awsm-renderer 0.3.0

//! Bind group layout + recreation for the material classify pass.
//!
//! Single bind group:
//!   0 visibility_data_tex — uint texture (per-pixel material id).
//!   1 material_mesh_metas — `storage[RO]` mesh-meta table.
//!   2 materials_data      — `storage[RO]` material payload (for shader_id).
//!   3 classify_output     — `storage[RW]` (atomic) per-`shader_id` buckets.

use std::borrow::Cow;

use awsm_renderer_core::bind_groups::{
    BindGroupDescriptor, BindGroupEntry, BindGroupLayoutResource, BindGroupResource,
    BufferBindingLayout, BufferBindingType, TextureBindingLayout,
};
use awsm_renderer_core::buffers::BufferBinding;
use awsm_renderer_core::texture::{TextureSampleType, TextureViewDimension};

use crate::bind_group_layout::{BindGroupLayoutCacheKey, BindGroupLayoutCacheKeyEntry};
use crate::bind_groups::{AwsmBindGroupError, BindGroupRecreateContext};
use crate::error::Result;
use crate::{bind_group_layout::BindGroupLayoutKey, render_passes::RenderPassInitContext};

/// Bind group layout + cached bind group for the classify pass.
pub struct MaterialClassifyBindGroups {
    pub multisampled_bind_group_layout_key: BindGroupLayoutKey,
    pub singlesampled_bind_group_layout_key: BindGroupLayoutKey,
    bind_group: Option<web_sys::GpuBindGroup>,
}

impl MaterialClassifyBindGroups {
    /// Creates the bind group layouts for the classify pass. The
    /// bind group itself is built lazily via [`Self::recreate`] when
    /// the renderer's `BindGroups::mark_create` event fires (e.g. on
    /// the first frame, on viewport resize, when classify buffers are
    /// recreated).
    pub async fn new(ctx: &mut RenderPassInitContext<'_>) -> Result<Self> {
        let multisampled_bind_group_layout_key = create_bind_group_layout_key(ctx, true).await?;
        let singlesampled_bind_group_layout_key = create_bind_group_layout_key(ctx, false).await?;

        Ok(Self {
            multisampled_bind_group_layout_key,
            singlesampled_bind_group_layout_key,
            bind_group: None,
        })
    }

    /// Returns the live classify bind group. Errors if
    /// [`Self::recreate`] hasn't been called yet this session.
    pub fn get_bind_group(
        &self,
    ) -> std::result::Result<&web_sys::GpuBindGroup, AwsmBindGroupError> {
        self.bind_group
            .as_ref()
            .ok_or_else(|| AwsmBindGroupError::NotFound("Material Classify".to_string()))
    }

    /// (Re)builds the classify bind group against the current
    /// classify buffer + visibility view + mesh-meta + materials
    /// buffers. Called from [`crate::bind_groups::BindGroups`] in
    /// response to a `MaterialClassifyResourcesChange` event.
    pub fn recreate(&mut self, ctx: &BindGroupRecreateContext<'_>) -> Result<()> {
        let msaa = ctx.anti_aliasing.msaa_sample_count.is_some();
        let layout_key = if msaa {
            self.multisampled_bind_group_layout_key
        } else {
            self.singlesampled_bind_group_layout_key
        };
        let mut entries = vec![
            BindGroupEntry::new(
                0,
                BindGroupResource::TextureView(Cow::Borrowed(
                    &ctx.render_texture_views.visibility_data,
                )),
            ),
            BindGroupEntry::new(
                1,
                BindGroupResource::Buffer(BufferBinding::new(
                    ctx.meshes.meta.material_gpu_buffer(),
                )),
            ),
            BindGroupEntry::new(
                2,
                BindGroupResource::Buffer(BufferBinding::new(&ctx.materials.gpu_buffer)),
            ),
            BindGroupEntry::new(
                3,
                BindGroupResource::Buffer(BufferBinding::new(
                    &ctx.material_classify_buffers.buffer,
                )),
            ),
        ];

        // Priority 3 — bind the edge buffers (args + data) + edge-layout
        // uniform when MSAA is on AND the device supports the full
        // edge_resolve dispatch wiring. On unsupported devices (or
        // non-MSAA frames) `material_edge_buffers` is None and the
        // classify pass's bind-group layout was built without slots
        // 4/5/6 to match. The split between args_buffer and data_buffer
        // is required so neither buffer is simultaneously bound as
        // Storage(read-write) and used as Indirect (WebGPU rejects that
        // combo within a single compute pass's sync scope).
        if msaa {
            if let (Some(edge_buffers), Some(edge_layout_uniform)) =
                (ctx.material_edge_buffers, ctx.material_edge_layout_uniform)
            {
                // 4: args_buffer (atomic counters + per-shader
                //    workgroup_count_x cells; storage RW for classify,
                //    read by indirect dispatch downstream).
                entries.push(BindGroupEntry::new(
                    4,
                    BindGroupResource::Buffer(BufferBinding::new(&edge_buffers.args_buffer)),
                ));
                // 5: edge_layout uniform.
                entries.push(BindGroupEntry::new(
                    5,
                    BindGroupResource::Buffer(BufferBinding::new(edge_layout_uniform)),
                ));
                // 6: data_buffer (edge_to_xy + edge_slot_map +
                //    accumulator + sample lists; storage RW for the
                //    shader writes).
                entries.push(BindGroupEntry::new(
                    6,
                    BindGroupResource::Buffer(BufferBinding::new(&edge_buffers.data_buffer)),
                ));
                // 7: depth_tex (multisampled depth view) for per-sample
                //    depth variance silhouette detection.
                entries.push(BindGroupEntry::new(
                    7,
                    BindGroupResource::TextureView(Cow::Borrowed(&ctx.render_texture_views.depth)),
                ));
                // 8: camera uniform (for view-space depth conversion
                //    in main's edge_mask_depth_msaa / edge_mask_neighbors
                //    threshold comparison).
                entries.push(BindGroupEntry::new(
                    8,
                    BindGroupResource::Buffer(BufferBinding::new(&ctx.camera.gpu_buffer)),
                ));
                // 9: normal_tangent_tex — per-sample world-space normal
                //    (packed) for the normal-discontinuity check in
                //    edge_mask_neighbors.
                entries.push(BindGroupEntry::new(
                    9,
                    BindGroupResource::TextureView(Cow::Borrowed(
                        &ctx.render_texture_views.normal_tangent,
                    )),
                ));
            }
            // else: edge bindings absent — layout was built without
            // them too, so the bind group is valid with just the 4
            // base entries.
        }

        let descriptor = BindGroupDescriptor::new(
            ctx.bind_group_layouts.get(layout_key)?,
            Some("Material Classify"),
            entries,
        );
        self.bind_group = Some(ctx.gpu.create_bind_group(&descriptor.into()));
        Ok(())
    }
}

/// Returns true when the device + features support the Priority-3
/// edge-emission bindings. **Must match** the same check used in
/// `AwsmRenderer::build()` for the edge buffer allocation (i.e.
/// `crate::edge_resolve_supported`) so the classify layout includes
/// edge bindings iff the renderer allocates the edge buffers. Post
/// the macOS-compatible 4-group layout fold (`6ca750a`), the
/// bind-group cap dropped; only the storage-buffer cap remains.
fn edge_emit_supported(ctx: &RenderPassInitContext<'_>) -> bool {
    crate::edge_resolve_supported(ctx.gpu)
}

async fn create_bind_group_layout_key(
    ctx: &mut RenderPassInitContext<'_>,
    multisampled_geometry: bool,
) -> Result<BindGroupLayoutKey> {
    let mut entries = vec![
        // visibility_data — uint texture; MSAA variant is multisampled.
        BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Texture(
                TextureBindingLayout::new()
                    .with_view_dimension(TextureViewDimension::N2d)
                    .with_sample_type(TextureSampleType::Uint)
                    .with_multisampled(multisampled_geometry),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        },
        // material_mesh_metas — storage RO.
        BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Buffer(
                BufferBindingLayout::new().with_binding_type(BufferBindingType::ReadOnlyStorage),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        },
        // materials_data — storage RO.
        BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Buffer(
                BufferBindingLayout::new().with_binding_type(BufferBindingType::ReadOnlyStorage),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        },
        // classify_output — storage RW (atomics).
        BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Buffer(
                BufferBindingLayout::new().with_binding_type(BufferBindingType::Storage),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        },
    ];

    // Priority 3 — MSAA edge emission. Adds three bindings when the
    // device supports the full Stage 3 dispatch wiring:
    //   4: args_buffer (storage RW — atomic counters + per-shader
    //      workgroup_count_x cells).
    //   5: edge_layout (uniform — host-uploaded offsets).
    //   6: data_buffer (storage RW — edge_to_xy + edge_slot_map +
    //      accumulator + sample lists).
    // Splitting args + data into two buffers sidesteps the WebGPU
    // validation rule that a single buffer can't be Indirect-readable
    // and Storage(read-write) inside one compute pass.
    if multisampled_geometry && edge_emit_supported(ctx) {
        // 4: args_buffer
        entries.push(BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Buffer(
                BufferBindingLayout::new().with_binding_type(BufferBindingType::Storage),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        });
        // 5: edge_layout uniform
        entries.push(BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Buffer(
                BufferBindingLayout::new().with_binding_type(BufferBindingType::Uniform),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        });
        // 6: data_buffer
        entries.push(BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Buffer(
                BufferBindingLayout::new().with_binding_type(BufferBindingType::Storage),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        });
        // 7: depth_tex — multisampled depth texture used by classify to
        //    detect mesh-vs-mesh in-pixel silhouettes via per-sample
        //    depth variance (matches main's edge_mask_depth_msaa).
        entries.push(BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Texture(
                TextureBindingLayout::new()
                    .with_view_dimension(TextureViewDimension::N2d)
                    .with_sample_type(TextureSampleType::Depth)
                    .with_multisampled(true),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        });
        // 8: camera uniform for view-space depth conversion.
        entries.push(BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Buffer(
                BufferBindingLayout::new().with_binding_type(BufferBindingType::Uniform),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        });
        // 9: normal_tangent_tex (per-sample) for the normal-discontinuity
        //    leg of edge_mask_neighbors — catches same-mesh in-pixel
        //    silhouettes at tile-facet boundaries where depth/coverage/
        //    mat_meta don't differ but neighboring facets have rotated
        //    surface normals (e.g. the platform's top-front-edge
        //    diagonal in MorphStressTest).
        entries.push(BindGroupLayoutCacheKeyEntry {
            resource: BindGroupLayoutResource::Texture(
                TextureBindingLayout::new()
                    .with_view_dimension(TextureViewDimension::N2d)
                    .with_sample_type(TextureSampleType::UnfilterableFloat)
                    .with_multisampled(true),
            ),
            visibility_vertex: false,
            visibility_fragment: false,
            visibility_compute: true,
        });
    }

    Ok(ctx
        .bind_group_layouts
        .get_key(ctx.gpu, BindGroupLayoutCacheKey { entries })?)
}