//! PERF-2 P3 go/no-go (`docs/perf2_campaign.md`): the batched f32 red-black
//! V-cycle on the device against the CPU one. Runs only with
//! `--features cuda` on a CUDA host, and only when asked (`--ignored`):
//!
//! `RTX_CUDA_ARCH=sm_120 cargo test --release -p rtx-cfd --features cuda --test gpu_vcycle_bench -- --ignored --nocapture`
//!
//! Checks first that the K = 1 device V-cycle reproduces the CPU f32
//! red-black V-cycle (same algorithm; FMA contraction and summation order
//! differ, so to 1e-4 of the correction's scale), then times one V-cycle
//! per march at K = 1, 8, 16 (including the residual upload and the
//! correction download) against the CPU's serial f64 red-black V-cycle.
#![cfg(feature = "cuda")]

use cudarc::driver::{CudaContext, CudaSlice, LaunchConfig, PushKernelArg};
use cudarc::nvrtc::{CompileOptions, compile_ptx_with_opts};
use rtx_cfd::solvers::incompressible::{
    LevelExport, MgSmoother, MultigridParameters, PoissonProblem, export_hierarchy,
    vcycle_f32_reference,
};
use std::sync::Arc;
use std::time::Instant;

const KERNELS: &str = include_str!("../src/kernels/cuda/mg_vcycle.cu");

/// The anchor-sized masked operator (ny 62 over the 2.5 × 0.41 channel with
/// a cylinder-sized hole), the same construction as the Poisson pins.
fn problem(nx: usize, ny: usize, seed: u64) -> PoissonProblem {
    let mut p = PoissonProblem::new(nx, ny);
    let (dx, dy, dt) = (2.5 / nx as f64, 0.41 / ny as f64, 3.24e-4);
    let (ae, an) = (dt * dy / dx, dt * dx / dy);
    let hole = |i: usize, j: usize| {
        let (x, y) = ((i as f64 + 0.5) * dx, (j as f64 + 0.5) * dy);
        (x - 0.2).powi(2) + (y - 0.2).powi(2) < 0.08 * 0.08
    };
    for j in 0..ny {
        for i in 0..nx {
            let idx = j * nx + i;
            if hole(i, j) {
                p.active[idx] = false;
                continue;
            }
            if i + 1 < nx && !hole(i + 1, j) {
                p.ae[idx] = ae;
            }
            if i > 0 && !hole(i - 1, j) {
                p.aw[idx] = ae;
            }
            if j + 1 < ny && !hole(i, j + 1) {
                p.an[idx] = an;
            }
            if j > 0 && !hole(i, j - 1) {
                p.as_[idx] = an;
            }
            if i + 1 == nx {
                p.extra_diag[idx] = 2.0 * ae;
            }
        }
    }
    let mut state = seed | 1;
    for idx in 0..nx * ny {
        state ^= state << 13;
        state ^= state >> 7;
        state ^= state << 17;
        p.rhs[idx] = if p.active[idx] {
            1e-3 * ((state >> 11) as f64 / (1u64 << 53) as f64 - 0.5)
        } else {
            0.0
        };
    }
    p
}

struct DeviceLevel {
    n: usize,
    nx: i32,
    cells: CudaSlice<u32>,
    red: CudaSlice<u32>,
    black: CudaSlice<u32>,
    coarse_of: CudaSlice<u32>,
    children_ptr: CudaSlice<u32>,
    children_idx: CudaSlice<u32>,
    n_cells: usize,
    n_red: usize,
    n_black: usize,
    ae: CudaSlice<f32>,
    aw: CudaSlice<f32>,
    an: CudaSlice<f32>,
    as_: CudaSlice<f32>,
    ap: CudaSlice<f32>,
    b: CudaSlice<f32>,
    x: CudaSlice<f32>,
    r: CudaSlice<f32>,
}

fn cfg(n_items: usize, k: usize) -> LaunchConfig {
    LaunchConfig {
        grid_dim: ((n_items as u32).div_ceil(256).max(1), k as u32, 1),
        block_dim: (256, 1, 1),
        shared_mem_bytes: 0,
    }
}

#[test]
#[ignore]
fn batched_device_vcycle_go_no_go() {
    let (nx, ny) = (372usize, 62usize);
    let prob = problem(nx, ny, 3);
    let params = MultigridParameters {
        smoother: MgSmoother::RedBlack,
        ..MultigridParameters::default()
    };
    let levels: Vec<LevelExport> = export_hierarchy(&prob, &params);
    let depth = levels.len();
    println!(
        "  hierarchy: {depth} levels, cells per level {:?}",
        levels.iter().map(|l| l.cells.len()).collect::<Vec<_>>()
    );

    // Device.
    let ctx = CudaContext::new(0).expect("CUDA context");
    let stream = ctx.default_stream();
    let arch = std::env::var("RTX_CUDA_ARCH").unwrap_or_else(|_| "sm_120".to_string());
    let ptx = compile_ptx_with_opts(
        KERNELS,
        CompileOptions {
            arch: Some(Box::leak(arch.clone().into_boxed_str())),
            ..Default::default()
        },
    )
    .expect("nvrtc");
    let module = ctx.load_module(ptx).expect("module");
    let f_half = module.load_function("mg_rb_half").unwrap();
    let f_res = module.load_function("mg_residual").unwrap();
    let f_restrict = module.load_function("mg_restrict").unwrap();
    let f_prolong = module.load_function("mg_prolong").unwrap();
    let f_zero = module.load_function("mg_zero").unwrap();
    let f_coarsest = module.load_function("mg_coarsest").unwrap();

    // The CPU reference V-cycle on the fine right-hand side.
    let n0 = nx * ny;
    let r_host: Vec<f64> = prob.rhs.clone();
    let mut z_ref = vec![0.0; n0];
    vcycle_f32_reference(&prob, &params, &r_host, &mut z_ref);
    let z_scale = z_ref.iter().fold(0.0_f64, |m, v| m.max(v.abs()));

    // The CPU f64 serial red-black V-cycle's wall time (the comparison point).
    let t0 = Instant::now();
    let reps = 20;
    for _ in 0..reps {
        let mut z = vec![0.0; n0];
        vcycle_f32_reference(&prob, &params, &r_host, &mut z);
    }
    let cpu_ms = t0.elapsed().as_secs_f64() * 1e3 / reps as f64;
    println!("  CPU f32 red-black V-cycle (incl. hierarchy build): {cpu_ms:.3} ms");

    let sweeps = 2usize;
    for &k in &[1usize, 8, 16] {
        // Upload: coefficients replicated K times, index lists shared.
        let mut dev: Vec<DeviceLevel> = Vec::with_capacity(depth);
        for l in &levels {
            let n = l.nx * l.ny;
            let rep = |v: &[f32]| -> Vec<f32> {
                let mut out = Vec::with_capacity(v.len() * k);
                for _ in 0..k {
                    out.extend_from_slice(v);
                }
                out
            };
            let up_u = |v: &[u32]| stream.memcpy_stod(v).unwrap();
            let up_f = |v: &[f32]| stream.memcpy_stod(v).unwrap();
            dev.push(DeviceLevel {
                n,
                nx: l.nx as i32,
                cells: up_u(&l.cells),
                red: up_u(&l.red),
                black: up_u(&l.black),
                coarse_of: up_u(if l.coarse_of.is_empty() { &[0u32][..] } else { &l.coarse_of }),
                children_ptr: up_u(if l.children_ptr.is_empty() { &[0u32][..] } else { &l.children_ptr }),
                children_idx: up_u(if l.children_idx.is_empty() { &[0u32][..] } else { &l.children_idx }),
                n_cells: l.cells.len(),
                n_red: l.red.len(),
                n_black: l.black.len(),
                ae: up_f(&rep(&l.ae)),
                aw: up_f(&rep(&l.aw)),
                an: up_f(&rep(&l.an)),
                as_: up_f(&rep(&l.as_)),
                ap: up_f(&rep(&l.ap)),
                b: stream.alloc_zeros::<f32>(n * k).unwrap(),
                x: stream.alloc_zeros::<f32>(n * k).unwrap(),
                r: stream.alloc_zeros::<f32>(n * k).unwrap(),
            });
        }
        stream.synchronize().unwrap();

        let r_f32: Vec<f32> = {
            let one: Vec<f32> = r_host.iter().map(|&v| v as f32).collect();
            let mut out = Vec::with_capacity(n0 * k);
            for _ in 0..k {
                out.extend_from_slice(&one);
            }
            out
        };
        let mut z_out = vec![0.0f32; n0 * k];

        let vcycle = |dev: &mut Vec<DeviceLevel>, r_f32: &[f32], z_out: &mut [f32]| {
            // Upload the residual into level 0's b.
            stream.memcpy_htod(r_f32, &mut dev[0].b).unwrap();
            let half = |lv: &mut DeviceLevel, colour: u8| {
                let (list, n_list) = if colour == 0 { (&lv.red, lv.n_red) } else { (&lv.black, lv.n_black) };
                let n_i = lv.n as i32;
                let n_list_i = n_list as i32;
                unsafe {
                    stream
                        .launch_builder(&f_half)
                        .arg(&n_list_i)
                        .arg(list)
                        .arg(&n_i)
                        .arg(&lv.ae)
                        .arg(&lv.aw)
                        .arg(&lv.an)
                        .arg(&lv.as_)
                        .arg(&lv.ap)
                        .arg(&lv.b)
                        .arg(&mut lv.x)
                        .arg(&lv.nx)
                        .launch(cfg(n_list, k))
                        .unwrap();
                }
            };
            let smooth = |lv: &mut DeviceLevel| {
                for _ in 0..sweeps {
                    half(lv, 0);
                    half(lv, 1);
                    half(lv, 1);
                    half(lv, 0);
                }
            };
            // Down.
            for l in 0..depth - 1 {
                {
                    let lv = &mut dev[l];
                    let (n_i, n_cells_i) = (lv.n as i32, lv.n_cells as i32);
                    unsafe {
                        stream
                            .launch_builder(&f_zero)
                            .arg(&n_cells_i)
                            .arg(&lv.cells)
                            .arg(&n_i)
                            .arg(&mut lv.x)
                            .launch(cfg(lv.n_cells, k))
                            .unwrap();
                    }
                    smooth(lv);
                    unsafe {
                        stream
                            .launch_builder(&f_res)
                            .arg(&n_cells_i)
                            .arg(&lv.cells)
                            .arg(&n_i)
                            .arg(&lv.ae)
                            .arg(&lv.aw)
                            .arg(&lv.an)
                            .arg(&lv.as_)
                            .arg(&lv.ap)
                            .arg(&lv.b)
                            .arg(&lv.x)
                            .arg(&mut lv.r)
                            .arg(&lv.nx)
                            .launch(cfg(lv.n_cells, k))
                            .unwrap();
                    }
                }
                let (fine, coarse) = dev.split_at_mut(l + 1);
                let (lf, lc) = (&fine[l], &mut coarse[0]);
                let (n_c_cells_i, n_f_i, n_c_i) = (lc.n_cells as i32, lf.n as i32, lc.n as i32);
                unsafe {
                    stream
                        .launch_builder(&f_restrict)
                        .arg(&n_c_cells_i)
                        .arg(&lc.cells)
                        .arg(&lf.children_ptr)
                        .arg(&lf.children_idx)
                        .arg(&n_f_i)
                        .arg(&n_c_i)
                        .arg(&lf.r)
                        .arg(&mut lc.b)
                        .launch(cfg(lc.n_cells, k))
                        .unwrap();
                }
            }
            // Coarsest.
            {
                let lv = &mut dev[depth - 1];
                let (k_i, n_cells_i, n_i, sw_i) = (k as i32, lv.n_cells as i32, lv.n as i32, 50i32);
                unsafe {
                    stream
                        .launch_builder(&f_coarsest)
                        .arg(&k_i)
                        .arg(&n_cells_i)
                        .arg(&lv.cells)
                        .arg(&n_i)
                        .arg(&lv.ae)
                        .arg(&lv.aw)
                        .arg(&lv.an)
                        .arg(&lv.as_)
                        .arg(&lv.ap)
                        .arg(&lv.b)
                        .arg(&mut lv.x)
                        .arg(&lv.nx)
                        .arg(&sw_i)
                        .launch(LaunchConfig {
                            grid_dim: ((k as u32).div_ceil(32), 1, 1),
                            block_dim: (32, 1, 1),
                            shared_mem_bytes: 0,
                        })
                        .unwrap();
                }
            }
            // Up.
            for l in (0..depth - 1).rev() {
                let (fine, coarse) = dev.split_at_mut(l + 1);
                let (lf, lc) = (&mut fine[l], &coarse[0]);
                let (n_cells_i, n_f_i, n_c_i) = (lf.n_cells as i32, lf.n as i32, lc.n as i32);
                unsafe {
                    stream
                        .launch_builder(&f_prolong)
                        .arg(&n_cells_i)
                        .arg(&lf.cells)
                        .arg(&lf.coarse_of)
                        .arg(&n_f_i)
                        .arg(&n_c_i)
                        .arg(&mut lf.x)
                        .arg(&lc.x)
                        .launch(cfg(lf.n_cells, k))
                        .unwrap();
                }
                smooth(lf);
            }
            stream.memcpy_dtoh(&dev[0].x, z_out).unwrap();
            stream.synchronize().unwrap();
        };

        // Correctness at K = 1 (and every batch member at K > 1).
        vcycle(&mut dev, &r_f32, &mut z_out);
        let mut worst = 0.0_f64;
        for m in 0..k {
            for idx in 0..n0 {
                let d = (z_out[m * n0 + idx] as f64 - z_ref[idx]).abs();
                worst = worst.max(d);
            }
        }
        println!("  K = {k}: device V-cycle vs CPU f32 reference: max |Δz| {worst:.3e} on a scale of {z_scale:.3e}");
        assert!(worst < 1e-4 * z_scale, "device V-cycle disagrees with the CPU one");

        // Timing: 50 V-cycles, per march.
        let reps = 50;
        vcycle(&mut dev, &r_f32, &mut z_out);
        let t0 = Instant::now();
        for _ in 0..reps {
            vcycle(&mut dev, &r_f32, &mut z_out);
        }
        let ms = t0.elapsed().as_secs_f64() * 1e3 / reps as f64;
        println!(
            "  K = {k}: {ms:.3} ms per batched V-cycle = {:.3} ms per march (CPU serial {cpu_ms:.3} ms; ratio {:.2}×)",
            ms / k as f64,
            cpu_ms / (ms / k as f64)
        );
    }
}