|
FireSync API
|
Represents a Cuda managed memory allocator object.
For devices that can support concurrent access to managed memory (both device and host can safely access the memory simultaneously), this allocator will allocate memory with global attachment. There are no special requirements when using this kind of managed memory.
For devices that cannot support concurrent access to managed memory (device and host cannot safely access the memory simultaneously), this allocator will allocate memory with host attachment. In this state, the memory can be accessed safely from any host (CPU) threads but cannot be accessed by a Cuda device. In order to allow a Cuda device to access this memory, the memory must first be attached to a stream. Once attached, this memory can only be accessed by host threads or by the device while executing operations on the specified stream. In this state, it is the caller's responsibility to ensure that the memory is not accessed simultaneously by the host and the device. When device operations with this memory have been completed, the memory can be detached from the stream, once again enabling unrestricted access from host threads.
Allocator traits can be accessed (via kAlloc_Traits) to determine whether concurrent managed access is supported. The presence of the kALLOC_TRAIT_SERIAL trait indicates that concurrent managed access is not supported. Alternatively, the kCudaDevice_Properties method can be used to determine this information directly from device properties. However, if both cases must be supported, then the simplest course of action is to assume that concurrent managed access is not available, and to perform the necessary attach/detach operations to ensure safe access. For systems with concurrent managed access, these operations will have no effect.
Public Member Functions | |
| kAlloc | kCudaManagedAlloc_Instance () |
| Gets the singleton Cuda managed memory allocator instance. More... | |
