Open VKL API¶
To access the Open VKL API you first need to include the Open VKL header. For C99 or C++:
#include <openvkl/openvkl.h>
For the Intel SPMD Program Compiler (ISPC):
#include <openvkl/openvkl.isph>
This documentation will discuss the C99/C++ API. The ISPC version has the same functionality and flavor. Looking at the headers, the vklTutorialISPC example, and this documentation should be enough to figure it out.
Initialization and shutdown¶
To use the API, one of the implemented backends must be loaded. Currently the only one that exists is the ISPC driver. ISPC in the name here just refers to the implementation language – it can also be used from the C99/C++ APIs. To load the module that implements the ISPC driver:
vklLoadModule("ispc_driver");
The driver then needs to be instantiated:
VKLDriver driver = vklNewDriver("ispc");
By default, the ISPC driver selects the maximum supported SIMD width (and associated ISA) for the system. Optionally, a specific width may be requested using the ispc_4, ispc_8, or ispc_16 driver names. Note that the system must support the given width (SSE4.1 for 4-wide, AVX for 8-wide, and AVX512 for 16-wide).
Once a driver is created, you can call
void vklDriverSetInt(VKLDriver, const char *name, int val);
void vklDriverSetString(VKLDriver, const char *name, const char *val);
to set parameters on the driver. The following parameters are understood by all drivers:
Type |
Name |
Description |
|---|---|---|
int |
logLevel |
logging level; valid values are |
string |
logOutput |
convenience for setting where log messages go; valid values are |
string |
errorOutput |
convenience for setting where error messages go; valid values are |
int |
numThreads |
number of threads which Open VKL can use |
int |
flushDenormals |
sets the |
Once parameters are set, the driver must be committed with
vklCommitDriver(driver);
Finally, to use the newly committed driver, you must call
vklSetCurrentDriver(driver);
Users can change parameters on a driver after initialization. In this case the driver would need to be re-committed. If changes are made to the driver that is already set as the current driver, it does not need to be set as current again. The currently set driver can be retrieved at any time by calling
VKLDriver driver = vklGetCurrentDriver();
Open VKL provides vector-wide versions for several APIs. To determine the native vector width for the given driver, call:
int width = vklGetNativeSIMDWidth();
When the application is finished with Open VKL or shutting down, call the shutdown function:
vklShutdown();
Environment variables¶
The generic driver parameters can be overridden via environment variables for easy changes to Open VKL’s behavior without needing to change the application (variables are prefixed by convention with “OPENVKL_”):
Variable |
Description |
|---|---|
OPENVKL_LOG_LEVEL |
logging level; valid values are |
OPENVKL_LOG_OUTPUT |
convenience for setting where log messages go; valid values are |
OPENVKL_ERROR_OUTPUT |
convenience for setting where error messages go; valid values are |
OPENVKL_THREADS |
number of threads which Open VKL can use |
OPENVKL_FLUSH_DENORMALS |
sets the |
Note that these environment variables take precedence over values set through the vklDriverSet*() functions.
Error handling and log messages¶
The following errors are currently used by Open VKL:
Name |
Description |
|---|---|
VKL_NO_ERROR |
no error occurred |
VKL_UNKNOWN_ERROR |
an unknown error occurred |
VKL_INVALID_ARGUMENT |
an invalid argument was specified |
VKL_INVALID_OPERATION |
the operation is not allowed for the specified object |
VKL_OUT_OF_MEMORY |
there is not enough memory to execute the command |
VKL_UNSUPPORTED_CPU |
the CPU is not supported (minimum ISA is SSE4.1) |
These error codes are either directly returned by some API functions, or are recorded to be later queried by the application via
VKLError vklDriverGetLastErrorCode(VKLDriver);
A more descriptive error message can be queried by calling
const char* vklDriverGetLastErrorMsg(VKLDriver);
Alternatively, the application can also register a callback function of type
typedef void (*VKLErrorCallback)(void *, VKLError, const char* message);
via
void vklDriverSetErrorCallback(VKLDriver, VKLErrorFunc, void *);
to get notified when errors occur. Applications may be interested in messages which Open VKL emits, whether for debugging or logging events. Applications can register a callback function of type
typedef void (*VKLLogCallback)(void *, const char* message);
via
void vklDriverSetLogCallback(VKLDriver, VKLLogCallback, void *);
which Open VKL will use to emit log messages. Applications can clear either callback by passing nullptr instead of an actual function pointer. By default, Open VKL uses cout and cerr to emit log and error messages, respectively. The last parameter to vklDriverSetErrorCallback and vklDriverSetLogCallback is a user data pointer. Open VKL passes this pointer to the callback functions as the first parameter. Note that in addition to setting the above callbacks, this behavior can be changed via the driver parameters and environment variables described previously.
Basic data types¶
Open VKL defines 3-component vectors of integer and vector types:
typedef struct
{
int x, y, z;
} vkl_vec3i;
typedef struct
{
float x, y, z;
} vkl_vec3f;
Vector versions of these are also defined in structure-of-array format for 4, 8, and 16 wide types.
typedef struct
{
float x[WIDTH];
float y[WIDTH];
float z[WIDTH];
} vkl_vvec3f##WIDTH;
typedef struct
{
float lower[WIDTH], upper[WIDTH];
} vkl_vrange1f##WIDTH;
1-D range and 3-D ranges are defined as ranges and boxes, with no vector versions:
typedef struct
{
float lower, upper;
} vkl_range1f;
typedef struct
{
vkl_vec3f lower, upper;
} vkl_box3f;
Object model¶
Objects in Open VKL are exposed to the APIs as handles with internal reference counting for lifetime determination. Objects are created with particular type’s vklNew... API entry point. For example, vklNewData and vklNewVolume.
In general, modifiable parameters to objects are modified using vklSet... functions based on the type of the parameter being set. The parameter name is passed as a string. Below are all variants of vklSet....
void vklSetBool(VKLObject object, const char *name, int b);
void vklSetFloat(VKLObject object, const char *name, float x);
void vklSetVec3f(VKLObject object, const char *name, float x, float y, float z);
void vklSetInt(VKLObject object, const char *name, int x);
void vklSetVec3i(VKLObject object, const char *name, int x, int y, int z);
void vklSetData(VKLObject object, const char *name, VKLData data);
void vklSetString(VKLObject object, const char *name, const char *s);
void vklSetVoidPtr(VKLObject object, const char *name, void *v);
The exception to this rule is the VKLValueSelector object (described in the iterators section below), which has object-specific set methods. The reason for this is to align the C99/C++ API with the ISPC API, which can’t use a parameter method due to language limitations.
After parameters have been set, vklCommit must be called on the object to make them take effect.
Open VKL uses reference counting to manage the lifetime of all objects. Therefore one cannot explicitly “delete” any object. Instead, one can indicate the application does not need or will not access the given object anymore by calling
void vklRelease(VKLObject);
This decreases the object’s reference count. If the count reaches 0 the object will automatically be deleted.
Managed data¶
Large data is passed to Open VKL via a VKLData handle created with vklNewData:
VKLData vklNewData(size_t numItems,
VKLDataType dataType,
const void *source,
VKLDataCreationFlags dataCreationFlags,
size_t byteStride);
Types accepted are listed in VKLDataType.h; basic types (UCHAR, INT, UINT, LONG, ULONG, FLOAT) exist as both scalar and chunked formats. The types accepted vary per volume at the moment; read the volume section below for specifics.
Data objects can be created as Open VKL owned (dataCreationFlags = VKL_DATA_DEFAULT), in which the library will make a copy of the data for its use, or shared (dataCreationFlags = VKL_DATA_SHARED_BUFFER), which will try to use the passed pointer for usage. The library is allowed to copy data when a volume is committed.
The distance between consecutive elements in source is given in bytes with byteStride. If the provided byteStride is zero, then it will be determined automatically as sizeof(type). Open VKL owned data will be compacted into a naturally-strided array on copy, regardless of the original byteStride.
As with other object types, when data objects are no longer needed they should be released via vklRelease.
Observers¶
Volumes and samplers in Open VKL may provide observers to communicate data back to the application. Observers may be created with
VKLObserver vklNewSamplerObserver(VKLSampler sampler,
const char *type);
VKLObserver vklNewVolumeObserver(VKLVolume volume,
const char *type);
The object passed to vklNew*Observer must already be committed. Valid observer type strings are defined by volume implementations (see section ‘Volume types’ below).
vklNew*Observer returns NULL on failure.
To access the underlying data, an observer must first be mapped using
const void * vklMapObserver(VKLObserver observer);
If this fails, the function returns NULL. vklMapObserver may fail on observers that are already mapped. On success, the application may query the underlying type and the number of elements in the buffer using
VKLDataType vklGetObserverElementType(VKLObserver observer);
size_t vklGetObserverNumElements(VKLObserver observer);
On failure, these functions return VKL_UNKNOWN and 0, respectively. Possible data types are defined by the volume that provides the observer , as are the semantics of the observation. See section ‘Volume types’ for details.
The pointer returned by vklMapObserver may be cast to the type corresponding to the value returned by vklGetObserverElementType to access the observation. For example, if vklGetObserverElementType returns VKL_FLOAT, then the pointer returned by vklMapObserver may be cast to const float * to access up to vklGetObserverNumElements consecutive values of type float.
Once the application has finished processing the observation, it should unmap the observer using
void vklUnmapObserver(VKLObserver observer);
so that the observer may be mapped again.
When an observer is no longer needed, it should be released using vklRelease.
The observer API is not thread safe, and these functions should not be called concurrently on the same object.
Volume types¶
Open VKL currently supports structured volumes on regular and spherical grids; unstructured volumes with tetrahedral, wedge, pyramid, and hexaderal primitive types; adaptive mesh refinement (AMR) volumes; sparse VDB volumes; and particle volumes. These volumes are created with vklNewVolume with the appropriate type string.
In addition to the usual vklSet...() and vklCommit() APIs, the volume bounding box can be queried:
vkl_box3f vklGetBoundingBox(VKLVolume volume);
The number of attributes in a volume can also be queried:
unsigned int vklGetNumAttributes(VKLVolume volume);
Finally, the value range of the volume (first attribute only) can be queried:
vkl_range1f vklGetValueRange(VKLVolume volume);
Structured Volumes¶
Structured volumes only need to store the values of the samples, because their addresses in memory can be easily computed from a 3D position. The dimensions for all structured volume types are in units of vertices, not cells. For example, a volume with dimensions \((x, y, z)\) will have \((x-1, y-1, z-1)\) cells in each dimension. Voxel data provided is assumed vertex-centered, so \(x*y*z\) values must be provided.
Structured Regular Volumes¶
A common type of structured volumes are regular grids, which are created by passing a type string of "structuredRegular" to vklNewVolume. The parameters understood by structured regular volumes are summarized in the table below.
Type |
Name |
Default |
Description |
|---|---|---|---|
vec3i |
dimensions |
number of voxels in each dimension \((x, y, z)\) |
|
VKLData VKLData[] |
data |
VKLData object(s) of voxel data, supported types are: |
|
|
|||
|
|||
|
|||
|
|||
|
|||
Multiple attributes are supported through passing an array of VKLData objects. |
|||
vec3f |
gridOrigin |
\((0, 0, 0)\) |
origin of the grid in object space |
vec3f |
gridSpacing |
\((1, 1, 1)\) |
size of the grid cells in object space |
int |
temporallyStructuredNumTimesteps |
for temporally structured variation, number of timesteps per voxel |
|
uint32[] uint64[] |
temporallyUnstructuredIndices |
for temporally unstructured variation, indices to |
|
float[] |
temporallyUnstructuredTimes |
for temporally unstructured variation, time values corresponding to values in |
Structured regular volumes support two forms of temporal variation: temporally structured and temporally unstructured. When one of these modes is enabled, the volume can be sampled at different times. In both modes, time is assumed to vary between zero and one. This can be useful for implementing renderers with motion blur, for example.
Temporally structured variation is configured through the temporallyStructuredNumTimesteps parameter. This specifies how many time steps (at least two) are provided for all voxels. Therefore, for a volume with dimensions \((x, y, z)\), each attribute must have \(x*y*z * temporallyStructuredNumTimesteps\) values provided in its data array. The values are assumed evenly spaced over times \([0, 1]\).
Temporally unstructured variation is configured through the temporallyUnstructuredIndices and temporallyUnstructuredTimes parameters, and supports differing time step counts and sample times per voxel. temporallyUnstructuredIndices specifies the index ranges for each voxel’s values in data, such that values for the the \(i\)th voxel can be found in the indices \([temporallyUnstructuredIndices[i], temporallyUnstructuredIndices[i+1])\). Therefore temporallyUnstructuredIndices must have \(x*y*z + 1\) values. temporallyUnstructuredTimes specifies the times corresponding to the sample values in each attribute’s data array; the time values for each voxel must be between zero and one and strictly increasing: \(t0 < t1 < ... < tN\). To return a value at sample time t, \(t0 <= t <= tN\), Open VKL will interpolate linearly from the two nearest time steps. Time values outside this range are clamped \([t0, tN]\).
The following additional parameters can be set both on "structuredRegular" volumes and their sampler objects. Sampler object parameters default to volume parameters.
Type |
Name |
Default |
Description |
|---|---|---|---|
int |
filter |
|
The filter used for reconstructing the field. Use |
int |
gradientFilter |
|
The filter used for reconstructing the field during gradient computations. Use |
Structured Spherical Volumes¶
Structured spherical volumes are also supported, which are created by passing a type string of "structuredSpherical" to vklNewVolume. The grid dimensions and parameters are defined in terms of radial distance (\(r\)), inclination angle (\(\theta\)), and azimuthal angle (\(\phi\)), conforming with the ISO convention for spherical coordinate systems. The coordinate system and parameters understood by structured spherical volumes are summarized below.
Structured spherical volume coordinate system: radial distance (\(r\)), inclination angle (\(\theta\)), and azimuthal angle (\(\phi\)).¶
Type |
Name |
Default |
Description |
|---|---|---|---|
vec3i |
dimensions |
number of voxels in each dimension \((r, \theta, \phi)\) |
|
VKLData VKLData[] |
data |
VKLData object(s) of voxel data, supported types are: |
|
|
|||
|
|||
|
|||
|
|||
|
|||
Multiple attributes are supported through passing an array of VKLData objects. |
|||
vec3f |
gridOrigin |
\((0, 0, 0)\) |
origin of the grid in units of \((r, \theta, \phi)\); angles in degrees |
vec3f |
gridSpacing |
\((1, 1, 1)\) |
size of the grid cells in units of \((r, \theta, \phi)\); angles in degrees |
These grid parameters support flexible specification of spheres, hemispheres, spherical shells, spherical wedges, and so forth. The grid extents (computed as \([gridOrigin, gridOrigin + (dimensions - 1) * gridSpacing]\)) however must be constrained such that:
\(r \geq 0\)
\(0 \leq \theta \leq 180\)
\(0 \leq \phi \leq 360\)
The following additional parameters can be set both on "structuredSpherical" volumes and their sampler objects. Sampler object parameters default to volume parameters.
Type |
Name |
Default |
Description |
|---|---|---|---|
int |
filter |
|
The filter used for reconstructing the field. Use |
int |
gradientFilter |
|
The filter used for reconstructing the field during gradient computations. Use |
Adaptive Mesh Refinement (AMR) Volumes¶
Open VKL currently supports block-structured (Berger-Colella) AMR volumes. Volumes are specified as a list of blocks, which exist at levels of refinement in potentially overlapping regions. Blocks exist in a tree structure, with coarser refinement level blocks containing finer blocks. The cell width is equal for all blocks at the same refinement level, though blocks at a coarser level have a larger cell width than finer levels.
There can be any number of refinement levels and any number of blocks at any level of refinement.
Blocks are defined by three parameters: their bounds, the refinement level in which they reside, and the scalar data contained within each block.
Note that cell widths are defined per refinement level, not per block.
AMR volumes are created by passing the type string "amr" to vklNewVolume, and have the following parameters:
Type |
Name |
Default |
Description |
|---|---|---|---|
float[] |
cellWidth |
[data] array of each level’s cell width |
|
box3i[] |
block.bounds |
[data] array of each block’s bounds (in voxels) |
|
int[] |
block.level |
[data] array of each block’s refinement level |
|
VKLData[] |
block.data |
[data] array of each block’s VKLData object containing the actual scalar voxel data. Currently only |
|
vec3f |
gridOrigin |
\((0, 0, 0)\) |
origin of the grid in object space |
vec3f |
gridSpacing |
\((1, 1, 1)\) |
size of the grid cells in object space |
Note that the gridOrigin and gridSpacing parameters act just like the structured volume equivalent, but they only modify the root (coarsest level) of refinement.
The following additional parameters can be set both on "amr" volumes and their sampler objects. Sampler object parameters default to volume parameters.
Type |
Name |
Default |
Description |
|---|---|---|---|
|
method |
|
|
|
|||
|
|||
|
Open VKL’s AMR implementation was designed to cover Berger-Colella [1] and Chombo [2] AMR data. The method parameter above determines the interpolation method used when sampling the volume.
VKL_AMR_CURRENTfinds the finest refinement level at that cell and interpolates through this “current” levelVKL_AMR_FINESTwill interpolate at the closest existing cell in the volume-wide finest refinement level regardless of the sample cell’s levelVKL_AMR_OCTANTinterpolates through all available refinement levels at that cell. This method avoids discontinuities at refinement level boundaries at the cost of performance
Details and more information can be found in the publication for the implementation [3].
Berger, and P. Colella. “Local adaptive mesh refinement for shock hydrodynamics.” Journal of Computational Physics 82.1 (1989): 64-84. DOI: 10.1016/0021-9991(89)90035-1
Adams, P. Colella, D. T. Graves, J.N. Johnson, N.D. Keen, T. J. Ligocki. D. F. Martin. P.W. McCorquodale, D. Modiano. P.O. Schwartz, T.D. Sternberg and B. Van Straalen, Chombo Software Package for AMR Applications - Design Document, Lawrence Berkeley National Laboratory Technical Report LBNL-6616E.
Wald, C. Brownlee, W. Usher, and A. Knoll. CPU volume rendering of adaptive mesh refinement data. SIGGRAPH Asia 2017 Symposium on Visualization on - SA ’17, 18(8), 1–8. DOI: 10.1145/3139295.3139305
Unstructured Volumes¶
Unstructured volumes can have their topology and geometry freely defined. Geometry can be composed of tetrahedral, hexahedral, wedge or pyramid cell types. The data format used is compatible with VTK and consists of multiple arrays: vertex positions and values, vertex indices, cell start indices, cell types, and cell values.
Sampled cell values can be specified either per-vertex (vertex.data) or per-cell (cell.data). If both arrays are set, cell.data takes precedence.
Similar to a mesh, each cell 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, if specified. The index order for a tetrahedron is the same as VTK_TETRA: bottom triangle counterclockwise, then the top vertex.
For hexahedral cells, each hexahedron is formed by a group of eight indices into the vertices and data values. Vertex ordering is the same as VTK_HEXAHEDRON: four bottom vertices counterclockwise, then top four counterclockwise.
For wedge cells, each wedge is formed by a group of six indices into the vertices and data values. Vertex ordering is the same as VTK_WEDGE: three bottom vertices counterclockwise, then top three counterclockwise.
For pyramid cells, each cell is formed by a group of five indices into the vertices and data values. Vertex ordering is the same as VTK_PYRAMID: four bottom vertices counterclockwise, then the top vertex.
To maintain VTK data compatibility, the index array may be specified with cell sizes interleaved with vertex indices in the following format: \(n, id_1, ..., id_n, m, id_1, ..., id_m\). This alternative index array layout can be enabled through the indexPrefixed flag (in which case, the cell.type parameter should be omitted).
A binary bounding volume hierarchy (BVH) is used internally to accelerate interval iteration. Intervals are found by intersecting BVH nodes up to a maximum level of the tree, configurable by the maxIteratorDepth parameter. Larger values of maxIteratorDepth lead to smaller individual intervals (up to leaf node intersections), yielding potentially more efficient space-skipping behavior and tighter bounds on returned interval metadata.
Unstructured volumes are created by passing the type string "unstructured" to vklNewVolume, and have the following parameters:
Type |
Name |
Default |
Description |
|---|---|---|---|
vec3f[] |
vertex.position |
[data] array of vertex positions |
|
float[] |
vertex.data |
[data] array of vertex data values to be sampled |
|
uint32[] / uint64[] |
index |
[data] array of indices (into the vertex array(s)) that form cells |
|
bool |
indexPrefixed |
false |
indicates that the |
uint32[] / uint64[] |
cell.index |
[data] array of locations (into the index array), specifying the first index of each cell |
|
float[] |
cell.data |
[data] array of cell data values to be sampled |
|
uint8[] |
cell.type |
[data] array of cell types (VTK compatible). Supported types are: |
|
|
|||
|
|||
|
|||
|
|||
bool |
hexIterative |
false |
hexahedron interpolation method, defaults to fast non-iterative version which could have rendering inaccuracies may appear if hex is not parallelepiped |
bool |
precomputedNormals |
false |
whether to accelerate by precomputing, at a cost of 12 bytes/face |
The following additional parameters can be set both on unstructured volumes and their sampler objects (sampler object parameters default to volume parameters).
Type |
Name |
Default |
Description |
|---|---|---|---|
int |
maxIteratorDepth |
6 |
Do not descend further than to this BVH depth during interval iteration. |
VDB Volumes¶
VDB volumes implement a data structure that is very similar to the data structure outlined in Museth [1].
The data structure is a hierarchical regular grid at its core: Nodes are regular grids, and each grid cell may either store a constant value (this is called a tile), or child pointers.
Nodes in VDB trees are wide: Nodes on the first level have a resolution of 32^3 voxels by default, on the next level 16^3, and on the leaf level 8^3 voxels. All nodes on a given level have the same resolution. This makes it easy to find the node containing a coordinate using shift operations (cp. [1]).
VDB leaf nodes are implicit in Open VKL: they are stored as pointers to user-provided data.
Structure of "vdb" volumes in the default configuration¶
VDB volumes interpret input data as constant cells (which are then potentially filtered). This is in contrast to structuredRegular volumes, which have a vertex-centered interpretation.
The VDB implementation in Open VKL follows the following goals:
Efficient data structure traversal on vector architectures.
Enable the use of industry-standard .vdb files created through the OpenVDB library.
Compatibility with OpenVDB on a leaf data level, so that .vdb files may be loaded with minimal overhead.
VDB volumes are created by passing the type string "vdb" to vklNewVolume, and have the following parameters:
Type |
Name |
Default |
Description |
|---|---|---|---|
float[] |
indexToObject |
1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0 |
An array of 12 values of type |
uint32[] |
node.format |
For each input node, the data format. Currently supported are |
|
uint32[] |
node.level |
For each input node, the level on which this node exists. Tiles may exist on levels [1, |
|
vec3i[] |
node.origin |
For each input node, the node origin index. |
|
VKLData[] |
node.data |
For each input node, the attribute data. Single-attribute volumes may have one array provided per node, while multi-attribute volumes require an array per attribute for each node. Nodes with format |
The level, origin, format, and data parameters must have the same size, and there must be at least one valid node or commit() will fail.
The following additional parameters can be set both on vdb volumes and their sampler objects (sampler object parameters default to volume parameters).
Type |
Name |
Default |
Description |
|---|---|---|---|
int |
filter |
|
The filter used for reconstructing the field. Use |
int |
gradientFilter |
|
The filter used for reconstructing the field during gradient computations. Use |
int |
maxSamplingDepth |
|
Do not descend further than to this depth during sampling. |
int |
maxIteratorDepth |
|
Do not descend further than to this depth during iteration. |
VDB sampler objects support the following observers:
Name |
Buffer Type |
Description |
|---|---|---|
LeafNodeAccess |
uint32[] |
This observer returns an array with as many entries as input nodes were passed. If the input node i was accessed during traversal, then the ith entry in this array has a nonzero value. This can be used for on-demand loading of leaf nodes. |
Major differences to OpenVDB¶
Open VKL implements sampling in ISPC, and can exploit wide SIMD architectures.
VDB volumes in Open VKL are read-only once committed, and designed for rendering only. Authoring or manipulating datasets is not in the scope of this implementation.
The only supported field type is
VKL_FLOATat this point. Other field types may be supported in the future.The root level in Open VKL has a single node with resolution 64^3 (cp. [1]. OpenVDB uses a hash map, instead).
Open VKL supports four-level vdb volumes. The resolution of each level can be configured at compile time using CMake variables.
VKL_VDB_LOG_RESOLUTION_0sets the base 2 logarithm of the root level resolution. This variable defaults to 6, which means that the root level has a resolution of \((2^6)^3 = 64^3\).VKL_VDB_LOG_RESOLUTION_1andVKL_VDB_LOG_RESOLUTION_2default to 5 and 4, respectively. This matches the default Open VDB resolution for inner levels.VKL_VDB_LOG_RESOLUTION_3set the base 2 logarithm of the leaf level resolution, and defaults to 3. Therefore, leaf nodes have a resolution of \(8^3\) voxels. Again, this matches the Open VDB default. The default settings lead to a domain resolution of \(2^18^3=262144^3\) voxels.
Loading OpenVDB .vdb files¶
Files generated with OpenVDB can be loaded easily since Open VKL vdb volumes implement the same leaf data layout. This means that OpenVDB leaf data pointers can be passed to Open VKL using shared data buffers, avoiding copy operations.
An example of this can be found in vdb_util/include/openvkl/OpenVdbGrid.h, where the class OpenVdbFloatGrid encapsulates the necessary operations. This class is also accessible through the vklExamples application using the -file and -field command line arguments.
To use this example feature, compile Open VKL with OpenVDB_ROOT pointing to the OpenVDB prefix.
Museth, K. VDB: High-Resolution Sparse Volumes with Dynamic Topology. ACM Transactions on Graphics 32(3), 2013. DOI: 10.1145/2487228.2487235
Particle Volumes¶
Particle volumes consist of a set of points in space. Each point has a position, a radius, and a weight typically associated with an attribute. A radial basis function defines the contribution of that particle. Currently, we use the Gaussian radial basis function,
phi(P) = w * exp( -0.5 * ((P - p) / r)^2 )
where P is the particle position, p is the sample position, r is the radius and w is the weight.
At each sample, the scalar field value is then computed as the sum of each radial basis function phi, for each particle that overlaps it.
The Open VKL implementation is similar to direct evaluation of samples in Reda et al.[2]. It uses an Embree-built BVH with a custom traversal, similar to the method in [1].
Similar to unstructured volumes, a binary BVH is used internally to accelerate interval iteration. Intervals are found by intersecting BVH nodes up to a maximum level of the tree, configurable by the maxIteratorDepth parameter.
Particle volumes are created by passing the type string "particle" to vklNewVolume, and have the following parameters:
Type |
Name |
Default |
Description |
|---|---|---|---|
vec3f[] |
particle.position |
[data] array of particle positions |
|
float[] |
particle.radius |
[data] array of particle radii |
|
float[] |
particle.weight |
null |
[data] (optional) array of particle weights, specifying the height of the kernel. |
float |
radiusSupportFactor |
3.0 |
The multipler of the particle radius required for support. Larger radii ensure smooth results at the cost of performance. In the Gaussian kernel, the the radius is one standard deviation (sigma), so a |
float |
clampMaxCumulativeValue |
0 |
The maximum cumulative value possible, set by user. All cumulative values will be clamped to this, and further traversal (RBF summation) of particle contributions will halt when this value is reached. A value of zero or less turns this off. |
bool |
estimateValueRanges |
true |
Enable heuristic estimation of value ranges which are used in internal acceleration structures for interval and hit iterators, as well as for determining the volume’s overall value range. When set to |
The following additional parameters can be set both on particle volumes and their sampler objects (sampler object parameters default to volume parameters).
Type |
Name |
Default |
Description |
|---|---|---|---|
int |
maxIteratorDepth |
6 |
Do not descend further than to this BVH depth during interval iteration. |
Knoll, A., Wald, I., Navratil, P., Bowen, A., Reda, K., Papka, M.E. and Gaither, K. (2014), RBF Volume Ray Casting on Multicore and Manycore CPUs. Computer Graphics Forum, 33: 71-80. doi:10.1111/cgf.12363
Reda, A. Knoll, K. Nomura, M. E. Papka, A. E. Johnson and J. Leigh, “Visualizing large-scale atomistic simulations in ultra-resolution immersive environments,” 2013 IEEE Symposium on Large-Scale Data Analysis and Visualization (LDAV), Atlanta, GA, 2013, pp. 59-65.
Sampler Objects¶
Computing the value of a volume at an object space coordinate is done using the sampling API, and sampler objects. Sampler objects can be created using
VKLSampler vklNewSampler(VKLVolume volume);
Sampler objects may then be parametrized with traversal parameters. Available parameters are defined by volumes, and are a subset of the volume parameters. As an example, filter can be set on both vdb volumes and their sampler objects. The volume parameter is used as the default for sampler objects. The sampler object parameter provides an override per ray. More detail on parameters can be found in the sections on volumes. Use vklCommit() to commit parameters to the sampler object.
Sampling¶
The scalar API takes a volume and coordinate, and returns a float value. NaN is returned for probe points outside the volume. The attribute index selects the scalar attribute of interest; not all volumes support multiple attributes. The time value, which must be between 0 and 1, specifies the sampling time. For temporally constant volumes, this value has no effect.
float vklComputeSample(VKLSampler sampler,
const vkl_vec3f *objectCoordinates,
unsigned int attributeIndex,
float time);
Vector versions allow sampling at 4, 8, or 16 positions at once. Depending on the machine type and Open VKL driver implementation, these can give greater performance. An active lane mask valid is passed in as an array of integers; set 0 for lanes to be ignored, -1 for active lanes. An array of time values corresponding to each object coordinate may be provided; a NULL value indicates all times are zero.
void vklComputeSample4(const int *valid,
VKLSampler sampler,
const vkl_vvec3f4 *objectCoordinates,
float *samples,
unsigned int attributeIndex,
const float *times);
void vklComputeSample8(const int *valid,
VKLSampler sampler,
const vkl_vvec3f8 *objectCoordinates,
float *samples,
unsigned int attributeIndex,
const float *times);
void vklComputeSample16(const int *valid,
VKLSampler sampler,
const vkl_vvec3f16 *objectCoordinates,
float *samples,
unsigned int attributeIndex,
const float *times);
A stream version allows sampling an arbitrary number of positions at once. While the vector version requires coordinates to be provided in a structure-of-arrays layout, the stream version allows coordinates to be provided in an array-of-structures layout. Thus, the stream API can be used to avoid reformatting of data by the application. As with the vector versions, the stream API can give greater performance than the scalar API.
void vklComputeSampleN(VKLSampler sampler,
unsigned int N,
const vkl_vec3f *objectCoordinates,
float *samples,
unsigned int attributeIndex,
const float *times);
All of the above sampling APIs can be used, regardless of the driver’s native SIMD width.
Sampling Multiple Attributes¶
Open VKL provides additional APIs for sampling multiple scalar attributes in a single call through the vklComputeSampleM*() interfaces. Beyond convenience, these can give improved performance relative to the single attribute sampling APIs. As with the single attribute APIs, sampling time values may be specified; note that these are provided per object coordinate only (rather than separately per attribute).
A scalar API supports sampling M attributes specified by attributeIndices on a single object space coordinate:
void vklComputeSampleM(VKLSampler sampler,
const vkl_vec3f *objectCoordinates,
float *samples,
unsigned int M,
const unsigned int *attributeIndices,
float time);
Vector versions allow sampling at 4, 8, or 16 positions at once across the M attributes:
void vklComputeSampleM4(const int *valid,
VKLSampler sampler,
const vkl_vvec3f4 *objectCoordinates,
float *samples,
unsigned int M,
const unsigned int *attributeIndices,
const float *times);
void vklComputeSampleM8(const int *valid,
VKLSampler sampler,
const vkl_vvec3f8 *objectCoordinates,
float *samples,
unsigned int M,
const unsigned int *attributeIndices,
const float *times);
void vklComputeSampleM16(const int *valid,
VKLSampler sampler,
const vkl_vvec3f16 *objectCoordinates,
float *samples,
unsigned int M,
const unsigned int *attributeIndices,
const float *times);
The [4, 8, 16] * M sampled values are populated in the samples array in a structure-of-arrays layout, with all values for each attribute provided in sequence. That is, sample values s_m,n for the mth attribute and nth object coordinate will be populated as
samples = [s_0,0, s_0,1, ..., s_0,N-1,
s_1,0, s_1,1, ..., s_1,N-1,
...,
s_M-1,0, s_M-1,1, ..., s_M-1,N-1]
A stream version allows sampling an arbitrary number of positions at once across the M attributes. As with single attribute stream sampling, the N coordinates are provided in an array-of-structures layout.
void vklComputeSampleMN(VKLSampler sampler,
unsigned int N,
const vkl_vec3f *objectCoordinates,
float *samples,
unsigned int M,
const unsigned int *attributeIndices,
const float *times);
The M * N sampled values are populated in the samples array in an array-of-structures layout, with all attribute values for each coordinate provided in sequence as
samples = [s_0,0, s_1,0, ..., s_M-1,0,
s_0,1, s_1,1, ..., s_M-1,1,
...,
s_0,N-1, s_1,N-1, ..., s_M-1,N-1]
All of the above sampling APIs can be used, regardless of the driver’s native SIMD width.
Gradients¶
In a very similar API to vklComputeSample, vklComputeGradient queries the value gradient at an object space coordinate. Again, a scalar API, now returning a vec3f instead of a float. NaN values are returned for points outside the volume. The time value, which must be between 0 and 1, specifies the sampling time. For temporally constant volumes, this value has no effect.
vkl_vec3f vklComputeGradient(VKLSampler sampler,
const vkl_vec3f *objectCoordinates,
unsigned int attributeIndex,
float time);
Vector versions are also provided:
void vklComputeGradient4(const int *valid,
VKLSampler sampler,
const vkl_vvec3f4 *objectCoordinates,
vkl_vvec3f4 *gradients,
unsigned int attributeIndex,
const float *times);
void vklComputeGradient8(const int *valid,
VKLSampler sampler,
const vkl_vvec3f8 *objectCoordinates,
vkl_vvec3f8 *gradients,
unsigned int attributeIndex,
const float *times);
void vklComputeGradient16(const int *valid,
VKLSampler sampler,
const vkl_vvec3f16 *objectCoordinates,
vkl_vvec3f16 *gradients,
unsigned int attributeIndex,
const float *times);
Finally, a stream version is provided:
void vklComputeGradientN(VKLSampler sampler,
unsigned int N,
const vkl_vec3f *objectCoordinates,
vkl_vec3f *gradients,
unsigned int attributeIndex,
const float *times);
All of the above gradient APIs can be used, regardless of the driver’s native SIMD width.
Iterators¶
Open VKL has APIs to search for particular volume values along a ray. Queries can be for ranges of volume values (vklIterateInterval) or for particular values (vklIterateHit). Only the first volume attribute is currently considered in the iterator APIs.
The desired values are set in a VKLValueSelector, which needs to be created, filled in with values, and then committed.
VKLValueSelector vklNewValueSelector(VKLVolume volume);
void vklValueSelectorSetRanges(VKLValueSelector valueSelector,
size_t numRanges,
const vkl_range1f *ranges);
void vklValueSelectorSetValues(VKLValueSelector valueSelector,
size_t numValues,
const float *values);
To query an interval, a VKLIntervalIterator of scalar or vector width must be initialized with vklInitIntervalIterator.
VKLIntervalIterator vklInitIntervalIterator(VKLSampler sampler,
const vkl_vec3f *origin,
const vkl_vec3f *direction,
const vkl_range1f *tRange,
VKLValueSelector valueSelector,
void *buffer);
VKLIntervalIterator4 vklInitIntervalIterator4(const int *valid,
VKLSampler sampler,
const vkl_vvec3f4 *origin,
const vkl_vvec3f4 *direction,
const vkl_vrange1f4 *tRange,
VKLValueSelector valueSelector,
void *buffer);
VKLIntervalIterator8 vklInitIntervalIterator8(const int *valid,
VKLSampler sampler,
const vkl_vvec3f8 *origin,
const vkl_vvec3f8 *direction,
const vkl_vrange1f8 *tRange,
VKLValueSelector valueSelector,
void *buffer);
VKLIntervalIterator16 vklInitIntervalIterator16(const int *valid,
VKLSampler sampler,
const vkl_vvec3f16 *origin,
const vkl_vvec3f16 *direction,
const vkl_vrange1f16 *tRange,
VKLValueSelector valueSelector,
void *buffer);
Open VKL places the iterator struct into a user-provided buffer, and the returned handle is essentially a pointer into this buffer. This means that the iterator handle must not be used after the buffer ceases to exist. Copying iterator buffers is currently not supported.
The required size, in bytes, of the buffer can be queried with
size_t vklGetIntervalIteratorSize(VKLSampler sampler);
size_t vklGetIntervalIteratorSize4(VKLSampler sampler);
size_t vklGetIntervalIteratorSize8(VKLSampler sampler);
size_t vklGetIntervalIteratorSize16(VKLSampler sampler);
The values these functions return may change depending on the parameters set on sampler.
Open VKL also provides a conservative maximum size over all volume types as a preprocessor definition (VKL_MAX_INTERVAL_ITERATOR_SIZE). This is particularly useful for stack-based allocation in ISPC. Open VKL will attempt to detect the native vector width using TARGET_WIDTH, which is defined in recent versions of ISPC.
Intervals can then be processed by calling vklIterateInterval as long as the returned lane masks indicates that the iterator is still within the volume:
int vklIterateInterval(VKLIntervalIterator iterator,
VKLInterval *interval);
void vklIterateInterval4(const int *valid,
VKLIntervalIterator4 iterator,
VKLInterval4 *interval,
int *result);
void vklIterateInterval8(const int *valid,
VKLIntervalIterator8 iterator,
VKLInterval8 *interval,
int *result);
void vklIterateInterval16(const int *valid,
VKLIntervalIterator16 iterator,
VKLInterval16 *interval,
int *result);
The intervals returned have a t-value range, a value range, and a nominalDeltaT which is approximately the step size (in units of ray direction) that should be used to walk through the interval, if desired. The number and length of intervals returned is volume type implementation dependent. There is currently no way of requesting a particular splitting.
typedef struct
{
vkl_range1f tRange;
vkl_range1f valueRange;
float nominalDeltaT;
} VKLInterval;
typedef struct
{
vkl_vrange1f4 tRange;
vkl_vrange1f4 valueRange;
float nominalDeltaT[4];
} VKLInterval4;
typedef struct
{
vkl_vrange1f8 tRange;
vkl_vrange1f8 valueRange;
float nominalDeltaT[8];
} VKLInterval8;
typedef struct
{
vkl_vrange1f16 tRange;
vkl_vrange1f16 valueRange;
float nominalDeltaT[16];
} VKLInterval16;
Querying for particular values is done using a VKLHitIterator in much the same fashion. This API could be used, for example, to find isosurfaces. In contrast to interval iterators, time value(s) may be provided to specify the sampling time. These values must be between 0 and 1; for the vector versions, a NULL value indicates all times are zero. For temporally constant volumes, the time values have no effect. Again, a user allocated buffer must be provided, and a VKLHitIterator of the desired width must be initialized:
VKLHitIterator vklInitHitIterator(VKLSampler sampler,
const vkl_vec3f *origin,
const vkl_vec3f *direction,
const vkl_range1f *tRange,
float time,
VKLValueSelector valueSelector,
void *buffer);
VKLHitIterator4 vklInitHitIterator4(const int *valid,
VKLSampler sampler,
const vkl_vvec3f4 *origin,
const vkl_vvec3f4 *direction,
const vkl_vrange1f4 *tRange,
const float *times,
VKLValueSelector valueSelector,
void *buffer);
VKLHitIterator8 vklInitHitIterator8(const int *valid,
VKLSampler sampler,
const vkl_vvec3f8 *origin,
const vkl_vvec3f8 *direction,
const vkl_vrange1f8 *tRange,
const float *times,
VKLValueSelector valueSelector,
void *buffer);
VKLHitIterator16 vklInitHitIterator16(const int *valid,
VKLSampler sampler,
const vkl_vvec3f16 *origin,
const vkl_vvec3f16 *direction,
const vkl_vrange1f16 *tRange,
const float *times,
VKLValueSelector valueSelector,
void *buffer);
Buffer size can be queried with
size_t vklGetHitIteratorSize(VKLSampler sampler);
size_t vklGetHitIteratorSize4(VKLSampler sampler);
size_t vklGetHitIteratorSize8(VKLSampler sampler);
size_t vklGetHitIteratorSize16(VKLSampler sampler);
Open VKL also provides the macro VKL_MAX_HIT_ITERATOR_SIZE as a conservative estimate.
Hits are then queried by looping a call to vklIterateHit as long as the returned lane mask indicates that the iterator is still within the volume.
int vklIterateHit(VKLHitIterator iterator, VKLHit *hit);
void vklIterateHit4(const int *valid,
VKLHitIterator4 iterator,
VKLHit4 *hit,
int *result);
void vklIterateHit8(const int *valid,
VKLHitIterator8 iterator,
VKLHit8 *hit,
int *result);
void vklIterateHit16(const int *valid,
VKLHitIterator16 iterator,
VKLHit16 *hit,
int *result);
Returned hits consist of a t-value, a volume value (equal to one of the requested values specified in the value selector), and an (object space) epsilon value estimating the error of the intersection:
typedef struct
{
float t;
float sample;
float epsilon;
} VKLHit;
typedef struct
{
float t[4];
float sample[4];
float epsilon[4];
} VKLHit4;
typedef struct
{
float t[8];
float sample[8];
float epsilon[8];
} VKLHit8;
typedef struct
{
float t[16];
float sample[16];
float epsilon[16];
} VKLHit16;
For both interval and hit iterators, only the vector-wide API for the native SIMD width (determined via vklGetNativeSIMDWidth can be called. The scalar versions are always valid. This restriction will likely be lifted in the future.
Performance Recommendations¶
MXCSR control and status register¶
It is strongly recommended to have the Flush to Zero and Denormals are Zero mode of the MXCSR control and status register enabled for each thread before calling the sampling, gradient, or interval API functions. Otherwise, under some circumstances special handling of denormalized floating point numbers can significantly reduce application and Open VKL performance. The driver parameter flushDenormals or environment variable OPENVKL_FLUSH_DENORMALS can be used to toggle this mode; by default it is enabled. Alternatively, when using Open VKL together with the Intel® Threading Building Blocks, it is sufficient to execute the following code at the beginning of the application main thread (before the creation of the tbb::task_scheduler_init object):
#include <xmmintrin.h>
#include <pmmintrin.h>
...
_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
If using a different tasking system, make sure each thread calling into Open VKL has the proper mode set.
Iterator Allocation¶
vklInitIntervalIterator and vklInitHitIterator expect a user allocated buffer. While this buffer can be allocated by any means, we expect iterators to be used in inner loops and advise against heap allocation in that case. Applications may provide high performance memory pools, but as a preferred alternative we recommend stack allocated buffers.
In C99, variable length arrays provide an easy way to achieve this:
const size_t bufferSize = vklGetIntervalIteratorSize(sampler);
char buffer[bufferSize];
Note that the call to vklGetIntervalIteratorSize or vklGetHitIteratorSize should not appear in an inner loop as it is relatively costly. The return value depends on the volume type, target architecture, and parameters to sampler.
In C++, variable length arrays are not part of the standard. Here, users may rely on alloca and similar functions:
#include <alloca.h>
const size_t bufferSize = vklGetIntervalIteratorSize(sampler);
char *buffer = alloca(bufferSize);
Users should understand the implications of alloca. In particular, alloca does check available stack space and may result in stack overflow. buffer also becomes invalid at the end of the scope. As one consequence, it cannot be returned from a function. On Windows, _malloca is a safer option that performs additional error checking, but requires the use of _freea.
In ISPC, variable length or alloca do not exist. Applications may instead rely on the VKL_MAX_INTERVAL_ITERATOR_SIZE and VKL_MAX_HIT_ITERATOR_SIZE macros:
uniform unsigned int8 buffer[VKL_MAX_INTERVAL_ITERATOR_SIZE];
These values are majorants over all drivers and volume types. Note that Open VKL attempts to detect the target SIMD width using TARGET_WIDTH, returning smaller buffer sizes for narrow architectures. However, Open VKL may fall back to the largest buffer size over all targets.
Multi-attribute Volume Data Layout¶
Open VKL provides flexible managed data APIs that allow applications to specify input data in various formats and layouts. When shared buffers are used (dataCreationFlags = VKL_DATA_SHARED_BUFFER), Open VKL will use the application-owned memory directly, respecting the input data layout. Shared buffers therefore allow applications to strategically select the best layout for multi-attribute volume data and expected sampling behavior.
For volume attributes that are sampled individually (e.g. using vklComputeSample[4,8,16,N]()), it is recommended to use a structure-of-arrays layout. That is, each attribute’s data should be compact in contiguous memory. This can be accomplished by simply using Open VKL owned data objects (dataCreationFlags = VKL_DATA_DEFAULT), or by using a natural byteStride for shared buffers.
For volume attributes that are sampled simultaneously (e.g. using vklComputeSampleM[4,8,16,N]()), it is recommended to use an array-of-structures layout. That is, data for these attributes should be provided per voxel in a contiguous layout. This is accomplished using shared buffers for each attribute with appropriate byte strides. For example, for a three attribute structured volume representing a velocity field, the data can be provided as:
// used in Open VKL shared buffers, so must not be freed by application
std::vector<vkl_vec3f> velocities(numVoxels);
for (auto &v : velocities) {
v.x = ...;
v.y = ...;
v.z = ...;
}
std::vector<VKLData> attributes;
attributes.push_back(vklNewData(velocities.size(),
VKL_FLOAT,
&velocities[0].x,
VKL_DATA_SHARED_BUFFER,
sizeof(vkl_vec3f)));
attributes.push_back(vklNewData(velocities.size(),
VKL_FLOAT,
&velocities[0].y,
VKL_DATA_SHARED_BUFFER,
sizeof(vkl_vec3f)));
attributes.push_back(vklNewData(velocities.size(),
VKL_FLOAT,
&velocities[0].z,
VKL_DATA_SHARED_BUFFER,
sizeof(vkl_vec3f)));
VKLData attributesData =
vklNewData(attributes.size(), VKL_DATA, attributes.data());
for (auto &attribute : attributes)
vklRelease(attribute);
VKLVolume volume = vklNewVolume("structuredRegular");
vklSetData(volume, "data", attributesData);
vklRelease(attributesData);
// set other volume parameters...
vklCommit(volume);
These are general recommendations for common scenarios; it is still recommended to evaluate performance of different volume data layouts for your application’s particular use case.