Merge pull request #528 from wheremyfoodat/bleh

arm64: Implement DPH/DPHI in shader recompiler, enable shader recompiler by default on Android & a64 desktop

include/config.hpp

@@ -7,7 +7,7 @@
 // Remember to initialize every field here to its default value otherwise bad things will happen
 struct EmulatorConfig {
 	// Only enable the shader JIT by default on platforms where it's completely tested
-#ifdef PANDA3DS_X64_HOST
+#if defined(PANDA3DS_X64_HOST) || defined(PANDA3DS_ARM64_HOST)
 	static constexpr bool shaderJitDefault = true;
 #else
 	static constexpr bool shaderJitDefault = false;

src/core/PICA/dynapica/shader_rec_emitter_arm64.cpp

@@ -144,8 +144,8 @@ void ShaderEmitter::compileInstruction(const PICAShader& shaderUnit) {
 		case ShaderOpcodes::CMP2: recCMP(shaderUnit, instruction); break;
 		case ShaderOpcodes::DP3: recDP3(shaderUnit, instruction); break;
 		case ShaderOpcodes::DP4: recDP4(shaderUnit, instruction); break;
-		// case ShaderOpcodes::DPH:
-		// case ShaderOpcodes::DPHI: recDPH(shaderUnit, instruction); break;
+		case ShaderOpcodes::DPH:
+		case ShaderOpcodes::DPHI: recDPH(shaderUnit, instruction); break;
 		case ShaderOpcodes::END: recEND(shaderUnit, instruction); break;
 		case ShaderOpcodes::EX2: recEX2(shaderUnit, instruction); break;
 		case ShaderOpcodes::FLR: recFLR(shaderUnit, instruction); break;
@@ -533,6 +533,39 @@ void ShaderEmitter::recDP4(const PICAShader& shader, u32 instruction) {
 	storeRegister(src1Vec, shader, dest, operandDescriptor);
 }
 
+void ShaderEmitter::recDPH(const PICAShader& shader, u32 instruction) {
+	const bool isDPHI = (instruction >> 26) == ShaderOpcodes::DPHI;
+
+	const u32 operandDescriptor = shader.operandDescriptors[instruction & 0x7f];
+	const u32 src1 = isDPHI ? getBits<14, 5>(instruction) : getBits<12, 7>(instruction);
+	const u32 src2 = isDPHI ? getBits<7, 7>(instruction) : getBits<7, 5>(instruction);
+	const u32 idx = getBits<19, 2>(instruction);
+	const u32 dest = getBits<21, 5>(instruction);
+	const u32 writeMask = getBits<0, 4>(operandDescriptor);
+
+	// TODO: Safe multiplication equivalent (Multiplication is not IEEE compliant on the PICA)
+	loadRegister<1>(src1Vec, shader, src1, isDPHI ? 0 : idx, operandDescriptor);
+	loadRegister<2>(src2Vec, shader, src2, isDPHI ? idx : 0, operandDescriptor);
+	// // Attach 1.0 to the w component of src1
+	MOV(src1Vec.Selem()[3], onesVector.Selem()[0]);
+
+	// Now perform a DP4
+	// Do a piecewise multiplication of the vectors first
+	if constexpr (useSafeMUL) {
+		emitSafeMUL(src1Vec, src2Vec, scratch1Vec);
+	} else {
+		FMUL(src1Vec.S4(), src1Vec.S4(), src2Vec.S4());
+	}
+	FADDP(src1Vec.S4(), src1Vec.S4(), src1Vec.S4());  // Now add the adjacent components together
+	FADDP(src1Vec.toS(), src1Vec.toD().S2());         // Again for the bottom 2 lanes. Now the bottom lane contains the dot product
+
+	if (writeMask != 0x8) {                     // Copy bottom lane to all lanes if we're not simply writing back x
+		DUP(src1Vec.S4(), src1Vec.Selem()[0]);  // src1Vec = src1Vec.xxxx
+	}
+
+	storeRegister(src1Vec, shader, dest, operandDescriptor);
+}
+
 oaknut::Label ShaderEmitter::emitLog2Func() {
 	oaknut::Label funcStart;
 

src/pandroid/app/src/main/java/com/panda3ds/pandroid/data/config/GlobalConfig.java

@@ -21,7 +21,7 @@ public class GlobalConfig {
 
     public static DataModel data;
 
-    public static final Key<Boolean> KEY_SHADER_JIT = new Key<>("emu.shader_jit", false);
+    public static final Key<Boolean> KEY_SHADER_JIT = new Key<>("emu.shader_jit", true);
     public static final Key<Boolean> KEY_PICTURE_IN_PICTURE = new Key<>("app.behavior.pictureInPicture", false);
     public static final Key<Boolean> KEY_SHOW_PERFORMANCE_OVERLAY = new Key<>("dev.performanceOverlay", false);
     public static final Key<Boolean> KEY_LOGGER_SERVICE = new Key<>("dev.loggerService", false);