dopt.cuda

This is the main interface for the dopt CUDA backend.

The APIs in this module allow users to evaluate operation graphs on GPUs through the use of CUDA. There is also functionality to register CUDA implementations of custom operations.

In future, this module will also have an interface allowing the user to register their own optimisation passes to be called when constructing a plan.

Modules

basic module dopt.cuda.basic: Undocumented in source.
math module dopt.cuda.math: Undocumented in source.
nnet module dopt.cuda.nnet: Undocumented in source.
nvrtc module dopt.cuda.nvrtc: Undocumented in source.
random module dopt.cuda.random: Undocumented in source.

Members

Aliases

CUDAKernelCtr alias CUDAKernelCtr = CUDAKernel delegate(Operation op): Undocumented in source.

Classes

CUDABuffer class CUDABuffer: A class that encapsulates the CUDA memory allocation/deallocation process.
CUDAPlan class CUDAPlan: A Plan stores all the resources (preallocated buffers, custom CUDA kernels) required to evaluate nodes from the Operation graph.

Functions

cudaType string cudaType(DataType t): Undocumented in source. Be warned that the author may not have intended to support it.
deregisterCUDAKernel void deregisterCUDAKernel(string opType): Deregisters a kernel constructor associated with the given operation type.
evaluateCUDA DeviceBuffer[] evaluateCUDA(Operation[] ops, DeviceBuffer[Operation] args): Used for performing a one-off evaluation of a set of operations.
evaluateCUDA DeviceBuffer evaluateCUDA(Operation op, DeviceBuffer[Operation] args): A convenience overload that evaluates a single operation and returns a single DeviceBuffer.
listCUDAOperations string[] listCUDAOperations(): Provides a list of all operation types supported by the CUDA backend.
registerCUDAKernel void registerCUDAKernel(string opName, CUDAKernelCtr kernelCtr): Registers a CUDA kernel constructor for a given operation type.

Interfaces

CUDAKernel interface CUDAKernel: Provides a common interface for CUDA kernels.