luisa_compute/examples/denoiser.rs

use image::Rgb;
use luisa::runtime::extension::DenoiserInput;
use luisa::runtime::DeviceExtensions;
use luisa_compute_api_types::denoiser_ext::{ImageColorSpace, ImageFormat, PrefilterMode};
use luisa_compute_api_types::StreamTag;
use rand::Rng;
use std::time::Instant;
use winit::event::{Event as WinitEvent, WindowEvent};
use winit::event_loop::{ControlFlow, EventLoop};

use luisa::lang::types::vector::alias::*;
use luisa::lang::types::vector::*;
use luisa::prelude::*;
use luisa::rtx::{
    offset_ray_origin, Accel, AccelBuildRequest, AccelOption, AccelVar, Index, Ray, RayComps,
};
use luisa_compute as luisa;
use std::{env::current_exe, process::exit};

fn main() {
    luisa::init_logger_verbose();
    let args: Vec<String> = std::env::args().collect();
    assert!(
        args.len() <= 2,
        "Usage: {} <backend>. <backend>: cpu, cuda, dx, metal, remote",
        args[0]
    );

    let ctx = Context::new(current_exe().unwrap());
    let device = ctx.create_device(if args.len() == 2 {
        args[1].as_str()
    } else {
        "cpu"
    });
    if cfg!(not(feature = "oidn")) {
        eprintln!("Please enable oidn feature to run this example");
        exit(1);
    } else {
        run_pt(device);
    }
}

#[derive(Value, Clone, Copy)]
#[repr(C)]
#[value_new(pub)]
pub struct Onb {
    tangent: Float3,
    binormal: Float3,
    normal: Float3,
}

impl OnbExpr {
    #[tracked]
    fn to_world(&self, v: Expr<Float3>) -> Expr<Float3> {
        self.tangent * v.x + self.binormal * v.y + self.normal * v.z
    }
}

const CBOX_OBJ: &'static str = "
# The original Cornell Box in OBJ format.
# Note that the real box is not a perfect cube, so
# the faces are imperfect in this data set.
#
# Created by Guedis Cardenas and Morgan McGuire at Williams College, 2011
# Released into the Public Domain.
#
# http://graphics.cs.williams.edu/data
# http://www.graphics.cornell.edu/online/box/data.html
#

mtllib CornellBox-Original.mtl

g floor
v  -1.01  0.00   0.99
v   1.00  0.00   0.99
v   1.00  0.00  -1.04
v  -0.99  0.00  -1.04
f -4 -3 -2 -1

g ceiling
v  -1.02  1.99   0.99
v  -1.02  1.99  -1.04
v   1.00  1.99  -1.04
v   1.00  1.99   0.99
f -4 -3 -2 -1

g backWall
v  -0.99  0.00  -1.04
v   1.00  0.00  -1.04
v   1.00  1.99  -1.04
v  -1.02  1.99  -1.04
f -4 -3 -2 -1

g rightWall
v	1.00  0.00  -1.04
v	1.00  0.00   0.99
v	1.00  1.99   0.99
v	1.00  1.99  -1.04
f -4 -3 -2 -1

g leftWall
v  -1.01  0.00   0.99
v  -0.99  0.00  -1.04
v  -1.02  1.99  -1.04
v  -1.02  1.99   0.99
f -4 -3 -2 -1

g shortBox

# Top Face
v	0.53  0.60   0.75
v	0.70  0.60   0.17
v	0.13  0.60   0.00
v  -0.05  0.60   0.57
f -4 -3 -2 -1

# Left Face
v  -0.05  0.00   0.57
v  -0.05  0.60   0.57
v   0.13  0.60   0.00
v   0.13  0.00   0.00
f -4 -3 -2 -1

# Front Face
v	0.53  0.00   0.75
v	0.53  0.60   0.75
v  -0.05  0.60   0.57
v  -0.05  0.00   0.57
f -4 -3 -2 -1

# Right Face
v	0.70  0.00   0.17
v	0.70  0.60   0.17
v	0.53  0.60   0.75
v	0.53  0.00   0.75
f -4 -3 -2 -1

# Back Face
v	0.13  0.00   0.00
v	0.13  0.60   0.00
v	0.70  0.60   0.17
v	0.70  0.00   0.17
f -4 -3 -2 -1

# Bottom Face
v	0.53  0.00   0.75
v	0.70  0.00   0.17
v	0.13  0.00   0.00
v  -0.05  0.00   0.57
f -4 -3 -2 -1

g tallBox

# Top Face
v	-0.53  1.20   0.09
v	 0.04  1.20  -0.09
v	-0.14  1.20  -0.67
v	-0.71  1.20  -0.49
f -4 -3 -2 -1

# Left Face
v	-0.53  0.00   0.09
v	-0.53  1.20   0.09
v	-0.71  1.20  -0.49
v	-0.71  0.00  -0.49
f -4 -3 -2 -1

# Back Face
v	-0.71  0.00  -0.49
v	-0.71  1.20  -0.49
v	-0.14  1.20  -0.67
v	-0.14  0.00  -0.67
f -4 -3 -2 -1

# Right Face
v	-0.14  0.00  -0.67
v	-0.14  1.20  -0.67
v	 0.04  1.20  -0.09
v	 0.04  0.00  -0.09
f -4 -3 -2 -1

# Front Face
v	 0.04  0.00  -0.09
v	 0.04  1.20  -0.09
v	-0.53  1.20   0.09
v	-0.53  0.00   0.09
f -4 -3 -2 -1

# Bottom Face
v	-0.53  0.00   0.09
v	 0.04  0.00  -0.09
v	-0.14  0.00  -0.67
v	-0.71  0.00  -0.49
f -4 -3 -2 -1

g light
v	-0.24  1.98   0.16
v	-0.24  1.98  -0.22
v	 0.23  1.98  -0.22
v	 0.23  1.98   0.16
f -4 -3 -2 -1";

pub const SPP_PER_DISPATCH: u32 = 1;

fn run_pt(device: Device) {
    let mut buf = std::io::BufReader::new(CBOX_OBJ.as_bytes());
    let (models, _) = tobj::load_obj_buf(
        &mut buf,
        &tobj::LoadOptions {
            triangulate: true,
            ..Default::default()
        },
        |p| tobj::load_mtl(p),
    )
    .unwrap();

    let vertex_heap = device.create_bindless_array(65536);
    let index_heap = device.create_bindless_array(65536);
    let mut vertex_buffers: Vec<Buffer<[f32; 3]>> = vec![];
    let mut index_buffers: Vec<Buffer<Index>> = vec![];
    let accel = device.create_accel(AccelOption::default());
    let stream = device.create_stream(StreamTag::Graphics);
    stream.with_scope(|s| {
        let mut cmds = vec![];
        for (index, model) in models.iter().enumerate() {
            let vertex_buffer = device.create_buffer(model.mesh.positions.len() / 3);
            let index_buffer = device.create_buffer(model.mesh.indices.len() / 3);
            let mesh = device.create_mesh(
                vertex_buffer.view(..),
                index_buffer.view(..),
                AccelOption::default(),
            );
            cmds.push(vertex_buffer.view(..).copy_from_async(unsafe {
                let vertex_ptr = model.mesh.positions.as_ptr();
                std::slice::from_raw_parts(
                    vertex_ptr as *const [f32; 3],
                    model.mesh.positions.len() / 3,
                )
            }));
            cmds.push(index_buffer.view(..).copy_from_async(unsafe {
                let index_ptr = model.mesh.indices.as_ptr();
                std::slice::from_raw_parts(index_ptr as *const Index, model.mesh.indices.len() / 3)
            }));

            vertex_buffers.push(vertex_buffer);
            index_buffers.push(index_buffer);
            vertex_heap.emplace_buffer_async(index, vertex_buffers.last().unwrap());
            index_heap.emplace_buffer_async(index, index_buffers.last().unwrap());
            cmds.push(mesh.build_async(AccelBuildRequest::ForceBuild));
            accel.push_mesh(&mesh, Mat4::identity(), 255, true);
        }
        cmds.push(vertex_heap.update_async());
        cmds.push(index_heap.update_async());
        cmds.push(accel.build_async(AccelBuildRequest::ForceBuild));
        s.submit(cmds);
        s.synchronize();
    });

    // use create_kernel_async to compile multiple kernels in parallel
    let path_tracer = Kernel::<
        fn(Tex2d<Float4>, Tex2d<Float4>, Tex2d<Float4>, Tex2d<u32>, Accel, Uint2),
    >::new_async(
        &device,
        &track!(|image: Tex2dVar<Float4>,
                 albedo_image: Tex2dVar<Float4>,
                 normal_image: Tex2dVar<Float4>,
                 seed_image: Tex2dVar<u32>,
                 accel: AccelVar,
                 resolution: Expr<Uint2>| {
            set_block_size([16u32, 16u32, 1u32]);
            let cbox_materials = ([
                Float3::new(0.725f32, 0.710f32, 0.680f32), // floor
                Float3::new(0.725f32, 0.710f32, 0.680f32), // ceiling
                Float3::new(0.725f32, 0.710f32, 0.680f32), // back wall
                Float3::new(0.140f32, 0.450f32, 0.091f32), // right wall
                Float3::new(0.630f32, 0.065f32, 0.050f32), // left wall
                Float3::new(0.725f32, 0.710f32, 0.680f32), // short box
                Float3::new(0.725f32, 0.710f32, 0.680f32), // tall box
                Float3::new(0.000f32, 0.000f32, 0.000f32), // light
            ])
            .expr();

            let lcg = |state: Var<u32>| -> Expr<f32> {
                let lcg = Callable::<fn(Var<u32>) -> Expr<f32>>::new_static(|state: Var<u32>| {
                    const LCG_A: u32 = 1664525u32;
                    const LCG_C: u32 = 1013904223u32;
                    *state = LCG_A * state + LCG_C;
                    (state & 0x00ffffffu32).as_f32() * (1.0f32 / 0x01000000u32 as f32)
                });
                lcg.call(state)
            };

            let make_ray =
                |o: Expr<Float3>, d: Expr<Float3>, tmin: Expr<f32>, tmax: Expr<f32>| -> Expr<Ray> {
                    Ray::from_comps_expr(RayComps {
                        orig: o.into(),
                        tmin: tmin,
                        dir: d.into(),
                        tmax: tmax,
                    })
                };

            let generate_ray = |p: Expr<Float2>| -> Expr<Ray> {
                let fov = escape!({
                    const FOV: f32 = 27.8f32 * std::f32::consts::PI / 180.0f32;
                    FOV
                });
                let origin = Float3::expr(-0.01f32, 0.995f32, 5.0f32);

                let pixel = origin
                    + Float3::expr(
                        p.x * escape!(f32::tan(0.5f32 * fov)),
                        p.y * escape!(f32::tan(0.5f32 * fov)),
                        -1.0f32,
                    );
                let direction = (pixel - origin).normalize();
                make_ray(origin, direction, 0.0f32.expr(), f32::MAX.expr())
            };

            let balanced_heuristic =
                |pdf_a: Expr<f32>, pdf_b: Expr<f32>| pdf_a / luisa::max(pdf_a + pdf_b, 1e-4f32);

            let make_onb = |normal: Expr<Float3>| -> Expr<Onb> {
                let binormal = if normal.x.abs() > normal.z.abs() {
                    Float3::expr(-normal.y, normal.x, 0.0f32)
                } else {
                    Float3::expr(0.0f32, -normal.z, normal.y)
                };

                let tangent = binormal.cross(normal).normalize();
                Onb::new_expr(tangent, binormal, normal)
            };

            let cosine_sample_hemisphere = |u: Expr<Float2>| {
                let r = u.x.sqrt();
                let phi = 2.0f32 * std::f32::consts::PI * u.y;
                Float3::expr(r * phi.cos(), r * phi.sin(), (1.0f32 - u.x).sqrt())
            };

            let coord = dispatch_id().xy();
            let frame_size = luisa::min(resolution.x, resolution.y).as_f32();
            let state = Var::<u32>::zeroed();
            state.store(seed_image.read(coord));

            let rx = lcg(state);
            let ry = lcg(state);

            let pixel = (coord.as_float2() + Float2::expr(rx, ry)) / frame_size * 2.0f32 - 1.0f32;

            let radiance = Var::<Float3>::zeroed();
            radiance.store(Float3::expr(0.0f32, 0.0f32, 0.0f32));
            for _ in 0..SPP_PER_DISPATCH as u32 {
                let init_ray = generate_ray(pixel * Float2::expr(1.0f32, -1.0f32));
                let ray = Var::<Ray>::zeroed();
                ray.store(init_ray);

                let beta = Var::<Float3>::zeroed();
                beta.store(Float3::expr(1.0f32, 1.0f32, 1.0f32));
                let pdf_bsdf = Var::<f32>::zeroed();
                pdf_bsdf.store(0.0f32);

                let light_position = Float3::expr(-0.24f32, 1.98f32, 0.16f32);
                let light_u = Float3::expr(-0.24f32, 1.98f32, -0.22f32) - light_position;
                let light_v = Float3::expr(0.23f32, 1.98f32, 0.16f32) - light_position;
                let light_emission = Float3::expr(17.0f32, 12.0f32, 4.0f32);
                let light_area = light_u.cross(light_v).length();
                let light_normal = light_u.cross(light_v).normalize();

                let depth = Var::<u32>::zeroed();
                while depth < 10u32 {
                    let hit = accel.trace_closest(**ray);

                    if !hit.valid() {
                        break;
                    }

                    let vertex_buffer = vertex_heap.var().buffer::<[f32; 3]>(hit.inst_id);
                    let triangle = index_heap
                        .var()
                        .buffer::<Index>(hit.inst_id)
                        .read(hit.prim_id);

                    let p0: Expr<Float3> = vertex_buffer.read(triangle[0]).into();
                    let p1: Expr<Float3> = vertex_buffer.read(triangle[1]).into();
                    let p2: Expr<Float3> = vertex_buffer.read(triangle[2]).into();

                    let p = p0 * (1.0f32 - hit.u - hit.v) + p1 * hit.u + p2 * hit.v;
                    let n = (p1 - p0).cross(p2 - p0).normalize();

                    let origin: Expr<Float3> = (**ray.orig).into();
                    let direction: Expr<Float3> = (**ray.dir).into();
                    let cos_wi = -direction.dot(n);
                    if cos_wi < 1e-4f32 {
                        break;
                    }
                    let pp = offset_ray_origin(p, n);
                    let albedo = cbox_materials.read(hit.inst_id);
                    if depth == 0 {
                        albedo_image
                            .write(coord, Float4::expr(albedo.x, albedo.y, albedo.z, 1.0f32));
                        normal_image.write(coord, Float4::expr(n.x, n.y, n.z, 1.0f32));
                    }
                    // hit light
                    if hit.inst_id == 7u32 {
                        if depth == 0u32 {
                            radiance.store(radiance + light_emission);
                        } else {
                            let pdf_light = (p - origin).length_squared() / (light_area * cos_wi);
                            let mis_weight = balanced_heuristic(**pdf_bsdf, pdf_light);
                            radiance.store(radiance + mis_weight * beta * light_emission);
                        }
                        break;
                    } else {
                        // sample light
                        let ux_light = lcg(state);
                        let uy_light = lcg(state);
                        let p_light = light_position + ux_light * light_u + uy_light * light_v;

                        let pp_light = offset_ray_origin(p_light, light_normal);
                        let d_light = (pp - pp_light).length();
                        let wi_light = (pp_light - pp).normalize();
                        let shadow_ray =
                            make_ray(offset_ray_origin(pp, n), wi_light, 0.0f32.expr(), d_light);
                        let occluded = accel.trace_any(shadow_ray);
                        let cos_wi_light = wi_light.dot(n);
                        let cos_light = -light_normal.dot(wi_light);

                        if !occluded && cos_wi_light > 1e-4f32 && cos_light > 1e-4f32 {
                            let pdf_light = (d_light * d_light) / (light_area * cos_light);
                            let pdf_bsdf = cos_wi_light * std::f32::consts::FRAC_1_PI;
                            let mis_weight = balanced_heuristic(pdf_light, pdf_bsdf);
                            let bsdf = albedo * std::f32::consts::FRAC_1_PI * cos_wi_light;
                            *radiance += beta * bsdf * mis_weight * light_emission
                                / luisa::max(pdf_light, 1e-4f32);
                        }
                    }
                    // sample BSDF
                    let onb = make_onb(n);
                    let ux = lcg(state);
                    let uy = lcg(state);

                    let new_direction =
                        onb.to_world(cosine_sample_hemisphere(Float2::expr(ux, uy)));
                    let cos_wi = new_direction.dot(n).abs();
                    *ray = make_ray(pp, new_direction, 0.0f32.expr(), std::f32::MAX.expr());
                    *beta *= albedo;
                    pdf_bsdf.store(cos_wi * std::f32::consts::FRAC_1_PI);

                    // russian roulette
                    let l = Float3::expr(0.212671f32, 0.715160f32, 0.072169f32).dot(beta);
                    if l == 0.0f32 {
                        break;
                    }
                    let q = luisa::max(l, 0.05f32);
                    let r = lcg(state);
                    if r > q {
                        break;
                    }
                    *beta = beta / q;

                    *depth += 1;
                }
            }
            radiance.store(radiance / SPP_PER_DISPATCH as f32);
            seed_image.write(coord, state);
            if radiance.is_nan().any() {
                radiance.store(Float3::expr(0.0f32, 0.0f32, 0.0f32));
            }
            let radiance = radiance.clamp(
                Float3::splat_expr(0.0f32.expr()),
                Float3::splat_expr(30.0f32.expr()),
            );
            let old = image.read(dispatch_id().xy());
            let spp = old.w;
            let radiance = radiance + old.xyz();
            image.write(
                dispatch_id().xy(),
                Float4::expr(radiance.x, radiance.y, radiance.z, spp + 1.0f32),
            );
        }),
    );
    let img_w = 1024u32;
    let img_h = 1024u32;

    let color_buf = device.create_buffer::<Float4>((img_w * img_h) as usize);
    let albedo_buf = device.create_buffer::<Float4>((img_w * img_h) as usize);
    let normal_buf = device.create_buffer::<Float4>((img_w * img_h) as usize);
    let output_buf = device.create_buffer::<Float4>((img_w * img_h) as usize);
    let ext = device
        .denoiser_ext()
        .unwrap_or_else(|| panic!("denoiser not available on current device"));
    let mut denoiser = ext.create(&device.default_stream());
    {
        let mut inputs = DenoiserInput::new(img_w, img_h);
        inputs.push_noisy_image(
            &color_buf.view(..),
            &output_buf.view(..),
            ImageFormat::Float3,
            ImageColorSpace::Hdr,
            None,
        );
        inputs.push_feature_image(
            "albedo",
            &albedo_buf.view(..),
            ImageFormat::Float3,
            ImageColorSpace::Hdr,
        );
        inputs.push_feature_image(
            "normal",
            &normal_buf.view(..),
            ImageFormat::Float3,
            ImageColorSpace::Hdr,
        );
        inputs.noisy_features(false);
        inputs.prefilter_mode(PrefilterMode::None);
        denoiser.init(inputs);
    }
    let acc_to_hdr = Kernel::<fn(Tex2d<Float4>, Tex2d<Float4>)>::new_async(
        &device,
        &track!(|acc, hdr| {
            set_block_size([16, 16, 1]);
            let coord = dispatch_id().xy();
            let radiance = acc.read(coord);
            let spp = radiance.w;
            let radiance = radiance.xyz() / spp;
            hdr.write(
                coord,
                Float4::expr(radiance.x, radiance.y, radiance.z, 1.0f32),
            );
        }),
    );
    let display = Kernel::<fn(Tex2d<Float4>, Tex2d<Float4>)>::new_async(
        &device,
        &track!(|hdr, display| {
            set_block_size([16, 16, 1]);
            let coord = dispatch_id().xy();
            let radiance = hdr.read(coord).xyz();

            let r = 1.055f32 * radiance.powf(1.0 / 2.4f32) - 0.055;

            let srgb = radiance.lt(0.0031308).select(radiance * 12.92, r);
            display.write(coord, Float4::expr(srgb.x, srgb.y, srgb.z, 1.0f32));
        }),
    );

    let albedo_img = device.create_tex2d::<Float4>(PixelStorage::Float4, img_w, img_h, 1);
    let normal_img = device.create_tex2d::<Float4>(PixelStorage::Float4, img_w, img_h, 1);
    let hdr_img = device.create_tex2d::<Float4>(PixelStorage::Float4, img_w, img_h, 1);
    let acc_img = device.create_tex2d::<Float4>(PixelStorage::Float4, img_w, img_h, 1);
    let seed_img = device.create_tex2d::<u32>(PixelStorage::Int1, img_w, img_h, 1);
    {
        let mut rng = rand::thread_rng();
        let seed_buffer = (0..img_w * img_h)
            .map(|_| rng.gen::<u32>())
            .collect::<Vec<_>>();
        seed_img.view(0).copy_from(&seed_buffer);
    }
    let event_loop = EventLoop::new().unwrap();
    let window = winit::window::WindowBuilder::new()
        .with_title("Luisa Compute Rust - Ray Tracing")
        .with_inner_size(winit::dpi::LogicalSize::new(img_w, img_h))
        .with_resizable(false)
        .build(&event_loop)
        .unwrap();
    let swapchain = device.create_swapchain(
        &window,
        &device.default_stream(),
        img_w,
        img_h,
        false,
        false,
        3,
    );
    let display_img = device.create_tex2d::<Float4>(swapchain.pixel_storage(), img_w, img_h, 1);
    event_loop.set_control_flow(ControlFlow::Poll);
    event_loop
        .run(move |event, elwt| {
            match event {
                WinitEvent::WindowEvent {
                    event: WindowEvent::CloseRequested,
                    window_id,
                } if window_id == window.id() => {
                    // FIXME: support half4 pixel storage
                    let mut img_buffer = vec![[0u8; 4]; (img_w * img_h) as usize];
                    {
                        let scope = device.default_stream().scope();
                        scope.submit([display_img.view(0).copy_to_async(&mut img_buffer)]);
                    }
                    {
                        let img = image::RgbImage::from_fn(img_w, img_h, |x, y| {
                            let i = x + y * img_w;
                            let px = img_buffer[i as usize];
                            Rgb([px[0], px[1], px[2]])
                        });
                        img.save("cbox.png").unwrap();
                    }
                    elwt.exit();
                }
                WinitEvent::AboutToWait => {
                    window.request_redraw();
                }
                WinitEvent::WindowEvent {
                    event: WindowEvent::RedrawRequested,
                    window_id,
                } if window_id == window.id() => {
                    let tic = Instant::now();
                    {
                        let scope = device.default_stream().scope();
                        scope.present(&swapchain, &display_img);
                        scope.submit([
                            path_tracer.dispatch_async(
                                [img_w, img_h, 1],
                                &acc_img,
                                &albedo_img,
                                &normal_img,
                                &seed_img,
                                &accel,
                                &Uint2::new(img_w, img_h),
                            ),
                            acc_to_hdr.dispatch_async([img_w, img_h, 1], &acc_img, &hdr_img),
                            hdr_img.view(0).copy_to_buffer_async(&color_buf.view(..)),
                            albedo_img
                                .view(0)
                                .copy_to_buffer_async(&albedo_buf.view(..)),
                            normal_img
                                .view(0)
                                .copy_to_buffer_async(&normal_buf.view(..)),
                        ]);
                        denoiser.execute(true);
                        scope.submit([
                            hdr_img.view(0).copy_from_buffer_async(&output_buf.view(..)),
                            display.dispatch_async([img_w, img_h, 1], &hdr_img, &display_img),
                        ]);
                    }
                    let toc = Instant::now();
                    let elapsed = (toc - tic).as_secs_f32();
                    log::info!(
                        "time: {}ms {}ms/spp",
                        elapsed * 1e3,
                        elapsed * 1e3 / SPP_PER_DISPATCH as f32
                    );
                    window.request_redraw();
                }
                _ => (),
            }
        })
        .unwrap();
}