[参考译文] MSP430FR5969：Q-math 库是否改善了此 MCU 的能耗？

admin

Other Parts Discussed in Thread: MSP430FR5969

请注意，本文内容源自机器翻译，可能存在语法或其它翻译错误，仅供参考。如需获取准确内容，请参阅链接中的英语原文或自行翻译。

https://e2e.ti.com/support/microcontrollers/msp-low-power-microcontrollers-group/msp430/f/msp-low-power-microcontroller-forum/1035536/msp430fr5969-does-q-math-library-improve-energy-consumption-on-this-mcu

器件型号：MSP430FR5969

大家好、

在过去的几天里、我一直在寻找一种在 MSP430FR5969 MCU 上高效实现矩阵乘法的方法、而这种方法浪费了太多的计算功率、同时也保持了低功耗。通过互联网、我们发现了两个名为 Qmath 和 IQmath 的库、这两个库都有望实现更快的乘法时间和更低的功耗。

遗憾的是、当我们对它们进行测试时、我们无法发现功耗的任何显著变化、也无法发现处理时间。我希望大家能给我一个手、因为我们可能缺少库的一些配置。

Comparison of consumptions

这是我使用 Qmath 库的代码(由于平均平方误差更小而配置 Q4)。

#include <msp430.h>
#include <stdint.h>

/* Select the global Q value and include the Qmath header file. */
#define GLOBAL_Q 4
#include "QmathLib.h"

#define sz 16
#define lpc 15

int main(void){
  WDTCTL = WDTPW | WDTHOLD;       // Stop WDT

  P1OUT = 0;
  P1DIR = 0xFF;

  P2OUT = 0;
  P2DIR = 0xFF;

  P3OUT = 0;
  P3DIR = 0xFF;

  P4OUT = 0;
  P4DIR = 0xFF;

  PJOUT = 0;
  PJDIR = 0xFFFF;

  PM5CTL0 &= ~LOCKLPM5;

  float fa[sz][sz] = {
  3.0451312,  -2.6170964,  -0.36237785, 0.67903620,  0.54625326,  -1.4386408,  1.1592991, -0.054826848,  -0.038200974,  -0.095277585,  -2.2473435,  3.3342791, -0.30391535, -1.0859760,  -0.34301928, 0.41210720,
  -2.6170964,  5.0436597, -2.0817051,  -0.79343557, 0.15403587,  1.5043120, -2.2707827,  1.2366960, -0.014145052,  0.049386151, 1.7008784, -4.7957544,  3.3069613, 0.57205218,  -0.67790008, -0.34301928,
  -0.36237785, -2.0817051,  4.9076262, -2.2785275,  -0.87535220, 0.58374476,  1.2477649, -2.2913878,  1.2353808, -0.0016905917, 0.38163254,  1.0874399, -4.5344720,  3.4830644, 0.57205218,  -1.0859760,
  0.67903620,  -0.79343557, -2.2785275,  4.6777167, -2.3003640,  0.0097080851,  0.048614282, 1.2007601, -2.3131311,  1.1977108, -0.11479994, 0.63922209,  1.2034756, -4.5344720,  3.3069613, -0.30391535,
  0.54625326,  0.15403587,  -0.87535220, -2.3003640,  4.9336734, -2.2747848,  -0.099792719,  0.039780665, 1.1884731, -2.3556819,  1.0750339, 0.21016976,  0.63922209,  1.0874399, -4.7957544,  3.3342791,
  -1.4386408,  1.5043120, 0.58374476,  0.0097080851,  -2.2747848,  1.8861172, -0.22406493, 0.013938670, 0.098963775, 1.3059657, -2.6704209,  1.0750339, -0.11479994, 0.38163254,  1.7008784, -2.2473435,
  1.1592991, -2.2707827,  1.2477649, 0.048614282, -0.099792719,  -0.22406493, 0.57083476,  -0.30296537, -0.026575994,  0.022921033, 1.3059657, -2.3556819,  1.1977108, -0.0016905917, 0.049386151, -0.095277585,
  -0.054826848,  1.2366960, -2.2913878,  1.2007601, 0.039780665, 0.013938670, -0.30296537, 0.56730592,  -0.30373546, -0.026575994,  0.098963775, 1.1884731, -2.3131311,  1.2353808, -0.014145052,  -0.038200974,
  -0.038200974,  -0.014145052,  1.2353808, -2.3131311,  1.1884731, 0.098963775, -0.026575994,  -0.30373546, 0.56730592,  -0.30296537, 0.013938670, 0.039780665, 1.2007601, -2.2913878,  1.2366960, -0.054826848,
  -0.095277585,  0.049386151, -0.0016905917, 1.1977108, -2.3556819,  1.3059657, 0.022921033, -0.026575994,  -0.30296537, 0.57083476,  -0.22406493, -0.099792719,  0.048614282, 1.2477649, -2.2707827,  1.1592991,
  -2.2473435,  1.7008784, 0.38163254,  -0.11479994, 1.0750339, -2.6704209,  1.3059657, 0.098963775, 0.013938670, -0.22406493, 1.8861172, -2.2747848,  0.0097080851,  0.58374476,  1.5043120, -1.4386408,
  3.3342791, -4.7957544,  1.0874399, 0.63922209,  0.21016976,  1.0750339, -2.3556819,  1.1884731, 0.039780665, -0.099792719,  -2.2747848,  4.9336734, -2.3003640,  -0.87535220, 0.15403587,  0.54625326,
  -0.30391535, 3.3069613, -4.5344720,  1.2034756, 0.63922209,  -0.11479994, 1.1977108, -2.3131311,  1.2007601, 0.048614282, 0.0097080851,  -2.3003640,  4.6777167, -2.2785275,  -0.79343557, 0.67903620,
  -1.0859760,  0.57205218,  3.4830644, -4.5344720,  1.0874399, 0.38163254,  -0.0016905917, 1.2353808, -2.2913878,  1.2477649, 0.58374476,  -0.87535220, -2.2785275,  4.9076262, -2.0817051,  -0.36237785,
  -0.34301928, -0.67790008, 0.57205218,  3.3069613, -4.7957544,  1.7008784, 0.049386151, -0.014145052,  1.2366960 -2.2707827,  1.5043120, 0.15403587,  -0.79343557, -2.0817051,  5.0436597, -2.6170964,
  0.41210720,  -0.34301928, -1.0859760,  -0.30391535, 3.3342791, -2.2473435, -0.095277585, -0.038200974, -0.054826848, 1.1592991, -1.4386408,  0.54625326,  0.67903620,  -0.36237785, -2.6170964,  3.0451312};

  float fb[sz] = {2.3, 5.1, 1.05, 2.33, 3.56, 2.3, 5.1, 1.05, 2.33, 3.56, 2.3, 5.1, 1.05, 2.33, 3.56, 1.1};
  float fout[sz] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};;
  _q out[sz];
  int n;
  int m;

  while(1){
    // Matrix-Vector multiplication
    for(n = lpc; n>=0; n--){
        out[n] = _Q(0);
        for(m = lpc; m>=0; m--)
            out[n] += _Qmpy(_Q(fa[n][m]), _Q(fb[m]));

        fout[n] = _QtoF(out[n]);
    }
    __delay_cycles(8000000);
    }
}

下面是使用通用 C 算术运算符实现的相同代码：

#include <msp430.h>
#include <stdint.h>

#define sz  16
#define lpc  15

int main(void){
  WDTCTL = WDTPW | WDTHOLD;       // Stop WDT

  P1OUT = 0;
  P1DIR = 0xFF;

  P2OUT = 0;
  P2DIR = 0xFF;

  P3OUT = 0;
  P3DIR = 0xFF;

  P4OUT = 0;
  P4DIR = 0xFF;

  PJOUT = 0;
  PJDIR = 0xFFFF;

  PM5CTL0 &= ~LOCKLPM5;

  float fa[sz][sz] = {
                          3.0451312,  -2.6170964,  -0.36237785, 0.67903620,  0.54625326,  -1.4386408,  1.1592991, -0.054826848,  -0.038200974,  -0.095277585,  -2.2473435,  3.3342791, -0.30391535, -1.0859760,  -0.34301928, 0.41210720,
                          -2.6170964,  5.0436597, -2.0817051,  -0.79343557, 0.15403587,  1.5043120, -2.2707827,  1.2366960, -0.014145052,  0.049386151, 1.7008784, -4.7957544,  3.3069613, 0.57205218,  -0.67790008, -0.34301928,
                          -0.36237785, -2.0817051,  4.9076262, -2.2785275,  -0.87535220, 0.58374476,  1.2477649, -2.2913878,  1.2353808, -0.0016905917, 0.38163254,  1.0874399, -4.5344720,  3.4830644, 0.57205218,  -1.0859760,
                          0.67903620,  -0.79343557, -2.2785275,  4.6777167, -2.3003640,  0.0097080851,  0.048614282, 1.2007601, -2.3131311,  1.1977108, -0.11479994, 0.63922209,  1.2034756, -4.5344720,  3.3069613, -0.30391535,
                          0.54625326,  0.15403587,  -0.87535220, -2.3003640,  4.9336734, -2.2747848,  -0.099792719,  0.039780665, 1.1884731, -2.3556819,  1.0750339, 0.21016976,  0.63922209,  1.0874399, -4.7957544,  3.3342791,
                          -1.4386408,  1.5043120, 0.58374476,  0.0097080851,  -2.2747848,  1.8861172, -0.22406493, 0.013938670, 0.098963775, 1.3059657, -2.6704209,  1.0750339, -0.11479994, 0.38163254,  1.7008784, -2.2473435,
                          1.1592991, -2.2707827,  1.2477649, 0.048614282, -0.099792719,  -0.22406493, 0.57083476,  -0.30296537, -0.026575994,  0.022921033, 1.3059657, -2.3556819,  1.1977108, -0.0016905917, 0.049386151, -0.095277585,
                          -0.054826848,  1.2366960, -2.2913878,  1.2007601, 0.039780665, 0.013938670, -0.30296537, 0.56730592,  -0.30373546, -0.026575994,  0.098963775, 1.1884731, -2.3131311,  1.2353808, -0.014145052,  -0.038200974,
                          -0.038200974,  -0.014145052,  1.2353808, -2.3131311,  1.1884731, 0.098963775, -0.026575994,  -0.30373546, 0.56730592,  -0.30296537, 0.013938670, 0.039780665, 1.2007601, -2.2913878,  1.2366960, -0.054826848,
                          -0.095277585,  0.049386151, -0.0016905917, 1.1977108, -2.3556819,  1.3059657, 0.022921033, -0.026575994,  -0.30296537, 0.57083476,  -0.22406493, -0.099792719,  0.048614282, 1.2477649, -2.2707827,  1.1592991,
                          -2.2473435,  1.7008784, 0.38163254,  -0.11479994, 1.0750339, -2.6704209,  1.3059657, 0.098963775, 0.013938670, -0.22406493, 1.8861172, -2.2747848,  0.0097080851,  0.58374476,  1.5043120, -1.4386408,
                          3.3342791, -4.7957544,  1.0874399, 0.63922209,  0.21016976,  1.0750339, -2.3556819,  1.1884731, 0.039780665, -0.099792719,  -2.2747848,  4.9336734, -2.3003640,  -0.87535220, 0.15403587,  0.54625326,
                          -0.30391535, 3.3069613, -4.5344720,  1.2034756, 0.63922209,  -0.11479994, 1.1977108, -2.3131311,  1.2007601, 0.048614282, 0.0097080851,  -2.3003640,  4.6777167, -2.2785275,  -0.79343557, 0.67903620,
                          -1.0859760,  0.57205218,  3.4830644, -4.5344720,  1.0874399, 0.38163254,  -0.0016905917, 1.2353808, -2.2913878,  1.2477649, 0.58374476,  -0.87535220, -2.2785275,  4.9076262, -2.0817051,  -0.36237785,
                          -0.34301928, -0.67790008, 0.57205218,  3.3069613, -4.7957544,  1.7008784, 0.049386151, -0.014145052,  1.2366960 -2.2707827,  1.5043120, 0.15403587,  -0.79343557, -2.0817051,  5.0436597, -2.6170964,
                          0.41210720,  -0.34301928, -1.0859760,  -0.30391535, 3.3342791, -2.2473435, -0.095277585, -0.038200974, -0.054826848, 1.1592991, -1.4386408,  0.54625326,  0.67903620,  -0.36237785, -2.6170964,  3.0451312};

  float fb[sz] = {2.3, 5.1, 1.05, 2.33, 3.56, 2.3, 5.1, 1.05, 2.33, 3.56, 2.3, 5.1, 1.05, 2.33, 3.56, 1.1};
  float fout[sz] = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
  int n;
  int m;

  while(1){
    for(n = lpc; n>=0; n--){
        for(m = lpc; m>=0; m--)
            fout[n] += fa[n][m]*fb[m];
    }
    __delay_cycles(8000000);
  }
}

此外、我的一些变量看起来会随着这两种实现而溢出。是否有关于如何避免这种情况的建议？

提前感谢、
Julio。

4 年多前

0 admin 4 年多前

TI__Guru**** 2539500 points

请注意，本文内容源自机器翻译，可能存在语法或其它翻译错误，仅供参考。如需获取准确内容，请参阅链接中的英语原文或自行翻译。

[引用 userid="433182" URL"~/support/microcontrollers/msp-low-power-microcontrollers-group/msp430/f/msp-low-power-microcontroller-forum/1035536/msp430fr5969-does-q-math-library-improve-energy-consumption-on-this-mcu "]此外，我的一些变量似乎会在这两种实现中溢出。

查看 main()中的代码，它在栈上使用了1088个字节来保存浮点变量，这可能会导致栈溢出。尝试以下任一操作：

使浮点变量数组成为全局变量
增加堆栈大小-在 CCS Project Properties under Build -> MSP430 Linker -> Basic Options -> Set C system stack size (--stack_size、-stack)

[引用 userid="433182" URL"~/support/microcontrollers/msp-low-power-microcontrollers-group/msp430/f/msp-low-power-microcontroller-forum/1035536/msp430fr5969-does-q-math-library-improve-energy-consumption-on-this-mcu "]我希望大家能给我一个手，因为我们可能缺少库上的配置。

使用 Qmath 库的代码仍使用浮点进行输入和输出、循环必须在浮点值和 Q 值之间进行转换。即、代码对浮点值和 Q 值之间的转换以及 Q 值的实际矢量矩阵乘法进行计时。是否可以更改代码以将转换移至 while (1)循环之外？

0 admin 4 年多前

TI__Guru**** 2539500 points

请注意，本文内容源自机器翻译，可能存在语法或其它翻译错误，仅供参考。如需获取准确内容，请参阅链接中的英语原文或自行翻译。

为了获得最佳效率、您最多只需跳过 Qmath 库的使用。这个实现的问题是、在每次乘法后、结果在被加至结果之前被移动4位。在总和完成后等待和执行移位会更好。

为了达到最佳效果、您需要直接使用硬件乘法器及其乘法累加功能。您可能需要查看 DSP 库。如果能耗非常重要、那么您可能需要具有低功耗加速器硬件的器件。

如果您要使用定点、请以该格式存储数据。浮点至定点转换似乎是库中的函数、但让预处理器为您执行此操作并不困难。

MSP 低功耗微控制器（参考译文帖）

MSP 低功耗微控制器（参考译文帖）(Read Only)

[参考译文] MSP430FR5969：Q-math 库是否改善了此 MCU 的能耗？