mod_compute_hh_trafo_real.F90 22 KB
Newer Older
1
2
module compute_hh_trafo_real
#include "config-f90.h"
3
  use elpa_mpi
4
5
6
7
8
9
10
11
12
13
14
  implicit none

#ifdef WITH_OPENMP
  public compute_hh_trafo_real_cpu_openmp
#else
  public compute_hh_trafo_real_cpu
#endif

  contains

#ifdef WITH_OPENMP
15
       subroutine compute_hh_trafo_real_cpu_openmp(a, stripe_width, a_dim2, stripe_count, max_threads, l_nev,         &
16
17
                                                   a_off, nbw, max_blk_size, bcast_buffer, kernel_flops, kernel_time, &
                                                   off, ncols, istripe,                                               &
18
                                                   my_thread, thread_width,  THIS_REAL_ELPA_KERNEL)
19
#else
20
       subroutine compute_hh_trafo_real_cpu       (a, stripe_width, a_dim2, stripe_count,                              &
21
22
23
24
25
26
27
28
29
30
31
32
33
                                                   a_off, nbw, max_blk_size, bcast_buffer,  kernel_flops, kernel_time, &
                                                   off, ncols, istripe, last_stripe_width,                             &
                                                   THIS_REAL_ELPA_KERNEL)
#endif


         use precision
         use elpa2_utilities
         use single_hh_trafo_real
#if defined(WITH_REAL_GENERIC_SIMPLE_KERNEL)
         use real_generic_simple_kernel, only : double_hh_trafo_generic_simple
#endif

34
35
36
#if defined(WITH_REAL_GENERIC_KERNEL) && !(defined(DESPERATELY_WANT_ASSUMED_SIZE))
        use real_generic_kernel, only : double_hh_trafo_generic
#endif
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

#if defined(WITH_REAL_BGP_KERNEL)
         use real_bgp_kernel, only : double_hh_trafo_bgp
#endif

#if defined(WITH_REAL_BGQ_KERNEL)
         use real_bgq_kernel, only : double_hh_trafo_bgq
#endif
#ifdef HAVE_DETAILED_TIMINGS
         use timings
#endif
         implicit none
         real(kind=rk), intent(inout) :: kernel_time
         integer(kind=lik)            :: kernel_flops
         integer(kind=ik), intent(in) :: nbw, max_blk_size
         real(kind=rk)                :: bcast_buffer(nbw,max_blk_size)
         integer(kind=ik), intent(in) :: a_off

         integer(kind=ik), intent(in) :: stripe_width,a_dim2,stripe_count

#ifndef WITH_OPENMP
         integer(kind=ik), intent(in) :: last_stripe_width
         real(kind=rk)                :: a(stripe_width,a_dim2,stripe_count)
#else
61
         integer(kind=ik), intent(in) :: max_threads, l_nev, thread_width
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
         real(kind=rk)                :: a(stripe_width,a_dim2,stripe_count,max_threads)
#endif
         integer(kind=ik), intent(in) :: THIS_REAL_ELPA_KERNEL

         ! Private variables in OMP regions (my_thread) should better be in the argument list!
         integer(kind=ik)             :: off, ncols, istripe
#ifdef WITH_OPENMP
         integer(kind=ik)             :: my_thread, noff
#endif
         integer(kind=ik)             :: j, nl, jj, jjj
         real(kind=rk)                :: w(nbw,6), ttt

#ifdef HAVE_DETAILED_TIMINGS
#ifdef WITH_OPENMP
         call timer%start("compute_hh_trafo_real_cpu_openmp")
#else
         call timer%start("compute_hh_trafo_real_cpu")
#endif
#endif
         ttt = mpi_wtime()

#ifndef WITH_OPENMP
         nl = merge(stripe_width, last_stripe_width, istripe<stripe_count)
#else

         if (istripe<stripe_count) then
           nl = stripe_width
         else
           noff = (my_thread-1)*thread_width + (istripe-1)*stripe_width
           nl = min(my_thread*thread_width-noff, l_nev-noff)
           if (nl<=0) then
#ifdef HAVE_DETAILED_TIMINGS
#ifdef WITH_OPENMP
             call timer%stop("compute_hh_trafo_real_cpu_openmp")
#else
             call timer%stop("compute_hh_trafo_real_cpu")
#endif
#endif
             return
           endif
         endif
#endif

#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
         if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX_BLOCK2 .or. &
107
             THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX2_BLOCK2 .or. &
108
             THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE_BLOCK2 .or. &
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
             THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC    .or. &
             THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC_SIMPLE .or. &
             THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE .or.        &
             THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_BGP .or.        &
             THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_BGQ) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */

           !FORTRAN CODE / X86 INRINISIC CODE / BG ASSEMBLER USING 2 HOUSEHOLDER VECTORS
#if defined(WITH_REAL_GENERIC_KERNEL)
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */

             do j = ncols, 2, -2
               w(:,1) = bcast_buffer(1:nbw,j+off)
               w(:,2) = bcast_buffer(1:nbw,j+off-1)

#ifdef WITH_OPENMP
127
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
128
               call double_hh_trafo_generic(a(1,j+off+a_off-1,istripe,my_thread), w, &
129
130
                                            nbw, nl, stripe_width, nbw)

131
#else
132
133
               call double_hh_trafo_generic(a(1:stripe_width,j+off+a_off-1:j+off+a_off+nbw-1, &
                                              istripe,my_thread), w(1:nbw,1:6), &
134
                                              nbw, nl, stripe_width, nbw)
135
136
137
138
139
140
141
142
143
#endif

#else /* WITH_OPENMP */

#ifdef DESPERATELY_WANT_ASSUMED_SIZE
               call double_hh_trafo_generic(a(1,j+off+a_off-1,istripe),w, &
                                            nbw, nl, stripe_width, nbw)

#else
144
145
               call double_hh_trafo_generic(a(1:stripe_width,j+off+a_off-1:j+off+a_off+nbw-1,istripe),w(1:nbw,1:6), &
                                            nbw, nl, stripe_width, nbw)
146
#endif
147
148
#endif /* WITH_OPENMP */

149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
             enddo

#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_GENERIC_KERNEL */


#if defined(WITH_REAL_GENERIC_SIMPLE_KERNEL)
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC_SIMPLE) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
             do j = ncols, 2, -2
               w(:,1) = bcast_buffer(1:nbw,j+off)
               w(:,2) = bcast_buffer(1:nbw,j+off-1)
#ifdef WITH_OPENMP
165
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
166
167
168
               call double_hh_trafo_generic_simple(a(1,j+off+a_off-1,istripe,my_thread), &
                                                     w, nbw, nl, stripe_width, nbw)
#else
169
170
171
172
173
174
175
               call double_hh_trafo_generic_simple(a(1:stripe_width,j+off+a_off-1:j+off+a_off-1+nbw,istripe,my_thread), &
                                                     w, nbw, nl, stripe_width, nbw)

#endif

#else /* WITH_OPENMP */
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
176
177
               call double_hh_trafo_generic_simple(a(1,j+off+a_off-1,istripe), &
                                                     w, nbw, nl, stripe_width, nbw)
178
179
180
181
#else
               call double_hh_trafo_generic_simple(a(1:stripe_width,j+off+a_off-1:j+off+a_off-1+nbw,istripe), &
                                                     w, nbw, nl, stripe_width, nbw)

182
#endif
183
184
185

#endif /* WITH_OPENMP */

186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
             enddo
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_GENERIC_SIMPLE_KERNEL */


#if defined(WITH_REAL_SSE_KERNEL)
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
             do j = ncols, 2, -2
               w(:,1) = bcast_buffer(1:nbw,j+off)
               w(:,2) = bcast_buffer(1:nbw,j+off-1)
#ifdef WITH_OPENMP
               call double_hh_trafo(a(1,j+off+a_off-1,istripe,my_thread), w, nbw, nl, &
                                      stripe_width, nbw)
#else
               call double_hh_trafo(a(1,j+off+a_off-1,istripe), w, nbw, nl, &
                                      stripe_width, nbw)
#endif
             enddo
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_SSE_KERNEL */

213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
#if defined(WITH_REAL_SSE_BLOCK2_KERNEL)
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE_BLOCK2) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
             do j = ncols, 2, -2
               w(:,1) = bcast_buffer(1:nbw,j+off)
               w(:,2) = bcast_buffer(1:nbw,j+off-1)
#ifdef WITH_OPENMP
               call double_hh_trafo_real_sse_2hv(a(1,j+off+a_off-1,istripe,my_thread), &
                                                       w, nbw, nl, stripe_width, nbw)
#else
               call double_hh_trafo_real_sse_2hv(a(1,j+off+a_off-1,istripe), &
                                                       w, nbw, nl, stripe_width, nbw)
#endif
             enddo
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_SSE_BLOCK2_KERNEL */
232

233
#if defined(WITH_REAL_AVX_BLOCK2_KERNEL) || defined(WITH_REAL_AVX2_BLOCK2_KERNEL)
234
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
235
236
           if ((THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX_BLOCK2) .or. &
               (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX2_BLOCK2))  then
237
238
239
240
241
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
             do j = ncols, 2, -2
               w(:,1) = bcast_buffer(1:nbw,j+off)
               w(:,2) = bcast_buffer(1:nbw,j+off-1)
#ifdef WITH_OPENMP
242
               call double_hh_trafo_real_avx_avx2_2hv(a(1,j+off+a_off-1,istripe,my_thread), &
243
244
                                                       w, nbw, nl, stripe_width, nbw)
#else
245
               call double_hh_trafo_real_avx_avx2_2hv(a(1,j+off+a_off-1,istripe), &
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
                                                       w, nbw, nl, stripe_width, nbw)
#endif
             enddo
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_AVX_BLOCK2_KERNEL */

#if defined(WITH_REAL_BGP_KERNEL)
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_BGP) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
             do j = ncols, 2, -2
               w(:,1) = bcast_buffer(1:nbw,j+off)
               w(:,2) = bcast_buffer(1:nbw,j+off-1)
#ifdef WITH_OPENMP
               call double_hh_trafo_bgp(a(1,j+off+a_off-1,istripe,my_thread), w, nbw, nl, &
                                          stripe_width, nbw)
#else
               call double_hh_trafo_bgp(a(1,j+off+a_off-1,istripe), w, nbw, nl, &
                                          stripe_width, nbw)
#endif
             enddo
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_BGP_KERNEL */


#if defined(WITH_REAL_BGQ_KERNEL)
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_BGQ) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
             do j = ncols, 2, -2
               w(:,1) = bcast_buffer(1:nbw,j+off)
               w(:,2) = bcast_buffer(1:nbw,j+off-1)
#ifdef WITH_OPENMP
               call double_hh_trafo_bgq(a(1,j+off+a_off-1,istripe,my_thread), w, nbw, nl, &
                                          stripe_width, nbw)
#else
               call double_hh_trafo_bgq(a(1,j+off+a_off-1,istripe), w, nbw, nl, &
                                          stripe_width, nbw)
#endif
             enddo
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
           endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_BGQ_KERNEL */


!#if defined(WITH_AVX_SANDYBRIDGE)
297
!              call double_hh_trafo_real_avx_avx2_2hv(a(1,j+off+a_off-1,istripe), w, nbw, nl, stripe_width, nbw)
298
299
300
!#endif

#ifdef WITH_OPENMP
301
           if (j==1) call single_hh_trafo_real_cpu_openmp(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe,my_thread), &
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
                                      bcast_buffer(1:nbw,off+1), nbw, nl,     &
                                      stripe_width)
#else
           if (j==1) call single_hh_trafo_real_cpu(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe),           &
                                      bcast_buffer(1:nbw,off+1), nbw, nl,     &
                                      stripe_width)
#endif


#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
         endif !
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */



317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
#if defined(WITH_REAL_SSE_BLOCK4_KERNEL)
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
         if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE_BLOCK4) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
           ! X86 INTRINSIC CODE, USING 4 HOUSEHOLDER VECTORS
           do j = ncols, 4, -4
             w(:,1) = bcast_buffer(1:nbw,j+off)
             w(:,2) = bcast_buffer(1:nbw,j+off-1)
             w(:,3) = bcast_buffer(1:nbw,j+off-2)
             w(:,4) = bcast_buffer(1:nbw,j+off-3)
#ifdef WITH_OPENMP
             call quad_hh_trafo_real_sse_4hv(a(1,j+off+a_off-3,istripe,my_thread), w, &
                                                  nbw, nl, stripe_width, nbw)
#else
             call quad_hh_trafo_real_sse_4hv(a(1,j+off+a_off-3,istripe), w, &
                                                  nbw, nl, stripe_width, nbw)
#endif
           enddo
           do jj = j, 2, -2
             w(:,1) = bcast_buffer(1:nbw,jj+off)
             w(:,2) = bcast_buffer(1:nbw,jj+off-1)
#ifdef WITH_OPENMP
             call double_hh_trafo_real_sse_2hv(a(1,jj+off+a_off-1,istripe,my_thread), &
                                                    w, nbw, nl, stripe_width, nbw)
#else
             call double_hh_trafo_real_sse_2hv(a(1,jj+off+a_off-1,istripe), &
                                                    w, nbw, nl, stripe_width, nbw)
#endif
           enddo
#ifdef WITH_OPENMP
           if (jj==1) call single_hh_trafo_real_cpu_openmp(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe,my_thread), &
                                          bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
#else
           if (jj==1) call single_hh_trafo_real_cpu(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe), &
                                          bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
#endif
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
         endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_SSE_BLOCK4_KERNEL */

358
#if defined(WITH_REAL_AVX_BLOCK4_KERNEL) || defined(WITH_REAL_AVX2_BLOCK4_KERNEL)
359
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
360
361
         if ((THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX_BLOCK4) .or. &
             (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX2_BLOCK4)) then
362
363
364
365
366
367
368
369
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
           ! X86 INTRINSIC CODE, USING 4 HOUSEHOLDER VECTORS
           do j = ncols, 4, -4
             w(:,1) = bcast_buffer(1:nbw,j+off)
             w(:,2) = bcast_buffer(1:nbw,j+off-1)
             w(:,3) = bcast_buffer(1:nbw,j+off-2)
             w(:,4) = bcast_buffer(1:nbw,j+off-3)
#ifdef WITH_OPENMP
370
             call quad_hh_trafo_real_avx_avx2_4hv(a(1,j+off+a_off-3,istripe,my_thread), w, &
371
372
                                                  nbw, nl, stripe_width, nbw)
#else
373
             call quad_hh_trafo_real_avx_avx2_4hv(a(1,j+off+a_off-3,istripe), w, &
374
375
376
377
378
379
380
                                                  nbw, nl, stripe_width, nbw)
#endif
           enddo
           do jj = j, 2, -2
             w(:,1) = bcast_buffer(1:nbw,jj+off)
             w(:,2) = bcast_buffer(1:nbw,jj+off-1)
#ifdef WITH_OPENMP
381
             call double_hh_trafo_real_avx_avx2_2hv(a(1,jj+off+a_off-1,istripe,my_thread), &
382
383
                                                    w, nbw, nl, stripe_width, nbw)
#else
384
             call double_hh_trafo_real_avx_avx2_2hv(a(1,jj+off+a_off-1,istripe), &
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
                                                    w, nbw, nl, stripe_width, nbw)
#endif
           enddo
#ifdef WITH_OPENMP
           if (jj==1) call single_hh_trafo_real_cpu_openmp(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe,my_thread), &
                                          bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
#else
           if (jj==1) call single_hh_trafo_real_cpu(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe), &
                                          bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
#endif
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
         endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_AVX_BLOCK4_KERNEL */

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
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
#if defined(WITH_REAL_SSE_BLOCK6_KERNEL)
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
         if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE_BLOCK6) then
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
           ! X86 INTRINSIC CODE, USING 6 HOUSEHOLDER VECTORS
           do j = ncols, 6, -6
             w(:,1) = bcast_buffer(1:nbw,j+off)
             w(:,2) = bcast_buffer(1:nbw,j+off-1)
             w(:,3) = bcast_buffer(1:nbw,j+off-2)
             w(:,4) = bcast_buffer(1:nbw,j+off-3)
             w(:,5) = bcast_buffer(1:nbw,j+off-4)
             w(:,6) = bcast_buffer(1:nbw,j+off-5)
#ifdef WITH_OPENMP
             call hexa_hh_trafo_real_sse_6hv(a(1,j+off+a_off-5,istripe,my_thread), w, &
                                                  nbw, nl, stripe_width, nbw)
#else
             call hexa_hh_trafo_real_sse_6hv(a(1,j+off+a_off-5,istripe), w, &
                                                  nbw, nl, stripe_width, nbw)
#endif
           enddo
           do jj = j, 4, -4
             w(:,1) = bcast_buffer(1:nbw,jj+off)
             w(:,2) = bcast_buffer(1:nbw,jj+off-1)
             w(:,3) = bcast_buffer(1:nbw,jj+off-2)
             w(:,4) = bcast_buffer(1:nbw,jj+off-3)
#ifdef WITH_OPENMP
             call quad_hh_trafo_real_sse_4hv(a(1,jj+off+a_off-3,istripe,my_thread), w, &
                                                  nbw, nl, stripe_width, nbw)
#else
             call quad_hh_trafo_real_sse_4hv(a(1,jj+off+a_off-3,istripe), w, &
                                                  nbw, nl, stripe_width, nbw)
#endif
           enddo
           do jjj = jj, 2, -2
             w(:,1) = bcast_buffer(1:nbw,jjj+off)
             w(:,2) = bcast_buffer(1:nbw,jjj+off-1)
#ifdef WITH_OPENMP
             call double_hh_trafo_real_sse_2hv(a(1,jjj+off+a_off-1,istripe,my_thread), &
                                                    w, nbw, nl, stripe_width, nbw)
#else
             call double_hh_trafo_real_sse_2hv(a(1,jjj+off+a_off-1,istripe), &
                                                    w, nbw, nl, stripe_width, nbw)
#endif
           enddo
#ifdef WITH_OPENMP
           if (jjj==1) call single_hh_trafo_real_cpu_openmp(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe,my_thread), &
                                           bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
#else
           if (jjj==1) call single_hh_trafo_real_cpu(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe), &
                                           bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
#endif
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
         endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_SSE_BLOCK4_KERNEL */
455

456
#if defined(WITH_REAL_AVX_BLOCK6_KERNEL) || defined(WITH_REAL_AVX2_BLOCK6_KERNEL)
457
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
458
459
         if ((THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX_BLOCK6) .or. &
             (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX2_BLOCK6)) then
460
461
462
463
464
465
466
467
468
469
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
           ! X86 INTRINSIC CODE, USING 6 HOUSEHOLDER VECTORS
           do j = ncols, 6, -6
             w(:,1) = bcast_buffer(1:nbw,j+off)
             w(:,2) = bcast_buffer(1:nbw,j+off-1)
             w(:,3) = bcast_buffer(1:nbw,j+off-2)
             w(:,4) = bcast_buffer(1:nbw,j+off-3)
             w(:,5) = bcast_buffer(1:nbw,j+off-4)
             w(:,6) = bcast_buffer(1:nbw,j+off-5)
#ifdef WITH_OPENMP
470
             call hexa_hh_trafo_real_avx_avx2_6hv(a(1,j+off+a_off-5,istripe,my_thread), w, &
471
472
                                                  nbw, nl, stripe_width, nbw)
#else
473
             call hexa_hh_trafo_real_avx_avx2_6hv(a(1,j+off+a_off-5,istripe), w, &
474
475
476
477
478
479
480
481
482
                                                  nbw, nl, stripe_width, nbw)
#endif
           enddo
           do jj = j, 4, -4
             w(:,1) = bcast_buffer(1:nbw,jj+off)
             w(:,2) = bcast_buffer(1:nbw,jj+off-1)
             w(:,3) = bcast_buffer(1:nbw,jj+off-2)
             w(:,4) = bcast_buffer(1:nbw,jj+off-3)
#ifdef WITH_OPENMP
483
             call quad_hh_trafo_real_avx_avx2_4hv(a(1,jj+off+a_off-3,istripe,my_thread), w, &
484
485
                                                  nbw, nl, stripe_width, nbw)
#else
486
             call quad_hh_trafo_real_avx_avx2_4hv(a(1,jj+off+a_off-3,istripe), w, &
487
488
489
490
491
492
493
                                                  nbw, nl, stripe_width, nbw)
#endif
           enddo
           do jjj = jj, 2, -2
             w(:,1) = bcast_buffer(1:nbw,jjj+off)
             w(:,2) = bcast_buffer(1:nbw,jjj+off-1)
#ifdef WITH_OPENMP
494
             call double_hh_trafo_real_avx_avx2_2hv(a(1,jjj+off+a_off-1,istripe,my_thread), &
495
496
                                                    w, nbw, nl, stripe_width, nbw)
#else
497
             call double_hh_trafo_real_avx_avx2_2hv(a(1,jjj+off+a_off-1,istripe), &
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
                                                    w, nbw, nl, stripe_width, nbw)
#endif
           enddo
#ifdef WITH_OPENMP
           if (jjj==1) call single_hh_trafo_real_cpu_openmp(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe,my_thread), &
                                           bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
#else
           if (jjj==1) call single_hh_trafo_real_cpu(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe), &
                                           bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
#endif
#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
         endif
#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
#endif /* WITH_REAL_AVX_BLOCK4_KERNEL */

#ifdef WITH_OPENMP
         if (my_thread==1) then
#endif
           kernel_flops = kernel_flops + 4*int(nl,8)*int(ncols,8)*int(nbw,8)
           kernel_time = kernel_time + mpi_wtime()-ttt
#ifdef WITH_OPENMP
         endif
#endif
#ifdef HAVE_DETAILED_TIMINGS
#ifdef WITH_OPENMP
         call timer%stop("compute_hh_trafo_real_cpu_openmp")
#else
         call timer%stop("compute_hh_trafo_real_cpu")
#endif
#endif

#ifdef WITH_OPENMP
       end subroutine compute_hh_trafo_real_cpu_openmp
#else
       end subroutine compute_hh_trafo_real_cpu
#endif

end module