Struct GpuAccelerator 

pub struct GpuAccelerator { /* private fields */ }

GPU-accelerated vector search engine.

Upload vectors once, then run many searches against them. If GPU initialization fails, callers should fall back to CPU SIMD.

Vectors are automatically split into chunks when they exceed the device’s max_storage_buffer_binding_size (typically 128 MB). Searches dispatch against each chunk and merge results transparently.

Implementations

impl GpuAccelerator

pub fn is_available() -> bool

Check if any GPU is available on this system.

pub fn new() -> Result<Self>

Initialize the best available GPU device.

Requests the adapter’s maximum buffer limits so that chunking only kicks in when truly necessary.

pub fn device_info(&self) -> &DeviceInfo

pub fn max_storage_buffer_binding_size(&self) -> u32

pub fn upload_vectors(&mut self, vectors: &[f32], dim: usize) -> Result<()>

Upload a flat array of vectors to GPU memory. Automatically splits into chunks when data exceeds the binding limit.

pub fn upload_norms(&mut self, norms: &[f32]) -> Result<()>

Upload pre-computed L2 norms, split to match vector chunk layout.

pub fn cosine_search( &self, query: &[f32], k: usize, ) -> Result<Vec<(usize, f32)>>

Cosine similarity search: returns top-k (index, score) pairs, highest first. Dispatches against each vector chunk and merges results.

pub fn batch_cosine_search( &self, queries: &[Vec<f32>], k: usize, ) -> Result<Vec<Vec<(usize, f32)>>>

Batch cosine search: multiple queries at once.

pub fn l2_search(&self, query: &[f32], k: usize) -> Result<Vec<(usize, f32)>>

L2 distance search: returns top-k (index, distance) pairs, smallest first.

pub fn compute_norms(&self) -> Result<Vec<f32>>

Compute L2 norms for all uploaded vectors on GPU.

pub fn compute_norms_gpu(&self, vectors: &[f32], dim: usize) -> Result<Vec<f32>>

Compute L2 norms from raw vectors (not previously uploaded). Handles chunking automatically for large inputs.

pub fn batch_dot_product( &self, queries_flat: &[f32], num_queries: usize, ) -> Result<Vec<f32>>

Batch dot product: queries [Q×D] × vectors [N×D] -> flat [Q×N] scores.

pub fn distance_matrix( &self, queries: &[f32], vectors: &[f32], dim: usize, ) -> Result<Vec<Vec<f32>>>

Compute L2 distance matrix: queries × vectors -> Q×N distances. Uses 16×16 workgroup tiling for cache efficiency.

pub fn f16_to_f32_batch(&self, f16_bits: &[u16]) -> Result<Vec<f32>>

Convert f16 values (as raw u16 bits) to f32 on the GPU.

pub fn f32_to_f16_batch(&self, values: &[f32]) -> Result<Vec<u16>>

Convert f32 values to f16 (as raw u16 bits) on the GPU.

pub fn vector_count(&self) -> usize

pub fn dimension(&self) -> usize

pub fn chunk_count(&self) -> usize