Struct CutlassActor 

pub struct CutlassActor { /* private fields */ }

Host-side actor. Holds an Arc<CutlassInner> so messages can be processed from a worker thread without locking the actor itself after construction.

Implementations

impl CutlassActor

pub fn new(plan_cache_capacity: usize) -> Self

Examples found in repository

examples/cutlass_gemm_fp8.rs (line 11)

fn main() {
    let actor = CutlassActor::new(16);
    let req = GemmRequest::<F8E4m3>::new(GemmShape::new(4096, 4096, 4096), SmArch::Sm90a)
        .with_epilogue(GemmEpilogue::LinearReLU {
            alpha: 1.0,
            beta: 0.0,
        });

    println!("plan key: {:?}", req.plan_key());
    let (src, name) = req.render_cu();
    println!("kernel:   {name}");
    println!("--- generated .cu ---");
    println!("{src}");

    actor.handle(CutlassMsg::Gemm(Box::new(req.clone())));
    actor.handle(CutlassMsg::Gemm(Box::new(req)));

    println!("dispatched: {}", actor.inner().dispatched());
    println!("plan cache len: {}", actor.inner().plan_cache.len());
}

pub fn prebuilt_active() -> bool

true when the crate was built with cutlass-prebuilt and nvcc was found at build time, so libatomr_cutlass_prebuilt.a is statically linked into the binary. Phase 6.1 ships one canonical GEMM placeholder cell (proves the wiring); Phase 6.2 expands the cell matrix and routes hits through the prebuilt symbol table before falling back to NVRTC.

pub fn inner(&self) -> Arc<CutlassInner>

Examples found in repository

examples/cutlass_gemm_fp8.rs (line 27)

fn main() {
    let actor = CutlassActor::new(16);
    let req = GemmRequest::<F8E4m3>::new(GemmShape::new(4096, 4096, 4096), SmArch::Sm90a)
        .with_epilogue(GemmEpilogue::LinearReLU {
            alpha: 1.0,
            beta: 0.0,
        });

    println!("plan key: {:?}", req.plan_key());
    let (src, name) = req.render_cu();
    println!("kernel:   {name}");
    println!("--- generated .cu ---");
    println!("{src}");

    actor.handle(CutlassMsg::Gemm(Box::new(req.clone())));
    actor.handle(CutlassMsg::Gemm(Box::new(req)));

    println!("dispatched: {}", actor.inner().dispatched());
    println!("plan cache len: {}", actor.inner().plan_cache.len());
}

pub fn handle(&self, msg: CutlassMsg)

Synchronously process a message. The real production path runs through atomr_core::actor::Actor::handle; this method is the host-only fast path that the unit tests exercise.

Examples found in repository

examples/cutlass_gemm_fp8.rs (line 24)

fn main() {
    let actor = CutlassActor::new(16);
    let req = GemmRequest::<F8E4m3>::new(GemmShape::new(4096, 4096, 4096), SmArch::Sm90a)
        .with_epilogue(GemmEpilogue::LinearReLU {
            alpha: 1.0,
            beta: 0.0,
        });

    println!("plan key: {:?}", req.plan_key());
    let (src, name) = req.render_cu();
    println!("kernel:   {name}");
    println!("--- generated .cu ---");
    println!("{src}");

    actor.handle(CutlassMsg::Gemm(Box::new(req.clone())));
    actor.handle(CutlassMsg::Gemm(Box::new(req)));

    println!("dispatched: {}", actor.inner().dispatched());
    println!("plan cache len: {}", actor.inner().plan_cache.len());
}