elpa2_kernels_complex_avx-avx2_2hv.c 50.4 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
#include "config-f90.h"
63

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

#define __forceinline __attribute__((always_inline))

69 70
#ifdef HAVE_AVX2

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

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

83 84
#endif

85
//Forward declaration
Andreas Marek's avatar
Andreas Marek committed
86 87 88 89
static __forceinline void hh_trafo_complex_kernel_8_AVX_2hv(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);
static __forceinline void hh_trafo_complex_kernel_6_AVX_2hv(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);
static __forceinline void hh_trafo_complex_kernel_4_AVX_2hv(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);
static __forceinline void hh_trafo_complex_kernel_2_AVX_2hv(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s);
90

91 92 93 94 95 96 97 98 99 100 101 102 103
/*
!f>#ifdef HAVE_AVX
!f> interface
!f>   subroutine double_hh_trafo_complex_avx_avx2_2hv(q, hh, pnb, pnq, pldq, pldh) bind(C, name="double_hh_trafo_complex_avx_avx2_2hv")
!f>     use, intrinsic :: iso_c_binding
!f>     integer(kind=c_int)     :: pnb, pnq, pldq, pldh
!f>     complex(kind=c_double)     :: q(*)
!f>     complex(kind=c_double)     :: hh(pnb,2)
!f>   end subroutine
!f> end interface
!f>#endif
*/

Andreas Marek's avatar
Andreas Marek committed
104
void double_hh_trafo_complex_avx_avx2_2hv(double complex* q, double complex* hh, int* pnb, int* pnq, int* pldq, int* pldh)
105 106 107 108 109 110 111
{
	int i;
	int nb = *pnb;
	int nq = *pldq;
	int ldq = *pldq;
	int ldh = *pldh;

Andreas Marek's avatar
Andreas Marek committed
112
	double complex s = conj(hh[(ldh)+1])*1.0;
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
	for (i = 2; i < nb; i++)
	{
		s += hh[i-1] * conj(hh[(i+ldh)]);
	}

#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
}

Andreas Marek's avatar
Andreas Marek committed
143
static __forceinline void hh_trafo_complex_kernel_8_AVX_2hv(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s)
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164
{
	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]);
165
#ifndef __ELPA_USE_FMA__
166 167 168 169 170 171 172 173 174 175
	// 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);
176 177
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
178 179 180 181
#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);
182 183
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
184 185 186 187
#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);
188 189
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
190 191 192 193
#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);
194 195
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_add_pd(y4, _mm256_FMSUBADD_pd(h2_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
196 197 198 199 200 201 202 203 204 205 206 207 208
#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]);
209
#ifndef __ELPA_USE_FMA__
210 211 212 213 214
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
215 216
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
217 218 219 220
#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);
221 222
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
223 224 225 226
#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);
227 228
#ifdef __ELPA_USE_FMA__
		x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
229 230 231 232
#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);
233 234
#ifdef __ELPA_USE_FMA__
		x4 = _mm256_add_pd(x4, _mm256_FMSUBADD_pd(h1_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
235 236 237 238 239 240
#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]);
241
#ifndef __ELPA_USE_FMA__
242 243 244 245 246
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
247 248
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
249 250 251 252
#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);
253 254
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
255 256 257 258
#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);
259 260
#ifdef __ELPA_USE_FMA__
		y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
261 262 263 264
#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);
265 266
#ifdef __ELPA_USE_FMA__
		y4 = _mm256_add_pd(y4, _mm256_FMSUBADD_pd(h2_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
267 268 269 270 271 272 273
#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]);
274
#ifndef __ELPA_USE_FMA__
275 276 277 278 279 280 281 282 283 284
	// 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);
285 286
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
287 288 289 290
#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);
291 292
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
293 294 295 296
#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);
297 298
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
299 300 301 302
#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);
303 304
#ifdef __ELPA_USE_FMA__
	x4 = _mm256_add_pd(x4, _mm256_FMSUBADD_pd(h1_real, q4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
305 306 307 308 309 310 311 312 313 314
#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);
315 316
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
317 318 319 320
#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);
321 322
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
323 324 325 326
#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);
327 328
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
329 330 331 332
#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);
333 334
#ifdef __ELPA_USE_FMA__
	x4 = _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5));
335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351
#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);
352 353
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
354 355 356 357 358 359 360 361
#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);
362 363
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
364 365 366 367
#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);
368 369
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
370 371 372 373
#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);
374 375
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_FMADDSUB_pd(h1_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
376 377 378 379
#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);
380 381
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_FMADDSUB_pd(h1_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5));
382 383 384 385 386
#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);
387 388
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMADDSUB_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
389 390 391 392
#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);
393 394
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMADDSUB_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
395 396 397 398
#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);
399 400
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMADDSUB_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
401 402 403 404
#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);
405 406
#ifdef __ELPA_USE_FMA__
	y4 = _mm256_add_pd(y4, _mm256_FMADDSUB_pd(h2_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439
#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);
440 441
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
442 443 444 445 446
#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_
447
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
448 449 450 451
#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);
452 453
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
454 455 456 457
#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);
458 459
#ifdef __ELPA_USE_FMA__
	q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h2_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479
#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);
480 481
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
482 483 484 485
#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);
486 487
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
488 489 490 491
#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);
492 493
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
494 495 496 497
#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);
498 499
#ifdef __ELPA_USE_FMA__
		q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
500 501 502 503 504 505 506 507
#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);
508 509
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
510 511 512 513
#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);
514 515
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
516 517 518 519
#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);
520 521
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
522 523 524 525
#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);
526 527
#ifdef __ELPA_USE_FMA__
		q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h2_real, y4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
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(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);
546 547
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
548 549 550 551
#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);
552 553
#ifdef __ELPA_USE_FMA__
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
554 555 556 557
#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);
558 559
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
560 561 562 563
#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);
564 565
#ifdef __ELPA_USE_FMA__
	q4 = _mm256_add_pd(q4, _mm256_FMADDSUB_pd(h1_real, x4, _mm256_shuffle_pd(tmp4, tmp4, 0x5)));
566 567 568 569 570 571 572 573 574 575
#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);
}

Andreas Marek's avatar
Andreas Marek committed
576
static __forceinline void hh_trafo_complex_kernel_6_AVX_2hv(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s)
577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596
{
	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]);
597
#ifndef __ELPA_USE_FMA__
598 599 600 601 602 603 604 605 606
	// 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);
607 608
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
609 610 611 612
#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);
613 614
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
615 616 617 618
#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);
619 620
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
621 622 623 624 625 626 627 628 629 630 631 632
#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]);
633
#ifndef __ELPA_USE_FMA__
634 635 636 637 638
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
639 640
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
641 642 643 644
#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);
645 646
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
647 648 649 650
#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);
651 652
#ifdef __ELPA_USE_FMA__
		x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
653 654 655 656 657 658
#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]);
659
#ifndef __ELPA_USE_FMA__
660 661 662 663 664
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
665 666
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
667 668 669 670
#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);
671 672
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
673 674 675 676
#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);
677 678
#ifdef __ELPA_USE_FMA__
		y3 = _mm256_add_pd(y3, _mm256_FMSUBADD_pd(h2_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
679 680 681 682 683 684 685
#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]);
686
#ifndef __ELPA_USE_FMA__
687 688 689 690 691 692 693 694 695
	// 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);
696 697
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
698 699 700 701
#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);
702 703
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
704 705 706 707
#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);
708 709
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_add_pd(x3, _mm256_FMSUBADD_pd(h1_real, q3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
710 711 712 713 714 715 716 717 718 719
#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);
720 721
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
722 723 724 725
#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);
726 727
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
728 729 730 731
#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);
732 733
#ifdef __ELPA_USE_FMA__
	x3 = _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750
#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);
751 752
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
753 754 755 756 757 758 759 760
#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);
761 762
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
763 764 765 766
#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);
767 768
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
769 770 771 772
#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);
773 774
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_FMADDSUB_pd(h1_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5));
775 776 777 778 779
#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);
780 781
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMADDSUB_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
782 783 784 785
#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);
786 787
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMADDSUB_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
788 789 790 791
#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);
792 793
#ifdef __ELPA_USE_FMA__
	y3 = _mm256_add_pd(y3, _mm256_FMADDSUB_pd(h2_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821
#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);
822 823
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
824 825 826 827 828
#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_
829
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
830 831 832 833
#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);
834 835
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853
#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);
854 855
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
856 857 858 859
#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);
860 861
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
862 863 864 865
#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);
866 867
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
868 869 870 871 872 873 874 875
#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);
876 877
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h2_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
878 879 880 881
#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);
882 883
#ifdef __ELPA_USE_FMA__
		q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h2_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
884 885 886 887
#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);
888 889
#ifdef __ELPA_USE_FMA__
		q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h2_real, y3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905
#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);
906 907
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_add_pd(q1, _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
908 909 910 911
#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);
912 913
#ifdef __ELPA_USE_FMA__
	q2 = _mm256_add_pd(q2, _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
914 915 916 917
#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);
918 919
#ifdef __ELPA_USE_FMA__
	q3 = _mm256_add_pd(q3, _mm256_FMADDSUB_pd(h1_real, x3, _mm256_shuffle_pd(tmp3, tmp3, 0x5)));
920 921 922 923 924 925 926 927 928
#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);
}

Andreas Marek's avatar
Andreas Marek committed
929
static __forceinline void hh_trafo_complex_kernel_4_AVX_2hv(double complex* q, double complex* hh, int nb, int ldq, int ldh, double complex s)
930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948
{
	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]);
949
#ifndef __ELPA_USE_FMA__
950 951 952 953 954 955 956 957
	// 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);
958 959
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
960 961 962 963
#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);
964 965
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
966 967 968 969 970 971 972 973 974 975 976
#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]);
977
#ifndef __ELPA_USE_FMA__
978 979 980 981 982
		// conjugate
		h1_imag = _mm256_xor_pd(h1_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h1_imag, q1);
983 984
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
985 986 987 988
#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);
989 990
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
991 992 993 994 995 996
#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]);
997
#ifndef __ELPA_USE_FMA__
998 999 1000 1001 1002
		// conjugate
		h2_imag = _mm256_xor_pd(h2_imag, sign);
#endif

		tmp1 = _mm256_mul_pd(h2_imag, q1);
1003 1004
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_add_pd(y1, _mm256_FMSUBADD_pd(h2_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1005 1006 1007 1008
#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);
1009 1010
#ifdef __ELPA_USE_FMA__
		y2 = _mm256_add_pd(y2, _mm256_FMSUBADD_pd(h2_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1011 1012 1013 1014 1015 1016 1017
#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]);
1018
#ifndef __ELPA_USE_FMA__
1019 1020 1021 1022 1023 1024 1025 1026
	// 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);
1027 1028
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_add_pd(x1, _mm256_FMSUBADD_pd(h1_real, q1, _mm256_shuffle_pd(tmp1, tmp1, 0x5)));
1029 1030 1031 1032
#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);
1033 1034
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_add_pd(x2, _mm256_FMSUBADD_pd(h1_real, q2, _mm256_shuffle_pd(tmp2, tmp2, 0x5)));
1035 1036 1037 1038 1039 1040 1041 1042 1043 1044
#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);
1045 1046
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMADDSUB_pd(h1_real, x1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1047 1048 1049 1050
#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);
1051 1052
#ifdef __ELPA_USE_FMA__
	x2 = _mm256_FMADDSUB_pd(h1_real, x2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069
#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);
1070 1071
#ifdef __ELPA_USE_FMA__
	tmp2 = _mm256_FMADDSUB_pd(h2_real, tmp2, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1072 1073 1074 1075 1076 1077 1078 1079
#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);
1080 1081
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMADDSUB_pd(h1_real, y1, _mm256_shuffle_pd(tmp1, tmp1, 0x5));
1082 1083 1084 1085
#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);
1086 1087
#ifdef __ELPA_USE_FMA__
	y2 = _mm256_FMADDSUB_pd(h1_real, y2, _mm256_shuffle_pd(tmp2, tmp2, 0x5));
1088 1089 1090 1091 1092
#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);