Merge remote-tracking branch 'upstream/no-light-cache'

CMakeLists.txt

@@ -1219,6 +1219,8 @@ add_library(GPU OBJECT
 	GPU/Common/GPUDebugInterface.h
 	GPU/Common/VertexDecoderCommon.cpp
 	GPU/Common/VertexDecoderCommon.h
+	GPU/Common/TransformCommon.cpp
+	GPU/Common/TransformCommon.h
 	GPU/Common/IndexGenerator.cpp
 	GPU/Common/IndexGenerator.h
 	GPU/Common/TextureDecoder.cpp

GPU/Common/TransformCommon.cpp

@@ -0,0 +1,181 @@
+
+// Copyright (c) 2013- PPSSPP Project.
+
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, version 2.0 or later versions.
+
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License 2.0 for more details.
+
+// A copy of the GPL 2.0 should have been included with the program.
+// If not, see http://www.gnu.org/licenses/
+
+// Official git repository and contact information can be found at
+// https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
+
+#include <stdio.h>
+
+#include "GPU/GPUState.h"
+#include "GPU/Common/TransformCommon.h"
+
+// Check for max first as clamping to max is more common than min when lighting.
+inline float clamp(float in, float min, float max) {
+	return in > max ? max : (in < min ? min : in);
+}
+
+Lighter::Lighter(int vertType) {
+	if (!gstate.isLightingEnabled())
+		return;
+
+	doShadeMapping_ = gstate.getUVGenMode() == GE_TEXMAP_ENVIRONMENT_MAP;
+	materialEmissive.GetFromRGB(gstate.materialemissive);
+	materialEmissive.a = 0.0f;
+	globalAmbient.GetFromRGB(gstate.ambientcolor);
+	globalAmbient.GetFromA(gstate.ambientalpha);
+	materialAmbient.GetFromRGB(gstate.materialambient);
+	materialAmbient.GetFromA(gstate.materialalpha);
+	materialDiffuse.GetFromRGB(gstate.materialdiffuse);
+	materialDiffuse.a = 1.0f;
+	materialSpecular.GetFromRGB(gstate.materialspecular);
+	materialSpecular.a = 1.0f;
+	specCoef_ = getFloat24(gstate.materialspecularcoef);
+	// viewer_ = Vec3f(-gstate.viewMatrix[9], -gstate.viewMatrix[10], -gstate.viewMatrix[11]);
+	bool hasColor = (vertType & GE_VTYPE_COL_MASK) != 0;
+	materialUpdate_ = hasColor ? (gstate.materialupdate & 7) : 0;
+
+	for (int l = 0; l < 4; l++) {
+		int i = l * 3;
+		if (gstate.isLightChanEnabled(l)) {
+			lpos[i] = getFloat24(gstate.lpos[i]);
+			lpos[i + 1] = getFloat24(gstate.lpos[i + 1]);
+			lpos[i + 2] = getFloat24(gstate.lpos[i + 2]);
+			ldir[i] = getFloat24(gstate.ldir[i]);
+			ldir[i + 1] = getFloat24(gstate.ldir[i + 1]);
+			ldir[i + 2] = getFloat24(gstate.ldir[i + 2]);
+			latt[i] = getFloat24(gstate.latt[i]);
+			latt[i + 1] = getFloat24(gstate.latt[i + 1]);
+			latt[i + 2] = getFloat24(gstate.latt[i + 2]);
+			for (int t = 0; t < 3; t++) {
+				u32 data = gstate.lcolor[l * 3 + t] & 0xFFFFFF;
+				float r = (float)(data & 0xff) * (1.0f / 255.0f);
+				float g = (float)((data >> 8) & 0xff) * (1.0f / 255.0f);
+				float b = (float)(data >> 16) * (1.0f / 255.0f);
+				lcolor[t][l][0] = r;
+				lcolor[t][l][1] = g;
+				lcolor[t][l][2] = b;
+			}
+		}
+	}
+}
+
+void Lighter::Light(float colorOut0[4], float colorOut1[4], const float colorIn[4], const Vec3f &pos, const Vec3f &norm) {
+	Color4 in(colorIn);
+
+	const Color4 *ambient;
+	if (materialUpdate_ & 1)
+		ambient = &in;
+	else
+		ambient = &materialAmbient;
+
+	const Color4 *diffuse;
+	if (materialUpdate_ & 2)
+		diffuse = &in;
+	else
+		diffuse = &materialDiffuse;
+
+	const Color4 *specular;
+	if (materialUpdate_ & 4)
+		specular = &in;
+	else
+		specular = &materialSpecular;
+
+	Color4 lightSum0 = globalAmbient * *ambient + materialEmissive;
+	Color4 lightSum1(0, 0, 0, 0);
+
+	for (int l = 0; l < 4; l++) {
+		// can we skip this light?
+		if (!gstate.isLightChanEnabled(l))
+			continue;
+
+		GELightType type = gstate.getLightType(l);
+
+		Vec3f toLight(0, 0, 0);
+		Vec3f lightDir(0, 0, 0);
+
+		if (type == GE_LIGHTTYPE_DIRECTIONAL)
+			toLight = Vec3f(&lpos[l * 3]);  // lightdir is for spotlights
+		else
+			toLight = Vec3f(&lpos[l * 3]) - pos;
+
+		bool doSpecular = gstate.isUsingSpecularLight(l);
+		bool poweredDiffuse = gstate.isUsingPoweredDiffuseLight(l);
+
+		float distanceToLight = toLight.Length();
+		float dot = 0.0f;
+		float angle = 0.0f;
+		float lightScale = 0.0f;
+
+		if (distanceToLight > 0.0f) {
+			toLight /= distanceToLight;
+			dot = Dot(toLight, norm);
+		}
+		// Clamp dot to zero.
+		if (dot < 0.0f) dot = 0.0f;
+
+		if (poweredDiffuse)
+			dot = powf(dot, specCoef_);
+
+		// Attenuation
+		switch (type) {
+		case GE_LIGHTTYPE_DIRECTIONAL:
+			lightScale = 1.0f;
+			break;
+		case GE_LIGHTTYPE_POINT:
+			lightScale = clamp(1.0f / (latt[l * 3] + latt[l * 3 + 1] * distanceToLight + latt[l * 3 + 2] * distanceToLight*distanceToLight), 0.0f, 1.0f);
+			break;
+		case GE_LIGHTTYPE_SPOT:
+		case GE_LIGHTTYPE_UNKNOWN:
+			lightDir = Vec3f(&ldir[l * 3]);
+			angle = Dot(toLight.Normalized(), lightDir.Normalized());
+			if (angle >= lcutoff[l])
+				lightScale = clamp(1.0f / (latt[l * 3] + latt[l * 3 + 1] * distanceToLight + latt[l * 3 + 2] * distanceToLight*distanceToLight), 0.0f, 1.0f) * powf(angle, lconv[l]);
+			break;
+		default:
+			// ILLEGAL
+			break;
+		}
+
+		Color4 lightDiff(lcolor[1][l], 0.0f);
+		Color4 diff = (lightDiff * *diffuse) * dot;
+
+		// Real PSP specular
+		Vec3f toViewer(0, 0, 1);
+		// Better specular
+		// Vec3f toViewer = (viewer - pos).Normalized();
+
+		if (doSpecular) {
+			Vec3f halfVec = (toLight + toViewer);
+			halfVec.Normalize();
+
+			dot = Dot(halfVec, norm);
+			if (dot > 0.0f) {
+				Color4 lightSpec(lcolor[2][l], 0.0f);
+				lightSum1 += (lightSpec * *specular * (powf(dot, specCoef_) * lightScale));
+			}
+		}
+
+		if (gstate.isLightChanEnabled(l)) {
+			Color4 lightAmbient(lcolor[0][l], 0.0f);
+			lightSum0 += (lightAmbient * *ambient + diff) * lightScale;
+		}
+	}
+
+	// 4?
+	for (int i = 0; i < 4; i++) {
+		colorOut0[i] = lightSum0[i] > 1.0f ? 1.0f : lightSum0[i];
+		colorOut1[i] = lightSum1[i] > 1.0f ? 1.0f : lightSum1[i];
+	}
+}

GPU/Common/TransformCommon.h

@@ -0,0 +1,93 @@
+// Copyright (c) 2014- PPSSPP Project.
+
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, version 2.0 or later versions.
+
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License 2.0 for more details.
+
+// A copy of the GPL 2.0 should have been included with the program.
+// If not, see http://www.gnu.org/licenses/
+
+// Official git repository and contact information can be found at
+// https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
+
+#pragma once
+
+#include <cstring>
+#include "base/basictypes.h"
+#include "Common/Log.h"
+#include "Common/CommonTypes.h"
+#include "Core/Reporting.h"
+#include "GPU/ge_constants.h"
+#include "GPU/Math3D.h"
+
+struct Color4 {
+	float r, g, b, a;
+
+	Color4() : r(0), g(0), b(0), a(0) { }
+	Color4(float _r, float _g, float _b, float _a = 1.0f)
+		: r(_r), g(_g), b(_b), a(_a) {
+	}
+	Color4(const float in[4]) { r = in[0]; g = in[1]; b = in[2]; a = in[3]; }
+	Color4(const float in[3], float alpha) { r = in[0]; g = in[1]; b = in[2]; a = alpha; }
+
+	const float &operator [](int i) const { return *(&r + i); }
+
+	Color4 operator *(float f) const {
+		return Color4(f*r, f*g, f*b, f*a);
+	}
+	Color4 operator *(const Color4 &c) const {
+		return Color4(r*c.r, g*c.g, b*c.b, a*c.a);
+	}
+	Color4 operator +(const Color4 &c) const {
+		return Color4(r + c.r, g + c.g, b + c.b, a + c.a);
+	}
+	void operator +=(const Color4 &c) {
+		r += c.r;
+		g += c.g;
+		b += c.b;
+		a += c.a;
+	}
+	void GetFromRGB(u32 col) {
+		b = ((col >> 16) & 0xff) * (1.0f / 255.0f);
+		g = ((col >> 8) & 0xff) * (1.0f / 255.0f);
+		r = ((col >> 0) & 0xff) * (1.0f / 255.0f);
+	}
+	void GetFromA(u32 col) {
+		a = (col & 0xff) * (1.0f / 255.0f);
+	}
+};
+
+// Convenient way to do precomputation to save the parts of the lighting calculation
+// that's common between the many vertices of a draw call.
+class Lighter {
+public:
+	Lighter(int vertType);
+	void Light(float colorOut0[4], float colorOut1[4], const float colorIn[4], const Vec3f &pos, const Vec3f &normal);
+
+private:
+	Color4 globalAmbient;
+	Color4 materialEmissive;
+	Color4 materialAmbient;
+	Color4 materialDiffuse;
+	Color4 materialSpecular;
+	float specCoef_;
+	// Vec3f viewer_;
+	bool doShadeMapping_;
+	int materialUpdate_;
+
+	// Converted light parameters
+public:
+	float lpos[12];  // Used by shade UV mapping
+private:
+	float ldir[12];
+	float latt[12];
+	float lcutoff[4];
+	float lconv[4];
+	float lcolor[3][4][3];
+};
+

GPU/Directx9/GPU_DX9.cpp

@@ -953,89 +953,46 @@ void DIRECTX9_GPU::ExecuteOp(u32 op, u32 diff) {
 	case GE_CMD_LIGHTTYPE2:
 	case GE_CMD_LIGHTTYPE3:
 		break;
-
 	case GE_CMD_LX0:case GE_CMD_LY0:case GE_CMD_LZ0:
-	case GE_CMD_LX1:case GE_CMD_LY1:case GE_CMD_LZ1:
-	case GE_CMD_LX2:case GE_CMD_LY2:case GE_CMD_LZ2:
-	case GE_CMD_LX3:case GE_CMD_LY3:case GE_CMD_LZ3:
-		{
-			int n = cmd - GE_CMD_LX0;
-			int l = n / 3;
-			int c = n % 3;
-			gstate_c.lightpos[l][c] = getFloat24(data);
-			if (diff)
-				shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
-		break;
-
 	case GE_CMD_LDX0:case GE_CMD_LDY0:case GE_CMD_LDZ0:
-	case GE_CMD_LDX1:case GE_CMD_LDY1:case GE_CMD_LDZ1:
-	case GE_CMD_LDX2:case GE_CMD_LDY2:case GE_CMD_LDZ2:
-	case GE_CMD_LDX3:case GE_CMD_LDY3:case GE_CMD_LDZ3:
-		{
-			int n = cmd - GE_CMD_LDX0;
-			int l = n / 3;
-			int c = n % 3;
-			gstate_c.lightdir[l][c] = getFloat24(data);
-			if (diff)
-				shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
-		break;
-
 	case GE_CMD_LKA0:case GE_CMD_LKB0:case GE_CMD_LKC0:
+	case GE_CMD_LKS0:  // spot coef ("conv")
+	case GE_CMD_LKO0: // light angle ("cutoff")
+	case GE_CMD_LAC0:
+	case GE_CMD_LDC0:
+	case GE_CMD_LSC0:
+		shaderManager_->DirtyUniform(DIRTY_LIGHT0);
+		break;
+
+	case GE_CMD_LX1:case GE_CMD_LY1:case GE_CMD_LZ1:
+	case GE_CMD_LDX1:case GE_CMD_LDY1:case GE_CMD_LDZ1:
 	case GE_CMD_LKA1:case GE_CMD_LKB1:case GE_CMD_LKC1:
-	case GE_CMD_LKA2:case GE_CMD_LKB2:case GE_CMD_LKC2:
-	case GE_CMD_LKA3:case GE_CMD_LKB3:case GE_CMD_LKC3:
-		{
-			int n = cmd - GE_CMD_LKA0;
-			int l = n / 3;
-			int c = n % 3;
-			gstate_c.lightatt[l][c] = getFloat24(data);
-			if (diff)
-				shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
-		break;
-
-	case GE_CMD_LKS0:
 	case GE_CMD_LKS1:
-	case GE_CMD_LKS2:
-	case GE_CMD_LKS3:
-		{
-			int l = cmd - GE_CMD_LKS0;
-			gstate_c.lightspotCoef[l] = getFloat24(data);
-			if (diff)
-				shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
-		break;
-
-	case GE_CMD_LKO0:
 	case GE_CMD_LKO1:
-	case GE_CMD_LKO2:
+	case GE_CMD_LAC1:
-	case GE_CMD_LKO3:
+	case GE_CMD_LDC1:
-		{
+	case GE_CMD_LSC1:
-			int l = cmd - GE_CMD_LKO0;
+		shaderManager_->DirtyUniform(DIRTY_LIGHT1);
-			gstate_c.lightangle[l] = getFloat24(data);
-			if (diff)
-				shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
 		break;
-
+	case GE_CMD_LX2:case GE_CMD_LY2:case GE_CMD_LZ2:
-	case GE_CMD_LAC0:case GE_CMD_LAC1:case GE_CMD_LAC2:case GE_CMD_LAC3:
+	case GE_CMD_LDX2:case GE_CMD_LDY2:case GE_CMD_LDZ2:
-	case GE_CMD_LDC0:case GE_CMD_LDC1:case GE_CMD_LDC2:case GE_CMD_LDC3:
+	case GE_CMD_LKA2:case GE_CMD_LKB2:case GE_CMD_LKC2:
-	case GE_CMD_LSC0:case GE_CMD_LSC1:case GE_CMD_LSC2:case GE_CMD_LSC3:
+	case GE_CMD_LKS2:
-		{
+	case GE_CMD_LKO2:
-			float r = (float)(data & 0xff)/255.0f;
+	case GE_CMD_LAC2:
-			float g = (float)((data>>8) & 0xff)/255.0f;
+	case GE_CMD_LDC2:
-			float b = (float)(data>>16)/255.0f;
+	case GE_CMD_LSC2:
-
+		shaderManager_->DirtyUniform(DIRTY_LIGHT2);
-			int l = (cmd - GE_CMD_LAC0) / 3;
+		break;
-			int t = (cmd - GE_CMD_LAC0) % 3;
+	case GE_CMD_LX3:case GE_CMD_LY3:case GE_CMD_LZ3:
-			gstate_c.lightColor[t][l][0] = r;
+	case GE_CMD_LDX3:case GE_CMD_LDY3:case GE_CMD_LDZ3:
-			gstate_c.lightColor[t][l][1] = g;
+	case GE_CMD_LKA3:case GE_CMD_LKB3:case GE_CMD_LKC3:
-			gstate_c.lightColor[t][l][2] = b;
+	case GE_CMD_LKS3:
-			if (diff)
+	case GE_CMD_LKO3:
-				shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
+	case GE_CMD_LAC3:
-		}
+	case GE_CMD_LDC3:
+	case GE_CMD_LSC3:
+		shaderManager_->DirtyUniform(DIRTY_LIGHT3);
 		break;
 
 	case GE_CMD_VIEWPORTX1:

GPU/Directx9/ShaderManagerDX9.cpp

@@ -208,7 +208,7 @@ void LinkedShaderDX9::SetFloat(D3DXHANDLE uniform, float value) {
 
 // Utility
 void LinkedShaderDX9::SetColorUniform3(D3DXHANDLE uniform, u32 color) {
-	const float col[3] = {
+	const float col[4] = {
 		((color & 0xFF)) / 255.0f,
 		((color & 0xFF00) >> 8) / 255.0f,
 		((color & 0xFF0000) >> 16) / 255.0f
@@ -216,6 +216,12 @@ void LinkedShaderDX9::SetColorUniform3(D3DXHANDLE uniform, u32 color) {
 	SetFloatArray(uniform, col, 4);
 }
 
+void LinkedShaderDX9::SetFloat24Uniform3(D3DXHANDLE uniform, const u32 data[3]) {
+	const u32 col[4] = {
+		data[0] >> 8, data[1] >> 8, data[2] >> 8,
+	};
+	SetFloatArray(uniform, (const float *)&col[0], 4);
+}
 
 void LinkedShaderDX9::SetColorUniform3Alpha(D3DXHANDLE uniform, u32 color, u8 alpha) {
 	const float col[4] = {
@@ -452,42 +458,30 @@ void LinkedShaderDX9::updateUniforms() {
 	}
 	for (int i = 0; i < 4; i++) {
 		if (dirtyUniforms & (DIRTY_LIGHT0 << i)) {
+			if (u_lightpos[i] != 0) {
 				if (gstate.isDirectionalLight(i)) {
 					// Prenormalize
-				float x = gstate_c.lightpos[i][0];
+					float x = getFloat24(gstate.lpos[i * 3 + 0]);
-				float y = gstate_c.lightpos[i][1];
+					float y = getFloat24(gstate.lpos[i * 3 + 1]);
-				float z = gstate_c.lightpos[i][2];
+					float z = getFloat24(gstate.lpos[i * 3 + 2]);
 					float len = sqrtf(x*x + y*y + z*z);
 					if (len == 0.0f)
 						len = 1.0f;
 					else
 						len = 1.0f / len;
 					float vec[3] = { x * len, y * len, z * len };
-				if (u_lightpos[i] != 0) 
 					SetFloatArray(u_lightpos[i], vec, 3);
 				} else {
-				if (u_lightpos[i] != 0) 
+					SetFloat24Uniform3(u_lightpos[i], &gstate.lpos[i * 3]);
-					SetFloatArray(u_lightpos[i],  gstate_c.lightpos[i], 3);
 				}
-			if (u_lightdir[i] != 0) 
+			}
-				SetFloatArray(u_lightdir[i],  gstate_c.lightdir[i], 3);
+			if (u_lightdir[i] != 0) SetFloat24Uniform3(u_lightdir[i], &gstate.ldir[i * 3]);
-			if (u_lightatt[i] != 0) 
+			if (u_lightatt[i] != 0) SetFloat24Uniform3(u_lightatt[i], &gstate.latt[i * 3]);
-				SetFloatArray(u_lightatt[i],  gstate_c.lightatt[i], 3);
+			if (u_lightangle[i] != 0) SetFloat(u_lightangle[i], getFloat24(gstate.lcutoff[i]));
-				
+			if (u_lightspotCoef[i] != 0) SetFloat(u_lightspotCoef[i], getFloat24(gstate.lconv[i]));
-			if (u_lightangle[i] != 0) 				
+			if (u_lightambient[i] != 0) SetColorUniform3(u_lightambient[i], gstate.lcolor[i * 3]);
-				SetFloat(u_lightangle[i],  gstate_c.lightangle[i]);
+			if (u_lightdiffuse[i] != 0) SetColorUniform3(u_lightdiffuse[i], gstate.lcolor[i * 3 + 1]);
-
+			if (u_lightspecular[i] != 0) SetColorUniform3(u_lightspecular[i], gstate.lcolor[i * 3 + 2]);
-			if (u_lightspotCoef[i] != 0) 
-				SetFloat(u_lightspotCoef[i],  gstate_c.lightspotCoef[i]);
-
-			if (u_lightambient[i] != 0) 				
-				SetFloatArray(u_lightambient[i],  gstate_c.lightColor[0][i], 3);
-
-			if (u_lightdiffuse[i] != 0) 
-				SetFloatArray(u_lightdiffuse[i],  gstate_c.lightColor[1][i], 3);
-				
-			if (u_lightspecular[i] != 0) 
-				SetFloatArray(u_lightspecular[i],  gstate_c.lightColor[2][i], 3);
 		}
 	}
 

GPU/Directx9/ShaderManagerDX9.h

@@ -41,6 +41,8 @@ protected:
 	void SetMatrix(D3DXHANDLE uniform, const float* pMatrix);
 	void SetFloatArray(D3DXHANDLE uniform, const float* pArray, int len);
 	void SetFloat(D3DXHANDLE uniform, float value);
+	void SetFloat24Uniform3(D3DXHANDLE uniform, const u32 data[3]);
+
 public:
 	LinkedShaderDX9(VSShader *vs, PSShader *fs, u32 vertType, bool useHWTransform);
 	~LinkedShaderDX9();

GPU/Directx9/TransformPipelineDX9.cpp

@@ -32,6 +32,7 @@
 #include "GPU/ge_constants.h"
 
 #include "GPU/Common/TextureDecoder.h"
+#include "GPU/Common/TransformCommon.h"
 #include "GPU/Directx9/StateMappingDX9.h"
 #include "GPU/Directx9/TextureCacheDX9.h"
 #include "GPU/Directx9/TransformPipelineDX9.h"
@@ -130,169 +131,6 @@ void TransformDrawEngineDX9::DestroyDeviceObjects() {
 	ClearTrackedVertexArrays();
 }
 
-namespace {
-using namespace DX9;
-
-// Convenient way to do precomputation to save the parts of the lighting calculation
-// that's common between the many vertices of a draw call.
-class Lighter {
-public:
-	Lighter(int vertType);
-	void Light(float colorOut0[4], float colorOut1[4], const float colorIn[4], const Vec3f &pos, const Vec3f &normal);
-
-private:
-	Color4 globalAmbient;
-	Color4 materialEmissive;
-	Color4 materialAmbient;
-	Color4 materialDiffuse;
-	Color4 materialSpecular;
-	float specCoef_;
-	// Vec3f viewer_;
-	bool doShadeMapping_;
-	int materialUpdate_;
-};
-
-Lighter::Lighter(int vertType) {
-	doShadeMapping_ = gstate.getUVGenMode() == GE_TEXMAP_ENVIRONMENT_MAP;
-	materialEmissive.GetFromRGB(gstate.materialemissive);
-	materialEmissive.a = 0.0f;
-	globalAmbient.GetFromRGB(gstate.ambientcolor);
-	globalAmbient.GetFromA(gstate.ambientalpha);
-	materialAmbient.GetFromRGB(gstate.materialambient);
-	materialAmbient.GetFromA(gstate.materialalpha);
-	materialDiffuse.GetFromRGB(gstate.materialdiffuse);
-	materialDiffuse.a = 1.0f;
-	materialSpecular.GetFromRGB(gstate.materialspecular);
-	materialSpecular.a = 1.0f;
-	specCoef_ = getFloat24(gstate.materialspecularcoef);
-	// viewer_ = Vec3f(-gstate.viewMatrix[9], -gstate.viewMatrix[10], -gstate.viewMatrix[11]);
-	bool hasColor = (vertType & GE_VTYPE_COL_MASK) != 0;
-	materialUpdate_ = hasColor ? gstate.materialupdate & 7 : 0;
-}
-
-void Lighter::Light(float colorOut0[4], float colorOut1[4], const float colorIn[4], const Vec3f &pos, const Vec3f &norm)
-{
-	// Color are in dx format
-	Color4 in;
-
-	in.a = colorIn[0];
-	in.r = colorIn[1];
-	in.g = colorIn[2];
-	in.b = colorIn[3];
-
-
-	const Color4 *ambient;
-	if (materialUpdate_ & 1)
-		ambient = ∈
-	else
-		ambient = &materialAmbient;
-
-	const Color4 *diffuse;
-	if (materialUpdate_ & 2)
-		diffuse = ∈
-	else
-		diffuse = &materialDiffuse;
-
-	const Color4 *specular;
-	if (materialUpdate_ & 4)
-		specular = ∈
-	else
-		specular = &materialSpecular;
-
-	Color4 lightSum0 = globalAmbient * *ambient + materialEmissive;
-	Color4 lightSum1(0, 0, 0, 0);
-
-	for (int l = 0; l < 4; l++)
-	{
-		// can we skip this light?
-		if (!gstate.isLightChanEnabled(l))
-			continue;
-
-		GELightType type = gstate.getLightType(l);
-
-		Vec3f toLight(0,0,0);
-		Vec3f lightDir(0,0,0);
-
-		if (type == GE_LIGHTTYPE_DIRECTIONAL)
-			toLight = Vec3f(gstate_c.lightpos[l]);  // lightdir is for spotlights
-		else
-			toLight = Vec3f(gstate_c.lightpos[l]) - pos;
-
-		bool doSpecular = gstate.isUsingSpecularLight(l);
-		bool poweredDiffuse = gstate.isUsingPoweredDiffuseLight(l);
-
-		float distanceToLight = toLight.Length();
-		float dot = 0.0f;
-		float angle = 0.0f;
-		float lightScale = 0.0f;
-
-		if (distanceToLight > 0.0f) {
-			toLight /= distanceToLight;
-			dot = Dot(toLight, norm);
-		}
-		// Clamp dot to zero.
-		if (dot < 0.0f) dot = 0.0f;
-
-		if (poweredDiffuse)
-			dot = powf(dot, specCoef_);
-
-		// Attenuation
-		switch (type) {
-		case GE_LIGHTTYPE_DIRECTIONAL:
-			lightScale = 1.0f;
-			break;
-		case GE_LIGHTTYPE_POINT:
-			lightScale = clamp(1.0f / (gstate_c.lightatt[l][0] + gstate_c.lightatt[l][1]*distanceToLight + gstate_c.lightatt[l][2]*distanceToLight*distanceToLight), 0.0f, 1.0f);
-			break;
-		case GE_LIGHTTYPE_SPOT:
-		case GE_LIGHTTYPE_UNKNOWN:
-			lightDir = gstate_c.lightdir[l];
-			angle = Dot(toLight.Normalized(), lightDir.Normalized());
-			if (angle >= gstate_c.lightangle[l])
-				lightScale = clamp(1.0f / (gstate_c.lightatt[l][0] + gstate_c.lightatt[l][1]*distanceToLight + gstate_c.lightatt[l][2]*distanceToLight*distanceToLight), 0.0f, 1.0f) * powf(angle, gstate_c.lightspotCoef[l]);
-			break;
-		default:
-			// ILLEGAL
-			break;
-		}
-
-		Color4 lightDiff(gstate_c.lightColor[1][l], 0.0f);
-		Color4 diff = (lightDiff * *diffuse) * dot;
-
-		// Real PSP specular
-		Vec3f toViewer(0,0,1);
-		// Better specular
-		// Vec3f toViewer = (viewer - pos).Normalized();
-
-		if (doSpecular)
-		{
-			Vec3f halfVec = (toLight + toViewer);
-			halfVec.Normalize();
-
-			dot = Dot(halfVec, norm);
-			if (dot > 0.0f)
-			{
-				Color4 lightSpec(gstate_c.lightColor[2][l], 0.0f);
-				lightSum1 += (lightSpec * *specular * (powf(dot, specCoef_) * lightScale));
-			}
-		}
-
-		if (gstate.isLightChanEnabled(l))
-		{
-			Color4 lightAmbient(gstate_c.lightColor[0][l], 0.0f);
-			lightSum0 += (lightAmbient * *ambient + diff) * lightScale;
-		}
-	}
-
-	// 4?
-	for (int i = 0; i < 4; i++) {
-		colorOut0[i] = lightSum0[i] > 1.0f ? 1.0f : lightSum0[i];
-		colorOut1[i] = lightSum1[i] > 1.0f ? 1.0f : lightSum1[i];
-	}
-}
-
-}  // namespace
-
 struct DeclTypeInfo {
 	u32 type;
 	const char * name;
@@ -747,8 +585,8 @@ void TransformDrawEngineDX9::SoftwareTransformAndDraw(
 				case GE_TEXMAP_ENVIRONMENT_MAP:
 					// Shade mapping - use two light sources to generate U and V.
 					{
-						Vec3f lightpos0 = Vec3f(gstate_c.lightpos[gstate.getUVLS0()]).Normalized();
+						Vec3f lightpos0 = Vec3f(&lighter.lpos[gstate.getUVLS0()]).Normalized();
-						Vec3f lightpos1 = Vec3f(gstate_c.lightpos[gstate.getUVLS1()]).Normalized();
+						Vec3f lightpos1 = Vec3f(&lighter.lpos[gstate.getUVLS1()]).Normalized();
 
 						uv[0] = (1.0f + Dot(lightpos0, normal))/2.0f;
 						uv[1] = (1.0f - Dot(lightpos1, normal))/2.0f;

GPU/GLES/GLES_GPU.cpp

@@ -1498,81 +1498,45 @@ void GLES_GPU::ExecuteOpInternal(u32 op, u32 diff) {
 		break;
 
 	case GE_CMD_LX0:case GE_CMD_LY0:case GE_CMD_LZ0:
-	case GE_CMD_LX1:case GE_CMD_LY1:case GE_CMD_LZ1:
-	case GE_CMD_LX2:case GE_CMD_LY2:case GE_CMD_LZ2:
-	case GE_CMD_LX3:case GE_CMD_LY3:case GE_CMD_LZ3:
-		{
-			int n = cmd - GE_CMD_LX0;
-			int l = n / 3;
-			int c = n % 3;
-			gstate_c.lightpos[l][c] = getFloat24(data);
-			shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
-		break;
-
 	case GE_CMD_LDX0:case GE_CMD_LDY0:case GE_CMD_LDZ0:
-	case GE_CMD_LDX1:case GE_CMD_LDY1:case GE_CMD_LDZ1:
-	case GE_CMD_LDX2:case GE_CMD_LDY2:case GE_CMD_LDZ2:
-	case GE_CMD_LDX3:case GE_CMD_LDY3:case GE_CMD_LDZ3:
-		{
-			int n = cmd - GE_CMD_LDX0;
-			int l = n / 3;
-			int c = n % 3;
-			gstate_c.lightdir[l][c] = getFloat24(data);
-			shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
-		break;
-
 	case GE_CMD_LKA0:case GE_CMD_LKB0:case GE_CMD_LKC0:
+	case GE_CMD_LKS0:  // spot coef ("conv")
+	case GE_CMD_LKO0: // light angle ("cutoff")
+	case GE_CMD_LAC0:
+	case GE_CMD_LDC0:
+	case GE_CMD_LSC0:
+		shaderManager_->DirtyUniform(DIRTY_LIGHT0);
+		break;
+
+	case GE_CMD_LX1:case GE_CMD_LY1:case GE_CMD_LZ1:
+	case GE_CMD_LDX1:case GE_CMD_LDY1:case GE_CMD_LDZ1:
 	case GE_CMD_LKA1:case GE_CMD_LKB1:case GE_CMD_LKC1:
-	case GE_CMD_LKA2:case GE_CMD_LKB2:case GE_CMD_LKC2:
-	case GE_CMD_LKA3:case GE_CMD_LKB3:case GE_CMD_LKC3:
-		{
-			int n = cmd - GE_CMD_LKA0;
-			int l = n / 3;
-			int c = n % 3;
-			gstate_c.lightatt[l][c] = getFloat24(data);
-			shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
-		break;
-
-	case GE_CMD_LKS0:
 	case GE_CMD_LKS1:
-	case GE_CMD_LKS2:
-	case GE_CMD_LKS3:
-		{
-			int l = cmd - GE_CMD_LKS0;
-			gstate_c.lightspotCoef[l] = getFloat24(data);
-			shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
-		break;
-
-	case GE_CMD_LKO0:
 	case GE_CMD_LKO1:
-	case GE_CMD_LKO2:
+	case GE_CMD_LAC1:
-	case GE_CMD_LKO3:
+	case GE_CMD_LDC1:
-		{
+	case GE_CMD_LSC1:
-			int l = cmd - GE_CMD_LKO0;
+		shaderManager_->DirtyUniform(DIRTY_LIGHT1);
-			gstate_c.lightangle[l] = getFloat24(data);
-			shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
-		}
 		break;
-
+	case GE_CMD_LX2:case GE_CMD_LY2:case GE_CMD_LZ2:
-	case GE_CMD_LAC0:case GE_CMD_LAC1:case GE_CMD_LAC2:case GE_CMD_LAC3:
+	case GE_CMD_LDX2:case GE_CMD_LDY2:case GE_CMD_LDZ2:
-	case GE_CMD_LDC0:case GE_CMD_LDC1:case GE_CMD_LDC2:case GE_CMD_LDC3:
+	case GE_CMD_LKA2:case GE_CMD_LKB2:case GE_CMD_LKC2:
-	case GE_CMD_LSC0:case GE_CMD_LSC1:case GE_CMD_LSC2:case GE_CMD_LSC3:
+	case GE_CMD_LKS2:
-		{
+	case GE_CMD_LKO2:
-			float r = (float)(data & 0xff) * (1.0f / 255.0f);
+	case GE_CMD_LAC2:
-			float g = (float)((data >> 8) & 0xff) * (1.0f / 255.0f);
+	case GE_CMD_LDC2:
-			float b = (float)(data >> 16) * (1.0f / 255.0f);
+	case GE_CMD_LSC2:
-
+		shaderManager_->DirtyUniform(DIRTY_LIGHT2);
-			int l = (cmd - GE_CMD_LAC0) / 3;
+		break;
-			int t = (cmd - GE_CMD_LAC0) % 3;
+	case GE_CMD_LX3:case GE_CMD_LY3:case GE_CMD_LZ3:
-			gstate_c.lightColor[t][l][0] = r;
+	case GE_CMD_LDX3:case GE_CMD_LDY3:case GE_CMD_LDZ3:
-			gstate_c.lightColor[t][l][1] = g;
+	case GE_CMD_LKA3:case GE_CMD_LKB3:case GE_CMD_LKC3:
-			gstate_c.lightColor[t][l][2] = b;
+	case GE_CMD_LKS3:
-			shaderManager_->DirtyUniform(DIRTY_LIGHT0 << l);
+	case GE_CMD_LKO3:
-		}
+	case GE_CMD_LAC3:
+	case GE_CMD_LDC3:
+	case GE_CMD_LSC3:
+		shaderManager_->DirtyUniform(DIRTY_LIGHT3);
 		break;
 
 	case GE_CMD_VIEWPORTX1:

GPU/GLES/ShaderManager.cpp

@@ -308,6 +308,13 @@ static void SetColorUniform3ExtraFloat(int uniform, u32 color, float extra) {
 	glUniform4fv(uniform, 1, col);
 }
 
+static void SetFloat24Uniform3(int uniform, const u32 data[3]) {
+	const u32 col[3] = {
+		data[0] << 8, data[1] << 8, data[2] << 8
+	};
+	glUniform3fv(uniform, 1, (const GLfloat *)&col[0]);
+}
+
 static void SetMatrix4x3(int uniform, const float *m4x3) {
 	float m4x4[16];
 	ConvertMatrix4x3To4x4(m4x4, m4x3);
@@ -532,28 +539,30 @@ void LinkedShader::UpdateUniforms(u32 vertType) {
 
 	for (int i = 0; i < 4; i++) {
 		if (dirty & (DIRTY_LIGHT0 << i)) {
+			if (u_lightpos[i] != -1) {
 				if (gstate.isDirectionalLight(i)) {
 					// Prenormalize
-				float x = gstate_c.lightpos[i][0];
+					float x = getFloat24(gstate.lpos[i * 3 + 0]);
-				float y = gstate_c.lightpos[i][1];
+					float y = getFloat24(gstate.lpos[i * 3 + 1]);
-				float z = gstate_c.lightpos[i][2];
+					float z = getFloat24(gstate.lpos[i * 3 + 2]);
 					float len = sqrtf(x*x + y*y + z*z);
 					if (len == 0.0f)
 						len = 1.0f;
 					else
 						len = 1.0f / len;
 					float vec[3] = { x * len, y * len, z * len };
-				if (u_lightpos[i] != -1) glUniform3fv(u_lightpos[i], 1, vec);
+					glUniform3fv(u_lightpos[i], 1, vec);
 				} else {
-				if (u_lightpos[i] != -1) glUniform3fv(u_lightpos[i], 1, gstate_c.lightpos[i]);
+					SetFloat24Uniform3(u_lightpos[i], &gstate.lpos[i * 3]);
 				}
-			if (u_lightdir[i] != -1) glUniform3fv(u_lightdir[i], 1, gstate_c.lightdir[i]);
+			}
-			if (u_lightatt[i] != -1) glUniform3fv(u_lightatt[i], 1, gstate_c.lightatt[i]);
+			if (u_lightdir[i] != -1) SetFloat24Uniform3(u_lightdir[i], &gstate.ldir[i * 3]);
-			if (u_lightangle[i] != -1) glUniform1f(u_lightangle[i], gstate_c.lightangle[i]);
+			if (u_lightatt[i] != -1) SetFloat24Uniform3(u_lightatt[i], &gstate.latt[i * 3]);
-			if (u_lightspotCoef[i] != -1) glUniform1f(u_lightspotCoef[i], gstate_c.lightspotCoef[i]);
+			if (u_lightangle[i] != -1) glUniform1f(u_lightangle[i], getFloat24(gstate.lcutoff[i]));
-			if (u_lightambient[i] != -1) glUniform3fv(u_lightambient[i], 1, gstate_c.lightColor[0][i]);
+			if (u_lightspotCoef[i] != -1) glUniform1f(u_lightspotCoef[i], getFloat24(gstate.lconv[i]));
-			if (u_lightdiffuse[i] != -1) glUniform3fv(u_lightdiffuse[i], 1, gstate_c.lightColor[1][i]);
+			if (u_lightambient[i] != -1) SetColorUniform3(u_lightambient[i], gstate.lcolor[i * 3]);
-			if (u_lightspecular[i] != -1) glUniform3fv(u_lightspecular[i], 1, gstate_c.lightColor[2][i]);
+			if (u_lightdiffuse[i] != -1) SetColorUniform3(u_lightdiffuse[i], gstate.lcolor[i * 3 + 1]);
+			if (u_lightspecular[i] != -1) SetColorUniform3(u_lightspecular[i], gstate.lcolor[i * 3 + 2]);
 		}
 	}
 }

GPU/GLES/SoftwareTransform.cpp

@@ -22,6 +22,7 @@
 #include "GPU/GPUState.h"
 #include "GPU/Math3D.h"
 #include "GPU/Common/VertexDecoderCommon.h"
+#include "GPU/Common/TransformCommon.h"
 #include "GPU/GLES/ShaderManager.h"
 #include "GPU/GLES/TransformPipeline.h"
 
@@ -44,160 +45,6 @@
 
 extern const GLuint glprim[8];
 
-// Check for max first as clamping to max is more common than min when lighting.
-inline float clamp(float in, float min, float max) {
-	return in > max ? max : (in < min ? min : in);
-}
-
-// Convenient way to do precomputation to save the parts of the lighting calculation
-// that's common between the many vertices of a draw call.
-class Lighter {
-public:
-	Lighter(int vertType);
-	void Light(float colorOut0[4], float colorOut1[4], const float colorIn[4], const Vec3f &pos, const Vec3f &normal);
-
-private:
-	Color4 globalAmbient;
-	Color4 materialEmissive;
-	Color4 materialAmbient;
-	Color4 materialDiffuse;
-	Color4 materialSpecular;
-	float specCoef_;
-	// Vec3f viewer_;
-	bool doShadeMapping_;
-	int materialUpdate_;
-};
-
-Lighter::Lighter(int vertType) {
-	if (!gstate.isLightingEnabled())
-		return;
-
-	doShadeMapping_ = gstate.getUVGenMode() == GE_TEXMAP_ENVIRONMENT_MAP;
-	materialEmissive.GetFromRGB(gstate.materialemissive);
-	materialEmissive.a = 0.0f;
-	globalAmbient.GetFromRGB(gstate.ambientcolor);
-	globalAmbient.GetFromA(gstate.ambientalpha);
-	materialAmbient.GetFromRGB(gstate.materialambient);
-	materialAmbient.GetFromA(gstate.materialalpha);
-	materialDiffuse.GetFromRGB(gstate.materialdiffuse);
-	materialDiffuse.a = 1.0f;
-	materialSpecular.GetFromRGB(gstate.materialspecular);
-	materialSpecular.a = 1.0f;
-	specCoef_ = getFloat24(gstate.materialspecularcoef);
-	// viewer_ = Vec3f(-gstate.viewMatrix[9], -gstate.viewMatrix[10], -gstate.viewMatrix[11]);
-	bool hasColor = (vertType & GE_VTYPE_COL_MASK) != 0;
-	materialUpdate_ = hasColor ? gstate.materialupdate & 7 : 0;
-}
-
-void Lighter::Light(float colorOut0[4], float colorOut1[4], const float colorIn[4], const Vec3f &pos, const Vec3f &norm) {
-	Color4 in(colorIn);
-
-	const Color4 *ambient;
-	if (materialUpdate_ & 1)
-		ambient = &in;
-	else
-		ambient = &materialAmbient;
-
-	const Color4 *diffuse;
-	if (materialUpdate_ & 2)
-		diffuse = &in;
-	else
-		diffuse = &materialDiffuse;
-
-	const Color4 *specular;
-	if (materialUpdate_ & 4)
-		specular = &in;
-	else
-		specular = &materialSpecular;
-
-	Color4 lightSum0 = globalAmbient * *ambient + materialEmissive;
-	Color4 lightSum1(0, 0, 0, 0);
-
-	for (int l = 0; l < 4; l++) {
-		// can we skip this light?
-		if (!gstate.isLightChanEnabled(l))
-			continue;
-
-		GELightType type = gstate.getLightType(l);
-		
-		Vec3f toLight(0,0,0);
-		Vec3f lightDir(0,0,0);
-		
-		if (type == GE_LIGHTTYPE_DIRECTIONAL)
-			toLight = Vec3f(gstate_c.lightpos[l]);  // lightdir is for spotlights
-		else
-			toLight = Vec3f(gstate_c.lightpos[l]) - pos;
-
-		bool doSpecular = gstate.isUsingSpecularLight(l);
-		bool poweredDiffuse = gstate.isUsingPoweredDiffuseLight(l);
-		
-		float distanceToLight = toLight.Length();
-		float dot = 0.0f;
-		float angle = 0.0f;
-		float lightScale = 0.0f;
-		
-		if (distanceToLight > 0.0f) {
-			toLight /= distanceToLight;
-			dot = Dot(toLight, norm);
-		}
-		// Clamp dot to zero.
-		if (dot < 0.0f) dot = 0.0f;
-
-		if (poweredDiffuse)
-			dot = powf(dot, specCoef_);
-
-		// Attenuation
-		switch (type) {
-		case GE_LIGHTTYPE_DIRECTIONAL:
-			lightScale = 1.0f;
-			break;
-		case GE_LIGHTTYPE_POINT:
-			lightScale = clamp(1.0f / (gstate_c.lightatt[l][0] + gstate_c.lightatt[l][1]*distanceToLight + gstate_c.lightatt[l][2]*distanceToLight*distanceToLight), 0.0f, 1.0f);
-			break;
-		case GE_LIGHTTYPE_SPOT:
-		case GE_LIGHTTYPE_UNKNOWN:
-			lightDir = gstate_c.lightdir[l];
-			angle = Dot(toLight.Normalized(), lightDir.Normalized());
-			if (angle >= gstate_c.lightangle[l])
-				lightScale = clamp(1.0f / (gstate_c.lightatt[l][0] + gstate_c.lightatt[l][1]*distanceToLight + gstate_c.lightatt[l][2]*distanceToLight*distanceToLight), 0.0f, 1.0f) * powf(angle, gstate_c.lightspotCoef[l]);
-			break;
-		default:
-			// ILLEGAL
-			break;
-		}
-
-		Color4 lightDiff(gstate_c.lightColor[1][l], 0.0f);
-		Color4 diff = (lightDiff * *diffuse) * dot;
-
-		// Real PSP specular
-		Vec3f toViewer(0,0,1);
-		// Better specular
-		// Vec3f toViewer = (viewer - pos).Normalized();
-
-		if (doSpecular) {
-			Vec3f halfVec = (toLight + toViewer);
-			halfVec.Normalize();
-
-			dot = Dot(halfVec, norm);
-			if (dot > 0.0f) {
-				Color4 lightSpec(gstate_c.lightColor[2][l], 0.0f);
-				lightSum1 += (lightSpec * *specular * (powf(dot, specCoef_) * lightScale));
-			}
-		}
-
-		if (gstate.isLightChanEnabled(l)) {
-			Color4 lightAmbient(gstate_c.lightColor[0][l], 0.0f);
-			lightSum0 += (lightAmbient * *ambient + diff) * lightScale;
-		}
-	}
-
-	// 4?
-	for (int i = 0; i < 4; i++) {
-		colorOut0[i] = lightSum0[i] > 1.0f ? 1.0f : lightSum0[i];
-		colorOut1[i] = lightSum1[i] > 1.0f ? 1.0f : lightSum1[i];
-	}
-}
-
 // The verts are in the order:  BR BL TL TR
 static void SwapUVs(TransformedVertex &a, TransformedVertex &b) {
 	float tempu = a.u;
@@ -498,8 +345,8 @@ void TransformDrawEngine::SoftwareTransformAndDraw(
 			case GE_TEXMAP_ENVIRONMENT_MAP:
 				// Shade mapping - use two light sources to generate U and V.
 				{
-					Vec3f lightpos0 = Vec3f(gstate_c.lightpos[gstate.getUVLS0()]).Normalized();
+					Vec3f lightpos0 = Vec3f(&lighter.lpos[gstate.getUVLS0() * 3]).Normalized();
-					Vec3f lightpos1 = Vec3f(gstate_c.lightpos[gstate.getUVLS1()]).Normalized();
+					Vec3f lightpos1 = Vec3f(&lighter.lpos[gstate.getUVLS1() * 3]).Normalized();
 
 					uv[0] = (1.0f + Dot(lightpos0, normal))/2.0f;
 					uv[1] = (1.0f + Dot(lightpos1, normal))/2.0f;

GPU/GLES/TransformPipeline.h

@@ -244,41 +244,3 @@ private:
 
 	UVScale *uvScale;
 };
-
-// Only used by SW transform
-struct Color4 {
-	float r, g, b, a;
-
-	Color4() : r(0), g(0), b(0), a(0) { }
-	Color4(float _r, float _g, float _b, float _a=1.0f)
-		: r(_r), g(_g), b(_b), a(_a) {
-	}
-	Color4(const float in[4]) {r=in[0];g=in[1];b=in[2];a=in[3];}
-	Color4(const float in[3], float alpha) {r=in[0];g=in[1];b=in[2];a=alpha;}
-
-	const float &operator [](int i) const {return *(&r + i);}
-
-	Color4 operator *(float f) const {
-		return Color4(f*r,f*g,f*b,f*a);
-	}
-	Color4 operator *(const Color4 &c) const {
-		return Color4(r*c.r,g*c.g,b*c.b,a*c.a);
-	}
-	Color4 operator +(const Color4 &c) const {
-		return Color4(r+c.r,g+c.g,b+c.b,a+c.a);
-	}
-	void operator +=(const Color4 &c) {
-		r+=c.r;
-		g+=c.g;
-		b+=c.b;
-		a+=c.a;
-	}
-	void GetFromRGB(u32 col) {
-		b = ((col>>16) & 0xff)/255.0f;
-		g = ((col>>8) & 0xff)/255.0f;
-		r = ((col>>0) & 0xff)/255.0f;
-	}
-	void GetFromA(u32 col) {
-		a = (col&0xff)/255.0f;
-	}
-};

GPU/GPU.vcxproj

@@ -171,6 +171,7 @@
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">true</ExcludedFromBuild>
     </ClInclude>
+    <ClInclude Include="Common\TransformCommon.h" />
     <ClInclude Include="Common\VertexDecoderCommon.h" />
     <ClInclude Include="Debugger\Breakpoints.h" />
     <ClInclude Include="Debugger\Stepping.h" />
@@ -222,6 +223,7 @@
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">true</ExcludedFromBuild>
     </ClCompile>
+    <ClCompile Include="Common\TransformCommon.cpp" />
     <ClCompile Include="Common\VertexDecoderCommon.cpp" />
     <ClCompile Include="Debugger\Breakpoints.cpp" />
     <ClCompile Include="Debugger\Stepping.cpp" />

GPU/GPU.vcxproj.filters

@@ -162,6 +162,9 @@
     <ClInclude Include="Common\TextureDecoderNEON.h">
       <Filter>Common</Filter>
     </ClInclude>
+    <ClInclude Include="Common\TransformCommon.h">
+      <Filter>Common</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="Math3D.cpp">
@@ -302,6 +305,9 @@
     <ClCompile Include="GLES\VertexDecoderX86.cpp">
       <Filter>GLES</Filter>
     </ClCompile>
+    <ClCompile Include="Common\TransformCommon.cpp">
+      <Filter>Common</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <None Include="CMakeLists.txt" />

GPU/GPUState.cpp

@@ -270,7 +270,7 @@ struct GPUStateCache_v0
 };
 
 void GPUStateCache::DoState(PointerWrap &p) {
-	auto s = p.Section("GPUStateCache", 0, 1);
+	auto s = p.Section("GPUStateCache", 0, 2);
 	if (!s) {
 		// Old state, this was not versioned.
 		GPUStateCache_v0 old;
@@ -302,12 +302,19 @@ void GPUStateCache::DoState(PointerWrap &p) {
 		p.Do(flipTexture);
 	}
 
-	p.Do(lightpos);
+	if (s < 2) {
-	p.Do(lightdir);
+		float l12[12];
-	p.Do(lightatt);
+		float l4[3];
-	p.Do(lightColor);
+		p.Do(l12);  // lightpos
-	p.Do(lightangle);
+		p.Do(l12);  // lightdir
-	p.Do(lightspotCoef);
+		p.Do(l12);  // lightattr
+		p.Do(l12);  // lightcol0
+		p.Do(l12);  // lightcol1
+		p.Do(l12);  // lightcol2
+		p.Do(l4);    // lightangle
+		p.Do(l4);  // lightspot
+	}
+
 	p.Do(morphWeights);
 
 	p.Do(curTextureWidth);

GPU/GPUState.h

@@ -458,12 +458,6 @@ struct GPUStateCache
 	UVScale uv;
 	bool flipTexture;
 
-	float lightpos[4][3];
-	float lightdir[4][3];
-	float lightatt[4][3];
-	float lightColor[3][4][3];  // Ambient Diffuse Specular
-	float lightangle[4]; // spotlight cone angle (cosine)
-	float lightspotCoef[4]; // spotlight dropoff
 	float morphWeights[8];
 
 	u32 curTextureWidth;

GPU/Null/NullGpu.cpp

@@ -385,7 +385,6 @@ void NullGPU::ExecuteOp(u32 op, u32 diff) {
 			int c = n % 3;
 			float val = getFloat24(data);
 			DEBUG_LOG(G3D,"DL Light %i %c pos: %f", l, c+'X', val);
-			gstate_c.lightpos[l][c] = val;
 		}
 		break;
 
@@ -399,7 +398,6 @@ void NullGPU::ExecuteOp(u32 op, u32 diff) {
 			int c = n % 3;
 			float val = getFloat24(data);
 			DEBUG_LOG(G3D,"DL Light %i %c dir: %f", l, c+'X', val);
-			gstate_c.lightdir[l][c] = val;
 		}
 		break;
 
@@ -413,7 +411,6 @@ void NullGPU::ExecuteOp(u32 op, u32 diff) {
 			int c = n % 3;
 			float val = getFloat24(data);
 			DEBUG_LOG(G3D,"DL Light %i %c att: %f", l, c+'X', val);
-			gstate_c.lightatt[l][c] = val;
 		}
 		break;
 
@@ -428,9 +425,7 @@ void NullGPU::ExecuteOp(u32 op, u32 diff) {
 
 			int l = (cmd - GE_CMD_LAC0) / 3;
 			int t = (cmd - GE_CMD_LAC0) % 3;
-			gstate_c.lightColor[t][l][0] = r;
+			// DEBUG_LOG(G3D, "DL Light color %i %c att: %f", l, c + 'X', val);
-			gstate_c.lightColor[t][l][1] = g;
-			gstate_c.lightColor[t][l][2] = b;
 	}
 		break;
 

Qt/Core.pro

@@ -62,6 +62,7 @@ SOURCES += $$P/Core/*.cpp \ # Core
 	$$P/GPU/Common/IndexGenerator.cpp \
 	$$P/GPU/Common/TextureDecoder.cpp \
 	$$P/GPU/Common/VertexDecoderCommon.cpp \
+	$$P/GPU/Common/TransformCommon.cpp \
 	$$P/GPU/Common/PostShader.cpp \
 	$$P/ext/libkirk/*.c \ # Kirk
 	$$P/ext/xxhash.c \ # xxHash

android/jni/Android.mk

@@ -137,6 +137,7 @@ EXEC_AND_LIB_FILES := \
   $(SRC)/GPU/GeDisasm.cpp \
   $(SRC)/GPU/Common/IndexGenerator.cpp.arm \
   $(SRC)/GPU/Common/VertexDecoderCommon.cpp.arm \
+  $(SRC)/GPU/Common/TransformCommon.cpp.arm \
   $(SRC)/GPU/Common/TextureDecoder.cpp \
   $(SRC)/GPU/Common/PostShader.cpp \
   $(SRC)/GPU/Debugger/Breakpoints.cpp \