Data graphs

Data graph pipelines encapsulate computational graphs that operate on whole resources (e.g. tensors). The graph operations that form these computational graphs are executed by graph processing engines.

Pipeline Creation

Sessions

Graph pipelines execute within data graph pipeline sessions that provide a context for their execution as well as binding points for the memory they need (e.g. transient storage).

Dispatch

Properties

Processing Engines

Data graph pipelines execute on one or more data graph processing engines. Each queue family of a VkPhysicalDevice that supports VK_QUEUE_DATA_GRAPH_BIT_ARM may provide access to processing engines. Processing engines may be foreign. Foreign processing engines require the use of external memory and semaphores to exchange data and synchronize with data graph pipelines that target them. Each processing engine may support a selection of graph operations. There exists a default processing engine that enables implementations to support some operations without tying support to specific processing engines. This allows implementations to avoid being specific about the processing engines that they use internally or to remove the necessity for applications to choose the specific processing engine used to execute a given set of operations.

Operations

Data graph processing engines execute data graph operations. Operations may be of a fixed-function nature, configurable or not, or may , for example, be provided as part of a SPIR-V extended instruction set.

Data Graph Operations Requirements

Physical devices that support the dataGraphShaderModule feature must support an operation whose VkPhysicalDeviceDataGraphOperationSupportARM::operationType is VK_PHYSICAL_DEVICE_DATA_GRAPH_OPERATION_TYPE_SPIRV_EXTENDED_INSTRUCTION_SET_ARM and name and version are as described in the following table for at least one processing engine in at least one queue family when specific extensions are supported at a specific version:

Physical devices that support VK_ARM_data_graph_optical_flow must support the following VkPhysicalDeviceDataGraphOperationSupportARM for at least one processing engine in at least one queue family:

• operationType: VK_PHYSICAL_DEVICE_DATA_GRAPH_OPERATION_TYPE_OPTICAL_FLOW_ARM
• name: OpticalFlow
• version: 1

TOSA

Optical flow

An optical flow fixed-function node in a graph calculates the forward optical flow between an input image and a reference image. An output flow vector (Vx, Vy) means that current pixel (x, y) of the input image can be found at location (x+Vx, y+Vy) in the reference image.

An optical flow fixed function nodes have the following graph node connection points:

• VK_DATA_GRAPH_PIPELINE_NODE_CONNECTION_TYPE_OPTICAL_FLOW_INPUT_ARM
• VK_DATA_GRAPH_PIPELINE_NODE_CONNECTION_TYPE_OPTICAL_FLOW_REFERENCE_ARM
• VK_DATA_GRAPH_PIPELINE_NODE_CONNECTION_TYPE_OPTICAL_FLOW_HINT_ARM
• VK_DATA_GRAPH_PIPELINE_NODE_CONNECTION_TYPE_OPTICAL_FLOW_FLOW_VECTOR_ARM
• VK_DATA_GRAPH_PIPELINE_NODE_CONNECTION_TYPE_OPTICAL_FLOW_COST_ARM

Image formats

Pipeline creation

Pipeline dispatch