extern "C" magma_int_t
magma_dsgesv_gpu(char trans, magma_int_t n, magma_int_t nrhs,
double *dA, magma_int_t ldda,
magma_int_t *ipiv, magma_int_t *dipiv,
double *dB, magma_int_t lddb,
double *dX, magma_int_t lddx,
double *dworkd, float *dworks,
magma_int_t *iter, magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
DSGESV computes the solution to a real system of linear equations
A * X = B or A' * X = B
where A is an N-by-N matrix and X and B are N-by-NRHS matrices.
DSGESV first attempts to factorize the matrix in real SINGLE PRECISION
and use this factorization within an iterative refinement procedure
to produce a solution with real DOUBLE PRECISION norm-wise backward error
quality (see below). If the approach fails the method switches to a
real DOUBLE PRECISION factorization and solve.
The iterative refinement is not going to be a winning strategy if
the ratio real SINGLE PRECISION performance over real DOUBLE PRECISION
performance is too small. A reasonable strategy should take the
number of right-hand sides and the size of the matrix into account.
This might be done with a call to ILAENV in the future. Up to now, we
always try iterative refinement.
The iterative refinement process is stopped if
ITER > ITERMAX
or for all the RHS we have:
RNRM < SQRT(N)*XNRM*ANRM*EPS*BWDMAX
where
o ITER is the number of the current iteration in the iterative
refinement process
o RNRM is the infinity-norm of the residual
o XNRM is the infinity-norm of the solution
o ANRM is the infinity-operator-norm of the matrix A
o EPS is the machine epsilon returned by DLAMCH('Epsilon')
The value ITERMAX and BWDMAX are fixed to 30 and 1.0D+00 respectively.
Arguments
=========
TRANS (input) CHARACTER*1
Specifies the form of the system of equations:
= 'N': A * X = B (No transpose)
= 'T': A'* X = B (Transpose)
= 'C': A'* X = B (Conjugate transpose = Transpose)
N (input) INTEGER
The number of linear equations, i.e., the order of the
matrix A. N >= 0.
NRHS (input) INTEGER
The number of right hand sides, i.e., the number of columns
of the matrix B. NRHS >= 0.
dA (input or input/output) DOUBLE PRECISION array on the GPU, dimension (ldda,N)
On entry, the N-by-N coefficient matrix A.
On exit, if iterative refinement has been successfully used
(info.EQ.0 and ITER.GE.0, see description below), A is
unchanged. If double precision factorization has been used
(info.EQ.0 and ITER.LT.0, see description below), then the
array dA contains the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
ldda (input) INTEGER
The leading dimension of the array dA. ldda >= max(1,N).
IPIV (output) INTEGER array, dimension (N)
The pivot indices that define the permutation matrix P;
row i of the matrix was interchanged with row IPIV(i).
Corresponds either to the single precision factorization
(if info.EQ.0 and ITER.GE.0) or the double precision
factorization (if info.EQ.0 and ITER.LT.0).
dIPIV (output) INTEGER array on the GPU, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was moved to row IPIV(i).
dB (input) DOUBLE PRECISION array on the GPU, dimension (lddb,NRHS)
The N-by-NRHS right hand side matrix B.
lddb (input) INTEGER
The leading dimension of the array dB. lddb >= max(1,N).
dX (output) DOUBLE PRECISION array on the GPU, dimension (lddx,NRHS)
If info = 0, the N-by-NRHS solution matrix X.
lddx (input) INTEGER
The leading dimension of the array dX. lddx >= max(1,N).
dworkd (workspace) DOUBLE PRECISION array on the GPU, dimension (N*NRHS)
This array is used to hold the residual vectors.
dworks (workspace) SINGLE PRECISION array on the GPU, dimension (N*(N+NRHS))
This array is used to store the real single precision matrix
and the right-hand sides or solutions in single precision.
iter (output) INTEGER
< 0: iterative refinement has failed, double precision
factorization has been performed
-1 : the routine fell back to full precision for
implementation- or machine-specific reasons
-2 : narrowing the precision induced an overflow,
the routine fell back to full precision
-3 : failure of SGETRF
-31: stop the iterative refinement after the 30th iteration
> 0: iterative refinement has been successfully used.
Returns the number of iterations
info (output) INTEGER
= 0: successful exit
< 0: if info = -i, the i-th argument had an illegal value
> 0: if info = i, U(i,i) computed in DOUBLE PRECISION is
exactly zero. The factorization has been completed,
but the factor U is exactly singular, so the solution
could not be computed.
===================================================================== */
#define dB(i,j) (dB + (i) + (j)*lddb)
#define dX(i,j) (dX + (i) + (j)*lddx)
#define dR(i,j) (dR + (i) + (j)*lddr)
double c_neg_one = MAGMA_D_NEG_ONE;
double c_one = MAGMA_D_ONE;
magma_int_t ione = 1;
double *dR;
float *dSA, *dSX;
double Xnrmv, Rnrmv;
double Anrm, Xnrm, Rnrm, cte, eps;
magma_int_t i, j, iiter, lddsa, lddr;
/* Check arguments */
*iter = 0;
*info = 0;
if ( n < 0 )
*info = -1;
else if ( nrhs < 0 )
*info = -2;
else if ( ldda < max(1,n))
*info = -4;
else if ( lddb < max(1,n))
*info = -8;
else if ( lddx < max(1,n))
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if ( n == 0 || nrhs == 0 )
return *info;
lddsa = n;
lddr = n;
dSA = dworks;
dSX = dSA + lddsa*n;
dR = dworkd;
eps = lapackf77_dlamch("Epsilon");
Anrm = magmablas_dlange('I', n, n, dA, ldda, (double*)dworkd );
cte = Anrm * eps * pow((double)n, 0.5) * BWDMAX;
/*
* Convert to single precision
*/
//magmablas_dlag2s( n, nrhs, dB, lddb, dSX, lddsx, info ); // done inside dsgetrs with pivots
if (*info != 0) {
*iter = -2;
goto FALLBACK;
}
magmablas_dlag2s( n, n, dA, ldda, dSA, lddsa, info );
if (*info != 0) {
*iter = -2;
goto FALLBACK;
}
// factor dSA in single precision
magma_sgetrf_gpu( n, n, dSA, lddsa, ipiv, info );
if (*info != 0) {
*iter = -3;
goto FALLBACK;
}
// Generate parallel pivots
{
magma_int_t *newipiv;
magma_imalloc_cpu( &newipiv, n );
if ( newipiv == NULL ) {
*iter = -3;
goto FALLBACK;
}
swp2pswp( trans, n, ipiv, newipiv );
magma_setvector( n, sizeof(magma_int_t), newipiv, 1, dipiv, 1 );
magma_free_cpu( newipiv );
}
// solve dSA*dSX = dB in single precision
// converts dB to dSX and applies pivots, solves, then converts result back to dX
magma_dsgetrs_gpu( trans, n, nrhs, dSA, lddsa, dipiv, dB, lddb, dX, lddx, dSX, info );
// residual dR = dB - dA*dX in double precision
magmablas_dlacpy( MagmaUpperLower, n, nrhs, dB, lddb, dR, lddr );
if ( nrhs == 1 ) {
magma_dgemv( trans, n, n,
c_neg_one, dA, ldda,
dX, 1,
c_one, dR, 1 );
}
else {
magma_dgemm( trans, MagmaNoTrans, n, nrhs, n,
c_neg_one, dA, ldda,
dX, lddx,
c_one, dR, lddr );
}
// TODO: use MAGMA_D_ABS( dX(i,j) ) instead of dlange?
for( j=0; j < nrhs; j++ ) {
i = magma_idamax( n, dX(0,j), 1) - 1;
magma_dgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1 );
Xnrm = lapackf77_dlange( "F", &ione, &ione, &Xnrmv, &ione, NULL );
i = magma_idamax ( n, dR(0,j), 1 ) - 1;
magma_dgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1 );
Rnrm = lapackf77_dlange( "F", &ione, &ione, &Rnrmv, &ione, NULL );
if ( Rnrm > Xnrm*cte ) {
goto REFINEMENT;
}
}
*iter = 0;
return *info;
REFINEMENT:
for( iiter=1; iiter < ITERMAX; ) {
*info = 0;
// convert residual dR to single precision dSX
// solve dSA*dSX = R in single precision
// convert result back to double precision dR
// it's okay that dR is used for both dB input and dX output.
magma_dsgetrs_gpu( trans, n, nrhs, dSA, lddsa, dipiv, dR, lddr, dR, lddr, dSX, info );
if (*info != 0) {
*iter = -3;
goto FALLBACK;
}
// Add correction and setup residual
// dX += dR --and--
// dR = dB
// This saves going through dR a second time (if done with one more kernel).
// -- not really: first time is read, second time is write.
for( j=0; j < nrhs; j++ ) {
magmablas_daxpycp( n, dR(0,j), dX(0,j), dB(0,j) );
}
// residual dR = dB - dA*dX in double precision
if ( nrhs == 1 ) {
magma_dgemv( trans, n, n,
c_neg_one, dA, ldda,
dX, 1,
c_one, dR, 1 );
}
else {
magma_dgemm( trans, MagmaNoTrans, n, nrhs, n,
c_neg_one, dA, ldda,
dX, lddx,
c_one, dR, lddr );
}
/* Check whether the nrhs normwise backward errors satisfy the
* stopping criterion. If yes, set ITER=IITER>0 and return. */
for( j=0; j < nrhs; j++ ) {
i = magma_idamax( n, dX(0,j), 1) - 1;
magma_dgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1 );
Xnrm = lapackf77_dlange( "F", &ione, &ione, &Xnrmv, &ione, NULL );
i = magma_idamax ( n, dR(0,j), 1 ) - 1;
magma_dgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1 );
Rnrm = lapackf77_dlange( "F", &ione, &ione, &Rnrmv, &ione, NULL );
if ( Rnrm > Xnrm*cte ) {
goto L20;
}
}
/* If we are here, the nrhs normwise backward errors satisfy
* the stopping criterion, we are good to exit. */
*iter = iiter;
return *info;
L20:
iiter++;
}
/* If we are at this place of the code, this is because we have
* performed ITER=ITERMAX iterations and never satisified the
* stopping criterion. Set up the ITER flag accordingly and follow
* up on double precision routine. */
*iter = -ITERMAX - 1;
FALLBACK:
/* Single-precision iterative refinement failed to converge to a
* satisfactory solution, so we resort to double precision. */
magma_dgetrf_gpu( n, n, dA, ldda, ipiv, info );
if (*info == 0) {
magmablas_dlacpy( MagmaUpperLower, n, nrhs, dB, lddb, dX, lddx );
magma_dgetrs_gpu( trans, n, nrhs, dA, ldda, ipiv, dX, lddx, info );
}
return *info;
}
/**
Purpose
-------
Solves a system of linear equations
A * X = B
where A is a general N-by-N matrix and X and B are N-by-NRHS matrices.
The LU decomposition with partial pivoting and row interchanges is
used to factor A as
A = P * L * U,
where P is a permutation matrix, L is unit lower triangular, and U is
upper triangular. The factored form of A is then used to solve the
system of equations A * X = B.
Arguments
---------
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in]
nrhs INTEGER
The number of right hand sides, i.e., the number of columns
of the matrix B. NRHS >= 0.
@param[in,out]
A REAL array, dimension (LDA,N).
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[out]
ipiv INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
@param[in,out]
B REAL array, dimension (LDB,NRHS)
On entry, the right hand side matrix B.
On exit, the solution matrix X.
@param[in]
ldb INTEGER
The leading dimension of the array B. LDB >= max(1,N).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_sgesv_driver
********************************************************************/
extern "C" magma_int_t
magma_sgesv(
magma_int_t n, magma_int_t nrhs,
float *A, magma_int_t lda,
magma_int_t *ipiv,
float *B, magma_int_t ldb,
magma_int_t *info)
{
magma_int_t ngpu, ldda, lddb;
*info = 0;
if (n < 0) {
*info = -1;
} else if (nrhs < 0) {
*info = -2;
} else if (lda < max(1,n)) {
*info = -4;
} else if (ldb < max(1,n)) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (n == 0 || nrhs == 0) {
return *info;
}
/* If single-GPU and allocation suceeds, use GPU interface. */
ngpu = magma_num_gpus();
float *dA, *dB;
if ( ngpu > 1 ) {
goto CPU_INTERFACE;
}
ldda = ((n+31)/32)*32;
lddb = ldda;
if ( MAGMA_SUCCESS != magma_smalloc( &dA, ldda*n )) {
goto CPU_INTERFACE;
}
if ( MAGMA_SUCCESS != magma_smalloc( &dB, lddb*nrhs )) {
magma_free( dA );
goto CPU_INTERFACE;
}
magma_ssetmatrix( n, n, A, lda, dA, ldda );
magma_sgetrf_gpu( n, n, dA, ldda, ipiv, info );
if ( *info == MAGMA_ERR_DEVICE_ALLOC ) {
magma_free( dA );
magma_free( dB );
goto CPU_INTERFACE;
}
magma_sgetmatrix( n, n, dA, ldda, A, lda );
if ( *info == 0 ) {
magma_ssetmatrix( n, nrhs, B, ldb, dB, lddb );
magma_sgetrs_gpu( MagmaNoTrans, n, nrhs, dA, ldda, ipiv, dB, lddb, info );
magma_sgetmatrix( n, nrhs, dB, lddb, B, ldb );
}
magma_free( dA );
magma_free( dB );
return *info;
CPU_INTERFACE:
/* If multi-GPU or allocation failed, use CPU interface and LAPACK.
* Faster to use LAPACK for getrs than to copy A to GPU. */
magma_sgetrf( n, n, A, lda, ipiv, info );
if ( *info == 0 ) {
lapackf77_sgetrs( MagmaNoTransStr, &n, &nrhs, A, &lda, ipiv, B, &ldb, info );
}
return *info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgetrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float error;
float *h_A;
magmaFloat_ptr d_A;
magma_int_t *ipiv;
magma_int_t M, N, n2, lda, ldda, info, min_mn;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% version %d\n", (int) opts.version );
if ( opts.check == 2 ) {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |Ax-b|/(N*|A|*|x|)\n");
}
else {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |PA-LU|/(N*|A|)\n");
}
printf("%%========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
gflops = FLOPS_SGETRF( M, N ) / 1e9;
TESTING_MALLOC_CPU( ipiv, magma_int_t, min_mn );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
init_matrix( opts, M, N, h_A, lda );
cpu_time = magma_wtime();
lapackf77_sgetrf( &M, &N, h_A, &lda, ipiv, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_sgetrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
init_matrix( opts, M, N, h_A, lda );
if ( opts.version == 2 ) {
// no pivoting versions, so set ipiv to identity
for (magma_int_t i=0; i < min_mn; ++i ) {
ipiv[i] = i+1;
}
}
magma_ssetmatrix( M, N, h_A, lda, d_A, ldda );
gpu_time = magma_wtime();
if ( opts.version == 1 ) {
magma_sgetrf_gpu( M, N, d_A, ldda, ipiv, &info);
}
else if ( opts.version == 2 ) {
magma_sgetrf_nopiv_gpu( M, N, d_A, ldda, &info);
}
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sgetrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Check the factorization
=================================================================== */
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f)",
(int) M, (int) N, gpu_perf, gpu_time );
}
if ( opts.check == 2 ) {
magma_sgetmatrix( M, N, d_A, ldda, h_A, lda );
error = get_residual( opts, M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else if ( opts.check ) {
magma_sgetmatrix( M, N, d_A, ldda, h_A, lda );
error = get_LU_error( opts, M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf(" --- \n");
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgetrf
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float *h_A, *h_R, *work;
magmaFloat_ptr d_A, dwork;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t N, n2, lda, ldda, info, lwork, ldwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float tmp;
float error, rwork[1];
magma_int_t *ipiv;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
// need looser bound (3000*eps instead of 30*eps) for tests
// TODO: should compute ||I - A*A^{-1}|| / (n*||A||*||A^{-1}||)
opts.tolerance = max( 3000., opts.tolerance );
float tol = opts.tolerance * lapackf77_slamch("E");
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / (N*||A||_F)\n");
printf("=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = ((N+31)/32)*32;
ldwork = N * magma_get_sgetri_nb( N );
gflops = FLOPS_SGETRI( N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_sgetri( &N, NULL, &lda, NULL, &tmp, &lwork, &info );
if (info != 0)
printf("lapackf77_sgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lwork = int( MAGMA_S_REAL( tmp ));
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_CPU( work, float, lwork );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_PIN( h_R, float, n2 );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( dwork, float, ldwork );
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
error = lapackf77_slange( "f", &N, &N, h_A, &lda, rwork ); // norm(A)
/* Factor the matrix. Both MAGMA and LAPACK will use this factor. */
magma_ssetmatrix( N, N, h_A, lda, d_A, ldda );
magma_sgetrf_gpu( N, N, d_A, ldda, ipiv, &info );
magma_sgetmatrix( N, N, d_A, ldda, h_A, lda );
if ( info != 0 )
printf("magma_sgetrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// check for exact singularity
//h_A[ 10 + 10*lda ] = MAGMA_S_MAKE( 0.0, 0.0 );
//magma_ssetmatrix( N, N, h_A, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_sgetri_gpu( N, d_A, ldda, ipiv, dwork, ldwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_sgetri_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_sgetmatrix( N, N, d_A, ldda, h_R, lda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_sgetri( &N, h_A, &lda, ipiv, work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_sgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
blasf77_saxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
error = lapackf77_slange( "f", &N, &N, h_R, &lda, rwork ) / (N*error);
printf( "%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf( "%5d --- ( --- ) %7.2f (%7.2f) ---\n",
(int) N, gpu_perf, gpu_time );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
DSGESV computes the solution to a real system of linear equations
A * X = B, A**T * X = B, or A**H * X = B,
where A is an N-by-N matrix and X and B are N-by-NRHS matrices.
DSGESV first attempts to factorize the matrix in real SINGLE PRECISION
and use this factorization within an iterative refinement procedure
to produce a solution with real DOUBLE PRECISION norm-wise backward error
quality (see below). If the approach fails the method switches to a
real DOUBLE PRECISION factorization and solve.
The iterative refinement is not going to be a winning strategy if
the ratio real SINGLE PRECISION performance over real DOUBLE PRECISION
performance is too small. A reasonable strategy should take the
number of right-hand sides and the size of the matrix into account.
This might be done with a call to ILAENV in the future. Up to now, we
always try iterative refinement.
The iterative refinement process is stopped if
ITER > ITERMAX
or for all the RHS we have:
RNRM < SQRT(N)*XNRM*ANRM*EPS*BWDMAX
where
o ITER is the number of the current iteration in the iterative
refinement process
o RNRM is the infinity-norm of the residual
o XNRM is the infinity-norm of the solution
o ANRM is the infinity-operator-norm of the matrix A
o EPS is the machine epsilon returned by DLAMCH('Epsilon')
The value ITERMAX and BWDMAX are fixed to 30 and 1.0D+00 respectively.
Arguments
---------
@param[in]
trans magma_trans_t
Specifies the form of the system of equations:
- = MagmaNoTrans: A * X = B (No transpose)
- = MagmaTrans: A**T * X = B (Transpose)
- = MagmaConjTrans: A**H * X = B (Conjugate transpose)
@param[in]
n INTEGER
The number of linear equations, i.e., the order of the
matrix A. N >= 0.
@param[in]
nrhs INTEGER
The number of right hand sides, i.e., the number of columns
of the matrix B. NRHS >= 0.
@param[in,out]
dA DOUBLE PRECISION array on the GPU, dimension (ldda,N)
On entry, the N-by-N coefficient matrix A.
On exit, if iterative refinement has been successfully used
(info.EQ.0 and ITER.GE.0, see description below), A is
unchanged. If double precision factorization has been used
(info.EQ.0 and ITER.LT.0, see description below), then the
array dA contains the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
@param[in]
ldda INTEGER
The leading dimension of the array dA. ldda >= max(1,N).
@param[out]
ipiv INTEGER array, dimension (N)
The pivot indices that define the permutation matrix P;
row i of the matrix was interchanged with row IPIV(i).
Corresponds either to the single precision factorization
(if info.EQ.0 and ITER.GE.0) or the double precision
factorization (if info.EQ.0 and ITER.LT.0).
@param[out]
dipiv INTEGER array on the GPU, dimension (N)
The pivot indices; for 1 <= i <= N, after permuting, row i of the
matrix was moved to row dIPIV(i).
Note this is different than IPIV, where interchanges
are applied one-after-another.
@param[in]
dB DOUBLE PRECISION array on the GPU, dimension (lddb,NRHS)
The N-by-NRHS right hand side matrix B.
@param[in]
lddb INTEGER
The leading dimension of the array dB. lddb >= max(1,N).
@param[out]
dX DOUBLE PRECISION array on the GPU, dimension (lddx,NRHS)
If info = 0, the N-by-NRHS solution matrix X.
@param[in]
lddx INTEGER
The leading dimension of the array dX. lddx >= max(1,N).
@param
dworkd (workspace) DOUBLE PRECISION array on the GPU, dimension (N*NRHS)
This array is used to hold the residual vectors.
@param
dworks (workspace) SINGLE PRECISION array on the GPU, dimension (N*(N+NRHS))
This array is used to store the real single precision matrix
and the right-hand sides or solutions in single precision.
@param[out]
iter INTEGER
- < 0: iterative refinement has failed, double precision
factorization has been performed
+ -1 : the routine fell back to full precision for
implementation- or machine-specific reasons
+ -2 : narrowing the precision induced an overflow,
the routine fell back to full precision
+ -3 : failure of SGETRF
+ -31: stop the iterative refinement after the 30th iteration
- > 0: iterative refinement has been successfully used.
Returns the number of iterations
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if info = -i, the i-th argument had an illegal value
- > 0: if info = i, U(i,i) computed in DOUBLE PRECISION is
exactly zero. The factorization has been completed,
but the factor U is exactly singular, so the solution
could not be computed.
@ingroup magma_dgesv_driver
********************************************************************/
extern "C" magma_int_t
magma_dsgesv_gpu(magma_trans_t trans, magma_int_t n, magma_int_t nrhs,
double *dA, magma_int_t ldda,
magma_int_t *ipiv, magma_int_t *dipiv,
double *dB, magma_int_t lddb,
double *dX, magma_int_t lddx,
double *dworkd, float *dworks,
magma_int_t *iter, magma_int_t *info)
{
#define dB(i,j) (dB + (i) + (j)*lddb)
#define dX(i,j) (dX + (i) + (j)*lddx)
#define dR(i,j) (dR + (i) + (j)*lddr)
double c_neg_one = MAGMA_D_NEG_ONE;
double c_one = MAGMA_D_ONE;
magma_int_t ione = 1;
double *dR;
float *dSA, *dSX;
double Xnrmv, Rnrmv;
double Anrm, Xnrm, Rnrm, cte, eps;
magma_int_t i, j, iiter, lddsa, lddr;
/* Check arguments */
*iter = 0;
*info = 0;
if ( n < 0 )
*info = -1;
else if ( nrhs < 0 )
*info = -2;
else if ( ldda < max(1,n))
*info = -4;
else if ( lddb < max(1,n))
*info = -8;
else if ( lddx < max(1,n))
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if ( n == 0 || nrhs == 0 )
return *info;
lddsa = n;
lddr = n;
dSA = dworks;
dSX = dSA + lddsa*n;
dR = dworkd;
eps = lapackf77_dlamch("Epsilon");
Anrm = magmablas_dlange(MagmaInfNorm, n, n, dA, ldda, (double*)dworkd );
cte = Anrm * eps * pow((double)n, 0.5) * BWDMAX;
/*
* Convert to single precision
*/
//magmablas_dlag2s( n, nrhs, dB, lddb, dSX, lddsx, info ); // done inside dsgetrs with pivots
if (*info != 0) {
*iter = -2;
goto FALLBACK;
}
magmablas_dlag2s( n, n, dA, ldda, dSA, lddsa, info );
if (*info != 0) {
*iter = -2;
goto FALLBACK;
}
// factor dSA in single precision
magma_sgetrf_gpu( n, n, dSA, lddsa, ipiv, info );
if (*info != 0) {
*iter = -3;
goto FALLBACK;
}
// Generate parallel pivots
{
magma_int_t *newipiv;
magma_imalloc_cpu( &newipiv, n );
if ( newipiv == NULL ) {
*iter = -3;
goto FALLBACK;
}
swp2pswp( trans, n, ipiv, newipiv );
magma_setvector( n, sizeof(magma_int_t), newipiv, 1, dipiv, 1 );
magma_free_cpu( newipiv );
}
// solve dSA*dSX = dB in single precision
// converts dB to dSX and applies pivots, solves, then converts result back to dX
magma_dsgetrs_gpu( trans, n, nrhs, dSA, lddsa, dipiv, dB, lddb, dX, lddx, dSX, info );
// residual dR = dB - dA*dX in double precision
magmablas_dlacpy( MagmaUpperLower, n, nrhs, dB, lddb, dR, lddr );
if ( nrhs == 1 ) {
magma_dgemv( trans, n, n,
c_neg_one, dA, ldda,
dX, 1,
c_one, dR, 1 );
}
else {
magma_dgemm( trans, MagmaNoTrans, n, nrhs, n,
c_neg_one, dA, ldda,
dX, lddx,
c_one, dR, lddr );
}
// TODO: use MAGMA_D_ABS( dX(i,j) ) instead of dlange?
for( j=0; j < nrhs; j++ ) {
i = magma_idamax( n, dX(0,j), 1) - 1;
magma_dgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1 );
Xnrm = lapackf77_dlange( "F", &ione, &ione, &Xnrmv, &ione, NULL );
i = magma_idamax ( n, dR(0,j), 1 ) - 1;
magma_dgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1 );
Rnrm = lapackf77_dlange( "F", &ione, &ione, &Rnrmv, &ione, NULL );
if ( Rnrm > Xnrm*cte ) {
goto REFINEMENT;
}
}
*iter = 0;
return *info;
REFINEMENT:
for( iiter=1; iiter < ITERMAX; ) {
*info = 0;
// convert residual dR to single precision dSX
// solve dSA*dSX = R in single precision
// convert result back to double precision dR
// it's okay that dR is used for both dB input and dX output.
magma_dsgetrs_gpu( trans, n, nrhs, dSA, lddsa, dipiv, dR, lddr, dR, lddr, dSX, info );
if (*info != 0) {
*iter = -3;
goto FALLBACK;
}
// Add correction and setup residual
// dX += dR --and--
// dR = dB
// This saves going through dR a second time (if done with one more kernel).
// -- not really: first time is read, second time is write.
for( j=0; j < nrhs; j++ ) {
magmablas_daxpycp( n, dR(0,j), dX(0,j), dB(0,j) );
}
// residual dR = dB - dA*dX in double precision
if ( nrhs == 1 ) {
magma_dgemv( trans, n, n,
c_neg_one, dA, ldda,
dX, 1,
c_one, dR, 1 );
}
else {
magma_dgemm( trans, MagmaNoTrans, n, nrhs, n,
c_neg_one, dA, ldda,
dX, lddx,
c_one, dR, lddr );
}
/* Check whether the nrhs normwise backward errors satisfy the
* stopping criterion. If yes, set ITER=IITER > 0 and return. */
for( j=0; j < nrhs; j++ ) {
i = magma_idamax( n, dX(0,j), 1) - 1;
magma_dgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1 );
Xnrm = lapackf77_dlange( "F", &ione, &ione, &Xnrmv, &ione, NULL );
i = magma_idamax ( n, dR(0,j), 1 ) - 1;
magma_dgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1 );
Rnrm = lapackf77_dlange( "F", &ione, &ione, &Rnrmv, &ione, NULL );
if ( Rnrm > Xnrm*cte ) {
goto L20;
}
}
/* If we are here, the nrhs normwise backward errors satisfy
* the stopping criterion, we are good to exit. */
*iter = iiter;
return *info;
L20:
iiter++;
}
/* If we are at this place of the code, this is because we have
* performed ITER=ITERMAX iterations and never satisified the
* stopping criterion. Set up the ITER flag accordingly and follow
* up on double precision routine. */
*iter = -ITERMAX - 1;
FALLBACK:
/* Single-precision iterative refinement failed to converge to a
* satisfactory solution, so we resort to double precision. */
magma_dgetrf_gpu( n, n, dA, ldda, ipiv, info );
if (*info == 0) {
magmablas_dlacpy( MagmaUpperLower, n, nrhs, dB, lddb, dX, lddx );
magma_dgetrs_gpu( trans, n, nrhs, dA, ldda, ipiv, dX, lddx, info );
}
return *info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgetri
*/
int main( int argc, char** argv )
{
TESTING_INIT();
// constants
const float c_zero = MAGMA_S_ZERO;
const float c_one = MAGMA_S_ONE;
const float c_neg_one = MAGMA_S_NEG_ONE;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float *h_A, *h_Ainv, *h_R, *work;
magmaFloat_ptr d_A, dwork;
magma_int_t N, n2, lda, ldda, info, lwork, ldwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float tmp;
float error, rwork[1];
magma_int_t *ipiv;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% N CPU Gflop/s (sec) GPU Gflop/s (sec) ||I - A*A^{-1}||_1 / (N*cond(A))\n");
printf("%%===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
ldwork = N * magma_get_sgetri_nb( N );
gflops = FLOPS_SGETRI( N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_sgetri( &N, NULL, &lda, NULL, &tmp, &lwork, &info );
if (info != 0) {
printf("lapackf77_sgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = magma_int_t( MAGMA_S_REAL( tmp ));
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_CPU( work, float, lwork );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_CPU( h_Ainv, float, n2 );
TESTING_MALLOC_CPU( h_R, float, n2 );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( dwork, float, ldwork );
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
/* Factor the matrix. Both MAGMA and LAPACK will use this factor. */
magma_ssetmatrix( N, N, h_A, lda, d_A, ldda, opts.queue );
magma_sgetrf_gpu( N, N, d_A, ldda, ipiv, &info );
magma_sgetmatrix( N, N, d_A, ldda, h_Ainv, lda, opts.queue );
if (info != 0) {
printf("magma_sgetrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
// check for exact singularity
//h_Ainv[ 10 + 10*lda ] = MAGMA_S_MAKE( 0.0, 0.0 );
//magma_ssetmatrix( N, N, h_Ainv, lda, d_A, ldda, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_sgetri_gpu( N, d_A, ldda, ipiv, dwork, ldwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sgetri_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_sgetri( &N, h_Ainv, &lda, ipiv, work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_sgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
printf( "%5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf( "%5d --- ( --- ) %7.2f (%7.2f)",
(int) N, gpu_perf, gpu_time );
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.check ) {
magma_sgetmatrix( N, N, d_A, ldda, h_Ainv, lda, opts.queue );
// compute 1-norm condition number estimate, following LAPACK's zget03
float normA, normAinv, rcond;
normA = lapackf77_slange( "1", &N, &N, h_A, &lda, rwork );
normAinv = lapackf77_slange( "1", &N, &N, h_Ainv, &lda, rwork );
if ( normA <= 0 || normAinv <= 0 ) {
rcond = 0;
error = 1 / (tol/opts.tolerance); // == 1/eps
}
else {
rcond = (1 / normA) / normAinv;
// R = I
// R -= A*A^{-1}
// err = ||I - A*A^{-1}|| / ( N ||A||*||A^{-1}|| ) = ||R|| * rcond / N, using 1-norm
lapackf77_slaset( "full", &N, &N, &c_zero, &c_one, h_R, &lda );
blasf77_sgemm( "no", "no", &N, &N, &N,
&c_neg_one, h_A, &lda,
h_Ainv, &lda,
&c_one, h_R, &lda );
error = lapackf77_slange( "1", &N, &N, h_R, &lda, rwork );
error = error * rcond / N;
}
bool okay = (error < tol);
status += ! okay;
printf( " %8.2e %s\n",
error, (okay ? "ok" : "failed"));
}
else {
printf( "\n" );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_Ainv );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
