batch_outer_product_expr.hpp

//=======================================================================
// Copyright (c) 2014-2023 Baptiste Wicht
// Distributed under the terms of the MIT License.
// (See accompanying file LICENSE or copy at
//  http://opensource.org/licenses/MIT)
//=======================================================================

#pragma once

#include "etl/expr/base_temporary_expr.hpp"

//Include the implementations
#include "etl/impl/std/outer.hpp"
#include "etl/impl/blas/outer.hpp"
#include "etl/impl/cublas/outer.hpp"
#include "etl/impl/vec/outer.hpp"

namespace etl {

template <etl_expr A, etl_expr B>
struct batch_outer_product_expr : base_temporary_expr_bin<batch_outer_product_expr<A, B>, A, B> {
    using value_type  = value_t<A>;                               
    using this_type   = batch_outer_product_expr<A, B>;           
    using base_type   = base_temporary_expr_bin<this_type, A, B>; 
    using left_traits = decay_traits<A>;                          

    static constexpr auto storage_order = left_traits::storage_order; 

    static constexpr bool gpu_computable = cublas_enabled;

    explicit batch_outer_product_expr(A a, B b) : base_type(a, b) {
        //Nothing else to init
    }

    // Assignment functions

    template <typename C>
    static constexpr etl::outer_impl select_default_batch_outer_impl(bool no_gpu) {
        if (cublas_enabled && !no_gpu) {
            return etl::outer_impl::CUBLAS;
        }

        if (cblas_enabled) {
            return etl::outer_impl::BLAS;
        }

        if (vec_enabled) {
            return etl::outer_impl::VEC;
        }

        return etl::outer_impl::STD;
    }

#ifdef ETL_MANUAL_SELECT

    template <typename C>
    static etl::outer_impl select_batch_outer_impl() {
        if (local_context().outer_selector.forced) {
            auto forced = local_context().outer_selector.impl;

            switch (forced) {
                //BLAS cannot always be used
                case outer_impl::BLAS:
                    if (!cblas_enabled) {
                        std::cerr << "Forced selection to BLAS outer implementation, but not possible for this expression" << std::endl;
                        return select_default_batch_outer_impl<C>(local_context().cpu);
                    }

                    return forced;

                //CUBLAS cannot always be used
                case outer_impl::CUBLAS:
                    if (!cublas_enabled || local_context().cpu) {
                        std::cerr << "Forced selection to CUBLAS outer implementation, but not possible for this expression" << std::endl;
                        return select_default_batch_outer_impl<C>(local_context().cpu);
                    }

                    return forced;

                //VEC cannot always be used
                case outer_impl::VEC:
                    if (!vec_enabled) {
                        std::cerr << "Forced selection to VEC outer implementation, but not possible for this expression" << std::endl;
                        return select_default_batch_outer_impl<C>(local_context().cpu);
                    }

                    return forced;

                //In other cases, simply use the forced impl
                default:
                    return forced;
            }
        }

        return select_default_batch_outer_impl<C>(local_context().cpu);
    }

#else

    template <typename C>
    static constexpr etl::outer_impl select_batch_outer_impl() {
        return select_default_batch_outer_impl<C>(false);
    }

#endif

    template <etl_expr C>
    void assign_to(C&& c) const {
        inc_counter("temp:assign");

        auto& a = this->a();
        auto& b = this->b();

        constexpr_select auto impl = select_batch_outer_impl<C>();

        if
            constexpr_select(impl == etl::outer_impl::STD) {
                inc_counter("impl:std");
                etl::impl::standard::batch_outer(smart_forward(a), smart_forward(b), c);
            }
        else if
            constexpr_select(impl == etl::outer_impl::BLAS) {
                inc_counter("impl:blas");
                etl::impl::blas::batch_outer(smart_forward(a), smart_forward(b), c);
            }
        else if
            constexpr_select(impl == etl::outer_impl::CUBLAS) {
                inc_counter("impl:cublas");
                etl::impl::cublas::batch_outer(smart_forward_gpu(a), smart_forward_gpu(b), c);
            }
        else if
            constexpr_select(impl == etl::outer_impl::VEC) {
                inc_counter("impl:vec");
                etl::impl::vec::batch_outer(smart_forward(a), smart_forward(b), c);
            }
        else {
            cpp_unreachable("Invalid batch_outer selection");
        }
    }

    template <typename L>
    void assign_add_to(L&& lhs) const {
        std_add_evaluate(*this, lhs);
    }

    template <typename L>
    void assign_sub_to(L&& lhs) const {
        std_sub_evaluate(*this, lhs);
    }

    template <typename L>
    void assign_mul_to(L&& lhs) const {
        std_mul_evaluate(*this, lhs);
    }

    template <typename L>
    void assign_div_to(L&& lhs) const {
        std_div_evaluate(*this, lhs);
    }

    template <typename L>
    void assign_mod_to(L&& lhs) const {
        std_mod_evaluate(*this, lhs);
    }

    friend std::ostream& operator<<(std::ostream& os, const batch_outer_product_expr& expr) {
        return os << "batch_outer(" << expr._a << ", " << expr._b << ")";
    }
};

template <typename A, typename B>
struct etl_traits<etl::batch_outer_product_expr<A, B>> {
    using expr_t       = etl::batch_outer_product_expr<A, B>; 
    using left_expr_t  = std::decay_t<A>;                     
    using right_expr_t = std::decay_t<B>;                     
    using left_traits  = etl_traits<left_expr_t>;             
    using right_traits = etl_traits<right_expr_t>;            
    using value_type   = value_t<A>;                          

    static constexpr bool is_etl         = true;                                          
    static constexpr bool is_transformer = false;                                         
    static constexpr bool is_view        = false;                                         
    static constexpr bool is_magic_view  = false;                                         
    static constexpr bool is_fast        = left_traits::is_fast && right_traits::is_fast; 
    static constexpr bool is_linear      = false;                                         
    static constexpr bool is_thread_safe = true;                                          
    static constexpr bool is_value       = false;                                         
    static constexpr bool is_direct      = true;                                          
    static constexpr bool is_generator   = false;                                         
    static constexpr bool is_padded      = false;                                         
    static constexpr bool is_aligned     = true;                                          
    static constexpr bool is_temporary   = true;                                          
    static constexpr order storage_order = left_traits::storage_order;                    
    static constexpr bool gpu_computable = is_gpu_t<value_type> && cuda_enabled;          

    template <vector_mode_t V>
    static constexpr bool vectorizable = true;

    template <size_t DD>
    static constexpr size_t dim() {
        return DD == 0 ? decay_traits<A>::template dim<1>() : decay_traits<B>::template dim<1>();
    }

    static size_t dim(const expr_t& e, size_t d) {
        if (d == 0) {
            return etl::dim(e._a, 1);
        } else {
            return etl::dim(e._b, 1);
        }
    }

    static size_t size(const expr_t& e) {
        return etl::dim(e._a, 1) * etl::dim(e._b, 1);
    }

    static constexpr size_t size() {
        return decay_traits<A>::template dim<1>() * decay_traits<B>::template dim<1>();
    }

    static constexpr size_t dimensions() {
        return 2;
    }

    static constexpr int complexity() noexcept {
        return -1;
    }
};

template <typename A, typename B>
batch_outer_product_expr<detail::build_type<A>, detail::build_type<B>> batch_outer(A&& a, B&& b) {
    return batch_outer_product_expr<detail::build_type<A>, detail::build_type<B>>{a, b};
}

template <typename A, typename B, typename C>
auto batch_outer(A&& a, B&& b, C&& c) {
    c = batch_outer(a, b);
    return c;
}

} //end of namespace etl