Big nasty wip push

Source/Core/Common/Hash.h

@@ -17,3 +17,104 @@ u64 GetHashHiresTexture(const u8 *src, int len, u32 samples);
 u64 GetMurmurHash3(const u8 *src, int len, u32 samples);
 u64 GetHash64(const u8 *src, int len, u32 samples);
 void SetHash64Function(bool useHiresTextures);
+
+#if _M_SSE >= 0x402
+
+template <size_t Sz>
+inline u32 GetCRC32_SmallSize(u64 cur, u8 const * b);
+
+template <> inline u32 GetCRC32_SmallSize<0>(u64 cur, u8 const *) { return u32(cur); }
+
+template <> inline u32 GetCRC32_SmallSize<1>(u64 cur, u8 const *b) { 
+	u32 tmp = u32(cur);
+	tmp = _mm_crc32_u8(tmp, *(u8 const *)(b+0));
+	return tmp; 
+}
+
+template <> inline u32 GetCRC32_SmallSize<2>(u64 cur, u8 const *b) { 
+	u32 tmp = u32(cur);
+	tmp = _mm_crc32_u16(tmp, *(u16 const *)(b+0));
+	return tmp; 
+}
+
+template <> inline u32 GetCRC32_SmallSize<3>(u64 cur, u8 const *b) { 
+	u32 tmp = u32(cur);
+	tmp = _mm_crc32_u16(tmp, *(u16 const *)(b+0));
+	tmp = _mm_crc32_u8(tmp, *(u8 const *)(b+2));
+	return tmp; 
+}
+
+template <> inline u32 GetCRC32_SmallSize<4>(u64 cur, u8 const *b) { 
+	u32 tmp = u32(cur);
+	tmp = _mm_crc32_u32(tmp, *(u32 const *)(b+0));
+	return tmp; 
+}
+
+template <> inline u32 GetCRC32_SmallSize<5>(u64 cur, u8 const *b) { 
+	u32 tmp = u32(cur);
+	tmp = _mm_crc32_u32(tmp, *(u32 const *)(b+0));
+	tmp = _mm_crc32_u8(tmp, *(u8 const *)(b+4));
+	return tmp; 
+}
+
+template <> inline u32 GetCRC32_SmallSize<6>(u64 cur, u8 const *b) { 
+	u32 tmp = u32(cur);
+	tmp = _mm_crc32_u32(tmp, *(u32 const *)(b+0));
+	tmp = _mm_crc32_u16(tmp, *(u16 const *)(b+4));
+	return tmp; 
+}
+
+template <> inline u32 GetCRC32_SmallSize<7>(u64 cur, u8 const *b) { 
+	u32 tmp = u32(cur);
+	tmp = _mm_crc32_u32(tmp, *(u32 const *)(b+0));
+	tmp = _mm_crc32_u16(tmp, *(u16 const *)(b+4));
+	tmp = _mm_crc32_u8(tmp, *(u8 const *)(b+6));
+	return tmp; 
+}
+
+template <> inline u32 GetCRC32_SmallSize<8>(u64 cur, u8 const *b) {
+	return u32(_mm_crc32_u64(cur, *(u64*)b));
+}
+
+template <size_t Sz>
+inline u32 GetCRC32_SmallSize(u64 cur, u8 const * b) {
+	return GetCRC32_SmallSize<Sz-8>( _mm_crc32_u64(cur, *(u64*)b),b+8);
+}
+
+template <typename T>
+u32 GetCRC32_CT(T const & st) {
+	return GetCRC32_SmallSize<sizeof(st)>(sizeof(st), (u8 const*)&st);
+#if 0
+	u32 sz { sizeof(st) };
+	u8 const * ptr{(u8 const *)&st};
+
+	u64 crc{sz};
+
+	while( sz >= 8 ) {
+		crc = _mm_crc32_u64(crc, *(u64*)ptr);
+		ptr += 8;
+		sz -= 8;
+	}
+	u32 crc32 = u32(crc);
+	if( sz >= 4 ) {
+		crc32 = _mm_crc32_u32(crc32, *(u32*)ptr);
+		ptr += 4;
+		sz -= 4;
+	}
+	if( sz >= 2 ) {
+		crc32 = _mm_crc32_u16(crc32, *(u16*)ptr);
+		ptr += 2;
+		sz -= 2;
+	}
+	if( sz ) {
+		crc32 = _mm_crc32_u8(crc32, *(u8*)ptr);
+	}
+
+	return crc32;
+#endif
+}
+#else 
+u64 GetCRC32_CT(const u8 *src, int len) {
+	return GetHash64(src,len, 0);
+}
+#endif

Source/Core/Common/Log.h

@@ -118,7 +118,7 @@ void GenericLog(LOGTYPES_LEVELS level, LOGTYPES_TYPE type,
 	if (!(_a_)) {\
 		ERROR_LOG(_t_, "Error...\n\n  Line: %d\n  File: %s\n  Time: %s\n\nIgnore and continue?", \
 					   __LINE__, __FILE__, __TIME__); \
-		if (!PanicYesNo("*** Assertion (see log)***\n")) {Crash();} \
+		if (!PanicYesNo("Assertion (see log) : \n" #_a_ "\n")) {Crash();} \
 	}
 #define _dbg_assert_msg_(_t_, _a_, ...)\
 	if (!(_a_)) {\

Source/Core/Common/Thread.cpp

@@ -15,8 +15,14 @@
 #include <process.h>
 #endif
 
+bool g_skipYield{true};
 namespace Common
 {
+	void YieldCPU()
+	{
+		if (!g_skipYield)
+			std::this_thread::yield();
+	}
 
 int CurrentThreadId()
 {

Source/Core/Common/Thread.h

@@ -74,10 +74,7 @@ void SwitchCurrentThread(); // On Linux, this is equal to sleep 1ms
 // Use this function during a spin-wait to make the current thread
 // relax while another thread is working. This may be more efficient
 // than using events because event functions use kernel calls.
-inline void YieldCPU()
-{
-	std::this_thread::yield();
-}
+void YieldCPU();
 
 void SetCurrentThreadName(const char *name);
 

Source/Core/Core/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp

@@ -248,7 +248,14 @@ void Interpreter::fnegx(UGeckoInstruction _inst)
 
 void Interpreter::fselx(UGeckoInstruction _inst)
 {
+	//double a, b, c,d;
+	//a = rPS0(_inst.FA);
+	//b = rPS0(_inst.FB);
+	//c = rPS0(_inst.FC);
 	rPS0(_inst.FD) = (rPS0(_inst.FA) >= -0.0) ? rPS0(_inst.FC) : rPS0(_inst.FB);
+	//d = rPS0(_inst.FD);
+	//NOTICE_LOG(VIDEO,"%lf = fsel(%lf %lf %lf)",d,a,b,c );
+
 	// This is a binary instruction. Does not alter FPSCR
 	if (_inst.Rc) Helper_UpdateCR1();
 }

Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp

@@ -199,6 +199,9 @@ void Jit64::fsign(UGeckoInstruction inst)
 
 void Jit64::fselx(UGeckoInstruction inst)
 {
+	FallBackToInterpreter(inst);
+	return;
+
 	INSTRUCTION_START
 	JITDISABLE(bJITFloatingPointOff)
 
@@ -215,12 +218,14 @@ void Jit64::fselx(UGeckoInstruction inst)
 
 	fpr.Lock(a, b, c, d);
 	MOVSD(XMM0, fpr.R(a));
+	PXOR(XMM0, M((void*)&psSignBits2));
+
 	PXOR(XMM1, R(XMM1));
 	// XMM0 = XMM0 < 0 ? all 1s : all 0s
-	CMPSD(XMM0, R(XMM1), LT);
+	CMPSD(XMM0, R(XMM1), NLT);
 	MOVSD(XMM1, R(XMM0));
-	PAND(XMM0, fpr.R(b));
-	PANDN(XMM1, fpr.R(c));
+	PAND(XMM0, fpr.R(c));
+	PANDN(XMM1, fpr.R(b));
 	POR(XMM0, R(XMM1));
 	fpr.BindToRegister(d, false);
 	MOVSD(fpr.RX(d), R(XMM0));

Source/Core/DolphinWX/DolphinWX.vcxproj.user

@@ -1,9 +1,13 @@
-<?xml version="1.0" encoding="utf-8"?>
+<?xml version="1.0" encoding="utf-8"?>
 <Project ToolsVersion="12.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <PropertyGroup>
     <!--For some stupid reason this has to be in the .user file...-->
     <LocalDebuggerCommand>$(BinaryOutputDir)$(TargetFileName)</LocalDebuggerCommand>
     <LocalDebuggerWorkingDirectory>$(BinaryOutputDir)</LocalDebuggerWorkingDirectory>
     <DebuggerFlavor>WindowsLocalDebugger</DebuggerFlavor>
   </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <LocalDebuggerCommandArguments>
+    </LocalDebuggerCommandArguments>
+  </PropertyGroup>
 </Project>

Source/Core/DolphinWX/Frame.cpp

@@ -913,7 +913,7 @@ void CFrame::OnKeyDown(wxKeyEvent& event)
 		{
 			OSDChoice = 1;
 			// Toggle native resolution
-			if (++g_Config.iEFBScale > SCALE_4X)
+			if (++g_Config.iEFBScale > SCALE_7X)
 				g_Config.iEFBScale = SCALE_AUTO;
 		}
 		else if (IsHotkey(event, HK_TOGGLE_AR))

Source/Core/DolphinWX/VideoConfigDiag.cpp

@@ -365,7 +365,13 @@ VideoConfigDiag::VideoConfigDiag(wxWindow* parent, const std::string &title, con
 	{
 	const wxString efbscale_choices[] = { _("Auto (Window Size)"), _("Auto (Multiple of 640x528)"),
 		_("1x Native (640x528)"), _("1.5x Native (960x792)"), _("2x Native (1280x1056)"),
-		_("2.5x Native (1600x1320)"), _("3x Native (1920x1584)"), _("4x Native (2560x2112)") };
+		_("2.5x Native (1600x1320)"), _("3x Native (1920x1584)"), _("4x Native (2560x2112)"),
+		_("5x Native (2560x2112)"),
+		_("6x Native (2560x2112)"),
+		_("7x Native (2560x2112)"),
+
+
+	};
 
 	wxChoice *const choice_efbscale = CreateChoice(page_enh,
 		vconfig.iEFBScale, wxGetTranslation(internal_res_desc), sizeof(efbscale_choices)/sizeof(*efbscale_choices), efbscale_choices);
@@ -514,7 +520,11 @@ VideoConfigDiag::VideoConfigDiag(wxWindow* parent, const std::string &title, con
 	wxGridSizer* const szr_other = new wxGridSizer(2, 5, 5);
 	szr_other->Add(CreateCheckBox(page_hacks, _("Disable Destination Alpha"), wxGetTranslation(disable_dstalpha_desc), vconfig.bDstAlphaPass));
 	szr_other->Add(CreateCheckBox(page_hacks, _("OpenMP Texture Decoder"), wxGetTranslation(omp_desc), vconfig.bOMPDecoder));
-	szr_other->Add(CreateCheckBox(page_hacks, _("Fast Depth Calculation"), wxGetTranslation(fast_depth_calc_desc), vconfig.bFastDepthCalc));
+
+	if (vconfig.backend_info.APIType != API_D3D)
+	{
+		szr_other->Add(CreateCheckBox(page_hacks, _("Fast Depth Calculation"), wxGetTranslation(fast_depth_calc_desc), vconfig.bFastDepthCalc));
+	}
 
 	wxStaticBoxSizer* const group_other = new wxStaticBoxSizer(wxVERTICAL, page_hacks, _("Other"));
 	group_other->Add(szr_other, 1, wxEXPAND | wxLEFT | wxRIGHT | wxBOTTOM, 5);

Source/Core/VideoBackends/D3D/D3D.vcxproj

@@ -55,6 +55,7 @@
     <ClCompile Include="NativeVertexFormat.cpp" />
     <ClCompile Include="PerfQuery.cpp" />
     <ClCompile Include="PixelShaderCache.cpp" />
+    <ClCompile Include="PSTextureDecoder.cpp" />
     <ClCompile Include="PSTextureEncoder.cpp" />
     <ClCompile Include="Render.cpp" />
     <ClCompile Include="stdafx.cpp">
@@ -81,6 +82,7 @@
     <ClInclude Include="main.h" />
     <ClInclude Include="PerfQuery.h" />
     <ClInclude Include="PixelShaderCache.h" />
+    <ClInclude Include="PSTextureDecoder.h" />
     <ClInclude Include="PSTextureEncoder.h" />
     <ClInclude Include="Render.h" />
     <ClInclude Include="stdafx.h" />

Source/Core/VideoBackends/D3D/D3D.vcxproj.filters

@@ -65,6 +65,9 @@
     <ClCompile Include="LineAndPointGeometryShader.cpp">
       <Filter>Render</Filter>
     </ClCompile>
+    <ClCompile Include="PSTextureDecoder.cpp">
+      <Filter>Render</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="D3DBase.h">
@@ -131,5 +134,8 @@
     <ClInclude Include="LineAndPointGeometryShader.h">
       <Filter>Render</Filter>
     </ClInclude>
+    <ClInclude Include="PSTextureDecoder.h">
+      <Filter>Render</Filter>
+    </ClInclude>
   </ItemGroup>
 </Project>