-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathSSEspmv.cpp
1294 lines (1066 loc) · 64.1 KB
/
SSEspmv.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
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
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
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
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
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
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
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
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
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
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
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
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
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
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <iostream>
#include <algorithm>
#include <iterator>
#include <mmintrin.h> // MMX
#include <xmmintrin.h> // SSE
#include <emmintrin.h> // SSE 2
#include <pmmintrin.h> // SSE 3
#ifndef AMD
#include <tmmintrin.h> // SSSE 3
#include <smmintrin.h> // SSE 4.1
#include <nmmintrin.h> // SSE 4.2
#include <wmmintrin.h> // SSE ?? (AES)
#endif
#ifdef ICC
#include <nmmintrin.h>
#else
#include <ammintrin.h> // SSE4A (amd's popcount)
#endif
using namespace std;
const uint64_t masktable64[64] = {0x8000000000000000, 0x4000000000000000, 0x2000000000000000, 0x1000000000000000,
0x0800000000000000, 0x0400000000000000, 0x0200000000000000, 0x0100000000000000,
0x0080000000000000, 0x0040000000000000, 0x0020000000000000, 0x0010000000000000,
0x0008000000000000, 0x0004000000000000, 0x0002000000000000, 0x0001000000000000,
0x0000800000000000, 0x0000400000000000, 0x0000200000000000, 0x0000100000000000,
0x0000080000000000, 0x0000040000000000, 0x0000020000000000, 0x0000010000000000,
0x0000008000000000, 0x0000004000000000, 0x0000002000000000, 0x0000001000000000,
0x0000000800000000, 0x0000000400000000, 0x0000000200000000, 0x0000000100000000,
0x0000000080000000, 0x0000000040000000, 0x0000000020000000, 0x0000000010000000,
0x0000000008000000, 0x0000000004000000, 0x0000000002000000, 0x0000000001000000,
0x0000000000800000, 0x0000000000400000, 0x0000000000200000, 0x0000000000100000,
0x0000000000080000, 0x0000000000040000, 0x0000000000020000, 0x0000000000010000,
0x0000000000008000, 0x0000000000004000, 0x0000000000002000, 0x0000000000001000,
0x0000000000000800, 0x0000000000000400, 0x0000000000000200, 0x0000000000000100,
0x0000000000000080, 0x0000000000000040, 0x0000000000000020, 0x0000000000000010,
0x0000000000000008, 0x0000000000000004, 0x0000000000000002, 0x0000000000000001 };
const unsigned short masktable16[16] = {0x8000, 0x4000, 0x2000, 0x1000, 0x0800, 0x0400, 0x0200, 0x0100,
0x0080, 0x0040, 0x0020, 0x0010, 0x0008, 0x0004, 0x0002, 0x0001 };
//---------------------------------------
// Type Definitions
//---------------------------------------
typedef signed char ssp_s8;
typedef unsigned char ssp_u8;
typedef signed short ssp_s16;
typedef unsigned short ssp_u16;
typedef signed int ssp_s32;
typedef unsigned int ssp_u32;
typedef float ssp_f32;
typedef double ssp_f64;
typedef signed long long ssp_s64;
typedef unsigned long long ssp_u64;
typedef union
{
__m128 f;
__m128d d;
__m128i i;
__m64 m64[ 2];
ssp_u64 u64[ 2];
ssp_s64 s64[ 2];
ssp_f64 f64[ 2];
ssp_u32 u32[ 4];
ssp_s32 s32[ 4];
ssp_f32 f32[ 4];
ssp_u16 u16[ 8];
ssp_s16 s16[ 8];
ssp_u8 u8 [16];
ssp_s8 s8 [16];
} ssp_m128;
/**
* \SSE4_1{SSE2,_mm_blendv_pd}
* ISSUE: Do not short-circuit, i.e. loads 'a' regardless of the mask value
* Question: Does the original blendv_pd (in SSE4.1) short-circuit?
*/
inline __m128d ssp_blendv_pd_SSE2( __m128d a, __m128d b, __m128d mask )
{
ssp_m128 A, B, Mask;
A.d = a;
B.d = b;
Mask.d = mask;
// _MM_SHUFFLE(z,y,x,w) does not select anything, this macro just creates a mask
// expands to the following value: (z<<6) | (y<<4) | (x<<2) | w
Mask.i = _mm_shuffle_epi32( Mask.i, _MM_SHUFFLE(3, 3, 1, 1) );
Mask.i = _mm_srai_epi32 ( Mask.i, 31 );
B.i = _mm_and_si128( B.i, Mask.i );
A.i = _mm_andnot_si128( Mask.i, A.i );
A.i = _mm_or_si128( A.i, B.i );
return A.d;
}
#ifdef AMD
#define _mm_blendv_pd ssp_blendv_pd_SSE2
#endif
#ifdef ICC
#define __builtin_popcountll _mm_popcnt_u64
#define __builtin_popcount _mm_popcnt_u32
#endif
// 16-bit reversal table
const unsigned char BitReverseTable64[] =
{
0x0, 0x20, 0x10, 0x30, 0x8, 0x28, 0x18, 0x38,
0x4, 0x24, 0x14, 0x34, 0xc, 0x2c, 0x1c, 0x3c,
0x2, 0x22, 0x12, 0x32, 0xa, 0x2a, 0x1a, 0x3a,
0x6, 0x26, 0x16, 0x36, 0xe, 0x2e, 0x1e, 0x3e,
0x1, 0x21, 0x11, 0x31, 0x9, 0x29, 0x19, 0x39,
0x5, 0x25, 0x15, 0x35, 0xd, 0x2d, 0x1d, 0x3d,
0x3, 0x23, 0x13, 0x33, 0xb, 0x2b, 0x1b, 0x3b,
0x7, 0x27, 0x17, 0x37, 0xf, 0x2f, 0x1f, 0x3f
};
// reverse 16-bit value, 6 bits at time
unsigned short BitReverse(unsigned short v)
{
unsigned short c = (BitReverseTable64[v & 0x3f] << 10) |
(BitReverseTable64[(v >> 6) & 0x3f] << 4) |
(BitReverseTable64[(v >> 12) & 0x0f] >> 2);
return c;
}
inline void atomicallyIncrementDouble(volatile double *target, const double by){
asm volatile(
"movq %0, %%rax \n\t" // rax = *(%0)
"xorpd %%xmm0, %%xmm0 \n\t" // xmm0 = [0.0,0.0]
"movsd %1, %%xmm0\n\t" // xmm0[lo] = *(%1)
"1:\n\t"
// rax (containing *target) was last set at startup or by a failed cmpxchg
"movq %%rax, %%xmm1\n\t" // xmm1[lo] = rax
"addsd %%xmm0, %%xmm1\n\t" // xmm1 = xmm0 + xmm1 = by + xmm1
"movq %%xmm1, %%r8 \n\t" // r8 = xmm1[lo]
"lock cmpxchgq %%r8, %0\n\t" // if(*(%0)==rax){ZF=1;*(%0)=r8}else{ZF=0;rax=*(%0);}
"jnz 1b\n\t" // jump back if failed (ZF=0)
: "=m"(*target) // outputs
: "m"(by) // inputs
: "cc", "memory", "%rax", "%r8", "%xmm0", "%xmm1" // clobbered
);
return;
}
void symcsr(const double * __restrict V, const uint64_t * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, const double * __restrict XT, double * Y, double * YT, unsigned lowmask, unsigned nlowbits)
{
static const size_t NMortonRows64[] =
{
0, 1, 0, 1, 2, 3, 2, 3, 0, 1, 0, 1, 2, 3, 2, 3,
4, 5, 4, 5, 6, 7, 6, 7, 4, 5, 4, 5, 6, 7, 6, 7,
0, 1, 0, 1, 2, 3, 2, 3, 0, 1, 0, 1, 2, 3, 2, 3,
4, 5, 4, 5, 6, 7, 6, 7, 4, 5, 4, 5, 6, 7, 6, 7
};
static const size_t NMortonCols64[] =
{
0, 0, 1, 1, 0, 0, 1, 1, 2, 2, 3, 3, 2, 2, 3, 3,
0, 0, 1, 1, 0, 0, 1, 1, 2, 2, 3, 3, 2, 2, 3, 3,
4, 4, 5, 5, 4, 4, 5, 5, 6, 6, 7, 7, 6, 6, 7, 7,
4, 4, 5, 5, 4, 4, 5, 5, 6, 6, 7, 7, 6, 6, 7, 7
};
for(unsigned i=0; i<nrb;++i)
{
const unsigned Ci = bot[i] & lowmask;
const unsigned Ri = (bot[i] >> nlowbits) & lowmask;
uint64_t mask = M[i];
for(size_t j=0; j<64; ++j)
{
if(mask & masktable64[j])
{ atomicallyIncrementDouble(&Y[Ri+NMortonRows64[j]], (*V) * X[Ci+NMortonCols64[j]]);
atomicallyIncrementDouble(&YT[Ci+NMortonCols64[j]], (*V) * XT[Ri+NMortonRows64[j]]);
++V;
}
}
}
}
void symcsr(const double * __restrict V, const unsigned short * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, const double * __restrict XT, double * Y, double * YT, unsigned lowmask, unsigned nlowbits)
{
static const size_t NMortonRows16[] = { 0, 1, 0, 1, 2, 3, 2, 3, 0, 1, 0, 1, 2, 3, 2, 3 };
static const size_t NMortonCols16[] = { 0, 0, 1, 1, 0, 0, 1, 1, 2, 2, 3, 3, 2, 2, 3, 3 };
for(unsigned i=0; i<nrb;++i)
{
const unsigned Ci = bot[i] & lowmask;
const unsigned Ri = (bot[i] >> nlowbits) & lowmask;
unsigned short mask = M[i];
for(size_t j=0; j<16; ++j)
{
if(mask & masktable16[j])
{ atomicallyIncrementDouble(&Y[Ri+NMortonRows16[j]], (*V) * X[Ci+NMortonCols16[j]]);
atomicallyIncrementDouble(&YT[Ci+NMortonCols16[j]], (*V) * XT[Ri+NMortonRows16[j]]);
++V;
}
}
}
}
void symcsr(const double * __restrict V, const unsigned char * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, const double * __restrict XT, double * Y, double * YT, unsigned lowmask, unsigned nlowbits)
{
for(unsigned i=0; i<nrb;++i)
{
const unsigned Ci = bot[i] & lowmask;
const unsigned Ri = (bot[i] >> nlowbits) & lowmask;
unsigned char mask = M[i];
if(mask & 0x8)
{
atomicallyIncrementDouble(&Y[Ri+0], (*V) * X[Ci+0]);
atomicallyIncrementDouble(&YT[Ci+0], (*V) * XT[Ri+0]);
V++;
}
if(mask & 0x4)
{
atomicallyIncrementDouble(&Y[Ri+1], (*V) * X[Ci+0]);
atomicallyIncrementDouble(&YT[Ci+0], (*V) * XT[Ri+1]);
V++;
}
if(mask & 0x2)
{
atomicallyIncrementDouble(&Y[Ri+0], (*V) * X[Ci+1]);
atomicallyIncrementDouble(&YT[Ci+1], (*V) * XT[Ri+0]);
V++;
}
if(mask & 0x1)
{
atomicallyIncrementDouble(&Y[Ri+1], (*V) * X[Ci+1]);
atomicallyIncrementDouble(&YT[Ci+1], (*V) * XT[Ri+1]);
V++;
}
}
}
/**
* Symmetric SpMV inner kernel using bitmasked register blocks
* 2-by-2 potentially diagonal case (X == XT and Y == YT)
* We can still use the __restrict keyword because we only use one alias for both X and XT
**/
void SSEsym(const double * __restrict V, const unsigned char * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, double * __restrict Y, unsigned lowmask, unsigned nlbits)
{
const double * __restrict _V = V-1;
// use popcnt to index into nonzero stream
// use blendv where 1 = zero
for(unsigned ind=0;ind<nrb;++ind)
{
const unsigned Ci = bot[ind] & lowmask;
const unsigned Ri = (bot[ind] >> nlbits) & lowmask;
const uint64_t m64 = (uint64_t) M[ind]; // upcast to 64 bit, fill-in zeros from left
const uint64_t Zi = ((~m64) << 60); // a 1 denotes a zero
const uint64_t Zil = Zi << 1;
#ifdef AMD
__m128i Z01QW = _mm_unpacklo_epi64 (_mm_loadl_epi64((__m128i*)&Zi), _mm_loadl_epi64((__m128i*)&Zil));
#else
__m128i Z01QW = _mm_insert_epi64(_mm_loadl_epi64((__m128i*)&Zi),Zil,1);
#endif
__m128i Z23QW = _mm_slli_epi64(Z01QW, 2);
__m128d Y01QW = _mm_loadu_pd(&Y[Ri]);
//--------------------------------------------------------------------------
__m128d X00QW = _mm_loaddup_pd(&X[0+Ci]); // load and duplicate a double into 128-bit registers.
__m128d X11QW = _mm_loaddup_pd(&X[1+Ci]);
__m128d X01QW = _mm_loadu_pd(&X[Ri]); // the transpose of X aliases X itself
// {0,2} 02 {0,1}
// {1,3} <- 13 {2,3}
__m128d A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0x8)])), _mm_setzero_pd(),(__m128d)Z01QW); // ERROR here ! [invalid read of _V, debug with sym matrix]
__m128d A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xE)])), _mm_setzero_pd(),(__m128d)Z23QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW,A01QW),Y01QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW,A23QW),Y01QW);
__m128d Y00QW = _mm_mul_pd(X01QW, A01QW);
__m128d Y11QW = _mm_mul_pd(X01QW, A23QW);
//--------------------------------------------------------------------------
_V += __builtin_popcount(M[ind] & 0x0F);
//--------------------------------------------------------------------------
ssp_m128 yt0, yt1;
yt0.d = Y00QW;
yt1.d = Y11QW;
_mm_store_pd(&Y[Ri],Y01QW);
// The additional Y_T updates should come after we stored Y[Ri] back, otherwise they will be lost
Y[Ci+0] += yt0.f64[0] + yt0.f64[1];
Y[Ci+1] += yt1.f64[0] + yt1.f64[1];
}
}
/**
* Symmetric SpMV inner kernel using bitmasked register blocks
* 2-by-2 general case (X != XT and Y != YT)
* assumes strict-aliasing on X and Y
**/
void SSEsym(const double * __restrict V, const unsigned char * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, const double * __restrict XT, double * Y, double * YT, unsigned lowmask, unsigned nlbits)
{
const double * __restrict _V = V-1;
// use popcnt to index into nonzero stream
// use blendv where 1 = zero
for(unsigned ind=0;ind<nrb;++ind)
{
const unsigned Ci = bot[ind] & lowmask;
const unsigned Ri = (bot[ind] >> nlbits) & lowmask;
const uint64_t m64 = (uint64_t) M[ind]; // upcast to 64 bit, fill-in zeros from left
const uint64_t Zi = ((~m64) << 60); // a 1 denotes a zero
const uint64_t Zil = Zi << 1;
#ifdef AMD
__m128i Z01QW = _mm_unpacklo_epi64 (_mm_loadl_epi64((__m128i*)&Zi), _mm_loadl_epi64((__m128i*)&Zil));
#else
__m128i Z01QW = _mm_insert_epi64(_mm_loadl_epi64((__m128i*)&Zi),Zil,1);
#endif
__m128i Z23QW = _mm_slli_epi64(Z01QW, 2);
__m128d Y01QW = _mm_loadu_pd(&Y[Ri]);
//--------------------------------------------------------------------------
__m128d X00QW = _mm_loaddup_pd(&X[0+Ci]); // load and duplicate a double into 128-bit registers.
__m128d X11QW = _mm_loaddup_pd(&X[1+Ci]);
__m128d X01QW = _mm_loadu_pd(&XT[Ri]); // use the transpose of X
// {0,2} 02 {0,1}
// {1,3} <- 13 {2,3}
__m128d A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0x8)])), _mm_setzero_pd(),(__m128d)Z01QW);
__m128d A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xE)])), _mm_setzero_pd(),(__m128d)Z23QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW,A01QW),Y01QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW,A23QW),Y01QW);
__m128d YT0QW = _mm_mul_pd(X01QW, A01QW);
__m128d YT1QW = _mm_mul_pd(X01QW, A23QW);
//--------------------------------------------------------------------------
_V += __builtin_popcount(M[ind] & 0x0F);
//--------------------------------------------------------------------------
ssp_m128 yt0, yt1;
yt0.d = YT0QW;
yt1.d = YT1QW;
YT[Ci+0] += yt0.f64[0] + yt0.f64[1];
YT[Ci+1] += yt1.f64[0] + yt1.f64[1];
_mm_store_pd(&Y[Ri],Y01QW);
}
}
/**
* SpMV (usually used as a subroutine) using bitmasked register blocks
* This version works only with double values, unsigned indices, and 2x2 register blocks
* @param[in] nbr number of register blocks for this compressed sparse block only
* @param[in] bot the local part of the bottom array, i.e. {lower row bits}.{higher row bits}
* \attention SSEspmv should only be called within a single compressed sparse block and
* X and Y should already be partially indexed by the higher order bits
* We don't need any template specialization based on the register block size
* because for different block sizes, M's type differs, hence creating overloaded definitions
**/
void SSEspmv(const double * __restrict V, const unsigned char * __restrict M, const unsigned * __restrict bot, const unsigned nrb, const double * __restrict X, double * Y, unsigned lcmask, unsigned lrmask, unsigned clbits)
{
const double * __restrict _V = V-1;
// use popcnt to index into nonzero stream
// use blendv where 1 = zero
for(unsigned ind=0;ind<nrb;++ind)
{
const unsigned Ci = bot[ind] & lcmask;
const unsigned Ri = (bot[ind] >> clbits) & lrmask;
const uint64_t m64 = (uint64_t) M[ind]; // upcast to 64 bit, fill-in zeros from left
const uint64_t Zi = ((~m64) << 60); // a 1 denotes a zero
const uint64_t Zil = Zi << 1;
#ifdef AMD
__m128i Z01QW = _mm_unpacklo_epi64 (_mm_loadl_epi64((__m128i*)&Zi), _mm_loadl_epi64((__m128i*)&Zil));
#else
__m128i Z01QW = _mm_insert_epi64(_mm_loadl_epi64((__m128i*)&Zi),Zil,1);
#endif
__m128i Z23QW = _mm_slli_epi64(Z01QW, 2);
__m128d Y01QW = _mm_loadu_pd(&Y[Ri]);
//--------------------------------------------------------------------------
__m128d X00QW = _mm_loaddup_pd(&X[0+Ci]); // load and duplicate a double into 128-bit registers.
__m128d X11QW = _mm_loaddup_pd(&X[1+Ci]);
// {0,2} 02 {0,1}
// {1,3} <- 13 {2,3}
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW,_mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0x8)])),_mm_setzero_pd(),(__m128d)Z01QW)),Y01QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW,_mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xE)])),_mm_setzero_pd(),(__m128d)Z23QW)),Y01QW);
//--------------------------------------------------------------------------
_V += __builtin_popcount(M[ind] & 0x0F);
//--------------------------------------------------------------------------
_mm_store_pd(&Y[Ri],Y01QW);
}
}
// 8x8 version, using uint64_t for M
// Possibly aliasing (Y=YT or X=XT) version for the blocks right on the diagonal
void SSEsym(const double * __restrict V, const uint64_t * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, double * Y, unsigned lowmask, unsigned nlbits)
{
const double * __restrict _V = V-1;
for(unsigned ind=0;ind<nrb;++ind)
{
const unsigned Ci = bot[ind] & lowmask;
const unsigned Ri = (bot[ind] >> nlbits) & lowmask;
const uint64_t Zi = ~M[ind]; // a 1 denotes a zero
const uint64_t Zil = Zi << 1;
#ifdef AMD
__m128i Z01QW = _mm_unpacklo_epi64 (_mm_loadl_epi64((__m128i*)&Zi), _mm_loadl_epi64((__m128i*)&Zil));
#else
__m128i Z01QW = _mm_insert_epi64(_mm_loadl_epi64((__m128i*)&Zi),Zil,1);
#endif
__m128i Z23QW = _mm_slli_epi64(Z01QW, 2);
__m128i Z45QW = _mm_slli_epi64(Z01QW, 4);
__m128i Z67QW = _mm_slli_epi64(Z01QW, 6);
__m128d Y01QW = _mm_loadu_pd(&Y[Ri]);
__m128d Y23QW = _mm_loadu_pd(&Y[Ri+2]);
__m128d Y45QW = _mm_loadu_pd(&Y[Ri+4]);
__m128d Y67QW = _mm_loadu_pd(&Y[Ri+6]);
//--------------------------------------------------------------------------
__m128d X00QW = _mm_loaddup_pd(&X[0+Ci]); // load and duplicate a double into 128-bit registers.
__m128d X11QW = _mm_loaddup_pd(&X[1+Ci]);
__m128d X22QW = _mm_loaddup_pd(&X[2+Ci]);
__m128d X33QW = _mm_loaddup_pd(&X[3+Ci]);
__m128d X01QW = _mm_loadu_pd(&X[Ri]); // the transpose of X aliases X itself
__m128d X23QW = _mm_loadu_pd(&X[Ri+2]);
__m128d X45QW = _mm_loadu_pd(&X[Ri+4]);
__m128d X67QW = _mm_loadu_pd(&X[Ri+6]);
__m128d A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0x8000000000000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
__m128d A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xE000000000000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
__m128d A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xF800000000000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
__m128d A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFE00000000000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
__m128d Y00QW = _mm_mul_pd(X01QW, A01QW);
__m128d Y11QW = _mm_mul_pd(X01QW, A23QW);
Y00QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), Y00QW);
Y11QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), Y11QW);
// reuse variables for the second half of the first quadrand
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFF80000000000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFE0000000000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFF8000000000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFE000000000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
__m128d Y22QW = _mm_mul_pd(X01QW, A01QW); // the transpose (lower-triangular) updates
__m128d Y33QW = _mm_mul_pd(X01QW, A23QW);
Y22QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), Y22QW);
Y33QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), Y33QW);
// Second quadrand
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFF800000000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFE00000000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFF80000000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFE0000000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y45QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y67QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y45QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y45QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y67QW); Z67QW=_mm_slli_epi64(Z67QW,8);
Y00QW = _mm_add_pd(_mm_mul_pd(X45QW, A01QW), Y00QW); // the transpose updates
Y11QW = _mm_add_pd(_mm_mul_pd(X45QW, A23QW), Y11QW);
Y00QW = _mm_add_pd(_mm_mul_pd(X67QW, A45QW), Y00QW);
Y11QW = _mm_add_pd(_mm_mul_pd(X67QW, A67QW), Y11QW);
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFF8000000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFE000000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFF800000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFE00000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y45QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y67QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y45QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y45QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y67QW); Z67QW=_mm_slli_epi64(Z67QW,8);
Y22QW = _mm_add_pd(_mm_mul_pd(X45QW, A01QW), Y22QW); // the transpose updates
Y33QW = _mm_add_pd(_mm_mul_pd(X45QW, A23QW), Y33QW);
Y22QW = _mm_add_pd(_mm_mul_pd(X67QW, A45QW), Y22QW);
Y33QW = _mm_add_pd(_mm_mul_pd(X67QW, A67QW), Y33QW);
// Reuse registers (e.g., X00QW <- X44QW)
X00QW = _mm_loaddup_pd(&X[4+Ci]);
X11QW = _mm_loaddup_pd(&X[5+Ci]);
X22QW = _mm_loaddup_pd(&X[6+Ci]);
X33QW = _mm_loaddup_pd(&X[7+Ci]);
// Third quadrand
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFF80000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFE0000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFF8000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFE000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
__m128d Y44QW = _mm_mul_pd(X01QW, A01QW);
__m128d Y55QW = _mm_mul_pd(X01QW, A23QW);
Y44QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), Y44QW);
Y55QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), Y55QW);
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFF800000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFE00000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFF80000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFE0000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
__m128d Y66QW = _mm_mul_pd(X01QW, A01QW);
__m128d Y77QW = _mm_mul_pd(X01QW, A23QW);
Y66QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), Y66QW);
Y77QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), Y77QW);
// Fourth quadrand
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFF8000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFE000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFF800)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFE00)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y45QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y67QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y45QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y45QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y67QW); Z67QW=_mm_slli_epi64(Z67QW,8);
Y44QW = _mm_add_pd(_mm_mul_pd(X45QW, A01QW), Y44QW); // the transpose updates
Y55QW = _mm_add_pd(_mm_mul_pd(X45QW, A23QW), Y55QW);
Y44QW = _mm_add_pd(_mm_mul_pd(X67QW, A45QW), Y44QW);
Y55QW = _mm_add_pd(_mm_mul_pd(X67QW, A67QW), Y55QW);
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFF80)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFFE0)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFFF8)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFFFE)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y45QW); // no need to shift ZxxQW
Y67QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y45QW);
Y67QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y67QW);
Y66QW = _mm_add_pd(_mm_mul_pd(X45QW, A01QW), Y66QW); // the transpose updates
Y77QW = _mm_add_pd(_mm_mul_pd(X45QW, A23QW), Y77QW);
Y66QW = _mm_add_pd(_mm_mul_pd(X67QW, A45QW), Y66QW);
Y77QW = _mm_add_pd(_mm_mul_pd(X67QW, A67QW), Y77QW);
//--------------------------------------------------------------------------
_V += __builtin_popcountll(M[ind]);
//--------------------------------------------------------------------------
_mm_store_pd(&Y[Ri],Y01QW);
_mm_store_pd(&Y[Ri+2],Y23QW);
_mm_store_pd(&Y[Ri+4],Y45QW);
_mm_store_pd(&Y[Ri+6],Y67QW);
// These mirror updates come after the stores, otherwise we lose the updates
ssp_m128 yt0, yt1, yt2, yt3,yt4,yt5,yt6,yt7;
yt0.d = Y00QW;
yt1.d = Y11QW;
yt2.d = Y22QW;
yt3.d = Y33QW;
yt4.d = Y44QW;
yt5.d = Y55QW;
yt6.d = Y66QW;
yt7.d = Y77QW;
Y[Ci+0] += yt0.f64[0] + yt0.f64[1];
Y[Ci+1] += yt1.f64[0] + yt1.f64[1];
Y[Ci+2] += yt2.f64[0] + yt2.f64[1];
Y[Ci+3] += yt3.f64[0] + yt3.f64[1];
Y[Ci+4] += yt4.f64[0] + yt4.f64[1];
Y[Ci+5] += yt5.f64[0] + yt5.f64[1];
Y[Ci+6] += yt6.f64[0] + yt6.f64[1];
Y[Ci+7] += yt7.f64[0] + yt7.f64[1];
}
}
// 8x8 version, using uint64_t for M
// No aliasing between Y and YT
void SSEsym(const double * __restrict V, const uint64_t * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, const double * __restrict XT, double * __restrict Y, double * __restrict YT, unsigned lowmask, unsigned nlbits)
{
const double * __restrict _V = V-1;
for(unsigned ind=0;ind<nrb;++ind)
{
const unsigned Ci = bot[ind] & lowmask;
const unsigned Ri = (bot[ind] >> nlbits) & lowmask;
const uint64_t Zi = ~M[ind]; // a 1 denotes a zero
const uint64_t Zil = Zi << 1;
#ifdef AMD
__m128i Z01QW = _mm_unpacklo_epi64 (_mm_loadl_epi64((__m128i*)&Zi), _mm_loadl_epi64((__m128i*)&Zil));
#else
__m128i Z01QW = _mm_insert_epi64(_mm_loadl_epi64((__m128i*)&Zi),Zil,1);
#endif
__m128i Z23QW = _mm_slli_epi64(Z01QW, 2);
__m128i Z45QW = _mm_slli_epi64(Z01QW, 4);
__m128i Z67QW = _mm_slli_epi64(Z01QW, 6);
__m128d Y01QW = _mm_loadu_pd(&Y[Ri]);
__m128d Y23QW = _mm_loadu_pd(&Y[Ri+2]);
__m128d Y45QW = _mm_loadu_pd(&Y[Ri+4]);
__m128d Y67QW = _mm_loadu_pd(&Y[Ri+6]);
//--------------------------------------------------------------------------
__m128d X00QW = _mm_loaddup_pd(&X[0+Ci]); // load and duplicate a double into 128-bit registers.
__m128d X11QW = _mm_loaddup_pd(&X[1+Ci]);
__m128d X22QW = _mm_loaddup_pd(&X[2+Ci]);
__m128d X33QW = _mm_loaddup_pd(&X[3+Ci]);
__m128d X01QW = _mm_loadu_pd(&XT[Ri]);
__m128d X23QW = _mm_loadu_pd(&XT[Ri+2]);
__m128d X45QW = _mm_loadu_pd(&XT[Ri+4]);
__m128d X67QW = _mm_loadu_pd(&XT[Ri+6]);
__m128d A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0x8000000000000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
__m128d A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xE000000000000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
__m128d A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xF800000000000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
__m128d A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFE00000000000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
__m128d YT0QW = _mm_mul_pd(X01QW, A01QW);
__m128d YT1QW = _mm_mul_pd(X01QW, A23QW);
YT0QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), YT0QW);
YT1QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), YT1QW);
// reuse variables for the second half of the first quadrand
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFF80000000000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFE0000000000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFF8000000000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFE000000000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
__m128d YT2QW = _mm_mul_pd(X01QW, A01QW); // the transpose (lower-triangular) updates
__m128d YT3QW = _mm_mul_pd(X01QW, A23QW);
YT2QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), YT2QW);
YT3QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), YT3QW);
// Second quadrand
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFF800000000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFE00000000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFF80000000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFE0000000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y45QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y67QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y45QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y45QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y67QW); Z67QW=_mm_slli_epi64(Z67QW,8);
YT0QW = _mm_add_pd(_mm_mul_pd(X45QW, A01QW), YT0QW); // the transpose updates
YT1QW = _mm_add_pd(_mm_mul_pd(X45QW, A23QW), YT1QW);
YT0QW = _mm_add_pd(_mm_mul_pd(X67QW, A45QW), YT0QW);
YT1QW = _mm_add_pd(_mm_mul_pd(X67QW, A67QW), YT1QW);
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFF8000000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFE000000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFF800000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFE00000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y45QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y67QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y45QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y45QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y67QW); Z67QW=_mm_slli_epi64(Z67QW,8);
YT2QW = _mm_add_pd(_mm_mul_pd(X45QW, A01QW), YT2QW); // the transpose updates
YT3QW = _mm_add_pd(_mm_mul_pd(X45QW, A23QW), YT3QW);
YT2QW = _mm_add_pd(_mm_mul_pd(X67QW, A45QW), YT2QW);
YT3QW = _mm_add_pd(_mm_mul_pd(X67QW, A67QW), YT3QW);
ssp_m128 yt0, yt1, yt2, yt3;
yt0.d = YT0QW;
yt1.d = YT1QW;
yt2.d = YT2QW;
yt3.d = YT3QW;
YT[Ci+0] += yt0.f64[0] + yt0.f64[1];
YT[Ci+1] += yt1.f64[0] + yt1.f64[1];
YT[Ci+2] += yt2.f64[0] + yt2.f64[1];
YT[Ci+3] += yt3.f64[0] + yt3.f64[1];
// Reuse registers (e.g., X00QW <- X44QW)
X00QW = _mm_loaddup_pd(&X[4+Ci]);
X11QW = _mm_loaddup_pd(&X[5+Ci]);
X22QW = _mm_loaddup_pd(&X[6+Ci]);
X33QW = _mm_loaddup_pd(&X[7+Ci]);
// Third quadrand
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFF80000000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFE0000000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFF8000000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFE000000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
YT0QW = _mm_mul_pd(X01QW, A01QW); // reuse Y(1:4) registers for Y(5:8)
YT1QW = _mm_mul_pd(X01QW, A23QW);
YT0QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), YT0QW);
YT1QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), YT1QW);
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFF800000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFE00000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFF80000)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFE0000)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
YT2QW = _mm_mul_pd(X01QW, A01QW);
YT3QW = _mm_mul_pd(X01QW, A23QW);
YT2QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), YT2QW);
YT3QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), YT3QW);
// Fourth quadrand
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFF8000)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFE000)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFF800)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFE00)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y45QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y67QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y45QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y45QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y67QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y67QW); Z67QW=_mm_slli_epi64(Z67QW,8);
YT0QW = _mm_add_pd(_mm_mul_pd(X45QW, A01QW), YT0QW); // the transpose updates
YT1QW = _mm_add_pd(_mm_mul_pd(X45QW, A23QW), YT1QW);
YT0QW = _mm_add_pd(_mm_mul_pd(X67QW, A45QW), YT0QW);
YT1QW = _mm_add_pd(_mm_mul_pd(X67QW, A67QW), YT1QW);
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFF80)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFFE0)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFFF8)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcountll(M[ind]&0xFFFFFFFFFFFFFFFE)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y45QW); // no need to shift ZxxQW
Y67QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y67QW);
Y45QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y45QW);
Y67QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y67QW);
YT2QW = _mm_add_pd(_mm_mul_pd(X45QW, A01QW), YT2QW); // the transpose updates
YT3QW = _mm_add_pd(_mm_mul_pd(X45QW, A23QW), YT3QW);
YT2QW = _mm_add_pd(_mm_mul_pd(X67QW, A45QW), YT2QW);
YT3QW = _mm_add_pd(_mm_mul_pd(X67QW, A67QW), YT3QW);
//--------------------------------------------------------------------------
_V += __builtin_popcountll(M[ind]);
//--------------------------------------------------------------------------
_mm_store_pd(&Y[Ri],Y01QW);
_mm_store_pd(&Y[Ri+2],Y23QW);
_mm_store_pd(&Y[Ri+4],Y45QW);
_mm_store_pd(&Y[Ri+6],Y67QW);
yt0.d = YT0QW;
yt1.d = YT1QW;
yt2.d = YT2QW;
yt3.d = YT3QW;
YT[Ci+4] += yt0.f64[0] + yt0.f64[1];
YT[Ci+5] += yt1.f64[0] + yt1.f64[1];
YT[Ci+6] += yt2.f64[0] + yt2.f64[1];
YT[Ci+7] += yt3.f64[0] + yt3.f64[1];
}
}
// Possibly aliasing (Y=YT or X=XT) version for the blocks right on the diagonal
void SSEsym(const double * __restrict V, const unsigned short * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, double * Y, unsigned lowmask, unsigned nlbits)
{
const double * __restrict _V = V-1;
for(unsigned ind=0;ind<nrb;++ind)
{
const unsigned Ci = bot[ind] & lowmask;
const unsigned Ri = (bot[ind] >> nlbits) & lowmask;
const uint64_t m64 = (uint64_t) M[ind]; // upcast to 64 bit, fill-in zeros from left
const uint64_t Zi = ((~m64) << 48); // a 1 denotes a zero
const uint64_t Zil = Zi << 1;
#ifdef AMD
__m128i Z01QW = _mm_unpacklo_epi64 (_mm_loadl_epi64((__m128i*)&Zi), _mm_loadl_epi64((__m128i*)&Zil));
#else
__m128i Z01QW = _mm_insert_epi64(_mm_loadl_epi64((__m128i*)&Zi),Zil,1);
#endif
__m128i Z23QW = _mm_slli_epi64(Z01QW, 2);
__m128i Z45QW = _mm_slli_epi64(Z01QW, 4);
__m128i Z67QW = _mm_slli_epi64(Z01QW, 6);
__m128d Y01QW = _mm_loadu_pd(&Y[Ri]);
__m128d Y23QW = _mm_loadu_pd(&Y[Ri+2]);
//--------------------------------------------------------------------------
__m128d X00QW = _mm_loaddup_pd(&X[0+Ci]); // load and duplicate a double into 128-bit registers.
__m128d X11QW = _mm_loaddup_pd(&X[1+Ci]);
__m128d X22QW = _mm_loaddup_pd(&X[2+Ci]);
__m128d X33QW = _mm_loaddup_pd(&X[3+Ci]);
__m128d X01QW = _mm_loadu_pd(&X[Ri]); // the transpose of X aliases X itself
__m128d X23QW = _mm_loadu_pd(&X[Ri+2]);
__m128d A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0x8000)])), _mm_setzero_pd(),(__m128d)Z01QW);
__m128d A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xE000)])), _mm_setzero_pd(),(__m128d)Z23QW);
__m128d A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xF800)])), _mm_setzero_pd(),(__m128d)Z45QW);
__m128d A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xFE00)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
__m128d Y00QW = _mm_mul_pd(X01QW, A01QW);
__m128d Y11QW = _mm_mul_pd(X01QW, A23QW);
Y00QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), Y00QW);
Y11QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), Y11QW);
// reuse variables for the second half of A
A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xFF80)])), _mm_setzero_pd(),(__m128d)Z01QW);
A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xFFE0)])), _mm_setzero_pd(),(__m128d)Z23QW);
A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xFFF8)])), _mm_setzero_pd(),(__m128d)Z45QW);
A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xFFFE)])), _mm_setzero_pd(),(__m128d)Z67QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X22QW, A01QW), Y01QW); // the shifts on ZxxQW are unnecessary after this point
Y23QW = _mm_add_pd(_mm_mul_pd(X22QW, A45QW), Y23QW);
Y01QW = _mm_add_pd(_mm_mul_pd(X33QW, A23QW), Y01QW);
Y23QW = _mm_add_pd(_mm_mul_pd(X33QW, A67QW), Y23QW);
__m128d Y22QW = _mm_mul_pd(X01QW, A01QW);
__m128d Y33QW = _mm_mul_pd(X01QW, A23QW);
Y22QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), Y22QW);
Y33QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), Y33QW);
//--------------------------------------------------------------------------
_V += __builtin_popcount(M[ind]);
//--------------------------------------------------------------------------
_mm_store_pd(&Y[Ri],Y01QW);
_mm_store_pd(&Y[Ri+2],Y23QW);
// These mirror updates come after the stores, otherwise we lose the updates
ssp_m128 yt0, yt1, yt2, yt3;
yt0.d = Y00QW;
yt1.d = Y11QW;
yt2.d = Y22QW;
yt3.d = Y33QW;
Y[Ci+0] += yt0.f64[0] + yt0.f64[1];
Y[Ci+1] += yt1.f64[0] + yt1.f64[1];
Y[Ci+2] += yt2.f64[0] + yt2.f64[1];
Y[Ci+3] += yt3.f64[0] + yt3.f64[1];
}
}
void SSEsym(const double * __restrict V, const unsigned short * __restrict M, const unsigned * __restrict bot, const unsigned nrb,
const double * __restrict X, const double * __restrict XT, double * Y, double * YT, unsigned lowmask, unsigned nlbits)
{
const double * __restrict _V = V-1;
// use popcnt to index into nonzero stream
// use blendv where 1 = zero
for(unsigned ind=0;ind<nrb;++ind)
{
const unsigned Ci = bot[ind] & lowmask;
const unsigned Ri = (bot[ind] >> nlbits) & lowmask;
const uint64_t m64 = (uint64_t) M[ind]; // upcast to 64 bit, fill-in zeros from left
const uint64_t Zi = ((~m64) << 48); // a 1 denotes a zero
const uint64_t Zil = Zi << 1;
#ifdef AMD
__m128i Z01QW = _mm_unpacklo_epi64 (_mm_loadl_epi64((__m128i*)&Zi), _mm_loadl_epi64((__m128i*)&Zil));
#else
__m128i Z01QW = _mm_insert_epi64(_mm_loadl_epi64((__m128i*)&Zi),Zil,1);
#endif
__m128i Z23QW = _mm_slli_epi64(Z01QW, 2);
__m128i Z45QW = _mm_slli_epi64(Z01QW, 4);
__m128i Z67QW = _mm_slli_epi64(Z01QW, 6);
__m128d Y01QW = _mm_loadu_pd(&Y[Ri]);
__m128d Y23QW = _mm_loadu_pd(&Y[Ri+2]);
//--------------------------------------------------------------------------
__m128d X00QW = _mm_loaddup_pd(&X[0+Ci]); // load and duplicate a double into 128-bit registers.
__m128d X11QW = _mm_loaddup_pd(&X[1+Ci]);
__m128d X22QW = _mm_loaddup_pd(&X[2+Ci]);
__m128d X33QW = _mm_loaddup_pd(&X[3+Ci]);
__m128d X01QW = _mm_loadu_pd(&XT[Ri]);
__m128d X23QW = _mm_loadu_pd(&XT[Ri+2]);
__m128d A01QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0x8000)])), _mm_setzero_pd(),(__m128d)Z01QW);
__m128d A23QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xE000)])), _mm_setzero_pd(),(__m128d)Z23QW);
__m128d A45QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xF800)])), _mm_setzero_pd(),(__m128d)Z45QW);
__m128d A67QW = _mm_blendv_pd((__m128d)_mm_loadu_ps((float*)&(_V[__builtin_popcount(M[ind]&0xFE00)])), _mm_setzero_pd(),(__m128d)Z67QW);
// Operations rescheduled for maximum parallelism (they follow a 1-3-2-4 order)
// {0,2} 02** {0,1}
// {1,3} <- 13** {2,3}
// * **** *
// * **** *
// * **** {4,5}
// * <- **** {6,7}
// {4,6} 46** *
// {5,7} 57** *
Y01QW = _mm_add_pd(_mm_mul_pd(X00QW, A01QW), Y01QW); Z01QW=_mm_slli_epi64(Z01QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X00QW, A45QW), Y23QW); Z45QW=_mm_slli_epi64(Z45QW,8);
Y01QW = _mm_add_pd(_mm_mul_pd(X11QW, A23QW), Y01QW); Z23QW=_mm_slli_epi64(Z23QW,8);
Y23QW = _mm_add_pd(_mm_mul_pd(X11QW, A67QW), Y23QW); Z67QW=_mm_slli_epi64(Z67QW,8);
__m128d YT0QW = _mm_mul_pd(X01QW, A01QW);
__m128d YT1QW = _mm_mul_pd(X01QW, A23QW);
YT0QW = _mm_add_pd(_mm_mul_pd(X23QW, A45QW), YT0QW);
YT1QW = _mm_add_pd(_mm_mul_pd(X23QW, A67QW), YT1QW);
// write YT back (Safe since we know that Y is not an alias to YT)
ssp_m128 yt0, yt1;
yt0.d = YT0QW;
yt1.d = YT1QW;
YT[Ci+0] += yt0.f64[0] + yt0.f64[1];
YT[Ci+1] += yt1.f64[0] + yt1.f64[1];