Add accurate shader multiplication option

src/config.cpp

@@ -62,6 +62,7 @@ void EmulatorConfig::load() {
 
 			shaderJitEnabled = toml::find_or<toml::boolean>(gpu, "EnableShaderJIT", shaderJitDefault);
 			vsyncEnabled = toml::find_or<toml::boolean>(gpu, "EnableVSync", true);
+			accurateShaderMul = toml::find_or<toml::boolean>(gpu, "AccurateShaderMultiplication", false);
 		}
 	}
 
@@ -125,6 +126,7 @@ void EmulatorConfig::save() {
 	data["GPU"]["EnableShaderJIT"] = shaderJitEnabled;
 	data["GPU"]["Renderer"] = std::string(Renderer::typeToString(rendererType));
 	data["GPU"]["EnableVSync"] = vsyncEnabled;
+	data["GPU"]["AccurateShaderMultiplication"] = accurateShaderMul;
 	data["Audio"]["DSPEmulation"] = std::string(Audio::DSPCore::typeToString(dspType));
 	data["Audio"]["EnableAudio"] = audioEnabled;
 

src/core/PICA/dynapica/shader_rec.cpp

@@ -16,7 +16,7 @@ void ShaderJIT::prepare(PICAShader& shaderUnit) {
 	auto it = cache.find(hash);
 
 	if (it == cache.end()) { // Block has not been compiled yet
-		auto emitter = std::make_unique<ShaderEmitter>();
+		auto emitter = std::make_unique<ShaderEmitter>(accurateMul);
 		emitter->compile(shaderUnit);
 		// Get pointer to callbacks
 		entrypointCallback = emitter->getInstructionCallback(shaderUnit.entrypoint);

src/core/PICA/dynapica/shader_rec_emitter_arm64.cpp

@@ -7,9 +7,6 @@ using namespace Helpers;
 using namespace oaknut;
 using namespace oaknut::util;
 
-// TODO: Expose safe/unsafe optimizations to the user
-constexpr bool useSafeMUL = true;
-
 // Similar to the x64 recompiler, we use an odd internal ABI, which abuses the fact that we'll very rarely be calling C++ functions
 // So to avoid pushing and popping, we'll be making use of volatile registers as much as possible
 static constexpr QReg src1Vec = Q1;

src/core/PICA/dynapica/shader_rec_emitter_x64.cpp

@@ -12,9 +12,6 @@ using namespace Xbyak;
 using namespace Xbyak::util;
 using namespace Helpers;
 
-// TODO: Expose safe/unsafe optimizations to the user
-constexpr bool useSafeMUL = false;
-
 // The shader recompiler uses quite an odd internal ABI
 // We make use of the fact that in regular conditions, we should pretty much never be calling C++ code from recompiled shader code
 // This allows us to establish an ABI that's optimized for this sort of workflow, statically allocating volatile host registers

src/core/PICA/gpu.cpp

@@ -64,6 +64,8 @@ void GPU::reset() {
 	regs.fill(0);
 	shaderUnit.reset();
 	shaderJIT.reset();
+	shaderJIT.setAccurateMul(config.accurateShaderMul);
+
 	std::memset(vram, 0, vramSize);
 	lightingLUT.fill(0);
 	lightingLUTDirty = true;

src/libretro_core.cpp

@@ -146,6 +146,7 @@ static bool FetchVariableBool(std::string key, bool def) {
 static void configInit() {
 	static const retro_variable values[] = {
 		{"panda3ds_use_shader_jit", "Enable shader JIT; enabled|disabled"},
+		{"panda3ds_accurate_shader_mul", "Enable accurate shader multiplication; disabled|enabled"},
 		{"panda3ds_use_vsync", "Enable VSync; enabled|disabled"},
 		{"panda3ds_dsp_emulation", "DSP emulation; Null|HLE|LLE"},
 		{"panda3ds_use_audio", "Enable audio; disabled|enabled"},
@@ -153,7 +154,7 @@ static void configInit() {
 		{"panda3ds_write_protect_virtual_sd", "Write protect virtual SD card; disabled|enabled"},
 		{"panda3ds_battery_level", "Battery percentage; 5|10|20|30|50|70|90|100"},
 		{"panda3ds_use_charger", "Charger plugged; enabled|disabled"},
-		{nullptr, nullptr}
+		{nullptr, nullptr},
 	};
 
 	envCallbacks(RETRO_ENVIRONMENT_SET_VARIABLES, (void*)values);
@@ -171,6 +172,7 @@ static void configUpdate() {
 	config.audioEnabled = FetchVariableBool("panda3ds_use_audio", false);
 	config.sdCardInserted = FetchVariableBool("panda3ds_use_virtual_sd", true);
 	config.sdWriteProtected = FetchVariableBool("panda3ds_write_protect_virtual_sd", false);
+	config.accurateShaderMul = FetchVariableBool("panda3ds_accurate_shader_mul", false);
 	config.discordRpcEnabled = false;
 
 	config.save();