Struct PirlsRowBackend 

pub struct PirlsRowBackend { /* private fields */ }

Process-wide cache of compiled per-family modules.

Implementations

impl PirlsRowBackend

pub const fn compiled() -> bool

pub fn probe() -> Result<&'static Self, GpuError>

pub fn module_for( &self, family: PirlsRowFamily, curvature: CurvatureMode, ) -> Result<Arc<CudaModule>, GpuError>

Compile (or fetch from cache) the final-row kernel module for (family, curvature). Writes all 9 output fields.

pub fn module_for_solve( &self, family: PirlsRowFamily, curvature: CurvatureMode, ) -> Result<Arc<CudaModule>, GpuError>

Compile (or fetch from cache) the solve-row kernel module for (family, curvature). Writes only grad_eta, w_solver, deviance, status — used on every hot Newton iteration.

pub fn module_for_ladder( &self, family: PirlsRowFamily, curvature: CurvatureMode, ) -> Result<Arc<CudaModule>, GpuError>

Compile (or fetch from cache) the alpha-ladder kernel module for (family, curvature). Evaluates all ALPHA_LADDER_LEN step sizes in a single launch, accumulating objective[] and status[] per alpha slot.

pub fn module_for_jit( &self, spec: &JitFamilySpec, curvature: CurvatureMode, ) -> Result<Arc<CudaModule>, GpuError>

Stage 6: JIT-compile and cache a custom-family row module.

The kernel name is pirls_row_jit_{spec.spec_id} so multiple distinct JIT specs in the same process get distinct cached modules. The cache key is (spec_id, curvature) which mirrors the built-in (family, curvature) cache and reuses the same HashMap-of-Arc<CudaModule> (but with a synthetic ModuleKey derived from the spec_id).

Note: a fresh (spec_id, curvature) recompiles via NVRTC the first time; subsequent fits in the same process hit the cache. Spec changes (different body) must use a different spec_id so that the cache does NOT return a stale module.