gmatrix (version 0.3)

gmatrix-package: Making GPU power available in R

Description

This package implements a general framework for utilizing R to harness the power of NVIDIA GPU's. The gmatrix and gvector classes allow for easy management of the separate device and host memory spaces. Numerous numerical operations are implemented for these objects on the GPU. These operations include matrix multiplication, addition, subtraction, the kronecker product, the outer product, comparison operators, logical operators, trigonometric functions, indexing, sorting, random number generation and many more.

Details

Package:

gmatrix

Type:

Package

Version:

0.1

Date:

2013-5-10

License:

GPL-2

LazyLoad:

yes

Depends:

methods

To move an R object to the GPU use the g function. To move the object back the the host use the h function. Standard operations such as matrix multiplication or addition may be performed on GPU objects as with any matrix or vector object in R. The gmatrix and gvector classes contain a type slot which determines if the data is stored as a "double", "single", "integer" or "logical." The gmatrix and gvector objects store only a pointer to an object on the GPU device. Thus, a line such as y<-x will duplicate the pointer but not the actual data on the GPU. One should use instead y<-gdup(x). Numerous random number generators such as grnorm have been implemented in this package. This package is intended to work with the companion package rcula which implements matrix factorization and inversions methods.

References

None

See Also

rcula-package, g, h, gmatrix-class, gvector-class

Examples

#Look at available GPU devices (use setDevice to change the device)
listDevices()

#Creating GPU objects
g_x=g(1:10) #transfer a sequence to the GPU
g_y=gseq(1,10) #create a sequence on the GPU
g_A=gmatrix(grnorm(100),10,10,dup=FALSE) #create a random 10x10 matrix

#look at and change the properties of these objects
type(g_x)
type(g_A)
type(g_x)="d" #'d' may be used as short form of"double"
length(g_x)
dim(g_A)
nrow(g_A)

#Perform some simple calculations
g_ans=g_x %*% log(abs(g_A)) %*% sin(-g_y)
g_rowsum= rowSums(g_A)

#Move data back to the host
h_ans=h(g_x)
h_roswum =h(g_A)