Migration from CLAPACK 22.0.0

[sbrugnoli]

Hi,

I've been asked to migrate some code that uses CLAPACK 22.0.0 to LAPACK-3.4.1 and so far everything compiles and links correctly. But my problem comes when I compare in unit tests the outputs between CLAPACK and LAPACK: for the same given inputs I get different results. The functions in question in CLAPACK 22.0.0 are dgelsd_, dgeqrf_, dtrtri_ and dstev_. I have tried to migrate them to LAPACK 3.4.1 functions: LAPACK_dgelsd, LAPACK_dgeqrf, LAPACK_dtrtri, LAPACK_dstev and also to LAPACKE_dgelsd_work, LAPACKE_dgeqrf_work, LAPACKE_dtrtri_work, LAPACKE_dstev_work.

I get the best results (the ones that are similar to CLAPACK 22.0.0) with LAPACKE_*_work with column major, but I still get some different values in array WORK and IWORK. Should I worry about theses differences?
For example for array WORK I only get one element in the array with a different value:

CLAPACK-22.0.0(dgelsd_) -> WORK[9] = 1
LAPACK-3.4.1(LAPACKE_dgelsd_work) -> WORK[9] = 32

For all the other functions the significant differences are similar.

I'm compiling with GCC 4.7.2 and the makefile flags are the following:
- CLAPACK 22.0.0:
CFLAGS=-c -Wall -I./clapack/lib -I./clapack/include -Wno-deprecated
LDFLAGS=-L./clapack/lib -llapack -lblas -lf2c

- LAPACK 3.4.1:
CFLAGS=-c -Wall -I./lapack-3.4.1 -I./lapack-3.4.1/lapacke/include -Wno-deprecated
LDFLAGS=-L./lapack-3.4.1 -llapack -llapacke -lrefblas -DLAPACK_ILP64 -lgfortran

Thanks,
Sebastian.

[sbrugnoli]

Hi,

I have managed to solve some of the issues, but for function dgelsd_ in CLAPACK 22.0 I still cannot explain the differences with LAPACK_dgelsd (or LAPACKE_dgelsd_work with column major). The test case has been ove simplified and reduced to:

Code: Select all

#ifdef LP
#define lp_int lapack_int
extern "C" lapack_int LAPACKE_dgelsd_work( int matrix_order, lapack_int m, lapack_int n,
         lapack_int nrhs, double* a, lapack_int lda,
         double* b, lapack_int ldb, double* s, double rcond,
         lapack_int* rank, double* work, lapack_int lwork, lapack_int* iwork);
#else
extern "C" int dgelsd_(integer* m, integer* n, integer* nrhs, double* a,
         integer* lda, double* b, integer* ldb, double* s,
         double* rcond, integer* rank, double* work,
         integer* lwork, integer* iwork, integer *info );

#define lp_int integer
#endif

void T1_dgelsd(lp_int m, lp_int n)
 {
     lp_int nhrs = 1;
     lp_int lda = m;
     lp_int ldb = m;
     lp_int rank = 1;                   
     lp_int lwork = 1024 * m;
     lp_int* iwork = new lp_int[lwork];
     lp_int info = 0;

     double* a = new double[m*n];
     double* b = new double[m]; 
     double* s = new double[min(m, n)];

     double rcond = -1.0;
     double* work = new double[lwork];

     //Initialize input parameters
     int desp = 0;
     for(int j = 1; j <= n; j++) {
         desp = (j - 1)*m;
         for(int i = 1; i <= m; i++) a[desp + i - 1] = i * j;
     }

     for(int i = 1; i <= m; i++) b[i-1] = i;
     for(int i = 1; i <= m; i++) s[i-1] = 0;
     for(int i = 1; i <= lwork; i++) iwork[i-1] = 0;

     std::cout << "\n===========================" << std::endl;
     std::cout << " T1_dgelsd - TEST1" << std::endl;
     std::cout << "===========================" << std::endl;

     std::cout << "INPUT DATA => " << std::endl;
     report_dgelsd(&info, &m, &n, &nhrs, a, &lda, b, &ldb, s, &rcond, &rank, work, &lwork, iwork);

     #ifdef LP
        LAPACKE_dgelsd_work(LAPACK_COL_MAJOR, m, n, nhrs, a, lda, b, ldb, s, rcond, &rank, work, lwork, iwork);
     #else
        dgelsd_( &m, &n, &nhrs, a, &lda, b, &ldb, s, &rcond, &rank, work, &lwork, iwork, &info);
     #endif

     std::cout << "OUTPUT => " << std::endl;
     report_dgelsd(&info, &m, &n, &nhrs, a, &lda, b, &ldb, s, &rcond, &rank, work, &lwork, iwork);
 }


Test is called with inputs parametes:

Code: Select all

T1_dgelsd(4, 3); // A = [1 2 3 4 2 4 6 8 3 6 9 12] and B = [1 2 3 4]


The input matrix clearly has rank 1, and I get this for dgelsd_ but for LAPACKE_dgelsd_work ir return a rank value of 2!! The output is for both functions:

CLAPACK 22.0.0 ----------- LAPACK-3.4.1
m = 4 ----------------------- m = 4
n = 3 ------------------------ n = 3
nhrs = 1 -------------------- nhrs = 1
A[0] = 5.47723 ------------ A[0] = -5.47723
A[1] = 0 -------------------- A[1] = 0
A[2] = 0 -------------------- A[2] = 0
A[3] = -0.89341 ----------- A[3] = 0.617548
A[4] = -19.7484 ----------- A[4] = 19.7484
A[5] = 2.79784e-15 ------- A[5] = 1.12886e-15
A[6] = 0.122366 ----------- A[6] = 0.331679
A[7] = 0.579796 ----------- A[7] = 0.579796
A[8] = 0.535184 ----------- A[8] = 0.535184
A[9] = 6.7235e-16 -------- A[9] = -2.84902e-15
A[10] = 6.30467e-16 ----- A[10] = -1.5626e-15
A[11] = 7.58139 ----------- A[11] = -0.131902
lda = 4 ---------------------- lda = 4
B[0] = 0.0714286 ---------- B[0] = 0.103441
B[1] = 0.142857 ------------ B[1] = 0.476801
B[2] = 0.214286 ------------ B[2] = -0.0190139
B[3] = -2.16038e-31 ------- B[3] = -2.16038e-31
ldb = 4 ----------------------- ldb = 4
S[0] = 20.4939 ------------- S[0] = 20.4939
S[1] = 1.10799e-15 -------- S[1] = 3.26017e-15
S[2] = 4.25486e-16 -------- S[2] = 1.44604e-16
rcond = -1 ------------------- rcond = -1
rank = 1 --------------------- rank = 2

The rest of the output file has been omitted.

My makefile is now as follows:

Code: Select all

CC=g++
CFLAGS=-c -Wall -I./lapack-3.4.1 -I./lapack-3.4.1/lapacke/include -Wno-deprecated -DHAVE_LAPACK_CONFIG_H -DLAPACK_ILP64
LDFLAGS=-L./lapack-3.4.1 -llapacke -llapack -lrefblas -lgfortran
SOURCES=main.cpp
OBJECTS=$(SOURCES:.cpp=.o)
EXECUTABLE=test_lp

all: $(SOURCES) $(EXECUTABLE)

$(EXECUTABLE): $(OBJECTS)
     $(CC) $(OBJECTS) -o $@ $(LDFLAGS)

 .cpp.o:
    $(CC) $(CFLAGS) $< -o $@

clean:
    rm -f *.o test_lp


The test case has been executed in the same machine, bit not necessarly in the same core of the chip.

Thanks,
Sebastian.

[Julien Langou]

Hey Sebastian,

I believe both codes are working properly. (At least: both codes are doing what they are supposed to be doing.)

CLAPACK computes S[2] as 3.26017e-15 and S[1] as 20.4939. In this case, 3.26017e-15/20.4939 = 1.5908e-16 is just above machine epsilon, and so CLAPACK decides that the matrix has rank two. LAPACK computes S[2] as 1.10799e-15 and S[1] as 20.4939. In this case, 1.10799e-15/20.4939 = 5.4064e-17 is just below machine epsilon, and so LAPACK decides that the matrix has rank one.

Of course this results in a huge change on the solution X !!!!

Well, that's what it is . . . S[2] is very hard to compute with high relative accuracy and so the numerical rank of your matrix might be one or two depending on what value you get for S[2] and the criterion.

The best is to change RCOND in input of DGELSD. I assume you use a default value for RCOND. Why don't you set RCOND to 1e-10 or something like this? Even 1e-15 or 5e-16 would do the trick in your case. The default is machine epsilon.

Cheers,
Julien.

[sbrugnoli]

Hi Julien,

Thanks for your reply. Following your advise, I changed RCOND to 5e-16 and now B matrix consistently has the same values for all unit tests . But I'm not sure why using machine epsilon would lead to different results for matrix B. I don't have previous experience with scientific programming, but if I'm correctly understanding, changing RCOND to a bigger value makes the calculation less sentitive to small variations, losing accuracy in the name of reproducibility? Am I right?

Another question I have, is it good practice to use machine epsilon (RCOND = -1.0)? Is there a criteria to define RCOND?

Thanks,
Sebastian.

[sbrugnoli]

I found the answer in an older post: viewtopic.php?f=2&t=1956&p=5536&hilit=epsilon#p5536