Merge pull request #13706 from hrydgard/vfpu-sincos-double-compat

Add back the old implementation of vfpu_sin/cos/sincos.

Core/Compatibility.cpp

@@ -75,6 +75,7 @@ void Compatibility::CheckSettings(IniFile &iniFile, const std::string &gameID) {
 	CheckSetting(iniFile, gameID, "MemstickFixedFree", &flags_.MemstickFixedFree);
 	CheckSetting(iniFile, gameID, "DateLimited", &flags_.DateLimited);
 	CheckSetting(iniFile, gameID, "ReinterpretFramebuffers", &flags_.ReinterpretFramebuffers);
+	CheckSetting(iniFile, gameID, "DoublePrecisionSinCos", &flags_.DoublePrecisionSinCos);
 }
 
 void Compatibility::CheckSetting(IniFile &iniFile, const std::string &gameID, const char *option, bool *flag) {

Core/Compatibility.h

@@ -73,6 +73,7 @@ struct CompatFlags {
 	bool MemstickFixedFree;
 	bool DateLimited;
 	bool ReinterpretFramebuffers;
+	bool DoublePrecisionSinCos;
 };
 
 class IniFile;

Core/MIPS/MIPSVFPUUtils.cpp

@@ -895,3 +895,86 @@ float vfpu_rsqrt(float a) {
 
 	return val.f;
 }
+
+float vfpu_sin_single(float angle) {
+	angle -= floorf(angle * 0.25f) * 4.f;
+	if (angle == 0.0f || angle == 2.0f) {
+		return 0.0f;
+	} else if (angle == 1.0f) {
+		return 1.0f;
+	} else if (angle == 3.0f) {
+		return -1.0f;
+	}
+	angle *= (float)M_PI_2;
+	return sinf(angle);
+}
+
+float vfpu_cos_single(float angle) {
+	angle -= floorf(angle * 0.25f) * 4.f;
+	if (angle == 1.0f || angle == 3.0f) {
+		return 0.0f;
+	} else if (angle == 0.0f) {
+		return 1.0f;
+	} else if (angle == 2.0f) {
+		return -1.0f;
+	}
+	angle *= (float)M_PI_2;
+	return cosf(angle);
+}
+
+void vfpu_sincos_single(float angle, float &sine, float &cosine) {
+	angle -= floorf(angle * 0.25f) * 4.f;
+	if (angle == 0.0f) {
+		sine = 0.0f;
+		cosine = 1.0f;
+	} else if (angle == 1.0f) {
+		sine = 1.0f;
+		cosine = 0.0f;
+	} else if (angle == 2.0f) {
+		sine = 0.0f;
+		cosine = -1.0f;
+	} else if (angle == 3.0f) {
+		sine = -1.0f;
+		cosine = 0.0f;
+	} else {
+		angle *= (float)M_PI_2;
+#if defined(__linux__)
+		sincosf(angle, &sine, &cosine);
+#else
+		sine = sinf(angle);
+		cosine = cosf(angle);
+#endif
+	}
+}
+
+float vfpu_sin_double(float angle) {
+	return (float)sin((double)angle * M_PI_2);
+}
+
+float vfpu_cos_double(float angle) {
+	return (float)cos((double)angle * M_PI_2);
+}
+
+void vfpu_sincos_double(float angle_f, float &sine, float &cosine) {
+	double angle = (double)angle_f * M_PI_2;
+#if defined(__linux__)
+	double d_sine;
+	double d_cosine;
+	sincos(angle, &d_sine, &d_cosine);
+	sine = (float)d_sine;
+	cosine = (float)d_cosine;
+#else
+	sine = (float)sin(angle);
+	cosine = (float)cos(angle);
+#endif
+}
+
+float (*vfpu_sin)(float);
+float (*vfpu_cos)(float);
+void (*vfpu_sincos)(float, float&, float&);
+
+void InitVFPUSinCos(bool useDoublePrecision) {
+	vfpu_sin = useDoublePrecision ? vfpu_sin_double : vfpu_sin_single;
+	vfpu_cos = useDoublePrecision ? vfpu_cos_double : vfpu_cos_single;
+	vfpu_sincos = useDoublePrecision ? vfpu_sincos_double : vfpu_sincos_single;
+}

Core/MIPS/MIPSVFPUUtils.h

@@ -35,44 +35,29 @@ inline int Xpose(int v) {
 #endif
 
 // The VFPU uses weird angles where 4.0 represents a full circle. This makes it possible to return
-// exact 1.0/-1.0 values at certain angles. We get close enough for #2921 and #12900 by computing
-// things in double precision, multiplying the input by pi/2.
+// exact 1.0/-1.0 values at certain angles. We currently just scale, and special case the cardinal directions.
+//
+// Stepping down to [0, 2pi) helps, but we also check common exact-result values.
+// TODO: cos(1) and sin(2) should be -0.0, but doing that gives wrong results (possibly from floorf.)
+//
+// We also try an alternative solution, computing things in double precision, multiplying the input by pi/2.
+// This fixes #12900 (Hitman Reborn 2) but breaks #13705 (Cho Aniki Zero) and #13671 (Hajime no Ippo).
+// #2921 is still fine. So the alt solution (vfpu_sin_double etc) are behind a compat flag.
 //
 // A better solution would be to tailor some sine approximation for the 0..90 degrees range, compute
 // modulo manually and mirror that around the circle. Also correctly special casing for inf/nan inputs
 // and just trying to match it as closely as possible to the real PSP.
 //
-// Stepping down to [0, 2pi) helps, but we also check common exact-result values.
-// TODO: cos(1) and sin(2) should be -0.0, but doing that gives wrong results (possibly from floorf.)
-
 // Messing around with the modulo functions? try https://www.desmos.com/calculator.
 
-inline float vfpu_sin(float angle) {
-	return (float)sin((double)angle * M_PI_2);
-}
-
-inline float vfpu_cos(float angle) {
-	return (float)cos((double)angle * M_PI_2);
-}
+extern float (*vfpu_sin)(float);
+extern float (*vfpu_cos)(float);
+extern void (*vfpu_sincos)(float, float&, float&);
 
 inline float vfpu_asin(float angle) {
 	return asinf(angle) / M_PI_2;
 }
 
-inline void vfpu_sincos(float angle_f, float &sine, float &cosine) {
-	double angle = (double)angle_f * M_PI_2;
-#if defined(__linux__)
-	double d_sine;
-	double d_cosine;
-	sincos(angle, &d_sine, &d_cosine);
-	sine = (float)d_sine;
-	cosine = (float)d_cosine;
-#else
-	sine = (float)sin(angle);
-	cosine = (float)cos(angle);
-#endif
-}
-
 inline float vfpu_clamp(float v, float min, float max) {
 	// Note: NAN is preserved, and -0.0 becomes +0.0 if min=+0.0.
 	return v >= max ? max : (v <= min ? min : v);
@@ -230,3 +215,4 @@ int GetVectorOverlap(int reg1, VectorSize size1, int reg2, VectorSize size2);
 bool GetVFPUCtrlMask(int reg, u32 *mask);
 
 float Float16ToFloat32(unsigned short l);
+void InitVFPUSinCos(bool useDoublePrecision);

Core/System.cpp

@@ -44,6 +44,7 @@
 #include "Core/HDRemaster.h"
 #include "Core/MIPS/MIPS.h"
 #include "Core/MIPS/MIPSAnalyst.h"
+#include "Core/MIPS/MIPSVFPUUtils.h"
 #include "Core/Debugger/SymbolMap.h"
 #include "Core/Host.h"
 #include "Core/System.h"
@@ -66,7 +67,6 @@
 #include "Common/LogManager.h"
 #include "Common/ExceptionHandlerSetup.h"
 #include "Core/HLE/sceAudiocodec.h"
-
 #include "GPU/GPUState.h"
 #include "GPU/GPUInterface.h"
 #include "GPU/Debugger/RecordFormat.h"
@@ -287,6 +287,8 @@ bool CPU_Init() {
 	// likely to collide with any commercial ones.
 	coreParameter.compat.Load(discID);
 
+	InitVFPUSinCos(coreParameter.compat.flags().DoublePrecisionSinCos);
+
 	HLEPlugins::Init();
 	if (!Memory::Init()) {
 		// We're screwed.

assets/compat.ini

@@ -808,3 +808,6 @@ ULKS46087 = true
 # ULJM05533 = true
 # NPJH50006 = true
 
+[DoublePrecisionSinCos]
+# Hitman Reborn Battle Arena 2 (#12900)
+ULJS00218 = true