Deep.Net


ICudaUOp

Namespace: SymTensor.Compiler.Cuda
Parent Module: CudaExecUnitTypes

A unified op that can be compiled to CUDA execution items.

Instance members

Instance memberDescription
ExecItems(arg1 arg2 arg3)
Signature: CudaCompileEnvT -> ExecItemsForOpArgs -> ExecItemsHelpersT -> CudaExecItemT list
Modifiers: abstract

Returns the execution items for the op. It must read from the given source manikin and write to the target manikin. Additional memory may be allocated for temporary results.

SrcReqs(arg1 arg2 arg3)
Signature: CudaCompileEnvT -> SrcReqsArgs -> SrcReqsHelpersT -> ChannelReqsT list
Modifiers: abstract

Computes desired source manikins given desired target manikin. There is no guarantee that the desired source manikins will be used. Also, it is not obligatory to use the requested target manikin.

TrgtGivenSrcs(arg1 arg2 arg3)
Signature: CudaCompileEnvT -> TrgtGivenSrcsArgs -> TrgtGivenSrcsHelpersT -> ChannelManikinsAndSharedT
Modifiers: abstract

Computes the definitive target manikin of an op given its source manikins. It is not obligatory to use the requested target manikin.

Fork me on GitHub