add and update samples for CUDA 11.6

Samples/5_Domain_Specific/HSOpticalFlow/.vscode/c_cpp_properties.json

@@ -0,0 +1,18 @@
+{
+    "configurations": [
+        {
+            "name": "Linux",
+            "includePath": [
+                "${workspaceFolder}/**",
+                "${workspaceFolder}/../../../Common"
+            ],
+            "defines": [],
+            "compilerPath": "/usr/local/cuda/bin/nvcc",
+            "cStandard": "gnu17",
+            "cppStandard": "gnu++14",
+            "intelliSenseMode": "linux-gcc-x64",
+            "configurationProvider": "ms-vscode.makefile-tools"
+        }
+    ],
+    "version": 4
+}

Samples/5_Domain_Specific/HSOpticalFlow/.vscode/extensions.json

@@ -0,0 +1,7 @@
+{
+    "recommendations": [
+        "nvidia.nsight-vscode-edition",
+        "ms-vscode.cpptools",
+        "ms-vscode.makefile-tools"
+    ]
+}

Samples/5_Domain_Specific/HSOpticalFlow/.vscode/launch.json

@@ -0,0 +1,10 @@
+{
+    "configurations": [
+        {
+            "name": "CUDA C++: Launch",
+            "type": "cuda-gdb",
+            "request": "launch",
+            "program": "${workspaceFolder}/HSOpticalFlow"
+        }
+    ]
+}

Samples/5_Domain_Specific/HSOpticalFlow/.vscode/tasks.json

@@ -0,0 +1,15 @@
+{
+    "version": "2.0.0",
+    "tasks": [
+        {
+            "label": "sample",
+            "type": "shell",
+            "command": "make dbg=1",
+            "problemMatcher": ["$nvcc"],
+            "group": {
+                "kind": "build",
+                "isDefault": true
+            }
+        }
+    ]
+}

Samples/5_Domain_Specific/HSOpticalFlow/HSOpticalFlow_vs2017.sln

@@ -0,0 +1,20 @@
+
+Microsoft Visual Studio Solution File, Format Version 12.00
+# Visual Studio 2017
+Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "HSOpticalFlow", "HSOpticalFlow_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
+EndProject
+Global
+	GlobalSection(SolutionConfigurationPlatforms) = preSolution
+		Debug|x64 = Debug|x64
+		Release|x64 = Release|x64
+	EndGlobalSection
+	GlobalSection(ProjectConfigurationPlatforms) = postSolution
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
+	EndGlobalSection
+	GlobalSection(SolutionProperties) = preSolution
+		HideSolutionNode = FALSE
+	EndGlobalSection
+EndGlobal

Samples/5_Domain_Specific/HSOpticalFlow/HSOpticalFlow_vs2017.vcxproj

@@ -0,0 +1,122 @@
+<?xml version="1.0" encoding="utf-8"?>
+<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+  <PropertyGroup>
+    <CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
+  </PropertyGroup>
+  <ItemGroup Label="ProjectConfigurations">
+    <ProjectConfiguration Include="Debug|x64">
+      <Configuration>Debug</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|x64">
+      <Configuration>Release</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+  </ItemGroup>
+  <PropertyGroup Label="Globals">
+    <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
+    <RootNamespace>HSOpticalFlow_vs2017</RootNamespace>
+    <ProjectName>HSOpticalFlow</ProjectName>
+    <CudaToolkitCustomDir />
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(WindowsTargetPlatformVersion)'==''">
+    <LatestTargetPlatformVersion>$([Microsoft.Build.Utilities.ToolLocationHelper]::GetLatestSDKTargetPlatformVersion('Windows', '10.0'))</LatestTargetPlatformVersion>
+    <WindowsTargetPlatformVersion Condition="'$(WindowsTargetPlatformVersion)' == ''">$(LatestTargetPlatformVersion)</WindowsTargetPlatformVersion>
+    <TargetPlatformVersion>$(WindowsTargetPlatformVersion)</TargetPlatformVersion>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
+  <PropertyGroup>
+    <ConfigurationType>Application</ConfigurationType>
+    <CharacterSet>MultiByte</CharacterSet>
+    <PlatformToolset>v141</PlatformToolset>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)'=='Debug'">
+    <UseDebugLibraries>true</UseDebugLibraries>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)'=='Release'">
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
+  <ImportGroup Label="ExtensionSettings">
+    <Import Project="$(CUDAPropsPath)\CUDA 11.6.props" />
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets">
+    <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
+  </ImportGroup>
+  <PropertyGroup Label="UserMacros" />
+  <PropertyGroup>
+    <IntDir>$(Platform)/$(Configuration)/</IntDir>
+    <IncludePath>$(IncludePath)</IncludePath>
+    <CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
+    <CodeAnalysisRules />
+    <CodeAnalysisRuleAssemblies />
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Platform)'=='x64'">
+    <OutDir>../../../bin/win64/$(Configuration)/</OutDir>
+  </PropertyGroup>
+  <ItemDefinitionGroup>
+    <ClCompile>
+      <WarningLevel>Level3</WarningLevel>
+      <PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../../Common;</AdditionalIncludeDirectories>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
+      <OutputFile>$(OutDir)/HSOpticalFlow.exe</OutputFile>
+    </Link>
+    <CudaCompile>
+      <CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;compute_86,sm_86;</CodeGeneration>
+      <AdditionalOptions>-Xcompiler "/wd 4819"  --threads 0 </AdditionalOptions>
+      <Include>./;../../../Common</Include>
+      <Defines>WIN32</Defines>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
+    <ClCompile>
+      <Optimization>Disabled</Optimization>
+      <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
+    </ClCompile>
+    <Link>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
+    </Link>
+    <CudaCompile>
+      <Runtime>MTd</Runtime>
+      <TargetMachinePlatform>64</TargetMachinePlatform>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
+    <ClCompile>
+      <Optimization>MaxSpeed</Optimization>
+      <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
+    </ClCompile>
+    <Link>
+      <GenerateDebugInformation>false</GenerateDebugInformation>
+      <LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
+    </Link>
+    <CudaCompile>
+      <Runtime>MT</Runtime>
+      <TargetMachinePlatform>64</TargetMachinePlatform>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemGroup>
+    <CudaCompile Include="flowCUDA.cu" />
+    <ClCompile Include="flowGold.cpp" />
+    <ClCompile Include="main.cpp" />
+    <ClInclude Include="common.h" />
+    <ClInclude Include="flowCUDA.h" />
+    <ClInclude Include="flowGold.h" />
+    <None Include="addKernel.cuh" />
+    <None Include="derivativesKernel.cuh" />
+    <None Include="downscaleKernel.cuh" />
+    <None Include="solverKernel.cuh" />
+    <None Include="upscaleKernel.cuh" />
+    <None Include="warpingKernel.cuh" />
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+    <Import Project="$(CUDAPropsPath)\CUDA 11.6.targets" />
+  </ImportGroup>
+</Project>

Samples/5_Domain_Specific/HSOpticalFlow/HSOpticalFlow_vs2019.sln

@@ -0,0 +1,20 @@
+
+Microsoft Visual Studio Solution File, Format Version 12.00
+# Visual Studio 2019
+Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "HSOpticalFlow", "HSOpticalFlow_vs2019.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
+EndProject
+Global
+	GlobalSection(SolutionConfigurationPlatforms) = preSolution
+		Debug|x64 = Debug|x64
+		Release|x64 = Release|x64
+	EndGlobalSection
+	GlobalSection(ProjectConfigurationPlatforms) = postSolution
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
+	EndGlobalSection
+	GlobalSection(SolutionProperties) = preSolution
+		HideSolutionNode = FALSE
+	EndGlobalSection
+EndGlobal

Samples/5_Domain_Specific/HSOpticalFlow/HSOpticalFlow_vs2019.vcxproj

@@ -0,0 +1,118 @@
+<?xml version="1.0" encoding="utf-8"?>
+<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+  <PropertyGroup>
+    <CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
+  </PropertyGroup>
+  <ItemGroup Label="ProjectConfigurations">
+    <ProjectConfiguration Include="Debug|x64">
+      <Configuration>Debug</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|x64">
+      <Configuration>Release</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+  </ItemGroup>
+  <PropertyGroup Label="Globals">
+    <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
+    <RootNamespace>HSOpticalFlow_vs2019</RootNamespace>
+    <ProjectName>HSOpticalFlow</ProjectName>
+    <CudaToolkitCustomDir />
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
+  <PropertyGroup>
+    <ConfigurationType>Application</ConfigurationType>
+    <CharacterSet>MultiByte</CharacterSet>
+    <PlatformToolset>v142</PlatformToolset>
+    <WindowsTargetPlatformVersion>10.0</WindowsTargetPlatformVersion>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)'=='Debug'">
+    <UseDebugLibraries>true</UseDebugLibraries>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)'=='Release'">
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
+  <ImportGroup Label="ExtensionSettings">
+    <Import Project="$(CUDAPropsPath)\CUDA 11.6.props" />
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets">
+    <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
+  </ImportGroup>
+  <PropertyGroup Label="UserMacros" />
+  <PropertyGroup>
+    <IntDir>$(Platform)/$(Configuration)/</IntDir>
+    <IncludePath>$(IncludePath)</IncludePath>
+    <CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
+    <CodeAnalysisRules />
+    <CodeAnalysisRuleAssemblies />
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Platform)'=='x64'">
+    <OutDir>../../../bin/win64/$(Configuration)/</OutDir>
+  </PropertyGroup>
+  <ItemDefinitionGroup>
+    <ClCompile>
+      <WarningLevel>Level3</WarningLevel>
+      <PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../../Common;</AdditionalIncludeDirectories>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
+      <OutputFile>$(OutDir)/HSOpticalFlow.exe</OutputFile>
+    </Link>
+    <CudaCompile>
+      <CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;compute_86,sm_86;</CodeGeneration>
+      <AdditionalOptions>-Xcompiler "/wd 4819"  --threads 0 </AdditionalOptions>
+      <Include>./;../../../Common</Include>
+      <Defines>WIN32</Defines>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
+    <ClCompile>
+      <Optimization>Disabled</Optimization>
+      <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
+    </ClCompile>
+    <Link>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
+    </Link>
+    <CudaCompile>
+      <Runtime>MTd</Runtime>
+      <TargetMachinePlatform>64</TargetMachinePlatform>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
+    <ClCompile>
+      <Optimization>MaxSpeed</Optimization>
+      <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
+    </ClCompile>
+    <Link>
+      <GenerateDebugInformation>false</GenerateDebugInformation>
+      <LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
+    </Link>
+    <CudaCompile>
+      <Runtime>MT</Runtime>
+      <TargetMachinePlatform>64</TargetMachinePlatform>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemGroup>
+    <CudaCompile Include="flowCUDA.cu" />
+    <ClCompile Include="flowGold.cpp" />
+    <ClCompile Include="main.cpp" />
+    <ClInclude Include="common.h" />
+    <ClInclude Include="flowCUDA.h" />
+    <ClInclude Include="flowGold.h" />
+    <None Include="addKernel.cuh" />
+    <None Include="derivativesKernel.cuh" />
+    <None Include="downscaleKernel.cuh" />
+    <None Include="solverKernel.cuh" />
+    <None Include="upscaleKernel.cuh" />
+    <None Include="warpingKernel.cuh" />
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+    <Import Project="$(CUDAPropsPath)\CUDA 11.6.targets" />
+  </ImportGroup>
+</Project>

Samples/5_Domain_Specific/HSOpticalFlow/HSOpticalFlow_vs2022.sln

@@ -0,0 +1,20 @@
+
+Microsoft Visual Studio Solution File, Format Version 12.00
+# Visual Studio 2022
+Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "HSOpticalFlow", "HSOpticalFlow_vs2022.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
+EndProject
+Global
+	GlobalSection(SolutionConfigurationPlatforms) = preSolution
+		Debug|x64 = Debug|x64
+		Release|x64 = Release|x64
+	EndGlobalSection
+	GlobalSection(ProjectConfigurationPlatforms) = postSolution
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
+		{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
+	EndGlobalSection
+	GlobalSection(SolutionProperties) = preSolution
+		HideSolutionNode = FALSE
+	EndGlobalSection
+EndGlobal

Samples/5_Domain_Specific/HSOpticalFlow/HSOpticalFlow_vs2022.vcxproj

@@ -0,0 +1,118 @@
+<?xml version="1.0" encoding="utf-8"?>
+<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+  <PropertyGroup>
+    <CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
+  </PropertyGroup>
+  <ItemGroup Label="ProjectConfigurations">
+    <ProjectConfiguration Include="Debug|x64">
+      <Configuration>Debug</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|x64">
+      <Configuration>Release</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+  </ItemGroup>
+  <PropertyGroup Label="Globals">
+    <ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
+    <RootNamespace>HSOpticalFlow_vs2022</RootNamespace>
+    <ProjectName>HSOpticalFlow</ProjectName>
+    <CudaToolkitCustomDir />
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
+  <PropertyGroup>
+    <ConfigurationType>Application</ConfigurationType>
+    <CharacterSet>MultiByte</CharacterSet>
+    <PlatformToolset>v143</PlatformToolset>
+    <WindowsTargetPlatformVersion>10.0</WindowsTargetPlatformVersion>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)'=='Debug'">
+    <UseDebugLibraries>true</UseDebugLibraries>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)'=='Release'">
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
+  <ImportGroup Label="ExtensionSettings">
+    <Import Project="$(CUDAPropsPath)\CUDA 11.6.props" />
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets">
+    <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
+  </ImportGroup>
+  <PropertyGroup Label="UserMacros" />
+  <PropertyGroup>
+    <IntDir>$(Platform)/$(Configuration)/</IntDir>
+    <IncludePath>$(IncludePath)</IncludePath>
+    <CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
+    <CodeAnalysisRules />
+    <CodeAnalysisRuleAssemblies />
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Platform)'=='x64'">
+    <OutDir>../../../bin/win64/$(Configuration)/</OutDir>
+  </PropertyGroup>
+  <ItemDefinitionGroup>
+    <ClCompile>
+      <WarningLevel>Level3</WarningLevel>
+      <PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../../Common;</AdditionalIncludeDirectories>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
+      <OutputFile>$(OutDir)/HSOpticalFlow.exe</OutputFile>
+    </Link>
+    <CudaCompile>
+      <CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;compute_86,sm_86;</CodeGeneration>
+      <AdditionalOptions>-Xcompiler "/wd 4819"  --threads 0 </AdditionalOptions>
+      <Include>./;../../../Common</Include>
+      <Defines>WIN32</Defines>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
+    <ClCompile>
+      <Optimization>Disabled</Optimization>
+      <RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
+    </ClCompile>
+    <Link>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
+    </Link>
+    <CudaCompile>
+      <Runtime>MTd</Runtime>
+      <TargetMachinePlatform>64</TargetMachinePlatform>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
+    <ClCompile>
+      <Optimization>MaxSpeed</Optimization>
+      <RuntimeLibrary>MultiThreaded</RuntimeLibrary>
+    </ClCompile>
+    <Link>
+      <GenerateDebugInformation>false</GenerateDebugInformation>
+      <LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
+    </Link>
+    <CudaCompile>
+      <Runtime>MT</Runtime>
+      <TargetMachinePlatform>64</TargetMachinePlatform>
+    </CudaCompile>
+  </ItemDefinitionGroup>
+  <ItemGroup>
+    <CudaCompile Include="flowCUDA.cu" />
+    <ClCompile Include="flowGold.cpp" />
+    <ClCompile Include="main.cpp" />
+    <ClInclude Include="common.h" />
+    <ClInclude Include="flowCUDA.h" />
+    <ClInclude Include="flowGold.h" />
+    <None Include="addKernel.cuh" />
+    <None Include="derivativesKernel.cuh" />
+    <None Include="downscaleKernel.cuh" />
+    <None Include="solverKernel.cuh" />
+    <None Include="upscaleKernel.cuh" />
+    <None Include="warpingKernel.cuh" />
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+    <Import Project="$(CUDAPropsPath)\CUDA 11.6.targets" />
+  </ImportGroup>
+</Project>

Samples/5_Domain_Specific/HSOpticalFlow/Makefile

@@ -0,0 +1,346 @@
+################################################################################
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+#
+################################################################################
+#
+# Makefile project only supported on Mac OS X and Linux Platforms)
+#
+################################################################################
+
+# Location of the CUDA Toolkit
+CUDA_PATH ?= /usr/local/cuda
+
+##############################
+# start deprecated interface #
+##############################
+ifeq ($(x86_64),1)
+    $(info WARNING - x86_64 variable has been deprecated)
+    $(info WARNING - please use TARGET_ARCH=x86_64 instead)
+    TARGET_ARCH ?= x86_64
+endif
+ifeq ($(ARMv7),1)
+    $(info WARNING - ARMv7 variable has been deprecated)
+    $(info WARNING - please use TARGET_ARCH=armv7l instead)
+    TARGET_ARCH ?= armv7l
+endif
+ifeq ($(aarch64),1)
+    $(info WARNING - aarch64 variable has been deprecated)
+    $(info WARNING - please use TARGET_ARCH=aarch64 instead)
+    TARGET_ARCH ?= aarch64
+endif
+ifeq ($(ppc64le),1)
+    $(info WARNING - ppc64le variable has been deprecated)
+    $(info WARNING - please use TARGET_ARCH=ppc64le instead)
+    TARGET_ARCH ?= ppc64le
+endif
+ifneq ($(GCC),)
+    $(info WARNING - GCC variable has been deprecated)
+    $(info WARNING - please use HOST_COMPILER=$(GCC) instead)
+    HOST_COMPILER ?= $(GCC)
+endif
+ifneq ($(abi),)
+    $(error ERROR - abi variable has been removed)
+endif
+############################
+# end deprecated interface #
+############################
+
+# architecture
+HOST_ARCH   := $(shell uname -m)
+TARGET_ARCH ?= $(HOST_ARCH)
+ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
+    ifneq ($(TARGET_ARCH),$(HOST_ARCH))
+        ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
+            TARGET_SIZE := 64
+        else ifneq (,$(filter $(TARGET_ARCH),armv7l))
+            TARGET_SIZE := 32
+        endif
+    else
+        TARGET_SIZE := $(shell getconf LONG_BIT)
+    endif
+else
+    $(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
+endif
+
+# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
+ifeq ($(HOST_ARCH),aarch64)
+    ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux 2>/dev/null))
+        HOST_ARCH := sbsa
+        TARGET_ARCH := sbsa
+    endif
+endif
+
+ifneq ($(TARGET_ARCH),$(HOST_ARCH))
+    ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa x86_64-ppc64le))
+        $(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
+    endif
+endif
+
+# When on native aarch64 system with userspace of 32-bit, change TARGET_ARCH to armv7l
+ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_SIZE),aarch64-aarch64-32)
+    TARGET_ARCH = armv7l
+endif
+
+# operating system
+HOST_OS   := $(shell uname -s 2>/dev/null | tr "[:upper:]" "[:lower:]")
+TARGET_OS ?= $(HOST_OS)
+ifeq (,$(filter $(TARGET_OS),linux darwin qnx android))
+    $(error ERROR - unsupported value $(TARGET_OS) for TARGET_OS!)
+endif
+
+# host compiler
+ifeq ($(TARGET_OS),darwin)
+    ifeq ($(shell expr `xcodebuild -version | grep -i xcode | awk '{print $$2}' | cut -d'.' -f1` \>= 5),1)
+        HOST_COMPILER ?= clang++
+    endif
+else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
+    ifeq ($(HOST_ARCH)-$(TARGET_ARCH),x86_64-armv7l)
+        ifeq ($(TARGET_OS),linux)
+            HOST_COMPILER ?= arm-linux-gnueabihf-g++
+        else ifeq ($(TARGET_OS),qnx)
+            ifeq ($(QNX_HOST),)
+                $(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
+            endif
+            ifeq ($(QNX_TARGET),)
+                $(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
+            endif
+            export QNX_HOST
+            export QNX_TARGET
+            HOST_COMPILER ?= $(QNX_HOST)/usr/bin/arm-unknown-nto-qnx6.6.0eabi-g++
+        else ifeq ($(TARGET_OS),android)
+            HOST_COMPILER ?= arm-linux-androideabi-g++
+        endif
+    else ifeq ($(TARGET_ARCH),aarch64)
+        ifeq ($(TARGET_OS), linux)
+            HOST_COMPILER ?= aarch64-linux-gnu-g++
+        else ifeq ($(TARGET_OS),qnx)
+            ifeq ($(QNX_HOST),)
+                $(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
+            endif
+            ifeq ($(QNX_TARGET),)
+                $(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
+            endif
+            export QNX_HOST
+            export QNX_TARGET
+            HOST_COMPILER ?= $(QNX_HOST)/usr/bin/q++
+        else ifeq ($(TARGET_OS), android)
+            HOST_COMPILER ?= aarch64-linux-android-clang++
+        endif
+    else ifeq ($(TARGET_ARCH),sbsa)
+        HOST_COMPILER ?= aarch64-linux-gnu-g++
+    else ifeq ($(TARGET_ARCH),ppc64le)
+        HOST_COMPILER ?= powerpc64le-linux-gnu-g++
+    endif
+endif
+HOST_COMPILER ?= g++
+NVCC          := $(CUDA_PATH)/bin/nvcc -ccbin $(HOST_COMPILER)
+
+# internal flags
+NVCCFLAGS   := -m${TARGET_SIZE}
+CCFLAGS     :=
+LDFLAGS     :=
+
+# build flags
+ifeq ($(TARGET_OS),darwin)
+    LDFLAGS += -rpath $(CUDA_PATH)/lib
+    CCFLAGS += -arch $(HOST_ARCH)
+else ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_OS),x86_64-armv7l-linux)
+    LDFLAGS += --dynamic-linker=/lib/ld-linux-armhf.so.3
+    CCFLAGS += -mfloat-abi=hard
+else ifeq ($(TARGET_OS),android)
+    LDFLAGS += -pie
+    CCFLAGS += -fpie -fpic -fexceptions
+endif
+
+ifneq ($(TARGET_ARCH),$(HOST_ARCH))
+    ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
+        ifneq ($(TARGET_FS),)
+            GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
+            ifeq ($(GCCVERSIONLTEQ46),1)
+                CCFLAGS += --sysroot=$(TARGET_FS)
+            endif
+            LDFLAGS += --sysroot=$(TARGET_FS)
+            LDFLAGS += -rpath-link=$(TARGET_FS)/lib
+            LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib
+            LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/arm-linux-gnueabihf
+        endif
+    endif
+    ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
+        ifneq ($(TARGET_FS),)
+            GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
+            ifeq ($(GCCVERSIONLTEQ46),1)
+                CCFLAGS += --sysroot=$(TARGET_FS)
+            endif
+            LDFLAGS += --sysroot=$(TARGET_FS)
+            LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L$(TARGET_FS)/lib
+            LDFLAGS += -rpath-link=$(TARGET_FS)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
+            LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L$(TARGET_FS)/usr/lib
+            LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L$(TARGET_FS)/usr/lib/aarch64-linux-gnu
+            LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
+            CCFLAGS += -isystem=$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include/libdrm
+            CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
+        endif
+    endif
+    ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
+        NVCCFLAGS += -D_QNX_SOURCE
+        NVCCFLAGS += --qpp-config 8.3.0,gcc_ntoaarch64le
+        CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
+        LDFLAGS += -lsocket
+        LDFLAGS += -L/usr/lib/aarch64-qnx-gnu
+        CCFLAGS += "-Wl\,-rpath-link\,/usr/lib/aarch64-qnx-gnu"
+        ifdef TARGET_OVERRIDE
+            LDFLAGS += -lslog2
+        endif
+
+        ifneq ($(TARGET_FS),)
+            LDFLAGS += -L$(TARGET_FS)/usr/lib
+            CCFLAGS += "-Wl\,-rpath-link\,$(TARGET_FS)/usr/lib"
+            LDFLAGS += -L$(TARGET_FS)/usr/libnvidia
+            CCFLAGS += "-Wl\,-rpath-link\,$(TARGET_FS)/usr/libnvidia"
+            CCFLAGS += -I$(TARGET_FS)/../include
+        endif
+    endif
+endif
+
+ifdef TARGET_OVERRIDE # cuda toolkit targets override
+    NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
+endif
+
+# Install directory of different arch
+CUDA_INSTALL_TARGET_DIR :=
+ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
+    CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
+else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
+    CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
+else ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
+    CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
+else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
+    CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
+else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
+    CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux-androideabi/
+else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-qnx)
+    CUDA_INSTALL_TARGET_DIR = targets/ARMv7-linux-QNX/
+else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
+    CUDA_INSTALL_TARGET_DIR = targets/aarch64-qnx/
+else ifeq ($(TARGET_ARCH),ppc64le)
+    CUDA_INSTALL_TARGET_DIR = targets/ppc64le-linux/
+endif
+
+# Debug build flags
+ifeq ($(dbg),1)
+      NVCCFLAGS += -g -G
+      BUILD_TYPE := debug
+else
+      BUILD_TYPE := release
+endif
+
+ALL_CCFLAGS :=
+ALL_CCFLAGS += $(NVCCFLAGS)
+ALL_CCFLAGS += $(EXTRA_NVCCFLAGS)
+ALL_CCFLAGS += $(addprefix -Xcompiler ,$(CCFLAGS))
+ALL_CCFLAGS += $(addprefix -Xcompiler ,$(EXTRA_CCFLAGS))
+
+SAMPLE_ENABLED := 1
+
+ALL_LDFLAGS :=
+ALL_LDFLAGS += $(ALL_CCFLAGS)
+ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
+ALL_LDFLAGS += $(addprefix -Xlinker ,$(EXTRA_LDFLAGS))
+
+# Common includes and paths for CUDA
+INCLUDES  := -I../../../Common
+LIBRARIES :=
+
+################################################################################
+
+# Gencode arguments
+ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64 sbsa))
+SMS ?= 53 61 70 72 75 80 86 87
+else
+SMS ?= 35 37 50 52 60 61 70 75 80 86
+endif
+
+ifeq ($(SMS),)
+$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)
+SAMPLE_ENABLED := 0
+endif
+
+ifeq ($(GENCODE_FLAGS),)
+# Generate SASS code for each SM architecture listed in $(SMS)
+$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))
+
+# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility
+HIGHEST_SM := $(lastword $(sort $(SMS)))
+ifneq ($(HIGHEST_SM),)
+GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
+endif
+endif
+
+ALL_CCFLAGS += --threads 0 --std=c++11
+
+ifeq ($(SAMPLE_ENABLED),0)
+EXEC ?= @echo "[@]"
+endif
+
+################################################################################
+
+# Target rules
+all: build
+
+build: HSOpticalFlow
+
+check.deps:
+ifeq ($(SAMPLE_ENABLED),0)
+	@echo "Sample will be waived due to the above missing dependencies"
+else
+	@echo "Sample is ready - all dependencies have been met"
+endif
+
+flowCUDA.o:flowCUDA.cu
+	$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
+
+flowGold.o:flowGold.cpp
+	$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
+
+main.o:main.cpp
+	$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
+
+HSOpticalFlow: flowCUDA.o flowGold.o main.o
+	$(EXEC) $(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)
+	$(EXEC) mkdir -p ../../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
+	$(EXEC) cp $@ ../../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
+
+run: build
+	$(EXEC) ./HSOpticalFlow
+
+testrun: build
+
+clean:
+	rm -f HSOpticalFlow flowCUDA.o flowGold.o main.o
+	rm -rf ../../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/HSOpticalFlow
+
+clobber: clean

Samples/5_Domain_Specific/HSOpticalFlow/NsightEclipse.xml

@@ -0,0 +1,84 @@
+<?xml version="1.0" encoding="UTF-8"?> 
+<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
+<entry>
+  <name>HSOpticalFlow</name>
+  <cuda_api_list>
+    <toolkit>cudaMalloc</toolkit>
+    <toolkit>cudaMemset</toolkit>
+    <toolkit>cudaFree</toolkit>
+    <toolkit>cudaMemcpy</toolkit>
+  </cuda_api_list>
+  <description><![CDATA[Variational optical flow estimation example.  Uses textures for image operations. Shows how simple PDE solver can be accelerated with CUDA.]]></description>
+  <devicecompilation>whole</devicecompilation>
+  <files>
+    <file>data\frame10.ppm</file>
+    <file>data\frame11.ppm</file>
+  </files>
+  <includepaths>
+    <path>./</path>
+    <path>../</path>
+    <path>../../../Common</path>
+  </includepaths>
+  <keyconcepts>
+    <concept level="advanced">Image Processing</concept>
+    <concept level="advanced">Data Parallel Algorithms</concept>
+  </keyconcepts>
+  <keywords>
+    <keyword>optical flow</keyword>
+    <keyword>motion estimation</keyword>
+    <keyword>texture</keyword>
+    <keyword>shared memory</keyword>
+  </keywords>
+  <libraries>
+  </libraries>
+  <librarypaths>
+  </librarypaths>
+  <nsight_eclipse>true</nsight_eclipse>
+  <primary_file>main.cpp</primary_file>
+  <scopes>
+    <scope>1:CUDA Advanced Topics</scope>
+    <scope>2:Image Processing</scope>
+  </scopes>
+  <sm-arch>sm35</sm-arch>
+  <sm-arch>sm37</sm-arch>
+  <sm-arch>sm50</sm-arch>
+  <sm-arch>sm52</sm-arch>
+  <sm-arch>sm53</sm-arch>
+  <sm-arch>sm60</sm-arch>
+  <sm-arch>sm61</sm-arch>
+  <sm-arch>sm70</sm-arch>
+  <sm-arch>sm72</sm-arch>
+  <sm-arch>sm75</sm-arch>
+  <sm-arch>sm80</sm-arch>
+  <sm-arch>sm86</sm-arch>
+  <sm-arch>sm87</sm-arch>
+  <supported_envs>
+    <env>
+      <arch>x86_64</arch>
+      <platform>linux</platform>
+    </env>
+    <env>
+      <platform>windows7</platform>
+    </env>
+    <env>
+      <arch>x86_64</arch>
+      <platform>macosx</platform>
+    </env>
+    <env>
+      <arch>arm</arch>
+    </env>
+    <env>
+      <arch>sbsa</arch>
+    </env>
+    <env>
+      <arch>ppc64le</arch>
+      <platform>linux</platform>
+    </env>
+  </supported_envs>
+  <supported_sm_architectures>
+    <include>all</include>
+  </supported_sm_architectures>
+  <title>Optical Flow</title>
+  <type>exe</type>
+  <whitepaper>doc\OpticalFlow.pdf</whitepaper>
+</entry>

Samples/5_Domain_Specific/HSOpticalFlow/README.md

@@ -0,0 +1,71 @@
+# HSOpticalFlow - Optical Flow
+
+## Description
+
+Variational optical flow estimation example.  Uses textures for image operations. Shows how simple PDE solver can be accelerated with CUDA.
+
+## Key Concepts
+
+Image Processing, Data Parallel Algorithms
+
+## Supported SM Architectures
+
+[SM 3.5 ](https://developer.nvidia.com/cuda-gpus)  [SM 3.7 ](https://developer.nvidia.com/cuda-gpus)  [SM 5.0 ](https://developer.nvidia.com/cuda-gpus)  [SM 5.2 ](https://developer.nvidia.com/cuda-gpus)  [SM 5.3 ](https://developer.nvidia.com/cuda-gpus)  [SM 6.0 ](https://developer.nvidia.com/cuda-gpus)  [SM 6.1 ](https://developer.nvidia.com/cuda-gpus)  [SM 7.0 ](https://developer.nvidia.com/cuda-gpus)  [SM 7.2 ](https://developer.nvidia.com/cuda-gpus)  [SM 7.5 ](https://developer.nvidia.com/cuda-gpus)  [SM 8.0 ](https://developer.nvidia.com/cuda-gpus)  [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)  [SM 8.7 ](https://developer.nvidia.com/cuda-gpus)
+
+## Supported OSes
+
+Linux, Windows
+
+## Supported CPU Architecture
+
+x86_64, ppc64le, armv7l
+
+## CUDA APIs involved
+
+### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
+cudaMalloc, cudaMemset, cudaFree, cudaMemcpy
+
+## Prerequisites
+
+Download and install the [CUDA Toolkit 11.6](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
+
+## Build and Run
+
+### Windows
+The Windows samples are built using the Visual Studio IDE. Solution files (.sln) are provided for each supported version of Visual Studio, using the format:
+```
+*_vs<version>.sln - for Visual Studio <version>
+```
+Each individual sample has its own set of solution files in its directory:
+
+To build/examine all the samples at once, the complete solution files should be used. To build/examine a single sample, the individual sample solution files should be used.
+> **Note:** Some samples require that the Microsoft DirectX SDK (June 2010 or newer) be installed and that the VC++ directory paths are properly set up (**Tools > Options...**). Check DirectX Dependencies section for details."
+
+### Linux
+The Linux samples are built using makefiles. To use the makefiles, change the current directory to the sample directory you wish to build, and run make:
+```
+$ cd <sample_dir>
+$ make
+```
+The samples makefiles can take advantage of certain options:
+*  **TARGET_ARCH=<arch>** - cross-compile targeting a specific architecture. Allowed architectures are x86_64, ppc64le, armv7l.
+    By default, TARGET_ARCH is set to HOST_ARCH. On a x86_64 machine, not setting TARGET_ARCH is the equivalent of setting TARGET_ARCH=x86_64.<br/>
+`$ make TARGET_ARCH=x86_64` <br/> `$ make TARGET_ARCH=ppc64le` <br/> `$ make TARGET_ARCH=armv7l` <br/>
+    See [here](http://docs.nvidia.com/cuda/cuda-samples/index.html#cross-samples) for more details.
+*   **dbg=1** - build with debug symbols
+    ```
+    $ make dbg=1
+    ```
+*   **SMS="A B ..."** - override the SM architectures for which the sample will be built, where `"A B ..."` is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use `SMS="50 60"`.
+    ```
+    $ make SMS="50 60"
+    ```
+
+*  **HOST_COMPILER=<host_compiler>** - override the default g++ host compiler. See the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements) for a list of supported host compilers.
+```
+    $ make HOST_COMPILER=g++
+```
+
+## References (for more details)
+
+[whitepaper](./doc/OpticalFlow.pdf)

Samples/5_Domain_Specific/HSOpticalFlow/addKernel.cuh

@@ -0,0 +1,60 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "common.h"
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief add two vectors of size _count_
+///
+/// CUDA kernel
+/// \param[in]  op1   term one
+/// \param[in]  op2   term two
+/// \param[in]  count vector size
+/// \param[out] sum   result
+///////////////////////////////////////////////////////////////////////////////
+__global__ void AddKernel(const float *op1, const float *op2, int count,
+                          float *sum) {
+  const int pos = threadIdx.x + blockIdx.x * blockDim.x;
+
+  if (pos >= count) return;
+
+  sum[pos] = op1[pos] + op2[pos];
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief add two vectors of size _count_
+/// \param[in]  op1   term one
+/// \param[in]  op2   term two
+/// \param[in]  count vector size
+/// \param[out] sum   result
+///////////////////////////////////////////////////////////////////////////////
+static void Add(const float *op1, const float *op2, int count, float *sum) {
+  dim3 threads(256);
+  dim3 blocks(iDivUp(count, threads.x));
+
+  AddKernel<<<blocks, threads>>>(op1, op2, count, sum);
+}

Samples/5_Domain_Specific/HSOpticalFlow/common.h

@@ -0,0 +1,76 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+///////////////////////////////////////////////////////////////////////////////
+// Header for common includes and utility functions
+///////////////////////////////////////////////////////////////////////////////
+
+#ifndef COMMON_H
+#define COMMON_H
+
+///////////////////////////////////////////////////////////////////////////////
+// Common includes
+///////////////////////////////////////////////////////////////////////////////
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <time.h>
+#include <memory.h>
+#include <math.h>
+
+#include <helper_cuda.h>
+
+///////////////////////////////////////////////////////////////////////////////
+// Common constants
+///////////////////////////////////////////////////////////////////////////////
+const int StrideAlignment = 32;
+
+///////////////////////////////////////////////////////////////////////////////
+// Common functions
+///////////////////////////////////////////////////////////////////////////////
+
+// Align up n to the nearest multiple of m
+inline int iAlignUp(int n, int m = StrideAlignment) {
+  int mod = n % m;
+
+  if (mod)
+    return n + m - mod;
+  else
+    return n;
+}
+
+// round up n/m
+inline int iDivUp(int n, int m) { return (n + m - 1) / m; }
+
+// swap two values
+template <typename T>
+inline void Swap(T &a, T &b) {
+  T t = a;
+  a = b;
+  b = t;
+}
+#endif

Samples/5_Domain_Specific/HSOpticalFlow/derivativesKernel.cuh

@@ -0,0 +1,144 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "common.h"
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief compute image derivatives
+///
+/// CUDA kernel, relies heavily on texture unit
+/// \param[in]  width   image width
+/// \param[in]  height  image height
+/// \param[in]  stride  image stride
+/// \param[out] Ix      x derivative
+/// \param[out] Iy      y derivative
+/// \param[out] Iz      temporal derivative
+///////////////////////////////////////////////////////////////////////////////
+__global__ void ComputeDerivativesKernel(int width, int height, int stride,
+                                         float *Ix, float *Iy, float *Iz,
+                                         cudaTextureObject_t texSource,
+                                         cudaTextureObject_t texTarget) {
+  const int ix = threadIdx.x + blockIdx.x * blockDim.x;
+  const int iy = threadIdx.y + blockIdx.y * blockDim.y;
+
+  const int pos = ix + iy * stride;
+
+  if (ix >= width || iy >= height) return;
+
+  float dx = 1.0f / (float)width;
+  float dy = 1.0f / (float)height;
+
+  float x = ((float)ix + 0.5f) * dx;
+  float y = ((float)iy + 0.5f) * dy;
+
+  float t0, t1;
+  // x derivative
+  t0 = tex2D<float>(texSource, x - 2.0f * dx, y);
+  t0 -= tex2D<float>(texSource, x - 1.0f * dx, y) * 8.0f;
+  t0 += tex2D<float>(texSource, x + 1.0f * dx, y) * 8.0f;
+  t0 -= tex2D<float>(texSource, x + 2.0f * dx, y);
+  t0 /= 12.0f;
+
+  t1 = tex2D<float>(texTarget, x - 2.0f * dx, y);
+  t1 -= tex2D<float>(texTarget, x - 1.0f * dx, y) * 8.0f;
+  t1 += tex2D<float>(texTarget, x + 1.0f * dx, y) * 8.0f;
+  t1 -= tex2D<float>(texTarget, x + 2.0f * dx, y);
+  t1 /= 12.0f;
+
+  Ix[pos] = (t0 + t1) * 0.5f;
+
+  // t derivative
+  Iz[pos] = tex2D<float>(texTarget, x, y) - tex2D<float>(texSource, x, y);
+
+  // y derivative
+  t0 = tex2D<float>(texSource, x, y - 2.0f * dy);
+  t0 -= tex2D<float>(texSource, x, y - 1.0f * dy) * 8.0f;
+  t0 += tex2D<float>(texSource, x, y + 1.0f * dy) * 8.0f;
+  t0 -= tex2D<float>(texSource, x, y + 2.0f * dy);
+  t0 /= 12.0f;
+
+  t1 = tex2D<float>(texTarget, x, y - 2.0f * dy);
+  t1 -= tex2D<float>(texTarget, x, y - 1.0f * dy) * 8.0f;
+  t1 += tex2D<float>(texTarget, x, y + 1.0f * dy) * 8.0f;
+  t1 -= tex2D<float>(texTarget, x, y + 2.0f * dy);
+  t1 /= 12.0f;
+
+  Iy[pos] = (t0 + t1) * 0.5f;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief compute image derivatives
+///
+/// \param[in]  I0  source image
+/// \param[in]  I1  tracked image
+/// \param[in]  w   image width
+/// \param[in]  h   image height
+/// \param[in]  s   image stride
+/// \param[out] Ix  x derivative
+/// \param[out] Iy  y derivative
+/// \param[out] Iz  temporal derivative
+///////////////////////////////////////////////////////////////////////////////
+static void ComputeDerivatives(const float *I0, const float *I1, int w, int h,
+                               int s, float *Ix, float *Iy, float *Iz) {
+  dim3 threads(32, 6);
+  dim3 blocks(iDivUp(w, threads.x), iDivUp(h, threads.y));
+
+  cudaTextureObject_t texSource, texTarget;
+  cudaResourceDesc texRes;
+  memset(&texRes, 0, sizeof(cudaResourceDesc));
+
+  texRes.resType = cudaResourceTypePitch2D;
+  texRes.res.pitch2D.devPtr = (void *)I0;
+  texRes.res.pitch2D.desc = cudaCreateChannelDesc<float>();
+  texRes.res.pitch2D.width = w;
+  texRes.res.pitch2D.height = h;
+  texRes.res.pitch2D.pitchInBytes = s * sizeof(float);
+
+  cudaTextureDesc texDescr;
+  memset(&texDescr, 0, sizeof(cudaTextureDesc));
+
+  texDescr.normalizedCoords = true;
+  texDescr.filterMode = cudaFilterModeLinear;
+  texDescr.addressMode[0] = cudaAddressModeMirror;
+  texDescr.addressMode[1] = cudaAddressModeMirror;
+  texDescr.readMode = cudaReadModeElementType;
+
+  checkCudaErrors(
+      cudaCreateTextureObject(&texSource, &texRes, &texDescr, NULL));
+  memset(&texRes, 0, sizeof(cudaResourceDesc));
+  texRes.resType = cudaResourceTypePitch2D;
+  texRes.res.pitch2D.devPtr = (void *)I1;
+  texRes.res.pitch2D.desc = cudaCreateChannelDesc<float>();
+  texRes.res.pitch2D.width = w;
+  texRes.res.pitch2D.height = h;
+  texRes.res.pitch2D.pitchInBytes = s * sizeof(float);
+  checkCudaErrors(
+      cudaCreateTextureObject(&texTarget, &texRes, &texDescr, NULL));
+
+  ComputeDerivativesKernel<<<blocks, threads>>>(w, h, s, Ix, Iy, Iz, texSource,
+                                                texTarget);
+}

Samples/5_Domain_Specific/HSOpticalFlow/downscaleKernel.cuh

@@ -0,0 +1,98 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "common.h"
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief downscale image
+///
+/// CUDA kernel, relies heavily on texture unit
+/// \param[in]  width   image width
+/// \param[in]  height  image height
+/// \param[in]  stride  image stride
+/// \param[out] out     result
+///////////////////////////////////////////////////////////////////////////////
+__global__ void DownscaleKernel(int width, int height, int stride, float *out,
+                                cudaTextureObject_t texFine) {
+  const int ix = threadIdx.x + blockIdx.x * blockDim.x;
+  const int iy = threadIdx.y + blockIdx.y * blockDim.y;
+
+  if (ix >= width || iy >= height) {
+    return;
+  }
+
+  float dx = 1.0f / (float)width;
+  float dy = 1.0f / (float)height;
+
+  float x = ((float)ix + 0.5f) * dx;
+  float y = ((float)iy + 0.5f) * dy;
+
+  out[ix + iy * stride] = 0.25f * (tex2D<float>(texFine, x - dx * 0.25f, y) +
+                                   tex2D<float>(texFine, x + dx * 0.25f, y) +
+                                   tex2D<float>(texFine, x, y - dy * 0.25f) +
+                                   tex2D<float>(texFine, x, y + dy * 0.25f));
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief downscale image
+///
+/// \param[in]  src     image to downscale
+/// \param[in]  width   image width
+/// \param[in]  height  image height
+/// \param[in]  stride  image stride
+/// \param[out] out     result
+///////////////////////////////////////////////////////////////////////////////
+static void Downscale(const float *src, int width, int height, int stride,
+                      int newWidth, int newHeight, int newStride, float *out) {
+  dim3 threads(32, 8);
+  dim3 blocks(iDivUp(newWidth, threads.x), iDivUp(newHeight, threads.y));
+
+  cudaTextureObject_t texFine;
+  cudaResourceDesc texRes;
+  memset(&texRes, 0, sizeof(cudaResourceDesc));
+
+  texRes.resType = cudaResourceTypePitch2D;
+  texRes.res.pitch2D.devPtr = (void *)src;
+  texRes.res.pitch2D.desc = cudaCreateChannelDesc<float>();
+  texRes.res.pitch2D.width = width;
+  texRes.res.pitch2D.height = height;
+  texRes.res.pitch2D.pitchInBytes = stride * sizeof(float);
+
+  cudaTextureDesc texDescr;
+  memset(&texDescr, 0, sizeof(cudaTextureDesc));
+
+  texDescr.normalizedCoords = true;
+  texDescr.filterMode = cudaFilterModeLinear;
+  texDescr.addressMode[0] = cudaAddressModeMirror;
+  texDescr.addressMode[1] = cudaAddressModeMirror;
+  texDescr.readMode = cudaReadModeElementType;
+
+  checkCudaErrors(cudaCreateTextureObject(&texFine, &texRes, &texDescr, NULL));
+
+  DownscaleKernel<<<blocks, threads>>>(newWidth, newHeight, newStride, out,
+                                       texFine);
+}

Samples/5_Domain_Specific/HSOpticalFlow/flowCUDA.cu

@@ -0,0 +1,216 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "common.h"
+
+// include kernels
+#include "downscaleKernel.cuh"
+#include "upscaleKernel.cuh"
+#include "warpingKernel.cuh"
+#include "derivativesKernel.cuh"
+#include "solverKernel.cuh"
+#include "addKernel.cuh"
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief method logic
+///
+/// handles memory allocations, control flow
+/// \param[in]  I0           source image
+/// \param[in]  I1           tracked image
+/// \param[in]  width        images width
+/// \param[in]  height       images height
+/// \param[in]  stride       images stride
+/// \param[in]  alpha        degree of displacement field smoothness
+/// \param[in]  nLevels      number of levels in a pyramid
+/// \param[in]  nWarpIters   number of warping iterations per pyramid level
+/// \param[in]  nSolverIters number of solver iterations (Jacobi iterations)
+/// \param[out] u            horizontal displacement
+/// \param[out] v            vertical displacement
+///////////////////////////////////////////////////////////////////////////////
+void ComputeFlowCUDA(const float *I0, const float *I1, int width, int height,
+                     int stride, float alpha, int nLevels, int nWarpIters,
+                     int nSolverIters, float *u, float *v) {
+  printf("Computing optical flow on GPU...\n");
+
+  // pI0 and pI1 will hold device pointers
+  const float **pI0 = new const float *[nLevels];
+  const float **pI1 = new const float *[nLevels];
+
+  int *pW = new int[nLevels];
+  int *pH = new int[nLevels];
+  int *pS = new int[nLevels];
+
+  // device memory pointers
+  float *d_tmp;
+  float *d_du0;
+  float *d_dv0;
+  float *d_du1;
+  float *d_dv1;
+
+  float *d_Ix;
+  float *d_Iy;
+  float *d_Iz;
+
+  float *d_u;
+  float *d_v;
+  float *d_nu;
+  float *d_nv;
+
+  const int dataSize = stride * height * sizeof(float);
+
+  checkCudaErrors(cudaMalloc(&d_tmp, dataSize));
+  checkCudaErrors(cudaMalloc(&d_du0, dataSize));
+  checkCudaErrors(cudaMalloc(&d_dv0, dataSize));
+  checkCudaErrors(cudaMalloc(&d_du1, dataSize));
+  checkCudaErrors(cudaMalloc(&d_dv1, dataSize));
+
+  checkCudaErrors(cudaMalloc(&d_Ix, dataSize));
+  checkCudaErrors(cudaMalloc(&d_Iy, dataSize));
+  checkCudaErrors(cudaMalloc(&d_Iz, dataSize));
+
+  checkCudaErrors(cudaMalloc(&d_u, dataSize));
+  checkCudaErrors(cudaMalloc(&d_v, dataSize));
+  checkCudaErrors(cudaMalloc(&d_nu, dataSize));
+  checkCudaErrors(cudaMalloc(&d_nv, dataSize));
+
+  // prepare pyramid
+
+  int currentLevel = nLevels - 1;
+  // allocate GPU memory for input images
+  checkCudaErrors(cudaMalloc(pI0 + currentLevel, dataSize));
+  checkCudaErrors(cudaMalloc(pI1 + currentLevel, dataSize));
+
+  checkCudaErrors(cudaMemcpy((void *)pI0[currentLevel], I0, dataSize,
+                             cudaMemcpyHostToDevice));
+  checkCudaErrors(cudaMemcpy((void *)pI1[currentLevel], I1, dataSize,
+                             cudaMemcpyHostToDevice));
+
+  pW[currentLevel] = width;
+  pH[currentLevel] = height;
+  pS[currentLevel] = stride;
+
+  for (; currentLevel > 0; --currentLevel) {
+    int nw = pW[currentLevel] / 2;
+    int nh = pH[currentLevel] / 2;
+    int ns = iAlignUp(nw);
+
+    checkCudaErrors(
+        cudaMalloc(pI0 + currentLevel - 1, ns * nh * sizeof(float)));
+    checkCudaErrors(
+        cudaMalloc(pI1 + currentLevel - 1, ns * nh * sizeof(float)));
+
+    Downscale(pI0[currentLevel], pW[currentLevel], pH[currentLevel],
+              pS[currentLevel], nw, nh, ns, (float *)pI0[currentLevel - 1]);
+
+    Downscale(pI1[currentLevel], pW[currentLevel], pH[currentLevel],
+              pS[currentLevel], nw, nh, ns, (float *)pI1[currentLevel - 1]);
+
+    pW[currentLevel - 1] = nw;
+    pH[currentLevel - 1] = nh;
+    pS[currentLevel - 1] = ns;
+  }
+
+  checkCudaErrors(cudaMemset(d_u, 0, stride * height * sizeof(float)));
+  checkCudaErrors(cudaMemset(d_v, 0, stride * height * sizeof(float)));
+
+  // compute flow
+  for (; currentLevel < nLevels; ++currentLevel) {
+    for (int warpIter = 0; warpIter < nWarpIters; ++warpIter) {
+      checkCudaErrors(cudaMemset(d_du0, 0, dataSize));
+      checkCudaErrors(cudaMemset(d_dv0, 0, dataSize));
+
+      checkCudaErrors(cudaMemset(d_du1, 0, dataSize));
+      checkCudaErrors(cudaMemset(d_dv1, 0, dataSize));
+
+      // on current level we compute optical flow
+      // between frame 0 and warped frame 1
+      WarpImage(pI1[currentLevel], pW[currentLevel], pH[currentLevel],
+                pS[currentLevel], d_u, d_v, d_tmp);
+
+      ComputeDerivatives(pI0[currentLevel], d_tmp, pW[currentLevel],
+                         pH[currentLevel], pS[currentLevel], d_Ix, d_Iy, d_Iz);
+
+      for (int iter = 0; iter < nSolverIters; ++iter) {
+        SolveForUpdate(d_du0, d_dv0, d_Ix, d_Iy, d_Iz, pW[currentLevel],
+                       pH[currentLevel], pS[currentLevel], alpha, d_du1, d_dv1);
+
+        Swap(d_du0, d_du1);
+        Swap(d_dv0, d_dv1);
+      }
+
+      // update u, v
+      Add(d_u, d_du0, pH[currentLevel] * pS[currentLevel], d_u);
+      Add(d_v, d_dv0, pH[currentLevel] * pS[currentLevel], d_v);
+    }
+
+    if (currentLevel != nLevels - 1) {
+      // prolongate solution
+      float scaleX = (float)pW[currentLevel + 1] / (float)pW[currentLevel];
+
+      Upscale(d_u, pW[currentLevel], pH[currentLevel], pS[currentLevel],
+              pW[currentLevel + 1], pH[currentLevel + 1], pS[currentLevel + 1],
+              scaleX, d_nu);
+
+      float scaleY = (float)pH[currentLevel + 1] / (float)pH[currentLevel];
+
+      Upscale(d_v, pW[currentLevel], pH[currentLevel], pS[currentLevel],
+              pW[currentLevel + 1], pH[currentLevel + 1], pS[currentLevel + 1],
+              scaleY, d_nv);
+
+      Swap(d_u, d_nu);
+      Swap(d_v, d_nv);
+    }
+  }
+
+  checkCudaErrors(cudaMemcpy(u, d_u, dataSize, cudaMemcpyDeviceToHost));
+  checkCudaErrors(cudaMemcpy(v, d_v, dataSize, cudaMemcpyDeviceToHost));
+
+  // cleanup
+  for (int i = 0; i < nLevels; ++i) {
+    checkCudaErrors(cudaFree((void *)pI0[i]));
+    checkCudaErrors(cudaFree((void *)pI1[i]));
+  }
+
+  delete[] pI0;
+  delete[] pI1;
+  delete[] pW;
+  delete[] pH;
+  delete[] pS;
+
+  checkCudaErrors(cudaFree(d_tmp));
+  checkCudaErrors(cudaFree(d_du0));
+  checkCudaErrors(cudaFree(d_dv0));
+  checkCudaErrors(cudaFree(d_du1));
+  checkCudaErrors(cudaFree(d_dv1));
+  checkCudaErrors(cudaFree(d_Ix));
+  checkCudaErrors(cudaFree(d_Iy));
+  checkCudaErrors(cudaFree(d_Iz));
+  checkCudaErrors(cudaFree(d_nu));
+  checkCudaErrors(cudaFree(d_nv));
+  checkCudaErrors(cudaFree(d_u));
+  checkCudaErrors(cudaFree(d_v));
+}

Samples/5_Domain_Specific/HSOpticalFlow/flowCUDA.h

@@ -0,0 +1,43 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef FLOW_CUDA_H
+#define FLOW_CUDA_H
+
+void ComputeFlowCUDA(
+    const float *I0,   // source frame
+    const float *I1,   // tracked frame
+    int width,         // frame width
+    int height,        // frame height
+    int stride,        // row access stride
+    float alpha,       // smoothness coefficient
+    int nLevels,       // number of levels in pyramid
+    int nWarpIters,    // number of warping iterations per pyramid level
+    int nSolverIters,  // number of solver iterations (for linear system)
+    float *u,          // output horizontal flow
+    float *v);         // output vertical flow
+#endif

Samples/5_Domain_Specific/HSOpticalFlow/flowGold.cpp

@@ -0,0 +1,462 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "common.h"
+#include "flowGold.h"
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief host texture fetch
+///
+/// read from arbitrary position within image using bilinear interpolation
+/// out of range coords are mirrored
+/// \param[in]  t   texture raw data
+/// \param[in]  w   texture width
+/// \param[in]  h   texture height
+/// \param[in]  s   texture stride
+/// \param[in]  x   x coord of the point to fetch value at
+/// \param[in]  y   y coord of the point to fetch value at
+/// \return fetched value
+///////////////////////////////////////////////////////////////////////////////
+inline float Tex2D(const float *t, int w, int h, int s, float x, float y) {
+  // integer parts in floating point format
+  float intPartX, intPartY;
+
+  // get fractional parts of coordinates
+  float dx = fabsf(modff(x, &intPartX));
+  float dy = fabsf(modff(y, &intPartY));
+
+  // assume pixels are squares
+  // one of the corners
+  int ix0 = (int)intPartX;
+  int iy0 = (int)intPartY;
+
+  // mirror out-of-range position
+  if (ix0 < 0) ix0 = abs(ix0 + 1);
+
+  if (iy0 < 0) iy0 = abs(iy0 + 1);
+
+  if (ix0 >= w) ix0 = w * 2 - ix0 - 1;
+
+  if (iy0 >= h) iy0 = h * 2 - iy0 - 1;
+
+  // corner which is opposite to (ix0, iy0)
+  int ix1 = ix0 + 1;
+  int iy1 = iy0 + 1;
+
+  if (ix1 >= w) ix1 = w * 2 - ix1 - 1;
+
+  if (iy1 >= h) iy1 = h * 2 - iy1 - 1;
+
+  float res = t[ix0 + iy0 * s] * (1.0f - dx) * (1.0f - dy);
+  res += t[ix1 + iy0 * s] * dx * (1.0f - dy);
+  res += t[ix0 + iy1 * s] * (1.0f - dx) * dy;
+  res += t[ix1 + iy1 * s] * dx * dy;
+
+  return res;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief host texture fetch
+///
+/// read specific texel value
+/// out of range coords are mirrored
+/// \param[in]  t   texture raw data
+/// \param[in]  w   texture width
+/// \param[in]  h   texture height
+/// \param[in]  s   texture stride
+/// \param[in]  x   x coord of the point to fetch value at
+/// \param[in]  y   y coord of the point to fetch value at
+/// \return fetched value
+///////////////////////////////////////////////////////////////////////////////
+inline float Tex2Di(const float *src, int w, int h, int s, int x, int y) {
+  if (x < 0) x = abs(x + 1);
+
+  if (y < 0) y = abs(y + 1);
+
+  if (x >= w) x = w * 2 - x - 1;
+
+  if (y >= h) y = h * 2 - y - 1;
+
+  return src[x + y * s];
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief resize image
+/// \param[in]  src         image to downscale
+/// \param[in]  width       image width
+/// \param[in]  height      image height
+/// \param[in]  stride      image stride
+/// \param[in]  newWidth    image new width
+/// \param[in]  newHeight   image new height
+/// \param[in]  newStride   image new stride
+/// \param[out] out         downscaled image data
+///////////////////////////////////////////////////////////////////////////////
+static void Downscale(const float *src, int width, int height, int stride,
+                      int newWidth, int newHeight, int newStride, float *out) {
+  for (int i = 0; i < newHeight; ++i) {
+    for (int j = 0; j < newWidth; ++j) {
+      const int srcX = j * 2;
+      const int srcY = i * 2;
+      // average 4 neighbouring pixels
+      float sum;
+      sum = Tex2Di(src, width, height, stride, srcX + 0, srcY + 0);
+      sum += Tex2Di(src, width, height, stride, srcX + 0, srcY + 1);
+      sum += Tex2Di(src, width, height, stride, srcX + 1, srcY + 0);
+      sum += Tex2Di(src, width, height, stride, srcX + 1, srcY + 1);
+      // normalize
+      sum *= 0.25f;
+      out[j + i * newStride] = sum;
+    }
+  }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief upscale one component of a displacement field
+/// \param[in]  src         field component to upscale
+/// \param[in]  width       field current width
+/// \param[in]  height      field current height
+/// \param[in]  stride      field current stride
+/// \param[in]  newWidth    field new width
+/// \param[in]  newHeight   field new height
+/// \param[in]  newStride   field new stride
+/// \param[in]  scale       value scale factor (multiplier)
+/// \param[out] out         upscaled field component
+///////////////////////////////////////////////////////////////////////////////
+static void Upscale(const float *src, int width, int height, int stride,
+                    int newWidth, int newHeight, int newStride, float scale,
+                    float *out) {
+  for (int i = 0; i < newHeight; ++i) {
+    for (int j = 0; j < newWidth; ++j) {
+      // position within smaller image
+      float x = ((float)j - 0.5f) * 0.5f;
+      float y = ((float)i - 0.5f) * 0.5f;
+
+      out[j + i * newStride] = Tex2D(src, width, height, stride, x, y) * scale;
+    }
+  }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief warp image with provided vector field
+///
+/// For each output pixel there is a vector which tells which pixel
+/// from a source image should be mapped to this particular output
+/// pixel.
+/// It is assumed that images and the vector field have the same stride and
+/// resolution.
+/// \param[in]  src source image
+/// \param[in]  w   width
+/// \param[in]  h   height
+/// \param[in]  s   stride
+/// \param[in]  u   horizontal displacement
+/// \param[in]  v   vertical displacement
+/// \param[out] out warped image
+///////////////////////////////////////////////////////////////////////////////
+static void WarpImage(const float *src, int w, int h, int s, const float *u,
+                      const float *v, float *out) {
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; ++j) {
+      const int pos = j + i * s;
+      // warped coords
+      float x = (float)j + u[pos];
+      float y = (float)i + v[pos];
+
+      out[pos] = Tex2D(src, w, h, s, x, y);
+    }
+  }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief computes image derivatives for a pair of images
+/// \param[in]  I0  source image
+/// \param[in]  I1  tracked image
+/// \param[in]  w   images width
+/// \param[in]  h   images height
+/// \param[in]  s   images stride
+/// \param[out] Ix  x derivative
+/// \param[out] Iy  y derivative
+/// \param[out] Iz  temporal derivative
+///////////////////////////////////////////////////////////////////////////////
+static void ComputeDerivatives(const float *I0, const float *I1, int w, int h,
+                               int s, float *Ix, float *Iy, float *Iz) {
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; ++j) {
+      const int pos = j + i * s;
+      float t0, t1;
+      // derivative filter is (1, -8, 0, 8, -1)/12
+      // x derivative
+      t0 = Tex2Di(I0, w, h, s, j - 2, i);
+      t0 -= Tex2Di(I0, w, h, s, j - 1, i) * 8.0f;
+      t0 += Tex2Di(I0, w, h, s, j + 1, i) * 8.0f;
+      t0 -= Tex2Di(I0, w, h, s, j + 2, i);
+      t0 /= 12.0f;
+
+      t1 = Tex2Di(I1, w, h, s, j - 2, i);
+      t1 -= Tex2Di(I1, w, h, s, j - 1, i) * 8.0f;
+      t1 += Tex2Di(I1, w, h, s, j + 1, i) * 8.0f;
+      t1 -= Tex2Di(I1, w, h, s, j + 2, i);
+      t1 /= 12.0f;
+
+      // spatial derivatives are averaged
+      Ix[pos] = (t0 + t1) * 0.5f;
+
+      // t derivative
+      Iz[pos] = I1[pos] - I0[pos];
+
+      // y derivative
+      t0 = Tex2Di(I0, w, h, s, j, i - 2);
+      t0 -= Tex2Di(I0, w, h, s, j, i - 1) * 8.0f;
+      t0 += Tex2Di(I0, w, h, s, j, i + 1) * 8.0f;
+      t0 -= Tex2Di(I0, w, h, s, j, i + 2);
+      t0 /= 12.0f;
+
+      t1 = Tex2Di(I1, w, h, s, j, i - 2);
+      t1 -= Tex2Di(I1, w, h, s, j, i - 1) * 8.0f;
+      t1 += Tex2Di(I1, w, h, s, j, i + 1) * 8.0f;
+      t1 -= Tex2Di(I1, w, h, s, j, i + 2);
+      t1 /= 12.0f;
+
+      Iy[pos] = (t0 + t1) * 0.5f;
+    }
+  }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief one iteration of classical Horn-Schunck method
+///
+/// It is one iteration of Jacobi method for a corresponding linear system
+/// \param[in]  du0     current horizontal displacement approximation
+/// \param[in]  dv0     current vertical displacement approximation
+/// \param[in]  Ix      image x derivative
+/// \param[in]  Iy      image y derivative
+/// \param[in]  Iz      temporal derivative
+/// \param[in]  w       width
+/// \param[in]  h       height
+/// \param[in]  s       stride
+/// \param[in]  alpha   degree of smoothness
+/// \param[out] du1     new horizontal displacement approximation
+/// \param[out] dv1     new vertical displacement approximation
+///////////////////////////////////////////////////////////////////////////////
+static void SolveForUpdate(const float *du0, const float *dv0, const float *Ix,
+                           const float *Iy, const float *Iz, int w, int h,
+                           int s, float alpha, float *du1, float *dv1) {
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; ++j) {
+      const int pos = j + i * s;
+      int left, right, up, down;
+
+      // handle borders
+      if (j != 0)
+        left = pos - 1;
+      else
+        left = pos;
+
+      if (j != w - 1)
+        right = pos + 1;
+      else
+        right = pos;
+
+      if (i != 0)
+        down = pos - s;
+      else
+        down = pos;
+
+      if (i != h - 1)
+        up = pos + s;
+      else
+        up = pos;
+
+      float sumU = (du0[left] + du0[right] + du0[up] + du0[down]) * 0.25f;
+      float sumV = (dv0[left] + dv0[right] + dv0[up] + dv0[down]) * 0.25f;
+
+      float frac = (Ix[pos] * sumU + Iy[pos] * sumV + Iz[pos]) /
+                   (Ix[pos] * Ix[pos] + Iy[pos] * Iy[pos] + alpha);
+
+      du1[pos] = sumU - Ix[pos] * frac;
+      dv1[pos] = sumV - Iy[pos] * frac;
+    }
+  }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief method logic
+///
+/// handles memory allocation and control flow
+/// \param[in]  I0           source image
+/// \param[in]  I1           tracked image
+/// \param[in]  width        images width
+/// \param[in]  height       images height
+/// \param[in]  stride       images stride
+/// \param[in]  alpha        degree of displacement field smoothness
+/// \param[in]  nLevels      number of levels in a pyramid
+/// \param[in]  nWarpIters   number of warping iterations per pyramid level
+/// \param[in]  nSolverIters number of solver iterations (Jacobi iterations)
+/// \param[out] u            horizontal displacement
+/// \param[out] v            vertical displacement
+///////////////////////////////////////////////////////////////////////////////
+void ComputeFlowGold(const float *I0, const float *I1, int width, int height,
+                     int stride, float alpha, int nLevels, int nWarpIters,
+                     int nSolverIters, float *u, float *v) {
+  printf("Computing optical flow on CPU...\n");
+
+  float *u0 = u;
+  float *v0 = v;
+
+  const float **pI0 = new const float *[nLevels];
+  const float **pI1 = new const float *[nLevels];
+
+  int *pW = new int[nLevels];
+  int *pH = new int[nLevels];
+  int *pS = new int[nLevels];
+
+  const int pixelCountAligned = height * stride;
+
+  float *tmp = new float[pixelCountAligned];
+  float *du0 = new float[pixelCountAligned];
+  float *dv0 = new float[pixelCountAligned];
+  float *du1 = new float[pixelCountAligned];
+  float *dv1 = new float[pixelCountAligned];
+  float *Ix = new float[pixelCountAligned];
+  float *Iy = new float[pixelCountAligned];
+  float *Iz = new float[pixelCountAligned];
+  float *nu = new float[pixelCountAligned];
+  float *nv = new float[pixelCountAligned];
+
+  // prepare pyramid
+  int currentLevel = nLevels - 1;
+  pI0[currentLevel] = I0;
+  pI1[currentLevel] = I1;
+
+  pW[currentLevel] = width;
+  pH[currentLevel] = height;
+  pS[currentLevel] = stride;
+
+  for (; currentLevel > 0; --currentLevel) {
+    int nw = pW[currentLevel] / 2;
+    int nh = pH[currentLevel] / 2;
+    int ns = iAlignUp(nw);
+    pI0[currentLevel - 1] = new float[ns * nh];
+    pI1[currentLevel - 1] = new float[ns * nh];
+
+    Downscale(pI0[currentLevel], pW[currentLevel], pH[currentLevel],
+              pS[currentLevel], nw, nh, ns, (float *)pI0[currentLevel - 1]);
+
+    Downscale(pI1[currentLevel], pW[currentLevel], pH[currentLevel],
+              pS[currentLevel], nw, nh, ns, (float *)pI1[currentLevel - 1]);
+
+    pW[currentLevel - 1] = nw;
+    pH[currentLevel - 1] = nh;
+    pS[currentLevel - 1] = ns;
+  }
+
+  // initial approximation
+  memset(u, 0, stride * height * sizeof(float));
+  memset(v, 0, stride * height * sizeof(float));
+
+  // compute flow
+  for (; currentLevel < nLevels; ++currentLevel) {
+    for (int warpIter = 0; warpIter < nWarpIters; ++warpIter) {
+      memset(du0, 0, pixelCountAligned * sizeof(float));
+      memset(dv0, 0, pixelCountAligned * sizeof(float));
+
+      memset(du1, 0, pixelCountAligned * sizeof(float));
+      memset(dv1, 0, pixelCountAligned * sizeof(float));
+
+      WarpImage(pI1[currentLevel], pW[currentLevel], pH[currentLevel],
+                pS[currentLevel], u, v, tmp);
+
+      // on current level we compute optical flow
+      // between frame 0 and warped frame 1
+      ComputeDerivatives(pI0[currentLevel], tmp, pW[currentLevel],
+                         pH[currentLevel], pS[currentLevel], Ix, Iy, Iz);
+
+      for (int iter = 0; iter < nSolverIters; ++iter) {
+        SolveForUpdate(du0, dv0, Ix, Iy, Iz, pW[currentLevel], pH[currentLevel],
+                       pS[currentLevel], alpha, du1, dv1);
+        Swap(du0, du1);
+        Swap(dv0, dv1);
+      }
+
+      // update u, v
+      for (int i = 0; i < pH[currentLevel] * pS[currentLevel]; ++i) {
+        u[i] += du0[i];
+        v[i] += dv0[i];
+      }
+    }  // end for (int warpIter = 0; warpIter < nWarpIters; ++warpIter)
+
+    if (currentLevel != nLevels - 1) {
+      // prolongate solution
+      float scaleX = (float)pW[currentLevel + 1] / (float)pW[currentLevel];
+
+      Upscale(u, pW[currentLevel], pH[currentLevel], pS[currentLevel],
+              pW[currentLevel + 1], pH[currentLevel + 1], pS[currentLevel + 1],
+              scaleX, nu);
+
+      float scaleY = (float)pH[currentLevel + 1] / (float)pH[currentLevel];
+
+      Upscale(v, pW[currentLevel], pH[currentLevel], pS[currentLevel],
+              pW[currentLevel + 1], pH[currentLevel + 1], pS[currentLevel + 1],
+              scaleY, nv);
+
+      Swap(u, nu);
+      Swap(v, nv);
+    }
+  }  // end for (; currentLevel < nLevels; ++currentLevel)
+
+  if (u != u0) {
+    // solution is not in the specified array
+    // copy
+    memcpy(u0, u, pixelCountAligned * sizeof(float));
+    memcpy(v0, v, pixelCountAligned * sizeof(float));
+    Swap(u, nu);
+    Swap(v, nv);
+  }
+
+  // cleanup
+  // last level is not being freed here
+  // because it refers to input images
+  for (int i = 0; i < nLevels - 1; ++i) {
+    delete[] pI0[i];
+    delete[] pI1[i];
+  }
+
+  delete[] pI0;
+  delete[] pI1;
+  delete[] pW;
+  delete[] pH;
+  delete[] pS;
+  delete[] tmp;
+  delete[] du0;
+  delete[] dv0;
+  delete[] du1;
+  delete[] dv1;
+  delete[] Ix;
+  delete[] Iy;
+  delete[] Iz;
+  delete[] nu;
+  delete[] nv;
+}

Samples/5_Domain_Specific/HSOpticalFlow/flowGold.h

@@ -0,0 +1,44 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef FLOW_GOLD_H
+#define FLOW_GOLD_H
+
+void ComputeFlowGold(
+    const float *I0,  // source frame
+    const float *I1,  // tracked frame
+    int width,        // frame width
+    int height,       // frame height
+    int stride,       // row access stride
+    float alpha,      // smoothness coefficient
+    int nLevels,      // number of levels in pyramid
+    int nWarpIters,   // number of warping iterations per pyramid level
+    int nIters,       // number of solver iterations (for linear system)
+    float *u,         // output horizontal flow
+    float *v);        // output vertical flow
+
+#endif

Samples/5_Domain_Specific/HSOpticalFlow/main.cpp

@@ -0,0 +1,237 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+const static char *const sSDKsample = "HSOpticalFlow";
+
+// CPU-GPU discrepancy threshold for self-test
+const float THRESHOLD = 0.05f;
+
+#include <cuda_runtime.h>
+
+#include "common.h"
+#include "flowGold.h"
+#include "flowCUDA.h"
+
+#include <helper_functions.h>
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief save optical flow in format described on vision.middlebury.edu/flow
+/// \param[in] name output file name
+/// \param[in] w    optical flow field width
+/// \param[in] h    optical flow field height
+/// \param[in] s    optical flow field row stride
+/// \param[in] u    horizontal displacement
+/// \param[in] v    vertical displacement
+///////////////////////////////////////////////////////////////////////////////
+void WriteFloFile(const char *name, int w, int h, int s, const float *u,
+                  const float *v) {
+  FILE *stream;
+  stream = fopen(name, "wb");
+
+  if (stream == 0) {
+    printf("Could not save flow to \"%s\"\n", name);
+    return;
+  }
+
+  float data = 202021.25f;
+  fwrite(&data, sizeof(float), 1, stream);
+  fwrite(&w, sizeof(w), 1, stream);
+  fwrite(&h, sizeof(h), 1, stream);
+
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; ++j) {
+      const int pos = j + i * s;
+      fwrite(u + pos, sizeof(float), 1, stream);
+      fwrite(v + pos, sizeof(float), 1, stream);
+    }
+  }
+
+  fclose(stream);
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief
+/// load 4-channel unsigned byte image
+/// and convert it to single channel FP32 image
+/// \param[out] img_data pointer to raw image data
+/// \param[out] img_w    image width
+/// \param[out] img_h    image height
+/// \param[out] img_s    image row stride
+/// \param[in]  name     image file name
+/// \param[in]  exePath  executable file path
+/// \return true if image is successfully loaded or false otherwise
+///////////////////////////////////////////////////////////////////////////////
+bool LoadImageAsFP32(float *&img_data, int &img_w, int &img_h, int &img_s,
+                     const char *name, const char *exePath) {
+  printf("Loading \"%s\" ...\n", name);
+  char *name_ = sdkFindFilePath(name, exePath);
+
+  if (!name_) {
+    printf("File not found\n");
+    return false;
+  }
+
+  unsigned char *data = 0;
+  unsigned int w = 0, h = 0;
+  bool result = sdkLoadPPM4ub(name_, &data, &w, &h);
+
+  if (result == false) {
+    printf("Invalid file format\n");
+    return false;
+  }
+
+  img_w = w;
+  img_h = h;
+  img_s = iAlignUp(img_w);
+
+  img_data = new float[img_s * h];
+
+  // source is 4 channel image
+  const int widthStep = 4 * img_w;
+
+  for (int i = 0; i < img_h; ++i) {
+    for (int j = 0; j < img_w; ++j) {
+      img_data[j + i * img_s] = ((float)data[j * 4 + i * widthStep]) / 255.0f;
+    }
+  }
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief compare given flow field with gold (L1 norm)
+/// \param[in] width    optical flow field width
+/// \param[in] height   optical flow field height
+/// \param[in] stride   optical flow field row stride
+/// \param[in] h_uGold  horizontal displacement, gold
+/// \param[in] h_vGold  vertical displacement, gold
+/// \param[in] h_u      horizontal displacement
+/// \param[in] h_v      vertical displacement
+/// \return true if discrepancy is lower than a given threshold
+///////////////////////////////////////////////////////////////////////////////
+bool CompareWithGold(int width, int height, int stride, const float *h_uGold,
+                     const float *h_vGold, const float *h_u, const float *h_v) {
+  float error = 0.0f;
+
+  for (int i = 0; i < height; ++i) {
+    for (int j = 0; j < width; ++j) {
+      const int pos = j + i * stride;
+      error += fabsf(h_u[pos] - h_uGold[pos]) + fabsf(h_v[pos] - h_vGold[pos]);
+    }
+  }
+
+  error /= (float)(width * height);
+
+  printf("L1 error : %.6f\n", error);
+
+  return (error < THRESHOLD);
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// application entry point
+///////////////////////////////////////////////////////////////////////////////
+int main(int argc, char **argv) {
+  // welcome message
+  printf("%s Starting...\n\n", sSDKsample);
+
+  // pick GPU
+  findCudaDevice(argc, (const char **)argv);
+
+  // find images
+  const char *const sourceFrameName = "frame10.ppm";
+  const char *const targetFrameName = "frame11.ppm";
+
+  // image dimensions
+  int width;
+  int height;
+  // row access stride
+  int stride;
+
+  // flow is computed from source image to target image
+  float *h_source;  // source image, host memory
+  float *h_target;  // target image, host memory
+
+  // load image from file
+  if (!LoadImageAsFP32(h_source, width, height, stride, sourceFrameName,
+                       argv[0])) {
+    exit(EXIT_FAILURE);
+  }
+
+  if (!LoadImageAsFP32(h_target, width, height, stride, targetFrameName,
+                       argv[0])) {
+    exit(EXIT_FAILURE);
+  }
+
+  // allocate host memory for CPU results
+  float *h_uGold = new float[stride * height];
+  float *h_vGold = new float[stride * height];
+
+  // allocate host memory for GPU results
+  float *h_u = new float[stride * height];
+  float *h_v = new float[stride * height];
+
+  // smoothness
+  // if image brightness is not within [0,1]
+  // this paramter should be scaled appropriately
+  const float alpha = 0.2f;
+
+  // number of pyramid levels
+  const int nLevels = 5;
+
+  // number of solver iterations on each level
+  const int nSolverIters = 500;
+
+  // number of warping iterations
+  const int nWarpIters = 3;
+
+  ComputeFlowGold(h_source, h_target, width, height, stride, alpha, nLevels,
+                  nWarpIters, nSolverIters, h_uGold, h_vGold);
+
+  ComputeFlowCUDA(h_source, h_target, width, height, stride, alpha, nLevels,
+                  nWarpIters, nSolverIters, h_u, h_v);
+
+  // compare results (L1 norm)
+  bool status =
+      CompareWithGold(width, height, stride, h_uGold, h_vGold, h_u, h_v);
+
+  WriteFloFile("FlowGPU.flo", width, height, stride, h_u, h_v);
+
+  WriteFloFile("FlowCPU.flo", width, height, stride, h_uGold, h_vGold);
+
+  // free resources
+  delete[] h_uGold;
+  delete[] h_vGold;
+
+  delete[] h_u;
+  delete[] h_v;
+
+  delete[] h_source;
+  delete[] h_target;
+
+  // report self-test status
+  exit(status ? EXIT_SUCCESS : EXIT_FAILURE);
+}

Samples/5_Domain_Specific/HSOpticalFlow/solverKernel.cuh

@@ -0,0 +1,181 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "common.h"
+#include <cooperative_groups.h>
+
+namespace cg = cooperative_groups;
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief one iteration of classical Horn-Schunck method, CUDA kernel.
+///
+/// It is one iteration of Jacobi method for a corresponding linear system.
+/// Template parameters are describe CTA size
+/// \param[in]  du0     current horizontal displacement approximation
+/// \param[in]  dv0     current vertical displacement approximation
+/// \param[in]  Ix      image x derivative
+/// \param[in]  Iy      image y derivative
+/// \param[in]  Iz      temporal derivative
+/// \param[in]  w       width
+/// \param[in]  h       height
+/// \param[in]  s       stride
+/// \param[in]  alpha   degree of smoothness
+/// \param[out] du1     new horizontal displacement approximation
+/// \param[out] dv1     new vertical displacement approximation
+///////////////////////////////////////////////////////////////////////////////
+template <int bx, int by>
+__global__ void JacobiIteration(const float *du0, const float *dv0,
+                                const float *Ix, const float *Iy,
+                                const float *Iz, int w, int h, int s,
+                                float alpha, float *du1, float *dv1) {
+  // Handle to thread block group
+  cg::thread_block cta = cg::this_thread_block();
+
+  volatile __shared__ float du[(bx + 2) * (by + 2)];
+  volatile __shared__ float dv[(bx + 2) * (by + 2)];
+
+  const int ix = threadIdx.x + blockIdx.x * blockDim.x;
+  const int iy = threadIdx.y + blockIdx.y * blockDim.y;
+
+  // position within global memory array
+  const int pos = min(ix, w - 1) + min(iy, h - 1) * s;
+
+  // position within shared memory array
+  const int shMemPos = threadIdx.x + 1 + (threadIdx.y + 1) * (bx + 2);
+
+  // Load data to shared memory.
+  // load tile being processed
+  du[shMemPos] = du0[pos];
+  dv[shMemPos] = dv0[pos];
+
+  // load necessary neighbouring elements
+  // We clamp out-of-range coordinates.
+  // It is equivalent to mirroring
+  // because we access data only one step away from borders.
+  if (threadIdx.y == 0) {
+    // beginning of the tile
+    const int bsx = blockIdx.x * blockDim.x;
+    const int bsy = blockIdx.y * blockDim.y;
+    // element position within matrix
+    int x, y;
+    // element position within linear array
+    // gm - global memory
+    // sm - shared memory
+    int gmPos, smPos;
+
+    x = min(bsx + threadIdx.x, w - 1);
+    // row just below the tile
+    y = max(bsy - 1, 0);
+    gmPos = y * s + x;
+    smPos = threadIdx.x + 1;
+    du[smPos] = du0[gmPos];
+    dv[smPos] = dv0[gmPos];
+
+    // row above the tile
+    y = min(bsy + by, h - 1);
+    smPos += (by + 1) * (bx + 2);
+    gmPos = y * s + x;
+    du[smPos] = du0[gmPos];
+    dv[smPos] = dv0[gmPos];
+  } else if (threadIdx.y == 1) {
+    // beginning of the tile
+    const int bsx = blockIdx.x * blockDim.x;
+    const int bsy = blockIdx.y * blockDim.y;
+    // element position within matrix
+    int x, y;
+    // element position within linear array
+    // gm - global memory
+    // sm - shared memory
+    int gmPos, smPos;
+
+    y = min(bsy + threadIdx.x, h - 1);
+    // column to the left
+    x = max(bsx - 1, 0);
+    smPos = bx + 2 + threadIdx.x * (bx + 2);
+    gmPos = x + y * s;
+
+    // check if we are within tile
+    if (threadIdx.x < by) {
+      du[smPos] = du0[gmPos];
+      dv[smPos] = dv0[gmPos];
+      // column to the right
+      x = min(bsx + bx, w - 1);
+      gmPos = y * s + x;
+      smPos += bx + 1;
+      du[smPos] = du0[gmPos];
+      dv[smPos] = dv0[gmPos];
+    }
+  }
+
+  cg::sync(cta);
+
+  if (ix >= w || iy >= h) return;
+
+  // now all necessary data are loaded to shared memory
+  int left, right, up, down;
+  left = shMemPos - 1;
+  right = shMemPos + 1;
+  up = shMemPos + bx + 2;
+  down = shMemPos - bx - 2;
+
+  float sumU = (du[left] + du[right] + du[up] + du[down]) * 0.25f;
+  float sumV = (dv[left] + dv[right] + dv[up] + dv[down]) * 0.25f;
+
+  float frac = (Ix[pos] * sumU + Iy[pos] * sumV + Iz[pos]) /
+               (Ix[pos] * Ix[pos] + Iy[pos] * Iy[pos] + alpha);
+
+  du1[pos] = sumU - Ix[pos] * frac;
+  dv1[pos] = sumV - Iy[pos] * frac;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief one iteration of classical Horn-Schunck method, CUDA kernel wrapper.
+///
+/// It is one iteration of Jacobi method for a corresponding linear system.
+/// \param[in]  du0     current horizontal displacement approximation
+/// \param[in]  dv0     current vertical displacement approximation
+/// \param[in]  Ix      image x derivative
+/// \param[in]  Iy      image y derivative
+/// \param[in]  Iz      temporal derivative
+/// \param[in]  w       width
+/// \param[in]  h       height
+/// \param[in]  s       stride
+/// \param[in]  alpha   degree of smoothness
+/// \param[out] du1     new horizontal displacement approximation
+/// \param[out] dv1     new vertical displacement approximation
+///////////////////////////////////////////////////////////////////////////////
+static void SolveForUpdate(const float *du0, const float *dv0, const float *Ix,
+                           const float *Iy, const float *Iz, int w, int h,
+                           int s, float alpha, float *du1, float *dv1) {
+  // CTA size
+  dim3 threads(32, 6);
+  // grid size
+  dim3 blocks(iDivUp(w, threads.x), iDivUp(h, threads.y));
+
+  JacobiIteration<32, 6><<<blocks, threads>>>(du0, dv0, Ix, Iy, Iz, w, h, s,
+                                              alpha, du1, dv1);
+}

Samples/5_Domain_Specific/HSOpticalFlow/upscaleKernel.cuh

@@ -0,0 +1,96 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "common.h"
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief upscale one component of a displacement field, CUDA kernel
+/// \param[in]  width   field width
+/// \param[in]  height  field height
+/// \param[in]  stride  field stride
+/// \param[in]  scale   scale factor (multiplier)
+/// \param[out] out     result
+///////////////////////////////////////////////////////////////////////////////
+__global__ void UpscaleKernel(int width, int height, int stride, float scale,
+                              float *out, cudaTextureObject_t texCoarse) {
+  const int ix = threadIdx.x + blockIdx.x * blockDim.x;
+  const int iy = threadIdx.y + blockIdx.y * blockDim.y;
+
+  if (ix >= width || iy >= height) return;
+
+  float x = ((float)ix + 0.5f) / (float)width;
+  float y = ((float)iy + 0.5f) / (float)height;
+
+  // exploit hardware interpolation
+  // and scale interpolated vector to match next pyramid level resolution
+  out[ix + iy * stride] = tex2D<float>(texCoarse, x, y) * scale;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief upscale one component of a displacement field, kernel wrapper
+/// \param[in]  src         field component to upscale
+/// \param[in]  width       field current width
+/// \param[in]  height      field current height
+/// \param[in]  stride      field current stride
+/// \param[in]  newWidth    field new width
+/// \param[in]  newHeight   field new height
+/// \param[in]  newStride   field new stride
+/// \param[in]  scale       value scale factor (multiplier)
+/// \param[out] out         upscaled field component
+///////////////////////////////////////////////////////////////////////////////
+static void Upscale(const float *src, int width, int height, int stride,
+                    int newWidth, int newHeight, int newStride, float scale,
+                    float *out) {
+  dim3 threads(32, 8);
+  dim3 blocks(iDivUp(newWidth, threads.x), iDivUp(newHeight, threads.y));
+
+  cudaTextureObject_t texCoarse;
+  cudaResourceDesc texRes;
+  memset(&texRes, 0, sizeof(cudaResourceDesc));
+
+  texRes.resType = cudaResourceTypePitch2D;
+  texRes.res.pitch2D.devPtr = (void *)src;
+  texRes.res.pitch2D.desc = cudaCreateChannelDesc<float>();
+  texRes.res.pitch2D.width = width;
+  texRes.res.pitch2D.height = height;
+  texRes.res.pitch2D.pitchInBytes = stride * sizeof(float);
+
+  cudaTextureDesc texDescr;
+  memset(&texDescr, 0, sizeof(cudaTextureDesc));
+
+  texDescr.normalizedCoords = true;
+  texDescr.filterMode = cudaFilterModeLinear;
+  texDescr.addressMode[0] = cudaAddressModeMirror;
+  texDescr.addressMode[1] = cudaAddressModeMirror;
+  texDescr.readMode = cudaReadModeElementType;
+
+  checkCudaErrors(
+      cudaCreateTextureObject(&texCoarse, &texRes, &texDescr, NULL));
+
+  UpscaleKernel<<<blocks, threads>>>(newWidth, newHeight, newStride, scale, out,
+                                     texCoarse);
+}

Samples/5_Domain_Specific/HSOpticalFlow/warpingKernel.cuh

@@ -0,0 +1,100 @@
+/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *  * Neither the name of NVIDIA CORPORATION nor the names of its
+ *    contributors may be used to endorse or promote products derived
+ *    from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+ * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+ * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+ * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "common.h"
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief warp image with a given displacement field, CUDA kernel.
+/// \param[in]  width   image width
+/// \param[in]  height  image height
+/// \param[in]  stride  image stride
+/// \param[in]  u       horizontal displacement
+/// \param[in]  v       vertical displacement
+/// \param[out] out     result
+///////////////////////////////////////////////////////////////////////////////
+__global__ void WarpingKernel(int width, int height, int stride, const float *u,
+                              const float *v, float *out,
+                              cudaTextureObject_t texToWarp) {
+  const int ix = threadIdx.x + blockIdx.x * blockDim.x;
+  const int iy = threadIdx.y + blockIdx.y * blockDim.y;
+
+  const int pos = ix + iy * stride;
+
+  if (ix >= width || iy >= height) return;
+
+  float x = ((float)ix + u[pos] + 0.5f) / (float)width;
+  float y = ((float)iy + v[pos] + 0.5f) / (float)height;
+
+  out[pos] = tex2D<float>(texToWarp, x, y);
+}
+
+///////////////////////////////////////////////////////////////////////////////
+/// \brief warp image with provided vector field, CUDA kernel wrapper.
+///
+/// For each output pixel there is a vector which tells which pixel
+/// from a source image should be mapped to this particular output
+/// pixel.
+/// It is assumed that images and the vector field have the same stride and
+/// resolution.
+/// \param[in]  src source image
+/// \param[in]  w   width
+/// \param[in]  h   height
+/// \param[in]  s   stride
+/// \param[in]  u   horizontal displacement
+/// \param[in]  v   vertical displacement
+/// \param[out] out warped image
+///////////////////////////////////////////////////////////////////////////////
+static void WarpImage(const float *src, int w, int h, int s, const float *u,
+                      const float *v, float *out) {
+  dim3 threads(32, 6);
+  dim3 blocks(iDivUp(w, threads.x), iDivUp(h, threads.y));
+
+  cudaTextureObject_t texToWarp;
+  cudaResourceDesc texRes;
+  memset(&texRes, 0, sizeof(cudaResourceDesc));
+
+  texRes.resType = cudaResourceTypePitch2D;
+  texRes.res.pitch2D.devPtr = (void *)src;
+  texRes.res.pitch2D.desc = cudaCreateChannelDesc<float>();
+  texRes.res.pitch2D.width = w;
+  texRes.res.pitch2D.height = h;
+  texRes.res.pitch2D.pitchInBytes = s * sizeof(float);
+
+  cudaTextureDesc texDescr;
+  memset(&texDescr, 0, sizeof(cudaTextureDesc));
+
+  texDescr.normalizedCoords = true;
+  texDescr.filterMode = cudaFilterModeLinear;
+  texDescr.addressMode[0] = cudaAddressModeMirror;
+  texDescr.addressMode[1] = cudaAddressModeMirror;
+  texDescr.readMode = cudaReadModeElementType;
+
+  checkCudaErrors(
+      cudaCreateTextureObject(&texToWarp, &texRes, &texDescr, NULL));
+
+  WarpingKernel<<<blocks, threads>>>(w, h, s, u, v, out, texToWarp);
+}