CachingDeviceAllocator.h

Go to the documentation of this file.

#ifndef HeterogenousCore_CUDAUtilities_src_CachingDeviceAllocator_h
#define HeterogenousCore_CUDAUtilities_src_CachingDeviceAllocator_h

/******************************************************************************
 * Copyright (c) 2011, Duane Merrill.  All rights reserved.
 * Copyright (c) 2011-2018, NVIDIA CORPORATION.  All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *     * Redistributions of source code must retain the above copyright
 *       notice, this list of conditions and the following disclaimer.
 *     * Redistributions in binary form must reproduce the above copyright
 *       notice, this list of conditions and the following disclaimer in the
 *       documentation and/or other materials provided with the distribution.
 *     * Neither the name of the NVIDIA CORPORATION nor the
 *       names of its contributors may be used to endorse or promote products
 *       derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 ******************************************************************************/

/******************************************************************************
 * Simple caching allocator for device memory allocations. The allocator is
 * thread-safe and capable of managing device allocations on multiple devices.
 ******************************************************************************/

#include <cmath>
#include <map>
#include <set>
#include <mutex>

#include "HeterogeneousCore/CUDAUtilities/interface/cudaCheck.h"
#include "HeterogeneousCore/CUDAUtilities/interface/deviceAllocatorStatus.h"

namespace notcub {

  /******************************************************************************
 * CachingDeviceAllocator (host use)
 ******************************************************************************/

  struct CachingDeviceAllocator {
    //---------------------------------------------------------------------
    // Constants
    //---------------------------------------------------------------------

    static const unsigned int INVALID_BIN = (unsigned int)-1;

    static const size_t INVALID_SIZE = (size_t)-1;

#ifndef DOXYGEN_SHOULD_SKIP_THIS  // Do not document

    static const int INVALID_DEVICE_ORDINAL = -1;

    //---------------------------------------------------------------------
    // Type definitions and helper types
    //---------------------------------------------------------------------

    struct BlockDescriptor {
      void *d_ptr;                     // Device pointer
      size_t bytes;                    // Size of allocation in bytes
      size_t bytesRequested;           // CMS: requested allocatoin size (for monitoring only)
      unsigned int bin;                // Bin enumeration
      int device;                      // device ordinal
      cudaStream_t associated_stream;  // Associated associated_stream
      cudaEvent_t ready_event;  // Signal when associated stream has run to the point at which this block was freed

      // Constructor (suitable for searching maps for a specific block, given its pointer and device)
      BlockDescriptor(void *d_ptr, int device)
          : d_ptr(d_ptr),
            bytes(0),
            bytesRequested(0),  // CMS
            bin(INVALID_BIN),
            device(device),
            associated_stream(nullptr),
            ready_event(nullptr) {}

      // Constructor (suitable for searching maps for a range of suitable blocks, given a device)
      BlockDescriptor(int device)
          : d_ptr(nullptr),
            bytes(0),
            bytesRequested(0),  // CMS
            bin(INVALID_BIN),
            device(device),
            associated_stream(nullptr),
            ready_event(nullptr) {}

      // Comparison functor for comparing device pointers
      static bool PtrCompare(const BlockDescriptor &a, const BlockDescriptor &b) {
        if (a.device == b.device)
          return (a.d_ptr < b.d_ptr);
        else
          return (a.device < b.device);
      }

      // Comparison functor for comparing allocation sizes
      static bool SizeCompare(const BlockDescriptor &a, const BlockDescriptor &b) {
        if (a.device == b.device)
          return (a.bytes < b.bytes);
        else
          return (a.device < b.device);
      }
    };

    typedef bool (*Compare)(const BlockDescriptor &, const BlockDescriptor &);

    // CMS: Moved TotalBytes to deviceAllocatorStatus.h

    typedef std::multiset<BlockDescriptor, Compare> CachedBlocks;

    typedef std::multiset<BlockDescriptor, Compare> BusyBlocks;

    // CMS: Moved definition to deviceAllocatorStatus.h
    using GpuCachedBytes = cms::cuda::allocator::GpuCachedBytes;

    //---------------------------------------------------------------------
    // Utility functions
    //---------------------------------------------------------------------

    static unsigned int IntPow(unsigned int base, unsigned int exp) {
      unsigned int retval = 1;
      while (exp > 0) {
        if (exp & 1) {
          retval = retval * base;  // multiply the result by the current base
        }
        base = base * base;  // square the base
        exp = exp >> 1;      // divide the exponent in half
      }
      return retval;
    }

    void NearestPowerOf(unsigned int &power, size_t &rounded_bytes, unsigned int base, size_t value) {
      power = 0;
      rounded_bytes = 1;

      if (value * base < value) {
        // Overflow
        power = sizeof(size_t) * 8;
        rounded_bytes = size_t(0) - 1;
        return;
      }

      while (rounded_bytes < value) {
        rounded_bytes *= base;
        power++;
      }
    }

    //---------------------------------------------------------------------
    // Fields
    //---------------------------------------------------------------------

    // CMS: use std::mutex instead of cub::Mutex, declare mutable
    mutable std::mutex mutex;  

    unsigned int bin_growth;  
    unsigned int min_bin;     
    unsigned int max_bin;     

    size_t min_bin_bytes;     
    size_t max_bin_bytes;     
    size_t max_cached_bytes;  

    const bool
        skip_cleanup;  
    bool debug;        

    GpuCachedBytes cached_bytes;  
    CachedBlocks cached_blocks;   
    BusyBlocks live_blocks;       

#endif  // DOXYGEN_SHOULD_SKIP_THIS

    //---------------------------------------------------------------------
    // Methods
    //---------------------------------------------------------------------

    CachingDeviceAllocator(
        unsigned int bin_growth,                 
        unsigned int min_bin = 1,                
        unsigned int max_bin = INVALID_BIN,      
        size_t max_cached_bytes = INVALID_SIZE,  
        bool skip_cleanup =
            false,  
        bool debug = false)  
        : bin_growth(bin_growth),
          min_bin(min_bin),
          max_bin(max_bin),
          min_bin_bytes(IntPow(bin_growth, min_bin)),
          max_bin_bytes(IntPow(bin_growth, max_bin)),
          max_cached_bytes(max_cached_bytes),
          skip_cleanup(skip_cleanup),
          debug(debug),
          cached_blocks(BlockDescriptor::SizeCompare),
          live_blocks(BlockDescriptor::PtrCompare) {}

    CachingDeviceAllocator(bool skip_cleanup = false, bool debug = false)
        : bin_growth(8),
          min_bin(3),
          max_bin(7),
          min_bin_bytes(IntPow(bin_growth, min_bin)),
          max_bin_bytes(IntPow(bin_growth, max_bin)),
          max_cached_bytes((max_bin_bytes * 3) - 1),
          skip_cleanup(skip_cleanup),
          debug(debug),
          cached_blocks(BlockDescriptor::SizeCompare),
          live_blocks(BlockDescriptor::PtrCompare) {}

    cudaError_t SetMaxCachedBytes(size_t max_cached_bytes) {
      // Lock
      // CMS: use RAII instead of (un)locking explicitly
      std::unique_lock mutex_locker(mutex);

      if (debug)
        // CMS: use raw printf
        printf("Changing max_cached_bytes (%lld -> %lld)\n",
               (long long)this->max_cached_bytes,
               (long long)max_cached_bytes);

      this->max_cached_bytes = max_cached_bytes;

      // Unlock (redundant, kept for style uniformity)
      mutex_locker.unlock();

      return cudaSuccess;
    }

    cudaError_t DeviceAllocate(
        int device,                            
        void **d_ptr,                          
        size_t bytes,                          
        cudaStream_t active_stream = nullptr)  
    {
      // CMS: use RAII instead of (un)locking explicitly
      std::unique_lock<std::mutex> mutex_locker(mutex, std::defer_lock);
      *d_ptr = nullptr;
      int entrypoint_device = INVALID_DEVICE_ORDINAL;
      cudaError_t error = cudaSuccess;

      if (device == INVALID_DEVICE_ORDINAL) {
        // CMS: throw exception on error
        cudaCheck(error = cudaGetDevice(&entrypoint_device));
        device = entrypoint_device;
      }

      // Create a block descriptor for the requested allocation
      bool found = false;
      BlockDescriptor search_key(device);
      search_key.bytesRequested = bytes;  // CMS
      search_key.associated_stream = active_stream;
      NearestPowerOf(search_key.bin, search_key.bytes, bin_growth, bytes);

      if (search_key.bin > max_bin) {
        // Bin is greater than our maximum bin: allocate the request
        // exactly and give out-of-bounds bin.  It will not be cached
        // for reuse when returned.
        search_key.bin = INVALID_BIN;
        search_key.bytes = bytes;
      } else {
        // Search for a suitable cached allocation: lock
        mutex_locker.lock();

        if (search_key.bin < min_bin) {
          // Bin is less than minimum bin: round up
          search_key.bin = min_bin;
          search_key.bytes = min_bin_bytes;
        }

        // Iterate through the range of cached blocks on the same device in the same bin
        CachedBlocks::iterator block_itr = cached_blocks.lower_bound(search_key);
        while ((block_itr != cached_blocks.end()) && (block_itr->device == device) &&
               (block_itr->bin == search_key.bin)) {
          // To prevent races with reusing blocks returned by the host but still
          // in use by the device, only consider cached blocks that are
          // either (from the active stream) or (from an idle stream)
          if ((active_stream == block_itr->associated_stream) ||
              (cudaEventQuery(block_itr->ready_event) != cudaErrorNotReady)) {
            // Reuse existing cache block.  Insert into live blocks.
            found = true;
            search_key = *block_itr;
            search_key.associated_stream = active_stream;
            live_blocks.insert(search_key);

            // Remove from free blocks
            cached_bytes[device].free -= search_key.bytes;
            cached_bytes[device].live += search_key.bytes;
            cached_bytes[device].liveRequested += search_key.bytesRequested;  // CMS

            if (debug)
              // CMS: improved debug message
              // CMS: use raw printf
              printf(
                  "\tDevice %d reused cached block at %p (%lld bytes) for stream %lld, event %lld (previously "
                  "associated with stream %lld, event %lld).\n",
                  device,
                  search_key.d_ptr,
                  (long long)search_key.bytes,
                  (long long)search_key.associated_stream,
                  (long long)search_key.ready_event,
                  (long long)block_itr->associated_stream,
                  (long long)block_itr->ready_event);

            cached_blocks.erase(block_itr);

            break;
          }
          block_itr++;
        }

        // Done searching: unlock
        mutex_locker.unlock();
      }

      // Allocate the block if necessary
      if (!found) {
        // Set runtime's current device to specified device (entrypoint may not be set)
        if (device != entrypoint_device) {
          // CMS: throw exception on error
          cudaCheck(error = cudaGetDevice(&entrypoint_device));
          cudaCheck(error = cudaSetDevice(device));
        }

        // Attempt to allocate
        // CMS: silently ignore errors and retry or pass them to the caller
        if ((error = cudaMalloc(&search_key.d_ptr, search_key.bytes)) == cudaErrorMemoryAllocation) {
          // The allocation attempt failed: free all cached blocks on device and retry
          if (debug)
            // CMS: use raw printf
            printf(
                "\tDevice %d failed to allocate %lld bytes for stream %lld, retrying after freeing cached allocations",
                device,
                (long long)search_key.bytes,
                (long long)search_key.associated_stream);

          error = cudaSuccess;  // Reset the error we will return
          cudaGetLastError();   // Reset CUDART's error

          // Lock
          mutex_locker.lock();

          // Iterate the range of free blocks on the same device
          BlockDescriptor free_key(device);
          CachedBlocks::iterator block_itr = cached_blocks.lower_bound(free_key);

          while ((block_itr != cached_blocks.end()) && (block_itr->device == device)) {
            // No need to worry about synchronization with the device: cudaFree is
            // blocking and will synchronize across all kernels executing
            // on the current device

            // Free device memory and destroy stream event.
            // CMS: silently ignore errors and pass them to the caller
            if ((error = cudaFree(block_itr->d_ptr)))
              break;
            if ((error = cudaEventDestroy(block_itr->ready_event)))
              break;

            // Reduce balance and erase entry
            cached_bytes[device].free -= block_itr->bytes;

            if (debug)
              // CMS: use raw printf
              printf(
                  "\tDevice %d freed %lld bytes.\n\t\t  %lld available blocks cached (%lld bytes), %lld live blocks "
                  "(%lld bytes) outstanding.\n",
                  device,
                  (long long)block_itr->bytes,
                  (long long)cached_blocks.size(),
                  (long long)cached_bytes[device].free,
                  (long long)live_blocks.size(),
                  (long long)cached_bytes[device].live);

            cached_blocks.erase(block_itr);

            block_itr++;
          }

          // Unlock
          mutex_locker.unlock();

          // Return under error
          if (error)
            return error;

          // Try to allocate again
          // CMS: throw exception on error
          cudaCheck(error = cudaMalloc(&search_key.d_ptr, search_key.bytes));
        }

        // Create ready event
        // CMS: throw exception on error
        cudaCheck(error = cudaEventCreateWithFlags(&search_key.ready_event, cudaEventDisableTiming));

        // Insert into live blocks
        mutex_locker.lock();
        live_blocks.insert(search_key);
        cached_bytes[device].live += search_key.bytes;
        cached_bytes[device].liveRequested += search_key.bytesRequested;  // CMS
        mutex_locker.unlock();

        if (debug)
          // CMS: improved debug message
          // CMS: use raw printf
          printf("\tDevice %d allocated new device block at %p (%lld bytes associated with stream %lld, event %lld).\n",
                 device,
                 search_key.d_ptr,
                 (long long)search_key.bytes,
                 (long long)search_key.associated_stream,
                 (long long)search_key.ready_event);

        // Attempt to revert back to previous device if necessary
        if ((entrypoint_device != INVALID_DEVICE_ORDINAL) && (entrypoint_device != device)) {
          // CMS: throw exception on error
          cudaCheck(error = cudaSetDevice(entrypoint_device));
        }
      }

      // Copy device pointer to output parameter
      *d_ptr = search_key.d_ptr;

      if (debug)
        // CMS: use raw printf
        printf("\t\t%lld available blocks cached (%lld bytes), %lld live blocks outstanding(%lld bytes).\n",
               (long long)cached_blocks.size(),
               (long long)cached_bytes[device].free,
               (long long)live_blocks.size(),
               (long long)cached_bytes[device].live);

      return error;
    }

    cudaError_t DeviceAllocate(
        void **d_ptr,                          
        size_t bytes,                          
        cudaStream_t active_stream = nullptr)  
    {
      return DeviceAllocate(INVALID_DEVICE_ORDINAL, d_ptr, bytes, active_stream);
    }

    cudaError_t DeviceFree(int device, void *d_ptr) {
      int entrypoint_device = INVALID_DEVICE_ORDINAL;
      cudaError_t error = cudaSuccess;
      // CMS: use RAII instead of (un)locking explicitly
      std::unique_lock<std::mutex> mutex_locker(mutex, std::defer_lock);

      if (device == INVALID_DEVICE_ORDINAL) {
        // CMS: throw exception on error
        cudaCheck(error = cudaGetDevice(&entrypoint_device));
        device = entrypoint_device;
      }

      // Lock
      mutex_locker.lock();

      // Find corresponding block descriptor
      bool recached = false;
      BlockDescriptor search_key(d_ptr, device);
      BusyBlocks::iterator block_itr = live_blocks.find(search_key);
      if (block_itr != live_blocks.end()) {
        // Remove from live blocks
        search_key = *block_itr;
        live_blocks.erase(block_itr);
        cached_bytes[device].live -= search_key.bytes;
        cached_bytes[device].liveRequested -= search_key.bytesRequested;  // CMS

        // Keep the returned allocation if bin is valid and we won't exceed the max cached threshold
        if ((search_key.bin != INVALID_BIN) && (cached_bytes[device].free + search_key.bytes <= max_cached_bytes)) {
          // Insert returned allocation into free blocks
          recached = true;
          cached_blocks.insert(search_key);
          cached_bytes[device].free += search_key.bytes;

          if (debug)
            // CMS: improved debug message
            // CMS: use raw printf
            printf(
                "\tDevice %d returned %lld bytes at %p from associated stream %lld, event %lld.\n\t\t %lld available "
                "blocks cached (%lld bytes), %lld live blocks outstanding. (%lld bytes)\n",
                device,
                (long long)search_key.bytes,
                d_ptr,
                (long long)search_key.associated_stream,
                (long long)search_key.ready_event,
                (long long)cached_blocks.size(),
                (long long)cached_bytes[device].free,
                (long long)live_blocks.size(),
                (long long)cached_bytes[device].live);
        }
      }

      // First set to specified device (entrypoint may not be set)
      if (device != entrypoint_device) {
        // CMS: throw exception on error
        cudaCheck(error = cudaGetDevice(&entrypoint_device));
        cudaCheck(error = cudaSetDevice(device));
      }

      if (recached) {
        // Insert the ready event in the associated stream (must have current device set properly)
        // CMS: throw exception on error
        cudaCheck(error = cudaEventRecord(search_key.ready_event, search_key.associated_stream));
      }

      // Unlock
      mutex_locker.unlock();

      if (!recached) {
        // Free the allocation from the runtime and cleanup the event.
        // CMS: throw exception on error
        cudaCheck(error = cudaFree(d_ptr));
        cudaCheck(error = cudaEventDestroy(search_key.ready_event));

        if (debug)
          // CMS: improved debug message
          printf(
              "\tDevice %d freed %lld bytes at %p from associated stream %lld, event %lld.\n\t\t  %lld available "
              "blocks cached (%lld bytes), %lld live blocks (%lld bytes) outstanding.\n",
              device,
              (long long)search_key.bytes,
              d_ptr,
              (long long)search_key.associated_stream,
              (long long)search_key.ready_event,
              (long long)cached_blocks.size(),
              (long long)cached_bytes[device].free,
              (long long)live_blocks.size(),
              (long long)cached_bytes[device].live);
      }

      // Reset device
      if ((entrypoint_device != INVALID_DEVICE_ORDINAL) && (entrypoint_device != device)) {
        // CMS: throw exception on error
        cudaCheck(error = cudaSetDevice(entrypoint_device));
      }

      return error;
    }

    cudaError_t DeviceFree(void *d_ptr) { return DeviceFree(INVALID_DEVICE_ORDINAL, d_ptr); }

    cudaError_t FreeAllCached() {
      cudaError_t error = cudaSuccess;
      int entrypoint_device = INVALID_DEVICE_ORDINAL;
      int current_device = INVALID_DEVICE_ORDINAL;
      // CMS: use RAII instead of (un)locking explicitly
      std::unique_lock<std::mutex> mutex_locker(mutex);

      while (!cached_blocks.empty()) {
        // Get first block
        CachedBlocks::iterator begin = cached_blocks.begin();

        // Get entry-point device ordinal if necessary
        if (entrypoint_device == INVALID_DEVICE_ORDINAL) {
          // CMS: silently ignore errors and pass them to the caller
          if ((error = cudaGetDevice(&entrypoint_device)))
            break;
        }

        // Set current device ordinal if necessary
        if (begin->device != current_device) {
          // CMS: silently ignore errors and pass them to the caller
          if ((error = cudaSetDevice(begin->device)))
            break;
          current_device = begin->device;
        }

        // Free device memory
        // CMS: silently ignore errors and pass them to the caller
        if ((error = cudaFree(begin->d_ptr)))
          break;
        if ((error = cudaEventDestroy(begin->ready_event)))
          break;

        // Reduce balance and erase entry
        cached_bytes[current_device].free -= begin->bytes;

        if (debug)
          printf(
              "\tDevice %d freed %lld bytes.\n\t\t  %lld available blocks cached (%lld bytes), %lld live blocks (%lld "
              "bytes) outstanding.\n",
              current_device,
              (long long)begin->bytes,
              (long long)cached_blocks.size(),
              (long long)cached_bytes[current_device].free,
              (long long)live_blocks.size(),
              (long long)cached_bytes[current_device].live);

        cached_blocks.erase(begin);
      }

      mutex_locker.unlock();

      // Attempt to revert back to entry-point device if necessary
      if (entrypoint_device != INVALID_DEVICE_ORDINAL) {
        // CMS: throw exception on error
        cudaCheck(error = cudaSetDevice(entrypoint_device));
      }

      return error;
    }

    // CMS: give access to cache allocation status
    GpuCachedBytes CacheStatus() const {
      std::unique_lock mutex_locker(mutex);
      return cached_bytes;
    }

    // CMS: make the destructor not virtual
    ~CachingDeviceAllocator() {
      if (!skip_cleanup)
        FreeAllCached();
    }
  };
  // end group UtilMgmt

}  // namespace notcub

#endif