Crate gam_gpu 

Re-exports

pub use cpu_traits::MatrixLocation;

pub use device::GpuDeviceInfo;

pub use device_runtime::GpuRuntime;

pub use gpu_error::GpuError;

pub use memory::DeviceBuffer;

pub use memory::DeviceCsrMatrix;

pub use memory::DeviceMatrix;

pub use memory::DeviceVector;

pub use policy::GpuDispatchPolicy;

pub use policy::GpuMixedPrecisionPolicy;

pub use pool::balanced_partition;

pub use pool::scatter_batched;

pub use profile::GpuExecutionTelemetry;

pub use profile::KernelStat;

pub use profile::KernelStatsSnapshot;

Modules

backend_probe

Shared CUDA backend-probe contract for every cudarc-backed module under src/gpu/*.

blas

Device BLAS surface for the cudarc-backed dense kernels.

calibration

cpu_traits

device

device_cache

Shared host-side scaffolding for every cudarc-backed module under src/gpu/* and src/solver/gpu/*.

device_runtime

driver

Shared CUDA driver presence/loading helpers used by every cuBLAS / cuSPARSE / cuSOLVER routing module.

gpu_error

Typed error for the src/gpu/* modules.

kernels

Domain-specific GPU kernels live in their owning algorithm crates.

linalg_dispatch

Automatic GPU dispatch shim for dense linear algebra hot kernels.

memory

numerics_device

Shared device-side probit numerics for NVRTC kernels.

numerics_host

Host-side scalar special functions shared by the CPU parity references of the GPU backends.

policy

pool

Multi-GPU device pool.

profile

solver

cuSOLVER-backed dense solver kernels for the GPU HAL.

Macros

gpu_bail

return Err(GpuError::DriverCallFailed { reason: format!(...) }).

gpu_err

Build a GpuError::DriverCallFailed { reason: format!(...) } value.

Structs

GpuDecision

A backend-selection decision for a single hot kernel.

Enums

CudaBackendStatus

GpuEligibility

Joint eligibility state for a GPU kernel at the call site.

GpuKernel

GpuMode

Fail-closed GPU residency mode (issue #1017).

GpuPolicy

User-facing GPU backend policy.

Functions

configure_global_policy

Configure the process-wide policy before solver kernels are selected. If a previous explicit configuration already set the policy, the first value wins so concurrent fits cannot race policy changes. Reads of global_policy() never claim the slot, so the very first explicit configuration always sticks even if dispatch code observed the default Auto beforehand.

cuda_selected

True when direct solver GPU entry points should be attempted.

decide

Decide whether a GPU kernel may run. This is deliberately conservative: with no compiled vendor backend, auto returns CPU fallback and force returns an error at the call site through GpuDecision::require_supported.

global_policy

gpu_mode

Read the process-wide GPU residency mode. Defaults to GpuMode::Auto without claiming the slot, mirroring global_policy so an incidental read never locks the mode against a later explicit set_gpu_mode.

log_backend_inventory_once

Emit the roadmap-visible kernels at startup/debug time without affecting numerical execution. This keeps backend coverage auditable as real device kernels are added incrementally.

set_gpu_mode

Configure the process-wide GPU residency mode. First-writer-wins so concurrent fits cannot race the contract; a redundant late call is ignored.

try_cholesky_batched_lower_inplace

try_cholesky_lower_inplace

try_fast_ab

try_fast_ab_broadcast_b_batched

try_fast_abt_strided_batched

try_fast_atb_on_ordinal

try_fast_atv

try_fast_av

try_solve_lower_triangular_matrix

try_solve_upper_triangular_matrix