Struct DeviceReduction 

pub struct DeviceReduction<T: FvmScalar, B: AcceleratorBackend> { /* private fields */ }

Reusable one-scalar output allocation for vector reductions.

Implementations

impl<T: FvmScalar, B: AcceleratorBackend> DeviceReduction<T, B>

pub fn new(backend: &B, input_len: usize) -> Result<Self, AcceleratorError>

Allocate a reduction workspace for vectors of exactly input_len values.

pub fn download(&self, backend: &B) -> Result<T, AcceleratorError>

Download the most recently computed scalar.

impl<T: FvmScalar> DeviceReduction<T, CpuBackend>

pub fn sum( &mut self, input: &<CpuBackend as AcceleratorBackend>::Buffer<T>, ) -> Result<T, AcceleratorError>

Deterministic sum in input order.

pub fn dot( &mut self, lhs: &<CpuBackend as AcceleratorBackend>::Buffer<T>, rhs: &<CpuBackend as AcceleratorBackend>::Buffer<T>, ) -> Result<T, AcceleratorError>

Deterministic dot product in input order.

pub fn l2_norm( &mut self, input: &<CpuBackend as AcceleratorBackend>::Buffer<T>, ) -> Result<T, AcceleratorError>

Euclidean norm.

pub fn max_abs( &mut self, input: &<CpuBackend as AcceleratorBackend>::Buffer<T>, ) -> Result<T, AcceleratorError>

Maximum absolute value, or zero for an empty vector.