elpa2_kernels_real_avx-avx2_4hv.c 78.9 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
// --------------------------------------------------------------------------------------------------

#include <x86intrin.h>

#define __forceinline __attribute__((always_inline)) static

#ifdef __USE_AVX128__
#undef __AVX__
#endif

71 72 73 74
#ifdef __FMA4__
#define __ELPA_USE_FMA__
#define _mm256_FMA_pd(a,b,c) _mm256_macc_pd(a,b,c)
#define _mm256_NFMA_pd(a,b,c) _mm256_nmacc_pd(a,b,c)
Andreas Marek's avatar
Andreas Marek committed
75
#define _mm256_FMSUB_pd(a,b,c) _mm256_msub(a,b,c)
76 77 78 79 80 81
#endif

#ifdef __AVX2__
#define __ELPA_USE_FMA__
#define _mm256_FMA_pd(a,b,c) _mm256_fmadd_pd(a,b,c)
#define _mm256_NFMA_pd(a,b,c) _mm256_fnmadd_pd(a,b,c)
Andreas Marek's avatar
Andreas Marek committed
82
#define _mm256_FMSUB_pd(a,b,c) _mm256_fmsub_pd(a,b,c)
83 84
#endif

85 86 87 88 89 90 91 92 93 94 95
//Forward declaration
#ifdef __AVX__
__forceinline void hh_trafo_kernel_4_AVX_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4);
__forceinline void hh_trafo_kernel_8_AVX_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4);
__forceinline void hh_trafo_kernel_12_AVX_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4);
#else
__forceinline void hh_trafo_kernel_2_SSE_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4);
__forceinline void hh_trafo_kernel_4_SSE_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4);
__forceinline void hh_trafo_kernel_6_SSE_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4);
#endif

Andreas Marek's avatar
Andreas Marek committed
96
void quad_hh_trafo_real_sse_avx_4hv_(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh);
97 98 99 100
#if 0
void quad_hh_trafo_fast_(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh);
#endif

Andreas Marek's avatar
Andreas Marek committed
101
void quad_hh_trafo_real_sse_avx_4hv_(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh)
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282
{
	int i;
	int nb = *pnb;
	int nq = *pldq;
	int ldq = *pldq;
	int ldh = *pldh;

	// calculating scalar products to compute
	// 4 householder vectors simultaneously
	double s_1_2 = hh[(ldh)+1];
	double s_1_3 = hh[(ldh*2)+2];
	double s_2_3 = hh[(ldh*2)+1];
	double s_1_4 = hh[(ldh*3)+3];
	double s_2_4 = hh[(ldh*3)+2];
	double s_3_4 = hh[(ldh*3)+1];

	// calculate scalar product of first and fourth householder vector
	// loop counter = 2
	s_1_2 += hh[2-1] * hh[(2+ldh)];
	s_2_3 += hh[(ldh)+2-1] * hh[2+(ldh*2)];
	s_3_4 += hh[(ldh*2)+2-1] * hh[2+(ldh*3)];

	// loop counter = 3
	s_1_2 += hh[3-1] * hh[(3+ldh)];
	s_2_3 += hh[(ldh)+3-1] * hh[3+(ldh*2)];
	s_3_4 += hh[(ldh*2)+3-1] * hh[3+(ldh*3)];

	s_1_3 += hh[3-2] * hh[3+(ldh*2)];
	s_2_4 += hh[(ldh*1)+3-2] * hh[3+(ldh*3)];

	#pragma ivdep
	for (i = 4; i < nb; i++)
	{
		s_1_2 += hh[i-1] * hh[(i+ldh)];
		s_2_3 += hh[(ldh)+i-1] * hh[i+(ldh*2)];
		s_3_4 += hh[(ldh*2)+i-1] * hh[i+(ldh*3)];

		s_1_3 += hh[i-2] * hh[i+(ldh*2)];
		s_2_4 += hh[(ldh*1)+i-2] * hh[i+(ldh*3)];

		s_1_4 += hh[i-3] * hh[i+(ldh*3)];
	}

//	printf("s_1_2: %f\n", s_1_2);
//	printf("s_1_3: %f\n", s_1_3);
//	printf("s_2_3: %f\n", s_2_3);
//	printf("s_1_4: %f\n", s_1_4);
//	printf("s_2_4: %f\n", s_2_4);
//	printf("s_3_4: %f\n", s_3_4);

	// Production level kernel calls with padding
#ifdef __AVX__
	for (i = 0; i < nq-8; i+=12)
	{
		hh_trafo_kernel_12_AVX_4hv(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
	}
	if (nq == i)
	{
		return;
	}
	else
	{
		if (nq-i > 4)
		{
			hh_trafo_kernel_8_AVX_4hv(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
		}
		else
		{
			hh_trafo_kernel_4_AVX_4hv(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
		}
	}
#else
	for (i = 0; i < nq-4; i+=6)
	{
		hh_trafo_kernel_6_SSE_4hv(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
	}
	if (nq == i)
	{
		return;
	}
	else
	{
		if (nq-i > 2)
		{
			hh_trafo_kernel_4_SSE_4hv(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
		}
		else
		{
			hh_trafo_kernel_2_SSE_4hv(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
		}
	}
#endif
}

#if 0
void quad_hh_trafo_fast_(double* q, double* hh, int* pnb, int* pnq, int* pldq, int* pldh)
{
	int i;
	int nb = *pnb;
	int nq = *pldq;
	int ldq = *pldq;
	int ldh = *pldh;

	// calculating scalar products to compute
	// 4 householder vectors simultaneously
	double s_1_2 = hh[(ldh)+1];
	double s_1_3 = hh[(ldh*2)+2];
	double s_2_3 = hh[(ldh*2)+1];
	double s_1_4 = hh[(ldh*3)+3];
	double s_2_4 = hh[(ldh*3)+2];
	double s_3_4 = hh[(ldh*3)+1];

	// calculate scalar product of first and fourth householder vector
	// loop counter = 2
	s_1_2 += hh[2-1] * hh[(2+ldh)];
	s_2_3 += hh[(ldh)+2-1] * hh[2+(ldh*2)];
	s_3_4 += hh[(ldh*2)+2-1] * hh[2+(ldh*3)];

	// loop counter = 3
	s_1_2 += hh[3-1] * hh[(3+ldh)];
	s_2_3 += hh[(ldh)+3-1] * hh[3+(ldh*2)];
	s_3_4 += hh[(ldh*2)+3-1] * hh[3+(ldh*3)];

	s_1_3 += hh[3-2] * hh[3+(ldh*2)];
	s_2_4 += hh[(ldh*1)+3-2] * hh[3+(ldh*3)];

	#pragma ivdep
	for (i = 4; i < nb; i++)
	{
		s_1_2 += hh[i-1] * hh[(i+ldh)];
		s_2_3 += hh[(ldh)+i-1] * hh[i+(ldh*2)];
		s_3_4 += hh[(ldh*2)+i-1] * hh[i+(ldh*3)];

		s_1_3 += hh[i-2] * hh[i+(ldh*2)];
		s_2_4 += hh[(ldh*1)+i-2] * hh[i+(ldh*3)];

		s_1_4 += hh[i-3] * hh[i+(ldh*3)];
	}

	// Production level kernel calls with padding
#ifdef __AVX__
	for (i = 0; i < nq; i+=12)
	{
		hh_trafo_kernel_12_AVX_4hv(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
	}
#else
	for (i = 0; i < nq; i+=6)
	{
		hh_trafo_kernel_6_SSE_4hv(&q[i], hh, nb, ldq, ldh, s_1_2, s_1_3, s_2_3, s_1_4, s_2_4, s_3_4);
	}
#endif
}
#endif

#ifdef __AVX__
/**
 * Unrolled kernel that computes
 * 12 rows of Q simultaneously, a
 * matrix vector product with two householder
 * vectors + a rank 1 update is performed
 */
__forceinline void hh_trafo_kernel_12_AVX_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4)
{
	/////////////////////////////////////////////////////
	// Matrix Vector Multiplication, Q [12 x nb+3] * hh
	// hh contains four householder vectors
	/////////////////////////////////////////////////////
	int i;

	__m256d a1_1 = _mm256_load_pd(&q[ldq*3]);
	__m256d a2_1 = _mm256_load_pd(&q[ldq*2]);
	__m256d a3_1 = _mm256_load_pd(&q[ldq]);
	__m256d a4_1 = _mm256_load_pd(&q[0]);

	__m256d h_2_1 = _mm256_broadcast_sd(&hh[ldh+1]);
	__m256d h_3_2 = _mm256_broadcast_sd(&hh[(ldh*2)+1]);
	__m256d h_3_1 = _mm256_broadcast_sd(&hh[(ldh*2)+2]);
	__m256d h_4_3 = _mm256_broadcast_sd(&hh[(ldh*3)+1]);
	__m256d h_4_2 = _mm256_broadcast_sd(&hh[(ldh*3)+2]);
	__m256d h_4_1 = _mm256_broadcast_sd(&hh[(ldh*3)+3]);

283 284 285 286 287 288 289
#ifdef __ELPA_USE_FMA__
	register __m256d w1 = _mm256_FMA_pd(a3_1, h_4_3, a4_1);
	w1 = _mm256_FMA_pd(a2_1, h_4_2, w1);
	w1 = _mm256_FMA_pd(a1_1, h_4_1, w1);
	register __m256d z1 = _mm256_FMA_pd(a2_1, h_3_2, a3_1);
	z1 = _mm256_FMA_pd(a1_1, h_3_1, z1);
	register __m256d y1 = _mm256_FMA_pd(a1_1, h_2_1, a2_1);
290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
	register __m256d x1 = a1_1;
#else
	register __m256d w1 = _mm256_add_pd(a4_1, _mm256_mul_pd(a3_1, h_4_3));
	w1 = _mm256_add_pd(w1, _mm256_mul_pd(a2_1, h_4_2));
	w1 = _mm256_add_pd(w1, _mm256_mul_pd(a1_1, h_4_1));
	register __m256d z1 = _mm256_add_pd(a3_1, _mm256_mul_pd(a2_1, h_3_2));
	z1 = _mm256_add_pd(z1, _mm256_mul_pd(a1_1, h_3_1));
	register __m256d y1 = _mm256_add_pd(a2_1, _mm256_mul_pd(a1_1, h_2_1));
	register __m256d x1 = a1_1;
#endif

	__m256d a1_2 = _mm256_load_pd(&q[(ldq*3)+4]);
	__m256d a2_2 = _mm256_load_pd(&q[(ldq*2)+4]);
	__m256d a3_2 = _mm256_load_pd(&q[ldq+4]);
	__m256d a4_2 = _mm256_load_pd(&q[0+4]);

306 307 308 309 310 311 312
#ifdef __ELPA_USE_FMA__
	register __m256d w2 = _mm256_FMA_pd(a3_2, h_4_3, a4_2);
	w2 = _mm256_FMA_pd(a2_2, h_4_2, w2);
	w2 = _mm256_FMA_pd(a1_2, h_4_1, w2);
	register __m256d z2 = _mm256_FMA_pd(a2_2, h_3_2, a3_2);
	z2 = _mm256_FMA_pd(a1_2, h_3_1, z2);
	register __m256d y2 = _mm256_FMA_pd(a1_2, h_2_1, a2_2);
313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328
	register __m256d x2 = a1_2;
#else
	register __m256d w2 = _mm256_add_pd(a4_2, _mm256_mul_pd(a3_2, h_4_3));
	w2 = _mm256_add_pd(w2, _mm256_mul_pd(a2_2, h_4_2));
	w2 = _mm256_add_pd(w2, _mm256_mul_pd(a1_2, h_4_1));
	register __m256d z2 = _mm256_add_pd(a3_2, _mm256_mul_pd(a2_2, h_3_2));
	z2 = _mm256_add_pd(z2, _mm256_mul_pd(a1_2, h_3_1));
	register __m256d y2 = _mm256_add_pd(a2_2, _mm256_mul_pd(a1_2, h_2_1));
	register __m256d x2 = a1_2;
#endif

	__m256d a1_3 = _mm256_load_pd(&q[(ldq*3)+8]);
	__m256d a2_3 = _mm256_load_pd(&q[(ldq*2)+8]);
	__m256d a3_3 = _mm256_load_pd(&q[ldq+8]);
	__m256d a4_3 = _mm256_load_pd(&q[0+8]);

329 330 331 332 333 334 335
#ifdef __ELPA_USE_FMA__
	register __m256d w3 = _mm256_FMA_pd(a3_3, h_4_3, a4_3);
	w3 = _mm256_FMA_pd(a2_3, h_4_2, w3);
	w3 = _mm256_FMA_pd(a1_3, h_4_1, w3);
	register __m256d z3 = _mm256_FMA_pd(a2_3, h_3_2, a3_3);
	z3 = _mm256_FMA_pd(a1_3, h_3_1, z3);
	register __m256d y3 = _mm256_FMA_pd(a1_3, h_2_1, a2_3);
336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361
	register __m256d x3 = a1_3;
#else
	register __m256d w3 = _mm256_add_pd(a4_3, _mm256_mul_pd(a3_3, h_4_3));
	w3 = _mm256_add_pd(w3, _mm256_mul_pd(a2_3, h_4_2));
	w3 = _mm256_add_pd(w3, _mm256_mul_pd(a1_3, h_4_1));
	register __m256d z3 = _mm256_add_pd(a3_3, _mm256_mul_pd(a2_3, h_3_2));
	z3 = _mm256_add_pd(z3, _mm256_mul_pd(a1_3, h_3_1));
	register __m256d y3 = _mm256_add_pd(a2_3, _mm256_mul_pd(a1_3, h_2_1));
	register __m256d x3 = a1_3;
#endif

	__m256d q1;
	__m256d q2;
	__m256d q3;

	__m256d h1;
	__m256d h2;
	__m256d h3;
	__m256d h4;

	for(i = 4; i < nb; i++)
	{
		h1 = _mm256_broadcast_sd(&hh[i-3]);
		q1 = _mm256_load_pd(&q[i*ldq]);
		q2 = _mm256_load_pd(&q[(i*ldq)+4]);
		q3 = _mm256_load_pd(&q[(i*ldq)+8]);
362 363 364 365
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_FMA_pd(q1, h1, x1);
		x2 = _mm256_FMA_pd(q2, h1, x2);
		x3 = _mm256_FMA_pd(q3, h1, x3);
366 367 368 369 370 371 372
#else
		x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
		x2 = _mm256_add_pd(x2, _mm256_mul_pd(q2,h1));
		x3 = _mm256_add_pd(x3, _mm256_mul_pd(q3,h1));
#endif

		h2 = _mm256_broadcast_sd(&hh[ldh+i-2]);
373 374 375 376
#ifdef __ELPA_USE_FMA__
		y1 = _mm256_FMA_pd(q1, h2, y1);
		y2 = _mm256_FMA_pd(q2, h2, y2);
		y3 = _mm256_FMA_pd(q3, h2, y3);
377 378 379 380 381 382 383
#else
		y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
		y2 = _mm256_add_pd(y2, _mm256_mul_pd(q2,h2));
		y3 = _mm256_add_pd(y3, _mm256_mul_pd(q3,h2));
#endif

		h3 = _mm256_broadcast_sd(&hh[(ldh*2)+i-1]);
384 385 386 387
#ifdef __ELPA_USE_FMA__
		z1 = _mm256_FMA_pd(q1, h3, z1);
		z2 = _mm256_FMA_pd(q2, h3, z2);
		z3 = _mm256_FMA_pd(q3, h3, z3);
388 389 390 391 392 393 394
#else
		z1 = _mm256_add_pd(z1, _mm256_mul_pd(q1,h3));
		z2 = _mm256_add_pd(z2, _mm256_mul_pd(q2,h3));
		z3 = _mm256_add_pd(z3, _mm256_mul_pd(q3,h3));
#endif

		h4 = _mm256_broadcast_sd(&hh[(ldh*3)+i]);
395 396 397 398
#ifdef __ELPA_USE_FMA__
		w1 = _mm256_FMA_pd(q1, h4, w1);
		w2 = _mm256_FMA_pd(q2, h4, w2);
		w3 = _mm256_FMA_pd(q3, h4, w3);
399 400 401 402 403 404 405 406 407 408 409 410 411
#else
		w1 = _mm256_add_pd(w1, _mm256_mul_pd(q1,h4));
		w2 = _mm256_add_pd(w2, _mm256_mul_pd(q2,h4));
		w3 = _mm256_add_pd(w3, _mm256_mul_pd(q3,h4));
#endif
	}

	h1 = _mm256_broadcast_sd(&hh[nb-3]);

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

412 413 414 415
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMA_pd(q1, h1, x1);
	x2 = _mm256_FMA_pd(q2, h1, x2);
	x3 = _mm256_FMA_pd(q3, h1, x3);
416 417 418 419 420 421 422 423
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
	x2 = _mm256_add_pd(x2, _mm256_mul_pd(q2,h1));
	x3 = _mm256_add_pd(x3, _mm256_mul_pd(q3,h1));
#endif

	h2 = _mm256_broadcast_sd(&hh[ldh+nb-2]);
#ifdef __FMA4_
424 425 426
	y1 = _mm256_FMA_pd(q1, h2, y1);
	y2 = _mm256_FMA_pd(q2, h2, y2);
	y3 = _mm256_FMA_pd(q3, h2, y3);
427 428 429 430 431 432 433
#else
	y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
	y2 = _mm256_add_pd(y2, _mm256_mul_pd(q2,h2));
	y3 = _mm256_add_pd(y3, _mm256_mul_pd(q3,h2));
#endif

	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+nb-1]);
434 435 436 437
#ifdef __ELPA_USE_FMA__
	z1 = _mm256_FMA_pd(q1, h3, z1);
	z2 = _mm256_FMA_pd(q2, h3, z2);
	z3 = _mm256_FMA_pd(q3, h3, z3);
438 439 440 441 442 443 444 445 446 447 448 449
#else
	z1 = _mm256_add_pd(z1, _mm256_mul_pd(q1,h3));
	z2 = _mm256_add_pd(z2, _mm256_mul_pd(q2,h3));
	z3 = _mm256_add_pd(z3, _mm256_mul_pd(q3,h3));
#endif

	h1 = _mm256_broadcast_sd(&hh[nb-2]);

	q1 = _mm256_load_pd(&q[(nb+1)*ldq]);
	q2 = _mm256_load_pd(&q[((nb+1)*ldq)+4]);
	q3 = _mm256_load_pd(&q[((nb+1)*ldq)+8]);

450 451 452 453
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMA_pd(q1, h1, x1);
	x2 = _mm256_FMA_pd(q2, h1, x2);
	x3 = _mm256_FMA_pd(q3, h1, x3);
454 455 456 457 458 459 460 461
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
	x2 = _mm256_add_pd(x2, _mm256_mul_pd(q2,h1));
	x3 = _mm256_add_pd(x3, _mm256_mul_pd(q3,h1));
#endif

	h2 = _mm256_broadcast_sd(&hh[(ldh*1)+nb-1]);

462 463 464 465
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMA_pd(q1, h2, y1);
	y2 = _mm256_FMA_pd(q2, h2, y2);
	y3 = _mm256_FMA_pd(q3, h2, y3);
466 467 468 469 470 471 472 473 474 475 476 477
#else
	y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
	y2 = _mm256_add_pd(y2, _mm256_mul_pd(q2,h2));
	y3 = _mm256_add_pd(y3, _mm256_mul_pd(q3,h2));
#endif

	h1 = _mm256_broadcast_sd(&hh[nb-1]);

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

478 479 480 481
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMA_pd(q1, h1, x1);
	x2 = _mm256_FMA_pd(q2, h1, x2);
	x3 = _mm256_FMA_pd(q3, h1, x3);
482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
	x2 = _mm256_add_pd(x2, _mm256_mul_pd(q2,h1));
	x3 = _mm256_add_pd(x3, _mm256_mul_pd(q3,h1));
#endif

	/////////////////////////////////////////////////////
	// Rank-1 update of Q [12 x nb+3]
	/////////////////////////////////////////////////////

	__m256d tau1 = _mm256_broadcast_sd(&hh[0]);

	h1 = tau1;
	x1 = _mm256_mul_pd(x1, h1);
	x2 = _mm256_mul_pd(x2, h1);
	x3 = _mm256_mul_pd(x3, h1);

	__m256d tau2 = _mm256_broadcast_sd(&hh[ldh]);
	__m256d vs_1_2 = _mm256_broadcast_sd(&s_1_2);

	h1 = tau2;
	h2 = _mm256_mul_pd(h1, vs_1_2);
504 505 506 507
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMSUB_pd(y1, h1, _mm256_mul_pd(x1,h2));
	y2 = _mm256_FMSUB_pd(y2, h1, _mm256_mul_pd(x2,h2));
	y3 = _mm256_FMSUB_pd(y3, h1, _mm256_mul_pd(x3,h2));
508 509 510 511 512 513 514 515 516 517 518 519 520
#else
	y1 = _mm256_sub_pd(_mm256_mul_pd(y1,h1), _mm256_mul_pd(x1,h2));
	y2 = _mm256_sub_pd(_mm256_mul_pd(y2,h1), _mm256_mul_pd(x2,h2));
	y3 = _mm256_sub_pd(_mm256_mul_pd(y3,h1), _mm256_mul_pd(x3,h2));
#endif

	__m256d tau3 = _mm256_broadcast_sd(&hh[ldh*2]);
	__m256d vs_1_3 = _mm256_broadcast_sd(&s_1_3);
	__m256d vs_2_3 = _mm256_broadcast_sd(&s_2_3);

	h1 = tau3;
	h2 = _mm256_mul_pd(h1, vs_1_3);
	h3 = _mm256_mul_pd(h1, vs_2_3);
521 522 523 524
#ifdef __ELPA_USE_FMA__
	z1 = _mm256_FMSUB_pd(z1, h1, _mm256_FMA_pd(y1, h3, _mm256_mul_pd(x1,h2)));
	z2 = _mm256_FMSUB_pd(z2, h1, _mm256_FMA_pd(y2, h3, _mm256_mul_pd(x2,h2)));
	z3 = _mm256_FMSUB_pd(z3, h1, _mm256_FMA_pd(y3, h3, _mm256_mul_pd(x3,h2)));
525 526 527 528 529 530 531 532 533 534 535 536 537 538 539
#else
	z1 = _mm256_sub_pd(_mm256_mul_pd(z1,h1), _mm256_add_pd(_mm256_mul_pd(y1,h3), _mm256_mul_pd(x1,h2)));
	z2 = _mm256_sub_pd(_mm256_mul_pd(z2,h1), _mm256_add_pd(_mm256_mul_pd(y2,h3), _mm256_mul_pd(x2,h2)));
	z3 = _mm256_sub_pd(_mm256_mul_pd(z3,h1), _mm256_add_pd(_mm256_mul_pd(y3,h3), _mm256_mul_pd(x3,h2)));
#endif

	__m256d tau4 = _mm256_broadcast_sd(&hh[ldh*3]);
	__m256d vs_1_4 = _mm256_broadcast_sd(&s_1_4);
	__m256d vs_2_4 = _mm256_broadcast_sd(&s_2_4);
	__m256d vs_3_4 = _mm256_broadcast_sd(&s_3_4);

	h1 = tau4;
	h2 = _mm256_mul_pd(h1, vs_1_4);
	h3 = _mm256_mul_pd(h1, vs_2_4);
	h4 = _mm256_mul_pd(h1, vs_3_4);
540 541 542 543
#ifdef __ELPA_USE_FMA__
	w1 = _mm256_FMSUB_pd(w1, h1, _mm256_FMA_pd(z1, h4, _mm256_FMA_pd(y1, h3, _mm256_mul_pd(x1,h2))));
	w2 = _mm256_FMSUB_pd(w2, h1, _mm256_FMA_pd(z2, h4, _mm256_FMA_pd(y2, h3, _mm256_mul_pd(x2,h2))));
	w3 = _mm256_FMSUB_pd(w3, h1, _mm256_FMA_pd(z3, h4, _mm256_FMA_pd(y3, h3, _mm256_mul_pd(x3,h2))));
544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563
#else
	w1 = _mm256_sub_pd(_mm256_mul_pd(w1,h1), _mm256_add_pd(_mm256_mul_pd(z1,h4), _mm256_add_pd(_mm256_mul_pd(y1,h3), _mm256_mul_pd(x1,h2))));
	w2 = _mm256_sub_pd(_mm256_mul_pd(w2,h1), _mm256_add_pd(_mm256_mul_pd(z2,h4), _mm256_add_pd(_mm256_mul_pd(y2,h3), _mm256_mul_pd(x2,h2))));
	w3 = _mm256_sub_pd(_mm256_mul_pd(w3,h1), _mm256_add_pd(_mm256_mul_pd(z3,h4), _mm256_add_pd(_mm256_mul_pd(y3,h3), _mm256_mul_pd(x3,h2))));
#endif

	q1 = _mm256_load_pd(&q[0]);
	q2 = _mm256_load_pd(&q[4]);
	q3 = _mm256_load_pd(&q[8]);
	q1 = _mm256_sub_pd(q1, w1);
	q2 = _mm256_sub_pd(q2, w2);
	q3 = _mm256_sub_pd(q3, w3);
	_mm256_store_pd(&q[0],q1);
	_mm256_store_pd(&q[4],q2);
	_mm256_store_pd(&q[8],q3);

	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+1]);
	q1 = _mm256_load_pd(&q[ldq]);
	q2 = _mm256_load_pd(&q[ldq+4]);
	q3 = _mm256_load_pd(&q[ldq+8]);
564 565 566 567
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_sub_pd(q1, _mm256_FMA_pd(w1, h4, z1));
	q2 = _mm256_sub_pd(q2, _mm256_FMA_pd(w2, h4, z2));
	q3 = _mm256_sub_pd(q3, _mm256_FMA_pd(w3, h4, z3));
568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd(z1, _mm256_mul_pd(w1, h4)));
	q2 = _mm256_sub_pd(q2, _mm256_add_pd(z2, _mm256_mul_pd(w2, h4)));
	q3 = _mm256_sub_pd(q3, _mm256_add_pd(z3, _mm256_mul_pd(w3, h4)));
#endif
	_mm256_store_pd(&q[ldq],q1);
	_mm256_store_pd(&q[ldq+4],q2);
	_mm256_store_pd(&q[ldq+8],q3);

	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+2]);
	q1 = _mm256_load_pd(&q[ldq*2]);
	q2 = _mm256_load_pd(&q[(ldq*2)+4]);
	q3 = _mm256_load_pd(&q[(ldq*2)+8]);
	q1 = _mm256_sub_pd(q1, y1);
	q2 = _mm256_sub_pd(q2, y2);
	q3 = _mm256_sub_pd(q3, y3);
584 585 586 587
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(w1, h4, q1);
	q2 = _mm256_NFMA_pd(w2, h4, q2);
	q3 = _mm256_NFMA_pd(w3, h4, q3);
588 589 590 591 592 593
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(w1, h4));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(w2, h4));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(w3, h4));
#endif
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+1]);
594 595 596 597
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(z1, h3, q1);
	q2 = _mm256_NFMA_pd(z2, h3, q2);
	q3 = _mm256_NFMA_pd(z3, h3, q3);
598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(z1, h3));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(z2, h3));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(z3, h3));
#endif
	_mm256_store_pd(&q[ldq*2],q1);
	_mm256_store_pd(&q[(ldq*2)+4],q2);
	_mm256_store_pd(&q[(ldq*2)+8],q3);

	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+3]);
	q1 = _mm256_load_pd(&q[ldq*3]);
	q2 = _mm256_load_pd(&q[(ldq*3)+4]);
	q3 = _mm256_load_pd(&q[(ldq*3)+8]);
	q1 = _mm256_sub_pd(q1, x1);
	q2 = _mm256_sub_pd(q2, x2);
	q3 = _mm256_sub_pd(q3, x3);
614 615 616 617
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(w1, h4, q1);
	q2 = _mm256_NFMA_pd(w2, h4, q2);
	q3 = _mm256_NFMA_pd(w3, h4, q3);
618 619 620 621 622 623
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(w1, h4));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(w2, h4));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(w3, h4));
#endif
	h2 = _mm256_broadcast_sd(&hh[ldh+1]);
624 625 626 627
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(y1, h2, q1);
	q2 = _mm256_NFMA_pd(y2, h2, q2);
	q3 = _mm256_NFMA_pd(y3, h2, q3);
628 629 630 631 632 633
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(y1, h2));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(y2, h2));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(y3, h2));
#endif
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+2]);
634 635 636 637
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(z1, h3, q1);
	q2 = _mm256_NFMA_pd(z2, h3, q2);
	q3 = _mm256_NFMA_pd(z3, h3, q3);
638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(z1, h3));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(z2, h3));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(z3, h3));
#endif
	_mm256_store_pd(&q[ldq*3], q1);
	_mm256_store_pd(&q[(ldq*3)+4], q2);
	_mm256_store_pd(&q[(ldq*3)+8], q3);

	for (i = 4; i < nb; i++)
	{
		h1 = _mm256_broadcast_sd(&hh[i-3]);

		q1 = _mm256_load_pd(&q[i*ldq]);
		q2 = _mm256_load_pd(&q[(i*ldq)+4]);
		q3 = _mm256_load_pd(&q[(i*ldq)+8]);

655 656 657 658
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_NFMA_pd(x1, h1, q1);
		q2 = _mm256_NFMA_pd(x2, h1, q2);
		q3 = _mm256_NFMA_pd(x3, h1, q3);
659 660 661 662 663 664 665
#else
		q1 = _mm256_sub_pd(q1, _mm256_mul_pd(x1,h1));
		q2 = _mm256_sub_pd(q2, _mm256_mul_pd(x2,h1));
		q3 = _mm256_sub_pd(q3, _mm256_mul_pd(x3,h1));
#endif

		h2 = _mm256_broadcast_sd(&hh[ldh+i-2]);
666 667 668 669
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_NFMA_pd(y1, h2, q1);
		q2 = _mm256_NFMA_pd(y2, h2, q2);
		q3 = _mm256_NFMA_pd(y3, h2, q3);
670 671 672 673 674 675 676
#else
		q1 = _mm256_sub_pd(q1, _mm256_mul_pd(y1,h2));
		q2 = _mm256_sub_pd(q2, _mm256_mul_pd(y2,h2));
		q3 = _mm256_sub_pd(q3, _mm256_mul_pd(y3,h2));
#endif

		h3 = _mm256_broadcast_sd(&hh[(ldh*2)+i-1]);
677 678 679 680
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_NFMA_pd(z1, h3, q1);
		q2 = _mm256_NFMA_pd(z2, h3, q2);
		q3 = _mm256_NFMA_pd(z3, h3, q3);
681 682 683 684 685 686 687
#else
		q1 = _mm256_sub_pd(q1, _mm256_mul_pd(z1,h3));
		q2 = _mm256_sub_pd(q2, _mm256_mul_pd(z2,h3));
		q3 = _mm256_sub_pd(q3, _mm256_mul_pd(z3,h3));
#endif

		h4 = _mm256_broadcast_sd(&hh[(ldh*3)+i]);
688 689 690 691
#ifdef __ELPA_USE_FMA__
		q1 = _mm256_NFMA_pd(w1, h4, q1);
		q2 = _mm256_NFMA_pd(w2, h4, q2);
		q3 = _mm256_NFMA_pd(w3, h4, q3);
692 693 694 695 696 697 698 699 700 701 702 703 704 705 706
#else
		q1 = _mm256_sub_pd(q1, _mm256_mul_pd(w1,h4));
		q2 = _mm256_sub_pd(q2, _mm256_mul_pd(w2,h4));
		q3 = _mm256_sub_pd(q3, _mm256_mul_pd(w3,h4));
#endif

		_mm256_store_pd(&q[i*ldq],q1);
		_mm256_store_pd(&q[(i*ldq)+4],q2);
		_mm256_store_pd(&q[(i*ldq)+8],q3);
	}

	h1 = _mm256_broadcast_sd(&hh[nb-3]);
	q1 = _mm256_load_pd(&q[nb*ldq]);
	q2 = _mm256_load_pd(&q[(nb*ldq)+4]);
	q3 = _mm256_load_pd(&q[(nb*ldq)+8]);
707 708 709 710
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(x1, h1, q1);
	q2 = _mm256_NFMA_pd(x2, h1, q2);
	q3 = _mm256_NFMA_pd(x3, h1, q3);
711 712 713 714 715 716
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(x1,h1));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(x2,h1));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(x3,h1));
#endif
	h2 = _mm256_broadcast_sd(&hh[ldh+nb-2]);
717 718 719 720
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(y1, h2, q1);
	q2 = _mm256_NFMA_pd(y2, h2, q2);
	q3 = _mm256_NFMA_pd(y3, h2, q3);
721 722 723 724 725 726
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(y1,h2));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(y2,h2));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(y3,h2));
#endif
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+nb-1]);
727 728 729 730
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(z1, h3, q1);
	q2 = _mm256_NFMA_pd(z2, h3, q2);
	q3 = _mm256_NFMA_pd(z3, h3, q3);
731 732 733 734 735 736 737 738 739 740 741 742 743
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(z1,h3));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(z2,h3));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(z3,h3));
#endif
	_mm256_store_pd(&q[nb*ldq],q1);
	_mm256_store_pd(&q[(nb*ldq)+4],q2);
	_mm256_store_pd(&q[(nb*ldq)+8],q3);

	h1 = _mm256_broadcast_sd(&hh[nb-2]);
	q1 = _mm256_load_pd(&q[(nb+1)*ldq]);
	q2 = _mm256_load_pd(&q[((nb+1)*ldq)+4]);
	q3 = _mm256_load_pd(&q[((nb+1)*ldq)+8]);
744 745 746 747
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(x1, h1, q1);
	q2 = _mm256_NFMA_pd(x2, h1, q2);
	q3 = _mm256_NFMA_pd(x3, h1, q3);
748 749 750 751 752 753
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(x1,h1));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(x2,h1));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(x3,h1));
#endif
	h2 = _mm256_broadcast_sd(&hh[ldh+nb-1]);
754 755 756 757
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(y1, h2, q1);
	q2 = _mm256_NFMA_pd(y2, h2, q2);
	q3 = _mm256_NFMA_pd(y3, h2, q3);
758 759 760 761 762 763 764 765 766 767 768 769 770
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(y1,h2));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(y2,h2));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(y3,h2));
#endif
	_mm256_store_pd(&q[(nb+1)*ldq],q1);
	_mm256_store_pd(&q[((nb+1)*ldq)+4],q2);
	_mm256_store_pd(&q[((nb+1)*ldq)+8],q3);

	h1 = _mm256_broadcast_sd(&hh[nb-1]);
	q1 = _mm256_load_pd(&q[(nb+2)*ldq]);
	q2 = _mm256_load_pd(&q[((nb+2)*ldq)+4]);
	q3 = _mm256_load_pd(&q[((nb+2)*ldq)+8]);
771 772 773 774
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(x1, h1, q1);
	q2 = _mm256_NFMA_pd(x2, h1, q2);
	q3 = _mm256_NFMA_pd(x3, h1, q3);
775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(x1,h1));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(x2,h1));
	q3 = _mm256_sub_pd(q3, _mm256_mul_pd(x3,h1));
#endif
	_mm256_store_pd(&q[(nb+2)*ldq],q1);
	_mm256_store_pd(&q[((nb+2)*ldq)+4],q2);
	_mm256_store_pd(&q[((nb+2)*ldq)+8],q3);
}

/**
 * Unrolled kernel that computes
 * 8 rows of Q simultaneously, a
 * matrix vector product with two householder
 * vectors + a rank 1 update is performed
 */
__forceinline void hh_trafo_kernel_8_AVX_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4)
{
	/////////////////////////////////////////////////////
	// Matrix Vector Multiplication, Q [4 x nb+3] * hh
	// hh contains four householder vectors
	/////////////////////////////////////////////////////
	int i;

	__m256d a1_1 = _mm256_load_pd(&q[ldq*3]);
	__m256d a2_1 = _mm256_load_pd(&q[ldq*2]);
	__m256d a3_1 = _mm256_load_pd(&q[ldq]);
	__m256d a4_1 = _mm256_load_pd(&q[0]);

	__m256d h_2_1 = _mm256_broadcast_sd(&hh[ldh+1]);
	__m256d h_3_2 = _mm256_broadcast_sd(&hh[(ldh*2)+1]);
	__m256d h_3_1 = _mm256_broadcast_sd(&hh[(ldh*2)+2]);
	__m256d h_4_3 = _mm256_broadcast_sd(&hh[(ldh*3)+1]);
	__m256d h_4_2 = _mm256_broadcast_sd(&hh[(ldh*3)+2]);
	__m256d h_4_1 = _mm256_broadcast_sd(&hh[(ldh*3)+3]);

811 812 813 814 815 816 817
#ifdef __ELPA_USE_FMA__
	__m256d w1 = _mm256_FMA_pd(a3_1, h_4_3, a4_1);
	w1 = _mm256_FMA_pd(a2_1, h_4_2, w1);
	w1 = _mm256_FMA_pd(a1_1, h_4_1, w1);
	__m256d z1 = _mm256_FMA_pd(a2_1, h_3_2, a3_1);
	z1 = _mm256_FMA_pd(a1_1, h_3_1, z1);
	__m256d y1 = _mm256_FMA_pd(a1_1, h_2_1, a2_1);
818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833
	__m256d x1 = a1_1;
#else
	__m256d w1 = _mm256_add_pd(a4_1, _mm256_mul_pd(a3_1, h_4_3));
	w1 = _mm256_add_pd(w1, _mm256_mul_pd(a2_1, h_4_2));
	w1 = _mm256_add_pd(w1, _mm256_mul_pd(a1_1, h_4_1));
	__m256d z1 = _mm256_add_pd(a3_1, _mm256_mul_pd(a2_1, h_3_2));
	z1 = _mm256_add_pd(z1, _mm256_mul_pd(a1_1, h_3_1));
	__m256d y1 = _mm256_add_pd(a2_1, _mm256_mul_pd(a1_1, h_2_1));
	__m256d x1 = a1_1;
#endif

	__m256d a1_2 = _mm256_load_pd(&q[(ldq*3)+4]);
	__m256d a2_2 = _mm256_load_pd(&q[(ldq*2)+4]);
	__m256d a3_2 = _mm256_load_pd(&q[ldq+4]);
	__m256d a4_2 = _mm256_load_pd(&q[0+4]);

834 835 836 837 838 839 840
#ifdef __ELPA_USE_FMA__
	__m256d w2 = _mm256_FMA_pd(a3_2, h_4_3, a4_2);
	w2 = _mm256_FMA_pd(a2_2, h_4_2, w2);
	w2 = _mm256_FMA_pd(a1_2, h_4_1, w2);
	__m256d z2 = _mm256_FMA_pd(a2_2, h_3_2, a3_2);
	z2 = _mm256_FMA_pd(a1_2, h_3_1, z2);
	__m256d y2 = _mm256_FMA_pd(a1_2, h_2_1, a2_2);
841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867
	__m256d x2 = a1_2;
#else
	__m256d w2 = _mm256_add_pd(a4_2, _mm256_mul_pd(a3_2, h_4_3));
	w2 = _mm256_add_pd(w2, _mm256_mul_pd(a2_2, h_4_2));
	w2 = _mm256_add_pd(w2, _mm256_mul_pd(a1_2, h_4_1));
	__m256d z2 = _mm256_add_pd(a3_2, _mm256_mul_pd(a2_2, h_3_2));
	z2 = _mm256_add_pd(z2, _mm256_mul_pd(a1_2, h_3_1));
	__m256d y2 = _mm256_add_pd(a2_2, _mm256_mul_pd(a1_2, h_2_1));
	__m256d x2 = a1_2;
#endif

	__m256d q1;
	__m256d q2;

	__m256d h1;
	__m256d h2;
	__m256d h3;
	__m256d h4;

	for(i = 4; i < nb; i++)
	{
		h1 = _mm256_broadcast_sd(&hh[i-3]);
		h2 = _mm256_broadcast_sd(&hh[ldh+i-2]);
		h3 = _mm256_broadcast_sd(&hh[(ldh*2)+i-1]);
		h4 = _mm256_broadcast_sd(&hh[(ldh*3)+i]);

		q1 = _mm256_load_pd(&q[i*ldq]);
868 869 870 871 872
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_FMA_pd(q1, h1, x1);
		y1 = _mm256_FMA_pd(q1, h2, y1);
		z1 = _mm256_FMA_pd(q1, h3, z1);
		w1 = _mm256_FMA_pd(q1, h4, w1);
873 874 875 876 877 878 879 880
#else
		x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
		y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
		z1 = _mm256_add_pd(z1, _mm256_mul_pd(q1,h3));
		w1 = _mm256_add_pd(w1, _mm256_mul_pd(q1,h4));
#endif

		q2 = _mm256_load_pd(&q[(i*ldq)+4]);
881 882 883 884 885
#ifdef __ELPA_USE_FMA__
		x2 = _mm256_FMA_pd(q2, h1, x2);
		y2 = _mm256_FMA_pd(q2, h2, y2);
		z2 = _mm256_FMA_pd(q2, h3, z2);
		w2 = _mm256_FMA_pd(q2, h4, w2);
886 887 888 889 890 891 892 893 894 895 896 897 898 899 900
#else
		x2 = _mm256_add_pd(x2, _mm256_mul_pd(q2,h1));
		y2 = _mm256_add_pd(y2, _mm256_mul_pd(q2,h2));
		z2 = _mm256_add_pd(z2, _mm256_mul_pd(q2,h3));
		w2 = _mm256_add_pd(w2, _mm256_mul_pd(q2,h4));
#endif
	}

	h1 = _mm256_broadcast_sd(&hh[nb-3]);
	h2 = _mm256_broadcast_sd(&hh[ldh+nb-2]);
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+nb-1]);

	q1 = _mm256_load_pd(&q[nb*ldq]);
	q2 = _mm256_load_pd(&q[(nb*ldq)+4]);

901 902 903 904 905 906 907
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMA_pd(q1, h1, x1);
	x2 = _mm256_FMA_pd(q2, h1, x2);
	y1 = _mm256_FMA_pd(q1, h2, y1);
	y2 = _mm256_FMA_pd(q2, h2, y2);
	z1 = _mm256_FMA_pd(q1, h3, z1);
	z2 = _mm256_FMA_pd(q2, h3, z2);
908 909 910 911 912 913 914 915 916 917 918 919 920 921 922
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
	x2 = _mm256_add_pd(x2, _mm256_mul_pd(q2,h1));
	y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
	y2 = _mm256_add_pd(y2, _mm256_mul_pd(q2,h2));
	z1 = _mm256_add_pd(z1, _mm256_mul_pd(q1,h3));
	z2 = _mm256_add_pd(z2, _mm256_mul_pd(q2,h3));
#endif

	h1 = _mm256_broadcast_sd(&hh[nb-2]);
	h2 = _mm256_broadcast_sd(&hh[(ldh*1)+nb-1]);

	q1 = _mm256_load_pd(&q[(nb+1)*ldq]);
	q2 = _mm256_load_pd(&q[((nb+1)*ldq)+4]);

923 924 925 926 927
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMA_pd(q1, h1, x1);
	x2 = _mm256_FMA_pd(q2, h1, x2);
	y1 = _mm256_FMA_pd(q1, h2, y1);
	y2 = _mm256_FMA_pd(q2, h2, y2);
928 929 930 931 932 933 934 935 936 937 938 939
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
	x2 = _mm256_add_pd(x2, _mm256_mul_pd(q2,h1));
	y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
	y2 = _mm256_add_pd(y2, _mm256_mul_pd(q2,h2));
#endif

	h1 = _mm256_broadcast_sd(&hh[nb-1]);

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

940 941 942
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMA_pd(q1, h1, x1);
	x2 = _mm256_FMA_pd(q2, h1, x2);
943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
	x2 = _mm256_add_pd(x2, _mm256_mul_pd(q2,h1));
#endif

	/////////////////////////////////////////////////////
	// Rank-1 update of Q [8 x nb+3]
	/////////////////////////////////////////////////////

	__m256d tau1 = _mm256_broadcast_sd(&hh[0]);
	__m256d tau2 = _mm256_broadcast_sd(&hh[ldh]);
	__m256d tau3 = _mm256_broadcast_sd(&hh[ldh*2]);
	__m256d tau4 = _mm256_broadcast_sd(&hh[ldh*3]);

	__m256d vs_1_2 = _mm256_broadcast_sd(&s_1_2);
	__m256d vs_1_3 = _mm256_broadcast_sd(&s_1_3);
	__m256d vs_2_3 = _mm256_broadcast_sd(&s_2_3);
	__m256d vs_1_4 = _mm256_broadcast_sd(&s_1_4);
	__m256d vs_2_4 = _mm256_broadcast_sd(&s_2_4);
	__m256d vs_3_4 = _mm256_broadcast_sd(&s_3_4);

	h1 = tau1;
	x1 = _mm256_mul_pd(x1, h1);
	x2 = _mm256_mul_pd(x2, h1);

	h1 = tau2;
	h2 = _mm256_mul_pd(h1, vs_1_2);
970 971 972
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMSUB_pd(y1, h1, _mm256_mul_pd(x1,h2));
	y2 = _mm256_FMSUB_pd(y2, h1, _mm256_mul_pd(x2,h2));
973 974 975 976 977 978 979 980
#else
	y1 = _mm256_sub_pd(_mm256_mul_pd(y1,h1), _mm256_mul_pd(x1,h2));
	y2 = _mm256_sub_pd(_mm256_mul_pd(y2,h1), _mm256_mul_pd(x2,h2));
#endif

	h1 = tau3;
	h2 = _mm256_mul_pd(h1, vs_1_3);
	h3 = _mm256_mul_pd(h1, vs_2_3);
981 982 983
#ifdef __ELPA_USE_FMA__
	z1 = _mm256_FMSUB_pd(z1, h1, _mm256_FMA_pd(y1, h3, _mm256_mul_pd(x1,h2)));
	z2 = _mm256_FMSUB_pd(z2, h1, _mm256_FMA_pd(y2, h3, _mm256_mul_pd(x2,h2)));
984 985 986 987 988 989 990 991 992
#else
	z1 = _mm256_sub_pd(_mm256_mul_pd(z1,h1), _mm256_add_pd(_mm256_mul_pd(y1,h3), _mm256_mul_pd(x1,h2)));
	z2 = _mm256_sub_pd(_mm256_mul_pd(z2,h1), _mm256_add_pd(_mm256_mul_pd(y2,h3), _mm256_mul_pd(x2,h2)));
#endif

	h1 = tau4;
	h2 = _mm256_mul_pd(h1, vs_1_4);
	h3 = _mm256_mul_pd(h1, vs_2_4);
	h4 = _mm256_mul_pd(h1, vs_3_4);
993 994 995
#ifdef __ELPA_USE_FMA__
	w1 = _mm256_FMSUB_pd(w1, h1, _mm256_FMA_pd(z1, h4, _mm256_FMA_pd(y1, h3, _mm256_mul_pd(x1,h2))));
	w2 = _mm256_FMSUB_pd(w2, h1, _mm256_FMA_pd(z2, h4, _mm256_FMA_pd(y2, h3, _mm256_mul_pd(x2,h2))));
996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010
#else
	w1 = _mm256_sub_pd(_mm256_mul_pd(w1,h1), _mm256_add_pd(_mm256_mul_pd(z1,h4), _mm256_add_pd(_mm256_mul_pd(y1,h3), _mm256_mul_pd(x1,h2))));
	w2 = _mm256_sub_pd(_mm256_mul_pd(w2,h1), _mm256_add_pd(_mm256_mul_pd(z2,h4), _mm256_add_pd(_mm256_mul_pd(y2,h3), _mm256_mul_pd(x2,h2))));
#endif

	q1 = _mm256_load_pd(&q[0]);
	q2 = _mm256_load_pd(&q[4]);
	q1 = _mm256_sub_pd(q1, w1);
	q2 = _mm256_sub_pd(q2, w2);
	_mm256_store_pd(&q[0],q1);
	_mm256_store_pd(&q[4],q2);

	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+1]);
	q1 = _mm256_load_pd(&q[ldq]);
	q2 = _mm256_load_pd(&q[ldq+4]);
1011 1012 1013
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_sub_pd(q1, _mm256_FMA_pd(w1, h4, z1));
	q2 = _mm256_sub_pd(q2, _mm256_FMA_pd(w2, h4, z2));
1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd(z1, _mm256_mul_pd(w1, h4)));
	q2 = _mm256_sub_pd(q2, _mm256_add_pd(z2, _mm256_mul_pd(w2, h4)));
#endif
	_mm256_store_pd(&q[ldq],q1);
	_mm256_store_pd(&q[ldq+4],q2);

	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+1]);
	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+2]);
	q1 = _mm256_load_pd(&q[ldq*2]);
	q2 = _mm256_load_pd(&q[(ldq*2)+4]);
1025 1026 1027 1028 1029 1030 1031
#ifdef __ELPA_USE_FMA__
        q1 = _mm256_sub_pd(q1, y1);
        q1 = _mm256_NFMA_pd(z1, h3, q1);
        q1 = _mm256_NFMA_pd(w1, h4, q1);
        q2 = _mm256_sub_pd(q2, y2);
        q2 = _mm256_NFMA_pd(z2, h3, q2);
        q2 = _mm256_NFMA_pd(w2, h4, q2);
1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd(y1, _mm256_add_pd(_mm256_mul_pd(z1, h3), _mm256_mul_pd(w1, h4))));
	q2 = _mm256_sub_pd(q2, _mm256_add_pd(y2, _mm256_add_pd(_mm256_mul_pd(z2, h3), _mm256_mul_pd(w2, h4))));
#endif
	_mm256_store_pd(&q[ldq*2],q1);
	_mm256_store_pd(&q[(ldq*2)+4],q2);

	h2 = _mm256_broadcast_sd(&hh[ldh+1]);
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+2]);
	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+3]);
	q1 = _mm256_load_pd(&q[ldq*3]);
	q2 = _mm256_load_pd(&q[(ldq*3)+4]);
1044 1045 1046 1047 1048 1049 1050 1051 1052
#ifdef __ELPA_USE_FMA__
        q1 = _mm256_sub_pd(q1, x1);
        q1 = _mm256_NFMA_pd(y1, h2, q1);
        q1 = _mm256_NFMA_pd(z1, h3, q1);
        q1 = _mm256_NFMA_pd(w1, h4, q1);
        q2 = _mm256_sub_pd(q2, x2);
        q2 = _mm256_NFMA_pd(y2, h2, q2);
        q2 = _mm256_NFMA_pd(z2, h3, q2);
        q2 = _mm256_NFMA_pd(w2, h4, q2);
1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd(x1, _mm256_add_pd(_mm256_mul_pd(y1, h2), _mm256_add_pd(_mm256_mul_pd(z1, h3), _mm256_mul_pd(w1, h4)))));
	q2 = _mm256_sub_pd(q2, _mm256_add_pd(x2, _mm256_add_pd(_mm256_mul_pd(y2, h2), _mm256_add_pd(_mm256_mul_pd(z2, h3), _mm256_mul_pd(w2, h4)))));
#endif
	_mm256_store_pd(&q[ldq*3], q1);
	_mm256_store_pd(&q[(ldq*3)+4], q2);

	for (i = 4; i < nb; i++)
	{
		h1 = _mm256_broadcast_sd(&hh[i-3]);
		h2 = _mm256_broadcast_sd(&hh[ldh+i-2]);
		h3 = _mm256_broadcast_sd(&hh[(ldh*2)+i-1]);
		h4 = _mm256_broadcast_sd(&hh[(ldh*3)+i]);

1067
#ifdef __ELPA_USE_FMA__
1068
		q1 = _mm256_load_pd(&q[i*ldq]);
1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079
		q2 = _mm256_load_pd(&q[(i*ldq)+4]);
                q1 = _mm256_NFMA_pd(x1, h1, q1);
                q1 = _mm256_NFMA_pd(y1, h2, q1);
                q1 = _mm256_NFMA_pd(z1, h3, q1);
                q1 = _mm256_NFMA_pd(w1, h4, q1);
                q2 = _mm256_NFMA_pd(x2, h1, q2);
                q2 = _mm256_NFMA_pd(y2, h2, q2);
                q2 = _mm256_NFMA_pd(z2, h3, q2);
                q2 = _mm256_NFMA_pd(w2, h4, q2);
		_mm256_store_pd(&q[i*ldq],q1);
		_mm256_store_pd(&q[(i*ldq)+4],q2);
1080
#else
1081
		q1 = _mm256_load_pd(&q[i*ldq]);
1082 1083 1084 1085 1086 1087
		q1 = _mm256_sub_pd(q1, _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(w1, h4), _mm256_mul_pd(z1, h3)), _mm256_add_pd(_mm256_mul_pd(x1,h1), _mm256_mul_pd(y1, h2))));
		_mm256_store_pd(&q[i*ldq],q1);

		q2 = _mm256_load_pd(&q[(i*ldq)+4]);
		q2 = _mm256_sub_pd(q2, _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(w2, h4), _mm256_mul_pd(z2, h3)), _mm256_add_pd(_mm256_mul_pd(x2,h1), _mm256_mul_pd(y2, h2))));
		_mm256_store_pd(&q[(i*ldq)+4],q2);
1088
#endif
1089 1090 1091 1092 1093 1094 1095
	}

	h1 = _mm256_broadcast_sd(&hh[nb-3]);
	h2 = _mm256_broadcast_sd(&hh[ldh+nb-2]);
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+nb-1]);
	q1 = _mm256_load_pd(&q[nb*ldq]);
	q2 = _mm256_load_pd(&q[(nb*ldq)+4]);
1096 1097 1098 1099 1100 1101 1102
#ifdef __ELPA_USE_FMA__
        q1 = _mm256_NFMA_pd(x1, h1, q1);
        q1 = _mm256_NFMA_pd(y1, h2, q1);
        q1 = _mm256_NFMA_pd(z1, h3, q1);
        q2 = _mm256_NFMA_pd(x2, h1, q2);
        q2 = _mm256_NFMA_pd(y2, h2, q2);
        q2 = _mm256_NFMA_pd(z2, h3, q2);
1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(z1, h3), _mm256_mul_pd(y1, h2)) , _mm256_mul_pd(x1, h1)));
	q2 = _mm256_sub_pd(q2, _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(z2, h3), _mm256_mul_pd(y2, h2)) , _mm256_mul_pd(x2, h1)));
#endif
	_mm256_store_pd(&q[nb*ldq],q1);
	_mm256_store_pd(&q[(nb*ldq)+4],q2);

	h1 = _mm256_broadcast_sd(&hh[nb-2]);
	h2 = _mm256_broadcast_sd(&hh[ldh+nb-1]);
	q1 = _mm256_load_pd(&q[(nb+1)*ldq]);
	q2 = _mm256_load_pd(&q[((nb+1)*ldq)+4]);
1114 1115 1116 1117 1118
#ifdef __ELPA_USE_FMA__
        q1 = _mm256_NFMA_pd(x1, h1, q1);
        q1 = _mm256_NFMA_pd(y1, h2, q1);
        q2 = _mm256_NFMA_pd(x2, h1, q2);
        q2 = _mm256_NFMA_pd(y2, h2, q2);
1119 1120 1121 1122 1123 1124 1125 1126 1127 1128
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd( _mm256_mul_pd(y1, h2) , _mm256_mul_pd(x1, h1)));
	q2 = _mm256_sub_pd(q2, _mm256_add_pd( _mm256_mul_pd(y2, h2) , _mm256_mul_pd(x2, h1)));
#endif
	_mm256_store_pd(&q[(nb+1)*ldq],q1);
	_mm256_store_pd(&q[((nb+1)*ldq)+4],q2);

	h1 = _mm256_broadcast_sd(&hh[nb-1]);
	q1 = _mm256_load_pd(&q[(nb+2)*ldq]);
	q2 = _mm256_load_pd(&q[((nb+2)*ldq)+4]);
1129 1130 1131
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_NFMA_pd(x1, h1, q1);
	q2 = _mm256_NFMA_pd(x2, h1, q2);
1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165
#else
	q1 = _mm256_sub_pd(q1, _mm256_mul_pd(x1, h1));
	q2 = _mm256_sub_pd(q2, _mm256_mul_pd(x2, h1));
#endif
	_mm256_store_pd(&q[(nb+2)*ldq],q1);
	_mm256_store_pd(&q[((nb+2)*ldq)+4],q2);
}

/**
 * Unrolled kernel that computes
 * 4 rows of Q simultaneously, a
 * matrix vector product with two householder
 * vectors + a rank 1 update is performed
 */
__forceinline void hh_trafo_kernel_4_AVX_4hv(double* q, double* hh, int nb, int ldq, int ldh, double s_1_2, double s_1_3, double s_2_3, double s_1_4, double s_2_4, double s_3_4)
{
	/////////////////////////////////////////////////////
	// Matrix Vector Multiplication, Q [4 x nb+3] * hh
	// hh contains four householder vectors
	/////////////////////////////////////////////////////
	int i;

	__m256d a1_1 = _mm256_load_pd(&q[ldq*3]);
	__m256d a2_1 = _mm256_load_pd(&q[ldq*2]);
	__m256d a3_1 = _mm256_load_pd(&q[ldq]);
	__m256d a4_1 = _mm256_load_pd(&q[0]);

	__m256d h_2_1 = _mm256_broadcast_sd(&hh[ldh+1]);
	__m256d h_3_2 = _mm256_broadcast_sd(&hh[(ldh*2)+1]);
	__m256d h_3_1 = _mm256_broadcast_sd(&hh[(ldh*2)+2]);
	__m256d h_4_3 = _mm256_broadcast_sd(&hh[(ldh*3)+1]);
	__m256d h_4_2 = _mm256_broadcast_sd(&hh[(ldh*3)+2]);
	__m256d h_4_1 = _mm256_broadcast_sd(&hh[(ldh*3)+3]);

1166 1167 1168 1169 1170 1171 1172
#ifdef __ELPA_USE_FMA__
	__m256d w1 = _mm256_FMA_pd(a3_1, h_4_3, a4_1);
	w1 = _mm256_FMA_pd(a2_1, h_4_2, w1);
	w1 = _mm256_FMA_pd(a1_1, h_4_1, w1);
	__m256d z1 = _mm256_FMA_pd(a2_1, h_3_2, a3_1);
	z1 = _mm256_FMA_pd(a1_1, h_3_1, z1);
	__m256d y1 = _mm256_FMA_pd(a1_1, h_2_1, a2_1);
1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198
	__m256d x1 = a1_1;
#else
	__m256d w1 = _mm256_add_pd(a4_1, _mm256_mul_pd(a3_1, h_4_3));
	w1 = _mm256_add_pd(w1, _mm256_mul_pd(a2_1, h_4_2));
	w1 = _mm256_add_pd(w1, _mm256_mul_pd(a1_1, h_4_1));
	__m256d z1 = _mm256_add_pd(a3_1, _mm256_mul_pd(a2_1, h_3_2));
	z1 = _mm256_add_pd(z1, _mm256_mul_pd(a1_1, h_3_1));
	__m256d y1 = _mm256_add_pd(a2_1, _mm256_mul_pd(a1_1, h_2_1));
	__m256d x1 = a1_1;
#endif

	__m256d q1;

	__m256d h1;
	__m256d h2;
	__m256d h3;
	__m256d h4;

	for(i = 4; i < nb; i++)
	{
		h1 = _mm256_broadcast_sd(&hh[i-3]);
		h2 = _mm256_broadcast_sd(&hh[ldh+i-2]);
		h3 = _mm256_broadcast_sd(&hh[(ldh*2)+i-1]);
		h4 = _mm256_broadcast_sd(&hh[(ldh*3)+i]);

		q1 = _mm256_load_pd(&q[i*ldq]);
1199 1200 1201 1202 1203
#ifdef __ELPA_USE_FMA__
		x1 = _mm256_FMA_pd(q1, h1, x1);
		y1 = _mm256_FMA_pd(q1, h2, y1);
		z1 = _mm256_FMA_pd(q1, h3, z1);
		w1 = _mm256_FMA_pd(q1, h4, w1);
1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216
#else
		x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
		y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
		z1 = _mm256_add_pd(z1, _mm256_mul_pd(q1,h3));
		w1 = _mm256_add_pd(w1, _mm256_mul_pd(q1,h4));
#endif
	}

	h1 = _mm256_broadcast_sd(&hh[nb-3]);
	h2 = _mm256_broadcast_sd(&hh[ldh+nb-2]);
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+nb-1]);
	q1 = _mm256_load_pd(&q[nb*ldq]);
#ifdef _FMA4__
1217 1218 1219
	x1 = _mm256_FMA_pd(q1, h1, x1);
	y1 = _mm256_FMA_pd(q1, h2, y1);
	z1 = _mm256_FMA_pd(q1, h3, z1);
1220 1221 1222 1223 1224 1225 1226 1227 1228
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
	y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
	z1 = _mm256_add_pd(z1, _mm256_mul_pd(q1,h3));
#endif

	h1 = _mm256_broadcast_sd(&hh[nb-2]);
	h2 = _mm256_broadcast_sd(&hh[(ldh*1)+nb-1]);
	q1 = _mm256_load_pd(&q[(nb+1)*ldq]);
1229 1230 1231
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMA_pd(q1, h1, x1);
	y1 = _mm256_FMA_pd(q1, h2, y1);
1232 1233 1234 1235 1236 1237 1238
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
	y1 = _mm256_add_pd(y1, _mm256_mul_pd(q1,h2));
#endif

	h1 = _mm256_broadcast_sd(&hh[nb-1]);
	q1 = _mm256_load_pd(&q[(nb+2)*ldq]);
1239 1240
#ifdef __ELPA_USE_FMA__
	x1 = _mm256_FMA_pd(q1, h1, x1);
1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265
#else
	x1 = _mm256_add_pd(x1, _mm256_mul_pd(q1,h1));
#endif

	/////////////////////////////////////////////////////
	// Rank-1 update of Q [4 x nb+3]
	/////////////////////////////////////////////////////

	__m256d tau1 = _mm256_broadcast_sd(&hh[0]);
	__m256d tau2 = _mm256_broadcast_sd(&hh[ldh]);
	__m256d tau3 = _mm256_broadcast_sd(&hh[ldh*2]);
	__m256d tau4 = _mm256_broadcast_sd(&hh[ldh*3]);

	__m256d vs_1_2 = _mm256_broadcast_sd(&s_1_2);
	__m256d vs_1_3 = _mm256_broadcast_sd(&s_1_3);
	__m256d vs_2_3 = _mm256_broadcast_sd(&s_2_3);
	__m256d vs_1_4 = _mm256_broadcast_sd(&s_1_4);
	__m256d vs_2_4 = _mm256_broadcast_sd(&s_2_4);
	__m256d vs_3_4 = _mm256_broadcast_sd(&s_3_4);

	h1 = tau1;
	x1 = _mm256_mul_pd(x1, h1);

	h1 = tau2;
	h2 = _mm256_mul_pd(h1, vs_1_2);
1266 1267
#ifdef __ELPA_USE_FMA__
	y1 = _mm256_FMSUB_pd(y1, h1, _mm256_mul_pd(x1,h2));
1268 1269 1270 1271 1272 1273 1274
#else
	y1 = _mm256_sub_pd(_mm256_mul_pd(y1,h1), _mm256_mul_pd(x1,h2));
#endif

	h1 = tau3;
	h2 = _mm256_mul_pd(h1, vs_1_3);
	h3 = _mm256_mul_pd(h1, vs_2_3);
1275 1276
#ifdef __ELPA_USE_FMA__
	z1 = _mm256_FMSUB_pd(z1, h1, _mm256_FMA_pd(y1, h3, _mm256_mul_pd(x1,h2)));
1277 1278 1279 1280 1281 1282 1283 1284
#else
	z1 = _mm256_sub_pd(_mm256_mul_pd(z1,h1), _mm256_add_pd(_mm256_mul_pd(y1,h3), _mm256_mul_pd(x1,h2)));
#endif

	h1 = tau4;
	h2 = _mm256_mul_pd(h1, vs_1_4);
	h3 = _mm256_mul_pd(h1, vs_2_4);
	h4 = _mm256_mul_pd(h1, vs_3_4);
1285 1286
#ifdef __ELPA_USE_FMA__
	w1 = _mm256_FMSUB_pd(w1, h1, _mm256_FMA_pd(z1, h4, _mm256_FMA_pd(y1, h3, _mm256_mul_pd(x1,h2))));
1287 1288 1289 1290 1291 1292 1293 1294 1295 1296
#else
	w1 = _mm256_sub_pd(_mm256_mul_pd(w1,h1), _mm256_add_pd(_mm256_mul_pd(z1,h4), _mm256_add_pd(_mm256_mul_pd(y1,h3), _mm256_mul_pd(x1,h2))));
#endif

	q1 = _mm256_load_pd(&q[0]);
	q1 = _mm256_sub_pd(q1, w1);
	_mm256_store_pd(&q[0],q1);

	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+1]);
	q1 = _mm256_load_pd(&q[ldq]);
1297 1298
#ifdef __ELPA_USE_FMA__
	q1 = _mm256_sub_pd(q1, _mm256_FMA_pd(w1, h4, z1));
1299 1300 1301 1302 1303 1304 1305 1306
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd(z1, _mm256_mul_pd(w1, h4)));
#endif
	_mm256_store_pd(&q[ldq],q1);

	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+1]);
	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+2]);
	q1 = _mm256_load_pd(&q[ldq*2]);
1307 1308 1309 1310
#ifdef __ELPA_USE_FMA__
        q1 = _mm256_sub_pd(q1, y1);
        q1 = _mm256_NFMA_pd(z1, h3, q1);
        q1 = _mm256_NFMA_pd(w1, h4, q1);
1311 1312 1313 1314 1315 1316 1317 1318 1319
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd(y1, _mm256_add_pd(_mm256_mul_pd(z1, h3), _mm256_mul_pd(w1, h4))));
#endif
	_mm256_store_pd(&q[ldq*2],q1);

	h2 = _mm256_broadcast_sd(&hh[ldh+1]);
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+2]);
	h4 = _mm256_broadcast_sd(&hh[(ldh*3)+3]);
	q1 = _mm256_load_pd(&q[ldq*3]);
1320 1321 1322 1323 1324
#ifdef __ELPA_USE_FMA__
        q1 = _mm256_sub_pd(q1, x1);
        q1 = _mm256_NFMA_pd(y1, h2, q1);
        q1 = _mm256_NFMA_pd(z1, h3, q1);
        q1 = _mm256_NFMA_pd(w1, h4, q1);
1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337
#else
	q1 = _mm256_sub_pd(q1, _mm256_add_pd(x1, _mm256_add_pd(_mm256_mul_pd(y1, h2), _mm256_add_pd(_mm256_mul_pd(z1, h3), _mm256_mul_pd(w1, h4)))));
#endif
	_mm256_store_pd(&q[ldq*3], q1);

	for (i = 4; i < nb; i++)
	{
		h1 = _mm256_broadcast_sd(&hh[i-3]);
		h2 = _mm256_broadcast_sd(&hh[ldh+i-2]);
		h3 = _mm256_broadcast_sd(&hh[(ldh*2)+i-1]);
		h4 = _mm256_broadcast_sd(&hh[(ldh*3)+i]);

		q1 = _mm256_load_pd(&q[i*ldq]);
1338 1339 1340 1341 1342
#ifdef __ELPA_USE_FMA__
                q1 = _mm256_NFMA_pd(x1, h1, q1);
                q1 = _mm256_NFMA_pd(y1, h2, q1);
                q1 = _mm256_NFMA_pd(z1, h3, q1);
                q1 = _mm256_NFMA_pd(w1, h4, q1);
1343 1344 1345 1346 1347 1348 1349 1350 1351 1352
#else
		q1 = _mm256_sub_pd(q1, _mm256_add_pd(_mm256_add_pd(_mm256_mul_pd(w1, h4), _mm256_mul_pd(z1, h3)), _mm256_add_pd(_mm256_mul_pd(x1,h1), _mm256_mul_pd(y1, h2))));
#endif
		_mm256_store_pd(&q[i*ldq],q1);
	}

	h1 = _mm256_broadcast_sd(&hh[nb-3]);
	h2 = _mm256_broadcast_sd(&hh[ldh+nb-2]);
	h3 = _mm256_broadcast_sd(&hh[(ldh*2)+nb-1]);
	q1 = _mm256_load_pd(&q[nb*ldq]);
1353 1354 1355 1356
#ifdef __ELPA_USE_FMA__
        q1 = _mm256_NFMA_pd(x1, h1, q1);
        q1 = _mm256_NFMA_pd(y1, h2, q1);
        q1 = _mm256_NFMA_pd(z1, h3, q1);
1357 1358 1359 1360 1361 1362 1363