exr 1.74.2

//! Data structures that represent a complete exr image.
//! Contains generic structs that must be nested to obtain a complete image
//! type.
//!
//!
//! For example, an rgba image containing multiple layers
//! can be represented using `Image<Layers<SpecificChannels<MyPixelStorage>>>`.
//! An image containing a single layer with arbitrary channels and no deep data
//! can be represented using `Image<Layer<AnyChannels<FlatSamples>>>`.
//!
//!
//! These and other predefined types are included in this module as
//! 1. `PixelImage`: A single layer, fixed set of arbitrary channels.
//! 1. `PixelLayersImage`: Multiple layers, fixed set of arbitrary channels.
//! 1. `RgbaImage`: A single layer, fixed set of channels: rgb, optional a.
//! 1. `RgbaLayersImage`: Multiple layers, fixed set of channels: rgb, optional
//!    a.
//! 1. `FlatImage`: Multiple layers, any channels, no deep data.
//! 1. `AnyImage`: All supported data (multiple layers, arbitrary channels, no
//!    deep data yet)
//!
//! You can also use your own types inside an image,
//! for example if you want to use a custom sample storage.
//!
//! This is the high-level interface for the pixels of an image.
//! See `exr::blocks` module for a low-level interface.

pub mod crop;
pub mod pixel_vec;
pub mod read;
pub mod recursive;
pub mod write;
// pub mod channel_groups;

use half::f16;
use smallvec::SmallVec;

use crate::{
    compression::Compression,
    error::Error,
    math::{RoundingMode, Vec2},
    meta::{
        attribute::{LineOrder, Text},
        header::{ImageAttributes, LayerAttributes},
    },
};

/// Don't do anything
pub(crate) const fn ignore_progress(_progress: f64) {}

/// This image type contains all supported exr features and can represent almost
/// any image. It currently does not support deep data yet.
pub type AnyImage = Image<Layers<AnyChannels<Levels<FlatSamples>>>>;

/// This image type contains the most common exr features and can represent
/// almost any plain image. Does not contain resolution levels. Does not support
/// deep data.
pub type FlatImage = Image<Layers<AnyChannels<FlatSamples>>>;

/// This image type contains multiple layers, with each layer containing a
/// user-defined type of pixels.
pub type PixelLayersImage<Storage, Channels> = Image<Layers<SpecificChannels<Storage, Channels>>>;

/// This image type contains a single layer containing a user-defined type of
/// pixels.
pub type PixelImage<Storage, Channels> = Image<Layer<SpecificChannels<Storage, Channels>>>;

/// This image type contains multiple layers, with each layer containing a
/// user-defined type of rgba pixels.
pub type RgbaLayersImage<Storage> = PixelLayersImage<Storage, RgbaChannels>;

/// This image type contains a single layer containing a user-defined type of
/// rgba pixels.
pub type RgbaImage<Storage> = PixelImage<Storage, RgbaChannels>;

/// Contains information about the channels in an rgba image, in the order
/// `(red, green, blue, alpha)`. The alpha channel is not required. May be
/// `None` if the image did not contain an alpha channel.
pub type RgbaChannels =
    (ChannelDescription, ChannelDescription, ChannelDescription, Option<ChannelDescription>);

/// Contains information about the channels in an rgb image, in the order `(red,
/// green, blue)`.
pub type RgbChannels = (ChannelDescription, ChannelDescription, ChannelDescription);

/// The complete exr image.
/// `Layers` can be either a single `Layer` or `Layers`.
#[derive(Debug, Clone, PartialEq)]
pub struct Image<Layers> {
    /// Attributes that apply to the whole image file.
    /// These attributes appear in each layer of the file.
    /// Excludes technical meta data.
    /// Each layer in this image also has its own attributes.
    pub attributes: ImageAttributes,

    /// The layers contained in the image file.
    /// Can be either a single `Layer` or a list of layers.
    pub layer_data: Layers,
}

/// A list of layers. `Channels` can be `SpecificChannels` or `AnyChannels`.
pub type Layers<Channels> = SmallVec<[Layer<Channels>; 2]>;

/// A single Layer, including fancy attributes and compression settings.
/// `Channels` can be either `SpecificChannels` or `AnyChannels`
#[derive(Debug, Clone, PartialEq)]
pub struct Layer<Channels> {
    /// The actual pixel data. Either `SpecificChannels` or `AnyChannels`
    pub channel_data: Channels,

    /// Attributes that apply to this layer.
    /// May still contain attributes that should be considered global for an
    /// image file. Excludes technical meta data: Does not contain data
    /// window size, line order, tiling, or compression attributes.
    /// The image also has attributes, which do not differ per layer.
    pub attributes: LayerAttributes,

    /// The pixel resolution of this layer.
    /// See `layer.attributes` for more attributes, like for example layer
    /// position.
    pub size: Vec2<usize>,

    /// How the pixels are split up and compressed.
    pub encoding: Encoding,
}

/// How the pixels are split up and compressed.
#[derive(Copy, Clone, Debug, PartialEq)]
pub struct Encoding {
    /// How the pixel data of all channels in this layer is compressed. May be
    /// `Compression::Uncompressed`. See `layer.attributes` for more
    /// attributes.
    pub compression: Compression,

    /// Describes how the pixels of this layer are divided into smaller blocks.
    /// Either splits the image into its scan lines or splits the image into
    /// tiles of the specified size. A single block can be loaded without
    /// processing all bytes of a file.
    pub blocks: Blocks,

    /// In what order the tiles of this header occur in the file.
    /// Does not change any actual image orientation.
    /// See `layer.attributes` for more attributes.
    pub line_order: LineOrder,
}

/// How the image pixels are split up into separate blocks.
#[derive(Copy, Clone, Debug, PartialEq, Eq)]
pub enum Blocks {
    /// The image is divided into scan line blocks.
    /// The number of scan lines in a block depends on the compression method.
    ScanLines,

    /// The image is divided into tile blocks.
    /// Also specifies the size of each tile in the image
    /// and whether this image contains multiple resolution levels.
    ///
    /// The inner `Vec2` describes the size of each tile.
    /// Stays the same number of pixels across all levels.
    Tiles(Vec2<usize>),
}

/// A grid of pixels. The pixels are written to your custom pixel storage.
///
/// `PixelStorage` can be anything, from a flat `Vec<f16>` to
/// `Vec<Vec<AnySample>>`, as desired. In order to write this image to a file,
/// your `PixelStorage` must implement [`GetPixel`].
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct SpecificChannels<Pixels, ChannelsDescription> {
    /// A description of the channels in the file, as opposed to the channels in
    /// memory. Should always be a tuple containing `ChannelDescription`s,
    /// one description for each channel.
    pub channels: ChannelsDescription, /* TODO this is awkward. can this be not a type parameter
                                        * please? maybe vec<option<chan_info>> ?? */

    /// Your custom pixel storage
    // TODO should also support `Levels<YourStorage>`, where levels are desired!
    pub pixels: Pixels, // TODO rename to "pixels"?
}

/// A dynamic list of arbitrary channels.
/// `Samples` can currently only be `FlatSamples` or `Levels<FlatSamples>`.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct AnyChannels<Samples> {
    /// This list must be sorted alphabetically, by channel name.
    /// Use `AnyChannels::sorted` for automatic sorting.
    pub list: SmallVec<[AnyChannel<Samples>; 4]>,
}

/// A single arbitrary channel.
/// `Samples` can currently only be `FlatSamples` or `Levels<FlatSamples>`
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct AnyChannel<Samples> {
    /// One of "R", "G", or "B" most of the time.
    pub name: Text,

    /// The actual pixel data.
    /// Can be `FlatSamples` or `Levels<FlatSamples>`.
    pub sample_data: Samples,

    /// This attribute only tells lossy compression methods
    /// whether this value should be quantized exponentially or linearly.
    ///
    /// Should be `false` for red, green, blue and luma channels, as they are
    /// not perceived linearly. Should be `true` for hue, chroma,
    /// saturation, and alpha channels.
    pub quantize_linearly: bool,

    /// How many of the samples are skipped compared to the other channels in
    /// this layer.
    ///
    /// Can be used for chroma subsampling for manual lossy data compression.
    /// Values other than 1 are allowed only in flat, scan-line based images.
    /// If an image is deep or tiled, the sampling rates for all of its channels
    /// must be 1.
    pub sampling: Vec2<usize>,
}

/// One or multiple resolution levels of the same image.
/// `Samples` can be `FlatSamples`.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum Levels<Samples> {
    /// A single image without smaller versions of itself.
    /// If you only want to handle exclusively this case, use `Samples`
    /// directly, and not `Levels<Samples>`.
    Singular(Samples),

    /// Contains uniformly scaled smaller versions of the original.
    Mip {
        /// Whether to round up or down when calculating Mip/Rip levels.
        rounding_mode: RoundingMode,

        /// The smaller versions of the original.
        level_data: LevelMaps<Samples>,
    },

    /// Contains any possible combination of smaller versions of the original.
    Rip {
        /// Whether to round up or down when calculating Mip/Rip levels.
        rounding_mode: RoundingMode,

        /// The smaller versions of the original.
        level_data: RipMaps<Samples>,
    },
}

/// A list of resolution levels. `Samples` can currently only be `FlatSamples`.
// or `DeepAndFlatSamples` (not yet implemented).
pub type LevelMaps<Samples> = Vec<Samples>;

/// In addition to the full resolution image,
/// this layer also contains smaller versions,
/// and each smaller version has further versions with varying aspect ratios.
/// `Samples` can currently only be `FlatSamples`.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct RipMaps<Samples> {
    /// A flattened list containing the individual levels
    pub map_data: LevelMaps<Samples>,

    /// The number of levels that were generated along the x-axis and y-axis.
    pub level_count: Vec2<usize>,
}

// TODO deep data
// #[derive(Clone, PartialEq)]
// pub enum DeepAndFlatSamples {
// Deep(DeepSamples),
// Flat(FlatSamples)
// }

/// A vector of non-deep values (one value per pixel per channel).
/// Stores row after row in a single vector.
/// The precision of all values is either `f16`, `f32` or `u32`.
///
/// Since this is close to the pixel layout in the byte file,
/// this will most likely be the fastest storage.
/// Using a different storage, for example `SpecificChannels`,
/// will probably be slower.
#[derive(Clone, PartialEq)] // debug is implemented manually
pub enum FlatSamples {
    /// A vector of non-deep `f16` values.
    F16(Vec<f16>),

    /// A vector of non-deep `f32` values.
    F32(Vec<f32>),

    /// A vector of non-deep `u32` values.
    U32(Vec<u32>),
}

// #[derive(Clone, PartialEq)]
// pub enum DeepSamples {
// F16(Vec<Vec<f16>>),
// F32(Vec<Vec<f32>>),
// U32(Vec<Vec<u32>>),
// }

use std::{marker::PhantomData, ops::Not};

use crate::{
    block::samples::{Sample, *},
    error::Result,
    image::{
        recursive::{IntoRecursive, NoneMore, Recursive},
        validate_results::ValidationOptions,
        write::{channels::*, layers::WritableLayers, samples::WritableSamples},
    },
    io::Data,
    meta::{attribute::*, mip_map_levels, rip_map_levels},
};

impl<Channels> Layer<Channels> {
    /// Sometimes called "data window"
    pub fn absolute_bounds(&self) -> IntegerBounds {
        IntegerBounds::new(self.attributes.layer_position, self.size)
    }
}

impl<SampleStorage, Channels> SpecificChannels<SampleStorage, Channels> {
    /// Create some pixels with channel information.
    /// The `Channels` must be a tuple containing either `ChannelDescription` or
    /// `Option<ChannelDescription>`. The length of the tuple dictates the
    /// number of channels in the sample storage.
    pub const fn new(channels: Channels, source_samples: SampleStorage) -> Self
    where
        SampleStorage: GetPixel,
        SampleStorage::Pixel: IntoRecursive,
        Channels: Sync + Clone + IntoRecursive,
        <Channels as IntoRecursive>::Recursive:
            WritableChannelsDescription<<SampleStorage::Pixel as IntoRecursive>::Recursive>,
    {
        Self {
            channels,
            pixels: source_samples,
        }
    }
}

/// Convert this type into one of the known sample types.
/// Also specify the preferred native type, which dictates the default sample
/// type in the image.
pub trait IntoSample: IntoNativeSample {
    /// The native sample types that this type should be converted to.
    const PREFERRED_SAMPLE_TYPE: SampleType;
}

impl IntoSample for f16 {
    const PREFERRED_SAMPLE_TYPE: SampleType = SampleType::F16;
}
impl IntoSample for f32 {
    const PREFERRED_SAMPLE_TYPE: SampleType = SampleType::F32;
}
impl IntoSample for u32 {
    const PREFERRED_SAMPLE_TYPE: SampleType = SampleType::U32;
}

/// Used to construct a `SpecificChannels`.
/// Call `with_named_channel` as many times as desired,
/// and then call `with_pixels` to define the colors.
#[derive(Debug)]
pub struct SpecificChannelsBuilder<RecursiveChannels, RecursivePixel> {
    channels: RecursiveChannels,
    px: PhantomData<RecursivePixel>,
}

/// This check can be executed at compile time
/// if the channel names are `&'static str` and the compiler is smart enough.
pub trait CheckDuplicates {
    /// Check for duplicate channel names.
    fn already_contains(&self, name: &Text) -> bool;
}

impl CheckDuplicates for NoneMore {
    fn already_contains(&self, _: &Text) -> bool {
        false
    }
}

impl<Inner: CheckDuplicates> CheckDuplicates for Recursive<Inner, ChannelDescription> {
    fn already_contains(&self, name: &Text) -> bool {
        &self.value.name == name || self.inner.already_contains(name)
    }
}

impl SpecificChannels<(), ()> {
    /// Start building some specific channels. On the result of this function,
    /// call `with_named_channel` as many times as desired,
    /// and then call `with_pixels` to define the colors.
    pub fn build() -> SpecificChannelsBuilder<NoneMore, NoneMore> {
        SpecificChannelsBuilder {
            channels: NoneMore,
            px: Default::default(),
        }
    }
}

impl<RecursiveChannels: CheckDuplicates, RecursivePixel>
    SpecificChannelsBuilder<RecursiveChannels, RecursivePixel>
{
    /// Add another channel to this image. Does not add the actual pixels,
    /// but instead only declares the presence of the channel.
    /// Panics if the name contains unsupported characters.
    /// Panics if a channel with the same name already exists.
    /// Use `Text::new_or_none()` to manually handle these cases.
    /// Use `with_channel_details` instead if you want to specify more options
    /// than just the name of the channel. The generic parameter can usually
    /// be inferred from the closure in `with_pixels`.
    pub fn with_channel<Sample: IntoSample>(
        self,
        name: impl Into<Text>,
    ) -> SpecificChannelsBuilder<
        Recursive<RecursiveChannels, ChannelDescription>,
        Recursive<RecursivePixel, Sample>,
    > {
        self.with_channel_details::<Sample>(ChannelDescription::named(
            name,
            Sample::PREFERRED_SAMPLE_TYPE,
        ))
    }

    /// Add another channel to this image. Does not add the actual pixels,
    /// but instead only declares the presence of the channel.
    /// Use `with_channel` instead if you only want to specify the name of the
    /// channel. Panics if a channel with the same name already exists.
    /// The generic parameter can usually be inferred from the closure in
    /// `with_pixels`.
    pub fn with_channel_details<Sample: Into<Sample>>(
        self,
        channel: ChannelDescription,
    ) -> SpecificChannelsBuilder<
        Recursive<RecursiveChannels, ChannelDescription>,
        Recursive<RecursivePixel, Sample>,
    > {
        // duplicate channel names are checked later, but also check now to make sure
        // there are no problems with the `SpecificChannelsWriter`
        assert!(
            self.channels.already_contains(&channel.name).not(),
            "channel name `{}` is duplicate",
            channel.name
        );

        SpecificChannelsBuilder {
            channels: Recursive::new(self.channels, channel),
            px: PhantomData,
        }
    }

    /// Specify the actual pixel contents of the image.
    /// You can pass a closure that returns a color for each pixel
    /// (`Fn(Vec2<usize>) -> Pixel`), or you can pass your own image if it
    /// implements `GetPixel`. The pixel type must be a tuple with the
    /// correct number of entries, depending on the number of channels.
    /// The tuple entries can be either `f16`, `f32`, `u32` or `Sample`.
    /// Use `with_pixel_fn` instead of this function, to get extra type safety
    /// for your pixel closure.
    pub fn with_pixels<Pixels>(
        self,
        get_pixel: Pixels,
    ) -> SpecificChannels<Pixels, RecursiveChannels>
    where
        Pixels: GetPixel,
        <Pixels as GetPixel>::Pixel: IntoRecursive<Recursive = RecursivePixel>,
    {
        SpecificChannels {
            channels: self.channels,
            pixels: get_pixel,
        }
    }

    /// Specify the contents of the image.
    /// The pixel type must be a tuple with the correct number of entries,
    /// depending on the number of channels. The tuple entries can be either
    /// `f16`, `f32`, `u32` or `Sample`. Use `with_pixels` instead of this
    /// function, if you want to pass an object that is not a closure.
    ///
    /// Usually, the compiler can infer the type of the pixel (for example,
    /// `f16,f32,f32`) from the closure. If that's not possible, you can
    /// specify the type of the channels when declaring the channel (for
    /// example, `with_named_channel::<f32>("R")`).
    pub fn with_pixel_fn<Pixel, Pixels>(
        self,
        get_pixel: Pixels,
    ) -> SpecificChannels<Pixels, RecursiveChannels>
    where
        Pixels: Sync + Fn(Vec2<usize>) -> Pixel,
        Pixel: IntoRecursive<Recursive = RecursivePixel>,
    {
        SpecificChannels {
            channels: self.channels,
            pixels: get_pixel,
        }
    }
}

impl<SampleStorage>
    SpecificChannels<
        SampleStorage,
        (ChannelDescription, ChannelDescription, ChannelDescription, ChannelDescription),
    >
{
    /// Create an image with red, green, blue, and alpha channels.
    /// You can pass a closure that returns a color for each pixel
    /// (`Fn(Vec2<usize>) -> (R,G,B,A)`), or you can pass your own image if
    /// it implements `GetPixel<Pixel=(R,G,B,A)>`. Each of `R`, `G`, `B` and
    /// `A` can be either `f16`, `f32`, `u32`, or `Sample`.
    pub fn rgba<R, G, B, A>(source_samples: SampleStorage) -> Self
    where
        R: IntoSample,
        G: IntoSample,
        B: IntoSample,
        A: IntoSample,
        SampleStorage: GetPixel<Pixel = (R, G, B, A)>,
    {
        Self {
            channels: (
                ChannelDescription::named("R", R::PREFERRED_SAMPLE_TYPE),
                ChannelDescription::named("G", G::PREFERRED_SAMPLE_TYPE),
                ChannelDescription::named("B", B::PREFERRED_SAMPLE_TYPE),
                ChannelDescription::named("A", A::PREFERRED_SAMPLE_TYPE),
            ),
            pixels: source_samples,
        }
    }
}

impl<SampleStorage>
    SpecificChannels<SampleStorage, (ChannelDescription, ChannelDescription, ChannelDescription)>
{
    /// Create an image with red, green, and blue channels.
    /// You can pass a closure that returns a color for each pixel
    /// (`Fn(Vec2<usize>) -> (R,G,B)`), or you can pass your own image if it
    /// implements `GetPixel<Pixel=(R,G,B)>`. Each of `R`, `G` and `B` can
    /// be either `f16`, `f32`, `u32`, or `Sample`.
    pub fn rgb<R, G, B>(source_samples: SampleStorage) -> Self
    where
        R: IntoSample,
        G: IntoSample,
        B: IntoSample,
        SampleStorage: GetPixel<Pixel = (R, G, B)>,
    {
        Self {
            channels: (
                ChannelDescription::named("R", R::PREFERRED_SAMPLE_TYPE),
                ChannelDescription::named("G", G::PREFERRED_SAMPLE_TYPE),
                ChannelDescription::named("B", B::PREFERRED_SAMPLE_TYPE),
            ),
            pixels: source_samples,
        }
    }
}

/// A list of samples representing a single pixel.
/// Does not heap allocate for images with 8 or fewer channels.
pub type FlatSamplesPixel = SmallVec<[Sample; 8]>;

// TODO also deep samples?
impl Layer<AnyChannels<FlatSamples>> {
    /// Use `samples_at` if you can borrow from this layer
    pub fn sample_vec_at(&self, position: Vec2<usize>) -> FlatSamplesPixel {
        self.samples_at(position).collect()
    }

    /// Lookup all channels of a single pixel in the image
    pub const fn samples_at(&self, position: Vec2<usize>) -> FlatSampleIterator<'_> {
        FlatSampleIterator {
            layer: self,
            channel_index: 0,
            position,
        }
    }
}

/// Iterate over all channels of a single pixel in the image
#[derive(Debug, Copy, Clone, PartialEq)]
pub struct FlatSampleIterator<'s> {
    layer: &'s Layer<AnyChannels<FlatSamples>>,
    channel_index: usize,
    position: Vec2<usize>,
}

impl Iterator for FlatSampleIterator<'_> {
    type Item = Sample;

    fn next(&mut self) -> Option<Self::Item> {
        if self.channel_index < self.layer.channel_data.list.len() {
            let channel = &self.layer.channel_data.list[self.channel_index];
            let sample = channel
                .sample_data
                .value_by_flat_index(self.position.flat_index_for_size(self.layer.size));
            self.channel_index += 1;
            Some(sample)
        } else {
            None
        }
    }

    fn nth(&mut self, pos: usize) -> Option<Self::Item> {
        self.channel_index += pos;
        self.next()
    }

    fn size_hint(&self) -> (usize, Option<usize>) {
        let remaining = self.layer.channel_data.list.len().saturating_sub(self.channel_index);
        (remaining, Some(remaining))
    }
}

impl ExactSizeIterator for FlatSampleIterator<'_> {}

impl<SampleData> AnyChannels<SampleData> {
    /// A new list of arbitrary channels. Sorts the list to make it
    /// alphabetically stable.
    pub fn sort(mut list: SmallVec<[AnyChannel<SampleData>; 4]>) -> Self {
        list.sort_unstable_by_key(|channel| channel.name.clone()); // TODO no clone?
        Self {
            list,
        }
    }
}

// FIXME check content size of layer somewhere??? before writing?
impl<LevelSamples> Levels<LevelSamples> {
    /// Get a resolution level by index, sorted by size, decreasing.
    pub fn get_level(&self, level: Vec2<usize>) -> Result<&LevelSamples> {
        match self {
            Self::Singular(block) => {
                debug_assert_eq!(
                    level,
                    Vec2(0, 0),
                    "singular image cannot write leveled blocks bug"
                );
                Ok(block)
            }

            Self::Mip {
                level_data,
                ..
            } => {
                debug_assert_eq!(
                    level.x(),
                    level.y(),
                    "mip map levels must be equal on x and y bug"
                );
                level_data.get(level.x()).ok_or_else(|| {
                    Error::invalid(format!(
                        "mip level index {} out of range (max: {})",
                        level.x(),
                        level_data.len().saturating_sub(1)
                    ))
                })
            }

            Self::Rip {
                level_data,
                ..
            } => level_data
                .get_by_level(level)
                .ok_or_else(|| Error::invalid(format!("rip level index {level:?} not found"))),
        }
    }

    /// Get a resolution level by index, sorted by size, decreasing.
    // TODO storage order for RIP maps?
    pub fn get_level_mut(&mut self, level: Vec2<usize>) -> Result<&mut LevelSamples> {
        match self {
            Self::Singular(ref mut block) => {
                debug_assert_eq!(
                    level,
                    Vec2(0, 0),
                    "singular image cannot write leveled blocks bug"
                );
                Ok(block)
            }

            Self::Mip {
                level_data,
                ..
            } => {
                debug_assert_eq!(
                    level.x(),
                    level.y(),
                    "mip map levels must be equal on x and y bug"
                );
                let max_level = level_data.len().saturating_sub(1);
                let level_index = level.x();
                level_data.get_mut(level_index).ok_or_else(|| {
                    Error::invalid(format!(
                        "mip level index {level_index} out of range (max: {max_level})"
                    ))
                })
            }

            Self::Rip {
                level_data,
                ..
            } => level_data
                .get_by_level_mut(level)
                .ok_or_else(|| Error::invalid(format!("rip level index {level:?} not found"))),
        }
    }

    /// Get a slice of all resolution levels, sorted by size, decreasing.
    pub fn levels_as_slice(&self) -> &[LevelSamples] {
        match self {
            Self::Singular(data) => std::slice::from_ref(data),
            Self::Mip {
                level_data,
                ..
            } => level_data,
            Self::Rip {
                level_data,
                ..
            } => &level_data.map_data,
        }
    }

    /// Get a mutable slice of all resolution levels, sorted by size,
    /// decreasing.
    pub fn levels_as_slice_mut(&mut self) -> &mut [LevelSamples] {
        match self {
            Self::Singular(data) => std::slice::from_mut(data),
            Self::Mip {
                level_data,
                ..
            } => level_data,
            Self::Rip {
                level_data,
                ..
            } => &mut level_data.map_data,
        }
    }

    // TODO simplify working with levels in general! like level_size_by_index and
    // such

    // pub fn levels_with_size(&self, rounding: RoundingMode, max_resolution:
    // Vec2<usize>) -> Vec<(Vec2<usize>, &S)> { match self {
    // Levels::Singular(ref data) => vec![ (max_resolution, data) ],
    // Levels::Mip(ref maps) => mip_map_levels(rounding,
    // max_resolution).map(|(_index, size)| size).zip(maps).collect(),
    // Levels::Rip(ref rip_maps) => rip_map_levels(rounding,
    // max_resolution).map(|(_index, size)| size).zip(&rip_maps.map_data).collect(),
    // }
    // }

    /// Whether this stores multiple resolution levels.
    pub const fn level_mode(&self) -> LevelMode {
        match self {
            Self::Singular(_) => LevelMode::Singular,
            Self::Mip {
                ..
            } => LevelMode::MipMap,
            Self::Rip {
                ..
            } => LevelMode::RipMap,
        }
    }
}

impl<Samples> RipMaps<Samples> {
    /// Flatten the 2D level index to a one dimensional index.
    pub fn get_level_index(&self, level: Vec2<usize>) -> usize {
        level.flat_index_for_size(self.level_count)
    }

    /// Return a level by level index. Level `0` has the largest resolution.
    pub fn get_by_level(&self, level: Vec2<usize>) -> Option<&Samples> {
        self.map_data.get(self.get_level_index(level))
    }

    /// Return a mutable level reference by level index. Level `0` has the
    /// largest resolution.
    pub fn get_by_level_mut(&mut self, level: Vec2<usize>) -> Option<&mut Samples> {
        let index = self.get_level_index(level);
        self.map_data.get_mut(index)
    }
}

impl FlatSamples {
    /// The number of samples in the image. Should be the width times the
    /// height. Might vary when subsampling is used.
    pub fn len(&self) -> usize {
        match self {
            Self::F16(vec) => vec.len(),
            Self::F32(vec) => vec.len(),
            Self::U32(vec) => vec.len(),
        }
    }

    /// Views all samples in this storage as f32.
    /// Matches the underlying sample type again for every sample,
    /// match yourself if performance is critical! Does not allocate.
    pub fn values_as_f32(&self) -> impl '_ + Iterator<Item = f32> {
        self.values().map(super::block::samples::Sample::to_f32)
    }

    /// All samples in this storage as iterator.
    /// Matches the underlying sample type again for every sample,
    /// match yourself if performance is critical! Does not allocate.
    pub fn values(&self) -> impl '_ + Iterator<Item = Sample> {
        (0..self.len()).map(move |index| self.value_by_flat_index(index))
    }

    /// Lookup a single value, by flat index.
    /// The flat index can be obtained using `Vec2::flatten_for_width`
    /// which computes the index in a flattened array of pixel rows.
    pub fn value_by_flat_index(&self, index: usize) -> Sample {
        match self {
            Self::F16(vec) => Sample::F16(vec[index]),
            Self::F32(vec) => Sample::F32(vec[index]),
            Self::U32(vec) => Sample::U32(vec[index]),
        }
    }
}

impl<'s, ChannelData: 's> Layer<ChannelData> {
    /// Create a layer with the specified size, attributes, encoding and
    /// channels. The channels can be either `SpecificChannels` or
    /// `AnyChannels`.
    pub fn new(
        dimensions: impl Into<Vec2<usize>>,
        attributes: LayerAttributes,
        encoding: Encoding,
        channels: ChannelData,
    ) -> Self
    where
        ChannelData: WritableChannels<'s>,
    {
        Self {
            channel_data: channels,
            attributes,
            size: dimensions.into(),
            encoding,
        }
    }

    // TODO test pls wtf
    /// Panics for images with Scanline encoding.
    pub fn levels_with_resolution<'l, L>(
        &self,
        levels: &'l Levels<L>,
    ) -> Box<dyn 'l + Iterator<Item = (&'l L, Vec2<usize>)>> {
        match levels {
            Levels::Singular(level) => Box::new(std::iter::once((level, self.size))),

            Levels::Mip {
                rounding_mode,
                level_data,
            } => Box::new(
                level_data
                    .iter()
                    .zip(mip_map_levels(*rounding_mode, self.size).map(|(_index, size)| size)),
            ),

            Levels::Rip {
                rounding_mode,
                level_data,
            } => Box::new(
                level_data
                    .map_data
                    .iter()
                    .zip(rip_map_levels(*rounding_mode, self.size).map(|(_index, size)| size)),
            ),
        }
    }
}

impl Encoding {
    /// Run-length encoding with tiles of 64x64 pixels. This is the recommended
    /// default encoding. Almost as fast as uncompressed data, but optimizes
    /// single-colored areas such as mattes and masks.
    pub const FAST_LOSSLESS: Self = Self {
        compression: Compression::RLE,
        blocks: Blocks::Tiles(Vec2(64, 64)), // optimize for RLE compression
        line_order: LineOrder::Unspecified,
    };
    /// PIZ compression with tiles of 256x256 pixels. Small images, not too
    /// slow.
    pub const SMALL_FAST_LOSSLESS: Self = Self {
        compression: Compression::PIZ,
        blocks: Blocks::Tiles(Vec2(256, 256)),
        line_order: LineOrder::Unspecified,
    };
    /// ZIP compression with blocks of 16 lines. Slow, but produces small files
    /// without visible artefacts.
    pub const SMALL_LOSSLESS: Self = Self {
        compression: Compression::ZIP16,
        blocks: Blocks::ScanLines, /* largest possible, but also with high probability of
                                    * parallel workers */
        line_order: LineOrder::Increasing,
    };
    /// No compression. Massive space requirements.
    /// Fast, because it minimizes data shuffling and reallocation.
    pub const UNCOMPRESSED: Self = Self {
        compression: Compression::Uncompressed,
        blocks: Blocks::ScanLines,         // longest lines, faster memcpy
        line_order: LineOrder::Increasing, // presumably fastest?
    };
}

impl Default for Encoding {
    fn default() -> Self {
        Self::FAST_LOSSLESS
    }
}

impl<'s, LayerData: 's> Image<LayerData>
where
    LayerData: WritableLayers<'s>,
{
    /// Create an image with one or multiple layers. The layer can be a `Layer`,
    /// or `Layers` small vector, or `Vec<Layer>` or `&[Layer]`.
    pub const fn new(image_attributes: ImageAttributes, layer_data: LayerData) -> Self {
        Self {
            attributes: image_attributes,
            layer_data,
        }
    }
}

// explorable constructor alias
impl<'s, Channels: 's> Image<Layers<Channels>>
where
    Channels: WritableChannels<'s>,
{
    /// Create an image with multiple layers. The layer can be a `Vec<Layer>` or
    /// `Layers` (a small vector).
    pub fn from_layers(
        image_attributes: ImageAttributes,
        layer_data: impl Into<Layers<Channels>>,
    ) -> Self {
        Self::new(image_attributes, layer_data.into())
    }
}

impl<'s, ChannelData: 's> Image<Layer<ChannelData>>
where
    ChannelData: WritableChannels<'s>,
{
    /// Uses the display position and size to the channel position and size of
    /// the layer.
    pub fn from_layer(layer: Layer<ChannelData>) -> Self {
        let bounds = IntegerBounds::new(layer.attributes.layer_position, layer.size);
        Self::new(ImageAttributes::new(bounds), layer)
    }

    /// Uses empty attributes.
    pub fn from_encoded_channels(
        size: impl Into<Vec2<usize>>,
        encoding: Encoding,
        channels: ChannelData,
    ) -> Self {
        // layer name is not required for single-layer images
        Self::from_layer(Layer::new(size, LayerAttributes::default(), encoding, channels))
    }

    /// Uses empty attributes and fast compression.
    pub fn from_channels(size: impl Into<Vec2<usize>>, channels: ChannelData) -> Self {
        Self::from_encoded_channels(size, Encoding::default(), channels)
    }
}

impl Image<NoneMore> {
    /// Create an empty image, to be filled with layers later on. Add at least
    /// one layer to obtain a valid image. Call `with_layer(another_layer)`
    /// for each layer you want to add to this image.
    #[must_use]
    pub const fn empty(attributes: ImageAttributes) -> Self {
        Self {
            attributes,
            layer_data: NoneMore,
        }
    }
}

impl<'s, InnerLayers: 's> Image<InnerLayers>
where
    InnerLayers: WritableLayers<'s>,
{
    /// Add another layer to this image. The layer type does
    /// not have to equal the existing layers in this image.
    pub fn with_layer<NewChannels>(
        self,
        layer: Layer<NewChannels>,
    ) -> Image<Recursive<InnerLayers, Layer<NewChannels>>>
    where
        NewChannels: 's + WritableChannels<'s>,
    {
        Image {
            attributes: self.attributes,
            layer_data: Recursive::new(self.layer_data, layer),
        }
    }
}

impl<'s, SampleData: 's> AnyChannel<SampleData> {
    /// Create a new channel without subsampling.
    ///
    /// Automatically flags this channel for specialized compression
    /// if the name is "R", "G", "B", "Y", or "L",
    /// as they typically encode values that are perceived non-linearly.
    /// Construct the value yourself using `AnyChannel { .. }`, if you want to
    /// control this flag.
    pub fn new(name: impl Into<Text>, sample_data: SampleData) -> Self
    where
        SampleData: WritableSamples<'s>,
    {
        let name: Text = name.into();

        Self {
            quantize_linearly: ChannelDescription::guess_quantization_linearity(&name),
            name,
            sample_data,
            sampling: Vec2(1, 1),
        }
    }

    // /// This is the same as `AnyChannel::new()`, but additionally ensures that
    // the closure type is correct. pub fn from_closure<V>(name: Text,
    // sample_data: S) -> Self where S: Sync + Fn(Vec2<usize>) -> V, V:
    // InferSampleType + Data {
    // Self::new(name, sample_data)
    // }
}

impl std::fmt::Debug for FlatSamples {
    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        if self.len() <= 6 {
            match self {
                Self::F16(vec) => vec.fmt(formatter),
                Self::F32(vec) => vec.fmt(formatter),
                Self::U32(vec) => vec.fmt(formatter),
            }
        } else {
            match self {
                Self::F16(vec) => write!(formatter, "[f16; {}]", vec.len()),
                Self::F32(vec) => write!(formatter, "[f32; {}]", vec.len()),
                Self::U32(vec) => write!(formatter, "[u32; {}]", vec.len()),
            }
        }
    }
}

/// Compare the result of a round trip test with the original method.
/// Supports lossy compression methods.
// #[cfg(test)] TODO do not ship this code
pub mod validate_results {
    use std::ops::Not;

    use smallvec::Array;

    use crate::{
        block::samples::IntoNativeSample,
        image::write::samples::WritableSamples,
        prelude::{recursive::*, *},
    };

    /// Compare two objects, but with a few special quirks.
    /// Intended mainly for unit testing.
    pub trait ValidateResult {
        /// Compare self with the other. Panics if not equal.
        ///
        /// Exceptional behaviour:
        /// This does not work the other way around! This method is not
        /// symmetrical! Returns whether the result is correct for this
        /// image. For lossy compression methods, uses approximate
        /// equality. Intended for unit testing.
        ///
        /// Warning: If you use `SpecificChannels`, the comparison might be
        /// inaccurate for images with mixed compression methods. This
        /// is to be used with `AnyChannels` mainly.
        fn assert_equals_result(&self, result: &Self) {
            self.validate_result(result, ValidationOptions::default(), String::new).unwrap();
        }

        /// Like [`Self::assert_equals_result`], but uses approximate (lossy)
        /// comparison for floating point values. Panics if they are not
        /// approximately equal.
        ///
        /// This is the single definitive helper for "a bunch of floats are
        /// approximately equal" checks in tests: it applies the same
        /// adaptive tolerance (`0.06 * (|a| + |b|)`, floored at `0.1`)
        /// that whole-image lossy comparison uses. Works on `f32`/
        /// `f16`, on slices/`Vec`s of them, and on the whole image types.
        fn assert_approx_equals_result(&self, result: &Self) {
            self.validate_result(
                result,
                ValidationOptions {
                    allow_lossy: true,
                    nan_converted_to_zero: false,
                },
                String::new,
            )
            .unwrap();
        }

        /// Compare self with the other.
        /// Exceptional behaviour:
        /// - Any two NaN values are considered equal, regardless of bit
        ///   representation.
        /// - If a `lossy` is specified, any two values that differ only by a
        ///   small amount will be considered equal.
        /// - If `nan_to_zero` is true, and __self is NaN/Infinite and the other
        ///   value is zero, they are considered equal__ (because some
        ///   compression methods replace nan with zero)
        ///
        /// This does not work the other way around! This method is not
        /// symmetrical!
        fn validate_result(
            &self,
            lossy_result: &Self,
            options: ValidationOptions,
            // this is a lazy string, because constructing a string is only necessary in the case
            // of an error, but eats up memory and allocation time every time. this was
            // measured.
            context: impl Fn() -> String,
        ) -> ValidationResult;
    }

    /// Whether to do accurate or approximate comparison.
    #[derive(Default, Debug, Eq, PartialEq, Hash, Copy, Clone)]
    pub struct ValidationOptions {
        allow_lossy: bool,
        nan_converted_to_zero: bool,
    }

    /// If invalid, contains the error message.
    pub type ValidationResult = std::result::Result<(), String>;

    impl<C> ValidateResult for Image<C>
    where
        C: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            if self.attributes == other.attributes {
                self.layer_data.validate_result(&other.layer_data, options, || {
                    location() + "| image > layer data"
                })
            } else {
                Err(location() + "| image > attributes")
            }
        }
    }

    impl<S> ValidateResult for Layer<AnyChannels<S>>
    where
        AnyChannel<S>: ValidateResult,
        S: for<'a> WritableSamples<'a>,
    {
        fn validate_result(
            &self,
            other: &Self,
            _overridden: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            let location = || format!("{} (layer `{:?}`)", location(), self.attributes.layer_name);
            if self.attributes != other.attributes {
                Err(location() + " > attributes")
            } else if self.encoding != other.encoding {
                Err(location() + " > encoding")
            } else if self.size != other.size {
                Err(location() + " > size")
            } else if self.channel_data.list.len() != other.channel_data.list.len() {
                Err(location() + " > channel count")
            } else {
                for (own_chan, other_chan) in
                    self.channel_data.list.iter().zip(other.channel_data.list.iter())
                {
                    own_chan.validate_result(
                        other_chan,
                        ValidationOptions {
                            // no tolerance for lossless channels
                            allow_lossy: other
                                .encoding
                                .compression
                                .is_lossless_for(other_chan.sample_data.sample_type())
                                .not(),

                            // consider nan and zero equal if the compression method does not
                            // support nan
                            nan_converted_to_zero: other.encoding.compression.supports_nan().not(),
                        },
                        || format!("{} > channel `{}`", location(), own_chan.name),
                    )?;
                }
                Ok(())
            }
        }
    }

    impl<Px, Desc> ValidateResult for Layer<SpecificChannels<Px, Desc>>
    where
        SpecificChannels<Px, Desc>: ValidateResult,
    {
        /// This does an approximate comparison for all channels,
        /// even if some channels can be compressed without loss.
        fn validate_result(
            &self,
            other: &Self,
            _overridden: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            let location = || format!("{} (layer `{:?}`)", location(), self.attributes.layer_name);

            // TODO dedup with above
            if self.attributes != other.attributes {
                Err(location() + " > attributes")
            } else if self.encoding != other.encoding {
                Err(location() + " > encoding")
            } else if self.size != other.size {
                Err(location() + " > size")
            } else {
                let options = ValidationOptions {
                    // no tolerance for lossless channels
                    // pxr only looses data for f32 values, B44 only for f16, not other any other
                    // types
                    allow_lossy: other.encoding.compression.may_loose_data(), /* TODO check
                                                                               * specific channels
                                                                               * sample types */

                    // consider nan and zero equal if the compression method does not support nan
                    nan_converted_to_zero: other.encoding.compression.supports_nan().not(),
                };

                self.channel_data.validate_result(&other.channel_data, options, || {
                    location() + " > channel_data"
                })?;
                Ok(())
            }
        }
    }

    impl<S> ValidateResult for AnyChannels<S>
    where
        S: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            self.list.validate_result(&other.list, options, location)
        }
    }

    impl<S> ValidateResult for AnyChannel<S>
    where
        S: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            if self.name != other.name {
                Err(location() + " > name")
            } else if self.quantize_linearly != other.quantize_linearly {
                Err(location() + " > quantize_linearly")
            } else if self.sampling != other.sampling {
                Err(location() + " > sampling")
            } else {
                self.sample_data
                    .validate_result(&other.sample_data, options, || location() + " > sample_data")
            }
        }
    }

    impl<Pxs, Chans> ValidateResult for SpecificChannels<Pxs, Chans>
    where
        Pxs: ValidateResult,
        Chans: Eq,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            if self.channels == other.channels {
                self.pixels
                    .validate_result(&other.pixels, options, || location() + " > specific pixels")
            } else {
                Err(location() + " > specific channels")
            }
        }
    }

    impl<S> ValidateResult for Levels<S>
    where
        S: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            self.levels_as_slice()
                .validate_result(&other.levels_as_slice(), options, || location() + " > levels")
        }
    }

    impl ValidateResult for FlatSamples {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            use FlatSamples::*;
            match (self, other) {
                (F16(values), F16(other_values)) => {
                    values.as_slice().validate_result(&other_values.as_slice(), options, || {
                        location() + " > f16 samples"
                    })
                }
                (F32(values), F32(other_values)) => {
                    values.as_slice().validate_result(&other_values.as_slice(), options, || {
                        location() + " > f32 samples"
                    })
                }
                (U32(values), U32(other_values)) => {
                    values.as_slice().validate_result(&other_values.as_slice(), options, || {
                        location() + " > u32 samples"
                    })
                }
                (own, other) => Err(format!(
                    "{}: samples type mismatch. expected {:?}, found {:?}",
                    location(),
                    own.sample_type(),
                    other.sample_type()
                )),
            }
        }
    }

    impl<T> ValidateResult for &[T]
    where
        T: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            if self.len() == other.len() {
                for (index, (slf, other)) in self.iter().zip(other.iter()).enumerate() {
                    slf.validate_result(other, options, || {
                        format!("{} element [{}] of {}", location(), index, self.len())
                    })?;
                }
                Ok(())
            } else {
                Err(location() + " count")
            }
        }
    }

    impl<A: Array> ValidateResult for SmallVec<A>
    where
        A::Item: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            self.as_slice().validate_result(&other.as_slice(), options, location)
        }
    }

    impl<A> ValidateResult for Vec<A>
    where
        A: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            self.as_slice().validate_result(&other.as_slice(), options, location)
        }
    }

    impl<A, B, C, D> ValidateResult for (A, B, C, D)
    where
        A: Clone + ValidateResult,
        B: Clone + ValidateResult,
        C: Clone + ValidateResult,
        D: Clone + ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            self.clone().into_recursive().validate_result(
                &other.clone().into_recursive(),
                options,
                location,
            )
        }
    }

    impl<A, B, C> ValidateResult for (A, B, C)
    where
        A: Clone + ValidateResult,
        B: Clone + ValidateResult,
        C: Clone + ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            self.clone().into_recursive().validate_result(
                &other.clone().into_recursive(),
                options,
                location,
            )
        }
    }

    // // (low priority because it is only used in the tests)
    // TODO
    // impl<Tuple> SimilarToLossy for Tuple where
    // Tuple: Clone + IntoRecursive,
    // <Tuple as IntoRecursive>::Recursive: SimilarToLossy,
    // {
    // fn similar_to_lossy(&self, other: &Self, max_difference: f32) -> bool {
    // self.clone().into_recursive().similar_to_lossy(&other.clone().
    // into_recursive(), max_difference) } // TODO no clone?
    // }

    // implement for recursive types
    impl ValidateResult for NoneMore {
        fn validate_result(
            &self,
            _: &Self,
            _: ValidationOptions,
            _: impl Fn() -> String,
        ) -> ValidationResult {
            Ok(())
        }
    }

    impl<Inner, T> ValidateResult for Recursive<Inner, T>
    where
        Inner: ValidateResult,
        T: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            self.value
                .validate_result(&other.value, options, &location)
                .and_then(|()| self.inner.validate_result(&other.inner, options, &location))
        }
    }

    impl<S> ValidateResult for Option<S>
    where
        S: ValidateResult,
    {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            match (self, other) {
                (None, None) => Ok(()),
                (Some(value), Some(other)) => value.validate_result(other, options, location),
                _ => Err(location() + ": option mismatch"),
            }
        }
    }

    impl ValidateResult for f32 {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            if self == other
                || (self.is_nan() && other.is_nan())
                || (options.nan_converted_to_zero && !self.is_normal() && *other == 0.0)
            {
                return Ok(());
            }

            if options.allow_lossy {
                let epsilon = 0.06;
                let max_difference = 0.1;

                let adaptive_threshold = epsilon * (self.abs() + other.abs());
                let tolerance = adaptive_threshold.max(max_difference);
                let difference = (self - other).abs();

                return if difference <= tolerance {
                    Ok(())
                } else {
                    Err(format!(
                        "{}: expected ~{}, found {} (adaptive tolerance {})",
                        location(),
                        self,
                        other,
                        tolerance
                    ))
                };
            }

            Err(format!("{}: expected exactly {}, found {}", location(), self, other))
        }
    }

    impl ValidateResult for f16 {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            if self.to_bits() == other.to_bits() {
                Ok(())
            } else {
                self.to_f32().validate_result(&other.to_f32(), options, location)
            }
        }
    }

    impl ValidateResult for u32 {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            if self == other {
                Ok(())
            } else {
                // todo to float conversion resulting in nan/infinity?
                self.to_f32().validate_result(&other.to_f32(), options, location)
            }
        }
    }

    impl ValidateResult for Sample {
        fn validate_result(
            &self,
            other: &Self,
            options: ValidationOptions,
            location: impl Fn() -> String,
        ) -> ValidationResult {
            use Sample::*;
            match (self, other) {
                (F16(a), F16(b)) => a.validate_result(b, options, || location() + " (f16)"),
                (F32(a), F32(b)) => a.validate_result(b, options, || location() + " (f32)"),
                (U32(a), U32(b)) => a.validate_result(b, options, || location() + " (u32)"),
                (_, _) => Err(location() + ": sample type mismatch"),
            }
        }
    }

    #[cfg(test)]
    mod test_value_result {
        use std::{f32::consts::*, io::Cursor};

        use crate::{
            image::{
                pixel_vec::PixelVec,
                validate_results::{ValidateResult, ValidationOptions},
                FlatSamples,
            },
            meta::attribute::LineOrder::Increasing,
        };

        fn expect_valid<T>(original: &T, result: &T, allow_lossy: bool, nan_converted_to_zero: bool)
        where
            T: ValidateResult,
        {
            original
                .validate_result(
                    result,
                    ValidationOptions {
                        allow_lossy,
                        nan_converted_to_zero,
                    },
                    String::new,
                )
                .unwrap();
        }

        fn expect_invalid<T>(
            original: &T,
            result: &T,
            allow_lossy: bool,
            nan_converted_to_zero: bool,
        ) where
            T: ValidateResult,
        {
            assert!(original
                .validate_result(
                    result,
                    ValidationOptions {
                        allow_lossy,
                        nan_converted_to_zero
                    },
                    String::new
                )
                .is_err());
        }

        #[test]
        fn test_f32() {
            let original: &[f32] = &[0.0, 0.1, 0.2, 0.3, 0.4, 0.5, -20.4, f32::NAN];
            let lossy: &[f32] = &[0.0, 0.2, 0.2, 0.3, 0.4, 0.5, -20.5, f32::NAN];

            expect_valid(&original, &original, true, true);
            expect_valid(&original, &original, true, false);
            expect_valid(&original, &original, false, true);
            expect_valid(&original, &original, false, false);

            expect_invalid(&original, &lossy, false, false);
            expect_valid(&original, &lossy, true, false);

            expect_invalid(&original, &&original[..original.len() - 2], true, true);

            // test relative comparison with some large values
            expect_valid(&1_000_f32, &1_001_f32, true, false);
            expect_invalid(&1_000_f32, &1_200_f32, true, false);

            expect_valid(&10_000_f32, &10_100_f32, true, false);
            expect_invalid(&10_000_f32, &12_000_f32, true, false);

            expect_valid(&33_120_f32, &30_120_f32, true, false);
            expect_invalid(&33_120_f32, &20_120_f32, true, false);
        }

        #[test]
        fn test_nan() {
            let original: &[f32] = &[0.0, f32::NAN, f32::NAN];
            let lossy: &[f32] = &[0.0, f32::NAN, 0.0];

            expect_valid(&original, &lossy, true, true);
            expect_invalid(&lossy, &original, true, true);

            expect_valid(&lossy, &lossy, true, true);
            expect_valid(&lossy, &lossy, false, true);
        }

        #[test]
        fn test_error() {
            fn print_error<T: ValidateResult>(original: &T, lossy: &T, allow_lossy: bool) {
                let message = original
                    .validate_result(
                        lossy,
                        ValidationOptions {
                            allow_lossy,
                            ..Default::default()
                        },
                        String::new, // type_name::<T>().to_string()
                    )
                    .unwrap_err();

                println!("message: {message}");
            }

            let original: &[f32] = &[0.0, f32::NAN, f32::NAN];
            let lossy: &[f32] = &[0.0, f32::NAN, 0.0];
            print_error(&original, &lossy, false);

            print_error(&2.0, &1.0, true);
            print_error(&2.0, &1.0, false);

            print_error(
                &FlatSamples::F32(vec![0.1, 0.1]),
                &FlatSamples::F32(vec![0.1, 0.2]),
                false,
            );
            print_error(&FlatSamples::U32(vec![0, 0]), &FlatSamples::F32(vec![0.1, 0.2]), false);

            {
                let image = crate::prelude::read_all_data_from_file(
                    "tests/images/valid/openexr/MultiResolution/Kapaa.exr",
                )
                .unwrap();

                let mut mutated = image.clone();
                let samples = mutated
                    .layer_data
                    .first_mut()
                    .unwrap()
                    .channel_data
                    .list
                    .first_mut()
                    .unwrap()
                    .sample_data
                    .levels_as_slice_mut()
                    .first_mut()
                    .unwrap();

                match samples {
                    FlatSamples::F16(vals) => vals[100] = vals[1],
                    FlatSamples::F32(vals) => vals[100] = vals[1],
                    FlatSamples::U32(vals) => vals[100] = vals[1],
                }

                print_error(&image, &mutated, false);
            }

            // TODO check out more nested behaviour!
        }

        #[test]
        fn test_uncompressed() {
            use crate::prelude::*;

            let original_pixels: [(f32, f32, f32); 4] = [
                (0.0, -1.1, PI),
                (0.0, -1.1, TAU),
                (0.0, -1.1, f32::EPSILON),
                (f32::NAN, 10000.1, -1024.009),
            ];

            let mut file_bytes = Vec::new();
            let original_image = Image::from_encoded_channels(
                (2, 2),
                Encoding {
                    compression: Compression::Uncompressed,
                    line_order: Increasing, /* FIXME unspecified may be optimized to increasing,
                                             * which destroys test eq */
                    ..Encoding::default()
                },
                SpecificChannels::rgb(PixelVec::new(Vec2(2, 2), original_pixels.to_vec())),
            );

            original_image.write().to_buffered(Cursor::new(&mut file_bytes)).unwrap();

            let lossy_image = read()
                .no_deep_data()
                .largest_resolution_level()
                .rgb_channels(PixelVec::<(f32, f32, f32)>::constructor, PixelVec::set_pixel)
                .first_valid_layer()
                .all_attributes()
                .from_buffered(Cursor::new(&file_bytes))
                .unwrap();

            original_image.assert_equals_result(&original_image);
            lossy_image.assert_equals_result(&lossy_image);
            original_image.assert_equals_result(&lossy_image);
            lossy_image.assert_equals_result(&original_image);
        }

        #[test]
        fn test_compiles() {
            use crate::prelude::*;

            fn accepts_validatable_value(_: &impl ValidateResult) {}

            let object: Levels<FlatSamples> = Levels::Singular(FlatSamples::F32(Vec::default()));
            accepts_validatable_value(&object);

            let object: AnyChannels<Levels<FlatSamples>> = AnyChannels::sort(SmallVec::default());
            accepts_validatable_value(&object);

            let layer: Layer<AnyChannels<Levels<FlatSamples>>> =
                Layer::new((0, 0), Default::default(), Default::default(), object);
            accepts_validatable_value(&layer);

            let layers: Layers<AnyChannels<Levels<FlatSamples>>> = Default::default();
            accepts_validatable_value(&layers);

            let object: Image<Layer<AnyChannels<Levels<FlatSamples>>>> = Image::from_layer(layer);
            object.assert_equals_result(&object);
        }
    }

    #[test]
    fn test_nan_compression_attribute() {
        use std::io::Cursor;

        use crate::{
            image::pixel_vec::PixelVec,
            prelude::{Compression::*, LineOrder::Increasing, *},
        };

        let all_compression_methods = [Uncompressed, RLE, ZIP1, ZIP16, PXR24, PIZ, B44, B44A];

        let original_pixels: [(f32, f32, f16); 4] = [
            (f32::NAN, f32::from_bits(0x7fc01234), f16::from_bits(0x7E01)),
            (f32::NAN, f32::from_bits(0xffcabcde), f16::from_bits(0x7FFF)),
            (f32::NAN, f32::from_bits(0x7f800001), f16::from_bits(0xFE01)),
            (f32::NAN, f32::NAN, f16::NAN),
        ];

        assert!(
            original_pixels.iter().all(|&(a, b, c)| a.is_nan() && b.is_nan() && c.is_nan()),
            "test case has a bug"
        );

        for compression in all_compression_methods {
            let mut file_bytes = Vec::new();

            let original_image = Image::from_encoded_channels(
                (2, 2),
                Encoding {
                    compression,
                    line_order: Increasing,
                    ..Encoding::default()
                },
                SpecificChannels::rgb(PixelVec::new((2, 2), original_pixels.to_vec())),
            );

            let result = original_image.write().to_buffered(Cursor::new(&mut file_bytes));
            if let Err(Error::NotSupported(_)) = result {
                continue;
            }

            let reconstructed_image = read()
                .no_deep_data()
                .largest_resolution_level()
                .rgb_channels(PixelVec::<(f32, f32, f16)>::constructor, PixelVec::set_pixel)
                .first_valid_layer()
                .all_attributes()
                .from_buffered(Cursor::new(&file_bytes))
                .unwrap();

            assert_eq!(
                original_image.layer_data.channel_data.pixels.pixels.len(),
                reconstructed_image.layer_data.channel_data.pixels.pixels.len()
            );

            let was_nanness_preserved = reconstructed_image
                .layer_data
                .channel_data
                .pixels
                .pixels
                .iter()
                .all(|(r, g, b)| r.is_nan() && g.is_nan() && b.is_nan());

            assert_eq!(
                was_nanness_preserved,
                compression.supports_nan(),
                "{compression} nanness claims do not match real output"
            );

            let was_nan_pattern_preserved = reconstructed_image
                .layer_data
                .channel_data
                .pixels
                .pixels
                .iter()
                .zip(original_pixels.iter())
                .all(|((r2, g2, b2), (r1, g1, b1))| {
                    r2.to_bits() == r1.to_bits()
                        && g2.to_bits() == g1.to_bits()
                        && b2.to_bits() == b1.to_bits()
                });

            assert_eq!(
                was_nan_pattern_preserved,
                compression.preserves_nan_bits(),
                "{compression} nan bit claims do not match real output"
            );
        }
    }
}