char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Max Krasnyanskyhexagon: support for OP_CPY, host buffers now optional, hvx-utils refactoring and optimizations (#18822)cff777f22

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
1.1 KiB37 linesraw
1#ifndef HEX_FASTDIV_H
2#define HEX_FASTDIV_H
3
4// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
5// Precompute mp (m' in the paper) and L such that division
6// can be computed using a multiply (high 32b of 64b result)
7// and a shift:
8//
9// n/d = (mulhi(n, mp) + n) >> L;
10struct fastdiv_values {
11    uint32_t mp;
12    uint32_t l;
13};
14
15static inline struct fastdiv_values init_fastdiv_values(uint32_t d) {
16    struct fastdiv_values result = { 0, 0 };
17    // compute L = ceil(log2(d));
18    while (result.l < 32 && ((uint32_t) 1 << result.l) < d) {
19        ++(result.l);
20    }
21
22    result.mp = (uint32_t) (((uint64_t) 1 << 32) * (((uint64_t) 1 << result.l) - d) / d + 1);
23    return result;
24}
25
26static inline uint32_t fastdiv(uint32_t n, const struct fastdiv_values * vals) {
27    // Compute high 32 bits of n * mp
28    const uint32_t hi = (uint32_t) (((uint64_t) n * vals->mp) >> 32);  // mulhi(n, mp)
29    // add n, apply bit shift
30    return (hi + n) >> vals->l;
31}
32
33static inline uint32_t fastmodulo(uint32_t n, uint32_t d, const struct fastdiv_values * vals) {
34    return n - fastdiv(n, vals) * d;
35}
36
37#endif /* HEX_FASTDIV_H */