//! Calibration: my integer `jpegai_synthesis::decode_y` vs the float reference —
//! run the SAME latent through `synthesis16.onnx` (ort) and through `decode_y`,
//! compare the Y planes (max abs diff + PSNR). This is the PSNR pass that tells
//! whether the whole-net integer executor is numerically faithful (it exposes the
//! approximated gates). Needs `neural` (ort). Skipped if assets absent.
//!
//! env: WAI_JPEGAI_SYNTH16_Y = path to synthesis16.onnx (model_y),
//!      WAI_JPEGAI_Y_WEIGHTS = path to the exported 16-bit weights JSON.
#![cfg(all(feature = "neural", feature = "legacy_int"))]

use ort::session::{builder::GraphOptimizationLevel, Session};
use ort::value::Value;
use wai::jpegai_synthesis::{decode_y, Weights};

fn ort_synth(path: &str, x: Vec<f32>) -> (Vec<f32>, Vec<usize>) {
    let mut s = Session::builder(wai::ort_environment().unwrap()).unwrap()
        .with_intra_threads(1).unwrap().with_inter_threads(1).unwrap()
        .with_optimization_level(GraphOptimizationLevel::Level3).unwrap()
        .commit_from_file(path).unwrap();
    let name = s.inputs().first().unwrap().name().to_string();
    let arr = ndarray::Array4::from_shape_vec((1, 160, 16, 16), x).unwrap();
    let v = Value::from_array(arr).unwrap();
    let out = s.run(ort::inputs![name.as_str() => v]).unwrap();
    let view = out[0].try_extract_array::<f32>().unwrap();
    (view.iter().copied().collect(), view.shape().to_vec())
}

#[test]
fn synthesis_int_vs_float_psnr() {
    let onnx = match std::env::var("WAI_JPEGAI_SYNTH16_Y") {
        Ok(p) if std::path::Path::new(&p).exists() => p,
        _ => { eprintln!("skip: set WAI_JPEGAI_SYNTH16_Y to synthesis16.onnx"); return; }
    };
    let wjson = match std::env::var("WAI_JPEGAI_Y_WEIGHTS") {
        Ok(p) if std::path::Path::new(&p).exists() => p,
        _ => { eprintln!("skip: set WAI_JPEGAI_Y_WEIGHTS to the weights JSON"); return; }
    };
    // deterministic small latent [160,16,16] (typical post-LSBS magnitude)
    let n = 160 * 16 * 16;
    let lf: Vec<f32> = (0..n).map(|i| (((i % 97) as f32) - 48.0) / 96.0).collect();

    let (yf, shape) = ort_synth(&onnx, lf.clone());
    let (fmn, fmx) = yf.iter().fold((f32::MAX, f32::MIN), |(a, b), &v| (a.min(v), b.max(v)));
    eprintln!("float synth out shape = {:?}, len {}, range [{:.3}, {:.3}]", shape, yf.len(), fmn, fmx);
    eprintln!("float sample [0..6] = {:?}", &yf[..6]);

    let li: Vec<i64> = lf.iter().map(|&x| (x * 65536.0).round() as i64).collect();
    let v: serde_json::Value = serde_json::from_slice(&std::fs::read(&wjson).unwrap()).unwrap();
    let wt = Weights::from_json(&v).unwrap();
    let yi = decode_y(&li, &wt);

    assert_eq!(yi.len(), yf.len(), "output size mismatch");
    let (mut maxd, mut mse) = (0f64, 0f64);
    for i in 0..yi.len() {
        let f = yf[i].clamp(0.0, 255.0) as f64;
        let d = (yi[i] as f64 - f).abs();
        maxd = maxd.max(d);
        mse += d * d;
    }
    mse /= yi.len() as f64;
    let psnr = if mse > 0.0 { 20.0 * (255.0f64).log10() - 10.0 * mse.log10() } else { f64::INFINITY };
    eprintln!("INT vs FLOAT synthesis:  max|Δ| = {:.2}   MSE = {:.2}   PSNR = {:.2} dB", maxd, mse, psnr);
    // the integer decode reproduces the float ONNX synthesis to within the 16-bit
    // PTQ quantization (~1 LSB); PSNR ≈ 58 dB. Guard against regressions.
    assert!(psnr > 45.0, "synthesis PSNR {psnr:.2} dB regressed (expected ~58)");
}