USERS_GUIDE.md 25.5 KB
Newer Older
Andreas Marek's avatar
Andreas Marek committed
1
## Users guide for the *ELPA* library ##
Andreas Marek's avatar
Andreas Marek committed
2 3 4 5 6 7 8

This document provides the guide for using the *ELPA* library in user applications.

### Online and local documentation ###

Local documentation (via man pages) should be available (if *ELPA* has been installed with the documentation):

9
For example "man elpa_get_communicators" should provide the documentation for the *ELPA* function which sets
Andreas Marek's avatar
Andreas Marek committed
10 11
the necessary communicators.

12
Also a [online doxygen documentation] (http://elpa.mpcdf.mpg.de/html/Documentation/ELPA-2016.05.004/html/index.html)
Andreas Marek's avatar
Andreas Marek committed
13 14
for each *ELPA* release is available.

Andreas Marek's avatar
Andreas Marek committed
15
### General concept of the *ELPA* library ###
Andreas Marek's avatar
Andreas Marek committed
16

Andreas Marek's avatar
Andreas Marek committed
17 18 19 20
The *ELPA* library consists of two main parts:
- *ELPA 1stage* solver
- *ELPA 2stage* solver

21
Both variants of the *ELPA* solvers are available for real or complex singe and double precision valued matrices.
Andreas Marek's avatar
Andreas Marek committed
22

23
Thus *ELPA* provides the following user functions (see man pages or [online] (http://elpa.mpcdf.mpg.de/html/Documentation/ELPA-2016.05.004/html/index.html) for details):
Andreas Marek's avatar
Andreas Marek committed
24

25 26 27 28 29
- elpa_get_communicators                        : set the row / column communicators for *ELPA*
- elpa_solve_evp_complex_1stage_{single|double} : solve a {single|double} precision complex eigenvalue proplem with the *ELPA 1stage* solver
- elpa_solve_evp_real_1stage_{single|double}    : solve a {single|double} precision real eigenvalue proplem with the *ELPA 1stage* solver
- elpa_solve_evp_complex_2stage_{single|double} : solve a {single|double} precision complex eigenvalue proplem with the *ELPA 2stage* solver
- elpa_solve_evp_real_2stage_{single|double}    : solve a {single|double} precision real eigenvalue proplem with the *ELPA 2stage* solver
Andreas Marek's avatar
Andreas Marek committed
30

31 32
*NEW*

33 34
- elpa_solve_evp_real_{single|double}      : driver for the {single|double} precision real *ELPA 1stage* or *ELPA 2stage* solver
- elpa_solve_evp_complex_{single|double}   : driver for the {single|double} precision complex *ELPA 1stage* or *ELPA 2stage* solver
35 36 37



Andreas Marek's avatar
Andreas Marek committed
38 39 40 41 42 43 44 45 46 47
Furthermore *ELPA* provides the utility binary "print_available_elpa2_kernels": it tells the user
which *ELPA 2stage* compute kernels have been installed and which default kernels are set

If you want to solve an eigenvalue problem with *ELPA*, you have to decide whether you
want to use *ELPA 1stage* or *ELPA 2stage* solver. Normally, *ELPA 2stage* is the better
choice since it is faster, but there a matrix dimensions where *ELPA 1stage* is supperior.

Independent of the choice of the solver, the concept of calling *ELPA* is always the same:

#### MPI version of *ELPA* ####
Andreas Marek's avatar
Andreas Marek committed
48 49 50 51 52

In this case, *ELPA* relies on a BLACS distributed matrix.
To solve a Eigenvalue problem of this matrix with *ELPA*, one has

1. to include the *ELPA* header (C case) or module (Fortran)
53 54
2. to create row and column MPI communicators for ELPA (with "elpa_get_communicators")
3. to call to the *ELPA driver* or directly call *ELPA 1stage* or *ELPA 2stage* for the matrix.
Andreas Marek's avatar
Andreas Marek committed
55

Andreas Marek's avatar
Andreas Marek committed
56
Here is a very simple MPI code snippet for using *ELPA 1stage*: For the definition of all variables
Andreas Marek's avatar
Andreas Marek committed
57 58
please have a look at the man pages and/or the online documentation (see above). A full version
of a simple example program can be found in ./test_project/src.
Andreas Marek's avatar
Andreas Marek committed
59 60 61 62 63


   ! All ELPA routines need MPI communicators for communicating within
   ! rows or columns of processes, these are set in get_elpa_communicators

Andreas Marek's avatar
Andreas Marek committed
64
   success = elpa_get_communicators(mpi_comm_world, my_prow, my_pcol, &
Andreas Marek's avatar
Andreas Marek committed
65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84
                                    mpi_comm_rows, mpi_comm_cols)

   if (myid==0) then
     print '(a)','| Past split communicator setup for rows and columns.'
   end if

   ! Determine the necessary size of the distributed matrices,
   ! we use the Scalapack tools routine NUMROC for that.

   na_rows = numroc(na, nblk, my_prow, 0, np_rows)
   na_cols = numroc(na, nblk, my_pcol, 0, np_cols)

   !-------------------------------------------------------------------------------
   ! Calculate eigenvalues/eigenvectors

   if (myid==0) then
     print '(a)','| Entering one-step ELPA solver ... '
     print *
   end if

85
   success = elpa_solve_evp_real_1stage_{single|double} (na, nev, a, na_rows, ev, z, na_rows, nblk, &
Andreas Marek's avatar
Andreas Marek committed
86
                                   matrixCols, mpi_comm_rows, mpi_comm_cols)
Andreas Marek's avatar
Andreas Marek committed
87 88 89 90 91 92 93

   if (myid==0) then
     print '(a)','| One-step ELPA solver complete.'
     print *
   end if


Andreas Marek's avatar
Andreas Marek committed
94
#### Shared-memory version of *ELPA* ####
Andreas Marek's avatar
Andreas Marek committed
95

Andreas Marek's avatar
Andreas Marek committed
96
If the *ELPA* library has been compiled with the configure option "--with-mpi=0",
Andreas Marek's avatar
Andreas Marek committed
97 98 99 100
no MPI will be used.

Still the **same** call sequence as in the MPI case can be used (see above).

Andreas Marek's avatar
Andreas Marek committed
101
#### Setting the row and column communicators ####
Andreas Marek's avatar
Andreas Marek committed
102

Andreas Marek's avatar
Andreas Marek committed
103 104 105
SYNOPSIS
   FORTRAN INTERFACE
       use elpa1
106

107
       success = elpa_get_communicators (mpi_comm_global, my_prow, my_pcol, mpi_comm_rows, mpi_comm_cols)
108

Andreas Marek's avatar
Andreas Marek committed
109 110 111 112 113 114 115 116 117 118 119
       integer, intent(in)   mpi_comm_global:  global communicator for the calculation
       integer, intent(in)   my_prow:          row coordinate of the calling process in the process grid
       integer, intent(in)   my_pcol:          column coordinate of the calling process in the process grid
       integer, intent(out)  mpi_comm_row:     communicator for communication within rows of processes
       integer, intent(out)  mpi_comm_row:     communicator for communication within columns of processes

       integer               success:          return value indicating success or failure of the underlying MPI_COMM_SPLIT function

   C INTERFACE
       #include "elpa_generated.h"

120
       success = elpa_get_communicators (int mpi_comm_world, int my_prow, my_pcol, int *mpi_comm_rows, int *Pmpi_comm_cols);
Andreas Marek's avatar
Andreas Marek committed
121 122 123 124 125 126 127 128 129 130 131 132

       int mpi_comm_global:  global communicator for the calculation
       int my_prow:          row coordinate of the calling process in the process grid
       int my_pcol:          column coordinate of the calling process in the process grid
       int *mpi_comm_row:    pointer to the communicator for communication within rows of processes
       int *mpi_comm_row:    pointer to the communicator for communication within columns of processes

       int  success:         return value indicating success or failure of the underlying MPI_COMM_SPLIT function


#### Using *ELPA 1stage* ####

133
After setting up the *ELPA* row and column communicators (by calling elpa_get_communicators),
Andreas Marek's avatar
Andreas Marek committed
134 135 136 137 138
only the real or complex valued solver has to be called:

SYNOPSIS
   FORTRAN INTERFACE
       use elpa1
139
       success = elpa_solve_evp_real_1stage_{single|double} (na, nev, a(lda,matrixCols), ev(nev), q(ldq, matrixCols), ldq, nblk, matrixCols, mpi_comm_rows,
Andreas Marek's avatar
Andreas Marek committed
140 141 142 143 144 145
       mpi_comm_cols)

       With the definintions of the input and output variables:

       integer, intent(in)    na:            global dimension of quadratic matrix a to solve
       integer, intent(in)    nev:           number of eigenvalues to be computed; the first nev eigenvalules are calculated
146
       real*{4|8},  intent(inout) a:         locally distributed part of the matrix a. The local dimensions are lda x matrixCols
Andreas Marek's avatar
Andreas Marek committed
147
       integer, intent(in)    lda:           leading dimension of locally distributed matrix a
148 149
       real*{4|8},  intent(inout) ev:        on output the first nev computed eigenvalues
       real*{4|8},  intent(inout) q:         on output the first nev computed eigenvectors
Andreas Marek's avatar
Andreas Marek committed
150 151 152 153 154 155 156 157 158 159 160
       integer, intent(in)    ldq:           leading dimension of matrix q which stores the eigenvectors
       integer, intent(in)    nblk:          blocksize of block cyclic distributin, must be the same in both directions
       integer, intent(in)    matrixCols:    number of columns of locally distributed matrices a and q
       integer, intent(in)    mpi_comm_rows: communicator for communication in rows. Constructed with get_elpa_communicators(3)
       integer, intent(in)    mpi_comm_cols: communicator for communication in colums. Constructed with get_elpa_communicators(3)

       logical                success:       return value indicating success or failure

   C INTERFACE
       #include "elpa.h"

161
       success = elpa_solve_evp_real_1stage_{single|double} (int na, int nev,  double *a, int lda,  double *ev, double *q, int ldq, int nblk, int matrixCols, int
Andreas Marek's avatar
Andreas Marek committed
162 163 164 165 166 167
       mpi_comm_rows, int mpi_comm_cols);

       With the definintions of the input and output variables:

       int     na:            global dimension of quadratic matrix a to solve
       int     nev:           number of eigenvalues to be computed; the first nev eigenvalules are calculated
168
       {float|double} *a:     pointer to locally distributed part of the matrix a. The local dimensions are lda x matrixCols
Andreas Marek's avatar
Andreas Marek committed
169
       int     lda:           leading dimension of locally distributed matrix a
170 171
       {float|double} *ev:    pointer to memory containing on output the first nev computed eigenvalues
       {float|double} *q:     pointer to memory containing on output the first nev computed eigenvectors
Andreas Marek's avatar
Andreas Marek committed
172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
       int     ldq:           leading dimension of matrix q which stores the eigenvectors
       int     nblk:          blocksize of block cyclic distributin, must be the same in both directions
       int     matrixCols:    number of columns of locally distributed matrices a and q
       int     mpi_comm_rows: communicator for communication in rows. Constructed with get_elpa_communicators(3)
       int     mpi_comm_cols: communicator for communication in colums. Constructed with get_elpa_communicators(3)

       int     success:       return value indicating success (1) or failure (0)

DESCRIPTION
       Solve the real eigenvalue problem with the 1-stage solver. The ELPA communicators mpi_comm_rows and mpi_comm_cols are obtained with the
       get_elpa_communicators(3) function. The distributed quadratic marix a has global dimensions na x na, and a local size lda x matrixCols.
       The solver will compute the first nev eigenvalues, which will be stored on exit in ev. The eigenvectors corresponding to the eigenvalues
       will be stored in q. All memory of the arguments must be allocated outside the call to the solver.

   FORTRAN INTERFACE
       use elpa1
188
       success = elpa_solve_evp_complex_1stage_{single|double} (na, nev, a(lda,matrixCols), ev(nev), q(ldq, matrixCols), ldq, nblk, matrixCols, mpi_comm_rows,
Andreas Marek's avatar
Andreas Marek committed
189 190 191 192 193 194
       mpi_comm_cols)

       With the definintions of the input and output variables:

       integer,     intent(in)    na:            global dimension of quadratic matrix a to solve
       integer,     intent(in)    nev:           number of eigenvalues to be computed; the first nev eigenvalules are calculated
195
       complex*{8|16},  intent(inout) a:         locally distributed part of the matrix a. The local dimensions are lda x matrixCols
Andreas Marek's avatar
Andreas Marek committed
196
       integer,     intent(in)    lda:           leading dimension of locally distributed matrix a
197 198
       real*{4|8},      intent(inout) ev:        on output the first nev computed eigenvalues
       complex*{8|16},  intent(inout) q:         on output the first nev computed eigenvectors
Andreas Marek's avatar
Andreas Marek committed
199 200 201 202 203 204 205 206 207 208 209 210
       integer,     intent(in)    ldq:           leading dimension of matrix q which stores the eigenvectors
       integer,     intent(in)    nblk:          blocksize of block cyclic distributin, must be the same in both directions
       integer,     intent(in)    matrixCols:    number of columns of locally distributed matrices a and q
       integer,     intent(in)    mpi_comm_rows: communicator for communication in rows. Constructed with get_elpa_communicators(3)
       integer, intent(in)        mpi_comm_cols: communicator for communication in colums. Constructed with get_elpa_communicators(3)

       logical                    success:       return value indicating success or failure

   C INTERFACE
       #include "elpa.h"
       #include <complex.h>

211
       success = elpa_solve_evp_complex_1stage_{single|double} (int na, int nev,  double complex *a, int lda,  double *ev, double complex*q, int ldq, int nblk, int
Andreas Marek's avatar
Andreas Marek committed
212 213 214 215 216 217
       matrixCols, int mpi_comm_rows, int mpi_comm_cols);

       With the definintions of the input and output variables:

       int             na:            global dimension of quadratic matrix a to solve
       int             nev:           number of eigenvalues to be computed; the first nev eigenvalules are calculated
218
       {float|double} complex *a:     pointer to locally distributed part of the matrix a. The local dimensions are lda x matrixCols
Andreas Marek's avatar
Andreas Marek committed
219
       int             lda:           leading dimension of locally distributed matrix a
220 221
       {float|double}         *ev:    pointer to memory containing on output the first nev computed eigenvalues
       {float|double} complex *q:     pointer to memory containing on output the first nev computed eigenvectors
Andreas Marek's avatar
Andreas Marek committed
222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
       int             ldq:           leading dimension of matrix q which stores the eigenvectors
       int             nblk:          blocksize of block cyclic distributin, must be the same in both directions
       int             matrixCols:    number of columns of locally distributed matrices a and q
       int             mpi_comm_rows: communicator for communication in rows. Constructed with get_elpa_communicators(3)
       int             mpi_comm_cols: communicator for communication in colums. Constructed with get_elpa_communicators(3)

       int             success:       return value indicating success (1) or failure (0)

DESCRIPTION
       Solve the complex eigenvalue problem with the 1-stage solver. The ELPA communicators mpi_comm_rows and mpi_comm_cols are obtained with the
       get_elpa_communicators(3) function. The distributed quadratic marix a has global dimensions na x na, and a local size lda x matrixCols.
       The solver will compute the first nev eigenvalues, which will be stored on exit in ev. The eigenvectors corresponding to the eigenvalues
       will be stored in q. All memory of the arguments must be allocated outside the call to the solver.


The *ELPA 1stage* solver, does not need or accept any other parameters than in the above
specification.

#### Using *ELPA 2stage* ####

The *ELPA 2stage* solver can be used in the same manner, as the *ELPA 1stage* solver.
However, the 2 stage solver, can be used with different compute kernels, which offers
more possibilities for configuration.

It is recommended to first call the utillity program

elpa2_print_kernels

which will tell all the compute kernels that can be used with *ELPA 2stage*". It will
also give information, whether a kernel can be set via environment variables.

##### Using the default kernels #####

If no kernel is set either via an environment variable or the *ELPA 2stage API* then
the default kernels will be set.

##### Setting the *ELPA 2stage* compute kernels #####

260 261
##### Setting the *ELPA 2stage* compute kernels with environment variables#####

Andreas Marek's avatar
Andreas Marek committed
262 263 264 265
If the *ELPA* installation allows setting ther compute kernels with enviroment variables,
setting the variables "REAL_ELPA_KERNEL" and "COMPLEX_ELPA_KERNEL" will set the compute
kernels. The environment variable setting will take precedence over all other settings!

266 267 268 269 270 271
The utility program "elpa2_print_kernels" can list which kernels are available and which
would be choosen. This reflects, as well the setting of the default kernel or the settings
with the environment variables

##### Setting the *ELPA 2stage* compute kernels with API calls#####

Andreas Marek's avatar
Andreas Marek committed
272 273
It is also possible to set the *ELPA 2stage* compute kernels via the API.

274 275
As an example the API for ELPA real double-precision 2stage is shown:

Andreas Marek's avatar
Andreas Marek committed
276 277
SYNOPSIS
   FORTRAN INTERFACE
Andreas Marek's avatar
Andreas Marek committed
278 279
       use elpa1
       use elpa2
280 281
       success = elpa_solve_evp_real_2stage_double (na, nev, a(lda,matrixCols), ev(nev), q(ldq, matrixCols), ldq, nblk, matrixCols, mpi_comm_rows,
       mpi_comm_cols, mpi_comm_all, THIS_REAL_ELPA_KERNEL, useQR, useGPU)
Andreas Marek's avatar
Andreas Marek committed
282 283 284 285 286

       With the definintions of the input and output variables:

       integer, intent(in)            na:            global dimension of quadratic matrix a to solve
       integer, intent(in)            nev:           number of eigenvalues to be computed; the first nev eigenvalules are calculated
287
       real*{4|8},  intent(inout)         a:         locally distributed part of the matrix a. The local dimensions are lda x matrixCols
Andreas Marek's avatar
Andreas Marek committed
288
       integer, intent(in)            lda:           leading dimension of locally distributed matrix a
289 290
       real*{4|8},  intent(inout)         ev:        on output the first nev computed eigenvalues
       real*{4|8},  intent(inout)         q:         on output the first nev computed eigenvectors
Andreas Marek's avatar
Andreas Marek committed
291 292 293 294 295 296 297
       integer, intent(in)            ldq:           leading dimension of matrix q which stores the eigenvectors
       integer, intent(in)            nblk:          blocksize of block cyclic distributin, must be the same in both directions
       integer, intent(in)            matrixCols:    number of columns of locally distributed matrices a and q
       integer, intent(in)            mpi_comm_rows: communicator for communication in rows. Constructed with get_elpa_communicators(3)
       integer, intent(in)            mpi_comm_cols: communicator for communication in colums. Constructed with get_elpa_communicators(3)
       integer, intent(in)            mpi_comm_all:  communicator for all processes in the processor set involved in ELPA
       logical, intent(in), optional: useQR:         optional argument; switches to QR-decomposition if set to .true.
298
       logical, intent(in), optional: useGPU:        decide whether GPUs should be used ore not
Andreas Marek's avatar
Andreas Marek committed
299 300 301 302 303 304

      logical                        success:       return value indicating success or failure

   C INTERFACE
       #include "elpa.h"

305 306
       success = elpa_solve_evp_real_2stage_double (int na, int nev,  double *a, int lda,  double *ev, double *q, int ldq, int nblk, int matrixCols, int
       mpi_comm_rows, int mpi_comm_cols, int mpi_comm_all, int THIS_ELPA_REAL_KERNEL, int useQR, int useGPU);
Andreas Marek's avatar
Andreas Marek committed
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322

       With the definintions of the input and output variables:

       int     na:            global dimension of quadratic matrix a to solve
       int     nev:           number of eigenvalues to be computed; the first nev eigenvalules are calculated
       double *a:             pointer to locally distributed part of the matrix a. The local dimensions are lda x matrixCols
       int     lda:           leading dimension of locally distributed matrix a
       double *ev:            pointer to memory containing on output the first nev computed eigenvalues
       double *q:             pointer to memory containing on output the first nev computed eigenvectors
       int     ldq:           leading dimension of matrix q which stores the eigenvectors
       int     nblk:          blocksize of block cyclic distributin, must be the same in both directions
       int     matrixCols:    number of columns of locally distributed matrices a and q
       int     mpi_comm_rows: communicator for communication in rows. Constructed with get_elpa_communicators(3)
       int     mpi_comm_cols: communicator for communication in colums. Constructed with get_elpa_communicators(3)
       int     mpi_comm_all:  communicator for all processes in the processor set involved in ELPA
       int     useQR:         if set to 1 switch to QR-decomposition
323
       int     useGPU:        decide whether the GPU version should be used or not
Andreas Marek's avatar
Andreas Marek committed
324 325 326 327 328 329 330 331 332 333

       int     success:       return value indicating success (1) or failure (0)


DESCRIPTION
       Solve the real eigenvalue problem with the 2-stage solver. The ELPA communicators mpi_comm_rows and mpi_comm_cols are obtained with the
       get_elpa_communicators(3) function. The distributed quadratic marix a has global dimensions na x na, and a local size lda x matrixCols.
       The solver will compute the first nev eigenvalues, which will be stored on exit in ev. The eigenvectors corresponding to the eigenvalues
       will be stored in q. All memory of the arguments must be allocated outside the call to the solver.

334 335
##### Setting up *ELPA 1stage* or *ELPA 2stage* with the *ELPA driver interface* #####

336 337 338
Since release ELPA 2016.005.004 a driver routine allows to choose more easily which solver (1stage or 2stage) will be used.

As an exmple the real double-precision case is explained:
339 340 341 342 343 344 345

 SYNOPSIS

 FORTRAN INTERFACE

  use elpa

346
  success = elpa_solve_evp_real_double (na, nev, a(lda,matrixCols), ev(nev), q(ldq, matrixCols), ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, mpi_comm_all, THIS_REAL_ELPA_KERNEL=THIS_REAL_ELPA_KERNEL, useQR, useGPU, method=method)
347 348 349 350 351 352

  Generalized interface to the ELPA 1stage and 2stage solver for real-valued problems

  With the definintions of the input and output variables:


Andreas Marek's avatar
Andreas Marek committed
353
  integer, intent(in)            na:                    global dimension of quadratic matrix a to solve
354

Andreas Marek's avatar
Andreas Marek committed
355
  integer, intent(in)            nev:                   number of eigenvalues to be computed; the first nev eigenvalules are calculated
356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380

  real*8,  intent(inout)         a:                     locally distributed part of the matrix a. The local dimensions are lda x matrixCols

  integer, intent(in)            lda:                   leading dimension of locally distributed matrix a

  real*8,  intent(inout)         ev:                    on output the first nev computed eigenvalues"

  real*8,  intent(inout)         q:                     on output the first nev computed eigenvectors"

  integer, intent(in)            ldq:                   leading dimension of matrix q which stores the eigenvectors

  integer, intent(in)            nblk:                  blocksize of block cyclic distributin, must be the same in both directions

  integer, intent(in)            matrixCols:            number of columns of locally distributed matrices a and q

  integer, intent(in)            mpi_comm_rows:         communicator for communication in rows. Constructed with elpa_get_communicators

  integer, intent(in)            mpi_comm_cols:         communicator for communication in colums. Constructed with elpa_get_communicators

  integer, intent(in)            mpi_comm_all:          communicator for all processes in the processor set involved in ELPA

  integer, intent(in), optional: THIS_REAL_ELPA_KERNEL: optional argument, choose the compute kernel for 2-stage solver

  logical, intent(in), optional: useQR:                 optional argument; switches to QR-decomposition if set to .true.

381 382
  logical, intent(in), optional: useQPU:                decide whether the GPU version should be used or not

383 384 385 386 387 388 389 390 391
  character(*), optional         method:                use 1stage solver if "1stage", use 2stage solver if "2stage", (at the moment) use 2stage solver if "auto"

  logical                        success:               return value indicating success or failure


 C INTERFACE

 #include "elpa.h"

392
 success = elpa_solve_evp_real_double (int na, int nev, double *a, int lda, double *ev, double *q, int ldq, int nblk, int matrixCols, int mpi_comm_rows, int mpi_comm_cols, int mpi_comm_all, int THIS_ELPA_REAL_KERNEL, int useQR, int useGPU, char *method);"
393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425


 With the definintions of the input and output variables:"


 int     na:                    global dimension of quadratic matrix a to solve

 int     nev:                   number of eigenvalues to be computed; the first nev eigenvalules are calculated

 double *a:                     pointer to locally distributed part of the matrix a. The local dimensions are lda x matrixCols

 int     lda:                   leading dimension of locally distributed matrix a

 double *ev:                    pointer to memory containing on output the first nev computed eigenvalues

 double *q:                     pointer to memory containing on output the first nev computed eigenvectors

 int     ldq:                   leading dimension of matrix q which stores the eigenvectors

 int     nblk:                  blocksize of block cyclic distributin, must be the same in both directions

 int     matrixCols:            number of columns of locally distributed matrices a and q

 int     mpi_comm_rows:         communicator for communication in rows. Constructed with elpa_get_communicators

 int     mpi_comm_cols:         communicator for communication in colums. Constructed with elpa_get_communicators

 int     mpi_comm_all:          communicator for all processes in the processor set involved in ELPA

 int     THIS_ELPA_REAL_KERNEL: choose the compute kernel for 2-stage solver

 int     useQR:                 if set to 1 switch to QR-decomposition

426 427
 int     useGPU:                decide whether the GPU version should be used or not

428 429 430 431 432 433 434
 char   *method:                use 1stage solver if "1stage", use 2stage solver if "2stage", (at the moment) use 2stage solver if "auto"

 int     success:               return value indicating success (1) or failure (0)

 DESCRIPTION
 Solve the real eigenvalue problem. The value of method desides whether the 1stage or 2stage solver is used. The ELPA communicators mpi_comm_rows and mpi_comm_cols are obtained with the elpa_get_communicators function. The distributed quadratic marix a has global dimensions na x na, and a local size lda x matrixCols. The solver will compute the first nev eigenvalues, which will be stored on exit in ev. The eigenvectors corresponding to the eigenvalues will be stored in q. All memory of the arguments must be allocated outside the call to the solver.

435 436 437 438 439 440 441 442 443 444 445 446 447 448
##### Setting up the GPU version of *ELPA* 1 and 2 stage #####

Since release ELPA 2016.011.001.pre *ELPA* offers GPU support, IF *ELPA* has been build with the configure option "--enabble-gpu-support".

At run-time the GPU version can be used by setting the environment variable "ELPA_USE_GPU" to "yes", or by calling the *ELPA* functions
(elpa_solve_evp_real_{double|single}, elpa_solve_evp_real_1stage_{double|single}, elpa_solve_evp_real_2stage_{double|single}) with the
argument "useGPU = .true." or "useGPU = 1" for the Fortran and C case, respectively. Please, not that similiar to the choice of the
*ELPA* 2stage compute kernels, the enviroment variable takes precendence over the setting in the API call.

Further note that it is NOT allowed to define the usage of GPUs AND to EXPLICITLY set an ELPA 2stage compute kernel other than
"REAL_ELPA_KERNEL_GPU" or "COMPLEX_ELPA_KERNEL_GPU".