CUDAPlan

A Plan stores all the resources (preallocated buffers, custom CUDA kernels) required to evaluate nodes from the Operation graph.

An instance of Plan can be constructed using the compileCUDA function. The primary use case for a CUDAPlan is when the same set of operations are likely to be evaluated more than once. This prevents the dopt CUDA runtime from reallocating and optimising the CUDA kernels every time the same set of operations is to be executed.

class CUDAPlan : Plan {

long[string] profiler;

this(Operation[] outputs);

~this();

void cleanup();

void executeImpl(DeviceBuffer[Operation] args, DeviceBuffer[] rets);

}

Constructors

this this(Operation[] outputs): Undocumented in source.

Destructor

~this ~this(): Undocumented in source.

Members

Functions

cleanup void cleanup(): Releases CUDA resources associated with this plan.
executeImpl void executeImpl(DeviceBuffer[Operation] args, DeviceBuffer[] rets): Undocumented in source. Be warned that the author may not have intended to support it.

Variables

profiler long[string] profiler;: Undocumented in source.

Inherited Members

From Plan

execute DeviceBuffer[] execute(DeviceBuffer[Operation] args): Executes the plan.
execute void execute(DeviceBuffer[Operation] args, DeviceBuffer[] rets)
mOutputs Operation[] mOutputs;: Undocumented in source.
executeImpl void executeImpl(DeviceBuffer[Operation] args, DeviceBuffer[] rets): Undocumented in source.