-
-
Notifications
You must be signed in to change notification settings - Fork 8.7k
/
device_helpers.cuh
1642 lines (1450 loc) · 59.5 KB
/
device_helpers.cuh
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
/*!
* Copyright 2017-2022 XGBoost contributors
*/
#pragma once
#include <thrust/device_ptr.h>
#include <thrust/device_vector.h>
#include <thrust/device_malloc_allocator.h>
#include <thrust/iterator/discard_iterator.h>
#include <thrust/iterator/transform_output_iterator.h>
#include <thrust/system/cuda/error.h>
#include <thrust/system_error.h>
#include <thrust/execution_policy.h>
#include <thrust/transform_scan.h>
#include <thrust/logical.h>
#include <thrust/gather.h>
#include <thrust/unique.h>
#include <thrust/binary_search.h>
#include <rabit/rabit.h>
#include <cub/cub.cuh>
#include <cub/util_allocator.cuh>
#include <algorithm>
#include <chrono>
#include <numeric>
#include <sstream>
#include <string>
#include <vector>
#include <tuple>
#include "xgboost/logging.h"
#include "xgboost/host_device_vector.h"
#include "xgboost/span.h"
#include "xgboost/global_config.h"
#include "common.h"
#ifdef XGBOOST_USE_NCCL
#include "nccl.h"
#endif // XGBOOST_USE_NCCL
#if defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
#include "rmm/mr/device/per_device_resource.hpp"
#include "rmm/mr/device/thrust_allocator_adaptor.hpp"
#include "rmm/version_config.hpp"
#if !defined(RMM_VERSION_MAJOR) || !defined(RMM_VERSION_MINOR)
#error "Please use RMM version 0.18 or later"
#elif RMM_VERSION_MAJOR == 0 && RMM_VERSION_MINOR < 18
#error "Please use RMM version 0.18 or later"
#endif // !defined(RMM_VERSION_MAJOR) || !defined(RMM_VERSION_MINOR)
#endif // defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 600 || defined(__clang__)
#else // In device code and CUDA < 600
__device__ __forceinline__ double atomicAdd(double* address, double val) { // NOLINT
unsigned long long int* address_as_ull =
(unsigned long long int*)address; // NOLINT
unsigned long long int old = *address_as_ull, assumed; // NOLINT
do {
assumed = old;
old = atomicCAS(address_as_ull, assumed,
__double_as_longlong(val + __longlong_as_double(assumed)));
// Note: uses integer comparison to avoid hang in case of NaN (since NaN !=
// NaN)
} while (assumed != old);
return __longlong_as_double(old);
}
#endif
namespace dh {
// FIXME(jiamingy): Remove this once we get rid of cub submodule.
constexpr bool BuildWithCUDACub() {
#if defined(THRUST_IGNORE_CUB_VERSION_CHECK) && THRUST_IGNORE_CUB_VERSION_CHECK == 1
return false;
#else
return true;
#endif // defined(THRUST_IGNORE_CUB_VERSION_CHECK) && THRUST_IGNORE_CUB_VERSION_CHECK == 1
}
namespace detail {
template <size_t size>
struct AtomicDispatcher;
template <>
struct AtomicDispatcher<sizeof(uint32_t)> {
using Type = unsigned int; // NOLINT
static_assert(sizeof(Type) == sizeof(uint32_t), "Unsigned should be of size 32 bits.");
};
template <>
struct AtomicDispatcher<sizeof(uint64_t)> {
using Type = unsigned long long; // NOLINT
static_assert(sizeof(Type) == sizeof(uint64_t), "Unsigned long long should be of size 64 bits.");
};
} // namespace detail
} // namespace dh
// atomicAdd is not defined for size_t.
template <typename T = size_t,
std::enable_if_t<std::is_same<size_t, T>::value &&
!std::is_same<size_t, unsigned long long>::value> * = // NOLINT
nullptr>
XGBOOST_DEV_INLINE T atomicAdd(T *addr, T v) { // NOLINT
using Type = typename dh::detail::AtomicDispatcher<sizeof(T)>::Type;
Type ret = ::atomicAdd(reinterpret_cast<Type *>(addr), static_cast<Type>(v));
return static_cast<T>(ret);
}
namespace dh {
#ifdef XGBOOST_USE_NCCL
#define safe_nccl(ans) ThrowOnNcclError((ans), __FILE__, __LINE__)
inline ncclResult_t ThrowOnNcclError(ncclResult_t code, const char *file,
int line) {
if (code != ncclSuccess) {
std::stringstream ss;
ss << "NCCL failure :" << ncclGetErrorString(code);
if (code == ncclUnhandledCudaError) {
// nccl usually preserves the last error so we can get more details.
auto err = cudaPeekAtLastError();
ss << " " << thrust::system_error(err, thrust::cuda_category()).what();
}
ss << " " << file << "(" << line << ")";
LOG(FATAL) << ss.str();
}
return code;
}
#endif
inline int32_t CudaGetPointerDevice(void const *ptr) {
int32_t device = -1;
cudaPointerAttributes attr;
dh::safe_cuda(cudaPointerGetAttributes(&attr, ptr));
device = attr.device;
return device;
}
inline size_t AvailableMemory(int device_idx) {
size_t device_free = 0;
size_t device_total = 0;
safe_cuda(cudaSetDevice(device_idx));
dh::safe_cuda(cudaMemGetInfo(&device_free, &device_total));
return device_free;
}
inline int32_t CurrentDevice() {
int32_t device = 0;
safe_cuda(cudaGetDevice(&device));
return device;
}
inline size_t TotalMemory(int device_idx) {
size_t device_free = 0;
size_t device_total = 0;
safe_cuda(cudaSetDevice(device_idx));
dh::safe_cuda(cudaMemGetInfo(&device_free, &device_total));
return device_total;
}
/**
* \fn inline int MaxSharedMemory(int device_idx)
*
* \brief Maximum shared memory per block on this device.
*
* \param device_idx Zero-based index of the device.
*/
inline size_t MaxSharedMemory(int device_idx) {
int max_shared_memory = 0;
dh::safe_cuda(cudaDeviceGetAttribute
(&max_shared_memory, cudaDevAttrMaxSharedMemoryPerBlock,
device_idx));
return size_t(max_shared_memory);
}
/**
* \fn inline int MaxSharedMemoryOptin(int device_idx)
*
* \brief Maximum dynamic shared memory per thread block on this device
that can be opted into when using cudaFuncSetAttribute().
*
* \param device_idx Zero-based index of the device.
*/
inline size_t MaxSharedMemoryOptin(int device_idx) {
int max_shared_memory = 0;
dh::safe_cuda(cudaDeviceGetAttribute
(&max_shared_memory, cudaDevAttrMaxSharedMemoryPerBlockOptin,
device_idx));
return size_t(max_shared_memory);
}
inline void CheckComputeCapability() {
for (int d_idx = 0; d_idx < xgboost::common::AllVisibleGPUs(); ++d_idx) {
cudaDeviceProp prop;
safe_cuda(cudaGetDeviceProperties(&prop, d_idx));
std::ostringstream oss;
oss << "CUDA Capability Major/Minor version number: " << prop.major << "."
<< prop.minor << " is insufficient. Need >=3.5";
int failed = prop.major < 3 || (prop.major == 3 && prop.minor < 5);
if (failed) LOG(WARNING) << oss.str() << " for device: " << d_idx;
}
}
XGBOOST_DEV_INLINE void AtomicOrByte(unsigned int *__restrict__ buffer,
size_t ibyte, unsigned char b) {
atomicOr(&buffer[ibyte / sizeof(unsigned int)],
static_cast<unsigned int>(b)
<< (ibyte % (sizeof(unsigned int)) * 8));
}
template <typename T>
__device__ xgboost::common::Range GridStrideRange(T begin, T end) {
begin += blockDim.x * blockIdx.x + threadIdx.x;
xgboost::common::Range r(begin, end);
r.Step(gridDim.x * blockDim.x);
return r;
}
template <typename T>
__device__ xgboost::common::Range BlockStrideRange(T begin, T end) {
begin += threadIdx.x;
xgboost::common::Range r(begin, end);
r.Step(blockDim.x);
return r;
}
// Threadblock iterates over range, filling with value. Requires all threads in
// block to be active.
template <typename IterT, typename ValueT>
__device__ void BlockFill(IterT begin, size_t n, ValueT value) {
for (auto i : BlockStrideRange(static_cast<size_t>(0), n)) {
begin[i] = value;
}
}
/*
* Kernel launcher
*/
template <typename L>
__global__ void LaunchNKernel(size_t begin, size_t end, L lambda) {
for (auto i : GridStrideRange(begin, end)) {
lambda(i);
}
}
template <typename L>
__global__ void LaunchNKernel(int device_idx, size_t begin, size_t end,
L lambda) {
for (auto i : GridStrideRange(begin, end)) {
lambda(i, device_idx);
}
}
/* \brief A wrapper around kernel launching syntax, used to guard against empty input.
*
* - nvcc fails to deduce template argument when kernel is a template accepting __device__
* function as argument. Hence functions like `LaunchN` cannot use this wrapper.
*
* - With c++ initialization list `{}` syntax, you are forced to comply with the CUDA type
* specification.
*/
class LaunchKernel {
size_t shmem_size_;
cudaStream_t stream_;
dim3 grids_;
dim3 blocks_;
public:
LaunchKernel(uint32_t _grids, uint32_t _blk, size_t _shmem=0, cudaStream_t _s=nullptr) :
grids_{_grids, 1, 1}, blocks_{_blk, 1, 1}, shmem_size_{_shmem}, stream_{_s} {}
LaunchKernel(dim3 _grids, dim3 _blk, size_t _shmem=0, cudaStream_t _s=nullptr) :
grids_{_grids}, blocks_{_blk}, shmem_size_{_shmem}, stream_{_s} {}
template <typename K, typename... Args>
void operator()(K kernel, Args... args) {
if (XGBOOST_EXPECT(grids_.x * grids_.y * grids_.z == 0, false)) {
LOG(DEBUG) << "Skipping empty CUDA kernel.";
return;
}
kernel<<<grids_, blocks_, shmem_size_, stream_>>>(args...); // NOLINT
}
};
template <int ITEMS_PER_THREAD = 8, int BLOCK_THREADS = 256, typename L>
inline void LaunchN(size_t n, cudaStream_t stream, L lambda) {
if (n == 0) {
return;
}
const int GRID_SIZE =
static_cast<int>(xgboost::common::DivRoundUp(n, ITEMS_PER_THREAD * BLOCK_THREADS));
LaunchNKernel<<<GRID_SIZE, BLOCK_THREADS, 0, stream>>>( // NOLINT
static_cast<size_t>(0), n, lambda);
}
// Default stream version
template <int ITEMS_PER_THREAD = 8, int BLOCK_THREADS = 256, typename L>
inline void LaunchN(size_t n, L lambda) {
LaunchN<ITEMS_PER_THREAD, BLOCK_THREADS>(n, nullptr, lambda);
}
template <typename Container>
void Iota(Container array) {
LaunchN(array.size(), [=] __device__(size_t i) { array[i] = i; });
}
namespace detail {
/** \brief Keeps track of global device memory allocations. Thread safe.*/
class MemoryLogger {
// Information for a single device
struct DeviceStats {
size_t currently_allocated_bytes{ 0 };
size_t peak_allocated_bytes{ 0 };
size_t num_allocations{ 0 };
size_t num_deallocations{ 0 };
std::map<void *, size_t> device_allocations;
void RegisterAllocation(void *ptr, size_t n) {
device_allocations[ptr] = n;
currently_allocated_bytes += n;
peak_allocated_bytes =
std::max(peak_allocated_bytes, currently_allocated_bytes);
num_allocations++;
CHECK_GT(num_allocations, num_deallocations);
}
void RegisterDeallocation(void *ptr, size_t n, int current_device) {
auto itr = device_allocations.find(ptr);
if (itr == device_allocations.end()) {
LOG(WARNING) << "Attempting to deallocate " << n << " bytes on device "
<< current_device << " that was never allocated ";
}
num_deallocations++;
CHECK_LE(num_deallocations, num_allocations);
currently_allocated_bytes -= itr->second;
device_allocations.erase(itr);
}
};
DeviceStats stats_;
std::mutex mutex_;
public:
void RegisterAllocation(void *ptr, size_t n) {
if (!xgboost::ConsoleLogger::ShouldLog(xgboost::ConsoleLogger::LV::kDebug)) {
return;
}
std::lock_guard<std::mutex> guard(mutex_);
int current_device;
safe_cuda(cudaGetDevice(¤t_device));
stats_.RegisterAllocation(ptr, n);
}
void RegisterDeallocation(void *ptr, size_t n) {
if (!xgboost::ConsoleLogger::ShouldLog(xgboost::ConsoleLogger::LV::kDebug)) {
return;
}
std::lock_guard<std::mutex> guard(mutex_);
int current_device;
safe_cuda(cudaGetDevice(¤t_device));
stats_.RegisterDeallocation(ptr, n, current_device);
}
size_t PeakMemory() const {
return stats_.peak_allocated_bytes;
}
size_t CurrentlyAllocatedBytes() const {
return stats_.currently_allocated_bytes;
}
void Clear()
{
stats_ = DeviceStats();
}
void Log() {
if (!xgboost::ConsoleLogger::ShouldLog(xgboost::ConsoleLogger::LV::kDebug)) {
return;
}
std::lock_guard<std::mutex> guard(mutex_);
int current_device;
safe_cuda(cudaGetDevice(¤t_device));
LOG(CONSOLE) << "======== Device " << current_device << " Memory Allocations: "
<< " ========";
LOG(CONSOLE) << "Peak memory usage: "
<< stats_.peak_allocated_bytes / 1048576 << "MiB";
LOG(CONSOLE) << "Number of allocations: " << stats_.num_allocations;
}
};
} // namespace detail
inline detail::MemoryLogger &GlobalMemoryLogger() {
static detail::MemoryLogger memory_logger;
return memory_logger;
}
// dh::DebugSyncDevice(__FILE__, __LINE__);
inline void DebugSyncDevice(std::string file="", int32_t line = -1) {
if (file != "" && line != -1) {
auto rank = rabit::GetRank();
LOG(DEBUG) << "R:" << rank << ": " << file << ":" << line;
}
safe_cuda(cudaDeviceSynchronize());
safe_cuda(cudaGetLastError());
}
namespace detail {
#if defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
template <typename T>
using XGBBaseDeviceAllocator = rmm::mr::thrust_allocator<T>;
#else // defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
template <typename T>
using XGBBaseDeviceAllocator = thrust::device_malloc_allocator<T>;
#endif // defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
inline void ThrowOOMError(std::string const& err, size_t bytes) {
auto device = CurrentDevice();
auto rank = rabit::GetRank();
std::stringstream ss;
ss << "Memory allocation error on worker " << rank << ": " << err << "\n"
<< "- Free memory: " << AvailableMemory(device) << "\n"
<< "- Requested memory: " << bytes << std::endl;
LOG(FATAL) << ss.str();
}
/**
* \brief Default memory allocator, uses cudaMalloc/Free and logs allocations if verbose.
*/
template <class T>
struct XGBDefaultDeviceAllocatorImpl : XGBBaseDeviceAllocator<T> {
using SuperT = XGBBaseDeviceAllocator<T>;
using pointer = thrust::device_ptr<T>; // NOLINT
template<typename U>
struct rebind // NOLINT
{
using other = XGBDefaultDeviceAllocatorImpl<U>; // NOLINT
};
pointer allocate(size_t n) { // NOLINT
pointer ptr;
try {
ptr = SuperT::allocate(n);
dh::safe_cuda(cudaGetLastError());
} catch (const std::exception &e) {
ThrowOOMError(e.what(), n * sizeof(T));
}
GlobalMemoryLogger().RegisterAllocation(ptr.get(), n * sizeof(T));
return ptr;
}
void deallocate(pointer ptr, size_t n) { // NOLINT
GlobalMemoryLogger().RegisterDeallocation(ptr.get(), n * sizeof(T));
SuperT::deallocate(ptr, n);
}
#if defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
XGBDefaultDeviceAllocatorImpl()
: SuperT(rmm::cuda_stream_default, rmm::mr::get_current_device_resource()) {}
#endif // defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
};
/**
* \brief Caching memory allocator, uses cub::CachingDeviceAllocator as a back-end, unless
* RMM pool allocator is enabled. Does not initialise memory on construction.
*/
template <class T>
struct XGBCachingDeviceAllocatorImpl : XGBBaseDeviceAllocator<T> {
using SuperT = XGBBaseDeviceAllocator<T>;
using pointer = thrust::device_ptr<T>; // NOLINT
template<typename U>
struct rebind // NOLINT
{
using other = XGBCachingDeviceAllocatorImpl<U>; // NOLINT
};
cub::CachingDeviceAllocator& GetGlobalCachingAllocator() {
// Configure allocator with maximum cached bin size of ~1GB and no limit on
// maximum cached bytes
static cub::CachingDeviceAllocator *allocator = new cub::CachingDeviceAllocator(2, 9, 29);
return *allocator;
}
pointer allocate(size_t n) { // NOLINT
pointer thrust_ptr;
if (use_cub_allocator_) {
T* raw_ptr{nullptr};
auto errc = GetGlobalCachingAllocator().DeviceAllocate(reinterpret_cast<void **>(&raw_ptr),
n * sizeof(T));
if (errc != cudaSuccess) {
ThrowOOMError("Caching allocator", n * sizeof(T));
}
thrust_ptr = pointer(raw_ptr);
} else {
try {
thrust_ptr = SuperT::allocate(n);
dh::safe_cuda(cudaGetLastError());
} catch (const std::exception &e) {
ThrowOOMError(e.what(), n * sizeof(T));
}
}
GlobalMemoryLogger().RegisterAllocation(thrust_ptr.get(), n * sizeof(T));
return thrust_ptr;
}
void deallocate(pointer ptr, size_t n) { // NOLINT
GlobalMemoryLogger().RegisterDeallocation(ptr.get(), n * sizeof(T));
if (use_cub_allocator_) {
GetGlobalCachingAllocator().DeviceFree(ptr.get());
} else {
SuperT::deallocate(ptr, n);
}
}
#if defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
XGBCachingDeviceAllocatorImpl()
: SuperT(rmm::cuda_stream_default, rmm::mr::get_current_device_resource()),
use_cub_allocator_(!xgboost::GlobalConfigThreadLocalStore::Get()->use_rmm) {}
#endif // defined(XGBOOST_USE_RMM) && XGBOOST_USE_RMM == 1
XGBOOST_DEVICE void construct(T *) {} // NOLINT
private:
bool use_cub_allocator_{true};
};
} // namespace detail
// Declare xgboost allocators
// Replacement of allocator with custom backend should occur here
template <typename T>
using XGBDeviceAllocator = detail::XGBDefaultDeviceAllocatorImpl<T>;
/*! Be careful that the initialization constructor is a no-op, which means calling
* `vec.resize(n)` won't initialize the memory region to 0. Instead use
* `vec.resize(n, 0)`*/
template <typename T>
using XGBCachingDeviceAllocator = detail::XGBCachingDeviceAllocatorImpl<T>;
/** \brief Specialisation of thrust device vector using custom allocator. */
template <typename T>
using device_vector = thrust::device_vector<T, XGBDeviceAllocator<T>>; // NOLINT
template <typename T>
using caching_device_vector = thrust::device_vector<T, XGBCachingDeviceAllocator<T>>; // NOLINT
// Faster to instantiate than caching_device_vector and invokes no synchronisation
// Use this where vector functionality (e.g. resize) is not required
template <typename T>
class TemporaryArray {
public:
using AllocT = XGBCachingDeviceAllocator<T>;
using value_type = T; // NOLINT
explicit TemporaryArray(size_t n) : size_(n) { ptr_ = AllocT().allocate(n); }
TemporaryArray(size_t n, T val) : size_(n) {
ptr_ = AllocT().allocate(n);
this->fill(val);
}
~TemporaryArray() { AllocT().deallocate(ptr_, this->size()); }
void fill(T val) // NOLINT
{
int device = 0;
dh::safe_cuda(cudaGetDevice(&device));
auto d_data = ptr_.get();
LaunchN(this->size(), [=] __device__(size_t idx) { d_data[idx] = val; });
}
thrust::device_ptr<T> data() { return ptr_; } // NOLINT
size_t size() { return size_; } // NOLINT
private:
thrust::device_ptr<T> ptr_;
size_t size_;
};
/**
* \brief A double buffer, useful for algorithms like sort.
*/
template <typename T>
class DoubleBuffer {
public:
cub::DoubleBuffer<T> buff;
xgboost::common::Span<T> a, b;
DoubleBuffer() = default;
template <typename VectorT>
DoubleBuffer(VectorT *v1, VectorT *v2) {
a = xgboost::common::Span<T>(v1->data().get(), v1->size());
b = xgboost::common::Span<T>(v2->data().get(), v2->size());
buff = cub::DoubleBuffer<T>(a.data(), b.data());
}
size_t Size() const {
CHECK_EQ(a.size(), b.size());
return a.size();
}
cub::DoubleBuffer<T> &CubBuffer() { return buff; }
T *Current() { return buff.Current(); }
xgboost::common::Span<T> CurrentSpan() {
return xgboost::common::Span<T>{buff.Current(), Size()};
}
T *Other() { return buff.Alternate(); }
};
/**
* \brief Copies device span to std::vector.
*
* \tparam T Generic type parameter.
* \param [in,out] dst Copy destination.
* \param src Copy source. Must be device memory.
*/
template <typename T>
void CopyDeviceSpanToVector(std::vector<T> *dst, xgboost::common::Span<T> src) {
CHECK_EQ(dst->size(), src.size());
dh::safe_cuda(cudaMemcpyAsync(dst->data(), src.data(), dst->size() * sizeof(T),
cudaMemcpyDeviceToHost));
}
/**
* \brief Copies const device span to std::vector.
*
* \tparam T Generic type parameter.
* \param [in,out] dst Copy destination.
* \param src Copy source. Must be device memory.
*/
template <typename T>
void CopyDeviceSpanToVector(std::vector<T> *dst, xgboost::common::Span<const T> src) {
CHECK_EQ(dst->size(), src.size());
dh::safe_cuda(cudaMemcpyAsync(dst->data(), src.data(), dst->size() * sizeof(T),
cudaMemcpyDeviceToHost));
}
template <class HContainer, class DContainer>
void CopyToD(HContainer const &h, DContainer *d) {
if (h.empty()) {
d->clear();
return;
}
d->resize(h.size());
using HVT = std::remove_cv_t<typename HContainer::value_type>;
using DVT = std::remove_cv_t<typename DContainer::value_type>;
static_assert(std::is_same<HVT, DVT>::value,
"Host and device containers must have same value type.");
dh::safe_cuda(cudaMemcpyAsync(d->data().get(), h.data(), h.size() * sizeof(HVT),
cudaMemcpyHostToDevice));
}
// Keep track of pinned memory allocation
struct PinnedMemory {
void *temp_storage{nullptr};
size_t temp_storage_bytes{0};
~PinnedMemory() { Free(); }
template <typename T>
xgboost::common::Span<T> GetSpan(size_t size) {
size_t num_bytes = size * sizeof(T);
if (num_bytes > temp_storage_bytes) {
Free();
safe_cuda(cudaMallocHost(&temp_storage, num_bytes));
temp_storage_bytes = num_bytes;
}
return xgboost::common::Span<T>(static_cast<T *>(temp_storage), size);
}
template <typename T>
xgboost::common::Span<T> GetSpan(size_t size, T init) {
auto result = this->GetSpan<T>(size);
for (auto &e : result) {
e = init;
}
return result;
}
void Free() {
if (temp_storage != nullptr) {
safe_cuda(cudaFreeHost(temp_storage));
}
}
};
/*
* Utility functions
*/
/**
* @brief Helper function to perform device-wide sum-reduction, returns to the
* host
* @param in the input array to be reduced
* @param nVals number of elements in the input array
*/
template <typename T>
typename std::iterator_traits<T>::value_type SumReduction(T in, int nVals) {
using ValueT = typename std::iterator_traits<T>::value_type;
size_t tmpSize {0};
ValueT *dummy_out = nullptr;
dh::safe_cuda(cub::DeviceReduce::Sum(nullptr, tmpSize, in, dummy_out, nVals));
TemporaryArray<char> temp(tmpSize + sizeof(ValueT));
auto ptr = reinterpret_cast<ValueT *>(temp.data().get()) + 1;
dh::safe_cuda(cub::DeviceReduce::Sum(
reinterpret_cast<void *>(ptr), tmpSize, in,
reinterpret_cast<ValueT *>(temp.data().get()),
nVals));
ValueT sum;
dh::safe_cuda(cudaMemcpy(&sum, temp.data().get(), sizeof(ValueT),
cudaMemcpyDeviceToHost));
return sum;
}
constexpr std::pair<int, int> CUDAVersion() {
#if defined(__CUDACC_VER_MAJOR__)
return std::make_pair(__CUDACC_VER_MAJOR__, __CUDACC_VER_MINOR__);
#else
// clang/clang-tidy
return std::make_pair((CUDA_VERSION) / 1000, (CUDA_VERSION) % 100 / 10);
#endif // defined(__CUDACC_VER_MAJOR__)
}
constexpr std::pair<int32_t, int32_t> ThrustVersion() {
return std::make_pair(THRUST_MAJOR_VERSION, THRUST_MINOR_VERSION);
}
// Whether do we have thrust 1.x with x >= minor
template <int32_t minor>
constexpr bool HasThrustMinorVer() {
return (ThrustVersion().first == 1 && ThrustVersion().second >= minor) ||
ThrustVersion().first > 1;
}
namespace detail {
template <typename T>
using TypedDiscardCTK114 = thrust::discard_iterator<T>;
template <typename T>
class TypedDiscard : public thrust::discard_iterator<T> {
public:
using value_type = T; // NOLINT
};
} // namespace detail
template <typename T>
using TypedDiscard =
std::conditional_t<HasThrustMinorVer<12>(), detail::TypedDiscardCTK114<T>,
detail::TypedDiscard<T>>;
/**
* \class AllReducer
*
* \brief All reducer class that manages its own communication group and
* streams. Must be initialised before use. If XGBoost is compiled without NCCL
* this is a dummy class that will error if used with more than one GPU.
*/
class AllReducer {
bool initialised_ {false};
size_t allreduce_bytes_ {0}; // Keep statistics of the number of bytes communicated
size_t allreduce_calls_ {0}; // Keep statistics of the number of reduce calls
#ifdef XGBOOST_USE_NCCL
ncclComm_t comm_;
cudaStream_t stream_;
int device_ordinal_;
ncclUniqueId id_;
#endif
public:
AllReducer() = default;
/**
* \brief Initialise with the desired device ordinal for this communication
* group.
*
* \param device_ordinal The device ordinal.
*/
void Init(int _device_ordinal);
~AllReducer();
/**
* \brief Allreduce. Use in exactly the same way as NCCL but without needing
* streams or comms.
*
* \param sendbuff The sendbuff.
* \param recvbuff The recvbuff.
* \param count Number of elements.
*/
void AllReduceSum(const double *sendbuff, double *recvbuff, int count) {
#ifdef XGBOOST_USE_NCCL
CHECK(initialised_);
dh::safe_cuda(cudaSetDevice(device_ordinal_));
dh::safe_nccl(ncclAllReduce(sendbuff, recvbuff, count, ncclDouble, ncclSum, comm_, stream_));
allreduce_bytes_ += count * sizeof(double);
allreduce_calls_ += 1;
#endif
}
/**
* \brief Allgather implemented as grouped calls to Broadcast. This way we can accept
* different size of data on different workers.
* \param length_bytes Size of input data in bytes.
* \param segments Size of data on each worker.
* \param recvbuf Buffer storing the result of data from all workers.
*/
void AllGather(void const* data, size_t length_bytes,
std::vector<size_t>* segments, dh::caching_device_vector<char>* recvbuf);
void AllGather(uint32_t const* data, size_t length,
dh::caching_device_vector<uint32_t>* recvbuf) {
#ifdef XGBOOST_USE_NCCL
CHECK(initialised_);
size_t world = rabit::GetWorldSize();
recvbuf->resize(length * world);
safe_nccl(ncclAllGather(data, recvbuf->data().get(), length, ncclUint32,
comm_, stream_));
#endif // XGBOOST_USE_NCCL
}
/**
* \brief Allreduce. Use in exactly the same way as NCCL but without needing
* streams or comms.
*
* \param sendbuff The sendbuff.
* \param recvbuff The recvbuff.
* \param count Number of elements.
*/
void AllReduceSum(const float *sendbuff, float *recvbuff, int count) {
#ifdef XGBOOST_USE_NCCL
CHECK(initialised_);
dh::safe_cuda(cudaSetDevice(device_ordinal_));
dh::safe_nccl(ncclAllReduce(sendbuff, recvbuff, count, ncclFloat, ncclSum, comm_, stream_));
allreduce_bytes_ += count * sizeof(float);
allreduce_calls_ += 1;
#endif
}
/**
* \brief Allreduce. Use in exactly the same way as NCCL but without needing streams or comms.
*
* \param count Number of.
*
* \param sendbuff The sendbuff.
* \param recvbuff The recvbuff.
* \param count Number of.
*/
void AllReduceSum(const int64_t *sendbuff, int64_t *recvbuff, int count) {
#ifdef XGBOOST_USE_NCCL
CHECK(initialised_);
dh::safe_cuda(cudaSetDevice(device_ordinal_));
dh::safe_nccl(ncclAllReduce(sendbuff, recvbuff, count, ncclInt64, ncclSum, comm_, stream_));
#endif
}
void AllReduceSum(const uint32_t *sendbuff, uint32_t *recvbuff, int count) {
#ifdef XGBOOST_USE_NCCL
CHECK(initialised_);
dh::safe_cuda(cudaSetDevice(device_ordinal_));
dh::safe_nccl(ncclAllReduce(sendbuff, recvbuff, count, ncclUint32, ncclSum, comm_, stream_));
#endif
}
void AllReduceSum(const uint64_t *sendbuff, uint64_t *recvbuff, int count) {
#ifdef XGBOOST_USE_NCCL
CHECK(initialised_);
dh::safe_cuda(cudaSetDevice(device_ordinal_));
dh::safe_nccl(ncclAllReduce(sendbuff, recvbuff, count, ncclUint64, ncclSum, comm_, stream_));
#endif
}
// Specialization for size_t, which is implementation defined so it might or might not
// be one of uint64_t/uint32_t/unsigned long long/unsigned long.
template <typename T = size_t,
std::enable_if_t<std::is_same<size_t, T>::value &&
!std::is_same<size_t, unsigned long long>::value> // NOLINT
* = nullptr>
void AllReduceSum(const T *sendbuff, T *recvbuff, int count) { // NOLINT
#ifdef XGBOOST_USE_NCCL
CHECK(initialised_);
dh::safe_cuda(cudaSetDevice(device_ordinal_));
static_assert(sizeof(unsigned long long) == sizeof(uint64_t), ""); // NOLINT
dh::safe_nccl(ncclAllReduce(sendbuff, recvbuff, count, ncclUint64, ncclSum, comm_, stream_));
#endif
}
/**
* \fn void Synchronize()
*
* \brief Synchronizes the entire communication group.
*/
void Synchronize() {
#ifdef XGBOOST_USE_NCCL
dh::safe_cuda(cudaSetDevice(device_ordinal_));
dh::safe_cuda(cudaStreamSynchronize(stream_));
#endif
};
#ifdef XGBOOST_USE_NCCL
/**
* \fn ncclUniqueId GetUniqueId()
*
* \brief Gets the Unique ID from NCCL to be used in setting up interprocess
* communication
*
* \return the Unique ID
*/
ncclUniqueId GetUniqueId() {
static const int kRootRank = 0;
ncclUniqueId id;
if (rabit::GetRank() == kRootRank) {
dh::safe_nccl(ncclGetUniqueId(&id));
}
rabit::Broadcast(
static_cast<void*>(&id),
sizeof(ncclUniqueId),
static_cast<int>(kRootRank));
return id;
}
#endif
};
template <typename VectorT, typename T = typename VectorT::value_type,
typename IndexT = typename xgboost::common::Span<T>::index_type>
xgboost::common::Span<T> ToSpan(
VectorT &vec,
IndexT offset = 0,
IndexT size = std::numeric_limits<size_t>::max()) {
size = size == std::numeric_limits<size_t>::max() ? vec.size() : size;
CHECK_LE(offset + size, vec.size());
return {vec.data().get() + offset, size};
}
template <typename T>
xgboost::common::Span<T> ToSpan(thrust::device_vector<T>& vec,
size_t offset, size_t size) {
return ToSpan(vec, offset, size);
}
// thrust begin, similiar to std::begin
template <typename T>
thrust::device_ptr<T> tbegin(xgboost::HostDeviceVector<T>& vector) { // NOLINT
return thrust::device_ptr<T>(vector.DevicePointer());
}
template <typename T>
thrust::device_ptr<T> tend(xgboost::HostDeviceVector<T>& vector) { // // NOLINT
return tbegin(vector) + vector.Size();
}
template <typename T>
thrust::device_ptr<T const> tcbegin(xgboost::HostDeviceVector<T> const& vector) { // NOLINT
return thrust::device_ptr<T const>(vector.ConstDevicePointer());
}
template <typename T>
thrust::device_ptr<T const> tcend(xgboost::HostDeviceVector<T> const& vector) { // NOLINT
return tcbegin(vector) + vector.Size();
}
template <typename T>
XGBOOST_DEVICE thrust::device_ptr<T> tbegin(xgboost::common::Span<T>& span) { // NOLINT
return thrust::device_ptr<T>(span.data());
}
template <typename T>
XGBOOST_DEVICE thrust::device_ptr<T> tbegin(xgboost::common::Span<T> const& span) { // NOLINT
return thrust::device_ptr<T>(span.data());
}
template <typename T>
XGBOOST_DEVICE thrust::device_ptr<T> tend(xgboost::common::Span<T>& span) { // NOLINT
return tbegin(span) + span.size();
}
template <typename T>
XGBOOST_DEVICE thrust::device_ptr<T> tend(xgboost::common::Span<T> const& span) { // NOLINT
return tbegin(span) + span.size();
}
template <typename T>
XGBOOST_DEVICE auto trbegin(xgboost::common::Span<T> &span) { // NOLINT
return thrust::make_reverse_iterator(span.data() + span.size());
}
template <typename T>
XGBOOST_DEVICE auto trend(xgboost::common::Span<T> &span) { // NOLINT
return trbegin(span) + span.size();
}
template <typename T>
XGBOOST_DEVICE thrust::device_ptr<T const> tcbegin(xgboost::common::Span<T> const& span) { // NOLINT
return thrust::device_ptr<T const>(span.data());
}
template <typename T>
XGBOOST_DEVICE thrust::device_ptr<T const> tcend(xgboost::common::Span<T> const& span) { // NOLINT
return tcbegin(span) + span.size();
}
template <typename T>
XGBOOST_DEVICE auto tcrbegin(xgboost::common::Span<T> const &span) { // NOLINT
return thrust::make_reverse_iterator(span.data() + span.size());
}
template <typename T>
XGBOOST_DEVICE auto tcrend(xgboost::common::Span<T> const &span) { // NOLINT