Using CUDA

CUDA Compilation

Linux

To use CUDA in SOFA under Linux, you need to follow these steps:

Get the required files from NVIDIA by specifying your operating system, distribution, architecture, version and desired installer type
Set the environment variables CUDA_HOME to the location of cuda, for example: export CUDA_HOME="/usr/local/cuda"
Add cuda to your path: export PATH=${CUDA_HOME}/bin:${PATH}
In CMake, turn on the SOFA-PLUGIN_SOFACUDA option, and reconfigure
- New options relating to Cuda will appear, some with the prefix "SOFA-CUDA" and some with the prefix "CUDA". The SOFA-CUDA options include:
  - SOFA-CUDA_DOUBLE: Option to activate double-precision support in CUDA (requires GT200+ GPU and -arch sm_13 flag)
  - SOFA-CUDA_PRECISE: Option to use IEEE 754-compliant floating point operations
  - SOFA-CUDA_DOUBLE_PRECISE: Option to get double-precision for sqrt/div (requires compute capability >= 2 and CUDA_VERSION > 3.0) (with SOFA_GPU_CUDA_PRECISE and SOFA_GPU_CUDA_DOUBLE you get IEEE 754-compliant floating point operations for addition and multiplication only
  - SOFA-CUDA_CUBLAS: Option to activate cublas support in CUDA (requires SOFA_GPU_CUDA_DOUBLE)
  - SOFA-CUDA_CUDPP: Option to activate CUDPP (for RadixSort)
  - SOFA-CUDA_THRUST: Option to activate THRUST (for RadixSort) Note: THRUST is included in CUDA SDK 4.0+, it is recommended to use it if available
- If you want to use some of the more advanced features of Cuda, such as atomics for floats, you should can set the CUDA_NVCC_FLAGS option to --ptxas-options=-v -arch sm_12, changing sm_12 to whatever Compute Capabilities you graphics card supports. See (link) for a table of Compute Capabilities.
Configure, Generate, and compile

Windows

To use CUDA in SOFA under Windows, you need to follow these steps:

Get the required installer from NVIDIA by specifying your operating system, architecture, version and desired installer type.
Check that the path to Cuda has been added to your system path. (for example, C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.5\bin)
- Right click the My Computer icon on the desktop
- Select Properties
- Select Advanced System Settings from the left hand panel
- Select the Adcanced tab
- Press Environment Variables...
- Under System Variables, scroll until you find Path
- If the above path is not part of the Path variable, add a semicolon after the last path in the list, and add the above path.
In CMake, turn on the PLUGIN_SOFACUDA option, and reconfigure
- New options relating to Cuda will appear, some with the prefix "SOFA-CUDA" and some with the prefix "CUDA". The SOFA-CUDA options include:
  - SOFA-CUDA_DOUBLE: Option to activate double-precision support in CUDA (requires GT200+ GPU and -arch sm_13 flag)
  - The appropriate Cuda driver for your system
  - SOFA-CUDA_PRECISE: Option to use IEEE 754-compliant floating point operations
  - SOFA-CUDA_DOUBLE_PRECISE: Option to get double-precision for sqrt/div (requires compute capability >= 2 and CUDA_VERSION > 3.0) (with SOFA_GPU_CUDA_PRECISE and SOFA_GPU_CUDA_DOUBLE you get IEEE 754-compliant floating point operations for addition and multiplication only
  - SOFA-CUDA_CUBLAS: Option to activate cublas support in CUDA (requires SOFA_GPU_CUDA_DOUBLE)
  - SOFA-CUDA_CUDPP: Option to activate CUDPP (for RadixSort)
  - SOFA-CUDA_THRUST: Option to activate THRUST (for RadixSort) Note: THRUST is included in CUDA SDK 4.0+, it is recommended to use it if available
- If you want to use some of the more advanced features of Cuda, such as atomics for floats, you can set the CUDA_NVCC_FLAGS option to --ptxas-options=-v -arch sm_12, changing sm_12 to whatever Compute Capabilities you graphics card supports. See (link) for a table of Compute Capabilities.
Configure, Generate, and compile

SofaCUDA

SofaCUDA is a plugin that provides a number of Sofa components that have been implemented using CUDA, which allows the graphics card to be used for multithreaded programming. This can significantly improve the performance of your scene when working with large meshes. To include this plugin, first follow the compilation instructions define above.

CUDA scene file

To begin using the Cuda components, you need to change the template of the Mechanical Object in your scene. Consider the following scene:

<Node name="root" dt="0.04" showBehaviorModels="0" showCollisionModels="0" showMappings="0" showForceFields="1">
  <Node name="M1">
    <MeshVTKLoader name="volume" filename="mesh/raptorTetra_19409.vtu" onlyAttachedPoints="false" />
    <EulerImplicit rayleighMass="0.01" rayleighStiffness="0.01" />
    <CGLinearSolver iterations="25" tolerance="1e-6" threshold="1e-20"/>
    <MeshTopology src="@volume" />
    <MechanicalObject template="Vec3f" />
    <UniformMass vertexMass="0.01" />
    <BoxROI name="box0" box="-2.2 -1 -10 2.2  10  10" drawBoxes="1" />
    <BoxROI name="box1" box="-2.2 -1  -1 2.2 2.5 1.5" drawBoxes="1" />
    <IndexValueMapper name="ind_box0"                                      indices="@box0.tetrahedronIndices" value="100000" />
    <IndexValueMapper name="ind_box1" inputValues="@ind_box0.outputValues" indices="@box1.tetrahedronIndices" value="1000000" />
    <TetrahedronFEMForceField name="FEM" youngModulus="@ind_box1.outputValues" poissonRatio="0.4" listening="true" />
    <BoxROI name="box3" box="-2.2 -0.3 -9.2    2.2 0.110668 2.88584" drawBoxes="1" drawSize="2" />
    <FixedProjectiveConstraint indices="@box3.indices" />
    <BoxROI name="boxF" box="-2.2 -1 6.88 2.2  10  10" drawBoxes="true" />
    <ConstantForceField points="@boxF.indices" force="7.5 -6.63 -15" arrowSizeCoef="0.1" />
    <PlaneForceField normal="0 1 0" d="-0.2" stiffness="100"  draw="1" drawSize="20" />
  </Node>
</Node>

To use the Cuda components that are available for this scene, there are two steps: Add:

<RequiredPlugin name="CUDA computing" pluginName="SofaCUDA" />

near the top of the scene. This indicates that the SofaCUDA plugin is needed, and will automatically load the plugin if it is not already loaded in the Plugin Manager. Change the MechanicalObject template from Vec3f to CudaVec3f. This will cause all of the components which have been implemented in Cuda to use that implementation, rather than the standard cpu implementation. In this scene, this includes components such as BoxROI, TetrahedronFEMForceField and UniformMass. This leaves us with the scene:

<Node name="root" dt="0.04" showBehaviorModels="0" showCollisionModels="0" showMappings="0" showForceFields="1">
    <RequiredPlugin name="CUDA computing" pluginName="SofaCUDA" />
    <Node name="M1">
        <MeshVTKLoader name="volume" filename="mesh/raptorTetra_19409.vtu" onlyAttachedPoints="false" />
        <EulerImplicit rayleighMass="0.01" rayleighStiffness="0.01" />
        <CGLinearSolver iterations="25" tolerance="1e-6" threshold="1e-20"/>
        <MeshTopology src="@volume" />
        <MechanicalObject template="CudaVec3f" />
        <UniformMass vertexMass="0.01" />
        <BoxROI name="box0" box="-2.2 -1 -10 2.2  10  10" drawBoxes="1" />
        <BoxROI name="box1" box="-2.2 -1  -1 2.2 2.5 1.5" drawBoxes="1" />
        <IndexValueMapper name="ind_box0"                                      indices="@box0.tetrahedronIndices" value="100000" />
        <IndexValueMapper name="ind_box1" inputValues="@ind_box0.outputValues" indices="@box1.tetrahedronIndices" value="1000000" />
        <TetrahedronFEMForceField name="FEM" youngModulus="@ind_box1.outputValues" poissonRatio="0.4" listening="true" />
        <BoxROI name="box3" box="-2.2 -0.3 -9.2    2.2 0.110668 2.88584" drawBoxes="1" drawSize="2" />
        <FixedProjectiveConstraint indices="@box3.indices" />
        <BoxROI name="boxF" box="-2.2 -1 6.88 2.2  10  10" drawBoxes="true" />
        <ConstantForceField points="@boxF.indices" force="7.5 -6.63 -15" arrowSizeCoef="0.1" />
        <PlaneForceField normal="0 1 0" d="-0.2" stiffness="100"  draw="1" drawSize="20" />
    </Node>
</Node>

CUDA coding

How to access a data from the MechanicalState and use it in CUDA: n CUDA code, you cannot use the usual Accessor as for CPU coding. To access and use a vector from the MechanicalState inside your kernel, two steps are required: Reading access First, you need to access the vector:

const VecCoord& myPositions = this->getMState()->read(core::VecCoordId:: position())->getValue();

Then, to send the recovered vector to your kernel, you need to give as argument:

myPositions.deviceRead();

Writing-Reading access In a similar way, you need to write it as follows:

VecCoord& myPositions = *(this->getMState()->write(core::VecCoordId:: position()).beginEdit());
myPositions.deviceWrite();