Known Issues for the NAG Library CL Interface
This document reflects all reported and resolved issues that affect various releases of the NAG Library CL Interface up to Mark 27.1.
Some of these issues may have been fixed at intermediate "point" releases of the Library, while other fixes are scheduled
for incorporation at future releases. For library Marks where those fixes are not yet incorporated, a workaround for the known
issue is provided wherever possible.
To find the Mark and point release number of your library, call NAG function
a00aac( ).
Order the issues by
Synopsis | Overflow may occur if the function attempts to scale the polynomial coefficients. |
Description | In rare circumstances overflow may be observed if . |
Severity | Non-critical |
Issue Since Mark | 7 |
Workaround | Set argument . |
Synopsis | Multi-level wavelets cannot handle periodic end extension. |
Description | When and the multi-level wavelet transform functions do not work properly if is not a power of 2. |
Severity | Non-critical |
Issue Since Mark | 9 |
Fixed at Mark | 23 |
Workaround | The option combination of a multi-level wavelet transform using a periodic end extension is currently disallowed; a call to
the initialization function c09aac with this combination will return with an error code.
For multilevel analysis of periodic data, you are advised to experiment with the alternative end conditions; the periodic
property of the data can also be used to extend the data set in both directions to points that better suit the alternative
end condition (e.g., extend the data to next maximum or minimum).
|
Synopsis | Initialization and option setting does not work when using the long name. |
Description | Initialization and option setting for the sparse grid function nag_quad_md_sgq_multi_vec (d01esc) using nag_quad_opt_set (d01zkc) does not work when using the long name nag_quad_md_sgq_multi_vec in the option string.
It does work when using the short name d01esc in the option string.
|
Severity | Non-critical |
Issue Since Mark | 25 |
Fixed at Mark | 25.1 |
Workaround | Initializing and setting options for nag_quad_md_sgq_multi_vec (d01esc) via calls to nag_quad_opt_set (d01zkc) should use option strings containing the short name d01esc rather than the long name. |
Synopsis | Segmentation faults when optional parameter is set to a value greater than . |
Description | Segmentation faults or other array bound violation problems may occur if the value of (set via a call to d01zkc) is greater than , the maximum level of the underlying quadrature rule. |
Severity | Critical |
Issue Since Mark | 25 |
Fixed at Mark | 25.4 |
Workaround | Do not set to more than 9 when using Gauss–Patterson or more than 12 when using Clenshaw–Curtis. |
Synopsis | is not accepted as a valid option. |
Description | When setting the quadrature rule for d01esc using the d01zkc option setting function, the documented choice is not recognised as a valid option and an error is reported. |
Severity | Non-critical |
Issue Since Mark | 25 |
Fixed at Mark | 25.4 |
Workaround | The alternatives
may be used instead.
Note: Gauss-Patterson is the default choice for the quadrature rule in d01esc, so in general it will not be necessary to specify this option.
|
Synopsis | Stack size or thread safety problems may be observed with some d06 functions. |
Description | d06aac, d06abc and d06acc contain large local arrays that may cause stack size and/or thread safety problems. |
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 23.3 |
Workaround | Do not use these functions in a multithreaded environment. For serial execution, set stack size limit to 10MB or greater. |
Synopsis | Although the documented constraint on
is
, the actual required minimum for
is
.
For some small scale problems, setting
will give unpredictable results and could produce a segmentation fault.
|
Description | Although the documented constraint on
is
, the actual required minimum for
is
.
For some small scale problems, setting
will give unpredictable results and could produce a segmentation fault.
The problem is remedied by setting
and ensuring that the arrays
and
are correspondingly large enough.
|
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 26 |
Workaround | Set ; allocate the arrays and using this value of . |
Synopsis | d06acc returns error for some boundary meshes due to an internal scaling issue. |
Description | d06acc returns error for some boundary meshes due to an internal scaling issue. |
Severity | Non-critical |
Issue Since Mark | 7 |
Fixed at Mark | 28.4 |
Workaround | Scale input boundary mesh prior to calling d06acc so that and . |
Synopsis | The algorithm underlying interpolation functions e01sgc, e01shc, e01tgc and e01thc was modified at Mark 26 and Mark 26.1; different results will be obtained when using these functions than previously. |
Description | The algorithm underlying interpolation functions e01sgc, e01shc, e01tgc and e01thc was modified at Mark 26 to improve perceived deficiencies. In particular, at earlier library Marks the evaluation functions would not attempt to return any useful result if an evaluation point was not close enough to any of the original data points,
and this issue was addressed at Mark 26.
At Mark 26.1 further work was done on the functions because they had been found not to work well on gridded data sets (as opposed to the random data sets that they are primarily
intended for).
It should be noted that because of the various underlying changes to the functions, the precise results returned from Mark 26 onwards will not usually be identical to those before Mark 26.
|
Severity | Non-critical |
Issue Since Mark | 26 |
Fixed at Mark | 26.1 |
Workaround | Not applicable. |
Synopsis | e01shc will occasionally incorrectly identify a point as being outside the region defined by the interpolant. |
Description | e01shc will occasionally incorrectly identify a point as being outside the region defined by the interpolant. This leads to the function value being extrapolated rather than interpolated and can lead to incorrect results. |
Severity | Non-critical |
Issue Since Mark | 26.0 |
Fixed at Mark | 27.1 |
Workaround | None. |
Synopsis | Ill-conditioned data sets may cause e02gac to get stuck in an infinite loop. |
Description | Certain ill-conditioned data sets could cause e02gac to get stuck in an infinite loop. |
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 26 |
Workaround | As a workaround, it may be possible to avoid the infinite loop by reordering the points in the input data. |
Synopsis | e04dgc returns wrong gradient values in . |
Description | e04dgc prints gradient values correctly, but returns the wrong values in argument . |
Severity | Non-critical |
Issue Since Mark | 7 |
Fixed at Mark | 25 |
Workaround | In , when is set, push these values into the structure. Following the call to e04dgc the correct values for can then be obtained from . |
Synopsis | Internal buffer overflow in e04fcc. |
Description | When the grade of the optimization problem drops to zero, an internal buffer overflow occurs. This destroys some of the internal
state of the optimizer, typically causing it to stop prematurely.
Scope of the problem:
If the grade of the optimization problem is non-zero on exit from e04fcc, then the bug is not triggered and that particular optimization is unaffected. If the grade is zero on exit, then the optimization
is affected in all supported CL marks.
How the problem manifests:
Severity:
Since the solver is typically close to convergence when the grade drops to zero, the returned solution is usually pretty good.
The bug fix is unlikely to improve the results of e04fcc significantly.
|
Severity | Non-critical |
Issue Since Mark | 8 |
Fixed at Mark | 24.2 |
Workaround | There is no practical workaround. |
Synopsis | In very rare cases, the algorithm used by e04lbc may become trapped in an infinite loop. |
Description | The function might loop unnecessarily and finish with when a variable lying on the boundary is cyclically added and removed to/from free variables. This can happen only at points with indefinite Hessian and very small projected gradients when one variable is lying on the boundary and another one is very close to it. |
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 25 |
Workaround | Unfortunately there is no convenient workaround. |
Synopsis | and were not correctly filled by the presolver. |
Description | The arrays
and
were not correctly filled when the problem was entirely solved by the presolver. It now returns the correct values.
The optional parameter
now correctly writes the linear constraints dual variables when no bounds are defined on the variables.
|
Severity | Non-critical |
Issue Since Mark | 26.1 |
Fixed at Mark | 27 |
Workaround | Don't rely on to hold the primal and dual objective in this case and recompute it as and , respectively. |
Synopsis | e04mtc does not report the correct solution when or more columns are proportional to each other in the constraint matrix. |
Description | e04mtc does not report the correct solution when or more columns are proportional to each other in the constraint matrix. In such a case, the solution reported may be infeasible. |
Severity | Non-critical |
Issue Since Mark | 26.1 |
Fixed at Mark | 27 |
Workaround | A workaround is to disable the more complex presolve operations by setting the optional parameter . This may slow down the solver depending on the problem. |
Synopsis | In some very rare cases, the solution reported presents big violations on a small number of linear constraints. |
Description | In some very rare cases, the solution reported presents big violations on a small number of linear constraints. |
Severity | Non-critical |
Issue Since Mark | 26.1 |
Fixed at Mark | 27.1 |
Workaround | A workaround is to deactivate the more complex presolver operations with the optional parameter . |
Synopsis | In some very rare cases, e04mtc reports problem infeasibility for a feasible problem. |
Description | In some very rare cases, the solver reports problem infeasibility when there are numerical difficulties. |
Severity | Non-critical |
Issue Since Mark | 26.1 |
Fixed at Mark | 28.6 |
Workaround | Unfortunately there is no convenient workaround. |
Synopsis | Infeasible bounds defined by e04rjc of a variable are ignored and infeasibility is not reported. |
Description | When infeasible bounds are defined by e04rjc for a variable, instead of reporting problem infeasibility, the bounds are overridden and wrong solution may be reported. |
Severity | Non-critical |
Issue Since Mark | 26.1 |
Fixed at Mark | 27.1 |
Workaround | A workaround is to deactivate the more complex presolver operations with the optional parameter for e04mtc and for e04ptc. |
Synopsis | Internal file overflow. |
Description | If you set a in e04nqc, e04vhc and e04wdc and your total problem size ( , or , respectively) is greater than 80 you will experience an internal buffer overflow and possible program crash. |
Severity | Critical |
Issue Since Mark | 9.4 |
Fixed at Mark | 23 |
Workaround | Unfortunately there is no convenient workaround. The only way to avoid this crash is to not specify a or to have a small enough problem. |
Synopsis | Optional parameters and are not handled correctly. |
Description | Functions e04nrc, e04vkc and e04wec do not handle optional parameters and correctly. Specifying does not alter the behaviour of subsequent functions in the suite, and specifying erroneously reports an error. |
Severity | Non-critical |
Issue Since Mark | 8 |
Fixed at Mark | 27.3 |
Workaround | Function e04nsc should be used instead to set optional parameters or . |
Synopsis | e04stc returns Lagrangian multipliers in the wrong order. |
Description | The Lagrangian multipliers returned in
are in the wrong order:
|
Severity | Non-critical |
Issue Since Mark | 26 |
Fixed at Mark | 26.1 |
Workaround | The order described in the documentation is now used. |
Synopsis | An inner optimization step might be unnecessarily resolved in certain cases. |
Description | If the solver is run in the mode when some (or all) derivatives might be missing ( or ), however, all derivatives are provided, the solver might trigger an extra resolve of the inner optimization step if it detects numerical difficulties. This extra step normally switches derivative approximations to central differences and thus is aimed to improve stability of the problem, however, in this case it doesn't change anything (all derivatives are provided by the user and no derivative approximation takes place) and thus is unnecessary. |
Severity | Non-critical |
Issue Since Mark | 7 |
Fixed at Mark | 26.1 |
Workaround | You might want to set and if all derivatives are provided. |
Synopsis | A possible buffer overflow in the printing of the derivative checker of e04unc. |
Description | If is set to a full check of the objective function and/or the constraints ( , ...) and and the checked function has zero elements in the derivatives, internal buffer might overflow which might lead to a crash. |
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 26 |
Workaround | To avoid the problem, use . |
Synopsis | When the objective function has no separated linear part, using user-defined names for variables and constraints might lead to a crash. |
Description | When the objective function only has the nonlinear part defined without a separated linear part, the solver might crash when trying to read user-defined names for variables and constraints. |
Severity | Critical |
Issue Since Mark | 8 |
Fixed at Mark | 27.1 |
Workaround | Unfortunately there is no convenient workaround. The only way to avoid this crash is to not specify names for variables and constraints. |
Synopsis | Information about the last constraint might not be printed. |
Description | If the problem has a nonlinear objective function without a linear part and , the last constraint is not printed in the final information about the solution (Rows section). |
Severity | Non-critical |
Issue Since Mark | 8 |
Fixed at Mark | 26 |
Workaround | None. |
Synopsis | User-supplied character strings containing spaces may cause garbled error messages. |
Description | Some functions produce error messages containing character data that has been supplied through the argument by the user. An example is e04vhc, where the or can be referred to in error messages. Having spaces in these strings confuses the internal error-message splitter, which splits on spaces. Thus, error messages returned by the function may be garbled. |
Severity | Non-critical |
Issue Since Mark | 9 |
Fixed at Mark | 23 |
Workaround | Make sure user-provided character data is without spaces |
Synopsis | An unhelpful error exit is returned if e05ucc is called with incorrectly initialized optional parameter arrays and . |
Description | Function e05ucc returns if e05ucc is called without previously having called e05zkc with argument ‘Initialize = e05ucc’. |
Severity | Non-critical |
Issue Since Mark | 23 |
Fixed at Mark | 24 |
Workaround | Call e05zkc with argument ‘Initialize = e05ucc’ before calling e05ucc. |
Synopsis | Function f02wef may fail to compute any results, but with no error flag set. |
Description | Certain combinations of arguments and together with their associated output arrays and can cause f02wec to fail to compute any results, but with no error flag set. Specifically, the function documentation states that in some circumstances array argument may be a NULL pointer (in which case the left hand singular vectors, if required, are stored in array ). However, an auxiliary function called by f02wef checks whether is NULL, and if so f02wef silently fails. |
Severity | Critical |
Issue Since Mark | 25 |
Fixed at Mark | 26 |
Workaround | Always supply a non-NULL array argument even if the documentation for f02wef states that a NULL pointer is allowed. |
Synopsis | Function f02xec may fail to compute any results, but with no error flag set. |
Description | Certain combinations of arguments and together with their associated output arrays and can cause f02xec to fail to compute any results, but with no error flag set. Specifically, the function documentation states that in some circumstances array argument may be a NULL pointer (in which case the left hand singular vectors, if required, are stored in array ). However, an auxiliary function called by f02xec checks whether is NULL, and if so f02xec silently fails. |
Severity | Critical |
Issue Since Mark | 25 |
Fixed at Mark | 26 |
Workaround | Always supply a non-NULL array argument even if the documentation for f02xec states that a NULL pointer is allowed. |
Synopsis | Multithreaded versions of the functions f11bec, f11bsc, f11gec and f11gsc may produce slightly different results when run on multiple threads. |
Description | Multithreaded versions of the functions f11bec, f11bsc, f11gec and f11gsc may produce slightly different results when run on multiple threads, e.g., the number of iterations to solution and the computed matrix norms and termination criteria reported by the associated monitoring functions. A bug affecting f11bec and f11gec has been fixed, and parallel vector dot products have been modified in all functions to improve consistency of results. |
Severity | Non-critical |
Issue Since Mark | 26 |
Fixed at Mark | 27.1 |
Workaround | None. |
Synopsis | f16qec and f16tec reference diagonal elements when unit diagonal entries are assumed. |
Description | f16qec and f16tec reference and copy diagonal elements when unit diagonal entries are assumed. |
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 28 |
Workaround | Nothing needs to be done unless diagonal entries of the target matrix contain useful data prior to a call of f16qec or f16tec with , in which case the useful data should be saved and copied back to the diagonal of the target matrix after the call to either f16qec or f16tec. |
Synopsis | f16qfc, if called with that violates minimum contraints, will produce a segmentation fault. |
Description | f16qfc, if called with that violates minimum contraints, will produce a segmentation fault. |
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 28 |
Workaround | Call f16qfc with that meets the documented minimum contraint. |
Synopsis | f16rbc and f16ubc return if or is , instead of the correct norm. is incorrectly forced to be at least when . |
Description | f16rbc and f16ubc mistakenly make a quick return if or is , instead of computing the correct value for the requested norm. Also, is incorrectly forced to be at least when . |
Severity | Critical |
Issue Since Mark | 9.1 |
Fixed at Mark | 23 |
Workaround | If or is , use the general matrix-norm functions f16rac or f16uac, with the input matrix in full storage. If , make sure that . |
Synopsis | Incorrect Frobenius norm returned in some cases. |
Description | When calling one of the functions: f16rdc, f16rec, f16udc, f16uec, f16ufc and f16ugc with and , the returned norm can be incorrect. |
Severity | Critical |
Issue Since Mark | 23 |
Fixed at Mark | 28 |
Workaround | These functions will return the correct norm if the argument is set to and the argument is flipped, i.e., from to or vice versa. |
Synopsis | f16smc returns wrong update of when is stored in row major order and is to be conjugated. |
Description | When f16smc is called with and , is updated as though . |
Severity | Critical |
Issue Since Mark | 8 |
Fixed at Mark | 28 |
Workaround | Call f16smc with and conjugate prior to call. |
Synopsis | f16tac stops program execution when called with . |
Description | f16tac, when called with does not return error code , but terminates program execution. |
Severity | Critical |
Issue Since Mark | 8 |
Fixed at Mark | 28 |
Workaround | Call f16tac with . |
Synopsis | f16tfc returns incorrect results when computing a transposed copy of a matrix. |
Description | f16tfc returns incorrect results when computing a transposed copy of a matrix. |
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 28 |
Workaround | Call f01cwf with and ; for row-ordered matrices, and should be switched. |
Synopsis | The returned matrix is not a valid correlation matrix. |
Description | The algorithm computes an incorrect value for . Thus the returned matrix is not positive definite as stated, and is not a valid correlation matrix. |
Severity | Critical |
Issue Since Mark | 25 |
Fixed at Mark | 25.3 |
Workaround | Unfortunately there is no convenient workaround. |
Synopsis | When , output arguments and are not initialized. |
Description | When , output arguments and are not initialized. These values relate to a regression constant that is only relevant in the case. However, the code for should initialize them to . This was not done, allowing previously set values or random results to be erroneously returned. |
Severity | Non-critical |
Issue Since Mark | 7 |
Fixed at Mark | 27.3 |
Workaround | The safest solution is to manually set these to (but only in the case) immediately after calling this function. |
Synopsis | Incorrect results are returned when performing a Mallows type regression. |
Description | Incorrect results are returned when performing a Mallows type regression, averaging over residuals. |
Severity | Non-critical |
Issue Since Mark | 7 |
Fixed at Mark | 26.1 |
Workaround | None. |
Synopsis | Segmentation fault caused by access past the end of an array. |
Description | An error can occur when there are multiple blocks of random variables, at least one with a subject variable and at least one without. The error can only occur when the block with the subject variable occurs first in . |
Severity | Critical |
Issue Since Mark | 23 |
Fixed at Mark | 25 |
Workaround | Ensure that blocks without subject variables appear in before those with subject variables. |
Synopsis | In very rare cases, the function may become trapped in an infinite loop. |
Description | The function was affected by a bug in the underlying solver e04lbc (modified Newton method). In very rare cases the solver might get into an infinite loop. |
Severity | Critical |
Issue Since Mark | 9 |
Fixed at Mark | 25 |
Workaround | The bug can be avoided by switching to the other optimizer (SQP method e04ucc, ). |
Synopsis | A segmentation fault is likely to occur if a model with multiple random statements is supplied to the function, where at least one of those statements does not have a term. |
Description | A segmentation fault is likely to occur if a model with multiple random statements is supplied to the function, where at least one of those statements does not have a
term.
For example, a model specified using:
V1 + V2 / SUBJECT = V3 V4 + V5 / SUBJECT = V6would not trigger the error, but one specified using: V1 + V2 V4 + V5 / SUBJECT = V6would. The error is not triggered when there is only a single random statement, so a model specified using just
V1 + V2will not trigger the error. |
Severity | Critical |
Issue Since Mark | 27 |
Fixed at Mark | 27.1 |
Workaround | A workaround to this issue is to always supply a
term. If the required model is of the form:
V1 + V2 V4 + V5 / SUBJECT = V6then you can specify an equivalent model by using: V1 + V2 / SUBJECT = DUMMY V4 + V5 / SUBJECT = V6where the variable |
Synopsis | Returns incorrect results when and user supplied initial values for are being used. |
Description | If , the optional parameter is set, and the rows of array are not all identical, then the results returned by g02qgc are incorrect. |
Severity | Critical |
Issue Since Mark | 23 |
Fixed at Mark | 24 |
Workaround | Rather than call the function once with , call the function multiple times with , analysing a different value of on each call. |
Synopsis | Unexpected s in g02zkc. |
Description | g02zkc may report a error for some valid minimum abbreviations of option names supplied in the input argument , e.g., when using ‘DEF’ instead of ‘Defaults’. |
Severity | Non-critical |
Issue Since Mark | 23 |
Fixed at Mark | 26 |
Workaround | Use the full options name, e.g., specify ‘Defaults’ rather than ‘Def’. |
Synopsis | Memory leak reported from CL interface on Windows from checking tools. |
Description | A during-execution memory leak can be reported from checking tools on Windows when running a multi-threaded program calling d01xbc. |
Severity | Critical |
Issue Since Mark | 5 |
Fixed at Mark | 28 |
Workaround | Ignore warnings. |
Synopsis | The wrong value for is returned when is large. |
Description | In g08ckc and g08clc the value returned for the upper tail probability is wrong when the calculated Anderson-Darling test statistic is large. In the case of g08ckc, when the returned value of should be zero; in the case of g08clc, when the returned value of should be . |
Severity | Critical |
Issue Since Mark | 23 |
Workaround | Workaround for g08ckc:
Call g08ckc(...); If (aa2 > 153.4677) p = 0.0;Workaround for g08clc: Call g08clc(...); If (aa2 > 10.03) p = exp(-14.360135); |
Synopsis | The methods were implemented in reverse. |
Description | g10cac implements two methods of smoothing, . Unfortunately they were implemented in reverse, so if you ask for you get and vice versa. |
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 25 |
Workaround | Use if you want results for , and vice versa. |
Synopsis | g13fac may return a negative value as the estimate of the last parameter (i.e., ) for a subset of models. |
Description | g13fac can result in a negative value for the estimate of the last
parameter (i.e.,
) or, if
, the last
parameter (i.e.,
).
This issue only affects a subset of models that have normally distributed errors and do not include an asymmetry term.
If the function did not return a negative value as the estimate of the last
parameter (or, if
, the last
parameter), then that particular model was not affected by the issue.
|
Severity | Critical |
Issue Since Mark | 7 |
Fixed at Mark | 27 |
Workaround | None |
Synopsis | When the information returned in and/or may be incorrect. |
Description | The information returned in and/or may be incorrect in cases where and the underlying linear mixed effects regression model has a random variable, with a single level (so either binary or continuous), that only takes the value zero. |
Severity | Non-critical |
Issue Since Mark | 27.0 |
Workaround | The work around is to drop the term from the model, as it does not contribute. For example, if the random part of your model was specified as: V1 + V2 / SUBJECT=V3 and the variable V2 was a continuous variable, that only takes a value of zero in the data, then this is equivalent to re-specifying the model using: V1 / SUBJECT=V3. |
Synopsis | Thread Local Storage default limit was exceeded for delay loaded shared library. |
Description | A fair amount of thread local storage had been allocated by an auxiliary function which has now been updated to use a very small amount of thread local storage. Prior to the update, this only affected the
case where the shared version of the Nag Library was delay loaded, since this assumed a small default maximum amount of thread
local storage, which was in fact exceeded.
The issue had been present since the introduction of the auxiliary function at Mark 26.1.
From Mark 28.6, the amount of thread local storage used is very small and this is no longer an issue.
|
Severity | Non-critical |
Issue Since Mark | 26.1 |
Fixed at Mark | 28.6 |
Workaround | None. |