|
Uni20
Asynchronous tensor and runtime library.
|
Functions | |
| template<uni20::cublas::CublasScalar Scalar, class OutputHandle , class LhsHandle , class RhsHandle > requires std::convertible_to<OutputHandle, Scalar*> && std::convertible_to<LhsHandle, Scalar const*> && std::convertible_to<RhsHandle, Scalar const*> | |
| void | gemm (uni20::cublas::ExecutionLease &execution, blas::BlasWritableMatrix< Scalar, OutputHandle > output, Scalar alpha, blas::BlasReadableMatrix< Scalar, LhsHandle > lhs, blas::BlasReadableMatrix< Scalar, RhsHandle > rhs, Scalar beta) |
| Run provider-ready column-major GEMM through cuBLAS. | |
| void uni20::linalg::cublas::gemm | ( | uni20::cublas::ExecutionLease & | execution, |
| blas::BlasWritableMatrix< Scalar, OutputHandle > | output, | ||
| Scalar | alpha, | ||
| blas::BlasReadableMatrix< Scalar, LhsHandle > | lhs, | ||
| blas::BlasReadableMatrix< Scalar, RhsHandle > | rhs, | ||
| Scalar | beta | ||
| ) |
Run provider-ready column-major GEMM through cuBLAS.
Backend acceptance must be completed before acquiring execution. Unsupported transforms and invalid dimensions are logic errors here.