ai.rapids.cudf

Class CudaMemoryBuffer

java.lang.Object

ai.rapids.cudf.MemoryBuffer

ai.rapids.cudf.BaseDeviceMemoryBuffer

ai.rapids.cudf.CudaMemoryBuffer

All Implemented Interfaces:

AutoCloseable

public class CudaMemoryBuffer
extends BaseDeviceMemoryBuffer

This class represents data allocated using `cudaMalloc` directly instead of the default RMM memory resource. Closing this object will effectively release the memory held by the buffer. Note that because of reference counting if a buffer is sliced it may not actually result in the memory being released.

Nested Class Summary

Nested classes/interfaces inherited from class ai.rapids.cudf.MemoryBuffer

MemoryBuffer.EventHandler, MemoryBuffer.MemoryBufferCleaner

Field Summary

Fields inherited from class ai.rapids.cudf.MemoryBuffer

address, cleaner, closed, id, length, refCount

Constructor Summary

Constructors

Constructor and Description
CudaMemoryBuffer(long address, long lengthInBytes, Cuda.Stream stream) Wrap an existing CUDA allocation in a device memory buffer.

Method Summary

Modifier and Type	Method and Description
static CudaMemoryBuffer	allocate(long bytes) Allocate memory for use on the GPU.
static CudaMemoryBuffer	allocate(long bytes, Cuda.Stream stream) Allocate memory for use on the GPU.
CudaMemoryBuffer	slice(long offset, long len) Slice off a part of the device buffer.

Methods inherited from class ai.rapids.cudf.BaseDeviceMemoryBuffer

copyFromDeviceBufferAsync, copyFromHostBuffer, copyFromHostBuffer, copyFromHostBuffer, copyFromHostBuffer, copyFromHostBuffer, copyFromHostBufferAsync, copyFromHostBufferAsync, sliceWithCopy

Methods inherited from class ai.rapids.cudf.MemoryBuffer

addressOutOfBoundsCheck, close, copyFromMemoryBuffer, copyFromMemoryBufferAsync, getAddress, getEventHandler, getLength, getRefCount, incRefCount, noWarnLeakExpected, setEventHandler, toString

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

Constructor Detail

CudaMemoryBuffer

public CudaMemoryBuffer(long address,
                        long lengthInBytes,
                        Cuda.Stream stream)

Wrap an existing CUDA allocation in a device memory buffer. The CUDA allocation will be freed when the resulting device memory buffer instance frees its memory resource (i.e.: when its reference count goes to zero).

Parameters:

address - device address of the CUDA memory allocation

lengthInBytes - length of the CUDA allocation in bytes

stream - CUDA stream to use for synchronization when freeing the allocation

Method Detail

allocate

public static CudaMemoryBuffer allocate(long bytes)

Allocate memory for use on the GPU. You must close it when done.

Parameters:

bytes - size in bytes to allocate

Returns:

the buffer

allocate

public static CudaMemoryBuffer allocate(long bytes,
                                        Cuda.Stream stream)

Allocate memory for use on the GPU. You must close it when done.

Parameters:

bytes - size in bytes to allocate

stream - The stream in which to synchronize this command

Returns:

the buffer

slice

public final CudaMemoryBuffer slice(long offset,
                                    long len)

Slice off a part of the device buffer. Note that this is a zero copy operation and all slices must be closed along with the original buffer before the memory is released to RMM. So use this with some caution.

Specified by:

slice in class MemoryBuffer

Parameters:

offset - where to start the slice at.

len - how many bytes to slice

Returns:

a device buffer that will need to be closed independently from this buffer.