in cuda's driver_types.h have:
typedef __device_builtin__ struct custream_st *cudastream_t; and in cuda_runtime.h have, in many places, default-initialized stream parameters. example:
template<class t> static __inline__ __host__ cudaerror_t cudalaunchkernel( const t *func, dim3 griddim, dim3 blockdim, void **args, size_t sharedmem = 0, cudastream_t stream = 0 ) how safe assume default stream (cudastream) nullptr?
this documented case in multiple places:
kernel launches , host <-> device memory copies not specify stream parameter, or equivalently set stream parameter zero, issued default stream.
- for example,
cudamemcpyasync:
the copy can optionally associated stream passing non-zero stream argument.
it seems quite safe assume default stream equivalent (cudastream_t) 0
Comments
Post a Comment