pyarrow.cuda.Context#
- class pyarrow.cuda.Context(*args, **kwargs)#
Bases:
_Weakrefable
CUDA driver context.
- __init__()#
Create a CUDA driver context for a particular device.
If a CUDA context handle is passed, it is wrapped, otherwise a default CUDA context for the given device is requested.
Methods
Create a CUDA driver context for a particular device.
buffer_from_data
(self, data, ...)Create device buffer and initialize with data.
buffer_from_object
(self, obj)Create device buffer view of arbitrary object that references device accessible memory.
foreign_buffer
(self, address, size[, base])Create device buffer from address and size as a view.
from_numba
([context])Create a Context instance from a Numba CUDA context.
get_device_address
(self, uintptr_t address)Return the device address that is reachable from kernels running in the context
Return the number of GPU devices.
new_buffer
(self, int64_t nbytes)Return new device buffer.
open_ipc_buffer
(self, ipc_handle)Open existing CUDA IPC memory handle
synchronize
(self)Blocks until the device has completed all preceding requested tasks.
to_numba
(self)Convert Context to a Numba CUDA context.
Attributes
Return the number of allocated bytes.
Return context device number.
Return pointer to context handle.
- buffer_from_data(self, data, int64_t offset=0, int64_t size=-1)#
Create device buffer and initialize with data.
- Parameters:
- data{
CudaBuffer
,HostBuffer
,Buffer
,array-like
} Specify data to be copied to device buffer.
- offset
int
Specify the offset of input buffer for device data buffering. Default: 0.
- size
int
Specify the size of device buffer in bytes. Default: all (starting from input offset)
- data{
- Returns:
- cbuf
CudaBuffer
Device buffer with copied data.
- cbuf
- buffer_from_object(self, obj)#
Create device buffer view of arbitrary object that references device accessible memory.
When the object contains a non-contiguous view of device accessible memory then the returned device buffer will contain contiguous view of the memory, that is, including the intermediate data that is otherwise invisible to the input object.
- Parameters:
- obj{object,
Buffer
,HostBuffer
,CudaBuffer
, …} Specify an object that holds (device or host) address that can be accessed from device. This includes objects with types defined in pyarrow.cuda as well as arbitrary objects that implement the CUDA array interface as defined by numba.
- obj{object,
- Returns:
- cbuf
CudaBuffer
Device buffer as a view of device accessible memory.
- cbuf
- bytes_allocated#
Return the number of allocated bytes.
- device_number#
Return context device number.
- foreign_buffer(self, address, size, base=None)#
Create device buffer from address and size as a view.
The caller is responsible for allocating and freeing the memory. When address==size==0 then a new zero-sized buffer is returned.
- Parameters:
- address
int
Specify the starting address of the buffer. The address can refer to both device or host memory but it must be accessible from device after mapping it with get_device_address method.
- size
int
Specify the size of device buffer in bytes.
- base{
None
, object} Specify object that owns the referenced memory.
- address
- Returns:
- cbuf
CudaBuffer
Device buffer as a view of device reachable memory.
- cbuf
- static from_numba(context=None)#
Create a Context instance from a Numba CUDA context.
- Parameters:
- context{
numba.cuda.cudadrv.driver.Context
,None
} A Numba CUDA context instance. If None, the current Numba context is used.
- context{
- Returns:
- shared_context
pyarrow.cuda.Context
Context instance.
- shared_context
- get_device_address(self, uintptr_t address)#
Return the device address that is reachable from kernels running in the context
- Parameters:
- address
int
Specify memory address value
- address
- Returns:
- device_address
int
Device address accessible from device context
- device_address
Notes
The device address is defined as a memory address accessible by device. While it is often a device memory address but it can be also a host memory address, for instance, when the memory is allocated as host memory (using cudaMallocHost or cudaHostAlloc) or as managed memory (using cudaMallocManaged) or the host memory is page-locked (using cudaHostRegister).
- static get_num_devices()#
Return the number of GPU devices.
- handle#
Return pointer to context handle.
- new_buffer(self, int64_t nbytes)#
Return new device buffer.
- Parameters:
- nbytes
int
Specify the number of bytes to be allocated.
- nbytes
- Returns:
- buf
CudaBuffer
Allocated buffer.
- buf
- open_ipc_buffer(self, ipc_handle)#
Open existing CUDA IPC memory handle
- Parameters:
- ipc_handle
IpcMemHandle
Specify opaque pointer to CUipcMemHandle (driver API).
- ipc_handle
- Returns:
- buf
CudaBuffer
referencing device buffer
- buf
- synchronize(self)#
Blocks until the device has completed all preceding requested tasks.
- to_numba(self)#
Convert Context to a Numba CUDA context.
- Returns:
- context
numba.cuda.cudadrv.driver.Context
Numba CUDA context instance.
- context