perf: harden realtime localization pipeline

2026-08-28 18:58:48 +09:00
parent 35c33e360e
commit 256558e601
13 changed files with 765 additions and 84 deletions
@@ -24,6 +24,8 @@ const localizer = createStreamingLocalizer({
  scan: { xMin: -1, xMax: 1, yMin: -1, yMax: 1, z: 2, columns: 50, rows: 50 },
  bands: [{ id: "machine", minHz: 900, maxHz: 1100, color: "#65f2bf" }],
  stft: { fftSize: 512, hopSize: 256 },
+  maxCompensationDb: 24,
+  peakDetection: { thresholdRatio: 0.5, maximumDetections: 8 },
  analysisWindowSamples: 4_800,
  analysisHopSamples: 4_800,
  maximumBufferedSamples: 19_200,
@@ -38,20 +40,47 @@ const results = await localizer.analyzeAvailable({ signal: abortController.signa
 for (const result of results) {
  const { position, rawPower } = result.locations[0];
  drawMarker(position, rawPower, result.centerTime);
+  if (!result.deadlineMet) reportOverload(result.processingTimeMs, result.deadlineMs);
 }
 ```

+`locations`는 band 순서, 같은 band 안에서는 raw power 내림차순의 평평한 목록입니다. 각 항목은
+기존 `{ bandId, index, position, rawPower }`에 `rank`를 더하며 같은 band의 여러 local peak를
+포함할 수 있습니다. `peakDetection.thresholdRatio` 기본값은 최강 peak의 0.5배,
+`minimumPeakDistanceM` 기본값은 scan cell 대각선, `maximumDetections` 기본값은 8입니다.
+기존처럼 `locations[0]`만 읽으면 최강 후보를 그대로 얻습니다.
+
+Streaming 계산이 중단되거나 실패하면 해당 분석창과 sequence를 commit하지 않고 다음 호출에서 같은
+창을 재시도합니다. 성공한 뒤에만 hop·sequence·`completedAnalyses`를 commit하며, 그 뒤 호출되는
+`onResult` callback의 예외는 `failedAnalyses`에 포함하지 않습니다. 계산 중 capture overflow가 나면
+ring의 오래된 sample은 평소와 같이 drop telemetry에 기록됩니다.
+
+한 블록 결과의 `processingTimeMs`는 DAS와 local-peak 검출에 걸린 실제 시간을 기록합니다. Streaming
+결과는 `deadlineMs = analysisHopSamples / sampleRate × 1000`과 `deadlineMet`을 더하며,
+`getState()`는 최근/최대 처리 시간과 `deadlineMisses` 누계를 제공합니다. `deadlineMisses` 또는
+`droppedSamples`가 증가하면 grid·대역 bin·분석창·update rate를 장치 성능에 맞게 줄여야 합니다.
+
+깊은 응답 notch의 역필터가 sensor noise까지 폭증시키지 않도록 inverse response의 양의 boost는 기본
+24 dB로 제한합니다. Experiment의 `beamforming.maxCompensationDb` 또는 portable plan의 top-level
+`maxCompensationDb`를 0~120 dB로 조절할 수 있습니다. 이 값은 boost만 제한하며 attenuation과 phase
+보정은 유지합니다. 선택 bin 전체의 gain/phase calibration byte도 plan 생성 전에 산정하므로 lazy
+cache가 나중에 `maximumPlanBytes`를 넘지 않습니다.
+
 한 블록만 처리할 때는 `createLocalizationPlan()`과 `localizeMultichannel()`을 사용합니다.
 `getState()`은 absolute sample origin, buffered/dropped sample, 대기 window 수와 고정 input
 buffer byte 수를 돌려주므로 capture adapter가 지연을 감시할 수 있습니다. 실제 장치의 clock,
 channel framing과 transport는 하드웨어별 adapter 책임이며 이 모듈은 그 다음 단계부터 시작합니다.

-반복 CPU 분석 계획은 거리 행렬, bin calibration, 제한 내의 Float64 steering matrix와 임시 bin map을
-재사용합니다. 예를 들어 192채널·50×50·2 bins를 10회 분석하면 거리 계산은 4,800,000회에서
-480,000회로, steering point/microphone pair는 9,600,000회에서 960,000회로 줄어듭니다
-(sin/cos 호출 17,280,000회 회피). 이 경우 약 16.5 MiB를 유지하며, 기본 64 MiB steering 한도를
-넘는 설정은 steering cache만 끕니다. 자동 테스트는 cache 전후 raw/display Float32 map의 전 요소
-bit-exact 일치를 고정합니다.
+반복 CPU 분석 계획은 거리 행렬, bin calibration, Hann window, Float64 FFT workspace, 임시 bin map과
+예산 내의 Float64 steering matrix를 재사용합니다. 예를 들어 192채널·50×50·2 bins를 10회 분석하면
+거리 계산은 4,800,000회에서 480,000회로, steering point/microphone pair는 9,600,000회에서
+960,000회로 줄어듭니다(sin/cos 호출 17,280,000회 회피). 이 경우 약 16.5 MiB를 유지합니다.
+
+기본 64 MiB보다 전체 wide-band steering matrix가 크더라도 cache 전체를 끄지 않습니다. 선택 bin
+앞쪽부터 예산에 들어가는 만큼만 Float64로 유지하고 나머지는 고정 microphone workspace에서 계산합니다.
+`analysisCache`는 cached bin capacity, uncached 사용 횟수·pair 수, STFT transform·FFT butterfly·DAS
+complex MAC 수와 실제 workspace allocation/reuse 수를 제공합니다. 자동 테스트는 부분 cache와
+cache 없음 사이 raw/display Float32 map의 전 요소 bit-exact 일치를 고정합니다.

 ## 이동 음원과 분석 시간창

@@ -75,8 +104,10 @@ ground-truth 위치에서 peak가 생기는 회귀 테스트로 이 부호를

 마이크 `sensitivityDb`와 `phaseOffsetDeg`는 수신 channel에 편차로 적용하고, DAS에서는 알려진
 calibration 값의 역 gain/phase로 보정합니다. raw power map과 표시용 normalized map은 서로 다른
-buffer로 보존합니다. bin spectrum은 제한된 batch로 처리하며 point batch 사이에서 event loop를
-양보해 진행 callback과 `AbortSignal` 취소가 동작합니다.
+buffer로 보존합니다. bin spectrum은 제한된 batch로 처리합니다. cooperative yield는 지원 환경에서
+`scheduler.yield()`, 그 외에는 `MessageChannel`을 사용하고 timer task에도 주기적으로 기회를 주므로
+진행 callback과 `AbortSignal` 취소를 유지하면서 연속 `setTimeout(0)` clamp를 피합니다. portable
+plan의 기본 batch는 point 256개·STFT transform 128개이며 `computeOptions`로 조정할 수 있습니다.

 마이크별 `frequencyResponse`는 2~64개의 `{ frequencyHz, gainDb, phaseDeg? }` 점으로 정의합니다.
 `interpolateFrequencyResponse`는 dB와 wrapped phase를 구간 선형 보간하고 측정 범위 밖에서는 가장
@@ -112,6 +143,6 @@ Scene의 `acoustics`를 생략하거나 `{ "model": "free-field" }`로 두면
 snapshot 합성은 전체 signal과 가장 늦은 sparse path를 포함하는 delay FFT working set을
 512 MiB로 제한합니다. 실측·장시간 입력은 위 streaming localizer의 bounded ring buffer로
 전체 녹음을 메모리에 보관하지 않고 처리할 수 있습니다. CPU analysis plan도 기본 512 MiB
-hard cap을 할당 전에 검사하며 steering cache는 별도의 64 MiB 한도를 넘으면 자동으로 끕니다.
+hard cap을 할당 전에 검사하며 steering cache는 별도의 64 MiB 안에서 가능한 bin만 유지합니다.

 192채널·50×50 fixture는 브라우저 통합 테스트에서 실제 합성부터 DAS peak 검출까지 실행합니다.
@@ -0,0 +1,40 @@
+let messageChannelYield;
+let cooperativeTurn = 0;
+
+export const COOPERATIVE_TIMER_TURN_INTERVAL = 8;
+
+function createMessageChannelYield() {
+  if (typeof globalThis.MessageChannel !== "function") return undefined;
+  const channel = new globalThis.MessageChannel();
+  const queue = [];
+  channel.port1.onmessage = () => queue.shift()?.();
+  channel.port1.unref?.();
+  channel.port2.unref?.();
+  return () => new Promise((resolve) => {
+    queue.push(resolve);
+    channel.port2.postMessage(0);
+  });
+}
+
+/**
+ * Yield one cooperative task without relying on nested zero-delay timers.
+ * scheduler.yield() is preferred when the host implements it; MessageChannel
+ * keeps the fallback task-based and avoids browser timer clamping.
+ */
+export function cooperativeYield() {
+  cooperativeTurn += 1;
+  if (cooperativeTurn % COOPERATIVE_TIMER_TURN_INTERVAL === 0) {
+    return new Promise((resolve) => setTimeout(resolve, 0));
+  }
+  const hostScheduler = globalThis.scheduler;
+  if (typeof hostScheduler?.yield === "function") {
+    try {
+      return hostScheduler.yield();
+    } catch {
+      // A partial scheduler implementation must not make computation unusable.
+    }
+  }
+  messageChannelYield ??= createMessageChannelYield();
+  if (messageChannelYield) return messageChannelYield();
+  return new Promise((resolve) => setTimeout(resolve, 0));
+}
@@ -1,6 +1,11 @@
 import { dbToLinear, createScanPoints, distance3, findPeak, scanIndexToPosition, TAU } from "./math.js";
-import { interpolateFrequencyResponse } from "./frequency-response.js";
-import { binsForBand, extractStftBins, getFrameCount } from "./stft.js";
+import {
+  interpolateFrequencyResponse,
+  inverseResponseGainDb,
+  requireMaxCompensationDb,
+} from "./frequency-response.js";
+import { cooperativeYield } from "./cooperative-yield.js";
+import { binsForBand, createStftWorkspace, extractStftBins, getFrameCount } from "./stft.js";

 function throwIfAborted(signal) {
  if (signal?.aborted) {
@@ -10,10 +15,6 @@ function throwIfAborted(signal) {
  }
 }

-function yieldToEventLoop() {
-  return new Promise((resolve) => setTimeout(resolve, 0));
-}
-
 export const DEFAULT_MAXIMUM_CPU_STEERING_CACHE_BYTES = 64 * 1024 * 1024;
 export const DEFAULT_MAXIMUM_CPU_ANALYSIS_PLAN_BYTES = 512 * 1024 * 1024;
 const DAS_ANALYSIS_PLAN_KIND = "micarrsim.das-analysis-plan/1";
@@ -47,12 +48,20 @@ function createBandIndexLookup(bandBins) {
  return lookup;
 }

-function createBinCalibration(arrayDefinition, bin, sampleRate, fftSize, compensateFrequencyResponse) {
+function createBinCalibration(
+  arrayDefinition,
+  bin,
+  sampleRate,
+  fftSize,
+  compensateFrequencyResponse,
+  maximumCompensationDb,
+) {
  const microphoneCount = arrayDefinition.microphones.length;
  const gains = new Float64Array(microphoneCount);
  const phases = new Float64Array(microphoneCount);
  const frequencyHz = (bin * sampleRate) / fftSize;
  const response = {};
+  const maxCompensationDb = requireMaxCompensationDb(maximumCompensationDb);
  for (let microphoneIndex = 0; microphoneIndex < microphoneCount; microphoneIndex += 1) {
    const microphone = arrayDefinition.microphones[microphoneIndex];
    let responseGainDb = 0;
@@ -62,8 +71,11 @@ function createBinCalibration(arrayDefinition, bin, sampleRate, fftSize, compens
      responseGainDb = response.gainDb;
      responsePhaseDeg = response.phaseDeg;
    }
+    const responseCompensationDb = compensateFrequencyResponse
+      ? inverseResponseGainDb(responseGainDb, maxCompensationDb)
+      : 0;
    gains[microphoneIndex] = dbToLinear(
-      -(microphone.sensitivityDb ?? 0) - responseGainDb,
+      -(microphone.sensitivityDb ?? 0) + responseCompensationDb,
    );
    phases[microphoneIndex] =
      ((microphone.phaseOffsetDeg ?? 0) + responsePhaseDeg) * Math.PI / 180;
@@ -119,7 +131,9 @@ async function createSteeringMatrix(options) {
      outputImaginary: imaginary,
      outputOffset: offset,
    });
-    if ((pointIndex + 1) % pointBatchSize === 0) await yieldToEventLoop();
+    if ((pointIndex + 1) % pointBatchSize === 0 && pointIndex + 1 < pointCount) {
+      await cooperativeYield();
+    }
  }
  return { real, imaginary };
 }
@@ -136,6 +150,7 @@ export async function computeBinMap(options) {
    fftSize,
    speedOfSound,
    compensateFrequencyResponse = true,
+    maxCompensationDb,
    calibration: providedCalibration,
    steeringMatrix,
    output: providedOutput,
@@ -156,6 +171,7 @@ export async function computeBinMap(options) {
    sampleRate,
    fftSize,
    compensateFrequencyResponse,
+    maxCompensationDb,
  );
  const phaseRealValues = steeringMatrix?.real ?? steeringReal;
  const phaseImaginaryValues = steeringMatrix?.imaginary ?? steeringImaginary;
@@ -212,7 +228,9 @@ export async function computeBinMap(options) {
      accumulatedPower += (sumReal * sumReal + sumImaginary * sumImaginary) * microphoneScale * microphoneScale;
    }
    output[pointIndex] = accumulatedPower / frameCount;
-    if ((pointIndex + 1) % pointBatchSize === 0) await yieldToEventLoop();
+    if ((pointIndex + 1) % pointBatchSize === 0 && pointIndex + 1 < pointCount) {
+      await cooperativeYield();
+    }
  }
  return output;
 }
@@ -255,32 +273,45 @@ export function createDasAnalysisPlan(options) {
  const pointCount = experiment.scan.columns * experiment.scan.rows;
  const pointMicrophonePairs = pointCount * microphoneCount;
  const distanceMatrixBytes = pointMicrophonePairs * Float32Array.BYTES_PER_ELEMENT;
+  const stftWorkspaceBytes = fftSize * Float64Array.BYTES_PER_ELEMENT * 3;
+  const calibrationBytes =
+    selectedBins.length * microphoneCount * Float64Array.BYTES_PER_ELEMENT * 2;
  const retainedBaseBytes =
    distanceMatrixBytes +
    pointCount * Float32Array.BYTES_PER_ELEMENT +
-    microphoneCount * Float64Array.BYTES_PER_ELEMENT * 2;
+    microphoneCount * Float64Array.BYTES_PER_ELEMENT * 2 +
+    stftWorkspaceBytes;
+  const retainedNonSteeringBytes = retainedBaseBytes + calibrationBytes;
  if (
    !Number.isSafeInteger(pointMicrophonePairs) ||
    !Number.isSafeInteger(retainedBaseBytes) ||
-    retainedBaseBytes > maximumPlanBytes
+    !Number.isSafeInteger(calibrationBytes) ||
+    !Number.isSafeInteger(retainedNonSteeringBytes) ||
+    retainedNonSteeringBytes > maximumPlanBytes
  ) {
    throw new RangeError(
-      "CPU analysis plan requires " + retainedBaseBytes.toLocaleString() +
-      " base bytes, above maximumPlanBytes " + maximumPlanBytes.toLocaleString() + ".",
+      "CPU analysis plan requires " + retainedNonSteeringBytes.toLocaleString() +
+      " non-steering bytes including " + calibrationBytes.toLocaleString() +
+      " calibration bytes, above maximumPlanBytes " +
+      maximumPlanBytes.toLocaleString() + ".",
    );
  }
  const distances = createDistanceMatrix(arrayDefinition, experiment.scan);
  const binMap = new Float32Array(pointCount);
  const steeringReal = new Float64Array(microphoneCount);
  const steeringImaginary = new Float64Array(microphoneCount);
+  const stftWorkspace = createStftWorkspace(fftSize);
  const steeringBytesPerBin =
    pointCount * microphoneCount * Float64Array.BYTES_PER_ELEMENT * 2;
  const requestedSteeringCacheBytes = steeringBytesPerBin * selectedBins.length;
-  const steeringCacheEnabled =
-    Number.isSafeInteger(requestedSteeringCacheBytes) &&
-    requestedSteeringCacheBytes > 0 &&
-    requestedSteeringCacheBytes <= maximumSteeringCacheBytes &&
-    retainedBaseBytes + requestedSteeringCacheBytes <= maximumPlanBytes;
+  const steeringCacheBinCapacity = Math.min(
+    selectedBins.length,
+    Math.floor(maximumSteeringCacheBytes / steeringBytesPerBin),
+    Math.floor((maximumPlanBytes - retainedNonSteeringBytes) / steeringBytesPerBin),
+  );
+  const steeringCacheBytes = steeringCacheBinCapacity * steeringBytesPerBin;
+  const steeringCacheEnabled = steeringCacheBinCapacity > 0;
+  const steeringCachedBins = new Set(selectedBins.slice(0, steeringCacheBinCapacity));

  return {
    kind: DAS_ANALYSIS_PLAN_KIND,
@@ -297,13 +328,20 @@ export function createDasAnalysisPlan(options) {
    calibrations: new Map(),
    steeringMatrices: new Map(),
    steeringCacheEnabled,
+    steeringCacheBinCapacity,
+    steeringCachedBins,
    steeringBytesPerBin,
+    steeringCacheBytes,
    requestedSteeringCacheBytes,
+    retainedBaseBytes,
+    calibrationBytes,
+    retainedNonSteeringBytes,
    maximumSteeringCacheBytes,
    maximumPlanBytes,
    binMap,
    steeringReal,
    steeringImaginary,
+    stftWorkspace,
    busy: false,
    statistics: {
      analysisAttempts: 0,
@@ -315,8 +353,20 @@ export function createDasAnalysisPlan(options) {
      calibrationCacheHits: 0,
      steeringMatrixBuilds: 0,
      steeringCacheHits: 0,
+      uncachedSteeringUses: 0,
      steeringPointMicrophonePairs: 0,
+      uncachedSteeringPointMicrophonePairs: 0,
      avoidedSteeringPointMicrophonePairs: 0,
+      stftWindowBuilds: 1,
+      stftWorkspaceAllocations: 1,
+      stftWorkspaceUses: 0,
+      avoidedStftWindowBuilds: 0,
+      avoidedStftWorkspaceAllocations: 0,
+      stftTransforms: 0,
+      fftButterflies: 0,
+      beamformingComplexMacs: 0,
+      binMapAllocations: 1,
+      steeringWorkspaceAllocations: 1,
      binMapUses: 0,
      avoidedBinMapAllocations: 0,
      retainedBytes: retainedBaseBytes,
@@ -350,6 +400,7 @@ function planCalibration(plan, bin) {
    plan.sampleRate,
    plan.experiment.stft.fftSize,
    plan.experiment.beamforming.compensateFrequencyResponse !== false,
+    plan.experiment.beamforming.maxCompensationDb,
  );
  plan.calibrations.set(bin, calibration);
  plan.statistics.calibrationBuilds += 1;
@@ -358,7 +409,12 @@ function planCalibration(plan, bin) {
 }

 async function planSteeringMatrix(plan, bin, calibration, signal, pointBatchSize) {
-  if (!plan.steeringCacheEnabled) return undefined;
+  if (!plan.steeringCachedBins.has(bin)) {
+    plan.statistics.uncachedSteeringUses += 1;
+    plan.statistics.uncachedSteeringPointMicrophonePairs +=
+      plan.pointCount * plan.microphoneCount;
+    return undefined;
+  }
  let matrix = plan.steeringMatrices.get(bin);
  if (matrix) {
    plan.statistics.steeringCacheHits += 1;
@@ -393,12 +449,31 @@ function analysisCacheSummary(plan) {
    avoidedDistanceEvaluations: plan.statistics.avoidedDistanceEvaluations,
    calibrationBuilds: plan.statistics.calibrationBuilds,
    calibrationCacheHits: plan.statistics.calibrationCacheHits,
+    calibrationBytes: plan.calibrationBytes,
    steeringCacheEnabled: plan.steeringCacheEnabled,
+    steeringCacheBinCapacity: plan.steeringCacheBinCapacity,
+    steeringCachedBinCount: plan.steeringMatrices.size,
+    steeringUncachedBinCount: plan.selectedBins.length - plan.steeringCacheBinCapacity,
+    steeringCacheBytes: plan.steeringCacheBytes,
+    requestedSteeringCacheBytes: plan.requestedSteeringCacheBytes,
    steeringMatrixBuilds: plan.statistics.steeringMatrixBuilds,
    steeringCacheHits: plan.statistics.steeringCacheHits,
+    uncachedSteeringUses: plan.statistics.uncachedSteeringUses,
    steeringPointMicrophonePairs: plan.statistics.steeringPointMicrophonePairs,
+    uncachedSteeringPointMicrophonePairs:
+      plan.statistics.uncachedSteeringPointMicrophonePairs,
    avoidedSteeringPointMicrophonePairs:
      plan.statistics.avoidedSteeringPointMicrophonePairs,
+    stftWindowBuilds: plan.statistics.stftWindowBuilds,
+    stftWorkspaceAllocations: plan.statistics.stftWorkspaceAllocations,
+    stftWorkspaceUses: plan.statistics.stftWorkspaceUses,
+    avoidedStftWindowBuilds: plan.statistics.avoidedStftWindowBuilds,
+    avoidedStftWorkspaceAllocations: plan.statistics.avoidedStftWorkspaceAllocations,
+    stftTransforms: plan.statistics.stftTransforms,
+    fftButterflies: plan.statistics.fftButterflies,
+    beamformingComplexMacs: plan.statistics.beamformingComplexMacs,
+    binMapAllocations: plan.statistics.binMapAllocations,
+    steeringWorkspaceAllocations: plan.statistics.steeringWorkspaceAllocations,
    avoidedBinMapAllocations: plan.statistics.avoidedBinMapAllocations,
    retainedBytes: plan.statistics.retainedBytes,
  };
@@ -419,12 +494,13 @@ export function estimateDasWorkload(options) {
  const outputBytes = bandCount * pointCount * Float32Array.BYTES_PER_ELEMENT * 2;
  const binMapBytes = pointCount * Float32Array.BYTES_PER_ELEMENT;
  const spectrumBytesPerBin = frameCount * microphoneCount * Float32Array.BYTES_PER_ELEMENT * 2;
-  const fftWorkspaceBytes = fftSize * Float64Array.BYTES_PER_ELEMENT * 2;
-  const steeringCacheBytes = analysisPlan?.steeringCacheEnabled
-    ? analysisPlan.requestedSteeringCacheBytes
-    : 0;
+  const fftWorkspaceBytes = fftSize * Float64Array.BYTES_PER_ELEMENT * 3;
+  const steeringCacheBytes = analysisPlan?.steeringCacheBytes ?? 0;
+  const calibrationBytes = analysisPlan?.calibrationBytes ??
+    microphoneCount * Float64Array.BYTES_PER_ELEMENT * 2;
  const baseWorkingBytes =
-    distanceBytes + outputBytes + binMapBytes + fftWorkspaceBytes + steeringCacheBytes;
+    distanceBytes + outputBytes + binMapBytes + fftWorkspaceBytes +
+    steeringCacheBytes + calibrationBytes;
  const complexMacCount = pointCount * microphoneCount * frameCount * selectedBins.length;
  return {
    bandBins,
@@ -434,7 +510,10 @@ export function estimateDasWorkload(options) {
    pointCount,
    distanceBytes,
    outputBytes,
+    fftWorkspaceBytes,
    steeringCacheBytes,
+    calibrationBytes,
+    steeringCacheBinCapacity: analysisPlan?.steeringCacheBinCapacity ?? 0,
    spectrumBytesPerBin,
    baseWorkingBytes,
    complexMacCount,
@@ -537,16 +616,21 @@ async function computeDasMapInternal(options, analysisPlan) {

  let completedBins = 0;
  for (let batchStart = 0; batchStart < workload.selectedBins.length; batchStart += batchSize) {
-    throwIfAborted(signal);
-    await yieldToEventLoop();
    throwIfAborted(signal);
    const batchBins = workload.selectedBins.slice(batchStart, batchStart + batchSize);
+    if (analysisPlan) {
+      analysisPlan.statistics.stftWorkspaceUses += 1;
+      const repeatedUses = Math.max(0, analysisPlan.statistics.stftWorkspaceUses - 1);
+      analysisPlan.statistics.avoidedStftWindowBuilds = repeatedUses;
+      analysisPlan.statistics.avoidedStftWorkspaceAllocations = repeatedUses;
+    }
    const stft = await extractStftBins(channels, {
      fftSize: experiment.stft.fftSize,
      hopSize: experiment.stft.hopSize,
      bins: batchBins,
      signal,
      transformBatchSize: stftTransformBatchSize,
+      workspace: analysisPlan?.stftWorkspace,
      onProgress: ({ completedTransforms, totalTransforms }) =>
        onProgress?.({
          stage: "stft",
@@ -558,7 +642,15 @@ async function computeDasMapInternal(options, analysisPlan) {
          spectrumBatchCount,
        }),
    });
-    await yieldToEventLoop();
+    if (analysisPlan) {
+      const completedTransformCount = stft.frameCount * stft.microphoneCount;
+      analysisPlan.statistics.stftTransforms += completedTransformCount;
+      analysisPlan.statistics.fftButterflies +=
+        completedTransformCount *
+        (experiment.stft.fftSize / 2) *
+        Math.log2(experiment.stft.fftSize);
+    }
+    await cooperativeYield();
    throwIfAborted(signal);

    for (const bin of batchBins) {
@@ -585,6 +677,7 @@ async function computeDasMapInternal(options, analysisPlan) {
        speedOfSound: scene.speedOfSound,
        compensateFrequencyResponse:
          experiment.beamforming.compensateFrequencyResponse !== false,
+        maxCompensationDb: experiment.beamforming.maxCompensationDb,
        calibration,
        steeringMatrix,
        output: analysisPlan?.binMap,
@@ -593,6 +686,10 @@ async function computeDasMapInternal(options, analysisPlan) {
        signal,
        pointBatchSize,
      });
+      if (analysisPlan) {
+        analysisPlan.statistics.beamformingComplexMacs +=
+          workload.pointCount * workload.microphoneCount * stft.frameCount;
+      }
      for (const bandIndex of bandIndexesByBin.get(bin)) {
        const rawMap = rawMaps[bandIndex];
        for (let pointIndex = 0; pointIndex < workload.pointCount; pointIndex += 1) {
@@ -2,6 +2,33 @@ import { dbToLinear } from "./math.js";

 export const MAX_FREQUENCY_RESPONSE_POINTS = 64;
 export const MAX_FREQUENCY_RESPONSE_HZ = 1_000_000;
+export const DEFAULT_MAX_COMPENSATION_DB = 24;
+export const MAX_COMPENSATION_DB = 120;
+
+export function requireMaxCompensationDb(
+  value = DEFAULT_MAX_COMPENSATION_DB,
+  label = "maxCompensationDb",
+) {
+  if (!Number.isFinite(value) || value < 0 || value > MAX_COMPENSATION_DB) {
+    throw new RangeError(`${label} must lie inside 0..${MAX_COMPENSATION_DB} dB.`);
+  }
+  return value;
+}
+
+/**
+ * Return the inverse response gain in dB while limiting only positive boost.
+ * Attenuation remains exact because it cannot amplify a response notch or the
+ * microphone self-noise that occupies it.
+ */
+export function inverseResponseGainDb(responseGainDb, maxCompensationDb) {
+  if (!Number.isFinite(responseGainDb)) {
+    throw new RangeError("responseGainDb must be finite.");
+  }
+  return Math.min(
+    -responseGainDb,
+    requireMaxCompensationDb(maxCompensationDb),
+  );
+}

 function phaseAt(point) {
  return point.phaseDeg ?? 0;
@@ -1,13 +1,18 @@
 import { computeDasMap, createDasAnalysisPlan } from "./das.js";
+import { findLocalPeaks } from "./evaluation.js";
 import {
+  DEFAULT_MAX_COMPENSATION_DB,
  MAX_FREQUENCY_RESPONSE_HZ,
  MAX_FREQUENCY_RESPONSE_POINTS,
+  requireMaxCompensationDb,
 } from "./frequency-response.js";

 export const REALTIME_LOCALIZATION_PLAN_KIND = "micarrsim.realtime-localization-plan/1";
 export const DEFAULT_STREAM_RESULTS_PER_CALL = 1;
 export const MAX_STREAM_RESULTS_PER_CALL = 32;
 export const DEFAULT_MAXIMUM_STREAM_INPUT_BYTES = 256 * 1024 * 1024;
+export const DEFAULT_REALTIME_POINT_BATCH_SIZE = 256;
+export const DEFAULT_REALTIME_STFT_TRANSFORM_BATCH_SIZE = 128;

 function requirePositiveFinite(value, label) {
  if (!Number.isFinite(value) || value <= 0) {
@@ -31,6 +36,10 @@ function requirePowerOfTwo(value, label) {
  return value;
 }

+function monotonicNow() {
+  return globalThis.performance?.now?.() ?? Date.now();
+}
+
 function freezePosition(position, label) {
  if (!Array.isArray(position) || position.length !== 3 || !position.every(Number.isFinite)) {
    throw new TypeError(`${label} must be a finite [x, y, z] array.`);
@@ -155,6 +164,28 @@ function copyBands(bands, sampleRate) {
  }));
 }

+function copyPeakDetection(value, scan) {
+  const options = value ?? {};
+  if (!options || typeof options !== "object" || Array.isArray(options)) {
+    throw new TypeError("peakDetection must be an object.");
+  }
+  const thresholdRatio = options.thresholdRatio ?? 0.5;
+  if (!Number.isFinite(thresholdRatio) || thresholdRatio < 0 || thresholdRatio > 1) {
+    throw new RangeError("peakDetection.thresholdRatio must lie inside 0..1.");
+  }
+  const xStep = scan.columns > 1 ? (scan.xMax - scan.xMin) / (scan.columns - 1) : 0;
+  const yStep = scan.rows > 1 ? (scan.yMax - scan.yMin) / (scan.rows - 1) : 0;
+  const minimumPeakDistanceM = options.minimumPeakDistanceM ?? Math.hypot(xStep, yStep);
+  if (!Number.isFinite(minimumPeakDistanceM) || minimumPeakDistanceM < 0) {
+    throw new RangeError("peakDetection.minimumPeakDistanceM must be a non-negative finite number.");
+  }
+  const maximumDetections = options.maximumDetections ?? 8;
+  if (!Number.isInteger(maximumDetections) || maximumDetections < 1 || maximumDetections > 128) {
+    throw new RangeError("peakDetection.maximumDetections must be an integer from 1 through 128.");
+  }
+  return Object.freeze({ thresholdRatio, minimumPeakDistanceM, maximumDetections });
+}
+
 function validateChannels(channels, expectedChannelCount) {
  if (!Array.isArray(channels) || channels.length !== expectedChannelCount) {
    throw new Error(
@@ -182,13 +213,42 @@ function throwIfAborted(signal) {
  if (signal?.aborted) throw createAbortError(signal);
 }

-function locationsFromBands(bands) {
-  return bands.map((band) => ({
-    bandId: band.id,
-    index: band.peak.index,
-    position: [...band.peak.position],
-    rawPower: band.peak.rawPower,
-  }));
+export function detectBandPeaks(rawMap, scan, options = {}) {
+  const peakDetection = copyPeakDetection(options, scan);
+  const localPeaks = findLocalPeaks(rawMap, scan);
+  const threshold = (localPeaks[0]?.value ?? 0) * peakDetection.thresholdRatio;
+  const detections = [];
+  for (const peak of localPeaks) {
+    if (peak.value < threshold) continue;
+    if (
+      peakDetection.minimumPeakDistanceM > 0 &&
+      detections.some((selected) =>
+        Math.hypot(
+          selected.position[0] - peak.position[0],
+          selected.position[1] - peak.position[1],
+        ) <= peakDetection.minimumPeakDistanceM
+      )
+    ) {
+      continue;
+    }
+    detections.push({
+      rank: detections.length + 1,
+      index: peak.index,
+      position: [...peak.position],
+      rawPower: peak.value,
+    });
+    if (detections.length === peakDetection.maximumDetections) break;
+  }
+  return detections;
+}
+
+function locationsFromBands(bands, scan, peakDetection) {
+  return bands.flatMap((band) =>
+    detectBandPeaks(band.rawMap, scan, peakDetection).map((peak) => ({
+      bandId: band.id,
+      ...peak,
+    })),
+  );
 }

 /**
@@ -202,6 +262,7 @@ export function createLocalizationPlan(config) {
  const sampleRate = requirePositiveFinite(config.sampleRate, "sampleRate");
  const arrayDefinition = copyArrayDefinition(config.array ?? config.arrayDefinition);
  const scan = copyScan(config.scan);
+  const peakDetection = copyPeakDetection(config.peakDetection, scan);
  const frequencyBands = copyBands(config.bands ?? config.frequencyBands, sampleRate);
  const fftSize = requirePowerOfTwo(config.stft?.fftSize ?? config.fftSize ?? 512, "fftSize");
  const hopSize = requirePositiveInteger(
@@ -221,13 +282,22 @@ export function createLocalizationPlan(config) {
  if (typeof compensateFrequencyResponse !== "boolean") {
    throw new TypeError("compensateFrequencyResponse must be boolean.");
  }
+  const maxCompensationDb = requireMaxCompensationDb(
+    config.maxCompensationDb ??
+      config.beamforming?.maxCompensationDb ??
+      DEFAULT_MAX_COMPENSATION_DB,
+  );

  const scene = Object.freeze({ speedOfSound });
  const experiment = Object.freeze({
    stft: Object.freeze({ fftSize, hopSize, window: "hann" }),
    scan,
    frequencyBands,
-    beamforming: Object.freeze({ normalization, compensateFrequencyResponse }),
+    beamforming: Object.freeze({
+      normalization,
+      compensateFrequencyResponse,
+      maxCompensationDb,
+    }),
  });
  const dasPlan = createDasAnalysisPlan({
    sampleRate,
@@ -245,13 +315,19 @@ export function createLocalizationPlan(config) {
    arrayDefinition,
    scan,
    bands: frequencyBands,
+    peakDetection,
    stft: experiment.stft,
    speedOfSound,
    normalization,
+    maxCompensationDb,
    scene,
    experiment,
    dasPlan,
-    computeOptions: Object.freeze({ ...(config.computeOptions ?? {}) }),
+    computeOptions: Object.freeze({
+      pointBatchSize: DEFAULT_REALTIME_POINT_BATCH_SIZE,
+      stftTransformBatchSize: DEFAULT_REALTIME_STFT_TRANSFORM_BATCH_SIZE,
+      ...(config.computeOptions ?? {}),
+    }),
  });
 }

@@ -273,6 +349,7 @@ export async function localizeMultichannel(options) {
  const sampleCount = validateChannels(options.channels, plan.channelCount);
  if (sampleCount === 0) throw new RangeError("channels must contain at least one sample.");
  throwIfAborted(options.signal);
+  const processingStartedAt = monotonicNow();
  const das = await computeDasMap({
    ...plan.computeOptions,
    ...(options.computeOptions ?? {}),
@@ -285,12 +362,15 @@ export async function localizeMultichannel(options) {
    signal: options.signal,
    onProgress: options.onProgress,
  });
+  const locations = locationsFromBands(das.bands, plan.scan, plan.peakDetection);
+  const processingTimeMs = Math.max(0, monotonicNow() - processingStartedAt);
  return {
    ...das,
    sampleRate: plan.sampleRate,
    sampleCount,
    duration: sampleCount / plan.sampleRate,
-    locations: locationsFromBands(das.bands),
+    processingTimeMs,
+    locations,
  };
 }

@@ -298,8 +378,9 @@ export async function localizeMultichannel(options) {
 * A fixed-capacity chunk buffer for real microphone adapters.
 *
 * append() is synchronous and remains safe while an analysis is awaiting CPU
- * batches. analyzeAvailable() copies one window before yielding, consumes its
- * hop, and returns no more than maximumResultsPerCall maps.
+ * batches. analyzeAvailable() copies one window before yielding, commits its
+ * hop only after successful localization, and returns no more than
+ * maximumResultsPerCall maps.
 */
 export class StreamingDasLocalizer {
  constructor(config) {
@@ -319,6 +400,7 @@ export class StreamingDasLocalizer {
    if (this.analysisHopSamples > this.analysisWindowSamples) {
      throw new RangeError("analysisHopSamples must not exceed analysisWindowSamples.");
    }
+    this.analysisDeadlineMs = this.analysisHopSamples / this.plan.sampleRate * 1000;
    this.maximumBufferedSamples = requirePositiveInteger(
      config.maximumBufferedSamples ?? this.analysisWindowSamples * 4,
      "maximumBufferedSamples",
@@ -370,15 +452,37 @@ export class StreamingDasLocalizer {
    this.analysisAttempts = 0;
    this.completedAnalyses = 0;
    this.failedAnalyses = 0;
+    this.lastProcessingTimeMs = null;
+    this.maximumProcessingTimeMs = 0;
+    this.deadlineMisses = 0;
    this.nextSequence = 0;
+    this.pendingAnalysis = null;
    this.busy = false;
  }

  availableAnalysisCount() {
-    if (this.bufferedSamples < this.analysisWindowSamples) return 0;
-    return 1 + Math.floor(
-      (this.bufferedSamples - this.analysisWindowSamples) / this.analysisHopSamples,
+    const bufferEndSample = this.bufferStartSample + this.bufferedSamples;
+    if (!this.pendingAnalysis) {
+      if (this.bufferedSamples < this.analysisWindowSamples) return 0;
+      return 1 + Math.floor(
+        (this.bufferedSamples - this.analysisWindowSamples) / this.analysisHopSamples,
+      );
+    }
+
+    // The failed window remains in analysisChannels even if capture overflow
+    // advances the ring while localization is running. Count it once, then
+    // count windows that will remain after its successful hop commit.
+    const nextRingStart = Math.max(
+      this.bufferStartSample,
+      this.pendingAnalysis.startSample + this.analysisHopSamples,
    );
+    const futureSamples = bufferEndSample - nextRingStart;
+    const futureCount = futureSamples < this.analysisWindowSamples
+      ? 0
+      : 1 + Math.floor(
+        (futureSamples - this.analysisWindowSamples) / this.analysisHopSamples,
+      );
+    return 1 + futureCount;
  }

  getState() {
@@ -387,6 +491,7 @@ export class StreamingDasLocalizer {
      channelCount: this.plan.channelCount,
      analysisWindowSamples: this.analysisWindowSamples,
      analysisHopSamples: this.analysisHopSamples,
+      analysisDeadlineMs: this.analysisDeadlineMs,
      maximumBufferedSamples: this.maximumBufferedSamples,
      maximumInputBufferBytes: this.maximumInputBufferBytes,
      maximumResultsPerCall: this.maximumResultsPerCall,
@@ -398,6 +503,9 @@ export class StreamingDasLocalizer {
      analysisAttempts: this.analysisAttempts,
      completedAnalyses: this.completedAnalyses,
      failedAnalyses: this.failedAnalyses,
+      lastProcessingTimeMs: this.lastProcessingTimeMs,
+      maximumProcessingTimeMs: this.maximumProcessingTimeMs,
+      deadlineMisses: this.deadlineMisses,
      nextSequence: this.nextSequence,
      busy: this.busy,
      retainedInputBytes: this.retainedInputBytes,
@@ -481,39 +589,62 @@ export class StreamingDasLocalizer {
    try {
      while (
        results.length < maximumResults &&
-        this.bufferedSamples >= this.analysisWindowSamples
+        (this.pendingAnalysis || this.bufferedSamples >= this.analysisWindowSamples)
      ) {
        throwIfAborted(options.signal);
-        const startSample = this.bufferStartSample;
+        const retry = this.pendingAnalysis;
+        const startSample = retry?.startSample ?? this.bufferStartSample;
        const endSample = startSample + this.analysisWindowSamples;
-        this.copyAnalysisWindow();
-        this.consume(this.analysisHopSamples, false);
+        if (!retry) this.copyAnalysisWindow();
        this.analysisAttempts += 1;
+        let localized;
        try {
-          const localized = await localizeMultichannel({
+          localized = await localizeMultichannel({
            plan: this.plan,
            channels: this.analysisChannels,
            signal: options.signal,
            onProgress: options.onProgress,
            computeOptions: options.computeOptions,
          });
-          const result = {
-            sequence: this.nextSequence,
-            startSample,
-            endSample,
-            startTime: startSample / this.plan.sampleRate,
-            endTime: endSample / this.plan.sampleRate,
-            centerTime: (startSample + endSample) / (2 * this.plan.sampleRate),
-            ...localized,
-          };
-          this.nextSequence += 1;
-          this.completedAnalyses += 1;
-          results.push(result);
-          options.onResult?.(result);
        } catch (error) {
+          this.pendingAnalysis = { startSample, endSample };
          this.failedAnalyses += 1;
          throw error;
        }
+
+        const result = {
+          sequence: this.nextSequence,
+          startSample,
+          endSample,
+          startTime: startSample / this.plan.sampleRate,
+          endTime: endSample / this.plan.sampleRate,
+          centerTime: (startSample + endSample) / (2 * this.plan.sampleRate),
+          ...localized,
+          deadlineMs: this.analysisDeadlineMs,
+          deadlineMet: localized.processingTimeMs <= this.analysisDeadlineMs,
+        };
+        this.pendingAnalysis = null;
+        const commitThroughSample = startSample + this.analysisHopSamples;
+        this.consume(
+          Math.min(
+            this.bufferedSamples,
+            Math.max(0, commitThroughSample - this.bufferStartSample),
+          ),
+          false,
+        );
+        this.nextSequence += 1;
+        this.completedAnalyses += 1;
+        this.lastProcessingTimeMs = localized.processingTimeMs;
+        this.maximumProcessingTimeMs = Math.max(
+          this.maximumProcessingTimeMs,
+          localized.processingTimeMs,
+        );
+        if (!result.deadlineMet) this.deadlineMisses += 1;
+        results.push(result);
+
+        // Consumer failures must not turn a successfully committed analysis
+        // into a localization failure or make its window eligible for retry.
+        options.onResult?.(result);
      }
      return results;
    } finally {
@@ -536,7 +667,11 @@ export class StreamingDasLocalizer {
    this.analysisAttempts = 0;
    this.completedAnalyses = 0;
    this.failedAnalyses = 0;
+    this.lastProcessingTimeMs = null;
+    this.maximumProcessingTimeMs = 0;
+    this.deadlineMisses = 0;
    this.nextSequence = 0;
+    this.pendingAnalysis = null;
  }
 }

@@ -1,4 +1,5 @@
 import { fftInPlace } from "./fft.js";
+import { cooperativeYield } from "./cooperative-yield.js";

 function throwIfAborted(signal) {
  if (signal?.aborted) {
@@ -8,10 +9,6 @@ function throwIfAborted(signal) {
  }
 }

-function yieldToEventLoop() {
-  return new Promise((resolve) => setTimeout(resolve, 0));
-}
-
 export function hannWindow(size) {
  if (!Number.isInteger(size) || size < 2) throw new Error("Hann window 크기는 2 이상이어야 합니다.");
  const window = new Float64Array(size);
@@ -21,6 +18,32 @@ export function hannWindow(size) {
  return window;
 }

+const STFT_WORKSPACE_KIND = "micarrsim.stft-workspace/1";
+
+export function createStftWorkspace(fftSize) {
+  return {
+    kind: STFT_WORKSPACE_KIND,
+    fftSize,
+    window: hannWindow(fftSize),
+    real: new Float64Array(fftSize),
+    imaginary: new Float64Array(fftSize),
+  };
+}
+
+function resolveStftWorkspace(fftSize, workspace) {
+  if (workspace === undefined) return createStftWorkspace(fftSize);
+  if (
+    workspace?.kind !== STFT_WORKSPACE_KIND ||
+    workspace.fftSize !== fftSize ||
+    workspace.window?.length !== fftSize ||
+    workspace.real?.length !== fftSize ||
+    workspace.imaginary?.length !== fftSize
+  ) {
+    throw new RangeError("STFT workspace must match fftSize.");
+  }
+  return workspace;
+}
+
 export function getFrameCount(sampleCount, fftSize, hopSize) {
  if (!Number.isInteger(hopSize) || hopSize < 1 || hopSize > fftSize) {
    throw new Error("hopSize는 1 이상 fftSize 이하여야 합니다.");
@@ -29,7 +52,15 @@ export function getFrameCount(sampleCount, fftSize, hopSize) {
 }

 export async function extractStftBins(channels, options) {
-  const { fftSize, hopSize, bins, signal, onProgress, transformBatchSize = 32 } = options;
+  const {
+    fftSize,
+    hopSize,
+    bins,
+    signal,
+    onProgress,
+    transformBatchSize = 32,
+    workspace: providedWorkspace,
+  } = options;
  if (!Array.isArray(channels) || channels.length === 0) throw new Error("마이크 채널이 필요합니다.");
  const sampleCount = channels[0].length;
  if (channels.some((channel) => channel.length !== sampleCount)) throw new Error("마이크 채널 길이가 다릅니다.");
@@ -51,9 +82,8 @@ export async function extractStftBins(channels, options) {
    });
  }

-  const window = hannWindow(fftSize);
-  const real = new Float64Array(fftSize);
-  const imaginary = new Float64Array(fftSize);
+  const workspace = resolveStftWorkspace(fftSize, providedWorkspace);
+  const { window, real, imaginary } = workspace;
  const totalTransforms = frameCount * microphoneCount;
  let completedTransforms = 0;
  for (let frameIndex = 0; frameIndex < frameCount; frameIndex += 1) {
@@ -73,9 +103,12 @@ export async function extractStftBins(channels, options) {
        spectrum.imaginary[spectrumIndex] = imaginary[bin];
      }
      completedTransforms += 1;
-      if (completedTransforms % transformBatchSize === 0) {
+      if (
+        completedTransforms % transformBatchSize === 0 &&
+        completedTransforms < totalTransforms
+      ) {
        onProgress?.({ completedTransforms, totalTransforms });
-        await yieldToEventLoop();
+        await cooperativeYield();
        throwIfAborted(signal);
      }
    }
@@ -455,6 +455,14 @@ function validateExperimentRelationships(value, context = {}) {
        "주파수 응답 보정 사용 여부는 true 또는 false여야 합니다.",
      );
    }
+    if (value.beamforming.maxCompensationDb !== undefined) {
+      requireFinite(
+        value.beamforming.maxCompensationDb,
+        "$.beamforming.maxCompensationDb",
+        issues,
+        { min: 0, max: 120 },
+      );
+    }
  }
  if (value.evaluation !== undefined && requireRecord(value.evaluation, "$.evaluation", issues)) {
    const threshold = value.evaluation.detectionThresholdRatio;
@@ -3,7 +3,11 @@ import {
  estimateDasWorkload,
  finalizeDasBands,
 } from "../acoustics-core/das.js";
-import { interpolateFrequencyResponse } from "../acoustics-core/frequency-response.js";
+import {
+  interpolateFrequencyResponse,
+  inverseResponseGainDb,
+  requireMaxCompensationDb,
+} from "../acoustics-core/frequency-response.js";
 import { dbToLinear } from "../acoustics-core/math.js";
 import { extractStftBins } from "../acoustics-core/stft.js";
 import {
@@ -77,6 +81,7 @@ export function packStftBatch(stft, bins, options = {}) {
  const packed = new Float32Array(bins.length * valuesPerBin * 2);
  const microphones = options.microphones;
  const compensateFrequencyResponse = options.compensateFrequencyResponse !== false;
+  const maxCompensationDb = requireMaxCompensationDb(options.maxCompensationDb);
  const hasFrequencyResponse =
    compensateFrequencyResponse &&
    Array.isArray(microphones) &&
@@ -100,7 +105,9 @@ export function packStftBatch(stft, bins, options = {}) {
      microphones.forEach((microphone, microphoneIndex) => {
        if (!Array.isArray(microphone.frequencyResponse)) return;
        interpolateFrequencyResponse(microphone.frequencyResponse, frequencyHz, response);
-        const inverseGain = dbToLinear(-response.gainDb);
+        const inverseGain = dbToLinear(
+          inverseResponseGainDb(response.gainDb, maxCompensationDb),
+        );
        const responseOffset = (binIndex * stft.microphoneCount + microphoneIndex) * 2;
        inverseResponses[responseOffset] = inverseGain * Math.cos(response.phaseRadians);
        inverseResponses[responseOffset + 1] = -inverseGain * Math.sin(response.phaseRadians);
@@ -541,6 +548,7 @@ export async function computeDasWebGpu(options) {
        fftSize: experiment.stft.fftSize,
        compensateFrequencyResponse:
          experiment.beamforming.compensateFrequencyResponse !== false,
+        maxCompensationDb: experiment.beamforming.maxCompensationDb,
      });
      stftMs += now() - stftStartedAt;
      const binIndices = Uint32Array.from(batchBins);