GX2: Rework GX2Set*UniformReg

- Use cafeExportRegister() instead of legacy export
- Submit as a single PM4 packet
- Add logging for the special case of the size parameter (not sure if this is used by any game?)
- Add some extra validation and logging which may be helpful to homebrew devs

src/Cafe/OS/libs/gx2/GX2.cpp

@@ -396,16 +396,13 @@ void gx2_load()
 	osLib_addFunction("gx2", "GX2GetCurrentScanBuffer", gx2Export_GX2GetCurrentScanBuffer);
 
 	// shader stuff
-	//osLib_addFunction("gx2", "GX2SetVertexShader", gx2Export_GX2SetVertexShader);
 	osLib_addFunction("gx2", "GX2SetPixelShader", gx2Export_GX2SetPixelShader);
 	osLib_addFunction("gx2", "GX2SetGeometryShader", gx2Export_GX2SetGeometryShader);
 	osLib_addFunction("gx2", "GX2SetComputeShader", gx2Export_GX2SetComputeShader);
-	osLib_addFunction("gx2", "GX2SetVertexUniformReg", gx2Export_GX2SetVertexUniformReg);
 	osLib_addFunction("gx2", "GX2SetVertexUniformBlock", gx2Export_GX2SetVertexUniformBlock);
 	osLib_addFunction("gx2", "GX2RSetVertexUniformBlock", gx2Export_GX2RSetVertexUniformBlock);
 
 	osLib_addFunction("gx2", "GX2SetPixelUniformBlock", gx2Export_GX2SetPixelUniformBlock);
-	osLib_addFunction("gx2", "GX2SetPixelUniformReg", gx2Export_GX2SetPixelUniformReg);
 	osLib_addFunction("gx2", "GX2SetGeometryUniformBlock", gx2Export_GX2SetGeometryUniformBlock);
 	osLib_addFunction("gx2", "GX2SetShaderModeEx", gx2Export_GX2SetShaderModeEx);
 

src/Cafe/OS/libs/gx2/GX2.h

@@ -18,11 +18,9 @@ void gx2_load();
 void gx2Export_GX2SetPixelShader(PPCInterpreter_t* hCPU);
 void gx2Export_GX2SetGeometryShader(PPCInterpreter_t* hCPU);
 void gx2Export_GX2SetComputeShader(PPCInterpreter_t* hCPU);
-void gx2Export_GX2SetVertexUniformReg(PPCInterpreter_t* hCPU);
 void gx2Export_GX2SetVertexUniformBlock(PPCInterpreter_t* hCPU);
 void gx2Export_GX2RSetVertexUniformBlock(PPCInterpreter_t* hCPU);
 void gx2Export_GX2SetPixelUniformBlock(PPCInterpreter_t* hCPU);
-void gx2Export_GX2SetPixelUniformReg(PPCInterpreter_t* hCPU);
 void gx2Export_GX2SetGeometryUniformBlock(PPCInterpreter_t* hCPU);
 void gx2Export_GX2SetShaderModeEx(PPCInterpreter_t* hCPU);
 void gx2Export_GX2CalcGeometryShaderInputRingBufferSize(PPCInterpreter_t* hCPU);

src/Cafe/OS/libs/gx2/GX2_Shader.cpp

@@ -417,6 +417,37 @@ namespace GX2
 		}
 	}
 
+	void _GX2SubmitUniformReg(uint32 offsetRegBase, uint32 aluRegisterOffset, uint32be* dataWords, uint32 sizeInU32s)
+	{
+		if(aluRegisterOffset&0x8000)
+		{
+			cemuLog_logDebug(LogType::Force, "_GX2SubmitUniformReg(): Unhandled loop const special case or invalid offset");
+			return;
+		}
+		if((aluRegisterOffset+sizeInU32s) > 0x400)
+		{
+			cemuLog_logOnce(LogType::APIErrors, "GX2SetVertexUniformReg values are out of range (offset {} + size {} must be equal or smaller than 0x400)", aluRegisterOffset, sizeInU32s);
+		}
+		if( (sizeInU32s&3) != 0)
+		{
+			cemuLog_logOnce(LogType::APIErrors, "GX2Set*UniformReg must be called with a size that is a multiple of 4 (size: {:})", sizeInU32s);
+			sizeInU32s &= ~3;
+		}
+		GX2ReserveCmdSpace(2 + sizeInU32s);
+		gx2WriteGather_submit(pm4HeaderType3(IT_SET_ALU_CONST, 1 + sizeInU32s), offsetRegBase + aluRegisterOffset);
+		gx2WriteGather_submitU32AsLEArray((uint32*)dataWords, sizeInU32s);
+	}
+
+	void GX2SetVertexUniformReg(uint32 offset, uint32 sizeInU32s, uint32be* values)
+	{
+		_GX2SubmitUniformReg(0x400, offset, values, sizeInU32s);
+	}
+
+	void GX2SetPixelUniformReg(uint32 offset, uint32 sizeInU32s, uint32be* values)
+	{
+		_GX2SubmitUniformReg(0, offset, values, sizeInU32s);
+	}
+
 	void GX2ShaderInit()
 	{
 		cafeExportRegister("gx2", GX2CalcFetchShaderSizeEx, LogType::GX2);
@@ -428,5 +459,8 @@ namespace GX2
 		cafeExportRegister("gx2", GX2GetPixelShaderStackEntries, LogType::GX2);
 		cafeExportRegister("gx2", GX2SetFetchShader, LogType::GX2);
 		cafeExportRegister("gx2", GX2SetVertexShader, LogType::GX2);
+
+		cafeExportRegister("gx2", GX2SetVertexUniformReg, LogType::GX2);
+		cafeExportRegister("gx2", GX2SetPixelUniformReg, LogType::GX2);
 	}
 }

src/Cafe/OS/libs/gx2/GX2_shader_legacy.cpp

@@ -270,41 +270,6 @@ void gx2Export_GX2SetComputeShader(PPCInterpreter_t* hCPU)
 	osLib_returnFromFunction(hCPU, 0);
 }
 
-void _GX2SubmitUniformReg(uint32 aluRegisterOffset, MPTR virtualAddress, uint32 count)
-{
-	uint32* dataWords = (uint32*)memory_getPointerFromVirtualOffset(virtualAddress);
-	GX2ReserveCmdSpace(2 + (count / 0xFF) * 2 + count);
-	// write PM4 command(s)
-	uint32 currentRegisterOffset = aluRegisterOffset;
-	while (count > 0)
-	{
-		uint32 subCount = std::min(count, 0xFFu); // a single command can write at most 0xFF values
-		gx2WriteGather_submit(pm4HeaderType3(IT_SET_ALU_CONST, 1 + subCount),
-			currentRegisterOffset);
-		gx2WriteGather_submitU32AsLEArray(dataWords, subCount);
-
-		dataWords += subCount;
-		count -= subCount;
-		currentRegisterOffset += subCount;
-	}
-}
-
-void gx2Export_GX2SetVertexUniformReg(PPCInterpreter_t* hCPU)
-{
-	cemuLog_log(LogType::GX2, "GX2SetVertexUniformReg(0x{:08x},0x{:x},0x{:08x})", hCPU->gpr[3], hCPU->gpr[4], hCPU->gpr[5]);
-	_GX2SubmitUniformReg(hCPU->gpr[3] + 0x400, hCPU->gpr[5], hCPU->gpr[4]);
-	cemu_assert_debug((hCPU->gpr[3] + hCPU->gpr[4]) <= 0x400);
-	osLib_returnFromFunction(hCPU, 0);
-}
-
-void gx2Export_GX2SetPixelUniformReg(PPCInterpreter_t* hCPU)
-{
-	cemuLog_log(LogType::GX2, "GX2SetPixelUniformReg(0x{:08x},0x{:x},0x{:08x})", hCPU->gpr[3], hCPU->gpr[4], hCPU->gpr[5]);
-	_GX2SubmitUniformReg(hCPU->gpr[3], hCPU->gpr[5], hCPU->gpr[4]);
-	cemu_assert_debug((hCPU->gpr[3] + hCPU->gpr[4]) <= 0x400);
-	osLib_returnFromFunction(hCPU, 0);
-}
-
 void _GX2SubmitUniformBlock(uint32 registerBase, uint32 index, MPTR virtualAddress, uint32 size)
 {
 	GX2ReserveCmdSpace(9);

src/Cafe/OS/libs/nn_act/nn_act.cpp

@@ -543,8 +543,6 @@ void nnActExport_GetDefaultAccount(PPCInterpreter_t* hCPU)
 void nnActExport_GetSlotNo(PPCInterpreter_t* hCPU)
 {
 	// id of active account
-	// uint8 GetSlotNo(void);
-	cemuLog_logDebug(LogType::Force, "nn_act.GetSlotNo()");
 	osLib_returnFromFunction(hCPU, 1); // 1 is the first slot (0 is invalid)
 }