@luma.gl/shadertools 9.3.6 → 9.4.0-alpha.1

package/src/modules/math/fp64/fp64-arithmetic-wgsl.ts

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: MIT
 // Copyright (c) vis.gl contributors
 
+/** WGSL source for fp64 arithmetic helpers and raw binary64-to-Float32 subtraction. */
 export const fp64arithmeticWGSL = /* wgsl */ `\
 struct Fp64ArithmeticUniforms {
   ONE: f32,
@@ -10,11 +11,278 @@ struct Fp64ArithmeticUniforms {
 
 @group(0) @binding(auto) var<uniform> fp64arithmetic : Fp64ArithmeticUniforms;
 
+struct Fp64Bits {
+  sign: u32,
+  exponent: i32,
+  significand: vec2u,
+  isZero: bool,
+  isInf: bool,
+  isNan: bool,
+};
+
 fn fp64_nan(seed: f32) -> f32 {
   let nanBits = 0x7fc00000u | select(0u, 1u, seed < 0.0);
   return bitcast<f32>(nanBits);
 }
 
+fn fp64_u64_is_zero(value: vec2u) -> bool {
+  return value.x == 0u && value.y == 0u;
+}
+
+fn fp64_u64_compare(a: vec2u, b: vec2u) -> i32 {
+  if (a.x != b.x) {
+    return select(-1, 1, a.x > b.x);
+  }
+  if (a.y != b.y) {
+    return select(-1, 1, a.y > b.y);
+  }
+  return 0;
+}
+
+fn fp64_u64_add(a: vec2u, b: vec2u) -> vec2u {
+  let low = a.y + b.y;
+  let carry = select(0u, 1u, low < a.y);
+  return vec2u(a.x + b.x + carry, low);
+}
+
+fn fp64_u64_sub(a: vec2u, b: vec2u) -> vec2u {
+  let borrow = select(0u, 1u, a.y < b.y);
+  return vec2u(a.x - b.x - borrow, a.y - b.y);
+}
+
+fn fp64_u64_shift_left(value: vec2u, shift: u32) -> vec2u {
+  if (shift == 0u) {
+    return value;
+  }
+  if (shift < 32u) {
+    return vec2u((value.x << shift) | (value.y >> (32u - shift)), value.y << shift);
+  }
+  if (shift == 32u) {
+    return vec2u(value.y, 0u);
+  }
+  if (shift < 64u) {
+    return vec2u(value.y << (shift - 32u), 0u);
+  }
+  return vec2u(0u);
+}
+
+fn fp64_u64_shift_right(value: vec2u, shift: u32) -> vec2u {
+  if (shift == 0u) {
+    return value;
+  }
+  if (shift < 32u) {
+    return vec2u(value.x >> shift, (value.y >> shift) | (value.x << (32u - shift)));
+  }
+  if (shift == 32u) {
+    return vec2u(0u, value.x);
+  }
+  if (shift < 64u) {
+    return vec2u(0u, value.x >> (shift - 32u));
+  }
+  return vec2u(0u);
+}
+
+fn fp64_u64_get_bit(value: vec2u, bitIndex: u32) -> bool {
+  if (bitIndex >= 64u) {
+    return false;
+  }
+  if (bitIndex >= 32u) {
+    return ((value.x >> (bitIndex - 32u)) & 1u) != 0u;
+  }
+  return ((value.y >> bitIndex) & 1u) != 0u;
+}
+
+fn fp64_u64_has_bits_below(value: vec2u, bitCount: u32) -> bool {
+  if (bitCount == 0u) {
+    return false;
+  }
+  if (bitCount >= 64u) {
+    return !fp64_u64_is_zero(value);
+  }
+  if (bitCount > 32u) {
+    let highBitCount = bitCount - 32u;
+    let highMask = (1u << highBitCount) - 1u;
+    return value.y != 0u || (value.x & highMask) != 0u;
+  }
+  if (bitCount == 32u) {
+    return value.y != 0u;
+  }
+  let lowMask = (1u << bitCount) - 1u;
+  return (value.y & lowMask) != 0u;
+}
+
+fn fp64_u64_shift_right_sticky(value: vec2u, shift: u32) -> vec2u {
+  var shifted = fp64_u64_shift_right(value, shift);
+  if (fp64_u64_has_bits_below(value, shift)) {
+    shifted.y = shifted.y | 1u;
+  }
+  return shifted;
+}
+
+fn fp64_u64_count_leading_zeros(value: vec2u) -> u32 {
+  if (value.x != 0u) {
+    return countLeadingZeros(value.x);
+  }
+  return 32u + countLeadingZeros(value.y);
+}
+
+fn fp64_round_shift_right_to_u32(value: vec2u, shift: u32) -> u32 {
+  if (shift == 0u) {
+    return value.y;
+  }
+
+  let truncated = fp64_u64_shift_right(value, shift);
+  var rounded = truncated.y;
+  let guard = fp64_u64_get_bit(value, shift - 1u);
+  let hasTrailingBits = fp64_u64_has_bits_below(value, shift - 1u);
+  if (guard && (hasTrailingBits || (rounded & 1u) == 1u)) {
+    rounded = rounded + 1u;
+  }
+  return rounded;
+}
+
+fn fp64_make_f32_bits_from_u64(sign: u32, significand: vec2u, baseExponent: i32) -> u32 {
+  if (fp64_u64_is_zero(significand)) {
+    return sign << 31u;
+  }
+
+  let leadingZeros = fp64_u64_count_leading_zeros(significand);
+  let mostSignificantBit = 63u - leadingZeros;
+  var exponent = baseExponent + i32(mostSignificantBit);
+
+  if (exponent > 127) {
+    return (sign << 31u) | 0x7f800000u;
+  }
+
+  if (exponent >= -126) {
+    let shift = i32(mostSignificantBit) - 23;
+    var significand24: u32;
+    if (shift > 0) {
+      significand24 = fp64_round_shift_right_to_u32(significand, u32(shift));
+    } else {
+      significand24 = fp64_u64_shift_left(significand, u32(-shift)).y;
+    }
+
+    if (significand24 >= 0x1000000u) {
+      significand24 = significand24 >> 1u;
+      exponent = exponent + 1;
+      if (exponent > 127) {
+        return (sign << 31u) | 0x7f800000u;
+      }
+    }
+
+    return (sign << 31u) | (u32(exponent + 127) << 23u) | (significand24 & 0x7fffffu);
+  }
+
+  let scaleExponent = baseExponent + 149;
+  var mantissa: u32;
+  if (scaleExponent >= 0) {
+    mantissa = fp64_u64_shift_left(significand, u32(scaleExponent)).y;
+  } else {
+    mantissa = fp64_round_shift_right_to_u32(significand, u32(-scaleExponent));
+  }
+
+  if (mantissa >= 0x800000u) {
+    return (sign << 31u) | 0x00800000u;
+  }
+  return (sign << 31u) | mantissa;
+}
+
+fn fp64_decode_bits(bits: vec2u) -> Fp64Bits {
+  let sign = bits.x >> 31u;
+  let exponentBits = (bits.x >> 20u) & 0x7ffu;
+  let fractionHigh = bits.x & 0xfffffu;
+  let fractionLow = bits.y;
+  let fraction = vec2u(fractionHigh, fractionLow);
+
+  if (exponentBits == 0x7ffu) {
+    let isInf = fp64_u64_is_zero(fraction);
+    return Fp64Bits(sign, 0, vec2u(0u), false, isInf, !isInf);
+  }
+
+  if (exponentBits == 0u) {
+    let isZero = fp64_u64_is_zero(fraction);
+    return Fp64Bits(sign, -1022, fraction, isZero, false, false);
+  }
+
+  return Fp64Bits(sign, i32(exponentBits) - 1023, vec2u((1u << 20u) | fractionHigh, fractionLow), false, false, false);
+}
+
+fn fp64_finite_magnitude_compare(a: Fp64Bits, b: Fp64Bits) -> i32 {
+  if (a.exponent != b.exponent) {
+    return select(-1, 1, a.exponent > b.exponent);
+  }
+  return fp64_u64_compare(a.significand, b.significand);
+}
+
+fn fp64_add_aligned_magnitudes_to_f32_bits(sign: u32, larger: Fp64Bits, smaller: Fp64Bits) -> u32 {
+  let largeSignificand = fp64_u64_shift_left(larger.significand, 3u);
+  let smallSignificand = fp64_u64_shift_right_sticky(
+    fp64_u64_shift_left(smaller.significand, 3u),
+    u32(larger.exponent - smaller.exponent)
+  );
+  let resultSignificand = fp64_u64_add(largeSignificand, smallSignificand);
+  return fp64_make_f32_bits_from_u64(sign, resultSignificand, larger.exponent - 55);
+}
+
+fn fp64_sub_aligned_magnitudes_to_f32_bits(sign: u32, larger: Fp64Bits, smaller: Fp64Bits) -> u32 {
+  let largeSignificand = fp64_u64_shift_left(larger.significand, 3u);
+  let smallSignificand = fp64_u64_shift_right_sticky(
+    fp64_u64_shift_left(smaller.significand, 3u),
+    u32(larger.exponent - smaller.exponent)
+  );
+  let resultSignificand = fp64_u64_sub(largeSignificand, smallSignificand);
+  return fp64_make_f32_bits_from_u64(sign, resultSignificand, larger.exponent - 55);
+}
+
+// Subtract two raw binary64 values and round the exact result once to f32.
+// The input words are canonical high/low words: .x contains sign/exponent/high
+// fraction bits, and .y contains the low 32 fraction bits.
+fn sub_fp64u32_to_f32_bits(aBits: vec2u, bBits: vec2u) -> u32 {
+  let a = fp64_decode_bits(aBits);
+  let b = fp64_decode_bits(bBits);
+  let bSubtractionSign = b.sign ^ 1u;
+
+  if (a.isNan || b.isNan) {
+    return 0x7fc00000u;
+  }
+  if (a.isInf && b.isInf) {
+    if (a.sign == bSubtractionSign) {
+      return (a.sign << 31u) | 0x7f800000u;
+    }
+    return 0x7fc00000u;
+  }
+  if (a.isInf) {
+    return (a.sign << 31u) | 0x7f800000u;
+  }
+  if (b.isInf) {
+    return (bSubtractionSign << 31u) | 0x7f800000u;
+  }
+  if (a.isZero && b.isZero) {
+    return select(0u, 0x80000000u, a.sign == 1u && b.sign == 0u);
+  }
+
+  let magnitudeComparison = fp64_finite_magnitude_compare(a, b);
+  if (a.sign == bSubtractionSign) {
+    if (magnitudeComparison >= 0) {
+      return fp64_add_aligned_magnitudes_to_f32_bits(a.sign, a, b);
+    }
+    return fp64_add_aligned_magnitudes_to_f32_bits(a.sign, b, a);
+  }
+
+  if (magnitudeComparison == 0) {
+    return 0u;
+  }
+  if (magnitudeComparison > 0) {
+    return fp64_sub_aligned_magnitudes_to_f32_bits(a.sign, a, b);
+  }
+  return fp64_sub_aligned_magnitudes_to_f32_bits(bSubtractionSign, b, a);
+}
+
+fn sub_fp64u32_to_f32(aBits: vec2u, bBits: vec2u) -> f32 {
+  return bitcast<f32>(sub_fp64u32_to_f32_bits(aBits, bBits));
+}
+
 fn fp64_runtime_zero() -> f32 {
   return fp64arithmetic.ONE * 0.0;
 }