Fixed really stupid mistake.

- Whoops, accidentally took x to the 4th power in MATH_Fast_Invert().

dreamcast/Makefile

@@ -79,6 +79,9 @@ OBJS_TEXCONV += \
 	../vendor/librw/src/d3d-x/d3d.texconv.o \
 	../vendor/librw/src/d3d-x/d3d8.texconv.o \
 	../vendor/librw/src/d3d-x/d3d8render.texconv.o \
+	../vendor/librw/src/bmp.texconv.o \
+	../vendor/librw/src/png.texconv.o \
+	../vendor/librw/src/lodepng/lodepng.texconv.o	
 
 # Add compilation units to this list to explicity compile them with
 # -O3 optimizations, while the rest get the default (-Os) treatment
@@ -316,6 +319,22 @@ TXD_OPTS_NEWS = 512 512
 TXD_OPTS_SPLASH1 = 512 512
 TXD_OPTS_SPLASH2 = 512 512
 TXD_OPTS_SPLASH3 = 512 512
+TXD_OPTS_frontend = 512 512 \
+	--delete-tex "fe_arrows4" \
+	--delete-tex "fe_arrows2" \
+	--delete-tex "fe_arrows1" \
+	--delete-tex "fe_controllersh" \
+	--delete-tex "fe_controller" \
+	\
+	--include-tex assets/dc_ps2d.png dc_ps2d \
+	--include-tex assets/dc_ps2f.png dc_ps2f \
+	--include-tex assets/dc_xboxd.png dc_xboxd \
+	--include-tex assets/dc_xboxf.png dc_xboxf \
+	--include-tex assets/ps4_d.png ps4_d \
+	--include-tex assets/ps4_f.png ps4_f \
+	--include-tex assets/xbox_d.png xbox_d \
+	--include-tex assets/xbox_f.png xbox_f
+
 DEFAULT_RES = 512
 
 PVR_ENCODER ?= PVRTEX

dreamcast/texconv.cpp

@@ -3,6 +3,7 @@
 #include <cstring>
 
 #include <set>
+#include <vector>
 
 #include <iostream>
 using namespace std;
@@ -238,6 +239,40 @@ PluginAttach(void)
 
 const char* currentFile;
 
+namespace rw {
+	Image* readBMP(const char *filename);
+	Image* readPNG(const char *filename);
+}
+
+void InsertImage(RwTexDictionary* texDict, const char* file, const char* texName) {
+	RwTexture *tex;
+	RwRaster *raster;
+	RwInt32 width, height, depth, format;
+	RwImage *image = rw::readBMP(file);
+	if (!image) {
+		image = rw::readPNG(file);
+	}
+
+	assert(image);
+
+	RwImageFindRasterFormat(image, rwRASTERTYPETEXTURE, &width, &height, &depth, &format);
+	raster = RwRasterCreate(width, height, depth, format);
+	RwRasterSetFromImage(raster, image);
+
+	tex = RwTextureCreate(raster);
+	RwTextureSetName(tex, texName);
+
+	RwTextureSetFilterMode(tex, rwFILTERLINEAR);
+
+	RwTexDictionaryAddTexture(texDict, tex);
+
+	RwImageDestroy(image);
+}
+
+std::vector<std::pair<const char*, const char*>> ImagesToAdd;
+std::vector<const char*> ImagesToRemove;
+bool listTextures = false;
+
 int main(int argc, const char** argv) {
 	if (argc >= 5) {
 	    int width = atoi(argv[3]);
@@ -247,7 +282,7 @@ int main(int argc, const char** argv) {
         if(height >= 16 && height <= 1024)
 		    rw::dc::maxRasterHeight = height;
 	}
-    for (int i = 0; i < argc; i++) {
+    for (int i = 5; i < argc; i++) {
         if (argv[i] != nullptr) {
             // Downsample Parameter
             if (strcmp(argv[i], "-d") == 0 || strcmp(argv[i], "-D") == 0) {
@@ -271,6 +306,22 @@ int main(int argc, const char** argv) {
                     }
                 }
             }
+
+			if (strcmp(argv[i], "--include-tex") == 0) {
+				assert(i + 2 < argc);
+				ImagesToAdd.emplace_back(argv[i+1], argv[i+2]);
+				i += 2;
+			}
+
+			if (strcmp(argv[i], "--delete-tex") == 0) {
+				assert(i + 1 < argc);
+				ImagesToRemove.emplace_back(argv[i+1]);
+				i += 1;
+			}
+
+			if (strcmp(argv[i], "--list-tex") == 0) {
+				listTextures = true;
+			}
         }
     }
 
@@ -298,6 +349,31 @@ int main(int argc, const char** argv) {
 
 		RwStreamClose(stream, nil);
 
+		if (listTextures) {
+			fprintf(stdout, "Incoming textures:\n");
+			FORLIST(lnk, texDict->textures) {
+				auto tex = rw::Texture::fromDict(lnk);
+				fprintf(stdout, "texture: '%s'\n", tex->name);
+			}
+		}
+
+		for (auto&& removedTextureName: ImagesToRemove) {
+			auto removedTexture = texDict->find(removedTextureName);
+			assert(removedTexture);
+			texDict->remove(removedTexture);
+		}
+		for (auto&& extraTexture: ImagesToAdd) {
+			InsertImage(texDict, extraTexture.first, extraTexture.second);
+		}
+
+		if (listTextures) {
+			fprintf(stdout, "Processing textures:\n");
+			FORLIST(lnk, texDict->textures) {
+				auto tex = rw::Texture::fromDict(lnk);
+				fprintf(stdout, "texture: '%s'\n", tex->name);
+			}
+		}
+
 		auto streamOut = RwStreamOpen(rwSTREAMFILENAME, rwSTREAMWRITE, argv[2]);
 		assert(streamOut && "failed to open output");
 

src/animation/AnimBlendNode.cpp

@@ -140,7 +140,7 @@ CAnimBlendNode::GetCurrentTranslation(CVector &trans, float weight)
 	float blend = association->GetBlendAmount(weight);
 	if(blend > 0.0f){
 		auto kfAdt = sequence->GetDeltaTime(frameA);
-		float t = (kfAdt - remainingTime)/kfAdt;
+		float t = kfAdt == 0.0f ? 0.0f : (kfAdt - remainingTime)/kfAdt;
 		if(sequence->type & CAnimBlendSequence::KF_TRANS){
 			auto kfAt = sequence->GetTranslation(frameA);
 			auto kfBt = sequence->GetTranslation(frameB);

src/audio/AudioLogic.cpp

@@ -3789,6 +3789,10 @@ cAudioManager::ProcessPedOneShots(cPedParams &params)
 						if (iSound > 60)
 							iSound = 21;
 					}
+					// In some cases the left and right channels have different loop points
+					// This looks like a data file issue where the left and right channels have different loop points
+					// This is a hot fix to always have the correct loop point for each channel individually
+					SET_LOOP_OFFSETS(m_sQueueSample.m_nSampleIndex)
 					AddSampleToRequestedQueue();
 				}
 			}

src/audio/sampman.h

@@ -174,6 +174,10 @@ public:
 	
 	void UpdateEffectsVolume(void);
 
+#ifdef DC_SH4
+	void UpdateChannelVolume(uint32 nChannel);
+#endif
+
 	void SetEffectsMasterVolume(uint8 nVolume);
 	void SetMusicMasterVolume  (uint8 nVolume);
 	void SetEffectsFadeVolume  (uint8 nVolume);

src/audio/sampman_dc.cpp

@@ -151,7 +151,6 @@ void aica_snd_sfx_freq_vol(int chn, int freq, int vol) {
 
 cSampleManager SampleManager;
 bool8 _bSampmanInitialised = FALSE;
-bool _dcAudioInitialized = false;
 
 uint32 BankStartOffset[MAX_SFX_BANKS];
 char SampleBankDescFilename[] = "sfx/sfx_all.dsc";
@@ -462,7 +461,7 @@ cSampleManager::Initialise(void)
 
 	assert(fdPedSfx >= 0);
 
-	_dcAudioInitialized = true;
+	_bSampmanInitialised = true;
 	return TRUE;
 }
 
@@ -485,7 +484,14 @@ char cSampleManager::GetCDAudioDriveLetter(void)
 void
 cSampleManager::UpdateEffectsVolume(void)
 {
-	// TODO
+	if(_bSampmanInitialised) {
+		std::lock_guard<std::mutex> lk(channel_mtx);
+		for (int i = 0; i < MAXCHANNELS+MAX2DCHANNELS; i++) {
+			if (channels[i].ch != -1) {
+				UpdateChannelVolume(i);
+			}
+		}
+	}
 }
 
 
@@ -851,9 +857,10 @@ cSampleManager::InitialiseChannel(uint32 nChannel, uint32 nSfx, uint8 nBank)
 	return TRUE;
 }
 
-void updateVol(uint32 nChannel) {
+void cSampleManager::UpdateChannelVolume(uint32 nChannel) {
 
 	auto newVol = channels[nChannel].emittingVol * channels[nChannel].attenuationVol / 255;
+	newVol = m_nEffectsFadeVolume * newVol * m_nEffectsVolume >> 14;
 	// newVol = 255;
 	// printf("updateVol(nChannel: %d) vol: %d, newVol: %d\n", nChannel, channels[nChannel].vol, newVol);
 	if (channels[nChannel].vol != newVol) {
@@ -876,7 +883,7 @@ cSampleManager::SetChannelVolume(uint32 nChannel, uint32 nVolume)
 	channels[nChannel].emittingVol = linearlize_volume(nVolume);// nVolume * 255 / MAX_VOLUME;
 	channels[nChannel].attenuationVol = 255;
 
-	updateVol(nChannel);
+	UpdateChannelVolume(nChannel);
 	verbosef("SetChannelVolume(nChannel: %d) vol: %d\n", nChannel, nVolume);
 }
 
@@ -1132,7 +1139,8 @@ cSampleManager::StartStreamedFile(uint8 nFile, uint32 nPos, uint8 nStream)
 	debugf("StartStreamedFile(%d, %d, %d)\n", nFile, nPos, nStream);
 	uint32_t seek_aligned = 0;
 	if (nPos) {
-		uint32 seek_bytes = nPos * streams[nStream].rate / (streams[nStream].stereo ? 1000: 2000);
+		uint64_t seek_bytes = (uint64_t)nPos * streams[nStream].rate / (streams[nStream].stereo ? 1000: 2000);
+		assert(seek_bytes <= INT32_MAX);
 		seek_aligned = seek_bytes & ~(streams[nStream].stereo ? (STREAM_STAGING_READ_SIZE_STEREO-1) : (STREAM_STAGING_READ_SIZE_MONO-1));
 	}
 	PreloadStreamedFile(nFile, nStream, seek_aligned);
@@ -1164,7 +1172,9 @@ cSampleManager::GetStreamedFilePosition(uint8 nStream)
 	ASSERT( nStream < MAX_STREAMS );
 	int32 rv = 0;
 
-	return streams[nStream].played_samples * 1000 / streams[nStream].rate;
+	int64_t rv64 = (int64_t)streams[nStream].played_samples * 1000 / streams[nStream].rate;
+	assert(rv64 <= INT32_MAX);
+	rv = (int32)rv64;
 	// if(streams[nStream].fd >= 0) {
 	// 	rv = fs_tell(streams[nStream].fd);
 	// }
@@ -1179,6 +1189,7 @@ cSampleManager::SetStreamedVolumeAndPan(uint8 nVolume, uint8 nPan, uint8 nEffect
 	if (nVolume > MAX_VOLUME)
 		nVolume = MAX_VOLUME;
 	nVolume = linearlize_volume(nVolume); //nVolume * 255 / MAX_VOLUME;
+	nVolume = m_nMusicFadeVolume * nVolume * m_nMusicVolume >> 14;
 	if (streams[nStream].vol != nVolume || streams[nStream].nPan != nPan) {
 		streams[nStream].vol = nVolume;
 		streams[nStream].nPan = nPan;
@@ -1207,7 +1218,10 @@ cSampleManager::GetStreamedFileLength(uint8 nFile)
 	WavHeader hdr;
 	assert(fs_read(fd, &hdr, sizeof(hdr)) == sizeof(hdr));
 
-	rv = hdr.dataSize * 2000 / hdr.numOfChan / hdr.samplesPerSec;
+	uint64_t rv64 = (uint64_t)hdr.dataSize * 2000 / hdr.numOfChan / hdr.samplesPerSec;
+
+	assert(rv64 <= INT32_MAX);
+	rv = (int32)rv64;
 
 	fs_close(fd);
 
@@ -1287,9 +1301,17 @@ void  cSampleManager::SetChannelEmittingVolume(uint32 nChannel, uint32 nVolume)
 	if (nVolume > MAX_VOLUME)
 		nVolume = MAX_VOLUME;
 
+	// reduce channel volume when JB.MP3 or S4_BDBD.MP3 playing
+	if (   MusicManager.GetMusicMode()    == MUSICMODE_CUTSCENE
+		&& MusicManager.GetNextTrack() != STREAMED_SOUND_NEWS_INTRO
+		&& MusicManager.GetNextTrack() != STREAMED_SOUND_CUTSCENE_SAL4_BDBD )
+	{
+		nVolume /= 4;
+	}
+
 	std::lock_guard<std::mutex> lk(channel_mtx);
 	channels[nChannel].emittingVol = linearlize_volume(nVolume); // nVolume * 255 / MAX_VOLUME;
-	updateVol(nChannel);
+	UpdateChannelVolume(nChannel);
 }
 
 float calculatePan(float x, float z) {
@@ -1334,7 +1356,7 @@ void  cSampleManager::SetChannel3DPosition    (uint32 nChannel, float fX, float
 		channels[nChannel].fY = fY;
 		channels[nChannel].fZ = fZ;
 		channels[nChannel].attenuationVol = calculateAttenuation(channels[nChannel].fX, channels[nChannel].fY, channels[nChannel].fZ, channels[nChannel].distMin, channels[nChannel].distMax) * 255;
-		updateVol(nChannel);
+		UpdateChannelVolume(nChannel);
 	}
 
 	SetChannelPan(nChannel, calculatePan(-fX, fZ) * 63 + 64);
@@ -1347,7 +1369,7 @@ void  cSampleManager::SetChannel3DDistances   (uint32 nChannel, float fMax, floa
 	channels[nChannel].distMin = fMin;
 	channels[nChannel].distMax = fMax;
 	channels[nChannel].attenuationVol = calculateAttenuation(channels[nChannel].fX, channels[nChannel].fY, channels[nChannel].fZ, channels[nChannel].distMin, channels[nChannel].distMax) * 255;
-	updateVol(nChannel);
+	UpdateChannelVolume(nChannel);
 }
 #endif
 

src/control/Script4.cpp

@@ -1721,8 +1721,8 @@ int8 CRunningScript::ProcessCommands900To999(int32 command)
 			return 0;
 		int attempts;
 		int model = -1;
-		int index = CGeneral::GetRandomNumberInRange(0, 50);
-		for (attempts = 0; attempts < 50; attempts++) {
+		int index = CGeneral::GetRandomNumberInRange(0, MAXVEHICLESLOADED);
+		for (attempts = 0; attempts < MAXVEHICLESLOADED; attempts++) {
 			if (model != -1)
 				break;
 			model = CStreaming::ms_vehiclesLoaded[index];

src/core/Pad.cpp

@@ -2785,22 +2785,22 @@ bool CPad::WeaponJustDown(void)
 		case 0:	//Xbox Mode
 			if (CPad::GetPad(0)->IsDualAnalog)
 			{
-				if (NewState.RightTrigger > 128)
+				if (NewState.RightTrigger > 128 && !(OldState.RightTrigger > 128))
 					return true;			
 			}
 			else
 			{
-				if (NewState.RightTrigger > 128)
+				if (NewState.RightTrigger > 128 && !(OldState.RightTrigger > 128))
 					return true;			
 			}
 		case 1:	//PS2 Mode
 			if (CPad::GetPad(0)->IsDualAnalog)
 			{
-				return	NewState.B;
+				return	NewState.B && !OldState.B;
 			}
 			else
 			{
-				return	NewState.B;
+				return	NewState.B && !OldState.B;
 			}
 	}
 

src/entities/Entity.cpp

@@ -83,6 +83,11 @@ CEntity::~CEntity(void)
 	ResolveReferences();
 }
 
+bool
+CEntity::IsFence(void) {
+	return IsObject() && ::IsFence(static_cast<CObject *>(this)->GetModelIndex());
+}
+
 void
 CEntity::SetModelIndex(uint32 id)
 {

src/entities/Entity.h

@@ -131,6 +131,7 @@ public:
 	bool IsPed(void) { return m_type == ENTITY_TYPE_PED; }
 	bool IsObject(void) { return m_type == ENTITY_TYPE_OBJECT; }
 	bool IsDummy(void) { return m_type == ENTITY_TYPE_DUMMY; }
+	bool IsFence(void);
 
 	RpAtomic *GetAtomic(void) {
 		assert(RwObjectGetType(m_rwObject) == rpATOMIC);

src/entities/Physical.cpp

@@ -59,7 +59,7 @@ CPhysical::CPhysical(void)
 	bInfiniteMass = false;
 	bIsInWater = false;
 	bHitByTrain = false;
-	bSkipLineCol = false;
+	bSkipLineCol = IsFence();
 
 	m_fDistanceTravelled = 0.0f;
 	m_treadable[PATH_CAR] = nil;
@@ -1461,6 +1461,10 @@ CPhysical::ProcessCollisionSectorList(CPtrList *lists)
 			skipCollision = false;
 			altcollision = false;
 
+			if(A->IsFence() && B->IsFence()) {
+				skipCollision = true;
+				A->bSkipLineCol = true;
+			}
 			if(B->IsBuilding())
 				skipCollision = false;
 			else if(IsStreetLight(A->GetModelIndex()) &&
@@ -1922,12 +1926,12 @@ CPhysical::ProcessCollision(void)
 				n = NUMSTEPS(0.09f);
 				step = savedTimeStep / n;
 			}
-		}else if(responsecase == COLLRESPONSE_SMALLBOX || responsecase == COLLRESPONSE_FENCEPART){
+		}else if(responsecase == COLLRESPONSE_SMALLBOX){
 			if(distSq >= sq(0.15f)){
 				n = NUMSTEPS(0.15f);
 				step = savedTimeStep / n;
 			}
-		}else{
+		}else if(responsecase != COLLRESPONSE_FENCEPART){
 			if(distSq >= sq(0.3f)){
 				n = NUMSTEPS(0.3f);
 				step = savedTimeStep / n;

src/vmu/vmu.cpp

@@ -10,7 +10,7 @@
 #       include <dc/sound/sound.h>
 #   endif
 
-extern bool _dcAudioInitialized;
+extern bool _bSampmanInitialised;
 
 // ====== STATIC METHODS =====
 
@@ -86,7 +86,7 @@ void VmuProfiler::run() {
 
 #ifdef DC_SH4
         if(auto *dev = maple_enum_type(0, MAPLE_FUNC_MEMCARD); 
-            dev && _dcAudioInitialized && updated_) 
+            dev && _bSampmanInitialised && updated_) 
         {
             pvr_stats_t pvrStats;   pvr_get_stats(&pvrStats);
             uint32_t    sramStats = snd_mem_available();

src/weapons/BulletInfo.cpp

@@ -126,6 +126,28 @@ void CBulletInfo::Update(void)
 						}
 						pPed->InflictDamage(pBullet->m_pSource, pBullet->m_eWeaponType, pBullet->m_nDamage, (ePedPieceTypes)point.pieceB, pPed->GetLocalDirection(pPed->GetPosition() - point.point));
 						CEventList::RegisterEvent(pPed->m_nPedType == PEDTYPE_COP ? EVENT_SHOOT_COP : EVENT_SHOOT_PED, EVENT_ENTITY_PED, pPed, (CPed*)pBullet->m_pSource, 1000);
+
+						if (CGame::nastyGame) {
+							CVector vecParticleDirection = (point.point - pPed->GetPosition()) * 0.01f;
+							vecParticleDirection.z = 0.01f;
+							if (pPed->GetIsOnScreen()) {
+								for (int j = 0; j < NUM_PED_BLOOD_PARTICLES; j++)
+									CParticle::AddParticle(PARTICLE_BLOOD_SMALL, point.point + BLOOD_PARTICLE_OFFSET, vecParticleDirection);
+							}
+							if (pPed->GetPedState() == PED_DEAD) {
+								CAnimBlendAssociation* pAnim;
+								if (RpAnimBlendClumpGetFirstAssociation(pPed->GetClump(), ASSOC_FRONTAL))
+									pAnim = CAnimManager::BlendAnimation(pPed->GetClump(), ASSOCGRP_STD, ANIM_STD_HIT_FLOOR_FRONT, 8.0f);
+								else
+									pAnim = CAnimManager::BlendAnimation(pPed->GetClump(), ASSOCGRP_STD, ANIM_STD_HIT_FLOOR, 8.0f);
+								if (pAnim) {
+									pAnim->SetCurrentTime(0.0f);
+									pAnim->flags |= ASSOC_RUNNING;
+									pAnim->flags &= ~ASSOC_FADEOUTWHENDONE;
+								}
+							}
+						}
+
 						pBullet->m_bInUse = false;
 #ifdef SQUEEZE_PERFORMANCE
 						bulletInfoInUse--;
@@ -136,31 +158,6 @@ void CBulletInfo::Update(void)
 						bAddSound = false;
 					}
 				}
-				if (CGame::nastyGame) {
-					CVector vecParticleDirection = (point.point - pPed->GetPosition()) * 0.01f;
-					vecParticleDirection.z = 0.01f;
-					if (pPed->GetIsOnScreen()) {
-						for (int j = 0; j < NUM_PED_BLOOD_PARTICLES; j++)
-							CParticle::AddParticle(PARTICLE_BLOOD_SMALL, point.point + BLOOD_PARTICLE_OFFSET, vecParticleDirection);
-					}
-					if (pPed->GetPedState() == PED_DEAD) {
-						CAnimBlendAssociation* pAnim;
-						if (RpAnimBlendClumpGetFirstAssociation(pPed->GetClump(), ASSOC_FRONTAL))
-							pAnim = CAnimManager::BlendAnimation(pPed->GetClump(), ASSOCGRP_STD, ANIM_STD_HIT_FLOOR_FRONT, 8.0f);
-						else
-							pAnim = CAnimManager::BlendAnimation(pPed->GetClump(), ASSOCGRP_STD, ANIM_STD_HIT_FLOOR, 8.0f);
-						if (pAnim) {
-							pAnim->SetCurrentTime(0.0f);
-							pAnim->flags |= ASSOC_RUNNING;
-							pAnim->flags &= ~ASSOC_FADEOUTWHENDONE;
-						}
-					}
-					pBullet->m_bInUse = false;
-#ifdef SQUEEZE_PERFORMANCE
-					bulletInfoInUse--;
-#endif
-					vecNewPos = point.point;
-				}
 			}
 			else if (pHitEntity->IsVehicle()) {
 				CVehicle* pVehicle = (CVehicle*)pHitEntity;

vendor/librw/src/dc/rwdc.cpp

@@ -164,8 +164,18 @@ struct alignas(32) pvr_vertex32_ut {
 static_assert(sizeof(pvr_vertex16_t) == 32, "pvr_vertex16_t size mismatch");
 static_assert(alignof(pvr_vertex16_t) == 32, "pvr_vertex16_t alignof mismatch");
 
+#define MATH_Very_Fast_Invert(x) ({ 1.0f / sqrtf((x) * (x)); })
 
-#define MATH_Fast_Invert(x) ({ (((x) < 0.0f)? -1.0f : 1.0f) * frsqrt((x) * (x)); }) 
+__always_inline float MATH_Fast_Invert(float x) {
+	bool neg = 0;
+
+	if(x < 0.0f)
+	    neg = true;
+
+	x = MATH_Very_Fast_Invert(x); // 1.0f / sqrt(x^2)
+
+	return (neg)? -x : x;
+}
 
 #define logf(...) // printf(__VA_ARGS__)
 
@@ -221,19 +231,21 @@ static pvr_dr_state_t drState;
 				frchg
 
 				fmov        @%[mtx]+, dr0
-
 				fldi0 		fr12
-				fldi0 		fr13
 
 				fmov        @%[mtx]+, dr2
-				fmov        @%[mtx]+, dr4
-				fmov        @%[mtx]+, dr6
-				fmov        @%[mtx]+, dr8
-				fmov        @%[mtx]+, dr10
+				fldi0 	    fr13
 
+				fmov        @%[mtx]+, dr4
 				fldi0	    fr3
+
+				fmov        @%[mtx]+, dr6
 				fldi0	    fr7
+
+				fmov        @%[mtx]+, dr8
 				fldi0	    fr11
+
+				fmov        @%[mtx]+, dr10
 				fmov        dr12, dr14
 
 				fschg
@@ -305,6 +317,8 @@ void rw_mat_load_4x4(rw::Matrix* mtx) {
 #define mat_identity(a)
 #define pvr_fog_table_color(a,r,g,b)
 #define pvr_fog_table_linear(s,e)
+#define pvr_fog_table_exp(d)
+#define pvr_fog_table_custom(d)
 #endif
 
 #define mat_trans_single3_nomod(x_, y_, z_, x2, y2, z2) do { \
@@ -416,6 +430,66 @@ void leave_oix_() {
 	#endif
 }
 
+void enter_ocr_() {
+	#if defined(DC_SH4)
+	auto mask = irq_disable();
+	dcache_purge_all();
+	volatile uint32_t * CCN_CCR = (uint32_t *)0xFF00001C;
+	*CCN_CCR |= (1 << 5); // enable OCR (ORA)
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+
+	irq_restore(mask);
+	#endif
+}
+
+void leave_ocr_() {
+	#if defined(DC_SH4)
+	auto mask = irq_disable();
+	dcache_inval_range(0x92000000, 8192);
+	dcache_purge_all();
+	volatile uint32_t * CCN_CCR = (uint32_t *)0xFF00001C;
+	*CCN_CCR &= ~( 1 << 5); // disable OCR (ORA)
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	__asm__ __volatile__ ("nop");
+	irq_restore(mask);
+	#endif
+}
+
 #if defined(DC_SH4)
 #define FLUSH_TA_DATA(src) do { __asm__ __volatile__("ocbwb @%0" : : "r" (src) : "memory"); } while(0)
 #else
@@ -604,7 +678,7 @@ struct atomic_context_t {
 __always_inline void DCE_RenderSubmitVertex(const pvr_vertex_t *v, uint32_t flags) {
     auto *sq  = reinterpret_cast<uint32_t *>(pvr_dr_target(drState));
     auto *src = reinterpret_cast<const uint32_t *>(v);
-    float sz  = MATH_Fast_Invert(v->z);
+    float sz  = MATH_Very_Fast_Invert(v->z);
     float sx  = v->x * sz;
     float sy  = v->y * sz;
     
@@ -631,7 +705,7 @@ __always_inline void DCE_RenderSubmitVertexIM3D(float x, float y, float w,
 {
     auto *sq   = reinterpret_cast<uint32_t *>(pvr_dr_target(drState));
     auto *uv32 = reinterpret_cast<const uint32_t *>(uv);
-    float sz   = MATH_Fast_Invert(w);
+    float sz   = MATH_Very_Fast_Invert(w);
     float sx   = x * sz;
     float sy   = y * sz;
 
@@ -663,7 +737,10 @@ void malloc_stats() { }
 #define UNIMPL_LOGV(...)
 #endif
 
+Camera* rwdcCam;
+
 void beginUpdate(Camera* cam)  {
+	rwdcCam = cam;
 	float view[16], proj[16];
 
 	// View Matrix
@@ -1098,6 +1175,7 @@ static bool doAlphaTest;
 
 static uint8_t fogFuncPvr = PVR_FOG_DISABLE;
 static uint32_t fogColor = 0;
+static float fogStart = 0.0f;
 static uint32 cullModePvr;
 
 static inline unsigned pvrCullMode(uint32_t cullMode) {
@@ -1198,8 +1276,9 @@ setRenderState(int32 state, void *pvalue)
 	 	fogFuncPvr = value ? PVR_FOG_TABLE : PVR_FOG_DISABLE;
 	 	break;
 	 case FOGCOLOR:
+#if !defined(DC_TEXCONV)		
         // Set fog color when state changes
-        if(fogColor != value) {
+        if(fogColor != value || fogStart != RwCameraGetFogDistance(rwdcCam)) {
             fogColor = value;
             RGBA c;
             c.red = value;
@@ -1207,9 +1286,22 @@ setRenderState(int32 state, void *pvalue)
             c.blue = value>>16;
             c.alpha = value>>24;
             pvr_fog_table_color(c.alpha / 255.0f, c.red / 255.0f, c.green  / 255.0f, c.blue  / 255.0f);
-            pvr_fog_table_linear(50.0f, 450.0f);
+
+			fogStart = RwCameraGetFogDistance(rwdcCam);
+			float fogEnd = RwCameraGetFarClipPlane(rwdcCam);
+			float fogIntensity[129];
+			uint8_t idx = 0;
+			float startIntensity = (-fogStart) / (fogEnd - fogStart);  //interpolate between start and end to get initial intensity
+			float step = (1.0f - startIntensity) / 129; // we have 129 entries, create a step such that start + (step*129) = 1.0
+			for(int i = 128; i >= 0; i--) {
+				fogIntensity[i] = startIntensity + (idx++ * step);
+			}
+			pvr_fog_far_depth(fogEnd);
+			pvr_fog_table_custom(fogIntensity);
         }
+#endif
 	 	break;
+
 	// case CULLMODE:
 	// 	if(rwStateCache.cullmode != value){
 	// 		rwStateCache.cullmode = value;
@@ -1409,7 +1501,7 @@ void im2DRenderPrimitive(PrimitiveType primType, void *vertices, int32_t numVert
 			pvrVert->flags = flags;
 			pvrVert->x 	   = gtaVert.x;
 			pvrVert->y	   = gtaVert.y;
-			pvrVert->z 	   = MATH_Fast_Invert(gtaVert.w); // this is perfect for almost every case...
+			pvrVert->z 	   = MATH_Very_Fast_Invert(gtaVert.w); // this is perfect for almost every case...
 			pvrVert->u 	   = gtaVert.u;
 			pvrVert->v 	   = gtaVert.v;
 			pvrVert->argb  = (gtaVert.a << 24) |
@@ -2342,6 +2434,36 @@ __attribute__ ((noinline)) void submitMeshlet(uint8_t* OCR, const int8_t* indexD
 	} while(--indexCount);
 }
 
+template<bool textured>
+__attribute__ ((noinline)) void submitMeshletFallback(uint8_t* OCR, const int8_t* indexData, uint32_t indexCount) {
+	struct SQBUF {
+		union {
+			uint32_t flags;
+			uint64_t data[4];
+			uint8_t data8[32];
+		};
+	};
+
+	SQBUF* sq = (SQBUF*)pvr_dr_target(drState);
+
+	static_assert(sizeof(SQBUF) == 32);
+ 
+	do {
+		auto idx = *indexData++;
+		auto flags = idx & 0x80 ? PVR_CMD_VERTEX_EOL : PVR_CMD_VERTEX;
+		auto lookup_idx = idx & 0x7F;
+
+		auto src = (SQBUF*)(OCR +  lookup_idx * 64);
+		src[0].flags = flags;
+		*sq = src[0];
+		pvr_dr_commit(sq);
+		if (textured) {
+			*sq = src[1];
+			pvr_dr_commit(sq);
+		}
+	} while(--indexCount);
+}
+
 
 #if defined(DC_SH4)
 template<>
@@ -2421,7 +2543,7 @@ __attribute__ ((noinline)) void submitMeshlet<true>(uint8_t* OCR, const int8_t*
 
 // 8 kb in total
 #if defined(DC_SH4)
-uint8_t* OCR_SPACE = (uint8_t*)0x92000000;
+uint8_t* OCR_SPACE;
 #else
 uint8_t OCR_SPACE[32 * 256] __attribute__((aligned(32)));
 #endif
@@ -2652,6 +2774,159 @@ __attribute__ ((noinline)) void clipAndsubmitMeshlet(uint8_t* vertexData, const
 			}
 		};
 	} while(indexCount != 0);
+
+	#undef FILLVERT
+	#undef SUBMIT_VTX
+	#undef SUBMIT_INTERPOLATE
+}
+
+
+template<bool textured>
+__attribute__ ((noinline)) void clipAndsubmitMeshletFallback(uint8_t* vertexData, const int8_t* indexData, uint32_t indexCount) {
+
+	struct SQBUF {
+		union {
+			uint32_t flags;
+			uint64_t data[4];
+			uint8_t data8[32];
+		};
+	};
+
+	static_assert(sizeof(SQBUF) == 32);
+
+	SQBUF* sq = (SQBUF*)pvr_dr_target(drState);
+
+	constexpr int8_t VERTEX = 0;
+	constexpr int8_t VERTEX_EOL = 0x80;
+	
+	#define FILLVERT(n) \
+		do { \
+			auto idx = *indexData++; \
+			auto local_idx = idx & 0x7f; \
+			eol_now = idx & 0x80; \
+			auto local_ptr = (vertexData + local_idx * 64); \
+			vpp[n] = local_ptr; \
+			auto v = (const pvr_vertex64_t*)local_ptr; \
+			vismask >>= 1; \
+			if((textured?v->tex_z:v->o_b) >= -v->o_g) vismask |= 0b100;	\
+			indexCount--; \
+			currentCount++; \
+		} while(0)
+
+	#define SUBMIT_VTX(vid, eolf) \
+		do { \
+			auto src = (SQBUF*) vpp[vid]; \
+			src[0].flags = eolf ? PVR_CMD_VERTEX_EOL : PVR_CMD_VERTEX; \
+			*sq = src[0]; \
+			pvr_dr_commit(sq); \
+			if (textured) { \
+				*sq = src[1]; \
+				pvr_dr_commit(sq); \
+			} \
+		} while(0)
+
+	#define SUBMIT_INTERPOLATE(vid1, vid2, eolf) \
+		do { \
+			sq = (SQBUF*)interpolateAndSubmit<textured>(sq, vpp[vid1], vpp[vid2], eolf ? PVR_CMD_VERTEX_EOL : PVR_CMD_VERTEX); \
+		} while(0)
+
+	uint32_t vismask = 0;
+
+	uint8_t* vpp[3];
+
+	int8_t eol = 0;
+	int8_t eol_now = 0;
+
+	do {
+		uint32_t currentCount = -1;
+
+		FILLVERT(0);
+		FILLVERT(1);
+		FILLVERT(2);
+
+		if (vismask & 1) {
+			SUBMIT_VTX(0, VERTEX);
+			if (vismask & 2) {
+				// both first verts visible
+				SUBMIT_VTX(1, VERTEX);
+			} else {
+				// 0 visible, 1 hidden
+				SUBMIT_INTERPOLATE(0, 1, VERTEX);
+			}
+		} else if (vismask & 2) {
+			// 0 hidden, 1 visible
+			SUBMIT_INTERPOLATE(1, 0, VERTEX);
+			SUBMIT_VTX(1, VERTEX);
+		}
+
+		eol = 0;
+		// each remaining vertex of the strip
+		while(!eol) {
+			// "ring buffery" indices
+			uint8_t vertZeroIdx = (currentCount - 2) % 3;
+			uint8_t vertOneIdx = (currentCount - 1) % 3;
+			uint8_t vertTwoIdx = currentCount % 3;
+			//dcache_pref_block(&vph[vertZeroIdx]); not sure where to put this honestly -jaxyn
+
+			eol = eol_now;
+
+			if (!vismask) {
+				if (!eol) {
+					// "ring buffery" filling
+					FILLVERT(vertZeroIdx);
+				}
+				continue;
+			}
+
+			if (vismask == 7) {
+				// all visible
+				SUBMIT_VTX(vertTwoIdx, eol);
+				if (!eol) {
+					// "ring buffery" filling
+					FILLVERT(vertZeroIdx);
+				}
+				continue;
+			}
+
+			switch (vismask) {
+				case 1: // 0 visible, 1 and 2 hidden
+					// pause strip
+					SUBMIT_INTERPOLATE(vertZeroIdx, vertTwoIdx, VERTEX_EOL);
+					break;
+				case 3: // 0 and 1 visible, 2 hidden
+					SUBMIT_INTERPOLATE(vertZeroIdx, vertTwoIdx, VERTEX);
+					SUBMIT_VTX(vertOneIdx, VERTEX);
+				case 2: // 0 hidden, 1 visible, 2 hidden
+					SUBMIT_INTERPOLATE(vertOneIdx, vertTwoIdx, eol);
+					break;
+				case 4: // 0 and 1 hidden, 2 visible
+					SUBMIT_INTERPOLATE(vertTwoIdx, vertZeroIdx, VERTEX);
+					if (currentCount & 0x01) { // flip directionality
+				case 5: // 0 visible, 1 hidden, 2 visible
+						SUBMIT_VTX(vertTwoIdx, VERTEX);
+					}
+					SUBMIT_INTERPOLATE(vertTwoIdx, vertOneIdx, VERTEX);
+					SUBMIT_VTX(vertTwoIdx, eol);
+					break;
+				case 6: // 0 hidden, 1 and 2 visible
+					SUBMIT_INTERPOLATE(vertTwoIdx, vertZeroIdx, VERTEX);
+					SUBMIT_VTX(vertOneIdx, VERTEX);
+					SUBMIT_VTX(vertTwoIdx, eol);
+					break;
+				default:
+					break;
+			}
+
+			if (!eol) {
+				// "ring buffery" filling
+				FILLVERT(vertZeroIdx);
+			}
+		};
+	} while(indexCount != 0);
+
+	#undef FILLVERT
+	#undef SUBMIT_VTX
+	#undef SUBMIT_INTERPOLATE
 }
 
 
@@ -2918,16 +3193,26 @@ static constexpr void (*tnlMeshletDiffuseColorSelector[8])(uint8_t* dstCol, cons
 	&tnlMeshletDiffuseColor<4, true>,
 };
 
-static constexpr void (*submitMeshletSelector[2])(uint8_t* OCR, const int8_t* indexData, uint32_t indexCount) = {
+static void (*submitMeshletSelector[2])(uint8_t* OCR, const int8_t* indexData, uint32_t indexCount) = {
 	&submitMeshlet<false>,
 	&submitMeshlet<true>,
 };
 
-static constexpr void (*clipAndsubmitMeshletSelector[2])(uint8_t* OCR, const int8_t* indexData, uint32_t indexCount) = {
+static void (*submitMeshletSelectorFallback[2])(uint8_t* OCR, const int8_t* indexData, uint32_t indexCount) = {
+	&submitMeshletFallback<false>,
+	&submitMeshletFallback<true>,
+};
+
+static void (*clipAndsubmitMeshletSelector[2])(uint8_t* OCR, const int8_t* indexData, uint32_t indexCount) = {
 	&clipAndsubmitMeshlet<false>,
 	&clipAndsubmitMeshlet<true>,
 };
 
+static void (*clipAndsubmitMeshletSelectorFallback[2])(uint8_t* OCR, const int8_t* indexData, uint32_t indexCount) = {
+	&clipAndsubmitMeshletFallback<false>,
+	&clipAndsubmitMeshletFallback<true>,
+};
+
 static constexpr void(*tnlMeshletSkinVerticesSelector[4])(uint8_t *OCR, uint8_t *OCR_normal, const uint8_t* vertex, const uint8_t* normals, const uint8_t* skinWeights, const uint8_t* skinIndexes, int vertexCount, int vertexSize, Matrix* skinMatrices) = {
 	&tnlMeshletSkinVertices<false, false>,
 	&tnlMeshletSkinVertices<true , false>,
@@ -3887,19 +4172,19 @@ imageFindRasterFormat(Image *img, int32 type,
 		;
 
     if(downsampleMode >= HALF) {
-        if(height / 2 >= 64) {
+        if(height / 2 >= 16) {
             height /= 2;
         }
-        if(width / 2 >= 64) {
+        if(width / 2 >= 16) {
             width /= 2;
         }
     }
 
     if(downsampleMode >= QUARTER) {
-        if(height / 2 >= 32) {
+        if(height / 2 >= 16) {
             height /= 2;
         }
-        if(width / 2 >= 32) {
+        if(width / 2 >= 16) {
             width /= 2;
         }
     }
@@ -4362,6 +4647,34 @@ ObjPipeline* makeDefaultPipeline(void)
 static void*
 driverOpen(void *o, int32, int32)
 {
+	#if defined(DC_SH4)
+	OCR_SPACE = (uint8_t*)0x92000000;
+
+	bool has_oix = true;
+	enter_oix();
+	*(volatile uint8_t*)OCR_SPACE = 1;
+	if (*(volatile uint8_t*)OCR_SPACE != 1) {
+		has_oix = false;
+	}
+	leave_oix();
+
+	if (!has_oix) {
+		dbglog(DBG_CRITICAL, "You appear to be using an emulator that does not support OIX. Attempting fallback to OCR\n");
+		OCR_SPACE = (uint8_t*)0x7c001000;
+		enter_oix = (void(*)())(((uintptr_t)&enter_ocr_) - 0x8c000000 + 0xAc000000);
+		leave_oix = (void(*)())(((uintptr_t)&leave_ocr_) - 0x8c000000 + 0xAc000000);
+
+		for (size_t i = 0; i < ARRAY_SIZE(submitMeshletSelector); i++) {
+			submitMeshletSelector[i] = submitMeshletSelectorFallback[i];
+		}
+
+		for (size_t i = 0; i < ARRAY_SIZE(clipAndsubmitMeshletSelector); i++) {
+			clipAndsubmitMeshletSelector[i] = clipAndsubmitMeshletSelectorFallback[i];
+		}
+	}
+	#endif
+	
+
     pvr_init(&pvr_params);
 
 	fake_tex = pvr_mem_malloc(sizeof(fake_tex_data));