GSdx: couple of fixes

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@1447 96395faa-99c1-11dd-bbfe-3dabce05a288

plugins/GSdx/GSCrc.cpp

@@ -32,6 +32,7 @@ CRC::Game CRC::m_games[] =
 	{0xDDB59F46, KyuuketsuKitanMoonties, JP, PointListPalette},
 	{0xC8EE2562, PiaCarroteYoukosoGPGakuenPrincess, JP, PointListPalette},
 	{0x6CF94A43, KazokuKeikakuKokoroNoKizuna, JP, PointListPalette},
+	{0xEDAF602D, DuelSaviorDestiny, JP, PointListPalette},
 	{0xa39517ab, FFX, EU, 0},
 	{0xa39517ae, FFX, FR, 0},
 	{0x941bb7d9, FFX, DE, 0},
@@ -133,12 +134,13 @@ CRC::Game CRC::m_games[] =
 	{0x6BF11378, Onimusha3, US, 0},	
 	{0xF442260C, MajokkoALaMode2, JP, 0},
 	{0x14FE77F7, TalesOfAbyss, US, 0},
-	{0x045D77E9, TalesOfAbyss, US, 0}, // undub
+	{0x045D77E9, TalesOfAbyss, JPUNDUB, 0},
 	{0xAA5EC3A3, TalesOfAbyss, JP, 0}, 
 	{0xFB236A46, SonicUnleashed, US, 0},
 	{0x4C7BB3C8, SimpsonsGame, Unknown, 0},
 	{0x4C94B32C, SimpsonsGame, Unknown, 0},
 	{0xD71B57F4, Genji, Unknown, 0},
+	{0x23A97857, StarOcean3, JPUNDUB, 0},
 };
 
 hash_map<uint32, CRC::Game*> CRC::m_map;

plugins/GSdx/GSCrc.h

@@ -34,6 +34,7 @@ public:
 		KyuuketsuKitanMoonties,
 		PiaCarroteYoukosoGPGakuenPrincess,
 		KazokuKeikakuKokoroNoKizuna,
+		DuelSaviorDestiny,
 		FFX,
 		FFX2,
 		FFXII,
@@ -70,6 +71,7 @@ public:
 		SonicUnleashed,
 		SimpsonsGame,
 		Genji,
+		StarOcean3,
 		TitleCount,
 	};
 

plugins/GSdx/GSDevice9.cpp

@@ -502,7 +502,6 @@ GSTexture* GSDevice9::CreateRenderTarget(int w, int h, int format)
 GSTexture* GSDevice9::CreateDepthStencil(int w, int h, int format)
 {
 	return __super::CreateDepthStencil(w, h, format ? format : D3DFMT_D24S8);
-	// return __super::CreateDepthStencil(w, h, format ? format : D3DFMT_D32F_LOCKABLE); 
 }
 
 GSTexture* GSDevice9::CreateTexture(int w, int h, int format)

plugins/GSdx/GSRendererDX.h

@@ -195,18 +195,25 @@ public:
 
 		ps_cb.FogColor_AREF = GSVector4((int)env.FOGCOL.FCR, (int)env.FOGCOL.FCG, (int)env.FOGCOL.FCB, (int)context->TEST.AREF) / 255;
 
-		if(ps_sel.atst == 2 || ps_sel.atst == 5)
+		if(ps_sel.atst == ATST_LESS || ps_sel.atst == ATST_GEQUAL)
 		{
-			ps_cb.FogColor_AREF.a -= 0.9f / 255;
+			ps_cb.FogColor_AREF.a -= 1.0f / 255;
 		}
-		else if(ps_sel.atst == 3 || ps_sel.atst == 6)
+		else if(ps_sel.atst == ATST_LEQUAL || ps_sel.atst == ATST_GREATER)
 		{
-			ps_cb.FogColor_AREF.a += 0.9f / 255;
+			// example: 
+			// ATST = ATST_GREATER, AREF = 127
+			// alpha = (int)127.0 => fail
+			// alpha = (int)127.5 => fail (!)
+			// alpha = (int)128.0 => pass
+
+			ps_cb.FogColor_AREF.a += 1.0f / 255; 
 		}
 
 		if(tex)
 		{
 			ps_sel.bpp = tex->m_bpp;
+			ps_sel.rt = tex->m_target;
 
 			int w = tex->m_texture->GetWidth();
 			int h = tex->m_texture->GetHeight();

plugins/GSdx/GSState.cpp

@@ -890,7 +890,7 @@ void GSState::GIFRegHandlerHWREG(GIFReg* r)
 {
 	ASSERT(m_env.TRXDIR.XDIR == 0); // host => local
 
-	Write((uint8*)r, 8); // hunting ground
+	Write((uint8*)r, 8); // haunting ground
 }
 
 void GSState::GIFRegHandlerSIGNAL(GIFReg* r)
@@ -2211,6 +2211,26 @@ bool GSC_Genji(const GSFrameInfo& fi, int& skip)
 	return true;
 }
 
+bool GSC_StarOcean3(const GSFrameInfo& fi, int& skip)
+{
+	if(skip == 0)
+	{
+		if(fi.TME && fi.FBP == fi.TBP0 && fi.FPSM == PSM_PSMCT32 && fi.TPSM == PSM_PSMT4HH)
+		{
+			skip = 1000; //
+		}
+	}
+	else
+	{
+		if(!(fi.TME && fi.FBP == fi.TBP0 && fi.FPSM == PSM_PSMCT32 && fi.TPSM == PSM_PSMT4HH))
+		{
+			skip = 0;
+		}
+	}
+
+	return true;
+}
+
 bool GSState::IsBadFrame(int& skip)
 {
 	GSFrameInfo fi;
@@ -2258,6 +2278,7 @@ bool GSState::IsBadFrame(int& skip)
 		map[CRC::TalesOfAbyss] = GSC_TalesOfAbyss;
 		map[CRC::SonicUnleashed] = GSC_SonicUnleashed;
 		map[CRC::Genji] = GSC_Genji;
+		map[CRC::StarOcean3] = GSC_StarOcean3;
 	}
 
 	// TODO: just set gsc in SetGameCRC once

plugins/GSdx/GSTextureCache.cpp

@@ -54,9 +54,9 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const GIFRegTEX0& TEX0, con
 	
 	Source* src = NULL;
 
-	const hash_map<Source*, bool>& map = m_src.m_map[TEX0.TBP0 >> 5];
+	hash_map<Source*, bool>& m = m_src.m_map[TEX0.TBP0 >> 5];
 
-	for(hash_map<Source*, bool>::const_iterator i = map.begin(); i != map.end(); i++)
+	for(hash_map<Source*, bool>::iterator i = m.begin(); i != m.end(); i++)
 	{
 		Source* s = i->first;
 
@@ -242,9 +242,9 @@ void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const
 
 	if(!target)
 	{
-		const hash_map<Source*, bool>& map = m_src.m_map[bp >> 5];
+		const hash_map<Source*, bool>& m = m_src.m_map[bp >> 5];
 
-		for(hash_map<Source*, bool>::const_iterator i = map.begin(); i != map.end(); )
+		for(hash_map<Source*, bool>::const_iterator i = m.begin(); i != m.end(); )
 		{
 			hash_map<Source*, bool>::const_iterator j = i++;
 
@@ -267,9 +267,9 @@ void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const
 
 			if(page < MAX_PAGES)
 			{
-				const hash_map<Source*, bool>& map = m_src.m_map[page];
+				const hash_map<Source*, bool>& m = m_src.m_map[page];
 
-				for(hash_map<Source*, bool>::const_iterator i = map.begin(); i != map.end(); )
+				for(hash_map<Source*, bool>::const_iterator i = m.begin(); i != m.end(); )
 				{
 					hash_map<Source*, bool>::const_iterator j = i++;
 
@@ -279,10 +279,8 @@ void GSTextureCache::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, const
 					{
 						if(!s->m_target)
 						{
-							s->m_blocks -= s->m_valid[page].count;
+							s->m_valid[page] = 0;
-
+							s->m_complete = false;
-							s->m_valid[page].block = 0;
-							s->m_valid[page].count = 0;
 
 							found = true;
 						}
@@ -485,6 +483,7 @@ GSTextureCache::Source::Source(GSRenderer* r)
 	, m_initpalette(false)
 	, m_bpp(0)
 	, m_target(false)
+	, m_complete(false)
 {
 	memset(m_valid, 0, sizeof(m_valid));
 
@@ -509,7 +508,7 @@ void GSTextureCache::Source::Update(const GIFRegTEX0& TEX0, const GIFRegTEXA& TE
 {
 	__super::Update();
 
-	if(m_target)
+	if(m_complete || m_target)
 	{
 		return;
 	}
@@ -517,17 +516,17 @@ void GSTextureCache::Source::Update(const GIFRegTEX0& TEX0, const GIFRegTEXA& TE
 	m_TEX0 = TEX0;
 	m_TEXA = TEXA;
 
-	if(m_blocks == m_total_blocks)
-	{
-		return;
-	}
-
 	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[m_TEX0.PSM];
 
 	GSVector2i bs = psm.bs;
 
 	GSVector4i r = rect.ralign<GSVector4i::Outside>(bs);
 
+	if(r.eq(GSVector4i(0, 0, 1 << m_TEX0.TW, 1 << m_TEX0.TH)))
+	{
+		m_complete = true; // lame, but better than nothing
+	}
+
 	uint32 bp = m_TEX0.TBP0;
 	uint32 bw = m_TEX0.TBW;
 
@@ -548,11 +547,12 @@ void GSTextureCache::Source::Update(const GIFRegTEX0& TEX0, const GIFRegTEXA& TE
 				uint32 row = block >> 5;
 				uint32 col = 1 << (block & 31);
 
-				if((m_valid[row].block & col) == 0)
+				if((m_valid[row] & col) == 0)
 				{
-					if(!repeating) m_valid[row].block |= col;
+					if(!repeating)
-
+					{
-					m_valid[row].count++;
+						m_valid[row] |= col;
+					}
 
 					Write(GSVector4i(x, y, x + bs.x, y + bs.y));
 
@@ -579,14 +579,12 @@ void GSTextureCache::Source::Update(const GIFRegTEX0& TEX0, const GIFRegTEXA& TE
 						uint32 row = block >> 5;
 						uint32 col = 1 << (block & 31);
 
-						m_valid[row].block |= col;
+						m_valid[row] |= col;
 					}
 				}
 			}
 		}
 
-		m_blocks += blocks;
-
 		m_renderer->m_perfmon.Put(GSPerfMon::Unswizzle, bs.x * bs.y * sizeof(uint32) * blocks);
 
 		Flush(m_write.count);
@@ -785,9 +783,7 @@ void GSTextureCache::SourceMap::Add(Source* s, const GIFRegTEX0& TEX0)
 
 	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[TEX0.PSM];
 
-	GSVector2i bs = psm.bs;
+	GSVector2i bs = (bp & 31) ? psm.pgs : psm.bs;
-
-	int blocks = 0;
 
 	for(int y = 0; y < th; y += bs.y)
 	{
@@ -795,19 +791,33 @@ void GSTextureCache::SourceMap::Add(Source* s, const GIFRegTEX0& TEX0)
 
 		for(int x = 0; x < tw; x += bs.x)
 		{
-			uint32 block = base + psm.blockOffset[x >> 3];
+			uint32 page = (base + psm.blockOffset[x >> 3]) >> 5;
 
-			if(block < MAX_BLOCKS)
+			if(page < MAX_PAGES)
 			{
-				m_map[block >> 5][s] = true;
+				m_pages[page >> 5] |= 1 << (page & 31);
-
-				blocks++;
 			}
 		}
 	}
 
-	s->m_blocks = 0;
+	for(int i = 0; i < countof(m_pages); i++)
-	s->m_total_blocks = blocks;
+	{
+		if(m_pages[i])
+		{
+			hash_map<Source*, bool>* m = &m_map[i << 5];
+			uint32 p = m_pages[i];
+
+			for(int j = 0; j < 32; j++)
+			{
+				if(p & (1 << j))
+				{
+					m[j][s] = true;
+				}
+			}
+
+			m_pages[i] = 0;
+		}
+	}
 }
 
 void GSTextureCache::SourceMap::RemoveAll()
@@ -819,7 +829,7 @@ void GSTextureCache::SourceMap::RemoveAll()
 
 	m_surfaces.clear();
 
-	for(int i = 0; i < MAX_PAGES; i++)
+	for(uint32 i = 0; i < countof(m_map); i++)
 	{
 		m_map[i].clear();
 	}
@@ -829,7 +839,7 @@ void GSTextureCache::SourceMap::RemoveAt(Source* s)
 {
 	m_surfaces.erase(s);
 
-	for(int i = 0; i < countof(m_map); i++)
+	for(uint32 i = s->m_TEX0.TBP0 >> 5; i < countof(m_map); i++)
 	{
 		m_map[i].erase(s);
 	}

plugins/GSdx/GSTextureCache.h

@@ -58,11 +58,11 @@ public:
 	public:
 		GSTexture* m_palette;
 		bool m_initpalette;
-		uint32 m_blocks, m_total_blocks;
+		uint32 m_valid[MAX_PAGES]; // each uint32 bits map to the 32 blocks of that page
-		struct {uint32 block, count;} m_valid[MAX_PAGES]; // each uint32 bits map to the 32 blocks of that page
 		uint32* m_clut;
 		int m_bpp;
 		bool m_target;
+		bool m_complete;
 
 	public:
 		explicit Source(GSRenderer* renderer);
@@ -95,9 +95,10 @@ protected:
 	{
 		hash_map<Source*, bool> m_surfaces;
 		hash_map<Source*, bool> m_map[MAX_PAGES];
+		uint32 m_pages[16];
 		bool m_used;
 
-		SourceMap() : m_used(false) {}
+		SourceMap() : m_used(false) {memset(m_pages, 0, sizeof(m_pages));}
 
 		void Add(Source* s, const GIFRegTEX0& TEX0);
 		void RemoveAll();

plugins/GSdx/GSTextureCache10.cpp

@@ -203,8 +203,14 @@ bool GSTextureCache10::Source10::Create(Target* dst)
 		m_initpalette = true;
 		break;
 	case PSM_PSMT4HL:
+		m_bpp = 4;
+		m_palette = m_renderer->m_dev->CreateTexture(256, 1);
+		m_initpalette = true;
+		break;
 	case PSM_PSMT4HH:
-		ASSERT(0); // TODO
+		m_bpp = 5;
+		m_palette = m_renderer->m_dev->CreateTexture(256, 1);
+		m_initpalette = true;
 		break;
 	}
 

plugins/GSdx/GSTextureCache11.cpp

@@ -205,8 +205,14 @@ bool GSTextureCache11::Source11::Create(Target* dst)
 		m_initpalette = true;
 		break;
 	case PSM_PSMT4HL:
+		m_bpp = 4;
+		m_palette = m_renderer->m_dev->CreateTexture(256, 1);
+		m_initpalette = true;
+		break;
 	case PSM_PSMT4HH:
-		ASSERT(0); // TODO
+		m_bpp = 5;
+		m_palette = m_renderer->m_dev->CreateTexture(256, 1);
+		m_initpalette = true;
 		break;
 	}
 

plugins/GSdx/GSTextureCache9.cpp

@@ -203,8 +203,14 @@ bool GSTextureCache9::Source9::Create(Target* dst)
 		m_initpalette = true;
 		break;
 	case PSM_PSMT4HL:
+		m_bpp = 4;
+		m_palette = m_renderer->m_dev->CreateTexture(256, 1);
+		m_initpalette = true;
+		break;
 	case PSM_PSMT4HH:
-		ASSERT(0); // TODO
+		m_bpp = 5;
+		m_palette = m_renderer->m_dev->CreateTexture(256, 1);
+		m_initpalette = true;
 		break;
 	}
 

plugins/GSdx/GSTextureCacheSW.cpp

@@ -177,7 +177,6 @@ void GSTextureCacheSW::InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, cons
 					if(GSUtil::HasSharedBits(BITBLTBUF.DPSM, t->m_TEX0.PSM))
 					{
 						t->m_valid[page] = 0;
-
 						t->m_complete = false;
 					}
 				}
@@ -237,7 +236,7 @@ bool GSTextureCacheSW::GSTexture::Update(const GIFRegTEX0& TEX0, const GIFRegTEX
 
 	GSVector4i r = rect.ralign<GSVector4i::Outside>(s);
 
-	if(r.left == 0 && r.top == 0 && r.right == tw && r.bottom == th)
+	if(r.eq(GSVector4i(0, 0, tw, th)))
 	{
 		m_complete = true; // lame, but better than nothing
 	}

plugins/GSdx/GSTextureFX10.cpp

@@ -288,7 +288,7 @@ void GSTextureFX10::UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSampl
 			m_ps_ss[ssel] = ss0;
 		}
 
-		if(sel.bpp == 3)
+		if(sel.bpp >= 3)
 		{
 			ss1 = m_palette_ss;
 		}

plugins/GSdx/GSTextureFX11.cpp

@@ -292,7 +292,7 @@ void GSTextureFX11::UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSampl
 			m_ps_ss[ssel] = ss0;
 		}
 
-		if(sel.bpp == 3)
+		if(sel.bpp >= 3)
 		{
 			ss1 = m_palette_ss;
 		}

plugins/GSdx/GSUtil.cpp

@@ -97,21 +97,6 @@ bool GSUtil::HasCompatibleBits(uint32 spsm, uint32 dpsm)
 	return s_maps.CompatibleBitsField[spsm][dpsm];
 }
 
-bool GSUtil::IsRectInRect(const GSVector4i& inner, const GSVector4i& outer)
-{
-	return outer.left <= inner.left && inner.right <= outer.right && outer.top <= inner.top && inner.bottom <= outer.bottom;
-}
-
-bool GSUtil::IsRectInRectH(const GSVector4i& inner, const GSVector4i& outer)
-{
-	return outer.top <= inner.top && inner.bottom <= outer.bottom;
-}
-
-bool GSUtil::IsRectInRectV(const GSVector4i& inner, const GSVector4i& outer)
-{
-	return outer.left <= inner.left && inner.right <= outer.right;
-}
-
 bool GSUtil::CheckDirectX()
 {
 	if(HINSTANCE hDll = LoadLibrary(format("d3dx9_%d.dll", D3DX_SDK_VERSION).c_str()))

plugins/GSdx/GSUtil.h

@@ -32,10 +32,6 @@ public:
 	static bool HasSharedBits(uint32 sbp, uint32 spsm, uint32 dbp, uint32 dpsm);
 	static bool HasCompatibleBits(uint32 spsm, uint32 dpsm);
 
-	static bool IsRectInRect(const GSVector4i& inner, const GSVector4i& outer);
-	static bool IsRectInRectH(const GSVector4i& inner, const GSVector4i& outer);
-	static bool IsRectInRectV(const GSVector4i& inner, const GSVector4i& outer);
-	
 	static uint32 EncodePSM(uint32 psm)
 	{
 		switch(psm)

plugins/GSdx/GSVector.h

@@ -24,18 +24,6 @@ public:
 		this->y = y;
 	}
 
-	GSVector2T(const GSVector2T& v) 
-	{
-		*this = v;
-	}
-
-	const GSVector2T& operator = (const GSVector2T& v) 
-	{
-		_mm_storel_epi64((__m128i*)this, _mm_loadl_epi64((__m128i*)&v));
-
-		return *this;
-	}
-
 	bool operator == (const GSVector2T& v) const
 	{
 		return x == v.x && y == v.y;

plugins/GSdx/res/tfx.fx

@@ -481,7 +481,7 @@ float4 sample(float2 tc, float w)
 
 		float4 t00, t01, t10, t11;
 
-		if(PS_BPP == 3) // 8HP
+		if(PS_BPP == 3) // 8H
 		{
 			float4 a;
 
@@ -495,6 +495,38 @@ float4 sample(float2 tc, float w)
 			t10 = Palette.Sample(PaletteSampler, a.z);
 			t11 = Palette.Sample(PaletteSampler, a.w);
 		}
+		else if(PS_BPP == 4) // 4HL
+		{
+			float4 a;
+
+			a.x = Texture.Sample(TextureSampler, uv.xy).a;
+			a.y = Texture.Sample(TextureSampler, uv.zy).a;
+			a.z = Texture.Sample(TextureSampler, uv.xw).a;
+			a.w = Texture.Sample(TextureSampler, uv.zw).a;
+
+			a = fmod(a, 1.0f / 16);
+
+			t00 = Palette.Sample(PaletteSampler, a.x);
+			t01 = Palette.Sample(PaletteSampler, a.y);
+			t10 = Palette.Sample(PaletteSampler, a.z);
+			t11 = Palette.Sample(PaletteSampler, a.w);
+		}
+		else if(PS_BPP == 5) // 4HH
+		{
+			float4 a;
+
+			a.x = Texture.Sample(TextureSampler, uv.xy).a;
+			a.y = Texture.Sample(TextureSampler, uv.zy).a;
+			a.z = Texture.Sample(TextureSampler, uv.xw).a;
+			a.w = Texture.Sample(TextureSampler, uv.zw).a;
+
+			a = fmod(a * 16, 1.0f / 16);
+
+			t00 = Palette.Sample(PaletteSampler, a.x);
+			t01 = Palette.Sample(PaletteSampler, a.y);
+			t10 = Palette.Sample(PaletteSampler, a.z);
+			t11 = Palette.Sample(PaletteSampler, a.w);
+		}
 		else
 		{
 			t00 = Texture.Sample(TextureSampler, uv.xy);
@@ -658,6 +690,42 @@ float4 sample(float2 tc, float w)
 			t10 = tex1D(Palette, a.z);
 			t11 = tex1D(Palette, a.w);
 		}
+		else if(PS_BPP == 4) // 4HL
+		{
+			float4 a;
+
+			a.x = tex2D(Texture, uv.xy).a;
+			a.y = tex2D(Texture, uv.zy).a;
+			a.z = tex2D(Texture, uv.xw).a;
+			a.w = tex2D(Texture, uv.zw).a;
+			
+			if(PS_RT == 1) a *= 0.5;
+			
+			a = fmod(a, 1.0f / 16);
+
+			t00 = tex1D(Palette, a.x);
+			t01 = tex1D(Palette, a.y);
+			t10 = tex1D(Palette, a.z);
+			t11 = tex1D(Palette, a.w);
+		}
+		else if(PS_BPP == 5) // 4HH
+		{
+			float4 a;
+
+			a.x = tex2D(Texture, uv.xy).a;
+			a.y = tex2D(Texture, uv.zy).a;
+			a.z = tex2D(Texture, uv.xw).a;
+			a.w = tex2D(Texture, uv.zw).a;
+
+			if(PS_RT == 1) a *= 0.5;
+			
+			a = fmod(a * 16, 1.0f / 16);
+
+			t00 = tex1D(Palette, a.x);
+			t01 = tex1D(Palette, a.y);
+			t10 = tex1D(Palette, a.z);
+			t11 = tex1D(Palette, a.w);
+		}
 		else
 		{
 			t00 = tex2D(Texture, uv.xy);

plugins/GSdx/xbyak/xbyak_util.h

@@ -167,7 +167,7 @@ struct EnableSetEip : public Gen {
 			Gen::call((void*)local::set_eip_to_esi);
 			break;
 		case Xbyak::Operand::EDI:
-			Gen::call((void*)local::set_eip_to_esi);
+			Gen::call((void*)local::set_eip_to_edi);
 			break;
 		case Xbyak::Operand::EBP:
 			Gen::call((void*)local::set_eip_to_ebp);