elpa2_kernels_complex_avx-avx2_2hv_double_precision.c 50.7 KB
Newer Older
1
2
//    This file is part of ELPA.
//
Andreas Marek's avatar
Andreas Marek committed
3
//    The ELPA library was originally created by the ELPA consortium,
4
5
//    consisting of the following organizations:
//
6
7
//    - Max Planck Computing and Data Facility (MPCDF), formerly known as
//      Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
8
9
10
//    - Bergische Universität Wuppertal, Lehrstuhl für angewandte
//      Informatik,
//    - Technische Universität München, Lehrstuhl für Informatik mit
Andreas Marek's avatar
Andreas Marek committed
11
12
//      Schwerpunkt Wissenschaftliches Rechnen ,
//    - Fritz-Haber-Institut, Berlin, Abt. Theorie,
13
//    - Max-Plack-Institut für Mathematik in den Naturwissenschaften,
Andreas Marek's avatar
Andreas Marek committed
14
15
//      Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition,
//      and
16
17
//    - IBM Deutschland GmbH
//
18
//    This particular source code file contains additions, changes and
Andreas Marek's avatar
Andreas Marek committed
19
//    enhancements authored by Intel Corporation which is not part of
20
//    the ELPA consortium.
21
22
//
//    More information can be found here:
23
//    http://elpa.mpcdf.mpg.de/
24
25
//
//    ELPA is free software: you can redistribute it and/or modify
Andreas Marek's avatar
Andreas Marek committed
26
27
//    it under the terms of the version 3 of the license of the
//    GNU Lesser General Public License as published by the Free
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
//    Software Foundation.
//
//    ELPA is distributed in the hope that it will be useful,
//    but WITHOUT ANY WARRANTY; without even the implied warranty of
//    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
//    GNU Lesser General Public License for more details.
//
//    You should have received a copy of the GNU Lesser General Public License
//    along with ELPA.  If not, see <http://www.gnu.org/licenses/>
//
//    ELPA reflects a substantial effort on the part of the original
//    ELPA consortium, and we ask you to respect the spirit of the
//    license that we chose: i.e., please contribute any changes you
//    may have back to the original ELPA library distribution, and keep
//    any derivatives of ELPA under the same license that we chose for
//    the original distribution, the GNU Lesser General Public License.
//
//
// --------------------------------------------------------------------------------------------------
//
// This file contains the compute intensive kernels for the Householder transformations.
// It should be compiled with the highest possible optimization level.
//
// On Intel Nehalem or Intel Westmere or AMD Magny Cours use -O3 -msse3
// On Intel Sandy Bridge use -O3 -mavx
//
// Copyright of the original code rests with the authors inside the ELPA
// consortium. The copyright of any additional modifications shall rest
// with their original authors, but shall adhere to the licensing terms
// distributed along with the original code in the file "COPYING".
//
// Author: Alexander Heinecke (alexander.heinecke@mytum.de)
60
// Adapted for building a shared-library by Andreas Marek, MPCDF (andreas.marek@mpcdf.mpg.de)
61
// --------------------------------------------------------------------------------------------------
62
#include "config-f90.h"
63

Andreas Marek's avatar
Andreas Marek committed
64
#include <complex.h>
65
66
67
68
#include <x86intrin.h>

#define __forceinline __attribute__((always_inline))

69
70
#ifdef HAVE_AVX2

71
72
73
74
75
76
77
78
79
80
81
#ifdef __FMA4__
#define __ELPA_USE_FMA__
#define _mm256_FMADDSUB_pd(a,b,c) _mm256_maddsub_pd(a,b,c)
#define _mm256_FMSUBADD_pd(a,b,c) _mm256_msubadd_pd(a,b,c)
#endif

#ifdef __AVX2__
#define __ELPA_USE_FMA__
#define _mm256_FMADDSUB_pd(a,b,c) _mm256_fmaddsub_pd(a,b,c)
#define _mm256_FMSUBADD_pd(a,b,c) _mm256_fmsubadd_pd(a,b,c)
#endif
82

83
84
#endif

85
//Forward declaration
86
87
88
89
static __forceinline void hh_trafo_complex_kernel_8_AVX_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);
static __forceinline void hh_trafo_complex_kernel_6_AVX_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);
static __forceinline void hh_trafo_complex_kernel_4_AVX_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);
static __forceinline void hh_trafo_complex_kernel_2_AVX_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);
90

91
/*
92
!f>#if defined(HAVE_AVX) || defined(HAVE_AVX2)
93
!f> interface
94
95
!f>   subroutine double_hh_trafo_complex_avx_avx2_2hv_double(q, hh, pnb, pnq, pldq, pldh) &
!f>                             bind(C, name="double_hh_trafo_complex_avx_avx2_2hv_double")
96
97
!f>     use, intrinsic :: iso_c_binding
!f>     integer(kind=c_int)     :: pnb, pnq, pldq, pldh
98
99
!f>     ! complex(kind=c_double_complex)     :: q(*)
!f>     type(c_ptr), value                   :: q
100
!f>     complex(kind=c_double_complex)     :: hh(pnb,2)
101
102
103
104
105
!f>   end subroutine
!f> end interface
!f>#endif
*/

106
void double_hh_trafo_complex_avx_avx2_2hv_double(double complex* q, double complex* hh, int* pnb, int* pnq, int* pldq, int* pldh)
107
108
109
110
111
112
113
{
	int i;
	int nb = *pnb;
	int nq = *pldq;
	int ldq = *pldq;
	int ldh = *pldh;

Andreas Marek's avatar
Andreas Marek committed
114
	double complex s = conj(hh[(ldh)+1])*1.0;
115
116
117
118
119
	for (i = 2; i < nb; i++)
	{
		s += hh[i-1] * conj(hh[(i+ldh)]);
	}

120
//#if 1
121
122
	for (i = 0; i < nq-4; i+=8)
	{
123
		hh_trafo_complex_kernel_8_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s);
124
	}
125
126
127
	if (nq-i == 0) {
          return;
	} else
128
	{
129
		hh_trafo_complex_kernel_4_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s);
130
	}
131
132
133
134
135
136
137
138
139
140
141
142
143
144
//#else
//	for (i = 0; i < nq-4; i+=6)
//	{
//		hh_trafo_complex_kernel_6_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s);
//	}
//	if (nq-i > 2)
//	{
//		hh_trafo_complex_kernel_4_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s);
//	}
//	else if (nq-i > 0)
//	{
//		hh_trafo_complex_kernel_2_AVX_2hv_double(&q[i], hh, nb, ldq, ldh, s);
//	}
//#endif
145
146
}

147
static __forceinline void hh_trafo_complex_kernel_8_AVX_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s)
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;
	double* s_dbl = (double*)(&s);

	__m256d x1, x2, x3, x4;
	__m256d y1, y2, y3, y4;
	__m256d q1, q2, q3, q4;
	__m256d h1_real, h1_imag, h2_real, h2_imag;
	__m256d tmp1, tmp2, tmp3, tmp4;
	int i=0;

	__m256d sign = (__m256d)_mm256_set_epi64x(0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000);

	x1 = _mm256_load_pd(&q_dbl[(2*ldq)+0]);
	x2 = _mm256_load_pd(&q_dbl[(2*ldq)+4]);
	x3 = _mm256_load_pd(&q_dbl[(2*ldq)+8]);
	x4 = _mm256_load_pd(&q_dbl[(2*ldq)+12]);

	h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+1)*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+1)*2)+1]);
169
#ifndef __ELPA_USE_FMA__
170
171
172
173
174
175
176
177
178
179
	// conjugate
	h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

	y1 = _mm256_load_pd(&q_dbl[0]);
	y2 = _mm256_load_pd(&q_dbl[4]);
	y3 = _mm256_load_pd(&q_dbl[8]);
	y4 = _mm256_load_pd(&q_dbl[12]);

	tmp1 = _mm256_mul_pd(h2_imag, x1);
180
181
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
182
183
184
185
#else
	y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, x2);
186
187
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
188
189
190
191
#else
	y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h2_imag, x3);
192
193
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
194
195
196
197
#else
	y3 = _mm256_add_pd(y3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
	tmp4 = _mm256_mul_pd(h2_imag, x4);
198
199
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_add_pd(y4, _mm256_FMSUBADD_pd(h2_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
200
201
202
203
204
205
206
207
208
209
210
211
212
#else
	y4 = _mm256_add_pd(y4, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif

	for (i = 2; i < nb; i++)
	{
		q1 = _mm256_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm256_load_pd(&q_dbl[(2*i*ldq)+4]);
		q3 = _mm256_load_pd(&q_dbl[(2*i*ldq)+8]);
		q4 = _mm256_load_pd(&q_dbl[(2*i*ldq)+12]);

		h1_real = _mm256_broadcast_sd(&hh_dbl[(i-1)*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[((i-1)*2)+1]);
213
#ifndef __ELPA_USE_FMA__
214
215
216
217
218
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
219
220
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
221
222
223
224
#else
		x1 = _mm256_add_pd(x1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h1_imag, q2);
225
226
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
227
228
229
230
#else
		x2 = _mm256_add_pd(x2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
		tmp3 = _mm256_mul_pd(h1_imag, q3);
231
232
#ifdef __ELPA_USE_FMA__
		x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
233
234
235
236
#else
		x3 = _mm256_add_pd(x3, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
		tmp4 = _mm256_mul_pd(h1_imag, q4);
237
238
#ifdef __ELPA_USE_FMA__
		x4 = _mm256_add_pd(x4, _mm256_FMSUBADD_pd(h1_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
239
240
241
242
243
244
#else
		x4 = _mm256_add_pd(x4, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif

		h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+i)*2]);
		h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+i)*2)+1]);
245
#ifndef __ELPA_USE_FMA__
246
247
248
249
250
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
251
252
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
253
254
255
256
#else
		y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h2_imag, q2);
257
258
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
259
260
261
262
#else
		y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
		tmp3 = _mm256_mul_pd(h2_imag, q3);
263
264
#ifdef __ELPA_USE_FMA__
		y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
265
266
267
268
#else
		y3 = _mm256_add_pd(y3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
		tmp4 = _mm256_mul_pd(h2_imag, q4);
269
270
#ifdef __ELPA_USE_FMA__
		y4 = _mm256_add_pd(y4, _mm256_FMSUBADD_pd(h2_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
271
272
273
274
275
276
277
#else
		y4 = _mm256_add_pd(y4, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif
	}

	h1_real = _mm256_broadcast_sd(&hh_dbl[(nb-1)*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[((nb-1)*2)+1]);
278
#ifndef __ELPA_USE_FMA__
279
280
281
282
283
284
285
286
287
288
	// conjugate
	h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

	q1 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+0]);
	q2 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+4]);
	q3 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+8]);
	q4 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+12]);

	tmp1 = _mm256_mul_pd(h1_imag, q1);
289
290
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
291
292
293
294
#else
	x1 = _mm256_add_pd(x1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, q2);
295
296
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
297
298
299
300
#else
	x2 = _mm256_add_pd(x2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h1_imag, q3);
301
302
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
303
304
305
306
#else
	x3 = _mm256_add_pd(x3, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
	tmp4 = _mm256_mul_pd(h1_imag, q4);
307
308
#ifdef __ELPA_USE_FMA__
	x4 = _mm256_add_pd(x4, _mm256_FMSUBADD_pd(h1_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
309
310
311
312
313
314
315
316
317
318
#else
	x4 = _mm256_add_pd(x4, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif

	h1_real = _mm256_broadcast_sd(&hh_dbl[0]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[1]);
	h1_real = _mm256_xor_pd(h1_real, sign);
	h1_imag = _mm256_xor_pd(h1_imag, sign);

	tmp1 = _mm256_mul_pd(h1_imag, x1);
319
320
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
321
322
323
324
#else
	x1 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, x2);
325
326
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
327
328
329
330
#else
	x2 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5));
#endif
	tmp3 = _mm256_mul_pd(h1_imag, x3);
331
332
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
333
334
335
336
#else
	x3 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5));
#endif
	tmp4 = _mm256_mul_pd(h1_imag, x4);
337
338
#ifdef __ELPA_USE_FMA__
	x4 = _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5));
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
#else
	x4 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x4), _mm256_shuffle_pd(tmp4, tmp4, 0x5));
#endif

	h1_real = _mm256_broadcast_sd(&hh_dbl[ldh*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[(ldh*2)+1]);
	h2_real = _mm256_broadcast_sd(&hh_dbl[ldh*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[(ldh*2)+1]);

	h1_real = _mm256_xor_pd(h1_real, sign);
	h1_imag = _mm256_xor_pd(h1_imag, sign);
	h2_real = _mm256_xor_pd(h2_real, sign);
	h2_imag = _mm256_xor_pd(h2_imag, sign);

	__m128d tmp_s_128 = _mm_loadu_pd(s_dbl);
	tmp2 = _mm256_broadcast_pd(&tmp_s_128);
	tmp1 = _mm256_mul_pd(h2_imag, tmp2);
356
357
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
358
359
360
361
362
363
364
365
#else
	tmp2 = _mm256_addsub_pd( _mm256_mul_pd(h2_real, tmp2), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	_mm_storeu_pd(s_dbl, _mm256_castpd256_pd128(tmp2));
	h2_real = _mm256_broadcast_sd(&s_dbl[0]);
	h2_imag = _mm256_broadcast_sd(&s_dbl[1]);

	tmp1 = _mm256_mul_pd(h1_imag, y1);
366
367
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
368
369
370
371
#else
	y1 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, y2);
372
373
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
374
375
376
377
#else
	y2 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5));
#endif
	tmp3 = _mm256_mul_pd(h1_imag, y3);
378
379
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_FMADDSUB_pd(h1_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
380
381
382
383
#else
	y3 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y3), _mm256_shuffle_pd(tmp3, tmp3, 0x5));
#endif
	tmp4 = _mm256_mul_pd(h1_imag, y4);
384
385
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_FMADDSUB_pd(h1_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5));
386
387
388
389
390
#else
	y4 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y4), _mm256_shuffle_pd(tmp4, tmp4, 0x5));
#endif

	tmp1 = _mm256_mul_pd(h2_imag, x1);
391
392
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMADDSUB_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
393
394
395
396
#else
	y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, x2);
397
398
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMADDSUB_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
399
400
401
402
#else
	y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h2_imag, x3);
403
404
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMADDSUB_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
405
406
407
408
#else
	y3 = _mm256_add_pd(y3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
	tmp4 = _mm256_mul_pd(h2_imag, x4);
409
410
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_add_pd(y4, _mm256_FMADDSUB_pd(h2_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
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
#else
	y4 = _mm256_add_pd(y4, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif

	q1 = _mm256_load_pd(&q_dbl[0]);
	q2 = _mm256_load_pd(&q_dbl[4]);
	q3 = _mm256_load_pd(&q_dbl[8]);
	q4 = _mm256_load_pd(&q_dbl[12]);

	q1 = _mm256_add_pd(q1, y1);
	q2 = _mm256_add_pd(q2, y2);
	q3 = _mm256_add_pd(q3, y3);
	q4 = _mm256_add_pd(q4, y4);

	_mm256_store_pd(&q_dbl[0], q1);
	_mm256_store_pd(&q_dbl[4], q2);
	_mm256_store_pd(&q_dbl[8], q3);
	_mm256_store_pd(&q_dbl[12], q4);

	h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+1)*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+1)*2)+1]);

	q1 = _mm256_load_pd(&q_dbl[(ldq*2)+0]);
	q2 = _mm256_load_pd(&q_dbl[(ldq*2)+4]);
	q3 = _mm256_load_pd(&q_dbl[(ldq*2)+8]);
	q4 = _mm256_load_pd(&q_dbl[(ldq*2)+12]);

	q1 = _mm256_add_pd(q1, x1);
	q2 = _mm256_add_pd(q2, x2);
	q3 = _mm256_add_pd(q3, x3);
	q4 = _mm256_add_pd(q4, x4);

	tmp1 = _mm256_mul_pd(h2_imag, y1);
444
445
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
446
447
448
449
#else
	q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, y2);
450
#ifdef __ELPA_USE_FMA_
451
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
452
453
454
455
#else
	q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h2_imag, y3);
456
457
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
458
459
460
461
#else
	q3 = _mm256_add_pd(q3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
	tmp4 = _mm256_mul_pd(h2_imag, y4);
462
463
#ifdef __ELPA_USE_FMA__
	q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h2_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
#else
	q4 = _mm256_add_pd(q4, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif

	_mm256_store_pd(&q_dbl[(ldq*2)+0], q1);
	_mm256_store_pd(&q_dbl[(ldq*2)+4], q2);
	_mm256_store_pd(&q_dbl[(ldq*2)+8], q3);
	_mm256_store_pd(&q_dbl[(ldq*2)+12], q4);

	for (i = 2; i < nb; i++)
	{
		q1 = _mm256_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm256_load_pd(&q_dbl[(2*i*ldq)+4]);
		q3 = _mm256_load_pd(&q_dbl[(2*i*ldq)+8]);
		q4 = _mm256_load_pd(&q_dbl[(2*i*ldq)+12]);

		h1_real = _mm256_broadcast_sd(&hh_dbl[(i-1)*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[((i-1)*2)+1]);

		tmp1 = _mm256_mul_pd(h1_imag, x1);
484
485
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
486
487
488
489
#else
		q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h1_imag, x2);
490
491
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
492
493
494
495
#else
		q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
		tmp3 = _mm256_mul_pd(h1_imag, x3);
496
497
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
498
499
500
501
#else
		q3 = _mm256_add_pd(q3, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
		tmp4 = _mm256_mul_pd(h1_imag, x4);
502
503
#ifdef __ELPA_USE_FMA__
		q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
504
505
506
507
508
509
510
511
#else
		q4 = _mm256_add_pd(q4, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif

		h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+i)*2]);
		h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+i)*2)+1]);

		tmp1 = _mm256_mul_pd(h2_imag, y1);
512
513
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
514
515
516
517
#else
		q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h2_imag, y2);
518
519
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
520
521
522
523
#else
		q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
		tmp3 = _mm256_mul_pd(h2_imag, y3);
524
525
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
526
527
528
529
#else
		q3 = _mm256_add_pd(q3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
		tmp4 = _mm256_mul_pd(h2_imag, y4);
530
531
#ifdef __ELPA_USE_FMA__
		q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h2_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
#else
		q4 = _mm256_add_pd(q4, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif

		_mm256_store_pd(&q_dbl[(2*i*ldq)+0], q1);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+4], q2);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+8], q3);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+12], q4);
	}
	h1_real = _mm256_broadcast_sd(&hh_dbl[(nb-1)*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[((nb-1)*2)+1]);

	q1 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+0]);
	q2 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+4]);
	q3 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+8]);
	q4 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+12]);

	tmp1 = _mm256_mul_pd(h1_imag, x1);
550
551
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
552
553
554
555
#else
	q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, x2);
556
557
#ifdef __ELPA_USE_FMA__
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
558
559
560
561
#else
	q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h1_imag, x3);
562
563
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
564
565
566
567
#else
	q3 = _mm256_add_pd(q3, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
	tmp4 = _mm256_mul_pd(h1_imag, x4);
568
569
#ifdef __ELPA_USE_FMA__
	q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
570
571
572
573
574
575
576
577
578
579
#else
	q4 = _mm256_add_pd(q4, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif

	_mm256_store_pd(&q_dbl[(2*nb*ldq)+0], q1);
	_mm256_store_pd(&q_dbl[(2*nb*ldq)+4], q2);
	_mm256_store_pd(&q_dbl[(2*nb*ldq)+8], q3);
	_mm256_store_pd(&q_dbl[(2*nb*ldq)+12], q4);
}

580
static __forceinline void hh_trafo_complex_kernel_6_AVX_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s)
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;
	double* s_dbl = (double*)(&s);

	__m256d x1, x2, x3;
	__m256d y1, y2, y3;
	__m256d q1, q2, q3;
	__m256d h1_real, h1_imag, h2_real, h2_imag;
	__m256d tmp1, tmp2, tmp3;
	int i=0;

	__m256d sign = (__m256d)_mm256_set_epi64x(0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000);

	x1 = _mm256_load_pd(&q_dbl[(2*ldq)+0]);
	x2 = _mm256_load_pd(&q_dbl[(2*ldq)+4]);
	x3 = _mm256_load_pd(&q_dbl[(2*ldq)+8]);

	h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+1)*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+1)*2)+1]);
601
#ifndef __ELPA_USE_FMA__
602
603
604
605
606
607
608
609
610
	// conjugate
	h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

	y1 = _mm256_load_pd(&q_dbl[0]);
	y2 = _mm256_load_pd(&q_dbl[4]);
	y3 = _mm256_load_pd(&q_dbl[8]);

	tmp1 = _mm256_mul_pd(h2_imag, x1);
611
612
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
613
614
615
616
#else
	y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, x2);
617
618
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
619
620
621
622
#else
	y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h2_imag, x3);
623
624
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
625
626
627
628
629
630
631
632
633
634
635
636
#else
	y3 = _mm256_add_pd(y3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif

	for (i = 2; i < nb; i++)
	{
		q1 = _mm256_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm256_load_pd(&q_dbl[(2*i*ldq)+4]);
		q3 = _mm256_load_pd(&q_dbl[(2*i*ldq)+8]);

		h1_real = _mm256_broadcast_sd(&hh_dbl[(i-1)*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[((i-1)*2)+1]);
637
#ifndef __ELPA_USE_FMA__
638
639
640
641
642
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
643
644
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
645
646
647
648
#else
		x1 = _mm256_add_pd(x1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h1_imag, q2);
649
650
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
651
652
653
654
#else
		x2 = _mm256_add_pd(x2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
		tmp3 = _mm256_mul_pd(h1_imag, q3);
655
656
#ifdef __ELPA_USE_FMA__
		x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
657
658
659
660
661
662
#else
		x3 = _mm256_add_pd(x3, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif

		h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+i)*2]);
		h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+i)*2)+1]);
663
#ifndef __ELPA_USE_FMA__
664
665
666
667
668
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
669
670
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
671
672
673
674
#else
		y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h2_imag, q2);
675
676
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
677
678
679
680
#else
		y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
		tmp3 = _mm256_mul_pd(h2_imag, q3);
681
682
#ifdef __ELPA_USE_FMA__
		y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
683
684
685
686
687
688
689
#else
		y3 = _mm256_add_pd(y3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif
	}

	h1_real = _mm256_broadcast_sd(&hh_dbl[(nb-1)*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[((nb-1)*2)+1]);
690
#ifndef __ELPA_USE_FMA__
691
692
693
694
695
696
697
698
699
	// conjugate
	h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

	q1 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+0]);
	q2 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+4]);
	q3 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+8]);

	tmp1 = _mm256_mul_pd(h1_imag, q1);
700
701
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
702
703
704
705
#else
	x1 = _mm256_add_pd(x1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, q2);
706
707
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
708
709
710
711
#else
	x2 = _mm256_add_pd(x2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h1_imag, q3);
712
713
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
714
715
716
717
718
719
720
721
722
723
#else
	x3 = _mm256_add_pd(x3, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif

	h1_real = _mm256_broadcast_sd(&hh_dbl[0]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[1]);
	h1_real = _mm256_xor_pd(h1_real, sign);
	h1_imag = _mm256_xor_pd(h1_imag, sign);

	tmp1 = _mm256_mul_pd(h1_imag, x1);
724
725
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
726
727
728
729
#else
	x1 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, x2);
730
731
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
732
733
734
735
#else
	x2 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5));
#endif
	tmp3 = _mm256_mul_pd(h1_imag, x3);
736
737
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
#else
	x3 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5));
#endif

	h1_real = _mm256_broadcast_sd(&hh_dbl[ldh*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[(ldh*2)+1]);
	h2_real = _mm256_broadcast_sd(&hh_dbl[ldh*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[(ldh*2)+1]);

	h1_real = _mm256_xor_pd(h1_real, sign);
	h1_imag = _mm256_xor_pd(h1_imag, sign);
	h2_real = _mm256_xor_pd(h2_real, sign);
	h2_imag = _mm256_xor_pd(h2_imag, sign);

	__m128d tmp_s_128 = _mm_loadu_pd(s_dbl);
	tmp2 = _mm256_broadcast_pd(&tmp_s_128);
	tmp1 = _mm256_mul_pd(h2_imag, tmp2);
755
756
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
757
758
759
760
761
762
763
764
#else
	tmp2 = _mm256_addsub_pd( _mm256_mul_pd(h2_real, tmp2), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	_mm_storeu_pd(s_dbl, _mm256_castpd256_pd128(tmp2));
	h2_real = _mm256_broadcast_sd(&s_dbl[0]);
	h2_imag = _mm256_broadcast_sd(&s_dbl[1]);

	tmp1 = _mm256_mul_pd(h1_imag, y1);
765
766
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
767
768
769
770
#else
	y1 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, y2);
771
772
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
773
774
775
776
#else
	y2 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5));
#endif
	tmp3 = _mm256_mul_pd(h1_imag, y3);
777
778
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_FMADDSUB_pd(h1_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
779
780
781
782
783
#else
	y3 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y3), _mm256_shuffle_pd(tmp3, tmp3, 0x5));
#endif

	tmp1 = _mm256_mul_pd(h2_imag, x1);
784
785
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMADDSUB_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
786
787
788
789
#else
	y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, x2);
790
791
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMADDSUB_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
792
793
794
795
#else
	y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h2_imag, x3);
796
797
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMADDSUB_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
#else
	y3 = _mm256_add_pd(y3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif

	q1 = _mm256_load_pd(&q_dbl[0]);
	q2 = _mm256_load_pd(&q_dbl[4]);
	q3 = _mm256_load_pd(&q_dbl[8]);

	q1 = _mm256_add_pd(q1, y1);
	q2 = _mm256_add_pd(q2, y2);
	q3 = _mm256_add_pd(q3, y3);

	_mm256_store_pd(&q_dbl[0], q1);
	_mm256_store_pd(&q_dbl[4], q2);
	_mm256_store_pd(&q_dbl[8], q3);

	h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+1)*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+1)*2)+1]);

	q1 = _mm256_load_pd(&q_dbl[(ldq*2)+0]);
	q2 = _mm256_load_pd(&q_dbl[(ldq*2)+4]);
	q3 = _mm256_load_pd(&q_dbl[(ldq*2)+8]);

	q1 = _mm256_add_pd(q1, x1);
	q2 = _mm256_add_pd(q2, x2);
	q3 = _mm256_add_pd(q3, x3);

	tmp1 = _mm256_mul_pd(h2_imag, y1);
826
827
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
828
829
830
831
832
#else
	q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, y2);
#ifdef __FMA4_
833
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
834
835
836
837
#else
	q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h2_imag, y3);
838
839
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
#else
	q3 = _mm256_add_pd(q3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif

	_mm256_store_pd(&q_dbl[(ldq*2)+0], q1);
	_mm256_store_pd(&q_dbl[(ldq*2)+4], q2);
	_mm256_store_pd(&q_dbl[(ldq*2)+8], q3);

	for (i = 2; i < nb; i++)
	{
		q1 = _mm256_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm256_load_pd(&q_dbl[(2*i*ldq)+4]);
		q3 = _mm256_load_pd(&q_dbl[(2*i*ldq)+8]);

		h1_real = _mm256_broadcast_sd(&hh_dbl[(i-1)*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[((i-1)*2)+1]);

		tmp1 = _mm256_mul_pd(h1_imag, x1);
858
859
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
860
861
862
863
#else
		q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h1_imag, x2);
864
865
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
866
867
868
869
#else
		q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
		tmp3 = _mm256_mul_pd(h1_imag, x3);
870
871
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
872
873
874
875
876
877
878
879
#else
		q3 = _mm256_add_pd(q3, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif

		h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+i)*2]);
		h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+i)*2)+1]);

		tmp1 = _mm256_mul_pd(h2_imag, y1);
880
881
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
882
883
884
885
#else
		q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h2_imag, y2);
886
887
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
888
889
890
891
#else
		q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
		tmp3 = _mm256_mul_pd(h2_imag, y3);
892
893
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
#else
		q3 = _mm256_add_pd(q3, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif

		_mm256_store_pd(&q_dbl[(2*i*ldq)+0], q1);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+4], q2);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+8], q3);
	}
	h1_real = _mm256_broadcast_sd(&hh_dbl[(nb-1)*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[((nb-1)*2)+1]);

	q1 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+0]);
	q2 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+4]);
	q3 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+8]);

	tmp1 = _mm256_mul_pd(h1_imag, x1);
910
911
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
912
913
914
915
#else
	q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, x2);
916
917
#ifdef __ELPA_USE_FMA__
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
918
919
920
921
#else
	q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	tmp3 = _mm256_mul_pd(h1_imag, x3);
922
923
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
924
925
926
927
928
929
930
931
932
#else
	q3 = _mm256_add_pd(q3, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x3), _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
#endif

	_mm256_store_pd(&q_dbl[(2*nb*ldq)+0], q1);
	_mm256_store_pd(&q_dbl[(2*nb*ldq)+4], q2);
	_mm256_store_pd(&q_dbl[(2*nb*ldq)+8], q3);
}

933
static __forceinline void hh_trafo_complex_kernel_4_AVX_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s)
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;
	double* s_dbl = (double*)(&s);

	__m256d x1, x2;
	__m256d y1, y2;
	__m256d q1, q2;
	__m256d h1_real, h1_imag, h2_real, h2_imag;
	__m256d tmp1, tmp2;
	int i=0;

	__m256d sign = (__m256d)_mm256_set_epi64x(0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000);

	x1 = _mm256_load_pd(&q_dbl[(2*ldq)+0]);
	x2 = _mm256_load_pd(&q_dbl[(2*ldq)+4]);

	h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+1)*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+1)*2)+1]);
953
#ifndef __ELPA_USE_FMA__
954
955
956
957
958
959
960
961
	// conjugate
	h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

	y1 = _mm256_load_pd(&q_dbl[0]);
	y2 = _mm256_load_pd(&q_dbl[4]);

	tmp1 = _mm256_mul_pd(h2_imag, x1);
962
963
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
964
965
966
967
#else
	y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, x2);
968
969
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
970
971
972
973
974
975
976
977
978
979
980
#else
	y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif

	for (i = 2; i < nb; i++)
	{
		q1 = _mm256_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm256_load_pd(&q_dbl[(2*i*ldq)+4]);

		h1_real = _mm256_broadcast_sd(&hh_dbl[(i-1)*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[((i-1)*2)+1]);
981
#ifndef __ELPA_USE_FMA__
982
983
984
985
986
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
987
988
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
989
990
991
992
#else
		x1 = _mm256_add_pd(x1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h1_imag, q2);
993
994
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
995
996
997
998
999
1000
#else
		x2 = _mm256_add_pd(x2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif

		h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+i)*2]);
		h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+i)*2)+1]);
1001
#ifndef __ELPA_USE_FMA__
1002
1003
1004
1005
1006
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
1007
1008
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1009
1010
1011
1012
#else
		y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h2_imag, q2);
1013
1014
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1015
1016
1017
1018
1019
1020
1021
#else
		y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif
	}

	h1_real = _mm256_broadcast_sd(&hh_dbl[(nb-1)*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[((nb-1)*2)+1]);
1022
#ifndef __ELPA_USE_FMA__
1023
1024
1025
1026
1027
1028
1029
1030
	// conjugate
	h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

	q1 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+0]);
	q2 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+4]);

	tmp1 = _mm256_mul_pd(h1_imag, q1);
1031
1032
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1033
1034
1035
1036
#else
	x1 = _mm256_add_pd(x1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, q2);
1037
1038
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
#else
	x2 = _mm256_add_pd(x2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif

	h1_real = _mm256_broadcast_sd(&hh_dbl[0]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[1]);
	h1_real = _mm256_xor_pd(h1_real, sign);
	h1_imag = _mm256_xor_pd(h1_imag, sign);

	tmp1 = _mm256_mul_pd(h1_imag, x1);
1049
1050
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1051
1052
1053
1054
#else
	x1 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, x2);
1055
1056
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
#else
	x2 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5));
#endif

	h1_real = _mm256_broadcast_sd(&hh_dbl[ldh*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[(ldh*2)+1]);
	h2_real = _mm256_broadcast_sd(&hh_dbl[ldh*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[(ldh*2)+1]);

	h1_real = _mm256_xor_pd(h1_real, sign);
	h1_imag = _mm256_xor_pd(h1_imag, sign);
	h2_real = _mm256_xor_pd(h2_real, sign);
	h2_imag = _mm256_xor_pd(h2_imag, sign);

	__m128d tmp_s_128 = _mm_loadu_pd(s_dbl);
	tmp2 = _mm256_broadcast_pd(&tmp_s_128);
	tmp1 = _mm256_mul_pd(h2_imag, tmp2);
1074
1075
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1076
1077
1078
1079
1080
1081
1082
1083
#else
	tmp2 = _mm256_addsub_pd( _mm256_mul_pd(h2_real, tmp2), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	_mm_storeu_pd(s_dbl, _mm256_castpd256_pd128(tmp2));
	h2_real = _mm256_broadcast_sd(&s_dbl[0]);
	h2_imag = _mm256_broadcast_sd(&s_dbl[1]);

	tmp1 = _mm256_mul_pd(h1_imag, y1);
1084
1085
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1086
1087
1088
1089
#else
	y1 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, y2);
1090
1091
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
1092
1093
1094
1095
1096
#else
	y2 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5));
#endif

	tmp1 = _mm256_mul_pd(h2_imag, x1);
1097
1098
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMADDSUB_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1099
1100
1101
1102
#else
	y1 = _mm256_add_pd(y1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, x2);
1103
1104
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMADDSUB_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
#else
	y2 = _mm256_add_pd(y2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif

	q1 = _mm256_load_pd(&q_dbl[0]);
	q2 = _mm256_load_pd(&q_dbl[4]);

	q1 = _mm256_add_pd(q1, y1);
	q2 = _mm256_add_pd(q2, y2);

	_mm256_store_pd(&q_dbl[0], q1);
	_mm256_store_pd(&q_dbl[4], q2);

	h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+1)*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+1)*2)+1]);

	q1 = _mm256_load_pd(&q_dbl[(ldq*2)+0]);
	q2 = _mm256_load_pd(&q_dbl[(ldq*2)+4]);

	q1 = _mm256_add_pd(q1, x1);
	q2 = _mm256_add_pd(q2, x2);

	tmp1 = _mm256_mul_pd(h2_imag, y1);
1128
1129
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1130
1131
1132
1133
1134
#else
	q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h2_imag, y2);
#ifdef __FMA4_
1135
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
#else
	q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif

	_mm256_store_pd(&q_dbl[(ldq*2)+0], q1);
	_mm256_store_pd(&q_dbl[(ldq*2)+4], q2);

	for (i = 2; i < nb; i++)
	{
		q1 = _mm256_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm256_load_pd(&q_dbl[(2*i*ldq)+4]);

		h1_real = _mm256_broadcast_sd(&hh_dbl[(i-1)*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[((i-1)*2)+1]);

		tmp1 = _mm256_mul_pd(h1_imag, x1);
1152
1153
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1154
1155
1156
1157
#else
		q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h1_imag, x2);
1158
1159
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1160
1161
1162
1163
1164
1165
1166
1167
#else
		q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif

		h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+i)*2]);
		h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+i)*2)+1]);

		tmp1 = _mm256_mul_pd(h2_imag, y1);
1168
1169
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1170
1171
1172
1173
#else
		q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
		tmp2 = _mm256_mul_pd(h2_imag, y2);
1174
1175
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
#else
		q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h2_real, y2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif

		_mm256_store_pd(&q_dbl[(2*i*ldq)+0], q1);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+4], q2);
	}
	h1_real = _mm256_broadcast_sd(&hh_dbl[(nb-1)*2]);
	h1_imag = _mm256_broadcast_sd(&hh_dbl[((nb-1)*2)+1]);

	q1 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+0]);
	q2 = _mm256_load_pd(&q_dbl[(2*nb*ldq)+4]);

	tmp1 = _mm256_mul_pd(h1_imag, x1);
1190
1191
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1192
1193
1194
1195
#else
	q1 = _mm256_add_pd(q1, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x1), _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
#endif
	tmp2 = _mm256_mul_pd(h1_imag, x2);
1196
1197
#ifdef __ELPA_USE_FMA__
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1198
1199
1200
1201
1202
1203
1204
1205
#else
	q2 = _mm256_add_pd(q2, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x2), _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
#endif

	_mm256_store_pd(&q_dbl[(2*nb*ldq)+0], q1);
	_mm256_store_pd(&q_dbl[(2*nb*ldq)+4], q2);
}

1206
static __forceinline void hh_trafo_complex_kernel_2_AVX_2hv_double(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s)
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;
	double* s_dbl = (double*)(&s);

	__m256d x1;
	__m256d y1;
	__m256d q1;
	__m256d h1_real, h1_imag, h2_real, h2_imag;
	__m256d tmp1;
	int i=0;

	__m256d sign = (__m256d)_mm256_set_epi64x(0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000);

	x1 = _mm256_load_pd(&q_dbl[(2*ldq)+0]);

	h2_real = _mm256_broadcast_sd(&hh_dbl[(ldh+1)*2]);
	h2_imag = _mm256_broadcast_sd(&hh_dbl[((ldh+1)*2)+1]);
1225
#ifndef __ELPA_USE_FMA__
1226
1227
1228
1229
1230
1231
1232
	// conjugate
	h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

	y1 = _mm256_load_pd(&q_dbl[0]);

	tmp1 = _mm256_mul_pd(h2_imag, x1);
1233
1234
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));