dyn_matrix_view.hpp

Go to the documentation of this file.

//=======================================================================
// 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/index.hpp"

namespace etl {

template <etl_expr T, size_t D>
requires(!is_dma<T>) struct dyn_matrix_view<T, D> final : iterable<dyn_matrix_view<T, D>, false>,
                                                          value_testable<dyn_matrix_view<T, D>>,
                                                          assignable<dyn_matrix_view<T, D>, value_t<T>> {
    using this_type            = dyn_matrix_view<T, D>;                                                
    using iterable_base_type   = iterable<this_type, false>;                                           
    using assignable_base_type = assignable<this_type, value_t<T>>;                                    
    using sub_type             = T;                                                                    
    using value_type           = value_t<sub_type>;                                                    
    using memory_type          = memory_t<sub_type>;                                                   
    using const_memory_type    = const_memory_t<sub_type>;                                             
    using return_type          = return_helper<sub_type, decltype(std::declval<sub_type>()[0])>;       
    using const_return_type    = const_return_helper<sub_type, decltype(std::declval<sub_type>()[0])>; 
    using iterator             = etl::iterator<this_type>;                                             
    using const_iterator       = etl::iterator<const this_type>;                                       

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

    using assignable_base_type::operator=;
    using iterable_base_type::begin;
    using iterable_base_type::end;

private:
    T sub;                            
    std::array<size_t, D> dimensions; 
    size_t _size;                     

    static constexpr order storage_order = decay_traits<sub_type>::storage_order; 

    friend struct etl_traits<dyn_matrix_view>;

public:
    template <typename... S>
    explicit dyn_matrix_view(sub_type sub, S... dims) : sub(sub), dimensions{{dims...}}, _size(etl::size(sub)) {}

    const_return_type operator[](size_t j) const {
        return sub[j];
    }

    return_type operator[](size_t j) {
        return sub[j];
    }

    const_return_type operator()(size_t j) const {
        return sub[j];
    }

    template <typename... S>
    const_return_type operator()(size_t f1, size_t f2, S... sizes) const {
        return sub[etl::dyn_index(*this, f1, f2, sizes...)];
    }

    value_type read_flat(size_t j) const noexcept {
        return sub.read_flat(j);
    }

    return_type operator()(size_t j) {
        return sub[j];
    }

    template <typename... S>
    return_type operator()(size_t f1, size_t f2, S... sizes) {
        return sub[etl::dyn_index(*this, f1, f2, sizes...)];
    }

    template <typename V = default_vec>
    auto load(size_t x) const noexcept {
        return sub.template load<V>(x);
    }

    template <typename V = default_vec>
    auto loadu(size_t x) const noexcept {
        return sub.template loadu<V>(x);
    }

    template <typename V = default_vec>
    void stream(vec_type<V> in, size_t i) noexcept {
        sub.template stream<V>(in, i);
    }

    template <typename V = default_vec>
    void store(vec_type<V> in, size_t i) noexcept {
        sub.template store<V>(in, i);
    }

    template <typename V = default_vec>
    void storeu(vec_type<V> in, size_t i) noexcept {
        sub.template storeu<V>(in, i);
    }

    template <typename E>
    bool alias(const E& rhs) const noexcept {
        return sub.alias(rhs);
    }

    // Assignment functions

    template <typename L>
    void assign_to(L&& lhs) const {
        std_assign_evaluate(*this, std::forward<L>(lhs));
    }

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

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

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

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

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

    // Internals

    void visit(detail::evaluator_visitor& visitor) const {
        sub.visit(visitor);
    }

    void ensure_cpu_up_to_date() const {
        // Need to ensure sub value
        sub.ensure_cpu_up_to_date();
    }

    void ensure_gpu_up_to_date() const {
        // Need to ensure both LHS and RHS
        sub.ensure_gpu_up_to_date();
    }

    friend std::ostream& operator<<(std::ostream& os, const dyn_matrix_view& v) {
        return os << "reshape[" << D << "D](" << v.sub << ")";
    }
};

template <etl_expr T, size_t D>
requires(is_dma<T>) struct dyn_matrix_view<T, D> final : iterable<dyn_matrix_view<T, D>, true>,
                                                         value_testable<dyn_matrix_view<T, D>>,
                                                         assignable<dyn_matrix_view<T, D>, value_t<T>> {
    using this_type            = dyn_matrix_view<T, D>;                                                
    using iterable_base_type   = iterable<this_type, true>;                                            
    using assignable_base_type = assignable<this_type, value_t<T>>;                                    
    using sub_type             = T;                                                                    
    using value_type           = value_t<sub_type>;                                                    
    using memory_type          = memory_t<sub_type>;                                                   
    using const_memory_type    = const_memory_t<sub_type>;                                             
    using return_type          = return_helper<sub_type, decltype(std::declval<sub_type>()[0])>;       
    using const_return_type    = const_return_helper<sub_type, decltype(std::declval<sub_type>()[0])>; 
    using iterator             = value_type*;                                                          
    using const_iterator       = const value_type*;                                                    

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

    using assignable_base_type::operator=;
    using iterable_base_type::begin;
    using iterable_base_type::end;

private:
    T sub;                            
    std::array<size_t, D> dimensions; 
    size_t _size;                     

    mutable memory_type memory; 

    static constexpr order storage_order = decay_traits<sub_type>::storage_order; 

    friend struct etl_traits<dyn_matrix_view>;

public:
    template <typename... S>
    explicit dyn_matrix_view(sub_type sub, S... dims) : sub(sub), dimensions{{dims...}}, _size(etl::size(sub)) {
        // Accessing the memory through fast sub views means evaluation
        if constexpr (decay_traits<sub_type>::is_temporary) {
            standard_evaluator::pre_assign_rhs(*this);
        }

        this->memory = this->sub.memory_start();

        cpp_assert(this->memory, "Memory should have been evaluated");
    }

    const_return_type operator[](size_t j) const {
        ensure_cpu_up_to_date();
        return memory[j];
    }

    return_type operator[](size_t j) {
        ensure_cpu_up_to_date();
        invalidate_gpu();
        return memory[j];
    }

    const_return_type operator()(size_t j) const requires (D == 1) {
        ensure_cpu_up_to_date();
        return memory[j];
    }

    return_type operator()(size_t j) requires (D == 1) {
        ensure_cpu_up_to_date();
        invalidate_gpu();
        return memory[j];
    }

    value_type read_flat(size_t j) const noexcept {
        ensure_cpu_up_to_date();
        return memory[j];
    }

    template <typename... S>
    const_return_type operator()(size_t f1, size_t f2, S... sizes) const {
        ensure_cpu_up_to_date();
        return memory[etl::dyn_index(*this, f1, f2, sizes...)];
    }

    template <typename... S>
    return_type operator()(size_t f1, size_t f2, S... sizes) {
        ensure_cpu_up_to_date();
        invalidate_gpu();
        return memory[etl::dyn_index(*this, f1, f2, sizes...)];
    }

    auto operator()(size_t i) noexcept requires (D > 1) {
        return etl::sub(*this, i);
    }

    auto operator()(size_t i) const noexcept requires (D > 1) {
        return etl::sub(*this, i);
    }

    template <typename V = default_vec>
    auto load(size_t x) const noexcept {
        return V::loadu(memory + x);
    }

    template <typename V = default_vec>
    auto loadu(size_t x) const noexcept {
        return V::loadu(memory + x);
    }

    template <typename V = default_vec>
    void stream(vec_type<V> in, size_t i) noexcept {
        return V::stream(memory + i, in);
    }

    template <typename V = default_vec>
    void store(vec_type<V> in, size_t i) noexcept {
        sub.template store<V>(in, i);
    }

    template <typename V = default_vec>
    void storeu(vec_type<V> in, size_t i) noexcept {
        return V::storeu(memory + i, in);
    }

    template <typename E>
    bool alias(const E& rhs) const noexcept {
        return sub.alias(rhs);
    }

    memory_type memory_start() noexcept {
        return memory;
    }

    const_memory_type memory_start() const noexcept {
        return memory;
    }

    memory_type memory_end() noexcept {
        return memory + _size;
    }

    const_memory_type memory_end() const noexcept {
        return memory + _size;
    }

    // Assignment functions

    template <typename L>
    void assign_to(L&& lhs) const {
        std_assign_evaluate(*this, std::forward<L>(lhs));
    }

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

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

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

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

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

    // Internals

    void visit(detail::evaluator_visitor& visitor) const {
        sub.visit(visitor);
    }

    value_type* gpu_memory() const noexcept {
        return sub.gpu_memory();
    }

    void gpu_evict() const noexcept {
        sub.gpu_evict();
    }

    void invalidate_cpu() const noexcept {
        sub.invalidate_cpu();
    }

    void invalidate_gpu() const noexcept {
        sub.invalidate_gpu();
    }

    void validate_cpu() const noexcept {
        sub.validate_cpu();
    }

    void validate_gpu() const noexcept {
        sub.validate_gpu();
    }

    void ensure_gpu_allocated() const {
        sub.ensure_gpu_allocated();
    }

    void ensure_gpu_up_to_date() const {
        sub.ensure_gpu_up_to_date();
    }

    void ensure_cpu_up_to_date() const {
        sub.ensure_cpu_up_to_date();
    }

    void gpu_copy_from(const value_type* gpu_memory) const {
        sub.gpu_copy_from(gpu_memory);
    }

    bool is_cpu_up_to_date() const noexcept {
        return sub.is_cpu_up_to_date();
    }

    bool is_gpu_up_to_date() const noexcept {
        return sub.is_gpu_up_to_date();
    }

    friend std::ostream& operator<<(std::ostream& os, const dyn_matrix_view& v) {
        return os << "reshape[" << D << "D](" << v.sub << ")";
    }
};

template <typename T, size_t D>
struct etl_traits<etl::dyn_matrix_view<T, D>> {
    using expr_t     = etl::dyn_matrix_view<T, D>;      
    using sub_expr_t = std::decay_t<T>;                 
    using sub_traits = etl_traits<sub_expr_t>;          
    using value_type = typename sub_traits::value_type; 

    static constexpr bool is_etl         = true;                       
    static constexpr bool is_transformer = false;                      
    static constexpr bool is_view        = true;                       
    static constexpr bool is_magic_view  = false;                      
    static constexpr bool is_linear      = sub_traits::is_linear;      
    static constexpr bool is_thread_safe = sub_traits::is_thread_safe; 
    static constexpr bool is_fast        = false;                      
    static constexpr bool is_value       = false;                      
    static constexpr bool is_direct      = sub_traits::is_direct;      
    static constexpr bool is_generator   = false;                      
    static constexpr bool is_padded      = false;                      
    static constexpr bool is_aligned     = false;                      
    static constexpr bool is_temporary   = sub_traits::is_temporary;   
    static constexpr bool gpu_computable = false;                      
    static constexpr order storage_order = sub_traits::storage_order;  

    template <vector_mode_t V>
    static constexpr bool vectorizable = sub_traits::template vectorizable<V>;

    static size_t size(const expr_t& v) {
        return v._size;
    }

    static size_t dim(const expr_t& v, size_t d) {
        return v.dimensions[d];
    }

    static constexpr size_t dimensions() {
        return D;
    }

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

} //end of namespace etl