4.12. Unified Addressing

This section describes the unified addressing functions of the low-level CUDA driver application programming interface.

Overview

CUDA devices can share a unified address space with the host. For these devices there is no distinction between a device pointer and a host pointer -- the same pointer value may be used to access memory from the host program and from a kernel running on the device (with exceptions enumerated below).

Supported Platforms

Whether or not a device supports unified addressing may be queried by calling cuDeviceGetAttribute() with the device attribute CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING.

Unified addressing is automatically enabled in 64-bit processes

Looking Up Information from Pointer Values

It is possible to look up information about the memory which backs a pointer value. For instance, one may want to know if a pointer points to host or device memory. As another example, in the case of device memory, one may want to know on which CUDA device the memory resides. These properties may be queried using the function cuPointerGetAttribute()

Since pointers are unique, it is not necessary to specify information about the pointers specified to the various copy functions in the CUDA API. The function cuMemcpy() may be used to perform a copy between two pointers, ignoring whether they point to host or device memory (making cuMemcpyHtoD(), cuMemcpyDtoD(), and cuMemcpyDtoH() unnecessary for devices supporting unified addressing). For multidimensional copies, the memory type CU_MEMORYTYPE_UNIFIED may be used to specify that the CUDA driver should infer the location of the pointer from its value.

Automatic Mapping of Host Allocated Host Memory

All host memory allocated in all contexts using cuMemAllocHost() and cuMemHostAlloc() is always directly accessible from all contexts on all devices that support unified addressing. This is the case regardless of whether or not the flags CU_MEMHOSTALLOC_PORTABLE and CU_MEMHOSTALLOC_DEVICEMAP are specified.

The pointer value through which allocated host memory may be accessed in kernels on all devices that support unified addressing is the same as the pointer value through which that memory is accessed on the host, so it is not necessary to call cuMemHostGetDevicePointer() to get the device pointer for these allocations.

Note that this is not the case for memory allocated using the flag CU_MEMHOSTALLOC_WRITECOMBINED, as discussed below.

Automatic Registration of Peer Memory

Upon enabling direct access from a context that supports unified addressing to another peer context that supports unified addressing using cuCtxEnablePeerAccess() all memory allocated in the peer context using cuMemAlloc() and cuMemAllocPitch() will immediately be accessible by the current context. The device pointer value through which any peer memory may be accessed in the current context is the same pointer value through which that memory may be accessed in the peer context.

Exceptions, Disjoint Addressing

Not all memory may be accessed on devices through the same pointer value through which they are accessed on the host. These exceptions are host memory registered using cuMemHostRegister() and host memory allocated using the flag CU_MEMHOSTALLOC_WRITECOMBINED. For these exceptions, there exists a distinct host and device address for the memory. The device address is guaranteed to not overlap any valid host pointer range and is guaranteed to have the same value across all contexts that support unified addressing.

This device address may be queried using cuMemHostGetDevicePointer() when a context using unified addressing is current. Either the host or the unified device pointer value may be used to refer to this memory through cuMemcpy() and similar functions using the CU_MEMORYTYPE_UNIFIED memory type.

Functions

CUresult cuMemAdvise ( CUdeviceptr devPtr, size_t count, CUmem_advise advice, CUdevice device )
Advise about the usage of a given memory range.
CUresult cuMemPrefetchAsync ( CUdeviceptr devPtr, size_t count, CUdevice dstDevice, CUstream hStream )
Prefetches memory to the specified destination device.
CUresult cuMemRangeGetAttribute ( void* data, size_t dataSize, CUmem_range_attribute attribute, CUdeviceptr devPtr, size_t count )
Query an attribute of a given memory range.
CUresult cuMemRangeGetAttributes ( void** data, size_t* dataSizes, CUmem_range_attribute* attributes, size_t numAttributes, CUdeviceptr devPtr, size_t count )
Query attributes of a given memory range.
CUresult cuPointerGetAttribute ( void* data, CUpointer_attribute attribute, CUdeviceptr ptr )
Returns information about a pointer.
CUresult cuPointerGetAttributes ( unsigned int  numAttributes, CUpointer_attribute* attributes, void** data, CUdeviceptr ptr )
Returns information about a pointer.
CUresult cuPointerSetAttribute ( const void* value, CUpointer_attribute attribute, CUdeviceptr ptr )
Set attributes on a previously allocated memory region.

Functions

CUresult cuMemAdvise ( CUdeviceptr devPtr, size_t count, CUmem_advise advice, CUdevice device )
Advise about the usage of a given memory range.
Parameters
devPtr
- Pointer to memory to set the advice for
count
- Size in bytes of the memory range
advice
- Advice to be applied for the specified memory range
device
- Device to apply the advice for
Description

Advise the Unified Memory subsystem about the usage pattern for the memory range starting at devPtr with a size of count bytes. The start address and end address of the memory range will be rounded down and rounded up respectively to be aligned to CPU page size before the advice is applied. The memory range must refer to managed memory allocated via cuMemAllocManaged or declared via __managed__ variables.

The advice parameter can take the following values:

  • CU_MEM_ADVISE_SET_READ_MOSTLY: This implies that the data is mostly going to be read from and only occasionally written to. Any read accesses from any processor to this region will create a read-only copy of at least the accessed pages in that processor's memory. Additionally, if cuMemPrefetchAsync is called on this region, it will create a read-only copy of the data on the destination processor. If any processor writes to this region, all copies of the corresponding page will be invalidated except for the one where the write occurred. The device argument is ignored for this advice. Note that for a page to be read-duplicated, the accessing processor must either be the CPU or a GPU that has a non-zero value for the device attribute CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS. Also, if a context is created on a device that does not have the device attribute CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS set, then read-duplication will not occur until all such contexts are destroyed.

  • CU_MEM_ADVISE_UNSET_READ_MOSTLY: Undoes the effect of CU_MEM_ADVISE_SET_READ_MOSTLY and also prevents the Unified Memory driver from attempting heuristic read-duplication on the memory range. Any read-duplicated copies of the data will be collapsed into a single copy. The location for the collapsed copy will be the preferred location if the page has a preferred location and one of the read-duplicated copies was resident at that location. Otherwise, the location chosen is arbitrary.

  • CU_MEM_ADVISE_SET_PREFERRED_LOCATION: This advice sets the preferred location for the data to be the memory belonging to device. Passing in CU_DEVICE_CPU for device sets the preferred location as host memory. If device is a GPU, then it must have a non-zero value for the device attribute CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS. Setting the preferred location does not cause data to migrate to that location immediately. Instead, it guides the migration policy when a fault occurs on that memory region. If the data is already in its preferred location and the faulting processor can establish a mapping without requiring the data to be migrated, then data migration will be avoided. On the other hand, if the data is not in its preferred location or if a direct mapping cannot be established, then it will be migrated to the processor accessing it. It is important to note that setting the preferred location does not prevent data prefetching done using cuMemPrefetchAsync. Having a preferred location can override the page thrash detection and resolution logic in the Unified Memory driver. Normally, if a page is detected to be constantly thrashing between for example host and device memory, the page may eventually be pinned to host memory by the Unified Memory driver. But if the preferred location is set as device memory, then the page will continue to thrash indefinitely. If CU_MEM_ADVISE_SET_READ_MOSTLY is also set on this memory region or any subset of it, then the policies associated with that advice will override the policies of this advice.

  • CU_MEM_ADVISE_UNSET_PREFERRED_LOCATION: Undoes the effect of CU_MEM_ADVISE_SET_PREFERRED_LOCATION and changes the preferred location to none.

  • CU_MEM_ADVISE_SET_ACCESSED_BY: This advice implies that the data will be accessed by device. Passing in CU_DEVICE_CPU for device will set the advice for the CPU. If device is a GPU, then the device attribute CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS must be non-zero. This advice does not cause data migration and has no impact on the location of the data per se. Instead, it causes the data to always be mapped in the specified processor's page tables, as long as the location of the data permits a mapping to be established. If the data gets migrated for any reason, the mappings are updated accordingly. This advice is recommended in scenarios where data locality is not important, but avoiding faults is. Consider for example a system containing multiple GPUs with peer-to-peer access enabled, where the data located on one GPU is occasionally accessed by peer GPUs. In such scenarios, migrating data over to the other GPUs is not as important because the accesses are infrequent and the overhead of migration may be too high. But preventing faults can still help improve performance, and so having a mapping set up in advance is useful. Note that on CPU access of this data, the data may be migrated to host memory because the CPU typically cannot access device memory directly. Any GPU that had the CU_MEM_ADVISE_SET_ACCESSED_BY flag set for this data will now have its mapping updated to point to the page in host memory. If CU_MEM_ADVISE_SET_READ_MOSTLY is also set on this memory region or any subset of it, then the policies associated with that advice will override the policies of this advice. Additionally, if the preferred location of this memory region or any subset of it is also device, then the policies associated with CU_MEM_ADVISE_SET_PREFERRED_LOCATION will override the policies of this advice.

  • CU_MEM_ADVISE_UNSET_ACCESSED_BY: Undoes the effect of CU_MEM_ADVISE_SET_ACCESSED_BY. Any mappings to the data from device may be removed at any time causing accesses to result in non-fatal page faults.

Note:
  • Note that this function may also return error codes from previous, asynchronous launches.

  • This function exhibits asynchronous behavior for most use cases.

  • This function uses standard default stream semantics.

See also:

cuMemcpy, cuMemcpyPeer, cuMemcpyAsync, cuMemcpy3DPeerAsync, cuMemPrefetchAsync, cudaMemAdvise

CUresult cuMemPrefetchAsync ( CUdeviceptr devPtr, size_t count, CUdevice dstDevice, CUstream hStream )
Prefetches memory to the specified destination device.
Parameters
devPtr
- Pointer to be prefetched
count
- Size in bytes
dstDevice
- Destination device to prefetch to
hStream
- Stream to enqueue prefetch operation
Description

Prefetches memory to the specified destination device. devPtr is the base device pointer of the memory to be prefetched and dstDevice is the destination device. count specifies the number of bytes to copy. hStream is the stream in which the operation is enqueued. The memory range must refer to managed memory allocated via cuMemAllocManaged or declared via __managed__ variables.

Passing in CU_DEVICE_CPU for dstDevice will prefetch the data to host memory. If dstDevice is a GPU, then the device attribute CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS must be non-zero. Additionally, hStream must be associated with a device that has a non-zero value for the device attribute CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS.

The start address and end address of the memory range will be rounded down and rounded up respectively to be aligned to CPU page size before the prefetch operation is enqueued in the stream.

If no physical memory has been allocated for this region, then this memory region will be populated and mapped on the destination device. If there's insufficient memory to prefetch the desired region, the Unified Memory driver may evict pages from other cuMemAllocManaged allocations to host memory in order to make room. Device memory allocated using cuMemAlloc or cuArrayCreate will not be evicted.

By default, any mappings to the previous location of the migrated pages are removed and mappings for the new location are only setup on dstDevice. The exact behavior however also depends on the settings applied to this memory range via cuMemAdvise as described below:

If CU_MEM_ADVISE_SET_READ_MOSTLY was set on any subset of this memory range, then that subset will create a read-only copy of the pages on dstDevice.

If CU_MEM_ADVISE_SET_PREFERRED_LOCATION was called on any subset of this memory range, then the pages will be migrated to dstDevice even if dstDevice is not the preferred location of any pages in the memory range.

If CU_MEM_ADVISE_SET_ACCESSED_BY was called on any subset of this memory range, then mappings to those pages from all the appropriate processors are updated to refer to the new location if establishing such a mapping is possible. Otherwise, those mappings are cleared.

Note that this API is not required for functionality and only serves to improve performance by allowing the application to migrate data to a suitable location before it is accessed. Memory accesses to this range are always coherent and are allowed even when the data is actively being migrated.

Note that this function is asynchronous with respect to the host and all work on other devices.

Note:
  • Note that this function may also return error codes from previous, asynchronous launches.

  • This function exhibits asynchronous behavior for most use cases.

  • This function uses standard default stream semantics.

See also:

cuMemcpy, cuMemcpyPeer, cuMemcpyAsync, cuMemcpy3DPeerAsync, cuMemAdvise, cudaMemPrefetchAsync

CUresult cuMemRangeGetAttribute ( void* data, size_t dataSize, CUmem_range_attribute attribute, CUdeviceptr devPtr, size_t count )
Query an attribute of a given memory range.
Parameters
data
- A pointers to a memory location where the result of each attribute query will be written to.
dataSize
- Array containing the size of data
attribute
- The attribute to query
devPtr
- Start of the range to query
count
- Size of the range to query
Description

Query an attribute about the memory range starting at devPtr with a size of count bytes. The memory range must refer to managed memory allocated via cuMemAllocManaged or declared via __managed__ variables.

The attribute parameter can take the following values:

  • CU_MEM_RANGE_ATTRIBUTE_READ_MOSTLY: If this attribute is specified, data will be interpreted as a 32-bit integer, and dataSize must be 4. The result returned will be 1 if all pages in the given memory range have read-duplication enabled, or 0 otherwise.

  • CU_MEM_RANGE_ATTRIBUTE_PREFERRED_LOCATION: If this attribute is specified, data will be interpreted as a 32-bit integer, and dataSize must be 4. The result returned will be a GPU device id if all pages in the memory range have that GPU as their preferred location, or it will be CU_DEVICE_CPU if all pages in the memory range have the CPU as their preferred location, or it will be CU_DEVICE_INVALID if either all the pages don't have the same preferred location or some of the pages don't have a preferred location at all. Note that the actual location of the pages in the memory range at the time of the query may be different from the preferred location.

  • CU_MEM_RANGE_ATTRIBUTE_ACCESSED_BY: If this attribute is specified, data will be interpreted as an array of 32-bit integers, and dataSize must be a non-zero multiple of 4. The result returned will be a list of device ids that had CU_MEM_ADVISE_SET_ACCESSED_BY set for that entire memory range. If any device does not have that advice set for the entire memory range, that device will not be included. If data is larger than the number of devices that have that advice set for that memory range, CU_DEVICE_INVALID will be returned in all the extra space provided. For ex., if dataSize is 12 (i.e. data has 3 elements) and only device 0 has the advice set, then the result returned will be { 0, CU_DEVICE_INVALID, CU_DEVICE_INVALID }. If data is smaller than the number of devices that have that advice set, then only as many devices will be returned as can fit in the array. There is no guarantee on which specific devices will be returned, however.

  • CU_MEM_RANGE_ATTRIBUTE_LAST_PREFETCH_LOCATION: If this attribute is specified, data will be interpreted as a 32-bit integer, and dataSize must be 4. The result returned will be the last location to which all pages in the memory range were prefetched explicitly via cuMemPrefetchAsync. This will either be a GPU id or CU_DEVICE_CPU depending on whether the last location for prefetch was a GPU or the CPU respectively. If any page in the memory range was never explicitly prefetched or if all pages were not prefetched to the same location, CU_DEVICE_INVALID will be returned. Note that this simply returns the last location that the applicaton requested to prefetch the memory range to. It gives no indication as to whether the prefetch operation to that location has completed or even begun.

Note:
  • Note that this function may also return error codes from previous, asynchronous launches.

  • This function exhibits asynchronous behavior for most use cases.

  • This function uses standard default stream semantics.

See also:

cuMemRangeGetAttributes, cuMemPrefetchAsync, cuMemAdvise, cudaMemRangeGetAttribute

CUresult cuMemRangeGetAttributes ( void** data, size_t* dataSizes, CUmem_range_attribute* attributes, size_t numAttributes, CUdeviceptr devPtr, size_t count )
Query attributes of a given memory range.
Parameters
data
- A two-dimensional array containing pointers to memory locations where the result of each attribute query will be written to.
dataSizes
- Array containing the sizes of each result
attributes
- An array of attributes to query (numAttributes and the number of attributes in this array should match)
numAttributes
- Number of attributes to query
devPtr
- Start of the range to query
count
- Size of the range to query
Description

Query attributes of the memory range starting at devPtr with a size of count bytes. The memory range must refer to managed memory allocated via cuMemAllocManaged or declared via __managed__ variables. The attributes array will be interpreted to have numAttributes entries. The dataSizes array will also be interpreted to have numAttributes entries. The results of the query will be stored in data.

The list of supported attributes are given below. Please refer to cuMemRangeGetAttribute for attribute descriptions and restrictions.

Note:

Note that this function may also return error codes from previous, asynchronous launches.

See also:

cuMemRangeGetAttribute, cuMemAdvisecuMemPrefetchAsync, cudaMemRangeGetAttributes

CUresult cuPointerGetAttribute ( void* data, CUpointer_attribute attribute, CUdeviceptr ptr )
Returns information about a pointer.
Parameters
data
- Returned pointer attribute value
attribute
- Pointer attribute to query
ptr
- Pointer
Description

The supported attributes are:

Returns in *data the CUcontext in which ptr was allocated or registered. The type of data must be CUcontext *.

If ptr was not allocated by, mapped by, or registered with a CUcontext which uses unified virtual addressing then CUDA_ERROR_INVALID_VALUE is returned.

Returns in *data the physical memory type of the memory that ptr addresses as a CUmemorytype enumerated value. The type of data must be unsigned int.

If ptr addresses device memory then *data is set to CU_MEMORYTYPE_DEVICE. The particular CUdevice on which the memory resides is the CUdevice of the CUcontext returned by the CU_POINTER_ATTRIBUTE_CONTEXT attribute of ptr.

If ptr addresses host memory then *data is set to CU_MEMORYTYPE_HOST.

If ptr was not allocated by, mapped by, or registered with a CUcontext which uses unified virtual addressing then CUDA_ERROR_INVALID_VALUE is returned.

If the current CUcontext does not support unified virtual addressing then CUDA_ERROR_INVALID_CONTEXT is returned.

Returns in *data the device pointer value through which ptr may be accessed by kernels running in the current CUcontext. The type of data must be CUdeviceptr *.

If there exists no device pointer value through which kernels running in the current CUcontext may access ptr then CUDA_ERROR_INVALID_VALUE is returned.

If there is no current CUcontext then CUDA_ERROR_INVALID_CONTEXT is returned.

Except in the exceptional disjoint addressing cases discussed below, the value returned in *data will equal the input value ptr.

Returns in *data the host pointer value through which ptr may be accessed by by the host program. The type of data must be void **. If there exists no host pointer value through which the host program may directly access ptr then CUDA_ERROR_INVALID_VALUE is returned.

Except in the exceptional disjoint addressing cases discussed below, the value returned in *data will equal the input value ptr.

Returns in *data two tokens for use with the nv-p2p.h Linux kernel interface. data must be a struct of type CUDA_POINTER_ATTRIBUTE_P2P_TOKENS.

ptr must be a pointer to memory obtained from :cuMemAlloc(). Note that p2pToken and vaSpaceToken are only valid for the lifetime of the source allocation. A subsequent allocation at the same address may return completely different tokens. Querying this attribute has a side effect of setting the attribute CU_POINTER_ATTRIBUTE_SYNC_MEMOPS for the region of memory that ptr points to.

A boolean attribute which when set, ensures that synchronous memory operations initiated on the region of memory that ptr points to will always synchronize. See further documentation in the section titled "API synchronization behavior" to learn more about cases when synchronous memory operations can exhibit asynchronous behavior.

Returns in *data a buffer ID which is guaranteed to be unique within the process. data must point to an unsigned long long.

ptr must be a pointer to memory obtained from a CUDA memory allocation API. Every memory allocation from any of the CUDA memory allocation APIs will have a unique ID over a process lifetime. Subsequent allocations do not reuse IDs from previous freed allocations. IDs are only unique within a single process.

Returns in *data a boolean that indicates whether the pointer points to managed memory or not.

Note that for most allocations in the unified virtual address space the host and device pointer for accessing the allocation will be the same. The exceptions to this are

  • user memory registered using cuMemHostRegister

  • host memory allocated using cuMemHostAlloc with the CU_MEMHOSTALLOC_WRITECOMBINED flag For these types of allocation there will exist separate, disjoint host and device addresses for accessing the allocation. In particular

  • The host address will correspond to an invalid unmapped device address (which will result in an exception if accessed from the device)

  • The device address will correspond to an invalid unmapped host address (which will result in an exception if accessed from the host). For these types of allocations, querying CU_POINTER_ATTRIBUTE_HOST_POINTER and CU_POINTER_ATTRIBUTE_DEVICE_POINTER may be used to retrieve the host and device addresses from either address.

Note:

Note that this function may also return error codes from previous, asynchronous launches.

See also:

cuPointerSetAttribute, cuMemAlloc, cuMemFree, cuMemAllocHost, cuMemFreeHost, cuMemHostAlloc, cuMemHostRegister, cuMemHostUnregister, cudaPointerGetAttributes

CUresult cuPointerGetAttributes ( unsigned int  numAttributes, CUpointer_attribute* attributes, void** data, CUdeviceptr ptr )
Returns information about a pointer.
Parameters
numAttributes
- Number of attributes to query
attributes
- An array of attributes to query (numAttributes and the number of attributes in this array should match)
data
- A two-dimensional array containing pointers to memory locations where the result of each attribute query will be written to.
ptr
- Pointer to query
Description

The supported attributes are (refer to cuPointerGetAttribute for attribute descriptions and restrictions):

Unlike cuPointerGetAttribute, this function will not return an error when the ptr encountered is not a valid CUDA pointer. Instead, the attributes are assigned default NULL values and CUDA_SUCCESS is returned.

If ptr was not allocated by, mapped by, or registered with a CUcontext which uses UVA (Unified Virtual Addressing), CUDA_ERROR_INVALID_CONTEXT is returned.

Note:

Note that this function may also return error codes from previous, asynchronous launches.

See also:

cuPointerGetAttribute, cuPointerSetAttribute, cudaPointerGetAttributes

CUresult cuPointerSetAttribute ( const void* value, CUpointer_attribute attribute, CUdeviceptr ptr )
Set attributes on a previously allocated memory region.
Parameters
value
- Pointer to memory containing the value to be set
attribute
- Pointer attribute to set
ptr
- Pointer to a memory region allocated using CUDA memory allocation APIs
Description

The supported attributes are:

A boolean attribute that can either be set (1) or unset (0). When set, the region of memory that ptr points to is guaranteed to always synchronize memory operations that are synchronous. If there are some previously initiated synchronous memory operations that are pending when this attribute is set, the function does not return until those memory operations are complete. See further documentation in the section titled "API synchronization behavior" to learn more about cases when synchronous memory operations can exhibit asynchronous behavior. value will be considered as a pointer to an unsigned integer to which this attribute is to be set.

Note:

Note that this function may also return error codes from previous, asynchronous launches.

See also:

cuPointerGetAttribute, cuPointerGetAttributes, cuMemAlloc, cuMemFree, cuMemAllocHost, cuMemFreeHost, cuMemHostAlloc, cuMemHostRegister, cuMemHostUnregister