//! What one frame of `UiRenderNode::draw` costs on the CPU, against the number
//! of layers it walks. Recording only: the pass is built and dropped without
//! being submitted, so this is the loop's cost and not the GPU's.
//!
//!     cargo test --release --test draw_cost -- --ignored --nocapture
//!
//! Wall time is the wrong number to read for anything under a few percent --
//! it varied by 2x between runs of one unchanged binary where instructions
//! retired varied by 0.1%. Count those instead:
//!
//!     perf stat -e instructions:u target/release/.../draw_cost-* --ignored
//!
//! That is how `PrimitiveRender` was measured against a match in the renderer:
//! 6 instructions per list drawn, against the ~5,400 wgpu spends recording
//! one.
//!
//! The instance is leaked deliberately. A Vulkan loader may unload the driver
//! when the last one drops, which can fault as a thread that used it exits --
//! and every test runs on a spawned thread.

use std::time::Instant;

use iris::prelude::*;
use iris_core::{
    GlyphPrimitive, MaskIdx, MoveIdx, PrimitiveInst, RectPrimitive, TextureHandle,
    TexturePrimitive, UiData, UiRegion, UiRenderNode, UiRenderState,
};
use wgpu::{Color as GpuColor, *};

const SIZE: u32 = 1024;
const FRAMES: u32 = 200;
/// Reported as the best of this many batches, since the mean moves by more
/// than the thing being measured.
const BATCHES: u32 = 8;

fn gpu() -> Option<(Device, Queue)> {
    // Probed rather than assumed: there may be no Vulkan adapter, and GL is
    // what is left when there is not.
    let all = Instance::new(InstanceDescriptor::new_without_display_handle());
    let instance = match pollster::block_on(all.request_adapter(&RequestAdapterOptions::default()))
    {
        Ok(_) => all,
        Err(_) => Instance::new(InstanceDescriptor {
            backends: Backends::GL,
            ..InstanceDescriptor::new_without_display_handle()
        }),
    };
    // Leaked rather than dropped: see the note at the top of the file.
    let instance: &'static Instance = Box::leak(Box::new(instance));
    let adapter =
        pollster::block_on(instance.request_adapter(&RequestAdapterOptions::default())).ok()?;
    println!("adapter: {:?}", adapter.get_info());
    pollster::block_on(adapter.request_device(&DeviceDescriptor::default())).ok()
}

fn config(format: TextureFormat) -> SurfaceConfiguration {
    SurfaceConfiguration {
        usage: TextureUsages::RENDER_ATTACHMENT,
        format,
        color_space: SurfaceColorSpace::Auto,
        width: SIZE,
        height: SIZE,
        present_mode: PresentMode::Fifo,
        desired_maximum_frame_latency: 2,
        alpha_mode: CompositeAlphaMode::Auto,
        view_formats: vec![],
    }
}

/// Every layer draws all three primitives, so the renderer takes a different
/// path for each list it walks -- which is the case a single-primitive layer
/// would never exercise. Images are bound per instance, so there are few.
fn fill(
    ui: &mut UiData,
    render: &mut UiRenderState,
    layers: usize,
    per_layer: usize,
) -> Vec<TextureHandle> {
    let rect = ui.primitives.kind::<RectPrimitive>();
    let glyph = ui.primitives.kind::<GlyphPrimitive>();
    let texture = ui.primitives.kind::<TexturePrimitive>();
    let id = ui.widgets.add_strong(Rect::new(UiColor::WHITE)).id();
    let handles: Vec<_> = (0..4)
        .map(|_| ui.textures.add(image::RgbaImage::new(4, 4)))
        .collect();

    let mut layer = 0;
    for _ in 0..layers {
        for _ in 0..per_layer {
            render.layers.write(
                layer,
                PrimitiveInst {
                    kind: rect,
                    id,
                    primitive: RectPrimitive::color(UiColor::WHITE),
                    region: UiRegion::FULL,
                    mask_idx: MaskIdx::NONE,
                    move_idx: MoveIdx::NONE,
                },
            );
            render.layers.write(
                layer,
                PrimitiveInst {
                    kind: glyph,
                    id,
                    primitive: GlyphPrimitive {
                        uv_min: vec2(0.0, 0.0),
                        uv_max: vec2(1.0, 1.0),
                        layer: 0,
                        color: UiColor::WHITE,
                        flags: 0,
                    },
                    region: UiRegion::FULL,
                    mask_idx: MaskIdx::NONE,
                    move_idx: MoveIdx::NONE,
                },
            );
        }
        for h in &handles[..2] {
            render.layers.write(
                layer,
                PrimitiveInst {
                    kind: texture,
                    id,
                    primitive: TexturePrimitive::from(h),
                    region: UiRegion::FULL,
                    mask_idx: MaskIdx::NONE,
                    move_idx: MoveIdx::NONE,
                },
            );
        }
        layer = render.layers.next(layer);
    }
    handles
}

fn frame_cost(device: &Device, queue: &Queue, layers: usize, per_layer: usize) -> f64 {
    let format = TextureFormat::Bgra8Unorm;
    let mut node = UiRenderNode::new(device, &config(format));
    let mut ui = UiData::default();
    let mut render = UiRenderState::new();
    let _handles = fill(&mut ui, &mut render, layers, per_layer);
    node.update(device, queue, &mut ui, &mut render);

    let target = device.create_texture(&TextureDescriptor {
        label: Some("draw cost"),
        size: Extent3d {
            width: SIZE,
            height: SIZE,
            depth_or_array_layers: 1,
        },
        mip_level_count: 1,
        sample_count: 1,
        dimension: TextureDimension::D2,
        format,
        usage: TextureUsages::RENDER_ATTACHMENT,
        view_formats: &[],
    });
    let view = target.create_view(&TextureViewDescriptor::default());

    let record = |frames: u32| {
        let start = Instant::now();
        for _ in 0..frames {
            let mut encoder = device.create_command_encoder(&CommandEncoderDescriptor::default());
            {
                let pass = &mut encoder.begin_render_pass(&RenderPassDescriptor {
                    color_attachments: &[Some(RenderPassColorAttachment {
                        view: &view,
                        resolve_target: None,
                        ops: Operations {
                            load: LoadOp::Clear(GpuColor::BLACK),
                            store: StoreOp::Store,
                        },
                        depth_slice: None,
                    })],
                    ..Default::default()
                });
                node.draw(pass);
            }
            drop(encoder.finish());
        }
        start.elapsed().as_secs_f64() / frames as f64
    };
    record(FRAMES / 4);
    (0..BATCHES)
        .map(|_| record(FRAMES))
        .fold(f64::MAX, f64::min)
}

#[test]
#[ignore = "measurement, not a check"]
fn draw_cost_by_layer_count() {
    let Some((device, queue)) = gpu() else {
        panic!("no wgpu device; see the this-machine-graphics notes");
    };
    println!(
        "layers, each 8 rects + 8 glyphs + 2 images: us/frame (us per layer), best of {BATCHES}"
    );
    let base = frame_cost(&device, &queue, 1, 8) * 1e6;
    for layers in [8, 64, 256, 1024] {
        let per_frame = frame_cost(&device, &queue, layers, 8) * 1e6;
        // Net of the empty pass, which is the same in any version of this.
        println!(
            "{layers:>5}: {per_frame:8.1} us ({:.3} us)",
            (per_frame - base).max(0.0) / layers as f64
        );
    }
}