Device API

The Device API allows communication to be initiated and performed from device (GPU) code. It is organized into the following areas:

• Host-Side Setup — Creating and configuring device communicators, querying properties, host-accessible device pointer functions, and related types.

• Memory and LSA — Load/store accessible (LSA) memory, barriers, pointer accessors, and multimem.

• GIN (GPU-Initiated Networking) — One-sided transfers, signals, counters, and network barriers.

• Reduce, Broadcast, and Fused Building Blocks — Building blocks for computation-fused kernels: reduce, copy (broadcast), and reduce-then-copy; used to implement algorithms such as AllReduce, AllGather, and ReduceScatter.

For an introduction and usage examples, see Device-Initiated Communication.

• Device API – Host-Side Setup
  • Host-Side Setup
  • Host-Accessible Device Pointer Functions
• Device API – Memory and LSA
  • LSA
  • Multimem
• Device API – GIN
  • GIN
• Device API – Remote Reduce and Copy: Building Blocks for Custom Communication Kernels
  • Compile-Time Requirements
  • API Overview
  • ReduceSum — N Sources to One Destination
  • Copy (Broadcast) — One Source to N Destinations
  • ReduceSumCopy
  • Lambda-Based (Custom Layouts)
  • Custom Reduction Operators