// Licensed to the Apache Software Foundation (ASF) under one // or more contributor license agreements. See the NOTICE file // distributed with this work for additional information // regarding copyright ownership. The ASF licenses this file // to you under the Apache License, Version 2.0 (the // "License"); you may not use this file except in compliance // with the License. You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, // software distributed under the License is distributed on an // "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY // KIND, either express or implied. See the License for the // specific language governing permissions and limitations // under the License. const f64 = new Float64Array(1); const u32 = new Uint32Array(f64.buffer); /** * Convert uint16 (logically a float16) to a JS float64. Inspired by numpy's `npy_half_to_double`: * https://github.com/numpy/numpy/blob/5a5987291dc95376bb098be8d8e5391e89e77a2c/numpy/core/src/npymath/halffloat.c#L29 * @param h {number} the uint16 to convert * @private * @ignore */ export function uint16ToFloat64(h: number) { const expo = (h & 0x7C00) >> 10; const sigf = (h & 0x03FF) / 1024; const sign = (-1) ** ((h & 0x8000) >> 15); switch (expo) { case 0x1F: return sign * (sigf ? Number.NaN : 1 / 0); case 0x00: return sign * (sigf ? 6.103515625e-5 * sigf : 0); } return sign * (2 ** (expo - 15)) * (1 + sigf); } /** * Convert a float64 to uint16 (assuming the float64 is logically a float16). Inspired by numpy's `npy_double_to_half`: * https://github.com/numpy/numpy/blob/5a5987291dc95376bb098be8d8e5391e89e77a2c/numpy/core/src/npymath/halffloat.c#L43 * @param d {number} The float64 to convert * @private * @ignore */ export function float64ToUint16(d: number) { if (d !== d) { return 0x7E00; } // NaN f64[0] = d; // Magic numbers: // 0x80000000 = 10000000 00000000 00000000 00000000 -- masks the 32nd bit // 0x7ff00000 = 01111111 11110000 00000000 00000000 -- masks the 21st-31st bits // 0x000fffff = 00000000 00001111 11111111 11111111 -- masks the 1st-20th bit const sign = (u32[1] & 0x80000000) >> 16 & 0xFFFF; let expo = (u32[1] & 0x7FF00000), sigf = 0x0000; if (expo >= 0x40F00000) { // // If exponent overflowed, the float16 is either NaN or Infinity. // Rules to propagate the sign bit: mantissa > 0 ? NaN : +/-Infinity // // Magic numbers: // 0x40F00000 = 01000000 11110000 00000000 00000000 -- 6-bit exponent overflow // 0x7C000000 = 01111100 00000000 00000000 00000000 -- masks the 27th-31st bits // // returns: // qNaN, aka 32256 decimal, 0x7E00 hex, or 01111110 00000000 binary // sNaN, aka 32000 decimal, 0x7D00 hex, or 01111101 00000000 binary // +inf, aka 31744 decimal, 0x7C00 hex, or 01111100 00000000 binary // -inf, aka 64512 decimal, 0xFC00 hex, or 11111100 00000000 binary // // If mantissa is greater than 23 bits, set to +Infinity like numpy if (u32[0] > 0) { expo = 0x7C00; } else { expo = (expo & 0x7C000000) >> 16; sigf = (u32[1] & 0x000FFFFF) >> 10; } } else if (expo <= 0x3F000000) { // // If exponent underflowed, the float is either signed zero or subnormal. // // Magic numbers: // 0x3F000000 = 00111111 00000000 00000000 00000000 -- 6-bit exponent underflow // sigf = 0x100000 + (u32[1] & 0x000FFFFF); sigf = 0x100000 + (sigf << ((expo >> 20) - 998)) >> 21; expo = 0; } else { // // No overflow or underflow, rebase the exponent and round the mantissa // Magic numbers: // 0x200 = 00000010 00000000 -- masks off the 10th bit // // Ensure the first mantissa bit (the 10th one) is 1 and round expo = (expo - 0x3F000000) >> 10; sigf = ((u32[1] & 0x000FFFFF) + 0x200) >> 10; } return sign | expo | sigf & 0xFFFF; }