Documentation

The following API documentation of OSPRay can also be found as a pdf document (3.6MB).

For a deeper explanation of the concepts, design, features and performance of OSPRay also have a look at the IEEE Vis 2016 paper “OSPRay – A CPU Ray Tracing Framework for Scientific Visualization” (49MB, or get the smaller version 1.8MB). Also available are the slides of the talk (5.2MB).

OSPRay API

To access the OSPRay API you first need to include the OSPRay header

#include "ospray/ospray.h"

where the API is compatible with C99 and C++.

Initialization

In order to use the API, OSPRay must be initialized with a “device”. A device is the object which implements the API. Creating and initializing a device can be done in either of two ways: command line arguments or manually instantiating a device.

Command Line Arguments

The first is to do so by giving OSPRay the command line from main() by calling

OSPError ospInit(int *argc, const char **argv);

OSPRay parses (and removes) its known command line parameters from your application’s main function. For an example see the tutorial. For possible error codes see section Error Handling and Status Messages. It is important to note that the arguments passed to ospInit() are processed in order they are listed. The following parameters (which are prefixed by convention with “--osp:”) are understood:

Command line parameters accepted by OSPRay’s ospInit.
Parameter Description
--osp:debug enables various extra checks and debug output, and disables multi-threading
--osp:numthreads <n> use n threads instead of per default using all detected hardware threads
--osp:loglevel <n> set logging level, default 0; increasing n means increasingly verbose log messages
--osp:verbose shortcut for --osp:loglevel 1
--osp:vv shortcut for --osp:loglevel 2
--osp:module:<name> load a module during initialization; equivalent to calling ospLoadModule(name)
--osp:mpi enables MPI mode for parallel rendering with the mpi_offload device, to be used in conjunction with mpirun; this will automatically load the “mpi” module if it is not yet loaded or linked
--osp:mpi-offload same as --osp:mpi
--osp:mpi-distributed same as --osp:mpi, but will create an mpi_distributed device instead; Note that this will likely require application changes to work properly
--osp:logoutput <dst> convenience for setting where status messages go; valid values for dst are cerr and cout
--osp:erroroutput <dst> convenience for setting where error messages go; valid values for dst are cerr and cout
--osp:device:<name> use name as the type of device for OSPRay to create; e.g. --osp:device:default gives you the default local device; Note if the device to be used is defined in a module, remember to pass --osp:module:<name> first
--osp:setaffinity <n> if 1, bind software threads to hardware threads; 0 disables binding; default is 1 on KNL and 0 otherwise

Manual Device Instantiation

The second method of initialization is to explicitly create the device yourself, and possibly set parameters. This method looks almost identical to how other objects are created and used by OSPRay (described in later sections). The first step is to create the device with

OSPDevice ospNewDevice(const char *type);

where the type string maps to a specific device implementation. OSPRay always provides the “default” device, which maps to a local CPU rendering device. If it is enabled in the build, you can also use “mpi” to access the MPI multi-node rendering device (see Parallel Rendering with MPI section for more information). Once a device is created, you can call

void ospDeviceSet1i(OSPDevice, const char *id, int val);

or

void ospDeviceSetString(OSPDevice, const char *id, const char *val);

to set parameters on the device. The following parameters can be set on all devices:

Parameters shared by all devices.
Type Name Description
int numThreads number of threads which OSPRay should use
int logLevel logging level
string logOutput convenience for setting where status messages go; valid values are cerr and cout
string errorOutput convenience for setting where error messages go; valid values are cerr and cout
int debug set debug mode; equivalent to logLevel=2 and numThreads=1
int setAffinity bind software threads to hardware threads if set to 1; 0 disables binding omitting the parameter will let OSPRay choose

Once parameters are set on the created device, the device must be committed with

void ospDeviceCommit(OSPDevice);

To use the newly committed device, you must call

void ospSetCurrentDevice(OSPDevice);

This then sets the given device as the object which will respond to all other OSPRay API calls.

Users can change parameters on the device after initialization (from either method above), by calling

OSPDevice ospGetCurrentDevice();

This function returns the handle to the device currently used to respond to OSPRay API calls, where users can set/change parameters and recommit the device. If changes are made to the device that is already set as the current device, it does not need to be set as current again.

Environment Variables

Finally, OSPRay’s generic device parameters can be overridden via environment variables for easy changes to OSPRay’s behavior without needing to change the application (variables are prefixed by convention with “OSPRAY_”):

Environment variables interpreted by OSPRay.
Variable Description
OSPRAY_THREADS equivalent to --osp:numthreads
OSPRAY_LOG_LEVEL equivalent to --osp:loglevel
OSPRAY_LOG_OUTPUT equivalent to --osp:logoutput
OSPRAY_ERROR_OUTPUT equivalent to --osp:erroroutput
OSPRAY_DEBUG equivalent to --osp:debug
OSPRAY_SET_AFFINITY equivalent to --osp:setaffinity

Error Handling and Status Messages

The following errors are currently used by OSPRay:

Possible error codes, i.e. valid named constants of type OSPError.
Name Description
OSP_NO_ERROR no error occurred
OSP_UNKNOWN_ERROR an unknown error occurred
OSP_INVALID_ARGUMENT an invalid argument was specified
OSP_INVALID_OPERATION the operation is not allowed for the specified object
OSP_OUT_OF_MEMORY there is not enough memory to execute the command
OSP_UNSUPPORTED_CPU the CPU is not supported (minimum ISA is SSE4.1)

These error codes are either directly return by some API functions, or are recorded to be later queried by the application via

OSPError ospDeviceGetLastErrorCode(OSPDevice);

A more descriptive error message can be queried by calling

const char* ospDeviceGetLastErrorMsg(OSPDevice);

Alternatively, the application can also register a callback function of type

typedef void (*OSPErrorFunc)(OSPError, const char* errorDetails);

via

void ospDeviceSetErrorFunc(OSPDevice, OSPErrorFunc);

to get notified when errors occur.

Applications may be interested in messages which OSPRay emits, whether for debugging or logging events. Applications can call

void ospDeviceSetStatusFunc(OSPDevice, OSPStatusFunc);

in order to register a callback function of type

typedef void (*OSPStatusFunc)(const char* messageText);

which OSPRay will use to emit status messages. By default, OSPRay uses a callback which does nothing, so any output desired by an application will require that a callback is provided. Note that callbacks for C++ std::cout and std::cerr can be alternatively set through ospInit() or the OSPRAY_LOG_OUTPUT environment variable.

Loading OSPRay Extensions at Runtime

OSPRay’s functionality can be extended via plugins, which are implemented in shared libraries. To load plugin name from libospray_module_<name>.so (on Linux and Mac OS X) or ospray_module_<name>.dll (on Windows) use

OSPError ospLoadModule(const char *name);

Modules are searched in OS-dependent paths. ospLoadModule returns OSP_NO_ERROR if the plugin could be successfully loaded.

Objects

All entities of OSPRay (the renderer, volumes, geometries, lights, cameras, …) are a specialization of OSPObject and share common mechanism to deal with parameters and lifetime.

An important aspect of object parameters is that parameters do not get passed to objects immediately. Instead, parameters are not visible at all to objects until they get explicitly committed to a given object via a call to

void ospCommit(OSPObject);

at which time all previously additions or changes to parameters are visible at the same time. If a user wants to change the state of an existing object (e.g., to change the origin of an already existing camera) it is perfectly valid to do so, as long as the changed parameters are recommitted.

The commit semantic allow for batching up multiple small changes, and specifies exactly when changes to objects will occur. This is important to ensure performance and consistency for devices crossing a PCI bus, or across a network. In our MPI implementation, for example, we can easily guarantee consistency among different nodes by MPI barrier’ing on every commit.

Note that OSPRay uses reference counting to manage the lifetime of all objects, so one cannot explicitly “delete” any object. Instead, to indicate that the application does not need and does not access the given object anymore, call

void ospRelease(OSPObject);

This decreases its reference count and if the count reaches 0 the object will automatically get deleted.

Parameters

Parameters allow to configure the behavior of and to pass data to objects. However, objects do not have an explicit interface for reasons of high flexibility and a more stable compile-time API. Instead, parameters are passed separately to objects in an arbitrary order, and unknown parameters will simply be ignored. The following functions allow adding various types of parameters with name id to a given object:

// add a C-string (zero-terminated char *) parameter
void ospSetString(OSPObject, const char *id, const char *s);

// add an object handle parameter to another object
void ospSetObject(OSPObject, const char *id, OSPObject object);

// add an untyped pointer -- this will *ONLY* work in local rendering!
void ospSetVoidPtr(OSPObject, const char *id, void *v);

// add scalar and vector integer and float parameters
void ospSetf  (OSPObject, const char *id, float x);
void ospSet1f (OSPObject, const char *id, float x);
void ospSet1i (OSPObject, const char *id, int32_t x);
void ospSet2f (OSPObject, const char *id, float x, float y);
void ospSet2fv(OSPObject, const char *id, const float *xy);
void ospSet2i (OSPObject, const char *id, int x, int y);
void ospSet2iv(OSPObject, const char *id, const int *xy);
void ospSet3f (OSPObject, const char *id, float x, float y, float z);
void ospSet3fv(OSPObject, const char *id, const float *xyz);
void ospSet3i (OSPObject, const char *id, int x, int y, int z);
void ospSet3iv(OSPObject, const char *id, const int *xyz);
void ospSet4f (OSPObject, const char *id, float x, float y, float z, float w);
void ospSet4fv(OSPObject, const char *id, const float *xyzw);

// additional functions to pass vector integer and float parameters in C++
void ospSetVec2f(OSPObject, const char *id, const vec2f &v);
void ospSetVec2i(OSPObject, const char *id, const vec2i &v);
void ospSetVec3f(OSPObject, const char *id, const vec3f &v);
void ospSetVec3i(OSPObject, const char *id, const vec3i &v);
void ospSetVec4f(OSPObject, const char *id, const vec4f &v);

Users can also remove parameters that have been explicitly set via an ospSet call. Any parameters which have been removed will go back to their default value during the next commit unless a new parameter was set after the parameter was removed. The following API function removes the named parameter from the given object:

void ospRemoveParam(OSPObject, const char *id);

Data

There is also the possibility to aggregate many values of the same type into an array, which then itself can be used as a parameter to objects. To create such a new data buffer, holding numItems elements of the given type, from the initialization data pointed to by source and optional creation flags, use

OSPData ospNewData(size_t numItems,
                   OSPDataType,
                   const void *source,
                   const uint32_t dataCreationFlags = 0);

The call returns an OSPData handle to the created array. The flag OSP_DATA_SHARED_BUFFER indicates that the buffer can be shared with the application. In this case the calling program guarantees that the source pointer will remain valid for the duration that this data array is being used. The enum type OSPDataType describes the different data types that can be represented in OSPRay; valid constants are listed in the table below.

Valid named constants for OSPDataType.
Type/Name Description
OSP_DEVICE API device object reference
OSP_VOID_PTR void pointer
OSP_DATA data reference
OSP_OBJECT generic object reference
OSP_CAMERA camera object reference
OSP_FRAMEBUFFER framebuffer object reference
OSP_LIGHT light object reference
OSP_MATERIAL material object reference
OSP_TEXTURE texture object reference
OSP_RENDERER renderer object reference
OSP_MODEL model object reference
OSP_GEOMETRY geometry object reference
OSP_VOLUME volume object reference
OSP_TRANSFER_FUNCTION transfer function object reference
OSP_PIXEL_OP pixel operation object reference
OSP_STRING C-style zero-terminated character string
OSP_CHAR 8 bit signed character scalar
OSP_UCHAR 8 bit unsigned character scalar
OSP_UCHAR[234] … and [234]-element vector
OSP_USHORT 16 bit unsigned integer scalar
OSP_INT 32 bit signed integer scalar
OSP_INT[234] … and [234]-element vector
OSP_UINT 32 bit unsigned integer scalar
OSP_UINT[234] … and [234]-element vector
OSP_LONG 64 bit signed integer scalar
OSP_LONG[234] … and [234]-element vector
OSP_ULONG 64 bit unsigned integer scalar
OSP_ULONG[234] … and [234]-element vector
OSP_FLOAT 32 bit single precision floating point scalar
OSP_FLOAT[234] … and [234]-element vector
OSP_FLOAT3A … and aligned 3-element vector
OSP_DOUBLE 64 bit double precision floating point scalar

To add a data array as parameter named id to another object call

void ospSetData(OSPObject, const char *id, OSPData);

Volumes

Volumes are volumetric datasets with discretely sampled values in 3D space, typically a 3D scalar field. To create a new volume object of given type type use

OSPVolume ospNewVolume(const char *type);

The call returns NULL if that type of volume is not known by OSPRay, or else an OSPVolume handle.

The common parameters understood by all volume variants are summarized in the table below.

Configuration parameters shared by all volume types.
Type Name Default Description
vec2f voxelRange minimum and maximum of the scalar values
bool gradientShadingEnabled false volume is rendered with surface shading wrt. to normalized gradient
bool preIntegration false use pre-integration for transfer function lookups
bool singleShade true shade only at the point of maximum intensity
bool adaptiveSampling true adapt ray step size based on opacity
float adaptiveScalar 15 modifier for adaptive step size
float adaptiveMaxSamplingRate 2 maximum sampling rate for adaptive sampling
float samplingRate 0.125 sampling rate of the volume (this is the minimum step size for adaptive sampling)
vec3f specular gray 0.3 specular color for shading
vec3f volumeClippingBoxLower disabled lower coordinate (in object-space) to clip the volume values
vec3f volumeClippingBoxUpper disabled upper coordinate (in object-space) to clip the volume values

Note that if voxelRange is not provided for a volume then OSPRay will compute it based on the voxel data, which may result in slower data updates.

Structured Volume

Structured volumes only need to store the values of the samples, because their addresses in memory can be easily computed from a 3D position. A common type of structured volumes are regular grids. OSPRay supports two variants that differ in how the volumetric data for the regular grids is specified.

The first variant shares the voxel data with the application. Such a volume type is created by passing the type string “shared_structured_volume” to ospNewVolume. The voxel data is laid out in memory in XYZ order and provided to the volume via a data buffer parameter named “voxelData”.

The second regular grid variant is optimized for rendering performance: data locality in memory is increased by arranging the voxel data in smaller blocks. This volume type is created by passing the type string “block_bricked_volume” to ospNewVolume. Because of this rearrangement of voxel data it cannot be shared the with the application anymore, but has to be transferred to OSPRay via

OSPError ospSetRegion(OSPVolume, void *source,
                      const vec3i &regionCoords,
                      const vec3i &regionSize);

The voxel data pointed to by source is copied into the given volume starting at position regionCoords, must be of size regionSize and be placed in memory in XYZ order. Note that OSPRay distinguishes between volume data and volume parameters. This function must be called only after all volume parameters (in particular dimensions and voxelType, see below) have been set and before ospCommit(volume) is called. If necessary then memory for the volume is allocated on the first call to this function.

The common parameters understood by both structured volume variants are summarized in the table below.

Additional configuration parameters for structured volumes.
Type Name Default Description
vec3i dimensions number of voxels in each dimension (x, y, z)
string voxelType data type of each voxel, currently supported are:
“uchar” (8 bit unsigned integer)
“short” (16 bit signed integer)
“ushort” (16 bit unsigned integer)
“float” (32 bit single precision floating point)
“double” (64 bit double precision floating point)
vec3f gridOrigin (0, 0, 0) origin of the grid in world-space
vec3f gridSpacing (1, 1, 1) size of the grid cells in world-space

Adaptive Mesh Refinement (AMR) Volume

AMR volumes are specified as a list of bricks, which are levels of refinement in potentially overlapping regions. There can be any number of refinement levels and any number of bricks at any level of refinement. An AMR volume type is created by passing the type string “amr_volume” to ospNewVolume.

Applications should first create an OSPData array which holds information about each brick. The following structure is used to populate this array (found in ospray.h):

struct amr_brick_info
{
  box3i bounds;
  int   refinementLevel;
  float cellWidth;
};

Then for each brick, the application should create an OSPData array of OSPData handles, where each handle is the data per-brick. Currently we only support float voxels.

Additional configuration parameters for AMR volumes.
Type Name Default Description
vec3f gridOrigin (0, 0, 0) origin of the grid in world-space
vec3f gridSpacing (1, 1, 1) size of the grid cells in world-space
string amrMethod current sampling method; valid values are “finest”, “current”, or “octant”
OSPData brickInfo array of info defining each brick
OSPData brickData array of handles to per-brick voxel data

Lastly, note that the gridOrigin and gridSpacing parameters act just like the structured volume equivalent, but they only modify the root (coarsest level) of refinement.

Unstructured Tetrahedral Volumes

Unstructured tetrahedral volumes are defined by three arrays: vertices, corresponding field values, and tetrahedra indices. A tetrahedral volume type is created by passing the type string “tetrahedral_volume” to ospNewVolume.

Similar to triangle mesh, each tetrahedron is formed by a group of indices into the vertices. For each vertex, the corresponding (by array index) data value will be used for sampling when rendering. Note that the index order for each tetrahedron does not matter, as OSPRay internally calculates vertex normals to ensure proper sampling and interpolation.

Additional configuration parameters for tetrahedral volumes.
Type Name Description
vec3f[] vertices data array of vertex positions
float[] field data array of vertex data values to be sampled
vec4i[] tetrahedra data array of tetrahedra indices (into vertices and field)

Transfer Function

Transfer functions map the scalar values of volumes to color and opacity and thus they can be used to visually emphasize certain features of the volume. To create a new transfer function of given type type use

OSPTransferFunction ospNewTransferFunction(const char *type);

The call returns NULL if that type of transfer functions is not known by OSPRay, or else an OSPTransferFunction handle to the created transfer function. That handle can be assigned to a volume as parameter “transferFunction” using ospSetObject.

One type of transfer function that is built-in in OSPRay is the linear transfer function, which interpolates between given equidistant colors and opacities. It is create by passing the string “piecewise_linear” to ospNewTransferFunction and it is controlled by these parameters:

Parameters accepted by the linear transfer function.
Type Name Description
vec3f[] colors data array of RGB colors
float[] opacities data array of opacities
vec2f valueRange domain (scalar range) this function maps from

Geometries

Geometries in OSPRay are objects that describe surfaces. To create a new geometry object of given type type use

OSPGeometry ospNewGeometry(const char *type);

The call returns NULL if that type of geometry is not known by OSPRay, or else an OSPGeometry handle.

Triangle Mesh

A traditional triangle mesh (indexed face set) geometry is created by calling ospNewGeometry with type string “triangles”. Once created, a triangle mesh recognizes the following parameters:

Parameters defining a triangle mesh geometry.
Type Name Description
vec3f(a)[] vertex data array of vertex positions
vec3f(a)[] vertex.normal data array of vertex normals
vec4f[] / vec3fa[] vertex.color data array of vertex colors (RGBA/RGB)
vec2f[] vertex.texcoord data array of vertex texture coordinates
vec3i(a)[] index data array of triangle indices (into vertex.*)

Spheres

A geometry consisting of individual spheres, each of which can have an own radius, is created by calling ospNewGeometry with type string “spheres”. The spheres will not be tessellated but rendered procedurally and are thus perfectly round. To allow a variety of sphere representations in the application this geometry allows a flexible way of specifying the data of center position and radius within a data array:

Parameters defining a spheres geometry.
Type Name Default Description
float radius 0.01 radius of all spheres (if offset_radius is not used)
OSPData spheres NULL memory holding the spatial data of all spheres
int bytes_per_sphere 16 size (in bytes) of each sphere within the spheres array
int offset_center 0 offset (in bytes) of each sphere’s “vec3f center” position (in object-space) within the spheres array
int offset_radius -1 offset (in bytes) of each sphere’s “float radius” within the spheres array (-1 means disabled and use radius)
vec4f[] / vec3f(a)[] color NULL data array of colors (RGBA/RGB), color is constant for each sphere
vec2f[] texcoord NULL data array of texture coordinates, coordinate is constant for each sphere

Cylinders

A geometry consisting of individual cylinders, each of which can have an own radius, is created by calling ospNewGeometry with type string “cylinders”. The cylinders will not be tessellated but rendered procedurally and are thus perfectly round. To allow a variety of cylinder representations in the application this geometry allows a flexible way of specifying the data of offsets for start position, end position and radius within a data array. All parameters are listed in the table below.

Parameters defining a cylinders geometry.
Type Name Default Description
float radius 0.01 radius of all cylinders (if offset_radius is not used)
OSPData cylinders NULL memory holding the spatial data of all cylinders
int bytes_per_cylinder 24 size (in bytes) of each cylinder within the cylinders array
int offset_v0 0 offset (in bytes) of each cylinder’s “vec3f v0” position (the start vertex, in object-space) within the cylinders array
int offset_v1 12 offset (in bytes) of each cylinder’s “vec3f v1” position (the end vertex, in object-space) within the cylinders array
int offset_radius -1 offset (in bytes) of each cylinder’s “float radius” within the cylinders array (-1 means disabled and use radius instead)
vec4f[] / vec3f(a)[] color NULL data array of colors (RGBA/RGB), color is constant for each cylinder
OSPData texcoord NULL data array of texture coordinates, in pairs (each a vec2f at vertex v0 and v1)

For texturing each cylinder is seen as a 1D primitive, i.e. a line segment: the 2D texture coordinates at its vertices v0 and v1 are linearly interpolated.

Streamlines

A geometry consisting of multiple streamlines of constant radius is created by calling ospNewGeometry with type string “streamlines”. The streamlines are internally assembled from connected (and rounded) cylinder segments and are thus perfectly round. The parameters defining this geometry are listed in the table below.

Parameters defining a streamlines geometry.
Type Name Description
float radius radius of all streamlines, default 0.01
vec3fa[] vertex data array of all vertices for all streamlines
vec4f[] vertex.color data array of corresponding vertex colors (RGBA)
int32[] index data array of indices to the first vertex of a link

Each streamline is specified by a set of (aligned) vec3fa control points in vertex; all vertices belonging to to the same logical streamline are connected via cylinders of a fixed radius radius, with additional spheres at each vertex to make for a smooth transition between the cylinders.

A streamlines geometry can contain multiple disjoint streamlines, each streamline is specified as a list of linear segments (or links) referenced via index: each entry e of the index array points the first vertex of a link (vertex[index[e]]) and the second vertex of the link is implicitly the directly following one (vertex[index[e]+1]). For example, two streamlines of vertices (A-B-C-D) and (E-F-G), respectively, would internally correspond to five links (A-B, B-C, C-D, E-F, and F-G), and would be specified via an array of vertices [A,B,C,D,E,F,G], plus an array of link indices [0,1,2,4,5].

Isosurfaces

OSPRay can directly render multiple isosurfaces of a volume without first tessellating them. To do so create an isosurfaces geometry by calling ospNewGeometry with type string “isosurfaces”. Each isosurface will be colored according to the provided volume’s transfer function.

Parameters defining an isosurfaces geometry.
Type Name Description
float[] isovalues data array of isovalues
OSPVolume volume handle of the volume to be isosurfaced

Slices

One tool to highlight interesting features of volumetric data is to visualize 2D cuts (or slices) by placing planes into the volume. Such a slices geometry is created by calling ospNewGeometry with type string “slices”. The planes are defined by the coefficients (a, b, c, d) of the plane equation ax + by + cz + d = 0. Each slice is colored according to the provided volume’s transfer function.

Parameters defining a slices geometry.
Type Name Description
vec4f[] planes data array with plane coefficients for all slices
OSPVolume volume handle of the volume that will be sliced

Instances

OSPRay supports instancing via a special type of geometry. Instances are created by transforming another given model modelToInstantiate with the given affine transformation transform by calling

OSPGeometry ospNewInstance(OSPModel modelToInstantiate, const affine3f &transform);

Renderer

A renderer is the central object for rendering in OSPRay. Different renderers implement different features and support different materials. To create a new renderer of given type type use

OSPRenderer ospNewRenderer(const char *type);

The call returns NULL if that type of renderer is not known, or else an OSPRenderer handle to the created renderer. General parameters of all renderers are

Parameters understood by all renderers.
Type Name Default Description
OSPModel model the model to render
OSPCamera camera the camera to be used for rendering
OSPLight[] lights data array with handles of the lights
float epsilon 10-6 ray epsilon to avoid self-intersections, relative to scene diameter
int spp 1 samples per pixel
int maxDepth 20 maximum ray recursion depth
float minContribution 0.001 sample contributions below this value will be neglected to speed-up rendering
float varianceThreshold 0 threshold for adaptive accumulation

OSPRay’s renderers support a feature called adaptive accumulation, which accelerates progressive rendering by stopping the rendering and refinement of image regions that have an estimated variance below the varianceThreshold. This feature requires a framebuffer with an OSP_FB_VARIANCE channel.

SciVis Renderer

The SciVis renderer is a fast ray tracer for scientific visualization which supports volume rendering and ambient occlusion (AO). It is created by passing the type string “scivis” or “raytracer” to ospNewRenderer. In addition to the general parameters understood by all renderers the SciVis renderer supports the following special parameters:

Special parameters understood by the SciVis renderer.
Type Name Default Description
bool shadowsEnabled false whether to compute (hard) shadows
int aoSamples 0 number of rays per sample to compute ambient occlusion
float aoDistance 1020 maximum distance to consider for ambient occlusion
bool aoTransparencyEnabled false whether object transparency is respected when computing ambient occlusion (slower)
bool oneSidedLighting true if true back-facing surfaces (wrt. light source) receive no illumination
float / vec3f / vec4f bgColor black, transparent background color and alpha (RGBA)
OSPTexture2D maxDepthTexture NULL screen-sized float texture with maximum far distance per pixel

Note that the intensity (and color) of AO is controlled via an ambient light. If aoSamples is zero (the default) then ambient lights cause ambient illumination (without occlusion).

Per default the background of the rendered image will be transparent black, i.e. the alpha channel holds the opacity of the rendered objects. This facilitates transparency-aware blending of the image with an arbitrary background image by the application. The parameter bgColor can be used to already blend with a constant background color (and alpha) during rendering.

The SciVis renderer supports depth composition with images of other renderers, for example to incorporate help geometries of a 3D UI that were rendered with OpenGL. The screen-sized texture maxDepthTexture must have format OSP_TEXTURE_R32F and flag OSP_TEXTURE_FILTER_NEAREST. The fetched values are used to limit the distance of primary rays, thus objects of other renderers can hide objects rendered by OSPRay.

Path Tracer

The path tracer supports soft shadows, indirect illumination and realistic materials. This renderer is created by passing the type string “pathtracer” to ospNewRenderer. In addition to the general parameters understood by all renderers the path tracer supports the following special parameters:

Special parameters understood by the path tracer.
Type Name Default Description
int rouletteDepth 5 ray recursion depth at which to start Russian roulette termination
float maxContribution samples are clamped to this value before they are accumulated into the framebuffer
OSPTexture2D backplate NULL texture image used as background, replacing visible lights in infinity (e.g. the HDRI light)

The path tracer requires that materials are assigned to geometries, otherwise surfaces are treated as completely black.

Model

Models are a container of scene data. They can hold the different geometries and volumes as well as references to (and instances of) other models. A model is associated with a single logical acceleration structure. To create an (empty) model call

OSPModel ospNewModel();

The call returns an OSPModel handle to the created model. To add an already created geometry or volume to a model use

void ospAddGeometry(OSPModel, OSPGeometry);
void ospAddVolume(OSPModel, OSPVolume);

An existing geometry or volume can be removed from a model with

void ospRemoveGeometry(OSPModel, OSPGeometry);
void ospRemoveVolume(OSPModel, OSPVolume);

Lights

To let the given renderer create a new light source of given type type use

OSPLight ospNewLight(OSPRenderer renderer, const char *type);

The call returns NULL if that type of camera is not known by the renderer, or else an OSPLight handle to the created light source. All light sources1 accept the following parameters:

Parameters accepted by the all lights.
Type Name Default Description
vec3f(a) color white color of the light
float intensity 1 intensity of the light (a factor)
bool isVisible true whether the light can be directly seen

The following light types are supported by most OSPRay renderers.

Directional Light / Distant Light

The distant light (or traditionally the directional light) is thought to be very far away (outside of the scene), thus its light arrives (almost) as parallel rays. It is created by passing the type string “distant” to ospNewLight. In addition to the general parameters understood by all lights the distant light supports the following special parameters:

Special parameters accepted by the distant light.
Type Name Description
vec3f(a) direction main emission direction of the distant light
float angularDiameter apparent size (angle in degree) of the light

Setting the angular diameter to a value greater than zero will result in soft shadows when the renderer uses stochastic sampling (like the path tracer). For instance, the apparent size of the sun is about 0.53°.

Point Light / Sphere Light

The sphere light (or the special case point light) is a light emitting uniformly in all directions. It is created by passing the type string “sphere” to ospNewLight. In addition to the general parameters understood by all lights the sphere light supports the following special parameters:

Special parameters accepted by the sphere light.
Type Name Description
vec3f(a) position the center of the sphere light, in world-space
float radius the size of the sphere light

Setting the radius to a value greater than zero will result in soft shadows when the renderer uses stochastic sampling (like the path tracer).

Spot Light

The spot light is a light emitting into a cone of directions. It is created by passing the type string “spot” to ospNewLight. In addition to the general parameters understood by all lights the spot light supports the special parameters listed in the table.

Special parameters accepted by the spot light.
Type Name Description
vec3f(a) position the center of the spot light, in world-space
vec3f(a) direction main emission direction of the spot
float openingAngle full opening angle (in degree) of the spot; outside of this cone is no illumination
float penumbraAngle size (angle in degree) of the “penumbra”, the region between the rim (of the illumination cone) and full intensity of the spot; should be smaller than half of openingAngle
float radius the size of the spot light, the radius of a disk with normal direction
Angles used by SpotLight.

Angles used by SpotLight.

Setting the radius to a value greater than zero will result in soft shadows when the renderer uses stochastic sampling (like the path tracer).

Quad Light

The quad2 light is a planar, procedural area light source emitting uniformly on one side into the half space. It is created by passing the type string “quad” to ospNewLight. In addition to the general parameters understood by all lights the spot light supports the following special parameters:

Special parameters accepted by the quad light.
Type Name Description
vec3f(a) position world-space position of one vertex of the quad light
vec3f(a) edge1 vector to one adjacent vertex
vec3f(a) edge2 vector to the other adjacent vertex
Defining a Quad Light.

Defining a Quad Light.

The emission side is determined by the cross product of edge1×edge2. Note that only renderers that use stochastic sampling (like the path tracer) will compute soft shadows from the quad light. Other renderers will just sample the center of the quad light, which results in hard shadows.

HDRI Light

The HDRI light is a textured light source surrounding the scene and illuminating it from infinity. It is created by passing the type string “hdri” to ospNewLight. In addition to the parameter intensity the HDRI light supports the following special parameters:

Special parameters accepted by the HDRI light.
Type Name Description
vec3f(a) up up direction of the light in world-space
vec3f(a) dir direction to which the center of the texture will be mapped to (analog to panoramic camera)
OSPTexture2D map environment map in latitude / longitude format
Orientation and Mapping of an HDRI Light.

Orientation and Mapping of an HDRI Light.

Note that the currently only the path tracer supports the HDRI light.

Ambient Light

The ambient light surrounds the scene and illuminates it from infinity with constant radiance (determined by combining the parameters color and intensity). It is created by passing the type string “ambient” to ospNewLight.

Note that the SciVis renderer uses ambient lights to control the color and intensity of the computed ambient occlusion (AO).

Emissive Objects

The path tracer will consider illumination by geometries which have a light emitting material assigned (for example the Luminous material).

Materials

Materials describe how light interacts with surfaces, they give objects their distinctive look. To let the given renderer create a new material of given type type call

OSPMaterial ospNewMaterial(OSPRenderer, const char *type);

The call returns NULL if the material type is not known by the renderer, or else an OSPMaterial handle to the created material. The handle can then be used to assign the material to a given geometry with

void ospSetMaterial(OSPGeometry, OSPMaterial);

OBJ Material

The OBJ material is the workhorse material supported by both the SciVis renderer and the path tracer. It offers widely used common properties like diffuse and specular reflection and is based on the MTL material format of Lightwave’s OBJ scene files. To create an OBJ material pass the type string “OBJMaterial” to ospNewMaterial. Its main parameters are

Main parameters of the OBJ material.
Type Name Default Description
vec3f Kd white 0.8 diffuse color
vec3f Ks black specular color
float Ns 10 shininess (Phong exponent), usually in [2–104]
float d opaque opacity
vec3f Tf black transparency filter color
OSPTexture2D map_Bump NULL normal map

In particular when using the path tracer it is important to adhere to the principle of energy conservation, i.e. that the amount of light reflected by a surface is not larger than the light arriving. Therefore the path tracer issues a warning and renormalizes the color parameters if the sum of Kd, Ks, and Tf is larger than one in any color channel. Similarly important to mention is that almost all materials of the real world reflect at most only about 80% of the incoming light. So even for a white sheet of paper or white wall paint do better not set Kd larger than 0.8; otherwise rendering times are unnecessary long and the contrast in the final images is low (for example, the corners of a white room would hardly be discernible).

Note that currently only the path tracer implements colored transparency with Tf.

Normal mapping can simulate small geometric features via the texture map_Bump. The normals n in the normal map are wrt. the local tangential shading coordinate system and are encoded as ½(n + 1), thus a texel (0.5, 0.5, 1)3 represents the unperturbed shading normal (0, 0, 1). Because of this encoding an sRGB gamma texture format is ignored and normals are always fetched as linear from a normal map. Note that the orientation of normal maps is important for a visually consistent look: by convention OSPRay uses a coordinate system with the origin in the lower left corner; thus a convexity will look green towards the top of the texture image (see also the example image of a normal map). If this is not the case flip the normal map vertically or invert its green channel.

Normal map representing an exalted square pyramidal frustum.

Normal map representing an exalted square pyramidal frustum.

All parameters (except Tf) can be textured by passing a texture handle, prefixed with “map_”. The fetched texels are multiplied by the respective parameter value. Texturing requires geometries with texture coordinates, e.g. a triangle mesh with vertex.texcoord provided. The color textures map_Kd and map_Ks are typically in one of the sRGB gamma encoded formats, whereas textures map_Ns and map_d are usually in a linear format (and only the first component is used). Additionally, all textures support texture transformations.

Rendering of a OBJ material with wood textures.

Rendering of a OBJ material with wood textures.

Metal

The path tracer offers a physical metal, supporting changing roughness and realistic color shifts at edges. To create a Metal material pass the type string “Metal” to ospNewMaterial. Its parameters are

Parameters of the Metal material.
Type Name Default Description
vec3f[] ior Aluminium data array of spectral samples of complex refractive index, each entry in the form (wavelength, eta, k), ordered by wavelength (which is in nm)
vec3f eta RGB complex refractive index, real part
vec3f k RGB complex refractive index, imaginary part
float roughness 0.1 roughness in [0–1], 0 is perfect mirror

The main appearance (mostly the color) of the Metal material is controlled by the physical parameters eta and k, the wavelength-dependent, complex index of refraction. These coefficients are quite counterintuitive but can be found in published measurements. For accuracy the index of refraction can be given as an array of spectral samples in ior, each sample a triplet of wavelength (in nm), eta, and k, ordered monotonically increasing by wavelength; OSPRay will then calculate the Fresnel in the spectral domain. Alternatively, eta and k can also be specified as approximated RGB coefficients; some examples are given in below table.

Index of refraction of selected metals as approximated RGB coefficients, based on data from https://refractiveindex.info/.
Metal eta k
Ag, Silver (0.051, 0.043, 0.041) (5.3, 3.6, 2.3)
Al, Aluminium (1.5, 0.98, 0.6) (7.6, 6.6, 5.4)
Au, Gold (0.07, 0.37, 1.5) (3.7, 2.3, 1.7)
Cr, Chromium (3.2, 3.1, 2.3) (3.3, 3.3, 3.1)
Cu, Copper (0.1, 0.8, 1.1) (3.5, 2.5, 2.4)

The roughness parameter controls the variation of microfacets and thus how polished the metal will look. The roughness can be modified by a texture map_roughness (texture transformations are supported as well) to create interesting edging effects.

Rendering of golden Metal material with textured roughness.

Rendering of golden Metal material with textured roughness.

Alloy

The path tracer offers an alloy material, which behaves similar to Metal, but allows for more intuitive and flexible control of the color. To create an Alloy material pass the type string “Alloy” to ospNewMaterial. Its parameters are

Parameters of the Alloy material.
Type Name Default Description
vec3f color white 0.9 reflectivity at normal incidence (0 degree)
vec3f edgeColor white reflectivity at grazing angle (90 degree)
float roughness 0.1 roughness in [0–1], 0 is perfect mirror

The main appearance of the Alloy material is controlled by the parameter color, while edgeColor influences the tint of reflections when seen at grazing angles (for real metals this is always 100% white). As in Metal the roughness parameter controls the variation of microfacets and thus how polished the alloy will look. All parameters can be textured by passing a texture handle, prefixed with “map_”; texture transformations are supported as well.

Rendering of a fictional Alloy material with textured color.

Rendering of a fictional Alloy material with textured color.

Glass

The path tracer offers a realistic a glass material, supporting refraction and volumetric attenuation (i.e. the transparency color varies with the geometric thickness). To create a Glass material pass the type string “Glass” to ospNewMaterial. Its parameters are

Parameters of the Glass material.
Type Name Default Description
float eta 1.5 index of refraction
vec3f attenuationColor white resulting color due to attenuation
float attenuationDistance 1 distance affecting attenuation

For convenience, the rather counterintuitive physical attenuation coefficients will be calculated from the user inputs in such a way, that the attenuationColor will be the result when white light traveled trough a glass of thickness attenuationDistance.

Rendering of a Glass material with orange attenuation.

Rendering of a Glass material with orange attenuation.

ThinGlass

The path tracer offers a thin glass material useful for objects with just a single surface, most prominently windows. It models a very thin, transparent slab, i.e. it behaves as if a second, virtual surface is parallel to the real geometric surface. The implementation accounts for multiple internal reflections between the interfaces (including attenuation), but neglects parallax effects due to its (virtual) thickness. To create a such a thin glass material pass the type string “ThinGlass” to ospNewMaterial. Its parameters are

Parameters of the ThinGlass material.
Type Name Default Description
float eta 1.5 index of refraction
vec3f attenuationColor white resulting color due to attenuation
float attenuationDistance 1 distance affecting attenuation
float thickness 1 virtual thickness

For convenience the attenuation is controlled the same way as with the Glass material. Additionally, the color due to attenuation can be modulated with a texture map_attenuationColor (texture transformations are supported as well). The thickness parameter sets the (virtual) thickness and allows for easy exchange of parameters with the (real) Glass material; internally just the ratio between attenuationDistance and thickness is used to calculate the resulting attenuation and thus the material appearance.

Rendering of a ThinGlass material with red attenuation.

Rendering of a ThinGlass material with red attenuation.

Example image of a colored window made with textured attenuation of the ThinGlass material.

Example image of a colored window made with textured attenuation of the ThinGlass material.

Luminous

The path tracer supports the Luminous material which emits light uniformly in all directions and which can thus be used to turn any geometric object into a light source. It is created by passing the type string “Luminous” to ospNewMaterial. The amount of constant radiance that is emitted is determined by combining the general parameters of lights: color and intensity.

Rendering of a yellow Luminous material.

Rendering of a yellow Luminous material.

Texture

To create a new 2D texture of size size (in pixels) and with the given format and flags use

OSPTexture2D ospNewTexture2D(const vec2i &size,
                             const OSPTextureFormat,
                             void *source = NULL,
                             const uint32_t textureCreationFlags = 0);

The call returns NULL if the texture could not be created with the given parameters, or else an OSPTexture2D handle to the created texture. The supported texture formats are:

Supported texture formats by ospNewTexture2D, i.e. valid constants of type OSPTextureFormat.
Name Description
OSP_TEXTURE_RGBA8 8 bit [0–255] linear components red, green, blue, alpha
OSP_TEXTURE_SRGBA 8 bit sRGB gamma encoded color components, and linear alpha
OSP_TEXTURE_RGBA32F 32 bit float components red, green, blue, alpha
OSP_TEXTURE_RGB8 8 bit [0–255] linear components red, green, blue
OSP_TEXTURE_SRGB 8 bit sRGB gamma encoded components red, green, blue
OSP_TEXTURE_RGB32F 32 bit float components red, green, blue
OSP_TEXTURE_R8 8 bit [0–255] linear single component
OSP_TEXTURE_R32F 32 bit float single component

The texel data addressed by source starts with the texels in the lower left corner of the texture image, like in OpenGL. Similar to data buffers the texel data can be shared by the application by specifying the OSP_TEXTURE_SHARED_BUFFER flag. Per default a texture fetch is filtered by performing bi-linear interpolation of the nearest 2×2 texels; if instead fetching only the nearest texel is desired (i.e. no filtering) then pass the OSP_TEXTURE_FILTER_NEAREST flag. Both texture creating flags can be combined with a bitwise OR.

Texture Transformations

All materials with textures also offer to manipulate the placement of these textures with the help of texture transformations. If so, this convention shall be used. The following parameters (prefixed with “texture_name.”) are combined into one transformation matrix:

Parameters to define texture coordinate transformations.
Type Name Description
vec4f transform interpreted as 2×2 matrix (linear part), column-major
float rotation angle in degree, counterclock-wise, around center
vec2f scale enlarge texture, relative to center (0.5, 0.5)
vec2f translation move texture in positive direction (right/up)

The transformations are applied in the given order. Rotation, scale and translation are interpreted “texture centric”, i.e. their effect seen by an user are relative to the texture (although the transformations are applied to the texture coordinates).

Cameras

To create a new camera of given type type use

OSPCamera ospNewCamera(const char *type);

The call returns NULL if that type of camera is not known, or else an OSPCamera handle to the created camera. All cameras accept these parameters:

Parameters accepted by all cameras.
Type Name Description
vec3f(a) pos position of the camera in world-space
vec3f(a) dir main viewing direction of the camera
vec3f(a) up up direction of the camera
float nearClip near clipping distance
vec2f imageStart start of image region (lower left corner)
vec2f imageEnd end of image region (upper right corner)

The camera is placed and oriented in the world with pos, dir and up. OSPRay uses a right-handed coordinate system. The region of the camera sensor that is rendered to the image can be specified in normalized screen-space coordinates with imageStart (lower left corner) and imageEnd (upper right corner). This can be used, for example, to crop the image, to achieve asymmetrical view frusta, or to horizontally flip the image to view scenes which are specified in a left-handed coordinate system. Note that values outside the default range of [0–1] are valid, which is useful to easily realize overscan or film gate, or to emulate a shifted sensor.

Perspective Camera

The perspective camera implements a simple thinlens camera for perspective rendering, supporting optionally depth of field and stereo rendering, but no motion blur. It is created by passing the type string “perspective” to ospNewCamera. In addition to the general parameters understood by all cameras the perspective camera supports the special parameters listed in the table below.

Parameters accepted by the perspective camera.
Type Name Description
float fovy the field of view (angle in degree) of the frame’s height
float aspect ratio of width by height of the frame
float apertureRadius size of the aperture, controls the depth of field
float focusDistance distance at where the image is sharpest when depth of field is enabled
bool architectural vertical edges are projected to be parallel
int stereoMode 0: no stereo (default), 1: left eye, 2: right eye, 3: side-by-side
float interpupillaryDistance distance between left and right eye when stereo is enabled

Note that when setting the aspect ratio a non-default image region (using imageStart & imageEnd) needs to be regarded.

In architectural photography it is often desired for aesthetic reasons to display the vertical edges of buildings or walls vertically in the image as well, regardless of how the camera is tilted. Enabling the architectural mode achieves this by internally leveling the camera parallel to the ground (based on the up direction) and then shifting the lens such that the objects in direction dir are centered in the image. If finer control of the lens shift is needed use imageStart & imageEnd. Because the camera is now effectively leveled its image plane and thus the plane of focus is oriented parallel to the front of buildings, the whole façade appears sharp, as can be seen in the example images below.

Example image created with the perspective camera, featuring depth of field.

Example image created with the perspective camera, featuring depth of field.

Enabling the architectural flag corrects the perspective projection distortion, resulting in parallel vertical edges.

Enabling the architectural flag corrects the perspective projection distortion, resulting in parallel vertical edges.

Example 3D stereo image using stereoMode side-by-side.

Example 3D stereo image using stereoMode side-by-side.

Orthographic Camera

The orthographic camera implements a simple camera with orthographic projection, without support for depth of field or motion blur. It is created by passing the type string “orthographic” to ospNewCamera. In addition to the general parameters understood by all cameras the orthographic camera supports the following special parameters:

Parameters accepted by the orthographic camera.
Type Name Description
float height size of the camera’s image plane in y, in world coordinates
float aspect ratio of width by height of the frame

For convenience the size of the camera sensor, and thus the extent of the scene that is captured in the image, can be controlled with the height parameter. The same effect can be achieved with imageStart and imageEnd, and both methods can be combined. In any case, the aspect ratio needs to be set accordingly to get an undistorted image.

Example image created with the orthographic camera.

Example image created with the orthographic camera.

Panoramic Camera

The panoramic camera implements a simple camera without support for motion blur. It captures the complete surrounding with a latitude / longitude mapping and thus the rendered images should best have a ratio of 2:1. A panoramic camera is created by passing the type string “panoramic” to ospNewCamera. It is placed and oriented in the scene by using the general parameters understood by all cameras.

Latitude / longitude map created with the panoramic camera.

Latitude / longitude map created with the panoramic camera.

Picking

To get the world-space position of the geometry (if any) seen at [0–1] normalized screen-space pixel coordinates screenPos use

void ospPick(OSPPickResult*, OSPRenderer, const vec2f &screenPos);

The result is returned in the provided OSPPickResult struct:

typedef struct {
    vec3f position; // the position of the hit point (in world-space)
    bool hit;       // whether or not a hit actually occurred
} OSPPickResult;

Note that ospPick considers exactly the same camera of the given renderer that is used to render an image, thus matching results can be expected. If the camera supports depth of field then the center of the lens and thus the center of the circle of confusion is used for picking.

Framebuffer

The framebuffer holds the rendered 2D image (and optionally auxiliary information associated with pixels). To create a new framebuffer object of given size size (in pixels), color format, and channels use

OSPFrameBuffer ospNewFrameBuffer(const vec2i &size,
                                 const OSPFrameBufferFormat format = OSP_FB_SRGBA,
                                 const uint32_t frameBufferChannels = OSP_FB_COLOR);

The parameter format describes the format the color buffer has on the host, and the format that ospMapFrameBuffer will eventually return. Valid values are:

Supported color formats of the framebuffer that can be passed to ospNewFrameBuffer, i.e. valid constants of type OSPFrameBufferFormat.
Name Description
OSP_FB_NONE framebuffer will not be mapped by the application
OSP_FB_RGBA8 8 bit [0–255] linear component red, green, blue, alpha
OSP_FB_SRGBA 8 bit sRGB gamma encoded color components, and linear alpha
OSP_FB_RGBA32F 32 bit float components red, green, blue, alpha

The parameter frameBufferChannels specifies which channels the framebuffer holds, and can be combined together by bitwise OR from the values of OSPFrameBufferChannel listed in the table below.

Framebuffer channels constants (of type OSPFrameBufferChannel), naming optional information the framebuffer can store. These values can be combined by bitwise OR when passed to ospNewFrameBuffer or ospClearFrameBuffer.
Name Description
OSP_FB_COLOR RGB color including alpha
OSP_FB_DEPTH euclidean distance to the camera (not to the image plane)
OSP_FB_ACCUM accumulation buffer for progressive refinement
OSP_FB_VARIANCE estimate of the current variance if OSP_FB_ACCUM is also present, see rendering

If a certain channel value is not specified, the given buffer channel will not be present. Note that ospray makes a very clear distinction between the external format of the framebuffer and the internal one: The external format is the format the user specifies in the format parameter; it specifies what color format OSPRay will eventually return the framebuffer to the application (when calling ospMapFrameBuffer): no matter what OSPRay uses internally, it will simply return a 2D array of pixels of that format, with possibly all kinds of reformatting, compression/decompression, etc, going on in-between the generation of the internal framebuffer and the mapping of the externally visible one.

In particular, OSP_FB_NONE is a perfectly valid pixel format for a framebuffer that an application will never map. For example, an application driving a display wall may well generate an intermediate framebuffer and eventually transfer its pixel to the individual displays using an OSPPixelOp pixel operation.

A framebuffer can be freed again using

void ospFreeFrameBuffer(OSPFrameBuffer);

Because OSPRay uses reference counting internally the framebuffer may not immediately be deleted at this time.

The application can map the given channel of a framebuffer – and thus access the stored pixel information – via

const void *ospMapFrameBuffer(OSPFrameBuffer,
                              const OSPFrameBufferChannel = OSP_FB_COLOR);

Note that only OSP_FB_COLOR or OSP_FB_DEPTH can be mapped. The origin of the screen coordinate system in OSPRay is the lower left corner (as in OpenGL), thus the first pixel addressed by the returned pointer is the lower left pixel of the image.

A previously mapped channel of a framebuffer can be unmapped by passing the received pointer mapped to

void ospUnmapFrameBuffer(const void *mapped, OSPFrameBuffer);

The individual channels of a framebuffer can be cleared with

void ospFrameBufferClear(OSPFrameBuffer, const uint32_t frameBufferChannels);

When selected, OSP_FB_COLOR will clear the color buffer to black (0, 0, 0, 0), OSP_FB_DEPTH will clear the depth buffer to inf, OSP_FB_ACCUM will clear the accumulation buffer to black, resets the accumulation counter accumID and also clears the variance buffer (if present) to inf.

Pixel Operation

A pixel operation are functions that are applied to every pixel that gets written into a framebuffer. Examples include post-processing, filtering, blending, tone mapping, or sending tiles to a display wall. To create a new pixel operation of given type type use

OSPPixelOp ospNewPixelOp(const char *type);

The call returns NULL if that type is not known, or else an OSPPixelOp handle to the created pixel operation.

To set a pixel operation to the given framebuffer use

void ospSetPixelOp(OSPFrameBuffer, OSPPixelOp);

Rendering

To render a frame into the given framebuffer with the given renderer use

float ospRenderFrame(OSPFrameBuffer, OSPRenderer,
                     const uint32_t frameBufferChannels = OSP_FB_COLOR);

The third parameter specifies what channel(s) of the framebuffer is written to4. What to render and how to render it depends on the renderer’s parameters. If the framebuffer supports accumulation (i.e. it was created with OSP_FB_ACCUM) then successive calls to ospRenderFrame will progressively refine the rendered image. If additionally the framebuffer has an OSP_FB_VARIANCE channel then ospRenderFrame returns an estimate of the current variance of the rendered image, otherwise inf is returned. The estimated variance can be used by the application as a quality indicator and thus to decide whether to stop or to continue progressive rendering.

Parallel Rendering with MPI

OSPRay has the ability to scale to multiple nodes in a cluster via MPI. This enables applications to take advantage of larger compute and memory resources when available.

Prerequisites for MPI Mode

In addition to the standard build requirements of OSPRay, you must have the following items available in your environment in order to build&run OSPRay in MPI mode:

Enabling the MPI Module in your Build

To build the MPI module the CMake option OSPRAY_MODULE_MPI must be enabled, which can be done directly on the command line (with -DOSPRAY_MODULE_MPI=ON) or through a configuration dialog (ccmake, cmake-gui), see also Compiling OSPRay.

This will trigger CMake to go look for an MPI implementation in your environment. You can then inspect the CMake value of MPI_LIBRARY to make sure that CMake found your MPI build environment correctly.

This will result in an OSPRay module being built. To enable using it, applications will need to either link libospray_module_mpi, or call

ospLoadModule("mpi");

before initializing OSPRay.

Modes of Using OSPRay’s MPI Features

OSPRay provides two ways of using MPI to scale up rendering: offload and distributed.

Offload Rendering

The “offload” rendering mode is where a single (not-distributed) calling application treats the OSPRay API the same as with local rendering. However, OSPRay uses multiple MPI connected nodes to evenly distribute frame rendering work, where each node contains a full copy of all scene data. This method is most effective for scenes which can fit into memory, but are very expensive to render: for example, path tracing with many samples-per-pixel is very compute heavy, making it a good situation to use the offload feature. This can be done with any application which already uses OSPRay for local rendering without the need for any code changes.

When doing MPI offload rendering, applications can optionally enable dynamic load balancing, which can be beneficial in certain contexts. This load balancing refers to the distribution of tile rendering work across nodes: thread-level load balancing on each node is still dynamic with the thread tasking system. The options for enabling/controlling the dynamic load balacing features on the mpi_offload device are found in the table below, which can be changed while the application is running. Please note that these options will likely only pay off for scenes which have heavy rendering load (e.g. path tracing a non-trivial scene) and have a lot of variance in how expensive each tile is to render.

Parameters specific to the mpi_offload device
Type Name Default Description
bool dynamicLoadBalancer false whether to use dynamic load balancing

Distributed Rendering

The “distributed” rendering mode is where a MPI distributed application (such as a scientific simulation) uses OSPRay collectively to render frames. In this case, the API expects all calls (both created objects and parameters) to be the same on every application rank, except each rank can specify arbitrary geometries and volumes. Each renderer will have its own limitations on the topology of the data (i.e. overlapping data regions, concave data, etc.), but the API calls will only differ for scene objects. Thus all other calls (i.e. setting camera, creating framebuffer, rendering frame, etc.) will all be assumed to be identical, but only rendering a frame and committing the model must be in lock-step. This mode targets using all available aggregate memory for very large scenes and for “in-situ” visualization where the data is already distributed by a simulation app.

Running an Application with the “offload” Device

As an example, our sample viewer can be run as a single application which offloads rendering work to multiple MPI processes running on multiple machines.

The example apps are setup to be launched in two different setups. In either setup, the application must initialize OSPRay with the offload device. This can be done by creating an “mpi_offload” device and setting it as the current device (via the ospSetCurrentDevice() function), or passing either “--osp:mpi” or “--osp:mpi-offload” as a command line parameter to ospInit(). Note that passing a command line parameter will automatically call ospLoadModule("mpi") to load the MPI module, while the application will have to load the module explicitly if using ospNewDevice().

Single MPI Launch

OSPRay is initialized with the ospInit() function call which takes command line arguments in and configures OSPRay based on what it finds. In this setup, the app is launched across all ranks, but workers will never return from ospInit(), essentially turning the application into a worker process for OSPRay. Here’s an example of running the ospVolumeViewer data-replicated, using c1-c4 as compute nodes and localhost the process running the viewer itself:

mpirun -perhost 1 -hosts localhost,c1,c2,c3,c4 ./ospExampleViewer <scene file> --osp:mpi

Separate Application&Worker Launches

The second option is to explicitly launch the app on rank 0 and worker ranks on the other nodes. This is done by running ospray_mpi_worker on worker nodes and the application on the display node. Here’s the same example above using this syntax:

mpirun -perhost 1 -hosts localhost ./ospExampleViewer <scene file> --osp:mpi \
  : -hosts c1,c2,c3,c4 ./ospray_mpi_worker

This method of launching the application and OSPRay worker separately works best for applications which do not immediately call ospInit() in their main() function, or for environments where application dependencies (such as GUI libraries) may not be available on compute nodes.

Running an Application with the “distributed” Device

Applications using the new distributed device should initialize OSPRay by creating (and setting current) an “mpi_distributed” device or pass "--osp:mpi-distributed" as a command line argument to ospInit(). Note that due to the semantic differences the distributed device gives the OSPRay API, it is not expected for applications which can already use the offload device to correctly use the distributed device without changes to the application.

Examples

Tutorial

A minimal working example demonstrating how to use OSPRay can be found at apps/ospTutorial.c5. On Linux build it in the build directory with

gcc -std=c99 ../apps/ospTutorial.c -I ../ospray/include -I .. \
./libospray.so -Wl,-rpath,. -o ospTutorial

On Windows build it in the “build_directory\$Configuration” with

cl ..\..\apps\ospTutorial.c -I ..\..\ospray\include -I ..\.. ospray.lib

Running ospTutorial will create two images of two triangles, rendered with the Scientific Visualization renderer with full Ambient Occlusion. The first image firstFrame.ppm shows the result after one call to ospRenderFrame – jagged edges and noise in the shadow can be seen. Calling ospRenderFrame multiple times enables progressive refinement, resulting in antialiased edges and converged shadows, shown after ten frames in the second image accumulatedFrames.ppm.

First frame.

First frame.

After accumulating ten frames.

After accumulating ten frames.

Example Viewer

OSPRay also includes an exemplary viewer application ospExampleViewerSg, showcasing all features of OSPRay. The Example Viewer uses the ImGui library for user interface controls. The viewer is based on a prototype OSPRay scenegraph interface where its nodes are displayed in the GUI and can be manipulated interactively. For instance, simply run it as ospExampleViewerSg teapot.obj.

This application also functions as an OSPRay state debugger – invalid values will be shown in red up the hierarchy and won’t change the viewer until corrected. You can also add new nodes where appropriate: for example, when “lights” is expanded right clicking on “lights” and typing in a light type, such as “point”, will add it to the scene. Similarly, right clicking on “world” and creating an “Importer” node will add a new scene importer from a file. Changing the filename to an appropriate file will load the scene and propagate the resulting state.

Screenshot of ospExampleViewerSg

Screenshot of ospExampleViewerSg

Distributed Viewer

The application ospDistribViewerDemo demonstrates how to write a distributed SciVis style interactive renderer using the distributed MPI device. Note that because OSPRay uses sort-last compositing it is up to the user to ensure that the data distribution across the nodes is suitable. Specifically, each nodes’ data must be convex and disjoint. This renderer supports multiple volumes and geometries per node. To ensure they are composited correctly you specify a list of bounding regions to the model, within these regions can be arbitrary volumes/geometries and each rank can have as many regions as needed. As long as the regions are disjoint/convex the data will be rendered correctly. In this demo we either generate a volume, or load a RAW volume file if one is passed on the commandline.

Loading a RAW Volume

To load a RAW volume you must specify the filename (-f <file>), the data type (-dtype <dtype>), the dimensions (-dims <x> <y> <z>) and the value range for the transfer function (-range <min> <max>). For example, to run on the CSAFE dataset from the demos page you would pass the following arguments:

mpirun -np <n> ./ospDistribViewerDemo \
    -f <path to csafe>/csafe-heptane-302-volume.raw \
    -dtype uchar -dims 302 302 302 -range 0 255

The volume file will then be chunked up into an x×y×z grid such that n = xyz. See loadVolume in gensv/generateSciVis.cpp for an example of how to properly load a volume distributed across ranks with correct specification of brick positions and ghost voxels for interpolation at boundaries. If no volume file data is passed a volume will be generated instead, in that case see makeVolume.

Geometry

The viewer can also display some randomly generated sphere geometry if you pass -spheres <n> where n is the number of spheres to generate per-node. These spheres will be generated inside the bounding box of the region’s volume data.

In the case that you have geometry crossing the boundary of nodes and are replicating it on both nodes to render (ghost zones, etc.) the region will be used by the renderer to clip rays against allowing to split the object between the two nodes, with each rendering half. This will keep the regions rendered by each rank disjoint and thus avoid any artifacts. For example, if a sphere center is on the border between two nodes, each would render half the sphere and the halves would be composited to produce the final complete sphere in the image.

App-initialized MPI

Passing the -appMPI flag will have the application initialize MPI instead of letting OSPRay do it internally when creating the MPI distributed device. In this case OSPRay will not finalize MPI when cleaning up the device, allowing the application to use OSPRay for some work, shut it down and recreate everything later if needed for additional computation, without accidentally shutting down its MPI communication.

Interactive Viewer

Rank 0 will open an interactive window with GLFW and display the rendered image. When the application state needs to update (e.g. camera or transfer function changes), this information is broadcasted out to the other nodes to update their scene data.

Demos

Several ready-to-run demos, models and data sets for OSPRay can be found at the OSPRay Demos and Examples page.


  1. The HDRI Light is an exception, it knows about intensity, but not about color.

  2. actually a parallelogram

  3. respectively (127, 127, 255) for 8 bit textures

  4. This is currently not implemented, i.e. all channels of the framebuffer are always updated.

  5. A C++ version that uses the C++ conveniance wrappers of OSPRay’s C99 API via include/ospray/ospray_cpp.h is available at apps/ospTutorial.cpp.