RFoptions: Setting control arguments

In RandomFieldsUtils: Utilities for the Simulation and Analysis of Random Fields and Genetic Data

Setting control arguments

Description

RFoptions sets and returns control arguments for the analysis and the simulation of random fields

Usage

 RFoptions(..., no.class=FALSE, install.control=NULL)

Arguments

...	arguments in tag = value form, or a list of tagged values. See ‘Details’ for options in package RandomFieldsUtils.
no.class	logical. If TRUE the list is returned without class specification.
install.control	list. See Details, Part 2

.

Details

The subsections below comment on
1. basic: Basic options
2. install.control
3. installNrun: Options for installation and running
4. solve: Options for solving linear systems
5. Reserved words

1. Basic options

asList

logical. Lists of arguments are treated slightly different from non-lists. If asList=FALSE they are treated the same way as non-lists. This options being set to FALSE after calling RFoptions it should be set as first element of a list.

Default: TRUE

cores

Number of cores for multicore algorithms; currently only used for the Cholesky decomposition.

Default : 1 if the package has been compiled with standard flags of CRAN and 0.75 * cores() + 0.25 * cpus() else.

Note that cores has not effect if set locally in this package or in package miraculix.

cPrintlevel

cPrintlevel is automatically set to printlevel when printlevel is changed. Standard users will never use a value higher than 3.

0 : no messages
1 : messages and warnings when the user's input looks odd
2 : messages (and internal errors) documenting the choice of the simulation method
3 : further user relevant informations
4 : information on recursive function calls
5 : function flow information of central functions
6 : errors that are internally treated
7 : details on building up the covariance structure
8 : details on taking the square root of the covariance matrix
9 : details on intermediate calculations
10 : further details on intermediate calculations

Note that printlevel works on the R level whereas cPrintlevel works on the C level. cPrintlevel should be changed only globally.

Default: 1

efficient

logical. If TRUE then always the most time efficient code is used.

Default: TRUE. It is strongly recommended to retain this value.

helpinfo

logical. If TRUE then additional information is printed for more efficient programming in R.

Default: TRUE

printlevel

If printlevel<=0 there is not any output on the screen. The higher the number the more tracing information is given. Standard users will never use a value higher than 3.

0 : no messages
1 : important (error) messages and warnings
2 : less important messages
3 : details, but still for the user
4 : recursive call tracing
5 : function flow information of large functions
6 : errors that are internally treated
7 : details on intermediate calculations
8 : further details on intermediate calculations

Default: 1

seed

integer (currently only used by the package RandomFields). If NULL or NA set.seed is not called. Otherwise, set.seed(seed) is set before any simulations are performed.

If the argument is set locally, i.e., within a function, it has the usual local effect. If it is set globally, i.e. by RFoptions the seed is fixed for all subsequent calls.

If the number of simulations n is greater than one and if RFoptions(seed=seed) is set, the ith simulation is started with the seed ‘seed+i-1’.

skipchecks

logical. If TRUE, several checks whether the given parameter values and the dimension are within the allowed range is skipped. Do not change the value of this variable except you really know what you do.

Default: FALSE

verbose

logical. If FALSE it identical to printlevel = 1 else to printlevel = 2.

bigendian

logical. Read only.

2. install.control: Details on argument install.control install.control may contain any argument of install.packages except type. This options is currently tailored for MS and Linux on Intel machines, only. The argument configure.args may not contain 'CXX_FLAGS' which should be passed as an extra argument with the list.

Note that if this argument is given (even with value NULL), an immediate installation takes place. In case the user tries to force to install 0 packages, an overview over the packages is given. If the user is asked whether re-installation shall take place, user can pass arguments to install.packages, e.g., "quiet=FALSE".

If install.control is given, no further argument may be passed to RFoptions.

Additional components of install.control and special behaviours are:

path

the path to the locally saved tar balls

verbose, quiet

They affect also the behaviour of RFoptions.

force

TRUE

reinstallation of all attached libraries based on and including RandomFieldsUtils. I.e., RFoptions(install.control=list(force=TRUE)) is the strongest form of forcing reinstallation.

FALSE

In case some packages have to be re-installed the user will be asked.

not given

reinstallation of the attached libraries based on and including RandomFieldsUtils that have not been tried yet in the current session.

pkgs=NULL

brief overview over the installed packages based on RandomFieldsUtils

CROSS

logical or character. CROSS is passed to ‘configure.ac’.

"noflag"

No extra compiler flag is set with respect to SIMD. This is the default.

TRUE

each file is compiled with its specific SIMD/AVX compiler flags; this guarantees the compatiblity on a plattform with different sets of kernels. No SIND/AVX flag should be given by the user. Cross-compilation supported; no check is performed whether the code would run on the compiling CPU.

"nosimd"

It is assumed that no SIMD is available and the flag "-no-sse2" is set (if possible).

"sse2"

Same behaviour as TRUE, but all CPUs have at least "sse2" available.

"sse3", "ssse3", "sse41", "avx", "avx2"

Alternatives to "sse2". Giving the highest guaranteed SIMD recognition leads to the most efficient code.

FALSE

each file is compiled with all SIMD/AVX flags recognized by both the CPU and the compiler (no cross-compilation); users may add their own SIMD/AVX flags. This might lead to faster code, but which is not downwards compatible.

NA

Same as FALSE except that the flag -mno-sse2 is set when no SIMD is needed or used.

This option can be set to "avx" interactively if install="ask".

CC_FLAGS

character. Flags passed to ‘configure.ac’.

SIMD_FLAGS

character. A subset of "sse2", "sse3", "ssse3", "sse41", "avx", "avx2", "arch=xxx", etc. which will be tried instead of default flags. SIMD_FLAGS is passed to ‘configure.ac’.

LOCAL_ONLY

logical. If TRUE, the web is not searched for the latest version of the package.

MEM_IS_ALIGNED

logical. If TRUE, then the memory is assumed to be aligned. If FALSE then the SIMD load commands _mm_*load_* are replaced by _mm_*loadu_*. If given, then force is set to TRUE.

USE_GPU

logical. Force or hinder the compilation for the GPU

3. installNrun: Options for installing and for determining basic behaviour

install

character. Only used by linux systems and alike including macOS The default by CRAN is that SIMD/AVX cannot be used at full extend. install determines what the action if the compiled version does not use the full CPU capacities. Since the use of GPU is heavily hardware dependent, its auto-recompilation is only performed in tow line of an AVX re-compilation. The users usually use

"no"

no re-installation

"ask"

asks whether the library should be reinstalled, using the full capacity of the CPU according to the package.

"install"

performs the auto-recompilation without asking. Note that only the command RFoptions(install.control=list(force=TRUE)) forces re-compilation of the currently loaded packages that are based on RandomFieldsUtils.

Note that, in each session, a package can be reinstalled only. This feature avoids trying to run jobs that cannot be done (e.g.\ due to missing programs of the OS). See argument install.control for overwriting this behaviour.

Default: at starting point it is "ask" or "no", but the value may change during the session.

installPackages

logical. Read only. Indicates whether packages are left to be re-installed. RFoptions(install="no") sets it to FALSE. RFoptions(install="no", install="ask") sets it to TRUE.

kahanCorrection

obsolete. logical. If TRUE, the Kahan summation algorithm is used for calculating scalar products.

Default: false

la_mode

determines

LA_AUTO, "auto"

If a graphic card is recognized, LA_GPU is used. In all other cases the default is primarily LA_R. Only on linux systems, the package peforms a simple speed test and takes LA_INTERN if it is faster than LA_R; the time, hence the choice, depends also on the number of cores used.

LA_INTERN, "intern"

mostly own algorithms, often based on SIMD/AVX. This option is of interest only if no advanced BLAS/LAPACK has been compiled into R

LA_R, "R"

BLAS/LAPACK implementation used by R

LA_GPU, "GPU"

This option is available when the package has been compiled with nvcc.

LA_QUERY, "query"

Request on currently used set-up

Default: LA_AUTO

mem_is_alignedlogical. Read only. See MEM_IS_ALIGNED in install.control.

warn_parallel

Logical. RandomFieldsUtils and packages using it, such as RandomFields and miraculix, should now be prepared for parallelization using package parallel, for instance. Internal OMP parallelization of RandomFieldsUtils is done, but only at a view points of the subsequent packages.

As a few parts cannot be in parallel technically or from a logical point of view, a hint or a warning is given, if such a point is not accessed adequately. These messages can be turned off by warn_parallel = FALSE.

Default: TRUE.

warn_unknown_option

integer.

0,1,-1

Unknown options are all ignored. If the value is positive, a hint is delivered whenever an unknown option is ignored.

-2,2

Unknown options that start with a capital letter are ignored. All others lead to an error. (Note that all RFoptions start with a minor letter.) If the value is positive, a hint is delivered whenever an unknown option is ignored.

3,-3

Unknown options that consists of a single capital letter are ignored. All others lead to an error. (Note that all RFoptions start with a minor letter.) If the value is positive, a hint is delivered whenever an unknown option is ignored.

4

(and other values) Any unknown option leads to an error.

Default for RandomFieldsUtils: 3

Default for RandomFields: 1

4. solve: Options for solving linear systems

det_as_log

eigen2zero

When the svd or eigen decomposition is calculated, all values with modulus less than or equal to eigen2zero are set to zero.

Default: 1e-12

max_chol

integer. Maximum number of rows of a matrix in a Cholesky decomposition

Default: 16384

max_svd

integer. Maximum number of rows of a matrix in a svd decomposition

Default: 10000

pivot_partialdet

logical. If TRUE then in case of low-rank matrices the determinant is calculated only in the part with positive eigenvalues

pivot

Type of pivoting for the Cholesky decomposition. Possible values are

PIVOT_NONE, "no"

No pivoting.

PIVOT_AUTO, "auto"

If the matrix has a size greater than 3x3 and Choleskey fails without pivoting, privoting is done. For matrices of size less than 4x4, no pivoting and no checks are performed. See also PIVOT_DO

PIVOT_DO, "do"

Do always pivoting. NOTE: privoted Cholesky decomposition yields only very approximately an upper triangular matrix L, but still L^t L = M holds true.

PIVOT_IDX, "idx"

uses the same pivoting as in the previous pivoted decomposition. This option becomes relevant only when simulations with different parameters or different models shall be performed with the same seed so that also the pivoting must be coupled.

Default: PIVOT_NONE

pivot_actual_size

integer. Genuine dimension of the linear mapping given by a matrix in cholx. This is a very rarely used option when pivoting with pivot=PIVOT_IDX.

pivot_check

logical. Only used in pivoted Cholesky decomposition. If TRUE and a numerically zero diagonal element is detected, it is checked whether the offdiagonal elements are numerically zero as well. (See also pivot_max_deviation and pivot_max_reldeviation.) If NA then only a warning is given.

Default: TRUE

pivot_idx

vector of integer. Sequence of pivoting indices in pivoted Cholesky decomposition. Note that pivot_idx[1] gives the number of indices that will be used. The vector must have at least the length pivot_idx[1] + 1.

Default: NULL

pivot_relerror

positive number. Tolerance for (numerically) negative eigenvalues and for (numerically) overdetermined systems appearing in the pivoted Cholesky decomposition.

Default: 1e-11

pivot_max_deviation

positive number. Together with pivot_max_reldeviation it determines when the rest of the matrix (eigenvalues) in the pivoted Cholesky decomposition are considered as zero.

Default: 1e-10

pseudoinverse

logical. In case of a singular matrix M, shall the pseudo inverse be returned for solvex(M)?

Default: FALSE

pivot_max_reldeviation

positive number. Together with pivot_max_deviation it determines when the rest of the matrix (eigenvalues) in the pivoted Cholesky decomposition are considered as zero.

Default: 1e-10

solve_method

vector of at most 3 integers that gives the sequence of methods in order to inverse a matrix or to calculate its square root: "cholesky", "svd", "lu", "eigen" "sparse", "method undefined". In the latter case, the algorithm decides which method might suit best.

Note that if use_spam is not false the algorithm checks whether a sparse matrix algorithm should be used and which is then tried first.

Default: "method undefined".

spam_factor

integer. See argument spam_sample_n.

Default: 4294967

spam_min_n

integer vector of size 2. The minimal size for a matrix to apply a sparse matrix algorithms automatically. The second value is used in case the GPU is activated.

Default: c(400, 4000)

spam_min_p (spam_min_p)

a numbers in (0,1) giving the proportion of zero above which an sparse matrix algorithm is used. The second value is used in case the GPU is activated.

Default: 0.8 (0.9)

spam_pivot

integer. Pivoting algorithm for sparse matrices:

PIVOT_NONE

No pivoting

PIVOTSPARSE_MMD

PIVOTSPARSE_RCM

See package spam for details.

Default: PIVOTSPARSE_MMD

spam_sample_n (spam_sample_n_GPU)

Whether a matrix is sparse or not is tested by a ‘random’ sample of size spam_sample_n; The selection of the sample is iteratively obtained by multiplying the index by spam_factor modulo the size of the matrix.

Default: 500 (10000).

spam_tol

largest absolute value being considered as zero. Default: DBL_EPSILON

svdtol

Internal. When the svd decomposition is used for calculating the square root of a matrix then the absolute componentwise difference between this matrix and the square of the square root must be less than svdtol. No check is performed if svdtol is not positive.

Default: 0

use_spam

Should the package spam (sparse matrices) be used for matrix calculations? If TRUE spam is always used. If FALSE, it is never used. If NA its use is determined by the size and the sparsity of the matrix.

Default: NA.

5. Reserved Words

list_

list_ usually equals the output of RFoptions(). This argument is used to reset the RFoptions. Some of the options behave differently if passed through list_. E.g. a warning counter is not reset. The argument list_ cannot be combined with any other arguments.

getoptions_

string vector of prefixes that indicate classes of options. In this package they can be "basic" and "solve". (E.g. package RandomFields has many more classes of options.) The given classes of options are then returned by RFoptions(). Note that the values are the previous values.

getoptions_ must always be the very first argument.

saveoptions_

string vector of prefixes. Same as for getoptions_, except that important classes are always returned and thus should not be given. Hence saveoptions_ is often a convenient short cut for getoptions_. The class always included in this package is "basic", in package RandomFields these are the two classes "basic" and "general".

saveoptions_ must always be the very first argument. In particular, it may not given at the same time with getoptions_.

local_

logical. This options is allowed only when advanced packages are used, see RandomFields.

warnUnknown_

integer. Same as option warn_unknown_option, except that its value overwrites the value of warn_unknown_option in the current command RFoptions. This options must be placed between CODE and getoptions_, if the latter are used.

Value

NULL if any argument is given, and the full list of arguments, otherwise.

Examples

n <- 10
M <- matrix(1, ncol=n, nrow=n)

## Not run: 
try(chol(M)) ## error, as M is not strictly positive definite
try(cholx(M)) ## also fails

## End(Not run)



RFoptions(la_mode=LA_INTERN, pivot=PIVOT_AUTO)
cholx(M) ## works
RFoptions(la_mode=LA_R)


RFoptions(solve_method="svd", pseudoinverse=TRUE)
solvex(M)
RFoptions(solve_method="method undefined", pseudoinverse=FALSE)

RandomFieldsUtils documentation built on April 19, 2022, 5:09 p.m.