🏡 index : cold-air-networks/mirror-webgpu-cts.git

import { anyOf } from '../../../../../util/compare.js';
import { FP } from '../../../../../util/floating_point.js';
import { Case } from '../../case.js';
import { makeCaseCache } from '../../case_cache.js';

// stepInterval's return value can't always be interpreted as a single acceptance
// interval, valid result may be 0.0 or 1.0 or both of them, but will never be a
// value in interval (0.0, 1.0).
// See the comment block on stepInterval for more details
const makeCase = (trait: 'f32' | 'f16' | 'abstract', edge: number, x: number): Case => {
  const FPTrait = FP[trait];
  edge = FPTrait.quantize(edge);
  x = FPTrait.quantize(x);
  const expected = FPTrait.stepInterval(edge, x);

  // [0, 0], [1, 1], or [-∞, +∞] cases
  if (expected.isPoint() || !expected.isFinite()) {
    return { input: [FPTrait.scalarBuilder(edge), FPTrait.scalarBuilder(x)], expected };
  }

  // [0, 1] case, valid result is either 0.0 or 1.0.
  const zeroInterval = FPTrait.toInterval(0);
  const oneInterval = FPTrait.toInterval(1);
  return {
    input: [FPTrait.scalarBuilder(edge), FPTrait.scalarBuilder(x)],
    expected: anyOf(zeroInterval, oneInterval),
  };
};

// Cases: [f32|f16|abstract]
const cases = (['f32', 'f16', 'abstract'] as const)
  .map(trait => ({
    [`${trait}`]: () => {
      return FP[trait]
        .sparseScalarRange()
        .flatMap(edge => FP[trait].sparseScalarRange().map(x => makeCase(trait, edge, x)));
    },
  }))
  .reduce((a, b) => ({ ...a, ...b }), {});

export const d = makeCaseCache('step', cases);