atomr-accel-cuda 0.10.0

GPU acceleration via the actor model. Wraps NVIDIA CUDA libraries (cuBLAS, cuDNN, cuFFT, cuRAND, cuSOLVER, cuSPARSE, cuTENSOR, cuBLASLt, NVRTC, NCCL) as supervised atomr actors with generation-validated buffers and a uniform async surface.
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

//! [`PipelineStage`] trait.

use std::sync::Arc;

use cudarc::driver::{CudaEvent, CudaStream};

use crate::error::GpuError;

/// One stage in a multi-stream GPU pipeline.
///
/// Implementations enqueue their kernel onto `stream` synchronously
/// (no host wait) and return a `CudaEvent` marking the completion of
/// that stage's GPU work, plus the typed output. The executor
/// arranges that the next stage's `wait_for` is the previous stage's
/// returned event, so cross-stage synchronization is on-device only.
pub trait PipelineStage: Send + 'static {
    type In: Send + 'static;
    type Out: Send + 'static;

    fn enqueue(
        &mut self,
        stream: &Arc<CudaStream>,
        wait_for: Option<&CudaEvent>,
        input: Self::In,
    ) -> Result<(CudaEvent, Self::Out), GpuError>;
}