elpa2_kernels_complex_avx-avx2_1hv.cpp 35.9 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

extern "C" {

//Forward declaration
#ifdef __AVX__
static  __forceinline void hh_trafo_complex_kernel_12_AVX_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq);
static  __forceinline void hh_trafo_complex_kernel_8_AVX_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq);
static  __forceinline void hh_trafo_complex_kernel_4_AVX_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq);

#else
static __forceinline void hh_trafo_complex_kernel_6_SSE_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq);
static __forceinline void hh_trafo_complex_kernel_4_SSE_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq);
static __forceinline void hh_trafo_complex_kernel_2_SSE_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq);
#endif

#if 0
static __forceinline void hh_trafo_complex_kernel_4_C_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq)
{
	std::complex<double> x0;
	std::complex<double> x1;
	std::complex<double> x2;
	std::complex<double> x3;
	std::complex<double> h0;
	std::complex<double> tau0;
	int i=0;

	x0 = q[0];
	x1 = q[1];
	x2 = q[2];
	x3 = q[3];

	for (i = 1; i < nb; i++)
	{
		h0 = conj(hh[i]);
		x0 += (q[(i*ldq)+0] * h0);
		x1 += (q[(i*ldq)+1] * h0);
		x2 += (q[(i*ldq)+2] * h0);
		x3 += (q[(i*ldq)+3] * h0);
	}

	tau0 = hh[0];

	h0 = (-1.0)*tau0;

	x0 *= h0;
	x1 *= h0;
	x2 *= h0;
	x3 *= h0;

	q[0] += x0;
	q[1] += x1;
	q[2] += x2;
	q[3] += x3;

	for (i = 1; i < nb; i++)
	{
		h0 = hh[i];
		q[(i*ldq)+0] += (x0*h0);
		q[(i*ldq)+1] += (x1*h0);
		q[(i*ldq)+2] += (x2*h0);
		q[(i*ldq)+3] += (x3*h0);
	}
}
#endif // if 0

Andreas Marek's avatar
Andreas Marek committed
148
void single_hh_trafo_complex_sse_avx_1hv_(std::complex<double>* q, std::complex<double>* hh, int* pnb, int* pnq, int* pldq)
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
197
198
199
200
201
202
203
204
205
206
207
208
209
{
	int i;
	int nb = *pnb;
	int nq = *pldq;
	int ldq = *pldq;
	//int ldh = *pldh;

#ifdef __AVX__
	for (i = 0; i < nq-8; i+=12)
	{
		hh_trafo_complex_kernel_12_AVX_1hv(&q[i], hh, nb, ldq);
	}
	if (nq-i > 4)
	{
		hh_trafo_complex_kernel_8_AVX_1hv(&q[i], hh, nb, ldq);
	}
	else if (nq-i > 0)
	{
		hh_trafo_complex_kernel_4_AVX_1hv(&q[i], hh, nb, ldq);
	}
#else
	for (i = 0; i < nq-4; i+=6)
	{
		hh_trafo_complex_kernel_6_SSE_1hv(&q[i], hh, nb, ldq);
	}
	if (nq-i > 2)
	{
		hh_trafo_complex_kernel_4_SSE_1hv(&q[i], hh, nb, ldq);
	}
	else if (nq-i > 0)
	{
		hh_trafo_complex_kernel_2_SSE_1hv(&q[i], hh, nb, ldq);
	}
#endif
}

#ifdef __AVX__
 static __forceinline void hh_trafo_complex_kernel_12_AVX_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq)
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;

	__m256d x1, x2, x3, x4, x5, x6;
	__m256d q1, q2, q3, q4, q5, q6;
	__m256d h1_real, h1_imag;
	__m256d tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
	int i=0;

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

	x1 = _mm256_load_pd(&q_dbl[0]);
	x2 = _mm256_load_pd(&q_dbl[4]);
	x3 = _mm256_load_pd(&q_dbl[8]);
	x4 = _mm256_load_pd(&q_dbl[12]);
	x5 = _mm256_load_pd(&q_dbl[16]);
	x6 = _mm256_load_pd(&q_dbl[20]);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm256_broadcast_sd(&hh_dbl[i*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[(i*2)+1]);
210
#ifndef __ELPA_USE_FMA__		
211
212
213
214
215
216
217
218
219
220
221
222
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		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]);
		q5 = _mm256_load_pd(&q_dbl[(2*i*ldq)+16]);
		q6 = _mm256_load_pd(&q_dbl[(2*i*ldq)+20]);

		tmp1 = _mm256_mul_pd(h1_imag, q1);
223
224
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
225
226
227
228
#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);
229
230
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
231
232
233
234
#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);
235
236
#ifdef __ELPA_USE_FMA__
		x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
237
238
239
240
#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);
241
242
#ifdef __ELPA_USE_FMA__
		x4 = _mm256_add_pd(x4, _mm256_FMSUBADD_pd(h1_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
243
244
245
246
#else
		x4 = _mm256_add_pd(x4, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif
		tmp5 = _mm256_mul_pd(h1_imag, q5);
247
248
#ifdef __ELPA_USE_FMA__
		x5 = _mm256_add_pd(x5, _mm256_FMSUBADD_pd(h1_real, q5, _mm256_shuffle_pd(tmp5, tmp5, 0x5)));
249
250
251
252
#else
		x5 = _mm256_add_pd(x5, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q5), _mm256_shuffle_pd(tmp5, tmp5, 0x5)));
#endif
		tmp6 = _mm256_mul_pd(h1_imag, q6);
253
254
#ifdef __ELPA_USE_FMA__
		x6 = _mm256_add_pd(x6, _mm256_FMSUBADD_pd(h1_real, q6, _mm256_shuffle_pd(tmp6, tmp6, 0x5)));
255
256
257
258
259
260
261
262
263
264
265
#else
		x6 = _mm256_add_pd(x6, _mm256_addsub_pd( _mm256_mul_pd(h1_real, q6), _mm256_shuffle_pd(tmp6, tmp6, 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);
266
267
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
268
269
270
271
#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);
272
273
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
274
275
276
277
#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);
278
279
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
280
281
282
283
#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);
284
285
#ifdef __ELPA_USE_FMA__
	x4 = _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5));
286
287
288
289
#else
	x4 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x4), _mm256_shuffle_pd(tmp4, tmp4, 0x5));
#endif
	tmp5 = _mm256_mul_pd(h1_imag, x5);
290
291
#ifdef __ELPA_USE_FMA__
	x5 = _mm256_FMADDSUB_pd(h1_real, x5, _mm256_shuffle_pd(tmp5, tmp5, 0x5));
292
293
294
295
#else
	x5 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x5), _mm256_shuffle_pd(tmp5, tmp5, 0x5));
#endif
	tmp6 = _mm256_mul_pd(h1_imag, x6);
296
297
#ifdef __ELPA_USE_FMA__
	x6 = _mm256_FMADDSUB_pd(h1_real, x6, _mm256_shuffle_pd(tmp6, tmp6, 0x5));
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
#else
	x6 = _mm256_addsub_pd( _mm256_mul_pd(h1_real, x6), _mm256_shuffle_pd(tmp6, tmp6, 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]);
	q5 = _mm256_load_pd(&q_dbl[16]);
	q6 = _mm256_load_pd(&q_dbl[20]);

	q1 = _mm256_add_pd(q1, x1);
	q2 = _mm256_add_pd(q2, x2);
	q3 = _mm256_add_pd(q3, x3);
	q4 = _mm256_add_pd(q4, x4);
	q5 = _mm256_add_pd(q5, x5);
	q6 = _mm256_add_pd(q6, x6);

	_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);
	_mm256_store_pd(&q_dbl[16], q5);
	_mm256_store_pd(&q_dbl[20], q6);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm256_broadcast_sd(&hh_dbl[i*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[(i*2)+1]);

		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]);
		q5 = _mm256_load_pd(&q_dbl[(2*i*ldq)+16]);
		q6 = _mm256_load_pd(&q_dbl[(2*i*ldq)+20]);

		tmp1 = _mm256_mul_pd(h1_imag, x1);
336
337
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
338
339
340
341
#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);
342
343
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
344
345
346
347
#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);
348
349
#ifdef __ELPA_USE_FMA__	
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
350
351
352
353
#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);
354
355
#ifdef __ELPA_USE_FMA__
		q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
356
357
358
359
#else
		q4 = _mm256_add_pd(q4, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x4), _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
#endif
		tmp5 = _mm256_mul_pd(h1_imag, x5);
360
361
#ifdef __ELPA_USE_FMA__
		q5 = _mm256_add_pd(q5, _mm256_FMADDSUB_pd(h1_real, x5, _mm256_shuffle_pd(tmp5, tmp5, 0x5)));
362
363
364
365
#else
		q5 = _mm256_add_pd(q5, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x5), _mm256_shuffle_pd(tmp5, tmp5, 0x5)));
#endif
		tmp6 = _mm256_mul_pd(h1_imag, x6);
366
367
#ifdef __ELPA_USE_FMA__
		q6 = _mm256_add_pd(q6, _mm256_FMADDSUB_pd(h1_real, x6, _mm256_shuffle_pd(tmp6, tmp6, 0x5)));
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
#else
		q6 = _mm256_add_pd(q6, _mm256_addsub_pd( _mm256_mul_pd(h1_real, x6), _mm256_shuffle_pd(tmp6, tmp6, 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);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+16], q5);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+20], q6);
	}
}

static __forceinline void hh_trafo_complex_kernel_8_AVX_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq)
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;

	__m256d x1, x2, x3, x4;
	__m256d q1, q2, q3, q4;
	__m256d h1_real, h1_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[0]);
	x2 = _mm256_load_pd(&q_dbl[4]);
	x3 = _mm256_load_pd(&q_dbl[8]);
	x4 = _mm256_load_pd(&q_dbl[12]);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm256_broadcast_sd(&hh_dbl[i*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[(i*2)+1]);
403
#ifndef __ELPA_USE_FMA__		
404
405
406
407
408
409
410
411
412
413
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		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]);

		tmp1 = _mm256_mul_pd(h1_imag, q1);
414
415
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
416
417
418
419
#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);
420
421
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
422
423
424
425
#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);
426
427
#ifdef __ELPA_USE_FMA__
		x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
428
429
430
431
#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);
432
433
#ifdef __ELPA_USE_FMA__
		x4 = _mm256_add_pd(x4, _mm256_FMSUBADD_pd(h1_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
434
435
436
437
438
439
440
441
442
443
444
#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);
445
446
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
447
448
449
450
#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);
451
452
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
453
454
455
456
#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);
457
458
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
459
460
461
462
#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);
463
464
#ifdef __ELPA_USE_FMA__
	x4 = _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5));
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
#else
	x4 = _mm256_addsub_pd( _mm256_mul_pd(h1_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, x1);
	q2 = _mm256_add_pd(q2, x2);
	q3 = _mm256_add_pd(q3, x3);
	q4 = _mm256_add_pd(q4, x4);

	_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);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm256_broadcast_sd(&hh_dbl[i*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[(i*2)+1]);

		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]);

		tmp1 = _mm256_mul_pd(h1_imag, x1);
495
496
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
497
498
499
500
#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);
501
502
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
503
504
505
506
#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);
507
508
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
509
510
511
512
#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);
513
514
#ifdef __ELPA_USE_FMA__
		q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
#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*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);
	}
}

static __forceinline void hh_trafo_complex_kernel_4_AVX_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq)
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;

	__m256d x1, x2;
	__m256d q1, q2;
	__m256d h1_real, h1_imag;
	__m256d tmp1, tmp2;
	int i=0;

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

	x1 = _mm256_load_pd(&q_dbl[0]);
	x2 = _mm256_load_pd(&q_dbl[4]);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm256_broadcast_sd(&hh_dbl[i*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[(i*2)+1]);
546
#ifndef __ELPA_USE_FMA__
547
548
549
550
551
552
553
554
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

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

		tmp1 = _mm256_mul_pd(h1_imag, q1);
555
556
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
557
558
559
560
#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);
561
562
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
563
564
565
566
567
568
569
570
571
572
573
#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);
574
575
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
576
577
578
579
#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);
580
581
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
#else
	x2 = _mm256_addsub_pd( _mm256_mul_pd(h1_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, x1);
	q2 = _mm256_add_pd(q2, x2);

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

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm256_broadcast_sd(&hh_dbl[i*2]);
		h1_imag = _mm256_broadcast_sd(&hh_dbl[(i*2)+1]);

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

		tmp1 = _mm256_mul_pd(h1_imag, x1);
604
605
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
606
607
608
609
#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);
610
611
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
#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*i*ldq)+0], q1);
		_mm256_store_pd(&q_dbl[(2*i*ldq)+4], q2);
	}
}

#else
static __forceinline void hh_trafo_complex_kernel_6_SSE_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq)
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;

	__m128d x1, x2, x3, x4, x5, x6;
	__m128d q1, q2, q3, q4, q5, q6;
	__m128d h1_real, h1_imag;
	__m128d tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
	int i=0;

	__m128d sign = (__m128d)_mm_set_epi64x(0x8000000000000000, 0x8000000000000000);

	x1 = _mm_load_pd(&q_dbl[0]);
	x2 = _mm_load_pd(&q_dbl[2]);
	x3 = _mm_load_pd(&q_dbl[4]);
	x4 = _mm_load_pd(&q_dbl[6]);
	x5 = _mm_load_pd(&q_dbl[8]);
	x6 = _mm_load_pd(&q_dbl[10]);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
		h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
646
#ifndef __ELPA_USE_FMA__		
647
648
649
650
651
652
653
654
655
656
657
658
		// conjugate
		h1_imag = _mm_xor_pd(h1_imag, sign);
#endif

		q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
		q3 = _mm_load_pd(&q_dbl[(2*i*ldq)+4]);
		q4 = _mm_load_pd(&q_dbl[(2*i*ldq)+6]);
		q5 = _mm_load_pd(&q_dbl[(2*i*ldq)+8]);
		q6 = _mm_load_pd(&q_dbl[(2*i*ldq)+10]);

		tmp1 = _mm_mul_pd(h1_imag, q1);
659
#ifdef __ELPA_USE_FMA__
660
661
662
663
664
		x1 = _mm_add_pd(x1, _mm_msubadd_pd(h1_real, q1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
		x1 = _mm_add_pd(x1, _mm_addsub_pd( _mm_mul_pd(h1_real, q1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
		tmp2 = _mm_mul_pd(h1_imag, q2);
665
#ifdef __ELPA_USE_FMA__
666
667
668
669
670
		x2 = _mm_add_pd(x2, _mm_msubadd_pd(h1_real, q2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
		x2 = _mm_add_pd(x2, _mm_addsub_pd( _mm_mul_pd(h1_real, q2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
		tmp3 = _mm_mul_pd(h1_imag, q3);
671
#ifdef __ELPA_USE_FMA__
672
673
674
675
676
		x3 = _mm_add_pd(x3, _mm_msubadd_pd(h1_real, q3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#else
		x3 = _mm_add_pd(x3, _mm_addsub_pd( _mm_mul_pd(h1_real, q3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#endif
		tmp4 = _mm_mul_pd(h1_imag, q4);
677
#ifdef __ELPA_USE_FMA__
678
679
680
681
682
		x4 = _mm_add_pd(x4, _mm_msubadd_pd(h1_real, q4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#else
		x4 = _mm_add_pd(x4, _mm_addsub_pd( _mm_mul_pd(h1_real, q4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#endif
		tmp5 = _mm_mul_pd(h1_imag, q5);
683
#ifdef __ELPA_USE_FMA__
684
685
686
687
688
		x5 = _mm_add_pd(x5, _mm_msubadd_pd(h1_real, q5, _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1))));
#else
		x5 = _mm_add_pd(x5, _mm_addsub_pd( _mm_mul_pd(h1_real, q5), _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1))));
#endif
		tmp6 = _mm_mul_pd(h1_imag, q6);
689
#ifdef __ELPA_USE_FMA__
690
691
692
693
694
695
696
697
698
699
700
701
		x6 = _mm_add_pd(x6, _mm_msubadd_pd(h1_real, q6, _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1))));
#else
		x6 = _mm_add_pd(x6, _mm_addsub_pd( _mm_mul_pd(h1_real, q6), _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1))));
#endif
	}

	h1_real = _mm_loaddup_pd(&hh_dbl[0]);
	h1_imag = _mm_loaddup_pd(&hh_dbl[1]);
	h1_real = _mm_xor_pd(h1_real, sign);
	h1_imag = _mm_xor_pd(h1_imag, sign);

	tmp1 = _mm_mul_pd(h1_imag, x1);
702
#ifdef __ELPA_USE_FMA__	
703
704
705
706
707
	x1 = _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#else
	x1 = _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#endif
	tmp2 = _mm_mul_pd(h1_imag, x2);
708
#ifdef __ELPA_USE_FMA__
709
710
711
712
713
	x2 = _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#else
	x2 = _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#endif
	tmp3 = _mm_mul_pd(h1_imag, x3);
714
#ifdef __ELPA_USE_FMA__
715
716
717
718
719
	x3 = _mm_maddsub_pd(h1_real, x3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1)));
#else
	x3 = _mm_addsub_pd( _mm_mul_pd(h1_real, x3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1)));
#endif
	tmp4 = _mm_mul_pd(h1_imag, x4);
720
#ifdef __ELPA_USE_FMA__
721
722
723
724
725
	x4 = _mm_maddsub_pd(h1_real, x4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1)));
#else
	x4 = _mm_addsub_pd( _mm_mul_pd(h1_real, x4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1)));
#endif
	tmp5 = _mm_mul_pd(h1_imag, x5);
726
#ifdef __ELPA_USE_FMA__
727
728
729
730
731
	x5 = _mm_maddsub_pd(h1_real, x5, _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1)));
#else
	x5 = _mm_addsub_pd( _mm_mul_pd(h1_real, x5), _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1)));
#endif
	tmp6 = _mm_mul_pd(h1_imag, x6);
732
#ifdef __ELPA_USE_FMA__
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
	x6 = _mm_maddsub_pd(h1_real, x6, _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1)));
#else
	x6 = _mm_addsub_pd( _mm_mul_pd(h1_real, x6), _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1)));
#endif

	q1 = _mm_load_pd(&q_dbl[0]);
	q2 = _mm_load_pd(&q_dbl[2]);
	q3 = _mm_load_pd(&q_dbl[4]);
	q4 = _mm_load_pd(&q_dbl[6]);
	q5 = _mm_load_pd(&q_dbl[8]);
	q6 = _mm_load_pd(&q_dbl[10]);

	q1 = _mm_add_pd(q1, x1);
	q2 = _mm_add_pd(q2, x2);
	q3 = _mm_add_pd(q3, x3);
	q4 = _mm_add_pd(q4, x4);
	q5 = _mm_add_pd(q5, x5);
	q6 = _mm_add_pd(q6, x6);

	_mm_store_pd(&q_dbl[0], q1);
	_mm_store_pd(&q_dbl[2], q2);
	_mm_store_pd(&q_dbl[4], q3);
	_mm_store_pd(&q_dbl[6], q4);
	_mm_store_pd(&q_dbl[8], q5);
	_mm_store_pd(&q_dbl[10], q6);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
		h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);

		q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
		q3 = _mm_load_pd(&q_dbl[(2*i*ldq)+4]);
		q4 = _mm_load_pd(&q_dbl[(2*i*ldq)+6]);
		q5 = _mm_load_pd(&q_dbl[(2*i*ldq)+8]);
		q6 = _mm_load_pd(&q_dbl[(2*i*ldq)+10]);

		tmp1 = _mm_mul_pd(h1_imag, x1);
772
#ifdef __ELPA_USE_FMA__
773
774
775
776
777
		q1 = _mm_add_pd(q1, _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
		q1 = _mm_add_pd(q1, _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
		tmp2 = _mm_mul_pd(h1_imag, x2);
778
#ifdef __ELPA_USE_FMA__
779
780
781
782
783
		q2 = _mm_add_pd(q2, _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
		q2 = _mm_add_pd(q2, _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
		tmp3 = _mm_mul_pd(h1_imag, x3);
784
#ifdef __ELPA_USE_FMA__
785
786
787
788
789
		q3 = _mm_add_pd(q3, _mm_maddsub_pd(h1_real, x3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#else
		q3 = _mm_add_pd(q3, _mm_addsub_pd( _mm_mul_pd(h1_real, x3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#endif
		tmp4 = _mm_mul_pd(h1_imag, x4);
790
#ifdef __ELPA_USE_FMA__
791
792
793
794
795
		q4 = _mm_add_pd(q4, _mm_maddsub_pd(h1_real, x4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#else
		q4 = _mm_add_pd(q4, _mm_addsub_pd( _mm_mul_pd(h1_real, x4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#endif
		tmp5 = _mm_mul_pd(h1_imag, x5);
796
#ifdef __ELPA_USE_FMA__
797
798
799
800
801
		q5 = _mm_add_pd(q5, _mm_maddsub_pd(h1_real, x5, _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1))));
#else
		q5 = _mm_add_pd(q5, _mm_addsub_pd( _mm_mul_pd(h1_real, x5), _mm_shuffle_pd(tmp5, tmp5, _MM_SHUFFLE2(0,1))));
#endif
		tmp6 = _mm_mul_pd(h1_imag, x6);
802
#ifdef __ELPA_USE_FMA__
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
		q6 = _mm_add_pd(q6, _mm_maddsub_pd(h1_real, x6, _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1))));
#else
		q6 = _mm_add_pd(q6, _mm_addsub_pd( _mm_mul_pd(h1_real, x6), _mm_shuffle_pd(tmp6, tmp6, _MM_SHUFFLE2(0,1))));
#endif

		_mm_store_pd(&q_dbl[(2*i*ldq)+0], q1);
		_mm_store_pd(&q_dbl[(2*i*ldq)+2], q2);
		_mm_store_pd(&q_dbl[(2*i*ldq)+4], q3);
		_mm_store_pd(&q_dbl[(2*i*ldq)+6], q4);
		_mm_store_pd(&q_dbl[(2*i*ldq)+8], q5);
		_mm_store_pd(&q_dbl[(2*i*ldq)+10], q6);
	}
}

static __forceinline void hh_trafo_complex_kernel_4_SSE_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq)
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;

	__m128d x1, x2, x3, x4;
	__m128d q1, q2, q3, q4;
	__m128d h1_real, h1_imag;
	__m128d tmp1, tmp2, tmp3, tmp4;
	int i=0;

	__m128d sign = (__m128d)_mm_set_epi64x(0x8000000000000000, 0x8000000000000000);

	x1 = _mm_load_pd(&q_dbl[0]);
	x2 = _mm_load_pd(&q_dbl[2]);
	x3 = _mm_load_pd(&q_dbl[4]);
	x4 = _mm_load_pd(&q_dbl[6]);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
		h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
839
#ifndef __ELPA_USE_FMA__
840
841
842
843
844
845
846
847
848
849
		// conjugate
		h1_imag = _mm_xor_pd(h1_imag, sign);
#endif

		q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
		q3 = _mm_load_pd(&q_dbl[(2*i*ldq)+4]);
		q4 = _mm_load_pd(&q_dbl[(2*i*ldq)+6]);

		tmp1 = _mm_mul_pd(h1_imag, q1);
850
#ifdef __ELPA_USE_FMA__
851
852
853
854
855
		x1 = _mm_add_pd(x1, _mm_msubadd_pd(h1_real, q1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
		x1 = _mm_add_pd(x1, _mm_addsub_pd( _mm_mul_pd(h1_real, q1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
		tmp2 = _mm_mul_pd(h1_imag, q2);
856
#ifdef __ELPA_USE_FMA__
857
858
859
860
861
		x2 = _mm_add_pd(x2, _mm_msubadd_pd(h1_real, q2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
		x2 = _mm_add_pd(x2, _mm_addsub_pd( _mm_mul_pd(h1_real, q2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
		tmp3 = _mm_mul_pd(h1_imag, q3);
862
#ifdef __ELPA_USE_FMA__
863
864
865
866
867
		x3 = _mm_add_pd(x3, _mm_msubadd_pd(h1_real, q3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#else
		x3 = _mm_add_pd(x3, _mm_addsub_pd( _mm_mul_pd(h1_real, q3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#endif
		tmp4 = _mm_mul_pd(h1_imag, q4);
868
#ifdef __ELPA_USE_FMA__
869
870
871
872
873
874
875
876
877
878
879
880
		x4 = _mm_add_pd(x4, _mm_msubadd_pd(h1_real, q4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#else
		x4 = _mm_add_pd(x4, _mm_addsub_pd( _mm_mul_pd(h1_real, q4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#endif
	}

	h1_real = _mm_loaddup_pd(&hh_dbl[0]);
	h1_imag = _mm_loaddup_pd(&hh_dbl[1]);
	h1_real = _mm_xor_pd(h1_real, sign);
	h1_imag = _mm_xor_pd(h1_imag, sign);

	tmp1 = _mm_mul_pd(h1_imag, x1);
881
#ifdef __ELPA_USE_FMA__
882
883
884
885
886
	x1 = _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#else
	x1 = _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#endif
	tmp2 = _mm_mul_pd(h1_imag, x2);
887
#ifdef __ELPA_USE_FMA__
888
889
890
891
892
	x2 = _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#else
	x2 = _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#endif
	tmp3 = _mm_mul_pd(h1_imag, x3);
893
#ifdef __ELPA_USE_FMA__
894
895
896
897
898
	x3 = _mm_maddsub_pd(h1_real, x3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1)));
#else
	x3 = _mm_addsub_pd( _mm_mul_pd(h1_real, x3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1)));
#endif
	tmp4 = _mm_mul_pd(h1_imag, x4);
899
#ifdef __ELPA_USE_FMA__
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
	x4 = _mm_maddsub_pd(h1_real, x4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1)));
#else
	x4 = _mm_addsub_pd( _mm_mul_pd(h1_real, x4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1)));
#endif

	q1 = _mm_load_pd(&q_dbl[0]);
	q2 = _mm_load_pd(&q_dbl[2]);
	q3 = _mm_load_pd(&q_dbl[4]);
	q4 = _mm_load_pd(&q_dbl[6]);

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

	_mm_store_pd(&q_dbl[0], q1);
	_mm_store_pd(&q_dbl[2], q2);
	_mm_store_pd(&q_dbl[4], q3);
	_mm_store_pd(&q_dbl[6], q4);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
		h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);

		q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);
		q3 = _mm_load_pd(&q_dbl[(2*i*ldq)+4]);
		q4 = _mm_load_pd(&q_dbl[(2*i*ldq)+6]);

		tmp1 = _mm_mul_pd(h1_imag, x1);
931
#ifdef __ELPA_USE_FMA__
932
933
934
935
936
		q1 = _mm_add_pd(q1, _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
		q1 = _mm_add_pd(q1, _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
		tmp2 = _mm_mul_pd(h1_imag, x2);
937
#ifdef __ELPA_USE_FMA__
938
939
940
941
942
		q2 = _mm_add_pd(q2, _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
		q2 = _mm_add_pd(q2, _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
		tmp3 = _mm_mul_pd(h1_imag, x3);
943
#ifdef __ELPA_USE_FMA__
944
945
946
947
948
		q3 = _mm_add_pd(q3, _mm_maddsub_pd(h1_real, x3, _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#else
		q3 = _mm_add_pd(q3, _mm_addsub_pd( _mm_mul_pd(h1_real, x3), _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0,1))));
#endif
		tmp4 = _mm_mul_pd(h1_imag, x4);
949
#ifdef __ELPA_USE_FMA__
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
		q4 = _mm_add_pd(q4, _mm_maddsub_pd(h1_real, x4, _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#else
		q4 = _mm_add_pd(q4, _mm_addsub_pd( _mm_mul_pd(h1_real, x4), _mm_shuffle_pd(tmp4, tmp4, _MM_SHUFFLE2(0,1))));
#endif

		_mm_store_pd(&q_dbl[(2*i*ldq)+0], q1);
		_mm_store_pd(&q_dbl[(2*i*ldq)+2], q2);
		_mm_store_pd(&q_dbl[(2*i*ldq)+4], q3);
		_mm_store_pd(&q_dbl[(2*i*ldq)+6], q4);
	}
}

static __forceinline void hh_trafo_complex_kernel_2_SSE_1hv(std::complex<double>* q, std::complex<double>* hh, int nb, int ldq)
{
	double* q_dbl = (double*)q;
	double* hh_dbl = (double*)hh;

	__m128d x1, x2;
	__m128d q1, q2;
	__m128d h1_real, h1_imag;
	__m128d tmp1, tmp2;
	int i=0;

	__m128d sign = (__m128d)_mm_set_epi64x(0x8000000000000000, 0x8000000000000000);

	x1 = _mm_load_pd(&q_dbl[0]);
	x2 = _mm_load_pd(&q_dbl[2]);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
		h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);
982
#ifndef __ELPA_USE_FMA__	
983
984
985
986
987
988
989
990
		// conjugate
		h1_imag = _mm_xor_pd(h1_imag, sign);
#endif

		q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);

		tmp1 = _mm_mul_pd(h1_imag, q1);
991
#ifdef __ELPA_USE_FMA__
992
993
994
995
996
		x1 = _mm_add_pd(x1, _mm_msubadd_pd(h1_real, q1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
		x1 = _mm_add_pd(x1, _mm_addsub_pd( _mm_mul_pd(h1_real, q1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
		tmp2 = _mm_mul_pd(h1_imag, q2);
997
#ifdef __ELPA_USE_FMA__
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
		x2 = _mm_add_pd(x2, _mm_msubadd_pd(h1_real, q2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
		x2 = _mm_add_pd(x2, _mm_addsub_pd( _mm_mul_pd(h1_real, q2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif
	}

	h1_real = _mm_loaddup_pd(&hh_dbl[0]);
	h1_imag = _mm_loaddup_pd(&hh_dbl[1]);
	h1_real = _mm_xor_pd(h1_real, sign);
	h1_imag = _mm_xor_pd(h1_imag, sign);

	tmp1 = _mm_mul_pd(h1_imag, x1);
1010
#ifdef __ELPA_USE_FMA__
1011
1012
1013
1014
1015
	x1 = _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#else
	x1 = _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1)));
#endif
	tmp2 = _mm_mul_pd(h1_imag, x2);
1016
#ifdef __ELPA_USE_FMA__
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
	x2 = _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#else
	x2 = _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1)));
#endif

	q1 = _mm_load_pd(&q_dbl[0]);
	q2 = _mm_load_pd(&q_dbl[2]);

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

	_mm_store_pd(&q_dbl[0], q1);
	_mm_store_pd(&q_dbl[2], q2);

	for (i = 1; i < nb; i++)
	{
		h1_real = _mm_loaddup_pd(&hh_dbl[i*2]);
		h1_imag = _mm_loaddup_pd(&hh_dbl[(i*2)+1]);

		q1 = _mm_load_pd(&q_dbl[(2*i*ldq)+0]);
		q2 = _mm_load_pd(&q_dbl[(2*i*ldq)+2]);

		tmp1 = _mm_mul_pd(h1_imag, x1);
1040
#ifdef __ELPA_USE_FMA__
1041
1042
1043
1044
1045
		q1 = _mm_add_pd(q1, _mm_maddsub_pd(h1_real, x1, _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#else
		q1 = _mm_add_pd(q1, _mm_addsub_pd( _mm_mul_pd(h1_real, x1), _mm_shuffle_pd(tmp1, tmp1, _MM_SHUFFLE2(0,1))));
#endif
		tmp2 = _mm_mul_pd(h1_imag, x2);
1046
#ifdef __ELPA_USE_FMA__
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
		q2 = _mm_add_pd(q2, _mm_maddsub_pd(h1_real, x2, _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#else
		q2 = _mm_add_pd(q2, _mm_addsub_pd( _mm_mul_pd(h1_real, x2), _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0,1))));
#endif

		_mm_store_pd(&q_dbl[(2*i*ldq)+0], q1);
		_mm_store_pd(&q_dbl[(2*i*ldq)+2], q2);
	}
}
#endif
} // extern C