OpenCL - Interface allowing R to use OpenCL
NEWS for the OpenCL package
0.2 (under development)
o OpenCL context and command queue can be persisted, allowing to keep data
between calls. The context also remembers whether to default to single-
or double-precision for numeric vectors.
o Data can stay on the OpenCL device (GPU) between kernel calls. This is
extremely valuable when working with discrete GPUs connected over a
relatively slow PCIe connection.
o A single-precision data type is no longer required. The conversion takes
place when transferring the data to the OpenCL device. On the R side,
data remains in numeric vectors.
o Kernels are executed asynchronously and possibly out-of-order, if the
OpenCL implementation allows it. Synchronization need not to be done
manually and happens without the user knowing: OpenCL events
corresponding to a kernel execution are attached to the output buffer.
Following kernel executions having the buffer as input then wait for the
event, hence for the preceding kernel execution to finish. Likewise,
reads from buffers wait on the attached event as well.
o OpenCL device information is amended by maximum frequency. Also the list
of extensions is broken down to make it easier searchable.
o By default, we choose GPU devices. CPU devices usually don't make a lot
of sense. Also, if there are multiple GPU devices available - think of
a notebook with integrated and discrete GPU - we try to choose the
o There are now several tests covering most of the functionality.
0.1-4 (under development)
o devices with very long extensions strings could cause error
on retrieval. Fixed with larger static buffer.
(Thanks to Valerio Aimale again)
o Improve error reporting by always including the OpenCL
o fix a bug causing device enumeration to use the default
device for device count regradless of the specified type.
This affects only systems with more than one type of device.
(Thanks to Valerio Aimale for reporting)
o added dim argument to oclRun() which allows multidimensional
indexing (up to 3d) in the kernel. The dimensions can be
obtained in the kernel via get_global_size() and the index
values with get_global_id(). Note that using index vectors
instead of multidimensional indexing may perform better
depending on the device. The default is to use single
dimension (dim=size) which is the same as previous versions of
o add precision="best" in oclSimpleKernel which switches
automatically to double-precision if supported by the device
o kernels objects are now less cryptic - they implement
print(), names() and $ methods for access to their attributes.
o add the support for asynchronous calls, i.e., execution
parallel to the CPU or multiple parallel GPU operations.
This is done by using x <- oclRun(..., wait=FALSE) to
dispatch the kernel and then oclResult(x) to collect the
o minor cleanup
o improve memory management and clean up on error in oclRun()
o use CL_MEM_USE_HOST_PTR instead of clEnqueueWriteBuffer() for
better performance on large input vectors
o add support for native single precision representation
(see ?clFloat and native.result argument in oclRun())
o added INSTALL file with links to common OpenCL implementations
o first public release
includes support for single and double precision computations
as well as simple kernels (one output vector, arbitrary input)