Implement GPUREG_VSH_OUTMAP_MASK

include/PICA/gpu.hpp

@@ -6,6 +6,7 @@
 #include "PICA/pica_vertex.hpp"
 #include "PICA/regs.hpp"
 #include "PICA/shader_unit.hpp"
+#include "compiler_builtins.hpp"
 #include "config.hpp"
 #include "helpers.hpp"
 #include "logger.hpp"
@@ -35,6 +36,12 @@ class GPU {
 
 	std::array<vec4f, 16> immediateModeAttributes;  // Vertex attributes uploaded via immediate mode submission
 	std::array<PICA::Vertex, 3> immediateModeVertices;
+
+	// Pointers for the output registers as arranged after GPUREG_VSH_OUTMAP_MASK is applied
+	std::array<Floats::f24*, 16> vsOutputRegisters;
+	// Previous value for GPUREG_VSH_OUTMAP_MASK
+	u32 oldVsOutputMask;
+
 	uint immediateModeVertIndex;
 	uint immediateModeAttrIndex;  // Index of the immediate mode attribute we're uploading
 
@@ -167,4 +174,28 @@ class GPU {
 	// We have them in the end of the struct for cache locality reasons. Tl;dr we want the more commonly used things to be packed in the start
 	// Of the struct, instead of externalRegs being in the middle
 	ExternalRegisters externalRegs;
+
+	ALWAYS_INLINE void setVsOutputMask(u32 val) {
+		val &= 0xffff;
+	
+		// Avoid recomputing this if not necessary 
+		if (oldVsOutputMask != val) [[unlikely]] {
+			oldVsOutputMask = val;
+
+			uint count = 0;
+			// See which registers are actually enabled and ignore the disabled ones
+			for (int i = 0; i < 16; i++) {
+				if (val & 1) {
+					vsOutputRegisters[count++] = &shaderUnit.vs.outputs[i][0];
+				}
+
+				val >>= 1;
+			}
+
+			// For the others, map the index to a vs output directly (TODO: What does hw actually do?)
+			for (; count < 16; count++) {
+				vsOutputRegisters[count] = &shaderUnit.vs.outputs[count][0];
+			}
+		}
+	}
 };

include/PICA/regs.hpp

@@ -143,6 +143,7 @@ namespace PICA {
 			VertexIntUniform3 = 0x2B4,
 
 			VertexShaderEntrypoint = 0x2BA,
+			VertexShaderOutputMask = 0x2BD,
 			VertexShaderTransferEnd = 0x2BF,
 			VertexFloatUniformIndex = 0x2C0,
 			VertexFloatUniformData0 = 0x2C1,

src/core/PICA/gpu.cpp

@@ -77,6 +77,9 @@ void GPU::reset() {
 
 	fixedAttrBuff.fill(0);
 
+	oldVsOutputMask = 0;
+	setVsOutputMask(0xFFFF);
+
 	for (auto& e : attributeInfo) {
 		e.offset = 0;
 		e.size = 0;
@@ -134,6 +137,8 @@ void GPU::drawArrays() {
 		shaderJIT.prepare(shaderUnit.vs);
 	}
 
+	setVsOutputMask(regs[PICA::InternalRegs::VertexShaderOutputMask]);
+
 	// Base address for vertex attributes
 	// The vertex base is always on a quadword boundary because the PICA does weird alignment shit any time possible
 	const u32 vertexBase = ((regs[PICA::InternalRegs::VertexAttribLoc] >> 1) & 0xfffffff) * 16;
@@ -329,7 +334,7 @@ void GPU::drawArrays() {
 
 			for (int j = 0; j < 4; j++) {  // pls unroll
 				const u32 mapping = (config >> (j * 8)) & 0x1F;
-				out.raw[mapping] = shaderUnit.vs.outputs[i][j];
+				out.raw[mapping] = vsOutputRegisters[i][j];
 			}
 		}
 	}
@@ -338,6 +343,8 @@ void GPU::drawArrays() {
 }
 
 PICA::Vertex GPU::getImmediateModeVertex() {
+	setVsOutputMask(regs[PICA::InternalRegs::VertexShaderOutputMask]);
+
 	PICA::Vertex v;
 	const int totalAttrCount = (regs[PICA::InternalRegs::VertexShaderAttrNum] & 0xf) + 1;
 
@@ -356,7 +363,7 @@ PICA::Vertex GPU::getImmediateModeVertex() {
 
 		for (int j = 0; j < 4; j++) {  // pls unroll
 			const u32 mapping = (config >> (j * 8)) & 0x1F;
-			v.raw[mapping] = shaderUnit.vs.outputs[i][j];
+			v.raw[mapping] = vsOutputRegisters[i][j];
 		}
 	}