conv_4d_valid_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, size_t S1, size_t S2, size_t P1, size_t P2, bool Flipped>
struct conv_4d_valid_expr : base_temporary_expr_bin<conv_4d_valid_expr<A, B, S1, S2, P1, P2, Flipped>, A, B> {
    using value_type  = value_t<A>;                                        
    using this_type   = conv_4d_valid_expr<A, B, S1, S2, P1, P2, 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 = cudnn_enabled && impl::cudnn::conv_possible_<A, B>;

    explicit conv_4d_valid_expr(A a, B b) : base_type(a, b) {
        //Nothing else to init
    }

    // Assignment functions

    template <etl_4d I, etl_4d K, etl_4d 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, I>(), "Invalid dimensions for conv4_valid");
            static_assert(etl::dim<1, C>() == etl::dim<0, K>(), "Invalid dimensions for conv4_valid");
            static_assert(etl::dim<1, I>() == etl::dim<1, K>(), "Invalid dimensions for conv4_valid");

            static_assert(etl::dim<2, C>() == (etl::dim<2, I>() - etl::dim<2, K>() + 2 * P1) / S1 + 1, "Invalid dimensions for conv4_valid");
            static_assert(etl::dim<3, C>() == (etl::dim<3, I>() - etl::dim<3, K>() + 2 * P2) / S2 + 1, "Invalid dimensions for conv4_valid");
        } else {
            cpp_assert(etl::dim(conv, 0) == etl::dim(input, 0), "Invalid dimensions for conv4_valid");
            cpp_assert(etl::dim(conv, 1) == etl::dim(kernel, 0), "Invalid dimensions for conv4_valid");
            cpp_assert(etl::dim(input, 1) == etl::dim(kernel, 1), "Invalid dimensions for conv4_valid");

            cpp_assert(etl::dim(conv, 2) == (etl::dim(input, 2) - etl::dim(kernel, 2) + 2 * P1) / S1 + 1, "Invalid dimensions for conv4_valid");
            cpp_assert(etl::dim(conv, 3) == (etl::dim(input, 3) - etl::dim(kernel, 3) + 2 * P2) / S2 + 1, "Invalid dimensions for conv4_valid");
        }
    }

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

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

        check(a, b, c);

        if constexpr (Flipped) {
            detail::conv4_valid_flipped_impl<S1, S2, P1, P2>::apply(a, b, c);
        } else {
            detail::conv4_valid_impl<S1, S2, P1, P2>::apply(a, b, c);
        }
    }

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

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

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

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

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

    friend std::ostream& operator<<(std::ostream& os, const conv_4d_valid_expr& expr) {
        return os << "conv4_valid(" << expr._a << ", " << expr._b << ")";
    }
};

template <etl_expr A, etl_expr B, size_t S1, size_t S2, size_t P1, size_t P2, bool Flipped>
struct etl_traits<etl::conv_4d_valid_expr<A, B, S1, S2, P1, P2, Flipped>> {
    using expr_t       = etl::conv_4d_valid_expr<A, B, S1, S2, P1, P2, 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 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 ? etl::dim<0, A>()
                       : DD == 1 ? etl::dim<0, B>()
                                 : DD == 2 ? (etl::dim<2, A>() - etl::dim<2, B>() + 2 * P1) / S1 + 1 : (etl::dim<3, A>() - etl::dim<3, B>() + 2 * P2) / S2 + 1;
    }

    static size_t dim(const expr_t& e, size_t d) {
        if (d == 0) {
            return etl::dim(e._a, 0);
        } else if (d == 1) {
            return etl::dim(e._b, 0);
        } else if (d == 2) {
            return (etl::dim(e._a, 2) - etl::dim(e._b, 2) + 2 * P1) / S1 + 1;
        } else {
            return (etl::dim(e._a, 3) - etl::dim(e._b, 3) + 2 * P2) / S2 + 1;
        }
    }

    static size_t size(const expr_t& e) {
        return etl::dim(e._a, 0) * etl::dim(e._b, 0) * ((etl::dim(e._a, 2) - etl::dim(e._b, 2) + 2 * P1) / S1 + 1)
               * ((etl::dim(e._a, 3) - etl::dim(e._b, 3) + 2 * P2) / S2 + 1);
    }

    static constexpr size_t size() {
        return etl::dim<0, A>() * etl::dim<0, B>() * ((etl::dim<2, A>() - etl::dim<2, B>() + 2 * P1) / S1 + 1)
               * ((etl::dim<3, A>() - etl::dim<3, B>() + 2 * P2) / S2 + 1);
    }

    static constexpr size_t dimensions() {
        return 4;
    }

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

template <size_t S1 = 1, size_t S2 = 1, size_t P1 = 0, size_t P2 = 0, etl_expr A, etl_expr B>
conv_4d_valid_expr<detail::build_type<A>, detail::build_type<B>, S1, S2, P1, P2, false> conv_4d_valid(A&& a, B&& b) {
    return conv_4d_valid_expr<detail::build_type<A>, detail::build_type<B>, S1, S2, P1, P2, false>{a, b};
}

template <size_t S1 = 1, size_t S2 = 1, size_t P1 = 0, size_t P2 = 0, etl_expr A, etl_expr B, etl_expr C>
auto conv_4d_valid(A&& a, B&& b, C&& c) {
    c = conv_4d_valid<S1, S2, P1, P2>(a, b);

    return c;
}

template <size_t S1 = 1, size_t S2 = 1, size_t P1 = 0, size_t P2 = 0, etl_expr A, etl_expr B>
conv_4d_valid_expr<detail::build_type<A>, detail::build_type<B>, S1, S2, P1, P2, true> conv_4d_valid_flipped(A&& a, B&& b) {
    return conv_4d_valid_expr<detail::build_type<A>, detail::build_type<B>, S1, S2, P1, P2, true>{a, b};
}

template <size_t S1 = 1, size_t S2 = 1, size_t P1 = 0, size_t P2 = 0, etl_expr A, etl_expr B, etl_expr C>
auto conv_4d_valid_flipped(A&& a, B&& b, C&& c) {
    c = conv_4d_valid_flipped<S1, S2, P1, P2>(a, b);

    return c;
}

} //end of namespace etl