Module linalg_dispatch 

Automatic GPU dispatch shim for dense linear algebra hot kernels.

Every try_* entry point in this module is invoked unconditionally from gam_linalg::faer_ndarray before the CPU fast-path runs. The decision to send the kernel to a device is fully automatic and never requires a user-facing flag — it depends only on:

1. Lossless runtime resolution returning an available device.
2. The kernel being large enough to amortize launch/PCIe overhead, per the thresholds in policy::GpuDispatchPolicy.
3. cudarc successfully dynamically loading libcuda at process startup via its fallback-dynamic-loading feature. When the loader fails (no driver, no toolkit installed), Auto receives typed absence and every try_* returns None so the caller falls through to the existing faer CPU kernel. Probe faults and Required absence fail loudly instead of being reclassified as an optional decline.

The wiring lives here so solver/pirls.rs and the family Hessian assemblers can stay backend-agnostic: they call gam_linalg::faer_ndarray::fast_* and get GPU acceleration automatically whenever it is profitable.

Structs

CudaGemmDispatch

ResidentDesignGram

#1017 Phase 3: a device-resident design matrix for repeated Xᵀ·diag(w)·X Gram evaluations that uploads X to the device ONCE.

Enums

DispatchOp

Discriminator used by route_through_gpu to apply the right size threshold from super::policy::GpuDispatchPolicy.

Functions

route_through_gpu

Returns Some(runtime) when a device is available and policy admits the operation. Auto applies calibrated profitability thresholds; Required deliberately bypasses those thresholds because a CPU continuation would violate the requested execution policy.

route_through_gpu_with_policy

Per-request counterpart of route_through_gpu. This is the device seam for solvers whose policy is part of an immutable fit request rather than the legacy process-wide configuration.

try_cholesky_batched_lower_inplace

try_cholesky_batched_lower_inplace_with_policy

try_cholesky_lower_inplace

try_fast_ab

try_fast_ab_broadcast_b_batched

try_fast_abt_strided_batched

try_fast_abt_strided_batched_with_policy

try_fast_atb

try_fast_atb_on_ordinal

Aᵀ·B on a specific device ordinal, for pool-tiled callers that already own the ordinal (the worker thread has bound that ordinal’s context). Semantics are identical to try_fast_atb — a is m×k, b is m×n, output is the k×n product aᵀ·b — but the kernel is pinned to ordinal instead of the probe-selected primary device. Auto returns None only when CUDA is absent or the shape is below policy threshold, so the caller can run its CPU path. A post-admission backend failure is fatal under every policy; Required also makes pre-admission absence fatal. f64 only.

try_fast_atv

try_fast_av

try_fast_joint_hessian_2x2

try_fast_spectral_leverage_diagonal

GPU-offloaded spectral leverage diagonal h[i] = ‖(X G)_{i,:}‖².

try_fast_xt_diag_x

try_fast_xt_diag_y

try_solve_lower_triangular_matrix

try_solve_upper_triangular_matrix