From 8a13b8c878917f4d6f59ddf683c80718340e0656 Mon Sep 17 00:00:00 2001
From: wheremyfoodat <44909372+wheremyfoodat@users.noreply.github.com>
Date: Fri, 30 Jun 2023 02:02:56 +0300
Subject: [PATCH] [Shader JIT] Compile MAD to FMA when possible

---
 .../PICA/dynapica/shader_rec_emitter_x64.hpp  |  2 ++
 .../PICA/dynapica/shader_rec_emitter_x64.cpp  | 27 ++++++++++++-------
 2 files changed, 20 insertions(+), 9 deletions(-)

diff --git a/include/PICA/dynapica/shader_rec_emitter_x64.hpp b/include/PICA/dynapica/shader_rec_emitter_x64.hpp
index 4ce2942a..3f33da06 100644
--- a/include/PICA/dynapica/shader_rec_emitter_x64.hpp
+++ b/include/PICA/dynapica/shader_rec_emitter_x64.hpp
@@ -36,6 +36,7 @@ class ShaderEmitter : public Xbyak::CodeGenerator {
 
 	bool haveSSE4_1 = false;  // Shows if the CPU supports SSE4.1
 	bool haveAVX = false;     // Shows if the CPU supports AVX (NOT AVX2, NOT AVX512. Regular AVX)
+	bool haveFMA3 = false;    // Shows if the CPU supports FMA3
 
 	// Compile all instructions from [current recompiler PC, end)
 	void compileUntil(const PICAShader& shaderUnit, u32 endPC);
@@ -112,6 +113,7 @@ public:
 
 		haveSSE4_1 = cpu.has(Xbyak::util::Cpu::tSSE41);
 		haveAVX = cpu.has(Xbyak::util::Cpu::tAVX);
+		haveFMA3 = cpu.has(Xbyak::util::Cpu::tFMA);
 
 		if (!cpu.has(Xbyak::util::Cpu::tSSE3)) {
 			Helpers::panic("This CPU does not support SSE3. Please use the shader interpreter instead");
diff --git a/src/core/PICA/dynapica/shader_rec_emitter_x64.cpp b/src/core/PICA/dynapica/shader_rec_emitter_x64.cpp
index cecd740c..8ccc4838 100644
--- a/src/core/PICA/dynapica/shader_rec_emitter_x64.cpp
+++ b/src/core/PICA/dynapica/shader_rec_emitter_x64.cpp
@@ -580,17 +580,26 @@ void ShaderEmitter::recMAD(const PICAShader& shader, u32 instruction) {
 	loadRegister<3>(src3_xmm, shader, src3, 0, operandDescriptor);
 
 	// TODO: Implement safe PICA mul
-	// Multiply src1 * src2
-	if (haveAVX) {
-		vmulps(scratch1, src1_xmm, src2_xmm);
-	} else {
-		movaps(scratch1, src1_xmm);
-		mulps(scratch1, src2_xmm);
+	// If we have FMA3, optimize MAD to use FMA
+	if (haveFMA3) {
+		vfmadd213ps(src1_xmm, src2_xmm, src3_xmm);
+		storeRegister(src1_xmm, shader, dest, operandDescriptor);
 	}
+	
+	// If we don't have FMA3, do a multiplication and addition
+	else {
+		// Multiply src1 * src2
+		if (haveAVX) {
+			vmulps(scratch1, src1_xmm, src2_xmm);
+		} else {
+			movaps(scratch1, src1_xmm);
+			mulps(scratch1, src2_xmm);
+		}
 
-	// Add src3
-	addps(scratch1, src3_xmm);
-	storeRegister(scratch1, shader, dest, operandDescriptor);
+		// Add src3
+		addps(scratch1, src3_xmm);
+		storeRegister(scratch1, shader, dest, operandDescriptor);
+	}
 }
 
 void ShaderEmitter::recCMP(const PICAShader& shader, u32 instruction) {