chromium/third_party/eigen3/src/Eigen/src/Core/products/GeneralMatrixVector.h

// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2008-2016 Gael Guennebaud <[email protected]>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.

#ifndef EIGEN_GENERAL_MATRIX_VECTOR_H
#define EIGEN_GENERAL_MATRIX_VECTOR_H

// IWYU pragma: private
#include "../InternalHeaderCheck.h"

namespace Eigen {

namespace internal {

enum GEMVPacketSizeType { … };

template <int N, typename T1, typename T2, typename T3>
struct gemv_packet_cond { … };

gemv_packet_cond<GEMVPacketFull, T1, T2, T3>;

gemv_packet_cond<GEMVPacketHalf, T1, T2, T3>;

template <typename LhsScalar, typename RhsScalar, int PacketSize_ = GEMVPacketFull>
class gemv_traits { … };

/* Optimized col-major matrix * vector product:
 * This algorithm processes the matrix per vertical panels,
 * which are then processed horizontally per chunck of 8*PacketSize x 1 vertical segments.
 *
 * Mixing type logic: C += alpha * A * B
 *  |  A  |  B  |alpha| comments
 *  |real |cplx |cplx | no vectorization
 *  |real |cplx |real | alpha is converted to a cplx when calling the run function, no vectorization
 *  |cplx |real |cplx | invalid, the caller has to do tmp: = A * B; C += alpha*tmp
 *  |cplx |real |real | optimal case, vectorization possible via real-cplx mul
 *
 * The same reasoning apply for the transposed case.
 */
general_matrix_vector_product<Index, LhsScalar, LhsMapper, ColMajor, ConjugateLhs, RhsScalar, RhsMapper, ConjugateRhs, Version>;

template <typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar,
          typename RhsMapper, bool ConjugateRhs, int Version>
EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void
general_matrix_vector_product<Index, LhsScalar, LhsMapper, ColMajor, ConjugateLhs, RhsScalar, RhsMapper, ConjugateRhs,
                              Version>::run(Index rows, Index cols, const LhsMapper& alhs, const RhsMapper& rhs,
                                            ResScalar* res, Index resIncr, RhsScalar alpha) { … }

/* Optimized row-major matrix * vector product:
 * This algorithm processes 4 rows at once that allows to both reduce
 * the number of load/stores of the result by a factor 4 and to reduce
 * the instruction dependency. Moreover, we know that all bands have the
 * same alignment pattern.
 *
 * Mixing type logic:
 *  - alpha is always a complex (or converted to a complex)
 *  - no vectorization
 */
general_matrix_vector_product<Index, LhsScalar, LhsMapper, RowMajor, ConjugateLhs, RhsScalar, RhsMapper, ConjugateRhs, Version>;

template <typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar,
          typename RhsMapper, bool ConjugateRhs, int Version>
EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void
general_matrix_vector_product<Index, LhsScalar, LhsMapper, RowMajor, ConjugateLhs, RhsScalar, RhsMapper, ConjugateRhs,
                              Version>::run(Index rows, Index cols, const LhsMapper& alhs, const RhsMapper& rhs,
                                            ResScalar* res, Index resIncr, ResScalar alpha) { … }

}  // end namespace internal

}  // end namespace Eigen

#endif  // EIGEN_GENERAL_MATRIX_VECTOR_H