Typescriptify auto optimizer in libSquoosh

libsquoosh/src/auto-optimizer.ts

@@ -0,0 +1,115 @@
+import { instantiateEmscriptenWasm } from './emscripten-utils.js';
+
+import visdif from '../../codecs/visdif/visdif.js';
+import visdifWasm from 'asset-url:../../codecs/visdif/visdif.wasm';
+import type ImageData from './image_data';
+
+interface VisDiff {
+  distance: (data: Uint8ClampedArray) => number;
+  delete: () => void;
+}
+
+interface VisdiffConstructor {
+  new (data: Uint8ClampedArray, width: number, height: number): VisDiff;
+}
+
+interface VisDiffModule extends EmscriptenWasm.Module {
+  VisDiff: VisdiffConstructor;
+}
+
+type VisDiffModuleFactory = EmscriptenWasm.ModuleFactory<VisDiffModule>;
+
+interface BinarySearchParams {
+  min?: number;
+  max?: number;
+  epsilon?: number;
+  maxRounds?: number;
+}
+
+interface AutoOptimizeParams extends BinarySearchParams {
+  butteraugliDistanceGoal?: number;
+}
+
+interface AutoOptimizeResult {
+  bitmap: ImageData;
+  binary: Uint8Array;
+  quality: number;
+}
+
+// `measure` is a (async) function that takes exactly one numeric parameter and
+// returns a value. The function is assumed to be monotonic (an increase in `parameter`
+// will result in an increase in the return value. The function uses binary search
+// to find `parameter` such that `measure` returns `measureGoal`, within an error
+// of `epsilon`. It will use at most `maxRounds` attempts.
+export async function binarySearch(
+  measureGoal: number,
+  measure: (val: number) => Promise<number>,
+  { min = 0, max = 100, epsilon = 0.1, maxRounds = 6 }: BinarySearchParams = {},
+) {
+  let parameter = (max - min) / 2 + min;
+  let delta = (max - min) / 4;
+  let value;
+  let round = 1;
+  while (true) {
+    value = await measure(parameter);
+    if (Math.abs(value - measureGoal) < epsilon || round >= maxRounds) {
+      return { parameter, round, value };
+    }
+    if (value > measureGoal) {
+      parameter -= delta;
+    } else if (value < measureGoal) {
+      parameter += delta;
+    }
+    delta /= 2;
+    round++;
+  }
+}
+
+export async function autoOptimize(
+  bitmapIn: ImageData,
+  encode: (
+    bitmap: ImageData,
+    quality: number,
+  ) => Promise<Uint8Array> | Uint8Array,
+  decode: (binary: Uint8Array) => Promise<ImageData> | ImageData,
+  {
+    butteraugliDistanceGoal = 1.4,
+    ...binarySearchParams
+  }: AutoOptimizeParams = {},
+): Promise<AutoOptimizeResult> {
+  const { VisDiff } = await instantiateEmscriptenWasm(
+    visdif as VisDiffModuleFactory,
+    visdifWasm,
+  );
+
+  const comparator = new VisDiff(
+    bitmapIn.data,
+    bitmapIn.width,
+    bitmapIn.height,
+  );
+
+  // We're able to do non null assertion because
+  // we know that binarySearch will set these values
+  let bitmapOut!: ImageData;
+  let binaryOut!: Uint8Array;
+
+  // Increasing quality means _decrease_ in Butteraugli distance.
+  // `binarySearch` assumes that increasing `parameter` will
+  // increase the metric value. So multipliy Butteraugli values by -1.
+  const { parameter } = await binarySearch(
+    -1 * butteraugliDistanceGoal,
+    async (quality) => {
+      binaryOut = await encode(bitmapIn, quality);
+      bitmapOut = await decode(binaryOut);
+      return -1 * comparator.distance(bitmapOut.data);
+    },
+    binarySearchParams,
+  );
+  comparator.delete();
+
+  return {
+    bitmap: bitmapOut,
+    binary: binaryOut,
+    quality: parameter,
+  };
+}