【STM32H7S78-DK】Evaluation + DSP calculation speed evaluation

dql2016 · Published on 2024-10-10 21:28

【STM32H7S78-DK】Evaluation + DSP calculation speed evaluation [Copy link]

This post was last edited by dql2016 on 2024-10-10 22:33

In the previous post, the DSP library was successfully added for mathematical calculations. This post adds the open source performance evaluation library perf_counter (

链接已隐藏，如需查看请登录或者注册

) of the silly kid to test the computing performance and measure the DSP performance by calculating the time.

perf_counter is a C-based module that provides the following functions:

Accurately measure system performance
Accurately measure function execution time
Accurately measure interrupt response latency
Provide blocking or non-blocking delay services accurate to the us level
Improving the random number properties of pseudo-random numbers
Provide system timestamp
……

It uses SysTick but does not occupy SysTick, and supports all Cortex-M processor bare metal environments and RTOS environments, as well as compiler platforms such as KEIL, IAR, and GCC. After downloading the source code package, you will get the following files:

For bare metal, just add the 4 files shown in the figure to the project:

First, add a reference to the header file in main.c:

#include"perf_counter.h"

Then initialize the perf_counter library:

init_cycle_counter(false);

or

init_cycle_counter(true);

If the user does not initialize SysTick himself , pass false to the function init_cycle_counter() .

If the user initializes SysTick himself , pass true to the function init_cycle_counter() .

The project produced by this project initializes SysTick, so the entry parameter is true

Call perfc_port_insert_to_system_timer_insert_ovf_handler() in the systick interrupt:

Block the printing of RMS calculation in the previous post

//DSP库求均方根测试
static void DSP_RMS(void)
{
    float32_t pSrc[10] = {0.7060f, 0.0318f, 0.2769f, 0.0462f, 0.0971f, 0.8235f, 0.6948f, 0.3171f,0.9502f, 0.0344f};
    float32_t pResult;
    uint32_t pIndex;

    q31_t pSrc1[10];
    q31_t pResult1;

    q15_t pSrc2[10];
    q15_t pResult2;

    //printf("******** stm32h7s78-dk eeworld dsp test ***********\r\n");

    arm_rms_f32(pSrc, 10, &pResult);
    //printf("arm_rms_f32 : pResult = %f\r\n", pResult);

    /*****************************************************************/
    for(pIndex = 0; pIndex < 10; pIndex++)
    {
         pSrc1[pIndex] = rand();
    }
    arm_rms_q31(pSrc1, 10, &pResult1);
    //printf("arm_rms_q31 : pResult = %d\r\n", pResult1);

    /*****************************************************************/
    for(pIndex = 0; pIndex < 10; pIndex++)
    {
         pSrc2[pIndex] = rand()%32768;
    }
    arm_rms_q15(pSrc2, 10, &pResult2);
    //printf("arm_rms_q15 : pResult = %d\r\n", pResult2);
    //printf("******************************************************************\r\n");
}

Measure the time it takes to call this function:

	start_cycle_counter();

	DSP_RMS();

	int64_t lCycleUsed = stop_cycle_counter();
	printf("cycle counter = %lld\n",lCycleUsed);

If printf needs to print 64-bit data of type int64 or uint64, it is necessary to enable support for the complete standard output library:

Compile and download the program directly, and the output calculation print result is 0. The answer is found in the STM32 official forum:

https://community.st.com/t5/stm32cubeide-mcus/wrong-result-when-printing-a-int64-t-value-using-stm32cubeide/td-p/148913

STM32CubeIDE uses a reduced library by default in order to reduce code size, as it's usually preferred for embedded development. Part of this tradeoff is dropping support for long long ints in printf-type functions.

> What can I do to print a 64-bit int value?

Change the runtime library from reduced to standard.

The previously configured standard output library was streamlined, resulting in incorrect output.

After modification, the print is correct:

The main frequency of stm32h7s7 is 600MHz, 1/600 000 000 (s) * 35630 (cycle) = 5.938*10^-5 (s) = 59.38us

It can be seen that the calculation speed is still very fast. A total of 10 data arm_rms_f32, arm_rms_q31, and arm_rms_q15 are calculated.

stm32cubeide_workspace_dsp_perf_counter.zip (17.42 MB, downloads: 0)