geqrf_batch.hpp

Go to the documentation of this file.

// Copyright (c) 2021--present, The Simons Foundation
// This file is part of TRIQS/nda and is licensed under the Apache License, Version 2.0.
// SPDX-License-Identifier: Apache-2.0
// See LICENSE in the root of this distribution for details.

 
#pragma once
 
#include "./geqrf.hpp"
#include "./interface/cxx_interface.hpp"
#include "../basic_array.hpp"
#include "../basic_functions.hpp"
#include "../blas/interface/cxx_interface.hpp"
#include "../blas/tools.hpp"
#include "../concepts.hpp"
#include "../declarations.hpp"
#include "../device.hpp"
#include "../layout_transforms.hpp"
#include "../macros.hpp"
#include "../mem/address_space.hpp"
#include "../traits.hpp"
 
#include <algorithm>
#include <type_traits>
#include <utility>
 
namespace nda::lapack {

  template <BlasArray<3> A, BlasArrayFor<A, 2> TAU, BlasArrayFor<A, 1> W = vector_value_t<A>>
    requires(has_F_layout<A, TAU>)
  int geqrf_batch(A &&a, TAU &&tau, [[maybe_unused]] W &&work = vector_value_t<A>{}) { // NOLINT (temporary views are allowed here)
    // check the dimensions of the input/output arrays/views and resize if necessary
    auto const [m, n, n_b] = a.shape();
    resize_or_check_if_view(tau, {std::min(m, n), n_b});
 
    // arrays/views must be LAPACK compatible
    EXPECTS(a.indexmap().min_stride() == 1);
    EXPECTS(tau.indexmap().min_stride() == 1);
 
#if defined(__has_feature)
#if __has_feature(memory_sanitizer)
    tau = get_value_t<A>{0};
#endif
#endif
 
    // perform actual library call(s)
    int info = 0;
    if constexpr (mem::have_device_compatible_addr_space<A, TAU, W>) {
      // get pointers to each matrix/vector in the batch
      auto a_ptrs   = to_device(batch_ptrs(a));
      auto tau_ptrs = to_device(batch_ptrs(tau));
 
      blas::device::geqrf_batch(m, n, a_ptrs.data(), get_ld(a(range::all, range::all, 0)), tau_ptrs.data(), info, n_b);
    } else {
      // for host, fall back to looping over batches
      for (int i = 0; i < n_b; ++i) {
        auto a_i       = a(range::all, range::all, i);
        auto tau_i     = tau(range::all, i);
        int local_info = geqrf(a_i, tau_i, work);
        if (local_info != 0 && info == 0) info = local_info;
      }
    }
    return info;
  }

  template <BlasArray<3> A, BlasArrayFor<A, 2> TAU, BlasArrayFor<A, 1> W = vector_value_t<A>>
    requires(has_F_layout<A, TAU>)
  int geqrf(A &&a, TAU &&tau, W &&work = vector_value_t<A>{}) { // NOLINT (temporary views are allowed here)
    return geqrf_batch(std::forward<A>(a), std::forward<TAU>(tau), std::forward<W>(work));
  }
 
} // namespace nda::lapack