gpuError_t MallocAsync(void **ptr, size_t size, gpuStream_t stream) {

LockGuardPtr<std::mutex> lock(mtx_);

if (UNLIKELY(NeedRecord() && cur_size_.load() + size > limit_size_)) {

return gpuErrorOutOfMemory;

// return cudaErrorMemoryAllocation directly here.

return gpuErrorOutOfMemory;

}

}

/**

}

void FreeAsync(void *ptr, size_t size, gpuStream_t stream) {

// Purposefully allow cudaErrorCudartUnloading, because

// that is returned if you ever call cudaFree after the

// driver has already shutdown. This happens only if the

"testing, should not use for release."));

return nullptr;

#endif

}

bool GetMemInfo(size_t *avail,

const int dev_id_;

const uint64_t limit_size_;

std::atomic<uint64_t> cur_size_{0};

cudaMemPool_t memPool_;

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

static std::once_flag once_flag_;

std::set<void *> gpu_ptrs; // just for testing

}; // NOLINT

std::once_flag RecordedGpuMallocHelper::once_flag_;

std::once_flag RecordedGpuMallocHelper::set_cudamempoolattr_once_flag_;

gpuError_t RecordedGpuMalloc(void **ptr,

size_t size,