Qfplib和IQmath的对比测试

Qfplib的介绍页面

https://www.quinapalus.com/qfplib.html
该浮点库使用针对cortex-m3优化的汇编代码实现了qfp_fadd，qfp_fsub，qfp_fmul，qfp_fdiv等函数，用以替代编译器内置的软浮点实现，和大多数数学库一样，Qfplib也需要你将代码中的符号运算替换成函数调用。
下面使用cortex-m4内核的单片机进行测试，虽然介绍页面有如下阐述It will also run on Cortex-M4 microcontrollers but is not optimised for these devices.，但是m4能支持m3的所有指令，将qfplib-m3.s文件中的.cpu cortex-m3改成.cpu cortex-m4直接进行编译也是可以通过的（不改也行）。

测试平台

目标芯片：AT32F415 Cortex-M4(No FPU)，实际主频：144MHz，SRAM：32KB，FLASH等待4个周期。
编译器：arm gcc 13.2.Rel1，优化等级：O2
IQmath定点数使用iq15，Qfplib使用float

测试代码

#define CONST_PI_VAL 3.1415926f

#define VECTOR_SIZE  600

static _iq15 ResultVector[VECTOR_SIZE];
static _iq15 VectorA[VECTOR_SIZE];
static _iq15 VectorB[VECTOR_SIZE];

static float ResultVectorF[VECTOR_SIZE];
static float VectorAF[VECTOR_SIZE];
static float VectorBF[VECTOR_SIZE];

// IQ定点数 数组初始化
static void BenchmarkVectorIQArrayInit(void) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        VectorA[index] = _IQ15(1.0221f) * index;
        VectorB[index] = _IQ15(2.127f) * index;
        ResultVector[index] = 0;
    }
}

// 浮点数 数组初始化
static void BenchmarkVectorQfpArrayInit(void) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        VectorAF[index] = qfp_fmul(1.0221f, index);
        VectorBF[index] = qfp_fmul(2.127f, index);
        ResultVectorF[index] = 0.0f;
    }
}

// IQ定点数累加
static void VectorIQAdd(_iq15 *vectorA, _iq15 *vectorB, _iq15 *result) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        result[index] = vectorA[index]+ vectorB[index];
    }
}

// 浮点数累加
static void VectorQfpAdd(float *vectorA, float *vectorB, float *result) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        result[index] = qfp_fadd(vectorA[index], vectorB[index]);
    }
}

// IQ定点数乘法
static void VectorIQMultiply(_iq15 *vectorA, _iq15 *vectorB, _iq15 *result) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        result[index] = _IQ15mpy(vectorA[index], vectorB[index]);
    }
}

// 浮点数乘法
static void VectorQfpMultiply(float *vectorA, float *vectorB, float *result ) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++){
        result[index] = qfp_fmul(vectorA[index], vectorB[index]);
    }
}

// IQ定点数乘加
static void VectorIQScale(_iq15 *vectorA, _iq15 *vectorB, _iq15 *result) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        result[index] = _IQ15mpy(vectorA[index], _IQ15(CONST_PI_VAL));
        result[index] += vectorB[index];
    }
}

// 浮点数乘加
static void VectorQfpScale(float *vectorA, float *vectorB, float *result) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        result[index] = qfp_fmul(vectorA[index], CONST_PI_VAL);
        result[index] = qfp_fadd(result[index], vectorB[index]);
    }
}

// IQ定点数除法
static void VectorIQDiv(_iq15 *vectorA, _iq15 *vectorB, _iq15 *result) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        result[index] = _IQ15div(vectorA[index], vectorB[index]);
    }
}

// 浮点数除法
static void VectorQfpDiv(float *vectorA, float *vectorB, float *result) {
    unsigned int index = 0u;
    for(index = 0; index < VECTOR_SIZE; index++) {
        result[index] = qfp_fdiv(vectorA[index], vectorB[index]);
    }
}

int main(void) {
    uint32_t start, end, diff;

    BenchmarkVectorIQArrayInit();
    BenchmarkVectorQfpArrayInit();

    DWT->CYCCNT = 0x0;

    start = DWT->CYCCNT;
    VectorIQAdd(VectorA, VectorB, ResultVector);
    end = DWT->CYCCNT;
    diff = end - start;
    rt_kprintf("iq add, elapse:%d\n", diff);

    start = DWT->CYCCNT;
    VectorQfpAdd(VectorAF, VectorBF, ResultVectorF);
    end = DWT->CYCCNT;
    diff = end - start;
    rt_kprintf("qfp add, elapse:%d\n", diff);

    start = DWT->CYCCNT;
    VectorIQMultiply(VectorA, VectorB, ResultVector);
    end = DWT->CYCCNT;
    diff = end - start;
    rt_kprintf("iq mpy, elapse:%d\n", diff);

    start = DWT->CYCCNT;
    VectorQfpMultiply(VectorAF, VectorBF, ResultVectorF);
    end = DWT->CYCCNT;
    diff = end - start;
    rt_kprintf("qfp mpy, elapse:%d\n", diff);

    start = DWT->CYCCNT;
    VectorIQDiv(VectorA, VectorB, ResultVector);
    end = DWT->CYCCNT;
    diff = end - start;
    rt_kprintf("iq div, elapse:%d\n", diff);

    start = DWT->CYCCNT;
    VectorQfpDiv(VectorAF, VectorBF, ResultVectorF);
    end = DWT->CYCCNT;
    diff = end - start;
    rt_kprintf("qfp div, elapse:%d\n", diff);

    start = DWT->CYCCNT;
    VectorIQScale(VectorA, VectorB, ResultVector);
    end = DWT->CYCCNT;
    diff = end - start;
    rt_kprintf("iq scale, elapse:%d\n", diff);

    start = DWT->CYCCNT;
    VectorQfpScale(VectorAF, VectorBF, ResultVectorF);
    end = DWT->CYCCNT;
    diff = end - start;
    rt_kprintf("qfp scale, elapse:%d\n", diff);


    while(1) {
        osDelay(1000);
    }

    return 0;
}

原始日志

iq add, elapse:6023
qfp add, elapse:40991
iq mpy, elapse:11463
qfp mpy, elapse:32473
iq div, elapse:83931
qfp div, elapse:49770
iq scale, elapse:14447
qfp scale, elapse:71038

iq add, elapse:6023
qfp add, elapse:40991
iq mpy, elapse:11463
qfp mpy, elapse:32473
iq div, elapse:83931
qfp div, elapse:49770
iq scale, elapse:14447
qfp scale, elapse:71038

iq add, elapse:6023
qfp add, elapse:40991
iq mpy, elapse:11463
qfp mpy, elapse:32473
iq div, elapse:83931
qfp div, elapse:49770
iq scale, elapse:14447
qfp scale, elapse:71038

iq add, elapse:6023
qfp add, elapse:40991
iq mpy, elapse:11463
qfp mpy, elapse:32473
iq div, elapse:83931
qfp div, elapse:49770
iq scale, elapse:14447
qfp scale, elapse:71038

iq add, elapse:6023
qfp add, elapse:40991
iq mpy, elapse:11463
qfp mpy, elapse:32473
iq div, elapse:83931
qfp div, elapse:49770
iq scale, elapse:14447
qfp scale, elapse:71038

定点数加法和软浮点数加法

定点数乘法和软浮点数乘法

定点数除法和软浮点数除法

定点数除法和软浮点数乘加

总结

Qfp浮点库相比于IQ定点库性能提升