interpreter_impl.cpp 58.5 KB
Newer Older
1
#include "./interpreter_impl.h"
2

3 4
#include "range/v3/all.hpp"

5
#include "megbrain/common.h"
6 7
#include "megbrain/imperative/opr_utility.h"
#include "megbrain/imperative/ops/autogen.h"
8 9
#include "megbrain/imperative/ops/backward_graph.h"
#include "megbrain/imperative/ops/opr_attr.h"
10
#include "megbrain/imperative/ops/utility.h"
11 12
#include "megbrain/imperative/utils/to_string.h"

13
#include "../blob_manager_impl.h"
14 15
#include "../event_pool.h"
#include "../op_trait.h"
16
#include "megbrain/imperative/backtrace.h"
17 18 19 20 21
using namespace mgb;
using namespace imperative;
using namespace interpreter;
using namespace interpreter::intl;

22
namespace {
M
Megvii Engine Team 已提交
23 24 25 26 27 28 29 30
auto tinfo_to_tid(SmallVector<TensorInfo*> tinfo) {
    SmallVector<uint64_t> tid;
    for (auto* ptinfo : tinfo) {
        tid.push_back(ptinfo->id);
    }
    return tid;
};
}  // namespace
31

32
namespace mgb {
M
Megvii Engine Team 已提交
33
using namespace profiler;
34 35
}

36 37 38 39 40
#if defined(_WIN32) || defined(_WIN64)
#define SYMBOL_EXPORT __declspec(dllexport)
#else
#define SYMBOL_EXPORT __attribute__((visibility("default")))
#endif
41 42 43 44 45 46 47

namespace mgb {

/**
 * USAGE
 *
 *   header:
48
 *     namespace mgb { void imperative_log_profile(const char* message); }
49 50 51 52 53
 *
 *   code:
 *     mgb::imperative_log_profile("MY MESSAGE");
 *
 **/
54
SYMBOL_EXPORT
55
void imperative_log_profile_begin(const char* message) {
56
    MGB_RECORD_EVENT(CustomEvent, std::string{message});
57 58
}

59
SYMBOL_EXPORT
60
void imperative_log_profile_end(const char* message) {
61
    MGB_RECORD_EVENT(CustomFinishEvent, std::string{message});
62 63
}

64
SYMBOL_EXPORT
M
Megvii Engine Team 已提交
65
void imperative_log_profile(const char* message) {
66 67 68 69
    imperative_log_profile_begin(message);
    imperative_log_profile_end(message);
}

70 71 72 73
SYMBOL_EXPORT
void imperative_log_profile_begin(const char* message, const char* device) {
    auto comp_node = CompNode::load(device);
    MGB_RECORD_EVENT(CustomEvent, std::string{message}, {}, comp_node);
M
Megvii Engine Team 已提交
74 75
    MGB_RECORD_EVENT(
            RecordDeviceEvent, EventPool::with_timer().alloc_shared(comp_node));
76 77 78 79 80
}

SYMBOL_EXPORT
void imperative_log_profile_end(const char* message, const char* device) {
    auto comp_node = CompNode::load(device);
M
Megvii Engine Team 已提交
81 82
    MGB_RECORD_EVENT(
            RecordDeviceEvent, EventPool::with_timer().alloc_shared(comp_node));
83 84 85
    MGB_RECORD_EVENT(CustomFinishEvent, std::string{message}, {}, comp_node);
}

M
Megvii Engine Team 已提交
86
}  // namespace mgb
87

88 89 90 91 92 93 94 95 96 97 98 99 100 101
std::thread::id ChannelImpl::get_worker_tid() {
    return m_worker_state.tid;
}

ChannelImpl::ChannelState& ChannelImpl::get_channel_state() {
    assert_in_channel();
    return m_channel_state;
}

ChannelImpl::WorkerState& ChannelImpl::get_worker_state() {
    assert_in_worker();
    return m_worker_state;
}

102 103 104
void ChannelImpl::WorkQueue::on_async_queue_worker_thread_start() {
    sys::set_thread_name("worker");
    m_owner->m_worker_state.tid = std::this_thread::get_id();
105
    auto custom_allocator = [&](CompNode device, size_t size) {
106 107 108
        auto blob = Blob::make(device, size);
        m_owner->alloc_tensor_with_evict(blob.get());
        return blob->storage();
109 110
    };
    OpDef::set_allocator(custom_allocator);
111 112
}

113
// Do not use m_xxx_state directly
114 115 116
#define m_channel_state
#define m_worker_state

117
std::unique_ptr<Interpreter::Channel> InterpreterImpl::create_channel() {
118 119 120 121 122 123 124 125 126 127 128 129
    auto ret = std::make_unique<ChannelImpl>();
#if !(defined(_WIN32) || defined(_WIN64))
    auto disable_channels = [](void) -> void {
        for (ChannelImpl* channel : ChannelImpl::m_all_active_channels) {
            if (channel->worker_started()) {
                channel->update_status_to_forked();
            }
        }
    };
    pthread_atfork(nullptr, nullptr, static_cast<void (*)(void)>(disable_channels));
#endif
    return ret;
130 131 132 133 134 135 136
}

Interpreter& Interpreter::inst() {
    static InterpreterImpl inst_;
    return inst_;
}

137
Handle ChannelImpl::put(const HostTensorND& value, bool no_cache) {
138
    MGB_LOCK_GUARD(m_spin);
139
    assert_available();
140 141 142 143 144
    std::optional<StackManager::Guard> guard;
    if (Profiler::is_profiling()) {
        auto& state = get_channel_state();
        guard.emplace("Put", &state.stack_manager);
    }
145
    auto info = put_impl(value, no_cache);
M
Megvii Engine Team 已提交
146
    return reinterpret_cast<Handle>(info);
147 148 149
}

TensorInfo* ChannelImpl::put_impl(const HostTensorND& value, bool no_cache) {
150 151 152 153 154
    if (value.empty()) {
        auto layout = value.layout();
        layout.init_contiguous_stride();
        const_cast<HostTensorND&>(value).reset(value.storage(), layout);
    }
155
    auto info = alloc();
156 157 158 159 160 161
    constexpr int size_threshold = TensorShape::MAX_NDIM;
    init(info, {value.layout(), value.comp_node()});
    if (value.layout().total_nr_elems() <= size_threshold) {
        info->h_value = value;
        info->desc.value = value.proxy_to_default_cpu();
    }
162 163 164 165 166 167 168 169 170 171
    if (Profiler::is_profiling()) {
        m_worker.add_task(
                {Profiler::next_id(), Put{info, value, no_cache},
                 get_channel_state().stack_manager.dump()});
    } else {
        m_worker.add_task({
                Profiler::next_id(),
                Put{info, value, no_cache},
        });
    }
172 173

    if (get_channel_state().options.async_level == 0) {
174
        sync_impl();
175
        info->desc.comp_node.sync();
176 177
        auto err = info->desc.comp_node.check_async_error();
        mgb_assert(!err, "%s", err->what());
178
    }
179 180 181
    return info;
}

182
Handle ChannelImpl::put(const DeviceTensorND& data, const HostTensorND& hvalue) {
183
    MGB_LOCK_GUARD(m_spin);
184
    assert_available();
M
Megvii Engine Team 已提交
185
    return reinterpret_cast<Handle>(put_impl(data, hvalue));
186
}
M
Megvii Engine Team 已提交
187 188
TensorInfo* ChannelImpl::put_impl(
        const DeviceTensorND& data, const HostTensorND& hvalue) {
189 190 191 192 193
    std::optional<StackManager::Guard> guard;
    if (Profiler::is_profiling()) {
        auto& state = get_channel_state();
        guard.emplace("Put", &state.stack_manager);
    }
M
Megvii Engine Team 已提交
194
    auto info = alloc();
195
    MGB_RECORD_EVENT(TensorCommandEvent, info->id, TensorCommandKind::Put);
196
    constexpr int size_threshold = TensorShape::MAX_NDIM;
197
    init(info, {data.layout(), data.comp_node()});
198 199 200
    if ((!hvalue.empty()) && info->desc.layout.total_nr_elems() <= size_threshold) {
        info->desc.value = hvalue.proxy_to_default_cpu();
    }
201
    info->ptr = Tensor::make(data, hvalue);
M
Megvii Engine Team 已提交
202 203 204
    MGB_RECORD_EVENT(
            TensorProduceEvent, info->id, info->desc.layout, info->desc.comp_node,
            data.raw_ptr());
205
    info->status = TensorInfo::Produced;
206
    MGB_RECORD_EVENT(TensorCommandFinishEvent, info->id, TensorCommandKind::Put);
M
Megvii Engine Team 已提交
207 208 209
    return info;
}

210
void ChannelImpl::del(Handle handle) {
211
    MGB_LOCK_GUARD(m_spin);
M
Megvii Engine Team 已提交
212
    if (!check_available()) {
213 214
        return;
    }
215 216 217 218
    del_impl(handle);
}

void ChannelImpl::del_impl(Handle handle) {
219 220 221
    mgb_assert(m_valid_handle.count(handle), "invalid handle: %p", handle);
    auto* info = reinterpret_cast<TensorInfo*>(handle);
    m_valid_handle.erase(handle);
222 223 224 225 226 227 228 229 230 231
    if (Profiler::is_profiling()) {
        m_worker.add_task(
                {Profiler::next_id(), Del{info},
                 get_channel_state().stack_manager.dump()});
    } else {
        m_worker.add_task({
                Profiler::next_id(),
                Del{info},
        });
    }
232 233
}

234
void ChannelImpl::drop(Handle handle) {
235
    MGB_LOCK_GUARD(m_spin);
236
    assert_available();
237 238
    auto& state = get_channel_state();
    if (state.options.enable_drop) {
M
Megvii Engine Team 已提交
239 240
        mgb_assert(
                m_valid_handle.find(handle) != m_valid_handle.end(),
241
                "invalid handle: %p", handle);
242
        auto* info = reinterpret_cast<TensorInfo*>(handle);
243 244 245 246 247 248 249 250 251 252
        if (Profiler::is_profiling()) {
            m_worker.add_task(
                    {Profiler::next_id(), Drop{info},
                     get_channel_state().stack_manager.dump()});
        } else {
            m_worker.add_task({
                    Profiler::next_id(),
                    Drop{info},
            });
        }
253 254 255
    }
}

256
void ChannelImpl::dispatch_default_cpu(
M
Megvii Engine Team 已提交
257
        std::shared_ptr<OpDef> op, const SmallVector<TensorInfo*>& input_infos,
258 259
        const SmallVector<LogicalTensorDesc>& input_descs,
        SmallVector<Handle>* outputs) {
260
    auto& state = get_channel_state();
261

262 263 264 265
    std::optional<StackManager::Guard> guard;
    if (Profiler::is_profiling()) {
        guard.emplace(op->trait()->make_name(*op), &state.stack_manager);
    }
266

M
Megvii Engine Team 已提交
267 268
    auto [output_descs, validated] =
            OpDef::infer_output_attrs_fallible(*op, input_descs);
269
    MGB_RECORD_EVENT(ShapeInferEvent, validated);
270

271 272
    SmallVector<DeviceTensorND> input_tensornds;
    CompNode output_cn;
273 274
    {
        MGB_LOCK_GUARD(m_mutex);
275
        for (auto&& info : input_infos) {
276
            auto input_cn = info->desc.comp_node;
277
            if (!output_cn.valid()) {
278 279 280 281 282 283
                output_cn = input_cn;
            } else {
                mgb_assert(output_cn == input_cn, "cannot decide output comp node");
            }

            if (info->ptr && info->ptr->try_get_value()) {
M
Megvii Engine Team 已提交
284 285
                input_tensornds.emplace_back(
                        info->ptr->get_value().proxy_to_default_cpu());
286
            } else {
287
                // We assign h_value before drop ptr
288 289
                mgb_assert(!info->h_value.empty(), "inp->h_value is empty!");
                input_tensornds.emplace_back(info->h_value.proxy_to_default_cpu());
290 291 292 293 294 295 296 297 298
            }
        }
    }

    SmallVector<DeviceTensorND> output_tensornds;
    for (auto&& desc : output_descs) {
        // TODO: may conflict with condtake, which need alloc inside
        mgb_assert(!desc.layout.is_empty());
        // use HostTensorND alloc_host for cuda pinned memory
M
Megvii Engine Team 已提交
299 300
        output_tensornds.emplace_back(
                HostTensorND(output_cn, desc.layout).proxy_to_default_cpu());
301 302
    }

303
    uint64_t op_id = Profiler::next_id();
304

305 306 307 308 309 310 311 312 313
    if (op->trait()->apply_on_device_tensornd) {
        OpDef::apply_on_device_tensornd(*op, input_tensornds, &output_tensornds);
    } else {
        // proxy to apply_on_physical_tensor
        SmallVector<TensorPtr> input_tensors;
        for (auto&& input_tensornd : input_tensornds) {
            input_tensors.push_back(Tensor::make(
                    input_tensornd, HostTensorND::make_proxy(input_tensornd)));
        }
314 315
        auto output_tensors = OpDef::apply_on_physical_tensor(
                *op, input_tensors, output_descs, validated);
316 317 318 319
        for (size_t i = 0; i < output_tensors.size(); ++i) {
            output_tensornds[i].copy_from_fixlayout(output_tensors[i]->dev_tensor());
        }
    }
320 321 322

    SmallVector<TensorInfo*> output_infos;
    for (auto&& tensornd : output_tensornds) {
M
Megvii Engine Team 已提交
323 324
        HostTensorND host_tensornd =
                HostTensorND::make_proxy(tensornd).proxy_to_comp_node(output_cn);
325
        // use `put` for consistency
326
        auto info = reinterpret_cast<TensorInfo*>(put_impl(host_tensornd, false));
327
        mgb_assert(info->shape_valid());
328
        output_infos.push_back(info);
M
Megvii Engine Team 已提交
329
        outputs->push_back(reinterpret_cast<Handle>(info));
330
    }
331 332
    auto& bt = get_backtrace();
    auto op_info_getter = [op, bt] {
333 334
        std::unordered_map<std::string, std::string> op_info;
        auto props = OpDef::props(*op);
M
Megvii Engine Team 已提交
335
        for (auto&& [key, value] : props) {
336 337
            op_info[key] = value;
        }
338 339 340 341 342 343
        if (bt != nullptr) {
            if (bt->py_stack_info != nullptr)
                op_info["python_backtrace"] = bt->py_traceback();
            if (bt->trans_stack_info.size() > 0)
                op_info["transformation_backtrace"] = bt->transformation_traceback();
        }
344 345
        return op_info;
    };
M
Megvii Engine Team 已提交
346
    MGB_RECORD_EVENT(
347 348 349
            OpDispatchEvent, op_id, guard.value().name(), op_info_getter,
            tinfo_to_tid(input_infos), tinfo_to_tid(output_infos),
            state.stack_manager.dump());
350
}
351

352
void ChannelImpl::dispatch_kernel(
M
Megvii Engine Team 已提交
353
        std::shared_ptr<OpDef> op, const SmallVector<TensorInfo*>& input_infos,
354 355
        const SmallVector<LogicalTensorDesc>& input_descs,
        SmallVector<Handle>* outputs) {
356
    auto& state = get_channel_state();
357 358
    auto& options = state.options;

359 360 361 362
    std::optional<StackManager::Guard> guard;
    if (Profiler::is_profiling()) {
        guard.emplace(op->trait()->make_name(*op), &state.stack_manager);
    }
363

M
Megvii Engine Team 已提交
364 365
    auto [output_descs, validated] =
            OpDef::infer_output_attrs_fallible(*op, input_descs);
366
    MGB_RECORD_EVENT(ShapeInferEvent, validated);
367

368 369 370 371
    SmallVector<TensorInfo*> output_infos;
    output_infos.reserve(output_descs.size());

    outputs->reserve(output_descs.size());
372 373
    for (int i = 0; i < output_descs.size(); ++i) {
        auto&& desc = output_descs[i];
374
        auto info = alloc();
375
        init(info, std::move(desc));
376 377
        // make sure desc's value is consistent with h_value
        if (!info->desc.value.empty()) {
378 379
            info->h_value = HostTensorND::make_proxy(info->desc.value)
                                    .proxy_to_comp_node(info->desc.comp_node);
380
        }
381
        output_infos.push_back(info);
M
Megvii Engine Team 已提交
382
        outputs->push_back(reinterpret_cast<Handle>(info));
383
    }
384 385 386
    auto& bt = get_backtrace();
    ApplyOp cmd{Profiler::next_id(),     std::move(op), std::move(input_infos),
                std::move(output_infos), validated,     bt};
387
    if (Profiler::is_profiling()) {
388
        auto op_info_getter = [op = cmd.op, bt = cmd.bt] {
389 390 391 392 393
            std::unordered_map<std::string, std::string> op_info;
            auto props = OpDef::props(*op);
            for (auto&& [key, value] : props) {
                op_info[key] = value;
            }
394 395 396 397 398 399 400
            if (bt != nullptr) {
                if (bt->py_stack_info != nullptr)
                    op_info["python_backtrace"] = bt->py_traceback();
                if (bt->trans_stack_info.size() > 0)
                    op_info["transformation_backtrace"] =
                            bt->transformation_traceback();
            }
401 402
            return op_info;
        };
403
        MGB_RECORD_EVENT(
404 405 406
                OpDispatchEvent, cmd.id, guard.value().name(), op_info_getter,
                tinfo_to_tid(cmd.inputs), tinfo_to_tid(cmd.outputs),
                state.stack_manager.dump());
407
        m_worker.add_task(
408
                {Profiler::next_id(), std::move(cmd),
409 410 411 412
                 get_channel_state().stack_manager.dump()});
    } else {
        m_worker.add_task({
                Profiler::next_id(),
413
                std::move(cmd),
414 415
        });
    }
416
    if (!validated && options.async_level == 1) {
417
        sync_impl();
418
    } else if (options.async_level == 0) {
419
        sync_impl();
420
        // check device error
421
        for (auto&& oup : *outputs) {
422 423
            auto info = reinterpret_cast<TensorInfo*>(oup);
            info->ptr->comp_node().sync();
424 425
            auto err = info->ptr->comp_node().check_async_error();
            mgb_assert(!err, "%s", err->what());
426
        }
427
    }
428 429 430
}

SmallVector<Handle> ChannelImpl::apply_op(
M
Megvii Engine Team 已提交
431
        std::shared_ptr<OpDef> op, const SmallVector<Handle>& inputs) {
432
    MGB_LOCK_GUARD(m_spin);
433
    assert_available();
434
    auto* input = reinterpret_cast<TensorInfo*>(inputs[0]);
435
    if (op->same_type<GetVarShape>() && input->shape_valid()) {
436 437 438 439 440 441 442 443 444
        size_t ndim = input->desc.layout.ndim;
        auto& gvs = op->cast_final_safe<GetVarShape>();
        if (gvs.axis == MEGDNN_MAX_NDIM) {
            HostTensorND shape_tensor{input->desc.comp_node, {ndim}, dtype::Int32()};
            DeviceTensorND shape_tensor_device = shape_tensor.proxy_to_default_cpu();
            cg::copy_shape_to_tensor_value(shape_tensor_device, input->desc.layout);
            return {reinterpret_cast<Handle>(put_impl(shape_tensor, false))};
        }
    }
445 446 447 448
    return apply_op_impl(std::move(op), inputs);
}

SmallVector<Handle> ChannelImpl::apply_op_impl(
M
Megvii Engine Team 已提交
449
        std::shared_ptr<OpDef> op, const SmallVector<Handle>& inputs) {
450
    auto& state = get_channel_state();
451
    for (auto i : inputs) {
M
Megvii Engine Team 已提交
452 453 454
        mgb_assert(
                m_valid_handle.find(i) != m_valid_handle.end(), "invalid handle: %p",
                i);
455 456 457 458
    }
    SmallVector<TensorInfo*> input_infos;
    SmallVector<LogicalTensorDesc> input_descs;
    {
459
        MGB_LOCK_GUARD(m_info_spin);
460 461
        for (auto i : inputs) {
            auto info = reinterpret_cast<TensorInfo*>(i);
M
Megvii Engine Team 已提交
462 463 464
            mgb_assert(
                    !info->invalid,
                    "an input tensor is unusable due to previous error");
465 466 467 468 469 470
            input_infos.push_back(info);
            input_descs.push_back(info->desc);
        }
    }

    SmallVector<Handle> outputs;
471
    DispatchMode dispatch_mode = state.options.enable_host_compute
M
Megvii Engine Team 已提交
472 473
                                       ? OpDef::decide_dispatch_mode(*op, input_descs)
                                       : DispatchMode::KERNEL;
474
    switch (dispatch_mode) {
475 476 477 478 479 480 481 482 483
        case DEFAULT_CPU: {
            dispatch_default_cpu(op, input_infos, input_descs, &outputs);
            break;
        }
        case KERNEL: {
            dispatch_kernel(op, input_infos, input_descs, &outputs);
            break;
        }
    }
484 485 486
    return outputs;
}

487
HostTensorND ChannelImpl::get_value(Handle handle) {
488
    MGB_LOCK_GUARD(m_spin);
489
    assert_available();
M
Megvii Engine Team 已提交
490 491 492
    mgb_assert(
            m_valid_handle.find(handle) != m_valid_handle.end(), "invalid handle: %p",
            handle);
493
    auto info = reinterpret_cast<TensorInfo*>(handle);
494
    // donnot use info->value_fetched, it's unsafe
495
    mgb_assert(!info->invalid, "tensor is unusable due to previous error");
496
    return wait_tensor(info, TensorProp::HostValue)->get_value();
497 498
}

499
TensorShape ChannelImpl::get_shape(Handle handle) {
500
    MGB_LOCK_GUARD(m_spin);
501
    assert_available();
M
Megvii Engine Team 已提交
502 503 504
    mgb_assert(
            m_valid_handle.find(handle) != m_valid_handle.end(), "invalid handle: %p",
            handle);
505
    auto info = reinterpret_cast<TensorInfo*>(handle);
506
    if (info->shape_valid()) {
507 508
        return info->desc.layout;
    }
509
    TensorShape ret = wait_tensor(info, TensorProp::Shape)->layout();
510
    mgb_assert(ret.ndim > 0);
511 512 513
    return ret;
}

514
DType ChannelImpl::get_dtype(Handle handle) {
515
    MGB_LOCK_GUARD(m_spin);
516
    assert_available();
M
Megvii Engine Team 已提交
517 518 519
    mgb_assert(
            m_valid_handle.find(handle) != m_valid_handle.end(), "invalid handle: %p",
            handle);
520
    auto info = reinterpret_cast<TensorInfo*>(handle);
521
    MGB_RECORD_EVENT(TensorGetPropEvent, info->id, TensorProp::DType);
522 523 524 525 526
    auto ret = info->desc.layout.dtype;
    mgb_assert(ret.valid());
    return ret;
}

527
CompNode ChannelImpl::get_device(Handle handle) {
528
    MGB_LOCK_GUARD(m_spin);
529
    assert_available();
M
Megvii Engine Team 已提交
530 531 532
    mgb_assert(
            m_valid_handle.find(handle) != m_valid_handle.end(), "invalid handle: %p",
            handle);
533
    auto info = reinterpret_cast<TensorInfo*>(handle);
534
    MGB_RECORD_EVENT(TensorGetPropEvent, info->id, TensorProp::Device);
535 536 537 538 539
    auto ret = info->desc.comp_node;
    mgb_assert(ret.valid());
    return ret;
}

540
DeviceTensorND ChannelImpl::get_dev_tensor(Handle handle) {
541
    MGB_LOCK_GUARD(m_spin);
542
    assert_available();
M
Megvii Engine Team 已提交
543 544 545
    mgb_assert(
            m_valid_handle.find(handle) != m_valid_handle.end(), "invalid handle: %p",
            handle);
546
    auto info = reinterpret_cast<TensorInfo*>(handle);
547
    return wait_tensor(info, TensorProp::DevValue)->dev_tensor();
548 549 550
}

void ChannelImpl::sync() {
551
    MGB_LOCK_GUARD(m_spin);
552
    assert_available();
553 554 555 556
    sync_impl();
}

void ChannelImpl::sync_impl() {
557 558 559 560 561 562
    m_worker.wait_all_task_finish();
    MGB_LOCK_GUARD(m_mutex);
    check_worker_exc_unsafe();
}

void ChannelImpl::close() {
563
    MGB_LOCK_GUARD(m_spin);
564 565 566 567
    if (!check_available()) {
        return;
    }
    std::vector<Handle> valid_handles(m_valid_handle.begin(), m_valid_handle.end());
M
Megvii Engine Team 已提交
568
    for (auto* handle : valid_handles) {
569
        del_impl(handle);
570 571 572
    }
    mgb_assert(m_valid_handle.empty());
    mgb_log_debug("%ld tensor exists before channel close", (long)valid_handles.size());
573
    sync_impl();
574
    m_status = ChannelRunningStatus::CLOSED;
575 576
}

577
size_t ChannelImpl::get_option(std::string name) {
578
    MGB_LOCK_GUARD(m_spin);
579
    assert_available();
580 581
    auto& state = get_channel_state();
    return state.options.get_option(name);
582 583
}

584
void ChannelImpl::set_option(std::string name, size_t value) {
585
    MGB_LOCK_GUARD(m_spin);
586
    assert_available();
587 588
    auto& state = get_channel_state();
    state.options.set_option(name, value);
589 590 591 592 593 594 595 596 597
    // FIXME
    if (name == "enable_dtr_auto_drop" && value) {
        auto custom_allocator = [&](CompNode device, size_t size) {
            auto blob = Blob::make(device, size);
            alloc_tensor_with_evict(blob.get());
            return blob->storage();
        };
        BlobManager::inst()->set_allocator(custom_allocator);
    }
598 599 600 601 602 603 604 605 606 607
    if (Profiler::is_profiling()) {
        m_worker.add_task(
                {Profiler::next_id(), SetOption{name, value},
                 get_channel_state().stack_manager.dump()});
    } else {
        m_worker.add_task({
                Profiler::next_id(),
                SetOption{name, value},
        });
    }
608 609
}

610 611
void ChannelImpl::clear_candidates() {
    MGB_LOCK_GUARD(m_spin);
612
    assert_available();
613 614 615
    m_dtr.candidates.clear();
}

616
TensorInfo* ChannelImpl::alloc() {
617
    auto& state = get_channel_state();
M
Megvii Engine Team 已提交
618
    auto info = [this] {
619
        MGB_LOCK_GUARD(m_pool_spin);
620
        return m_pool.alloc();
621 622 623
    }();
    info->id = Profiler::next_id();
    if (Profiler::is_profiling()) {
624
        size_t tensor_id = state.stack_manager.current()->next_id("tensor");
M
Megvii Engine Team 已提交
625 626
        info->name =
                state.stack_manager.dump().to_string() + ssprintf(":%zu", tensor_id);
627
    }
628
    return info;
629 630
}

631
void ChannelImpl::init(TensorInfo* info, LogicalTensorDesc&& desc) {
M
Megvii Engine Team 已提交
632
    m_valid_handle.insert(reinterpret_cast<Handle>(info));
633
    MGB_RECORD_EVENT(TensorDeclareEvent, info->id, info->name);
634
    mgb_assert(desc.comp_node.valid(), "comp_node invalid");
635
    info->status = TensorInfo::Allocated;
636
    info->desc = std::move(desc);
637 638
}

M
Megvii Engine Team 已提交
639
void ChannelImpl::do_drop(TensorInfo* ptr, bool user = false) {
640 641
    if (!ptr->producer) {
        if (user) {
M
Megvii Engine Team 已提交
642 643 644 645
            mgb_log_warn(
                    "the input that produced tensor %p has been deleted, this drop "
                    "operation will be ignored",
                    ptr);
646 647 648 649 650 651 652
        }
        return;
    }
    if (ptr->evict_type != EvictType::NONE) {
        return;
    }
    ptr->evict_type = EvictType::DROP;
653
    ptr->status = TensorInfo::Dropped;
654 655 656
    release_tensor(ptr);
}

657
void ChannelImpl::free(TensorInfo* ptr) {
658 659
    auto& state = get_worker_state();
    if (state.options.enable_dtr_auto_drop) {
660 661 662 663 664 665 666 667 668 669 670 671 672 673 674
        // Evicting a tensor, rather than freeing it, can avoid pinning
        // potentially exploding amounts of memory and allow us to save
        // more memory.
        ptr->allow_delete = true;
        if (!ptr->ref_cnt) {
            recursive_free(ptr);
        } else {
            do_drop(ptr);
        }
    } else {
        real_free(ptr);
    }
}

void ChannelImpl::recursive_free(TensorInfo* ptr) {
675
    MGB_RECORD_EVENT(TensorCommandEvent, ptr->id, TensorCommandKind::RecFree);
676
    SmallVector<TensorInfo*> inps;
677 678 679 680 681 682 683 684 685 686 687 688 689
    if (ptr->producer) {
        for (auto i : ptr->producer->inputs) {
            if (i && --i->ref_cnt == 0) {
                inps.push_back(i);
            }
        }
    }
    real_free(ptr);
    for (auto i : inps) {
        if (i->allow_delete) {
            recursive_free(i);
        }
    }
690
    MGB_RECORD_EVENT(TensorCommandFinishEvent, ptr->id, TensorCommandKind::RecFree);
691 692 693
}

void ChannelImpl::real_free(TensorInfo* ptr) {
694 695
    auto& state = get_worker_state();
    if (ptr->size_exceeds_thd(state.options.dtr_evictee_minimum_size)) {
696 697 698 699
        m_dtr.erase_candidate(ptr);
    }
    detach_users(ptr);
    ptr->detach_producer();
700 701
    bool has_value = ptr->ptr != nullptr;
    if (has_value) {
702
        MGB_RECORD_EVENT(TensorReleaseEvent, ptr->id);
703
    }
704
    MGB_RECORD_EVENT(TensorEraseEvent, ptr->id, ptr->ptr_use_count);
705
    ptr->status = TensorInfo::Deleted;
706
    MGB_LOCK_GUARD(m_pool_spin);
707 708 709
    m_pool.free(ptr);
}

710 711 712 713 714 715 716
std::unordered_set<ChannelImpl*> ChannelImpl::m_all_active_channels{};
MGB_MUTEX ChannelImpl::m_all_active_channels_mutex{};

ChannelImpl::ChannelImpl() : m_worker(this) {
    MGB_LOCK_GUARD(m_all_active_channels_mutex);
    m_all_active_channels.emplace(this);
}
717

718 719
ChannelImpl::~ChannelImpl() {
    close();
720 721
    MGB_LOCK_GUARD(m_all_active_channels_mutex);
    m_all_active_channels.erase(this);
722
}
723

724
void ChannelImpl::produce_tensor(TensorInfo* dest, TensorPtr ptr) {
725
    auto& state = get_worker_state();
726
    MGB_LOCK_GUARD(m_mutex);
727
    MGB_LOCK_GUARD(m_info_spin);
728
    m_dtr.update_used_time(dest);
M
Megvii Engine Team 已提交
729 730
    MGB_RECORD_EVENT(
            TensorProduceEvent, dest->id, ptr->layout(), ptr->comp_node(),
731
            ptr->raw_ptr_not_for_readwrite());
732
    // update tensor desc for static infer
733
    dest->update_layout(ptr->layout());
734 735
    // in order to avoid performance impact,
    // memory forwarding is disabled when DTR is enabled
736
    if (state.options.enable_dtr_auto_drop || state.options.disable_memory_forwarding) {
737 738
        ptr->to_contiguous_inplace();
    }
739
    dest->desc.comp_node = ptr->comp_node();
740
    dest->memory = ptr->blob()->size();
741
    dest->ptr = std::move(ptr);
742
    dest->evict_type = EvictType::NONE;
743
    dest->status = TensorInfo::Produced;
744 745
    if (dest->pinned == 0 &&
        dest->size_exceeds_thd(state.options.dtr_evictee_minimum_size)) {
746 747
        m_dtr.insert_candidate(dest);
    }
748
    notify_tensor_unsafe(dest);
749 750
}

751
void ChannelImpl::release_tensor(TensorInfo* dest) {
752
    MGB_RECORD_EVENT(TensorReleaseEvent, dest->id);
753 754
    MGB_LOCK_GUARD(m_mutex);
    dest->ptr.reset();
755 756 757 758
    auto& state = get_worker_state();
    if (dest->size_exceeds_thd(state.options.dtr_evictee_minimum_size)) {
        m_dtr.erase_candidate(dest);
    }
759 760
}

761
void ChannelImpl::regenerate(TensorInfo* dest) {
762
    if (dest->evict_type == EvictType::DROP) {
M
Megvii Engine Team 已提交
763 764
        auto&& path = dest->producer;
        m_apply_stack.push(
765 766
                {ApplyOp{path->id, path->op, path->inputs, path->outputs}, 0, dest,
                 "dtr"});
M
Megvii Engine Team 已提交
767 768
        if (!m_applying)
            flush_apply_stack();
769 770 771
    }
}

772
void ChannelImpl::do_apply_op(const ApplyOp& cmd, std::string reason) {
773 774
    using namespace ranges;
    using namespace ranges::views;
775
    auto& state = get_worker_state();
M
Megvii Engine Team 已提交
776 777
    bool profiling_device =
            Profiler::is_profiling() && Profiler::get_option("profile_device", 0);
778
    uint64_t apply_id = cmd.id;
779
    SmallVector<TensorPtr> inputs;
780
    inputs.reserve(cmd.inputs.size());
781 782 783
    // refcnt == 1, owners: [TensorInfo::ptr]
    for (auto i : cmd.inputs) {
        mgb_assert(i->ptr, "Invalid input tensor ptr!");
784
        // refcnt ++, owners: [i->ptr, tensor_inputs]
785
        // tensor_inputs.push_back(i->ptr);
786
        inputs.push_back(i->ptr);
787
    }
M
Megvii Engine Team 已提交
788 789
    if (state.options.enable_dtr_auto_drop &&
        state.options.dtr_eviction_threshold > 0) {
790 791
        auto_evict(0);
    }
M
Megvii Engine Team 已提交
792
    auto apply_on_physical_tensor =
793
            [&](auto&& self, const OpDef& def, SmallVector<TensorPtr>&& inputs,
794 795
                SmallVector<LogicalTensorDesc>& output_descs,
                const bool& validated) -> SmallVector<TensorPtr> {
796
        if (def.trait()->make_forward_graph) {
797 798 799 800 801 802 803 804 805 806
            auto apply_functor = [&](std::shared_ptr<OpDef> op,
                                     SmallVector<TensorPtr> inputs,
                                     size_t nr_outputs) -> SmallVector<TensorPtr> {
                auto opname = op->trait()->make_name(*op);
                imperative_log_profile_begin(opname.c_str());
                auto outputs = self(self, *op, std::move(inputs), output_descs, false);
                imperative_log_profile_end(opname.c_str());
                return outputs;
            };
            auto const_functor = [&](TensorPtr value) -> TensorPtr { return value; };
807 808
            // apply recursivily
            SmallVector<LogicalTensorDesc> input_descs;
M
Megvii Engine Team 已提交
809
            for (auto&& input : inputs) {
810
                input_descs.push_back({{{}, input->dtype()}, input->comp_node()});
811
            }
812
            auto forward_graph = OpDef::make_forward_graph(def, input_descs);
813 814
            auto outputs = forward_graph.apply<TensorPtr>(
                    inputs, apply_functor, const_functor);
815 816
            return outputs;
        }
817 818 819 820 821 822 823 824 825 826
        // Check Input Layout
        // Get the input layout constraints, and if the constraint is not satisfied
        // inplace update the layout and blob to make the tensor contiguous
        auto&& constraints = OpDef::get_input_layout_constraint(def, inputs);
        for (size_t idx = 0; idx < inputs.size(); ++idx) {
            auto&& layout_checker = constraints[idx];
            if (layout_checker) {
                inputs[idx]->to_contiguous_inplace(layout_checker);
            }
        }
827
        auto outputs = OpDef::apply_on_physical_tensor(
828
                def, std::move(inputs), output_descs, validated);
829 830 831 832 833
        for (auto& o : outputs) {
            o->set_ready_event(
                    record_event(o->comp_node(), def.same_type<imperative::Barrier>()));
        }
        return outputs;
834
    };
835
    MGB_RECORD_EVENT(OpExecuteEvent, apply_id, {}, reason);
836 837 838 839
    SmallVector<std::pair<CompNode, uint64_t>> kernels;
    if (profiling_device) {
        // Collecting devices
        SmallVector<CompNode> devices;
840 841 842
        for (auto&& i : concat(cmd.inputs, cmd.outputs)) {
            if (i != nullptr && count(devices, i->desc.comp_node) == 0) {
                devices.push_back(i->desc.comp_node);
843
                kernels.push_back({i->desc.comp_node, Profiler::next_id()});
844 845 846
            }
        }
    }
M
Megvii Engine Team 已提交
847
    for (auto* input : cmd.inputs) {
848
        auto input_id = input->id;
849 850 851
        MGB_RECORD_EVENT(OpInputEvent, input_id);
        MGB_RECORD_EVENT(TensorUsageEvent, input_id);
        MGB_RECORD_EVENT(OpInputFinishEvent, input_id);
852 853
    }
    // Before wait
M
Megvii Engine Team 已提交
854
    // TODO: split operator wait and execute so that OpWait could be corrected recorded.
855
    // Before execute
M
Megvii Engine Team 已提交
856
    for (auto&& [device, kernel_id] : kernels) {
857
        MGB_RECORD_EVENT(KernelLaunchEvent, apply_id, kernel_id, device);
M
Megvii Engine Team 已提交
858
        MGB_RECORD_EVENT_IF(
859 860
                (Profiler::get_option("profile_device", 0)), RecordDeviceEvent,
                Timer::record_device(device));
861 862
    }
    // Apply op
863
    SmallVector<LogicalTensorDesc> output_descs;
864 865 866 867 868 869
    bool validated = cmd.validated;
    if (!state.options.enable_dtr_auto_drop) {
        for (auto i : cmd.outputs) {
            output_descs.push_back(i->desc);
        }
    } else {
870
        // i may be null
871
        validated = false;
872 873 874
        for (auto i : cmd.outputs) {
            output_descs.push_back({});
        }
875
    }
876
    // Here std::move is REQUIRED for removing duplicated references.
877
    auto outputs = apply_on_physical_tensor(
878
            apply_on_physical_tensor, *cmd.op, std::move(inputs), output_descs,
879
            validated);
880
    // After execute
M
Megvii Engine Team 已提交
881 882
    for (auto&& [device, kernel_id] : kernels) {
        MGB_RECORD_EVENT_IF(
883 884
                (Profiler::get_option("profile_device", 0)), RecordDeviceEvent,
                Timer::record_device(device));
885
        MGB_RECORD_EVENT(KernelLaunchFinishEvent, apply_id, kernel_id, device);
886 887
    }
    // End profiling operator
888 889
    mgb_assert(outputs.size() == cmd.outputs.size());
    for (size_t i = 0; i < outputs.size(); ++i) {
890
        auto output = cmd.outputs[i];
891
        if (mgb_unlikely(output == nullptr)) {
892 893
            MGB_RECORD_EVENT(OpOutputEvent, 0);
            MGB_RECORD_EVENT(OpOutputFinishEvent, 0);
894
        } else if (mgb_unlikely(output->ptr != nullptr)) {
895 896
            MGB_RECORD_EVENT(OpOutputEvent, output->id);
            MGB_RECORD_EVENT(OpOutputFinishEvent, output->id);
897
        } else {
898
            MGB_RECORD_EVENT(OpOutputEvent, output->id);
899
            produce_tensor(output, outputs[i]);
900
            MGB_RECORD_EVENT(OpOutputFinishEvent, output->id);
901
            sample_on_device(output->desc.comp_node, false);
902 903 904 905 906 907 908 909
        }
    }

    if (state.options.enable_dtr_auto_drop) {
        double estimate_compute_time = 0;
        for (auto i : cmd.inputs) {
            estimate_compute_time += i->memory;
        }
910
        for (auto i : outputs) {
911
            estimate_compute_time += i->blob()->size();
912 913 914 915 916 917 918
        }
        m_dtr.estimate_timestamp += estimate_compute_time / 1e8;
        for (auto i : cmd.outputs) {
            if (i != nullptr) {
                i->compute_time = estimate_compute_time;
            }
        }
919
        m_dtr.unpin(cmd.inputs, state);
920
    }
921
    MGB_RECORD_EVENT(OpExecuteFinishEvent, apply_id, {}, reason);
922
    // End profiling operator
923
}
924

925 926
void ChannelImpl::flush_apply_stack() {
    m_applying = true;
927
    auto& state = get_worker_state();
928
    while (!m_apply_stack.empty()) {
M
Megvii Engine Team 已提交
929 930
        auto& [cmd, idx, recomp, reason] =
                m_apply_stack.top();  // cmd.inputs[0~idx-1] is in memory
931 932 933 934 935
        if (idx == 0) {
            if (state.options.enable_dtr_auto_drop) {
                m_dtr.pin(cmd.inputs);
            }
            if (recomp) {
M
Megvii Engine Team 已提交
936 937
                MGB_RECORD_EVENT(
                        TensorCommandEvent, recomp->id, TensorCommandKind::ReGen);
938 939 940
            }
        }
        bool regen = false;
M
Megvii Engine Team 已提交
941
        for (size_t i = idx; i < cmd.inputs.size(); i++) {
942 943 944 945 946 947
            auto&& p = cmd.inputs[i];
            if (state.options.enable_dtr_auto_drop) {
                m_dtr.update_used_time(p);
            }
            if (!p->ptr && p->evict_type != EvictType::NONE) {
                idx = i + 1;
M
Megvii Engine Team 已提交
948
                regenerate(p);  // add ApplyOp to the stack
949 950 951 952
                regen = true;
                break;
            }
        }
M
Megvii Engine Team 已提交
953 954
        if (regen)
            continue;
955
        // the required input tensors are already in memory
M
Megvii Engine Team 已提交
956 957
        auto [cmd_backup, recomp_backup, reason_backup] =
                std::make_tuple(cmd, recomp, reason);
958
        m_apply_stack.pop();
959
        do_apply_op(cmd_backup, reason_backup);
960
        if (recomp_backup) {
M
Megvii Engine Team 已提交
961 962 963
            MGB_RECORD_EVENT(
                    TensorCommandFinishEvent, recomp_backup->id,
                    TensorCommandKind::ReGen);
964 965
            for (auto o : cmd_backup.outputs) {
                if (o) {
966 967 968 969
                    m_dtr.update_dsu_after_recompute(o);
                }
            }
        }
970
    }
971
    m_applying = false;
972 973
}

974
bool ChannelImpl::auto_evict(size_t force_num) {
975
    auto& state = get_worker_state();
976
    if (!m_dtr.comp_node.valid()) {
977
        return false;
978 979
    }
    size_t current_memory = m_dtr.comp_node.get_used_memory();
980
    size_t flag = false;
M
Megvii Engine Team 已提交
981 982 983
    while ((state.options.dtr_eviction_threshold > 0 &&
            current_memory > state.options.dtr_eviction_threshold) ||
           force_num > 0) {
984
        MGB_RECORD_EVENT(AutoEvictEvent);
985
        sample_on_device(m_dtr.comp_node, false);
986
        auto best = m_dtr.find_best_tensor(state.options.enable_dtr_sqrt_sampling);
987
        if (!best) {
988
            MGB_RECORD_EVENT(AutoEvictFinishEvent);
989 990 991 992
            break;
        }
        if (best->ptr.unique() && best->ptr->blob().unique()) {
            current_memory -= best->memory;
993
            if (force_num > 0) {
M
Megvii Engine Team 已提交
994
                force_num--;
995 996
            }
            flag = true;
997 998 999 1000
        }
        do_drop(best);
        if (best->evict_type == EvictType::DROP) {
            m_dtr.update_dsu_after_evict(best);
1001
        }
1002
        sample_on_device(m_dtr.comp_node, false);
1003
        MGB_RECORD_EVENT(AutoEvictFinishEvent);
1004
    }
1005
    return flag;
1006 1007
}

1008 1009
void ChannelImpl::detach_users(TensorInfo* dest) {
    SmallVector<TensorInfo::ComputePath*> users = dest->users;
M
Megvii Engine Team 已提交
1010
    for (auto* user : users) {
1011 1012
        SmallVector<TensorInfo*> outputs = user->outputs;
        SmallVector<TensorInfo*> inputs = user->inputs;
M
Megvii Engine Team 已提交
1013 1014 1015 1016 1017
        for (auto* output : outputs) {
            // When a `ComputePath` is detach from it's input,
            // there is no need to reserve it,
            // so we detach all output of this path
            // to decrease it's `ref_cnt` to zero.
1018 1019 1020 1021 1022
            if (output == nullptr) {
                continue;
            }
            regenerate(output);
            output->detach_producer();
M
Megvii Engine Team 已提交
1023 1024
            for (auto* input : inputs) {
                input->ref_cnt--;
1025
            }
1026
        }
1027
        // now user is dead
1028
    }
1029
    mgb_assert(dest->users.empty(), "ComputePath leaking");
1030 1031
}

1032
bool ChannelImpl::check_available() {
1033
    return m_status == ChannelRunningStatus::RUNING;
1034 1035
}

1036 1037 1038 1039 1040
TensorPtr ChannelImpl::wait_tensor(TensorInfo* info, TensorProp prop) {
    std::unique_lock<decltype(m_mutex)> lock(m_mutex);
    mgb_assert(!m_waitee, "duplicate waitee");
    m_waitee = info;
    m_waitee_id = Profiler::next_id();
1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052
    auto backtrace_getter = [bt = get_backtrace()]() {
        std::unordered_map<std::string, std::string> infos;
        if (bt != nullptr) {
            if (bt->py_stack_info != nullptr)
                infos["python_backtrace"] = bt->py_traceback();
            if (bt->trans_stack_info.size() > 0)
                infos["transformation_backtrace"] = bt->transformation_traceback();
        }
        return infos;
    };
    MGB_RECORD_EVENT(
            TensorWaitPropEvent, info->id, m_waitee_id, prop, backtrace_getter);
1053
    bool require_host = prop == TensorProp::HostValue;
1054
    bool require_dev = prop == TensorProp::DevValue;
M
Megvii Engine Team 已提交
1055
    auto host_available = [&] { return info->ptr && info->ptr->value_fetched(); };
1056
    auto dev_available = [&] { return info->ptr; };
1057
    bool wait_host = false;
1058
    bool wait_regen = false;
1059
    if (require_host && !host_available()) {
1060 1061
        // avoid dead lock
        lock.unlock();
1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
        if (Profiler::is_profiling()) {
            m_worker.add_task(
                    {Profiler::next_id(), GetValue{info},
                     get_channel_state().stack_manager.dump()});
        } else {
            m_worker.add_task({
                    Profiler::next_id(),
                    GetValue{info},
            });
        }
1072
        lock.lock();
1073
        wait_host = true;
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
    if (require_dev && !dev_available()) {
        lock.unlock();
        if (Profiler::is_profiling()) {
            m_worker.add_task(
                    {Profiler::next_id(), StartRegen{info},
                     get_channel_state().stack_manager.dump()});
        } else {
            m_worker.add_task({
                    Profiler::next_id(),
                    StartRegen{info},
            });
        }
        lock.lock();
        wait_regen = true;
    }
    if (require_dev) {
        m_cv.wait(lock, [&]() {
            check_worker_exc_unsafe();
            return dev_available();
        });
    } else {
        m_cv.wait(lock, [&]() {
            check_worker_exc_unsafe();
            return require_host ? host_available() : static_cast<bool>(info->ptr);
        });
    }
1101 1102
    MGB_RECORD_EVENT(
            TensorWaitPropFinishEvent, info->id, m_waitee_id, prop, backtrace_getter);
1103
    m_waitee = nullptr;
1104
    if (wait_host) {
1105 1106 1107
        auto err = info->ptr->comp_node().check_async_error();
        mgb_assert(!err, "%s", err->what());
    }
1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121
    if (wait_regen) {
        lock.unlock();
        if (Profiler::is_profiling()) {
            m_worker.add_task(
                    {Profiler::next_id(), StopRegen{info},
                     get_channel_state().stack_manager.dump()});
        } else {
            m_worker.add_task({
                    Profiler::next_id(),
                    StopRegen{info},
            });
        }
        lock.lock();
    }
1122 1123 1124 1125 1126
    return info->ptr;
}

void ChannelImpl::notify_tensor_unsafe(TensorInfo* info) {
    if (info == m_waitee) {
1127
        MGB_RECORD_EVENT(TensorNotifyPropEvent, info->id);
1128
        m_cv.notify_all();
1129
    }
1130 1131 1132 1133
}

std::unordered_set<TensorInfo*> ChannelImpl::collect_valid_tensors() {
    std::unordered_set<TensorInfo*> valid_tensors;
M
Megvii Engine Team 已提交
1134
    for (auto* handle : m_valid_handle) {
1135 1136
        auto* info = reinterpret_cast<TensorInfo*>(handle);
        valid_tensors.insert(info);
1137
    }
1138
    return valid_tensors;
1139 1140
}

1141
void ChannelImpl::alloc_tensor_with_evict(OwnedBlob* x) {
1142
    bool in_worker = (get_worker_tid() == std::this_thread::get_id());
1143 1144 1145 1146 1147 1148
    auto reserve_size = [&](size_t size) {
        if (!m_dtr.comp_node.valid()) {
            return false;
        }
        while (size > m_dtr.comp_node.get_max_block_size_available()) {
            bool evict_suc = auto_evict(1);
M
Megvii Engine Team 已提交
1149 1150
            if (!evict_suc)
                return false;
1151 1152 1153 1154
        }
        return true;
    };
    auto pre_level = set_log_level(LogLevel::NO_LOG);
1155 1156 1157
    if (in_worker) {
        reserve_size(x->size());
    }
1158
    if (!BlobManager::inst()->try_alloc_direct(x, x->size())) {
1159
        bool suc = false;
1160 1161 1162 1163 1164
        if (in_worker) {
            while (!suc) {
                if (!auto_evict(1)) {
                    break;
                }
1165 1166 1167
                if (BlobManager::inst()->try_alloc_direct(x, x->size())) {
                    suc = true;
                }
1168 1169 1170 1171
            }
        }
        if (!suc) {
            set_log_level(pre_level);
M
Megvii Engine Team 已提交
1172 1173 1174
            mgb_log_warn(
                    "reallocating all cuda memory to alleviate fragmentation, the "
                    "performance may be affected");
1175
            set_log_level(LogLevel::NO_LOG);
1176
            imperative_log_profile_begin("defrag");
1177
            BlobManager::inst()->defrag(x->comp_node());
1178
            imperative_log_profile_end("defrag");
1179 1180 1181
            mgb_assert(
                    BlobManager::inst()->try_alloc_direct(x, x->size()),
                    "allocation failed after defrag");
1182
        }
1183
    }
1184 1185 1186
    set_log_level(pre_level);
}

1187
void ChannelImpl::process_one_task(Command& icmd) {
1188 1189
    using namespace ranges;
    using namespace ranges::views;
1190
    auto& state = get_worker_state();
1191
    auto& options = state.options;
M
Megvii Engine Team 已提交
1192
    // TODO: remove std::visit for support osx 10.12
1193
    auto cmd_visitor = [&](const auto& cmd) {
M
Megvii Engine Team 已提交
1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210
        using T = std::decay_t<decltype(cmd)>;
        if constexpr (std::is_same_v<T, Put>) {
            MGB_RECORD_EVENT(TensorCommandEvent, cmd.dest->id, TensorCommandKind::Put);
            MGB_RECORD_EVENT_IF(
                    (Profiler::get_option("profile_device", 0)), RecordDeviceEvent,
                    Timer::record_device(cmd.value.comp_node()));
            auto value = cmd.no_cache ? std::make_shared<Tensor>(cmd.value)
                                      : Tensor::make(cmd.value);
            MGB_RECORD_EVENT_IF(
                    (Profiler::get_option("profile_device", 0)), RecordDeviceEvent,
                    Timer::record_device(cmd.value.comp_node()));
            produce_tensor(cmd.dest, std::move(value));
            MGB_RECORD_EVENT(
                    TensorCommandFinishEvent, cmd.dest->id, TensorCommandKind::Put);
            sample_on_device(cmd.dest->desc.comp_node, false);
        } else if constexpr (std::is_same_v<T, ApplyOp>) {
            for (auto& i : cmd.inputs) {
1211
                if (mgb_unlikely(i->invalid)) {
M
Megvii Engine Team 已提交
1212 1213 1214
                    MGB_LOCK_GUARD(m_mutex);
                    for (auto& i : cmd.outputs) {
                        i->invalid = true;
1215
                    }
M
Megvii Engine Team 已提交
1216 1217 1218
                    return;
                }
            }
1219 1220 1221 1222 1223 1224 1225 1226
            if (state.options.enable_dtr_auto_drop) {
                m_apply_stack.push({cmd, 0, nullptr, "cmd"});
                flush_apply_stack();
                for (size_t i = 0; i < cmd.outputs.size(); ++i) {
                    auto output = cmd.outputs[i];
                    if (output == nullptr) {
                        continue;
                    }
M
Megvii Engine Team 已提交
1227
                    output->dsu_ptr = std::make_shared<DsuNode>(output->compute_time);
1228
                }
1229 1230
            } else {
                do_apply_op(cmd, "cmd");
M
Megvii Engine Team 已提交
1231 1232 1233 1234 1235 1236 1237
            }
            if (state.options.enable_drop && state.options.record_computing_path) {
                auto is_inplace = [](std::tuple<TensorInfo*, TensorInfo*> tuple2) {
                    auto& input = std::get<0>(tuple2);
                    auto& output = std::get<1>(tuple2);
                    if (!input->ptr || !output->ptr) {
                        return false;
1238
                    }
M
Megvii Engine Team 已提交
1239 1240 1241 1242 1243 1244 1245
                    return input->ptr->blob()->storage() ==
                           output->ptr->blob()->storage();
                };
                // FIXME: do not use opname as identifier
                auto get_name = [](const OpDef& opdef) {
                    if (auto attr = opdef.try_cast_final<OprAttr>()) {
                        return attr->type.c_str();
1246
                    }
M
Megvii Engine Team 已提交
1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259
                    return opdef.dyn_typeinfo()->name;
                };

                auto is_cross_cn = [comp_node = m_dtr.comp_node](TensorInfo* info) {
                    return info->desc.comp_node != comp_node;
                };

                bool cross_cn = any_of(concat(cmd.inputs, cmd.outputs), is_cross_cn);
                bool inplace =
                        any_of(cartesian_product(cmd.inputs, cmd.outputs), is_inplace);

                if (!inplace && !cross_cn && !m_dtr.is_bad_op(get_name(*cmd.op))) {
                    TensorInfo::ComputePath::make(
1260
                            cmd.id, cmd.op, cmd.inputs, cmd.outputs);
M
Megvii Engine Team 已提交
1261 1262
                    size_t detach_cnt = 0;
                    if (!strcmp(get_name(*cmd.op), "BatchNorm") &&
1263
                        cmd.outputs.size() == 6) {
M
Megvii Engine Team 已提交
1264 1265
                        cmd.outputs[0]->detach_producer();  // detach running_mean
                        cmd.outputs[1]->detach_producer();  // detach running_var
1266
                        for (auto input : cmd.inputs) {
M
Megvii Engine Team 已提交
1267
                            input->ref_cnt -= 2;
1268 1269
                        }
                    }
M
Megvii Engine Team 已提交
1270 1271 1272 1273 1274 1275 1276
                    for (auto output : cmd.outputs) {
                        if (output->producer &&
                            !output->size_exceeds_thd(
                                    state.options.dtr_evictee_minimum_size)) {
                            output->detach_producer();
                            detach_cnt++;
                        }
1277
                    }
M
Megvii Engine Team 已提交
1278 1279
                    for (auto input : cmd.inputs) {
                        input->ref_cnt -= detach_cnt;
1280
                    }
1281
                }
1282
            }
M
Megvii Engine Team 已提交
1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
        } else if constexpr (std::is_same_v<T, Del>) {
            MGB_RECORD_EVENT(TensorCommandEvent, cmd.dest->id, TensorCommandKind::Del);
            CompNode device = cmd.dest->desc.comp_node;
            uint64_t tensor_id = cmd.dest->id;
            free(cmd.dest);
            MGB_RECORD_EVENT(
                    TensorCommandFinishEvent, tensor_id, TensorCommandKind::Del);
            sample_on_device(device, false);
        } else if constexpr (std::is_same_v<T, GetValue>) {
            if (cmd.dest->invalid)
                return;
            imperative_log_profile_begin("GetValue");
            if (!cmd.dest->ptr && cmd.dest->evict_type != EvictType::NONE) {
                regenerate(cmd.dest);
            }
            cmd.dest->ptr->fetch_value();
1299
            MGB_LOCK_GUARD(m_mutex);
M
Megvii Engine Team 已提交
1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316
            notify_tensor_unsafe(cmd.dest);
            imperative_log_profile_end("GetValue");
        } else if constexpr (std::is_same_v<T, Drop>) {
            if (cmd.dest->invalid)
                return;
            MGB_RECORD_EVENT(TensorCommandEvent, cmd.dest->id, TensorCommandKind::Drop);
            do_drop(cmd.dest, true);
            MGB_RECORD_EVENT(
                    TensorCommandFinishEvent, cmd.dest->id, TensorCommandKind::Drop);
        } else if constexpr (std::is_same_v<T, SetOption>) {
            options.set_option(cmd.key, cmd.value);
        } else if constexpr (std::is_same_v<T, StartProfile>) {
            MGB_RECORD_EVENT(StartProfileEvent);
            CompNode::sync_all();
            for (auto* info : cmd.capture_tensors) {
                MGB_RECORD_EVENT(TensorDeclareEvent, info->id, info->name);
                if (info->status == TensorInfo::Produced) {
1317
                    // TODO: handle drop
M
Megvii Engine Team 已提交
1318 1319 1320
                    MGB_RECORD_EVENT(
                            TensorProduceEvent, info->id, info->desc.layout,
                            info->desc.comp_node, info->ptr->dev_tensor().raw_ptr());
1321 1322
                }
            }
M
Megvii Engine Team 已提交
1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337
            CompNode::foreach ([&](CompNode device) {
                sample_on_device(device, true);
                MGB_RECORD_EVENT_IF(
                        (Profiler::get_option("profile_device", 0)), RecordDeviceEvent,
                        Timer::record_device(device));
            });
            MGB_RECORD_EVENT(StartProfileFinishEvent);
        } else if constexpr (std::is_same_v<T, StopProfile>) {
            MGB_RECORD_EVENT(StopProfileEvent);
            for (auto* info : cmd.escape_tensors) {
                bool has_value = info->status == TensorInfo::Produced;
                if (has_value) {
                    MGB_RECORD_EVENT(TensorReleaseEvent, info->id);
                }
                MGB_RECORD_EVENT(TensorEraseEvent, info->id);
1338
            }
M
Megvii Engine Team 已提交
1339 1340 1341 1342 1343 1344 1345
            CompNode::foreach (
                    [&](CompNode device) { sample_on_device(device, true); });
            MGB_RECORD_EVENT(StopProfileFinishEvent);
        } else if constexpr (std::is_same_v<T, PushScope>) {
            MGB_RECORD_EVENT(ScopeEvent, cmd.scope_name);
        } else if constexpr (std::is_same_v<T, PopScope>) {
            MGB_RECORD_EVENT(ScopeFinishEvent, cmd.scope_name);
1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356
        } else if constexpr (std::is_same_v<T, StartRegen>) {
            if (cmd.dest->invalid)
                return;
            cmd.dest->pin();
            if (!cmd.dest->ptr && cmd.dest->evict_type != EvictType::NONE) {
                regenerate(cmd.dest);
            }
            MGB_LOCK_GUARD(m_mutex);
            notify_tensor_unsafe(cmd.dest);
        } else if constexpr (std::is_same_v<T, StopRegen>) {
            cmd.dest->unpin();
M
Megvii Engine Team 已提交
1357 1358
        } else {
            static_assert(!std::is_same_v<T, T>);
1359
        }
M
Megvii Engine Team 已提交
1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386
    };
    std::visit(
            [&](const auto& cmd) {
                using T = std::decay_t<decltype(cmd)>;
                if (!options.catch_worker_execption) {
                    cmd_visitor(cmd);
                    return;
                }
                try {
                    cmd_visitor(cmd);
                } catch (...) {
                    MGB_LOCK_GUARD(m_mutex);
                    if constexpr (std::is_same_v<T, ApplyOp>) {
                        for (auto oup : cmd.outputs) {
                            oup->invalid = true;
                        }
                    } else if constexpr (std::is_same_v<T, Put>) {
                        cmd.dest->invalid = true;
                    }
                    m_worker_exc = std::current_exception();
                    MGB_RECORD_EVENT(WorkerExceptionEvent);
                    if (m_waitee) {
                        notify_tensor_unsafe(m_waitee);
                    }
                }
            },
            icmd.data);
1387 1388 1389 1390
}

void ChannelImpl::check_worker_exc_unsafe() {
    if (m_worker_exc) {
1391 1392
        // for reuse interpreter_for_py after some exception tests
        m_waitee = nullptr;
1393 1394
        std::exception_ptr exc;
        std::swap(exc, m_worker_exc);
1395 1396 1397 1398 1399
        try {
            std::rethrow_exception(exc);
        } catch (...) {
            throw AsyncError();
        }
1400 1401
    }
}
1402

1403
void ChannelImpl::start_profile() {
1404
    MGB_LOCK_GUARD(m_spin);
1405
    assert_available();
1406 1407
    auto capture_tensors = collect_valid_tensors();
    if (capture_tensors.size() > 0) {
1408 1409 1410 1411 1412 1413 1414 1415 1416 1417
        if (Profiler::is_profiling()) {
            m_worker.add_task(
                    {Profiler::next_id(), StartProfile{std::move(capture_tensors)},
                     get_channel_state().stack_manager.dump()});
        } else {
            m_worker.add_task({
                    Profiler::next_id(),
                    StartProfile{std::move(capture_tensors)},
            });
        }
1418
    }
1419 1420
}

1421
void ChannelImpl::stop_profile() {
1422
    MGB_LOCK_GUARD(m_spin);
1423
    assert_available();
1424 1425
    auto escape_tensors = collect_valid_tensors();
    if (escape_tensors.size() > 0) {
1426 1427 1428 1429 1430 1431 1432 1433 1434 1435
        if (Profiler::is_profiling()) {
            m_worker.add_task(
                    {Profiler::next_id(), StopProfile{std::move(escape_tensors)},
                     get_channel_state().stack_manager.dump()});
        } else {
            m_worker.add_task({
                    Profiler::next_id(),
                    StopProfile{std::move(escape_tensors)},
            });
        }
1436
    }
1437 1438 1439
}

void ChannelImpl::push_scope(std::string name) {
1440
    MGB_LOCK_GUARD(m_spin);
1441
    assert_available();
1442
    auto& state = get_channel_state();
1443
    state.stack_manager.enter(name);
1444
    MGB_RECORD_EVENT(ScopeEvent, name);
1445 1446 1447 1448 1449 1450 1451 1452 1453 1454
    if (Profiler::is_profiling()) {
        m_worker.add_task(
                {Profiler::next_id(), PushScope{name},
                 get_channel_state().stack_manager.dump()});
    } else {
        m_worker.add_task({
                Profiler::next_id(),
                PushScope{name},
        });
    }
1455 1456 1457
}

void ChannelImpl::pop_scope(std::string name) {
1458
    MGB_LOCK_GUARD(m_spin);
1459
    assert_available();
1460
    auto& state = get_channel_state();
1461
    state.stack_manager.exit(name);
1462
    MGB_RECORD_EVENT(ScopeFinishEvent, name);
1463 1464 1465 1466 1467 1468 1469 1470 1471 1472
    if (Profiler::is_profiling()) {
        m_worker.add_task(
                {Profiler::next_id(), PopScope{name},
                 get_channel_state().stack_manager.dump()});
    } else {
        m_worker.add_task({
                Profiler::next_id(),
                PopScope{name},
        });
    }
1473 1474
}

1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486
BackTraceInfoPtr& ChannelImpl::get_backtrace() {
    return m_bt;
}

void ChannelImpl::set_backtrace(BackTraceInfoPtr bt) {
    m_bt = std::move(bt);
}

void ChannelImpl::clear_backtrace() {
    m_bt = nullptr;
}

1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511
bool ChannelImpl::worker_started() const {
    return m_worker.worker_started();
}

void ChannelImpl::update_status_to_forked(void) {
    MGB_LOCK_GUARD(m_spin);
    m_status = ChannelRunningStatus::FORKED;
}

void ChannelImpl::assert_available() const {
    if (m_status == ChannelRunningStatus::RUNING) {
        return;
    } else if (m_status == ChannelRunningStatus::CLOSED) {
        mgb_assert(false, "Channel already closed");
    } else if (m_status == ChannelRunningStatus::FORKED) {
        mgb_assert(
                false,
                "your program is forked and megengine is be disabled in subprocess, if "
                "you want to use megengine in subprocess, please DO NOT setup and use "
                "megengine before fork");
    } else {
        mgb_assert(false, "impossible, Channel status is undefined");
    }
}

1512
void ChannelImpl::assert_in_channel() {
M
Megvii Engine Team 已提交
1513 1514 1515
    mgb_assert(
            get_worker_tid() != std::this_thread::get_id(),
            "this method cannot be called in worker thread");
1516 1517 1518
}

void ChannelImpl::assert_in_worker() {
M
Megvii Engine Team 已提交
1519 1520 1521
    mgb_assert(
            get_worker_tid() == std::this_thread::get_id(),
            "this method can only be called in worker thread");
1522 1523
}

1524
void ChannelImpl::sample_on_device(CompNode device, bool force) {
1525 1526 1527
    if (!Profiler::is_profiling()) {
        return;
    }
1528 1529
    if (!force) {
        thread_local int last_sample_id = 0;
1530
        int sample_rate = Profiler::get_option("sample_rate", 0);
1531 1532 1533 1534
        if (!sample_rate || ((++last_sample_id) % sample_rate != 0)) {
            return;
        }
    }
1535
    MGB_RECORD_EVENT(SampleDeviceEvent, device);
1536
    auto [total, free] = device.get_mem_status_bytes();
1537
    MGB_RECORD_EVENT(SampleDeviceFinishEvent, device, total, free);
1538 1539
}

1540 1541 1542
void ChannelImpl::DynamicSublinear::pin(const SmallVector<TensorInfo*>& vec) {
    for (auto i : vec) {
        i->pin();
1543
        erase_candidate(i);
1544 1545 1546
    }
}

1547 1548
void ChannelImpl::DynamicSublinear::unpin(
        const SmallVector<TensorInfo*>& vec, WorkerState& state) {
1549 1550
    for (auto i : vec) {
        i->unpin();
1551 1552 1553 1554 1555
        if (i->pinned == 0 &&
            i->size_exceeds_thd(state.options.dtr_evictee_minimum_size) &&
            i->cand_index == UINT_MAX) {
            insert_candidate(i);
        }
1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601
    }
}

void ChannelImpl::DynamicSublinear::update_dsu_after_recompute(TensorInfo* ptr) {
    auto&& dsu_fa = find_father(ptr->dsu_ptr);
    dsu_fa->t -= ptr->compute_time;
    ptr->dsu_ptr->parent.reset();
    ptr->dsu_ptr->t = ptr->compute_time;
}

void ChannelImpl::DynamicSublinear::update_dsu_after_evict(TensorInfo* ptr) {
    for (auto i : ptr->producer->inputs) {
        if (i->evict_type == EvictType::DROP) {
            merge(i->dsu_ptr, ptr->dsu_ptr);
        }
    }
    for (auto i : ptr->producer->outputs) {
        if (i && i->evict_type == EvictType::DROP) {
            merge(ptr->dsu_ptr, i->dsu_ptr);
        }
    }
}

double ChannelImpl::DynamicSublinear::estimate_neighbor_cost(TensorInfo* ptr) {
    double cost = 0;
    for (auto i : ptr->producer->inputs) {
        if (i->evict_type == EvictType::DROP) {
            double t = find_father(i->dsu_ptr)->t;
            if (t < i->compute_time) {
                t = i->compute_time;
            }
            cost += t;
        }
    }
    for (auto i : ptr->producer->outputs) {
        if (i && i->evict_type == EvictType::DROP) {
            double t = find_father(i->dsu_ptr)->t;
            if (t < i->compute_time) {
                t = i->compute_time;
            }
            cost += t;
        }
    }
    return cost;
}

M
Megvii Engine Team 已提交
1602 1603
TensorInfo* ChannelImpl::DynamicSublinear::find_best_tensor(
        bool enable_dtr_sqrt_sampling = false) {
1604 1605 1606
    if (candidates.empty())
        return nullptr;

1607 1608
    double min_msps = -1;
    TensorInfo* best = nullptr;
1609 1610
    size_t sz = 1;
    if (enable_dtr_sqrt_sampling) {
M
Megvii Engine Team 已提交
1611 1612
        while (sz * sz <= candidates.size())
            sz++;
1613
        sz--;
1614 1615 1616
    } else {
        sz = candidates.size();
    }
1617 1618 1619 1620 1621 1622 1623

    size_t ti = rand() % sz;
    for (size_t vi = 0; vi < sz; vi++) {
        if (!enable_dtr_sqrt_sampling) {
            ti = vi;
        }
        auto i = candidates[ti];
1624
        if (i->producer && i->ptr && i->evict_type == EvictType::NONE) {
1625
            double neighbor_cost = estimate_neighbor_cost(i);
M
Megvii Engine Team 已提交
1626 1627 1628 1629
            size_t begin_ptr =
                    reinterpret_cast<size_t>(i->ptr->blob()->storage().get());
            auto side_info = i->ptr->comp_node().get_free_left_and_right(
                    begin_ptr, begin_ptr + i->ptr->blob()->size());
1630
            double free_mem = side_info.first + side_info.second;
M
Megvii Engine Team 已提交
1631 1632
            double msps = i->eval_func(
                    neighbor_cost, free_mem, estimate_timestamp, 1.0, 1.0, 1.0, 1.0001);
1633 1634 1635 1636 1637
            if (min_msps < 0 || msps < min_msps) {
                min_msps = msps;
                best = i;
            }
        }
1638 1639 1640 1641 1642
        if (enable_dtr_sqrt_sampling) {
            ti += rand() % sz;
            if (ti > candidates.size())
                break;
        }
1643 1644 1645 1646
    }
    return best;
}

M
Megvii Engine Team 已提交
1647 1648
void ChannelImpl::DynamicSublinear::merge(
        std::shared_ptr<DsuNode>& x, std::shared_ptr<DsuNode>& y) {
1649 1650 1651 1652 1653 1654 1655 1656 1657
    auto&& f_x = find_father(x);
    auto&& f_y = find_father(y);
    if (f_x.get() == f_y.get()) {
        return;
    }
    f_y->t += f_x->t;
    f_x->parent = f_y;
}

M
Megvii Engine Team 已提交
1658 1659
std::shared_ptr<DsuNode> ChannelImpl::DynamicSublinear::find_father(
        std::shared_ptr<DsuNode>& x) {
1660 1661 1662 1663 1664 1665 1666 1667 1668
    if (x->is_root()) {
        return x;
    } else {
        auto&& fa = find_father(x->parent);
        return x->parent = fa;
    }
}

void ChannelImpl::DynamicSublinear::insert_candidate(TensorInfo* ptr) {
1669 1670 1671 1672 1673 1674
    // tensor to be inserted must be brand new
    mgb_assert(
            ptr->cand_index == UINT_MAX, "got wrong candidate index : %lu",
            ptr->cand_index);
    ptr->cand_index = candidates.size();
    candidates.push_back(ptr);
1675 1676 1677 1678 1679 1680
    if (!comp_node.valid()) {
        comp_node = ptr->ptr->comp_node();
    }
}

void ChannelImpl::DynamicSublinear::erase_candidate(TensorInfo* ptr) {
1681 1682 1683 1684 1685 1686
    // close dtr will just clear candidates, so nothing to erase
    if (candidates.empty()) {
        ptr->cand_index = UINT_MAX;
        return;
    }
    // some tensors may be erased already, just skip them
1687 1688 1689 1690 1691 1692
    if (ptr->cand_index != UINT_MAX) {
        std::swap(candidates[ptr->cand_index], candidates.back());
        candidates[ptr->cand_index]->cand_index = ptr->cand_index;
        candidates.pop_back();
        ptr->cand_index = UINT_MAX;
    }
1693 1694 1695 1696 1697
}

void ChannelImpl::DynamicSublinear::update_used_time(TensorInfo* ptr) {
    ptr->last_used_time = estimate_timestamp;
}