MCU进行双精度与单精度浮点数运算的误差比较测试

测试平台

MCU: STM32F767，启动硬件双精度浮点运算协处理器
IDE: Keil RVMDK V5.21.1.0

测试方法

浮点数0.1分别以单精度与双精度累加100000/1000000/10000000/100000000次，打印输出累加结果。

测试代码

    double dVal = 0;
    float fVal = 0;
    for(int i = 0; i < 100000001; i++){
        if((i == 100000) || (i == 1000000) || (i == 10000000) || (i == 100000000)){
            printf("Test floating point add %d times: dVal = %f, fVal = %f\r\n", i, dVal, fVal);
        }
        dVal += 0.1;
        fVal += 0.1f;
    }

测试结果

Test floating point add 100000 times: dVal = 10000.000000, fVal = 9998.556641
Test floating point add 1000000 times: dVal = 100000.000001, fVal = 100958.343750
Test floating point add 10000000 times: dVal = 999999.999839, fVal = 1087937.000000
Test floating point add 100000000 times: dVal = 9999999.981129, fVal = 2097152.000000

结论

单精度浮点运算损失的精度是巨大的，在某些要求严格的场合下造成的误差不可容忍，必须引起重视！