Low Level API (NvBlast)

Table of Contents

The low-level API is the core of Blast™ destruction. It is designed to be a minimal API that allows an experienced user to incorporate destruction into their application. Summarizing what the low-level API has, and doesn't have:

There is no physics representation. The low-level API is agnostic with respect to any physics engine, and furthermore does not have any notion of collision geometry. The NvBlastActor is an abstraction which is intended to correspond to a rigid body. However it is up to the user to implement that connection. The NvBlastActor references a list of visible chunk indices, which correspond to NvBlastChunk data in the asset. The NvBlastChunk contains a userData field which can be used to associate collision geometry with the actor based upon the visible chunks. The same is true for constraints created between actors. Bonds contain a userData field that can be used to inform the user that actors should have joints created at a particular location, but it is up to the user to create and manage physical joints between two actors.
There is no graphics representation. Just as there is no notion of collision geometry, there is also no notion of graphics geometry. The NvBlastChunk userData field (see the item above) can be used to associate graphics geometry with the actor based upon the visible chunks.
There is no notion of threading. The API is a collection of free functions which the user may call from appropriate threads. Blast™ guarantees that it is safe to operate on different actors from different threads.
There is no global memory manager, message handler, etc. All low-level API functions take an optional message function pointer argument in order to report warnings or errors. Memory is managed by the user, and functions that build objects require an appropriately-sized memory block to be passed in. A corresponding utility function that calculates the memory requirements is always present alongside such functions. Temporary storage needed by a function is always handled via user-supplied scratch space. For scratch, there is always a corresponding "RequiredScratch" function or documentation which lets the user know how much scratch space is needed based upon the function arguments.
Backwards-compatible, versioned, device-independent serialization is not handled by Blast™. There is however a Blast™ extension which does, see Serialization (NvBlastExtSerialization). However, a simple form of serialization may be performed on assets and familes (see Definitions) via simple memory copy. The data associated with these objects is available to the user, and may be copied and stored by the user. Simply casting a pointer to such a block of memory to the correct object type will produce a usable object for Blast™. (The only restriction is that the block must be 16-byte aligned.) Families contain a number of actors and so this form of deserialization recreates all actors in the family. This form of serialization may be used between two devices which have the same endianness, and contain Blast™ SDKs which use the same object format.
Single-actor serialization and deserialization is, however, supported. This is not as light-weight as family serialization, but may be a better serialization model for a particular application. To deserialize a single actor, one must have a family to hold the actor, created from the appropriate asset. If none exists already, the user may create an empty family. After that, all actors that had been in that family may be deserialized into it one-at-a-time, in any order.

Linking and Header Files

To use the low-level Blast™ SDK, the application need only inlclude the header NvBlast.h, found in the top-level include folder, and link against the appropriate version of the NvBlast library. Depending on the platform and configuration, various suffixes will be added to the library name. The general naming scheme is

NvBlast(config)(arch).(ext)

(config) is DEBUG, CHECKED, OR PROFILE for the corresponding configurations. For a release configuration there is no (config) suffix.

(arch) is _x86 or _x64 for Windows 32- and 64-bit builds, respectively, and empty for non-Windows platforms.

(ext) is .lib for static linking and .dll for dynamic linking on Windows. On XBoxOne it is .lib, and on PS4 it is .a.

Creating an Asset from a Descriptor (Authoring)

The NvBlastAsset is an opaque type pointing to an object constructed by Blast™ in memory allocated by the user. To create an asset from a descriptor, use the function NvBlastCreateAsset. See the function documentation for a description of its parameters.

N.B., there are strict rules for the ordering of chunks with an asset, and also conditions on the chunks marked as "support" (using the NvBlastChunkDesc::SupportFlag). See the function documentation for these conditions. NvBlastCreateAsset does not reorder chunks or modify support flags to meet these conditions. If the conditions are not met, NvBlastCreateAsset fails and returns NULL. However, Blast™ provides helper functions to reorder chunk descriptors and modify the support flags within those descriptors so that they are valid for asset creation. The helper functions return a mapping from the original chunk ordering to the new chunk ordering, so that corresponding adjustments or mappings may be made for graphics and other data the user associates with chunks.

Example code is given below. Throughout, we assume the user has defined a logging function called logFn, with the signature of NvBlastLog. In all cases, the log function is optional, and NULL may be passed in its place.

// Create chunk descriptors
std::vector<NvBlastChunkDesc> chunkDescs;
chunkDescs.resize( chunkCount );    // chunkCount > 0

chunkDescs[0].parentChunkIndex = UINT32_MAX;    // invalid index denotes a chunk hierarchy root
chunkDescs[0].centroid[0] = 0.0f;   // centroid position in asset-local space
chunkDescs[0].centroid[1] = 0.0f;
chunkDescs[0].centroid[2] = 0.0f;
chunkDescs[0].volume = 1.0f;    // Unit volume
chunkDescs[0].flags = NvBlastChunkDesc::NoFlags;
chunkDescs[0].userData = 0; // User-supplied ID.  For example, this can be the index of the chunkDesc.
                            // The userData can be left undefined.

chunkDescs[1].parentChunkIndex = 0; // child of chunk described by chunkDescs[0]
chunkDescs[1].centroid[0] = 2.0f;   // centroid position in asset-local space
chunkDescs[1].centroid[1] = 4.0f;
chunkDescs[1].centroid[2] = 6.0f;
chunkDescs[1].volume = 1.0f;    // Unit volume
chunkDescs[1].flags = NvBlastChunkDesc::SupportFlag; // This chunk should be represented in the support graph
chunkDescs[1].userData = 1;

// ... etc. for all chunks

// Create bond descriptors
std::vector<NvBlastBondDesc> bondDescs;
bondDescs.resize( bondCount );  // bondCount > 0

bondDescs[0].chunkIndices[0] = 1;   // chunkIndices refer to chunk descriptor indices for support chunks
bondDescs[0].chunkIndices[1] = 2;
bondDescs[0].bond.normal[0] = 1.0f; // normal in the +x direction
bondDescs[0].bond.normal[1] = 0.0f;
bondDescs[0].bond.normal[2] = 0.0f;
bondDescs[0].bond.area = 1.0f;  // unit area
bondDescs[0].bond.centroid[0] = 1.0f;   // centroid position in asset-local space
bondDescs[0].bond.centroid[1] = 2.0f;
bondDescs[0].bond.centroid[2] = 3.0f;
bondDescs[0].userData = 0;  // this can be used to tell the user more information about this
                                // bond for example to create a joint when this bond breaks

bondDescs[1].chunkIndices[0] = 1;
bondDescs[1].chunkIndices[1] = ~0;  // ~0 (UINT32_MAX) is the "invalid index."  This creates a world bond
// ... etc. for bondDescs[1], all other fields are filled in as usual

// ... etc. for all bonds

// Set the fields of the descriptor
NvBlastAssetDesc assetDesc;
assetDesc.chunkCount = chunkCount;
assetDesc.chunkDescs = chunkDescs.data();
assetDesc.bondCount = bondCount;
assetDesc.bondDescs = bondDescs.data();

// Now ensure the support coverage in the chunk descriptors is exact, and the chunks are correctly ordered
std::vector<char> scratch( chunkCount * sizeof(NvBlastChunkDesc) ); // This is enough scratch for both NvBlastEnsureAssetExactSupportCoverage and NvBlastReorderAssetDescChunks
NvBlastEnsureAssetExactSupportCoverage( chunkDescs.data(), chunkCount, scratch.data(), logFn );
std::vector<uint32_t> map(chunkCount);  // Will be filled with a map from the original chunk descriptor order to the new one
NvBlastReorderAssetDescChunks( chunkDescs.data(), chunkCount, bondDescs.data(), bondCount, map, true, scratch.data(), logFn );

// Create the asset
scratch.resize( NvBlastGetRequiredScratchForCreateAsset( &assetDesc ) );    // Provide scratch memory for asset creation
void* mem = malloc( NvBlastGetAssetMemorySize( &assetDesc ) );      // Allocate memory for the asset object
NvBlastAsset* asset = NvBlastCreateAsset( mem, &assetDesc, scratch.data(), logFn );

It should be noted that the geometric information (centroid, volume, area, normal) in chunks and bonds is only used by damage shader functions (see Damage Shaders (NvBlastExtShaders)). Depending on the shader, some, all, or none of the geometric information will be needed. The user may write damage shader functions that interpret this data in any way they wish.

Cloning an Asset

To clone an asset, one only needs to copy the memory associated with the NvBlastAsset.

uint32_t assetSize = NvBlastAssetGetSize( asset );

NvBlastAsset* newAsset = (NvBlastAsset*)malloc(assetSize);  // NOTE: the memory buffer MUST be 16-byte aligned!
memcpy( newAsset, asset, assetSize );   // this data may be copied into a buffer, stored to a file, etc.

N.B. the comment after the malloc call above. NvBlastAsset memory must be 16-byte aligned.

Releasing an Asset

Blast™ low-level does no internal allocation; since the memory is allocated by the user, one simply has to free the memory they've allocated. The asset pointer returned by NvBlastCreateAsset has the same numerical value as the mem block passed in (if the function is successful, or NULL otherwise). So releasing an asset with memory allocated by malloc is simply done with a call to free:

free( asset );

Creating Actors and Families

Actors live within a family created from asset data. To create an actor, one must first create a family. This family is used by the initial actor created from the asset, as well as all of the descendant actors created by recursively fracturing the initial actor. As with assets, family allocation is done by the user.

To create a family, use:

// Allocate memory for the family object - this depends on the asset being represented by the family.
void* mem = malloc( NvBlastAssetGetFamilyMemorySize( asset, logFn ) );

NvBlastFamily* family = NvBlastAssetCreateFamily( mem, asset, logFn );

When an actor is first created from an asset, it represents the root of the chunk hierarchy, that is the unfractured object. To create this actor, use:

// Set the fields of the descriptor
NvBlastActorDesc actorDesc;
actorDesc.asset = asset;    // point to a valid asset
actorDesc.initialBondHealth = 1.0f;  // this health value will be given to all bonds
actorDesc.initialChunkHealth = 1.0f; // this health value will be given to all lower-support chunks

// Provide scratch memory
std::vector<char> scratch( NvBlastFamilyGetRequiredScratchForCreateFirstActor( &actorDesc ) );

// Create the first actor
NvBlastActor* actor = NvBlastFamilyCreateFirstActor( family, &actorDesc, scratch.data(), logFn );   // ready to be associated with physics and graphics by the user

Copying Actors (Serialization and Deserialization)

There are two ways to copy NvBlastActors: cloning an NvBlastFamily, and single-actor serialization. Cloning an NvBlastFamily is extremely fast as it only requires a single memory copy. All actors in the family may be saved, loaded, or copied at once in this way.

Cloning a Family

To clone a family, use the family pointer which may be retrieved from any active actor in the family if needed, using the NvBlastActorGetFamily function:

const NvBlastFamily* family = NvBlastActorGetFamily( &actor, logFn );

Then the size of the family may be obtained using:

size_t size = NvBlastFamilyGetSize( family, logFn );

Now this memory may be copied, saved to disk, etc. To clone the family, for example, we can duplicate the memory:

std::vector<char> buffer( size );
NvBlastFamily* family2 = reinterpret_cast<NvBlastFamily*>( buffer.data() );
memcpy( family2, family, size );

N.B. If this data has been serialized from an external source, the family will not contain a valid reference to its associated asset. The user must set the family's asset. The family does however contain the asset's ID, to help the user match the correct asset to the family. So one way of restoring the asset to the family follows:

const NvBlastGUID guid = NvBlastFamilyGetAssetID( family2, logFn );
// ... here the user must retrieve the asset using the GUID or by some other means
NvBlastFamilySetAsset( family2, asset, logFn );

The data in family2 will contain the same actors as the original family. To access them, use:

uint32_t actorCount = NvBlastFamilyGetActorCount( family2, logFn );
std::vector<NvBlastActor*> actors( actorCount );
uint32_t actorsWritten = NvBlastFamilyGetActors( actors.data(), actorCount, family2, logFn );

In the code above, actorsWritten should equal actorCount.

Single Actor Serialization

To perform single-actor serialization, first find the buffer size required to store the serialization data:

size_t bufferSize = NvBlastActorGetSerializationSize( actor, logFn );

If you want to use an upper bound which will be large enough for any actor in a family, you may use:

size_t bufferSize = NvBlastAssetGetActorSerializationSizeUpperBound( asset, logFn );

Then create a buffer of that size and use NvBlastActorSerialize to write to the buffer:

std::vector<char> buffer( bufferSize );
size_t bytesWritten = NvBlastActorSerialize( buffer, bufferSize, actor, logFn );

To deserialize the buffer, an appropriate family must be created. It must not already hold a copy of the actor. It must be formed using the correct asset (the one that originally created the actor):

void* mem = malloc( NvBlastAssetGetFamilyMemorySize( asset, logFn ) );
NvBlastFamily* family = NvBlastAssetCreateFamily( mem, asset, logFn );

Then deserialize into the family:

NvBlastActor* newActor = NvBlastFamilyDeserializeActor( family, buffer.data(), logFn );

If newActor is not NULL, then the actor was successfully deserialized.

Deactivating an Actor

Actors may not be released in the usual sense of deallocation. This is because actors' memory is stored as a block within the owning family. The memory is only released when the family is released. However, one may deactivate an actor using NvBlastActorDeactivate. This clears the actor's chunk lists and marks it as invalid, effectively disassociating it from the family. The user should consider this actor to be destroyed.

bool success = NvBlastActorDeactivate( actor, logFn );

Releasing a family

As mentioned above, releasing an actor does not actually do any deallocation; it simply invalidates the actor within its family. To actually deallocate memory, you must deallocate the family. Note, this will invalidate all actors in the family. This is a fast way to delete all actors that were created from repeated fracturing of a single instance. As with NvBlastAsset, memory is allocated by the user, so to release a family with memory allocated by malloc, simply free that memory with free:

free( family );

The family will not be automatically released when all actors within it are invalidated using NvBlastActorDeactivate. However, the user may query the number of active actors in a family using

uint32_t actorCount = NvBlastFamilyGetActorCount( family, logFn );

Damage and Fracturing

Damaging and fracturing is a staged process. In a first step, a NvBlastDamageProgram creates lists of Bonds and Chunks to damage - so called Fracture Commands. The lists are created from input specific to the NvBlastDamageProgram.
NvBlastDamagePrograms are composed of a NvBlastGraphShaderFunction and a NvBlastSubgraphShaderFunction operating on support graphs (support chunks and bonds) and disconnected subsupport chunks respectively. An implementer can freely define the shader functions and parameters. Different functions can have the effect of emulating different physical materials.
Blast™ provides reference implementations of such functions in Damage Shaders (NvBlastExtShaders), see also NvBlastExtDamageShaders.h. The NvBlastDamageProgram is used through NvBlastActorGenerateFracture that will provide the necessary internal data for the NvBlastActor being processed. The shader functions see the internal data as NvBlastGraphShaderActor and NvBlastSubgraphShaderActor respectively.

The second stage is carried out with NvBlastActorApplyFracture. This function takes the previously generated Fracture Commands and applies them to the NvBlastActor. The result of every applied command is reported as a respective Fracture Event if requested.

Fracture Commands and Fracture Events both are represented by NvBlastFractureBuffers. The splitting of the actor into child actors is not done until the third stage, NvBlastActorSplit, is called. Fractures may be repeatedly applied to an actor before splitting.

The NvBlastActorGenerateFracture, NvBlastActorApplyFracture and NvBlastActorSplit functions are profiled in Profile configurations. This is done through a pointer to a NvBlastTimers struct passed into the functions. If this pointer is not NULL, then timing values will be accumulated in the referenced struct.

The following example illustrates the process:

// Step one: Generate Fracture Commands

// Damage programs (shader functions), material properties and damage description relate to each other.
// Together they define how actors will break by generating the desired set of Fracture Commands for Bonds and Chunks.
NvBlastDamageProgram damageProgram = { GraphShader, SubgraphShader };
NvBlastProgramParams programParams = { damageDescs, damageDescCount, materialProperties };

// Generating the set of Fracture Commands does not modify the NvBlastActor.
NvBlastActorGenerateFracture( fractureCommands, actor, damageProgram, &programParams, logFn, &timers );


// Step two: Apply Fracture Commands

// Applying Fracture Commands does modify the state of the NvBlastActor.
// The Fracture Events report the resulting state of each Bond or Chunk involved.
// Chunks fractured hard enough will also fracture their children, creating Fracture Events for each.
NvBlastActorApplyFracture( fractureEvents, actor, fractureCommands, logFn, &timers );


// Step three: Splitting

// The Actor may be split into all its smallest pieces.
uint32_t maxNewActorCount = NvBlastActorGetMaxActorCountForSplit( actor, logFn );
std::vector<NvBlastActor*> newActors( maxNewActorCount );

// Make this memory available to NvBlastSplitEvent.
NvBlastActorSplitEvent splitEvent;
splitEvent.newActors = newActors.data();

// Some temporary memory is necessary as well.
std::vector<char> scratch( NvBlastActorGetRequiredScratchForSplit( actor, logFn ) );

// New actors created are reported in splitEvent.newActors.
// If newActorCount != 0, then the old actor is deleted and is reported in splitEvent.deletedActor.
size_t newActorCount = NvBlastActorSplit( &splitEvent, actor, maxNewActorCount, scratch.data(), logFn, &timers );