cl.enqueue.kernel.unsafe: Execute an OpenCL kernel with less overhead
In caffeine-overload/rocl: Low Level OpenCL binding for R

Just like cl.enqueue.kernel, but without expensive calls to stopifnot.

1
2
3

cl.enqueue.kernel.unsafe(queue, kernel, dimensions, global_work_size,
  arguments, local_work_size = NULL, global_work_offset,
  waitlist = NULL, return_event = NULL)

`queue`	An opencl_command_queue object. The queue the operation will be put on.
`kernel`	The kernel to execute. An opencl_kernel object
`dimensions`	The dimensions of the kernel
`global_work_size`	The number of global threads in each dimension
`arguments`	A list of kernel arguments. Must be all opencl_buffer objects. Any names are ignored, the arguments are passed by position.
`local_work_size`	The number of local threads in each dimension inside a work-group. If left NULL, the OpenCL runtime will pick a value.
`global_work_offset`	The starting offset of the global thread ids in each dimension.
`waitlist`	OpenCL events that must complete before this operation can start. Not currently used, included for future use.
`return_event`	Schedule the kernel asynchronously and return an opencl event object. Not currently used, included for future use.

cl.enqueue.kernel makes calls to stopifnot to ensure the types of arguments. This introduces significant overhead, so this function is without any type-checking, at the risk of crashing R if the wrong types are given. Note that global_work_offset is mandatory here.

caffeine-overload/rocl documentation built on May 6, 2019, 7:02 p.m.