[libc-commits] [libc] [llvm] [libc][math][c23] Improve rsqrtf16() function (PR #160639)

[via libc-commits]

================
@@ -24,6 +24,193 @@
 namespace LIBC_NAMESPACE_DECL {
 namespace math {
 
+namespace rsqrtf16_internal {
+
+// Fixed-point computations below use Q29: the integer N represents
+// N * 2^-29.  Multiplying two Q29 values produces a Q58 value, so products are
+// shifted right by RSQRT_FRACTION_BITS to return to Q29.
+LIBC_INLINE_VAR constexpr int RSQRT_FRACTION_BITS = 29;
+LIBC_INLINE_VAR constexpr int64_t ONE = int64_t(1) << RSQRT_FRACTION_BITS;
+LIBC_INLINE_VAR constexpr int64_t THREE_HALVES = 3 * (ONE >> 1);
+
+// Midpoint lookup table for 1/sqrt(x) on 16 sub-intervals of [0.5;1).
+// Values are stored in Q29 fixed-point format.  The Newton step and exact
+// rounding below correct the seed before producing the final half result.
+LIBC_INLINE_VAR constexpr uint32_t RSQRT_APPROX[16] = {
----------------
lntue wrote:

Can this table be printed in hex instead?

https://github.com/llvm/llvm-project/pull/160639