-
Notifications
You must be signed in to change notification settings - Fork 2.3k
/
llvm_program.cpp
568 lines (488 loc) · 20.2 KB
/
llvm_program.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
#include "llvm_program.h"
#include "taichi/backends/cuda/cuda_driver.h"
#include "taichi/program/arch.h"
#include "taichi/platform/cuda/detect_cuda.h"
#include "taichi/math/arithmetic.h"
#include "taichi/runtime/llvm/mem_request.h"
#include "taichi/util/str.h"
#include "taichi/codegen/codegen.h"
#include "taichi/ir/statements.h"
#include "taichi/backends/cpu/cpu_device.h"
#include "taichi/backends/cuda/cuda_device.h"
#include "taichi/backends/cuda/cuda_device.h"
#if defined(TI_WITH_CUDA)
#include "taichi/backends/cuda/cuda_driver.h"
#include "taichi/backends/cuda/codegen_cuda.h"
#include "taichi/backends/cuda/cuda_context.h"
#endif
namespace taichi {
namespace lang {
namespace {
void assert_failed_host(const char *msg) {
TI_ERROR("Assertion failure: {}", msg);
}
void *taichi_allocate_aligned(MemoryPool *memory_pool,
std::size_t size,
std::size_t alignment) {
return memory_pool->allocate(size, alignment);
}
} // namespace
LlvmProgramImpl::LlvmProgramImpl(CompileConfig &config_,
KernelProfilerBase *profiler)
: ProgramImpl(config_) {
runtime_mem_info = Runtime::create(config_.arch);
if (config_.arch == Arch::cuda) {
if (!runtime_mem_info) {
TI_WARN("Taichi is not compiled with CUDA.");
config_.arch = host_arch();
} else if (!is_cuda_api_available()) {
TI_WARN("No CUDA driver API detected.");
config_.arch = host_arch();
} else if (!runtime_mem_info->detected()) {
TI_WARN("No CUDA device detected.");
config_.arch = host_arch();
} else {
// CUDA runtime created successfully
}
if (config_.arch != Arch::cuda) {
TI_WARN("Falling back to {}.", arch_name(host_arch()));
}
}
snode_tree_buffer_manager = std::make_unique<SNodeTreeBufferManager>(this);
thread_pool = std::make_unique<ThreadPool>(config->cpu_max_num_threads);
preallocated_device_buffer = nullptr;
llvm_runtime = nullptr;
llvm_context_host = std::make_unique<TaichiLLVMContext>(host_arch());
if (config_.arch == Arch::cuda) {
#if defined(TI_WITH_CUDA)
int num_SMs;
CUDADriver::get_instance().device_get_attribute(
&num_SMs, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, nullptr);
int query_max_block_dim;
CUDADriver::get_instance().device_get_attribute(
&query_max_block_dim, CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X, nullptr);
if (config_.max_block_dim == 0) {
config_.max_block_dim = query_max_block_dim;
}
if (config_.saturating_grid_dim == 0) {
// each SM can have 16-32 resident blocks
config_.saturating_grid_dim = num_SMs * 32;
}
#endif
}
if (arch_is_cpu(config->arch)) {
config_.max_block_dim = 1024;
device_ = std::make_unique<cpu::CpuDevice>();
}
if (config->kernel_profiler && runtime_mem_info) {
runtime_mem_info->set_profiler(profiler);
}
#if defined(TI_WITH_CUDA)
if (config_.arch == Arch::cuda) {
if (config_.kernel_profiler) {
CUDAContext::get_instance().set_profiler(profiler);
} else {
CUDAContext::get_instance().set_profiler(nullptr);
}
CUDAContext::get_instance().set_debug(config->debug);
device_ = std::make_unique<cuda::CudaDevice>();
}
#endif
}
void LlvmProgramImpl::initialize_host() {
// Note this cannot be placed inside LlvmProgramImpl constructor, see doc
// string for init_runtime_jit_module() for more details.
llvm_context_host->init_runtime_jit_module();
}
void LlvmProgramImpl::maybe_initialize_cuda_llvm_context() {
if (config->arch == Arch::cuda && llvm_context_device == nullptr) {
llvm_context_device = std::make_unique<TaichiLLVMContext>(Arch::cuda);
llvm_context_device->init_runtime_jit_module();
}
}
FunctionType LlvmProgramImpl::compile(Kernel *kernel,
OffloadedStmt *offloaded) {
if (!kernel->lowered()) {
kernel->lower();
}
auto codegen = KernelCodeGen::create(kernel->arch, kernel, offloaded);
return codegen->codegen();
}
void LlvmProgramImpl::synchronize() {
if (config->arch == Arch::cuda) {
#if defined(TI_WITH_CUDA)
CUDADriver::get_instance().stream_synchronize(nullptr);
#else
TI_ERROR("No CUDA support");
#endif
}
}
std::unique_ptr<llvm::Module>
LlvmProgramImpl::clone_struct_compiler_initial_context(
const std::vector<std::unique_ptr<SNodeTree>> &snode_trees_,
TaichiLLVMContext *tlctx) {
if (!snode_trees_.empty())
return tlctx->clone_struct_module();
return tlctx->clone_runtime_module();
}
void LlvmProgramImpl::initialize_llvm_runtime_snodes(const SNodeTree *tree,
StructCompiler *scomp,
uint64 *result_buffer) {
TaichiLLVMContext *tlctx = nullptr;
if (config->arch == Arch::cuda) {
#if defined(TI_WITH_CUDA)
tlctx = llvm_context_device.get();
#else
TI_NOT_IMPLEMENTED
#endif
} else {
tlctx = llvm_context_host.get();
}
auto *const runtime_jit = tlctx->runtime_jit_module;
// By the time this creator is called, "this" is already destroyed.
// Therefore it is necessary to capture members by values.
const auto snodes = scomp->snodes;
const int root_id = tree->root()->id;
TI_TRACE("Allocating data structure of size {} bytes", scomp->root_size);
std::size_t rounded_size =
taichi::iroundup(scomp->root_size, taichi_page_size);
Ptr root_buffer = snode_tree_buffer_manager->allocate(
runtime_jit, llvm_runtime, rounded_size, taichi_page_size, tree->id(),
result_buffer);
DeviceAllocation alloc{kDeviceNullAllocation};
if (config->arch == Arch::cuda) {
#if defined(TI_WITH_CUDA)
alloc = cuda_device()->import_memory(root_buffer, rounded_size);
#else
TI_NOT_IMPLEMENTED
#endif
} else {
alloc = cpu_device()->import_memory(root_buffer, rounded_size);
}
snode_tree_allocs_[tree->id()] = alloc;
bool all_dense = config->demote_dense_struct_fors;
for (int i = 0; i < (int)snodes.size(); i++) {
if (snodes[i]->type != SNodeType::dense &&
snodes[i]->type != SNodeType::place &&
snodes[i]->type != SNodeType::root) {
all_dense = false;
break;
}
}
runtime_jit->call<void *, std::size_t, int, int, int, std::size_t, Ptr>(
"runtime_initialize_snodes", llvm_runtime, scomp->root_size, root_id,
(int)snodes.size(), tree->id(), rounded_size, root_buffer, all_dense);
for (int i = 0; i < (int)snodes.size(); i++) {
if (is_gc_able(snodes[i]->type)) {
const auto snode_id = snodes[i]->id;
std::size_t node_size;
auto element_size = snodes[i]->cell_size_bytes;
if (snodes[i]->type == SNodeType::pointer) {
// pointer. Allocators are for single elements
node_size = element_size;
} else {
// dynamic. Allocators are for the chunks
node_size = sizeof(void *) + element_size * snodes[i]->chunk_size;
}
TI_TRACE("Initializing allocator for snode {} (node size {})", snode_id,
node_size);
auto rt = llvm_runtime;
runtime_jit->call<void *, int, std::size_t>(
"runtime_NodeAllocator_initialize", rt, snode_id, node_size);
TI_TRACE("Allocating ambient element for snode {} (node size {})",
snode_id, node_size);
runtime_jit->call<void *, int>("runtime_allocate_ambient", rt, snode_id,
node_size);
}
}
}
void LlvmProgramImpl::compile_snode_tree_types(
SNodeTree *tree,
std::vector<std::unique_ptr<SNodeTree>> &snode_trees) {
auto *const root = tree->root();
if (arch_is_cpu(config->arch)) {
auto host_module = clone_struct_compiler_initial_context(
snode_trees, llvm_context_host.get());
struct_compiler_ = std::make_unique<StructCompilerLLVM>(
host_arch(), this, std::move(host_module));
} else {
TI_ASSERT(config->arch == Arch::cuda);
auto device_module = clone_struct_compiler_initial_context(
snode_trees, llvm_context_device.get());
struct_compiler_ = std::make_unique<StructCompilerLLVM>(
Arch::cuda, this, std::move(device_module));
}
struct_compiler_->run(*root);
}
void LlvmProgramImpl::materialize_snode_tree(
SNodeTree *tree,
std::vector<std::unique_ptr<SNodeTree>> &snode_trees_,
uint64 *result_buffer) {
compile_snode_tree_types(tree, snode_trees_);
initialize_llvm_runtime_snodes(tree, struct_compiler_.get(), result_buffer);
}
uint64 LlvmProgramImpl::fetch_result_uint64(int i, uint64 *result_buffer) {
// TODO: We are likely doing more synchronization than necessary. Simplify the
// sync logic when we fetch the result.
synchronize();
uint64 ret;
auto arch = config->arch;
if (arch == Arch::cuda) {
#if defined(TI_WITH_CUDA)
CUDADriver::get_instance().memcpy_device_to_host(&ret, result_buffer + i,
sizeof(uint64));
#else
TI_NOT_IMPLEMENTED;
#endif
} else {
ret = result_buffer[i];
}
return ret;
}
std::size_t LlvmProgramImpl::get_snode_num_dynamically_allocated(
SNode *snode,
uint64 *result_buffer) {
TI_ASSERT(arch_uses_llvm(config->arch));
auto node_allocator =
runtime_query<void *>("LLVMRuntime_get_node_allocators", result_buffer,
llvm_runtime, snode->id);
auto data_list = runtime_query<void *>("NodeManager_get_data_list",
result_buffer, node_allocator);
return (std::size_t)runtime_query<int32>("ListManager_get_num_elements",
result_buffer, data_list);
}
void LlvmProgramImpl::print_list_manager_info(void *list_manager,
uint64 *result_buffer) {
auto list_manager_len = runtime_query<int32>("ListManager_get_num_elements",
result_buffer, list_manager);
auto element_size = runtime_query<int32>("ListManager_get_element_size",
result_buffer, list_manager);
auto elements_per_chunk =
runtime_query<int32>("ListManager_get_max_num_elements_per_chunk",
result_buffer, list_manager);
auto num_active_chunks = runtime_query<int32>(
"ListManager_get_num_active_chunks", result_buffer, list_manager);
auto size_MB = 1e-6f * num_active_chunks * elements_per_chunk * element_size;
fmt::print(
" length={:n} {:n} chunks x [{:n} x {:n} B] total={:.4f} MB\n",
list_manager_len, num_active_chunks, elements_per_chunk, element_size,
size_MB);
}
void LlvmProgramImpl::materialize_runtime(MemoryPool *memory_pool,
KernelProfilerBase *profiler,
uint64 **result_buffer_ptr) {
maybe_initialize_cuda_llvm_context();
std::size_t prealloc_size = 0;
TaichiLLVMContext *tlctx = nullptr;
if (config->arch == Arch::cuda) {
#if defined(TI_WITH_CUDA)
CUDADriver::get_instance().malloc(
(void **)result_buffer_ptr,
sizeof(uint64) * taichi_result_buffer_entries);
const auto total_mem = runtime_mem_info->get_total_memory();
if (config->device_memory_fraction == 0) {
TI_ASSERT(config->device_memory_GB > 0);
prealloc_size = std::size_t(config->device_memory_GB * (1UL << 30));
} else {
prealloc_size = std::size_t(config->device_memory_fraction * total_mem);
}
TI_ASSERT(prealloc_size <= total_mem);
TI_TRACE("Allocating device memory {:.2f} GB",
1.0 * prealloc_size / (1UL << 30));
Device::AllocParams preallocated_device_buffer_alloc_params;
preallocated_device_buffer_alloc_params.size = prealloc_size;
preallocated_device_buffer_alloc =
cuda_device()->allocate_memory(preallocated_device_buffer_alloc_params);
cuda::CudaDevice::AllocInfo preallocated_device_buffer_alloc_info =
cuda_device()->get_alloc_info(preallocated_device_buffer_alloc);
preallocated_device_buffer = preallocated_device_buffer_alloc_info.ptr;
CUDADriver::get_instance().memset(preallocated_device_buffer, 0,
prealloc_size);
tlctx = llvm_context_device.get();
#else
TI_NOT_IMPLEMENTED
#endif
} else {
*result_buffer_ptr = (uint64 *)memory_pool->allocate(
sizeof(uint64) * taichi_result_buffer_entries, 8);
tlctx = llvm_context_host.get();
}
auto *const runtime_jit = tlctx->runtime_jit_module;
// Starting random state for the program calculated using the random seed.
// The seed is multiplied by 2^20 so that two programs with different seeds
// will not have overlapping random states in any thread.
int starting_rand_state = config->random_seed * 1048576;
// Number of random states. One per CPU/CUDA thread.
int num_rand_states = 0;
if (config->arch == Arch::cuda) {
#if defined(TI_WITH_CUDA)
// It is important to make sure that every CUDA thread has its own random
// state so that we do not need expensive per-state locks.
num_rand_states = config->saturating_grid_dim * config->max_block_dim;
#else
TI_NOT_IMPLEMENTED
#endif
} else {
num_rand_states = config->cpu_max_num_threads;
}
TI_TRACE("Allocating {} random states (used by CUDA only)", num_rand_states);
runtime_jit->call<void *, void *, std::size_t, void *, int, int, void *,
void *, void *>(
"runtime_initialize", *result_buffer_ptr, memory_pool, prealloc_size,
preallocated_device_buffer, starting_rand_state, num_rand_states,
(void *)&taichi_allocate_aligned, (void *)std::printf,
(void *)std::vsnprintf);
TI_TRACE("LLVMRuntime initialized (excluding `root`)");
llvm_runtime = fetch_result<void *>(taichi_result_buffer_ret_value_id,
*result_buffer_ptr);
TI_TRACE("LLVMRuntime pointer fetched");
if (arch_use_host_memory(config->arch)) {
runtime_jit->call<void *>("runtime_get_mem_req_queue", llvm_runtime);
auto mem_req_queue = fetch_result<void *>(taichi_result_buffer_ret_value_id,
*result_buffer_ptr);
memory_pool->set_queue((MemRequestQueue *)mem_req_queue);
}
if (arch_use_host_memory(config->arch)) {
runtime_jit->call<void *, void *, void *>(
"LLVMRuntime_initialize_thread_pool", llvm_runtime, thread_pool.get(),
(void *)ThreadPool::static_run);
runtime_jit->call<void *, void *>("LLVMRuntime_set_assert_failed",
llvm_runtime, (void *)assert_failed_host);
}
if (arch_is_cpu(config->arch)) {
// Profiler functions can only be called on CPU kernels
runtime_jit->call<void *, void *>("LLVMRuntime_set_profiler", llvm_runtime,
profiler);
runtime_jit->call<void *, void *>(
"LLVMRuntime_set_profiler_start", llvm_runtime,
(void *)&KernelProfilerBase::profiler_start);
runtime_jit->call<void *, void *>(
"LLVMRuntime_set_profiler_stop", llvm_runtime,
(void *)&KernelProfilerBase::profiler_stop);
}
}
void LlvmProgramImpl::check_runtime_error(uint64 *result_buffer) {
synchronize();
auto tlctx = llvm_context_host.get();
if (llvm_context_device) {
// In case there is a standalone device context (e.g. CUDA without unified
// memory), use the device context instead.
tlctx = llvm_context_device.get();
}
auto *runtime_jit_module = tlctx->runtime_jit_module;
runtime_jit_module->call<void *>("runtime_retrieve_and_reset_error_code",
llvm_runtime);
auto error_code =
fetch_result<int64>(taichi_result_buffer_error_id, result_buffer);
if (error_code) {
std::string error_message_template;
// Here we fetch the error_message_template char by char.
// This is not efficient, but fortunately we only need to do this when an
// assertion fails. Note that we may not have unified memory here, so using
// "fetch_result" that works across device/host memroy is necessary.
for (int i = 0;; i++) {
runtime_jit_module->call<void *>("runtime_retrieve_error_message",
llvm_runtime, i);
auto c = fetch_result<char>(taichi_result_buffer_error_id, result_buffer);
error_message_template += c;
if (c == '\0') {
break;
}
}
if (error_code == 1) {
const auto error_message_formatted = format_error_message(
error_message_template,
[runtime_jit_module, result_buffer, this](int argument_id) {
runtime_jit_module->call<void *>(
"runtime_retrieve_error_message_argument", llvm_runtime,
argument_id);
return fetch_result<uint64>(taichi_result_buffer_error_id,
result_buffer);
});
TI_ERROR("Assertion failure: {}", error_message_formatted);
} else {
TI_NOT_IMPLEMENTED
}
}
}
void LlvmProgramImpl::finalize() {
if (runtime_mem_info)
runtime_mem_info->set_profiler(nullptr);
#if defined(TI_WITH_CUDA)
if (preallocated_device_buffer != nullptr) {
cuda_device()->dealloc_memory(preallocated_device_buffer_alloc);
}
#endif
}
void LlvmProgramImpl::print_memory_profiler_info(
std::vector<std::unique_ptr<SNodeTree>> &snode_trees_,
uint64 *result_buffer) {
TI_ASSERT(arch_uses_llvm(config->arch));
fmt::print("\n[Memory Profiler]\n");
std::locale::global(std::locale("en_US.UTF-8"));
// So that thousand separators are added to "{:n}" slots in fmtlib.
// E.g., 10000 is printed as "10,000".
// TODO: is there a way to set locale only locally in this function?
std::function<void(SNode *, int)> visit = [&](SNode *snode, int depth) {
auto element_list =
runtime_query<void *>("LLVMRuntime_get_element_lists", result_buffer,
llvm_runtime, snode->id);
if (snode->type != SNodeType::place) {
fmt::print("SNode {:10}\n", snode->get_node_type_name_hinted());
if (element_list) {
fmt::print(" active element list:");
print_list_manager_info(element_list, result_buffer);
auto node_allocator =
runtime_query<void *>("LLVMRuntime_get_node_allocators",
result_buffer, llvm_runtime, snode->id);
if (node_allocator) {
auto free_list = runtime_query<void *>("NodeManager_get_free_list",
result_buffer, node_allocator);
auto recycled_list = runtime_query<void *>(
"NodeManager_get_recycled_list", result_buffer, node_allocator);
auto free_list_len = runtime_query<int32>(
"ListManager_get_num_elements", result_buffer, free_list);
auto recycled_list_len = runtime_query<int32>(
"ListManager_get_num_elements", result_buffer, recycled_list);
auto free_list_used = runtime_query<int32>(
"NodeManager_get_free_list_used", result_buffer, node_allocator);
auto data_list = runtime_query<void *>("NodeManager_get_data_list",
result_buffer, node_allocator);
fmt::print(" data list: ");
print_list_manager_info(data_list, result_buffer);
fmt::print(
" Allocated elements={:n}; free list length={:n}; recycled list "
"length={:n}\n",
free_list_used, free_list_len, recycled_list_len);
}
}
}
for (const auto &ch : snode->ch) {
visit(ch.get(), depth + 1);
}
};
for (auto &a : snode_trees_) {
visit(a->root(), /*depth=*/0);
}
auto total_requested_memory = runtime_query<std::size_t>(
"LLVMRuntime_get_total_requested_memory", result_buffer, llvm_runtime);
fmt::print(
"Total requested dynamic memory (excluding alignment padding): {:n} B\n",
total_requested_memory);
}
cuda::CudaDevice *LlvmProgramImpl::cuda_device() {
if (config->arch != Arch::cuda) {
TI_ERROR("arch is not cuda");
}
return static_cast<cuda::CudaDevice *>(device_.get());
}
cpu::CpuDevice *LlvmProgramImpl::cpu_device() {
TI_ERROR_IF(!arch_is_cpu(config->arch), "arch is not cpu");
return static_cast<cpu::CpuDevice *>(device_.get());
}
DevicePtr LlvmProgramImpl::get_snode_tree_device_ptr(int tree_id) {
DeviceAllocation tree_alloc = snode_tree_allocs_[tree_id];
return tree_alloc.get_ptr();
}
} // namespace lang
} // namespace taichi