Add BC6 support to nvtt lib and utils.

- Use 3x3 eigensolver for initial fit in ZOH.  Slightly better perf and RMSE than power method.
- Remove use of double precision in ZOH - speeds up by 12%.
- Fixed RGBM encoding that was broken for HDR images.
- Use gamma-2.0 space for RGBM for HDR images (improves precision in darks).
- Use UNORM instead of TYPELESS formats when saving a DX10 .dds file.  The TYPELESS formats break most viewers.
- Cleaned up warnings in ZOH code.
- Command-line utils will warn if you give them an unrecognized parameter.
- Added VS2010 profiling results.

project/vc10/Nvidia.TextureTools/TextureTools.cs

@@ -31,6 +31,14 @@ namespace Nvidia.TextureTools
 		BC3n = DXT5n,
 		BC4,
 		BC5,
+
+		DXT1n,   // Not supported on CPU yet.
+		CTX1,    // Not supported on CPU yet.
+
+		BC6,
+		BC7,     // Not supported yet.
+
+		DXT1_Luma,
 	}
 	#endregion
 

project/vc10/bc6.psess

@@ -0,0 +1,105 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<VSPerformanceSession Version="1.00">
+  <Options>
+    <Solution>C:\nvtt\project\vc10\nvtt.sln</Solution>
+    <CollectionMethod>Sampling</CollectionMethod>
+    <AllocationMethod>None</AllocationMethod>
+    <LaunchRuntimeControl>true</LaunchRuntimeControl>
+    <AddReport>true</AddReport>
+    <ResourceBasedAnalysisSelected>false</ResourceBasedAnalysisSelected>
+    <ThreadBasedAnalysisSelected>false</ThreadBasedAnalysisSelected>
+    <UniqueReport>Timestamp</UniqueReport>
+    <SamplingMethod>Cycles</SamplingMethod>
+    <CycleCount>10000000</CycleCount>
+    <PageFaultCount>10</PageFaultCount>
+    <SysCallCount>10</SysCallCount>
+    <SamplingCounter Name="" ReloadValue="00000000000f4240" DisplayName="" />
+    <RelocateBinaries>false</RelocateBinaries>
+    <CollectOnHpcCluster>false</CollectOnHpcCluster>
+    <HPCSettings>
+      <NumberOfProcesses>0</NumberOfProcesses>
+      <TargetSelection>UnknownDisabled</TargetSelection>
+      <ProfileOnRank>0</ProfileOnRank>
+      <AdvancedProperties>
+        <Project>nvcompress</Project>
+        <CleanUp>true</CleanUp>
+        <MpiexecCommand>mpiexec.exe</MpiexecCommand>
+        <DeployCRT>true</DeployCRT>
+      </AdvancedProperties>
+    </HPCSettings>
+    <HardwareCounters EnableHWCounters="false" />
+    <EtwSettings />
+    <PdhSettings>
+      <PdhCountersEnabled>false</PdhCountersEnabled>
+      <PdhCountersRate>500</PdhCountersRate>
+      <PdhCounters>
+        <PdhCounter>\Memory\Pages/sec</PdhCounter>
+        <PdhCounter>\PhysicalDisk(_Total)\Avg. Disk Queue Length</PdhCounter>
+        <PdhCounter>\Processor(_Total)\% Processor Time</PdhCounter>
+      </PdhCounters>
+    </PdhSettings>
+  </Options>
+  <ExcludeSmallFuncs>true</ExcludeSmallFuncs>
+  <JScriptProfilingEnabled>false</JScriptProfilingEnabled>
+  <PreinstrumentEvent>
+    <InstrEventExclude>false</InstrEventExclude>
+  </PreinstrumentEvent>
+  <PostinstrumentEvent>
+    <InstrEventExclude>false</InstrEventExclude>
+  </PostinstrumentEvent>
+  <Binaries>
+    <ProjBinary>
+      <Path>C:\nvtt\project\vc10\Release.x64\bin\nvcompress.exe</Path>
+      <ArgumentTimestamp>01/01/0001 00:00:00</ArgumentTimestamp>
+      <Instrument>true</Instrument>
+      <Sample>true</Sample>
+      <ExternalWebsite>false</ExternalWebsite>
+      <InteractionProfilingEnabled>false</InteractionProfilingEnabled>
+      <IsLocalJavascript>false</IsLocalJavascript>
+      <LaunchProject>true</LaunchProject>
+      <OverrideProjectSettings>false</OverrideProjectSettings>
+      <LaunchMethod>Executable</LaunchMethod>
+      <ExecutablePath>C:\nvtt\project\vc10\\Release.x64\bin\nvcompress.exe</ExecutablePath>
+      <StartupDirectory>C:\nvtt\project\vc10\nvcompress\</StartupDirectory>
+      <Arguments>-bc6 -nomips C:\nvtt\data\testsuite\kodak\kodim01.png C:\nvtt\data\testsuite\kodak\kodim01_bc6.dds</Arguments>
+      <NetAppHost>IIS</NetAppHost>
+      <NetBrowser>InternetExplorer</NetBrowser>
+      <ExcludeSmallFuncs>true</ExcludeSmallFuncs>
+      <JScriptProfilingEnabled>false</JScriptProfilingEnabled>
+      <PreinstrumentEvent>
+        <InstrEventExclude>false</InstrEventExclude>
+      </PreinstrumentEvent>
+      <PostinstrumentEvent>
+        <InstrEventExclude>false</InstrEventExclude>
+      </PostinstrumentEvent>
+      <ProjRef>{88079E38-83AA-4E8A-B18A-66A78D1B058B}|nvcompress\nvcompress.vcxproj</ProjRef>
+      <ProjPath>C:\nvtt\project\vc10\nvcompress\nvcompress.vcxproj</ProjPath>
+      <ProjName>nvcompress</ProjName>
+    </ProjBinary>
+  </Binaries>
+  <Reports>
+    <Report>
+      <Path>C:\nvtt\project\vc10\bc6-01-eigen-solver.vsp</Path>
+    </Report>
+    <Report>
+      <Path>C:\nvtt\project\vc10\bc6-02-no-double.vsp</Path>
+    </Report>
+  </Reports>
+  <Launches>
+    <ProjBinary>
+      <Path>:PB:{88079E38-83AA-4E8A-B18A-66A78D1B058B}|nvcompress\nvcompress.vcxproj</Path>
+    </ProjBinary>
+  </Launches>
+  <RuntimeMarks>
+    <MarkName>Mark One</MarkName>
+    <MarkName>Mark Two</MarkName>
+    <MarkName>Mark Three</MarkName>
+    <MarkName>Mark Four</MarkName>
+    <MarkName>Mark Five</MarkName>
+    <MarkName>Mark Six</MarkName>
+    <MarkName>Mark Seven</MarkName>
+    <MarkName>Mark Eight</MarkName>
+    <MarkName>Mark Nine</MarkName>
+    <MarkName>Mark Ten</MarkName>
+  </RuntimeMarks>
+</VSPerformanceSession>

project/vc10/nvassemble/nvassemble.vcxproj

@@ -178,6 +178,9 @@
     <None Include="nvidia.ico" />
   </ItemGroup>
   <ItemGroup>
+    <ProjectReference Include="..\bc6h\bc6h.vcxproj">
+      <Project>{c33787e3-5564-4834-9fe3-a9020455a669}</Project>
+    </ProjectReference>
     <ProjectReference Include="..\nvcore\nvcore.vcxproj">
       <Project>{f143d180-d4c4-4037-b3de-be89a21c8d1d}</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>

project/vc10/nvcompress/nvcompress.vcxproj

@@ -352,6 +352,9 @@
     </None>
   </ItemGroup>
   <ItemGroup>
+    <ProjectReference Include="..\bc6h\bc6h.vcxproj">
+      <Project>{c33787e3-5564-4834-9fe3-a9020455a669}</Project>
+    </ProjectReference>
     <ProjectReference Include="..\nvcore\nvcore.vcxproj">
       <Project>{f143d180-d4c4-4037-b3de-be89a21c8d1d}</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>

project/vc10/nvddsinfo/nvddsinfo.vcxproj

@@ -178,6 +178,9 @@
     <None Include="nvidia.ico" />
   </ItemGroup>
   <ItemGroup>
+    <ProjectReference Include="..\bc6h\bc6h.vcxproj">
+      <Project>{c33787e3-5564-4834-9fe3-a9020455a669}</Project>
+    </ProjectReference>
     <ProjectReference Include="..\nvcore\nvcore.vcxproj">
       <Project>{f143d180-d4c4-4037-b3de-be89a21c8d1d}</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>

project/vc10/nvdecompress/nvdecompress.vcxproj

@@ -204,6 +204,9 @@
     </None>
   </ItemGroup>
   <ItemGroup>
+    <ProjectReference Include="..\bc6h\bc6h.vcxproj">
+      <Project>{c33787e3-5564-4834-9fe3-a9020455a669}</Project>
+    </ProjectReference>
     <ProjectReference Include="..\nvcore\nvcore.vcxproj">
       <Project>{f143d180-d4c4-4037-b3de-be89a21c8d1d}</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>

project/vc10/nvimgdiff/nvimgdiff.vcxproj

@@ -206,6 +206,9 @@
     <ResourceCompile Include="nvimgdiff.rc" />
   </ItemGroup>
   <ItemGroup>
+    <ProjectReference Include="..\bc6h\bc6h.vcxproj">
+      <Project>{c33787e3-5564-4834-9fe3-a9020455a669}</Project>
+    </ProjectReference>
     <ProjectReference Include="..\nvcore\nvcore.vcxproj">
       <Project>{f143d180-d4c4-4037-b3de-be89a21c8d1d}</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>

project/vc10/nvtt.sln

@@ -3,6 +3,7 @@ Microsoft Visual Studio Solution File, Format Version 11.00
 # Visual Studio 2010
 Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Solution Items", "Solution Items", "{A08D9CB4-C9D8-46EF-A74D-0EAB11FE34C5}"
 	ProjectSection(SolutionItems) = preProject
+		bc6.psess = bc6.psess
 		nvconfig.h = nvconfig.h
 	EndProjectSection
 EndProject
@@ -239,25 +240,31 @@ Global
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|Win32.ActiveCfg = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|Win32.Build.0 = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|x64.ActiveCfg = Debug|Any CPU
+		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|x64.Build.0 = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|Win32.ActiveCfg = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|Win32.Build.0 = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|x64.ActiveCfg = Release|Any CPU
+		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|x64.Build.0 = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|Win32.ActiveCfg = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|Win32.Build.0 = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|x64.ActiveCfg = Release|Any CPU
+		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|x64.Build.0 = Release|Any CPU
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|Win32.ActiveCfg = Debug|Win32
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|Win32.Build.0 = Debug|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|x64.ActiveCfg = Debug|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|x64.ActiveCfg = Debug|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|x64.Build.0 = Debug|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Win32.Build.0 = Debug|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Win32.ActiveCfg = Debug-CUDA|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|x64.ActiveCfg = Debug|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Win32.Build.0 = Debug-CUDA|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|x64.ActiveCfg = Debug-CUDA|x64
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|x64.Build.0 = Debug-CUDA|x64
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|Win32.ActiveCfg = Release|Win32
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|Win32.Build.0 = Release|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|x64.ActiveCfg = Release|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|x64.ActiveCfg = Release|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Win32.ActiveCfg = Release|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|x64.Build.0 = Release|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Win32.Build.0 = Release|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Win32.ActiveCfg = Release-CUDA|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|x64.ActiveCfg = Release|x64
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Win32.Build.0 = Release-CUDA|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|x64.Build.0 = Release|x64
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|x64.ActiveCfg = Release-CUDA|x64
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|x64.Build.0 = Release-CUDA|x64
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|Win32.ActiveCfg = Debug|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|Win32.Build.0 = Debug|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|x64.ActiveCfg = Debug|x64

project/vc10/nvtt/nvtt.vcxproj

@@ -148,7 +148,7 @@
       <PrecompiledHeader>
       </PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
-      <DebugInformationFormat>EditAndContinue</DebugInformationFormat>
+      <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
     </ClCompile>
     <Link>
       <AdditionalDependencies>cudart.lib;%(AdditionalDependencies)</AdditionalDependencies>
@@ -242,6 +242,7 @@
       </DataExecutionPrevention>
       <ImportLibrary>$(SolutionDir)\$(Configuration).$(Platform)\lib\$(ProjectName).lib</ImportLibrary>
       <TargetMachine>MachineX86</TargetMachine>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
     </Link>
     <PostBuildEvent>
       <Message>Copying header files...</Message>
@@ -283,6 +284,7 @@
       </DataExecutionPrevention>
       <ImportLibrary>$(SolutionDir)\$(Configuration).$(Platform)\lib\$(ProjectName).lib</ImportLibrary>
       <TargetMachine>MachineX64</TargetMachine>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
     </Link>
     <CudaCompile>
       <TargetMachinePlatform>64</TargetMachinePlatform>
@@ -302,7 +304,7 @@
       <PrecompiledHeader>
       </PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
-      <DebugInformationFormat>EditAndContinue</DebugInformationFormat>
+      <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
     </ClCompile>
     <Link>
       <AdditionalDependencies>%(AdditionalDependencies)</AdditionalDependencies>
@@ -355,8 +357,7 @@
       <PrecompiledHeader>
       </PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
-      <DebugInformationFormat>
+      <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
-      </DebugInformationFormat>
     </ClCompile>
     <Link>
       <AdditionalDependencies>%(AdditionalDependencies)</AdditionalDependencies>
@@ -369,6 +370,7 @@
       </DataExecutionPrevention>
       <ImportLibrary>$(SolutionDir)\$(Configuration).$(Platform)\lib\$(ProjectName).lib</ImportLibrary>
       <TargetMachine>MachineX86</TargetMachine>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
     </Link>
     <PostBuildEvent>
       <Message>Copying header files...</Message>
@@ -384,6 +386,7 @@
       <PreprocessorDefinitions>WIN32;NDEBUG;_WINDOWS;_USRDLL;NVTT_EXPORTS;NVTT_SHARED;__SSE2__;__SSE__;__MMX__;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
       <WarningLevel>Level3</WarningLevel>
+      <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
     </ClCompile>
     <Link>
       <AdditionalLibraryDirectories>$(CUDA_LIB_PATH)\..\lib64;%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
@@ -397,6 +400,7 @@
       <ImportLibrary>$(SolutionDir)\$(Configuration).$(Platform)\lib\$(ProjectName).lib</ImportLibrary>
       <TargetMachine>MachineX64</TargetMachine>
       <AdditionalDependencies>%(AdditionalDependencies)</AdditionalDependencies>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
     </Link>
   </ItemDefinitionGroup>
   <ItemGroup>
@@ -415,16 +419,7 @@
     <ClInclude Include="..\..\..\src\nvtt\CompressionOptions.h" />
     <ClInclude Include="..\..\..\src\nvtt\Compressor.h" />
     <ClInclude Include="..\..\..\src\nvtt\CompressorDX10.h" />
-    <ClInclude Include="..\..\..\src\nvtt\CompressorDX11.h">
+    <ClInclude Include="..\..\..\src\nvtt\CompressorDX11.h" />
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|x64'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|x64'">true</ExcludedFromBuild>
-    </ClInclude>
     <ClInclude Include="..\..\..\src\nvtt\CompressorDX9.h" />
     <ClInclude Include="..\..\..\src\nvtt\CompressorRGB.h" />
     <ClInclude Include="..\..\..\src\nvtt\Context.h" />
@@ -457,16 +452,7 @@
     <ClCompile Include="..\..\..\src\nvtt\ClusterFit.cpp" />
     <ClCompile Include="..\..\..\src\nvtt\CompressionOptions.cpp" />
     <ClCompile Include="..\..\..\src\nvtt\CompressorDX10.cpp" />
-    <ClCompile Include="..\..\..\src\nvtt\CompressorDX11.cpp">
+    <ClCompile Include="..\..\..\src\nvtt\CompressorDX11.cpp" />
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|x64'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|x64'">true</ExcludedFromBuild>
-    </ClCompile>
     <ClCompile Include="..\..\..\src\nvtt\CompressorDX9.cpp" />
     <ClCompile Include="..\..\..\src\nvtt\CompressorRGB.cpp" />
     <ClCompile Include="..\..\..\src\nvtt\Context.cpp" />

project/vc10/nvzoom/nvzoom.vcxproj

@@ -197,6 +197,9 @@
     <ResourceCompile Include="nvzoom.rc" />
   </ItemGroup>
   <ItemGroup>
+    <ProjectReference Include="..\bc6h\bc6h.vcxproj">
+      <Project>{c33787e3-5564-4834-9fe3-a9020455a669}</Project>
+    </ProjectReference>
     <ProjectReference Include="..\nvcore\nvcore.vcxproj">
       <Project>{f143d180-d4c4-4037-b3de-be89a21c8d1d}</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>

project/vc10/testsuite/testsuite.vcxproj

@@ -1,6 +1,14 @@
 <?xml version="1.0" encoding="utf-8"?>
 <Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <ItemGroup Label="ProjectConfigurations">
+    <ProjectConfiguration Include="Debug-CUDA|Win32">
+      <Configuration>Debug-CUDA</Configuration>
+      <Platform>Win32</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Debug-CUDA|x64">
+      <Configuration>Debug-CUDA</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
     <ProjectConfiguration Include="Debug|Win32">
       <Configuration>Debug</Configuration>
       <Platform>Win32</Platform>
@@ -9,6 +17,14 @@
       <Configuration>Debug</Configuration>
       <Platform>x64</Platform>
     </ProjectConfiguration>
+    <ProjectConfiguration Include="Release-CUDA|Win32">
+      <Configuration>Release-CUDA</Configuration>
+      <Platform>Win32</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release-CUDA|x64">
+      <Configuration>Release-CUDA</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
     <ProjectConfiguration Include="Release|Win32">
       <Configuration>Release</Configuration>
       <Platform>Win32</Platform>
@@ -30,19 +46,37 @@
     <CharacterSet>Unicode</CharacterSet>
     <WholeProgramOptimization>true</WholeProgramOptimization>
   </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|Win32'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <CharacterSet>Unicode</CharacterSet>
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+  </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'" Label="Configuration">
     <ConfigurationType>Application</ConfigurationType>
     <CharacterSet>Unicode</CharacterSet>
     <WholeProgramOptimization>true</WholeProgramOptimization>
   </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|x64'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <CharacterSet>Unicode</CharacterSet>
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+  </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'" Label="Configuration">
     <ConfigurationType>Application</ConfigurationType>
     <CharacterSet>Unicode</CharacterSet>
   </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|Win32'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'" Label="Configuration">
     <ConfigurationType>Application</ConfigurationType>
     <CharacterSet>Unicode</CharacterSet>
   </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|x64'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
   <ImportGroup Label="ExtensionSettings">
   </ImportGroup>
@@ -50,33 +84,61 @@
     <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
     <Import Project="$(SolutionDir)\nvtt.props" />
   </ImportGroup>
+  <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|Win32'" Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+    <Import Project="$(SolutionDir)\nvtt.props" />
+  </ImportGroup>
   <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'" Label="PropertySheets">
     <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
     <Import Project="$(SolutionDir)\nvtt.props" />
   </ImportGroup>
+  <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|x64'" Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+    <Import Project="$(SolutionDir)\nvtt.props" />
+  </ImportGroup>
   <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'" Label="PropertySheets">
     <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
     <Import Project="$(SolutionDir)\nvtt.props" />
   </ImportGroup>
+  <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|Win32'" Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+    <Import Project="$(SolutionDir)\nvtt.props" />
+  </ImportGroup>
   <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'" Label="PropertySheets">
     <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
     <Import Project="$(SolutionDir)\nvtt.props" />
   </ImportGroup>
+  <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|x64'" Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+    <Import Project="$(SolutionDir)\nvtt.props" />
+  </ImportGroup>
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup>
     <_ProjectFileVersion>10.0.30319.1</_ProjectFileVersion>
     <OutDir Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(SolutionDir)\$(Configuration).$(Platform)\bin\</OutDir>
+    <OutDir Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|Win32'">$(SolutionDir)\$(Configuration).$(Platform)\bin\</OutDir>
     <OutDir Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(SolutionDir)\$(Configuration).$(Platform)\bin\</OutDir>
+    <OutDir Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|x64'">$(SolutionDir)\$(Configuration).$(Platform)\bin\</OutDir>
     <IntDir Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(Configuration)\$(Platform)\</IntDir>
+    <IntDir Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|Win32'">$(Configuration)\$(Platform)\</IntDir>
     <IntDir Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(Configuration)\$(Platform)\</IntDir>
+    <IntDir Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|x64'">$(Configuration)\$(Platform)\</IntDir>
     <LinkIncremental Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</LinkIncremental>
+    <LinkIncremental Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|Win32'">true</LinkIncremental>
     <LinkIncremental Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</LinkIncremental>
+    <LinkIncremental Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|x64'">true</LinkIncremental>
     <OutDir Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(SolutionDir)\$(Configuration).$(Platform)\bin\</OutDir>
+    <OutDir Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|Win32'">$(SolutionDir)\$(Configuration).$(Platform)\bin\</OutDir>
     <OutDir Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(SolutionDir)\$(Configuration).$(Platform)\bin\</OutDir>
+    <OutDir Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|x64'">$(SolutionDir)\$(Configuration).$(Platform)\bin\</OutDir>
     <IntDir Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(Configuration)\$(Platform)\</IntDir>
+    <IntDir Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|Win32'">$(Configuration)\$(Platform)\</IntDir>
     <IntDir Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(Configuration)\$(Platform)\</IntDir>
+    <IntDir Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|x64'">$(Configuration)\$(Platform)\</IntDir>
     <LinkIncremental Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">false</LinkIncremental>
+    <LinkIncremental Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|Win32'">false</LinkIncremental>
     <LinkIncremental Condition="'$(Configuration)|$(Platform)'=='Release|x64'">false</LinkIncremental>
+    <LinkIncremental Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|x64'">false</LinkIncremental>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <ClCompile>
@@ -102,6 +164,30 @@
       <TargetMachine>MachineX86</TargetMachine>
     </Link>
   </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|Win32'">
+    <ClCompile>
+      <Optimization>Disabled</Optimization>
+      <AdditionalIncludeDirectories>$(SolutionDir);$(SolutionDir)\..\..\src;$(SolutionDir)\..\..\extern\poshlib;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <MinimalRebuild>true</MinimalRebuild>
+      <BasicRuntimeChecks>EnableFastChecks</BasicRuntimeChecks>
+      <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <DebugInformationFormat>EditAndContinue</DebugInformationFormat>
+    </ClCompile>
+    <Link>
+      <AdditionalDependencies>%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalLibraryDirectories>%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <SubSystem>Console</SubSystem>
+      <RandomizedBaseAddress>false</RandomizedBaseAddress>
+      <DataExecutionPrevention>
+      </DataExecutionPrevention>
+      <TargetMachine>MachineX86</TargetMachine>
+    </Link>
+  </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <ClCompile>
       <Optimization>Disabled</Optimization>
@@ -124,6 +210,28 @@
       </DataExecutionPrevention>
     </Link>
   </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug-CUDA|x64'">
+    <ClCompile>
+      <Optimization>Disabled</Optimization>
+      <AdditionalIncludeDirectories>$(SolutionDir);$(SolutionDir)\..\..\src;$(SolutionDir)\..\..\extern\poshlib;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <BasicRuntimeChecks>EnableFastChecks</BasicRuntimeChecks>
+      <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
+    </ClCompile>
+    <Link>
+      <AdditionalDependencies>%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalLibraryDirectories>%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <SubSystem>Console</SubSystem>
+      <RandomizedBaseAddress>false</RandomizedBaseAddress>
+      <DataExecutionPrevention>
+      </DataExecutionPrevention>
+    </Link>
+  </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
     <ClCompile>
       <AdditionalIncludeDirectories>$(SolutionDir);$(SolutionDir)\..\..\src;$(SolutionDir)\..\..\extern\poshlib;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
@@ -151,6 +259,33 @@
       <TargetMachine>MachineX86</TargetMachine>
     </Link>
   </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|Win32'">
+    <ClCompile>
+      <AdditionalIncludeDirectories>$(SolutionDir);$(SolutionDir)\..\..\src;$(SolutionDir)\..\..\extern\poshlib;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <StringPooling>true</StringPooling>
+      <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
+      <EnableEnhancedInstructionSet>StreamingSIMDExtensions2</EnableEnhancedInstructionSet>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <DebugInformationFormat>
+      </DebugInformationFormat>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
+    </ClCompile>
+    <Link>
+      <AdditionalDependencies>%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalLibraryDirectories>%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <SubSystem>Console</SubSystem>
+      <OptimizeReferences>true</OptimizeReferences>
+      <EnableCOMDATFolding>true</EnableCOMDATFolding>
+      <RandomizedBaseAddress>false</RandomizedBaseAddress>
+      <DataExecutionPrevention>
+      </DataExecutionPrevention>
+      <TargetMachine>MachineX86</TargetMachine>
+    </Link>
+  </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <ClCompile>
       <AdditionalIncludeDirectories>$(SolutionDir);$(SolutionDir)\..\..\src;$(SolutionDir)\..\..\extern\poshlib;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
@@ -177,6 +312,32 @@
       </DataExecutionPrevention>
     </Link>
   </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release-CUDA|x64'">
+    <ClCompile>
+      <AdditionalIncludeDirectories>$(SolutionDir);$(SolutionDir)\..\..\src;$(SolutionDir)\..\..\extern\poshlib;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <StringPooling>true</StringPooling>
+      <RuntimeLibrary>MultiThreadedDLL</RuntimeLibrary>
+      <EnableEnhancedInstructionSet>StreamingSIMDExtensions2</EnableEnhancedInstructionSet>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <DebugInformationFormat>
+      </DebugInformationFormat>
+      <MultiProcessorCompilation>true</MultiProcessorCompilation>
+    </ClCompile>
+    <Link>
+      <AdditionalDependencies>%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalLibraryDirectories>%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <SubSystem>Console</SubSystem>
+      <OptimizeReferences>true</OptimizeReferences>
+      <EnableCOMDATFolding>true</EnableCOMDATFolding>
+      <RandomizedBaseAddress>false</RandomizedBaseAddress>
+      <DataExecutionPrevention>
+      </DataExecutionPrevention>
+    </Link>
+  </ItemDefinitionGroup>
   <ItemGroup>
     <ClCompile Include="..\..\..\src\nvtt\tests\testsuite.cpp" />
   </ItemGroup>

project/vc9/Nvidia.TextureTools/TextureTools.cs

@@ -31,6 +31,14 @@ namespace Nvidia.TextureTools
 		BC3n = DXT5n,
 		BC4,
 		BC5,
+
+        DXT1n,   // Not supported on CPU yet.
+        CTX1,    // Not supported on CPU yet.
+
+        BC6,
+        BC7,     // Not supported yet.
+
+        DXT1_Luma,
     }
 	#endregion
 

project/vc9/nvtt.sln

@@ -16,6 +16,7 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "nvcompress", "nvcompress\nv
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D} = {F143D180-D4C4-4037-B3DE-BE89A21C8D1D}
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647} = {1AEB7681-57D8-48EE-813D-5C41CC38B647}
 		{4046F392-A18B-4C66-9639-3EABFFF5D531} = {4046F392-A18B-4C66-9639-3EABFFF5D531}
+		{C33787E3-5564-4834-9FE3-A9020455A669} = {C33787E3-5564-4834-9FE3-A9020455A669}
 		{50C465FE-B308-42BC-894D-89484482AF06} = {50C465FE-B308-42BC-894D-89484482AF06}
 	EndProjectSection
 EndProject
@@ -36,6 +37,7 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "nvddsinfo", "nvddsinfo\nvdd
 	ProjectSection(ProjectDependencies) = postProject
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D} = {F143D180-D4C4-4037-B3DE-BE89A21C8D1D}
 		{4046F392-A18B-4C66-9639-3EABFFF5D531} = {4046F392-A18B-4C66-9639-3EABFFF5D531}
+		{C33787E3-5564-4834-9FE3-A9020455A669} = {C33787E3-5564-4834-9FE3-A9020455A669}
 		{50C465FE-B308-42BC-894D-89484482AF06} = {50C465FE-B308-42BC-894D-89484482AF06}
 	EndProjectSection
 EndProject
@@ -43,6 +45,7 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "nvdecompress", "nvdecompres
 	ProjectSection(ProjectDependencies) = postProject
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D} = {F143D180-D4C4-4037-B3DE-BE89A21C8D1D}
 		{4046F392-A18B-4C66-9639-3EABFFF5D531} = {4046F392-A18B-4C66-9639-3EABFFF5D531}
+		{C33787E3-5564-4834-9FE3-A9020455A669} = {C33787E3-5564-4834-9FE3-A9020455A669}
 		{50C465FE-B308-42BC-894D-89484482AF06} = {50C465FE-B308-42BC-894D-89484482AF06}
 	EndProjectSection
 EndProject
@@ -50,6 +53,7 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "nvimgdiff", "nvimgdiff\nvim
 	ProjectSection(ProjectDependencies) = postProject
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D} = {F143D180-D4C4-4037-B3DE-BE89A21C8D1D}
 		{4046F392-A18B-4C66-9639-3EABFFF5D531} = {4046F392-A18B-4C66-9639-3EABFFF5D531}
+		{C33787E3-5564-4834-9FE3-A9020455A669} = {C33787E3-5564-4834-9FE3-A9020455A669}
 		{50C465FE-B308-42BC-894D-89484482AF06} = {50C465FE-B308-42BC-894D-89484482AF06}
 	EndProjectSection
 EndProject
@@ -57,6 +61,7 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "nvassemble", "nvassemble\nv
 	ProjectSection(ProjectDependencies) = postProject
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D} = {F143D180-D4C4-4037-B3DE-BE89A21C8D1D}
 		{4046F392-A18B-4C66-9639-3EABFFF5D531} = {4046F392-A18B-4C66-9639-3EABFFF5D531}
+		{C33787E3-5564-4834-9FE3-A9020455A669} = {C33787E3-5564-4834-9FE3-A9020455A669}
 		{50C465FE-B308-42BC-894D-89484482AF06} = {50C465FE-B308-42BC-894D-89484482AF06}
 	EndProjectSection
 EndProject
@@ -64,6 +69,7 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "nvzoom", "nvzoom\nvzoom.vcp
 	ProjectSection(ProjectDependencies) = postProject
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D} = {F143D180-D4C4-4037-B3DE-BE89A21C8D1D}
 		{4046F392-A18B-4C66-9639-3EABFFF5D531} = {4046F392-A18B-4C66-9639-3EABFFF5D531}
+		{C33787E3-5564-4834-9FE3-A9020455A669} = {C33787E3-5564-4834-9FE3-A9020455A669}
 		{50C465FE-B308-42BC-894D-89484482AF06} = {50C465FE-B308-42BC-894D-89484482AF06}
 	EndProjectSection
 EndProject
@@ -106,445 +112,302 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "hdrtest", "hdrtest\hdrtest.
 EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
-		Debug|Mixed Platforms = Debug|Mixed Platforms
 		Debug|Win32 = Debug|Win32
 		Debug|x64 = Debug|x64
-		Debug-CUDA|Mixed Platforms = Debug-CUDA|Mixed Platforms
 		Debug-CUDA|Win32 = Debug-CUDA|Win32
 		Debug-CUDA|x64 = Debug-CUDA|x64
-		Release|Mixed Platforms = Release|Mixed Platforms
 		Release|Win32 = Release|Win32
 		Release|x64 = Release|x64
-		Release-CUDA|Mixed Platforms = Release-CUDA|Mixed Platforms
 		Release-CUDA|Win32 = Release-CUDA|Win32
 		Release-CUDA|x64 = Release-CUDA|x64
 	EndGlobalSection
 	GlobalSection(ProjectConfigurationPlatforms) = postSolution
-		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug|Win32.ActiveCfg = Debug|Win32
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug|Win32.Build.0 = Debug|Win32
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug|x64.ActiveCfg = Debug|x64
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug|x64.Build.0 = Debug|x64
-		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug-CUDA|x64
-		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug-CUDA|Mixed Platforms.Build.0 = Debug-CUDA|x64
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug-CUDA|Win32.ActiveCfg = Debug-CUDA|Win32
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug-CUDA|Win32.Build.0 = Debug-CUDA|Win32
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug-CUDA|x64.ActiveCfg = Debug-CUDA|x64
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Debug-CUDA|x64.Build.0 = Debug-CUDA|x64
-		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release|Mixed Platforms.Build.0 = Release|x64
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release|Win32.ActiveCfg = Release|Win32
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release|Win32.Build.0 = Release|Win32
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release|x64.ActiveCfg = Release|x64
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release|x64.Build.0 = Release|x64
-		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release-CUDA|Mixed Platforms.ActiveCfg = Release-CUDA|x64
-		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release-CUDA|Mixed Platforms.Build.0 = Release-CUDA|x64
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release-CUDA|Win32.ActiveCfg = Release-CUDA|Win32
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release-CUDA|Win32.Build.0 = Release-CUDA|Win32
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release-CUDA|x64.ActiveCfg = Release-CUDA|x64
 		{1AEB7681-57D8-48EE-813D-5C41CC38B647}.Release-CUDA|x64.Build.0 = Release-CUDA|x64
-		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug|Win32.ActiveCfg = Debug|Win32
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug|Win32.Build.0 = Debug|Win32
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug|x64.ActiveCfg = Debug|x64
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug|x64.Build.0 = Debug|x64
-		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug-CUDA|x64
-		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug-CUDA|Mixed Platforms.Build.0 = Debug-CUDA|x64
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug-CUDA|Win32.ActiveCfg = Debug-CUDA|Win32
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug-CUDA|Win32.Build.0 = Debug-CUDA|Win32
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug-CUDA|x64.ActiveCfg = Debug-CUDA|x64
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Debug-CUDA|x64.Build.0 = Debug-CUDA|x64
-		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release|Mixed Platforms.Build.0 = Release|x64
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release|Win32.ActiveCfg = Release|Win32
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release|Win32.Build.0 = Release|Win32
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release|x64.ActiveCfg = Release|x64
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release|x64.Build.0 = Release|x64
-		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release-CUDA|Mixed Platforms.ActiveCfg = Release-CUDA|x64
-		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release-CUDA|Mixed Platforms.Build.0 = Release-CUDA|x64
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release-CUDA|Win32.ActiveCfg = Release-CUDA|Win32
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release-CUDA|Win32.Build.0 = Release-CUDA|Win32
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release-CUDA|x64.ActiveCfg = Release-CUDA|x64
 		{88079E38-83AA-4E8A-B18A-66A78D1B058B}.Release-CUDA|x64.Build.0 = Release-CUDA|x64
-		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug|Win32.ActiveCfg = Debug|Win32
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug|Win32.Build.0 = Debug|Win32
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug|x64.ActiveCfg = Debug|x64
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug|x64.Build.0 = Debug|x64
-		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release|Mixed Platforms.Build.0 = Release|x64
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release|Win32.ActiveCfg = Release|Win32
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release|Win32.Build.0 = Release|Win32
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release|x64.ActiveCfg = Release|x64
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release|x64.Build.0 = Release|x64
-		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{4046F392-A18B-4C66-9639-3EABFFF5D531}.Release-CUDA|x64.Build.0 = Release|x64
-		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug|Win32.ActiveCfg = Debug|Win32
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug|Win32.Build.0 = Debug|Win32
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug|x64.ActiveCfg = Debug|x64
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug|x64.Build.0 = Debug|x64
-		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release|Mixed Platforms.Build.0 = Release|x64
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release|Win32.ActiveCfg = Release|Win32
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release|Win32.Build.0 = Release|Win32
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release|x64.ActiveCfg = Release|x64
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release|x64.Build.0 = Release|x64
-		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{F143D180-D4C4-4037-B3DE-BE89A21C8D1D}.Release-CUDA|x64.Build.0 = Release|x64
-		{50C465FE-B308-42BC-894D-89484482AF06}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{50C465FE-B308-42BC-894D-89484482AF06}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{50C465FE-B308-42BC-894D-89484482AF06}.Debug|Win32.ActiveCfg = Debug|Win32
 		{50C465FE-B308-42BC-894D-89484482AF06}.Debug|Win32.Build.0 = Debug|Win32
 		{50C465FE-B308-42BC-894D-89484482AF06}.Debug|x64.ActiveCfg = Debug|x64
 		{50C465FE-B308-42BC-894D-89484482AF06}.Debug|x64.Build.0 = Debug|x64
-		{50C465FE-B308-42BC-894D-89484482AF06}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{50C465FE-B308-42BC-894D-89484482AF06}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{50C465FE-B308-42BC-894D-89484482AF06}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{50C465FE-B308-42BC-894D-89484482AF06}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{50C465FE-B308-42BC-894D-89484482AF06}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{50C465FE-B308-42BC-894D-89484482AF06}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{50C465FE-B308-42BC-894D-89484482AF06}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{50C465FE-B308-42BC-894D-89484482AF06}.Release|Mixed Platforms.Build.0 = Release|x64
 		{50C465FE-B308-42BC-894D-89484482AF06}.Release|Win32.ActiveCfg = Release|Win32
 		{50C465FE-B308-42BC-894D-89484482AF06}.Release|Win32.Build.0 = Release|Win32
 		{50C465FE-B308-42BC-894D-89484482AF06}.Release|x64.ActiveCfg = Release|x64
 		{50C465FE-B308-42BC-894D-89484482AF06}.Release|x64.Build.0 = Release|x64
-		{50C465FE-B308-42BC-894D-89484482AF06}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{50C465FE-B308-42BC-894D-89484482AF06}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{50C465FE-B308-42BC-894D-89484482AF06}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{50C465FE-B308-42BC-894D-89484482AF06}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{50C465FE-B308-42BC-894D-89484482AF06}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{50C465FE-B308-42BC-894D-89484482AF06}.Release-CUDA|x64.Build.0 = Release|x64
-		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug|Win32.ActiveCfg = Debug|Win32
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug|Win32.Build.0 = Debug|Win32
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug|x64.ActiveCfg = Debug|x64
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug|x64.Build.0 = Debug|x64
-		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release|Mixed Platforms.Build.0 = Release|x64
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release|Win32.ActiveCfg = Release|Win32
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release|Win32.Build.0 = Release|Win32
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release|x64.ActiveCfg = Release|x64
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release|x64.Build.0 = Release|x64
-		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{CE017322-01FC-4851-9C8B-64E9A8E26C38}.Release-CUDA|x64.Build.0 = Release|x64
-		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug|Win32.ActiveCfg = Debug|Win32
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug|Win32.Build.0 = Debug|Win32
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug|x64.ActiveCfg = Debug|x64
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug|x64.Build.0 = Debug|x64
-		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release|Mixed Platforms.Build.0 = Release|x64
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release|Win32.ActiveCfg = Release|Win32
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release|Win32.Build.0 = Release|Win32
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release|x64.ActiveCfg = Release|x64
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release|x64.Build.0 = Release|x64
-		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{841B73C5-C679-4EEF-A50A-7D6106642B49}.Release-CUDA|x64.Build.0 = Release|x64
-		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug|Win32.ActiveCfg = Debug|Win32
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug|Win32.Build.0 = Debug|Win32
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug|x64.ActiveCfg = Debug|x64
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug|x64.Build.0 = Debug|x64
-		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release|Mixed Platforms.Build.0 = Release|x64
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release|Win32.ActiveCfg = Release|Win32
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release|Win32.Build.0 = Release|Win32
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release|x64.ActiveCfg = Release|x64
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release|x64.Build.0 = Release|x64
-		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{75A0527D-BFC9-49C3-B46B-CD1A901D5927}.Release-CUDA|x64.Build.0 = Release|x64
-		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug|Win32.ActiveCfg = Debug|Win32
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug|Win32.Build.0 = Debug|Win32
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug|x64.ActiveCfg = Debug|x64
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug|x64.Build.0 = Debug|x64
-		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release|Mixed Platforms.Build.0 = Release|x64
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release|Win32.ActiveCfg = Release|Win32
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release|Win32.Build.0 = Release|Win32
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release|x64.ActiveCfg = Release|x64
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release|x64.Build.0 = Release|x64
-		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{05A59E8B-EA70-4F22-89E8-E0927BA13064}.Release-CUDA|x64.Build.0 = Release|x64
-		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug|Win32.ActiveCfg = Debug|Win32
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug|Win32.Build.0 = Debug|Win32
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug|x64.ActiveCfg = Debug|x64
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug|x64.Build.0 = Debug|x64
-		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release|Mixed Platforms.Build.0 = Release|x64
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release|Win32.ActiveCfg = Release|Win32
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release|Win32.Build.0 = Release|Win32
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release|x64.ActiveCfg = Release|x64
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release|x64.Build.0 = Release|x64
-		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{3BC6D760-91E8-4FFB-BD0E-F86F367AD8EA}.Release-CUDA|x64.Build.0 = Release|x64
-		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug|Win32.ActiveCfg = Debug|Win32
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug|Win32.Build.0 = Debug|Win32
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug|x64.ActiveCfg = Debug|x64
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug|x64.Build.0 = Debug|x64
-		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release|Mixed Platforms.Build.0 = Release|x64
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release|Win32.ActiveCfg = Release|Win32
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release|Win32.Build.0 = Release|Win32
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release|x64.ActiveCfg = Release|x64
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release|x64.Build.0 = Release|x64
-		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{51999D3E-EF22-4BDD-965F-4201034D3DCE}.Release-CUDA|x64.Build.0 = Release|x64
-		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug|Mixed Platforms.ActiveCfg = Debug|Any CPU
-		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug|Mixed Platforms.Build.0 = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug|Win32.ActiveCfg = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug|Win32.Build.0 = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug|x64.ActiveCfg = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug|x64.Build.0 = Debug|Any CPU
-		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|Any CPU
-		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|Win32.ActiveCfg = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|Win32.Build.0 = Debug|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|x64.ActiveCfg = Debug|Any CPU
-		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|Mixed Platforms.ActiveCfg = Release|Any CPU
+		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Debug-CUDA|x64.Build.0 = Debug|Any CPU
-		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|Mixed Platforms.Build.0 = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|Win32.ActiveCfg = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|Win32.Build.0 = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|x64.ActiveCfg = Release|Any CPU
-		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|Any CPU
+		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release|x64.Build.0 = Release|Any CPU
-		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|Mixed Platforms.Build.0 = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|Win32.ActiveCfg = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|Win32.Build.0 = Release|Any CPU
 		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|x64.ActiveCfg = Release|Any CPU
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|Mixed Platforms.ActiveCfg = Debug|Win32
+		{CAB55C39-8FA9-4912-98D9-E52669C8911D}.Release-CUDA|x64.Build.0 = Release|Any CPU
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|Mixed Platforms.Build.0 = Debug|Win32
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|Win32.ActiveCfg = Debug|Win32
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|Win32.Build.0 = Debug|Win32
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|x64.ActiveCfg = Debug|x64
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug|x64.Build.0 = Debug|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Win32.ActiveCfg = Debug-CUDA|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Win32.Build.0 = Debug-CUDA|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|x64.ActiveCfg = Debug-CUDA|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|Win32.Build.0 = Debug|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|x64.Build.0 = Debug-CUDA|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|x64.ActiveCfg = Debug|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|Mixed Platforms.ActiveCfg = Release|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|Mixed Platforms.Build.0 = Release|Win32
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|Win32.ActiveCfg = Release|Win32
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|Win32.Build.0 = Release|Win32
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|x64.ActiveCfg = Release|x64
 		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release|x64.Build.0 = Release|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Win32.ActiveCfg = Release-CUDA|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Mixed Platforms.Build.0 = Release|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Win32.Build.0 = Release-CUDA|Win32
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Win32.ActiveCfg = Release|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|x64.ActiveCfg = Release-CUDA|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|Win32.Build.0 = Release|Win32
+		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|x64.Build.0 = Release-CUDA|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|x64.ActiveCfg = Release|x64
-		{317B694E-B5C1-42A6-956F-FC12B69175A6}.Release-CUDA|x64.Build.0 = Release|x64
-		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|Mixed Platforms.ActiveCfg = Debug|Win32
-		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|Mixed Platforms.Build.0 = Debug|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|Win32.ActiveCfg = Debug|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|Win32.Build.0 = Debug|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|x64.ActiveCfg = Debug|x64
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug|x64.Build.0 = Debug|x64
-		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|Win32
-		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release|Mixed Platforms.Build.0 = Release|x64
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release|Win32.ActiveCfg = Release|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release|Win32.Build.0 = Release|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release|x64.ActiveCfg = Release|x64
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release|x64.Build.0 = Release|x64
-		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|Win32
-		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release-CUDA|Mixed Platforms.Build.0 = Release|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{7DCF280E-702B-49F3-84A7-AE7E146384D6}.Release-CUDA|x64.Build.0 = Release|x64
-		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug|Mixed Platforms.ActiveCfg = Debug|Win32
-		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug|Mixed Platforms.Build.0 = Debug|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug|Win32.ActiveCfg = Debug|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug|Win32.Build.0 = Debug|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug|x64.ActiveCfg = Debug|x64
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug|x64.Build.0 = Debug|x64
-		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|Win32
-		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{C33787E3-5564-4834-9FE3-A9020455A669}.Release|Mixed Platforms.ActiveCfg = Release|Win32
-		{C33787E3-5564-4834-9FE3-A9020455A669}.Release|Mixed Platforms.Build.0 = Release|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Release|Win32.ActiveCfg = Release|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Release|Win32.Build.0 = Release|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Release|x64.ActiveCfg = Release|x64
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Release|x64.Build.0 = Release|x64
-		{C33787E3-5564-4834-9FE3-A9020455A669}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|Win32
-		{C33787E3-5564-4834-9FE3-A9020455A669}.Release-CUDA|Mixed Platforms.Build.0 = Release|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{C33787E3-5564-4834-9FE3-A9020455A669}.Release-CUDA|x64.Build.0 = Release|x64
-		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug|Win32.ActiveCfg = Debug|Win32
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug|Win32.Build.0 = Debug|Win32
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug|x64.ActiveCfg = Debug|x64
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug|x64.Build.0 = Debug|x64
-		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release|Mixed Platforms.Build.0 = Release|x64
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release|Win32.ActiveCfg = Release|Win32
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release|Win32.Build.0 = Release|Win32
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release|x64.ActiveCfg = Release|x64
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release|x64.Build.0 = Release|x64
-		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{3DD3A43D-C6EA-460F-821B-6C339A03C5BB}.Release-CUDA|x64.Build.0 = Release|x64
-		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug|Win32.ActiveCfg = Debug|Win32
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug|Win32.Build.0 = Debug|Win32
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug|x64.ActiveCfg = Debug|x64
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug|x64.Build.0 = Debug|x64
-		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
-		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug-CUDA|Win32.Build.0 = Debug|Win32
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release|Mixed Platforms.Build.0 = Release|x64
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release|Win32.ActiveCfg = Release|Win32
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release|Win32.Build.0 = Release|Win32
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release|x64.ActiveCfg = Release|x64
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release|x64.Build.0 = Release|x64
-		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
-		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release-CUDA|Win32.ActiveCfg = Release|Win32
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release-CUDA|Win32.Build.0 = Release|Win32
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{CFB3FEAC-5720-4B16-9D7E-039DB180B641}.Release-CUDA|x64.Build.0 = Release|x64
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug|Mixed Platforms.ActiveCfg = Debug|x64
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug|Mixed Platforms.Build.0 = Debug|x64
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug|Win32.ActiveCfg = Debug|Win32
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug|Win32.Build.0 = Debug|Win32
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug|x64.ActiveCfg = Debug|x64
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug|x64.Build.0 = Debug|x64
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug-CUDA|Mixed Platforms.ActiveCfg = Debug|x64
+		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug-CUDA|Win32.ActiveCfg = Debug|Win32
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug-CUDA|Mixed Platforms.Build.0 = Debug|x64
+		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug-CUDA|Win32.Build.0 = Debug|Win32
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug-CUDA|Win32.ActiveCfg = Debug|x64
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug-CUDA|x64.ActiveCfg = Debug|x64
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Debug-CUDA|x64.Build.0 = Debug|x64
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release|Mixed Platforms.ActiveCfg = Release|x64
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release|Mixed Platforms.Build.0 = Release|x64
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release|Win32.ActiveCfg = Release|Win32
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release|Win32.Build.0 = Release|Win32
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release|x64.ActiveCfg = Release|x64
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release|x64.Build.0 = Release|x64
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release-CUDA|Mixed Platforms.ActiveCfg = Release|x64
+		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release-CUDA|Win32.ActiveCfg = Release|Win32
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release-CUDA|Mixed Platforms.Build.0 = Release|x64
+		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release-CUDA|Win32.Build.0 = Release|Win32
-		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release-CUDA|Win32.ActiveCfg = Release|x64
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release-CUDA|x64.ActiveCfg = Release|x64
 		{E493E368-A4CF-4A8D-99DD-E128CC3A27EF}.Release-CUDA|x64.Build.0 = Release|x64
 	EndGlobalSection

project/vc9/nvtt/nvtt.vcproj

@@ -916,138 +916,10 @@
 		<File
 			RelativePath="..\..\..\src\nvtt\CompressorDX11.cpp"
 			>
-			<FileConfiguration
-				Name="Debug-CUDA|Win32"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCLCompilerTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Debug-CUDA|x64"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCLCompilerTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Release-CUDA|Win32"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCLCompilerTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Release-CUDA|x64"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCLCompilerTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Debug|Win32"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCLCompilerTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Debug|x64"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCLCompilerTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Release|Win32"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCLCompilerTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Release|x64"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCLCompilerTool"
-				/>
-			</FileConfiguration>
 		</File>
 		<File
 			RelativePath="..\..\..\src\nvtt\CompressorDX11.h"
 			>
-			<FileConfiguration
-				Name="Debug-CUDA|Win32"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCustomBuildTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Debug-CUDA|x64"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCustomBuildTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Release-CUDA|Win32"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCustomBuildTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Release-CUDA|x64"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCustomBuildTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Debug|Win32"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCustomBuildTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Debug|x64"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCustomBuildTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Release|Win32"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCustomBuildTool"
-				/>
-			</FileConfiguration>
-			<FileConfiguration
-				Name="Release|x64"
-				ExcludedFromBuild="true"
-				>
-				<Tool
-					Name="VCCustomBuildTool"
-				/>
-			</FileConfiguration>
 		</File>
 		<File
 			RelativePath="..\..\..\src\nvtt\CompressorDX9.cpp"

project/vc9/testsuite/testsuite.vcproj

@@ -97,85 +97,6 @@
 				Name="VCPostBuildEventTool"
 			/>
 		</Configuration>
-		<Configuration
-			Name="Release|Win32"
-			OutputDirectory="$(ConfigurationName)\$(PlatformName)"
-			IntermediateDirectory="$(ConfigurationName)\$(PlatformName)"
-			ConfigurationType="1"
-			InheritedPropertySheets="$(SolutionDir)\nvtt.vsprops"
-			CharacterSet="1"
-			WholeProgramOptimization="1"
-			>
-			<Tool
-				Name="VCPreBuildEventTool"
-			/>
-			<Tool
-				Name="VCCustomBuildTool"
-			/>
-			<Tool
-				Name="VCXMLDataGeneratorTool"
-			/>
-			<Tool
-				Name="VCWebServiceProxyGeneratorTool"
-			/>
-			<Tool
-				Name="VCMIDLTool"
-			/>
-			<Tool
-				Name="VCCLCompilerTool"
-				AdditionalIncludeDirectories="..;..\..\..\src"
-				PreprocessorDefinitions="WIN32;NDEBUG;_CONSOLE"
-				StringPooling="true"
-				RuntimeLibrary="2"
-				EnableEnhancedInstructionSet="2"
-				UsePrecompiledHeader="0"
-				WarningLevel="3"
-			/>
-			<Tool
-				Name="VCManagedResourceCompilerTool"
-			/>
-			<Tool
-				Name="VCResourceCompilerTool"
-			/>
-			<Tool
-				Name="VCPreLinkEventTool"
-			/>
-			<Tool
-				Name="VCLinkerTool"
-				AdditionalDependencies="FreeImage.lib jpeg.lib libpng.lib tiff.lib zlib.lib"
-				OutputFile="$(SolutionDir)\$(ConfigurationName).$(PlatformName)\bin\$(ProjectName).exe"
-				LinkIncremental="1"
-				AdditionalLibraryDirectories="$(GnuWinDir)\lib;$(FreeImageDir)"
-				GenerateDebugInformation="true"
-				SubSystem="1"
-				OptimizeReferences="2"
-				EnableCOMDATFolding="2"
-				RandomizedBaseAddress="1"
-				DataExecutionPrevention="0"
-				TargetMachine="1"
-			/>
-			<Tool
-				Name="VCALinkTool"
-			/>
-			<Tool
-				Name="VCManifestTool"
-			/>
-			<Tool
-				Name="VCXDCMakeTool"
-			/>
-			<Tool
-				Name="VCBscMakeTool"
-			/>
-			<Tool
-				Name="VCFxCopTool"
-			/>
-			<Tool
-				Name="VCAppVerifierTool"
-			/>
-			<Tool
-				Name="VCPostBuildEventTool"
-			/>
-		</Configuration>
 		<Configuration
 			Name="Debug|x64"
 			OutputDirectory="$(PlatformName)\$(ConfigurationName)"
@@ -255,6 +176,85 @@
 				Name="VCPostBuildEventTool"
 			/>
 		</Configuration>
+		<Configuration
+			Name="Release|Win32"
+			OutputDirectory="$(ConfigurationName)\$(PlatformName)"
+			IntermediateDirectory="$(ConfigurationName)\$(PlatformName)"
+			ConfigurationType="1"
+			InheritedPropertySheets="$(SolutionDir)\nvtt.vsprops"
+			CharacterSet="1"
+			WholeProgramOptimization="1"
+			>
+			<Tool
+				Name="VCPreBuildEventTool"
+			/>
+			<Tool
+				Name="VCCustomBuildTool"
+			/>
+			<Tool
+				Name="VCXMLDataGeneratorTool"
+			/>
+			<Tool
+				Name="VCWebServiceProxyGeneratorTool"
+			/>
+			<Tool
+				Name="VCMIDLTool"
+			/>
+			<Tool
+				Name="VCCLCompilerTool"
+				AdditionalIncludeDirectories="..;..\..\..\src"
+				PreprocessorDefinitions="WIN32;NDEBUG;_CONSOLE"
+				StringPooling="true"
+				RuntimeLibrary="2"
+				EnableEnhancedInstructionSet="2"
+				UsePrecompiledHeader="0"
+				WarningLevel="3"
+			/>
+			<Tool
+				Name="VCManagedResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCPreLinkEventTool"
+			/>
+			<Tool
+				Name="VCLinkerTool"
+				AdditionalDependencies="FreeImage.lib jpeg.lib libpng.lib tiff.lib zlib.lib"
+				OutputFile="$(SolutionDir)\$(ConfigurationName).$(PlatformName)\bin\$(ProjectName).exe"
+				LinkIncremental="1"
+				AdditionalLibraryDirectories="$(GnuWinDir)\lib;$(FreeImageDir)"
+				GenerateDebugInformation="true"
+				SubSystem="1"
+				OptimizeReferences="2"
+				EnableCOMDATFolding="2"
+				RandomizedBaseAddress="1"
+				DataExecutionPrevention="0"
+				TargetMachine="1"
+			/>
+			<Tool
+				Name="VCALinkTool"
+			/>
+			<Tool
+				Name="VCManifestTool"
+			/>
+			<Tool
+				Name="VCXDCMakeTool"
+			/>
+			<Tool
+				Name="VCBscMakeTool"
+			/>
+			<Tool
+				Name="VCFxCopTool"
+			/>
+			<Tool
+				Name="VCAppVerifierTool"
+			/>
+			<Tool
+				Name="VCPostBuildEventTool"
+			/>
+		</Configuration>
 		<Configuration
 			Name="Release|x64"
 			OutputDirectory="$(PlatformName)\$(ConfigurationName)"
@@ -335,6 +335,322 @@
 				Name="VCPostBuildEventTool"
 			/>
 		</Configuration>
+		<Configuration
+			Name="Debug-CUDA|Win32"
+			OutputDirectory="$(ConfigurationName)"
+			IntermediateDirectory="$(ConfigurationName)"
+			ConfigurationType="1"
+			InheritedPropertySheets="$(SolutionDir)\nvtt.vsprops"
+			CharacterSet="1"
+			>
+			<Tool
+				Name="VCPreBuildEventTool"
+			/>
+			<Tool
+				Name="VCCustomBuildTool"
+			/>
+			<Tool
+				Name="VCXMLDataGeneratorTool"
+			/>
+			<Tool
+				Name="VCWebServiceProxyGeneratorTool"
+			/>
+			<Tool
+				Name="VCMIDLTool"
+			/>
+			<Tool
+				Name="VCCLCompilerTool"
+				Optimization="0"
+				AdditionalIncludeDirectories="..;..\..\..\src"
+				PreprocessorDefinitions="WIN32;_DEBUG;_CONSOLE"
+				MinimalRebuild="true"
+				BasicRuntimeChecks="3"
+				RuntimeLibrary="3"
+				UsePrecompiledHeader="0"
+				WarningLevel="3"
+				DebugInformationFormat="3"
+			/>
+			<Tool
+				Name="VCManagedResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCPreLinkEventTool"
+			/>
+			<Tool
+				Name="VCLinkerTool"
+				AdditionalDependencies="FreeImage.lib jpeg.lib libpng.lib tiff.lib zlib.lib"
+				OutputFile="$(SolutionDir)\$(ConfigurationName).$(PlatformName)\bin\$(ProjectName).exe"
+				LinkIncremental="2"
+				AdditionalLibraryDirectories="$(GnuWinDir)\lib;$(FreeImageDir)"
+				GenerateDebugInformation="true"
+				SubSystem="1"
+				RandomizedBaseAddress="1"
+				DataExecutionPrevention="0"
+				TargetMachine="1"
+			/>
+			<Tool
+				Name="VCALinkTool"
+			/>
+			<Tool
+				Name="VCManifestTool"
+			/>
+			<Tool
+				Name="VCXDCMakeTool"
+			/>
+			<Tool
+				Name="VCBscMakeTool"
+			/>
+			<Tool
+				Name="VCFxCopTool"
+			/>
+			<Tool
+				Name="VCAppVerifierTool"
+			/>
+			<Tool
+				Name="VCPostBuildEventTool"
+			/>
+		</Configuration>
+		<Configuration
+			Name="Debug-CUDA|x64"
+			OutputDirectory="$(PlatformName)\$(ConfigurationName)"
+			IntermediateDirectory="$(PlatformName)\$(ConfigurationName)"
+			ConfigurationType="1"
+			InheritedPropertySheets="$(SolutionDir)\nvtt.vsprops"
+			CharacterSet="1"
+			>
+			<Tool
+				Name="VCPreBuildEventTool"
+			/>
+			<Tool
+				Name="VCCustomBuildTool"
+			/>
+			<Tool
+				Name="VCXMLDataGeneratorTool"
+			/>
+			<Tool
+				Name="VCWebServiceProxyGeneratorTool"
+			/>
+			<Tool
+				Name="VCMIDLTool"
+				TargetEnvironment="3"
+			/>
+			<Tool
+				Name="VCCLCompilerTool"
+				Optimization="0"
+				AdditionalIncludeDirectories="..;..\..\..\src"
+				PreprocessorDefinitions="WIN32;_DEBUG;_CONSOLE"
+				MinimalRebuild="true"
+				BasicRuntimeChecks="3"
+				RuntimeLibrary="3"
+				UsePrecompiledHeader="0"
+				WarningLevel="3"
+				DebugInformationFormat="3"
+			/>
+			<Tool
+				Name="VCManagedResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCPreLinkEventTool"
+			/>
+			<Tool
+				Name="VCLinkerTool"
+				AdditionalDependencies="FreeImage.lib jpeg.lib libpng.lib tiff.lib zlib.lib"
+				OutputFile="$(SolutionDir)\$(ConfigurationName).$(PlatformName)\bin\$(ProjectName).exe"
+				LinkIncremental="2"
+				AdditionalLibraryDirectories="$(GnuWinDir)\lib;$(FreeImageDir)"
+				GenerateDebugInformation="true"
+				SubSystem="1"
+				RandomizedBaseAddress="1"
+				DataExecutionPrevention="0"
+				TargetMachine="17"
+			/>
+			<Tool
+				Name="VCALinkTool"
+			/>
+			<Tool
+				Name="VCManifestTool"
+			/>
+			<Tool
+				Name="VCXDCMakeTool"
+			/>
+			<Tool
+				Name="VCBscMakeTool"
+			/>
+			<Tool
+				Name="VCFxCopTool"
+			/>
+			<Tool
+				Name="VCAppVerifierTool"
+			/>
+			<Tool
+				Name="VCPostBuildEventTool"
+			/>
+		</Configuration>
+		<Configuration
+			Name="Release-CUDA|Win32"
+			OutputDirectory="$(ConfigurationName)"
+			IntermediateDirectory="$(ConfigurationName)"
+			ConfigurationType="1"
+			InheritedPropertySheets="$(SolutionDir)\nvtt.vsprops"
+			CharacterSet="1"
+			WholeProgramOptimization="1"
+			>
+			<Tool
+				Name="VCPreBuildEventTool"
+			/>
+			<Tool
+				Name="VCCustomBuildTool"
+			/>
+			<Tool
+				Name="VCXMLDataGeneratorTool"
+			/>
+			<Tool
+				Name="VCWebServiceProxyGeneratorTool"
+			/>
+			<Tool
+				Name="VCMIDLTool"
+			/>
+			<Tool
+				Name="VCCLCompilerTool"
+				AdditionalIncludeDirectories="..;..\..\..\src"
+				PreprocessorDefinitions="WIN32;NDEBUG;_CONSOLE"
+				StringPooling="true"
+				RuntimeLibrary="2"
+				EnableEnhancedInstructionSet="2"
+				UsePrecompiledHeader="0"
+				WarningLevel="3"
+			/>
+			<Tool
+				Name="VCManagedResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCPreLinkEventTool"
+			/>
+			<Tool
+				Name="VCLinkerTool"
+				AdditionalDependencies="FreeImage.lib jpeg.lib libpng.lib tiff.lib zlib.lib"
+				OutputFile="$(SolutionDir)\$(ConfigurationName).$(PlatformName)\bin\$(ProjectName).exe"
+				LinkIncremental="1"
+				AdditionalLibraryDirectories="$(GnuWinDir)\lib;$(FreeImageDir)"
+				GenerateDebugInformation="true"
+				SubSystem="1"
+				OptimizeReferences="2"
+				EnableCOMDATFolding="2"
+				RandomizedBaseAddress="1"
+				DataExecutionPrevention="0"
+				TargetMachine="1"
+			/>
+			<Tool
+				Name="VCALinkTool"
+			/>
+			<Tool
+				Name="VCManifestTool"
+			/>
+			<Tool
+				Name="VCXDCMakeTool"
+			/>
+			<Tool
+				Name="VCBscMakeTool"
+			/>
+			<Tool
+				Name="VCFxCopTool"
+			/>
+			<Tool
+				Name="VCAppVerifierTool"
+			/>
+			<Tool
+				Name="VCPostBuildEventTool"
+			/>
+		</Configuration>
+		<Configuration
+			Name="Release-CUDA|x64"
+			OutputDirectory="$(PlatformName)\$(ConfigurationName)"
+			IntermediateDirectory="$(PlatformName)\$(ConfigurationName)"
+			ConfigurationType="1"
+			InheritedPropertySheets="$(SolutionDir)\nvtt.vsprops"
+			CharacterSet="1"
+			WholeProgramOptimization="1"
+			>
+			<Tool
+				Name="VCPreBuildEventTool"
+			/>
+			<Tool
+				Name="VCCustomBuildTool"
+			/>
+			<Tool
+				Name="VCXMLDataGeneratorTool"
+			/>
+			<Tool
+				Name="VCWebServiceProxyGeneratorTool"
+			/>
+			<Tool
+				Name="VCMIDLTool"
+				TargetEnvironment="3"
+			/>
+			<Tool
+				Name="VCCLCompilerTool"
+				AdditionalIncludeDirectories="..;..\..\..\src"
+				PreprocessorDefinitions="WIN32;NDEBUG;_CONSOLE"
+				StringPooling="true"
+				RuntimeLibrary="2"
+				EnableEnhancedInstructionSet="2"
+				UsePrecompiledHeader="0"
+				WarningLevel="3"
+			/>
+			<Tool
+				Name="VCManagedResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCResourceCompilerTool"
+			/>
+			<Tool
+				Name="VCPreLinkEventTool"
+			/>
+			<Tool
+				Name="VCLinkerTool"
+				AdditionalDependencies="FreeImage.lib jpeg.lib libpng.lib tiff.lib zlib.lib"
+				OutputFile="$(SolutionDir)\$(ConfigurationName).$(PlatformName)\bin\$(ProjectName).exe"
+				LinkIncremental="1"
+				AdditionalLibraryDirectories="$(GnuWinDir)\lib;$(FreeImageDir)"
+				GenerateDebugInformation="true"
+				SubSystem="1"
+				OptimizeReferences="2"
+				EnableCOMDATFolding="2"
+				RandomizedBaseAddress="1"
+				DataExecutionPrevention="0"
+				TargetMachine="17"
+			/>
+			<Tool
+				Name="VCALinkTool"
+			/>
+			<Tool
+				Name="VCManifestTool"
+			/>
+			<Tool
+				Name="VCXDCMakeTool"
+			/>
+			<Tool
+				Name="VCBscMakeTool"
+			/>
+			<Tool
+				Name="VCFxCopTool"
+			/>
+			<Tool
+				Name="VCAppVerifierTool"
+			/>
+			<Tool
+				Name="VCPostBuildEventTool"
+			/>
+		</Configuration>
 	</Configurations>
 	<References>
 	</References>

src/nvimage/BlockDXT.cpp

@@ -26,6 +26,10 @@
 
 #include "nvcore/Stream.h"
 #include "nvcore/Utils.h" // swap
+#include "nvmath/Half.h"
+
+#include "nvtt/bc6h/zoh.h"
+#include "nvtt/bc6h/utils.h"
 
 
 using namespace nv;
@@ -610,6 +614,33 @@ void BlockCTX1::setIndices(int * idx)
 }
 
 
+/// Decode BC6 block.
+void BlockBC6::decodeBlock(ColorSet * set) const
+{
+	Tile tile(4, 4);
+	ZOH::decompress((const char *)data, tile);
+
+	// Convert ZOH's tile struct back to NVTT's, and convert half to float.
+	set->allocate(4, 4);
+	for (uint y = 0; y < 4; ++y)
+	{
+		for (uint x = 0; x < 4; ++x)
+		{
+			uint16 rHalf = Tile::float2half(tile.data[y][x].x);
+			uint16 gHalf = Tile::float2half(tile.data[y][x].y);
+			uint16 bHalf = Tile::float2half(tile.data[y][x].z);
+			set->colors[y * 4 + x].x = to_float(rHalf);
+			set->colors[y * 4 + x].y = to_float(gHalf);
+			set->colors[y * 4 + x].z = to_float(bHalf);
+			set->colors[y * 4 + x].w = 1.0f;
+
+			// Set indices in case someone uses them
+			set->indices[y * 4 + x] = y * 4 + x;
+		}
+	}
+}
+
+
 /// Flip CTX1 block vertically.
 inline void BlockCTX1::flip4()
 {
@@ -671,3 +702,8 @@ Stream & nv::operator<<(Stream & stream, BlockCTX1 & block)
     return stream;
 }
 
+Stream & nv::operator<<(Stream & stream, BlockBC6 & block)
+{
+    stream.serialize(&block, sizeof(block));
+    return stream;
+}

src/nvimage/BlockDXT.h

@@ -32,6 +32,7 @@
 namespace nv
 {
     struct ColorBlock;
+	struct ColorSet;
     class Stream;
 
 
@@ -212,6 +213,15 @@ namespace nv
         void flip2();
     };
 
+	/// BC6 block.
+	struct BlockBC6
+	{
+		uint8 data[16];		// Not even going to try to write a union for this thing.
+		void decodeBlock(ColorSet * set) const;
+	};
+
+	/// !!!UNDONE: BC7 block
+
 
     // Serialization functions.
     NVIMAGE_API Stream & operator<<(Stream & stream, BlockDXT1 & block);
@@ -222,6 +232,7 @@ namespace nv
     NVIMAGE_API Stream & operator<<(Stream & stream, BlockATI1 & block);
     NVIMAGE_API Stream & operator<<(Stream & stream, BlockATI2 & block);
     NVIMAGE_API Stream & operator<<(Stream & stream, BlockCTX1 & block);
+    NVIMAGE_API Stream & operator<<(Stream & stream, BlockBC6 & block);
 
 } // nv namespace
 

src/nvimage/DirectDrawSurface.cpp

@@ -950,7 +950,8 @@ bool DirectDrawSurface::isSupported() const
             header.header10.dxgiFormat == DXGI_FORMAT_BC2_UNORM ||
             header.header10.dxgiFormat == DXGI_FORMAT_BC3_UNORM ||
             header.header10.dxgiFormat == DXGI_FORMAT_BC4_UNORM ||
-            header.header10.dxgiFormat == DXGI_FORMAT_BC5_UNORM)
+            header.header10.dxgiFormat == DXGI_FORMAT_BC5_UNORM ||
+			header.header10.dxgiFormat == DXGI_FORMAT_BC6H_UF16)
         {
             return true;
         }
@@ -1340,13 +1341,12 @@ void DirectDrawSurface::readBlock(ColorBlock * rgba)
     if (header.hasDX10Header())
     {
         if (header.header10.dxgiFormat == DXGI_FORMAT_BC1_UNORM) fourcc = FOURCC_DXT1;
-        if (header.header10.dxgiFormat == DXGI_FORMAT_BC2_UNORM) fourcc = FOURCC_DXT3;
+        else if (header.header10.dxgiFormat == DXGI_FORMAT_BC2_UNORM) fourcc = FOURCC_DXT3;
-        if (header.header10.dxgiFormat == DXGI_FORMAT_BC3_UNORM) fourcc = FOURCC_DXT5;
+        else if (header.header10.dxgiFormat == DXGI_FORMAT_BC3_UNORM) fourcc = FOURCC_DXT5;
-        if (header.header10.dxgiFormat == DXGI_FORMAT_BC4_UNORM) fourcc = FOURCC_ATI1;
+        else if (header.header10.dxgiFormat == DXGI_FORMAT_BC4_UNORM) fourcc = FOURCC_ATI1;
-        if (header.header10.dxgiFormat == DXGI_FORMAT_BC5_UNORM) fourcc = FOURCC_ATI2;
+        else if (header.header10.dxgiFormat == DXGI_FORMAT_BC5_UNORM) fourcc = FOURCC_ATI2;
     }
 
-
     if (fourcc == FOURCC_DXT1)
     {
         BlockDXT1 block;
@@ -1389,6 +1389,31 @@ void DirectDrawSurface::readBlock(ColorBlock * rgba)
         *stream << block;
         block.decodeBlock(rgba);
     }
+	else if (header.hasDX10Header() && header.header10.dxgiFormat == DXGI_FORMAT_BC6H_UF16)
+	{
+		BlockBC6 block;
+		*stream << block;
+		ColorSet set;
+		block.decodeBlock(&set);
+
+		// Clamp to [0, 1] and round to 8-bit
+		for (int y = 0; y < 4; ++y)
+		{
+			for (int x = 0; x < 4; ++x)
+			{
+				Vector4 px = set.colors[y*4 + x];
+				rgba->color(x, y).setRGBA(
+									uint8(clamp(px.x, 0.0f, 1.0f) * 255.0f + 0.5f),
+									uint8(clamp(px.y, 0.0f, 1.0f) * 255.0f + 0.5f),
+									uint8(clamp(px.z, 0.0f, 1.0f) * 255.0f + 0.5f),
+									uint8(clamp(px.w, 0.0f, 1.0f) * 255.0f + 0.5f));
+			}
+		}
+	}
+	else
+	{
+		nvDebugCheck(false);
+	}
 
     // If normal flag set, convert to normal.
     if (header.pf.flags & DDPF_NORMAL)

src/nvmath/Fitting.cpp

@@ -13,7 +13,7 @@ using namespace nv;
 // @@ Move to EigenSolver.h
 
 // @@ We should be able to do something cheaper...
-static Vector3 estimatePrincipleComponent(const float * __restrict matrix)
+static Vector3 estimatePrincipalComponent(const float * __restrict matrix)
 {
 	const Vector3 row0(matrix[0], matrix[1], matrix[2]);
 	const Vector3 row1(matrix[1], matrix[3], matrix[4]);
@@ -36,7 +36,7 @@ static inline Vector3 firstEigenVector_PowerMethod(const float *__restrict matri
         return Vector3(0.0f);
     }
 
-    Vector3 v = estimatePrincipleComponent(matrix);
+    Vector3 v = estimatePrincipalComponent(matrix);
 
     const int NUM = 8;
     for (int i = 0; i < NUM; i++)
@@ -136,7 +136,7 @@ Vector3 nv::Fit::computeCovariance(int n, const Vector3 *__restrict points, cons
     return centroid;
 }
 
-Vector3 nv::Fit::computePrincipalComponent(int n, const Vector3 *__restrict points)
+Vector3 nv::Fit::computePrincipalComponent_PowerMethod(int n, const Vector3 *__restrict points)
 {
     float matrix[6];
     computeCovariance(n, points, matrix);
@@ -144,7 +144,7 @@ Vector3 nv::Fit::computePrincipalComponent(int n, const Vector3 *__restrict poin
     return firstEigenVector_PowerMethod(matrix);
 }
 
-Vector3 nv::Fit::computePrincipalComponent(int n, const Vector3 *__restrict points, const float *__restrict weights, Vector3::Arg metric)
+Vector3 nv::Fit::computePrincipalComponent_PowerMethod(int n, const Vector3 *__restrict points, const float *__restrict weights, Vector3::Arg metric)
 {
     float matrix[6];
     computeCovariance(n, points, weights, metric, matrix);
@@ -153,6 +153,42 @@ Vector3 nv::Fit::computePrincipalComponent(int n, const Vector3 *__restrict poin
 }
 
 
+
+static inline Vector3 firstEigenVector_EigenSolver(const float *__restrict matrix)
+{
+    if (matrix[0] == 0 && matrix[3] == 0 && matrix[5] == 0)
+    {
+        return Vector3(0.0f);
+    }
+
+    float eigenValues[3];
+    Vector3 eigenVectors[3];
+	if (!nv::Fit::eigenSolveSymmetric(matrix, eigenValues, eigenVectors))
+	{
+		return Vector3(0.0f);
+	}
+
+	return eigenVectors[0];
+}
+
+Vector3 nv::Fit::computePrincipalComponent_EigenSolver(int n, const Vector3 *__restrict points)
+{
+    float matrix[6];
+    computeCovariance(n, points, matrix);
+
+    return firstEigenVector_EigenSolver(matrix);
+}
+
+Vector3 nv::Fit::computePrincipalComponent_EigenSolver(int n, const Vector3 *__restrict points, const float *__restrict weights, Vector3::Arg metric)
+{
+    float matrix[6];
+    computeCovariance(n, points, weights, metric, matrix);
+
+    return firstEigenVector_EigenSolver(matrix);
+}
+
+
+
 Plane nv::Fit::bestPlane(int n, const Vector3 *__restrict points)
 {
     // compute the centroid and covariance
@@ -199,7 +235,7 @@ bool nv::Fit::isPlanar(int n, const Vector3 * points, float epsilon/*=NV_EPSILON
 static void EigenSolver_Tridiagonal(double mat[3][3],double * diag,double * subd);
 static bool EigenSolver_QLAlgorithm(double mat[3][3],double * diag,double * subd);
 
-bool nv::Fit::eigenSolveSymmetric(float matrix[6], float eigenValues[3], Vector3 eigenVectors[3])
+bool nv::Fit::eigenSolveSymmetric(const float matrix[6], float eigenValues[3], Vector3 eigenVectors[3])
 {
     nvDebugCheck(matrix != NULL && eigenValues != NULL && eigenVectors != NULL);
 

src/nvmath/Fitting.h

@@ -17,13 +17,16 @@ namespace nv
         Vector3 computeCovariance(int n, const Vector3 * points, float * covariance);
         Vector3 computeCovariance(int n, const Vector3 * points, const float * weights, const Vector3 & metric, float * covariance);
 
-        Vector3 computePrincipalComponent(int n, const Vector3 * points);
+        Vector3 computePrincipalComponent_PowerMethod(int n, const Vector3 * points);
-        Vector3 computePrincipalComponent(int n, const Vector3 * points, const float * weights, const Vector3 & metric);
+        Vector3 computePrincipalComponent_PowerMethod(int n, const Vector3 * points, const float * weights, const Vector3 & metric);
+
+        Vector3 computePrincipalComponent_EigenSolver(int n, const Vector3 * points);
+        Vector3 computePrincipalComponent_EigenSolver(int n, const Vector3 * points, const float * weights, const Vector3 & metric);
 
         Plane bestPlane(int n, const Vector3 * points);
         bool isPlanar(int n, const Vector3 * points, float epsilon = NV_EPSILON);
 
-        bool eigenSolveSymmetric (float matrix[6], float eigenValues[3], Vector3 eigenVectors[3]);
+        bool eigenSolveSymmetric (const float matrix[6], float eigenValues[3], Vector3 eigenVectors[3]);
 
 
         // Returns number of clusters [1-4].

src/nvmath/Half.h

@@ -23,7 +23,9 @@ namespace nv {
     // http://www.fox-toolkit.org/ftp/fasthalffloatconversion.pdf
     inline uint32 fast_half_to_float(uint16 h)
     {
-        nvDebugCheck(mantissa_table[0] == 0); // Make sure table was initialized.
+		// Initialize table if necessary.
+		if (mantissa_table[0] != 0)
+			half_init_tables();
 	    uint exp = h >> 10;
 	    return mantissa_table[offset_table[exp] + (h & 0x3ff)] + exponent_table[exp];
     }

src/nvtt/ClusterFit.cpp

@@ -57,14 +57,14 @@ void ClusterFit::setColourSet(const ColorSet * set)
         values[i] = set->colors[i].xyz();
     }
 
-    Vector3 principle = Fit::computePrincipalComponent(m_count, values, set->weights, metric);
+    Vector3 principal = Fit::computePrincipalComponent_PowerMethod(m_count, values, set->weights, metric);
 
     // build the list of values
     int order[16];
     float dps[16];
     for (uint i = 0; i < m_count; ++i)
     {
-        dps[i] = dot(values[i], principle);
+        dps[i] = dot(values[i], principal);
         order[i] = i;
     }
 

src/nvtt/CompressorDX11.cpp

@@ -24,8 +24,11 @@
 
 #include "CompressorDX11.h"
 
+#include <cstring>
 #include "nvtt.h"
 #include "CompressionOptions.h"
+#include "nvimage/ColorBlock.h"
+#include "nvmath/Half.h"
 
 #include "bc6h/zoh.h"
 #include "bc6h/utils.h"
@@ -52,7 +55,26 @@ void CompressorBC6::compressBlock(ColorSet & tile, AlphaMode alphaMode, const Co
         Utils::FORMAT = SIGNED_F16;
     }
 
-    ZOH::compress(tile, (char *)output);
+	// Convert NVTT's tile struct to ZOH's, and convert float to half.
+	Tile zohTile(tile.w, tile.h);
+	memset(zohTile.data, 0, sizeof(zohTile.data));
+	memset(zohTile.importance_map, 0, sizeof(zohTile.importance_map));
+	for (uint y = 0; y < tile.h; ++y)
+	{
+		for (uint x = 0; x < tile.w; ++x)
+		{
+			Vector3 color = tile.color(x, y).xyz();
+			uint16 rHalf = to_half(color.x);
+			uint16 gHalf = to_half(color.y);
+			uint16 bHalf = to_half(color.z);
+			zohTile.data[y][x].x = Tile::half2float(rHalf);
+			zohTile.data[y][x].y = Tile::half2float(gHalf);
+			zohTile.data[y][x].z = Tile::half2float(bHalf);
+			zohTile.importance_map[y][x] = 1.0f;
+		}
+	}
+
+    ZOH::compress(zohTile, (char *)output);
 }
 
 

src/nvtt/Context.cpp

@@ -485,32 +485,32 @@ bool Compressor::Private::outputHeader(nvtt::TextureType textureType, int w, int
             else
             {
                 if (compressionOptions.format == Format_DXT1 || compressionOptions.format == Format_DXT1a || compressionOptions.format == Format_DXT1n) {
-                    header.setDX10Format(70); // DXGI_FORMAT_BC1_TYPELESS
+                    header.setDX10Format(DXGI_FORMAT_BC1_UNORM);
                     if (compressionOptions.format == Format_DXT1a) header.setHasAlphaFlag(true);
                     if (isNormalMap) header.setNormalFlag(true);
                 }
                 else if (compressionOptions.format == Format_DXT3) {
-                    header.setDX10Format(73); // DXGI_FORMAT_BC2_TYPELESS
+                    header.setDX10Format(DXGI_FORMAT_BC2_UNORM);
                 }
                 else if (compressionOptions.format == Format_DXT5) {
-                    header.setDX10Format(76); // DXGI_FORMAT_BC3_TYPELESS
+                    header.setDX10Format(DXGI_FORMAT_BC3_UNORM);
                 }
                 else if (compressionOptions.format == Format_DXT5n) {
-                    header.setDX10Format(76); // DXGI_FORMAT_BC3_TYPELESS
+                    header.setDX10Format(DXGI_FORMAT_BC3_UNORM);
                     if (isNormalMap) header.setNormalFlag(true);
                 }
                 else if (compressionOptions.format == Format_BC4) {
-                    header.setDX10Format(79); // DXGI_FORMAT_BC4_TYPELESS
+                    header.setDX10Format(DXGI_FORMAT_BC4_UNORM);
                 }
                 else if (compressionOptions.format == Format_BC5) {
-                    header.setDX10Format(82); // DXGI_FORMAT_BC5_TYPELESS
+                    header.setDX10Format(DXGI_FORMAT_BC5_UNORM);
                     if (isNormalMap) header.setNormalFlag(true);
                 }
                 else if (compressionOptions.format == Format_BC6) {
-                    header.setDX10Format(94); // DXGI_FORMAT_BC6H_TYPELESS
+                    header.setDX10Format(DXGI_FORMAT_BC6H_UF16);
                 }
                 else if (compressionOptions.format == Format_BC7) {
-                    header.setDX10Format(97); // DXGI_FORMAT_BC7_TYPELESS
+                    header.setDX10Format(DXGI_FORMAT_BC7_UNORM);
                     if (isNormalMap) header.setNormalFlag(true);
                 }
                 else {
@@ -771,11 +771,12 @@ CompressorInterface * Compressor::Private::chooseCpuCompressor(const Compression
     }
     else if (compressionOptions.format == Format_BC6)
     {
-        // Not supported.
+        return new CompressorBC6;
     }
     else if (compressionOptions.format == Format_BC7)
     {
-        // Not supported.
+		// !!!UNDONE
+        //return new CompressorBC7;
     }
 
     return NULL;

src/nvtt/Surface.cpp

@@ -702,7 +702,12 @@ bool Surface::setImage(InputFormat format, int w, int h, int d, const void * r,
 // @@ Add support for compressed 3D textures.
 bool Surface::setImage2D(Format format, Decoder decoder, int w, int h, const void * data)
 {
-    if (format != nvtt::Format_BC1 && format != nvtt::Format_BC2 && format != nvtt::Format_BC3 && format != nvtt::Format_BC4 && format != nvtt::Format_BC5)
+    if (format != nvtt::Format_BC1 &&
+		format != nvtt::Format_BC2 &&
+		format != nvtt::Format_BC3 &&
+		format != nvtt::Format_BC4 &&
+		format != nvtt::Format_BC5 &&
+		format != nvtt::Format_BC6)
     {
         return false;
     }
@@ -723,6 +728,42 @@ bool Surface::setImage2D(Format format, Decoder decoder, int w, int h, const voi
     const uint8 * ptr = (const uint8 *)data;
 
     TRY {
+		if (format == nvtt::Format_BC6)
+		{
+			// BC6 format - decode directly to float
+
+			for (int y = 0; y < bh; y++)
+			{
+				for (int x = 0; x < bw; x++)
+				{
+					ColorSet colors;
+					const BlockBC6 * block = (const BlockBC6 *)ptr;
+					block->decodeBlock(&colors);
+
+					for (int yy = 0; yy < 4; yy++)
+					{
+						for (int xx = 0; xx < 4; xx++)
+						{
+							Vector4 rgba = colors.colors[yy*4 + xx];
+
+							if (x * 4 + xx < w && y * 4 + yy < h)
+							{
+								m->image->pixel(0, x*4 + xx, y*4 + yy, 0) = rgba.x;
+								m->image->pixel(1, x*4 + xx, y*4 + yy, 0) = rgba.y;
+								m->image->pixel(2, x*4 + xx, y*4 + yy, 0) = rgba.z;
+								m->image->pixel(3, x*4 + xx, y*4 + yy, 0) = rgba.w;
+							}
+						}
+					}
+
+					ptr += bs;
+				}
+			}
+		}
+		else
+		{
+			// Non-BC6 - decode to 8-bit, then convert to float
+
 			for (int y = 0; y < bh; y++)
 			{
 				for (int x = 0; x < bw; x++)
@@ -781,6 +822,10 @@ bool Surface::setImage2D(Format format, Decoder decoder, int w, int h, const voi
 						const BlockATI2 * block = (const BlockATI2 *)ptr;
 						block->decodeBlock(&colors, decoder == Decoder_D3D9);
 					}
+					else
+					{
+						nvDebugCheck(false);
+					}
 
 					for (int yy = 0; yy < 4; yy++)
 					{
@@ -802,6 +847,7 @@ bool Surface::setImage2D(Format format, Decoder decoder, int w, int h, const voi
 				}
 			}
 		}
+    }
     CATCH {
         return false;
     }
@@ -1455,7 +1501,7 @@ void Surface::scaleAlphaToCoverage(float coverage, float alphaRef/*= 0.5f*/)
 
 // Ideally you should compress/quantize the RGB and M portions independently.
 // Once you have M quantized, you would compute the corresponding RGB and quantize that.
-void Surface::toRGBM(float range/*= 1*/, float threshold/*= 0.25*/)
+void Surface::toRGBM(float range/*= 1*/, float threshold/*= 0.0f*/)
 {
     if (isNull()) return;
 
@@ -1471,17 +1517,17 @@ void Surface::toRGBM(float range/*= 1*/, float threshold/*= 0.25*/)
 
     const uint count = img->pixelCount();
     for (uint i = 0; i < count; i++) {
-        float R = nv::clamp(r[i], 0.0f, 1.0f);
+        float R = r[i];
-        float G = nv::clamp(g[i], 0.0f, 1.0f);
+        float G = g[i];
-        float B = nv::clamp(b[i], 0.0f, 1.0f);
+        float B = b[i];
 #if 1
-        float M = max(max(R, G), max(B, threshold));
+        float M = nv::clamp(max(max(R, G), max(B, threshold)), 0.0f, range);
 
-        r[i] = R / M;
+        r[i] = nv::clamp(R / M, 0.0f, 1.0f);
-        g[i] = G / M;
+        g[i] = nv::clamp(G / M, 0.0f, 1.0f);
-        b[i] = B / M;
+        b[i] = nv::clamp(B / M, 0.0f, 1.0f);
 
-        a[i] = (M - threshold) / (1 - threshold);
+        a[i] = (M - threshold) / (range - threshold);
 
 #else
 
@@ -1524,12 +1570,14 @@ void Surface::toRGBM(float range/*= 1*/, float threshold/*= 0.25*/)
 }
 
 
-void Surface::fromRGBM(float range/*= 1*/)
+void Surface::fromRGBM(float range/*= 1*/, float threshold/*= 0.0*/)
 {
     if (isNull()) return;
 
     detach();
 
+    threshold = ::clamp(threshold, 1e-6f, 1.0f);
+
 	FloatImage * img = m->image;
     float * r = img->channel(0);
     float * g = img->channel(1);
@@ -1538,7 +1586,7 @@ void Surface::fromRGBM(float range/*= 1*/)
 
     const uint count = img->pixelCount();
     for (uint i = 0; i < count; i++) {
-        float M = a[i] * range;
+        float M = a[i] * (range - threshold) + threshold;
 
         r[i] *= M;
         g[i] *= M;

src/nvtt/bc6h/shapes_two.h

@@ -19,7 +19,7 @@ See the License for the specific language governing permissions and limitations
 #define NSHAPES 64
 #define SHAPEBITS 6
 
-static int shapes[NSHAPES*16] = 
+static const int shapes[NSHAPES*16] = 
 {
 0, 0, 1, 1,   0, 0, 0, 1,   0, 1, 1, 1,   0, 0, 0, 1,   
 0, 0, 1, 1,   0, 0, 0, 1,   0, 1, 1, 1,   0, 0, 1, 1,   
@@ -105,7 +105,7 @@ static int shapes[NSHAPES*16] =
 
 #define	REGION(x,y,si)	shapes[((si)&3)*4+((si)>>2)*64+(x)+(y)*16]
 
-static int shapeindex_to_compressed_indices[NSHAPES*2] = 
+static const int shapeindex_to_compressed_indices[NSHAPES*2] = 
 {
 	0,15,  0,15,  0,15,  0,15,
 	0,15,  0,15,  0,15,  0,15,

src/nvtt/bc6h/tile.h

@@ -23,7 +23,7 @@ See the License for the specific language governing permissions and limitations
 //#define	USE_IMPORTANCE_MAP	1		// define this if you want to increase importance of some pixels in tile
 class Tile
 {
-private:
+public:
 	// NOTE: this returns the appropriately-clamped BIT PATTERN of the half as an INTEGRAL float value
 	static float half2float(uint16 h)
 	{
@@ -67,7 +67,6 @@ private:
 	}
 #endif
 
-public:
 	Tile() {};
 	~Tile(){};
 	Tile(int xs, int ys) {size_x = xs; size_y = ys;}

src/nvtt/bc6h/utils.cpp

@@ -18,8 +18,8 @@ See the License for the specific language governing permissions and limitations
 
 using namespace nv;
 
-static int denom7_weights_64[] = {0, 9, 18, 27, 37, 46, 55, 64};										// divided by 64
+static const int denom7_weights_64[] = {0, 9, 18, 27, 37, 46, 55, 64};										// divided by 64
-static int denom15_weights_64[] = {0, 4, 9, 13, 17, 21, 26, 30, 34, 38, 43, 47, 51, 55, 60, 64};		// divided by 64
+static const int denom15_weights_64[] = {0, 4, 9, 13, 17, 21, 26, 30, 34, 38, 43, 47, 51, 55, 60, 64};		// divided by 64
 
 /*static*/ Format Utils::FORMAT;
 
@@ -28,7 +28,8 @@ int Utils::lerp(int a, int b, int i, int denom)
 	nvDebugCheck (denom == 3 || denom == 7 || denom == 15);
 	nvDebugCheck (i >= 0 && i <= denom);
 
-	int round = 32, shift = 6, *weights;
+	int round = 32, shift = 6;
+	const int *weights;
 
 	switch(denom)
 	{
@@ -46,7 +47,8 @@ Vector3 Utils::lerp(const Vector3& a, const Vector3 &b, int i, int denom)
 	nvDebugCheck (denom == 3 || denom == 7 || denom == 15);
 	nvDebugCheck (i >= 0 && i <= denom);
 
-	int shift = 6, *weights;
+	int shift = 6;
+	const int *weights;
 
 	switch(denom)
 	{
@@ -57,7 +59,7 @@ Vector3 Utils::lerp(const Vector3& a, const Vector3 &b, int i, int denom)
 	}
 
 	// no need to round these as this is an exact division
-	return (a*weights[denom-i] +b*weights[i]) / float(1 << shift);
+	return (a*float(weights[denom-i]) +b*float(weights[i])) / float(1 << shift);
 }
 
 
@@ -254,117 +256,12 @@ int Utils::unquantize(int q, int prec)
 	return unq;
 }
 
-static int clamp(double r, double low, double high)
-{
-	if (r < low) return low;
-	else if (r > high) return high;
-	else return r;
-}
 
 
-// match the tonemapping function used by exrdisplay
-/*static void tonemap(const Vector3 &in, double exposure, Vector3 &out)
-{
-    double r,g,b;
-    unsigned short h;
-
-	// convert from bit pattern back to half and then to double
-	h = in.x;	r = h;
-	h = in.y;	g = h;
-	h = in.z;	b = h;
-
-    //  1) Compensate for fogging by subtracting defog
-    //     from the raw pixel values.
-    // Response: We work with defog of 0.0, so this is a no-op
-
-    //  2) Multiply the defogged pixel values by
-    //     2^(exposure + 2.47393).
-	double exposure_scale = pow(2.0, exposure + 2.47393);
-    r *= exposure_scale;
-    g *= exposure_scale;
-    b *= exposure_scale;
-
-    //  3) Values, which are now 1.0, are called "middle gray".
-    //     If defog and exposure are both set to 0.0, then
-    //     middle gray corresponds to a raw pixel value of 0.18.
-    //     In step 6, middle gray values will be mapped to an
-    //     intensity 3.5 f-stops below the display's maximum
-    //     intensity.
-    // Response: no apparent content.
-
-    //  4) Apply a knee function.  The knee function has two
-    //     parameters, kneeLow and kneeHigh.  Pixel values
-    //     below 2^kneeLow are not changed by the knee
-    //     function.  Pixel values above kneeLow are lowered
-    //     according to a logarithmic curve, such that the
-    //     value 2^kneeHigh is mapped to 2^3.5 (in step 6,
-    //     this value will be mapped to the the display's
-    //     maximum intensity).
-    // Response: kneeLow = 0.0 (2^0.0 => 1); kneeHigh = 5.0 (2^5 =>32)
-    if (r > 1.0)
-        r = 1.0 + log ((r-1.0) * 0.184874 + 1) / 0.184874;
-    if (g > 1.0)
-        g = 1.0 + log ((g-1.0) * 0.184874 + 1) / 0.184874;
-    if (b > 1.0)
-        b = 1.0 + log ((b-1.0) * 0.184874 + 1) / 0.184874;
-//
-//  5) Gamma-correct the pixel values, assuming that the
-//     screen's gamma is 0.4545 (or 1/2.2).
-    r = pow (r, 0.4545);
-    g = pow (g, 0.4545);
-    b = pow (b, 0.4545);
-
-//  6) Scale the values such that pixels middle gray
-//     pixels are mapped to 84.66 (or 3.5 f-stops below
-//     the display's maximum intensity).
-//
-//  7) Clamp the values to [0, 255].
-    r *= 84.66f;
-    g *= 84.66f;
-    b *= 84.66f;
-
-    out.x = clamp (r, 0, 255);
-    out.y = clamp (g, 0, 255);
-    out.z = clamp (b, 0, 255);
-}*/
-
-static void mpsnrmap(const Vector3 &in, int exposure, Vector3 &out)
-{
-    double r,g,b;
-    uint16 h;
-
-	// convert from bit pattern back to half and then to double
-	h = in.x;	r = h;
-	h = in.y;	g = h;
-	h = in.z;	b = h;
-
-	nvDebugCheck (exposure > -32 && exposure < 32);
-	if (exposure > 0)
-	{
-		r *= 1 << exposure;
-		g *= 1 << exposure;
-		b *= 1 << exposure;
-	}
-	else if (exposure < 0)
-	{
-		exposure = -exposure;
-		r /= 1 << exposure;
-		g /= 1 << exposure;
-		b /= 1 << exposure;
-	}
-    r = 255 * pow (r, 0.4545);
-    g = 255 * pow (g, 0.4545);
-    b = 255 * pow (b, 0.4545);
-
-    out.x = clamp (r, 0, 255);
-    out.y = clamp (g, 0, 255);
-    out.z = clamp (b, 0, 255);
-}
-
 // pick a norm!
 #define	NORM_EUCLIDEAN 1
 
-double Utils::norm(const Vector3 &a, const Vector3 &b)
+float Utils::norm(const Vector3 &a, const Vector3 &b)
 {
 #ifdef	NORM_EUCLIDEAN
 	return lengthSquared(a - b);
@@ -373,45 +270,6 @@ double Utils::norm(const Vector3 &a, const Vector3 &b)
 	Vector3 err = a - b;
 	return fabs(err.x) + fabs(err.y) + fabs(err.z);
 #endif
-#ifdef	NORM_EUCLIDEAN_EXPOSURE_UNWEIGHED
-	double toterr = 0;
-	Vector3 mapa, mapb, err;
-	for (int i=-6; i <= 6; i += 3)			// figure how many exposure samples needed. I'd argue if you take too many it's same as euclidean
-	{
-		tonemap(a, i, mapa);
-		tonemap(b, i, mapb);
-		err = mapa - mapb;
-		toterr += err * err;
-	}
-	return toterr;
-#endif
-#ifdef	NORM_EUCLIDEAN_EXPOSURE_WEIGHED
-	double toterr = 0;
-	Vector3 mapa, mapb, err;
-	double rwt = 0.299;
-	double gwt = 0.587;
-	double bwt = 0.114;
-	for (int i=-6; i <= 6; i += 3)			// figure how many exposure samples needed. I'd argue if you take too many it's same as euclidean
-	{
-		tonemap(a, i, mapa);
-		tonemap(b, i, mapb);
-		mapa.x *= rwt; mapa.y *= gwt; mapa.z *= bwt;
-		mapb.x *= rwt; mapb.y *= gwt; mapb.z *= bwt;
-		err = mapa - mapb;
-		toterr += err * err;
-	}
-	return toterr;
-#endif
-}
-
-double Utils::mpsnr_norm(const Vector3 &a, int exposure, const Vector3 &b)
-{
-	Vector3 mapa, mapb;
-
-	mpsnrmap(a, exposure, mapa);
-	mpsnrmap(b, exposure, mapb);
-
-	return lengthSquared(mapa - mapb);
 }
 
 // parse <name>[<start>{:<end>}]{,}	

src/nvtt/bc6h/utils.h

@@ -47,8 +47,8 @@ public:
     static ::Format FORMAT;     // this is a global -- we're either handling unsigned or unsigned half values
 
     // error metrics
-    static double norm(const nv::Vector3 &a, const nv::Vector3 &b);
+    static float norm(const nv::Vector3 &a, const nv::Vector3 &b);
-    static double mpsnr_norm(const nv::Vector3 &a, int exposure, const nv::Vector3 &b);
+    static float mpsnr_norm(const nv::Vector3 &a, int exposure, const nv::Vector3 &b);
 
     // conversion & clamp
     static int ushort_to_format(unsigned short input);

src/nvtt/bc6h/zoh.cpp

@@ -29,8 +29,8 @@ void ZOH::compress(const Tile &t, char *block)
 {
 	char oneblock[ZOH::BLOCKSIZE], twoblock[ZOH::BLOCKSIZE];
 
-	double mseone = ZOH::compressone(t, oneblock);
+	float mseone = ZOH::compressone(t, oneblock);
-	double msetwo = ZOH::compresstwo(t, twoblock);
+	float msetwo = ZOH::compresstwo(t, twoblock);
 
 	if (mseone <= msetwo)
 		memcpy(block, oneblock, ZOH::BLOCKSIZE);

src/nvtt/bc6h/zoh.h

@@ -55,16 +55,16 @@ public:
 	static void compress(const Tile &t, char *block);
 	static void decompress(const char *block, Tile &t);
 
-	static double compressone(const Tile &t, char *block);
+	static float compressone(const Tile &t, char *block);
-	static double compresstwo(const Tile &t, char *block);
+	static float compresstwo(const Tile &t, char *block);
 	static void decompressone(const char *block, Tile &t);
 	static void decompresstwo(const char *block, Tile &t);
 
-	static double refinetwo(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_TWO], char *block);
+	static float refinetwo(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_TWO], char *block);
-	static double roughtwo(const Tile &tile, int shape, FltEndpts endpts[NREGIONS_TWO]);
+	static float roughtwo(const Tile &tile, int shape, FltEndpts endpts[NREGIONS_TWO]);
 
-	static double refineone(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_ONE], char *block);
+	static float refineone(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_ONE], char *block);
-	static double roughone(const Tile &tile, int shape, FltEndpts endpts[NREGIONS_ONE]);
+	static float roughone(const Tile &tile, int shape, FltEndpts endpts[NREGIONS_ONE]);
 
 	static bool isone(const char *block);
 };

src/nvtt/bc6h/zohone.cpp

@@ -33,7 +33,7 @@ using namespace nv;
 
 #define	NSHAPES	1
 
-static int shapes[NSHAPES] =
+static const int shapes[NSHAPES] =
 {
     0x0000
 };	// only 1 shape
@@ -64,7 +64,7 @@ struct Pattern
 
 #define	NPATTERNS 4
 
-static Pattern patterns[NPATTERNS] =
+static const Pattern patterns[NPATTERNS] =
 {
     16,4,  16,4,  16,4,   1, 0x0f, 5, "bw[10],bw[11],bw[12],bw[13],bw[14],bw[15],bx[3:0],gw[10],gw[11],gw[12],gw[13],gw[14],gw[15],gx[3:0],rw[10],rw[11],rw[12],rw[13],rw[14],rw[15],rx[3:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
     12,8,  12,8,  12,8,   1, 0x0b, 5, "bw[10],bw[11],bx[7:0],gw[10],gw[11],gx[7:0],rw[10],rw[11],rx[7:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
@@ -73,7 +73,7 @@ static Pattern patterns[NPATTERNS] =
 };
 
 // mapping of mode to the corresponding index in pattern
-static int mode_to_pat[MAXMODES] = {
+static const int mode_to_pat[MAXMODES] = {
     -1,-1,-1,
     3,	// 0x03
     -1,-1,-1,
@@ -199,12 +199,13 @@ static void write_header(const ComprEndpts endpts[NREGIONS_ONE], const Pattern &
     int rw = endpts[0].A[0], rx = endpts[0].B[0];
     int gw = endpts[0].A[1], gx = endpts[0].B[1];
     int bw = endpts[0].A[2], bx = endpts[0].B[2];
-    int ptr = strlen(p.encoding);
+    int ptr = int(strlen(p.encoding));
     while (ptr)
     {
         Field field;
         int endbit, len;
 
+		// !!!UNDONE: get rid of string parsing!!!
         Utils::parse(p.encoding, ptr, field, endbit, len);
         switch(field)
         {
@@ -252,13 +253,14 @@ static void read_header(Bits &in, ComprEndpts endpts[NREGIONS_ONE], Pattern &p)
     gw = gx = 0;
     bw = bx = 0;
 
-    int ptr = strlen(p.encoding);
+    int ptr = int(strlen(p.encoding));
 
     while (ptr)
     {
         Field field;
         int endbit, len;
 
+		// !!!UNDONE: get rid of string parsing!!!
         Utils::parse(p.encoding, ptr, field, endbit, len);
 
         switch(field)
@@ -322,21 +324,21 @@ static void generate_palette_quantized(const IntEndpts &endpts, int prec, Vector
 
     // interpolate
     for (int i = 0; i < NINDICES; ++i)
-        palette[i].x = Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec);
+        palette[i].x = float(Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec));
 
     a = Utils::unquantize(endpts.A[1], prec);
     b = Utils::unquantize(endpts.B[1], prec);
 
     // interpolate
     for (int i = 0; i < NINDICES; ++i)
-        palette[i].y = Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec);
+        palette[i].y = float(Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec));
 
     a = Utils::unquantize(endpts.A[2], prec);
     b = Utils::unquantize(endpts.B[2], prec);
 
     // interpolate
     for (int i = 0; i < NINDICES; ++i)
-        palette[i].z = Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec);
+        palette[i].z = float(Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec));
 }
 
 // position 0 was compressed
@@ -382,17 +384,17 @@ void ZOH::decompressone(const char *block, Tile &t)
 }
 
 // given a collection of colors and quantized endpoints, generate a palette, choose best entries, and return a single toterr
-static double map_colors(const Vector3 colors[], const float importance[], int np, const IntEndpts &endpts, int prec)
+static float map_colors(const Vector3 colors[], const float importance[], int np, const IntEndpts &endpts, int prec)
 {
     Vector3 palette[NINDICES];
-    double toterr = 0;
+    float toterr = 0;
     Vector3 err;
 
     generate_palette_quantized(endpts, prec, palette);
 
     for (int i = 0; i < np; ++i)
     {
-        double err, besterr;
+        float err, besterr;
 
         besterr = Utils::norm(colors[i], palette[0]) * importance[i];
 
@@ -412,7 +414,7 @@ static double map_colors(const Vector3 colors[], const float importance[], int n
 
 // assign indices given a tile, shape, and quantized endpoints, return toterr for each region
 static void assign_indices(const Tile &tile, int shapeindex, IntEndpts endpts[NREGIONS_ONE], int prec, 
-                           int indices[Tile::TILE_H][Tile::TILE_W], double toterr[NREGIONS_ONE])
+                           int indices[Tile::TILE_H][Tile::TILE_W], float toterr[NREGIONS_ONE])
 {
     // build list of possibles
     Vector3 palette[NREGIONS_ONE][NINDICES];
@@ -429,7 +431,7 @@ static void assign_indices(const Tile &tile, int shapeindex, IntEndpts endpts[NR
 	for (int x = 0; x < tile.size_x; x++)
 	{
         int region = REGION(x,y,shapeindex);
-        double err, besterr;
+        float err, besterr;
 
         besterr = Utils::norm(tile.data[y][x], palette[region][0]);
         indices[y][x] = 0;
@@ -450,8 +452,8 @@ static void assign_indices(const Tile &tile, int shapeindex, IntEndpts endpts[NR
     }
 }
 
-static double perturb_one(const Vector3 colors[], const float importance[], int np, int ch, int prec, const IntEndpts &old_endpts, IntEndpts &new_endpts,
+static float perturb_one(const Vector3 colors[], const float importance[], int np, int ch, int prec, const IntEndpts &old_endpts, IntEndpts &new_endpts,
-                          double old_err, int do_b)
+                          float old_err, int do_b)
 {
     // we have the old endpoints: old_endpts
     // we have the perturbed endpoints: new_endpts
@@ -504,9 +506,9 @@ static double perturb_one(const Vector3 colors[], const float importance[], int
     return min_err;
 }
 
-static void optimize_one(const Vector3 colors[], const float importance[], int np, double orig_err, const IntEndpts &orig_endpts, int prec, IntEndpts &opt_endpts)
+static void optimize_one(const Vector3 colors[], const float importance[], int np, float orig_err, const IntEndpts &orig_endpts, int prec, IntEndpts &opt_endpts)
 {
-    double opt_err = orig_err;
+    float opt_err = orig_err;
     for (int ch = 0; ch < NCHANNELS; ++ch)
     {
         opt_endpts.A[ch] = orig_endpts.A[ch];
@@ -576,12 +578,12 @@ static void optimize_one(const Vector3 colors[], const float importance[], int n
     }
 }
 
-static void optimize_endpts(const Tile &tile, int shapeindex, const double orig_err[NREGIONS_ONE], 
+static void optimize_endpts(const Tile &tile, int shapeindex, const float orig_err[NREGIONS_ONE], 
                             const IntEndpts orig_endpts[NREGIONS_ONE], int prec, IntEndpts opt_endpts[NREGIONS_ONE])
 {
     Vector3 pixels[Tile::TILE_TOTAL];
     float importance[Tile::TILE_TOTAL];
-    double err = 0;
+    float err = 0;
 
     for (int region=0; region<NREGIONS_ONE; ++region)
     {
@@ -618,9 +620,9 @@ static void optimize_endpts(const Tile &tile, int shapeindex, const double orig_
                 emit compressed block with original data // to try to preserve maximum endpoint precision
 */
 
-double ZOH::refineone(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_ONE], char *block)
+float ZOH::refineone(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_ONE], char *block)
 {
-    double orig_err[NREGIONS_ONE], opt_err[NREGIONS_ONE], orig_toterr, opt_toterr;
+    float orig_err[NREGIONS_ONE], opt_err[NREGIONS_ONE], orig_toterr, opt_toterr;
     IntEndpts orig_endpts[NREGIONS_ONE], opt_endpts[NREGIONS_ONE];
     ComprEndpts compr_orig[NREGIONS_ONE], compr_opt[NREGIONS_ONE];
     int orig_indices[Tile::TILE_H][Tile::TILE_W], opt_indices[Tile::TILE_H][Tile::TILE_W];
@@ -668,21 +670,21 @@ static void generate_palette_unquantized(const FltEndpts endpts[NREGIONS_ONE], V
 }
 
 // generate a palette from unquantized endpoints, then pick best palette color for all pixels in each region, return toterr for all regions combined
-static double map_colors(const Tile &tile, int shapeindex, const FltEndpts endpts[NREGIONS_ONE])
+static float map_colors(const Tile &tile, int shapeindex, const FltEndpts endpts[NREGIONS_ONE])
 {
     // build list of possibles
     Vector3 palette[NREGIONS_ONE][NINDICES];
 
     generate_palette_unquantized(endpts, palette);
 
-    double toterr = 0;
+    float toterr = 0;
     Vector3 err;
 
     for (int y = 0; y < tile.size_y; y++)
 	for (int x = 0; x < tile.size_x; x++)
 	{
         int region = REGION(x,y,shapeindex);
-        double err, besterr;
+        float err, besterr;
 
         besterr = Utils::norm(tile.data[y][x], palette[region][0]) * tile.importance_map[y][x];
 
@@ -700,7 +702,7 @@ static double map_colors(const Tile &tile, int shapeindex, const FltEndpts endpt
     return toterr;
 }
 
-double ZOH::roughone(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS_ONE])
+float ZOH::roughone(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS_ONE])
 {
     for (int region=0; region<NREGIONS_ONE; ++region)
     {
@@ -742,7 +744,7 @@ double ZOH::roughone(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS
 
         mean /= float(np);
 
-        Vector3 direction = Fit::computePrincipalComponent(np, colors);
+        Vector3 direction = Fit::computePrincipalComponent_EigenSolver(np, colors);
 
         // project each pixel value along the principal direction
         float minp = FLT_MAX, maxp = -FLT_MAX;
@@ -767,11 +769,11 @@ double ZOH::roughone(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS
     return map_colors(tile, shapeindex, endpts);
 }
 
-double ZOH::compressone(const Tile &t, char *block)
+float ZOH::compressone(const Tile &t, char *block)
 {
     int shapeindex_best = 0;
     FltEndpts endptsbest[NREGIONS_ONE], tempendpts[NREGIONS_ONE];
-    double msebest = DBL_MAX;
+    float msebest = FLT_MAX;
 
     /*
 		collect the mse values that are within 5% of the best values
@@ -780,7 +782,7 @@ double ZOH::compressone(const Tile &t, char *block)
     // hack for now -- just use the best value WORK
     for (int i=0; i<NSHAPES && msebest>0.0; ++i)
     {
-        double mse = roughone(t, i, tempendpts);
+        float mse = roughone(t, i, tempendpts);
         if (mse < msebest)
         {
             msebest = mse;

src/nvtt/bc6h/zohtwo.cpp

@@ -90,7 +90,7 @@ struct Pattern
 
 #define	NPATTERNS 10
 
-static Pattern patterns[NPATTERNS] =
+static const Pattern patterns[NPATTERNS] =
 {
     11,5,5,5,	11,4,4,4,	11,4,4,4,	1,	0x02, 5, "d[4:0],bz[3],rz[4:0],bz[2],ry[4:0],by[3:0],bz[1],bw[10],bx[3:0],gz[3:0],bz[0],gw[10],gx[3:0],gy[3:0],rw[10],rx[4:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
     11,4,4,4,	11,5,5,5,	11,4,4,4,	1,	0x06, 5, "d[4:0],bz[3],gy[4],rz[3:0],bz[2],bz[0],ry[3:0],by[3:0],bz[1],bw[10],bx[3:0],gz[3:0],gw[10],gx[4:0],gy[3:0],gz[4],rw[10],rx[3:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
@@ -106,7 +106,7 @@ static Pattern patterns[NPATTERNS] =
 
 // mapping of mode to the corresponding index in pattern
 // UNUSED ZOH MODES are 0x13, 0x17, 0x1b, 0x1f -- return -2 for these
-static int mode_to_pat[MAXMODES] = {	
+static const int mode_to_pat[MAXMODES] = {	
     3,	// 0x00
     8,	// 0x01
     0,	// 0x02
@@ -258,12 +258,13 @@ static void write_header(const ComprEndpts endpts[NREGIONS_TWO], int shapeindex,
     int rw = endpts[0].A[0], rx = endpts[0].B[0], ry = endpts[1].A[0], rz = endpts[1].B[0];
     int gw = endpts[0].A[1], gx = endpts[0].B[1], gy = endpts[1].A[1], gz = endpts[1].B[1];
     int bw = endpts[0].A[2], bx = endpts[0].B[2], by = endpts[1].A[2], bz = endpts[1].B[2];
-    int ptr = strlen(p.encoding);
+    int ptr = int(strlen(p.encoding));
     while (ptr)
     {
         Field field;
         int endbit, len;
 
+		// !!!UNDONE: get rid of string parsing!!!
         Utils::parse(p.encoding, ptr, field, endbit, len);
         switch(field)
         {
@@ -313,13 +314,14 @@ static bool read_header(Bits &in, ComprEndpts endpts[NREGIONS_TWO], int &shapein
     gw = gx = gy = gz = 0;
     bw = bx = by = bz = 0;
 
-    int ptr = strlen(p.encoding);
+    int ptr = int(strlen(p.encoding));
 
     while (ptr)
     {
         Field field;
         int endbit, len;
 
+		// !!!UNDONE: get rid of string parsing!!!
         Utils::parse(p.encoding, ptr, field, endbit, len);
 
         switch(field)
@@ -394,21 +396,21 @@ static void generate_palette_quantized(const IntEndpts &endpts, int prec, Vector
 
     // interpolate
     for (int i = 0; i < NINDICES; ++i)
-        palette[i].x = Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec);
+        palette[i].x = float(Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec));
 
     a = Utils::unquantize(endpts.A[1], prec);
     b = Utils::unquantize(endpts.B[1], prec);
 
     // interpolate
     for (int i = 0; i < NINDICES; ++i)
-        palette[i].y = Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec);
+        palette[i].y = float(Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec));
 
     a = Utils::unquantize(endpts.A[2], prec);
     b = Utils::unquantize(endpts.B[2], prec);
 
     // interpolate
     for (int i = 0; i < NINDICES; ++i)
-        palette[i].z = Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec);
+        palette[i].z = float(Utils::finish_unquantize(PALETTE_LERP(a, b, i, DENOM), prec));
 }
 
 static void read_indices(Bits &in, int shapeindex, int indices[Tile::TILE_H][Tile::TILE_W])
@@ -470,17 +472,17 @@ void ZOH::decompresstwo(const char *block, Tile &t)
 }
 
 // given a collection of colors and quantized endpoints, generate a palette, choose best entries, and return a single toterr
-static double map_colors(const Vector3 colors[], const float importance[], int np, const IntEndpts &endpts, int prec)
+static float map_colors(const Vector3 colors[], const float importance[], int np, const IntEndpts &endpts, int prec)
 {
     Vector3 palette[NINDICES];
-    double toterr = 0;
+    float toterr = 0;
     Vector3 err;
 
     generate_palette_quantized(endpts, prec, palette);
 
     for (int i = 0; i < np; ++i)
     {
-        double err, besterr;
+        float err, besterr;
 
         besterr = Utils::norm(colors[i], palette[0]) * importance[i];
 
@@ -500,7 +502,7 @@ static double map_colors(const Vector3 colors[], const float importance[], int n
 
 // assign indices given a tile, shape, and quantized endpoints, return toterr for each region
 static void assign_indices(const Tile &tile, int shapeindex, IntEndpts endpts[NREGIONS_TWO], int prec, 
-                           int indices[Tile::TILE_H][Tile::TILE_W], double toterr[NREGIONS_TWO])
+                           int indices[Tile::TILE_H][Tile::TILE_W], float toterr[NREGIONS_TWO])
 {
     // build list of possibles
     Vector3 palette[NREGIONS_TWO][NINDICES];
@@ -517,7 +519,7 @@ static void assign_indices(const Tile &tile, int shapeindex, IntEndpts endpts[NR
 	for (int x = 0; x < tile.size_x; x++)
 	{
         int region = REGION(x,y,shapeindex);
-        double err, besterr;
+        float err, besterr;
 
         besterr = Utils::norm(tile.data[y][x], palette[region][0]);
         indices[y][x] = 0;
@@ -538,8 +540,8 @@ static void assign_indices(const Tile &tile, int shapeindex, IntEndpts endpts[NR
     }
 }
 
-static double perturb_one(const Vector3 colors[], const float importance[], int np, int ch, int prec, const IntEndpts &old_endpts, IntEndpts &new_endpts,
+static float perturb_one(const Vector3 colors[], const float importance[], int np, int ch, int prec, const IntEndpts &old_endpts, IntEndpts &new_endpts,
-                          double old_err, int do_b)
+                          float old_err, int do_b)
 {
     // we have the old endpoints: old_endpts
     // we have the perturbed endpoints: new_endpts
@@ -592,9 +594,9 @@ static double perturb_one(const Vector3 colors[], const float importance[], int
     return min_err;
 }
 
-static void optimize_one(const Vector3 colors[], const float importance[], int np, double orig_err, const IntEndpts &orig_endpts, int prec, IntEndpts &opt_endpts)
+static void optimize_one(const Vector3 colors[], const float importance[], int np, float orig_err, const IntEndpts &orig_endpts, int prec, IntEndpts &opt_endpts)
 {
-    double opt_err = orig_err;
+    float opt_err = orig_err;
     for (int ch = 0; ch < NCHANNELS; ++ch)
     {
         opt_endpts.A[ch] = orig_endpts.A[ch];
@@ -664,12 +666,12 @@ static void optimize_one(const Vector3 colors[], const float importance[], int n
     }
 }
 
-static void optimize_endpts(const Tile &tile, int shapeindex, const double orig_err[NREGIONS_TWO], 
+static void optimize_endpts(const Tile &tile, int shapeindex, const float orig_err[NREGIONS_TWO], 
                             const IntEndpts orig_endpts[NREGIONS_TWO], int prec, IntEndpts opt_endpts[NREGIONS_TWO])
 {
     Vector3 pixels[Tile::TILE_TOTAL];
     float importance[Tile::TILE_TOTAL];
-    double err = 0;
+    float err = 0;
 
     for (int region=0; region<NREGIONS_TWO; ++region)
     {
@@ -706,9 +708,9 @@ static void optimize_endpts(const Tile &tile, int shapeindex, const double orig_
                 emit compressed block with original data // to try to preserve maximum endpoint precision
 */
 
-double ZOH::refinetwo(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_TWO], char *block)
+float ZOH::refinetwo(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_TWO], char *block)
 {
-    double orig_err[NREGIONS_TWO], opt_err[NREGIONS_TWO], orig_toterr, opt_toterr;
+    float orig_err[NREGIONS_TWO], opt_err[NREGIONS_TWO], orig_toterr, opt_toterr;
     IntEndpts orig_endpts[NREGIONS_TWO], opt_endpts[NREGIONS_TWO];
     ComprEndpts compr_orig[NREGIONS_TWO], compr_opt[NREGIONS_TWO];
     int orig_indices[Tile::TILE_H][Tile::TILE_W], opt_indices[Tile::TILE_H][Tile::TILE_W];
@@ -755,21 +757,21 @@ static void generate_palette_unquantized(const FltEndpts endpts[NREGIONS_TWO], V
 }
 
 // generate a palette from unquantized endpoints, then pick best palette color for all pixels in each region, return toterr for all regions combined
-static double map_colors(const Tile &tile, int shapeindex, const FltEndpts endpts[NREGIONS_TWO])
+static float map_colors(const Tile &tile, int shapeindex, const FltEndpts endpts[NREGIONS_TWO])
 {
     // build list of possibles
     Vector3 palette[NREGIONS_TWO][NINDICES];
 
     generate_palette_unquantized(endpts, palette);
 
-    double toterr = 0;
+    float toterr = 0;
     Vector3 err;
 
     for (int y = 0; y < tile.size_y; y++)
 	for (int x = 0; x < tile.size_x; x++)
 	{
         int region = REGION(x,y,shapeindex);
-        double err, besterr;
+        float err, besterr;
 
         besterr = Utils::norm(tile.data[y][x], palette[region][0]) * tile.importance_map[y][x];
 
@@ -787,7 +789,7 @@ static double map_colors(const Tile &tile, int shapeindex, const FltEndpts endpt
     return toterr;
 }
 
-double ZOH::roughtwo(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS_TWO])
+float ZOH::roughtwo(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS_TWO])
 {
     for (int region=0; region<NREGIONS_TWO; ++region)
     {
@@ -827,7 +829,7 @@ double ZOH::roughtwo(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS
 
         mean /= float(np);
 
-        Vector3 direction = Fit::computePrincipalComponent(np, colors);
+        Vector3 direction = Fit::computePrincipalComponent_EigenSolver(np, colors);
 
         // project each pixel value along the principal direction
         float minp = FLT_MAX, maxp = -FLT_MAX;
@@ -852,11 +854,11 @@ double ZOH::roughtwo(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS
     return map_colors(tile, shapeindex, endpts);
 }
 
-double ZOH::compresstwo(const Tile &t, char *block)
+float ZOH::compresstwo(const Tile &t, char *block)
 {
     int shapeindex_best = 0;
     FltEndpts endptsbest[NREGIONS_TWO], tempendpts[NREGIONS_TWO];
-    double msebest = DBL_MAX;
+    float msebest = FLT_MAX;
 
     /*
     collect the mse values that are within 5% of the best values
@@ -865,7 +867,7 @@ double ZOH::compresstwo(const Tile &t, char *block)
     // hack for now -- just use the best value WORK
     for (int i=0; i<NSHAPES && msebest>0.0; ++i)
     {
-        double mse = roughtwo(t, i, tempendpts);
+        float mse = roughtwo(t, i, tempendpts);
         if (mse < msebest)
         {
             msebest = mse;

src/nvtt/nvtt.h

@@ -101,7 +101,7 @@ namespace nvtt
         Format_DXT1n,   // Not supported on CPU yet.
         Format_CTX1,    // Not supported on CPU yet.
 
-        Format_BC6,     // Not supported yet.
+        Format_BC6,
         Format_BC7,     // Not supported yet.
 
         Format_DXT1_Luma,
@@ -503,7 +503,7 @@ namespace nvtt
         NVTT_API void fill(float r, float g, float b, float a);
         NVTT_API void scaleAlphaToCoverage(float coverage, float alphaRef = 0.5f);
         NVTT_API void toRGBM(float range = 1.0f, float threshold = 0.0f);
-        NVTT_API void fromRGBM(float range = 1.0f);
+        NVTT_API void fromRGBM(float range = 1.0f, float threshold = 0.0f);
         NVTT_API void toLM(float range = 1.0f, float threshold = 0.0f);
         NVTT_API void toRGBE(int mantissaBits, int exponentBits);
         NVTT_API void fromRGBE(int mantissaBits, int exponentBits);

src/nvtt/tests/testsuite.cpp

@@ -187,6 +187,7 @@ enum Mode {
     Mode_BC5_Normal_Paraboloid,
     Mode_BC5_Normal_Quartic,
     //Mode_BC5_Normal_DualParaboloid,
+	Mode_BC6,
     Mode_Count
 };
 static const char * s_modeNames[] = {
@@ -205,6 +206,7 @@ static const char * s_modeNames[] = {
     "BC5-Normal-Paraboloid",        // Mode_BC5_Normal_Paraboloid,
     "BC5-Normal-Quartic",           // Mode_BC5_Normal_Quartic,
     //"BC5-Normal-DualParaboloid",    // Mode_BC5_Normal_DualParaboloid,
+	"BC6",			// Mode_BC6,
 };
 nvStaticCheck(NV_ARRAY_SIZE(s_modeNames) == Mode_Count);
 
@@ -214,11 +216,14 @@ struct Test {
     Mode modes[6];
 };
 static Test s_imageTests[] = {
-    {"Color", 3, {Mode_BC1, Mode_BC3_YCoCg, Mode_BC3_RGBM, Mode_BC3_LUVW}},
+    {"Color", 4, {Mode_BC1, Mode_BC3_YCoCg, Mode_BC3_RGBM, Mode_BC6, /*Mode_BC3_LUVW*/}},
     {"Alpha", 3, {Mode_BC1_Alpha, Mode_BC2_Alpha, Mode_BC3_Alpha}},
     //{"Normal", 3, {Mode_BC1_Normal, Mode_BC3_Normal, Mode_BC5_Normal}},
     {"Normal", 4, {Mode_BC5_Normal, Mode_BC5_Normal_Stereographic, Mode_BC5_Normal_Paraboloid, Mode_BC5_Normal_Quartic}},
     {"Lightmap", 4, {Mode_BC1, Mode_BC3_YCoCg, Mode_BC3_RGBM, Mode_BC3_RGBS}},
+	{"HDR", 2, {Mode_BC3_RGBM, Mode_BC6}},
+	//{"BC6", 1, {Mode_BC6}},	// temporary mode for testing
+	//{"BC7", 1, {Mode_BC7}},	// temporary mode for testing
 };
 const int s_imageTestCount = ARRAY_SIZE(s_imageTests);
 
@@ -247,8 +252,9 @@ static ImageSet s_imageSets[] = {
     {"Witness",     "witness",      s_witnessImageSet,      ARRAY_SIZE(s_witnessImageSet),      ImageType_RGB},     // 6
     {"Lightmap",    "lightmap",     s_witnessLmapImageSet,  ARRAY_SIZE(s_witnessLmapImageSet),  ImageType_HDR},     // 7
     {"Normal",      "id_tnmap",     s_normalMapImageSet,    ARRAY_SIZE(s_normalMapImageSet),    ImageType_Normal},  // 8
+	// !!!UNDONE: more HDR image sets
 };
-const int s_imageSetCount = sizeof(s_imageSets)/sizeof(s_imageSets[0]);
+const int s_imageSetCount = ARRAY_SIZE(s_imageSets);
 
 
 struct MyOutputHandler : public nvtt::OutputHandler
@@ -396,6 +402,10 @@ int main(int argc, char *argv[])
                 i++;
             }
         }
+		else
+		{
+			printf("Warning: unrecognized option \"%s\"\n", argv[i]);
+		}
     }
 
     // Validate inputs.
@@ -419,7 +429,7 @@ int main(int argc, char *argv[])
         for (int i = 0; i < s_imageSetCount; i++) {
             printf("    %i:           \t%s.\n", i, s_imageSets[i].name);
         }
-        printf("  -test [0:%d]    \tCompression tests to run.", s_imageTestCount);
+        printf("  -test [0:%d]    \tCompression tests to run.\n", s_imageTestCount);
         for (int i = 0; i < s_imageTestCount; i++) {
             printf("    %i:           \t%s.\n", i, s_imageTests[i].name);
         }
@@ -551,7 +561,7 @@ int main(int argc, char *argv[])
         if (t != test.count-1) graphWriter << ",";
     }
 
-    // Leyends.
+    // Legends.
     graphWriter << "&chdl=";
     for (int t = 0; t < test.count; t++)
     {
@@ -584,7 +594,7 @@ int main(int argc, char *argv[])
 
     for (int t = 0; t < test.count; t++)
     {
-        float totalTime = 0;
+        float totalCompressionTime = 0;
         float totalError = 0;
 
         Mode mode = test.modes[t];
@@ -602,6 +612,14 @@ int main(int argc, char *argv[])
         else if (mode == Mode_BC5_Normal || mode == Mode_BC5_Normal_Stereographic || mode == Mode_BC5_Normal_Paraboloid || mode == Mode_BC5_Normal_Quartic) {
             format = nvtt::Format_BC5;
         }
+		else if (mode == Mode_BC6)
+		{
+			format = nvtt::Format_BC6;
+		}
+		else
+		{
+			nvDebugCheck(false);
+		}
         
         compressionOptions.setFormat(format);
 
@@ -633,12 +651,12 @@ int main(int argc, char *argv[])
             if (img.isNormalMap()) {
                 img.normalizeNormalMap();
             }
-            if (set.type == ImageType_HDR) {
+            /*if (set.type == ImageType_HDR) {
                 img.scaleBias(0, 1.0f/4.0f, 0.0f); img.clamp(0);
                 img.scaleBias(1, 1.0f/4.0f, 0.0f); img.clamp(1);
                 img.scaleBias(2, 1.0f/4.0f, 0.0f); img.clamp(2);
                 img.toGamma(2);
-            }
+            }*/
 
             nvtt::Surface tmp = img;
             if (mode == Mode_BC1) {
@@ -672,7 +690,9 @@ int main(int argc, char *argv[])
             else if (mode == Mode_BC3_RGBM) {
                 tmp.setAlphaMode(nvtt::AlphaMode_None);
                 if (set.type == ImageType_HDR) {
-                    tmp.toRGBM(/*4*/);
+					// Transform to gamma-2.0 space before applying RGBM - helps a lot with banding in the darks.
+					tmp.toGamma(2.0f);
+                    tmp.toRGBM(3.0f);	// range of 3.0 in gamma-2.0 space == range of 9.0 in linear space
                 }
                 else {
                     tmp.toRGBM();
@@ -681,7 +701,7 @@ int main(int argc, char *argv[])
             else if (mode == Mode_BC3_LUVW) {
                 tmp.setAlphaMode(nvtt::AlphaMode_None);
                 if (set.type == ImageType_HDR) {
-                    tmp.toLUVW(/*4*/);
+                    tmp.toLUVW(8.0f);
                 }
                 else {
                     tmp.toLUVW();
@@ -725,7 +745,7 @@ int main(int argc, char *argv[])
 
             timer.stop();
             printf("  Time:  \t%.3f sec\n", timer.elapsed());
-            totalTime += timer.elapsed();
+            totalCompressionTime += timer.elapsed();
 
             nvtt::Surface img_out = outputHandler.decompress(mode, format, decoder);
             img_out.setAlphaMode(img.alphaMode());
@@ -757,7 +777,8 @@ int main(int argc, char *argv[])
             }
             else if (mode == Mode_BC3_RGBM) {
                 if (set.type == ImageType_HDR) {
-                    img_out.fromRGBM(/*4*/);
+                    img_out.fromRGBM(3.0f);
+					img_out.toLinear(2.0f);
                 }
                 else {
                     img_out.fromRGBM();
@@ -765,7 +786,7 @@ int main(int argc, char *argv[])
             }
             else if (mode == Mode_BC3_LUVW) {
                 if (set.type == ImageType_HDR) {
-                    img_out.fromLUVW(/*4*/);
+                    img_out.fromLUVW(8.0f);
                 }
                 else {
                     img_out.fromLUVW();
@@ -837,7 +858,7 @@ int main(int argc, char *argv[])
                 img_out.addChannel(residual_out, 2, 2, -1.0f); img_out.clamp(2);
             }
 
-            if (set.type == ImageType_HDR)
+            /*if (set.type == ImageType_HDR)
             {
                 Path outputFileName;
                 outputFileName.format("%s/%s", outPath, set.fileNames[i]);
@@ -848,7 +869,7 @@ int main(int argc, char *argv[])
                 {
                     printf("Error saving file '%s'.\n", outputFileName.str());
                 }
-            }
+            }*/
 
             // Output compressed image.
             Path outputFileName;
@@ -865,7 +886,6 @@ int main(int argc, char *argv[])
             float error;
             if (errorMode == ErrorMode_RMSE) {
                 error = nvtt::rmsError(img, img_out);
-                if (set.type == ImageType_HDR) error *= 4;
             }
             else if (errorMode == ErrorMode_CieLab) {
                 error = nvtt::cieLabError(img, img_out);
@@ -932,7 +952,7 @@ int main(int argc, char *argv[])
         totalError /= set.fileCount;
 
         printf("Total Results:\n");
-        printf("  Total Time:            \t%.3f sec\n", totalTime);
+        printf("  Total Compression Time:\t%.3f sec\n", totalCompressionTime);
         printf("  Average Error:         \t%.4f\n", totalError);
 
         if (t != test.count-1) graphWriter << "|";

src/nvtt/tools/assemble.cpp

@@ -83,6 +83,10 @@ int main(int argc, char *argv[])
 		{
 			files.append(argv[i]);
 		}
+		else
+		{
+			printf("Warning: unrecognized option \"%s\"\n", argv[i]);
+		}
 	}
 	
 	if (files.count() == 0)

src/nvtt/tools/compress.cpp

@@ -261,6 +261,15 @@ int main(int argc, char *argv[])
         {
             format = nvtt::Format_BC5;
         }
+        else if (strcmp("-bc6", argv[i]) == 0)
+        {
+            format = nvtt::Format_BC6;
+        }
+		// !!!UNDONE: add BC7 support
+        /*else if (strcmp("-bc7", argv[i]) == 0)
+        {
+            format = nvtt::Format_BC7;
+        }*/
 
         // Undocumented option. Mainly used for testing.
         else if (strcmp("-ext", argv[i]) == 0)
@@ -302,6 +311,10 @@ int main(int argc, char *argv[])
 
             break;
         }
+		else
+		{
+			printf("Warning: unrecognized option \"%s\"\n", argv[i]);
+		}
     }
 
     const uint version = nvtt::version();
@@ -314,7 +327,7 @@ int main(int argc, char *argv[])
 
     if (input.isNull())
     {
-        printf("usage: nvcompress [options] infile [outfile]\n\n");
+        printf("usage: nvcompress [options] infile [outfile.dds]\n\n");
 
         printf("Input options:\n");
         printf("  -color     \tThe input image is a color map (default).\n");
@@ -340,11 +353,13 @@ int main(int argc, char *argv[])
         printf("  -bc3     \tBC3 format (DXT5)\n");
         printf("  -bc3n    \tBC3 normal map format (DXT5nm)\n");
         printf("  -bc4     \tBC4 format (ATI1)\n");
-        printf("  -bc5     \tBC5 format (3Dc/ATI2)\n\n");
+        printf("  -bc5     \tBC5 format (3Dc/ATI2)\n");
+        printf("  -bc6     \tBC6 format\n");
+        //printf("  -bc7     \tBC7 format\n\n");
 
         printf("Output options:\n");
         printf("  -silent  \tDo not output progress messages\n");
-        printf("  -dds10   \tUse DirectX 10 DDS format\n\n");
+        printf("  -dds10   \tUse DirectX 10 DDS format (enabled by default for BC6/7)\n\n");
 
         return EXIT_FAILURE;
     }
@@ -581,6 +596,12 @@ int main(int argc, char *argv[])
     outputOptions.setOutputHandler(&outputHandler);
     outputOptions.setErrorHandler(&errorHandler);
 
+	// Automatically use dds10 if compressing to BC6 or BC7
+	if (format == nvtt::Format_BC6 || format == nvtt::Format_BC7)
+	{
+		dds10 = true;
+	}
+
     if (dds10)
     {
         outputOptions.setContainer(nvtt::Container_DDS10);

src/nvtt/tools/decompress.cpp

@@ -66,6 +66,8 @@ int main(int argc, char *argv[])
 			if (i+1 == argc) break;
 			i++;
 
+			// !!!UNDONE: Support at least one HDR output format
+
 #ifdef HAVE_PNG
 			if (strcmp("png", argv[i]) == 0) savePNG = true;
 			else 
@@ -92,13 +94,17 @@ int main(int argc, char *argv[])
 
 			break;
 		}
+		else
+		{
+			printf("Warning: unrecognized option \"%s\"\n", argv[i]);
+		}
 	}
 	
 	printf("NVIDIA Texture Tools - Copyright NVIDIA Corporation 2007\n\n");
 
 	if (input.isNull())
 	{
-		printf("usage: nvdecompress [options] infile [outfile]\n\n");
+		printf("usage: nvdecompress [options] infile.dds [outfile]\n\n");
 
 		printf("Note: the .tga or .png extension is forced on outfile\n\n");
 
@@ -112,6 +118,8 @@ int main(int argc, char *argv[])
  	}
 
 	// Load surface.
+	// !!! DirectDrawSurface API doesn't support float images, so BC6 will be converted to 8-bit on load.
+	// Should use nvtt::Surface instead.
 	nv::DirectDrawSurface dds(input.str());
 	if (!dds.isValid())
 	{

src/nvtt/tools/imgdiff.cpp

@@ -182,11 +182,10 @@ int main(int argc, char *argv[])
 		{
 			compareNormal = true;
 		}
-		if (strcmp("-alpha", argv[i]) == 0)
+		else if (strcmp("-alpha", argv[i]) == 0)
 		{
 			compareAlpha = true;
 		}
-
 		else if (argv[i][0] != '-')
 		{
 			input0 = argv[i];
@@ -197,6 +196,10 @@ int main(int argc, char *argv[])
 
 			break;
 		}
+		else
+		{
+			printf("Warning: unrecognized option \"%s\"\n", argv[i]);
+		}
 	}
 
 	if (input0.isNull() || input1.isNull())

src/nvtt/tools/resize.cpp

@@ -131,6 +131,10 @@ int main(int argc, char *argv[])
 
 			break;
 		}
+		else
+		{
+			printf("Warning: unrecognized option \"%s\"\n", argv[i]);
+		}
 	}
 
 	if (input.isNull() || output.isNull())