Struct GemmRequest 

pub struct GemmRequest<T: GemmSupported> {

    pub a: GpuRef<T>,
    pub b: GpuRef<T>,
    pub c: GpuRef<T>,
    pub m: i32,
    pub n: i32,
    pub k: i32,
    pub alpha: T,
    pub beta: T,
    pub trans_a: cublasOperation_t,
    pub trans_b: cublasOperation_t,
    pub lda: i32,
    pub ldb: i32,
    pub ldc: i32,
    pub reply: Sender<Result<(), GpuError>>,
}

Typed cuBLAS gemm request: C = α·op(A)·op(B) + β·C.

lda/ldb/ldc follow cuBLAS’s column-major convention (see cuBLAS docs). For the no-transpose case, lda = m, ldb = k, ldc = m.

Capability marker compile-fail

T: GemmSupported gates the dtype matrix. cuBLAS does not support i64 gemm, so building a GemmRequest::<i64> is rejected at compile time:

// Fails: i64 does not implement `GemmSupported`.
let _req = GemmRequest::<i64> {
    a, b, c,
    m: 1, n: 1, k: 1,
    alpha: 1, beta: 0,
    trans_a: cublasOperation_t::CUBLAS_OP_N,
    trans_b: cublasOperation_t::CUBLAS_OP_N,
    lda: 1, ldb: 1, ldc: 1,
    reply: tx,
};

Fields

a: GpuRef<T>

b: GpuRef<T>

c: GpuRef<T>

m: i32

n: i32

k: i32

alpha: T

beta: T

trans_a: cublasOperation_t

trans_b: cublasOperation_t

lda: i32

ldb: i32

ldc: i32

reply: Sender<Result<(), GpuError>>

Implementations

impl<T> GemmRequest<T>

where T: GemmSupported, GemmRequest<T>: GemmDispatch,

pub fn into_msg(self) -> BlasMsg

Box-and-wrap into a crate::kernel::BlasMsg::Gemm variant.

Trait Implementations

impl GemmDispatch for GemmRequest<f32>

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

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

fn dispatch(self: Box<Self>, ctx: &BlasDispatchCtx<'_>)

impl GemmDispatch for GemmRequest<f64>

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

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

fn dispatch(self: Box<Self>, ctx: &BlasDispatchCtx<'_>)