C++ Float32 与 Float16 互转

三采

已于 2023-12-15 14:56:18 修改

阅读量1.7k

点赞数

分类专栏： C++ 文章标签： C++

于 2023-07-20 10:35:06 首次发布

本文链接：https://blog.csdn.net/zywhehe/article/details/131824810

版权

C++ 专栏收录该内容

5 篇文章 0 订阅

订阅专栏

float32转float16

void Float32ToFloat16(uint16_t *result, float fp)
{
    if (fp > FP16_MAX) {
        (*result) = 0x7fff;
        return;
    } else if (fp < (-FP16_MAX)) {
        *result = 0xffff;
        return;
    } else if (fp < FP16_MIN && fp > (-FP16_MIN)) {
        *result = 0;
        return;
    }
    uint32_t fpBit = 0;
    int32_t ret = memcpy_s(&fpBit, sizeof(uint32_t), &fp, sizeof(float));
    LOG_IF_FAIL(ret == 0, "Float32ToFloat16 memcpy_s error %d", ret);
    uint32_t sign = ((fpBit >> 31) & 1);
    uint32_t exponent = ((fpBit >> 23) & 0xff);
    uint32_t fraction = ((fpBit >> 13) & 0x3ff);
    while(exponent < 0x71) {
        fraction >>= 1;
        exponent += 1;
    }
    exponent -= 0x70;
    *result = ((sign << 15) | (exponent << 10) | fraction);
    return;
}

之前用了ChatGPT生成，经提醒似乎是错的，已更新

float16转float32

void Float16ToFloat32(float *result, uint16_t halfFloat)
{
    // 符号+指数+尾数 FP16: 1+5+10 FP32: 1+8+23
    uint32_t sign = ((halfFloat >> 15) & 1);                // 15: 符号位
    uint32_t exponent = ((halfFloat >> 10) & 0x1f);         // 10: 去除尾数
    uint32_t fraction = ((halfFloat & 0x3ff) << 13);        // 13: fp32与fp16尾数位数差  0x3ff: 2^10-1,取位数
    if (exponent == 0x1f) {
        // nan或inf,除符号位外全1
        fraction = (fraction ? (sign = 0, 0x7fffff) : 0);   // 0x7fffff: 2^23-1
        exponent = 0xff;                                    // 0xff: 2^8-1
    } else if (!exponent) {
        // zero或denorm，指数为0
        if (fraction) {
            // not zero
            uint32_t m;
            exponent = 0x71;                                // 0x71: 2^7-2^4+1,指数移码差值+1
            do {
                m = (fraction & 0x400000);                  // 0x400000: 2^(23-1),取fraction的第一个非0位
                fraction <<= 1;
                --exponent;
            } while (!m);
            fraction &= 0x7fffff;                           // 0x7fffff: 2^23-1
        }
    } else {
        exponent += 0x70;                                   // 0x70: 2^7-2^4,指数移码差值
    }
    uint32_t temp = ((sign << 31) | (exponent << 23) | fraction);
    int32_t ret = memcpy_s(&result, sizeof(float), &temp, sizeof(uint32_t));
    LOG_IF_FAIL(ret == 0, "Float16ToFloat32 memcpy_s error %d", ret);
    return;
}