INNER CODE UNIT · Rust

nccl4rust_get_multimem_pointer

NVIDIA/nccl · contrib/nccl4rust/crates/nccl-device-sys/src/lib.rs:84

    pub fn nccl4rust_get_multimem_pointer(
        window: ncclWindow_t,
        offset: usize,
        mc_base_ptr: *mut c_void,
    ) -> *mut c_void;
    pub fn nccl4rust_get_lsa_multimem_pointer(
        window: ncclWindow_t,
        offset: usize,
        comm: *const ncclDevComm_t,
    ) -> *mut c_void;

    pub fn nccl4rust_lsa_barrier_thread(comm: *const ncclDevComm_t, index: u32, multimem: bool);
    pub fn nccl4rust_lsa_barrier_warp(comm: *const ncclDevComm_t, index: u32, multimem: bool);
    pub fn nccl4rust_lsa_barrier_cta(comm: *const ncclDevComm_t, index: u32, multimem: bool);

    /// LSA f32 reduce-sum-copy using all threads in the CTA.
    ///
    /// This primitive does not add entry/exit barriers. Callers must establish

View source record →

📰 Research Paper
Loading…
⏳ Fetching content…