concinnity-device 0.19.119

//! Projected (deferred) decals for the Vulkan backend. Each decal is drawn
//! as a unit cube (positions in `[-0.5, 0.5]^3`) transformed by its world
//! model matrix and the camera VP; the fragment shader samples the main
//! pass's depth attachment to reconstruct the world-space sample point at
//! each pixel and tests it against the decal's local bounding box,
//! stamping the texture onto whatever fills the box.
//!
//! Runs after the main HDR resolve and before SSR resolve / TAA, so
//! decals are reflected and tracked by the temporal history just like
//! the rest of the scene. Mirrors `src/directx/decal.rs` and
//! `src/metal/decal.rs`.

use ash::vk;
use concinnity_core::gfx::frustum::Frustum;
use concinnity_core::render::decal::DecalRecord;
use concinnity_core::render::error::{RenderError, RenderResult};
use concinnity_core::transform::mat4_inverse;
use std::cell::Cell;
// `DecalView` (per-frame, 144 bytes) is the layout struct shared with the other
// backends; the per-decal `DecalParams` (160 bytes, inside the 256-byte stride
// slot) rides the set's slots. Both mirror `shaders/decal.hlsl`.
use concinnity_core::render::uniforms::DecalView;

use super::allocator::{DeviceAllocator, PooledBuffer};
use super::context::VkContext;
use super::descriptor_layout::{Binding, PoolSizes};
use super::pipeline_desc::{Blend, GraphicsPipelineDesc, Raster};
use super::resources::{alloc_descriptor_sets, create_descriptor_set_layout, source_set_bindings};
use super::set_writes::SetWrites;
use super::texture::GpuImage;
use crate::vulkan::builtin_shaders::CompileProgram;
use crate::vulkan::owned::{
    OwnedDescriptorPool, OwnedFramebuffer, OwnedPipeline, OwnedPipelineLayout, OwnedRenderPass,
    OwnedSetLayout, VkDevice,
};

// Cap on the number of active decals: the descriptor pool reserves a
// fixed block of `MAX_DECALS` per-decal albedo sets at init, so runtime
// adds past this many return an error.
pub(in crate::vulkan) const MAX_DECALS: usize = 256;

// Eight unit-cube corners in [-0.5, 0.5]^3. Matches the DirectX / Metal
// vertex lists.
const CUBE_VERTS: [f32; 24] = [
    -0.5, -0.5, -0.5, 0.5, -0.5, -0.5, 0.5, 0.5, -0.5, -0.5, 0.5, -0.5, -0.5, -0.5, 0.5, 0.5, -0.5,
    0.5, 0.5, 0.5, 0.5, -0.5, 0.5, 0.5,
];

// 36 indices forming 12 triangles wound CCW outward. Matches the
// DirectX / Metal index list so the rasterized cube exactly mirrors the
// reference.
const CUBE_INDICES: [u16; 36] = [
    // -Z face                +Z face
    0, 2, 1, 0, 3, 2, 4, 5, 6, 4, 6, 7, // -Y                     +Y
    0, 1, 5, 0, 5, 4, 3, 6, 2, 3, 7, 6, // -X                     +X
    0, 4, 7, 0, 7, 3, 1, 2, 6, 1, 6, 5,
];

// Stride for the per-decal params uniform buffer ring. Vulkan's
// `minUniformBufferOffsetAlignment` is at most 256 bytes on every
// desktop GPU we target (spec-guaranteed upper bound), so a constant
// 256 byte stride keeps each dynamic-offset slot naturally aligned
// without querying the device.
const PARAMS_STRIDE: u64 = 256;

// Owned by `VkContext` exactly once: the decal pipeline + its dependent
// resources. The decal slot table lives on `VkContext` itself (mirroring the
// DirectX / Metal layout).
//
// Decal-pass descriptor sets follow a two-set layout:
//   * **set 0** (per-frame, FRAMES sets):
//       - binding 0: UNIFORM_BUFFER, `DecalView` (per-frame)
//       - binding 1: UNIFORM_BUFFER_DYNAMIC, `DecalParams` ring (per-frame,
//         MAX_DECALS slots; dynamic offset picks the per-decal slot)
//       - binding 2: SAMPLED_IMAGE, main depth view, read by texel (per-frame
//         so a future-frame depth swap doesn't break a binding allocated
//         from a sibling frame slot)
//   * **set 1** (per-decal, MAX_DECALS sets):
//       - binding 0: SAMPLED_IMAGE, decal albedo
//       - binding 1: SAMPLER, the albedo's sampler
pub(in crate::vulkan) struct DecalResources {
    pub(in crate::vulkan) render_pass: OwnedRenderPass,
    pub(in crate::vulkan) pipeline: OwnedPipeline,
    pub(in crate::vulkan) pipeline_layout: OwnedPipelineLayout,
    pub(in crate::vulkan) _view_set_layout: OwnedSetLayout,
    pub(in crate::vulkan) _albedo_set_layout: OwnedSetLayout,
    pub(in crate::vulkan) _descriptor_pool: OwnedDescriptorPool,

    // Unit-cube vertex + index buffers (shared across frames).
    pub(in crate::vulkan) vertex_buffer: PooledBuffer,
    pub(in crate::vulkan) index_buffer: PooledBuffer,

    // Per-frame view UBO (DecalView, 144 bytes). Persistently mapped.
    pub(in crate::vulkan) view_ubos: Vec<PooledBuffer>,

    // Per-frame per-decal params ring (PARAMS_STRIDE * MAX_DECALS bytes).
    // Persistently mapped; bound as UNIFORM_BUFFER_DYNAMIC with a
    // per-draw offset of `decal_id * PARAMS_STRIDE`.
    pub(in crate::vulkan) params_ubos: Vec<PooledBuffer>,

    // Per-frame view sets (binding 0 view UBO, 1 params dynamic, 2 depth).
    pub(in crate::vulkan) view_sets: Vec<vk::DescriptorSet>,
    // Per-decal albedo sets (binding 0 albedo, 1 its sampler). Indexed by
    // `decal_id`; tombstoned slots keep their last write, but the
    // executor only binds the set for visible records.
    pub(in crate::vulkan) albedo_sets: Vec<vk::DescriptorSet>,

    // One framebuffer per frame-in-flight slot, each binding its frame
    // slot's `hdr_resolve_images[i].view` as the sole color attachment.
    pub(in crate::vulkan) framebuffers: Vec<OwnedFramebuffer>,

    // Last-uploaded texture-pool slot per decal id. Used by
    // `rewrite_texture_slot` to detect which decal albedo sets need
    // re-writing when a streamed texture pool entry swaps.
    pub(in crate::vulkan) decal_texture_slots: Cell<[usize; MAX_DECALS]>,
}

// Vulkan handles needed to create the decal pass's GPU resources
// (buffers, one-shot transfer submits). Borrowed for the duration of
// `DecalResources::new`.
#[derive(Clone, Copy)]
pub(in crate::vulkan) struct DecalDeviceContext<'a> {
    pub(in crate::vulkan) alloc: &'a DeviceAllocator,
    pub(in crate::vulkan) device: &'a VkDevice,
    pub(in crate::vulkan) command_pool: vk::CommandPool,
    pub(in crate::vulkan) queue: vk::Queue,
}

// Render-target inputs the decal pass writes into / samples from: the
// resolved HDR color attachment (format + per-frame views), the main
// depth views, the albedo sampler, and the framebuffer extent.
#[derive(Clone, Copy)]
pub(in crate::vulkan) struct DecalPassTargets<'a> {
    pub(in crate::vulkan) hdr_format: vk::Format,
    pub(in crate::vulkan) hdr_resolve_views: &'a [vk::ImageView],
    pub(in crate::vulkan) depth_views: &'a [vk::ImageView],
    pub(in crate::vulkan) sampler: vk::Sampler,
    pub(in crate::vulkan) extent: vk::Extent2D,
}

impl DecalResources {
    // Build the decal pipeline + its dependent resources. Called
    // unconditionally from `VkContext::new` so runtime `add_decal`
    // works from a world that started empty; the cost is one pipeline
    // + small buffers.
    pub(in crate::vulkan) fn new(
        ctx: DecalDeviceContext,
        targets: DecalPassTargets,
        frames: usize,
        msaa: bool,
        hot_reload: bool,
    ) -> RenderResult<Self> {
        let DecalDeviceContext {
            alloc,
            device,
            command_pool,
            queue,
        } = ctx;
        let DecalPassTargets {
            hdr_format,
            hdr_resolve_views,
            depth_views,
            sampler,
            extent,
        } = targets;
        let render_pass = create_decal_render_pass(device, hdr_format)?;
        let view_set_layout = create_descriptor_set_layout(device, &view_set_bindings())?;
        let albedo_set_layout = create_descriptor_set_layout(device, &source_set_bindings(1))?;
        let pipeline_layout = create_decal_pipeline_layout(
            device,
            view_set_layout.handle(),
            albedo_set_layout.handle(),
        )?;

        let (vert_spv, frag_spv) = compile_decal_shaders(hot_reload, msaa)?;
        let pipeline = create_decal_pipeline(
            device,
            render_pass.handle(),
            pipeline_layout.handle(),
            &vert_spv,
            &frag_spv,
        )?;

        // Unit-cube vertex + index buffers (single device-local upload).
        let vertex_buffer = upload_static_buffer(
            alloc,
            device,
            command_pool,
            queue,
            bytemuck_cast(&CUBE_VERTS),
            vk::BufferUsageFlags::VERTEX_BUFFER,
        )?;
        let index_buffer = upload_static_buffer(
            alloc,
            device,
            command_pool,
            queue,
            bytemuck_cast(&CUBE_INDICES),
            vk::BufferUsageFlags::INDEX_BUFFER,
        )?;

        // Per-frame view UBOs (HOST_VISIBLE | HOST_COHERENT, persistently
        // mapped).
        let mut view_ubos = Vec::with_capacity(frames);
        for _ in 0..frames {
            view_ubos.push(alloc.create_buffer(
                std::mem::size_of::<DecalView>() as u64,
                vk::BufferUsageFlags::UNIFORM_BUFFER,
                vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT,
            )?);
        }

        // Per-frame per-decal params ring.
        let params_total = PARAMS_STRIDE * MAX_DECALS as u64;
        let mut params_ubos = Vec::with_capacity(frames);
        for _ in 0..frames {
            params_ubos.push(alloc.create_buffer(
                params_total,
                vk::BufferUsageFlags::UNIFORM_BUFFER,
                vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT,
            )?);
        }

        let descriptor_pool = create_decal_descriptor_pool(device, frames)?;

        // Per-frame view sets (one per frame slot).
        let view_layouts: Vec<_> = (0..frames).map(|_| view_set_layout.handle()).collect();
        let view_sets = alloc_descriptor_sets(device, descriptor_pool.handle(), &view_layouts)?;
        for (i, &set) in view_sets.iter().enumerate() {
            // The dynamic params binding windows one PARAMS_STRIDE slot at the
            // offset given at bind time.
            SetWrites::new(set)
                .uniform_buffer(
                    0,
                    view_ubos[i].buffer(),
                    std::mem::size_of::<DecalView>() as u64,
                )
                .buffer(
                    1,
                    vk::DescriptorType::UNIFORM_BUFFER_DYNAMIC,
                    params_ubos[i].buffer(),
                    0,
                    PARAMS_STRIDE,
                )
                .sampled_image(2, depth_views[i.min(depth_views.len().saturating_sub(1))])
                .apply(device);
        }

        // Per-decal albedo sets (MAX_DECALS sets, pre-allocated).
        let albedo_layouts: Vec<_> = (0..MAX_DECALS)
            .map(|_| albedo_set_layout.handle())
            .collect();
        let albedo_sets = alloc_descriptor_sets(device, descriptor_pool.handle(), &albedo_layouts)?;
        // The sampler never changes, so it is written once here; a decal's
        // albedo image is written when the decal takes the slot.
        for &set in &albedo_sets {
            SetWrites::new(set).sampler(1, sampler).apply(device);
        }

        // Per-frame framebuffers (one per frame slot binding that slot's
        // hdr_resolve view as the color attachment).
        let mut framebuffers = Vec::with_capacity(frames);
        for &view in hdr_resolve_views.iter().take(frames) {
            let attachments = [view];
            let fb_info = vk::FramebufferCreateInfo::default()
                .render_pass(render_pass.handle())
                .attachments(&attachments)
                .width(extent.width.max(1))
                .height(extent.height.max(1))
                .layers(1);
            let fb = device
                .create_framebuffer(&fb_info)
                .map_err(|e| super::error::map_vk_result(e, "decal framebuffer"))?;
            framebuffers.push(fb);
        }

        Ok(Self {
            render_pass,
            pipeline,
            pipeline_layout,
            _view_set_layout: view_set_layout,
            _albedo_set_layout: albedo_set_layout,
            _descriptor_pool: descriptor_pool,
            vertex_buffer,
            index_buffer,
            view_ubos,
            params_ubos,
            view_sets,
            albedo_sets,
            framebuffers,
            decal_texture_slots: Cell::new([usize::MAX; MAX_DECALS]),
        })
    }

    // Rebuild the framebuffers + re-point the per-frame view set's depth
    // binding after a swapchain resize. Called from
    // `VkContext::rebuild_swapchain`; same pattern as `SsrResources` /
    // `SsaoResources`. The pipeline, layouts, buffers, and
    // per-decal albedo sets all survive.
    pub(in crate::vulkan) fn rebuild(
        &mut self,
        device: &VkDevice,
        hdr_resolve_views: &[vk::ImageView],
        depth_views: &[vk::ImageView],
        extent: vk::Extent2D,
    ) -> RenderResult<()> {
        self.framebuffers.clear();
        for &view in hdr_resolve_views.iter().take(self.view_ubos.len()) {
            let attachments = [view];
            let fb_info = vk::FramebufferCreateInfo::default()
                .render_pass(self.render_pass.handle())
                .attachments(&attachments)
                .width(extent.width.max(1))
                .height(extent.height.max(1))
                .layers(1);
            let fb = device
                .create_framebuffer(&fb_info)
                .map_err(|e| super::error::map_vk_result(e, "decal framebuffer (rebuild)"))?;
            self.framebuffers.push(fb);
        }
        // Re-point each per-frame view set's depth binding (binding 2)
        // at the rebuilt depth view.
        for (i, &set) in self.view_sets.iter().enumerate() {
            SetWrites::new(set)
                .sampled_image(2, depth_views[i.min(depth_views.len().saturating_sub(1))])
                .apply(device);
        }
        Ok(())
    }

    // Destroy every GPU resource. Called from `VkContext::destroy` after
    // `wait_idle`; the pooled buffers retire through the allocator as their
    // fields clear.
    pub(in crate::vulkan) fn destroy(&mut self, _device: &VkDevice) {
        self.framebuffers.clear();
        self.view_ubos.clear();
        self.params_ubos.clear();
        self.vertex_buffer = PooledBuffer::null();
        self.index_buffer = PooledBuffer::null();
    }
}

// Render pass / pipeline construction

fn create_decal_render_pass(
    device: &VkDevice,
    format: vk::Format,
) -> RenderResult<OwnedRenderPass> {
    // One color attachment: the resolved HDR scene. The main pass left
    // it in SHADER_READ_ONLY_OPTIMAL; we want it in COLOR_ATTACHMENT
    // during the subpass, then SHADER_READ_ONLY_OPTIMAL again on exit so
    // SSR / TAA / bloom / composite can sample it. The subpass
    // dependencies + (initial=SHADER_READ_ONLY, final=SHADER_READ_ONLY)
    // do the round-trip transition without an explicit barrier from the
    // caller.
    let attachment = vk::AttachmentDescription::default()
        .format(format)
        .samples(vk::SampleCountFlags::TYPE_1)
        .load_op(vk::AttachmentLoadOp::LOAD)
        .store_op(vk::AttachmentStoreOp::STORE)
        .stencil_load_op(vk::AttachmentLoadOp::DONT_CARE)
        .stencil_store_op(vk::AttachmentStoreOp::DONT_CARE)
        .initial_layout(vk::ImageLayout::SHADER_READ_ONLY_OPTIMAL)
        .final_layout(vk::ImageLayout::SHADER_READ_ONLY_OPTIMAL);
    let color_ref = vk::AttachmentReference::default()
        .attachment(0)
        .layout(vk::ImageLayout::COLOR_ATTACHMENT_OPTIMAL);
    let subpass = vk::SubpassDescription::default()
        .pipeline_bind_point(vk::PipelineBindPoint::GRAPHICS)
        .color_attachments(std::slice::from_ref(&color_ref));
    // Entry: anyone sampling the resolved HDR (e.g. the main pass's resolve
    // attachment writer, which finished with this image in
    // SHADER_READ_ONLY) must complete before the decal subpass starts
    // writing it.
    let dep_in = vk::SubpassDependency::default()
        .src_subpass(vk::SUBPASS_EXTERNAL)
        .dst_subpass(0)
        .src_stage_mask(
            vk::PipelineStageFlags::COLOR_ATTACHMENT_OUTPUT
                | vk::PipelineStageFlags::FRAGMENT_SHADER,
        )
        .src_access_mask(vk::AccessFlags::SHADER_READ | vk::AccessFlags::COLOR_ATTACHMENT_WRITE)
        .dst_stage_mask(vk::PipelineStageFlags::COLOR_ATTACHMENT_OUTPUT)
        .dst_access_mask(
            vk::AccessFlags::COLOR_ATTACHMENT_WRITE | vk::AccessFlags::COLOR_ATTACHMENT_READ,
        );
    // Exit: any subsequent SSR / TAA / bloom / composite pass reading the
    // resolved HDR must wait for our writes to complete and become
    // available + visible to the fragment shader.
    let dep_out = vk::SubpassDependency::default()
        .src_subpass(0)
        .dst_subpass(vk::SUBPASS_EXTERNAL)
        .src_stage_mask(vk::PipelineStageFlags::COLOR_ATTACHMENT_OUTPUT)
        .src_access_mask(vk::AccessFlags::COLOR_ATTACHMENT_WRITE)
        .dst_stage_mask(vk::PipelineStageFlags::FRAGMENT_SHADER)
        .dst_access_mask(vk::AccessFlags::SHADER_READ);
    let deps = [dep_in, dep_out];
    let info = vk::RenderPassCreateInfo::default()
        .attachments(std::slice::from_ref(&attachment))
        .subpasses(std::slice::from_ref(&subpass))
        .dependencies(&deps);
    device
        .create_render_pass(&info)
        .map_err(|e| super::error::map_vk_result(e, "decal render pass"))
}

// Set 0, per frame: the view UBO, the per-decal params dynamic UBO and the
// main depth.
fn view_set_bindings() -> [Binding; 3] {
    use vk::DescriptorType as T;
    let vert_frag = vk::ShaderStageFlags::VERTEX | vk::ShaderStageFlags::FRAGMENT;
    [
        (0, T::UNIFORM_BUFFER, vert_frag),
        (1, T::UNIFORM_BUFFER_DYNAMIC, vert_frag),
        (2, T::SAMPLED_IMAGE, vk::ShaderStageFlags::FRAGMENT),
    ]
}

fn create_decal_pipeline_layout(
    device: &VkDevice,
    view_set_layout: vk::DescriptorSetLayout,
    albedo_set_layout: vk::DescriptorSetLayout,
) -> RenderResult<OwnedPipelineLayout> {
    let set_layouts = [view_set_layout, albedo_set_layout];
    let info = vk::PipelineLayoutCreateInfo::default().set_layouts(&set_layouts);
    device
        .create_pipeline_layout(&info)
        .map_err(|e| super::error::map_vk_result(e, "decal pipeline layout"))
}

fn create_decal_descriptor_pool(
    device: &VkDevice,
    frames: usize,
) -> RenderResult<OwnedDescriptorPool> {
    let frames = frames as u32;
    let max_decals = MAX_DECALS as u32;
    let sizes = PoolSizes::default()
        .sets(&view_set_bindings(), frames)
        .sets(&source_set_bindings(1), max_decals)
        .build();
    let info = vk::DescriptorPoolCreateInfo::default()
        .max_sets(frames + max_decals)
        .pool_sizes(&sizes);
    device
        .create_descriptor_pool(&info)
        .map_err(|e| super::error::map_vk_result(e, "decal descriptor pool"))
}

fn compile_decal_shaders(hot_reload: bool, msaa: bool) -> RenderResult<(Vec<u8>, Vec<u8>)> {
    let vert = super::builtin_shaders::DECAL_VERT.compile(hot_reload)?;
    let frag = super::builtin_shaders::DECAL_FRAG
        .at(msaa)
        .compile(hot_reload)?;
    Ok((vert, frag))
}

// Rebuild the decal graphics pipeline against the existing render pass +
// layout. Used by the Vulkan shader hot-reload path. The caller is
// responsible for destroying the previous pipeline only after this call
// succeeds.
pub(in crate::vulkan) fn rebuild_decal_pipeline(
    device: &VkDevice,
    decals: &DecalResources,
    msaa: bool,
    hot_reload: bool,
) -> RenderResult<OwnedPipeline> {
    let (vert_spv, frag_spv) = compile_decal_shaders(hot_reload, msaa)?;
    create_decal_pipeline(
        device,
        decals.render_pass.handle(),
        decals.pipeline_layout.handle(),
        &vert_spv,
        &frag_spv,
    )
}

// A unit cube with front faces culled, so the back faces still rasterize when
// the camera is inside the decal volume, alpha-blended into the SINGLE-SAMPLE
// resolved HDR regardless of the main pass's MSAA. Mirrors DirectX / Metal.
fn create_decal_pipeline(
    device: &VkDevice,
    render_pass: vk::RenderPass,
    layout: vk::PipelineLayout,
    vert_spv: &[u8],
    frag_spv: &[u8],
) -> RenderResult<OwnedPipeline> {
    let bindings = [vk::VertexInputBindingDescription::default()
        .binding(0)
        .stride(12) // vec3 position
        .input_rate(vk::VertexInputRate::VERTEX)];
    let attrs = [vk::VertexInputAttributeDescription::default()
        .location(0)
        .binding(0)
        .format(vk::Format::R32G32B32_SFLOAT)
        .offset(0)];
    GraphicsPipelineDesc {
        raster: Raster {
            cull: vk::CullModeFlags::FRONT,
            ..Raster::default()
        },
        vertex_bindings: &bindings,
        vertex_attributes: &attrs,
        ..GraphicsPipelineDesc::fullscreen(
            vert_spv,
            frag_spv,
            layout,
            render_pass,
            &[Blend::AlphaOver],
        )
    }
    .build(device, "decal")
}

// Helpers for upload + casting

fn upload_static_buffer(
    alloc: &DeviceAllocator,
    device: &VkDevice,
    command_pool: vk::CommandPool,
    queue: vk::Queue,
    data: &[u8],
    usage: vk::BufferUsageFlags,
) -> RenderResult<PooledBuffer> {
    let size = data.len() as vk::DeviceSize;
    let staging = alloc.create_buffer(
        size,
        vk::BufferUsageFlags::TRANSFER_SRC,
        vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT,
    )?;
    staging.write_bytes(0, data);
    let buf = alloc.create_buffer(
        size,
        usage | vk::BufferUsageFlags::TRANSFER_DST,
        vk::MemoryPropertyFlags::DEVICE_LOCAL,
    )?;
    super::texture::one_shot_submit(device, command_pool, queue, |cmd| {
        let region = vk::BufferCopy::default().size(size);
        // SAFETY: `cmd` is a command buffer in the recording state, and every handle and slice
        // these commands name is live for the call.
        unsafe {
            device.cmd_copy_buffer(
                cmd,
                staging.buffer(),
                buf.buffer(),
                std::slice::from_ref(&region),
            )
        };
    })?;
    Ok(buf)
}

fn bytemuck_cast<T: bytemuck::NoUninit>(s: &[T]) -> &[u8] {
    bytemuck::cast_slice(s)
}

// Encoder

impl VkContext {
    // Encode the projected-decal pass. Called between the main HDR
    // resolve and the SSR resolve so a decal is reflected by SSR and
    // tracked by TAA's history buffer like the rest of the scene.
    //
    // `vp` is the same jittered view-projection the main pass
    // rasterized with; the inverse drives the world-space reconstruction
    // in the fragment shader.
    pub(in crate::vulkan) fn encode_decals(
        &self,
        cmd: vk::CommandBuffer,
        frame_idx: usize,
        vp: [[f32; 4]; 4],
        frustum: &Frustum,
    ) {
        let decals = match &self.decal.resources {
            Some(s) => s,
            None => return,
        };
        // Frustum-cull first so a frame where every live decal lands
        // off-screen skips the pass, including the depth-transition
        // barriers. Peeking answers that without testing any decal twice.
        let mut visible = self.decal.set.visible(frustum).peekable();
        if visible.peek().is_none() {
            return;
        }

        let device = &self.hw.device;
        let extent = self.targets.render_extent;

        // Upload this frame's view UBO.
        let inv_vp = mat4_inverse(vp);
        let viewport_pix = [extent.width as f32, extent.height as f32];
        let view_uni = DecalView {
            vp,
            inv_vp,
            viewport: viewport_pix,
            _pad: [0.0; 2],
        };
        decals.view_ubos[frame_idx].write_val(0, &view_uni);

        // Main depth is already in SHADER_READ_ONLY for the fragment's sample:
        // the graph declares this pass's depth read and the executor emits the
        // transition ahead of this command buffer.
        let rp_begin = vk::RenderPassBeginInfo::default()
            .render_pass(decals.render_pass.handle())
            .framebuffer(decals.framebuffers[frame_idx].handle())
            .render_area(vk::Rect2D::default().extent(extent));

        // Negative-height viewport matches the main pass so the
        // rasterized pixel grid lines up with the depth attachment we're
        // sampling. Without this, the cube would be drawn Y-flipped
        // relative to the scene depth.
        let vp_state = vk::Viewport {
            x: 0.0,
            y: extent.height as f32,
            width: extent.width as f32,
            height: -(extent.height as f32),
            min_depth: 0.0,
            max_depth: 1.0,
        };
        let scissor = vk::Rect2D::default().extent(extent);

        // SAFETY: `cmd` is a command buffer in the recording state, and every handle and slice
        // these commands name is live for the call.
        unsafe {
            device.cmd_begin_render_pass(cmd, &rp_begin, vk::SubpassContents::INLINE);
            device.cmd_set_viewport(cmd, 0, std::slice::from_ref(&vp_state));
            device.cmd_set_scissor(cmd, 0, std::slice::from_ref(&scissor));
            device.cmd_bind_pipeline(
                cmd,
                vk::PipelineBindPoint::GRAPHICS,
                decals.pipeline.handle(),
            );
            device.cmd_bind_vertex_buffers(cmd, 0, &[decals.vertex_buffer.buffer()], &[0]);
            device.cmd_bind_index_buffer(
                cmd,
                decals.index_buffer.buffer(),
                0,
                vk::IndexType::UINT16,
            );
        }

        for decal in visible {
            // This frame's ring slot keeps what an earlier frame wrote, so a
            // decal whose record has not changed since is already uploaded.
            if decal.take_upload(frame_idx) {
                decals.params_ubos[frame_idx]
                    .write_val(decal.id * PARAMS_STRIDE as usize, decal.params);
            }
            let dynamic_offset = (decal.id as u64 * PARAMS_STRIDE) as u32;
            // SAFETY: `cmd` is a command buffer in the recording state, and every handle and slice
            // these commands name is live for the call.
            unsafe {
                device.cmd_bind_descriptor_sets(
                    cmd,
                    vk::PipelineBindPoint::GRAPHICS,
                    decals.pipeline_layout.handle(),
                    0,
                    std::slice::from_ref(&decals.view_sets[frame_idx]),
                    std::slice::from_ref(&dynamic_offset),
                );
                device.cmd_bind_descriptor_sets(
                    cmd,
                    vk::PipelineBindPoint::GRAPHICS,
                    decals.pipeline_layout.handle(),
                    1,
                    std::slice::from_ref(&decals.albedo_sets[decal.id]),
                    &[],
                );
                device.cmd_draw_indexed(cmd, 36, 1, 0, 0, 0);
            }
            self.inc_draw_calls(1);
        }

        // SAFETY: `cmd` is a command buffer in the recording state, and every handle and slice
        // these commands name is live for the call.
        unsafe {
            device.cmd_end_render_pass(cmd);
        }
    }
}

// Runtime mutation (RenderBackend::add_decal / remove_decal)

impl VkContext {
    // Append a runtime decal. Writes the per-decal albedo descriptor
    // into the reserved slot for `id`; the encoder reads it next frame.
    // Reuses tombstoned slots from a prior `remove_decal` before growing
    // the vec.
    pub(crate) fn add_decal(&mut self, record: DecalRecord) -> RenderResult<usize> {
        let last_tex = self.scene.textures.len().saturating_sub(1);
        let tex_idx = record.texture_slot.min(last_tex);

        let id = self.decal.set.insert(record).map_err(|_| {
            RenderError::Other(format!("add_decal: MAX_DECALS ({MAX_DECALS}) exceeded"))
        })?;

        // Write the albedo descriptor for this slot. The texture pool
        // entry is referenced live; a future eviction routes through
        // `rewrite_texture_slot` to re-point.
        let decals = self.decal.resources.as_ref().ok_or_else(|| {
            RenderError::Other("add_decal: decal pipeline unavailable".to_string())
        })?;
        write_albedo_image(
            &self.hw.device,
            decals.albedo_sets[id],
            self.scene.textures[tex_idx].view,
        );
        let mut slots = decals.decal_texture_slots.get();
        slots[id] = tex_idx;
        decals.decal_texture_slots.set(slots);
        Ok(id)
    }

    // Tombstone a runtime decal slot. The id becomes invalid; the next
    // `add_decal` may reuse it. Reached only through the bin's `cn debug`
    // runtime-mutation path (dead in the FFI lib, live in the bin).
    pub(crate) fn remove_decal(&mut self, decal_id: usize) -> RenderResult<()> {
        self.decal
            .set
            .remove(decal_id)
            .map_err(|e| RenderError::Other(format!("remove_decal: id {decal_id} {e}")))?;
        if let Some(decals) = &self.decal.resources {
            let mut slots = decals.decal_texture_slots.get();
            slots[decal_id] = usize::MAX;
            decals.decal_texture_slots.set(slots);
        }
        Ok(())
    }

    // Re-point every decal albedo set that pointed at texture-pool slot
    // `slot` to the new `GpuImage` at that slot. Called from the
    // streaming-texture path when an evicted slot is replaced. Walks
    // `decal.resources.decal_texture_slots` so a world with no decals pays
    // nothing.
    // Whether any live decal's albedo set samples texture-pool `slot`. The
    // streaming fast path checks this: decal sets are single-copy and bound
    // whenever the decal pass runs, so a swap of a slot they sample must
    // drain the device before rewriting.
    pub(in crate::vulkan) fn decal_samples_slot(&self, slot: usize) -> bool {
        match &self.decal.resources {
            Some(s) => s.decal_texture_slots.get().contains(&slot),
            None => false,
        }
    }

    pub(in crate::vulkan) fn rewrite_decal_albedo_slot(&self, slot: usize) {
        let decals = match &self.decal.resources {
            Some(s) => s,
            None => return,
        };
        let slots = decals.decal_texture_slots.get();
        let last_tex = self.scene.textures.len().saturating_sub(1);
        for (id, &tex_slot) in slots.iter().enumerate() {
            if tex_slot == slot {
                let view = self.scene.textures[tex_slot.min(last_tex)].view;
                write_albedo_image(&self.hw.device, decals.albedo_sets[id], view);
            }
        }
    }
}

// Point a decal's albedo set at `view`. Its sampler was written with the set.
fn write_albedo_image(device: &VkDevice, set: vk::DescriptorSet, view: vk::ImageView) {
    SetWrites::new(set).sampled_image(0, view).apply(device);
}

// Wire authored decals into the runtime state at init.

impl VkContext {
    // Push every world-authored `DecalRecord` through `add_decal` so its
    // albedo descriptor lands in the reserved slot. Called once from
    // `VkContext::new` after `decal.resources` is built.
    pub(in crate::vulkan) fn upload_initial_decals(
        &mut self,
        records: Vec<DecalRecord>,
    ) -> RenderResult<()> {
        if records.len() > MAX_DECALS {
            return Err(RenderError::Other(format!(
                "decals: {} authored decals exceed MAX_DECALS ({})",
                records.len(),
                MAX_DECALS
            )));
        }
        for rec in records {
            self.add_decal(rec)?;
        }
        Ok(())
    }
}

// Re-export GpuImage so the textures module compiles cleanly when the
// decal module is the only consumer of a few of its helpers.
type _Marker = GpuImage;