mirror of
https://github.com/ikawrakow/ik_llama.cpp.git
synced 2026-01-26 17:20:01 +00:00
1f9541172f
Use 3 bits for the exponent and 5 bits for the mantissa. This makes PPL to be the same as fp16 (but the previous version with 4 bits for the exponent and mantissa was good enough for any practical purposes).
47 lines
1.0 KiB
C
47 lines
1.0 KiB
C
#pragma once
|
|
|
|
#include <stdint.h>
|
|
|
|
typedef union {
|
|
float f;
|
|
uint32_t i;
|
|
} iq1bn_scale_t;
|
|
|
|
#ifdef __cplusplus
|
|
extern "C" {
|
|
#endif
|
|
|
|
#ifdef BITNET_IQ1BN_4x4
|
|
static inline float iq1bn_min_value(void) { return 1.9074e-06f; }
|
|
static inline float iq1bn_max_value(void) { return 0.12109f; }
|
|
#else
|
|
static inline float iq1bn_min_value(void) { return 0.000488281f; }
|
|
static inline float iq1bn_max_value(void) { return 0.123047f; }
|
|
#endif
|
|
|
|
static inline uint8_t iq1bn_float_to_fp8(float f) {
|
|
if (f <= iq1bn_min_value()) return 0;
|
|
if (f >= iq1bn_max_value()) return 255;
|
|
iq1bn_scale_t s;
|
|
s.f = f;
|
|
#ifdef BITNET_IQ1BN_4x4
|
|
return ((((s.i >> 23) + 132) & 0xf) << 4) | ((s.i >> 19) & 0xf);
|
|
#else
|
|
return ((s.i >> 18) & 0x1f) | (((s.i >> 23) - 116) << 5);
|
|
#endif
|
|
}
|
|
|
|
static inline float iq1bn_fp8_to_float(uint8_t fp8) {
|
|
iq1bn_scale_t s;
|
|
#ifdef BITNET_IQ1BN_4x4
|
|
s.i = ((((fp8 >> 4) | 0xf0) - 132) << 23) | ((fp8 & 0x0f) << 19);
|
|
#else
|
|
s.i = (((fp8 >> 5) + 116) << 23) | ((fp8 & 0x1f) << 18);
|
|
#endif
|
|
return s.f;
|
|
}
|
|
|
|
#ifdef __cplusplus
|
|
}
|
|
#endif
|