int alloc_part() {
    const uint s0 = atomicAdd(fcnt[1], 1u);
    for (uint k = 0u; k < 64u; ++k) {
        const uint s = (s0 + k) % uint(FPART_MAX);
        const uint word = 4u + uint(FP_MAX) * 24u + s * 12u + 4u;
        const uint previous = atomicAdd(fxwords[word], 0u);
        if (uintBitsToFloat(previous) <= 0.0 && atomicCompSwap(fxwords[word], previous, floatBitsToUint(3.0e38)) == previous) return int(s);
    }
    return -1;
}

void release_part(uint slot) {
    // Publish a free slot only after its previous owner has finished all writes.
    memoryBarrierBuffer();
    atomicExchange(fxwords[4u + uint(FP_MAX) * 24u + slot * 12u + 4u], 0u);
}