Skip to main content

Crate tenferro_cpu

Crate tenferro_cpu 

Source
Expand description

CPU execution engine and kernels.

There is one CPU backend: faer-backed tlinalg for linalg and cpueinsum/tcontract for contraction, linked unconditionally. The crate has no backend feature; its only two features, ledger and route-report, are off by default, gate host-check hooks that are #[doc(hidden)], and change no numerical path.

A CpuEngine owns one Pool built by the caller (or none) and the workspace its operations write. Every operation takes the engine as &mut:

use tenferro_cpu::{Pool, CpuEngine};
use tenferro_tensor::{Tensor, TensorElementwise, TensorRead};

let a = Tensor::from_vec_col_major(vec![2], vec![1.0_f64, 2.0])?;
let b = Tensor::from_vec_col_major(vec![2], vec![3.0_f64, 4.0])?;

let mut serial = CpuEngine::serial();
let c = serial.add_read(TensorRead::from_tensor(&a), TensorRead::from_tensor(&b))?;
assert_eq!(c.as_slice::<f64>()?, &[4.0, 6.0]);

let mut parallel = CpuEngine::new(Pool::builder().threads(2).build()?);
let d = parallel.add_read(TensorRead::from_tensor(&a), TensorRead::from_tensor(&b))?;
assert_eq!(d.as_slice::<f64>()?, &[4.0, 6.0]);

Re-exports§

pub use threadteam;
pub use faer;

Modules§

backend
Pool helpers shared by the CPU engine’s operations.

Macros§

same_variant_pair
Dispatch a same-variant pair of tensors to a typed kernel.
same_variant_unary
Dispatch a single tensor to a typed kernel and erase its result.

Structs§

AddOp
Ordinary CPU entry points that take a caller-provided destination and the caller’s own arithmetic instead of the typed pool. Addition.
BufferPool
Typed buffer pool keyed by element capacity and separated by scalar type.
BufferPoolStats
Snapshot of typed host buffers retained by a BufferPool.
CpuBatchPolicy
A batch strategy together with the CpuBatchThresholds Auto applies.
CpuBatchThresholds
The cost and fan-out model CpuBatchStrategy::Auto applies.
CpuEngine
A caller-held CPU execution engine.
CpuPlanCache
A shareable, bounded cache of immutable CPU numerical plans.
IndexedPlanCacheLimits
Limits for the CPU indexed-plan cache.
MulOp
Ordinary CPU entry points that take a caller-provided destination and the caller’s own arithmetic instead of the typed pool. Multiplication.
Pool
A fixed set of worker threads that Exec tokens place work on.
PoolBuilder
Configuration of a Pool.
SerialCtx
The serial context CpuEngine::with_guard lends: the engine’s workspace and batch policy, without a placement token.
SubOp
Ordinary CPU entry points that take a caller-provided destination and the caller’s own arithmetic instead of the typed pool. Subtraction.

Enums§

CpuBatchStrategy
How the items of one batched operation are split over a thread budget.
CpuEngineError
Typed refusals reported by a CpuEngine before an operation runs.

Traits§

BinaryScalarOp
Ordinary CPU entry points that take a caller-provided destination and the caller’s own arithmetic instead of the typed pool. A binary operation between two scalars of the same type.

Functions§

cpu_capabilities
Return the CPU backend operation capability descriptor table.
runtime_engine_id
Return the canonical CPU runtime engine identifier.
runtime_engine_registration
Build a runtime engine registration for a CpuEngine.
runtime_engine_registration_with_id
Build a runtime engine registration for a CpuEngine under engine_id.
runtime_hardware_class
Return the canonical CPU runtime hardware class.
scalar_binary_into
Ordinary CPU entry points that take a caller-provided destination and the caller’s own arithmetic instead of the typed pool. Apply a named binary operation elementwise into a caller-owned destination.
scalar_fold
Ordinary CPU entry points that take a caller-provided destination and the caller’s own arithmetic instead of the typed pool. Fold every element of a caller-owned tensor with a named associative operation, starting from init.
with_batch_policy
Run f on engine with policy as the effective CPU batch policy.