domains / families / devops / doca-gpi

doca-gpi

separate

Use this skill for DOCA GPI work — configuring a GPU Packet Initiator so device code issues RDMA operations straight out of GPU-resident buffers via doca_gpi objects: domain creation and channel limits, attribute objects, the doca_gpu_gpi_channel* handoff consumed inside device code, and DOCA_ERROR_* triage on doca_gpi_* entry points such as doca_gpi_gpu_channel_get. Also load it for implied asks like "how many channels fit in a GPI domain" or "device code issuing one-sided RDMA". Refuse and route to the proper owner for Send/Receive semantics (doca-gpunetio), host-managed QP lifecycle (doca-rdma), Arm-side origination (doca-rdmi), or generic CUDA tutelage.

When to use

• configuring a GPU Packet Initiator
• issuing RDMA operations from GPU-resident buffers
• domain creation and channel limits
• attribute objects
• doca_gpu_gpi_channel* handoff
• DOCA_ERROR_* triage on doca_gpi_* entry points
• sizing channels and work queues
• connecting GPI channels to remote peers

When not to use

• Send/Receive semantics
• host-managed QP lifecycle
• Arm-side origination
• generic CUDA tutelage
• higher-level Send/Receive Ethernet-shaped GPU NetIO API

Skill page

Source