CUDA support#

Contexts#

class CudaDeviceManager#

Public Functions

Result<std::shared_ptr<CudaDevice>> GetDevice(int device_number)#

Get a CudaDevice instance for a particular device.

Parameters:

device_number – [in] the CUDA device number

Result<std::shared_ptr<CudaContext>> GetContext(int device_number)#

Get the CUDA driver context for a particular device.

Parameters:

device_number – [in] the CUDA device number

Returns:

cached context

Result<std::shared_ptr<CudaContext>> GetSharedContext(int device_number, void *handle)#

Get the shared CUDA driver context for a particular device.

Parameters:

  • device_number – [in] the CUDA device number

  • handle – [in] CUDA context handle created by another library

Returns:

shared context

Result<std::shared_ptr<CudaHostBuffer>> AllocateHost(int device_number, int64_t nbytes)#

Allocate host memory with fast access to given GPU device.

Parameters:

  • device_number – [in] the CUDA device number

  • nbytes – [in] number of bytes

Returns:

Host buffer or Status

Status FreeHost(void *data, int64_t nbytes)#

Free host memory.

The given memory pointer must have been allocated with AllocateHost.

class CudaContext : public std::enable_shared_from_this<CudaContext>#

Object-oriented interface to the low-level CUDA driver API.

Public Functions

Result<std::unique_ptr<CudaBuffer>> Allocate(int64_t nbytes)#

Allocate CUDA memory on GPU device for this context.

Parameters:

nbytes – [in] number of bytes

Returns:

the allocated buffer

Status Free(void *device_ptr, int64_t nbytes)#

Release CUDA memory on GPU device for this context.

Parameters:

  • device_ptr – [in] the buffer address

  • nbytes – [in] number of bytes

Returns:

Status

Result<std::shared_ptr<CudaBuffer>> View(uint8_t *data, int64_t nbytes)#

Create a view of CUDA memory on GPU device of this context.

Note

The caller is responsible for allocating and freeing the memory as well as ensuring that the memory belongs to the CUDA context that this CudaContext instance holds.

Parameters:

  • data – [in] the starting device address

  • nbytes – [in] number of bytes

Returns:

the view buffer

Result<std::shared_ptr<CudaBuffer>> OpenIpcBuffer(const CudaIpcMemHandle &ipc_handle)#

Open existing CUDA IPC memory handle.

Parameters:

ipc_handle – [in] opaque pointer to CUipcMemHandle (driver API)

Returns:

a CudaBuffer referencing the IPC segment

Status CloseIpcBuffer(CudaBuffer *buffer)#

Close memory mapped with IPC buffer.

Parameters:

buffer – [in] a CudaBuffer referencing

Returns:

Status

Status Synchronize(void)#

Block until the all device tasks are completed.

void *handle() const#

Expose CUDA context handle to other libraries.

std::shared_ptr<CudaMemoryManager> memory_manager() const#

Return the default memory manager tied to this context’s device.

std::shared_ptr<CudaDevice> device() const#

Return the device instance associated with this context.

int device_number() const#

Return the logical device number.

Result<uintptr_t> GetDeviceAddress(uint8_t *addr)#

Return the device address that is reachable from kernels running in the context.

The device address is defined as a memory address accessible by device. While it is often a device memory address, 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).

Parameters:

addr – [in] device or host memory address

Returns:

the device address

Devices#

class CudaDevice : public arrow::Device#

Device implementation for CUDA.

Each CudaDevice instance is tied to a particular CUDA device (identified by its logical device number).

Public Functions

virtual const char *type_name() const override#

A shorthand for this device’s type.

The returned value is different for each device class, but is the same for all instances of a given class. It can be used as a replacement for RTTI.

virtual std::string ToString() const override#

A human-readable description of the device.

The returned value should be detailed enough to distinguish between different instances, where necessary.

virtual bool Equals(const Device&) const override#

Whether this instance points to the same device as another one.

virtual std::shared_ptr<MemoryManager> default_memory_manager() override#

Return a MemoryManager instance tied to this device.

The returned instance uses default parameters for this device type’s MemoryManager implementation. Some devices also allow constructing MemoryManager instances with non-default parameters.

inline virtual DeviceAllocationType device_type() const override#

Return the DeviceAllocationType of this device.

inline virtual int64_t device_id() const override#

A device ID to identify this device if there are multiple of this type.

If there is no “device_id” equivalent (such as for the main CPU device on non-numa systems) returns -1.

int device_number() const#

Return the device logical number.

std::string device_name() const#

Return the GPU model name.

int64_t total_memory() const#

Return total memory on this device.

int handle() const#

Return a raw CUDA device handle.

The returned value can be used to expose this device to other libraries. It should be interpreted as CUdevice.

Result<std::shared_ptr<CudaContext>> GetContext()#

Get a CUDA driver context for this device.

The returned context is associated with the primary CUDA context for the device. This is the recommended way of getting a context for a device, as it allows interoperating transparently with any library using the primary CUDA context API.

Result<std::shared_ptr<CudaContext>> GetSharedContext(void *handle)#

Get a CUDA driver context for this device, using an existing handle.

The handle is not owned: it will not be released when the CudaContext is destroyed. This function should only be used if you need interoperation with a library that uses a non-primary context.

Parameters:

handle – [in] CUDA context handle created by another library

Result<std::shared_ptr<CudaHostBuffer>> AllocateHostBuffer(int64_t size)#

Allocate a host-residing, GPU-accessible buffer.

The buffer is allocated using this device’s primary context.

Parameters:

size – [in] The buffer size in bytes

virtual Result<std::shared_ptr<Device::Stream>> MakeStream(unsigned int flags) override#

Create a CUstream wrapper in the current context.

virtual Result<std::shared_ptr<Device::Stream>> WrapStream(void *device_stream, Stream::release_fn_t release_fn) override#

Wrap a pointer to an existing stream.

Parameters:

  • device_stream – passed in stream (should be a CUstream*)

  • release_fn – destructor to free the stream. nullptr may be passed to indicate there is no destruction/freeing necessary.

Public Static Functions

static Result<std::shared_ptr<CudaDevice>> Make(int device_number)#

Return a CudaDevice instance for a particular device.

Parameters:

device_number – [in] the CUDA device number

class Stream : public arrow::Device::Stream#

EXPERIMENTAL: Wrapper for CUstreams.

Does not own the CUstream object which must be separately constructed and freed using cuStreamCreate and cuStreamDestroy (or equivalent). Default construction will use the cuda default stream, and does not allow construction from literal 0 or nullptr.

Public Functions

virtual Status WaitEvent(const Device::SyncEvent&) override#

Make the stream wait on the provided event.

Tells the stream that it should wait until the synchronization event is completed without blocking the CPU.

virtual Status Synchronize() const override#

Blocks the current thread until a stream’s remaining tasks are completed.

class SyncEvent : public arrow::Device::SyncEvent#

Public Functions

virtual Status Wait() override#

Block until the sync event is marked completed.

virtual Status Record(const Device::Stream&) override#

Record the wrapped event on the stream.

Once the stream completes the tasks previously added to it, it will trigger the event.

class CudaMemoryManager : public arrow::MemoryManager#

MemoryManager implementation for CUDA.

Public Functions

virtual Result<std::shared_ptr<io::RandomAccessFile>> GetBufferReader(std::shared_ptr<Buffer> buf) override#

Create a RandomAccessFile to read a particular buffer.

The given buffer must be tied to this MemoryManager.

See also the Buffer::GetReader shorthand.

virtual Result<std::shared_ptr<io::OutputStream>> GetBufferWriter(std::shared_ptr<Buffer> buf) override#

Create a OutputStream to write to a particular buffer.

The given buffer must be mutable and tied to this MemoryManager. The returned stream object writes into the buffer’s underlying memory (but it won’t resize it).

See also the Buffer::GetWriter shorthand.

virtual Result<std::unique_ptr<Buffer>> AllocateBuffer(int64_t size) override#

Allocate a (mutable) Buffer.

The buffer will be allocated in the device’s memory.

std::shared_ptr<CudaDevice> cuda_device() const#

The CudaDevice instance tied to this MemoryManager.

This is a useful shorthand returning a concrete-typed pointer, avoiding having to cast the device() result.

virtual Result<std::shared_ptr<Device::SyncEvent>> MakeDeviceSyncEvent() override#

Creates a wrapped CUevent.

Will call cuEventCreate and it will call cuEventDestroy internally when the event is destructed.

virtual Result<std::shared_ptr<Device::SyncEvent>> WrapDeviceSyncEvent(void *sync_event, Device::SyncEvent::release_fn_t release_sync_event) override#

Wraps an existing event into a sync event.

Parameters:

  • sync_event – the event to wrap, must be a CUevent*

  • release_sync_event – a function to call during destruction, nullptr or a no-op function can be passed to indicate ownership is maintained externally

Buffers#

class CudaBuffer : public arrow::Buffer#

An Arrow buffer located on a GPU device.

Be careful using this in any Arrow code which may not be GPU-aware

Public Functions

Status CopyToHost(const int64_t position, const int64_t nbytes, void *out) const#

Copy memory from GPU device to CPU host.

Parameters:

  • position – [in] start position inside buffer to copy bytes from

  • nbytes – [in] number of bytes to copy

  • out – [out] start address of the host memory area to copy to

Returns:

Status

Status CopyFromHost(const int64_t position, const void *data, int64_t nbytes)#

Copy memory to device at position.

Parameters:

  • position – [in] start position to copy bytes to

  • data – [in] the host data to copy

  • nbytes – [in] number of bytes to copy

Returns:

Status

Status CopyFromDevice(const int64_t position, const void *data, int64_t nbytes)#

Copy memory from device to device at position.

Note

It is assumed that both source and destination device memories have been allocated within the same context.

Parameters:

  • position – [in] start position inside buffer to copy bytes to

  • data – [in] start address of the device memory area to copy from

  • nbytes – [in] number of bytes to copy

Returns:

Status

Status CopyFromAnotherDevice(const std::shared_ptr<CudaContext> &src_ctx, const int64_t position, const void *data, int64_t nbytes)#

Copy memory from another device to device at position.

Parameters:

  • src_ctx – [in] context of the source device memory

  • position – [in] start position inside buffer to copy bytes to

  • data – [in] start address of the another device memory area to copy from

  • nbytes – [in] number of bytes to copy

Returns:

Status

virtual Result<std::shared_ptr<CudaIpcMemHandle>> ExportForIpc()#

Expose this device buffer as IPC memory which can be used in other processes.

Note

After calling this function, this device memory will not be freed when the CudaBuffer is destructed

Returns:

Handle or Status

Public Static Functions

static Result<std::shared_ptr<CudaBuffer>> FromBuffer(std::shared_ptr<Buffer> buffer)#

Convert back generic buffer into CudaBuffer.

Note

This function returns an error if the buffer isn’t backed by GPU memory

Parameters:

buffer – [in] buffer to convert

Returns:

CudaBuffer or Status

class CudaHostBuffer : public arrow::MutableBuffer#

Device-accessible CPU memory created using cudaHostAlloc.

Public Functions

Result<uintptr_t> GetDeviceAddress(const std::shared_ptr<CudaContext> &ctx)#

Return a device address the GPU can read this memory from.

Memory Input / Output#

class CudaBufferReader : public arrow::io::internal::RandomAccessFileConcurrencyWrapper<CudaBufferReader>#

File interface for zero-copy read from CUDA buffers.

CAUTION: reading to a Buffer returns a Buffer pointing to device memory. It will generally not be compatible with Arrow code expecting a buffer pointing to CPU memory. Reading to a raw pointer, though, copies device memory into the host memory pointed to.

Public Functions

virtual bool closed() const override#

Return whether the stream is closed.

virtual bool supports_zero_copy() const override#

Return true if InputStream is capable of zero copy Buffer reads.

Zero copy reads imply the use of Buffer-returning Read() overloads.

class CudaBufferWriter : public arrow::io::WritableFile#

File interface for writing to CUDA buffers, with optional buffering.

Public Functions

virtual Status Close() override#

Close writer and flush buffered bytes to GPU.

virtual bool closed() const override#

Return whether the stream is closed.

virtual Status Flush() override#

Flush buffered bytes to GPU.

virtual Status Write(const void *data, int64_t nbytes) override#

Write the given data to the stream.

This method always processes the bytes in full. Depending on the semantics of the stream, the data may be written out immediately, held in a buffer, or written asynchronously. In the case where the stream buffers the data, it will be copied. To avoid potentially large copies, use the Write variant that takes an owned Buffer.

virtual Result<int64_t> Tell() const override#

Return the position in this stream.

Status SetBufferSize(const int64_t buffer_size)#

Set CPU buffer size to limit calls to cudaMemcpy.

By default writes are unbuffered

Parameters:

buffer_size – [in] the size of CPU buffer to allocate

Returns:

Status

int64_t buffer_size() const#

Returns size of host (CPU) buffer, 0 for unbuffered.

int64_t num_bytes_buffered() const#

Returns number of bytes buffered on host.

IPC#

class CudaIpcMemHandle#

Public Functions

Result<std::shared_ptr<Buffer>> Serialize(MemoryPool *pool = default_memory_pool()) const#

Write CudaIpcMemHandle to a Buffer.

Parameters:

pool – [in] a MemoryPool to allocate memory from

Returns:

Buffer or Status

Public Static Functions

static Result<std::shared_ptr<CudaIpcMemHandle>> FromBuffer(const void *opaque_handle)#

Create CudaIpcMemHandle from opaque buffer (e.g.

from another process)

Parameters:

opaque_handle – [in] a CUipcMemHandle as a const void*

Returns:

Handle or Status

Result<std::shared_ptr<CudaBuffer>> SerializeRecordBatch(const RecordBatch &batch, CudaContext *ctx)#

Write record batch message to GPU device memory.

Parameters:

  • batch – [in] record batch to write

  • ctx – [in] CudaContext to allocate device memory from

Returns:

CudaBuffer or Status

Result<std::shared_ptr<RecordBatch>> ReadRecordBatch(const std::shared_ptr<Schema> &schema, const ipc::DictionaryMemo *dictionary_memo, const std::shared_ptr<CudaBuffer> &buffer, MemoryPool *pool = default_memory_pool())#

ReadRecordBatch specialized to handle metadata on CUDA device.

Parameters:

  • schema – [in] the Schema for the record batch

  • dictionary_memo – [in] DictionaryMemo which has any dictionaries. Can be nullptr if you are sure there are no dictionary-encoded fields

  • buffer – [in] a CudaBuffer containing the complete IPC message

  • pool – [in] a MemoryPool to use for allocating space for the metadata

Returns:

RecordBatch or Status