cudla

Note

The cuDLA bindings require a Jetson platform with DLA hardware (Xavier or Orin). cuDLA is not available on desktop GPUs.

Functions

cuDLA defines the following functions for DLA device management and inference.

cuda.bindings.cudla.get_version() → uint64_t

cuda.bindings.cudla.device_get_count() → uint64_t

cuda.bindings.cudla.create_device(uint64_t device, uint32_t flags) → intptr_t

cuda.bindings.cudla.destroy_device(intptr_t dev_handle)

cuda.bindings.cudla.mem_register(

intptr_t dev_handle,

intptr_t ptr,

size_t size,

uint32_t flags,

) → intptr_t

cuda.bindings.cudla.mem_unregister(intptr_t dev_handle, intptr_t dev_ptr)

cuda.bindings.cudla.module_load_from_memory(

intptr_t dev_handle,

p_module,

size_t module_size,

uint32_t flags,

) → intptr_t

cuda.bindings.cudla.module_get_attributes(intptr_t h_module, int attr_type)

Query module attributes, interpreting the cudlaModuleAttribute union based on the requested attribute type.

For count attributes (NUM_INPUT_TENSORS, NUM_OUTPUT_TENSORS, NUM_OUTPUT_TASK_STATISTICS), returns an int.

For descriptor attributes (INPUT_TENSOR_DESCRIPTORS, OUTPUT_TENSOR_DESCRIPTORS, OUTPUT_TASK_STATISTICS_DESCRIPTORS), returns a list of ModuleTensorDescriptor objects.

cuda.bindings.cudla.module_unload(intptr_t h_module, uint32_t flags)

cuda.bindings.cudla.submit_task(

intptr_t dev_handle,

intptr_t ptr_to_tasks,

uint32_t num_tasks,

intptr_t stream,

uint32_t flags,

)

cuda.bindings.cudla.device_get_attribute(intptr_t dev_handle, int attrib)

cuda.bindings.cudla.get_last_error(intptr_t dev_handle) → int

cuda.bindings.cudla.set_task_timeout_in_ms(intptr_t dev_handle, uint32_t timeout)

Types

class cuda.bindings.cudla.ExternalMemoryHandleDesc

Empty-initialize an instance of cudlaExternalMemoryHandleDesc_t.

See also

cudlaExternalMemoryHandleDesc_t

class cuda.bindings.cudla.ExternalSemaphoreHandleDesc

Empty-initialize an instance of cudlaExternalSemaphoreHandleDesc_t.

See also

cudlaExternalSemaphoreHandleDesc_t

class cuda.bindings.cudla.ModuleTensorDescriptor

Empty-initialize an instance of cudlaModuleTensorDescriptor.

See also

cudlaModuleTensorDescriptor

class cuda.bindings.cudla.Fence

Empty-initialize an instance of CudlaFence.

See also

CudlaFence

class cuda.bindings.cudla.DevAttribute

Empty-initialize an instance of cudlaDevAttribute.

See also

cudlaDevAttribute

class cuda.bindings.cudla.ModuleAttribute

Empty-initialize an instance of cudlaModuleAttribute.

See also

cudlaModuleAttribute

class cuda.bindings.cudla.WaitEvents

Empty-initialize an instance of cudlaWaitEvents.

See also

cudlaWaitEvents

class cuda.bindings.cudla.SignalEvents

Empty-initialize an instance of cudlaSignalEvents.

See also

cudlaSignalEvents

class cuda.bindings.cudla.Task

Empty-initialize an instance of cudlaTask.

See also

cudlaTask

Enums

class cuda.bindings.cudla.Status(

value,

names=<not given>,

*values,

module=None,

qualname=None,

type=None,

start=1,

boundary=None,

)

See cudlaStatus.

Success = 0

class cuda.bindings.cudla.Mode(

value,

names=<not given>,

*values,

module=None,

qualname=None,

type=None,

start=1,

boundary=None,

)

See cudlaMode.

CUDA_DLA = 0

STANDALONE = 1

class cuda.bindings.cudla.ModuleAttributeType(

value,

names=<not given>,

*values,

module=None,

qualname=None,

type=None,

start=1,

boundary=None,

)

See cudlaModuleAttributeType.

class cuda.bindings.cudla.FenceType(

value,

names=<not given>,

*values,

module=None,

qualname=None,

type=None,

start=1,

boundary=None,

)

See cudlaFenceType.

class cuda.bindings.cudla.ModuleLoadFlags(

value,

names=<not given>,

*values,

module=None,

qualname=None,

type=None,

start=1,

boundary=None,

)

See cudlaModuleLoadFlags.

class cuda.bindings.cudla.SubmissionFlags(

value,

names=<not given>,

*values,

module=None,

qualname=None,

type=None,

start=1,

boundary=None,

)

See cudlaSubmissionFlags.

class cuda.bindings.cudla.AccessPermissionFlags(

value,

names=<not given>,

*values,

module=None,

qualname=None,

type=None,

start=1,

boundary=None,

)

See cudlaAccessPermissionFlags.

class cuda.bindings.cudla.DevAttributeType(

value,

names=<not given>,

*values,

module=None,

qualname=None,

type=None,

start=1,

boundary=None,

)

See cudlaDevAttributeType.