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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
|
/*
* Copyright 2011-2021 Blender Foundation
*
* 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.
*/
#ifdef WITH_HIP
# include <climits>
# include <limits.h>
# include <stdio.h>
# include <stdlib.h>
# include <string.h>
# include "device/hip/device_impl.h"
# include "util/debug.h"
# include "util/foreach.h"
# include "util/log.h"
# include "util/map.h"
# include "util/md5.h"
# include "util/opengl.h"
# include "util/path.h"
# include "util/string.h"
# include "util/system.h"
# include "util/time.h"
# include "util/types.h"
# include "util/windows.h"
CCL_NAMESPACE_BEGIN
class HIPDevice;
bool HIPDevice::have_precompiled_kernels()
{
string fatbins_path = path_get("lib");
return path_exists(fatbins_path);
}
BVHLayoutMask HIPDevice::get_bvh_layout_mask() const
{
return BVH_LAYOUT_BVH2;
}
void HIPDevice::set_error(const string &error)
{
Device::set_error(error);
if (first_error) {
fprintf(stderr, "\nRefer to the Cycles GPU rendering documentation for possible solutions:\n");
fprintf(stderr,
"https://docs.blender.org/manual/en/latest/render/cycles/gpu_rendering.html\n\n");
first_error = false;
}
}
HIPDevice::HIPDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler)
: Device(info, stats, profiler), texture_info(this, "__texture_info", MEM_GLOBAL)
{
first_error = true;
hipDevId = info.num;
hipDevice = 0;
hipContext = 0;
hipModule = 0;
need_texture_info = false;
device_texture_headroom = 0;
device_working_headroom = 0;
move_texture_to_host = false;
map_host_limit = 0;
map_host_used = 0;
can_map_host = 0;
pitch_alignment = 0;
/* Initialize HIP. */
hipError_t result = hipInit(0);
if (result != hipSuccess) {
set_error(string_printf("Failed to initialize HIP runtime (%s)", hipewErrorString(result)));
return;
}
/* Setup device and context. */
result = hipGetDevice(&hipDevice, hipDevId);
if (result != hipSuccess) {
set_error(string_printf("Failed to get HIP device handle from ordinal (%s)",
hipewErrorString(result)));
return;
}
/* hipDeviceMapHost for mapping host memory when out of device memory.
* hipDeviceLmemResizeToMax for reserving local memory ahead of render,
* so we can predict which memory to map to host. */
hip_assert(hipDeviceGetAttribute(&can_map_host, hipDeviceAttributeCanMapHostMemory, hipDevice));
hip_assert(
hipDeviceGetAttribute(&pitch_alignment, hipDeviceAttributeTexturePitchAlignment, hipDevice));
unsigned int ctx_flags = hipDeviceLmemResizeToMax;
if (can_map_host) {
ctx_flags |= hipDeviceMapHost;
init_host_memory();
}
/* Create context. */
result = hipCtxCreate(&hipContext, ctx_flags, hipDevice);
if (result != hipSuccess) {
set_error(string_printf("Failed to create HIP context (%s)", hipewErrorString(result)));
return;
}
int major, minor;
hipDeviceGetAttribute(&major, hipDeviceAttributeComputeCapabilityMajor, hipDevId);
hipDeviceGetAttribute(&minor, hipDeviceAttributeComputeCapabilityMinor, hipDevId);
hipDevArchitecture = major * 100 + minor * 10;
/* Pop context set by hipCtxCreate. */
hipCtxPopCurrent(NULL);
}
HIPDevice::~HIPDevice()
{
texture_info.free();
hip_assert(hipCtxDestroy(hipContext));
}
bool HIPDevice::support_device(const uint /*kernel_features*/)
{
if (hipSupportsDevice(hipDevId)) {
return true;
}
else {
/* We only support Navi and above. */
hipDeviceProp_t props;
hipGetDeviceProperties(&props, hipDevId);
set_error(string_printf("HIP backend requires AMD RDNA graphics card or up, but found %s.",
props.name));
return false;
}
}
bool HIPDevice::check_peer_access(Device *peer_device)
{
if (peer_device == this) {
return false;
}
if (peer_device->info.type != DEVICE_HIP && peer_device->info.type != DEVICE_OPTIX) {
return false;
}
HIPDevice *const peer_device_hip = static_cast<HIPDevice *>(peer_device);
int can_access = 0;
hip_assert(hipDeviceCanAccessPeer(&can_access, hipDevice, peer_device_hip->hipDevice));
if (can_access == 0) {
return false;
}
// Ensure array access over the link is possible as well (for 3D textures)
hip_assert(hipDeviceGetP2PAttribute(
&can_access, hipDevP2PAttrHipArrayAccessSupported, hipDevice, peer_device_hip->hipDevice));
if (can_access == 0) {
return false;
}
// Enable peer access in both directions
{
const HIPContextScope scope(this);
hipError_t result = hipCtxEnablePeerAccess(peer_device_hip->hipContext, 0);
if (result != hipSuccess) {
set_error(string_printf("Failed to enable peer access on HIP context (%s)",
hipewErrorString(result)));
return false;
}
}
{
const HIPContextScope scope(peer_device_hip);
hipError_t result = hipCtxEnablePeerAccess(hipContext, 0);
if (result != hipSuccess) {
set_error(string_printf("Failed to enable peer access on HIP context (%s)",
hipewErrorString(result)));
return false;
}
}
return true;
}
bool HIPDevice::use_adaptive_compilation()
{
return DebugFlags().hip.adaptive_compile;
}
/* Common HIPCC flags which stays the same regardless of shading model,
* kernel sources md5 and only depends on compiler or compilation settings.
*/
string HIPDevice::compile_kernel_get_common_cflags(const uint kernel_features)
{
const int machine = system_cpu_bits();
const string source_path = path_get("source");
const string include_path = source_path;
string cflags = string_printf(
"-m%d "
"--use_fast_math "
"-DHIPCC "
"-I\"%s\"",
machine,
include_path.c_str());
if (use_adaptive_compilation()) {
cflags += " -D__KERNEL_FEATURES__=" + to_string(kernel_features);
}
return cflags;
}
string HIPDevice::compile_kernel(const uint kernel_features, const char *name, const char *base)
{
/* Compute kernel name. */
int major, minor;
hipDeviceGetAttribute(&major, hipDeviceAttributeComputeCapabilityMajor, hipDevId);
hipDeviceGetAttribute(&minor, hipDeviceAttributeComputeCapabilityMinor, hipDevId);
hipDeviceProp_t props;
hipGetDeviceProperties(&props, hipDevId);
/* gcnArchName can contain tokens after the arch name with features, ie.
* `gfx1010:sramecc-:xnack-` so we tokenize it to get the first part. */
char *arch = strtok(props.gcnArchName, ":");
if (arch == NULL) {
arch = props.gcnArchName;
}
/* Attempt to use kernel provided with Blender. */
if (!use_adaptive_compilation()) {
const string fatbin = path_get(string_printf("lib/%s_%s.fatbin", name, arch));
VLOG(1) << "Testing for pre-compiled kernel " << fatbin << ".";
if (path_exists(fatbin)) {
VLOG(1) << "Using precompiled kernel.";
return fatbin;
}
}
/* Try to use locally compiled kernel. */
string source_path = path_get("source");
const string source_md5 = path_files_md5_hash(source_path);
/* We include cflags into md5 so changing hip toolkit or changing other
* compiler command line arguments makes sure fatbin gets re-built.
*/
string common_cflags = compile_kernel_get_common_cflags(kernel_features);
const string kernel_md5 = util_md5_string(source_md5 + common_cflags);
const char *const kernel_ext = "genco";
std::string options;
# ifdef _WIN32
options.append("Wno-parentheses-equality -Wno-unused-value --hipcc-func-supp -ffast-math");
# else
options.append("Wno-parentheses-equality -Wno-unused-value --hipcc-func-supp -O3 -ffast-math");
# endif
# ifdef _DEBUG
options.append(" -save-temps");
# endif
options.append(" --amdgpu-target=").append(arch);
const string include_path = source_path;
const string fatbin_file = string_printf("cycles_%s_%s_%s", name, arch, kernel_md5.c_str());
const string fatbin = path_cache_get(path_join("kernels", fatbin_file));
VLOG(1) << "Testing for locally compiled kernel " << fatbin << ".";
if (path_exists(fatbin)) {
VLOG(1) << "Using locally compiled kernel.";
return fatbin;
}
# ifdef _WIN32
if (!use_adaptive_compilation() && have_precompiled_kernels()) {
if (!hipSupportsDevice(hipDevId)) {
set_error(
string_printf("HIP backend requires compute capability 10.1 or up, but found %d.%d. "
"Your GPU is not supported.",
major,
minor));
}
else {
set_error(
string_printf("HIP binary kernel for this graphics card compute "
"capability (%d.%d) not found.",
major,
minor));
}
return string();
}
# endif
/* Compile. */
const char *const hipcc = hipewCompilerPath();
if (hipcc == NULL) {
set_error(
"HIP hipcc compiler not found. "
"Install HIP toolkit in default location.");
return string();
}
const int hipcc_hip_version = hipewCompilerVersion();
VLOG(1) << "Found hipcc " << hipcc << ", HIP version " << hipcc_hip_version << ".";
if (hipcc_hip_version < 40) {
printf(
"Unsupported HIP version %d.%d detected, "
"you need HIP 4.0 or newer.\n",
hipcc_hip_version / 10,
hipcc_hip_version % 10);
return string();
}
double starttime = time_dt();
path_create_directories(fatbin);
source_path = path_join(path_join(source_path, "kernel"),
path_join("device", path_join(base, string_printf("%s.cpp", name))));
string command = string_printf("%s -%s -I %s --%s %s -o \"%s\"",
hipcc,
options.c_str(),
include_path.c_str(),
kernel_ext,
source_path.c_str(),
fatbin.c_str());
printf("Compiling %sHIP kernel ...\n%s\n",
(use_adaptive_compilation()) ? "adaptive " : "",
command.c_str());
# ifdef _WIN32
command = "call " + command;
# endif
if (system(command.c_str()) != 0) {
set_error(
"Failed to execute compilation command, "
"see console for details.");
return string();
}
/* Verify if compilation succeeded */
if (!path_exists(fatbin)) {
set_error(
"HIP kernel compilation failed, "
"see console for details.");
return string();
}
printf("Kernel compilation finished in %.2lfs.\n", time_dt() - starttime);
return fatbin;
}
bool HIPDevice::load_kernels(const uint kernel_features)
{
/* TODO(sergey): Support kernels re-load for HIP devices adaptive compile.
*
* Currently re-loading kernels will invalidate memory pointers.
*/
if (hipModule) {
if (use_adaptive_compilation()) {
VLOG(1) << "Skipping HIP kernel reload for adaptive compilation, not currently supported.";
}
return true;
}
/* check if hip init succeeded */
if (hipContext == 0)
return false;
/* check if GPU is supported */
if (!support_device(kernel_features)) {
return false;
}
/* get kernel */
const char *kernel_name = "kernel";
string fatbin = compile_kernel(kernel_features, kernel_name);
if (fatbin.empty())
return false;
/* open module */
HIPContextScope scope(this);
string fatbin_data;
hipError_t result;
if (path_read_text(fatbin, fatbin_data))
result = hipModuleLoadData(&hipModule, fatbin_data.c_str());
else
result = hipErrorFileNotFound;
if (result != hipSuccess)
set_error(string_printf(
"Failed to load HIP kernel from '%s' (%s)", fatbin.c_str(), hipewErrorString(result)));
if (result == hipSuccess) {
kernels.load(this);
reserve_local_memory(kernel_features);
}
return (result == hipSuccess);
}
void HIPDevice::reserve_local_memory(const uint kernel_features)
{
/* Together with hipDeviceLmemResizeToMax, this reserves local memory
* needed for kernel launches, so that we can reliably figure out when
* to allocate scene data in mapped host memory. */
size_t total = 0, free_before = 0, free_after = 0;
{
HIPContextScope scope(this);
hipMemGetInfo(&free_before, &total);
}
{
/* Use the biggest kernel for estimation. */
const DeviceKernel test_kernel = (kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) ?
DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE :
DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE;
/* Launch kernel, using just 1 block appears sufficient to reserve memory for all
* multiprocessors. It would be good to do this in parallel for the multi GPU case
* still to make it faster. */
HIPDeviceQueue queue(this);
void *d_path_index = nullptr;
void *d_render_buffer = nullptr;
int d_work_size = 0;
void *args[] = {&d_path_index, &d_render_buffer, &d_work_size};
queue.init_execution();
queue.enqueue(test_kernel, 1, args);
queue.synchronize();
}
{
HIPContextScope scope(this);
hipMemGetInfo(&free_after, &total);
}
VLOG(1) << "Local memory reserved " << string_human_readable_number(free_before - free_after)
<< " bytes. (" << string_human_readable_size(free_before - free_after) << ")";
# if 0
/* For testing mapped host memory, fill up device memory. */
const size_t keep_mb = 1024;
while (free_after > keep_mb * 1024 * 1024LL) {
hipDeviceptr_t tmp;
hip_assert(hipMalloc(&tmp, 10 * 1024 * 1024LL));
hipMemGetInfo(&free_after, &total);
}
# endif
}
void HIPDevice::init_host_memory()
{
/* Limit amount of host mapped memory, because allocating too much can
* cause system instability. Leave at least half or 4 GB of system
* memory free, whichever is smaller. */
size_t default_limit = 4 * 1024 * 1024 * 1024LL;
size_t system_ram = system_physical_ram();
if (system_ram > 0) {
if (system_ram / 2 > default_limit) {
map_host_limit = system_ram - default_limit;
}
else {
map_host_limit = system_ram / 2;
}
}
else {
VLOG(1) << "Mapped host memory disabled, failed to get system RAM";
map_host_limit = 0;
}
/* Amount of device memory to keep is free after texture memory
* and working memory allocations respectively. We set the working
* memory limit headroom lower so that some space is left after all
* texture memory allocations. */
device_working_headroom = 32 * 1024 * 1024LL; // 32MB
device_texture_headroom = 128 * 1024 * 1024LL; // 128MB
VLOG(1) << "Mapped host memory limit set to " << string_human_readable_number(map_host_limit)
<< " bytes. (" << string_human_readable_size(map_host_limit) << ")";
}
void HIPDevice::load_texture_info()
{
if (need_texture_info) {
/* Unset flag before copying, so this does not loop indefinitely if the copy below calls
* into 'move_textures_to_host' (which calls 'load_texture_info' again). */
need_texture_info = false;
texture_info.copy_to_device();
}
}
void HIPDevice::move_textures_to_host(size_t size, bool for_texture)
{
/* Break out of recursive call, which can happen when moving memory on a multi device. */
static bool any_device_moving_textures_to_host = false;
if (any_device_moving_textures_to_host) {
return;
}
/* Signal to reallocate textures in host memory only. */
move_texture_to_host = true;
while (size > 0) {
/* Find suitable memory allocation to move. */
device_memory *max_mem = NULL;
size_t max_size = 0;
bool max_is_image = false;
thread_scoped_lock lock(hip_mem_map_mutex);
foreach (HIPMemMap::value_type &pair, hip_mem_map) {
device_memory &mem = *pair.first;
HIPMem *cmem = &pair.second;
/* Can only move textures allocated on this device (and not those from peer devices).
* And need to ignore memory that is already on the host. */
if (!mem.is_resident(this) || cmem->use_mapped_host) {
continue;
}
bool is_texture = (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) &&
(&mem != &texture_info);
bool is_image = is_texture && (mem.data_height > 1);
/* Can't move this type of memory. */
if (!is_texture || cmem->array) {
continue;
}
/* For other textures, only move image textures. */
if (for_texture && !is_image) {
continue;
}
/* Try to move largest allocation, prefer moving images. */
if (is_image > max_is_image || (is_image == max_is_image && mem.device_size > max_size)) {
max_is_image = is_image;
max_size = mem.device_size;
max_mem = &mem;
}
}
lock.unlock();
/* Move to host memory. This part is mutex protected since
* multiple HIP devices could be moving the memory. The
* first one will do it, and the rest will adopt the pointer. */
if (max_mem) {
VLOG(1) << "Move memory from device to host: " << max_mem->name;
static thread_mutex move_mutex;
thread_scoped_lock lock(move_mutex);
any_device_moving_textures_to_host = true;
/* Potentially need to call back into multi device, so pointer mapping
* and peer devices are updated. This is also necessary since the device
* pointer may just be a key here, so cannot be accessed and freed directly.
* Unfortunately it does mean that memory is reallocated on all other
* devices as well, which is potentially dangerous when still in use (since
* a thread rendering on another devices would only be caught in this mutex
* if it so happens to do an allocation at the same time as well. */
max_mem->device_copy_to();
size = (max_size >= size) ? 0 : size - max_size;
any_device_moving_textures_to_host = false;
}
else {
break;
}
}
/* Unset flag before texture info is reloaded, since it should stay in device memory. */
move_texture_to_host = false;
/* Update texture info array with new pointers. */
load_texture_info();
}
HIPDevice::HIPMem *HIPDevice::generic_alloc(device_memory &mem, size_t pitch_padding)
{
HIPContextScope scope(this);
hipDeviceptr_t device_pointer = 0;
size_t size = mem.memory_size() + pitch_padding;
hipError_t mem_alloc_result = hipErrorOutOfMemory;
const char *status = "";
/* First try allocating in device memory, respecting headroom. We make
* an exception for texture info. It is small and frequently accessed,
* so treat it as working memory.
*
* If there is not enough room for working memory, we will try to move
* textures to host memory, assuming the performance impact would have
* been worse for working memory. */
bool is_texture = (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) && (&mem != &texture_info);
bool is_image = is_texture && (mem.data_height > 1);
size_t headroom = (is_texture) ? device_texture_headroom : device_working_headroom;
size_t total = 0, free = 0;
hipMemGetInfo(&free, &total);
/* Move textures to host memory if needed. */
if (!move_texture_to_host && !is_image && (size + headroom) >= free && can_map_host) {
move_textures_to_host(size + headroom - free, is_texture);
hipMemGetInfo(&free, &total);
}
/* Allocate in device memory. */
if (!move_texture_to_host && (size + headroom) < free) {
mem_alloc_result = hipMalloc(&device_pointer, size);
if (mem_alloc_result == hipSuccess) {
status = " in device memory";
}
}
/* Fall back to mapped host memory if needed and possible. */
void *shared_pointer = 0;
if (mem_alloc_result != hipSuccess && can_map_host) {
if (mem.shared_pointer) {
/* Another device already allocated host memory. */
mem_alloc_result = hipSuccess;
shared_pointer = mem.shared_pointer;
}
else if (map_host_used + size < map_host_limit) {
/* Allocate host memory ourselves. */
mem_alloc_result = hipHostMalloc(
&shared_pointer, size, hipHostMallocMapped | hipHostMallocWriteCombined);
assert((mem_alloc_result == hipSuccess && shared_pointer != 0) ||
(mem_alloc_result != hipSuccess && shared_pointer == 0));
}
if (mem_alloc_result == hipSuccess) {
hip_assert(hipHostGetDevicePointer(&device_pointer, shared_pointer, 0));
map_host_used += size;
status = " in host memory";
}
}
if (mem_alloc_result != hipSuccess) {
status = " failed, out of device and host memory";
set_error("System is out of GPU and shared host memory");
}
if (mem.name) {
VLOG(1) << "Buffer allocate: " << mem.name << ", "
<< string_human_readable_number(mem.memory_size()) << " bytes. ("
<< string_human_readable_size(mem.memory_size()) << ")" << status;
}
mem.device_pointer = (device_ptr)device_pointer;
mem.device_size = size;
stats.mem_alloc(size);
if (!mem.device_pointer) {
return NULL;
}
/* Insert into map of allocations. */
thread_scoped_lock lock(hip_mem_map_mutex);
HIPMem *cmem = &hip_mem_map[&mem];
if (shared_pointer != 0) {
/* Replace host pointer with our host allocation. Only works if
* HIP memory layout is the same and has no pitch padding. Also
* does not work if we move textures to host during a render,
* since other devices might be using the memory. */
if (!move_texture_to_host && pitch_padding == 0 && mem.host_pointer &&
mem.host_pointer != shared_pointer) {
memcpy(shared_pointer, mem.host_pointer, size);
/* A Call to device_memory::host_free() should be preceded by
* a call to device_memory::device_free() for host memory
* allocated by a device to be handled properly. Two exceptions
* are here and a call in OptiXDevice::generic_alloc(), where
* the current host memory can be assumed to be allocated by
* device_memory::host_alloc(), not by a device */
mem.host_free();
mem.host_pointer = shared_pointer;
}
mem.shared_pointer = shared_pointer;
mem.shared_counter++;
cmem->use_mapped_host = true;
}
else {
cmem->use_mapped_host = false;
}
return cmem;
}
void HIPDevice::generic_copy_to(device_memory &mem)
{
if (!mem.host_pointer || !mem.device_pointer) {
return;
}
/* If use_mapped_host of mem is false, the current device only uses device memory allocated by
* hipMalloc regardless of mem.host_pointer and mem.shared_pointer, and should copy data from
* mem.host_pointer. */
thread_scoped_lock lock(hip_mem_map_mutex);
if (!hip_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
const HIPContextScope scope(this);
hip_assert(
hipMemcpyHtoD((hipDeviceptr_t)mem.device_pointer, mem.host_pointer, mem.memory_size()));
}
}
void HIPDevice::generic_free(device_memory &mem)
{
if (mem.device_pointer) {
HIPContextScope scope(this);
thread_scoped_lock lock(hip_mem_map_mutex);
DCHECK(hip_mem_map.find(&mem) != hip_mem_map.end());
const HIPMem &cmem = hip_mem_map[&mem];
/* If cmem.use_mapped_host is true, reference counting is used
* to safely free a mapped host memory. */
if (cmem.use_mapped_host) {
assert(mem.shared_pointer);
if (mem.shared_pointer) {
assert(mem.shared_counter > 0);
if (--mem.shared_counter == 0) {
if (mem.host_pointer == mem.shared_pointer) {
mem.host_pointer = 0;
}
hipHostFree(mem.shared_pointer);
mem.shared_pointer = 0;
}
}
map_host_used -= mem.device_size;
}
else {
/* Free device memory. */
hip_assert(hipFree(mem.device_pointer));
}
stats.mem_free(mem.device_size);
mem.device_pointer = 0;
mem.device_size = 0;
hip_mem_map.erase(hip_mem_map.find(&mem));
}
}
void HIPDevice::mem_alloc(device_memory &mem)
{
if (mem.type == MEM_TEXTURE) {
assert(!"mem_alloc not supported for textures.");
}
else if (mem.type == MEM_GLOBAL) {
assert(!"mem_alloc not supported for global memory.");
}
else {
generic_alloc(mem);
}
}
void HIPDevice::mem_copy_to(device_memory &mem)
{
if (mem.type == MEM_GLOBAL) {
global_free(mem);
global_alloc(mem);
}
else if (mem.type == MEM_TEXTURE) {
tex_free((device_texture &)mem);
tex_alloc((device_texture &)mem);
}
else {
if (!mem.device_pointer) {
generic_alloc(mem);
}
generic_copy_to(mem);
}
}
void HIPDevice::mem_copy_from(device_memory &mem, size_t y, size_t w, size_t h, size_t elem)
{
if (mem.type == MEM_TEXTURE || mem.type == MEM_GLOBAL) {
assert(!"mem_copy_from not supported for textures.");
}
else if (mem.host_pointer) {
const size_t size = elem * w * h;
const size_t offset = elem * y * w;
if (mem.device_pointer) {
const HIPContextScope scope(this);
hip_assert(hipMemcpyDtoH(
(char *)mem.host_pointer + offset, (hipDeviceptr_t)mem.device_pointer + offset, size));
}
else {
memset((char *)mem.host_pointer + offset, 0, size);
}
}
}
void HIPDevice::mem_zero(device_memory &mem)
{
if (!mem.device_pointer) {
mem_alloc(mem);
}
if (!mem.device_pointer) {
return;
}
/* If use_mapped_host of mem is false, mem.device_pointer currently refers to device memory
* regardless of mem.host_pointer and mem.shared_pointer. */
thread_scoped_lock lock(hip_mem_map_mutex);
if (!hip_mem_map[&mem].use_mapped_host || mem.host_pointer != mem.shared_pointer) {
const HIPContextScope scope(this);
hip_assert(hipMemsetD8((hipDeviceptr_t)mem.device_pointer, 0, mem.memory_size()));
}
else if (mem.host_pointer) {
memset(mem.host_pointer, 0, mem.memory_size());
}
}
void HIPDevice::mem_free(device_memory &mem)
{
if (mem.type == MEM_GLOBAL) {
global_free(mem);
}
else if (mem.type == MEM_TEXTURE) {
tex_free((device_texture &)mem);
}
else {
generic_free(mem);
}
}
device_ptr HIPDevice::mem_alloc_sub_ptr(device_memory &mem, size_t offset, size_t /*size*/)
{
return (device_ptr)(((char *)mem.device_pointer) + mem.memory_elements_size(offset));
}
void HIPDevice::const_copy_to(const char *name, void *host, size_t size)
{
HIPContextScope scope(this);
hipDeviceptr_t mem;
size_t bytes;
hip_assert(hipModuleGetGlobal(&mem, &bytes, hipModule, name));
hip_assert(hipMemcpyHtoD(mem, host, size));
}
void HIPDevice::global_alloc(device_memory &mem)
{
if (mem.is_resident(this)) {
generic_alloc(mem);
generic_copy_to(mem);
}
const_copy_to(mem.name, &mem.device_pointer, sizeof(mem.device_pointer));
}
void HIPDevice::global_free(device_memory &mem)
{
if (mem.is_resident(this) && mem.device_pointer) {
generic_free(mem);
}
}
void HIPDevice::tex_alloc(device_texture &mem)
{
HIPContextScope scope(this);
string bind_name = mem.name;
size_t dsize = datatype_size(mem.data_type);
size_t size = mem.memory_size();
hipTextureAddressMode address_mode = hipAddressModeWrap;
switch (mem.info.extension) {
case EXTENSION_REPEAT:
address_mode = hipAddressModeWrap;
break;
case EXTENSION_EXTEND:
address_mode = hipAddressModeClamp;
break;
case EXTENSION_CLIP:
// TODO : (Arya) setting this to Mode Clamp instead of Mode Border because it's unsupported
// in hip
address_mode = hipAddressModeClamp;
break;
default:
assert(0);
break;
}
hipTextureFilterMode filter_mode;
if (mem.info.interpolation == INTERPOLATION_CLOSEST) {
filter_mode = hipFilterModePoint;
}
else {
filter_mode = hipFilterModeLinear;
}
/* Image Texture Storage */
hipArray_Format format;
switch (mem.data_type) {
case TYPE_UCHAR:
format = HIP_AD_FORMAT_UNSIGNED_INT8;
break;
case TYPE_UINT16:
format = HIP_AD_FORMAT_UNSIGNED_INT16;
break;
case TYPE_UINT:
format = HIP_AD_FORMAT_UNSIGNED_INT32;
break;
case TYPE_INT:
format = HIP_AD_FORMAT_SIGNED_INT32;
break;
case TYPE_FLOAT:
format = HIP_AD_FORMAT_FLOAT;
break;
case TYPE_HALF:
format = HIP_AD_FORMAT_HALF;
break;
default:
assert(0);
return;
}
HIPMem *cmem = NULL;
hArray array_3d = NULL;
size_t src_pitch = mem.data_width * dsize * mem.data_elements;
size_t dst_pitch = src_pitch;
if (!mem.is_resident(this)) {
thread_scoped_lock lock(hip_mem_map_mutex);
cmem = &hip_mem_map[&mem];
cmem->texobject = 0;
if (mem.data_depth > 1) {
array_3d = (hArray)mem.device_pointer;
cmem->array = array_3d;
}
else if (mem.data_height > 0) {
dst_pitch = align_up(src_pitch, pitch_alignment);
}
}
else if (mem.data_depth > 1) {
/* 3D texture using array, there is no API for linear memory. */
HIP_ARRAY3D_DESCRIPTOR desc;
desc.Width = mem.data_width;
desc.Height = mem.data_height;
desc.Depth = mem.data_depth;
desc.Format = format;
desc.NumChannels = mem.data_elements;
desc.Flags = 0;
VLOG(1) << "Array 3D allocate: " << mem.name << ", "
<< string_human_readable_number(mem.memory_size()) << " bytes. ("
<< string_human_readable_size(mem.memory_size()) << ")";
hip_assert(hipArray3DCreate((hArray *)&array_3d, &desc));
if (!array_3d) {
return;
}
HIP_MEMCPY3D param;
memset(¶m, 0, sizeof(HIP_MEMCPY3D));
param.dstMemoryType = hipMemoryTypeArray;
param.dstArray = array_3d;
param.srcMemoryType = hipMemoryTypeHost;
param.srcHost = mem.host_pointer;
param.srcPitch = src_pitch;
param.WidthInBytes = param.srcPitch;
param.Height = mem.data_height;
param.Depth = mem.data_depth;
hip_assert(hipDrvMemcpy3D(¶m));
mem.device_pointer = (device_ptr)array_3d;
mem.device_size = size;
stats.mem_alloc(size);
thread_scoped_lock lock(hip_mem_map_mutex);
cmem = &hip_mem_map[&mem];
cmem->texobject = 0;
cmem->array = array_3d;
}
else if (mem.data_height > 0) {
/* 2D texture, using pitch aligned linear memory. */
dst_pitch = align_up(src_pitch, pitch_alignment);
size_t dst_size = dst_pitch * mem.data_height;
cmem = generic_alloc(mem, dst_size - mem.memory_size());
if (!cmem) {
return;
}
hip_Memcpy2D param;
memset(¶m, 0, sizeof(param));
param.dstMemoryType = hipMemoryTypeDevice;
param.dstDevice = mem.device_pointer;
param.dstPitch = dst_pitch;
param.srcMemoryType = hipMemoryTypeHost;
param.srcHost = mem.host_pointer;
param.srcPitch = src_pitch;
param.WidthInBytes = param.srcPitch;
param.Height = mem.data_height;
hip_assert(hipDrvMemcpy2DUnaligned(¶m));
}
else {
/* 1D texture, using linear memory. */
cmem = generic_alloc(mem);
if (!cmem) {
return;
}
hip_assert(hipMemcpyHtoD(mem.device_pointer, mem.host_pointer, size));
}
/* Resize once */
const uint slot = mem.slot;
if (slot >= texture_info.size()) {
/* Allocate some slots in advance, to reduce amount
* of re-allocations. */
texture_info.resize(slot + 128);
}
/* Set Mapping and tag that we need to (re-)upload to device */
texture_info[slot] = mem.info;
need_texture_info = true;
if (mem.info.data_type != IMAGE_DATA_TYPE_NANOVDB_FLOAT &&
mem.info.data_type != IMAGE_DATA_TYPE_NANOVDB_FLOAT3) {
/* Bindless textures. */
hipResourceDesc resDesc;
memset(&resDesc, 0, sizeof(resDesc));
if (array_3d) {
resDesc.resType = hipResourceTypeArray;
resDesc.res.array.h_Array = array_3d;
resDesc.flags = 0;
}
else if (mem.data_height > 0) {
resDesc.resType = hipResourceTypePitch2D;
resDesc.res.pitch2D.devPtr = mem.device_pointer;
resDesc.res.pitch2D.format = format;
resDesc.res.pitch2D.numChannels = mem.data_elements;
resDesc.res.pitch2D.height = mem.data_height;
resDesc.res.pitch2D.width = mem.data_width;
resDesc.res.pitch2D.pitchInBytes = dst_pitch;
}
else {
resDesc.resType = hipResourceTypeLinear;
resDesc.res.linear.devPtr = mem.device_pointer;
resDesc.res.linear.format = format;
resDesc.res.linear.numChannels = mem.data_elements;
resDesc.res.linear.sizeInBytes = mem.device_size;
}
hipTextureDesc texDesc;
memset(&texDesc, 0, sizeof(texDesc));
texDesc.addressMode[0] = address_mode;
texDesc.addressMode[1] = address_mode;
texDesc.addressMode[2] = address_mode;
texDesc.filterMode = filter_mode;
texDesc.flags = HIP_TRSF_NORMALIZED_COORDINATES;
thread_scoped_lock lock(hip_mem_map_mutex);
cmem = &hip_mem_map[&mem];
hip_assert(hipTexObjectCreate(&cmem->texobject, &resDesc, &texDesc, NULL));
texture_info[slot].data = (uint64_t)cmem->texobject;
}
else {
texture_info[slot].data = (uint64_t)mem.device_pointer;
}
}
void HIPDevice::tex_free(device_texture &mem)
{
if (mem.device_pointer) {
HIPContextScope scope(this);
thread_scoped_lock lock(hip_mem_map_mutex);
DCHECK(hip_mem_map.find(&mem) != hip_mem_map.end());
const HIPMem &cmem = hip_mem_map[&mem];
if (cmem.texobject) {
/* Free bindless texture. */
hipTexObjectDestroy(cmem.texobject);
}
if (!mem.is_resident(this)) {
/* Do not free memory here, since it was allocated on a different device. */
hip_mem_map.erase(hip_mem_map.find(&mem));
}
else if (cmem.array) {
/* Free array. */
hipArrayDestroy(cmem.array);
stats.mem_free(mem.device_size);
mem.device_pointer = 0;
mem.device_size = 0;
hip_mem_map.erase(hip_mem_map.find(&mem));
}
else {
lock.unlock();
generic_free(mem);
}
}
}
unique_ptr<DeviceQueue> HIPDevice::gpu_queue_create()
{
return make_unique<HIPDeviceQueue>(this);
}
bool HIPDevice::should_use_graphics_interop()
{
/* Check whether this device is part of OpenGL context.
*
* Using HIP device for graphics interoperability which is not part of the OpenGL context is
* possible, but from the empiric measurements it can be considerably slower than using naive
* pixels copy. */
/* Disable graphics interop for now, because of driver bug in 21.40. See T92972 */
# if 0
HIPContextScope scope(this);
int num_all_devices = 0;
hip_assert(hipGetDeviceCount(&num_all_devices));
if (num_all_devices == 0) {
return false;
}
vector<hipDevice_t> gl_devices(num_all_devices);
uint num_gl_devices = 0;
hipGLGetDevices(&num_gl_devices, gl_devices.data(), num_all_devices, hipGLDeviceListAll);
for (hipDevice_t gl_device : gl_devices) {
if (gl_device == hipDevice) {
return true;
}
}
# endif
return false;
}
int HIPDevice::get_num_multiprocessors()
{
return get_device_default_attribute(hipDeviceAttributeMultiprocessorCount, 0);
}
int HIPDevice::get_max_num_threads_per_multiprocessor()
{
return get_device_default_attribute(hipDeviceAttributeMaxThreadsPerMultiProcessor, 0);
}
bool HIPDevice::get_device_attribute(hipDeviceAttribute_t attribute, int *value)
{
HIPContextScope scope(this);
return hipDeviceGetAttribute(value, attribute, hipDevice) == hipSuccess;
}
int HIPDevice::get_device_default_attribute(hipDeviceAttribute_t attribute, int default_value)
{
int value = 0;
if (!get_device_attribute(attribute, &value)) {
return default_value;
}
return value;
}
CCL_NAMESPACE_END
#endif
|