Author: SamoZ256

src/Cafe/HW/Latte/LegacyShaderDecompiler/LatteDecompiler.h

@@ -182,6 +182,8 @@ struct LatteDecompilerShader
 
 	// analyzer stage (pixel outputs)
 	uint32 pixelColorOutputMask{ 0 }; // from LSB to MSB, 1 bit per written output. 1 if written (indices of color attachments)
+	// analyzer stage (depth write)
+	bool depthWritten{ false };
 	// analyzer stage (geometry shader parameters/inputs)
 	uint32 ringParameterCount{ 0 };
 	uint32 ringParameterCountFromPrevStage{ 0 }; // used in geometry shader to hold VS ringParameterCount

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

@@ -287,15 +287,15 @@ void LatteDecompiler_analyzeTEXClause(LatteDecompilerShaderContext* shaderContex
 	LatteDecompilerShader* shader = shaderContext->shader;
 	for(auto& texInstruction : cfInstruction->instructionsTEX)
 	{
-		if( texInstruction.opcode == GPU7_TEX_INST_SAMPLE || 
-			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_L || 
-			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_LB || 
-			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_LZ || 
-			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_C || 
+		if( texInstruction.opcode == GPU7_TEX_INST_SAMPLE ||
+			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_L ||
+			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_LB ||
+			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_LZ ||
+			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_C ||
 			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_C_L ||
 			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_C_LZ ||
-			texInstruction.opcode == GPU7_TEX_INST_FETCH4 || 
-			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_G || 
+			texInstruction.opcode == GPU7_TEX_INST_FETCH4 ||
+			texInstruction.opcode == GPU7_TEX_INST_SAMPLE_G ||
 			texInstruction.opcode == GPU7_TEX_INST_LD )
 		{
 			if (texInstruction.textureFetch.textureIndex < 0 || texInstruction.textureFetch.textureIndex >= LATTE_NUM_MAX_TEX_UNITS)
@@ -313,7 +313,7 @@ void LatteDecompiler_analyzeTEXClause(LatteDecompilerShaderContext* shaderContex
 			shader->textureUnitSamplerAssignment[texInstruction.textureFetch.textureIndex] = texInstruction.textureFetch.samplerIndex;
 			if( texInstruction.opcode == GPU7_TEX_INST_SAMPLE_C || texInstruction.opcode == GPU7_TEX_INST_SAMPLE_C_L || texInstruction.opcode == GPU7_TEX_INST_SAMPLE_C_LZ)
 				shader->textureUsesDepthCompare[texInstruction.textureFetch.textureIndex] = true;
-			
+
 			bool useTexelCoords = false;
 			if (texInstruction.opcode == GPU7_TEX_INST_SAMPLE && (texInstruction.textureFetch.unnormalized[0] && texInstruction.textureFetch.unnormalized[1] && texInstruction.textureFetch.unnormalized[2] && texInstruction.textureFetch.unnormalized[3]))
 				useTexelCoords = true;
@@ -393,7 +393,7 @@ void LatteDecompiler_analyzeExport(LatteDecompilerShaderContext* shaderContext,
 		}
 		else if( cfInstruction->exportType == 0 && cfInstruction->exportArrayBase == 61 )
 		{
-			// writes pixel depth
+			shader->depthWritten = true;
 		}
 		else
 			debugBreakpoint();
@@ -419,7 +419,7 @@ void LatteDecompiler_analyzeExport(LatteDecompilerShaderContext* shaderContext,
 void LatteDecompiler_analyzeSubroutine(LatteDecompilerShaderContext* shaderContext, uint32 cfAddr)
 {
 	// analyze CF and clauses up to RET statement
-	
+
 	// todo - find cfInstruction index from cfAddr
 	cemu_assert_debug(false);
 
@@ -505,9 +505,9 @@ namespace LatteDecompiler
 			decompilerContext->hasUniformVarBlock = true;
 		else if (decompilerContext->shader->uniformMode == LATTE_DECOMPILER_UNIFORM_MODE_FULL_CFILE)
 			decompilerContext->hasUniformVarBlock = true;
-		
-		bool hasAnyViewportScaleDisabled = 
-			!decompilerContext->contextRegistersNew->PA_CL_VTE_CNTL.get_VPORT_X_SCALE_ENA() || 
+
+		bool hasAnyViewportScaleDisabled =
+			!decompilerContext->contextRegistersNew->PA_CL_VTE_CNTL.get_VPORT_X_SCALE_ENA() ||
 			!decompilerContext->contextRegistersNew->PA_CL_VTE_CNTL.get_VPORT_Y_SCALE_ENA() ||
 			!decompilerContext->contextRegistersNew->PA_CL_VTE_CNTL.get_VPORT_Z_SCALE_ENA();
 		// we currently only support all on/off. Individual component scaling is not supported
@@ -803,7 +803,7 @@ void LatteDecompiler_analyze(LatteDecompilerShaderContext* shaderContext, LatteD
 
 	for(sint32 i=0; i<LATTE_NUM_MAX_TEX_UNITS; i++)
 	{
-		if (!shaderContext->output->textureUnitMask[i]) 
+		if (!shaderContext->output->textureUnitMask[i])
 		{
 			// texture unit not used
 			shader->textureUnitDim[i] = (Latte::E_DIM)0xFF;

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

@@ -2441,11 +2441,12 @@ static void _emitTEXSampleTextureCode(LatteDecompilerShaderContext* shaderContex
 				src->add(")");
 			}
 		}
-		else if( texOpcode == GPU7_TEX_INST_SAMPLE_LZ || texOpcode == GPU7_TEX_INST_SAMPLE_C_LZ )
-		{
-		    // TODO: correct?
-			src->add(", level(0.0)");
-		}
+		// TODO: uncomment?
+		//else if( texOpcode == GPU7_TEX_INST_SAMPLE_LZ || texOpcode == GPU7_TEX_INST_SAMPLE_C_LZ )
+		//{
+		//    // TODO: correct?
+		//	src->add(", level(0.0)");
+		//}
 	}
 	// gradient parameters
 	if (texOpcode == GPU7_TEX_INST_SAMPLE_G)
@@ -3215,7 +3216,7 @@ static void _emitExportCode(LatteDecompilerShaderContext* shaderContext, LatteDe
 				cemu_assert_unimplemented(); // ukn
 			}
 
-			src->add("out.depth = ");
+			src->add("out.passDepth = ");
 			_emitExportGPRReadCode(shaderContext, cfInstruction, LATTE_DECOMPILER_DTYPE_FLOAT, 0);
 			src->add(".x");
 			src->add(";" _CRLF);

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

@@ -272,9 +272,9 @@ namespace LatteDecompiler
             }
 
             // generate depth output for pixel shader
-            if (decompilerContext->shader->pixelDepthOutputMask)
+            if (decompilerContext->shader->depthWritten)
             {
-                src->add("float passDepth [[depth(any)]];" _CRLF);
+                src->add("float passDepth [[depth]];" _CRLF);
             }
 
             src->add("};" _CRLF _CRLF);