elpa2_kernels_complex_sse_2hv.cpp 97.4 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
13
14
15
//      Schwerpunkt Wissenschaftliches Rechnen ,
//    - Fritz-Haber-Institut, Berlin, Abt. Theorie,
//    - Max-Plack-Institut für Mathematik in den Naturwissenschaftrn,
//      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
63
64
65
66
67
68
69
70
71
// --------------------------------------------------------------------------------------------------

#include <complex>
#include <x86intrin.h>

#define __forceinline __attribute__((always_inline))

#ifdef __USE_AVX128__
#undef __AVX__
#endif

72
73
74
75
76
77
78
79
80
81
82
#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
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196

extern "C" {

//Forward declaration
#ifdef __AVX__
static __forceinline void hh_trafo_complex_kernel_8_AVX_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s);
static __forceinline void hh_trafo_complex_kernel_6_AVX_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s);
static __forceinline void hh_trafo_complex_kernel_4_AVX_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s);
static __forceinline void hh_trafo_complex_kernel_2_AVX_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s);
#else
static __forceinline void hh_trafo_complex_kernel_4_SSE_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s);
static __forceinline void hh_trafo_complex_kernel_3_SSE_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s);
static __forceinline void hh_trafo_complex_kernel_2_SSE_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s);
static __forceinline void hh_trafo_complex_kernel_1_SSE_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s);
#endif

#if 0
static __forceinline void hh_trafo_complex_kernel_4_C_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s)
{
	std::complex<double> x1;
	std::complex<double> x2;
	std::complex<double> x3;
	std::complex<double> x4;
	std::complex<double> y1;
	std::complex<double> y2;
	std::complex<double> y3;
	std::complex<double> y4;
	std::complex<double> h1;
	std::complex<double> h2;
	std::complex<double> tau1;
	std::complex<double> tau2;
	int i=0;

	x1 = q[ldq+0];
	x2 = q[ldq+1];
	x3 = q[ldq+2];
	x4 = q[ldq+3];

	h2 = conj(hh[ldh+1]);

	y1 = q[0] + (x1*h2);
	y2 = q[1] + (x2*h2);
	y3 = q[2] + (x3*h2);
	y4 = q[3] + (x4*h2);

	for (i = 2; i < nb; i++)
	{
		h1 = conj(hh[i-1]);
		h2 = conj(hh[ldh+i]);

		x1 += (q[(i*ldq)+0] * h1);
		y1 += (q[(i*ldq)+0] * h2);
		x2 += (q[(i*ldq)+1] * h1);
		y2 += (q[(i*ldq)+1] * h2);
		x3 += (q[(i*ldq)+2] * h1);
		y3 += (q[(i*ldq)+2] * h2);
		x4 += (q[(i*ldq)+3] * h1);
		y4 += (q[(i*ldq)+3] * h2);
	}
	h1 = conj(hh[nb-1]);

	x1 += (q[(nb*ldq)+0] * h1);
	x2 += (q[(nb*ldq)+1] * h1);
	x3 += (q[(nb*ldq)+2] * h1);
	x4 += (q[(nb*ldq)+3] * h1);

	tau1 = hh[0];
	tau2 = hh[ldh];

	h1 = (-1.0)*tau1;

	x1 *= h1;
	x2 *= h1;
	x3 *= h1;
	x4 *= h1;

	h1 = (-1.0)*tau2;
	h2 = (-1.0)*tau2;
	h2 *= s;
	y1 = y1*h1 +x1*h2;
	y2 = y2*h1 +x2*h2;
	y3 = y3*h1 +x3*h2;
	y4 = y4*h1 +x4*h2;

	q[0] += y1;
	q[1] += y2;
	q[2] += y3;
	q[3] += y4;

	h2 = hh[ldh+1];
	q[ldq+0] += (x1 + (y1*h2));
	q[ldq+1] += (x2 + (y2*h2));
	q[ldq+2] += (x3 + (y3*h2));
	q[ldq+3] += (x4 + (y4*h2));

	for (i = 2; i < nb; i++)
	{
		h1 = hh[i-1];
		h2 = hh[ldh+i];

		q[(i*ldq)+0] += ((x1*h1) + (y1*h2));
		q[(i*ldq)+1] += ((x2*h1) + (y2*h2));
		q[(i*ldq)+2] += ((x3*h1) + (y3*h2));
		q[(i*ldq)+3] += ((x4*h1) + (y4*h2));
	}

	h1 = hh[nb-1];
	q[(nb*ldq)+0] += (x1*h1);
	q[(nb*ldq)+1] += (x2*h1);
	q[(nb*ldq)+2] += (x3*h1);
	q[(nb*ldq)+3] += (x4*h1);
}
#endif

Andreas Marek's avatar
Andreas Marek committed
197
void double_hh_trafo_complex_sse_avx_2hv_(std::complex<double>* q, std::complex<double>* hh, int* pnb, int* pnq, int* pldq, int* pldh)
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
{
	int i;
	int nb = *pnb;
	int nq = *pldq;
	int ldq = *pldq;
	int ldh = *pldh;

	std::complex<double> s = conj(hh[(ldh)+1])*1.0;
	for (i = 2; i < nb; i++)
	{
		s += hh[i-1] * conj(hh[(i+ldh)]);
	}

#ifdef __AVX__
#if 1
	for (i = 0; i < nq-4; i+=8)
	{
		hh_trafo_complex_kernel_8_AVX_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
	if (nq-i > 0)
	{
		hh_trafo_complex_kernel_4_AVX_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
#else
	for (i = 0; i < nq-4; i+=6)
	{
		hh_trafo_complex_kernel_6_AVX_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
	if (nq-i > 2)
	{
		hh_trafo_complex_kernel_4_AVX_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
	else if (nq-i > 0)
	{
		hh_trafo_complex_kernel_2_AVX_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
#endif
#else
#if 1
	for (i = 0; i < nq; i+=4)
	{
		hh_trafo_complex_kernel_4_SSE_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
#else
	for (i = 0; i < nq-2; i+=3)
	{
		hh_trafo_complex_kernel_3_SSE_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
	if (nq-i > 1)
	{
		hh_trafo_complex_kernel_2_SSE_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
	else if (nq-i > 0)
	{
		hh_trafo_complex_kernel_1_SSE_2hv(&q[i], hh, nb, ldq, ldh, s);
	}
#endif
#endif
}

#ifdef __AVX__
static __forceinline void hh_trafo_complex_kernel_8_AVX_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s)
{
	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]);
281
#ifndef __ELPA_USE_FMA__
282
283
284
285
286
287
288
289
290
291
	// 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);
292
293
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
294
295
296
297
#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);
298
299
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
300
301
302
303
#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);
304
305
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
306
307
308
309
#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);
310
311
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_add_pd(y4, _mm256_FMSUBADD_pd(h2_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
312
313
314
315
316
317
318
319
320
321
322
323
324
#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]);
325
#ifndef __ELPA_USE_FMA__
326
327
328
329
330
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
331
332
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
333
334
335
336
#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);
337
338
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
339
340
341
342
#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);
343
344
#ifdef __ELPA_USE_FMA__
		x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
345
346
347
348
#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);
349
350
#ifdef __ELPA_USE_FMA__
		x4 = _mm256_add_pd(x4, _mm256_FMSUBADD_pd(h1_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
351
352
353
354
355
356
#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]);
357
#ifndef __ELPA_USE_FMA__
358
359
360
361
362
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
363
364
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
365
366
367
368
#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);
369
370
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
371
372
373
374
#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);
375
376
#ifdef __ELPA_USE_FMA__
		y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
377
378
379
380
#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);
381
382
#ifdef __ELPA_USE_FMA__
		y4 = _mm256_add_pd(y4, _mm256_FMSUBADD_pd(h2_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
383
384
385
386
387
388
389
#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]);
390
#ifndef __ELPA_USE_FMA__
391
392
393
394
395
396
397
398
399
400
	// 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);
401
402
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
403
404
405
406
#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);
407
408
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
409
410
411
412
#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);
413
414
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
415
416
417
418
#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);
419
420
#ifdef __ELPA_USE_FMA__
	x4 = _mm256_add_pd(x4, _mm256_FMSUBADD_pd(h1_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
421
422
423
424
425
426
427
428
429
430
#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);
431
432
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
433
434
435
436
#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);
437
438
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
439
440
441
442
#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);
443
444
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
445
446
447
448
#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);
449
450
#ifdef __ELPA_USE_FMA__
	x4 = _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5));
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
#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);
468
469
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
470
471
472
473
474
475
476
477
#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);
478
479
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
480
481
482
483
#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);
484
485
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
486
487
488
489
#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);
490
491
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_FMADDSUB_pd(h1_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
492
493
494
495
#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);
496
497
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_FMADDSUB_pd(h1_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5));
498
499
500
501
502
#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);
503
504
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMADDSUB_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
505
506
507
508
#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);
509
510
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMADDSUB_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
511
512
513
514
#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);
515
516
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMADDSUB_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
517
518
519
520
#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);
521
522
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_add_pd(y4, _mm256_FMADDSUB_pd(h2_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
#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);
556
557
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
558
559
560
561
562
#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_
563
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
564
565
566
567
#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);
568
569
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
570
571
572
573
#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);
574
575
#ifdef __ELPA_USE_FMA__
	q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h2_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
#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);
596
597
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
598
599
600
601
#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);
602
603
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
604
605
606
607
#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);
608
609
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
610
611
612
613
#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);
614
615
#ifdef __ELPA_USE_FMA__
		q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
616
617
618
619
620
621
622
623
#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);
624
625
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
626
627
628
629
#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);
630
631
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
632
633
634
635
#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);
636
637
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
638
639
640
641
#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);
642
643
#ifdef __ELPA_USE_FMA__
		q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h2_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
#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);
662
663
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
664
665
666
667
#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);
668
669
#ifdef __ELPA_USE_FMA__
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
670
671
672
673
#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);
674
675
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
676
677
678
679
#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);
680
681
#ifdef __ELPA_USE_FMA__
	q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
#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);
}

static __forceinline void hh_trafo_complex_kernel_6_AVX_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s)
{
	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]);
713
#ifndef __ELPA_USE_FMA__
714
715
716
717
718
719
720
721
722
	// 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);
723
724
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
725
726
727
728
#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);
729
730
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
731
732
733
734
#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);
735
736
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
737
738
739
740
741
742
743
744
745
746
747
748
#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]);
749
#ifndef __ELPA_USE_FMA__
750
751
752
753
754
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
755
756
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
757
758
759
760
#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);
761
762
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
763
764
765
766
#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);
767
768
#ifdef __ELPA_USE_FMA__
		x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
769
770
771
772
773
774
#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]);
775
#ifndef __ELPA_USE_FMA__
776
777
778
779
780
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
781
782
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
783
784
785
786
#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);
787
788
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
789
790
791
792
#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);
793
794
#ifdef __ELPA_USE_FMA__
		y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
795
796
797
798
799
800
801
#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]);
802
#ifndef __ELPA_USE_FMA__
803
804
805
806
807
808
809
810
811
	// 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);
812
813
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
814
815
816
817
#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);
818
819
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
820
821
822
823
#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);
824
825
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
826
827
828
829
830
831
832
833
834
835
#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);
836
837
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
838
839
840
841
#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);
842
843
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
844
845
846
847
#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);
848
849
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
#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);
867
868
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
869
870
871
872
873
874
875
876
#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);
877
878
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
879
880
881
882
#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);
883
884
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
885
886
887
888
#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);
889
890
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_FMADDSUB_pd(h1_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
891
892
893
894
895
#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);
896
897
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMADDSUB_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
898
899
900
901
#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);
902
903
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMADDSUB_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
904
905
906
907
#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);
908
909
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMADDSUB_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
#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);
938
939
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
940
941
942
943
944
#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_
945
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
946
947
948
949
#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);
950
951
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
#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);
970
971
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
972
973
974
975
#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);
976
977
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
978
979
980
981
#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);
982
983
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
984
985
986
987
988
989
990
991
#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);
992
993
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
994
995
996
997
#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);
998
999
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1000
1001
1002
1003
#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);
1004
1005
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
#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);
1022
1023
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1024
1025
1026
1027
#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);
1028
1029
#ifdef __ELPA_USE_FMA__
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1030
1031
1032
1033
#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);
1034
1035
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
#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);
}

static __forceinline void hh_trafo_complex_kernel_4_AVX_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s)
{
	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]);
1065
#ifndef __ELPA_USE_FMA__
1066
1067
1068
1069
1070
1071
1072
1073
	// 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);
1074
1075
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1076
1077
1078
1079
#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);
1080
1081
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
#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]);
1093
#ifndef __ELPA_USE_FMA__
1094
1095
1096
1097
1098
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
1099
1100
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1101
1102
1103
1104
#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);
1105
1106
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1107
1108
1109
1110
1111
1112
#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]);
1113
#ifndef __ELPA_USE_FMA__
1114
1115
1116
1117
1118
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
1119
1120
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1121
1122
1123
1124
#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);
1125
1126
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1127
1128
1129
1130
1131
1132
1133
#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]);
1134
#ifndef __ELPA_USE_FMA__
1135
1136
1137
1138
1139
1140
1141
1142
	// 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);
1143
1144
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1145
1146
1147
1148
#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);
1149
1150
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
#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);
1161
1162
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1163
1164
1165
1166
#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);
1167
1168
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
#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);
1186
1187
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1188
1189
1190
1191
1192
1193
1194
1195
#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);
1196
1197
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1198
1199
1200
1201
#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);
1202
1203
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
1204
1205
1206
1207
1208
#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);
1209
1210
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMADDSUB_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1211
1212
1213
1214
#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);
1215
1216
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMADDSUB_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
#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);
1240
1241
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1242
1243
1244
1245
1246
#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_
1247
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
#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);
1264
1265
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1266
1267
1268
1269
#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);
1270
1271
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1272
1273
1274
1275
1276
1277
1278
1279
#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);
1280
1281
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1282
1283
1284
1285
#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);
1286
1287
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
#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);
1302
1303
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1304
1305
1306
1307
#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);
1308
1309
#ifdef __ELPA_USE_FMA__
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
#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);
}

static __forceinline void hh_trafo_complex_kernel_2_AVX_2hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq, int ldh, std::complex<double> s)
{
	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]);
1337
#ifndef __ELPA_USE_FMA__
1338
1339
1340
1341
1342
1343
1344
	// conjugate
	h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

	y1 = _mm256_load_pd(&q_dbl[0]);

	tmp1 = _mm256_mul_pd(h2_imag, x1);
1345
1346
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1347
1348
1349
1350
1351
1352
1353