interleaved.hpp

#pragma once
#include <cstddef>
#include <cstdint>
#include <iterator>
#include <ranges>
#include <span>
#include "argon.hpp"
#include "arm_simd/helpers/vec128.hpp"
 
#ifdef __ARM_FEATURE_MVE
#define simd mve
#else
#define simd neon
#endif
 
namespace argon::vectorize {
 
template <size_t Stride, typename ScalarType>
struct interleaved : public std::ranges::view_interface<interleaved<Stride, ScalarType>> {
  using intrinsic_type = simd::Vec128_t<ScalarType>;
  static constexpr size_t lanes = sizeof(intrinsic_type) / sizeof(ScalarType);
  // Round down to a whole number of iterations; each iteration spans lanes*Stride elements.
  static constexpr size_t vectorizeable_size(size_t size) { return size - (size % (lanes * Stride)); }
 
 public:
  struct Iterator {
    using iterator_category = std::forward_iterator_tag;
    using argon_type = Argon<ScalarType>;
    using value_type = std::array<argon_type, Stride>;
    using difference_type = std::ptrdiff_t;
 
    Iterator() = default;
    Iterator(ScalarType* ptr) : ptr{ptr}, vec{argon_type::template LoadInterleaved<Stride>(ptr)} {}
 
    value_type& operator*() { return vec; }
    value_type* operator->() { return &vec; }
    const value_type& operator*() const { return vec; }
    const value_type* operator->() const { return &vec; }
    Iterator& operator++() {
      argon::store_interleaved(ptr, vec);  // store before increment
      ptr += lanes * Stride;
      vec = argon_type::template LoadInterleaved<Stride>(ptr);
      return *this;
    }
 
    Iterator operator++(int) {
      Iterator tmp = *this;
      ++(*this);
      return tmp;
    }
 
    friend bool operator==(const Iterator& a, const Iterator& b) { return a.ptr == b.ptr; }
    friend bool operator==(const Iterator& a, const ScalarType* ptr) { return a.ptr == ptr; }
    friend bool operator!=(const Iterator& a, const Iterator& b) { return a.ptr != b.ptr; }
    friend bool operator!=(const Iterator& a, const ScalarType* ptr) { return a.ptr != ptr; }
 
   private:
    ScalarType* ptr = nullptr;
    value_type vec;
  };
  static_assert(std::input_or_output_iterator<Iterator>);
  struct ConstIterator {
    using iterator_category = std::forward_iterator_tag;
    using argon_type = Argon<ScalarType>;
    using value_type = std::array<argon_type, Stride>;
    using difference_type = std::ptrdiff_t;
 
    ConstIterator() = default;
    ConstIterator(const ScalarType* ptr) : ptr{ptr}, vec{argon_type::template LoadInterleaved<Stride>(ptr)} {}
 
    const value_type operator*() const { return vec; }
    ConstIterator& operator++() {
      ptr += lanes * Stride;
      vec = argon_type::template LoadInterleaved<Stride>(ptr);
      return *this;
    }
    ConstIterator operator++(int) {
      ConstIterator tmp = *this;
      ++(*this);
      return tmp;
    }
    friend bool operator==(const ConstIterator& a, const ConstIterator& b) { return a.ptr == b.ptr; }
    friend bool operator!=(const ConstIterator& a, const ConstIterator& b) { return a.ptr != b.ptr; }
 
   private:
    const ScalarType* ptr = nullptr;
    value_type vec;
  };
  static_assert(std::input_iterator<ConstIterator>);
 
  using iterator = Iterator;
  using const_iterator = ConstIterator;
 
  interleaved(ScalarType* start, ScalarType* end) : start_{start}, size_{vectorizeable_size(end - start)} {};
  interleaved(ScalarType* start, const size_t size) : start_{start}, size_{vectorizeable_size(size)} {};
  interleaved(const std::span<ScalarType> span) : start_{span.data()}, size_{vectorizeable_size(span.size())} {};
 
  iterator begin() const { return Iterator(start_); }
  const ScalarType* end() const { return start_ + size_; }
  const_iterator cbegin() const { return ConstIterator(start_); }
  const ScalarType* cend() const { return start_ + size_; }
  size_t size() const { return size_; }
 
 private:
  ScalarType* start_;
  size_t size_;
};
 
// template <size_t stride, std::ranges::contiguous_range R>
// interleaved(R&& r) -> interleaved<stride, std::ranges::range_value_t<R>>;
 
// MSVC's ranges implementation does not accept this view as a std::ranges::range
// (it works on GCC and Clang). TODO: revisit MSVC support for the interleaved
// read-modify-write view; until then, skip the concept checks there.
#if !defined(_MSC_VER) || defined(__clang__)
static_assert(std::ranges::range<interleaved<3, int32_t>>);
static_assert(std::ranges::view<interleaved<3, int32_t>>);
static_assert(std::movable<interleaved<3, int32_t>>);
static_assert(std::ranges::viewable_range<interleaved<3, int32_t>>);
#endif
 
}  // namespace argon::vectorize
 
#undef simd