* Copyright (c) 2026 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#ifndef OP_API_HIFLOAT4_H
#define OP_API_HIFLOAT4_H
#include <cstdint>
#include <limits>
#include <ostream>
namespace op {
* @brief HiFloat4 floating-point format
*
* HiFloat4 is a 4-bit floating-point format designed for high-efficiency
* neural network computations. It provides an optimal balance between
* dynamic range and precision for quantized inference and training scenarios.
*
* Format: 1 sign bit, 1 exponent bit, 2 mantissa bits
* +---+--+---+
* | S |E |MM |
* +---+--+---+
*
* Key characteristics:
* - Bias: 1
* - Max normal value: 1.75 (0x7 = 0 1 11, exp=1, man=3 -> 2^0 * 1.75)
* - Min normal value: -1.75 (0xF = 1 1 11, exp=1, man=3 -> -2^0 * 1.75)
* - Min positive normal: 2^0 = 1.0 (0x4 = 0 1 00)
* - Min positive denorm: 2^(1-bias) * (man/4) = 2^0 * 0.25 = 0.25 (0x1 = 0 0 01)
* - No NaN representation (all bit patterns are valid numbers)
* - No infinity representation
*
* This format is particularly suitable for:
* - Quantized neural network inference
* - Low-precision training with gradient scaling
* - Memory-bandwidth constrained applications
*/
struct HiFloat4 {
struct FromBitsTag {};
static constexpr FromBitsTag FromBits() { return FromBitsTag(); }
uint8_t value;
static constexpr uint8_t HIGHEST_VALUE = 0x07;
static constexpr uint8_t LOWEST_VALUE = 0x0F;
static constexpr uint8_t MIN_POS_NORMAL_VALUE = 0x04;
static constexpr uint8_t BITS_MASK = 0x0F;
static constexpr uint8_t EPSILON_VALUE = 0x01;
constexpr HiFloat4() : value(0) {}
constexpr HiFloat4(uint8_t bits, [[maybe_unused]] FromBitsTag fromBits) : value(bits & BITS_MASK) {}
HiFloat4(float v);
operator float() const;
operator double() const;
static constexpr HiFloat4 Epsilon()
{
return HiFloat4(EPSILON_VALUE, FromBits());
}
static constexpr HiFloat4 Highest()
{
return HiFloat4(HIGHEST_VALUE, FromBits());
}
static constexpr HiFloat4 Lowest()
{
return HiFloat4(LOWEST_VALUE, FromBits());
}
static constexpr HiFloat4 MinPositiveNormal()
{
return HiFloat4(MIN_POS_NORMAL_VALUE, FromBits());
}
bool IsZero() const;
bool IsNaN() const;
bool IsInf() const;
private:
union FP32 {
uint32_t u;
float f;
};
static float HiFloat4ToFloat(HiFloat4 fp4);
static HiFloat4 FloatToHiFloat4(float f);
};
inline std::ostream& operator<<(std::ostream& os, const HiFloat4& dt)
{
os << static_cast<float>(dt);
return os;
}
}
namespace std {
inline bool isinf(const op::HiFloat4& a [[maybe_unused]])
{
return false;
}
inline bool isnan(const op::HiFloat4& a) { return a.IsNaN(); }
inline bool isfinite(const op::HiFloat4& a) { return !a.IsNaN(); }
template <>
class numeric_limits<op::HiFloat4> {
public:
static constexpr bool has_infinity = false;
static constexpr bool has_quiet_NaN = false;
static constexpr bool has_signaling_NaN = false;
static constexpr bool is_bounded = true;
static constexpr bool is_exact = false;
static constexpr bool is_integer = false;
static constexpr bool is_iec559 = false;
static constexpr bool is_modulo = false;
static constexpr bool is_signed = true;
static constexpr bool is_specialized = true;
static constexpr int digits = 2;
static constexpr int digits10 = 0;
static constexpr int max_digits10 = 1;
static constexpr int min_exponent = 0;
static constexpr int min_exponent10 = 0;
static constexpr int max_exponent = 0;
static constexpr int max_exponent10 = 0;
static constexpr int radix = 2;
static constexpr float round_error() { return 0.5f; }
static constexpr float denorm_min()
{
return 0.25f;
}
static constexpr op::HiFloat4 min() { return op::HiFloat4::MinPositiveNormal(); }
static constexpr op::HiFloat4 lowest() { return op::HiFloat4::Lowest(); }
static constexpr op::HiFloat4 max() { return op::HiFloat4::Highest(); }
static constexpr op::HiFloat4 epsilon() { return op::HiFloat4::Epsilon(); }
static constexpr op::HiFloat4 infinity()
{
return op::HiFloat4::Highest();
}
static constexpr op::HiFloat4 quiet_NaN()
{
return op::HiFloat4(0x00, op::HiFloat4::FromBits());
}
static constexpr op::HiFloat4 signaling_NaN() { return quiet_NaN(); }
};
}
#endif