wehub-resource-sync
120 lines
4.6 KiB
C++
120 lines
4.6 KiB
C++
/* ******************************************************************************
|
|
*
|
|
*
|
|
* This program and the accompanying materials are made available under the
|
|
* terms of the Apache License, Version 2.0 which is available at
|
|
* https://www.apache.org/licenses/LICENSE-2.0.
|
|
*
|
|
* See the NOTICE file distributed with this work for additional
|
|
* information regarding copyright ownership.
|
|
* 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.
|
|
*
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
******************************************************************************/
|
|
|
|
//
|
|
// Created by raver119 on 30.11.17.
|
|
//
|
|
#include <exceptions/cuda_exception.h>
|
|
#include <execution/AffinityManager.h>
|
|
#include <execution/LaunchContext.h>
|
|
#include <helpers/logger.h>
|
|
#include <thread>
|
|
|
|
// NOTE: Removed thread_local to fix "cannot allocate memory in static TLS block" error
|
|
// when JavaCPP loads library via dlopen(). This error occurs because:
|
|
// 1. dlopen() loads the library at runtime (not at program startup)
|
|
// 2. thread_local variables require space in the static TLS block
|
|
// 3. The static TLS block has limited size and cannot be extended after program start
|
|
// 4. When loaded via dlopen(), the library's TLS requirements must fit in remaining space
|
|
// 5. If sanitizers or lifecycle tracking are enabled, TLS space may already be exhausted
|
|
//
|
|
// iOS/Apple/Android builds already avoid thread_local for similar reasons.
|
|
// For CPU builds, making this a non-thread-local global is acceptable because:
|
|
// - Each LaunchContext instance maintains its own thread-safe state
|
|
// - The global contextBuffers is used as a default/fallback buffer pool
|
|
// - Proper synchronization is handled at the LaunchContext level
|
|
sd::ContextBuffers contextBuffers = sd::ContextBuffers();
|
|
|
|
#if HAVE_ONEDNN
|
|
#include <dnnl.hpp>
|
|
#endif
|
|
|
|
namespace sd {
|
|
|
|
LaunchContext::~LaunchContext() {
|
|
#if HAVE_ONEDNN
|
|
// Intentionally NOT deleting _engine to avoid use-after-free during shutdown
|
|
// The LaunchContext objects are kept alive in a function-local static vector (contexts())
|
|
// to survive JVM shutdown. Deleting _engine here causes crashes because
|
|
// OneDNN's static cleanup may have already run, making the engine invalid.
|
|
// This is safe because LaunchContext instances are never destroyed during
|
|
// normal operation - only during process exit when memory cleanup doesn't matter.
|
|
// delete reinterpret_cast<dnnl::engine*>(_engine);
|
|
#endif
|
|
}
|
|
|
|
// Use pointer to prevent static destructor from running.
|
|
// The vector is intentionally leaked to avoid crashes during shutdown when:
|
|
// 1. An exception is thrown that references LaunchContext
|
|
// 2. JVM shuts down and runs static destructors
|
|
// 3. The vector's destructor tries to destroy stored contexts while they're still in use
|
|
// This is safe because LaunchContexts are only created during initialization and
|
|
// process exit cleans up all memory anyway.
|
|
std::vector<LaunchContext*>& LaunchContext::contexts() {
|
|
static std::vector<LaunchContext*>* _contexts = new std::vector<LaunchContext*>();
|
|
return *_contexts;
|
|
}
|
|
|
|
SD_MAP_IMPL<int, std::mutex*> LaunchContext::_deviceMutexes;
|
|
std::mutex LaunchContext::_mutex;
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
LaunchContext::LaunchContext() {
|
|
// default constructor, just to make clang/ranlib happy
|
|
_workspace = nullptr;
|
|
_deviceID = 0;
|
|
#if HAVE_ONEDNN
|
|
_engine = new dnnl::engine(dnnl::engine::kind::cpu, 0);
|
|
#endif
|
|
}
|
|
|
|
LaunchContext::LaunchContext(sd::Pointer cudaStream, sd::Pointer reductionPointer, sd::Pointer scalarPointer,
|
|
sd::Pointer allocationPointer) {}
|
|
|
|
static std::mutex _lock;
|
|
|
|
LaunchContext* LaunchContext::defaultContext() {
|
|
{
|
|
// synchronous block goes here
|
|
std::lock_guard<std::mutex> lock(_lock);
|
|
// TODO: we need it to be device-aware, but only once we add NUMA support for cpu
|
|
if (LaunchContext::contexts().empty())
|
|
LaunchContext::contexts().emplace_back(new LaunchContext());
|
|
}
|
|
|
|
// return context for current device
|
|
return LaunchContext::contexts().at(0);
|
|
}
|
|
|
|
std::mutex* LaunchContext::deviceMutex() { return &_mutex; }
|
|
|
|
void LaunchContext::swapContextBuffers(ContextBuffers& buffers) {
|
|
//
|
|
}
|
|
|
|
bool LaunchContext::isInitialized() { return true; }
|
|
|
|
void LaunchContext::releaseBuffers() {
|
|
//
|
|
}
|
|
|
|
sd::ErrorReference* LaunchContext::errorReference() { return contextBuffers.errorReference(); }
|
|
|
|
void* LaunchContext::engine() { return _engine; }
|
|
} // namespace sd
|