Expand description
Solver GPU kernels and dispatch wiring.
Process-wide GPU policy and direct device selection live in crate::gpu;
the domain compute kernels live in crate::gpu_kernels. This module owns
the solver-side device dispatch glue: the PIRLS CUDA kernel bodies
(pirls_gpu, [pirls_dispatch_wire]), their PIRLS-side host admission
shim ([pirls_host_dispatch], folded here from solver/pirls/gpu_dispatch.rs),
and the REML/arrow-Schur device entry points.
Modulesยง
- arrow_
schur_ gpu - Caller-facing thin wrapper around
crate::gpu_kernels::arrow_schur. - pirls_
gpu - reml_
gpu - Exact GPU REML evidence + derivative gradient.