mul.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

namespace etl {

template <typename T>
struct mul_binary_op;

// detect 1.0 * (x * y)

template <typename L, typename R>
struct is_axmy_left_impl {
    static constexpr bool value = false;
};

template <typename T0, typename T1, typename T2, typename LeftExpr, typename RightExpr>
struct is_axmy_left_impl<etl::scalar<T0>, binary_expr<T1, LeftExpr, etl::mul_binary_op<T2>, RightExpr>> {
    static constexpr bool value = true;
};

// detect (x * y) * 1.0

template <typename L, typename R>
struct is_axmy_right_impl {
    static constexpr bool value = false;
};

template <typename T0, typename T1, typename T2, typename LeftExpr, typename RightExpr>
struct is_axmy_right_impl<binary_expr<T1, LeftExpr, etl::mul_binary_op<T2>, RightExpr>, etl::scalar<T0>> {
    static constexpr bool value = true;
};

// detect (1.0 * x) * y

template <typename L, typename R>
struct is_axmy_left_left_impl {
    static constexpr bool value = false;
};

template <typename T0, typename T1, typename T2, typename RightExpr, typename R>
struct is_axmy_left_left_impl<binary_expr<T1, etl::scalar<T0>, etl::mul_binary_op<T2>, RightExpr>, R> {
    static constexpr bool value = true;
};

// detect (x * 1.0) * y

template <typename L, typename R>
struct is_axmy_left_right_impl {
    static constexpr bool value = false;
};

template <typename T0, typename T1, typename T2, typename LeftExpr, typename R>
struct is_axmy_left_right_impl<binary_expr<T1, LeftExpr, etl::mul_binary_op<T2>, etl::scalar<T0>>, R> {
    static constexpr bool value = !is_scalar<R>;
};

// detect x * (1.0 * y)

template <typename L, typename R>
struct is_axmy_right_left_impl {
    static constexpr bool value = false;
};

template <typename L, typename T0, typename T1, typename T2, typename RightExpr>
struct is_axmy_right_left_impl<L, binary_expr<T0, etl::scalar<T1>, etl::mul_binary_op<T2>, RightExpr>> {
    static constexpr bool value = !is_scalar<L>;
};

// detect x * (1.0 * y)

template <typename L, typename R>
struct is_axmy_right_right_impl {
    static constexpr bool value = false;
};

template <typename L, typename T0, typename T1, typename T2, typename RightExpr>
struct is_axmy_right_right_impl<L, binary_expr<T0, RightExpr, etl::mul_binary_op<T2>, etl::scalar<T1>>> {
    static constexpr bool value = true;
};

// Variable templates helper

template <typename L, typename R>
static constexpr bool is_axmy_left = is_axmy_left_impl<L, R>::value;

template <typename L, typename R>
static constexpr bool is_axmy_right = is_axmy_right_impl<L, R>::value;

template <typename L, typename R>
static constexpr bool is_axmy_left_left = is_axmy_left_left_impl<L, R>::value;

template <typename L, typename R>
static constexpr bool is_axmy_left_right = is_axmy_left_right_impl<L, R>::value;

template <typename L, typename R>
static constexpr bool is_axmy_right_left = is_axmy_right_left_impl<L, R>::value;

template <typename L, typename R>
static constexpr bool is_axmy_right_right = is_axmy_right_right_impl<L, R>::value;

template <typename L, typename R>
static constexpr bool is_axmy =
    is_axmy_left<L, R> || is_axmy_right<L, R> || is_axmy_left_left<L, R> || is_axmy_left_right<L, R> || is_axmy_right_left<L, R> || is_axmy_right_right<L, R>;

template <typename T>
struct mul_binary_op {
    static constexpr bool linear      = true;  
    static constexpr bool thread_safe = true;  
    static constexpr bool desc_func   = false; 

    template <vector_mode_t V>
    static constexpr bool vectorizable = V == vector_mode_t::AVX512 ? !is_complex_t<T> : true;

    template <typename L, typename R>
    static constexpr bool gpu_computable = ((!is_scalar<L> && !is_scalar<R>)&&((is_single_precision_t<T> && impl::egblas::has_saxmy_3)
                                                                               || (is_double_precision_t<T> && impl::egblas::has_daxmy_3)
                                                                               || (is_complex_single_t<T> && impl::egblas::has_caxmy_3)
                                                                               || (is_complex_double_t<T> && impl::egblas::has_zaxmy_3)))
                                           || ((is_scalar<L> != is_scalar<R>)&&((is_single_precision_t<T> && impl::egblas::has_scalar_smul)
                                                                                || (is_double_precision_t<T> && impl::egblas::has_scalar_dmul)
                                                                                || (is_complex_single_t<T> && impl::egblas::has_scalar_cmul)
                                                                                || (is_complex_double_t<T> && impl::egblas::has_scalar_zmul)));

    static constexpr int complexity() {
        return 2;
    }

    template <typename V = default_vec>
    using vec_type = typename V::template vec_type<T>;

    static constexpr T apply(const T& lhs, const T& rhs) noexcept {
        return lhs * rhs;
    }

    template <typename V = default_vec>
    static vec_type<V> load(const vec_type<V>& lhs, const vec_type<V>& rhs) noexcept {
        return V::mul(lhs, rhs);
    }

    template <typename L, typename R, typename Y>
    static auto gpu_compute_hint(const L& lhs, const R& rhs, Y& y) noexcept {
        auto t3 = force_temporary_gpu_dim_only(y);
        gpu_compute(lhs, rhs, t3);
        return t3;
    }

    template <typename L, typename R, typename YY>
    static YY& gpu_compute(const L& lhs, const R& rhs, YY& yy) noexcept {
        if constexpr (is_axmy_left<L, R>) {
            auto& rhs_lhs = rhs.get_lhs();
            auto& rhs_rhs = rhs.get_rhs();

            decltype(auto) x = smart_gpu_compute_hint(rhs_lhs, yy);
            decltype(auto) y = smart_gpu_compute_hint(rhs_rhs, yy);

            constexpr auto incx = gpu_inc<decltype(rhs_lhs)>;
            constexpr auto incy = gpu_inc<decltype(rhs_rhs)>;

            impl::egblas::axmy_3(etl::size(yy), lhs.value, x.gpu_memory(), incx, y.gpu_memory(), incy, yy.gpu_memory(), 1);
        } else if constexpr (is_axmy_right<L, R>) {
            auto& lhs_lhs = lhs.get_lhs();
            auto& lhs_rhs = lhs.get_rhs();

            decltype(auto) x = smart_gpu_compute_hint(lhs_lhs, yy);
            decltype(auto) y = smart_gpu_compute_hint(lhs_rhs, yy);

            constexpr auto incx = gpu_inc<decltype(lhs_lhs)>;
            constexpr auto incy = gpu_inc<decltype(lhs_rhs)>;

            impl::egblas::axmy_3(etl::size(yy), rhs.value, x.gpu_memory(), incx, y.gpu_memory(), incy, yy.gpu_memory(), 1);
        } else if constexpr (is_axmy_left_left<L, R>) {
            auto& lhs_lhs = lhs.get_lhs();
            auto& lhs_rhs = lhs.get_rhs();

            decltype(auto) x = smart_gpu_compute_hint(lhs_rhs, yy);
            decltype(auto) y = smart_gpu_compute_hint(rhs, yy);

            constexpr auto incx = gpu_inc<decltype(lhs_rhs)>;
            constexpr auto incy = gpu_inc<decltype(rhs)>;

            impl::egblas::axmy_3(etl::size(yy), lhs_lhs.value, x.gpu_memory(), incx, y.gpu_memory(), incy, yy.gpu_memory(), 1);
        } else if constexpr (is_axmy_left_right<L, R>) {
            auto& lhs_lhs = lhs.get_lhs();
            auto& lhs_rhs = lhs.get_rhs();

            decltype(auto) x = smart_gpu_compute_hint(lhs_lhs, yy);
            decltype(auto) y = smart_gpu_compute_hint(rhs, yy);

            constexpr auto incx = gpu_inc<decltype(lhs_lhs)>;
            constexpr auto incy = gpu_inc<decltype(rhs)>;

            impl::egblas::axmy_3(etl::size(yy), lhs_rhs.value, x.gpu_memory(), incx, y.gpu_memory(), incy, yy.gpu_memory(), 1);
        } else if constexpr (is_axmy_right_left<L, R>) {
            auto& rhs_lhs = rhs.get_lhs();
            auto& rhs_rhs = rhs.get_rhs();

            decltype(auto) x = smart_gpu_compute_hint(lhs, yy);
            decltype(auto) y = smart_gpu_compute_hint(rhs_rhs, yy);

            constexpr auto incx = gpu_inc<decltype(lhs)>;
            constexpr auto incy = gpu_inc<decltype(rhs_rhs)>;

            impl::egblas::axmy_3(etl::size(yy), rhs_lhs.value, x.gpu_memory(), incx, y.gpu_memory(), incy, yy.gpu_memory(), 1);
        } else if constexpr (is_axmy_right_right<L, R>) {
            auto& rhs_lhs = rhs.get_lhs();
            auto& rhs_rhs = rhs.get_rhs();

            decltype(auto) x = smart_gpu_compute_hint(lhs, yy);
            decltype(auto) y = smart_gpu_compute_hint(rhs_lhs, yy);

            constexpr auto incx = gpu_inc<decltype(lhs)>;
            constexpr auto incy = gpu_inc<decltype(rhs_lhs)>;

            impl::egblas::axmy_3(etl::size(yy), rhs_rhs.value, x.gpu_memory(), incx, y.gpu_memory(), incy, yy.gpu_memory(), 1);
        } else if constexpr (!is_scalar<L> && !is_scalar<R> && !is_axmy<L, R>) {
            decltype(auto) x = smart_gpu_compute_hint(lhs, yy);
            decltype(auto) y = smart_gpu_compute_hint(rhs, yy);

            constexpr auto incx = gpu_inc<decltype(lhs)>;
            constexpr auto incy = gpu_inc<decltype(rhs)>;

            impl::egblas::axmy_3(etl::size(yy), 1, x.gpu_memory(), incx, y.gpu_memory(), incy, yy.gpu_memory(), 1);
        } else if constexpr (is_scalar<L> && !is_scalar<R> && !is_axmy<L, R>) {
            smart_gpu_compute(rhs, yy);

            impl::egblas::scalar_mul(yy.gpu_memory(), etl::size(yy), 1, lhs.value);
        } else if constexpr (!is_scalar<L> && is_scalar<R> && !is_axmy<L, R>) {
            smart_gpu_compute(lhs, yy);

            impl::egblas::scalar_mul(yy.gpu_memory(), etl::size(yy), 1, rhs.value);
        }

        yy.validate_gpu();
        yy.invalidate_cpu();

        return yy;
    }

    static std::string desc() noexcept {
        return "*";
    }
};

} //end of namespace etl