Line data Source code
1 : !--------------------------------------------------------------------------------------------------!
2 : ! CP2K: A general program to perform molecular dynamics simulations !
3 : ! Copyright 2000-2026 CP2K developers group <https://cp2k.org> !
4 : ! !
5 : ! SPDX-License-Identifier: GPL-2.0-or-later !
6 : !--------------------------------------------------------------------------------------------------!
7 :
8 : ! **************************************************************************************************
9 : !> \brief Performance tests for basic tasks like matrix multiplies, copy, fft.
10 : !> \par History
11 : !> 30-Nov-2000 (JGH) added input
12 : !> 02-Jan-2001 (JGH) Parallel FFT
13 : !> 28-Feb-2002 (JGH) Clebsch-Gordon Coefficients
14 : !> 06-Jun-2003 (JGH) Real space grid test
15 : !> Eigensolver test (29.08.05,MK)
16 : !> \author JGH 6-NOV-2000
17 : ! **************************************************************************************************
18 : MODULE library_tests
19 :
20 : USE ai_coulomb_test, ONLY: eri_test
21 : USE cell_methods, ONLY: cell_create,&
22 : init_cell
23 : USE cell_types, ONLY: cell_release,&
24 : cell_type
25 : USE cg_test, ONLY: gaunt_coefficient_test
26 : USE cp_blacs_env, ONLY: cp_blacs_env_create,&
27 : cp_blacs_env_release,&
28 : cp_blacs_env_type
29 : USE cp_dbcsr_api, ONLY: dbcsr_reset_randmat_seed,&
30 : dbcsr_run_tests
31 : USE cp_eri_mme_interface, ONLY: cp_eri_mme_perf_acc_test
32 : USE cp_files, ONLY: close_file,&
33 : open_file
34 : USE cp_fm_basic_linalg, ONLY: cp_fm_gemm
35 : USE cp_fm_diag, ONLY: cp_fm_syevd,&
36 : cp_fm_syevx
37 : USE cp_fm_struct, ONLY: cp_fm_struct_create,&
38 : cp_fm_struct_get,&
39 : cp_fm_struct_release,&
40 : cp_fm_struct_type
41 : USE cp_fm_types, ONLY: cp_fm_create,&
42 : cp_fm_pilaenv,&
43 : cp_fm_release,&
44 : cp_fm_set_all,&
45 : cp_fm_set_submatrix,&
46 : cp_fm_to_fm,&
47 : cp_fm_type
48 : USE cp_log_handling, ONLY: cp_get_default_logger,&
49 : cp_logger_type
50 : USE cp_output_handling, ONLY: cp_print_key_finished_output,&
51 : cp_print_key_unit_nr
52 : USE cp_realspace_grid_init, ONLY: init_input_type
53 : USE dbm_tests, ONLY: dbm_run_tests
54 : USE fft_tools, ONLY: BWFFT,&
55 : FFT_RADIX_CLOSEST,&
56 : FWFFT,&
57 : fft3d,&
58 : fft_radix_operations,&
59 : finalize_fft,&
60 : init_fft
61 : USE global_types, ONLY: global_environment_type
62 : USE input_constants, ONLY: do_diag_syevd,&
63 : do_diag_syevx,&
64 : do_mat_random,&
65 : do_mat_read,&
66 : do_pwgrid_ns_fullspace,&
67 : do_pwgrid_ns_halfspace,&
68 : do_pwgrid_spherical
69 : USE input_section_types, ONLY: section_vals_get,&
70 : section_vals_get_subs_vals,&
71 : section_vals_type,&
72 : section_vals_val_get
73 : USE kinds, ONLY: dp
74 : USE machine, ONLY: m_flush,&
75 : m_walltime
76 : USE mathconstants, ONLY: gaussi
77 : USE message_passing, ONLY: mp_para_env_type
78 : USE minimax_exp, ONLY: validate_exp_minimax
79 : USE mp_perf_test, ONLY: mpi_perf_test
80 : USE parallel_gemm_api, ONLY: parallel_gemm
81 : USE parallel_rng_types, ONLY: UNIFORM,&
82 : rng_stream_type
83 : USE pw_grid_types, ONLY: FULLSPACE,&
84 : HALFSPACE,&
85 : pw_grid_type
86 : USE pw_grids, ONLY: pw_grid_create,&
87 : pw_grid_release
88 : USE pw_methods, ONLY: pw_transfer,&
89 : pw_zero
90 : USE pw_types, ONLY: pw_c1d_gs_type,&
91 : pw_c3d_rs_type,&
92 : pw_r3d_rs_type
93 : USE realspace_grid_types, ONLY: &
94 : realspace_grid_desc_type, realspace_grid_input_type, realspace_grid_type, rs_grid_create, &
95 : rs_grid_create_descriptor, rs_grid_print, rs_grid_release, rs_grid_release_descriptor, &
96 : rs_grid_zero, transfer_pw2rs, transfer_rs2pw
97 : USE shg_integrals_test, ONLY: shg_integrals_perf_acc_test
98 : USE time_frequency_grids, ONLY: test_least_square_ft
99 : #include "./base/base_uses.f90"
100 :
101 : IMPLICIT NONE
102 :
103 : PRIVATE
104 : PUBLIC :: lib_test
105 :
106 : INTEGER :: runtest(100)
107 : REAL(KIND=dp) :: max_memory
108 : REAL(KIND=dp), PARAMETER :: threshold = 1.0E-8_dp
109 : CHARACTER(len=*), PARAMETER, PRIVATE :: moduleN = 'library_tests'
110 :
111 : CONTAINS
112 :
113 : ! **************************************************************************************************
114 : !> \brief Master routine for tests
115 : !> \param root_section ...
116 : !> \param para_env ...
117 : !> \param globenv ...
118 : !> \par History
119 : !> none
120 : !> \author JGH 6-NOV-2000
121 : ! **************************************************************************************************
122 720 : SUBROUTINE lib_test(root_section, para_env, globenv)
123 :
124 : TYPE(section_vals_type), POINTER :: root_section
125 : TYPE(mp_para_env_type), POINTER :: para_env
126 : TYPE(global_environment_type), POINTER :: globenv
127 :
128 : CHARACTER(LEN=*), PARAMETER :: routineN = 'lib_test'
129 :
130 : INTEGER :: handle, iw
131 : LOGICAL :: explicit
132 : TYPE(cp_logger_type), POINTER :: logger
133 : TYPE(section_vals_type), POINTER :: cp_dbcsr_test_section, cp_fm_gemm_test_section, &
134 : dbm_test_section, eigensolver_section, eri_mme_test_section, pw_transfer_section, &
135 : rs_pw_transfer_section, shg_integrals_test_section
136 :
137 80 : CALL timeset(routineN, handle)
138 :
139 80 : logger => cp_get_default_logger()
140 80 : iw = cp_print_key_unit_nr(logger, root_section, "TEST%PROGRAM_RUN_INFO", extension=".log")
141 :
142 80 : IF (iw > 0) THEN
143 40 : WRITE (iw, '(T2,79("*"))')
144 40 : WRITE (iw, '(A,T31,A,T80,A)') ' *', ' PERFORMANCE TESTS ', '*'
145 40 : WRITE (iw, '(T2,79("*"))')
146 : END IF
147 : !
148 80 : CALL test_input(root_section, para_env)
149 : !
150 80 : IF (runtest(1) /= 0) CALL copy_test(para_env, iw)
151 : !
152 80 : IF (runtest(2) /= 0) CALL matmul_test(para_env, test_matmul=.TRUE., test_dgemm=.FALSE., iw=iw)
153 80 : IF (runtest(5) /= 0) CALL matmul_test(para_env, test_matmul=.FALSE., test_dgemm=.TRUE., iw=iw)
154 : !
155 80 : IF (runtest(3) /= 0) CALL fft_test(para_env, iw, globenv%fftw_plan_type, &
156 : globenv%fftw_wisdom_file_name, &
157 2 : globenv%fftw_wisdom_file_explicit)
158 : !
159 80 : IF (runtest(4) /= 0) CALL eri_test(iw)
160 : !
161 80 : IF (runtest(6) /= 0) CALL gaunt_coefficient_test()
162 : !
163 : ! runtest 7 has been deleted and can be recycled
164 : !
165 80 : IF (runtest(8) /= 0) CALL mpi_perf_test(para_env, runtest(8), iw)
166 : !
167 80 : IF (runtest(10) /= 0) CALL validate_exp_minimax(runtest(10), iw)
168 : !
169 80 : IF (runtest(11) /= 0) CALL test_least_square_ft(runtest(11), iw)
170 : !
171 :
172 80 : rs_pw_transfer_section => section_vals_get_subs_vals(root_section, "TEST%RS_PW_TRANSFER")
173 80 : CALL section_vals_get(rs_pw_transfer_section, explicit=explicit)
174 80 : IF (explicit) THEN
175 2 : CALL rs_pw_transfer_test(para_env, iw, globenv, rs_pw_transfer_section)
176 : END IF
177 :
178 80 : pw_transfer_section => section_vals_get_subs_vals(root_section, "TEST%PW_TRANSFER")
179 80 : CALL section_vals_get(pw_transfer_section, explicit=explicit)
180 80 : IF (explicit) THEN
181 10 : CALL pw_fft_test(para_env, iw, globenv, pw_transfer_section)
182 : END IF
183 :
184 80 : cp_fm_gemm_test_section => section_vals_get_subs_vals(root_section, "TEST%CP_FM_GEMM")
185 80 : CALL section_vals_get(cp_fm_gemm_test_section, explicit=explicit)
186 80 : IF (explicit) THEN
187 4 : CALL cp_fm_gemm_test(para_env, iw, cp_fm_gemm_test_section)
188 : END IF
189 :
190 80 : eigensolver_section => section_vals_get_subs_vals(root_section, "TEST%EIGENSOLVER")
191 80 : CALL section_vals_get(eigensolver_section, explicit=explicit)
192 80 : IF (explicit) THEN
193 2 : CALL eigensolver_test(para_env, iw, eigensolver_section)
194 : END IF
195 :
196 80 : eri_mme_test_section => section_vals_get_subs_vals(root_section, "TEST%ERI_MME_TEST")
197 80 : CALL section_vals_get(eri_mme_test_section, explicit=explicit)
198 80 : IF (explicit) THEN
199 8 : CALL cp_eri_mme_perf_acc_test(para_env, iw, eri_mme_test_section)
200 : END IF
201 :
202 80 : shg_integrals_test_section => section_vals_get_subs_vals(root_section, "TEST%SHG_INTEGRALS_TEST")
203 80 : CALL section_vals_get(shg_integrals_test_section, explicit=explicit)
204 80 : IF (explicit) THEN
205 4 : CALL shg_integrals_perf_acc_test(iw, shg_integrals_test_section)
206 : END IF
207 :
208 : ! DBCSR tests
209 80 : cp_dbcsr_test_section => section_vals_get_subs_vals(root_section, "TEST%CP_DBCSR")
210 80 : CALL section_vals_get(cp_dbcsr_test_section, explicit=explicit)
211 80 : IF (explicit) THEN
212 32 : CALL cp_dbcsr_tests(para_env, iw, cp_dbcsr_test_section)
213 : END IF
214 :
215 : ! DBM tests
216 80 : dbm_test_section => section_vals_get_subs_vals(root_section, "TEST%DBM")
217 80 : CALL section_vals_get(dbm_test_section, explicit=explicit)
218 80 : IF (explicit) THEN
219 14 : CALL run_dbm_tests(para_env, iw, dbm_test_section)
220 : END IF
221 :
222 80 : CALL cp_print_key_finished_output(iw, logger, root_section, "TEST%PROGRAM_RUN_INFO")
223 :
224 80 : CALL timestop(handle)
225 :
226 80 : END SUBROUTINE lib_test
227 :
228 : ! **************************************************************************************************
229 : !> \brief Reads input section &TEST ... &END
230 : !> \param root_section ...
231 : !> \param para_env ...
232 : !> \author JGH 30-NOV-2000
233 : !> \note
234 : !> I---------------------------------------------------------------------------I
235 : !> I SECTION: &TEST ... &END I
236 : !> I I
237 : !> I MEMORY max_memory I
238 : !> I COPY n I
239 : !> I MATMUL n I
240 : !> I FFT n I
241 : !> I ERI n I
242 : !> I PW_FFT n I
243 : !> I Clebsch-Gordon n -> Gaunt coefficients I
244 : !> I RS_GRIDS n I
245 : !> I MPI n I
246 : !> I RNG n -> Parallel random number generator I
247 : !> I---------------------------------------------------------------------------I
248 : ! **************************************************************************************************
249 160 : SUBROUTINE test_input(root_section, para_env)
250 : TYPE(section_vals_type), POINTER :: root_section
251 : TYPE(mp_para_env_type), POINTER :: para_env
252 :
253 : TYPE(section_vals_type), POINTER :: test_section
254 :
255 : !
256 : !..defaults
257 : ! using this style is not recommended, introduce sections instead (see e.g. cp_fm_gemm)
258 :
259 80 : runtest = 0
260 80 : test_section => section_vals_get_subs_vals(root_section, "TEST")
261 80 : CALL section_vals_val_get(test_section, "MEMORY", r_val=max_memory)
262 80 : CALL section_vals_val_get(test_section, 'COPY', i_val=runtest(1))
263 80 : CALL section_vals_val_get(test_section, 'MATMUL', i_val=runtest(2))
264 80 : CALL section_vals_val_get(test_section, 'DGEMM', i_val=runtest(5))
265 80 : CALL section_vals_val_get(test_section, 'FFT', i_val=runtest(3))
266 80 : CALL section_vals_val_get(test_section, 'ERI', i_val=runtest(4))
267 80 : CALL section_vals_val_get(test_section, 'CLEBSCH_GORDON', i_val=runtest(6))
268 80 : CALL section_vals_val_get(test_section, 'MPI', i_val=runtest(8))
269 80 : CALL section_vals_val_get(test_section, 'MINIMAX', i_val=runtest(10))
270 80 : CALL section_vals_val_get(test_section, 'LEAST_SQ_FT', i_val=runtest(11))
271 :
272 80 : CALL para_env%sync()
273 80 : END SUBROUTINE test_input
274 :
275 : ! **************************************************************************************************
276 : !> \brief Tests the performance to copy two vectors.
277 : !> \param para_env ...
278 : !> \param iw ...
279 : !> \par History
280 : !> none
281 : !> \author JGH 6-NOV-2000
282 : !> \note
283 : !> The results of these tests allow to determine the size of the cache
284 : !> of the CPU. This can be used to optimize the performance of the
285 : !> FFTSG library.
286 : ! **************************************************************************************************
287 2 : SUBROUTINE copy_test(para_env, iw)
288 : TYPE(mp_para_env_type), POINTER :: para_env
289 : INTEGER :: iw
290 :
291 : INTEGER :: i, j, len, ntim, siz
292 : REAL(KIND=dp) :: perf, t, tend, tstart
293 2 : REAL(KIND=dp), ALLOCATABLE, DIMENSION(:) :: ca, cb
294 :
295 : ! test for copy --> Cache size
296 :
297 2 : siz = ABS(runtest(1))
298 2 : IF (para_env%is_source()) WRITE (iw, '(//,A,/)') " Test of copy ( F95 ) "
299 34 : DO i = 6, 24
300 34 : len = 2**i
301 34 : IF (8.0_dp*REAL(len, KIND=dp) > max_memory*0.5_dp) EXIT
302 96 : ALLOCATE (ca(len))
303 64 : ALLOCATE (cb(len))
304 :
305 32 : CALL RANDOM_NUMBER(ca)
306 32 : ntim = NINT(1.e7_dp/REAL(len, KIND=dp))
307 32 : ntim = MAX(ntim, 1)
308 32 : ntim = MIN(ntim, siz*10000)
309 :
310 32 : tstart = m_walltime()
311 512524 : DO j = 1, ntim
312 315058412 : cb(:) = ca(:)
313 512524 : ca(1) = REAL(j, KIND=dp)
314 : END DO
315 32 : tend = m_walltime()
316 32 : t = tend - tstart + threshold
317 32 : IF (t > 0.0_dp) THEN
318 32 : perf = REAL(ntim, KIND=dp)*REAL(len, KIND=dp)*1.e-6_dp/t
319 : ELSE
320 0 : perf = 0.0_dp
321 : END IF
322 :
323 32 : IF (para_env%is_source()) THEN
324 16 : WRITE (iw, '(A,i2,i10,A,T59,F14.4,A)') " Copy test: Size = 2^", i, &
325 32 : len/1024, " Kwords", perf, " Mcopy/s"
326 : END IF
327 :
328 32 : DEALLOCATE (ca)
329 34 : DEALLOCATE (cb)
330 : END DO
331 2 : CALL para_env%sync()
332 2 : END SUBROUTINE copy_test
333 :
334 : ! **************************************************************************************************
335 : !> \brief Tests the performance of different kinds of matrix matrix multiply
336 : !> kernels for the BLAS and F95 intrinsic matmul.
337 : !> \param para_env ...
338 : !> \param test_matmul ...
339 : !> \param test_dgemm ...
340 : !> \param iw ...
341 : !> \par History
342 : !> none
343 : !> \author JGH 6-NOV-2000
344 : ! **************************************************************************************************
345 2 : SUBROUTINE matmul_test(para_env, test_matmul, test_dgemm, iw)
346 : TYPE(mp_para_env_type), POINTER :: para_env
347 : LOGICAL :: test_matmul, test_dgemm
348 : INTEGER :: iw
349 :
350 : INTEGER :: i, j, len, ntim, siz
351 : REAL(KIND=dp) :: perf, t, tend, tstart, xdum
352 2 : REAL(KIND=dp), ALLOCATABLE, DIMENSION(:, :) :: ma, mb, mc
353 :
354 : ! test for matrix multpies
355 :
356 2 : IF (test_matmul) THEN
357 2 : siz = ABS(runtest(2))
358 2 : IF (para_env%is_source()) WRITE (iw, '(//,A,/)') " Test of matmul ( F95 ) "
359 2 : DO i = 5, siz, 2
360 4 : len = 2**i + 1
361 4 : IF (8.0_dp*REAL(len*len, KIND=dp) > max_memory*0.3_dp) EXIT
362 16 : ALLOCATE (ma(len, len))
363 12 : ALLOCATE (mb(len, len))
364 12 : ALLOCATE (mc(len, len))
365 4 : mc = 0.0_dp
366 :
367 4 : CALL RANDOM_NUMBER(xdum)
368 35788 : ma = xdum
369 4 : CALL RANDOM_NUMBER(xdum)
370 35788 : mb = xdum
371 4 : ntim = NINT(1.e8_dp/(2.0_dp*REAL(len, KIND=dp)**3))
372 4 : ntim = MAX(ntim, 1)
373 4 : ntim = MIN(ntim, siz*200)
374 4 : tstart = m_walltime()
375 2832 : DO j = 1, ntim
376 2828 : mc(:, :) = MATMUL(ma, mb)
377 2832 : ma(1, 1) = REAL(j, KIND=dp)
378 : END DO
379 4 : tend = m_walltime()
380 4 : t = tend - tstart + threshold
381 4 : perf = REAL(ntim, KIND=dp)*2.0_dp*REAL(len, KIND=dp)**3*1.e-6_dp/t
382 4 : IF (para_env%is_source()) THEN
383 : WRITE (iw, '(A,i6,T59,F14.4,A)') &
384 2 : " Matrix multiply test: c = a * b Size = ", len, perf, " Mflop/s"
385 : END IF
386 4 : tstart = m_walltime()
387 2832 : DO j = 1, ntim
388 3895652 : mc(:, :) = mc + MATMUL(ma, mb)
389 2832 : ma(1, 1) = REAL(j, KIND=dp)
390 : END DO
391 4 : tend = m_walltime()
392 4 : t = tend - tstart + threshold
393 4 : IF (t > 0.0_dp) THEN
394 4 : perf = REAL(ntim, KIND=dp)*2.0_dp*REAL(len, KIND=dp)**3*1.e-6_dp/t
395 : ELSE
396 0 : perf = 0.0_dp
397 : END IF
398 :
399 4 : IF (para_env%is_source()) THEN
400 : WRITE (iw, '(A,i6,T59,F14.4,A)') &
401 2 : " Matrix multiply test: a = a * b Size = ", len, perf, " Mflop/s"
402 : END IF
403 :
404 4 : tstart = m_walltime()
405 2832 : DO j = 1, ntim
406 3895652 : mc(:, :) = mc + MATMUL(ma, TRANSPOSE(mb))
407 2832 : ma(1, 1) = REAL(j, KIND=dp)
408 : END DO
409 4 : tend = m_walltime()
410 4 : t = tend - tstart + threshold
411 4 : IF (t > 0.0_dp) THEN
412 4 : perf = REAL(ntim, KIND=dp)*2.0_dp*REAL(len, KIND=dp)**3*1.e-6_dp/t
413 : ELSE
414 0 : perf = 0.0_dp
415 : END IF
416 :
417 4 : IF (para_env%is_source()) THEN
418 : WRITE (iw, '(A,i6,T59,F14.4,A)') &
419 2 : " Matrix multiply test: c = a * b(T) Size = ", len, perf, " Mflop/s"
420 : END IF
421 :
422 4 : tstart = m_walltime()
423 2832 : DO j = 1, ntim
424 3895652 : mc(:, :) = mc + MATMUL(TRANSPOSE(ma), mb)
425 2832 : ma(1, 1) = REAL(j, KIND=dp)
426 : END DO
427 4 : tend = m_walltime()
428 4 : t = tend - tstart + threshold
429 4 : IF (t > 0.0_dp) THEN
430 4 : perf = REAL(ntim, KIND=dp)*2.0_dp*REAL(len, KIND=dp)**3*1.e-6_dp/t
431 : ELSE
432 0 : perf = 0.0_dp
433 : END IF
434 :
435 4 : IF (para_env%is_source()) THEN
436 : WRITE (iw, '(A,i6,T59,F14.4,A)') &
437 2 : " Matrix multiply test: c = a(T) * b Size = ", len, perf, " Mflop/s"
438 : END IF
439 :
440 4 : DEALLOCATE (ma)
441 4 : DEALLOCATE (mb)
442 4 : DEALLOCATE (mc)
443 : END DO
444 : END IF
445 :
446 : ! test for matrix multpies
447 2 : IF (test_dgemm) THEN
448 0 : siz = ABS(runtest(5))
449 0 : IF (para_env%is_source()) WRITE (iw, '(//,A,/)') " Test of matmul ( BLAS ) "
450 0 : DO i = 5, siz, 2
451 0 : len = 2**i + 1
452 0 : IF (8.0_dp*REAL(len*len, KIND=dp) > max_memory*0.3_dp) EXIT
453 0 : ALLOCATE (ma(len, len))
454 0 : ALLOCATE (mb(len, len))
455 0 : ALLOCATE (mc(len, len))
456 0 : mc = 0.0_dp
457 :
458 0 : CALL RANDOM_NUMBER(xdum)
459 0 : ma = xdum
460 0 : CALL RANDOM_NUMBER(xdum)
461 0 : mb = xdum
462 0 : ntim = NINT(1.e8_dp/(2.0_dp*REAL(len, KIND=dp)**3))
463 0 : ntim = MAX(ntim, 1)
464 0 : ntim = MIN(ntim, 1000)
465 :
466 0 : tstart = m_walltime()
467 0 : DO j = 1, ntim
468 0 : CALL dgemm("N", "N", len, len, len, 1.0_dp, ma, len, mb, len, 1.0_dp, mc, len)
469 : END DO
470 0 : tend = m_walltime()
471 0 : t = tend - tstart + threshold
472 0 : IF (t > 0.0_dp) THEN
473 0 : perf = REAL(ntim, KIND=dp)*2.0_dp*REAL(len, KIND=dp)**3*1.e-6_dp/t
474 : ELSE
475 0 : perf = 0.0_dp
476 : END IF
477 :
478 0 : IF (para_env%is_source()) THEN
479 : WRITE (iw, '(A,i6,T59,F14.4,A)') &
480 0 : " Matrix multiply test: c = a * b Size = ", len, perf, " Mflop/s"
481 : END IF
482 :
483 0 : tstart = m_walltime()
484 0 : DO j = 1, ntim
485 0 : CALL dgemm("N", "N", len, len, len, 1.0_dp, ma, len, mb, len, 1.0_dp, mc, len)
486 : END DO
487 0 : tend = m_walltime()
488 0 : t = tend - tstart + threshold
489 0 : IF (t > 0.0_dp) THEN
490 0 : perf = REAL(ntim, KIND=dp)*2.0_dp*REAL(len, KIND=dp)**3*1.e-6_dp/t
491 : ELSE
492 0 : perf = 0.0_dp
493 : END IF
494 :
495 0 : IF (para_env%is_source()) THEN
496 : WRITE (iw, '(A,i6,T59,F14.4,A)') &
497 0 : " Matrix multiply test: a = a * b Size = ", len, perf, " Mflop/s"
498 : END IF
499 :
500 0 : tstart = m_walltime()
501 0 : DO j = 1, ntim
502 0 : CALL dgemm("N", "T", len, len, len, 1.0_dp, ma, len, mb, len, 1.0_dp, mc, len)
503 : END DO
504 0 : tend = m_walltime()
505 0 : t = tend - tstart + threshold
506 0 : IF (t > 0.0_dp) THEN
507 0 : perf = REAL(ntim, KIND=dp)*2.0_dp*REAL(len, KIND=dp)**3*1.e-6_dp/t
508 : ELSE
509 0 : perf = 0.0_dp
510 : END IF
511 :
512 0 : IF (para_env%is_source()) THEN
513 : WRITE (iw, '(A,i6,T59,F14.4,A)') &
514 0 : " Matrix multiply test: c = a * b(T) Size = ", len, perf, " Mflop/s"
515 : END IF
516 :
517 0 : tstart = m_walltime()
518 0 : DO j = 1, ntim
519 0 : CALL dgemm("T", "N", len, len, len, 1.0_dp, ma, len, mb, len, 1.0_dp, mc, len)
520 : END DO
521 0 : tend = m_walltime()
522 0 : t = tend - tstart + threshold
523 0 : IF (t > 0.0_dp) THEN
524 0 : perf = REAL(ntim, KIND=dp)*2.0_dp*REAL(len, KIND=dp)**3*1.e-6_dp/t
525 : ELSE
526 0 : perf = 0.0_dp
527 : END IF
528 :
529 0 : IF (para_env%is_source()) THEN
530 : WRITE (iw, '(A,i6,T59,F14.4,A)') &
531 0 : " Matrix multiply test: c = a(T) * b Size = ", len, perf, " Mflop/s"
532 : END IF
533 :
534 0 : DEALLOCATE (ma)
535 0 : DEALLOCATE (mb)
536 0 : DEALLOCATE (mc)
537 : END DO
538 : END IF
539 :
540 2 : CALL para_env%sync()
541 :
542 2 : END SUBROUTINE matmul_test
543 :
544 : ! **************************************************************************************************
545 : !> \brief Tests the performance of all available FFT libraries for 3D FFTs
546 : !> \param para_env ...
547 : !> \param iw ...
548 : !> \param fftw_plan_type ...
549 : !> \param wisdom_file where FFTW3 should look to save/load wisdom
550 : !> \param export_wisdom whether the FFTW3 wisdom is saved to wisdom_file
551 : !> \par History
552 : !> none
553 : !> \author JGH 6-NOV-2000
554 : ! **************************************************************************************************
555 2 : SUBROUTINE fft_test(para_env, iw, fftw_plan_type, wisdom_file, export_wisdom)
556 :
557 : TYPE(mp_para_env_type), POINTER :: para_env
558 : INTEGER :: iw, fftw_plan_type
559 : CHARACTER(LEN=*), INTENT(IN) :: wisdom_file
560 : LOGICAL, INTENT(IN) :: export_wisdom
561 :
562 : INTEGER, PARAMETER :: ndate(3) = [12, 48, 96]
563 :
564 : INTEGER :: iall, it, j, len, n(3), ntim, radix_in, &
565 : radix_out, siz, stat
566 : COMPLEX(KIND=dp), DIMENSION(4, 4, 4) :: zz
567 2 : COMPLEX(KIND=dp), ALLOCATABLE, DIMENSION(:, :, :) :: ca, cb, cc
568 : CHARACTER(LEN=7) :: method
569 : REAL(KIND=dp) :: flops, perf, scale, t, tdiff, tend, &
570 : tstart
571 2 : REAL(KIND=dp), ALLOCATABLE, DIMENSION(:, :, :) :: ra
572 :
573 : ! test for 3d FFT
574 :
575 2 : IF (para_env%is_source()) WRITE (iw, '(//,A,/)') " Test of 3D-FFT "
576 2 : siz = ABS(runtest(3))
577 :
578 8 : DO iall = 1, 100
579 : SELECT CASE (iall)
580 : CASE DEFAULT
581 2 : EXIT
582 : CASE (1)
583 : CALL init_fft("FFTSG", alltoall=.FALSE., wisdom_file=wisdom_file, &
584 2 : pool_limit=10, plan_style=fftw_plan_type)
585 2 : method = "FFTSG "
586 : CASE (2)
587 2 : CYCLE
588 : CASE (3)
589 : CALL init_fft("FFTW3", alltoall=.FALSE., wisdom_file=wisdom_file, &
590 2 : pool_limit=10, plan_style=fftw_plan_type)
591 10 : method = "FFTW3 "
592 : END SELECT
593 16 : n = 4
594 4 : zz = 0.0_dp
595 4 : CALL fft3d(FWFFT, n, zz, status=stat)
596 4 : IF (stat == 0) THEN
597 16 : DO it = 1, 3
598 12 : radix_in = ndate(it)
599 12 : CALL fft_radix_operations(radix_in, radix_out, FFT_RADIX_CLOSEST)
600 12 : len = radix_out
601 48 : n = len
602 12 : IF (16.0_dp*REAL(len*len*len, KIND=dp) > max_memory*0.5_dp) EXIT
603 60 : ALLOCATE (ra(len, len, len))
604 60 : ALLOCATE (ca(len, len, len))
605 12 : CALL RANDOM_NUMBER(ra)
606 4035516 : ca(:, :, :) = ra
607 12 : CALL RANDOM_NUMBER(ra)
608 4035516 : ca(:, :, :) = ca + gaussi*ra
609 12 : flops = REAL(len**3, KIND=dp)*15.0_dp*LOG(REAL(len, KIND=dp))
610 12 : ntim = NINT(siz*1.e7_dp/flops)
611 12 : ntim = MAX(ntim, 1)
612 12 : ntim = MIN(ntim, 200)
613 12 : scale = 1.0_dp/REAL(len**3, KIND=dp)
614 12 : tstart = m_walltime()
615 884 : DO j = 1, ntim
616 872 : CALL fft3d(FWFFT, n, ca)
617 884 : CALL fft3d(BWFFT, n, ca)
618 : END DO
619 12 : tend = m_walltime()
620 12 : t = tend - tstart + threshold
621 12 : IF (t > 0.0_dp) THEN
622 12 : perf = REAL(ntim, KIND=dp)*2.0_dp*flops*1.e-6_dp/t
623 : ELSE
624 0 : perf = 0.0_dp
625 : END IF
626 :
627 12 : IF (para_env%is_source()) THEN
628 : WRITE (iw, '(T2,A,A,i6,T59,F14.4,A)') &
629 6 : ADJUSTR(method), " test (in-place) Size = ", len, perf, " Mflop/s"
630 : END IF
631 12 : DEALLOCATE (ca)
632 16 : DEALLOCATE (ra)
633 : END DO
634 4 : IF (para_env%is_source()) WRITE (iw, *)
635 : ! test if input data is preserved
636 4 : len = 24
637 16 : n = len
638 4 : ALLOCATE (ra(len, len, len))
639 4 : ALLOCATE (ca(len, len, len))
640 4 : ALLOCATE (cb(len, len, len))
641 4 : ALLOCATE (cc(len, len, len))
642 4 : CALL RANDOM_NUMBER(ra)
643 57700 : ca(:, :, :) = ra
644 4 : CALL RANDOM_NUMBER(ra)
645 57700 : ca(:, :, :) = ca + gaussi*ra
646 57700 : cc(:, :, :) = ca
647 4 : CALL fft3d(FWFFT, n, ca, cb)
648 57700 : tdiff = MAXVAL(ABS(ca - cc))
649 4 : IF (tdiff > 1.0E-12_dp) THEN
650 0 : IF (para_env%is_source()) THEN
651 0 : WRITE (iw, '(T2,A,A,A)') ADJUSTR(method), " FWFFT ", &
652 0 : " Input array is changed in out-of-place FFT !"
653 : END IF
654 : ELSE
655 4 : IF (para_env%is_source()) THEN
656 2 : WRITE (iw, '(T2,A,A,A)') ADJUSTR(method), " FWFFT ", &
657 4 : " Input array is not changed in out-of-place FFT !"
658 : END IF
659 : END IF
660 57700 : ca(:, :, :) = cc
661 4 : CALL fft3d(BWFFT, n, ca, cb)
662 57700 : tdiff = MAXVAL(ABS(ca - cc))
663 4 : IF (tdiff > 1.0E-12_dp) THEN
664 0 : IF (para_env%is_source()) THEN
665 0 : WRITE (iw, '(T2,A,A,A)') ADJUSTR(method), " BWFFT ", &
666 0 : " Input array is changed in out-of-place FFT !"
667 : END IF
668 : ELSE
669 4 : IF (para_env%is_source()) THEN
670 2 : WRITE (iw, '(T2,A,A,A)') ADJUSTR(method), " BWFFT ", &
671 4 : " Input array is not changed in out-of-place FFT !"
672 : END IF
673 : END IF
674 4 : IF (para_env%is_source()) WRITE (iw, *)
675 :
676 4 : DEALLOCATE (ra)
677 4 : DEALLOCATE (ca)
678 4 : DEALLOCATE (cb)
679 4 : DEALLOCATE (cc)
680 : END IF
681 4 : CALL finalize_fft(para_env, wisdom_file=wisdom_file, export_wisdom=export_wisdom)
682 : END DO
683 :
684 2 : END SUBROUTINE fft_test
685 :
686 : ! **************************************************************************************************
687 : !> \brief test rs_pw_transfer performance
688 : !> \param para_env ...
689 : !> \param iw ...
690 : !> \param globenv ...
691 : !> \param rs_pw_transfer_section ...
692 : !> \author Joost VandeVondele
693 : !> 9.2008 Randomise rs grid [Iain Bethune]
694 : !> (c) The Numerical Algorithms Group (NAG) Ltd, 2008 on behalf of the HECToR project
695 : ! **************************************************************************************************
696 2 : SUBROUTINE rs_pw_transfer_test(para_env, iw, globenv, rs_pw_transfer_section)
697 :
698 : TYPE(mp_para_env_type), POINTER :: para_env
699 : INTEGER :: iw
700 : TYPE(global_environment_type), POINTER :: globenv
701 : TYPE(section_vals_type), POINTER :: rs_pw_transfer_section
702 :
703 : CHARACTER(LEN=*), PARAMETER :: routineN = 'rs_pw_transfer_test'
704 :
705 : INTEGER :: halo_size, handle, i_loop, n_loop, ns_max
706 : INTEGER, DIMENSION(3) :: no, np
707 2 : INTEGER, DIMENSION(:), POINTER :: i_vals
708 : LOGICAL :: do_rs2pw
709 : REAL(KIND=dp) :: tend, tstart
710 : TYPE(cell_type), POINTER :: box
711 : TYPE(pw_grid_type), POINTER :: grid
712 : TYPE(pw_r3d_rs_type) :: ca
713 : TYPE(realspace_grid_desc_type), POINTER :: rs_desc
714 : TYPE(realspace_grid_input_type) :: input_settings
715 32 : TYPE(realspace_grid_type) :: rs_grid
716 : TYPE(section_vals_type), POINTER :: rs_grid_section
717 :
718 2 : CALL timeset(routineN, handle)
719 :
720 : !..set fft lib
721 : CALL init_fft(globenv%default_fft_library, alltoall=.FALSE., &
722 : pool_limit=globenv%fft_pool_scratch_limit, &
723 : wisdom_file=globenv%fftw_wisdom_file_name, &
724 2 : plan_style=globenv%fftw_plan_type)
725 :
726 : ! .. set cell (should otherwise be irrelevant)
727 2 : NULLIFY (box)
728 2 : CALL cell_create(box)
729 : box%hmat = RESHAPE([20.0_dp, 0.0_dp, 0.0_dp, 0.0_dp, 20.0_dp, 0.0_dp, &
730 26 : 0.0_dp, 0.0_dp, 20.0_dp], [3, 3])
731 2 : CALL init_cell(box)
732 :
733 : ! .. grid type and pw_grid
734 2 : NULLIFY (grid)
735 2 : CALL section_vals_val_get(rs_pw_transfer_section, "GRID", i_vals=i_vals)
736 8 : np = i_vals
737 2 : CALL pw_grid_create(grid, para_env, box%hmat, grid_span=FULLSPACE, npts=np, fft_usage=.TRUE., iounit=iw)
738 8 : no = grid%npts
739 :
740 2 : CALL ca%create(grid)
741 2 : CALL pw_zero(ca)
742 :
743 : ! .. rs input setting type
744 2 : CALL section_vals_val_get(rs_pw_transfer_section, "HALO_SIZE", i_val=halo_size)
745 2 : rs_grid_section => section_vals_get_subs_vals(rs_pw_transfer_section, "RS_GRID")
746 2 : ns_max = 2*halo_size + 1
747 2 : CALL init_input_type(input_settings, ns_max, rs_grid_section, 1, [-1, -1, -1])
748 :
749 : ! .. rs type
750 2 : NULLIFY (rs_desc)
751 2 : CALL rs_grid_create_descriptor(rs_desc, pw_grid=grid, input_settings=input_settings)
752 2 : CALL rs_grid_create(rs_grid, rs_desc)
753 2 : CALL rs_grid_print(rs_grid, iw)
754 2 : CALL rs_grid_zero(rs_grid)
755 :
756 : ! Put random values on the grid, so summation check will pick up errors
757 2 : CALL RANDOM_NUMBER(rs_grid%r)
758 :
759 2 : CALL section_vals_val_get(rs_pw_transfer_section, "N_loop", i_val=N_loop)
760 2 : CALL section_vals_val_get(rs_pw_transfer_section, "RS2PW", l_val=do_rs2pw)
761 :
762 : ! go for the real loops, sync to get max timings
763 2 : IF (para_env%is_source()) THEN
764 1 : WRITE (iw, '(T2,A)') ""
765 1 : WRITE (iw, '(T2,A)') "Timing rs_pw_transfer routine"
766 1 : WRITE (iw, '(T2,A)') ""
767 1 : WRITE (iw, '(T2,A)') "iteration time[s]"
768 : END IF
769 8 : DO i_loop = 1, N_loop
770 6 : CALL para_env%sync()
771 6 : tstart = m_walltime()
772 6 : IF (do_rs2pw) THEN
773 6 : CALL transfer_rs2pw(rs_grid, ca)
774 : ELSE
775 0 : CALL transfer_pw2rs(rs_grid, ca)
776 : END IF
777 6 : CALL para_env%sync()
778 6 : tend = m_walltime()
779 8 : IF (para_env%is_source()) THEN
780 3 : WRITE (iw, '(T2,I9,1X,F12.6)') i_loop, tend - tstart
781 : END IF
782 : END DO
783 :
784 : !cleanup
785 2 : CALL rs_grid_release(rs_grid)
786 2 : CALL rs_grid_release_descriptor(rs_desc)
787 2 : CALL ca%release()
788 2 : CALL pw_grid_release(grid)
789 2 : CALL cell_release(box)
790 : CALL finalize_fft(para_env, wisdom_file=globenv%fftw_wisdom_file_name, &
791 2 : export_wisdom=globenv%fftw_wisdom_file_explicit)
792 :
793 2 : CALL timestop(handle)
794 :
795 10 : END SUBROUTINE rs_pw_transfer_test
796 :
797 : ! **************************************************************************************************
798 : !> \brief Tests the performance of PW calls to FFT routines
799 : !> \param para_env ...
800 : !> \param iw ...
801 : !> \param globenv ...
802 : !> \param pw_transfer_section ...
803 : !> \par History
804 : !> JGH 6-Feb-2001 : Test and performance code
805 : !> Made input sensitive [Joost VandeVondele]
806 : !> \author JGH 1-JAN-2001
807 : ! **************************************************************************************************
808 10 : SUBROUTINE pw_fft_test(para_env, iw, globenv, pw_transfer_section)
809 :
810 : TYPE(mp_para_env_type), POINTER :: para_env
811 : INTEGER :: iw
812 : TYPE(global_environment_type), POINTER :: globenv
813 : TYPE(section_vals_type), POINTER :: pw_transfer_section
814 :
815 : REAL(KIND=dp), PARAMETER :: toler = 1.e-11_dp
816 :
817 : INTEGER :: blocked_id, grid_span, i_layout, i_rep, &
818 : ig, ip, itmp, n_loop, n_rep, nn, p, q
819 10 : INTEGER, ALLOCATABLE, DIMENSION(:, :) :: layouts
820 : INTEGER, DIMENSION(2) :: distribution_layout
821 : INTEGER, DIMENSION(3) :: no, np
822 10 : INTEGER, DIMENSION(:), POINTER :: i_vals
823 : LOGICAL :: debug, is_fullspace, odd, &
824 : pw_grid_layout_all, spherical
825 : REAL(KIND=dp) :: em, et, flops, gsq, perf, t, t_max, &
826 : t_min, tend, tstart
827 10 : REAL(KIND=dp), ALLOCATABLE, DIMENSION(:) :: t_end, t_start
828 : TYPE(cell_type), POINTER :: box
829 : TYPE(pw_c1d_gs_type) :: ca, cc
830 : TYPE(pw_c3d_rs_type) :: cb
831 : TYPE(pw_grid_type), POINTER :: grid
832 :
833 : !..set fft lib
834 :
835 : CALL init_fft(globenv%default_fft_library, alltoall=.FALSE., &
836 : pool_limit=globenv%fft_pool_scratch_limit, &
837 : wisdom_file=globenv%fftw_wisdom_file_name, &
838 10 : plan_style=globenv%fftw_plan_type)
839 :
840 : !..the unit cell (should not really matter, the number of grid points do)
841 10 : NULLIFY (box, grid)
842 10 : CALL cell_create(box)
843 : box%hmat = RESHAPE([10.0_dp, 0.0_dp, 0.0_dp, 0.0_dp, 8.0_dp, 0.0_dp, &
844 130 : 0.0_dp, 0.0_dp, 7.0_dp], [3, 3])
845 10 : CALL init_cell(box)
846 :
847 10 : CALL section_vals_get(pw_transfer_section, n_repetition=n_rep)
848 70 : DO i_rep = 1, n_rep
849 :
850 : ! how often should we do the transfer
851 60 : CALL section_vals_val_get(pw_transfer_section, "N_loop", i_rep_section=i_rep, i_val=N_loop)
852 180 : ALLOCATE (t_start(N_loop))
853 120 : ALLOCATE (t_end(N_loop))
854 :
855 : ! setup of the grids
856 60 : CALL section_vals_val_get(pw_transfer_section, "GRID", i_rep_section=i_rep, i_vals=i_vals)
857 240 : np = i_vals
858 :
859 60 : CALL section_vals_val_get(pw_transfer_section, "PW_GRID_BLOCKED", i_rep_section=i_rep, i_val=blocked_id)
860 60 : CALL section_vals_val_get(pw_transfer_section, "DEBUG", i_rep_section=i_rep, l_val=debug)
861 :
862 : CALL section_vals_val_get(pw_transfer_section, "PW_GRID_LAYOUT_ALL", i_rep_section=i_rep, &
863 60 : l_val=pw_grid_layout_all)
864 :
865 : ! prepare to loop over all or a specific layout
866 60 : IF (pw_grid_layout_all) THEN
867 : ! count layouts that fit
868 8 : itmp = 0
869 : ! start from 2, (/1,para_env%num_pe/) is not supported
870 16 : DO p = 2, para_env%num_pe
871 8 : q = para_env%num_pe/p
872 16 : IF (p*q == para_env%num_pe) THEN
873 8 : itmp = itmp + 1
874 : END IF
875 : END DO
876 : ! build list
877 24 : ALLOCATE (layouts(2, itmp))
878 8 : itmp = 0
879 16 : DO p = 2, para_env%num_pe
880 8 : q = para_env%num_pe/p
881 16 : IF (p*q == para_env%num_pe) THEN
882 8 : itmp = itmp + 1
883 24 : layouts(:, itmp) = [p, q]
884 : END IF
885 : END DO
886 : ELSE
887 52 : CALL section_vals_val_get(pw_transfer_section, "PW_GRID_LAYOUT", i_rep_section=i_rep, i_vals=i_vals)
888 52 : ALLOCATE (layouts(2, 1))
889 156 : layouts(:, 1) = i_vals
890 : END IF
891 :
892 120 : DO i_layout = 1, SIZE(layouts, 2)
893 :
894 180 : distribution_layout = layouts(:, i_layout)
895 :
896 60 : CALL section_vals_val_get(pw_transfer_section, "PW_GRID", i_rep_section=i_rep, i_val=itmp)
897 :
898 : ! from cp_control_utils
899 16 : SELECT CASE (itmp)
900 : CASE (do_pwgrid_spherical)
901 16 : spherical = .TRUE.
902 16 : is_fullspace = .FALSE.
903 : CASE (do_pwgrid_ns_fullspace)
904 28 : spherical = .FALSE.
905 28 : is_fullspace = .TRUE.
906 : CASE (do_pwgrid_ns_halfspace)
907 16 : spherical = .FALSE.
908 60 : is_fullspace = .FALSE.
909 : END SELECT
910 :
911 : ! from pw_env_methods
912 60 : IF (spherical) THEN
913 16 : grid_span = HALFSPACE
914 16 : spherical = .TRUE.
915 16 : odd = .TRUE.
916 44 : ELSE IF (is_fullspace) THEN
917 28 : grid_span = FULLSPACE
918 28 : spherical = .FALSE.
919 28 : odd = .FALSE.
920 : ELSE
921 16 : grid_span = HALFSPACE
922 16 : spherical = .FALSE.
923 16 : odd = .TRUE.
924 : END IF
925 :
926 : ! actual setup
927 : CALL pw_grid_create(grid, para_env, box%hmat, grid_span=grid_span, odd=odd, spherical=spherical, &
928 : blocked=blocked_id, npts=np, fft_usage=.TRUE., &
929 60 : rs_dims=distribution_layout, iounit=iw)
930 :
931 60 : IF (iw > 0) CALL m_flush(iw)
932 :
933 : ! note that the number of grid points might be different from what the user requested (fft-able needed)
934 240 : no = grid%npts
935 :
936 60 : CALL ca%create(grid)
937 60 : CALL cb%create(grid)
938 60 : CALL cc%create(grid)
939 :
940 : ! initialize data
941 60 : CALL pw_zero(ca)
942 60 : CALL pw_zero(cb)
943 60 : CALL pw_zero(cc)
944 60 : nn = SIZE(ca%array)
945 142042 : DO ig = 1, nn
946 141982 : gsq = grid%gsq(ig)
947 142042 : ca%array(ig) = EXP(-gsq)
948 : END DO
949 :
950 420 : flops = PRODUCT(no)*30.0_dp*LOG(REAL(MAXVAL(no), KIND=dp))
951 60 : tstart = m_walltime()
952 596 : DO ip = 1, n_loop
953 536 : CALL para_env%sync()
954 536 : t_start(ip) = m_walltime()
955 536 : CALL pw_transfer(ca, cb, debug)
956 536 : CALL pw_transfer(cb, cc, debug)
957 536 : CALL para_env%sync()
958 596 : t_end(ip) = m_walltime()
959 : END DO
960 60 : tend = m_walltime()
961 60 : t = tend - tstart + threshold
962 60 : IF (t > 0.0_dp) THEN
963 60 : perf = REAL(n_loop, KIND=dp)*2.0_dp*flops*1.e-6_dp/t
964 : ELSE
965 0 : perf = 0.0_dp
966 : END IF
967 :
968 142042 : em = MAXVAL(ABS(ca%array(:) - cc%array(:)))
969 60 : CALL para_env%max(em)
970 142042 : et = SUM(ABS(ca%array(:) - cc%array(:)))
971 60 : CALL para_env%sum(et)
972 596 : t_min = MINVAL(t_end - t_start)
973 596 : t_max = MAXVAL(t_end - t_start)
974 :
975 60 : IF (para_env%is_source()) THEN
976 30 : WRITE (iw, *)
977 30 : WRITE (iw, '(A,T67,E14.6)') " Parallel FFT Tests: Maximal Error ", em
978 30 : WRITE (iw, '(A,T67,E14.6)') " Parallel FFT Tests: Total Error ", et
979 : WRITE (iw, '(A,T67,F14.0)') &
980 30 : " Parallel FFT Tests: Performance [Mflops] ", perf
981 30 : WRITE (iw, '(A,T67,F14.6)') " Best time : ", t_min
982 30 : WRITE (iw, '(A,T67,F14.6)') " Worst time: ", t_max
983 30 : IF (iw > 0) CALL m_flush(iw)
984 : END IF
985 :
986 : ! need debugging ???
987 60 : IF (em > toler .OR. et > toler) THEN
988 0 : CPWARN("The FFT results are not accurate ... starting debug pw_transfer")
989 0 : CALL pw_transfer(ca, cb, .TRUE.)
990 0 : CALL pw_transfer(cb, cc, .TRUE.)
991 : END IF
992 :
993 : ! done with these grids
994 60 : CALL ca%release()
995 60 : CALL cb%release()
996 60 : CALL cc%release()
997 180 : CALL pw_grid_release(grid)
998 :
999 : END DO
1000 :
1001 : ! local arrays
1002 60 : DEALLOCATE (layouts)
1003 60 : DEALLOCATE (t_start)
1004 190 : DEALLOCATE (t_end)
1005 :
1006 : END DO
1007 :
1008 : ! cleanup
1009 10 : CALL cell_release(box)
1010 : CALL finalize_fft(para_env, wisdom_file=globenv%fftw_wisdom_file_name, &
1011 10 : export_wisdom=globenv%fftw_wisdom_file_explicit)
1012 :
1013 20 : END SUBROUTINE pw_fft_test
1014 :
1015 : ! **************************************************************************************************
1016 : !> \brief Tests the eigensolver library routines
1017 : !> \param para_env ...
1018 : !> \param iw ...
1019 : !> \param eigensolver_section ...
1020 : !> \par History
1021 : !> JGH 6-Feb-2001 : Test and performance code
1022 : !> \author JGH 1-JAN-2001
1023 : ! **************************************************************************************************
1024 2 : SUBROUTINE eigensolver_test(para_env, iw, eigensolver_section)
1025 :
1026 : TYPE(mp_para_env_type), POINTER :: para_env
1027 : INTEGER :: iw
1028 : TYPE(section_vals_type), POINTER :: eigensolver_section
1029 :
1030 : INTEGER :: diag_method, i, i_loop, i_rep, &
1031 : init_method, j, n, n_loop, n_rep, &
1032 : neig, unit_number
1033 : REAL(KIND=dp) :: t1, t2
1034 2 : REAL(KIND=dp), ALLOCATABLE, DIMENSION(:) :: eigenvalues
1035 2 : REAL(KIND=dp), ALLOCATABLE, DIMENSION(:, :) :: buffer
1036 : TYPE(cp_blacs_env_type), POINTER :: blacs_env
1037 : TYPE(cp_fm_struct_type), POINTER :: fmstruct
1038 : TYPE(cp_fm_type) :: eigenvectors, matrix, work
1039 2 : TYPE(rng_stream_type), ALLOCATABLE :: rng_stream
1040 :
1041 2 : IF (iw > 0) THEN
1042 1 : WRITE (UNIT=iw, FMT="(/,/,T2,A,/)") "EIGENSOLVER TEST"
1043 : END IF
1044 :
1045 : ! create blacs env corresponding to para_env
1046 2 : NULLIFY (blacs_env)
1047 : CALL cp_blacs_env_create(blacs_env=blacs_env, &
1048 2 : para_env=para_env)
1049 :
1050 : ! loop over all tests
1051 2 : CALL section_vals_get(eigensolver_section, n_repetition=n_rep)
1052 10 : DO i_rep = 1, n_rep
1053 :
1054 : ! parse section
1055 8 : CALL section_vals_val_get(eigensolver_section, "N", i_rep_section=i_rep, i_val=n)
1056 8 : CALL section_vals_val_get(eigensolver_section, "EIGENVALUES", i_rep_section=i_rep, i_val=neig)
1057 8 : CALL section_vals_val_get(eigensolver_section, "DIAG_METHOD", i_rep_section=i_rep, i_val=diag_method)
1058 8 : CALL section_vals_val_get(eigensolver_section, "INIT_METHOD", i_rep_section=i_rep, i_val=init_method)
1059 8 : CALL section_vals_val_get(eigensolver_section, "N_loop", i_rep_section=i_rep, i_val=n_loop)
1060 :
1061 : ! proper number of eigs
1062 8 : IF (neig < 0) neig = n
1063 8 : neig = MIN(neig, n)
1064 :
1065 : ! report
1066 8 : IF (iw > 0) THEN
1067 4 : WRITE (iw, *) "Matrix size", n
1068 4 : WRITE (iw, *) "Number of eigenvalues", neig
1069 4 : WRITE (iw, *) "Timing loops", n_loop
1070 2 : SELECT CASE (diag_method)
1071 : CASE (do_diag_syevd)
1072 2 : WRITE (iw, *) "Diag using syevd"
1073 : CASE (do_diag_syevx)
1074 4 : WRITE (iw, *) "Diag using syevx"
1075 : CASE DEFAULT
1076 : ! stop
1077 : END SELECT
1078 :
1079 4 : SELECT CASE (init_method)
1080 : CASE (do_mat_random)
1081 4 : WRITE (iw, *) "using random matrix"
1082 : CASE (do_mat_read)
1083 4 : WRITE (iw, *) "reading from file"
1084 : CASE DEFAULT
1085 : ! stop
1086 : END SELECT
1087 : END IF
1088 :
1089 : ! create matrix struct type
1090 8 : NULLIFY (fmstruct)
1091 : CALL cp_fm_struct_create(fmstruct=fmstruct, &
1092 : para_env=para_env, &
1093 : context=blacs_env, &
1094 : nrow_global=n, &
1095 8 : ncol_global=n)
1096 :
1097 : ! create all needed matrices, and buffers for the eigenvalues
1098 : CALL cp_fm_create(matrix=matrix, &
1099 : matrix_struct=fmstruct, &
1100 8 : name="MATRIX")
1101 8 : CALL cp_fm_set_all(matrix, 0.0_dp)
1102 :
1103 : CALL cp_fm_create(matrix=eigenvectors, &
1104 : matrix_struct=fmstruct, &
1105 8 : name="EIGENVECTORS")
1106 8 : CALL cp_fm_set_all(eigenvectors, 0.0_dp)
1107 :
1108 : CALL cp_fm_create(matrix=work, &
1109 : matrix_struct=fmstruct, &
1110 8 : name="WORK")
1111 8 : CALL cp_fm_set_all(matrix, 0.0_dp)
1112 :
1113 24 : ALLOCATE (eigenvalues(n))
1114 8 : eigenvalues = 0.0_dp
1115 16 : ALLOCATE (buffer(1, n))
1116 :
1117 : ! generate initial matrix, either by reading a file, or using random numbers
1118 8 : IF (para_env%is_source()) THEN
1119 4 : SELECT CASE (init_method)
1120 : CASE (do_mat_random)
1121 : rng_stream = rng_stream_type( &
1122 : name="rng_stream", &
1123 : distribution_type=UNIFORM, &
1124 4 : extended_precision=.TRUE.)
1125 : CASE (do_mat_read)
1126 : CALL open_file(file_name="MATRIX", &
1127 : file_action="READ", &
1128 : file_form="FORMATTED", &
1129 : file_status="OLD", &
1130 4 : unit_number=unit_number)
1131 : END SELECT
1132 : END IF
1133 :
1134 100 : DO i = 1, n
1135 92 : IF (para_env%is_source()) THEN
1136 : SELECT CASE (init_method)
1137 : CASE (do_mat_random)
1138 347 : DO j = i, n
1139 347 : buffer(1, j) = rng_stream%next() - 0.5_dp
1140 : END DO
1141 : !MK activate/modify for a diagonal dominant symmetric matrix:
1142 : !MK buffer(1,i) = 10.0_dp*buffer(1,i)
1143 : CASE (do_mat_read)
1144 46 : READ (UNIT=unit_number, FMT=*) buffer(1, 1:n)
1145 : END SELECT
1146 : END IF
1147 92 : CALL para_env%bcast(buffer)
1148 8 : SELECT CASE (init_method)
1149 : CASE (do_mat_random)
1150 : CALL cp_fm_set_submatrix(fm=matrix, &
1151 : new_values=buffer, &
1152 : start_row=i, &
1153 : start_col=i, &
1154 : n_rows=1, &
1155 : n_cols=n - i + 1, &
1156 : alpha=1.0_dp, &
1157 : beta=0.0_dp, &
1158 92 : transpose=.FALSE.)
1159 : CALL cp_fm_set_submatrix(fm=matrix, &
1160 : new_values=buffer, &
1161 : start_row=i, &
1162 : start_col=i, &
1163 : n_rows=n - i + 1, &
1164 : n_cols=1, &
1165 : alpha=1.0_dp, &
1166 : beta=0.0_dp, &
1167 92 : transpose=.TRUE.)
1168 : CASE (do_mat_read)
1169 : CALL cp_fm_set_submatrix(fm=matrix, &
1170 : new_values=buffer, &
1171 : start_row=i, &
1172 : start_col=1, &
1173 : n_rows=1, &
1174 : n_cols=n, &
1175 : alpha=1.0_dp, &
1176 : beta=0.0_dp, &
1177 92 : transpose=.FALSE.)
1178 : END SELECT
1179 : END DO
1180 :
1181 8 : DEALLOCATE (buffer)
1182 :
1183 8 : IF (para_env%is_source()) THEN
1184 0 : SELECT CASE (init_method)
1185 : CASE (do_mat_read)
1186 4 : CALL close_file(unit_number=unit_number)
1187 : END SELECT
1188 : END IF
1189 :
1190 88 : DO i_loop = 1, n_loop
1191 80 : eigenvalues = 0.0_dp
1192 80 : CALL cp_fm_set_all(eigenvectors, 0.0_dp)
1193 : CALL cp_fm_to_fm(source=matrix, &
1194 80 : destination=work)
1195 :
1196 : ! DONE, now testing
1197 80 : t1 = m_walltime()
1198 40 : SELECT CASE (diag_method)
1199 : CASE (do_diag_syevd)
1200 : CALL cp_fm_syevd(matrix=work, &
1201 : eigenvectors=eigenvectors, &
1202 40 : eigenvalues=eigenvalues)
1203 : CASE (do_diag_syevx)
1204 : CALL cp_fm_syevx(matrix=work, &
1205 : eigenvectors=eigenvectors, &
1206 : eigenvalues=eigenvalues, &
1207 : neig=neig, &
1208 80 : work_syevx=1.0_dp)
1209 : END SELECT
1210 80 : t2 = m_walltime()
1211 88 : IF (iw > 0) WRITE (iw, *) "Timing for loop ", i_loop, " : ", t2 - t1
1212 : END DO
1213 :
1214 8 : IF (iw > 0) THEN
1215 4 : WRITE (iw, *) "Eigenvalues: "
1216 4 : WRITE (UNIT=iw, FMT="(T3,5F14.6)") eigenvalues(1:neig)
1217 42 : WRITE (UNIT=iw, FMT="(T3,A4,F16.6)") "Sum:", SUM(eigenvalues(1:neig))
1218 4 : WRITE (iw, *) ""
1219 : END IF
1220 :
1221 : ! Clean up
1222 8 : DEALLOCATE (eigenvalues)
1223 8 : CALL cp_fm_release(matrix=work)
1224 8 : CALL cp_fm_release(matrix=eigenvectors)
1225 8 : CALL cp_fm_release(matrix=matrix)
1226 42 : CALL cp_fm_struct_release(fmstruct=fmstruct)
1227 :
1228 : END DO
1229 :
1230 2 : CALL cp_blacs_env_release(blacs_env=blacs_env)
1231 :
1232 4 : END SUBROUTINE eigensolver_test
1233 :
1234 : ! **************************************************************************************************
1235 : !> \brief Tests the parallel matrix multiply
1236 : !> \param para_env ...
1237 : !> \param iw ...
1238 : !> \param cp_fm_gemm_test_section ...
1239 : ! **************************************************************************************************
1240 4 : SUBROUTINE cp_fm_gemm_test(para_env, iw, cp_fm_gemm_test_section)
1241 :
1242 : TYPE(mp_para_env_type), POINTER :: para_env
1243 : INTEGER :: iw
1244 : TYPE(section_vals_type), POINTER :: cp_fm_gemm_test_section
1245 :
1246 : CHARACTER(LEN=1) :: transa, transb
1247 : INTEGER :: i_loop, i_rep, k, m, n, N_loop, n_rep, ncol_block, ncol_block_actual, &
1248 : ncol_global, np, nrow_block, nrow_block_actual, nrow_global
1249 4 : INTEGER, DIMENSION(:), POINTER :: grid_2d
1250 : LOGICAL :: force_blocksize, row_major, transa_p, &
1251 : transb_p
1252 : REAL(KIND=dp) :: t1, t2, t3, t4
1253 : TYPE(cp_blacs_env_type), POINTER :: blacs_env
1254 : TYPE(cp_fm_struct_type), POINTER :: fmstruct_a, fmstruct_b, fmstruct_c
1255 : TYPE(cp_fm_type) :: matrix_a, matrix_b, matrix_c
1256 :
1257 4 : CALL section_vals_get(cp_fm_gemm_test_section, n_repetition=n_rep)
1258 24 : DO i_rep = 1, n_rep
1259 :
1260 : ! how often should we do the multiply
1261 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "N_loop", i_rep_section=i_rep, i_val=N_loop)
1262 :
1263 : ! matrices def.
1264 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "K", i_rep_section=i_rep, i_val=k)
1265 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "N", i_rep_section=i_rep, i_val=n)
1266 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "M", i_rep_section=i_rep, i_val=m)
1267 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "transa", i_rep_section=i_rep, l_val=transa_p)
1268 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "transb", i_rep_section=i_rep, l_val=transb_p)
1269 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "nrow_block", i_rep_section=i_rep, i_val=nrow_block)
1270 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "ncol_block", i_rep_section=i_rep, i_val=ncol_block)
1271 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "ROW_MAJOR", i_rep_section=i_rep, l_val=row_major)
1272 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "GRID_2D", i_rep_section=i_rep, i_vals=grid_2d)
1273 20 : CALL section_vals_val_get(cp_fm_gemm_test_section, "FORCE_BLOCKSIZE", i_rep_section=i_rep, l_val=force_blocksize)
1274 20 : transa = "N"
1275 20 : transb = "N"
1276 20 : IF (transa_p) transa = "T"
1277 20 : IF (transb_p) transb = "T"
1278 :
1279 20 : IF (iw > 0) THEN
1280 10 : WRITE (iw, '(T2,A)') "----------- TESTING PARALLEL MATRIX MULTIPLY -------------"
1281 10 : WRITE (iw, '(T2,A)', ADVANCE="NO") "C = "
1282 10 : IF (transa_p) THEN
1283 2 : WRITE (iw, '(A)', ADVANCE="NO") "TRANSPOSE(A) x"
1284 : ELSE
1285 8 : WRITE (iw, '(A)', ADVANCE="NO") "A x "
1286 : END IF
1287 10 : IF (transb_p) THEN
1288 2 : WRITE (iw, '(A)') "TRANSPOSE(B) "
1289 : ELSE
1290 8 : WRITE (iw, '(A)') "B "
1291 : END IF
1292 10 : WRITE (iw, '(T2,A,T50,I5,A,I5)') 'requested block size', nrow_block, ' by ', ncol_block
1293 10 : WRITE (iw, '(T2,A,T50,I5)') 'number of repetitions of cp_fm_gemm ', n_loop
1294 10 : WRITE (iw, '(T2,A,T50,L5)') 'Row Major', row_major
1295 30 : WRITE (iw, '(T2,A,T50,2I7)') 'GRID_2D ', grid_2d
1296 10 : WRITE (iw, '(T2,A,T50,L5)') 'Force blocksize ', force_blocksize
1297 : ! check the return value of pilaenv, too small values limit the performance (assuming pdgemm is the vanilla variant)
1298 10 : np = cp_fm_pilaenv(0, 'D')
1299 10 : IF (np > 0) THEN
1300 10 : WRITE (iw, '(T2,A,T50,I5)') 'PILAENV blocksize', np
1301 : END IF
1302 : END IF
1303 :
1304 20 : NULLIFY (blacs_env)
1305 : CALL cp_blacs_env_create(blacs_env=blacs_env, &
1306 : para_env=para_env, &
1307 : row_major=row_major, &
1308 20 : grid_2d=grid_2d)
1309 :
1310 20 : NULLIFY (fmstruct_a)
1311 20 : IF (transa_p) THEN
1312 4 : nrow_global = m; ncol_global = k
1313 : ELSE
1314 16 : nrow_global = k; ncol_global = m
1315 : END IF
1316 : CALL cp_fm_struct_create(fmstruct=fmstruct_a, para_env=para_env, context=blacs_env, &
1317 : nrow_global=nrow_global, ncol_global=ncol_global, &
1318 20 : nrow_block=nrow_block, ncol_block=ncol_block, force_block=force_blocksize)
1319 20 : CALL cp_fm_struct_get(fmstruct_a, nrow_block=nrow_block_actual, ncol_block=ncol_block_actual)
1320 30 : IF (iw > 0) WRITE (iw, '(T2,A,I9,A,I9,A,I5,A,I5)') 'matrix A ', nrow_global, " by ", ncol_global, &
1321 20 : ' using blocks of ', nrow_block_actual, ' by ', ncol_block_actual
1322 :
1323 20 : IF (transb_p) THEN
1324 4 : nrow_global = n; ncol_global = m
1325 : ELSE
1326 16 : nrow_global = m; ncol_global = n
1327 : END IF
1328 20 : NULLIFY (fmstruct_b)
1329 : CALL cp_fm_struct_create(fmstruct=fmstruct_b, para_env=para_env, context=blacs_env, &
1330 : nrow_global=nrow_global, ncol_global=ncol_global, &
1331 20 : nrow_block=nrow_block, ncol_block=ncol_block, force_block=force_blocksize)
1332 20 : CALL cp_fm_struct_get(fmstruct_b, nrow_block=nrow_block_actual, ncol_block=ncol_block_actual)
1333 30 : IF (iw > 0) WRITE (iw, '(T2,A,I9,A,I9,A,I5,A,I5)') 'matrix B ', nrow_global, " by ", ncol_global, &
1334 20 : ' using blocks of ', nrow_block_actual, ' by ', ncol_block_actual
1335 :
1336 20 : NULLIFY (fmstruct_c)
1337 20 : nrow_global = k
1338 20 : ncol_global = n
1339 : CALL cp_fm_struct_create(fmstruct=fmstruct_c, para_env=para_env, context=blacs_env, &
1340 : nrow_global=nrow_global, ncol_global=ncol_global, &
1341 20 : nrow_block=nrow_block, ncol_block=ncol_block, force_block=force_blocksize)
1342 20 : CALL cp_fm_struct_get(fmstruct_c, nrow_block=nrow_block_actual, ncol_block=ncol_block_actual)
1343 30 : IF (iw > 0) WRITE (iw, '(T2,A,I9,A,I9,A,I5,A,I5)') 'matrix C ', nrow_global, " by ", ncol_global, &
1344 20 : ' using blocks of ', nrow_block_actual, ' by ', ncol_block_actual
1345 :
1346 20 : CALL cp_fm_create(matrix=matrix_a, matrix_struct=fmstruct_a, name="MATRIX A")
1347 20 : CALL cp_fm_create(matrix=matrix_b, matrix_struct=fmstruct_b, name="MATRIX B")
1348 20 : CALL cp_fm_create(matrix=matrix_c, matrix_struct=fmstruct_c, name="MATRIX C")
1349 :
1350 20 : CALL RANDOM_NUMBER(matrix_a%local_data)
1351 20 : CALL RANDOM_NUMBER(matrix_b%local_data)
1352 20 : CALL RANDOM_NUMBER(matrix_c%local_data)
1353 :
1354 20 : IF (iw > 0) CALL m_flush(iw)
1355 :
1356 20 : t1 = m_walltime()
1357 932 : DO i_loop = 1, N_loop
1358 912 : t3 = m_walltime()
1359 912 : CALL parallel_gemm(transa, transb, k, n, m, 1.0_dp, matrix_a, matrix_b, 0.0_dp, matrix_c)
1360 912 : t4 = m_walltime()
1361 932 : IF (iw > 0) THEN
1362 456 : WRITE (iw, '(T2,A,T50,F12.6)') "cp_fm_gemm timing: ", (t4 - t3)
1363 456 : CALL m_flush(iw)
1364 : END IF
1365 : END DO
1366 20 : t2 = m_walltime()
1367 :
1368 20 : IF (iw > 0) THEN
1369 10 : WRITE (iw, '(T2,A,T50,F12.6)') "average cp_fm_gemm timing: ", (t2 - t1)/N_loop
1370 10 : IF (t2 > t1) THEN
1371 10 : WRITE (iw, '(T2,A,T50,F12.6)') "cp_fm_gemm Gflops per MPI task: ", &
1372 20 : 2*REAL(m, kind=dp)*REAL(n, kind=dp)*REAL(k, kind=dp)*N_loop/MAX(0.001_dp, t2 - t1)/1.0E9_dp/para_env%num_pe
1373 : END IF
1374 : END IF
1375 :
1376 20 : CALL cp_fm_release(matrix=matrix_a)
1377 20 : CALL cp_fm_release(matrix=matrix_b)
1378 20 : CALL cp_fm_release(matrix=matrix_c)
1379 20 : CALL cp_fm_struct_release(fmstruct=fmstruct_a)
1380 20 : CALL cp_fm_struct_release(fmstruct=fmstruct_b)
1381 20 : CALL cp_fm_struct_release(fmstruct=fmstruct_c)
1382 144 : CALL cp_blacs_env_release(blacs_env=blacs_env)
1383 :
1384 : END DO
1385 :
1386 4 : END SUBROUTINE cp_fm_gemm_test
1387 :
1388 : ! **************************************************************************************************
1389 : !> \brief Tests the DBCSR interface.
1390 : !> \param para_env ...
1391 : !> \param iw ...
1392 : !> \param input_section ...
1393 : ! **************************************************************************************************
1394 32 : SUBROUTINE cp_dbcsr_tests(para_env, iw, input_section)
1395 :
1396 : TYPE(mp_para_env_type), POINTER :: para_env
1397 : INTEGER :: iw
1398 : TYPE(section_vals_type), POINTER :: input_section
1399 :
1400 : CHARACTER, DIMENSION(3) :: types
1401 : INTEGER :: data_type, i_rep, k, m, n, N_loop, &
1402 : n_rep, test_type
1403 32 : INTEGER, DIMENSION(:), POINTER :: bs_k, bs_m, bs_n, nproc
1404 : LOGICAL :: always_checksum, retain_sparsity, &
1405 : transa_p, transb_p
1406 : REAL(KIND=dp) :: alpha, beta, filter_eps, s_a, s_b, s_c
1407 :
1408 : ! ---------------------------------------------------------------------------
1409 :
1410 32 : NULLIFY (bs_m, bs_n, bs_k)
1411 32 : CALL section_vals_get(input_section, n_repetition=n_rep)
1412 32 : CALL dbcsr_reset_randmat_seed()
1413 78 : DO i_rep = 1, n_rep
1414 : ! how often should we do the multiply
1415 46 : CALL section_vals_val_get(input_section, "N_loop", i_rep_section=i_rep, i_val=N_loop)
1416 :
1417 : ! matrices def.
1418 46 : CALL section_vals_val_get(input_section, "TEST_TYPE", i_rep_section=i_rep, i_val=test_type)
1419 46 : CALL section_vals_val_get(input_section, "DATA_TYPE", i_rep_section=i_rep, i_val=data_type)
1420 46 : CALL section_vals_val_get(input_section, "K", i_rep_section=i_rep, i_val=k)
1421 46 : CALL section_vals_val_get(input_section, "N", i_rep_section=i_rep, i_val=n)
1422 46 : CALL section_vals_val_get(input_section, "M", i_rep_section=i_rep, i_val=m)
1423 46 : CALL section_vals_val_get(input_section, "transa", i_rep_section=i_rep, l_val=transa_p)
1424 46 : CALL section_vals_val_get(input_section, "transb", i_rep_section=i_rep, l_val=transb_p)
1425 : CALL section_vals_val_get(input_section, "bs_m", i_rep_section=i_rep, &
1426 46 : i_vals=bs_m)
1427 : CALL section_vals_val_get(input_section, "bs_n", i_rep_section=i_rep, &
1428 46 : i_vals=bs_n)
1429 : CALL section_vals_val_get(input_section, "bs_k", i_rep_section=i_rep, &
1430 46 : i_vals=bs_k)
1431 46 : CALL section_vals_val_get(input_section, "keepsparse", i_rep_section=i_rep, l_val=retain_sparsity)
1432 46 : CALL section_vals_val_get(input_section, "asparsity", i_rep_section=i_rep, r_val=s_a)
1433 46 : CALL section_vals_val_get(input_section, "bsparsity", i_rep_section=i_rep, r_val=s_b)
1434 46 : CALL section_vals_val_get(input_section, "csparsity", i_rep_section=i_rep, r_val=s_c)
1435 46 : CALL section_vals_val_get(input_section, "alpha", i_rep_section=i_rep, r_val=alpha)
1436 46 : CALL section_vals_val_get(input_section, "beta", i_rep_section=i_rep, r_val=beta)
1437 : CALL section_vals_val_get(input_section, "nproc", i_rep_section=i_rep, &
1438 46 : i_vals=nproc)
1439 : CALL section_vals_val_get(input_section, "atype", i_rep_section=i_rep, &
1440 46 : c_val=types(1))
1441 : CALL section_vals_val_get(input_section, "btype", i_rep_section=i_rep, &
1442 46 : c_val=types(2))
1443 : CALL section_vals_val_get(input_section, "ctype", i_rep_section=i_rep, &
1444 46 : c_val=types(3))
1445 : CALL section_vals_val_get(input_section, "filter_eps", &
1446 46 : i_rep_section=i_rep, r_val=filter_eps)
1447 46 : CALL section_vals_val_get(input_section, "ALWAYS_CHECKSUM", i_rep_section=i_rep, l_val=always_checksum)
1448 :
1449 : CALL dbcsr_run_tests(para_env%get_handle(), iw, nproc, &
1450 : [m, n, k], &
1451 : [transa_p, transb_p], &
1452 : bs_m, bs_n, bs_k, &
1453 : [s_a, s_b, s_c], &
1454 : alpha, beta, &
1455 : data_type=data_type, &
1456 : test_type=test_type, &
1457 : n_loops=n_loop, eps=filter_eps, retain_sparsity=retain_sparsity, &
1458 538 : always_checksum=always_checksum)
1459 : END DO
1460 32 : END SUBROUTINE cp_dbcsr_tests
1461 :
1462 : ! **************************************************************************************************
1463 : !> \brief Tests the DBM library.
1464 : !> \param para_env ...
1465 : !> \param iw ...
1466 : !> \param input_section ...
1467 : ! **************************************************************************************************
1468 14 : SUBROUTINE run_dbm_tests(para_env, iw, input_section)
1469 :
1470 : TYPE(mp_para_env_type), POINTER :: para_env
1471 : INTEGER :: iw
1472 : TYPE(section_vals_type), POINTER :: input_section
1473 :
1474 : INTEGER :: i_rep, k, m, n, N_loop, n_rep
1475 14 : INTEGER, DIMENSION(:), POINTER :: bs_k, bs_m, bs_n
1476 : LOGICAL :: always_checksum, retain_sparsity, &
1477 : transa_p, transb_p
1478 : REAL(KIND=dp) :: alpha, beta, filter_eps, s_a, s_b, s_c
1479 :
1480 : ! ---------------------------------------------------------------------------
1481 :
1482 14 : NULLIFY (bs_m, bs_n, bs_k)
1483 14 : CALL section_vals_get(input_section, n_repetition=n_rep)
1484 14 : CALL dbcsr_reset_randmat_seed()
1485 28 : DO i_rep = 1, n_rep
1486 14 : CALL section_vals_val_get(input_section, "N_loop", i_rep_section=i_rep, i_val=N_loop)
1487 14 : CALL section_vals_val_get(input_section, "K", i_rep_section=i_rep, i_val=k)
1488 14 : CALL section_vals_val_get(input_section, "N", i_rep_section=i_rep, i_val=n)
1489 14 : CALL section_vals_val_get(input_section, "M", i_rep_section=i_rep, i_val=m)
1490 14 : CALL section_vals_val_get(input_section, "transa", i_rep_section=i_rep, l_val=transa_p)
1491 14 : CALL section_vals_val_get(input_section, "transb", i_rep_section=i_rep, l_val=transb_p)
1492 14 : CALL section_vals_val_get(input_section, "bs_m", i_rep_section=i_rep, i_vals=bs_m)
1493 14 : CALL section_vals_val_get(input_section, "bs_n", i_rep_section=i_rep, i_vals=bs_n)
1494 14 : CALL section_vals_val_get(input_section, "bs_k", i_rep_section=i_rep, i_vals=bs_k)
1495 14 : CALL section_vals_val_get(input_section, "keepsparse", i_rep_section=i_rep, l_val=retain_sparsity)
1496 14 : CALL section_vals_val_get(input_section, "asparsity", i_rep_section=i_rep, r_val=s_a)
1497 14 : CALL section_vals_val_get(input_section, "bsparsity", i_rep_section=i_rep, r_val=s_b)
1498 14 : CALL section_vals_val_get(input_section, "csparsity", i_rep_section=i_rep, r_val=s_c)
1499 14 : CALL section_vals_val_get(input_section, "alpha", i_rep_section=i_rep, r_val=alpha)
1500 14 : CALL section_vals_val_get(input_section, "beta", i_rep_section=i_rep, r_val=beta)
1501 14 : CALL section_vals_val_get(input_section, "filter_eps", i_rep_section=i_rep, r_val=filter_eps)
1502 14 : CALL section_vals_val_get(input_section, "ALWAYS_CHECKSUM", i_rep_section=i_rep, l_val=always_checksum)
1503 :
1504 : CALL dbm_run_tests(mp_group=para_env, &
1505 : io_unit=iw, &
1506 : matrix_sizes=[m, n, k], &
1507 : trs=[transa_p, transb_p], &
1508 : bs_m=bs_m, &
1509 : bs_n=bs_n, &
1510 : bs_k=bs_k, &
1511 : sparsities=[s_a, s_b, s_c], &
1512 : alpha=alpha, &
1513 : beta=beta, &
1514 : n_loops=n_loop, &
1515 : eps=filter_eps, &
1516 : retain_sparsity=retain_sparsity, &
1517 154 : always_checksum=always_checksum)
1518 : END DO
1519 14 : END SUBROUTINE run_dbm_tests
1520 :
1521 8484 : END MODULE library_tests
|
