conv_2d_same_multi_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"

//Get the implementations
#include "etl/impl/conv.hpp"

namespace etl {

template <etl_expr A, etl_expr B, bool Flipped>
struct conv_2d_same_multi_expr : base_temporary_expr_bin<conv_2d_same_multi_expr<A, B, Flipped>, A, B> {
    using value_type  = value_t<A>;                               
    using this_type   = conv_2d_same_multi_expr<A, B, Flipped>;   
    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 = false;

    explicit conv_2d_same_multi_expr(A a, B b) : base_type(a, b) {
        //Nothing else to init
    }

    // Assignment functions

    template <etl_2d I, etl_3d K, etl_3d C>
    static void check([[maybe_unused]] const I& input, [[maybe_unused]] const K& kernel, [[maybe_unused]] const C& conv) {
        if constexpr (all_fast<A, B, C>) {
            static_assert(etl::dim<0, C>() == etl::dim<0, K>(), "Invalid dimensions for conv2_same_multi");
            static_assert(etl::dim<1, C>() == etl::dim<0, I>(), "Invalid dimensions for conv2_same_multi");
            static_assert(etl::dim<2, C>() == etl::dim<1, I>(), "Invalid dimensions for conv2_same_multi");
        } else {
            cpp_assert(etl::dim(conv, 0) == etl::dim(kernel, 0), "Invalid dimensions for conv2_same_multi");
            cpp_assert(etl::dim(conv, 1) == etl::dim(input, 1), "Invalid dimensions for conv2_same_multi");
            cpp_assert(etl::dim(conv, 2) == etl::dim(input, 2), "Invalid dimensions for conv2_same_multi");
        }
    }

    template <typename C>
    static constexpr etl::conv_multi_impl select_default_impl() {
        constexpr order input_order  = decay_traits<A>::storage_order;
        constexpr order kernel_order = decay_traits<B>::storage_order;
        constexpr order output_order = decay_traits<C>::storage_order;

        //Only the standard implementation is able to handle column major
        if (input_order == order::ColumnMajor || kernel_order == order::ColumnMajor || output_order == order::ColumnMajor) {
            return etl::conv_multi_impl::STD;
        }

        if (impl::vec::conv2_possible<vector_mode, A, B, C>) {
            return etl::conv_multi_impl::VEC;
        }

        return etl::conv_multi_impl::STD;
    }

#ifdef ETL_MANUAL_SELECT

    template <typename C>
    static inline etl::conv_multi_impl select_impl() {
        auto default_impl = select_default_impl<C>();

        if (local_context().conv_multi_selector.forced) {
            auto forced = local_context().conv_multi_selector.impl;

            switch (forced) {
                //VEC cannot always be used
                case conv_multi_impl::VEC:
                    if (!impl::vec::conv2_possible<vector_mode, A, B, C>) {
                        std::cerr << "Forced selection to VEC conv2_same_multi implementation, but not possible for this expression" << std::endl;
                        return default_impl;
                    }

                    return forced;

                    // Although it may be suboptimal the forced selection can
                    // always be achieved
                default:
                    return forced;
            }
        }

        return default_impl;
    }

#else

    template <typename C>
    static constexpr etl::conv_multi_impl select_impl() {
        return select_default_impl<C>();
    }

#endif

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

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

        check(input, kernel, conv);

        constexpr_select auto impl = select_impl<C>();

        if constexpr (Flipped) {
            if
                constexpr_select(impl == etl::conv_multi_impl::VEC) {
                    inc_counter("impl:vec");
                    impl::vec::conv2_same_multi_flipped(input, kernel, conv);
                }
            else if
                constexpr_select(impl == etl::conv_multi_impl::STD) {
                    inc_counter("impl:std");
                    impl::standard::conv2_same_multi_flipped(input, kernel, conv);
                }
            else {
                cpp_unreachable("Invalid conv implementation selection");
            }
        } else {
            if
                constexpr_select(impl == etl::conv_multi_impl::VEC) {
                    inc_counter("impl:vec");
                    impl::vec::conv2_same_multi(input, kernel, conv);
                }
            else if
                constexpr_select(impl == etl::conv_multi_impl::STD) {
                    inc_counter("impl:std");
                    impl::standard::conv2_same_multi(input, kernel, conv);
                }
            else {
                cpp_unreachable("Invalid conv implementation 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 conv_2d_same_multi_expr& expr) {
        return os << "conv2_same_multi(" << expr._a << ", " << expr._b << ")";
    }
};

template <typename A, typename B, bool Flipped>
struct etl_traits<etl::conv_2d_same_multi_expr<A, B, Flipped>> {
    using expr_t       = etl::conv_2d_same_multi_expr<A, B, Flipped>; 
    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        = all_fast<A, B>;                       
    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 bool gpu_computable = is_gpu_t<value_type> && cuda_enabled; 
    static constexpr order storage_order = left_traits::storage_order;           

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

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

    static size_t dim(const expr_t& e, size_t d) {
        if (d == 0) {
            return etl::dim(e._b, 0);
        } else if (d == 1) {
            return etl::dim(e._a, 0);
        } else if (d == 2) {
            return etl::dim(e._a, 1);
        }
    }

    static size_t size(const expr_t& e) {
        return (etl::dim(e._b, 0)) * (etl::dim(e._a, 0)) * (etl::dim(e._a, 1));
    }

    static constexpr size_t size() {
        return (etl::dim<0, B>()) * (etl::dim<0, A>()) * (etl::dim<1, A>());
    }

    static constexpr size_t dimensions() {
        return 3;
    }

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

template <etl_expr A, etl_expr B>
conv_2d_same_multi_expr<detail::build_type<A>, detail::build_type<B>, false> conv_2d_same_multi(A&& a, B&& b) {
    return conv_2d_same_multi_expr<detail::build_type<A>, detail::build_type<B>, false>{a, b};
}

template <etl_expr A, etl_expr B, etl_expr C>
auto conv_2d_same_multi(A&& a, B&& b, C&& c) {
    c = conv_2d_same_multi(a, b);

    return c;
}

template <etl_expr A, etl_expr B>
conv_2d_same_multi_expr<detail::build_type<A>, detail::build_type<B>, true> conv_2d_same_multi_flipped(A&& a, B&& b) {
    return conv_2d_same_multi_expr<detail::build_type<A>, detail::build_type<B>, true>{a, b};
}

template <etl_expr A, etl_expr B, etl_expr C>
auto conv_2d_same_multi_flipped(A&& a, B&& b, C&& c) {
    c = conv_2d_same_multi_flipped(a, b);

    return c;
}

} //end of namespace etl