Merge pull request #22399 from hrydgard/specular-lighting

Correct the specular lighting math: Fix view vector, use Mitchell approximation

Common/GPU/ShaderWriter.cpp

@@ -37,6 +37,8 @@ const char * const hlsl_preamble_fs =
 "#define inversesqrt rsqrt\n"
 "#define floatBitsToUint asuint\n"
 "#define uintBitsToFloat asfloat\n"
+"#define floatBitsToInt asint\n"
+"#define intBitsToFloat asfloat\n"
 "\n";
 
 static const char * const hlsl_d3d11_preamble_fs =
@@ -71,6 +73,8 @@ static const char * const hlsl_preamble_vs =
 "#define inversesqrt rsqrt\n"
 "#define floatBitsToUint asuint\n"
 "#define uintBitsToFloat asfloat\n"
+"#define floatBitsToInt asint\n"
+"#define intBitsToFloat asfloat\n"
 "\n";
 
 static const char * const semanticNames[] = {

GPU/Common/ShaderId.cpp

@@ -116,8 +116,22 @@ void ComputeVertexShaderID(VShaderID *id_out, u32 vertType, bool useHWTransform,
 			id.SetBits(VS_BIT_LS1, 2, gstate.getUVLS1());
 		}
 
+		if (doShadeMapping) {
+			// Shade mapping depends on the type of its lights and whether they do specular, even when
+			// they're off. The ubershader reads that from u_lightControl instead.
+			if (gstate_c.Use(GPU_USE_LIGHT_UBERSHADER)) {
+				id.SetBit(VS_BIT_LIGHT_UBERSHADER);
+			} else {
+				const int shadeLights[2] = { gstate.getUVLS0(), gstate.getUVLS1() };
+				for (int l : shadeLights) {
+					id.SetBits(VS_BIT_LIGHT0_COMP + 4 * l, 2, gstate.getLightComputation(l));
+					id.SetBits(VS_BIT_LIGHT0_TYPE + 4 * l, 2, gstate.getLightType(l));
+				}
+			}
+		}
+
 		if (gstate.isLightingEnabled()) {
-			// doShadeMapping is stored as UVGenMode, and light type doesn't matter for shade mapping.
+			// doShadeMapping is stored as UVGenMode.
 			id.SetBit(VS_BIT_LIGHTING_ENABLE);
 			if (gstate_c.Use(GPU_USE_LIGHT_UBERSHADER)) {
 				id.SetBit(VS_BIT_LIGHT_UBERSHADER);

GPU/Common/ShaderUniforms.cpp

@@ -12,6 +12,7 @@
 #include "GPU/GPUState.h"
 #include "GPU/Common/FramebufferManagerCommon.h"
 #include "GPU/Common/GPUStateUtils.h"
+#include "GPU/Common/TransformCommon.h"
 #include "GPU/Math3D.h"
 
 using namespace Lin;
@@ -222,7 +223,7 @@ void LightUpdateUniforms(UB_VS_Lights *ub, uint64_t dirtyUniforms) {
 		Uint8x3ToFloat4(ub->materialDiffuse, gstate.materialdiffuse);
 	}
 	if (dirtyUniforms & DIRTY_MATSPECULAR) {
-		Uint8x3ToFloat4_Alpha(ub->materialSpecular, gstate.materialspecular, std::max(0.0f, getFloat24(gstate.materialspecularcoef)));
+		Uint8x3ToFloat4_Alpha(ub->materialSpecular, gstate.materialspecular, std::max(0.0f, PSPSpecularCoef(getFloat24(gstate.materialspecularcoef))));
 	}
 	if (dirtyUniforms & DIRTY_MATEMISSIVE) {
 		// We're not touching the fourth f32 here, because we store an u32 of control bits in it.

GPU/Common/SoftwareTransformCommon.cpp

@@ -333,23 +333,11 @@ SoftwareTransformAction RunSoftwareTransform(SoftwareTransformParams &params, in
 			case GE_TEXMAP_ENVIRONMENT_MAP:
 				// Shade mapping - use two light sources to generate U and V.
 				{
-					auto getLPosFloat = [&](int l, int i) {
-						return getFloat24(gstate.lpos[l * 3 + i]);
-					};
-					auto getLPos = [&](int l) {
-						return Vec3f(getLPosFloat(l, 0), getLPosFloat(l, 1), getLPosFloat(l, 2));
-					};
-					auto calcShadingLPos = [&](int l) {
-						Vec3f pos = getLPos(l);
-						return pos.NormalizedOr001(cpu_info.bSSE4_1);
-					};
-
 					// Might not have lighting enabled, so don't use lighter.
-					Vec3f lightpos0 = calcShadingLPos(gstate.getUVLS0());
-					Vec3f lightpos1 = calcShadingLPos(gstate.getUVLS1());
-
-					uv[0] = (1.0f + Dot(lightpos0, worldnormal))/2.0f;
-					uv[1] = (1.0f + Dot(lightpos1, worldnormal))/2.0f;
+					const Vec3f viewDir = PSPViewDirection(gstate.viewMatrix);
+					const Vec3f worldpos(out[0], out[1], out[2]);
+					uv[0] = PSPShadeMapCoord(gstate.getUVLS0(), worldpos, worldnormal, viewDir);
+					uv[1] = PSPShadeMapCoord(gstate.getUVLS1(), worldpos, worldnormal, viewDir);
 					uv[2] = 1.0f;
 				}
 				break;

GPU/Common/TransformCommon.cpp

@@ -40,8 +40,8 @@ Lighter::Lighter(int vertType) {
 	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]);
+	specCoef_ = PSPSpecularCoef(getFloat24(gstate.materialspecularcoef));
+	viewDir_ = PSPViewDirection(gstate.viewMatrix);
 	bool hasColor = (vertType & GE_VTYPE_COL_MASK) != 0;
 	materialUpdate_ = hasColor ? (gstate.materialupdate & 7) : 0;
 
@@ -117,11 +117,13 @@ void Lighter::Light(float colorOut0[4], float colorOut1[4], const float colorIn[
 			toLight /= distanceToLight;
 			dot = Dot(toLight, norm);
 		}
+		// Specular only applies when the light is in front of the surface.
+		const bool facingLight = dot >= 0.0f;
 		// Clamp dot to zero.
 		if (dot < 0.0f) dot = 0.0f;
 
 		if (poweredDiffuse)
-			dot = powf(dot, specCoef_);
+			dot = PSPLightPow(dot, specCoef_);
 
 		// Attenuation
 		switch (type) {
@@ -136,7 +138,7 @@ void Lighter::Light(float colorOut0[4], float colorOut1[4], const float colorIn[
 			lightDir = ldir[l];
 			angle = Dot(toLight.NormalizedOr001(cpu_info.bSSE4_1), lightDir.NormalizedOr001(cpu_info.bSSE4_1));
 			if (angle >= lcutoff[l])
-				lightScale = clamp(1.0f / (latt[l].x + latt[l].y * distanceToLight + latt[l].z * distanceToLight*distanceToLight), 0.0f, 1.0f) * powf(angle, lconv[l]);
+				lightScale = clamp(1.0f / (latt[l].x + latt[l].y * distanceToLight + latt[l].z * distanceToLight*distanceToLight), 0.0f, 1.0f) * PSPLightPow(angle, lconv[l]);
 			break;
 		default:
 			// ILLEGAL
@@ -146,18 +148,13 @@ void Lighter::Light(float colorOut0[4], float colorOut1[4], const float colorIn[
 		Color4 lightDiff(lcolor[1][l], 0.0f);
 		Color4 diff = (lightDiff * *diffuse) * dot;
 
-		// Real PSP specular
-		static const Vec3f toViewer(0, 0, 1);
-		// Better specular
-		// Vec3f toViewer = (viewer - pos).NormalizedOr001(cpu_info.bSSE4_1);
-
-		if (doSpecular) {
-			Vec3f halfVec = (toLight + toViewer).NormalizedOr001(cpu_info.bSSE4_1);
+		if (doSpecular && facingLight) {
+			Vec3f halfVec = (toLight + viewDir_).NormalizedOr001(cpu_info.bSSE4_1);
 
 			dot = Dot(halfVec, norm);
 			if (dot > 0.0f) {
 				Color4 lightSpec(lcolor[2][l], 0.0f);
-				lightSum1 += (lightSpec * *specular * (powf(dot, specCoef_) * lightScale));
+				lightSum1 += (lightSpec * *specular * (PSPLightPow(dot, specCoef_) * lightScale));
 			}
 		}
 

GPU/Common/TransformCommon.h

@@ -17,11 +17,13 @@
 
 #pragma once
 
+#include <cmath>
 #include <cstring>
 
 #include "Common/CommonTypes.h"
 #include "GPU/Math3D.h"
 #include "GPU/GPU.h"
+#include "GPU/GPUState.h"
 
 struct Color4 {
 	float r, g, b, a;
@@ -60,6 +62,59 @@ struct Color4 {
 	}
 };
 
+// The GE's pow() for specular, powered diffuse and the spot exponent: 1 for e <= 0, else 0 for
+// x <= 0. Otherwise exp2(e * log2(x)) with log2 and exp2 each a straight line between powers of two
+// (Mitchell's approximation), which is what reading a float's bits as an integer gives: exponent
+// plus mantissa, scaled by 2^23. Matches hardware within one step of 255 (gpu/lighting/specular).
+inline float PSPLightPow(float x, float e) {
+	if (!(x > 0.0f)) {
+		return e > 0.0f ? 0.0f : 1.0f;
+	}
+	int32_t ix;
+	memcpy(&ix, &x, sizeof(ix));
+	float t = (e > 0.0f ? e : 0.0f) * (float)(ix - 0x3F800000) + 1065353216.0f;
+	// Also turns NaN into 0, and stays below infinity's bits.
+	t = t >= 0.0f ? (t < 2139095039.0f ? t : 2139095039.0f) : 0.0f;
+	int32_t iy = (int32_t)t;
+	float y;
+	memcpy(&y, &iy, sizeof(y));
+	return y;
+}
+
+// The GE only uses the top 4 bits of the specular coefficient's mantissa.
+inline float PSPSpecularCoef(float e) {
+	u32 bits;
+	memcpy(&bits, &e, sizeof(bits));
+	bits &= 0xFFF80000;
+	memcpy(&e, &bits, sizeof(bits));
+	return e;
+}
+
+// The viewer is at infinity along view space +z, so in world space it's the view matrix's third column.
+inline Vec3f PSPViewDirection(const float viewMatrix[12]) {
+	return Vec3f(viewMatrix[2], viewMatrix[5], viewMatrix[8]).NormalizedOr001(false);
+}
+
+inline Vec3f NormalizedOr000(const Vec3f &v) {
+	float len2 = v.Length2();
+	return len2 > 0.0f ? v * (1.0f / sqrtf(len2)) : Vec3f(0.0f, 0.0f, 0.0f);
+}
+
+// Shade mapping (environment map UV gen) coordinate from light l: (N.L + 1) / 2, with L the light's
+// direction as lighting sees it (a zero vector stays zero), or the half vector if the light does
+// specular. Lighting and light enables don't matter (gpu/lighting/shademap).
+inline float PSPShadeMapCoord(int l, const Vec3f &worldpos, const Vec3f &worldnormal, const Vec3f &viewDir) {
+	Vec3f L(getFloat24(gstate.lpos[l * 3]), getFloat24(gstate.lpos[l * 3 + 1]), getFloat24(gstate.lpos[l * 3 + 2]));
+	if (gstate.getLightType(l) != GE_LIGHTTYPE_DIRECTIONAL) {
+		L -= worldpos;
+	}
+	L = NormalizedOr000(L);
+	if (gstate.isUsingSpecularLight(l)) {
+		L = NormalizedOr000(L + viewDir);
+	}
+	return (Dot(L, worldnormal) + 1.0f) * 0.5f;
+}
+
 // 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 {
@@ -81,7 +136,7 @@ private:
 	Color4 materialDiffuse;
 	Color4 materialSpecular;
 	float specCoef_;
-	// Vec3f viewer_;
+	Vec3f viewDir_;
 	bool doShadeMapping_;
 	int materialUpdate_;
 

GPU/Common/VertexShaderGenerator.cpp

@@ -135,7 +135,9 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 	int matUpdate = id.Bits(VS_BIT_MATERIAL_UPDATE, 3);
 
 	bool lightUberShader = id.Bit(VS_BIT_LIGHT_UBERSHADER) && enableLighting;  // checking lighting here for the shader test's benefit, in reality if ubershader is set, lighting is set.
-	if (lightUberShader && !compat.bitwiseOps) {
+	// With the ubershader, shade mapping reads its lights' type and computation from u_lightControl.
+	bool shadeUberShader = id.Bit(VS_BIT_LIGHT_UBERSHADER) && doShadeMapping;
+	if ((lightUberShader || shadeUberShader) && !compat.bitwiseOps) {
 		*errorString = "Light ubershader requires bitwise ops in shader language";
 		return false;
 	}
@@ -349,7 +351,7 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 			WRITE(p, "uniform vec4 u_uvscaleoffset;\n");
 			*uniformMask |= DIRTY_UVSCALEOFFSET;
 
-			if (lightUberShader) {
+			if (lightUberShader || shadeUberShader) {
 				p.C("uniform uint u_lightControl;\n");
 				*uniformMask |= DIRTY_LIGHT_CONTROL;
 			}
@@ -428,6 +430,22 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 		WRITE(p, "   float len2 = dot(v, v);\n");
 		WRITE(p, "   return len2 == 0.0 ? vec3(0.0, 0.0, 1.0) : (v * inversesqrt(len2));\n");
 		WRITE(p, "}\n");
+		WRITE(p, "vec3 normalizeOr000(vec3 v) {\n");
+		WRITE(p, "   float len2 = dot(v, v);\n");
+		WRITE(p, "   return len2 == 0.0 ? splat3(0.0) : (v * inversesqrt(len2));\n");
+		WRITE(p, "}\n");
+		// The GE's pow for lighting: 1 for e <= 0, else 0 for x <= 0. Otherwise exp2(e * log2(x)) with
+		// log2 and exp2 each a straight line between powers of two, which is what reading a float's
+		// bits as an integer gives: exponent plus mantissa, scaled by 2^23. Without integers, a true
+		// pow is close enough.
+		WRITE(p, "float pspPow(float x, float e) {\n");
+		if (compat.bitwiseOps) {
+			WRITE(p, "   float t = max(e, 0.0) * float(floatBitsToInt(max(x, 1e-30)) - 0x3F800000) + 1065353216.0;\n");
+			WRITE(p, "   return x > 0.0 || e <= 0.0 ? intBitsToFloat(int(max(t, 0.0))) : 0.0;\n");
+		} else {
+			WRITE(p, "   return e <= 0.0 ? 1.0 : pow(max(x, 0.0), e);\n");
+		}
+		WRITE(p, "}\n");
 	}
 
 	if (ShaderLanguageIsOpenGL(compat.shaderLanguage) || compat.shaderLanguage == GLSL_VULKAN) {
@@ -490,6 +508,14 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 		} else {
 			WRITE(p, "  mediump vec3 worldnormal = normalizeOr001(mul(vec4(0.0, 0.0, %s1.0, 0.0), u_world).xyz);\n", flipNormal ? "-" : "");
 		}
+		if (enableLighting || doShadeMapping) {
+			// The viewer is at infinity along view space +z: in world space, the view matrix's third column.
+			if (compat.shaderLanguage == HLSL_D3D11) {
+				WRITE(p, "  mediump vec3 viewDir = normalizeOr001(vec3(u_view[0].z, u_view[1].z, u_view[2].z));\n");
+			} else {
+				WRITE(p, "  mediump vec3 viewDir = normalizeOr001(u_view[2].xyz);\n");
+			}
+		}
 
 		WRITE(p, "  vec4 viewPos = vec4(mul(vec4(worldpos, 1.0), u_view).xyz, 1.0);\n");
 		if (useSimpleStereo) {
@@ -643,7 +669,7 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 				p.C("      } else {\n");  // type must be 0x02 - GE_LIGHTTYPE_SPOT
 				p.F("        angle = dot(u_lightdir%s, toLight);\n", iStr);
 				p.F("        if (angle >= u_lightangle_spotCoef%s.x) {\n", iStr);
-				p.F("          lightScale = attenuation * (u_lightangle_spotCoef%s.y <= 0.0 ? 1.0 : pow(angle, u_lightangle_spotCoef%s.y));\n", iStr, iStr, iStr);
+				p.F("          lightScale = attenuation * pspPow(angle, u_lightangle_spotCoef%s.y);\n", iStr, iStr);
 				p.C("        } else {\n");
 				p.C("          lightScale = 0.0;\n");
 				p.C("        }\n");
@@ -653,17 +679,13 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 				p.C("    }\n");
 				p.C("    ldot = dot(toLight, worldnormal);\n");
 				p.C("    if (comp == 0x2u) {\n");  // GE_LIGHTCOMP_ONLYPOWDIFFUSE
-				p.C("      ldot = u_matspecular.a > 0.0 ? pow(max(ldot, 0.0), u_matspecular.a) : 1.0;\n");
+				p.C("      ldot = pspPow(ldot, u_matspecular.a);\n");
 				p.C("    }\n");
 				p.F("    diffuse = (u_lightdiffuse%s * diffuseColor) * max(ldot, 0.0);\n", iStr);
 				p.C("    if (comp == 0x1u && ldot >= 0.0) {\n");  // do specular. note - must allow for the >= case, since the u_matspecular.a <= 0.0 case relies on it.
-				p.C("      if (u_matspecular.a > 0.0) {\n");
-				p.C("        vec3 halfVec = toLight + vec3(0.0, 0.0, 1.0);\n");
-				p.C("        float halfInvLen = inversesqrt(dot(halfVec, halfVec));\n");
-				p.C("        ldot = pow(max(dot(halfVec, worldnormal) * halfInvLen, 0.0), u_matspecular.a);\n");
-				p.C("      } else {\n");
-				p.C("        ldot = 1.0;\n");
-				p.C("      }\n");
+				p.C("      vec3 halfVec = toLight + viewDir;\n");
+				p.C("      float halfInvLen = inversesqrt(dot(halfVec, halfVec));\n");
+				p.C("      ldot = pspPow(dot(halfVec, worldnormal) * halfInvLen, u_matspecular.a);\n");
 				p.F("      lightSum1 += u_lightspecular%s * specularColor * ldot * lightScale;\n", iStr);
 				p.C("    }\n");
 				p.F("    lightSum0.rgb += (u_lightambient%s * ambientColor.rgb + diffuse) * lightScale;\n", iStr);
@@ -703,11 +725,7 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 				if (poweredDiffuse) {
 					// pow(0.0, 0.0) may be undefined, but the PSP seems to treat it as 1.0.
 					// Seen in Tales of the World: Radiant Mythology (#2424.)
-					p.C("  if (u_matspecular.a > 0.0) {\n");
-					p.C("    ldot = pow(max(ldot, 0.0), u_matspecular.a);\n");
-					p.C("  } else {\n");
-					p.C("    ldot = 1.0;\n");
-					p.C("  }\n");
+					p.C("  ldot = pspPow(ldot, u_matspecular.a);\n");
 				}
 
 				const char *timesLightScale = " * lightScale";
@@ -724,7 +742,7 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 				case GE_LIGHTTYPE_UNKNOWN:
 					p.F("  angle = dot(u_lightdir%s, toLight);\n", iStr, iStr);
 					p.F("  if (angle >= u_lightangle_spotCoef%s.x) {\n", iStr);
-					p.F("    lightScale = clamp(1.0 / dot(u_lightatt%s, vec3(1.0, distance, distSq)), 0.0, 1.0) * (u_lightangle_spotCoef%s.y <= 0.0 ? 1.0 : pow(max(angle, 0.0), u_lightangle_spotCoef%s.y));\n", iStr, iStr, iStr);
+					p.F("    lightScale = clamp(1.0 / dot(u_lightatt%s, vec3(1.0, distance, distSq)), 0.0, 1.0) * pspPow(angle, u_lightangle_spotCoef%s.y);\n", iStr, iStr);
 					p.C("  } else {\n");
 					p.C("    lightScale = 0.0;\n");
 					p.C("  }\n");
@@ -737,13 +755,9 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 				p.F("  diffuse = (u_lightdiffuse%s * diffuseColor) * max(ldot, 0.0);\n", iStr);
 				if (doSpecular) {
 					p.C("  if (ldot >= 0.0) {\n");
-					p.C("    if (u_matspecular.a > 0.0) {\n");
-					p.C("      vec3 halfVec = toLight + vec3(0.0, 0.0, 1.0);\n");
-					p.C("      float halfInvLen = inversesqrt(dot(halfVec, halfVec));\n");
-					p.C("      ldot = pow(max(dot(halfVec, worldnormal) * halfInvLen, 0.0), u_matspecular.a);\n");
-					p.C("    } else {\n");
-					p.C("      ldot = 1.0;\n");
-					p.C("    }\n");
+					p.C("    vec3 halfVec = toLight + viewDir;\n");
+					p.C("    float halfInvLen = inversesqrt(dot(halfVec, halfVec));\n");
+					p.C("    ldot = pspPow(dot(halfVec, worldnormal) * halfInvLen, u_matspecular.a);\n");
 					p.C("    if (ldot > 0.0)\n");
 					p.F("      lightSum1 += u_lightspecular%s * specularColor * ldot %s;\n", iStr, timesLightScale);
 					p.C("  }\n");
@@ -851,9 +865,31 @@ bool GenerateVertexShader(const VShaderID &id, char *buffer, const ShaderLanguag
 						snprintf(ls0Str, sizeof(ls0Str), "%d", ls0);
 						snprintf(ls1Str, sizeof(ls1Str), "%d", ls1);
 					}
-					std::string lightFactor0 = StringFromFormat("(length(u_lightpos%s) == 0.0 ? worldnormal.z : dot(normalize(u_lightpos%s), worldnormal))", ls0Str, ls0Str);
-					std::string lightFactor1 = StringFromFormat("(length(u_lightpos%s) == 0.0 ? worldnormal.z : dot(normalize(u_lightpos%s), worldnormal))", ls1Str, ls1Str);
-					WRITE(p, "  %sv_texcoord = vec3(u_uvscaleoffset.xy * vec2(1.0 + %s, 1.0 + %s) * 0.5, 1.0);\n", compat.vsOutPrefix, lightFactor0.c_str(), lightFactor1.c_str());
+					// N.L with L the light vector as lighting sees it (zero stays zero), or the half vector
+					// if the light does specular. Whether lighting or the light is on doesn't matter.
+					auto shadeLight = [&](int ls, const char *lsStr, const char *name) {
+						if (shadeUberShader) {
+							p.F("  vec3 %s = u_lightpos%s;\n", name, lsStr);
+							p.F("  if (((u_lightControl >> 0x%02xu) & 0x3u) != 0x0u) %s = u_lightpos%s - worldpos;\n", 4 + 4 * ls + 2, name, lsStr);
+							p.F("  %s = normalizeOr000(%s);\n", name, name);
+							p.F("  if (((u_lightControl >> 0x%02xu) & 0x3u) == 0x1u) %s = normalizeOr000(%s + viewDir);\n", 4 + 4 * ls, name, name);
+							return;
+						}
+						GELightType type = static_cast<GELightType>(id.Bits(VS_BIT_LIGHT0_TYPE + 4 * ls, 2));
+						GELightComputation comp = static_cast<GELightComputation>(id.Bits(VS_BIT_LIGHT0_COMP + 4 * ls, 2));
+						if (type == GE_LIGHTTYPE_DIRECTIONAL) {
+							// Prenormalized.
+							p.F("  vec3 %s = u_lightpos%s;\n", name, lsStr);
+						} else {
+							p.F("  vec3 %s = normalizeOr000(u_lightpos%s - worldpos);\n", name, lsStr);
+						}
+						if (comp == GE_LIGHTCOMP_BOTH) {
+							p.F("  %s = normalizeOr000(%s + viewDir);\n", name, name);
+						}
+					};
+					shadeLight(ls0, ls0Str, "shadeL0");
+					shadeLight(ls1, ls1Str, "shadeL1");
+					WRITE(p, "  %sv_texcoord = vec3(u_uvscaleoffset.xy * vec2(1.0 + dot(shadeL0, worldnormal), 1.0 + dot(shadeL1, worldnormal)) * 0.5, 1.0);\n", compat.vsOutPrefix);
 				}
 				break;
 

GPU/GLES/ShaderManagerGLES.cpp

@@ -46,6 +46,7 @@
 #include "GPU/GPUState.h"
 #include "GPU/ge_constants.h"
 #include "GPU/Common/ShaderUniforms.h"
+#include "GPU/Common/TransformCommon.h"
 #include "GPU/GLES/ShaderManagerGLES.h"
 #include "GPU/GLES/DrawEngineGLES.h"
 
@@ -549,7 +550,7 @@ void LinkedShader::UpdateUniforms(const ShaderID &vsid, const ShaderLanguageDesc
 		SetColorUniform3(render_, &u_matemissive, gstate.materialemissive);
 	}
 	if (dirty & DIRTY_MATSPECULAR) {
-		SetColorUniform3ExtraFloat(render_, &u_matspecular, gstate.materialspecular, getFloat24(gstate.materialspecularcoef));
+		SetColorUniform3ExtraFloat(render_, &u_matspecular, gstate.materialspecular, PSPSpecularCoef(getFloat24(gstate.materialspecularcoef)));
 	}
 
 	for (int i = 0; i < 4; i++) {
@@ -855,7 +856,7 @@ enum class CacheDetectFlags {
 };
 
 #define CACHE_HEADER_MAGIC 0x83277592
-#define CACHE_VERSION 43
+#define CACHE_VERSION 44
 
 struct CacheHeader {
 	uint32_t magic;

GPU/Software/Lighting.cpp

@@ -21,6 +21,7 @@
 #include "Common/CPUDetect.h"
 #include "Common/Math/SIMDHeaders.h"
 #include "GPU/GPUState.h"
+#include "GPU/Common/TransformCommon.h"
 #include "GPU/Software/Lighting.h"
 
 #if PPSSPP_ARCH(SSE2)
@@ -49,7 +50,7 @@ static inline float pspLightPow(float v, float e) {
 		return 1.0f;
 	}
 	if (v > 0.0f) {
-		return pow(v, e);
+		return PSPLightPow(v, e);
 	}
 	// Negative stays negative, so let's just return the original.
 	return v;
@@ -179,7 +180,7 @@ void ComputeState(State *state, bool hasColor0) {
 	}
 
 	if (anyDiffuse || anySpecular) {
-		state->specularExp = gstate.getMaterialSpecularCoef();
+		state->specularExp = PSPSpecularCoef(gstate.getMaterialSpecularCoef());
 		if (state->specularExp <= 0.0f)
 			state->specularExp = 0.0f;
 		else if (std::isnan(state->specularExp))
@@ -193,20 +194,11 @@ void ComputeState(State *state, bool hasColor0) {
 	state->usesWorldNormal = gstate.getUVGenMode() == GE_TEXMAP_ENVIRONMENT_MAP || anyDiffuse || anySpecular;
 }
 
-static inline float GenerateLightCoord(VertexData &vertex, const WorldCoords &worldnormal, int light) {
-	// TODO: Should specular lighting should affect this, too?  Doesn't in GLES.
-	Vec3<float> L = GetLightVec(gstate.lpos, light);
-	// In other words, L.Length2() == 0.0f means Dot({0, 0, 1}, worldnormal).
-	float diffuse_factor = Dot(L.NormalizedOr001(cpu_info.bSSE4_1), worldnormal);
-
-	return (diffuse_factor + 1.0f) / 2.0f;
-}
-
-void GenerateLightST(VertexData &vertex, const WorldCoords &worldnormal) {
+void GenerateLightST(VertexData &vertex, const WorldCoords &worldpos, const WorldCoords &worldnormal, const Vec3f &viewDir) {
 	// Always calculate texture coords from lighting results if environment mapping is active
 	// This should be done even if lighting is disabled altogether.
-	vertex.texturecoords.s() = GenerateLightCoord(vertex, worldnormal, gstate.getUVLS0());
-	vertex.texturecoords.t() = GenerateLightCoord(vertex, worldnormal, gstate.getUVLS1());
+	vertex.texturecoords.s() = PSPShadeMapCoord(gstate.getUVLS0(), worldpos, worldnormal, viewDir);
+	vertex.texturecoords.t() = PSPShadeMapCoord(gstate.getUVLS1(), worldpos, worldnormal, viewDir);
 }
 
 #if defined(_M_SSE)
@@ -368,7 +360,7 @@ static void ProcessSIMD(VertexData &vertex, const WorldCoords &worldpos, const W
 		}
 
 		if (lstate.specular && diffuse_factor >= 0.0f) {
-			Vec3<float> H = L + Vec3<float>(0.f, 0.f, 1.f);
+			Vec3<float> H = L + state.viewDir;
 
 			float specular_factor = Dot33(H.NormalizedOr001(useSSE4), worldnormal);
 			specular_factor = pspLightPow(specular_factor, state.specularExp);

GPU/Software/Lighting.h

@@ -53,6 +53,7 @@ struct State {
 
 	Vec4<int> baseAmbientColorFactor;
 	float specularExp;
+	Vec3f viewDir;
 
 	struct {
 		bool colorForAmbient : 1;
@@ -67,7 +68,7 @@ struct State {
 
 void ComputeState(State *state, bool hasColor0);
 
-void GenerateLightST(VertexData &vertex, const WorldCoords &worldnormal);
+void GenerateLightST(VertexData &vertex, const WorldCoords &worldpos, const WorldCoords &worldnormal, const Vec3f &viewDir);
 void Process(VertexData &vertex, const WorldCoords &worldpos, const WorldCoords &worldnormal, const State &state);
 
 }

GPU/Software/Rasterizer.cpp

@@ -1159,7 +1159,7 @@ void DrawTriangleSlice(
 						int32x4_t sec = vsetq_lane_s32(0, sec_color[i].ivec, 3);
 						prim_color[i].ivec = vaddq_s32(prim_color[i].ivec, sec);
 #else
-						prim_color[i] = Vec4<int>(sec_color[i], 0);
+						prim_color[i] += Vec4<int>(sec_color[i], 0);
 #endif
 					}
 				}

GPU/Software/SoftGpu.cpp

@@ -111,7 +111,7 @@ const SoftwareCommandTableEntry softgpuCommandTable[] = {
 	{ GE_CMD_FOGENABLE, 0, SoftDirty::PIXEL_BASIC | SoftDirty::PIXEL_CACHED | SoftDirty::TRANSFORM_BASIC | SoftDirty::TRANSFORM_FOG | SoftDirty::TRANSFORM_MATRIX },
 	{ GE_CMD_TEXMODE, 0, SoftDirty::SAMPLER_BASIC | SoftDirty::SAMPLER_TEXLIST | SoftDirty::RAST_TEX },
 	// Currently this doesn't affect any state, but maybe it should.
-	{ GE_CMD_TEXSHADELS },
+	{ GE_CMD_TEXSHADELS, 0, SoftDirty::TRANSFORM_BASIC },
 	{ GE_CMD_SHADEMODE, 0, SoftDirty::RAST_BASIC },
 	{ GE_CMD_TEXFUNC, 0, SoftDirty::SAMPLER_BASIC },
 	{ GE_CMD_COLORTEST, 0, SoftDirty::PIXEL_BASIC | SoftDirty::PIXEL_CACHED },

GPU/Software/TransformUnit.cpp

@@ -28,6 +28,7 @@
 #include "GPU/Common/DrawEngineCommon.h"
 #include "GPU/Common/VertexDecoderCommon.h"
 #include "GPU/Common/SoftwareTransformCommon.h"
+#include "GPU/Common/TransformCommon.h"
 #include "GPU/Common/VertexReader.h"
 #include "GPU/GPUStateSIMDUtil.h"
 #include "Common/Math/SIMDHeaders.h"
@@ -259,6 +260,13 @@ void ComputeTransformState(TransformState *state, const VertexReader &vreader) {
 		} else {
 			state->lightingState.usesWorldNormal = state->uvGenMode == GE_TEXMAP_ENVIRONMENT_MAP;
 		}
+		if (state->uvGenMode == GE_TEXMAP_ENVIRONMENT_MAP) {
+			// Shade mapping uses the light vector as lighting sees it, which depends on position for other lights.
+			if (!gstate.isDirectionalLight(gstate.getUVLS0()) || !gstate.isDirectionalLight(gstate.getUVLS1())) {
+				canSkipWorldPos = false;
+			}
+		}
+		state->lightingState.viewDir = PSPViewDirection(gstate.viewMatrix);
 
 		float world[16];
 		float view[16];
@@ -446,7 +454,7 @@ ClipVertexData TransformUnit::ReadVertex(const VertexReader &vreader, const Tran
 			Vec3<float> stq = Vec3ByMatrix43(source, gstate.tgenMatrix);
 			vertex.v.texturecoords = Vec3Packedf(stq.x, stq.y, stq.z);
 		} else if (state.uvGenMode == GE_TEXMAP_ENVIRONMENT_MAP) {
-			Lighting::GenerateLightST(vertex.v, worldnormal);
+			Lighting::GenerateLightST(vertex.v, worldpos, worldnormal, state.lightingState.viewDir);
 		}
 
 		PROFILE_THIS_SCOPE("light");

GPU/Vulkan/ShaderManagerVulkan.cpp

@@ -369,7 +369,7 @@ enum class VulkanCacheDetectFlags {
 };
 
 #define CACHE_HEADER_MAGIC 0xff51f420 
-#define CACHE_VERSION 60
+#define CACHE_VERSION 61
 
 struct VulkanCacheHeader {
 	uint32_t magic;

test.py

@@ -230,6 +230,8 @@ tests_good = [
   "gpu/ge/intrsuspend",
   "gpu/ge/queue",
   "gpu/ge/queue2",
+  "gpu/lighting/shademap",
+  "gpu/lighting/specular",
   "gpu/primitives/indices",
   "gpu/primitives/invalidprim",
   "gpu/primitives/points",
@@ -510,17 +512,9 @@ known_failures = {
     # The ISA returns the canonical NaN (0x7fc00000) from every operation, never the operand's
     # NaN, so a negative or signaling NaN input loses its sign and payload. Everything else passes.
     "cpu/fpu/roundmode",
-    # The software renderer's output differs from the reference by the same amount on both of
-    # these architectures, despite them using completely different SIMD paths. Unexplained.
-    "gpu/clipping/homogeneous",
-    "gpu/commands/cull",
-    "gpu/primitives/triangles",
   ],
   "loongarch64": [
     "cpu/fpu/fpu",
-    "gpu/clipping/homogeneous",
-    "gpu/commands/cull",
-    "gpu/primitives/triangles",
   ],
 }