Class KernelShape

cpp/src/arrow/util/bpacking_simd_kernel_internal.h:283–309

Different sizes of a given kernel. When integers are bit-packed, they can spread over multiple bytes. For instance, integers packed over three bits quickly spread over two bytes (on the value `C` below) despite three bits being much smaller than a single byte. ``` |A|A|A|B|B|B|C|C| |C|D|D|D|E|E|E|F| ... ``` When the spread is smaller or equal to the unsigned integer to unpack to, we classify i

/// reduce the number of swizzles done in a kernel run. Other optimizations of interest
/// for small input sizes on BMI2 is to use ``_pdep_u64`` directly to perform unpacking.
struct KernelShape {
  const int simd_bit_size_;
  const int unpacked_bit_size_;
  const int packed_bit_size_;
  const int packed_max_spread_bytes_ = PackedMaxSpreadBytes(packed_bit_size_, 0);
  const bool is_oversized_ =
      PackedIsOversizedForSimd(simd_bit_size_, unpacked_bit_size_, packed_bit_size_);

  constexpr bool is_medium() const {
    return packed_max_spread_bytes() <= unpacked_byte_size();
  }
  constexpr bool is_large() const { return !is_medium() && !is_oversized(); }
  constexpr bool is_oversized() const { return is_oversized_; }

  /// Properties of an SIMD batch
  constexpr int simd_bit_size() const { return simd_bit_size_; }
  constexpr int simd_byte_size() const { return simd_bit_size_ / 8; }

  /// Properties of the unpacked integers
  constexpr int unpacked_bit_size() const { return unpacked_bit_size_; }
  constexpr int unpacked_byte_size() const { return unpacked_bit_size_ / 8; }
  constexpr int unpacked_per_simd() const { return simd_bit_size_ / unpacked_bit_size_; }

  /// Properties of the packed integers
  constexpr int packed_bit_size() const { return packed_bit_size_; }
  constexpr int packed_max_spread_bytes() const { return packed_max_spread_bytes_; }
};

/// Packing all useful and derived information about a kernel in a single type.
template <typename UnpackedUint, int kPackedBitSize, typename Arch>