Pipeline checker: Add support for checking for identical pipelines with different vertex formats

This turned out not to be very fruitful.

GPU/Common/DrawEngineCommon.cpp

@@ -999,7 +999,7 @@ bool DrawEngineCommon::DescribeCodePtr(const u8 *ptr, std::string &name) const {
 
 	if (found) {
 		char temp[256];
-		found->ToString(temp, false);
+		found->ToString(temp, sizeof(temp), false);
 		name = temp;
 		snprintf(temp, sizeof(temp), "_%08X", foundKey);
 		name += temp;

GPU/Common/VertexDecoderArm.cpp

@@ -259,7 +259,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 			// Reset the code ptr and return zero to indicate that we failed.
 			ResetCodePtr(GetOffset(start));
 			char temp[1024]{};
-			dec.ToString(temp, true);
+			dec.ToString(temp, sizeof(temp), true);
 			WARN_LOG(Log::G3D, "Could not compile vertex decoder, failed at step %s: %s", GetStepFunctionName(dec.steps_[i]), temp);
 			return 0;
 		}
@@ -287,7 +287,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 	/*
 	DisassembleArm(start, GetCodePtr() - start);
 	char temp[1024] = {0};
-	dec.ToString(temp, true);
+	dec.ToString(temp, sizeof(temp), true);
 	INFO_LOG(Log::G3D, "%s", temp);
 	*/
 

GPU/Common/VertexDecoderArm64.cpp

@@ -256,7 +256,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 			// Reset the code ptr (effectively undoing what we generated) and return zero to indicate that we failed.
 			ResetCodePtr(GetOffset(start));
 			char temp[1024]{};
-			dec.ToString(temp, true);
+			dec.ToString(temp, sizeof(temp), true);
 			WARN_LOG(Log::G3D, "Could not compile vertex decoder, failed at step %s: %s", GetStepFunctionName(dec.steps_[i]), temp);
 			return nullptr;
 		}
@@ -291,7 +291,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 
 	if (log) {
 		char temp[1024] = { 0 };
-		dec.ToString(temp, true);
+		dec.ToString(temp, sizeof(temp), true);
 		INFO_LOG(Log::JIT, "=== %s (%d bytes) ===", temp, (int)(GetCodePtr() - start));
 		std::vector<std::string> lines = DisassembleArm64(start, (int)(GetCodePtr() - start));
 		for (auto line : lines) {

GPU/Common/VertexDecoderCommon.cpp

@@ -23,6 +23,7 @@
 
 #include "Common/CommonTypes.h"
 #include "Common/Data/Convert/ColorConv.h"
+#include "Common/Data/Text/StringWriter.h"
 #include "Common/Math/CrossSIMD.h"
 #include "Common/Log.h"
 #include "Common/LogReporting.h"
@@ -84,6 +85,41 @@ static int DecFmtSize(u8 fmt) {
 	}
 }
 
+const char *DecFmtComponentToString(u8 fmt) {
+	switch (fmt) {
+	case DEC_NONE: return "NONE";
+	case DEC_FLOAT_1: return "FLOAT_1";
+	case DEC_FLOAT_2: return "FLOAT_2";
+	case DEC_FLOAT_3: return "FLOAT_3";
+	case DEC_FLOAT_4: return "FLOAT_4";
+	case DEC_S8_3: return "S8_3";
+	case DEC_S16_3: return "S16_3";
+	case DEC_U8_1:  return "U8_1";
+	case DEC_U8_2:  return "U8_2";
+	case DEC_U8_3:  return "U8_3";
+	case DEC_U8_4:  return "U8_4";
+	case DEC_U16_1: return "U16_1";
+	case DEC_U16_2: return "U16_2";
+	case DEC_U16_3: return "U16_3";
+	case DEC_U16_4: return "U16_4";
+	default: return "UNKNOWN";
+	}
+}
+
+std::string DecVtxFormat::ToString() const {
+	char buf[256];
+	StringWriter w(buf, sizeof(buf));
+	if (w0fmt) {
+		w.F("W0: %s ", DecFmtComponentToString(w0fmt));
+	}
+	w.F("W1: %s ", DecFmtComponentToString(w1fmt));
+	w.F("UV: %s ", DecFmtComponentToString(uvfmt));
+	w.F("C0: %s ", DecFmtComponentToString(c0fmt));
+	w.F("C1: %s ", DecFmtComponentToString(c1fmt));
+	w.F("N: %s", DecFmtComponentToString(nrmfmt));
+	return w.as_string();
+}
+
 void DecVtxFormat::ComputeID() {
 	id = w0fmt | (w1fmt << 4) | (uvfmt << 8) | (c0fmt << 12) | (c1fmt << 16) | (nrmfmt << 20);
 }
@@ -1413,7 +1449,7 @@ void VertexDecoder::SetVertexType(u32 fmt, const VertexDecoderOptions &options,
 
 	if (reportNoPos) {
 		char temp[256]{};
-		ToString(temp, true);
+		ToString(temp, sizeof(temp), true);
 		ERROR_LOG(Log::G3D, "Vertices without position found (and ignored): (%08x) %s", fmt_, temp);
 	}
 
@@ -1550,8 +1586,8 @@ void VertexDecoder::CompareToJit(const u8 *startPtr, u8 *decodedptr, int count,
 		controlReader.Goto(i);
 		jittedReader.Goto(i);
 		if (!DecodedVertsAreSimilar(controlReader, jittedReader)) {
-			char name[512]{};
+			char name[256]{};
-			ToString(name, true);
+			ToString(name, sizeof(name), true);
 			ERROR_LOG(Log::G3D, "Encountered vertexjit mismatch at %d/%d for %s", i, count, name);
 			if (morphcount > 1) {
 				printf("Morph:\n");
@@ -1605,26 +1641,30 @@ static const char * const idxnames[4] = { "-", "u8", "u16", "?" };
 static const char * const weightnames[4] = { "-", "u8", "u16", "f" };
 static const char * const colnames[8] = { "", "?", "?", "?", "565", "5551", "4444", "8888" };
 
-int VertexDecoder::ToString(char *output, bool spaces) const {
+// There's a direct ToString function for vertex types in the GPU debugger, GeDescribeVertexType.
-	char *start = output;
-	output += sprintf(output, "[%08x] ", fmt_);
-	output += sprintf(output, "P: %s ", posnames[pos]);
-	if (nrm)
-		output += sprintf(output, "N: %s ", nrmnames[nrm]);
-	if (col)
-		output += sprintf(output, "C: %s ", colnames[col]);
-	if (tc)
-		output += sprintf(output, "T: %s ", tcnames[tc]);
-	if (weighttype)
-		output += sprintf(output, "W: %s (%ix) ", weightnames[weighttype], nweights);
-	if (idx)
-		output += sprintf(output, "I: %s ", idxnames[idx]);
-	if (morphcount > 1)
-		output += sprintf(output, "Morph: %i ", morphcount);
-	if (throughmode)
-		output += sprintf(output, " (through)");
 
-	output += sprintf(output, " (%ib)", VertexSize());
+int VertexDecoder::ToString(char *buf, int bufSize, bool spaces) const {
+	StringWriter w(buf, bufSize);
+
+	char *start = buf;
+	w.F("[%08x] ", fmt_);
+	w.F("P: %s ", posnames[pos]);
+	if (nrm)
+		w.F("N: %s ", nrmnames[nrm]);
+	if (col)
+		w.F("C: %s ", colnames[col]);
+	if (tc)
+		w.F("T: %s ", tcnames[tc]);
+	if (weighttype)
+		w.F("W: %s (%ix) ", weightnames[weighttype], nweights);
+	if (idx)
+		w.F("I: %s ", idxnames[idx]);
+	if (morphcount > 1)
+		w.F("Morph: %i ", morphcount);
+	if (throughmode)
+		w.F(" (through)");
+
+	w.F(" (%ib)", VertexSize());
 
 	if (!spaces) {
 		size_t len = strlen(start);
@@ -1635,17 +1675,16 @@ int VertexDecoder::ToString(char *output, bool spaces) const {
 	}
 
 #ifdef _DEBUG
-	output += sprintf(output, " (%llu)", (long long)decodedCount);
+	w.F(" (%llu)", (long long)decodedCount);
 #endif
-
+	return (int)w.size();
-	return output - start;
 }
 
 std::string VertexDecoder::GetString(DebugShaderStringType stringType) const {
 	char buffer[256];
 	switch (stringType) {
 	case SHADER_STRING_SHORT_DESC:
-		ToString(buffer, true);
+		ToString(buffer, sizeof(buffer), true);
 		return std::string(buffer);
 	case SHADER_STRING_SOURCE_CODE:
 		{

GPU/Common/VertexDecoderCommon.h

@@ -46,7 +46,7 @@
 
 // Keep this in 4 bits.
 enum {
-	DEC_NONE,
+	DEC_NONE = 0,
 	DEC_FLOAT_1,
 	DEC_FLOAT_2,
 	DEC_FLOAT_3,
@@ -63,6 +63,8 @@ enum {
 	DEC_U16_4,
 };
 
+const char *DecFmtComponentToString(u8 fmt);
+
 // DecVtxFormat - vertex formats for PC
 // Kind of like a D3D VertexDeclaration.
 // No morph support, that is taken care of by the VertexDecoder.
@@ -81,6 +83,8 @@ struct DecVtxFormat {
 	void InitializeFromID(uint32_t id);
 
 	static u8 PosFmt() { return DEC_FLOAT_3; }
+
+	std::string ToString() const;
 };
 
 void GetIndexBounds(const void *inds, int count, u32 vertType, u16 *indexLowerBound, u16 *indexUpperBound);
@@ -153,6 +157,8 @@ inline GETexMapMode VertTypeIDUVGenMode(uint32_t vertType) {
 	return (GETexMapMode)((vertType >> 24) & 3);
 }
 
+std::string DecFmtIdToString(u32 id);
+
 class VertexDecoder {
 public:
 	// A jit cache is not mandatory.
@@ -258,7 +264,7 @@ public:
 	// output must be big for safety.
 	// Returns number of chars written.
 	// Ugly for speed.
-	int ToString(char *output, bool spaces) const;
+	int ToString(char *buf, int bufSize, bool spaces) const;
 
 	bool IsInSpace(const uint8_t *ptr) const {
 		return ptr >= (const uint8_t *)jitted_ && ptr < ((const uint8_t *)jitted_ + jittedSize_);

GPU/Common/VertexDecoderLoongArch64.cpp

@@ -297,7 +297,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
             // Reset the code ptr (effectively undoing what we generated) and return zero to indicate that we failed.
             ResetCodePtr(GetOffset(start));
             char temp[1024]{};
-            dec.ToString(temp, true);
+            dec.ToString(temp, sizeof(temp), true);
             WARN_LOG(Log::G3D, "Could not compile vertex decoder, failed at step %s: %s", GetStepFunctionName(dec.steps_[i]), temp);
             return nullptr;
         }
@@ -338,7 +338,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 
 	if (log) {
 		char temp[1024]{};
-		dec.ToString(temp, true);
+		dec.ToString(temp, sizeof(temp), true);
 		INFO_LOG(Log::JIT, "=== %s (%d bytes) ===", temp, (int)(GetCodePtr() - start));
 		std::vector<std::string> lines = DisassembleLA64(start, (int)(GetCodePtr() - start));
 		for (auto line : lines) {

GPU/Common/VertexDecoderRiscV.cpp

@@ -316,7 +316,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 			// Reset the code ptr (effectively undoing what we generated) and return zero to indicate that we failed.
 			ResetCodePtr(GetOffset(start));
 			char temp[1024]{};
-			dec.ToString(temp, true);
+			dec.ToString(temp, sizeof(temp), true);
 			WARN_LOG(Log::G3D, "Could not compile vertex decoder, failed at step %s: %s", GetStepFunctionName(dec.steps_[i]), temp);
 			return nullptr;
 		}
@@ -349,7 +349,7 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 
 	if (log) {
 		char temp[1024]{};
-		dec.ToString(temp, true);
+		dec.ToString(temp, sizeof(temp), true);
 		INFO_LOG(Log::JIT, "=== %s (%d bytes) ===", temp, (int)(GetCodePtr() - start));
 		std::vector<std::string> lines = DisassembleRV64(start, (int)(GetCodePtr() - start));
 		for (auto line : lines) {

GPU/Common/VertexDecoderX86.cpp

@@ -281,8 +281,8 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 			EndWrite();
 			// Reset the code ptr and return zero to indicate that we failed.
 			ResetCodePtr(GetOffset(start));
-			char temp[1024]{};
+			char temp[256]{};
-			dec.ToString(temp, true);
+			dec.ToString(temp, sizeof(temp), true);
 			WARN_LOG(Log::G3D, "Could not compile vertex decoder, failed at step %s: %s", GetStepFunctionName(dec.steps_[i]), temp);
 			return 0;
 		}

GPU/Vulkan/PipelineManagerVulkan.h

@@ -45,7 +45,7 @@ struct VulkanPipelineKey {
 	VulkanPipelineRasterStateKey raster;  // prim is included here
 	VShaderID vid;
 	FShaderID fid;
-	uint32_t vtxFmtId;
+	uint32_t vtxFmtId;  // Note: This is the decoded format ID, not the source format!
 	bool useHWTransform;
 
 	void ToString(std::string *str) const {

UI/DevScreens.cpp

@@ -510,6 +510,10 @@ const bool IsSameExceptForShaders(const VulkanPipelineKey &a, const VulkanPipeli
 	return a.useHWTransform == b.useHWTransform && a.raster == b.raster && a.vtxFmtId == b.vtxFmtId;
 }
 
+const bool IsSameExceptForVertexFormat(const VulkanPipelineKey &a, const VulkanPipelineKey &b) {
+	return a.useHWTransform == b.useHWTransform && a.raster == b.raster && a.vid == b.vid && a.fid == b.fid;
+}
+
 void ShaderListScreen::AddPipelineAnalysis(UI::LinearLayout *tabContent) {
 #if !PPSSPP_PLATFORM(UWP)
 	GPU_Vulkan *vgpu = dynamic_cast<GPU_Vulkan *>(gpu);
@@ -538,6 +542,15 @@ void ShaderListScreen::AddPipelineAnalysis(UI::LinearLayout *tabContent) {
 					toBeMerged += "FShader: " + keyA.fid.Description() + " vs " + keyB.fid.Description() + "\n";
 				}
 			}
+			if (IsSameExceptForVertexFormat(keyA, keyB)) {
+				if (keyA.vtxFmtId != 0 && keyB.vtxFmtId != 0 && keyA.vtxFmtId != keyB.vtxFmtId) {
+					DecVtxFormat fmtA;
+					DecVtxFormat fmtB;
+					fmtA.InitializeFromID(keyA.vtxFmtId);
+					fmtB.InitializeFromID(keyB.vtxFmtId);
+					toBeMerged += "VertexFmt: " + fmtA.ToString() + " vs " + fmtB.ToString() + "\n";
+				}
+			}
 		}
 		if (!toBeMerged.empty()) {
 			tabContent->Add(new UI::TextView(keys[i].GetDescription(SHADER_STRING_SHORT_DESC), FLAG_DYNAMIC_ASCII | FLAG_WRAP_TEXT, true));