// Float64 reference time loop with exactly the structure of the GPU kernels: an element pass
// (forces, sigma1 term, radiation accumulation), a hit pass (contact forces, release), a node
// pass (gather, contact, lagged rigid-mode projection, leapfrog update, partial sums for the
// next projection), and the exciter update. See THEORY.md.
import { elementForces, strainEnergies } from "../model/forces.ts";
import { HOLD_HEIGHT } from "../model/exciter.ts";
import { DECIMATION_SPAN } from "../model/radiation.ts";
import type { Setup } from "../model/setup.ts";
import type { ChunkResult, Diagnostics, SimulatorBackend } from "../model/simulator.ts";

export class CpuSimulator implements SimulatorBackend {
  readonly kind = "cpu" as const;
  private n = 0;                       // sub-step counter
  private readonly q: Float64Array; private readonly v: Float64Array; private readonly nodeAcc: Float64Array;
  private readonly vwPrev: Float64Array;   // projected transverse velocity of the previous step (for the radiated acceleration)
  private readonly ef: Float64Array; private readonly fc: Float64Array;
  private readonly zs: Float64Array; private readonly vs: Float64Array; private readonly Fk: Float64Array; private readonly F: Float64Array;
  private readonly Fmax: Float64Array; private readonly ncont: Uint32Array;
  private partialsPrev = new Float64Array(12); private partialsNew = new Float64Array(12);   // [0,6): Rm.(M v), [6,12): Rm.(M q)
  private readonly acc: Float64Array;   // (E, DECIMATION_SPAN) decimation accumulators
  private readonly history: Float64Array;
  private pending: { j: number; front: number; back: number }[] = [];
  samplesDone = 0;
  readonly maxChunk = 8192;
  readonly setup: Setup; readonly movieEvery: number;

  /** ``movieEvery`` > 0 captures w every that many audio samples. */
  constructor(setup: Setup, movieEvery = 0) {
    this.setup = setup; this.movieEvery = movieEvery;
    const N = setup.mesh.n, E = setup.mesh.e, K = setup.schedule.K, P = setup.schedule.P;
    this.q = new Float64Array(5 * N); this.v = new Float64Array(5 * N); this.nodeAcc = new Float64Array(N); this.vwPrev = new Float64Array(N);
    this.ef = new Float64Array(15 * E); this.fc = new Float64Array(N);
    this.zs = new Float64Array(K).fill(HOLD_HEIGHT); this.vs = new Float64Array(K);
    this.Fk = new Float64Array(K * P); this.F = new Float64Array(K); this.Fmax = new Float64Array(K); this.ncont = new Uint32Array(K);
    this.acc = new Float64Array(E * DECIMATION_SPAN); this.history = new Float64Array(E * setup.radiation.ringLength);
  }

  private step() {
    const { shell: sh, mesh, schedule: sc, radiation: rad, dt, substeps: m } = this.setup;
    const { tri } = mesh, N = mesh.n, E = mesh.e, n = this.n;
    const sigma0 = this.setup.config.material.sigma0;
    // ---- element pass
    elementForces(sh, this.q, this.v, this.ef);
    if (n >= 1) {
      const s = n - 1, j0 = Math.floor(s / m), kk = s - j0 * m, W = DECIMATION_SPAN, taps = rad.taps;
      if (kk === 0 && j0 > 0) this.finalizeSample(j0 - W / 2);   // slot 0 = sample j0 - (W/2 - 1) of the previous j0
      for (let e = 0; e < E; e++) {
        const a = (this.nodeAcc[tri[3 * e]] + this.nodeAcc[tri[3 * e + 1]] + this.nodeAcc[tri[3 * e + 2]]) / 3;
        for (let w = 0; w < W; w++) this.acc[e * W + w] += a * taps[w * m + kk];
      }
    }
    // ---- hit pass
    this.fc.fill(0);
    for (let k = 0; k < sc.K; k++) {
      let wmax = -Infinity;
      const wp = new Float64Array(sc.P);
      for (let p = 0; p < sc.P; p++) {
        const b = (k * sc.P + p) * 3;
        let w = sc.offsets[k * sc.P + p];
        for (let i = 0; i < 3; i++) w += sc.weights[b + i] * this.q[5 * sc.nodes[b + i] + 2];
        wp[p] = w; wmax = Math.max(wmax, w);
      }
      if (n === sc.releaseStep[k]) { this.zs[k] = wmax + sc.gap[k]; this.vs[k] = -sc.velocity[k]; }
      const active = n >= sc.releaseStep[k];
      let Ftot = 0;
      for (let p = 0; p < sc.P; p++) {
        const delta = Math.max(wp[p] - this.zs[k], 0);
        const f = active ? sc.kPoint[k] * Math.pow(delta, sc.alpha[k]) : 0;
        this.Fk[k * sc.P + p] = f; Ftot += f;
        if (f > 0) { const b = (k * sc.P + p) * 3; for (let i = 0; i < 3; i++) this.fc[sc.nodes[b + i]] -= f * sc.weights[b + i]; }
      }
      this.F[k] = Ftot;
    }
    // ---- node pass
    const { M, Minv, Rm, adjOffsets, adjEntries } = sh, L = 5 * N, c = this.partialsPrev;
    this.partialsNew.fill(0);
    const f = new Float64Array(5);
    for (let nd = 0; nd < N; nd++) {
      f.fill(0);
      for (let a = adjOffsets[nd]; a < adjOffsets[nd + 1]; a++) {
        const ent = adjEntries[a], e = (ent / 3) | 0, i = ent % 3;
        for (let d = 0; d < 5; d++) f[d] += this.ef[15 * e + 5 * i + d];
      }
      f[2] += this.fc[nd];
      for (let d = 0; d < 5; d++) {
        const idx = 5 * nd + d;
        // lagged projection: remove the rigid content that the previous step's update introduced
        let vd = this.v[idx], qd = this.q[idx];
        for (let k = 0; k < 6; k++) { const r = Rm[k * L + idx]; vd -= c[k] * r; qd -= c[6 + k] * r; }
        const v1 = (vd + dt * f[d] * Minv[idx]) / (1 + 2 * sigma0 * dt);
        // radiated acceleration from consecutive projected velocities, so that the rigid increment
        // that the lagged projection removes one step later never radiates
        if (d === 2) { this.nodeAcc[nd] = (vd - this.vwPrev[nd]) / dt; this.vwPrev[nd] = vd; }
        const q1 = qd + dt * v1;
        this.q[idx] = q1; this.v[idx] = v1;
        for (let k = 0; k < 6; k++) { const r = Rm[k * L + idx] * M[idx]; this.partialsNew[k] += r * v1; this.partialsNew[6 + k] += r * q1; }
      }
    }
    [this.partialsPrev, this.partialsNew] = [this.partialsNew, this.partialsPrev];
    // ---- exciters
    for (let k = 0; k < sc.K; k++) {
      this.vs[k] += (dt * this.F[k]) / sc.mass[k]; this.zs[k] += dt * this.vs[k];
      this.Fmax[k] = Math.max(this.Fmax[k], this.F[k]); if (this.F[k] > 0) this.ncont[k]++;
    }
    this.n++;
  }

  /** Slot 0 (audio sample j) of the accumulators is complete: store it, shift the slots, evaluate the listener pressure. */
  private finalizeSample(j: number) {
    const { radiation: rad, mesh } = this.setup, E = mesh.e, D = rad.ringLength, W = DECIMATION_SPAN;
    for (let e = 0; e < E; e++) {
      if (j >= 0) this.history[e * D + (j % D)] = this.acc[e * W];
      for (let w = 0; w < W - 1; w++) this.acc[e * W + w] = this.acc[e * W + w + 1];
      this.acc[e * W + W - 1] = 0;
    }
    if (j < 0) return;
    let front = 0, back = 0;
    for (let e = 0; e < E; e++) {
      const jf = j - rad.i0F[e], jb = j - rad.i0B[e];
      if (jf >= 1) front += rad.coefF[e] * (this.history[e * D + (jf % D)] * (1 - rad.frF[e]) + this.history[e * D + ((jf - 1) % D)] * rad.frF[e]);
      if (jb >= 1) back += rad.coefB[e] * (this.history[e * D + (jb % D)] * (1 - rad.frB[e]) + this.history[e * D + ((jb - 1) % D)] * rad.frB[e]);
    }
    this.pending.push({ j, front, back });
  }

  async runChunk(nSamples: number): Promise<ChunkResult> {
    const m = this.setup.substeps, N = this.setup.mesh.n, frames: Float32Array[] = [], frameSamples: number[] = [];
    for (let s = 0; s < nSamples * m; s++) {
      if (this.movieEvery > 0 && this.n % (m * this.movieEvery) === 0) {
        const w = new Float32Array(N);
        for (let nd = 0; nd < N; nd++) w[nd] = this.q[5 * nd + 2];
        frames.push(w); frameSamples.push(this.n / m);
      }
      this.step();
    }
    this.samplesDone += nSamples;
    const out = this.pending; this.pending = [];
    const firstSample = out.length ? out[0].j : this.samplesDone;
    return { firstSample, pressureFront: Float32Array.from(out, (o) => o.front), pressureBack: Float32Array.from(out, (o) => o.back), frames, frameSamples };
  }

  async diagnostics(): Promise<Diagnostics> {
    const { shell: sh, schedule: sc, dt } = this.setup, N = sh.mesh.n;
    let ke = 0, woh = 0;
    for (let i = 0; i < 5 * N; i++) ke += 0.5 * sh.M[i] * this.v[i] * this.v[i];
    for (let nd = 0; nd < N; nd++) woh = Math.max(woh, Math.abs(this.q[5 * nd + 2] / sh.hNode[nd]));
    const { membrane, bending } = strainEnergies(sh, this.q);
    let ex = 0;
    for (let k = 0; k < sc.K; k++) ex += 0.5 * sc.mass[k] * this.vs[k] * this.vs[k];
    return { time: this.n * dt, kinetic: ke, membrane, bending, exciters: ex, maxWoverH: woh,
             hits: Array.from({ length: sc.K }, (_, k) => ({ peakForce: this.Fmax[k], contactTime: this.ncont[k] * dt, velocity: this.vs[k], height: this.zs[k] })) };
  }

  /** M-weighted rigid-mode content of the current velocity (the lagged projection leaves one step's increment). */
  rigidContent(): number {
    const { shell: sh } = this.setup, L = 5 * sh.mesh.n;
    let worst = 0;
    for (let k = 0; k < 6; k++) { let c = 0; for (let i = 0; i < L; i++) c += sh.Rm[k * L + i] * sh.M[i] * this.v[i]; worst = Math.max(worst, Math.abs(c)); }
    return worst;
  }

  async displacement(): Promise<Float32Array> {
    const N = this.setup.mesh.n, w = new Float32Array(N);
    for (let nd = 0; nd < N; nd++) w[nd] = this.q[5 * nd + 2];
    return w;
  }

  dispose() {}
}