STM32G4 Math Operation Benchmark Test

Test Condition

MCU: STM32G431RB @170MHz
IDE: IAR V9.40
Optimization: -o3

Math Operation Benchmark test

OP times: 1M

Multiply

上述表格中Q15的测试数据为把相乘的结果赋值给int16_t,如果改为把结果赋值给int32_t，则所耗时间与Q31一致，更少。

Division

似乎同样的代码重新编译以后测试结果又不一样。。。

几点观察结论：

• 浮点数的乘除用时与操作数无关
• 浮点数的乘法用时比定点多 20% 左右
• 浮点数的除法所用时间在操作数较小时大约是定点数的两倍多，当被除数值接近极限时32位定点除法与浮点除法接近
• 盲猜在电机控制程序里，如果用全浮点数计算，所耗时间资源将会大约是定点数的150%左右

Test Code

volatile uint32_t sMulDurationUs;
volatile uint32_t iMulDurationUs;
volatile uint32_t fMulDurationUs;

volatile uint32_t sDivDurationUs;
volatile uint32_t iDivDurationUs;
volatile uint32_t fDivDurationUs;

void MathOpBenchmarkTest(void)
{

  volatile int16_t sa = 1;
  volatile int16_t sb = 2;
  volatile int16_t sc = 3;

  volatile int32_t ia = 1;
  volatile int32_t ib = 2;
  volatile int32_t ic = 3;

  volatile float32_t fa = 1;
  volatile float32_t fb = 2000;
  volatile float32_t fc = 3;

  volatile uint32_t startTick = DWT->CYCCNT;
  for (uint32_t i = 1000000; i > 0; i--) {
    sc = sa * sb;
  }
  volatile uint32_t stopTick = DWT->CYCCNT;
  sMulDurationUs = TimeMeasure(startTick, stopTick);

  startTick = DWT->CYCCNT;
  for (uint32_t i = 1000000; i > 0; i--) {
    ic = ia * ib;
  }
  stopTick = DWT->CYCCNT;
  iMulDurationUs = TimeMeasure(startTick, stopTick);

  startTick = DWT->CYCCNT;
  for (uint32_t i = 1000000; i > 0; i--) {
    fc = fa * fb;
  }
  stopTick = DWT->CYCCNT;
  fMulDurationUs = TimeMeasure(startTick, stopTick);

  startTick = DWT->CYCCNT;
  for (uint32_t i = 1000000; i > 0; i--) {
    sc = sa / sb;
  }
  stopTick = DWT->CYCCNT;
  sDivDurationUs = TimeMeasure(startTick, stopTick);

  startTick = DWT->CYCCNT;
  for (uint32_t i = 1000000; i > 0; i--) {
    ic = ia / ib;
  }
  stopTick = DWT->CYCCNT;
  iDivDurationUs = TimeMeasure(startTick, stopTick);

  startTick = DWT->CYCCNT;
  for (uint32_t i = 1000000; i > 0; i--) {
    fc = fa / fb;
  }
  stopTick = DWT->CYCCNT;
  fDivDurationUs = TimeMeasure(startTick, stopTick);

  sc++;
  ic++;
  fc++;
  
}