concinnity-device 0.19.16

// src/directx/init/effects.rs
//
// Post-process pipeline + target construction extracted from DxContext::new:
// bloom (always built), TAA history-resolve (gated on `taa_enabled`), and
// SSAO (gated on `ssao_settings.is_some()`). Each gated block pays zero cost
// when its setting is off.

use windows::Win32::Graphics::Direct3D12::*;

use crate::directx::allocator::{DeviceAllocator, PooledTexture};
use crate::directx::context::dump_on_err;
use crate::directx::post::bloom::{
    compile_bloom_shaders, create_bloom_mips, create_bloom_pso, create_bloom_root_signature,
    write_color_rtv,
};
use crate::directx::post::rt_reflections::{
    RtBuildContext, RtBuildInit, RtOutputDescriptors, RtReflectionsResources,
};
use crate::directx::post::ssao::{SsaoDescriptorHandles, SsaoDeviceCtx, SsaoResources};
use crate::directx::post::ssgi::{SsgiDescriptors, SsgiDevice, SsgiResources};
use crate::directx::post::ssr::{SsrInitInputs, SsrResources};
use crate::directx::post::taa::TaaResources;
use crate::directx::texture::{
    HDR_FORMAT, create_fallback_white_resource, write_hdr_srv, write_texture_srv,
};
use crate::directx::transient_pool::{TransientResourcePool, transient_slots};

pub(super) struct EffectsBundle {
    // Graph-owned transient render targets (the resources the aliasing planner
    // manages: `bloom_top` always, `ao_output` when SSAO is on). Built first so
    // the bloom chain and SSAO read their placed resources back by label.
    pub transient_pool: TransientResourcePool,
    pub bloom_mips: Vec<ID3D12Resource>,
    pub bloom_mip_rtvs: Vec<D3D12_CPU_DESCRIPTOR_HANDLE>,
    pub bloom_mip_srv_gpus: Vec<D3D12_GPU_DESCRIPTOR_HANDLE>,
    pub bloom_mip_extents: Vec<(u32, u32)>,
    pub bloom_root_sig: ID3D12RootSignature,
    pub bloom_pso_prefilter: ID3D12PipelineState,
    pub bloom_pso_downsample: ID3D12PipelineState,
    pub bloom_pso_upsample: ID3D12PipelineState,
    pub taa: Option<TaaResources>,
    pub ssao: Option<SsaoResources>,
    pub ssao_white: PooledTexture,
    pub ssao_white_srv_gpu: D3D12_GPU_DESCRIPTOR_HANDLE,
    pub ssr: Option<SsrResources>,
    pub ssgi: Option<SsgiResources>,
    // Built only when RT reflections are authored AND the GPU supports the DXR
    // tier AND the DXC compile succeeds; `None` otherwise (the graph then falls
    // back to SSR). The acceleration structure is built separately in mod.rs.
    pub rt_reflections: Option<RtReflectionsResources>,
}

// Slot handles for the bloom mip chain. The caller knows the descriptor-heap
// layout (RTV base, SRV slot base for the first mip) and supplies a closure
// per kind to mint per-mip handles. This keeps the heap layout in mod.rs.
pub(super) struct BloomSlots<'a> {
    pub rtv_for: &'a dyn Fn(usize) -> D3D12_CPU_DESCRIPTOR_HANDLE,
    pub srv_cpu_for: &'a dyn Fn(usize) -> D3D12_CPU_DESCRIPTOR_HANDLE,
    pub srv_gpu_for: &'a dyn Fn(usize) -> D3D12_GPU_DESCRIPTOR_HANDLE,
}

pub(super) struct TaaSlots {
    pub history_rtv: [D3D12_CPU_DESCRIPTOR_HANDLE; 2],
    pub history_srv: [(D3D12_CPU_DESCRIPTOR_HANDLE, D3D12_GPU_DESCRIPTOR_HANDLE); 2],
}

pub(super) struct SsaoSlots {
    pub ao_raw_rtv: D3D12_CPU_DESCRIPTOR_HANDLE,
    pub ao_raw_srv: (D3D12_CPU_DESCRIPTOR_HANDLE, D3D12_GPU_DESCRIPTOR_HANDLE),
    pub ao_rtv: D3D12_CPU_DESCRIPTOR_HANDLE,
    pub ao_srv: (D3D12_CPU_DESCRIPTOR_HANDLE, D3D12_GPU_DESCRIPTOR_HANDLE),
    pub white_srv: (D3D12_CPU_DESCRIPTOR_HANDLE, D3D12_GPU_DESCRIPTOR_HANDLE),
}

pub(super) struct SsrSlots {
    pub output_rtv: D3D12_CPU_DESCRIPTOR_HANDLE,
    pub output_srv: (D3D12_CPU_DESCRIPTOR_HANDLE, D3D12_GPU_DESCRIPTOR_HANDLE),
}

pub(super) struct SsgiSlots {
    pub gi_rtv: D3D12_CPU_DESCRIPTOR_HANDLE,
    pub gi_srv: (D3D12_CPU_DESCRIPTOR_HANDLE, D3D12_GPU_DESCRIPTOR_HANDLE),
}

pub(super) struct RtReflectionsSlots {
    pub output_rtv: D3D12_CPU_DESCRIPTOR_HANDLE,
    pub output_srv: (D3D12_CPU_DESCRIPTOR_HANDLE, D3D12_GPU_DESCRIPTOR_HANDLE),
}

// Render dimensions for the effect targets. Drawable (width/height) sizes the
// bloom mip chain; render-resolution sizes the TAA / SSAO / SSR targets upstream
// of the upscaler. They are equal when temporal upscaling is off.
#[derive(Clone, Copy)]
pub(super) struct EffectDimensions {
    pub width: u32,
    pub height: u32,
    pub render_width: u32,
    pub render_height: u32,
}

// The world-resolved feature settings that gate each pass. Each Option is None
// when the feature is off. `rt_supported` is the DXR-tier capability; RT builds
// only when both it and `rt_reflection_settings` are set.
pub(super) struct EffectSettings {
    pub ssao_settings: Option<crate::gfx::ssao::SsaoSettings>,
    pub ssr_settings: Option<crate::gfx::ssr::SsrSettings>,
    pub ssgi_settings: Option<crate::gfx::ssgi::SsgiSettings>,
    pub rt_reflection_settings: Option<crate::gfx::rt_reflections::RtReflectionSettings>,
    pub rt_supported: bool,
}

// Non-settings build flags: TAA on/off, whether the unified G-buffer pre-pass
// is built (which decides whether the pool places its colour targets), and the
// shader hot-reload toggle.
#[derive(Clone, Copy)]
pub(super) struct EffectFlags {
    pub taa_enabled: bool,
    pub gbuffer_enabled: bool,
    pub hot_reload: bool,
}

// Descriptor-heap slots for every effect pass: bloom, TAA, SSAO, SSR, SSGI, RT.
pub(super) struct EffectDescriptorSlots<'a> {
    pub bloom: BloomSlots<'a>,
    pub taa: TaaSlots,
    pub ssao: SsaoSlots,
    pub ssr: SsrSlots,
    pub ssgi: SsgiSlots,
    pub rt: RtReflectionsSlots,
}

pub(super) fn build_effects(
    alloc: &DeviceAllocator,
    info_queue: Option<&ID3D12InfoQueue>,
    dims: EffectDimensions,
    settings: EffectSettings,
    flags: EffectFlags,
    slots: EffectDescriptorSlots<'_>,
) -> Result<EffectsBundle, String> {
    let device = alloc.device();
    let EffectDimensions {
        width,
        height,
        render_width,
        render_height,
    } = dims;
    let EffectSettings {
        ssao_settings,
        ssr_settings,
        ssgi_settings,
        rt_reflection_settings,
        rt_supported,
    } = settings;
    let EffectFlags {
        taa_enabled,
        gbuffer_enabled,
        hot_reload,
    } = flags;
    let EffectDescriptorSlots {
        bloom,
        taa: taa_slots,
        ssao: ssao_slots,
        ssr: ssr_slots,
        ssgi: ssgi_slots,
        rt: rt_slots,
    } = slots;
    // Transient pool: the graph-owned transient render targets. `bloom_top`
    // (bloom mip 0) is always managed; `ao_output` is placed only when SSAO is
    // on (else `resource_for` returns None and the main pass binding 6 falls
    // back to `ssao_white`). Built first so the bloom chain and SSAO read their
    // placed resources back by label.
    let transient_pool = TransientResourcePool::build(
        device,
        alloc.queue(),
        &transient_slots(
            ssao_settings.is_some(),
            gbuffer_enabled,
            (render_width, render_height),
            (width, height),
        )?,
    )?;

    // Bloom mip chain + per-mip RTV/SRV writes. `mips[0]` (`bloom_top`) is the
    // pool's placed resource; the finer mips are committed.
    let bloom_top = transient_pool
        .resource_for("bloom_top")
        .ok_or("transient pool missing bloom_top")?
        .clone();
    let (bloom_mips, bloom_mip_extents) = create_bloom_mips(device, width, height, bloom_top)?;
    let mut bloom_mip_rtvs: Vec<D3D12_CPU_DESCRIPTOR_HANDLE> = Vec::with_capacity(bloom_mips.len());
    let mut bloom_mip_srv_gpus: Vec<D3D12_GPU_DESCRIPTOR_HANDLE> =
        Vec::with_capacity(bloom_mips.len());
    for (i, mip) in bloom_mips.iter().enumerate() {
        let rtv = (bloom.rtv_for)(i);
        write_color_rtv(device, mip, rtv);
        bloom_mip_rtvs.push(rtv);
        let srv_cpu = (bloom.srv_cpu_for)(i);
        write_hdr_srv(device, mip, srv_cpu);
        bloom_mip_srv_gpus.push((bloom.srv_gpu_for)(i));
    }

    // Bloom pipelines (prefilter / downsample / upsample). All three write
    // HDR_FORMAT mips; the upsample blends additively so each coarser mip
    // accumulates onto the finer one.
    let bloom_root_sig = dump_on_err(info_queue, create_bloom_root_signature(device))?;
    let bloom_shaders = compile_bloom_shaders(hot_reload)?;
    let bloom_pso_prefilter = dump_on_err(
        info_queue,
        create_bloom_pso(
            device,
            &bloom_root_sig,
            &bloom_shaders.vs,
            &bloom_shaders.prefilter_ps,
            HDR_FORMAT,
            false,
        ),
    )?;
    let bloom_pso_downsample = dump_on_err(
        info_queue,
        create_bloom_pso(
            device,
            &bloom_root_sig,
            &bloom_shaders.vs,
            &bloom_shaders.downsample_ps,
            HDR_FORMAT,
            false,
        ),
    )?;
    let bloom_pso_upsample = dump_on_err(
        info_queue,
        create_bloom_pso(
            device,
            &bloom_root_sig,
            &bloom_shaders.vs,
            &bloom_shaders.upsample_ps,
            HDR_FORMAT,
            true,
        ),
    )?;

    // TAA history-resolve resources. Sized at render-res; the motion it
    // reprojects through comes from the unified G-buffer pre-pass.
    let taa = if taa_enabled {
        Some(TaaResources::new(
            device,
            render_width,
            render_height,
            taa_slots.history_rtv,
            taa_slots.history_srv,
            info_queue,
            hot_reload,
        )?)
    } else {
        None
    };

    // SSAO: 1x1 white fallback always populated so the main pass binds a
    // pass-through occlusion when SSAO is off. The real SSAO targets sit in
    // the three slots before it.
    let ssao_white = create_fallback_white_resource(alloc)?;
    write_texture_srv(device, &ssao_white, ssao_slots.white_srv.0);
    let ssao_white_srv_gpu = ssao_slots.white_srv.1;
    let ssao = if let Some(settings) = ssao_settings {
        let ao_resource = transient_pool
            .resource_for("ao_output")
            .ok_or("transient pool missing ao_output while SSAO is enabled")?;
        Some(SsaoResources::new(
            SsaoDeviceCtx { device, info_queue },
            render_width,
            render_height,
            settings,
            SsaoDescriptorHandles {
                ao_raw_rtv: ssao_slots.ao_raw_rtv,
                ao_raw_srv: ssao_slots.ao_raw_srv,
                ao_rtv: ssao_slots.ao_rtv,
                ao_srv: ssao_slots.ao_srv,
            },
            ao_resource,
            hot_reload,
        )?)
    } else {
        None
    };

    // SSR: a fullscreen resolve reading the unified G-buffer pre-pass. The
    // resources are built whenever SSR *or* SSGI *or* RT reflections are on (all
    // reuse the G-buffer); `ssr_settings` (the resolve half) stays `None` for a
    // SSGI-only or RT-only build, which leaves the reserved output slot unwritten.
    let ssr =
        if ssr_settings.is_some() || ssgi_settings.is_some() || rt_reflection_settings.is_some() {
            Some(SsrResources::new(
                alloc,
                render_width,
                render_height,
                SsrInitInputs {
                    resolve_settings: ssr_settings,
                    output_rtv: ssr_slots.output_rtv,
                    output_srv: ssr_slots.output_srv,
                },
                info_queue,
                hot_reload,
            )?)
        } else {
            None
        };

    // SSGI: hemisphere-gather + depth-aware blur over the unified G-buffer
    // pre-pass. The gather target lives here; the composite blends straight
    // into the scene.
    let ssgi = if let Some(settings) = ssgi_settings {
        Some(SsgiResources::new(
            SsgiDevice { alloc, info_queue },
            render_width,
            render_height,
            settings,
            SsgiDescriptors {
                gi_rtv: ssgi_slots.gi_rtv,
                gi_srv: ssgi_slots.gi_srv,
            },
            hot_reload,
        )?)
    } else {
        None
    };

    // RT reflections: build the pipelines + output target only when authored AND
    // the GPU supports the DXR tier. The DXC compile can still fail (DXC DLL
    // absent / shader error); that is non-fatal -> leave `None` and the graph
    // falls back to SSR (whose resolve is also off in an RT-only world, so the
    // scene simply renders without reflections). The acceleration structure is
    // built separately in mod.rs and gates `rt_reflections_active` alongside this.
    let rt_reflections = match (rt_reflection_settings, rt_supported) {
        (Some(settings), true) => match RtReflectionsResources::new(
            RtBuildContext {
                alloc,
                width: render_width,
                height: render_height,
            },
            settings,
            RtOutputDescriptors {
                output_rtv: rt_slots.output_rtv,
                output_srv: rt_slots.output_srv,
            },
            RtBuildInit {
                info_queue,
                hot_reload,
            },
        ) {
            Ok(r) => Some(r),
            Err(e) => {
                tracing::warn!("RT reflections unavailable, falling back to SSR: {e}");
                None
            }
        },
        _ => None,
    };

    Ok(EffectsBundle {
        transient_pool,
        bloom_mips,
        bloom_mip_rtvs,
        bloom_mip_srv_gpus,
        bloom_mip_extents,
        bloom_root_sig,
        bloom_pso_prefilter,
        bloom_pso_downsample,
        bloom_pso_upsample,
        taa,
        ssao,
        ssao_white,
        ssao_white_srv_gpu,
        ssr,
        ssgi,
        rt_reflections,
    })
}