| bool batchlas::ops::geqrf::can_run | ( | const GeqrfChoice & | c, |
| const select::Device & | d, | ||
| const MV< T > & | A ) |
Correctness only (R3): each clause is its driver's check.
One (m, n, ld, stride) per launch and a wide view's off-panel update: natives take neither; the vendor (top-level (m, n) per item) takes no heterogeneous batch.
Definition at line 27 of file can_run.hh.
|
constexpr |
|
constexpr |
| using batchlas::ops::geqrf::MV = MatrixView<T, MatrixFormat::Dense> |
Definition at line 22 of file can_run.hh.
| using batchlas::ops::geqrf::GeqrfChoice = std::variant<Tiny, Cta, Blocked, Vendor> |
|
inlineconstexpr |
|
inlineconstexpr |
|
inlineconstexpr |
Rows: sq x grid_n, tall x grid_n x grid_aspect, wide x grid_wide_n x grid_wide_aspect; both sides of every transcribed threshold (the transcriber, tuned/README.md, spells the same grid).
|
inlineconstexpr |
|
inlineconstexpr |
|
inlineconstexpr |
Classes | |
| struct | Blocked |
| geqrf_blocked_dispatch + the public gemm, any m >= n More... | |
| struct | Cta |
| geqrf_cta_dispatch: the whole panel in SLM (geqrf_cta_fits) More... | |
| struct | overloaded |
| std::visit over a choice with one lambda per family. More... | |
| struct | Tiny |
| geqrf_tiny_dispatch: square, registers, n <= the tiny ceiling More... | |
| struct | Vendor |
| backend::geqrf_vendor; needs the factorization library More... | |