npu_info.cc 13.5 KB
Newer Older
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
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.

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. */

#include "paddle/fluid/platform/npu_info.h"
#include <algorithm>
#include <cstdlib>
#include <memory>

#include "gflags/gflags.h"

#include "paddle/fluid/platform/lock_guard_ptr.h"
#include "paddle/fluid/platform/macros.h"
#include "paddle/fluid/platform/monitor.h"
#include "paddle/fluid/string/split.h"

DECLARE_double(fraction_of_gpu_memory_to_use);
DECLARE_uint64(initial_gpu_memory_in_mb);
DECLARE_uint64(reallocate_gpu_memory_in_mb);
DECLARE_bool(enable_cublas_tensor_op_math);
DECLARE_uint64(gpu_memory_limit_mb);
L
Leo Chen 已提交
32
DECLARE_string(selected_npus);
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51

constexpr static float fraction_reserve_gpu_memory = 0.05f;

USE_NPU_MEM_STAT;

namespace paddle {
namespace platform {

static int GetNPUDeviceCountImpl() {
  uint32_t count;
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtGetDeviceCount(&count));
  return count;
}

int GetNPUDeviceCount() {
  static auto dev_cnt = GetNPUDeviceCountImpl();
  return dev_cnt;
}

52 53 54 55 56 57
int NPUCanAccessPeer(int src, int dst) {
  int can = 0;
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtDeviceCanAccessPeer(&can, src, dst));
  return can;
}

58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80
// For example, "1.0.1"
std::string GetNPURuntimeVersion(int id) {
  PADDLE_ENFORCE_LT(id, GetNPUDeviceCount(),
                    platform::errors::InvalidArgument(
                        "Device id must be less than NPU count, "
                        "but received id is: %d. NPU count is: %d.",
                        id, GetNPUDeviceCount()));
  int major = 0, minor = 0, patch = 0;
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtGetVersion(&major, &minor, &patch));
  return string::Sprintf("%d.%d.%d", major, minor, patch);
}

int GetCurrentNPUDeviceId() {
  int device_id;
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtGetDevice(&device_id));
  return device_id;
}

//! Get a list of device ids from environment variable or use all.
std::vector<int> GetSelectedNPUDevices() {
  // use user specified NPUs in single-node multi-process mode.
  std::vector<int> devices;
  if (!FLAGS_selected_gpus.empty()) {
L
Leo Chen 已提交
81
    auto devices_str = paddle::string::Split(FLAGS_selected_npus, ',');
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
    for (auto id : devices_str) {
      devices.push_back(atoi(id.c_str()));
    }
  } else {
    int count = GetNPUDeviceCount();
    for (int i = 0; i < count; ++i) {
      devices.push_back(i);
    }
  }
  return devices;
}

void SetNPUDeviceId(int id) {
  PADDLE_ENFORCE_LT(id, GetNPUDeviceCount(),
                    platform::errors::InvalidArgument(
                        "Device id must be less than NPU count, "
                        "but received id is: %d. NPU count is: %d.",
                        id, GetNPUDeviceCount()));
  // NOTE(zihqiu): It is recommended to call aclrtSetDevice and aclrtResetDevice
  // pairly.
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtSetDevice(id));
}

void ResetNPUDeviceId(int id) {
  PADDLE_ENFORCE_LT(id, GetNPUDeviceCount(),
                    platform::errors::InvalidArgument(
                        "Device id must be less than NPU count, "
                        "but received id is: %d. NPU count is: %d.",
                        id, GetNPUDeviceCount()));
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtResetDevice(id));
}

void NPUMemoryUsage(size_t *available, size_t *total) {
  size_t actual_available, actual_total;
  RecordedNPUMemGetInfo(available, total, &actual_available, &actual_total,
                        platform::GetCurrentNPUDeviceId());
}

size_t NPUAvailableMemToAlloc() {
  size_t total = 0;
  size_t available = 0;
  NPUMemoryUsage(&available, &total);
  size_t reserving =
      static_cast<size_t>(fraction_reserve_gpu_memory * available);
  // If available size is less than minimum chunk size, no usable memory exists
  size_t available_to_alloc = available - reserving;
  size_t min_chunk_size = NPUMinChunkSize();
  if (available_to_alloc < min_chunk_size) {
    available_to_alloc = 0;
  }
  VLOG(10) << "NPU usage " << (available >> 20) << "M/" << (total >> 20)
           << "M, " << (available_to_alloc >> 20) << "M available to allocate";
  return available_to_alloc;
}

size_t NPUMaxAllocSize() {
  return std::max(NPUInitAllocSize(), NPUReallocSize());
}

static size_t NPUAllocSize(bool realloc) {
  size_t available_to_alloc = NPUAvailableMemToAlloc();
  PADDLE_ENFORCE_GT(
      available_to_alloc, 0,
      platform::errors::ResourceExhausted("Not enough available NPU memory."));
  // If FLAGS_initial_gpu_memory_in_mb is 0, then initial memory will be
  // allocated by fraction
  size_t flag_mb = realloc ? FLAGS_reallocate_gpu_memory_in_mb
                           : FLAGS_initial_gpu_memory_in_mb;
  size_t alloc_bytes =
      (flag_mb > 0ul ? flag_mb << 20 : available_to_alloc *
                                           FLAGS_fraction_of_gpu_memory_to_use);
  PADDLE_ENFORCE_GE(
      available_to_alloc, alloc_bytes,
      platform::errors::ResourceExhausted("Not enough available NPU memory."));
  VLOG(10) << "Alloc size is " << (alloc_bytes >> 20)
           << " MiB, is it Re-alloc: " << realloc;
  return alloc_bytes;
}

size_t NPUInitAllocSize() { return NPUAllocSize(/* realloc = */ false); }

size_t NPUReallocSize() { return NPUAllocSize(/* realloc = */ true); }

size_t NPUMinChunkSize() {
  // Allow to allocate the minimum chunk size is 256 bytes.
  return 1 << 8;
}

size_t NPUMaxChunkSize() {
  size_t max_chunk_size = NPUMaxAllocSize();
  VLOG(10) << "Max chunk size " << (max_chunk_size >> 20) << "M";
  return max_chunk_size;
}

176
void NPUMemcpyAsync(void *dst, const void *src, size_t count,
177 178 179
                    enum aclrtMemcpyKind kind, aclrtStream stream,
                    size_t dst_max_count) {
  dst_max_count = dst_max_count ? dst_max_count : count;
180 181
  VLOG(4) << dst << " " << dst_max_count << " " << src << " " << count << " "
          << kind << " " << stream;
182 183 184 185 186 187 188 189 190
  PADDLE_ENFORCE_NPU_SUCCESS(
      aclrtMemcpyAsync(dst, dst_max_count, src, count, kind, stream));
}

void NPUMemcpySync(void *dst, const void *src, size_t count,
                   enum aclrtMemcpyKind kind, size_t dst_max_count) {
  // NOTE(zhiqiu):  The default max_count is count
  dst_max_count = dst_max_count ? dst_max_count : count;
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtMemcpy(dst, dst_max_count, src, count, kind));
191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
}

void NPUMemcpyPeerASync(void *dst, int dst_device, const void *src,
                        size_t count, enum aclrtMemcpyKind kind,
                        aclrtStream stream, size_t dst_max_count) {
  dst_max_count = dst_max_count ? dst_max_count : count;
  PADDLE_ENFORCE_NPU_SUCCESS(
      aclrtMemcpyAsync(dst, dst_max_count, src, count, kind, stream));
}

void NPUMemcpyPeerSync(void *dst, int dst_device, const void *src, size_t count,
                       enum aclrtMemcpyKind kind, size_t dst_max_count) {
  // NOTE(zhiqiu):  The default max_count is count
  dst_max_count = dst_max_count ? dst_max_count : count;
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtMemcpy(dst, dst_max_count, src, count, kind));
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
}

void NPUMemsetAsync(void *dst, int value, size_t count, aclrtStream stream,
                    size_t max_count) {
  max_count = max_count ? max_count : count;
  PADDLE_ENFORCE_NPU_SUCCESS(
      aclrtMemsetAsync(dst, max_count, value, count, stream));
}

void NPUStreamSync(aclrtStream stream) {
  PADDLE_ENFORCE_NPU_SUCCESS(aclrtSynchronizeStream(stream));
}

static void RaiseNonOutOfMemoryError(aclError *status) {
  if (*status == ACL_ERROR_BAD_ALLOC) {
    *status = ACL_ERROR_NONE;
  }
  PADDLE_ENFORCE_NPU_SUCCESS(*status);
}

class RecordedNPUMallocHelper {
 private:
  explicit RecordedNPUMallocHelper(int dev_id, uint64_t limit_size = 0)
      : dev_id_(dev_id), limit_size_(limit_size) {
    if (NeedRecord()) {
      mtx_.reset(new std::mutex());
    }
  }

  DISABLE_COPY_AND_ASSIGN(RecordedNPUMallocHelper);

 public:
  static RecordedNPUMallocHelper *Instance(int dev_id) {
    std::call_once(once_flag_, [] {
      int dev_cnt = GetNPUDeviceCount();
      instances_.reserve(dev_cnt);
      for (int i = 0; i < dev_cnt; ++i) {
        // NOTE(zhiqiu): share the flags with gpu, avoid more flags.
        instances_.emplace_back(
            new RecordedNPUMallocHelper(i, FLAGS_gpu_memory_limit_mb << 20));
      }
    });

    PADDLE_ENFORCE_GE(
        dev_id, 0,
        platform::errors::OutOfRange(
            "Device id must be not less than 0, but got %d.", dev_id));
    PADDLE_ENFORCE_LT(
        dev_id, instances_.size(),
        platform::errors::OutOfRange("Device id %d exceeds npu card number %d.",
                                     dev_id, instances_.size()));
    return instances_[dev_id].get();
  }

  /**
   * Try to allocate `size` npu memory. Only ACL_ERROR_BAD_ALLOC
   * or ACL_ERROR_NONE would be returned.
   */
  aclError Malloc(void **ptr, size_t size) {
    LockGuardPtr<std::mutex> lock(mtx_);
    if (UNLIKELY(NeedRecord() && cur_size_ + size > limit_size_)) {
      return ACL_ERROR_BAD_ALLOC;
    }

    NPUDeviceGuard guard(dev_id_);
    auto result = aclrtMalloc(ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
    if (result == ACL_ERROR_NONE) {
      if (NeedRecord()) {
        cur_size_ += size;
      }
      STAT_INT_ADD("STAT_npu" + std::to_string(dev_id_) + "_mem_size", size);
      return result;
    } else {
      RaiseNonOutOfMemoryError(&result);
      // Non out of memory error would be raised inside
      // RaiseNonOutOfMemoryError. Therefore, we can
      // return cudaErrorMemoryAllocation directly here.
      return ACL_ERROR_BAD_ALLOC;
    }
  }

  /**
   * Free gpu memory. Usually, free is not allowed to raise error.
   * If it does raise error, the process should be crashed.
   */
  void Free(void *ptr, size_t size) {
    NPUDeviceGuard guard(dev_id_);
    auto result = aclrtFree(ptr);
    PADDLE_ENFORCE_NPU_SUCCESS(result);
    if (NeedRecord()) {
      std::lock_guard<std::mutex> guard(*mtx_);
      cur_size_ -= size;
    }
    STAT_INT_SUB("STAT_npu" + std::to_string(dev_id_) + "_mem_size", size);
  }

  bool GetMemInfo(size_t *avail, size_t *total, size_t *actual_avail,
                  size_t *actual_total) {
    {
      NPUDeviceGuard guard(dev_id_);
      auto result = aclrtGetMemInfo(ACL_HBM_MEM, actual_avail, actual_total);
      if (result != ACL_ERROR_NONE) {
        *actual_avail = 0;
      }
      RaiseNonOutOfMemoryError(&result);
    }

    if (NeedRecord()) {
      std::lock_guard<std::mutex> guard(*mtx_);
      *avail = std::min(*actual_avail, limit_size_ - cur_size_);
      *total = std::min(*actual_total, limit_size_);
      return *total < *actual_total;
    } else {
      *avail = *actual_avail;
      *total = *actual_total;
      return false;
    }
  }

  inline bool NeedRecord() const { return limit_size_ != 0; }

  uint64_t RecordedSize() const {
    LockGuardPtr<std::mutex> lock(mtx_);
    return NeedRecord() ? cur_size_ : 0;
  }

  uint64_t LimitSize() const { return limit_size_; }

 private:
  const int dev_id_;
  const uint64_t limit_size_;
  uint64_t cur_size_{0};

  mutable std::unique_ptr<std::mutex> mtx_;

  static std::once_flag once_flag_;
  static std::vector<std::unique_ptr<RecordedNPUMallocHelper>> instances_;
};

std::once_flag RecordedNPUMallocHelper::once_flag_;
std::vector<std::unique_ptr<RecordedNPUMallocHelper>>
    RecordedNPUMallocHelper::instances_;

aclError RecordedNPUMalloc(void **ptr, size_t size, int dev_id) {
  return RecordedNPUMallocHelper::Instance(dev_id)->Malloc(ptr, size);
}

void RecordedNPUFree(void *p, size_t size, int dev_id) {
  return RecordedNPUMallocHelper::Instance(dev_id)->Free(p, size);
}

bool RecordedNPUMemGetInfo(size_t *avail, size_t *total, size_t *actual_avail,
                           size_t *actual_total, int dev_id) {
  return RecordedNPUMallocHelper::Instance(dev_id)->GetMemInfo(
      avail, total, actual_avail, actual_total);
}

uint64_t RecordedNPUMallocSize(int dev_id) {
  return RecordedNPUMallocHelper::Instance(dev_id)->RecordedSize();
}

bool IsNPUMallocRecorded(int dev_id) {
  return RecordedNPUMallocHelper::Instance(dev_id)->NeedRecord();
}

L
Leo Chen 已提交
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
AclInstance::~AclInstance() {}

AclInstance &AclInstance::Instance() {
  static AclInstance instance;
  return instance;
}

AclInstance::AclInstance() {
  PADDLE_ENFORCE_NPU_SUCCESS(aclInit(nullptr));
  VLOG(4) << "Call aclrtSetDevice ";
  // NOTE(zhiqiu): why set devices here?
  // Because ACL creates a default context which contains 2 streams
  // when calling aclrtSetDeviceId, so usually we do not need to
  // create contexts explicitly. And, for each device, aclrtSetDeviceId
  // need to call parily with aclrtResetDeviceId to destory the default
  // context. Here, we use this singleton and static instance to manage
  // the devices to make sure they will be resetted before program exit.
  devices_ = platform::GetSelectedNPUDevices();
  for (auto it = devices_.rbegin(); it != devices_.rend(); ++it) {
    SetNPUDeviceId(*it);
    VLOG(4) << "Call aclrtSetDevice " << *it;
  }
}

void AclInstance::Finalize() {
  // NOTE(zhiqiu): DO NOT perform finalize in destructor
  // to avoid problems caused by destructor order of static
  // object.
  for (size_t i = 0; i < devices_.size(); ++i) {
    auto status = aclrtResetDevice(devices_[i]);
    VLOG(4) << "Call aclrtResetDevice " << devices_[i]
            << " status = " << status;
  }
  auto status = aclFinalize();
  VLOG(4) << "Call aclFinalize, status = " << status;
}

408 409
}  // namespace platform
}  // namespace paddle