playcanvas
Version:
PlayCanvas WebGL game engine
103 lines (87 loc) • 4.03 kB
JavaScript
// Generic half-float packing with software fallback for subnormals
// Addresses vendor differences in packHalf2x16 subnormal handling (e.g., Adreno (TM) 750 on Samsung Galaxy S24)
var packHalfPS = /* glsl */ `
// Software pack of one f32 -> f16 (low 16 bits). Ties-to-even, full subnormals.
uint floatToHalf(float a) {
uint u = floatBitsToUint(a);
uint sign = (u >> 16u) & 0x8000u;
uint absu = u & 0x7FFFFFFFu;
uint man = u & 0x007FFFFFu;
int e32 = int((u >> 23u) & 0xFFu) - 127;
// NaN / Inf
if ((absu & 0x7F800000u) == 0x7F800000u) {
bool isnan = (man != 0u);
return sign | (isnan ? 0x7E00u : 0x7C00u);
}
// Overflow to Inf
if (e32 > 15) return sign | 0x7C00u;
// Normal half
if (e32 >= -14) {
uint he = uint(e32 + 15);
uint hm = man >> 13u;
uint rem = man & 0x1FFFu;
uint add = (rem > 0x1000u || (rem == 0x1000u && (hm & 1u) == 1u)) ? 1u : 0u;
hm = (hm + add) & 0x3FFu;
if ((hm & 0x400u) != 0u) {
hm = 0u; he = he + 1u;
if (he >= 31u) return sign | 0x7C00u;
}
return sign | (he << 10u) | hm;
}
// Subnormals
if (e32 >= -24) {
uint s = uint(-(e32 + 1));
uint mnorm = 0x00800000u | man;
uint hm = mnorm >> s;
uint mask = (1u << s) - 1u;
uint rem = mnorm & mask;
uint halfBt = 1u << (s - 1u);
uint add = (rem > halfBt || (rem == halfBt && (hm & 1u) == 1u)) ? 1u : 0u;
hm = hm + add;
if (hm >= 0x400u) return sign | (1u << 10u);
return sign | hm;
}
// Underflow to signed zero
return sign;
}
// Hybrid pack: software for subnormals, builtin for normal range
uint packHalf2x16Safe(vec2 v) {
// Convert the input floats to their 32-bit IEEE-754 bit patterns.
// We'll inspect the exponent bits directly to determine their numeric range.
uint u_x = floatBitsToUint(v.x);
uint u_y = floatBitsToUint(v.y);
// Extract the unbiased exponent for each component (float32 uses bias = 127).
// e32 = exponent - 127 ⇒ actual power of two for each value.
int e32_x = int((u_x >> 23u) & 0xFFu) - 127;
int e32_y = int((u_y >> 23u) & 0xFFu) - 127;
// -------------------------------------------------------------------------
// Detect values that would become *subnormal* (or zero) in float16.
//
// e32 < -14 ⇔ |value| < 2^-14 ≈ 6.1035e-5
//
// Many mobile GPUs (including Adreno and Mali) mishandle half-precision
// subnormals—typically flushing them to zero or rounding incorrectly.
// To preserve correct rounding and sign, we use the software conversion
// path (floatToHalf) for these small magnitudes.
//
// The software branch runs very rarely (<0.1% of typical values for
// normalized scene data) and costs only a few ALU instructions, so the
// performance impact is negligible while avoiding visible precision loss.
// -------------------------------------------------------------------------
if (e32_x < -14 || e32_y < -14) {
// Convert both components with the reference software routine
// and pack into a 32-bit uint: low 16 bits = x, high 16 bits = y.
return (floatToHalf(v.y) << 16u) | floatToHalf(v.x);
}
// Normal range: use the fast hardware builtin
return packHalf2x16(v);
}
// On non-Android platforms, use builtin directly (no subnormal workaround needed)
uint packHalf2x16Safe(vec2 v) {
return packHalf2x16(v);
}
`;
export { packHalfPS as default };