// Exact 32x32 -> two-word product without shaderInt64.
// Donor: cryptography/pqc/mldsa/mldsa_field.slang, mul32_wide/mulhi32.
// Adaptation: mechanical extraction; preserve the independently proved formula.
uint2 mul32_wide(uint a, uint b) {
uint a_lo = a & 0xFFFFu;
uint a_hi = a >> 16u;
uint b_lo = b & 0xFFFFu;
uint b_hi = b >> 16u;
uint lo_lo = a_lo * b_lo; // 0..32 bits
uint lo_hi = a_lo * b_hi; // 16..48 bits
uint hi_lo = a_hi * b_lo; // 16..48 bits
uint hi_hi = a_hi * b_hi; // 32..64 bits
uint mid = (lo_lo >> 16u) + (lo_hi & 0xFFFFu) + (hi_lo & 0xFFFFu);
uint carry = mid >> 16u;
uint result_lo = (lo_lo & 0xFFFFu) | ((mid & 0xFFFFu) << 16u);
uint result_hi = hi_hi + (lo_hi >> 16u) + (hi_lo >> 16u) + carry;
return uint2(result_lo, result_hi);
}
// Return high 32 bits of a*b.
uint mulhi32(uint a, uint b) {
return mul32_wide(a, b).y;
}