// Web Worker that renders the cymbal2 stroke. The MATLAB implementation in
// ../matlab remains the source of truth; what varies is who runs the time
// loop. When the browser exposes WebGPU, the worker executes the unmodified
// setup stage (grid, per-wavenumber inverses, mallet, stroke schedule)
// through numbl, hands the operators to a WGSL port of cymbal2_sim.m
// (./gpuSim.ts), and feeds the returned acceleration back through the
// unmodified cymbal2_output.m; this is 20x or more faster than interpreting
// the loop. Without WebGPU, or when cymbal2_sim.m has been edited in the app
// (the WGSL mirrors the shipped step, so edits there must fall back), the
// original interpreter path runs instead. Requests carry a quality tier
//   preview -> oversample 1, everything else as given: the render covers
//              the full duration, so the decay can be judged, and the
//              azimuthal grid is kept because hard strokes need the full
//              wavenumber band to carry the cascade, and coarsening it
//              destabilizes the integration
//   full    -> the parameters as given
// and the UI shows elapsed time, plus integration progress on the GPU path.
// A cheap linear-modes stage (cymbal2_modes per wavenumber) is cached on the
// geometry so the spectrum plot can overlay the mode table immediately.

import { executeCode, RTV, type RuntimeValue, type RuntimeTensor } from 'numbl';
import { hashMFiles, DEFAULT_MFILES, EXECUTED_BY_APP, type MFile } from '../mfiles';
import { MODES_KEYS, type Cymbal2Params } from '../params';
import { requestGpuDevice, runGpuSim, type GpuSetup } from './gpuSim';

export interface RenderRequest {
  type: 'render';
  id: number;
  params: Cymbal2Params;
  quality: 'preview' | 'full';
  mfiles: MFile[];
}

export interface RenderResponse {
  type: 'result';
  id: number;
  quality: 'preview' | 'full';
  y: Float32Array;               // final signal, [-1, 1]
  sampleRate: number;
  forceT: Float32Array;          // contact force history (max-pooled), s
  forceN: Float32Array;          // and its values, N
  modeF: Float32Array;           // linear mode frequencies, Hz
  modeM: Float32Array;           // nodal diameters of each
  diag: {
    strokesPlayed: number;
    meanV: number;
    contactVel: number;
    peakForceN: number;
    contactMs: number;
    contactEvents: number;
    peakXiMm: number;
    peakDispMm: number;
    peakDispThickness: number;
    peakAccel: number;
    strikeRadiusM: number;
    pickupRadiusM: number;
    RsM: number;
    nsteps: number;
    simRate: number;
    nDropped: number;
    durationS: number;
    renderMs: number;
    oversample: number;
    nTheta: number;
    backend: 'webgpu' | 'cpu';
    warnings: string[];
  };
}

export interface ProgressMessage {
  type: 'progress';
  id: number;
  stage: 'parsing' | 'modes' | 'integrate' | 'output';
  frac: number;
}

export interface ErrorMessage {
  type: 'error';
  id: number;
  message: string;
  file?: string;
  line?: number;
}

interface ModesEntry {
  modeF: Float32Array;
  modeM: Float32Array;
}

const modesCache = new Map<string, ModesEntry>();
const MODES_CACHE_MAX = 4;

function lruGet<T>(cache: Map<string, T>, key: string): T | undefined {
  const v = cache.get(key);
  if (v !== undefined) {
    cache.delete(key);
    cache.set(key, v);
  }
  return v;
}

function lruSet<T>(cache: Map<string, T>, key: string, value: T, max: number): void {
  cache.delete(key);
  cache.set(key, value);
  while (cache.size > max) {
    const oldest = cache.keys().next().value as string;
    cache.delete(oldest);
  }
}

function keyFor(params: Cymbal2Params, keys: readonly (keyof Cymbal2Params)[]): string {
  return keys.map((k) => String(params[k])).join('|');
}

// MATLAB source lines assigning every parameter onto a struct P.
function pLiteral(params: Cymbal2Params): string {
  const lines: string[] = [];
  for (const [k, v] of Object.entries(params)) {
    if (k === 'bit_depth') continue;
    if (typeof v === 'number') lines.push(`P.${k} = ${v};`);
    else if (typeof v === 'boolean') lines.push(`P.${k} = ${v ? 1 : 0};`);
  }
  return lines.join('\n');
}

// An error in the MATLAB source, carrying the file and line where numbl put
// it. Reported to the UI, which stops the render and shows nothing playable.
class CodeError extends Error {
  file?: string;
  line?: number;
  constructor(message: string, file?: string, line?: number) {
    super(message);
    this.name = 'CodeError';
    this.file = file;
    this.line = line;
  }
}

const SKIP_WARNING = /skipping (\S+\.m) \((?:syntax error at line (\d+)|parse error)\)/;

// numbl does not abort on a file it cannot parse: it drops that file with a
// warning and carries on with the rest, so a render could otherwise finish
// (or fail much later with a confusing 'undefined function') while the real
// fault is a typo in a file the stage never reached. Every numbl call runs
// with console.warn captured, and a dropped file becomes a CodeError. Only
// the files the app executes count: the CLI entry points (cymbal2_demo.m and
// the rest) are in the editor but play no part in a render, so a typo there
// must not stop one.
function skipWarningToError(warnings: string[]): CodeError | null {
  for (const w of warnings) {
    const hit = SKIP_WARNING.exec(w);
    if (!hit) continue;
    const file = hit[1];
    if (!EXECUTED_BY_APP.has(file)) continue;
    const line = hit[2] ? Number(hit[2]) : undefined;
    return new CodeError(
      line !== undefined
        ? 'syntax error; the render was stopped'
        : 'the file cannot be parsed; the render was stopped',
      file,
      line,
    );
  }
  return null;
}

function runStage(mfiles: MFile[], source: string, inputs?: Record<string, RuntimeValue>) {
  const warnings: string[] = [];
  const prevWarn = console.warn;
  console.warn = (...args: unknown[]) => {
    warnings.push(args.map(String).join(' '));
    prevWarn(...(args as []));
  };
  try {
    const r = executeCode(
      source,
      { initialVariableValues: inputs, displayResults: false },
      mfiles,
      'stage.m',
    );
    const skipped = skipWarningToError(warnings);
    if (skipped) throw skipped;
    return r;
  } catch (err) {
    // A file numbl dropped is the better explanation than whatever the run
    // failed on afterwards (typically a call to the missing function).
    const skipped = skipWarningToError(warnings);
    if (skipped) throw skipped;
    throw err;
  } finally {
    console.warn = prevWarn;
  }
}

// Parse every .m file before any expensive stage runs, so a syntax error is
// reported immediately instead of after a stage (or, if the modes table is
// cached, after the whole integration). Executing a trivial script with the
// files attached is enough: numbl pre-parses the whole workspace.
function checkMFiles(mfiles: MFile[]): void {
  runStage(mfiles, '1;');
}

function asScalar(v: RuntimeValue): number {
  if (typeof v === 'number') return v;
  const t = v as RuntimeTensor;
  if (t && t.kind === 'tensor' && t.data.length === 1) return t.data[0];
  if (typeof v === 'boolean') return v ? 1 : 0;
  throw new Error('expected scalar result from numbl stage');
}

function toF32(v: RuntimeValue): Float32Array {
  const t = v as RuntimeTensor;
  if (!t || t.kind !== 'tensor') throw new Error('expected tensor result from numbl stage');
  const out = new Float32Array(t.data.length);
  for (let i = 0; i < t.data.length; i++) out[i] = t.data[i];
  return out;
}

// Max-pool the contact force history to about 2000 points for the plot.
function poolForce(fhist: Float32Array, simRate: number): { tpool: Float32Array; fpool: Float32Array } {
  const nsim = fhist.length;
  const np0 = Math.min(2000, nsim);
  const blk = Math.ceil(nsim / np0);
  const np = Math.ceil(nsim / blk);
  const fpool = new Float32Array(np);
  const tpool = new Float32Array(np);
  for (let b = 0; b < np; b++) {
    let m = 0;
    const end = Math.min(nsim, (b + 1) * blk);
    for (let i = b * blk; i < end; i++) m = Math.max(m, fhist[i]);
    fpool[b] = m;
    tpool[b] = (b * blk + blk / 2) / simRate;
  }
  return { tpool, fpool };
}

self.onmessage = (ev: MessageEvent<RenderRequest>) => {
  const msg = ev.data;
  if (msg.type !== 'render') return;

  const warnings: string[] = [];
  const origWarn = console.warn;
  console.warn = (...args: unknown[]) => {
    warnings.push(args.map(String).join(' '));
  };

  render(msg)
    .catch((err) => {
      // runStage turns a dropped file into a CodeError itself; this catches
      // the same warning from anything that reached numbl another way.
      const e = (skipWarningToError(warnings) ?? err) as {
        message?: string;
        file?: string;
        line?: number;
      };
      const m: ErrorMessage = {
        type: 'error',
        id: msg.id,
        message: e?.message ? String(e.message) : String(err),
        file: typeof e?.file === 'string' ? e.file : undefined,
        line: typeof e?.line === 'number' ? e.line : undefined,
      };
      (self as unknown as Worker).postMessage(m);
    })
    .finally(() => {
      console.warn = origWarn;
    });
};

// Everything the response assembly needs, produced by either backend.
interface SimOutcome {
  y32: Float32Array;
  peakAccel: number;
  nDropped: number;
  tpool: Float32Array;
  fpool: Float32Array;
  backend: 'webgpu' | 'cpu';
  d: {
    strokesPlayed: number;
    meanV: number;
    contactVel: number;
    peakForceN: number;
    contactMs: number;
    contactEvents: number;
    peakXiMm: number;
    peakDispM: number;
    peakDispThickness: number;
    nsteps: number;
    simRate: number;
    strikeRadiusM: number;
    pickupRadiusM: number;
    RsM: number;
  };
}

// The stroke-schedule block from cymbal2_sim.m, verbatim, so the GPU render
// integrates exactly the schedule the interpreter would.
const SCHEDULE_SRC = `
K = max(1, round(P.strike_count));
tstr = P.pre_delay_ms * 1e-3 + (0:K - 1) / max(1e-3, P.strike_rate_hz);
vstr = zeros(1, K);
state = P.seed * 2654435761;
for i = 1:K
  state = mod(state * 1664525 + 1013904223, 4294967296);
  x = state / 4294967296;
  vstr(i) = max(0.05, P.strike_velocity * (1 + P.velocity_spread * (2 * x - 1)));
end
gap = P.strike_gap_m;
if gap <= 0
  gap = 1.05 * max(vstr)^2 / (2 * P.strike_accel) + 0.005;
end
travel = gap ./ vstr + vstr / (2 * P.strike_accel);
tlaunch = tstr - travel;
off = -min(0, min(tlaunch));
tlaunch = tlaunch + off;
duration = off + (K - 1) / max(1e-3, P.strike_rate_hz) + P.tail_s;
nsteps = floor(duration * P.sample_rate * P.oversample);
`;

async function renderOnGpu(
  mfiles: MFile[],
  params: Cymbal2Params,
  onProgress: (frac: number) => void,
): Promise<SimOutcome> {
  // A WebGPU device first: if there is none, fail before spending seconds on
  // the setup stage.
  const device = await requestGpuDevice();
  try {
    const r = runStage(mfiles, `
${pLiteral(params)}
S = cymbal2_system(P);
MB = cymbal2_mallet(P);
G = S.G;
o_D1 = G.D1; o_D2 = G.D2;
o_F = G.F; o_Gi = G.Gi;
o_Rd = G.Rd; o_R2d = G.R2d;
o_AiB = S.AiB; o_AiC = S.AiC; o_Ai = S.Ai; o_Fi = S.Fi;
o_ASiB = MB.ASiB; o_ASiC = MB.ASiC;
o_vM = MB.vM; o_vR = MB.vR; o_vK = MB.vK; o_tipcol = MB.tipcol;
${SCHEDULE_SRC}
o_tlaunch = tlaunch; o_vstr = vstr;
c_nrt = G.nrt; c_Nr = G.Nr; c_nth = G.nth; c_hr = G.hr; c_hth = G.hth;
c_ht = S.ht; c_Eh = S.Eh; c_Rs = S.Rs; c_rhoh = S.rhoh;
c_fsc = S.forceScale; c_nlr0 = S.eomRow0; c_nDropped = S.nDropped;
c_ls = S.ls; c_js = S.js; c_lp = S.lp; c_jp = S.jp;
c_strikeR = S.strikeRadiusM; c_pickupR = S.pickupRadiusM;
c_tipi = MB.tipi; c_nfree = MB.nfree; c_ps = MB.ps;
c_holdSteps = max(1, round(P.strike_hold_ms * 1e-3 / S.ht));
c_missDepth = gap + 0.06;
c_gap = gap; c_K = K; c_nsteps = nsteps;
c_simRate = P.sample_rate * P.oversample;
`);
    const v = r.variableValues;
    const s = (n: string) => asScalar(v[n]);
    const t = (n: string) => toF32(v[n]);
    const setup: GpuSetup = {
      ops: {
        D1: t('o_D1'), D2: t('o_D2'),
        F: t('o_F'), Gi: t('o_Gi'),
        Rd: t('o_Rd'), R2d: t('o_R2d'),
        AiB: t('o_AiB'), AiC: t('o_AiC'), Ai: t('o_Ai'), Fi: t('o_Fi'),
        ASiB: t('o_ASiB'), ASiC: t('o_ASiC'),
        vM: t('o_vM'), vR: t('o_vR'), vK: t('o_vK'), tipcol: t('o_tipcol'),
      },
      tlaunch: Array.from((v.o_tlaunch as RuntimeTensor).data),
      vstr: Array.from((v.o_vstr as RuntimeTensor).data),
      nrt: s('c_nrt'), Nr: s('c_Nr'), nth: s('c_nth'),
      hr: s('c_hr'), hth: s('c_hth'),
      ht: s('c_ht'), Eh: s('c_Eh'), Rs: s('c_Rs'), rhoh: s('c_rhoh'),
      fsc: s('c_fsc'), nlr0: s('c_nlr0'),
      ls: s('c_ls'), js: s('c_js'), lp: s('c_lp'), jp: s('c_jp'),
      tipi: s('c_tipi'), nfree: s('c_nfree'), ps: s('c_ps'),
      KM: params.head_stiffness, alM: params.head_alpha,
      aH: params.strike_accel, aHr: params.retract_accel,
      holdSteps: s('c_holdSteps'), missDepth: s('c_missDepth'),
      gap: s('c_gap'), K: s('c_K'), nsteps: s('c_nsteps'),
    };
    const simRate = s('c_simRate');

    const sim = await runGpuSim(device, setup, (done, n) => onProgress(done / n));

    // The output stage through the unmodified cymbal2_output.m.
    const accT = RTV.tensor(Float64Array.from(sim.acc), [1, sim.acc.length]);
    const ro = runStage(mfiles, `
${pLiteral(params)}
[y, peakAccel] = cymbal2_output(P, acc);
`, { acc: accT });

    // Contact bookkeeping from the force history, as cymbal2_sim does it.
    let peakF = 0, contactSteps = 0, events = 0, prev = 0;
    for (let i = 0; i < sim.fhist.length; i++) {
      const f = sim.fhist[i];
      if (f > peakF) peakF = f;
      if (f > 0) { contactSteps++; if (prev === 0) events++; }
      prev = f;
    }
    const { tpool, fpool } = poolForce(sim.fhist, simRate);
    const meanV = setup.vstr.reduce((a, b) => a + b, 0) / setup.vstr.length;

    return {
      y32: toF32(ro.variableValues.y),
      peakAccel: asScalar(ro.variableValues.peakAccel),
      nDropped: s('c_nDropped'),
      tpool,
      fpool,
      backend: 'webgpu',
      d: {
        strokesPlayed: sim.strokesPlayed,
        meanV,
        contactVel: sim.contactVel[0] ?? 0,
        peakForceN: peakF,
        contactMs: (1000 * contactSteps) / simRate,
        contactEvents: events,
        peakXiMm: sim.peakXi * 1000,
        peakDispM: sim.peakDisp,
        peakDispThickness: sim.peakDisp / params.thickness_m,
        nsteps: setup.nsteps,
        simRate,
        strikeRadiusM: s('c_strikeR'),
        pickupRadiusM: s('c_pickupR'),
        RsM: setup.Rs,
      },
    };
  } finally {
    device.destroy();
  }
}

function renderOnCpu(mfiles: MFile[], params: Cymbal2Params): SimOutcome {
  const r = runStage(mfiles, `
${pLiteral(params)}
S = cymbal2_system(P);
MB = cymbal2_mallet(P);
[acc, dinfo, fhist] = cymbal2_sim(P, S, MB);
[y, peakAccel] = cymbal2_output(P, acc);
% max-pool the contact force history to about 2000 points for the plot
nsim = length(fhist);
np = min(2000, nsim);
blk = ceil(nsim / np);
np = ceil(nsim / blk);
fpool = max(reshape([fhist, zeros(1, np * blk - nsim)], blk, np), [], 1);
tpool = ((0:np - 1) * blk + blk / 2) / dinfo.simRate;
d_strokes = dinfo.strokesPlayed;
d_meanV = dinfo.meanV;
d_cvel = dinfo.contactVel(1);
d_peakF = dinfo.peakForceN;
d_contactMs = dinfo.contactMs;
d_events = dinfo.contactEvents;
d_xi = dinfo.peakXiMm;
d_disp = dinfo.peakDispM;
d_dispH = dinfo.peakDispThickness;
d_nsteps = dinfo.nsteps;
d_simRate = dinfo.simRate;
d_dropped = S.nDropped;
d_strikeR = S.strikeRadiusM;
d_pickupR = S.pickupRadiusM;
d_Rs = S.Rs;
`);
  const v = r.variableValues;
  return {
    y32: toF32(v.y),
    peakAccel: asScalar(v.peakAccel),
    nDropped: asScalar(v.d_dropped),
    tpool: toF32(v.tpool),
    fpool: toF32(v.fpool),
    backend: 'cpu',
    d: {
      strokesPlayed: asScalar(v.d_strokes),
      meanV: asScalar(v.d_meanV),
      contactVel: asScalar(v.d_cvel),
      peakForceN: asScalar(v.d_peakF),
      contactMs: asScalar(v.d_contactMs),
      contactEvents: asScalar(v.d_events),
      peakXiMm: asScalar(v.d_xi),
      peakDispM: asScalar(v.d_disp),
      peakDispThickness: asScalar(v.d_dispH),
      nsteps: asScalar(v.d_nsteps),
      simRate: asScalar(v.d_simRate),
      strikeRadiusM: asScalar(v.d_strikeR),
      pickupRadiusM: asScalar(v.d_pickupR),
      RsM: asScalar(v.d_Rs),
    },
  };
}

async function render(msg: RenderRequest): Promise<void> {
  const { id, mfiles, quality } = msg;
  const params = { ...msg.params };
  if (quality === 'preview') {
    params.oversample = 1;
  }

  const post = (m: ProgressMessage) => (self as unknown as Worker).postMessage(m);

  // Cheap, and it must come first: a file that does not parse stops the
  // render here rather than partway through a stage.
  post({ type: 'progress', id, stage: 'parsing', frac: 0.01 });
  checkMFiles(mfiles);

  const codeKey = hashMFiles(mfiles);

  // ---- linear modes, cheap, cached on geometry ----
  const modesKey = codeKey + '|' + keyFor(params, MODES_KEYS);
  let modesEntry = lruGet(modesCache, modesKey);
  if (!modesEntry) {
    post({ type: 'progress', id, stage: 'modes', frac: 0.05 });
    const r = runStage(mfiles, `
${pLiteral(params)}
mF = [];
mM = [];
for mw = 0:min(10, floor(P.n_theta / 2))
  f = cymbal2_modes(P, mw);
  f = f(f > 0.5 & f < 4000);
  nk = min(4, length(f));
  mF = [mF; f(1:nk)];
  mM = [mM; mw * ones(nk, 1)];
end
`);
    modesEntry = {
      modeF: toF32(r.variableValues.mF),
      modeM: toF32(r.variableValues.mM),
    };
    lruSet(modesCache, modesKey, modesEntry, MODES_CACHE_MAX);
  }

  // ---- the time integration ----
  post({ type: 'progress', id, stage: 'integrate', frac: 0.1 });
  const t0 = performance.now();
  const warnings: string[] = [];

  // The WGSL mirrors the shipped cymbal2_sim.m, so an edited step falls back
  // to the interpreter, where the edit is live. Edits to the setup and output
  // stages flow through numbl on both paths.
  const simSource = mfiles.find((f) => f.name === 'cymbal2_sim.m')?.source;
  const simShipped = DEFAULT_MFILES.find((f) => f.name === 'cymbal2_sim.m')?.source;
  const simEdited = simSource !== simShipped;

  let out: SimOutcome | null = null;
  if (!simEdited) {
    try {
      out = await renderOnGpu(mfiles, params, (frac) =>
        post({ type: 'progress', id, stage: 'integrate', frac: 0.1 + 0.85 * frac }),
      );
    } catch (err) {
      const detail = err instanceof Error ? err.message : String(err);
      // An error in the MATLAB source (it carries a file, or names the
      // divergence guard) is a fault to report, not a reason to retry the
      // same code on the CPU. Only a genuine GPU problem falls back.
      const fromCode =
        err instanceof CodeError ||
        typeof (err as { file?: unknown }).file === 'string' ||
        (err as { name?: string }).name === 'RuntimeError';
      if (fromCode || /diverged/.test(detail)) throw err;
      // console.warn is intercepted during a render; use info so the reason
      // is visible in the devtools console.
      console.info(`cymbal2: WebGPU unavailable (${detail}); rendering on the CPU`);
    }
  }
  if (!out) {
    out = renderOnCpu(mfiles, params);
    if (simEdited) {
      // Not a fault: just say why the render took the slow path.
      warnings.push('cymbal2_sim.m differs from the shipped source, so the render used the interpreter (the GPU path mirrors the shipped step)');
    }
  }
  const renderMs = performance.now() - t0;

  post({ type: 'progress', id, stage: 'output', frac: 0.97 });
  const { y32, peakAccel, nDropped } = out;

  if (nDropped > 0) {
    warnings.push(
      `${nDropped} azimuthal columns were truncated: the radial grid cannot resolve their inner boundary layer; raise n_radial`,
    );
  }
  if (!params.normalize) {
    let peak = 0;
    for (let i = 0; i < y32.length; i++) peak = Math.max(peak, Math.abs(y32[i]));
    if (peak > 1.0) {
      warnings.push(
        `clipping: peak acceleration ${peakAccel.toExponential(2)} m/s² exceeds full_scale_accel`,
      );
    }
  }

  const resp: RenderResponse = {
    type: 'result',
    id,
    quality,
    y: y32,
    sampleRate: params.sample_rate,
    forceT: out.tpool,
    forceN: out.fpool,
    modeF: modesEntry.modeF.slice(),
    modeM: modesEntry.modeM.slice(),
    diag: {
      strokesPlayed: out.d.strokesPlayed,
      meanV: out.d.meanV,
      contactVel: out.d.contactVel,
      peakForceN: out.d.peakForceN,
      contactMs: out.d.contactMs,
      contactEvents: out.d.contactEvents,
      peakXiMm: out.d.peakXiMm,
      peakDispMm: out.d.peakDispM * 1000,
      peakDispThickness: out.d.peakDispThickness,
      peakAccel,
      strikeRadiusM: out.d.strikeRadiusM,
      pickupRadiusM: out.d.pickupRadiusM,
      RsM: out.d.RsM,
      nsteps: out.d.nsteps,
      simRate: out.d.simRate,
      nDropped,
      durationS: y32.length / params.sample_rate,
      renderMs,
      oversample: params.oversample,
      nTheta: params.n_theta,
      backend: out.backend,
      warnings,
    },
  };
  (self as unknown as Worker).postMessage(resp, [
    resp.y.buffer,
    resp.forceT.buffer,
    resp.forceN.buffer,
    resp.modeF.buffer,
    resp.modeM.buffer,
  ]);
}