OpenCL (version 0.2-10)

oclRun: Run a kernel using OpenCL.

Description

oclRun is used to execute code that has been compiled for OpenCL.

Usage

oclRun(kernel, size, ..., dim = size)

Value

The resulting buffer of length size.

Arguments

kernel

Kernel object as obtained from oclSimpleKernel

size

Length of the output vector

...

Additional arguments passed to the kernel

dim

Numeric vector describing the global work dimensions, i.e., the index range that the kernel will be run on. The kernel can use get_global_id(n) to obtain the (n + 1)-th dimension index and get_global_size(n) to get the dimension. OpenCL standard supports only up to three dimensions, you can use use index vectors as arguments if more dimensions are required. Note that dim is not necessarily the dimension of the result although it can be.

Author

Simon Urbanek, Aaron Puchert

Details

oclRun pushes kernel arguments, executes the kernel and retrieves the result. The kernel is expected to have either __global double * or __global float * type (write-only) as the first argument which will be used for the result and const unsigned int second argument denoting the result length. All other arguments are assumed to be read-only and will be filled according to the ... values. These can either be OpenCL buffers as generated by clBuffer for pointer arguments, or scalar values (vectors of length one) for scalar arguments. Only integer (int), and numeric (double or float) scalars and OpenCL buffers are supported as kernel arguments. The caller is responsible for matching the argument types according to the kernel in a way similar to .C and .Call.

Note that the kernel must match the input types as well, so typically as.clBuffer() should include the mode (e.g., "numeric") to match the kernel and/or explicit as.numeric() coercion should be used.

See Also

oclSimpleKernel, clBuffer

Examples

library(OpenCL)
## Only proceed if this machine has at least one OpenCL platform
if (length(oclPlatforms())) {
ctx = oclContext(precision="single")

code = c("
__kernel void dnorm(
  __global numeric* output,
 const unsigned int count,
  __global numeric* input,
 const numeric mu, const numeric sigma)
{
  size_t i = get_global_id(0);
  if(i < count)
      output[i] = exp((numeric) (-0.5 * ((input[i] - mu) / sigma) * ((input[i] - mu) / sigma)))
      / (sigma * sqrt((numeric) (2 * 3.14159265358979323846264338327950288 )) );
}")
k.dnorm <- oclSimpleKernel(ctx, "dnorm", code)
f <- function(x, mu=0, sigma=1)
  as.numeric(oclRun(k.dnorm, length(x), as.clBuffer(x, ctx, "numeric"), mu, sigma))

## expect differences since the above uses single-precision but
## it should be close enough
f(1:10/2) - dnorm(1:10/2)

## does the device support double-precision?
if (any("cl_khr_fp64" == oclInfo(attributes(ctx)$device)$exts)) {
  k.dnorm <- oclSimpleKernel(ctx, "dnorm", code, "double")
  f <- function(x, mu=0, sigma=1)
    as.numeric(oclRun(k.dnorm, length(x), as.clBuffer(x, ctx, "double"), mu, sigma))

  ## probably not identical, but close...
  f(1:10/2) - dnorm(1:10/2)
} else cat("\nSorry, your device doesn't support double-precision\n")

## Note that in practice you can use precision="best" in the first
## example which will pick "double" on devices that support it and
## "single" elsewhere
}