Struct SgemmOp 

pub struct SgemmOp { /* private fields */ }

SGEMM op for graph capture: C := alpha · A·B + beta · C, column-major, no transpose.

The op needs a cuBLAS handle on the captured stream, supplied by GraphRecordCtx::blas. If absent the op fails with GpuError::Unrecoverable.

record consumes the held GpuRefs on first invocation (matching the pre-trait closed-enum semantics where the boxed op was destructured by-move). A second record call on the same op returns GpuError::Unrecoverable.

Implementations

impl SgemmOp

pub fn new( a: GpuRef<f32>, b: GpuRef<f32>, c: GpuRef<f32>, m: i32, n: i32, k: i32, alpha: f32, beta: f32, ) -> Self

Trait Implementations

impl GraphOp for SgemmOp

fn record(&mut self, ctx: &mut GraphRecordCtx<'_>) -> Result<(), GpuError>

Record this op into the captured stream. Called once per op during graph build; the resulting CUDA graph is then instantiated and replayed.

fn op_name(&self) -> &'static str

Display name for telemetry / error messages. Defaults to a generic label so trivial impls don’t need to override.