Repository navigation
Expand file tree
/
Copy pathperf.cpp
More file actions
890 lines (800 loc) · 42.6 KB
/
Copy pathperf.cpp
File metadata and controls
890 lines (800 loc) · 42.6 KB
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
// Copyright © 2019-2023
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// ============================================================================
// perf.cpp — vortex2.h MPM performance-counter report.
//
// vx_device_dump_perf walks the per-core / cluster / cache MPM counters via
// the public vx_device_query + vx_device_mpm_query primitives and prints a
// formatted report. Profiling class is selected by VORTEX_PROFILING.
// ============================================================================
#include <common.h>
#include <vortex2.h>
#include "vortex2_internal.h" // Device / to_device — vx_device_mpm_query CP access
#include <algorithm>
#include <cinttypes>
#include <cmath>
#include <cstdarg>
#include <cstdio>
#include <cstdlib>
#include <sstream>
#include <string>
#include <tuple>
#include <vector>
// -----------------------------------------------------------------------------
// Helpers
// -----------------------------------------------------------------------------
class ProfilingMode {
public:
ProfilingMode() : mpm_class_(0) {
auto profiling_s = getenv("VORTEX_PROFILING");
if (profiling_s) {
mpm_class_ = std::atoi(profiling_s);
}
}
~ProfilingMode() {}
int mpm_class() const {
return mpm_class_;
}
private:
int mpm_class_;
};
static int get_profiling_mode() {
static ProfilingMode gProfilingMode;
return gProfilingMode.mpm_class();
}
static inline double safe_div(double num, double den) {
return (den == 0.0) ? 0.0 : (num / den);
}
static inline int calc_percent(uint64_t part, uint64_t total) {
return (total == 0) ? 0 : (int)std::lround(safe_div((double)part, (double)total) * 100.0);
}
static inline int calc_ratio(uint64_t misses, uint64_t accesses) {
if (accesses == 0)
return 0;
double miss_rate = safe_div((double)misses, (double)accesses);
return (int)std::lround((1.0 - miss_rate) * 100.0);
}
static inline int calc_utility(uint64_t useful, uint64_t stalls) {
return calc_percent(useful, useful + stalls);
}
static inline double bytes_to_GB(uint64_t bytes) {
return (double)bytes / 1e9;
}
static inline void perf_print_core(FILE *stream, int core_id, const char *fmt, ...) {
std::va_list args;
va_start(args, fmt);
std::fprintf(stream, "PERF: core%d: ", core_id);
std::vfprintf(stream, fmt, args);
std::fprintf(stream, "\n");
va_end(args);
}
static inline void perf_print(FILE *stream, const char *fmt, ...) {
std::va_list args;
va_start(args, fmt);
std::fprintf(stream, "PERF: ");
std::vfprintf(stream, fmt, args);
std::fprintf(stream, "\n");
va_end(args);
}
// -----------------------------------------------------------------------------
// Data structs
// -----------------------------------------------------------------------------
struct CoreCounters {
uint64_t cycles = 0;
uint64_t instrs = 0;
// scheduler
uint64_t sched_idle = 0;
uint64_t active_warps = 0;
uint64_t stalled_warps = 0;
uint64_t issued_warps = 0;
uint64_t issued_threads = 0;
// pipeline stalls
uint64_t stall_fetch = 0;
uint64_t stall_ibuf = 0;
uint64_t stall_scrb = 0;
uint64_t stall_opds = 0;
uint64_t stall_alu = 0;
uint64_t stall_fpu = 0;
uint64_t stall_lsu = 0;
uint64_t stall_sfu = 0;
uint64_t stall_tcu = 0;
// workload mix
uint64_t instr_alu = 0;
uint64_t instr_fpu = 0;
uint64_t instr_lsu = 0;
uint64_t instr_sfu = 0;
uint64_t instr_tcu = 0;
// branches
uint64_t branches = 0;
uint64_t divergence = 0;
// memory (front-end + LSU)
uint64_t ifetches = 0;
uint64_t ifetch_lt = 0;
uint64_t loads = 0;
uint64_t load_lt = 0;
uint64_t stores = 0;
// memory
uint64_t mem_reads = 0;
uint64_t mem_writes = 0;
};
struct CacheCounters {
uint64_t reads = 0;
uint64_t writes = 0;
uint64_t miss_r = 0;
uint64_t miss_w = 0;
uint64_t bank_st = 0;
uint64_t mshr_st = 0;
};
struct RtuCounters {
uint64_t rays = 0;
uint64_t hits = 0;
uint64_t box_tests = 0;
uint64_t tri_tests = 0;
uint64_t callbacks = 0;
uint64_t busy = 0;
CacheCounters rtcache;
RtuCounters& operator+=(const RtuCounters& rhs) {
rays += rhs.rays;
hits += rhs.hits;
box_tests += rhs.box_tests;
tri_tests += rhs.tri_tests;
callbacks += rhs.callbacks;
busy += rhs.busy;
rtcache.reads += rhs.rtcache.reads;
rtcache.miss_r += rhs.rtcache.miss_r;
return *this;
}
};
static vx_result_t query_rtu_counters(vx_device_h hdevice, uint32_t mpm_class,
uint32_t core_id, RtuCounters* c) {
const struct {
uint32_t addr;
uint64_t* value;
} fields[] = {
{VX_CSR_MPM_RTU_RAYS, &c->rays},
{VX_CSR_MPM_RTU_HITS, &c->hits},
{VX_CSR_MPM_RTU_BOX_TESTS, &c->box_tests},
{VX_CSR_MPM_RTU_TRI_TESTS, &c->tri_tests},
{VX_CSR_MPM_RTU_CALLBACKS, &c->callbacks},
{VX_CSR_MPM_RTU_BUSY, &c->busy},
{VX_CSR_MPM_RTCACHE_READS, &c->rtcache.reads},
{VX_CSR_MPM_RTCACHE_MISS_R, &c->rtcache.miss_r},
};
for (const auto& f : fields) {
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, f.addr, core_id, f.value), { return err; });
}
return VX_SUCCESS;
}
// Raw counters first (name=value), derived metrics in parentheses. `units` is
// the number of RTU sockets the counters were summed over, so the utilization
// stays a per-RTU average.
static void print_rtu_counters(FILE* stream, int core_id, const RtuCounters& c,
uint64_t cycles, uint64_t units) {
const double rays = (double)c.rays;
const double unit_cycles = (double)cycles * (double)units;
char buf[512];
std::snprintf(buf, sizeof(buf),
"rtu: rays=%" PRIu64 ", hits=%" PRIu64 " (rate=%d%%), box_tests=%" PRIu64 " (%.1f/ray)"
", tri_tests=%" PRIu64 " (%.1f/ray), callbacks=%" PRIu64 ", busy=%" PRIu64 " (util=%d%%)"
"; rtcache: reqs=%" PRIu64 " (%.1f/ray), miss=%" PRIu64 " (hit=%d%%)",
c.rays, c.hits, calc_percent(c.hits, c.rays),
c.box_tests, safe_div((double)c.box_tests, rays),
c.tri_tests, safe_div((double)c.tri_tests, rays),
c.callbacks,
c.busy, (int)std::lround(safe_div((double)c.busy, unit_cycles) * 100.0),
c.rtcache.reads, safe_div((double)c.rtcache.reads, rays),
c.rtcache.miss_r, calc_ratio(c.rtcache.miss_r, c.rtcache.reads));
if (core_id >= 0) {
perf_print_core(stream, core_id, "%s", buf);
} else {
perf_print(stream, "%s", buf);
}
}
struct Metric {
const char *label;
int value;
bool enabled;
};
static void print_metric_list(FILE *stream, const char *category, int core_id, const std::vector<Metric> &metrics) {
std::stringstream ss;
ss << category << ": ";
bool first = true;
for (const auto &m : metrics) {
if (m.enabled) {
if (!first)
ss << ", ";
ss << m.label << "=" << m.value << "%";
first = false;
}
}
if (core_id >= 0) {
perf_print_core(stream, core_id, "%s", ss.str().c_str());
} else {
perf_print(stream, "%s", ss.str().c_str());
}
}
// Read a single 64-bit MPM performance counter. `addr` is an MPM CSR in
// [VX_CSR_MPM_BASE, VX_CSR_MPM_BASE+32); `core_id == 0xffffffff` sums the
// counter across all cores. The 64-bit value is assembled from two 32-bit
// DCR half-reads (lo at csr_id, hi at csr_id+32), each packed with
// mpm_class|csr_id|core onto the DCR bus. Submits straight to the CP — the
// public vx_enqueue_dcr_read cannot carry the MPM tag.
extern "C" vx_result_t vx_device_mpm_query(vx_device_h dev, uint32_t mpm_class,
uint32_t addr, uint32_t core_id,
uint64_t* out_value) {
if (!dev) return VX_ERR_INVALID_HANDLE;
if (!out_value) return VX_ERR_INVALID_VALUE;
if (!(addr >= VX_CSR_MPM_BASE && addr < (VX_CSR_MPM_BASE + 32)))
return VX_ERR_INVALID_VALUE;
vx::Device* device = vx::to_device(dev);
const uint32_t csr_id = addr - VX_CSR_MPM_BASE;
const uint32_t clss_sh = mpm_class << (16 + 6);
auto read_one = [&](uint32_t cid, uint64_t* out) -> vx_result_t {
uint32_t lo = 0, hi = 0;
uint32_t tag_lo = clss_sh | ((csr_id << 16) | cid);
uint32_t tag_hi = clss_sh | (((csr_id + 32) << 16) | cid);
auto r = device->cp_submit_dcr_read(VX_DCR_BASE_MPM_VALUE, tag_lo, &lo);
if (r != VX_SUCCESS) return r;
r = device->cp_submit_dcr_read(VX_DCR_BASE_MPM_VALUE, tag_hi, &hi);
if (r != VX_SUCCESS) return r;
*out = ((uint64_t)hi << 32) | lo;
return VX_SUCCESS;
};
if (core_id == 0xffffffff) {
uint64_t num_cores = 0;
auto r = device->query_caps(VX_CAPS_NUM_CORES, &num_cores);
if (r != VX_SUCCESS) return r;
uint64_t sum = 0;
for (uint32_t i = 0; i < (uint32_t)num_cores; ++i) {
uint64_t cur = 0;
r = read_one(i, &cur);
if (r != VX_SUCCESS) return r;
sum += cur;
}
*out_value = sum;
return VX_SUCCESS;
}
return read_one(core_id, out_value);
}
extern "C" vx_result_t vx_device_dump_perf(vx_device_h hdevice, FILE *stream) {
if (nullptr == stream)
stream = stdout;
uint64_t num_cores = 0;
uint64_t num_clusters = 0;
uint64_t socket_size = 0;
uint64_t num_warps = 0;
uint64_t num_threads = 0;
uint64_t issue_width = 0;
uint64_t isa_flags = 0;
uint64_t clock_mhz = 0;
uint64_t peak_mem_bw_MBps = 0;
uint64_t vm_support = 0;
const auto mpm_class = get_profiling_mode();
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_NUM_CORES, &num_cores), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_NUM_CLUSTERS, &num_clusters), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_SOCKET_SIZE, &socket_size), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_NUM_WARPS, &num_warps), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_NUM_THREADS, &num_threads), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_ISSUE_WIDTH, &issue_width), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_ISA_FLAGS, &isa_flags), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_CLOCK_RATE, &clock_mhz), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_PEAK_MEM_BW, &peak_mem_bw_MBps), { return err; });
CHECK_ERR(vx_device_query(hdevice, VX_CAPS_VM_SUPPORT, &vm_support), { return err; });
// VM is a runtime device property — gate the MMU/TLB perf counters on it.
const bool vm_enabled = (vm_support != 0);
const bool icache_en = (isa_flags & VX_ISA_EXT_ICACHE) != 0;
const bool dcache_en = (isa_flags & VX_ISA_EXT_DCACHE) != 0;
const bool l2_en = (isa_flags & VX_ISA_EXT_L2CACHE) != 0;
const bool l3_en = (isa_flags & VX_ISA_EXT_L3CACHE) != 0;
const bool lmem_en = (isa_flags & VX_ISA_EXT_LMEM) != 0;
const bool fpu_en = (isa_flags & VX_ISA_STD_F) != 0;
const bool tcu_en = (isa_flags & VX_ISA_EXT_TCU) != 0;
const bool tex_en = (isa_flags & VX_ISA_EXT_TEX) != 0;
const bool raster_en = (isa_flags & VX_ISA_EXT_RASTER) != 0;
const bool om_en = (isa_flags & VX_ISA_EXT_OM) != 0;
const bool rtu_en = (isa_flags & VX_ISA_EXT_RTU) != 0;
std::vector<CoreCounters> cores(num_cores);
uint64_t total_instrs = 0;
uint64_t max_cycles = 0;
for (uint32_t core_id = 0; core_id < num_cores; ++core_id) {
auto &c = cores[core_id];
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MCYCLE, core_id, &c.cycles), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MINSTRET, core_id, &c.instrs), { return err; });
total_instrs += c.instrs;
max_cycles = std::max<uint64_t>(max_cycles, c.cycles);
}
switch (mpm_class) {
case VX_DCR_MPM_CLASS_BASE:
break;
case VX_DCR_MPM_CLASS_CORE: {
CoreCounters tot;
for (uint32_t core_id = 0; core_id < num_cores; ++core_id) {
auto &c = cores[core_id];
// Query Core Specifics
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_SCHED_IDLE, core_id, &c.sched_idle), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_ACTIVE_WARPS, core_id, &c.active_warps), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALLED_WARPS, core_id, &c.stalled_warps), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_ISSUED_WARPS, core_id, &c.issued_warps), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_ISSUED_THREADS, core_id, &c.issued_threads), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_FETCH, core_id, &c.stall_fetch), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_IBUF, core_id, &c.stall_ibuf), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_SCRB, core_id, &c.stall_scrb), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_OPDS, core_id, &c.stall_opds), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_ALU, core_id, &c.stall_alu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_FPU, core_id, &c.stall_fpu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_LSU, core_id, &c.stall_lsu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_SFU, core_id, &c.stall_sfu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STALL_TCU, core_id, &c.stall_tcu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_INSTR_ALU, core_id, &c.instr_alu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_INSTR_FPU, core_id, &c.instr_fpu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_INSTR_LSU, core_id, &c.instr_lsu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_INSTR_SFU, core_id, &c.instr_sfu), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_INSTR_TCU, core_id, &c.instr_tcu), { return err; });
// Branches
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_BRANCHES, core_id, &c.branches), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DIVERGENCE, core_id, &c.divergence), { return err; });
// Memory (front-end + LSU)
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_IFETCHES, core_id, &c.ifetches), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_IFETCH_LT, core_id, &c.ifetch_lt), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_LOADS, core_id, &c.loads), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_LOAD_LT, core_id, &c.load_lt), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_STORES, core_id, &c.stores), { return err; });
if (num_cores > 1) {
// Per-Core report
const uint64_t cycles = c.cycles;
const uint64_t cycles_wide = cycles * issue_width;
const double ipc = safe_div((double)c.instrs, (double)cycles);
// Scheduler report
const int idle_pct = calc_percent(c.sched_idle, cycles);
const double avg_occ = safe_div((double)c.active_warps, (double)cycles);
const int occ_pct = (int)std::lround(safe_div(avg_occ, (double)num_warps) * 100.0);
const double warp_eff = safe_div((double)c.issued_threads, (double)c.issued_warps);
const int warp_eff_pct = (int)std::lround(safe_div(warp_eff, (double)num_threads) * 100.0);
perf_print_core(stream, core_id, "scheduler: idle=%d%%, occupancy=%.1f (%d%%), simt_util=%.1f (%d%%)",
idle_pct, avg_occ, occ_pct, warp_eff, warp_eff_pct);
// Pipeline stalls report
std::vector<Metric> stall_metrics = {
{"fetch", calc_percent(c.stall_fetch, cycles), true},
{"ibuf", calc_percent(c.stall_ibuf, cycles), true},
{"scrb", calc_percent(c.stall_scrb, cycles_wide), true},
{"opds", calc_percent(c.stall_opds, cycles_wide), true},
{"alu", calc_percent(c.stall_alu, cycles_wide), true},
{"lsu", calc_percent(c.stall_lsu, cycles_wide), true},
{"sfu", calc_percent(c.stall_sfu, cycles_wide), true},
{"fpu", calc_percent(c.stall_fpu, cycles_wide), fpu_en},
{"tcu", calc_percent(c.stall_tcu, cycles_wide), tcu_en}
};
print_metric_list(stream, "stalls", core_id, stall_metrics);
// Instruction mix report
std::vector<Metric> mix_metrics = {
{"alu", calc_percent(c.instr_alu, c.instrs), true},
{"lsu", calc_percent(c.instr_lsu, c.instrs), true},
{"sfu", calc_percent(c.instr_sfu, c.instrs), true},
{"fpu", calc_percent(c.instr_fpu, c.instrs), fpu_en},
{"tcu", calc_percent(c.instr_tcu, c.instrs), tcu_en}
};
print_metric_list(stream, "inst_mix", core_id, mix_metrics);
// Branch report
int div_pct = calc_percent(c.divergence, c.branches);
perf_print_core(stream, core_id, "branches: total=%" PRIu64 ", divergent=%" PRIu64 " (rate=%d%%)",
c.branches, c.divergence, div_pct);
// Memory report
const double ifetch_avg_lt = safe_div((double)c.ifetch_lt, (double)c.ifetches);
const double load_avg_lt = safe_div((double)c.load_lt, (double)c.loads);
perf_print_core(stream, core_id, "memory: ifetch_lat=%.2f, load_lat=%.2f, loads=%" PRIu64 ", stores=%" PRIu64,
ifetch_avg_lt, load_avg_lt, c.loads, c.stores);
perf_print_core(stream, core_id, "instrs=%" PRIu64 ", cycles=%" PRIu64 ", IPC=%.3f", c.instrs, c.cycles, ipc);
}
// Accumulate
tot.cycles += c.cycles;
tot.instrs += c.instrs;
tot.sched_idle += c.sched_idle;
tot.active_warps += c.active_warps;
tot.stalled_warps += c.stalled_warps;
tot.issued_warps += c.issued_warps;
tot.issued_threads += c.issued_threads;
tot.stall_fetch += c.stall_fetch;
tot.stall_ibuf += c.stall_ibuf;
tot.stall_scrb += c.stall_scrb;
tot.stall_opds += c.stall_opds;
tot.stall_alu += c.stall_alu;
tot.stall_fpu += c.stall_fpu;
tot.stall_lsu += c.stall_lsu;
tot.stall_sfu += c.stall_sfu;
tot.stall_tcu += c.stall_tcu;
tot.instr_alu += c.instr_alu;
tot.instr_fpu += c.instr_fpu;
tot.instr_lsu += c.instr_lsu;
tot.instr_sfu += c.instr_sfu;
tot.instr_tcu += c.instr_tcu;
tot.branches += c.branches;
tot.divergence += c.divergence;
tot.ifetches += c.ifetches;
tot.ifetch_lt += c.ifetch_lt;
tot.loads += c.loads;
tot.load_lt += c.load_lt;
tot.stores += c.stores;
}
// Core Summary
uint64_t tot_cycles_wide = tot.cycles * issue_width;
// Scheduler report
const int tot_idle_pct = calc_percent(tot.sched_idle, tot.cycles);
const double tot_avg_occ = safe_div((double)tot.active_warps, (double)tot.cycles);
const int tot_occ_pct = (int)std::lround(safe_div(tot_avg_occ, (double)num_warps) * 100.0);
const double tot_warp_eff = safe_div((double)tot.issued_threads, (double)tot.issued_warps);
const int tot_warp_eff_pct = (int)std::lround(safe_div(tot_warp_eff, (double)num_threads) * 100.0);
perf_print(stream, "scheduler: idle=%d%%, occupancy=%.1f (%d%%), simt_util=%.1f (%d%%)",
tot_idle_pct, tot_avg_occ, tot_occ_pct, tot_warp_eff, tot_warp_eff_pct);
// Pipeline stalls report
std::vector<Metric> global_stalls = {
{"fetch", calc_percent(tot.stall_fetch, tot.cycles), true},
{"ibuf", calc_percent(tot.stall_ibuf, tot.cycles), true},
{"scrb", calc_percent(tot.stall_scrb, tot_cycles_wide), true},
{"opds", calc_percent(tot.stall_opds, tot_cycles_wide), true},
{"alu", calc_percent(tot.stall_alu, tot_cycles_wide), true},
{"lsu", calc_percent(tot.stall_lsu, tot_cycles_wide), true},
{"sfu", calc_percent(tot.stall_sfu, tot_cycles_wide), true},
{"fpu", calc_percent(tot.stall_fpu, tot_cycles_wide), fpu_en},
{"tcu", calc_percent(tot.stall_tcu, tot_cycles_wide), tcu_en}
};
print_metric_list(stream, "stalls", -1, global_stalls);
// Instruction mix report
std::vector<Metric> global_mix = {
{"alu", calc_percent(tot.instr_alu, tot.instrs), true},
{"lsu", calc_percent(tot.instr_lsu, tot.instrs), true},
{"sfu", calc_percent(tot.instr_sfu, tot.instrs), true},
{"fpu", calc_percent(tot.instr_fpu, tot.instrs), fpu_en},
{"tcu", calc_percent(tot.instr_tcu, tot.instrs), tcu_en}
};
print_metric_list(stream, "inst_mix", -1, global_mix);
// Branch report
int tot_div_pct = calc_percent(tot.divergence, tot.branches);
perf_print(stream, "branches: total=%" PRIu64 ", divergent=%" PRIu64 " (rate=%d%%)",
tot.branches, tot.divergence, tot_div_pct);
// Memory report
const double tot_ifetch_avg_lt = safe_div((double)tot.ifetch_lt, (double)tot.ifetches);
const double tot_load_avg_lt = safe_div((double)tot.load_lt, (double)tot.loads);
perf_print(stream, "memory: ifetch_lat=%.2f, load_lat=%.2f, loads=%" PRIu64 ", stores=%" PRIu64,
tot_ifetch_avg_lt, tot_load_avg_lt, tot.loads, tot.stores);
} break;
// Each cache level is now its own MPM class; --perf selects which to dump.
case VX_DCR_MPM_CLASS_ICACHE: {
if (icache_en) {
const uint64_t num_sockets = (num_cores + socket_size - 1) / socket_size;
for (uint32_t s = 0; s < num_sockets; ++s) {
uint32_t rep_core = s * (uint32_t)socket_size;
if (rep_core >= num_cores)
continue;
uint64_t r = 0, m = 0, st = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_ICACHE_READS, rep_core, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_ICACHE_MISS_R, rep_core, &m), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_ICACHE_MSHR_ST, rep_core, &st), { return err; });
perf_print_core(stream, rep_core, "icache: reads=%" PRIu64 ", miss=%" PRIu64 " (hit=%d%%), mshr_st=%" PRIu64 " (utility=%d%%)",
r, m, calc_ratio(m, r), st, calc_utility(m, st));
}
} else {
perf_print(stream, "icache: disabled");
}
} break;
case VX_DCR_MPM_CLASS_DCACHE: {
if (dcache_en) {
const uint64_t num_sockets = (num_cores + socket_size - 1) / socket_size;
for (uint32_t s = 0; s < num_sockets; ++s) {
uint32_t rep_core = s * (uint32_t)socket_size;
if (rep_core >= num_cores)
continue;
uint64_t r = 0, w = 0, mr = 0, mw = 0, ev = 0, bst = 0, mst = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DCACHE_READS, rep_core, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DCACHE_WRITES, rep_core, &w), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DCACHE_MISS_R, rep_core, &mr), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DCACHE_MISS_W, rep_core, &mw), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DCACHE_EVICTS, rep_core, &ev), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DCACHE_BANK_ST, rep_core, &bst), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DCACHE_MSHR_ST, rep_core, &mst), { return err; });
perf_print_core(stream, rep_core, "dcache: reqs=%" PRIu64 ", miss_r=%" PRIu64 " (hit=%d%%), miss_w=%" PRIu64 " (hit=%d%%), evicts=%" PRIu64 ", bank_st=%" PRIu64 " (utility=%d%%)",
r + w, mr, calc_ratio(mr, r), mw, calc_ratio(mw, w), ev, bst, calc_utility(r + w, bst));
}
} else {
perf_print(stream, "dcache: disabled");
}
} break;
case VX_DCR_MPM_CLASS_L2CACHE: {
if (l2_en) {
const uint64_t cores_per_cluster = (num_cores + num_clusters - 1) / num_clusters;
for (uint32_t c = 0; c < num_clusters; ++c) {
uint32_t rep_core = c * (uint32_t)cores_per_cluster;
if (rep_core >= num_cores)
continue;
uint64_t r = 0, w = 0, mr = 0, mw = 0, ev = 0, bst = 0, mst = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L2CACHE_READS, rep_core, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L2CACHE_WRITES, rep_core, &w), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L2CACHE_MISS_R, rep_core, &mr), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L2CACHE_MISS_W, rep_core, &mw), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L2CACHE_EVICTS, rep_core, &ev), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L2CACHE_BANK_ST, rep_core, &bst), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L2CACHE_MSHR_ST, rep_core, &mst), { return err; });
perf_print_core(stream, rep_core, "l2cache: reqs=%" PRIu64 ", miss_r=%" PRIu64 " (hit=%d%%), miss_w=%" PRIu64 " (hit=%d%%), evicts=%" PRIu64,
r + w, mr, calc_ratio(mr, r), mw, calc_ratio(mw, w), ev);
}
} else {
perf_print(stream, "l2cache: disabled");
}
} break;
case VX_DCR_MPM_CLASS_L3CACHE: {
if (l3_en) {
uint64_t r = 0, w = 0, mr = 0, mw = 0, ev = 0, bst = 0, mst = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L3CACHE_READS, 0, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L3CACHE_WRITES, 0, &w), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L3CACHE_MISS_R, 0, &mr), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L3CACHE_MISS_W, 0, &mw), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L3CACHE_EVICTS, 0, &ev), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L3CACHE_BANK_ST, 0, &bst), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_L3CACHE_MSHR_ST, 0, &mst), { return err; });
perf_print(stream, "l3cache: reqs=%" PRIu64 ", miss_r=%" PRIu64 " (hit=%d%%), miss_w=%" PRIu64 " (hit=%d%%), evicts=%" PRIu64,
r + w, mr, calc_ratio(mr, r), mw, calc_ratio(mw, w), ev);
} else {
perf_print(stream, "l3cache: disabled");
}
} break;
case VX_DCR_MPM_CLASS_MEM: {
// Per-core local memory + coalescer + VM/MMU.
for (uint32_t core_id = 0; core_id < num_cores; ++core_id) {
if (lmem_en) {
uint64_t r = 0, w = 0, bst = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_LMEM_READS, core_id, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_LMEM_WRITES, core_id, &w), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_LMEM_BANK_ST, core_id, &bst), { return err; });
perf_print_core(stream, core_id, "lmem: reqs=%" PRIu64 ", bank_stalls=%" PRIu64 " (utility=%d%%)",
r + w, bst, calc_utility(r + w, bst));
}
// The coalescer lives on the LSU->memory path; it exists when
// NUM_LSU_LANES > 1 && LSU_LINE_SIZE > XLENB, independent of the dcache.
// The counter is hard-zeroed in hardware when no coalescer is present.
{
uint64_t cm = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_COALESCER_MISS, core_id, &cm), { return err; });
perf_print_core(stream, core_id, "coalescer: misses=%" PRIu64, cm);
}
// VM/MMU (per-core; hardware sums icache + dcache MMU counters).
if (vm_enabled) {
uint64_t reads = 0, hits = 0, evicts = 0, walks = 0, lat = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TLB_READS, core_id, &reads), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TLB_HITS, core_id, &hits), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TLB_EVICTS, core_id, &evicts), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_PTW_WALKS, core_id, &walks), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_PTW_LATENCY, core_id, &lat), { return err; });
perf_print_core(stream, core_id, "vm: tlb_reads=%" PRIu64 ", hit=%d%%, evicts=%" PRIu64 ", ptw_walks=%" PRIu64 ", ptw_avg_lat=%.2f",
reads, calc_percent(hits, reads), evicts, walks, safe_div((double)lat, (double)walks));
}
}
// Global off-chip memory.
{
uint64_t r = 0, w = 0, lat = 0, bst = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_MEM_READS, 0, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_MEM_WRITES, 0, &w), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_MEM_LT, 0, &lat), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_MEM_BANK_ST, 0, &bst), { return err; });
double avg_lat = safe_div((double)lat, (double)r);
perf_print(stream, "memory: reqs=%" PRIu64 " (r=%" PRIu64 ", w=%" PRIu64 "), lat=%.1f cyc, bank_st=%" PRIu64 " (utility=%d%%)",
r + w, r, w, avg_lat, bst, calc_utility(r + w, bst));
}
} break;
case VX_DCR_MPM_CLASS_DXA: {
// DXA counters are cluster-level; query representative core per cluster.
uint64_t cores_per_cluster = (num_cores + num_clusters - 1) / num_clusters;
uint64_t tot_transfers = 0, tot_gmem_reads = 0, tot_gmem_dedup = 0;
uint64_t tot_lmem_writes = 0, tot_gmem_lt = 0, tot_noslot = 0;
for (uint32_t c = 0; c < num_clusters; ++c) {
uint32_t rep_core = (uint32_t)(c * cores_per_cluster);
if (rep_core >= num_cores)
break;
uint64_t transfers = 0, gmem_reads = 0, gmem_dedup = 0, lmem_writes = 0, gmem_lt = 0, noslot = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DXA_TRANSFERS, rep_core, &transfers), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DXA_GMEM_READS, rep_core, &gmem_reads), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DXA_GMEM_DEDUP, rep_core, &gmem_dedup), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DXA_LMEM_WRITES,rep_core, &lmem_writes),{ return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DXA_GMEM_LT, rep_core, &gmem_lt), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_DXA_NOSLOT_STALLS, rep_core, &noslot), { return err; });
tot_transfers += transfers;
tot_gmem_reads += gmem_reads;
tot_gmem_dedup += gmem_dedup;
tot_lmem_writes+= lmem_writes;
tot_gmem_lt += gmem_lt;
tot_noslot += noslot;
double avg_lat = safe_div((double)gmem_lt, (double)transfers);
int dedup_pct = calc_percent(gmem_dedup, gmem_reads + gmem_dedup);
perf_print_core(stream, rep_core, "dxa: transfers=%" PRIu64 ", gmem_reads=%" PRIu64 ", gmem_dedup=%" PRIu64 " (rate=%d%%), lmem_writes=%" PRIu64 ", avg_gmem_lat=%.1f, noslot_stalls=%" PRIu64,
transfers, gmem_reads, gmem_dedup, dedup_pct, lmem_writes, avg_lat, noslot);
}
{
double avg_lat = safe_div((double)tot_gmem_lt, (double)tot_transfers);
int dedup_pct = calc_percent(tot_gmem_dedup, tot_gmem_reads + tot_gmem_dedup);
perf_print(stream, "dxa: transfers=%" PRIu64 ", gmem_reads=%" PRIu64 ", gmem_dedup=%" PRIu64 " (rate=%d%%), lmem_writes=%" PRIu64 ", avg_gmem_lat=%.1f, noslot_stalls=%" PRIu64,
tot_transfers, tot_gmem_reads, tot_gmem_dedup, dedup_pct, tot_lmem_writes, avg_lat, tot_noslot);
}
} break;
case VX_DCR_MPM_CLASS_TCU: {
if (!tcu_en) {
perf_print(stream, "TCU not supported");
break;
}
uint64_t tot_tbuf_stalls = 0, tot_tbuf_cache_hits = 0, tot_lmem_reads = 0;
for (uint32_t core_id = 0; core_id < num_cores; ++core_id) {
uint64_t tbuf_stalls = 0, tbuf_cache_hits = 0, lmem_reads = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TCU_TBUF_STALLS, core_id, &tbuf_stalls), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TCU_TBUF_CACHE_HITS, core_id, &tbuf_cache_hits), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TCU_LMEM_READS, core_id, &lmem_reads), { return err; });
perf_print_core(stream, core_id,
"tcu: tbuf_stalls=%" PRIu64 ", tbuf_cache_hits=%" PRIu64 ", lmem_reads=%" PRIu64,
tbuf_stalls, tbuf_cache_hits, lmem_reads);
tot_tbuf_stalls += tbuf_stalls;
tot_tbuf_cache_hits += tbuf_cache_hits;
tot_lmem_reads += lmem_reads;
}
perf_print(stream,
"tcu: total_tbuf_stalls=%" PRIu64 ", total_tbuf_cache_hits=%" PRIu64 ", total_lmem_reads=%" PRIu64,
tot_tbuf_stalls, tot_tbuf_cache_hits, tot_lmem_reads);
} break;
case VX_DCR_MPM_CLASS_TEX: {
if (!tex_en) {
perf_print(stream, "TEX not supported");
break;
}
// Socket-scoped counters: every core of a socket reads the same values.
const uint64_t num_sockets = (num_cores + socket_size - 1) / socket_size;
uint64_t tot_reads = 0, tot_lat = 0, tot_st = 0;
CacheCounters tcache_tot;
for (uint32_t s = 0; s < num_sockets; ++s) {
uint32_t core_id = s * (uint32_t)socket_size;
if (core_id >= num_cores)
break;
uint64_t reads = 0, lat = 0, st = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TEX_READS, core_id, &reads), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TEX_LAT, core_id, &lat), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TEX_ST, core_id, &st), { return err; });
uint64_t r = 0, mr = 0, bst = 0, mst = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TCACHE_READS, core_id, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TCACHE_MISS_R, core_id, &mr), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TCACHE_BANK_ST, core_id, &bst), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_TCACHE_MSHR_ST, core_id, &mst), { return err; });
double avg_lat = safe_div((double)lat, (double)reads);
perf_print_core(stream, core_id,
"tex: reads=%" PRIu64 ", avg_lat=%.1f cyc, stalls=%" PRIu64 "; tcache: reqs=%" PRIu64 ", miss=%" PRIu64 " (hit=%d%%), bank_st=%" PRIu64 " (utility=%d%%)",
reads, avg_lat, st, r, mr, calc_ratio(mr, r), bst, calc_utility(r, bst));
tot_reads += reads; tot_lat += lat; tot_st += st;
tcache_tot.reads += r; tcache_tot.miss_r += mr; tcache_tot.bank_st += bst; tcache_tot.mshr_st += mst;
}
double tot_avg_lat = safe_div((double)tot_lat, (double)tot_reads);
perf_print(stream,
"tex: total_reads=%" PRIu64 ", avg_lat=%.1f cyc, total_stalls=%" PRIu64
"; tcache: total_reqs=%" PRIu64 ", miss=%" PRIu64 " (hit=%d%%), bank_st=%" PRIu64 " (utility=%d%%), mshr_st=%" PRIu64 " (utility=%d%%)",
tot_reads, tot_avg_lat, tot_st,
tcache_tot.reads, tcache_tot.miss_r, calc_ratio(tcache_tot.miss_r, tcache_tot.reads),
tcache_tot.bank_st, calc_utility(tcache_tot.reads, tcache_tot.bank_st),
tcache_tot.mshr_st, calc_utility(tcache_tot.miss_r, tcache_tot.mshr_st));
} break;
case VX_DCR_MPM_CLASS_RASTER: {
if (!raster_en) {
perf_print(stream, "RASTER not supported");
break;
}
// Cluster-scoped counters: every core of a cluster reads the same values.
const uint64_t cores_per_cluster = (num_cores + num_clusters - 1) / num_clusters;
uint64_t tot_reads = 0, tot_lat = 0, tot_st = 0;
CacheCounters rcache_tot;
for (uint32_t c = 0; c < num_clusters; ++c) {
uint32_t core_id = c * (uint32_t)cores_per_cluster;
if (core_id >= num_cores)
break;
uint64_t reads = 0, lat = 0, st = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_RASTER_READS, core_id, &reads), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_RASTER_LAT, core_id, &lat), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_RASTER_ST, core_id, &st), { return err; });
uint64_t r = 0, mr = 0, bst = 0, mst = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_RCACHE_READS, core_id, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_RCACHE_MISS_R, core_id, &mr), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_RCACHE_BANK_ST, core_id, &bst), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_RCACHE_MSHR_ST, core_id, &mst), { return err; });
double avg_lat = safe_div((double)lat, (double)reads);
perf_print_core(stream, core_id,
"raster: reads=%" PRIu64 ", avg_lat=%.1f cyc, stalls=%" PRIu64 "; rcache: reqs=%" PRIu64 ", miss=%" PRIu64 " (hit=%d%%), bank_st=%" PRIu64 " (utility=%d%%)",
reads, avg_lat, st, r, mr, calc_ratio(mr, r), bst, calc_utility(r, bst));
tot_reads += reads; tot_lat += lat; tot_st += st;
rcache_tot.reads += r; rcache_tot.miss_r += mr; rcache_tot.bank_st += bst; rcache_tot.mshr_st += mst;
}
double tot_avg_lat = safe_div((double)tot_lat, (double)tot_reads);
perf_print(stream,
"raster: total_reads=%" PRIu64 ", avg_lat=%.1f cyc, total_stalls=%" PRIu64
"; rcache: total_reqs=%" PRIu64 ", miss=%" PRIu64 " (hit=%d%%), bank_st=%" PRIu64 " (utility=%d%%), mshr_st=%" PRIu64 " (utility=%d%%)",
tot_reads, tot_avg_lat, tot_st,
rcache_tot.reads, rcache_tot.miss_r, calc_ratio(rcache_tot.miss_r, rcache_tot.reads),
rcache_tot.bank_st, calc_utility(rcache_tot.reads, rcache_tot.bank_st),
rcache_tot.mshr_st, calc_utility(rcache_tot.miss_r, rcache_tot.mshr_st));
} break;
case VX_DCR_MPM_CLASS_OM: {
if (!om_en) {
perf_print(stream, "OM not supported");
break;
}
// Cluster-scoped counters: every core of a cluster reads the same values.
const uint64_t cores_per_cluster = (num_cores + num_clusters - 1) / num_clusters;
uint64_t tot_reads = 0, tot_writes = 0, tot_lat = 0, tot_st = 0;
CacheCounters ocache_tot;
for (uint32_t c = 0; c < num_clusters; ++c) {
uint32_t core_id = c * (uint32_t)cores_per_cluster;
if (core_id >= num_cores)
break;
uint64_t reads = 0, writes = 0, lat = 0, st = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OM_READS, core_id, &reads), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OM_WRITES, core_id, &writes), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OM_LAT, core_id, &lat), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OM_ST, core_id, &st), { return err; });
uint64_t r = 0, w = 0, mr = 0, mw = 0, bst = 0, mst = 0;
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OCACHE_READS, core_id, &r), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OCACHE_WRITES, core_id, &w), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OCACHE_MISS_R, core_id, &mr), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OCACHE_MISS_W, core_id, &mw), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OCACHE_BANK_ST, core_id, &bst), { return err; });
CHECK_ERR(vx_device_mpm_query(hdevice, mpm_class, VX_CSR_MPM_OCACHE_MSHR_ST, core_id, &mst), { return err; });
double avg_lat = safe_div((double)lat, (double)reads);
perf_print_core(stream, core_id,
"om: reads=%" PRIu64 ", writes=%" PRIu64 ", avg_lat=%.1f cyc, stalls=%" PRIu64
"; ocache: reqs=%" PRIu64 ", miss_r=%" PRIu64 " (hit=%d%%), miss_w=%" PRIu64 " (hit=%d%%), bank_st=%" PRIu64 " (utility=%d%%)",
reads, writes, avg_lat, st,
r + w, mr, calc_ratio(mr, r), mw, calc_ratio(mw, w), bst, calc_utility(r + w, bst));
tot_reads += reads; tot_writes += writes; tot_lat += lat; tot_st += st;
ocache_tot.reads += r; ocache_tot.writes += w;
ocache_tot.miss_r += mr; ocache_tot.miss_w += mw;
ocache_tot.bank_st += bst; ocache_tot.mshr_st += mst;
}
double tot_avg_lat = safe_div((double)tot_lat, (double)tot_reads);
perf_print(stream,
"om: total_reads=%" PRIu64 ", total_writes=%" PRIu64 ", avg_lat=%.1f cyc, total_stalls=%" PRIu64
"; ocache: total_reqs=%" PRIu64 ", miss_r=%" PRIu64 " (hit=%d%%), miss_w=%" PRIu64 " (hit=%d%%), bank_st=%" PRIu64 " (utility=%d%%), mshr_st=%" PRIu64 " (utility=%d%%)",
tot_reads, tot_writes, tot_avg_lat, tot_st,
ocache_tot.reads + ocache_tot.writes,
ocache_tot.miss_r, calc_ratio(ocache_tot.miss_r, ocache_tot.reads),
ocache_tot.miss_w, calc_ratio(ocache_tot.miss_w, ocache_tot.writes),
ocache_tot.bank_st, calc_utility(ocache_tot.reads + ocache_tot.writes, ocache_tot.bank_st),
ocache_tot.mshr_st, calc_utility(ocache_tot.miss_r + ocache_tot.miss_w, ocache_tot.mshr_st));
} break;
case VX_DCR_MPM_CLASS_RTU: {
if (!rtu_en) {
perf_print(stream, "RTU not supported");
break;
}
// Socket-scoped counters: every core of a socket reads the same values.
const uint64_t num_sockets = (num_cores + socket_size - 1) / socket_size;
RtuCounters tot;
for (uint32_t s = 0; s < num_sockets; ++s) {
uint32_t rep_core = s * (uint32_t)socket_size;
if (rep_core >= num_cores)
break;
RtuCounters c;
CHECK_ERR(query_rtu_counters(hdevice, mpm_class, rep_core, &c), { return err; });
if (num_sockets > 1) {
print_rtu_counters(stream, rep_core, c, cores[rep_core].cycles, 1);
}
tot += c;
}
print_rtu_counters(stream, -1, tot, max_cycles, num_sockets);
} break;
default:
fprintf(stream, "Error: invalid profiling class: %d)", mpm_class);
return VX_ERR_INVALID_VALUE;
}
double global_ipc = safe_div((double)total_instrs, (double)max_cycles);
perf_print(stream, "instrs=%" PRIu64 ", cycles=%" PRIu64 ", IPC=%.3f", total_instrs, max_cycles, global_ipc);
std::fflush(stream);
return VX_SUCCESS;
}