Hello,
I am currently using Intel MKL 2018 Update 3 and it seems that the Sparse BLAS Level 2 and Level 3 routines have been deprecated. I used to use the mkl_?dnscsr function to convert a dense matrix to sparse format. However, I cannot find the corresponding routine from the Intel MKL Inspector-executor Sparse BLAS as mentioned in the reference manual. Can you please help? Thanks much.
Afshin
Dear Sir/Madam,
I am grateful for the development of Intel Visual Fortran Compiler for Windows. It helps my work a lot.
I would like to ask about RCI ISS in the newest Intel MKL solver. I know that in Intel MKL, there is an RCI conjugate gradient solver.
Hence, does it work well with EBE-PCG (element by element preconditioner conjugate gradient) method or can I build my code for EBE-PCG method using Intel MKL solver?
Thank you very much!
Tien-Dat
Hi)
LinX after MKL 11.2.2.010 have error!
w_lpk_p_11.2.2.010.zip c:\test>linpack_xeon64.exe Input data or print help ? Type [data]/help : Number of equations to solve (problem size): 8135 Leading dimension of array: 8136 Number of trials to run: 10 Data alignment value (in Kbytes): 4 Current date/time: Sun Aug 05 19:59:41 2018 CPU frequency: 3.399 GHz Number of CPUs: 1 Number of cores: 6 Number of threads: 12 Parameters are set to: Number of tests: 1 Number of equations to solve (problem size) : 8135 Leading dimension of array : 8136 Number of trials to run : 10 Data alignment value (in Kbytes) : 4 Maximum memory requested that can be used=1066524512, at the size=8135 =================== Timing linear equation system solver =================== Size LDA Align. Time(s) GFlops Residual Residual(norm) Check 8135 8136 4 2.411 148.9028 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.378 150.9828 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.262 158.7518 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.329 154.1468 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.496 143.8229 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.258 159.0080 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.358 152.2624 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.323 154.5865 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.427 147.9225 4.503162e-011 2.387924e-002 pass 8135 8136 4 2.346 153.0549 4.503162e-011 2.387924e-002 pass Performance Summary (GFlops) Size LDA Align. Average Maximal 8135 8136 4 152.3442 159.0080 Residual checks PASSED End of tests c:\test>
w_mklb_p_11.1.3.005.zip c:\test>linpack_xeon64.exe Input data or print help ? Type [data]/help : Number of equations to solve (problem size): 8135 Leading dimension of array: 8136 Number of trials to run: 10 Data alignment value (in Kbytes): 4 Current date/time: Sun Aug 05 19:56:41 2018 CPU frequency: 3.399 GHz Number of CPUs: 1 Number of cores: 6 Number of threads: 6 Parameters are set to: Number of tests: 1 Number of equations to solve (problem size) : 8135 Leading dimension of array : 8136 Number of trials to run : 10 Data alignment value (in Kbytes) : 4 Maximum memory requested that can be used=529719136, at the size=8135 =================== Timing linear equation system solver =================== Size LDA Align. Time(s) GFlops Residual Residual(norm) Check 8135 8136 4 2.225 161.3400 7.189260e-011 3.812300e-002 pass 8135 8136 4 2.094 171.4791 7.189260e-011 3.812300e-002 pass 8135 8136 4 1.935 185.5586 7.189260e-011 3.812300e-002 pass 8135 8136 4 2.024 177.3962 7.189260e-011 3.812300e-002 pass 8135 8136 4 2.230 160.9874 7.189260e-011 3.812300e-002 pass 8135 8136 4 2.053 174.8977 7.189260e-011 3.812300e-002 pass 8135 8136 4 2.233 160.7589 7.189260e-011 3.812300e-002 pass 8135 8136 4 1.717 209.0689 7.189260e-011 3.812300e-002 pass 8135 8136 4 2.127 168.7899 7.189260e-011 3.812300e-002 pass 8135 8136 4 2.128 168.7352 7.189260e-011 3.812300e-002 pass Performance Summary (GFlops) Size LDA Align. Average Maximal 8135 8136 4 173.9012 209.0689 Residual checks PASSED End of tests c:\test>
w_mklb_p_2018.2.010.zip c:\test>linpack_xeon64.exe Input data or print help ? Type [data]/help : Number of equations to solve (problem size): 8135 Leading dimension of array: 8136 Number of trials to run: 10 Data alignment value (in Kbytes): 4 Current date/time: Sun Aug 05 20:09:57 2018 CPU frequency: 3.339 GHz Number of CPUs: 1 Number of cores: 6 Number of threads: 6 Parameters are set to: Number of tests: 1 Number of equations to solve (problem size) : 8135 Leading dimension of array : 8136 Number of trials to run : 10 Data alignment value (in Kbytes) : 4 Maximum memory requested that can be used=529719136, at the size=8135 =================== Timing linear equation system solver =================== Size LDA Align. Time(s) GFlops Residual Residual(norm) Check 8135 8136 4 2.231 160.9364 6.790949e-011 3.601085e-002 pass 8135 8136 4 2.084 172.3152 6.790949e-011 3.601085e-002 pass 8135 8136 4 2.117 169.6332 6.790949e-011 3.601085e-002 pass 8135 8136 4 2.132 168.3802 6.790949e-011 3.601085e-002 pass 8135 8136 4 2.069 173.5236 6.790949e-011 3.601085e-002 pass 8135 8136 4 2.000 179.5242 6.790949e-011 3.601085e-002 pass 8135 8136 4 2.109 170.2193 6.790949e-011 3.601085e-002 pass 8135 8136 4 2.018 177.8865 6.790949e-011 3.601085e-002 pass 8135 8136 4 1.964 182.8301 6.790949e-011 3.601085e-002 pass 8135 8136 4 2.050 175.0993 6.790949e-011 3.601085e-002 pass Performance Summary (GFlops) Size LDA Align. Average Maximal 8135 8136 4 173.0348 182.8301 Residual checks PASSED End of tests
If run w_mklb_p_2018.2.010.zip as 12 Thread or as 5, 9,10, 11,12 Threads, in any shell will be this:
Intel(R) LINPACK 64-bit data - LinX 0.6.5
Current date/time: Sat Aug 18 00:36:31 2018
CPU frequency: 3.398 GHz
Number of CPUs: 1
Number of cores: 6
Number of threads: 12
Parameters are set to:
Number of tests: 1
Number of equations to solve (problem size) : 8135
Leading dimension of array : 8136
Number of trials to run : 10
Data alignment value (in Kbytes) : 4
Maximum memory requested that can be used=529657696, at the size=8135
=================== Timing linear equation system solver ===================
Size LDA Align. Time(s) GFlops Residual Residual(norm) Check
8135 8136 4 1.956 183.5389 6.526518e-011 3.460863e-002 pass
8135 8136 4 2.187 164.2005 6.619277e-011 3.510051e-002 pass
8135 8136 4 2.872 125.0139 6.526518e-011 3.460863e-002 pass
8135 8136 4 2.100 170.9365 5.353940e-011 2.839072e-002 pass
8135 8136 4 2.231 160.9549 7.749093e-011 4.109167e-002 pass
8135 8136 4 2.702 132.8717 6.938300e-011 3.679222e-002 pass
8135 8136 4 2.145 167.3753 6.537959e-011 3.466930e-002 pass
8135 8136 4 2.661 134.9388 6.526518e-011 3.460863e-002 pass
8135 8136 4 2.043 175.7050 6.526518e-011 3.460863e-002 pass
8135 8136 4 2.353 152.5804 6.553075e-011 3.474946e-002 pass
Performance Summary (GFlops)
Size LDA Align. Average Maximal
8135 8136 4 156.8116 183.5389
Residual checks PASSED
End of tests
Please, fix it!
In first I random , pointed out 524288 as data aligment
Hello,
I'm using Sparse BLAS routines in MKL.
For matrix-matrix multiplications using the mkl_sparse_spmm routine, I created matrix handles, and then execute the routine.
But, how can I get the result in the original CSR format? I mean, the mkl_sparse_spmm gives the result in the form of matrix handle (pointer),
so I need a further step to get the final result in CSR format.
Thanks.
Hello,
I would like to report a runtime error in feast zfeast_hcsrgv routine.
I have utlized the FEAST solver to solve generalized Hermitian eigenvalue problems using the zfeast_hcsrgv routine, but encountered an access violation during running the codes.
I couldn't find any problem in my input data.
An example code is enclosed, and both of two input files, "read_mat_info_1" and "read_mat_info_2", result in runtime errors during calling zfeast_hcsrgv.
Thanks.
| Attachment | Size |
|---|---|
Download feast_example_runtime_error_2.zip | 2.61 MB |
I have a problem with solving system of linear equations using Pardiso. I have application in which one of the processes loads dll where system of linear equations Ax=B is solved using Pardiso. There are two cases:
- First case - Application is started. Matrix A and B are populated with some values, and after that matrix A is factorized so that equation Ax=B can be solved. After resolving, I get solution x and write it to the file. Values in x are very suspicious, there is a lot of values which are larger than 10^4. They will be used later for some calculation in iterations.
- Second case - After execution in first case, I kill the process (in Task Manager) which loads dll (in which equations are solved), and process is restarted and the same dll is loaded. After executing my application again, same equation is solved, but x is not the same (and this new values seem better). I don't get it why solution is not the same in both cases when everything about matrix is the same?
I have logged all the matrices and vectors (and attached them). What can be seen is that when I compare A, x and B there is a difference in x, and A and B are the same in both cases. How can solver for two same inputs give two different solutions? I don't now what killing the process and loading dll did, but after that x is not same when equation is solved. Maybe there is some problem with memory or something similar.
Matrix A is sparse matrix and is indefinite. Matrix type is -2 and Pardiso call is as following:
call pardiso(matrix%handle, 1, 1, & !handle
matrix%matrixType, & !Matrix type
33, & !3 = solve
matrix%numRows, & !Dimension
matrix%values, & !Values
matrix%firstInRow, & !Row cummulative
matrix%columns, & !Column indices
[0], & !permutation (none)
1, & !Number of RHS equations (1)
opt, & !Options (default)
0, & !Message level (none)
rhs, & !RHS
solution, & !LHS
err)
Does anyone have idea what happened or how could I try to debug the problem?
| Attachment | Size |
|---|---|
| DownloadMatrices.7z | 13.88 KB |
I compile my program and it gives the following error:
forrtl: severe (174): SIGSEGV, segmentation fault occurred
I isolated the problem and it is occurring at the call of the mkl_sparse_d_mv function of the following function:
function dVdt(self, t, V)
class(MotorUnitPool), intent(inout) :: self
real(wp), intent(in) :: t
real(wp), intent(in) :: V(:)
real(wp), dimension(self%totalNumberOfCompartments) :: dVdt
integer :: i, j, stat
real(wp), dimension(:), allocatable :: matInt
allocate(matInt(self%totalNumberOfCompartments))
do i = 1, self%MUnumber
do j = 1, self%unit(i)%compNumber
self%iIonic((i-1)*self%unit(i)%compNumber+j) = self%unit(i)%Compartments(j)%computeCurrent(t, V((i-1)*self%unit(i)%compNumber+j))
end do
end do
stat = mkl_sparse_d_mv(self%spOperation, &
self%spAlpha, &
self%GSp, &
self%spDescr, &
V, &
self%spBeta, &
matInt)
dVdt = (self%iIonic + matInt + self%iInjected + self%EqCurrent_nA) * self%capacitanceInv
end function
I compile with gfortran without any problems. As compiler options to the ifort compiler i tried to use the option -heap_arrays, whith no success.
This function is part of a bigger program, with some files. If you want to see the whole software, you can find it in https://github.com/rnwatanabe/projectFR .
Thanks in advance,
Renato Watanabe
Hi all,
I would like to use PZGESVD of SCALAPACK to solve my problem distributedly. Before using PZGESVD, the matrix should be distributed to all the processes involved. Therefore, I want to test the usage of PDGEMR2D in advance.
However, I have some problem in using PDGEMR2D. It generated an error from which I cannot figure the cause.
The error is:
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
??MR2D:Bad submatrix:i=-1,j=-1,m=50,n=10,M=50,N=10
Assertion failed in c:\bt\479\private\mpich2\src\pm\smpd\smpd_handle_command.cpp(640): proc != 0
unable to read the cmd header on the left child context, Other MPI error, error stack:
ReadFailed(1298): An existing connection was forcibly closed by the remote host. (errno 10054).
Aborting: mpiexec on TEMFPC1005 failed to communicate with smpd on TEMFPC1005
Other MPI error, error stack:
ReadFailed(1298): An existing connection was forcibly closed by the remote host. (errno 10054)
The Code is:
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <iostream>
#include <iomanip>
#include <string>
#include <fstream>
#include <sstream>
#include <complex>
#include <algorithm>
#include <vector>
//#define MKL_Complex8 std::complex<float>
//#define MKL_Complex16 std::complex<double>
#include <mkl_blacs.h>
#include <mkl_scalapack.h>
#include <mkl_pblas.h>
#include <mkl.h>
#include "petsc.h"
#define MAX(a,b)((a)<(b)?(b):(a))
#define MIN(a,b)((a)>(b)?(b):(a))
using namespace std;
static char help[] = "Test SCALAPACK";
int main(int argc, char **argv)
{
PetscErrorCode ierr;
const MKL_INT intmkl_negone = -1, intmkl_zero = 0;
MKL_INT intmkl_rank, intmkl_size, intmkl_info, intmkl_ctxt, intmkl_nProcRows, intmkl_nProcCols, intmkl_myRow, intmkl_myCol;
MKL_INT intmkl_MA, intmkl_NA, intmkl_MBA, intmkl_NBA, intmkl_lldA, intmkl_nRowProc, intmkl_nColProc;
MKL_INT intmkl_MB, intmkl_NB, intmkl_MBB, intmkl_NBB, intmkl_lldB;
MKL_INT descA[9], descB[9];
PetscInt int_size, int_rank, int_numLocalRowMatA, int_numGlobalRowMatA, int_numGlobalColMatA;
Mat mat_A;
double *doubleAr_A, *doubleAr_B;
/* MPI initialization */
ierr = PetscInitialize(&argc, &argv, (char*)0, help); CHKERRQ(ierr);
MPI_Comm_size(PETSC_COMM_WORLD, &int_size);
MPI_Comm_rank(PETSC_COMM_WORLD, &int_rank);
/* Generate a random matrix */
int_numLocalRowMatA = 5;
int_numGlobalRowMatA = int_numLocalRowMatA*int_size;
int_numGlobalColMatA = 10;
/* initialize blacs */
BLACS_PINFO(&intmkl_rank, &intmkl_size);
intmkl_nProcRows = intmkl_size;
intmkl_nProcCols = 1;
BLACS_GET(&intmkl_negone, &intmkl_zero, &intmkl_ctxt);
BLACS_GRIDINIT(&intmkl_ctxt, "C", &intmkl_nProcRows, &intmkl_nProcCols);
BLACS_GRIDINFO(&intmkl_ctxt, &intmkl_nProcRows, &intmkl_nProcCols, &intmkl_myRow, &intmkl_myCol);
/* compute precise length of local pieces and allocate array on each process for parts of distributed matrices */
intmkl_MA = (MKL_INT)int_numGlobalRowMatA;
intmkl_NA = (MKL_INT)int_numGlobalColMatA;
intmkl_MBA = (MKL_INT)int_numLocalRowMatA;
intmkl_NBA = (MKL_INT)int_numGlobalColMatA;
intmkl_nRowProc = NUMROC(&intmkl_MA, &intmkl_MBA, &intmkl_myRow, &intmkl_zero, &intmkl_nProcRows);
intmkl_nColProc = NUMROC(&intmkl_NA, &intmkl_NBA, &intmkl_myCol, &intmkl_zero, &intmkl_nProcCols);
intmkl_lldA = MAX(1, intmkl_nRowProc);
DESCINIT(descA, &intmkl_MA, &intmkl_NA, &intmkl_MBA, &intmkl_NBA, &intmkl_zero, &intmkl_zero, &intmkl_ctxt, &intmkl_lldA, &intmkl_info);
std::cout << "MA = "<< intmkl_MA << "; NA = "<< intmkl_NA << "; intmkl_nRowProc = "<< intmkl_nRowProc << "; intmkl_nColProc = "<< intmkl_nColProc << "; lldA = "<< intmkl_lldA << std::endl;
doubleAr_A = (double*)mkl_calloc(intmkl_nRowProc*intmkl_nColProc, sizeof(double), 64);
for (int int_cnt1 = 0; int_cnt1 < intmkl_nRowProc*intmkl_nColProc; int_cnt1++) doubleAr_A[int_cnt1] = 1.0;
/* compute precise length of local pieces and allocate array on each process for parts of distributed matrices */
intmkl_MB = (MKL_INT)int_numGlobalRowMatA;
intmkl_NB = (MKL_INT)int_numGlobalColMatA;
intmkl_MBB = (MKL_INT)int_numLocalRowMatA;
intmkl_NBB = (MKL_INT)int_numGlobalColMatA;
intmkl_nRowProc = NUMROC(&intmkl_MB, &intmkl_MBB, &intmkl_myRow, &intmkl_zero, &intmkl_nProcRows);
intmkl_nColProc = NUMROC(&intmkl_NB, &intmkl_NBB, &intmkl_myCol, &intmkl_zero, &intmkl_nProcCols);
intmkl_lldB = MAX(1, intmkl_nRowProc);
DESCINIT(descB, &intmkl_MB, &intmkl_NB, &intmkl_MBB, &intmkl_NBB, &intmkl_zero, &intmkl_zero, &intmkl_ctxt, &intmkl_lldB, &intmkl_info);
std::cout << "MB = "<< intmkl_MB << "; NB = "<< intmkl_NB << "; intmkl_nRowProc = "<< intmkl_nRowProc << "; intmkl_nColProc = "<< intmkl_nColProc << "; lldB = "<< intmkl_lldB << std::endl;
doubleAr_B = (double*)mkl_calloc(intmkl_nRowProc*intmkl_nColProc, sizeof(double), 64);
/* copy value from matrix A to matrix B */
PDGEMR2D(&intmkl_MA, &intmkl_NA, doubleAr_A, &intmkl_zero, &intmkl_zero, descA, doubleAr_B, &intmkl_zero, &intmkl_zero, descB, &intmkl_ctxt);
/* destroy variables */
mkl_free(doubleAr_A);
mkl_free(doubleAr_B);
/* finalize blacs */
BLACS_GRIDEXIT(&intmkl_ctxt);
/* finalize petsc*/
PetscFinalize();
}
I would appreciate if you can have any advice or comment on solving this problem.
Thank a lot,
Hello folks,
I've a strange performance problem with PARDISO on Windows. Before I open a support call I'll hope to get some feedback in this forum.
I'm using Intel® Parallel Studio XE 2018 Update 3 Composer Edition for Fortran Windows*, Version 18.0.0040.
I have noticed that parallel processing in PARDISO in MKL version 2018.0.3 does not work at all and processing with only one thread is significantly slower than in version 2016.
Attached I've a small C++ test program and sample data to solve a small system multiple time.
When I run the program using the MKL DLLs from version 2018.0.3 I get following result:
>Release\pardiso.exe _data\mat.mm _data\b.mm Intel(R) Math Kernel Library Version 2018.0.3 Product Build 20180406 for 32-bit applications Solving matrix file _data\mat.mm with vector data _data\b.mm. Data: rows=445, cols=445, values=1339 MKL threads: 6 Performance: Loops=10000, Time=2.785514 sec
And now the funny stuff starts. The same program executed with MKL DLLs from version 2016 (11.3.3) create the following result:
>Release\pardiso.exe _data\mat.mm _data\b.mm Intel(R) Math Kernel Library Version 11.3.3 Product Build 20160413 for 32-bit applications Solving matrix file _data\mat.mm with vector data _data\b.mm. Data: rows=445, cols=445, values=1339 MKL threads: 6 Performance: Loops=10000, Time=1.171534 sec
And it's gonna get worse. The new PARDISO version 2018.0.3 uses a big amount of CPU time for multiple threads but it is slower compared with execution with only one single thread!
According to my understanding I've configured all stuff correct. And as it can be seen, using the old MKL stuff from 2016 it works fine.
| Attachment | Size |
|---|---|
| Downloadsample.zip | 28.61 KB |
I was trying to use DGESV to solve a set of linear equations, an MKL routine.
but it tells me its an unknown entry point.
Steve mentioned using a LINKER option, but I don't see any way to do that from Visual Studio.
I looked thru all the drop down menus.
at least if its there, it is sure deeply buried somewhere.
No problem using the routine, if I can attach it somehow.
any clues ?
n=INT_MAX: Intel MKL TRIG TRANSFORMS ERROR: Fatal error (error message=Intel MKL DFTI ERROR: Inconsistent configuration parameters
n=INT_MAX-1: Finishes fine
n=INT_MAX+1: Segmentation fault
I am using ILP64. Tracking through the FFTW interface, I I know the segfault occurs in s_init_trig_transform().
Compile line:
icc transform.cpp -std=c++11 -O3 -DMKL_ILP64 -xcore-avx2 -I${MKLROOT}/include/fftw -L${MKLROOT}/lib/intel64 -lmkl_intel_ilp64 -lmkl_intel_thread -lmkl_core -liomp5 -lpthread -lm -ldl -qopenmp
Minimal reproducible example:
#include <fftw3.h>
#include <fftw3_mkl.h>
#include <random>
#include <omp.h>
int main()
{
MKL_INT n = 2147483647; //INT_MAX
fftwf_init_threads();
float* inout = (float*)fftwf_malloc(sizeof(float)*n);
//Initialize inout array with random values
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_real_distribution<float> dist(0.0f, 1.0f);
#pragma omp parallel for
for (size_t i=0; i<n; ++i) {
inout[i] = dist(gen);
}
//FFTW Guru parameters
const fftwf_r2r_kind kind = FFTW_RODFT00;
int rank=1;
int howmany_rank=0; //Not used by MKL
fftwf_iodim64* howmany_dims; //Not used by MKL
fftwf_iodim64* dims = (fftwf_iodim64*)fftwf_malloc(rank*sizeof(fftwf_iodim64));
dims[0].n=n;
dims[0].is=1;
dims[0].os=1;
fftwf_plan_with_nthreads(omp_get_max_threads());
//in-place real-to-real sine transform
fftwf_plan p = fftwf_plan_guru64_r2r(rank, dims, howmany_rank, howmany_dims, inout, inout, &kind, FFTW_MEASURE);
fftwf_execute(p);
fftwf_destroy_plan(p);
fftwf_free(inout);
fftwf_cleanup_threads();
return 0;
}
Dear all,
I would like to ask for your help to solve this error when I used Pardiso for my Finite element analysis.
I am using Intel® Parallel Studio XE 2017 update 6 for Windows with Intel Fortran compiler integrated in Visual Studio 2015. My laptop has 8 Gb RAM, 4 processors 2.0 GHz.
I do not know why when I solved the equation Ax=b where A size [90,000 x 90,000] with the setting of Pardiso solver as the attached picture 3, there was an error relating to stack overflow as the attached pictures 1&2. When running only 45% of RAM memory was used.
Please help me find the reason and the way to fix this problem.
Thank you very much.
| Attachment | Size |
|---|---|
| Downloade.JPG | 136.82 KB |
| Downloaderror2.JPG | 218.27 KB |
| Downloaderror3.JPG | 52.85 KB |
Hi,
I am struggling with the recompilation of R using the Intel MKL. I have identified the problem to the following:
gfortran -c testf.f
gcc -c test.c
gcc -L${MKL_LIB_PATH} -o test test.o testf.o -lmkl_intel_ilp64 -lmkl_core -lmkl_sequential
When running the program, I get a sigfault 11 error message.
I tried debugging with gdb but couldn't really pinpoint the problem. It seems all the good libraries are being used (otool -L) points to the MKL libraries...
I am running out of options...
thanks,
Bernd
I went back to an earlier version (VS 2010) which does the Fortran source code routines fine -
But I still cannot make it attach an entry point routine in the MKL library.
Steve mentioned an option in the LINKER (project properties) to tell it to attach the MKL library,
but for some reason, that option is not available (at least not to me), I cant find it anywhere.
are you saying I should give up on Visual Studio entirely ?
I was kinda commited to using that for program development and debugging up till now - - - -
I am trying to determine whether the Intel Direct Sparse Solver for Clusters is a good parallel solver for our application. I have implemented the sparse solver in Fortran to solve a linear FEA problem. For the call to the sparse solver, I am seeing speed-up with increasing number of MPI processes, but not seeing good speed-up with increasing number of threads per MPI process.
In this case, the A matrix is generated from a finite-difference-type grid in distributed format (DCSR). Node ordering is such that distributing the matrix results in gaps in the sparse matrix storage of each part, similar to the example given here: https://software.intel.com/en-us/articles/intel-math-kernel-library-parallel-direct-sparse-solver-for-clusters
This case is a linear time-domain problem so we factorize the matrix once and then solve it many times with evolving boundary conditions. Should I expect to see good scaling over MPI processes and OpenMP threads in the solve phase of the direct sparse solver for clusters?
I have benchmarked a 10 million DOF model on a linux cluster with the number of MPI processes ranging from 2 to 128, with 1 process per hardware node, and the number of OpenMP threads per node ranging from 2 to 16. I see speed-up in increasing the number of MPI processes up to about 32 processes, but very little improvement in using more OpenMP threads. I see about the same speed-up on 8 or 16 threads as I see on 2 threads.
I am using the following iparm variables:
iparm(1) = 1 !no default values
iparm(2) = 2
iparm(10) = 8
iparm(18) = -1
iparm(27) = 0
iparm(28) = 1
iparm(40) = 2
iparm(41) = ibegin
iparm(42) = iend
iparm(60) = 1
Is the Direct Sparse Solver for Clusters suitable here and what issues should I look at to try to improve scaling with number of OpenMP threads? Thanks for your help.
hello,
i am new to computing on windows operating systems. i have pgfortran 18.4 installed and MKL 2018.3.210 installed, including the mkl_pgi_thread library.
i am trying to statically link to the libraries needed to use PARDISO_D_64 (mkl_intel_ilp64.lib mkl_pgi_thread.lib mkl_core.lib) but the best outcome i can get is:
pgfortran -i8 -pgf90libs -mp -o GEMINI GEMINI.obj MD.obj MR.obj M1.obj M2.obj MW.obj -I"C:\IntelMKL\compilers_and_libraries_2018.3.210\windows\mkl\include" -LC:\Inte
lMKL\compilers_and_libraries_2018.3.210\windows\mkl\lib\intel64_win\mkl_intel_ilp64.lib -LC:\IntelMKL\compilers_and_libraries_2018.3.210\windows\mkl\lib\intel64_win\mk
l_pgi_thread.lib -LC:\IntelMKL\compilers_and_libraries_2018.3.210\windows\mkl\lib\intel64_win\mkl_core.lib -LC:\IntelMKL\compilers_and_libraries_2018.3.210\windows\com
piler\lib\intel64_win
GEMINI.obj : error LNK2019: unresolved external symbol mkl_set_num_threads_ referenced in function MAIN_
GEMINI.obj : error LNK2019: unresolved external symbol mkl_set_dynamic_ referenced in function MAIN_
M1.obj : error LNK2019: unresolved external symbol pardiso_d_64_ referenced in function m1_s1_
GEMINI.exe : fatal error LNK1120: 3 unresolved externals
the main problem, i hope, is that i just don't know how to link to multiple libraries using pgfortran on windows. i previously compiled and executed this same code, without issue, on linux mint using gfortran.
i MKL link line advisor also states that MKL support in windows with pgi fortran is "limited." what does that mean? could it be that i just cannot use pgfortran to link to the pardiso routines? or, could it be that i need to add the "mkl service" to my code, which i found to not be necessary when i was working on linux with gfortran.
I've generated and received a license file for my downloaded free version of the Math Kernel Library, and I understand I can install it on other computers for my use only on other machines.
I have a work requirement (actually the primary reason for obtaining the MKL) to install the software for development in a closed area, upon a Windows 7 machine that is completely non-networked for security reasons. It does not have a Network Interface Card (NIC), and therefore has no MAC address ("Host ID").
When I attempt to install MKL I get the following error: "Package signature verification failed. Click Help for details."
Clicking "Help" attempts to open a web page, which of course cannot connect to anything.
Is it possible to do what I am trying to do?
Thanks
Hi everyone,
I've compiled HPL with Intel's MPI and Intel MKL but I'm confused about the best way to run it on a single node with two processors (12 cores each). Should I be using mpirun even though I only have a single node? If I try to set my PxQ grid to be 24 then I get an error saying that there are not enough processes. This is alleviated when I use mpirun. But I'm worried about communication overhead. Is it really necessary to use MPI when I've got a single node but with two processors? One of the other things I was confused about was what I should set the mkl option as when linking the MKL Library (cluster vs parallel). I've chosen parallel for the time being.
Any thoughts?
Thanks for Reading
Hi All;
Quick Question; How can i use GMRES for a preconditioned very large linear system without storing the entire coefficient matrix?. As i understand; GMRES asks for the RHS vector, the entire coefficient matrix and the Preconditioner. But for the memory limitation and for the compuational cost i will not be able to assemble and store the entire coefficient matrix. Any thought is very appreciated!.
Hello together,
I am a PhD student researching in the area of parallel programming. In my next research paper, I aim to present some high-performance (OpenCL) implementations for the Basic Linear Algebra Subroutines (BLAS) -- especially for the matrix multiplication routine GEMM -- on matrix sizes as used in the area of deep learning; my targeted hardware is Intel Xeon CPU. To strengthen my evaluation, I want to compare to the fastest state-of-the-art implementation for BLAS that targets Intel Xeon CPU.
My question is: Which is the currently fastest BLAS implementation for Intel Xeon CPU on matrix sizes as used in deep learning -- the Intel Math Kernel Library (MKL)?
Many thanks in advance.
Best,
Richard