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

import { kValue } from '../../../../../util/constants.js';
import { FP } from '../../../../../util/floating_point.js';
import { biasedRange, linearRange } from '../../../../../util/math.js';
import { makeCaseCache } from '../../case_cache.js';

// floor(ln(max f32 value)) = 88, so exp(88) will be within range of a f32, but exp(89) will not
// floor(ln(max f64 value)) = 709, so exp(709) can be handled by the testing framework, but exp(710) will misbehave
const f32_inputs = [
  0, // Returns 1 by definition
  -89, // Returns subnormal value
  kValue.f32.negative.min, // Closest to returning 0 as possible
  ...biasedRange(kValue.f32.negative.max, -88, 100),
  ...biasedRange(kValue.f32.positive.min, 88, 100),
  ...linearRange(89, 709, 10), // Overflows f32, but not f64
];

// floor(ln(max f16 value)) = 11, so exp(11) will be within range of a f16, but exp(12) will not
const f16_inputs = [
  0, // Returns 1 by definition
  -12, // Returns subnormal value
  kValue.f16.negative.min, // Closest to returning 0 as possible
  ...biasedRange(kValue.f16.negative.max, -11, 100),
  ...biasedRange(kValue.f16.positive.min, 11, 100),
  ...linearRange(12, 709, 10), // Overflows f16, but not f64
];

export const d = makeCaseCache('exp', {
  f32_const: () => {
    return FP.f32.generateScalarToIntervalCases(f32_inputs, 'finite', FP.f32.expInterval);
  },
  f32_non_const: () => {
    return FP.f32.generateScalarToIntervalCases(f32_inputs, 'unfiltered', FP.f32.expInterval);
  },
  f16_const: () => {
    return FP.f16.generateScalarToIntervalCases(f16_inputs, 'finite', FP.f16.expInterval);
  },
  f16_non_const: () => {
    return FP.f16.generateScalarToIntervalCases(f16_inputs, 'unfiltered', FP.f16.expInterval);
  },
  abstract: () => {
    // exp has an ulp accuracy, so is only expected to be as accurate as f32
    return FP.abstract.generateScalarToIntervalCases(f32_inputs, 'finite', FP.f32.expInterval);
  },
});