/* * Copyright 2021 Alyssa Rosenzweig * SPDX-License-Identifier: MIT */ #pragma once #include #include "util/macros.h" /* AGX includes an 8-bit floating-point format for small dyadic immediates, * consisting of 3 bits for the exponent, 4 bits for the mantissa, and 1-bit * for sign, in the usual order. Zero exponent has special handling. */ static inline float agx_minifloat_decode(uint8_t imm) { float sign = (imm & 0x80) ? -1.0 : 1.0; signed exp = (imm & 0x70) >> 4; unsigned mantissa = (imm & 0xF); if (exp) return ldexpf(sign * (float)(mantissa | 0x10), exp - 7); else return ldexpf(sign * ((float)mantissa), -6); } /* Encodes a float. Results are only valid if the float can be represented * exactly, if not the result of this function is UNDEFINED. However, it is * guaranteed that this function will not crash on out-of-spec inputs, so it is * safe to call on any input. signbit() is used to ensure -0.0 is handled * correctly. */ static inline uint8_t agx_minifloat_encode(float f) { unsigned sign = signbit(f) ? 0x80 : 0; f = fabsf(f); /* frac is in [0.5, 1) and f = frac * 2^exp */ int exp = 0; float frac = frexpf(f, &exp); if (f >= 0.25) { unsigned mantissa = (frac * 32.0); exp -= 5; /* 2^5 = 32 */ exp = CLAMP(exp + 7, 0, 7); return sign | (exp << 4) | (mantissa & 0xF); } else { unsigned mantissa = (f * 64.0f); return sign | mantissa; } } static inline bool agx_minifloat_exact(float f) { float f_ = agx_minifloat_decode(agx_minifloat_encode(f)); return memcmp(&f, &f_, sizeof(float)) == 0; }