1// These tests may take a long time!
 2// They are to prove that conversion from double to float of various functions in ggml.c doesn't affect the result.
 3// This is done by checking all finite (non-NaN, non-infinite) floats.
 4
 5#undef NDEBUG
 6#include <cassert>
 7#if !defined(__riscv) && !defined(__s390__) && !defined(__ARM_NEON)
 8#include <immintrin.h>
 9#endif
10#include <cmath>
11#include <cstdint>
12#include <cstring>
13
14#pragma GCC diagnostic push
15#pragma GCC diagnostic ignored "-Wdouble-promotion"
16
17// ggml.c::quantize_row_q4_0_ref
18inline static uint8_t round_orig(float v0) { return ((int8_t) (round(v0))) + 8; }
19
20// ggml.c::ggml_silu_f32
21inline static float silu_orig(float x) {
22    return x/(1.0 + exp(-x));
23}
24
25#pragma GCC diagnostic pop
26
27// ggml.c::quantize_row_q4_0_ref
28inline static uint8_t round_float(float v0) { return (int8_t)roundf(v0) + 8; }
29
30// ggml.c::ggml_silu_f32
31inline static float silu_float(float x) {
32    return x/(1.0f + expf(-x));
33}
34
35int main(void) {
36    uint32_t x = UINT32_MAX;
37    do {
38        float f;
39        memcpy(&f, &x, sizeof(x));
40        assert(!std::isfinite(f) || (round_orig(f) == round_float(f)));
41    } while (x--);
42
43#ifdef __F16C__
44    // GELU and SILU implementations are used with a FP16 lookup table.
45    // The original and float-only results are not equal for all inputs after converting to FP16.
46    // GELU is an approximation anyway (tanh), not tested here.
47    // For SILU, verify that the results are at least the closest floating point numbers, if the FP16 values don't match.
48    for (x = 0; x <= UINT16_MAX; x++) {
49        float f = _cvtsh_ss(x);
50        const float so = silu_orig(f);
51        const float sf = silu_float(f);
52        assert(   (_cvtss_sh(so, 0) == _cvtss_sh(sf, 0))
53               || (nextafterf(so, sf) == sf)
54               || (nextafterf(sf, so) == so));
55    }
56#endif
57}