-
Notifications
You must be signed in to change notification settings - Fork 36
/
Copy pathRCGpuCore.cu
917 lines (805 loc) · 29.2 KB
/
RCGpuCore.cu
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
// This file is a part of RCKangaroo software
// (c) 2024, RetiredCoder (RC)
// License: GPLv3, see "LICENSE.TXT" file
// https://github.com/RetiredC
#include "defs.h"
#include "RCGpuUtils.h"
//imp2 table points for KernelA
__device__ __constant__ u64 jmp2_table[8 * JMP_CNT];
#define BLOCK_CNT gridDim.x
#define BLOCK_X blockIdx.x
#define THREAD_X threadIdx.x
//coalescing
#define LOAD_VAL_256(dst, ptr, group) { *((int4*)&(dst)[0]) = *((int4*)&(ptr)[BLOCK_SIZE * 4 * BLOCK_CNT * (group)]); *((int4*)&(dst)[2]) = *((int4*)&(ptr)[2 * BLOCK_SIZE + BLOCK_SIZE * 4 * BLOCK_CNT * (group)]); }
#define SAVE_VAL_256(ptr, src, group) { *((int4*)&(ptr)[BLOCK_SIZE * 4 * BLOCK_CNT * (group)]) = *((int4*)&(src)[0]); *((int4*)&(ptr)[2 * BLOCK_SIZE + BLOCK_SIZE * 4 * BLOCK_CNT * (group)]) = *((int4*)&(src)[2]); }
extern __shared__ u64 LDS[];
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
#ifndef OLD_GPU
//this kernel performs main jumps
extern "C" __launch_bounds__(BLOCK_SIZE, 1)
__global__ void KernelA(const TKparams Kparams)
{
u64* L2x = Kparams.L2 + 2 * THREAD_X + 4 * BLOCK_SIZE * BLOCK_X;
u64* L2y = L2x + 4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE;
u64* L2s = L2y + 4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE;
//list of distances of performed jumps for KernelB
int4* jlist = (int4*)(Kparams.JumpsList + (u64)BLOCK_X * STEP_CNT * PNT_GROUP_CNT * BLOCK_SIZE / 4);
jlist += (THREAD_X / 32) * 32 * PNT_GROUP_CNT / 8;
//list of last visited points for KernelC
u64* x_last0 = Kparams.LastPnts + 2 * THREAD_X + 4 * BLOCK_SIZE * BLOCK_X;
u64* y_last0 = x_last0 + 4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE;
u64* jmp1_table = LDS; //32KB
u16* lds_jlist = (u16*)&LDS[8 * JMP_CNT]; //4KB, must be aligned 16bytes
int i = THREAD_X;
while (i < JMP_CNT)
{
*(int4*)&jmp1_table[8 * i + 0] = *(int4*)&Kparams.Jumps1[12 * i + 0];
*(int4*)&jmp1_table[8 * i + 2] = *(int4*)&Kparams.Jumps1[12 * i + 2];
*(int4*)&jmp1_table[8 * i + 4] = *(int4*)&Kparams.Jumps1[12 * i + 4];
*(int4*)&jmp1_table[8 * i + 6] = *(int4*)&Kparams.Jumps1[12 * i + 6];
i += BLOCK_SIZE;
}
__syncthreads();
__align__(16) u64 x[4], y[4], tmp[4], tmp2[4];
u64 dp_mask64 = ~((1ull << (64 - Kparams.DP)) - 1);
u16 jmp_ind;
//copy kangs from global to L2
u32 kang_ind = PNT_GROUP_CNT * (THREAD_X + BLOCK_X * BLOCK_SIZE);
for (u32 group = 0; group < PNT_GROUP_CNT; group++)
{
tmp[0] = Kparams.Kangs[(kang_ind + group) * 12 + 0];
tmp[1] = Kparams.Kangs[(kang_ind + group) * 12 + 1];
tmp[2] = Kparams.Kangs[(kang_ind + group) * 12 + 2];
tmp[3] = Kparams.Kangs[(kang_ind + group) * 12 + 3];
SAVE_VAL_256(L2x, tmp, group);
tmp[0] = Kparams.Kangs[(kang_ind + group) * 12 + 4];
tmp[1] = Kparams.Kangs[(kang_ind + group) * 12 + 5];
tmp[2] = Kparams.Kangs[(kang_ind + group) * 12 + 6];
tmp[3] = Kparams.Kangs[(kang_ind + group) * 12 + 7];
SAVE_VAL_256(L2y, tmp, group);
}
u32 L1S2 = Kparams.L1S2[BLOCK_X * BLOCK_SIZE + THREAD_X];
for (int step_ind = 0; step_ind < STEP_CNT; step_ind++)
{
__align__(16) u64 inverse[5];
u64* jmp_table;
__align__(16) u64 jmp_x[4];
__align__(16) u64 jmp_y[4];
//first group
LOAD_VAL_256(x, L2x, 0);
jmp_ind = x[0] % JMP_CNT;
jmp_table = ((L1S2 >> 0) & 1) ? jmp2_table : jmp1_table;
Copy_int4_x2(jmp_x, jmp_table + 8 * jmp_ind);
SubModP(inverse, x, jmp_x);
SAVE_VAL_256(L2s, inverse, 0);
//the rest
for (int group = 1; group < PNT_GROUP_CNT; group++)
{
LOAD_VAL_256(x, L2x, group);
jmp_ind = x[0] % JMP_CNT;
jmp_table = ((L1S2 >> group) & 1) ? jmp2_table : jmp1_table;
Copy_int4_x2(jmp_x, jmp_table + 8 * jmp_ind);
SubModP(tmp, x, jmp_x);
MulModP(inverse, inverse, tmp);
SAVE_VAL_256(L2s, inverse, group);
}
InvModP((u32*)inverse);
for (int group = PNT_GROUP_CNT - 1; group >= 0; group--)
{
__align__(16) u64 x0[4];
__align__(16) u64 y0[4];
__align__(16) u64 dxs[4];
LOAD_VAL_256(x0, L2x, group);
LOAD_VAL_256(y0, L2y, group);
jmp_ind = x0[0] % JMP_CNT;
jmp_table = ((L1S2 >> group) & 1) ? jmp2_table : jmp1_table;
Copy_int4_x2(jmp_x, jmp_table + 8 * jmp_ind);
Copy_int4_x2(jmp_y, jmp_table + 8 * jmp_ind + 4);
u32 inv_flag = (u32)y0[0] & 1;
if (inv_flag)
{
jmp_ind |= INV_FLAG;
NegModP(jmp_y);
}
if (group)
{
LOAD_VAL_256(tmp, L2s, group - 1);
SubModP(tmp2, x0, jmp_x);
MulModP(dxs, tmp, inverse);
MulModP(inverse, inverse, tmp2);
}
else
Copy_u64_x4(dxs, inverse);
SubModP(tmp2, y0, jmp_y);
MulModP(tmp, tmp2, dxs);
SqrModP(tmp2, tmp);
SubModP(x, tmp2, jmp_x);
SubModP(x, x, x0);
SAVE_VAL_256(L2x, x, group);
SubModP(y, x0, x);
MulModP(y, y, tmp);
SubModP(y, y, y0);
SAVE_VAL_256(L2y, y, group);
if (((L1S2 >> group) & 1) == 0) //normal mode, check L1S2 loop
{
u32 jmp_next = x[0] % JMP_CNT;
jmp_next |= ((u32)y[0] & 1) ? 0 : INV_FLAG; //inverted
L1S2 |= (jmp_ind == jmp_next) ? (1u << group) : 0; //loop L1S2 detected
}
else
{
L1S2 &= ~(1u << group);
jmp_ind |= JMP2_FLAG;
}
if ((x[3] & dp_mask64) == 0)
{
u32 kang_ind = (THREAD_X + BLOCK_X * BLOCK_SIZE) * PNT_GROUP_CNT + group;
u32 ind = atomicAdd(Kparams.DPTable + kang_ind, 1);
ind = min(ind, DPTABLE_MAX_CNT - 1);
int4* dst = (int4*)(Kparams.DPTable + Kparams.KangCnt + (kang_ind * DPTABLE_MAX_CNT + ind) * 4);
dst[0] = ((int4*)x)[0];
jmp_ind |= DP_FLAG;
}
lds_jlist[8 * THREAD_X + (group % 8)] = jmp_ind;
if ((group % 8) == 0)
st_cs_v4_b32(&jlist[(group / 8) * 32 + (THREAD_X % 32)], *(int4*)&lds_jlist[8 * THREAD_X]); //skip L2 cache
if (step_ind + MD_LEN >= STEP_CNT) //store last kangs to be able to find loop exit point
{
int n = step_ind + MD_LEN - STEP_CNT;
u64* x_last = x_last0 + n * 2 * (4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE);
u64* y_last = y_last0 + n * 2 * (4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE);
SAVE_VAL_256(x_last, x, group);
SAVE_VAL_256(y_last, y, group);
}
}
jlist += PNT_GROUP_CNT * BLOCK_SIZE / 8;
}
Kparams.L1S2[BLOCK_X * BLOCK_SIZE + THREAD_X] = L1S2;
//copy kangs from L2 to global
kang_ind = PNT_GROUP_CNT * (THREAD_X + BLOCK_X * BLOCK_SIZE);
for (u32 group = 0; group < PNT_GROUP_CNT; group++)
{
LOAD_VAL_256(tmp, L2x, group);
Kparams.Kangs[(kang_ind + group) * 12 + 0] = tmp[0];
Kparams.Kangs[(kang_ind + group) * 12 + 1] = tmp[1];
Kparams.Kangs[(kang_ind + group) * 12 + 2] = tmp[2];
Kparams.Kangs[(kang_ind + group) * 12 + 3] = tmp[3];
LOAD_VAL_256(tmp, L2y, group);
Kparams.Kangs[(kang_ind + group) * 12 + 4] = tmp[0];
Kparams.Kangs[(kang_ind + group) * 12 + 5] = tmp[1];
Kparams.Kangs[(kang_ind + group) * 12 + 6] = tmp[2];
Kparams.Kangs[(kang_ind + group) * 12 + 7] = tmp[3];
}
}
#else
#define LOAD_VAL_256_m(dst,p,i) { *((int4*)&(dst)[0]) = *((int4*)&(p)[4 * (i)]); *((int4*)&(dst)[2]) = *((int4*)&(p)[4 * (i) + 2]); }
#define SAVE_VAL_256_m(p,src,i) { *((int4*)&(p)[4 * (i)]) = *((int4*)&(src)[0]); *((int4*)&(p)[4 * (i) + 2]) = *((int4*)&(src)[2]); }
//this kernel performs main jumps for old cards
//not good but works
extern "C" __launch_bounds__(BLOCK_SIZE, 1)
__global__ void KernelA(const TKparams Kparams)
{
__align__(16) u64 Lx[4 * PNT_GROUP_CNT];
__align__(16) u64 Ly[4 * PNT_GROUP_CNT];
__align__(16) u64 Ls[4 * PNT_GROUP_CNT / 2]; //we store only half so need only half mem
//list of distances of performed jumps for KernelB
int4* jlist = (int4*)(Kparams.JumpsList + (u64)BLOCK_X * STEP_CNT * PNT_GROUP_CNT * BLOCK_SIZE / 4);
jlist += (THREAD_X / 32) * 32 * PNT_GROUP_CNT / 8;
//list of last visited points for KernelC
u64* x_last0 = Kparams.LastPnts + 2 * THREAD_X + 4 * BLOCK_SIZE * BLOCK_X;
u64* y_last0 = x_last0 + 4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE;
u64* jmp1_table = LDS; //32KB
u16* lds_jlist = (u16*)&LDS[8 * JMP_CNT]; //8KB, must be aligned 16bytes
int i = THREAD_X;
while (i < JMP_CNT)
{
*(int4*)&jmp1_table[8 * i + 0] = *(int4*)&Kparams.Jumps1[12 * i + 0];
*(int4*)&jmp1_table[8 * i + 2] = *(int4*)&Kparams.Jumps1[12 * i + 2];
*(int4*)&jmp1_table[8 * i + 4] = *(int4*)&Kparams.Jumps1[12 * i + 4];
*(int4*)&jmp1_table[8 * i + 6] = *(int4*)&Kparams.Jumps1[12 * i + 6];
i += BLOCK_SIZE;
}
__syncthreads();
__align__(16) u64 inverse[5];
__align__(16) u64 x[4], y[4], tmp[4], tmp2[4];
u64 dp_mask64 = ~((1ull << (64 - Kparams.DP)) - 1);
u16 jmp_ind;
//copy kangs from global to local
u32 kang_ind = PNT_GROUP_CNT * (THREAD_X + BLOCK_X * BLOCK_SIZE);
for (u32 group = 0; group < PNT_GROUP_CNT; group++)
{
tmp[0] = Kparams.Kangs[(kang_ind + group) * 12 + 0];
tmp[1] = Kparams.Kangs[(kang_ind + group) * 12 + 1];
tmp[2] = Kparams.Kangs[(kang_ind + group) * 12 + 2];
tmp[3] = Kparams.Kangs[(kang_ind + group) * 12 + 3];
SAVE_VAL_256_m(Lx, tmp, group);
tmp[0] = Kparams.Kangs[(kang_ind + group) * 12 + 4];
tmp[1] = Kparams.Kangs[(kang_ind + group) * 12 + 5];
tmp[2] = Kparams.Kangs[(kang_ind + group) * 12 + 6];
tmp[3] = Kparams.Kangs[(kang_ind + group) * 12 + 7];
SAVE_VAL_256_m(Ly, tmp, group);
}
u64 L1S2 = ((u64*)Kparams.L1S2)[BLOCK_X * BLOCK_SIZE + THREAD_X];
u64* jmp_table;
__align__(16) u64 jmp_x[4];
__align__(16) u64 jmp_y[4];
//preparations (first calc for inv)
for (int group = 0; group < PNT_GROUP_CNT; group++)
{
LOAD_VAL_256_m(x, Lx, group);
jmp_ind = x[0] % JMP_CNT;
jmp_table = ((L1S2 >> group) & 1) ? jmp2_table : jmp1_table;
Copy_int4_x2(jmp_x, jmp_table + 8 * jmp_ind);
SubModP(tmp, x, jmp_x);
if (group == 0)
{
Copy_u64_x4(inverse, tmp);
SAVE_VAL_256_m(Ls, tmp, 0);
}
else
{
MulModP(inverse, inverse, tmp);
if ((group & 1) == 0)
SAVE_VAL_256_m(Ls, inverse, group / 2);
}
}
//main loop
int g_beg = PNT_GROUP_CNT - 1; //start val
int g_end = -1; //first val after range
int g_inc = -1;
int s_mask = 1;
int jlast_add = 0;
__align__(16) u64 t_cache[4], x0_cache[4], jmpx_cached[4];
t_cache[0] = t_cache[1] = t_cache[2] = t_cache[3] = 0;
x0_cache[0] = x0_cache[1] = x0_cache[2] = x0_cache[3] = 0;
for (int step_ind = 0; step_ind < STEP_CNT; step_ind++)
{
__align__(16) u64 next_inv[4];
InvModP((u32*)inverse);
int group = g_beg;
bool cached = false;
while (group != g_end)
{
__align__(16) u64 dx[4], x0[4], y0[4], dx0[4];
if (cached)
{
Copy_u64_x4(x0, x0_cache);
}
else
{
LOAD_VAL_256_m(x0, Lx, group);
}
LOAD_VAL_256_m(y0, Ly, group);
jmp_ind = x0[0] % JMP_CNT;
jmp_table = ((L1S2 >> group) & 1) ? jmp2_table : jmp1_table;
if (cached)
{
Copy_u64_x4(jmp_x, jmpx_cached);
}
else
{
Copy_int4_x2(jmp_x, jmp_table + 8 * jmp_ind);
}
Copy_int4_x2(jmp_y, jmp_table + 8 * jmp_ind + 4);
u32 inv_flag = (u32)y0[0] & 1;
if (inv_flag)
{
jmp_ind |= INV_FLAG;
NegModP(jmp_y);
}
if (group == g_end - g_inc)
{
Copy_u64_x4(dx0, inverse);
}
else
{
if ((group & 1) == s_mask) //simple case
{
if (cached)
{
Copy_u64_x4(tmp, t_cache);
cached = false;
}
else
{
LOAD_VAL_256_m(tmp, Ls, (group + g_inc) / 2);
}
}
else //no s(-1), need to calc it
{
LOAD_VAL_256_m(t_cache, Ls, (group + g_inc + g_inc) / 2);
cached = true;
LOAD_VAL_256_m(x0_cache, Lx, group + g_inc);
u32 jmp_tmp = x0_cache[0] % JMP_CNT;
__align__(16) u64 dx2[4];
u64* jmp_table_tmp = ((L1S2 >> (group + g_inc)) & 1) ? jmp2_table : jmp1_table;
Copy_int4_x2(jmpx_cached, jmp_table_tmp + 8 * jmp_tmp);
SubModP(dx2, x0_cache, jmpx_cached);
MulModP(tmp, t_cache, dx2); //t = s(-1)
}
SubModP(dx, x0, jmp_x);
MulModP(dx0, tmp, inverse);
MulModP(inverse, inverse, dx);
}
SubModP(tmp2, y0, jmp_y);
MulModP(tmp, tmp2, dx0);
SqrModP(tmp2, tmp);
SubModP(x, tmp2, jmp_x);
SubModP(x, x, x0);
SAVE_VAL_256_m(Lx, x, group);
SubModP(y, x0, x);
MulModP(y, y, tmp);
SubModP(y, y, y0);
SAVE_VAL_256_m(Ly, y, group);
if (((L1S2 >> group) & 1) == 0) //normal mode, check L1S2 loop
{
u32 jmp_next = x[0] % JMP_CNT;
jmp_next |= ((u32)y[0] & 1) ? 0 : INV_FLAG; //inverted
L1S2 |= (jmp_ind == jmp_next) ? (1ull << group) : 0; //loop L1S2 detected
}
else
{
L1S2 &= ~(1ull << group);
jmp_ind |= JMP2_FLAG;
}
if ((x[3] & dp_mask64) == 0)
{
u32 kang_ind = (THREAD_X + BLOCK_X * BLOCK_SIZE) * PNT_GROUP_CNT + group;
u32 ind = atomicAdd(Kparams.DPTable + kang_ind, 1);
ind = min(ind, DPTABLE_MAX_CNT - 1);
int4* dst = (int4*)(Kparams.DPTable + Kparams.KangCnt + (kang_ind * DPTABLE_MAX_CNT + ind) * 4);
dst[0] = ((int4*)x)[0];
jmp_ind |= DP_FLAG;
}
lds_jlist[8 * THREAD_X + (group % 8)] = jmp_ind;
if (((group + jlast_add) % 8) == 0)
st_cs_v4_b32(&jlist[(group / 8) * 32 + (THREAD_X % 32)], *(int4*)&lds_jlist[8 * THREAD_X]); //skip L2 cache
if (step_ind + MD_LEN >= STEP_CNT) //store last kangs to be able to find loop exit point
{
int n = step_ind + MD_LEN - STEP_CNT;
u64* x_last = x_last0 + n * 2 * (4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE);
u64* y_last = y_last0 + n * 2 * (4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE);
SAVE_VAL_256(x_last, x, group);
SAVE_VAL_256(y_last, y, group);
}
//preps to calc next inv
jmp_ind = x[0] % JMP_CNT;
jmp_table = ((L1S2 >> group) & 1) ? jmp2_table : jmp1_table;
Copy_int4_x2(jmp_x, jmp_table + 8 * jmp_ind);
SubModP(dx, x, jmp_x);
if (group == g_beg)
{
Copy_u64_x4(next_inv, dx);
SAVE_VAL_256_m(Ls, dx, g_beg / 2);
}
else
{
MulModP(next_inv, next_inv, dx);
if ((group & 1) == s_mask)
{
SAVE_VAL_256_m(Ls, next_inv, group / 2);
}
}
group += g_inc;
} //group
jlist += PNT_GROUP_CNT * BLOCK_SIZE / 8;
Copy_u64_x4(inverse, next_inv);
if (g_inc < 0) //invert direction
{
g_beg = 0;
g_end = PNT_GROUP_CNT;
g_inc = 1;
s_mask = 0;
jlast_add = 1;
}
else
{
g_beg = PNT_GROUP_CNT - 1;
g_end = -1;
g_inc = -1;
s_mask = 1;
jlast_add = 0;
}
}
((u64*)Kparams.L1S2)[BLOCK_X * BLOCK_SIZE + THREAD_X] = L1S2;
//copy kangs from local to global
kang_ind = PNT_GROUP_CNT * (THREAD_X + BLOCK_X * BLOCK_SIZE);
for (u32 group = 0; group < PNT_GROUP_CNT; group++)
{
LOAD_VAL_256_m(tmp, Lx, group);
Kparams.Kangs[(kang_ind + group) * 12 + 0] = tmp[0];
Kparams.Kangs[(kang_ind + group) * 12 + 1] = tmp[1];
Kparams.Kangs[(kang_ind + group) * 12 + 2] = tmp[2];
Kparams.Kangs[(kang_ind + group) * 12 + 3] = tmp[3];
LOAD_VAL_256_m(tmp, Ly, group);
Kparams.Kangs[(kang_ind + group) * 12 + 4] = tmp[0];
Kparams.Kangs[(kang_ind + group) * 12 + 5] = tmp[1];
Kparams.Kangs[(kang_ind + group) * 12 + 6] = tmp[2];
Kparams.Kangs[(kang_ind + group) * 12 + 7] = tmp[3];
}
}
#endif
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
__device__ __forceinline__ void BuildDP(const TKparams& Kparams, int kang_ind, u64* d)
{
int ind = atomicAdd(Kparams.DPTable + kang_ind, 0x10000);
ind >>= 16;
if (ind >= DPTABLE_MAX_CNT)
return;
int4 rx = *(int4*)(Kparams.DPTable + Kparams.KangCnt + (kang_ind * DPTABLE_MAX_CNT + ind) * 4);
u32 pos = atomicAdd(Kparams.DPs_out, 1);
pos = min(pos, MAX_DP_CNT - 1);
u32* DPs = Kparams.DPs_out + 4 + pos * GPU_DP_SIZE / 4;
*(int4*)&DPs[0] = rx;
*(int4*)&DPs[4] = ((int4*)d)[0];
*(u64*)&DPs[8] = d[2];
DPs[10] = 3 * kang_ind / Kparams.KangCnt; //kang type
}
__device__ __forceinline__ bool ProcessJumpDistance(u32 step_ind, u32 d_cur, u64* d, u32 kang_ind, u64* jmp1_d, u64* jmp2_d, const TKparams& Kparams, u64* table, u32* cur_ind, u8 iter)
{
u64* jmp_d = (d_cur & JMP2_FLAG) ? jmp2_d : jmp1_d;
__align__(16) u64 jmp[3];
((int4*)(jmp))[0] = ((int4*)(jmp_d + 4 * (d_cur & JMP_MASK)))[0];
jmp[2] = *(jmp_d + 4 * (d_cur & JMP_MASK) + 2);
if (d_cur & INV_FLAG)
Sub192from192(d, jmp)
else
Add192to192(d, jmp);
//check in table
int found_ind = iter + MD_LEN - 4;
while (1)
{
if (table[found_ind % MD_LEN] == d[0])
break;
found_ind -= 2;
if (table[found_ind % MD_LEN] == d[0])
break;
found_ind -= 2;
if (table[found_ind % MD_LEN] == d[0])
break;
found_ind = iter;
if (table[found_ind] == d[0])
break;
found_ind = -1;
break;
}
table[iter] = d[0];
*cur_ind = (iter + 1) % MD_LEN;
if (found_ind < 0)
{
if (d_cur & DP_FLAG)
BuildDP(Kparams, kang_ind, d);
return false;
}
u32 LoopSize = (iter + MD_LEN - found_ind) % MD_LEN;
if (!LoopSize)
LoopSize = MD_LEN;
atomicAdd(Kparams.dbg_buf + LoopSize, 1); //dbg
//calc index in LastPnts
u32 ind_LastPnts = MD_LEN - 1 - ((STEP_CNT - 1 - step_ind) % LoopSize);
u32 ind = atomicAdd(Kparams.LoopedKangs, 1);
Kparams.LoopedKangs[2 + ind] = kang_ind | (ind_LastPnts << 28);
return true;
}
#define DO_ITER(iter) {\
u32 cur_dAB = jlist[THREAD_X]; \
u16 cur_dA = cur_dAB & 0xFFFF; \
u16 cur_dB = cur_dAB >> 16; \
if (!LoopedA) \
LoopedA = ProcessJumpDistance(step_ind, cur_dA, dA, kang_ind, jmp1_d, jmp2_d, Kparams, RegsA, &cur_indA, iter); \
if (!LoopedB) \
LoopedB = ProcessJumpDistance(step_ind, cur_dB, dB, kang_ind + 1, jmp1_d, jmp2_d, Kparams, RegsB, &cur_indB, iter); \
jlist += BLOCK_SIZE * PNT_GROUP_CNT / 2; \
step_ind++; \
}
//this kernel counts distances and detects loops Size>2
//Loops Level1 statistics for JMP_CNT=512: L1S2 = 1/1024 (so one loop every 1024 jumps), L1S4 = L1S2/1024, L1S6 = L1S4/256, L1S8 = L1S6/158, L1S10 = L1S8/82. L1S12 = L1S10/50.
// For RTX4090 at 8HG/s for 24 hours and JMP_CNT=512: jumps = 691200bln, L1S2 = 682bln, L1S4 = 666mln, L1S6 = 2.6mln, L1S8 = 16.5k, L1S10 = 201. L1S12 = 4.
// I don't see any reasons to catch L1S12 because we have 786432 kangs, if we lose 4 kangs every day, we lose 1460 kangs a year which is about 0.19%.
// This degradation depends only on speed of a single kangaroo, so it's about the same for all 40xx GPUs (50xx GPUs will have +20% clock speed may be).
// Since we lose kangs gradually, for a year we lose 0.19/2 = 0.1% of speed, so you should catch L1S12 only if you are going to solve same point for decades.
// Or you can check all kangs for L1S12 on CPU once a day and restart looped kangs.
// Level2 loops are very rare and they have even size too so they will be handled by the same code. We don't know what loop level we catch so we use JmpTable3 for escaping.
extern "C" __launch_bounds__(BLOCK_SIZE, 1)
__global__ void KernelB(const TKparams Kparams)
{
u64* jmp1_d = LDS; //16KB, 192bit jumps
u64* jmp2_d = LDS + 4 * JMP_CNT; //16KB, 192bit jumps
int i = THREAD_X;
while (i < JMP_CNT)
{
//192bits but we need align 128 so use 256
jmp1_d[4 * i + 0] = Kparams.Jumps1[12 * i + 8];
jmp1_d[4 * i + 1] = Kparams.Jumps1[12 * i + 9];
jmp1_d[4 * i + 2] = Kparams.Jumps1[12 * i + 10];
jmp2_d[4 * i + 0] = Kparams.Jumps2[12 * i + 8];
jmp2_d[4 * i + 1] = Kparams.Jumps2[12 * i + 9];
jmp2_d[4 * i + 2] = Kparams.Jumps2[12 * i + 10];
i += BLOCK_SIZE;
}
u32* jlist0 = (u32*)(Kparams.JumpsList + (u64)BLOCK_X * STEP_CNT * PNT_GROUP_CNT * BLOCK_SIZE / 4);
__syncthreads();
u64 RegsA[MD_LEN], RegsB[MD_LEN];
//we process two kangs at once
for (u32 gr_ind2 = 0; gr_ind2 < PNT_GROUP_CNT/2; gr_ind2++)
{
#pragma unroll
for (int i = 0; i < MD_LEN; i++)
{
RegsA[i] = Kparams.LoopTable[MD_LEN * BLOCK_SIZE * PNT_GROUP_CNT * BLOCK_X + 2 * MD_LEN * BLOCK_SIZE * gr_ind2 + i * BLOCK_SIZE + BLOCK_X];
RegsB[i] = Kparams.LoopTable[MD_LEN * BLOCK_SIZE * PNT_GROUP_CNT * BLOCK_X + 2 * MD_LEN * BLOCK_SIZE * gr_ind2 + (i + MD_LEN) * BLOCK_SIZE + BLOCK_X];
}
u32 cur_indA = 0;
u32 cur_indB = 0;
u32* jlist = jlist0 + gr_ind2 * BLOCK_SIZE;
//calc original kang_ind
u32 tind = (THREAD_X + gr_ind2 * BLOCK_SIZE); //0..3071
u32 warp_ind = tind / (32 * PNT_GROUP_CNT / 2); // 0..7
u32 thr_ind = (tind / 4) % 32; //index in warp 0..31
u32 g8_ind = (tind % (32 * PNT_GROUP_CNT / 2)) / 128; // 0..2
u32 gr_ind = 2 * (tind % 4); // 0, 2, 4, 6
u32 kang_ind = (BLOCK_X * BLOCK_SIZE) * PNT_GROUP_CNT;
kang_ind += (32 * warp_ind + thr_ind) * PNT_GROUP_CNT + 8 * g8_ind + gr_ind;
__align__(8) u64 dA[3], dB[3];
dA[0] = Kparams.Kangs[kang_ind * 12 + 8];
dA[1] = Kparams.Kangs[kang_ind * 12 + 9];
dA[2] = Kparams.Kangs[kang_ind * 12 + 10];
dB[0] = Kparams.Kangs[(kang_ind + 1) * 12 + 8];
dB[1] = Kparams.Kangs[(kang_ind + 1) * 12 + 9];
dB[2] = Kparams.Kangs[(kang_ind + 1) * 12 + 10];
bool LoopedA = false;
bool LoopedB = false;
u32 step_ind = 0;
while (step_ind < STEP_CNT)
{
DO_ITER(0);
DO_ITER(1);
DO_ITER(2);
DO_ITER(3);
DO_ITER(4);
DO_ITER(5);
DO_ITER(6);
DO_ITER(7);
DO_ITER(8);
DO_ITER(9);
}
Kparams.Kangs[kang_ind * 12 + 8] = dA[0];
Kparams.Kangs[kang_ind * 12 + 9] = dA[1];
Kparams.Kangs[kang_ind * 12 + 10] = dA[2];
Kparams.Kangs[(kang_ind + 1) * 12 + 8] = dB[0];
Kparams.Kangs[(kang_ind + 1) * 12 + 9] = dB[1];
Kparams.Kangs[(kang_ind + 1) * 12 + 10] = dB[2];
//store so cur_ind is 0 at next loading
#pragma unroll
for (int i = 0; i < MD_LEN; i++)
{
int ind = (i + MD_LEN - cur_indA) % MD_LEN;
Kparams.LoopTable[MD_LEN * BLOCK_SIZE * PNT_GROUP_CNT * BLOCK_X + 2 * MD_LEN * BLOCK_SIZE * gr_ind2 + ind * BLOCK_SIZE + BLOCK_X] = RegsA[i];
ind = (i + MD_LEN - cur_indB) % MD_LEN;
Kparams.LoopTable[MD_LEN * BLOCK_SIZE * PNT_GROUP_CNT * BLOCK_X + 2 * MD_LEN * BLOCK_SIZE * gr_ind2 + (ind + MD_LEN) * BLOCK_SIZE + BLOCK_X] = RegsB[i];
}
}
}
//this kernel performes single jump3 for looped kangs
extern "C" __launch_bounds__(BLOCK_SIZE, 1)
__global__ void KernelC(const TKparams Kparams)
{
u64* jmp3_table = LDS; //48KB
int i = THREAD_X;
while (i < JMP_CNT)
{
*(int4*)&jmp3_table[12 * i + 0] = *(int4*)&Kparams.Jumps3[12 * i + 0];
*(int4*)&jmp3_table[12 * i + 2] = *(int4*)&Kparams.Jumps3[12 * i + 2];
*(int4*)&jmp3_table[12 * i + 4] = *(int4*)&Kparams.Jumps3[12 * i + 4];
*(int4*)&jmp3_table[12 * i + 6] = *(int4*)&Kparams.Jumps3[12 * i + 6];
*(int4*)&jmp3_table[12 * i + 8] = *(int4*)&Kparams.Jumps3[12 * i + 8];
*(int4*)&jmp3_table[12 * i + 10] = *(int4*)&Kparams.Jumps3[12 * i + 10];
i += BLOCK_SIZE;
}
__syncthreads();
while (1)
{
u32 ind = atomicAdd(Kparams.LoopedKangs + 1, 1);
if (ind >= Kparams.LoopedKangs[0])
break;
u32 kang_ind = Kparams.LoopedKangs[2 + ind] & 0x0FFFFFFF;
u32 last_ind = Kparams.LoopedKangs[2 + ind] >> 28;
__align__(16) u64 x0[4], x[4];
__align__(16) u64 y0[4], y[4];
__align__(16) u64 jmp_x[4];
__align__(16) u64 jmp_y[4];
__align__(16) u64 inverse[5];
u64 tmp[4], tmp2[4];
u64* x_last0 = Kparams.LastPnts;
u64* y_last0 = x_last0 + 4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE;
u32 block_ind = kang_ind / (BLOCK_SIZE * PNT_GROUP_CNT);
u32 thr_ind = (kang_ind - block_ind * (BLOCK_SIZE * PNT_GROUP_CNT)) / PNT_GROUP_CNT;
u32 gr_ind = (kang_ind - block_ind * (BLOCK_SIZE * PNT_GROUP_CNT) - thr_ind * PNT_GROUP_CNT);
y_last0 += 2 * thr_ind + 4 * BLOCK_SIZE * block_ind;
x_last0 += 2 * thr_ind + 4 * BLOCK_SIZE * block_ind;
u64* x_last = x_last0 + last_ind * 2 * (4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE);
u64* y_last = y_last0 + last_ind * 2 * (4 * PNT_GROUP_CNT * BLOCK_CNT * BLOCK_SIZE);
LOAD_VAL_256(x0, x_last, gr_ind);
LOAD_VAL_256(y0, y_last, gr_ind);
u32 jmp_ind = x0[0] % JMP_CNT;
Copy_int4_x2(jmp_x, jmp3_table + 12 * jmp_ind);
Copy_int4_x2(jmp_y, jmp3_table + 12 * jmp_ind + 4);
SubModP(inverse, x0, jmp_x);
InvModP((u32*)inverse);
u32 inv_flag = y0[0] & 1;
if (inv_flag)
NegModP(jmp_y);
SubModP(tmp, y0, jmp_y);
MulModP(tmp2, tmp, inverse);
SqrModP(tmp, tmp2);
SubModP(x, tmp, jmp_x);
SubModP(x, x, x0);
SubModP(y, x0, x);
MulModP(y, y, tmp2);
SubModP(y, y, y0);
//save kang
Kparams.Kangs[kang_ind * 12 + 0] = x[0];
Kparams.Kangs[kang_ind * 12 + 1] = x[1];
Kparams.Kangs[kang_ind * 12 + 2] = x[2];
Kparams.Kangs[kang_ind * 12 + 3] = x[3];
Kparams.Kangs[kang_ind * 12 + 4] = y[0];
Kparams.Kangs[kang_ind * 12 + 5] = y[1];
Kparams.Kangs[kang_ind * 12 + 6] = y[2];
Kparams.Kangs[kang_ind * 12 + 7] = y[3];
//add distance
u64 d[3];
d[0] = Kparams.Kangs[kang_ind * 12 + 8];
d[1] = Kparams.Kangs[kang_ind * 12 + 9];
d[2] = Kparams.Kangs[kang_ind * 12 + 10];
if (inv_flag)
Sub192from192(d, jmp3_table + 12 * jmp_ind + 8)
else
Add192to192(d, jmp3_table + 12 * jmp_ind + 8);
Kparams.Kangs[kang_ind * 12 + 8] = d[0];
Kparams.Kangs[kang_ind * 12 + 9] = d[1];
Kparams.Kangs[kang_ind * 12 + 10] = d[2];
#ifndef OLD_GPU
atomicAnd(&Kparams.L1S2[block_ind * BLOCK_SIZE + thr_ind], ~(1u << gr_ind));
#else
atomicAnd(&((u64*)Kparams.L1S2)[block_ind * BLOCK_SIZE + thr_ind], ~(1ull << gr_ind));
#endif
}
}
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
#define GX_0 0x59F2815B16F81798ull
#define GX_1 0x029BFCDB2DCE28D9ull
#define GX_2 0x55A06295CE870B07ull
#define GX_3 0x79BE667EF9DCBBACull
#define GY_0 0x9C47D08FFB10D4B8ull
#define GY_1 0xFD17B448A6855419ull
#define GY_2 0x5DA4FBFC0E1108A8ull
#define GY_3 0x483ADA7726A3C465ull
__device__ __forceinline__ void AddPoints(u64* res_x, u64* res_y, u64* pnt1x, u64* pnt1y, u64* pnt2x, u64* pnt2y)
{
__align__(16) u64 tmp[4], tmp2[4], lambda[4], lambda2[4];
__align__(16) u64 inverse[5];
SubModP(inverse, pnt2x, pnt1x);
InvModP((u32*)inverse);
SubModP(tmp, pnt2y, pnt1y);
MulModP(lambda, tmp, inverse);
MulModP(lambda2, lambda, lambda);
SubModP(tmp, lambda2, pnt1x);
SubModP(res_x, tmp, pnt2x);
SubModP(tmp, pnt2x, res_x);
MulModP(tmp2, tmp, lambda);
SubModP(res_y, tmp2, pnt2y);
}
__device__ __forceinline__ void DoublePoint(u64* res_x, u64* res_y, u64* pntx, u64* pnty)
{
__align__(16) u64 tmp[4], tmp2[4], lambda[4], lambda2[4];
__align__(16) u64 inverse[5];
AddModP(inverse, pnty, pnty);
InvModP((u32*)inverse);
MulModP(tmp2, pntx, pntx);
AddModP(tmp, tmp2, tmp2);
AddModP(tmp, tmp, tmp2);
MulModP(lambda, tmp, inverse);
MulModP(lambda2, lambda, lambda);
SubModP(tmp, lambda2, pntx);
SubModP(res_x, tmp, pntx);
SubModP(tmp, pntx, res_x);
MulModP(tmp2, tmp, lambda);
SubModP(res_y, tmp2, pnty);
}
//this kernel calculates start points of kangs
extern "C" __launch_bounds__(BLOCK_SIZE, 1)
__global__ void KernelGen(const TKparams Kparams)
{
for (u32 group = 0; group < PNT_GROUP_CNT; group++)
{
__align__(16) u64 x0[4], y0[4], d[3];
__align__(16) u64 x[4], y[4];
__align__(16) u64 tx[4], ty[4];
__align__(16) u64 t2x[4], t2y[4];
u32 kang_ind = PNT_GROUP_CNT * (THREAD_X + BLOCK_X * BLOCK_SIZE) + group;
x0[0] = Kparams.Kangs[kang_ind * 12 + 0];
x0[1] = Kparams.Kangs[kang_ind * 12 + 1];
x0[2] = Kparams.Kangs[kang_ind * 12 + 2];
x0[3] = Kparams.Kangs[kang_ind * 12 + 3];
y0[0] = Kparams.Kangs[kang_ind * 12 + 4];
y0[1] = Kparams.Kangs[kang_ind * 12 + 5];
y0[2] = Kparams.Kangs[kang_ind * 12 + 6];
y0[3] = Kparams.Kangs[kang_ind * 12 + 7];
d[0] = Kparams.Kangs[kang_ind * 12 + 8];
d[1] = Kparams.Kangs[kang_ind * 12 + 9];
d[2] = Kparams.Kangs[kang_ind * 12 + 10];
tx[0] = GX_0; tx[1] = GX_1; tx[2] = GX_2; tx[3] = GX_3;
ty[0] = GY_0; ty[1] = GY_1; ty[2] = GY_2; ty[3] = GY_3;
bool first = true;
int n = 2;
while ((n >= 0) && !d[n])
n--;
if (n < 0)
continue; //error
int index = __clzll(d[n]);
for (int i = 0; i <= 64 * n + (63 - index); i++)
{
u8 v = (d[i / 64] >> (i % 64)) & 1;
if (v)
{
if (first)
{
first = false;
Copy_u64_x4(x, tx);
Copy_u64_x4(y, ty);
}
else
{
AddPoints(t2x, t2y, x, y, tx, ty);
Copy_u64_x4(x, t2x);
Copy_u64_x4(y, t2y);
}
}
DoublePoint(t2x, t2y, tx, ty);
Copy_u64_x4(tx, t2x);
Copy_u64_x4(ty, t2y);
}
if (!Kparams.IsGenMode)
if (kang_ind >= Kparams.KangCnt / 3)
{
AddPoints(t2x, t2y, x, y, x0, y0);
Copy_u64_x4(x, t2x);
Copy_u64_x4(y, t2y);
}
Kparams.Kangs[kang_ind * 12 + 0] = x[0];
Kparams.Kangs[kang_ind * 12 + 1] = x[1];
Kparams.Kangs[kang_ind * 12 + 2] = x[2];
Kparams.Kangs[kang_ind * 12 + 3] = x[3];
Kparams.Kangs[kang_ind * 12 + 4] = y[0];
Kparams.Kangs[kang_ind * 12 + 5] = y[1];
Kparams.Kangs[kang_ind * 12 + 6] = y[2];
Kparams.Kangs[kang_ind * 12 + 7] = y[3];
}
}
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void CallGpuKernelABC(TKparams Kparams)
{
KernelA <<< Kparams.BlockCnt, Kparams.BlockSize, Kparams.KernelA_LDS_Size >>> (Kparams);
KernelB <<< Kparams.BlockCnt, Kparams.BlockSize, Kparams.KernelB_LDS_Size >>> (Kparams);
KernelC <<< Kparams.BlockCnt, Kparams.BlockSize, Kparams.KernelC_LDS_Size >>> (Kparams);
}
void CallGpuKernelGen(TKparams Kparams)
{
KernelGen << < Kparams.BlockCnt, Kparams.BlockSize, 0 >> > (Kparams);
}
cudaError_t cuSetGpuParams(TKparams Kparams, u64* _jmp2_table)
{
cudaError_t err = cudaFuncSetAttribute(KernelA, cudaFuncAttributeMaxDynamicSharedMemorySize, Kparams.KernelA_LDS_Size);
if (err != cudaSuccess)
return err;
err = cudaFuncSetAttribute(KernelB, cudaFuncAttributeMaxDynamicSharedMemorySize, Kparams.KernelB_LDS_Size);
if (err != cudaSuccess)
return err;
err = cudaFuncSetAttribute(KernelC, cudaFuncAttributeMaxDynamicSharedMemorySize, Kparams.KernelC_LDS_Size);
if (err != cudaSuccess)
return err;
err = cudaMemcpyToSymbol(jmp2_table, _jmp2_table, JMP_CNT * 64);
if (err != cudaSuccess)
return err;
return cudaSuccess;
}