can_run.hh

Go to the documentation of this file.

#pragma once
 
/// @file
/// @brief geqrf's can_run, a header so host tests can pass a synthetic select::Device. No real test
/// device lacks a GPU, sub-group 32 or one CTA element of SLM. @ingroup api_selection_ops
 
#include <batchlas/blas/matrix.hh>
 
#include <src/ops/geqrf/choice.hh>
#include <src/select/select.hh>
#include <src/ops/geqrf/geqrf_native.hh>
 
#include <cstddef>
#include <cstdint>
#include <variant>
 
namespace batchlas::ops::geqrf {
 
using select::overloaded;
 
template <class T>
using MV = MatrixView<T, MatrixFormat::Dense>;
 
/// 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.
template <class T>
bool can_run(const GeqrfChoice& c, const select::Device& d, const MV<T>& A) {
    const std::int64_t m = A.rows(), n = A.cols();
    const bool native = d.is_gpu && d.has_sg32 && !A.is_heterogeneous() && m >= n && n >= 1 && A.batch_size() >= 1;
    const auto budget = static_cast<std::size_t>(d.slm_budget);
    return std::visit(overloaded{
        [&](Tiny) { return native && m == n && n <= sycl_geqrf::geqrf_tiny_max_n_for_slm<T>(budget); },
        [&](Cta) { return native && sycl_geqrf::geqrf_cta_fits<T>(static_cast<int>(m), static_cast<int>(n), budget); },
        // Blocked's panel leaf IS Cta's device function: it needs the tier present, not the fit.
        [&](Blocked) {
            return native && sycl_geqrf::geqrf_blocked_available<T>() &&
                   sycl_geqrf::geqrf_cta_max_elems_for_slm<T>(budget) >= 1;
        },
        [&](Vendor) { return d.has_vendor && !A.is_heterogeneous(); },
    }, c);
}
 
}  // namespace batchlas::ops::geqrf