scirs2_neural::performance::threading::distributed

Trait ProcessGroup

Source 
pub trait ProcessGroup: Send + Sync {
    // Required methods
    fn all_reduce(&self, tensor: &mut ArrayD<f32>) -> Result<()>;
    fn barrier(&self) -> Result<()>;
    fn broadcast(&self, tensor: &mut ArrayD<f32>, root: usize) -> Result<()>;
    fn get_rank(&self) -> usize;
    fn get_world_size(&self) -> usize;
}
Expand description

Process group trait for different communication backends

Required Methods§

Source

fn all_reduce(&self, tensor: &mut ArrayD<f32>) -> Result<()>

Perform all-reduce operation on tensor across all processes

Source

fn barrier(&self) -> Result<()>

Synchronize all processes

Source

fn broadcast(&self, tensor: &mut ArrayD<f32>, root: usize) -> Result<()>

Broadcast tensor from root process to all others

Source

fn get_rank(&self) -> usize

Get the rank of current process

Source

fn get_world_size(&self) -> usize

Get the total number of processes

Implementors§