Crate zer_compute 

zer-compute, hardware-accelerated backend for entity resolution.

Provides DeviceComparator and DeviceScorer as drop-in replacements for the CPU-only counterparts in zer-compare. Both implement the same zer_core::traits::Comparator and zer_core::traits::Scorer traits, so the rest of the pipeline is backend-agnostic.

Backend selection

use std::sync::Arc;
use zer_compute::{GpuBackend, DeviceComparator, DeviceScorer};
use zer_core::schema::{FieldKind, SchemaBuilder};

let schema = SchemaBuilder::new()
    .field("naam",  FieldKind::Name)
    .field("datum", FieldKind::Date)
    .build()
    .unwrap();

// Auto-detect: tries CUDA → AVX2 → CPU in order.
let backend    = Arc::new(GpuBackend::auto_detect());
let comparator = DeviceComparator::new(Arc::clone(&backend), &schema).unwrap();
let scorer     = DeviceScorer::new(Arc::clone(&backend));

Feature flags

Flag	Description
cuda	NVIDIA CUDA via cudarc, requires CUDA Toolkit 13.1+ and nvcc at build time
vulkan	Vulkan 1.3 compute via ash, requires slangc on PATH at build time
avx2	x86_64 AVX2 SIMD via std::arch, no external toolchain required
cpu	Explicit scalar CPU path backed by zer-compare (Rayon parallel)
debug-shaders	Embed debug info in compiled CUDA kernels for cuda-gdb / Nsight stepping

When no flag is set the crate compiles and runs normally using the always-available scalar CPU fallback backed by zer-compare.

Re-exports

pub use backend::BackendPreference;

pub use backend::DeviceBackend;

pub use backend::GpuBackend;

pub use batch_sizer::BatchSizer;

pub use comparator::DeviceComparator;

pub use error::GpuError;

pub use scorer::DeviceScorer;

Modules

backend

Backend abstraction, DeviceBackend enum, BackendPreference, and auto-detection logic.

batch_sizer

Auto-tunes the GPU batch size from available VRAM using the exact buffer layout.

comparator

DeviceComparator, implements the Comparator trait.

error

kernel

Core trait definitions for the kernel dispatch system.

kernels

Typed kernel descriptors, one module per compute operation.

scorer

DeviceScorer, implements the Scorer trait with GPU-accelerated EM.

soa