dspk-core 0.0.10

Core types for dspk: recordings, buffers, time, masks, windows, and the CubeCL compute layer with run-time device choice (wgpu, CUDA, HIP, CPU)
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

//! The compute layer (`compute` feature): the single entry point through which every crate reaches
//! a CubeCL runtime.
//!
//! - [`ComputeTask`] / [`ComputeTarget::run`](crate::device::ComputeTarget): run generic work on
//!   the selected runtime.
//! - [`LaunchGeometry`]: cube sizes and counts from the runtime's properties.
//! - [`tune`]: CubeCL autotuning for launch settings with no device-independent best value.
//! - [`mod@bench`]: device-synchronised timing.
//!
//! Algorithm crates (`dspk-ops`, `dspk-synapse`, …) take a `Client` and use these; they
//! never inspect or special-case the device.

pub mod bench;
pub mod launch;
pub mod limits;
pub mod pinning;
mod special;
mod target;
pub mod tune;

pub use crate::device::{ComputeError, ComputeTarget, RUNTIME_ENV, TEST_RUNTIMES_ENV};
pub use target::{configure_runtime, exit_now, open_device, shutdown_devices, CACHE_DIR_ENV, KERNEL_CACHE_ENV};
pub use limits::{device_elements, MAX_DEVICE_ELEMENTS};
pub use pinning::{pin_tuned_choices, tuned_choices_pinned, PinnedChoices};
pub use launch::{LaunchGeometry, RowShare, channel_position, element_position, row_position, row_share, sample_position};
pub use special::{negative_infinity, positive_infinity, NEGATIVE_INFINITY_BITS, POSITIVE_INFINITY_BITS};
pub use target::ComputeTask;