完善资料让更多小伙伴认识你,还能领取20积分哦, 立即完善>
转dsp系列教程
本期教程开始学习ARM官方的DSP库,这里我们先从基本数学函数开始。本期教程主要讲绝对值,加法,点乘和乘法四种运算。 8.1 绝对值(Vector Absolute Value) 8.2 求和(Vector Addition) 8.3 点乘(Vector Dot Product) 8.4 乘法(Vector Multiplication) 8.1 绝对值(Vector Absolute Value) 这部分函数主要用于求绝对值,公式描述如下: pDst[n] = abs(pSrc[n]), 0 <= n < blockSize. 特别注意,这部分函数支持目标指针和源指针指向相同的缓冲区。 8.1.1 arm_abs_f32 这个函数用于求32位浮点数的绝对值,源代码分析如下: [url=]复制代码[/url]
|
|
相关推荐
|
|
1. 在这里简单的跟大家介绍一下DSP库中函数的通用格式,后面就不再赘述了。
(1) 基本所有的函数都是可重入的。 (2) 大部分函数都支持一组数的计算,比如这个函数arm_abs_f32就可以计算一组数的绝对值。所以如果只是就几个数的绝对值,用这个库函数就没有什么优势了。 (3) 库函数基本是CM0,CM3和CM4都支持的(最新的DSP库已经添加CM7的支持)。 (4) 每组数据基本上都是以4个数为一个单位进行计算,不够四个再单独计算。 (5) 大部分函数都是配有f32,Q31,Q15和Q7四种格式。 2. 函数参数,支持输入一个数组进行计算绝对值。 3. 这部分代码是用于CM3和CM4内核。 4. 左移两位从而实现每4个数据为一组进行计算。 5. fabsf:这个函数不是用Cortex-M4F支持的DSP指令实现的,而是用C语言实现的,这个函数是被MDK封装起来的。 6. 切换到下一组数据。 7. 这部分代码用于CM0. 8. 用于不够4个数据的计算或者CM0内核。 8.1.2 arm_abs_q31 这个函数用于求32位定点数的绝对值,源代码分析如下: 复制代码 /** * @brief Q31 vector absolute value. * @param[in] *pSrc points to the input buffer * @param[out] *pDst points to the output buffer * @param[in] blockSize number of samples in each vector * @return none. * * Scaling and Overflow Behavior: (1) * par * The function uses saturating arithmetic. * The Q31 value -1 (0x80000000) will be saturated to the maximum allowable positive value 0x7FFFFFFF. */ void arm_abs_q31( q31_t * pSrc, q31_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counter */ q31_t in; /* Input value */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ q31_t in1, in2, in3, in4; /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = |A| */ /* Calculate absolute of input (if -1 then saturated to 0x7fffffff) and then store the results in the destination buffer. */ in1 = *pSrc++; in2 = *pSrc++; in3 = *pSrc++; in4 = *pSrc++; *pDst++ = (in1 > 0) ? in1 : (q31_t)__QSUB(0, in1); (2) *pDst++ = (in2 > 0) ? in2 : (q31_t)__QSUB(0, in2); *pDst++ = (in3 > 0) ? in3 : (q31_t)__QSUB(0, in3); *pDst++ = (in4 > 0) ? in4 : (q31_t)__QSUB(0, in4); /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; #endif /* #ifndef ARM_MATH_CM0_FAMILY */ while(blkCnt > 0u) { /* C = |A| */ /* Calculate absolute value of the input (if -1 then saturated to 0x7fffffff) and then store the results in the destination buffer. */ in = *pSrc++; *pDst++ = (in > 0) ? in : ((in == INT32_MIN) ? INT32_MAX : -in); /* Decrement the loop counter */ blkCnt--; } } 1. 这个函数使用了饱和运算,其实不光这个函数,后面很多函数都是使用了饱和运算的,关于什么是饱和运算,大家看Cortex-M3权威指南中文版的4.3.6 小节:汇编语言:饱和运算即可。 对于Q31格式的数据,饱和运算会使得数据0x80000000变成0x7fffffff(这个数比较特殊,算是特殊处理,记住即可)。 2. 这里重点说一下函数__QSUB,其实这个函数算是Cortex-M4/M3的一个指令,用于实现饱和减法。 比如函数:__QSUB(0, in1) 的作用就是实现0 – in1并返回结果。这里__QSUB实现的是32位数的饱和减法。还有__QSUB16和__QSUB8实现的是16位和8位数的减法。 |
|
|
|
|
|
8.1.3 arm_abs_q15
这个函数用于求15位定点数的绝对值,源代码分析如下: 复制代码 /** * @brief Q15 vector absolute value. * @param[in] *pSrc points to the input buffer * @param[out] *pDst points to the output buffer * @param[in] blockSize number of samples in each vector * @return none. * * Scaling and Overflow Behavior: * par * The function uses saturating arithmetic. * The Q15 value -1 (0x8000) will be saturated to the maximum allowable positive value 0x7FFF. (1) */ void arm_abs_q15( q15_t * pSrc, q15_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counter */ #ifndef ARM_MATH_CM0_FAMILY __SIMD32_TYPE *simd; (2) /* Run the below code for Cortex-M4 and Cortex-M3 */ q15_t in1; /* Input value1 */ q15_t in2; /* Input value2 */ /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ simd = __SIMD32_CONST(pDst); (3) while(blkCnt > 0u) { /* C = |A| */ /* Read two inputs */ in1 = *pSrc++; in2 = *pSrc++; /* Store the Absolute result in the destination buffer by packing the two values, in a single cycle */ #ifndef ARM_MATH_BIG_ENDIAN *simd++ = __PKHBT(((in1 > 0) ? in1 : (q15_t)__QSUB16(0, in1)), (4) ((in2 > 0) ? in2 : (q15_t)__QSUB16(0, in2)), 16); #else *simd++ = __PKHBT(((in2 > 0) ? in2 : (q15_t)__QSUB16(0, in2)), ((in1 > 0) ? in1 : (q15_t)__QSUB16(0, in1)), 16); #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ in1 = *pSrc++; in2 = *pSrc++; #ifndef ARM_MATH_BIG_ENDIAN *simd++ = __PKHBT(((in1 > 0) ? in1 : (q15_t)__QSUB16(0, in1)), ((in2 > 0) ? in2 : (q15_t)__QSUB16(0, in2)), 16); #else *simd++ = __PKHBT(((in2 > 0) ? in2 : (q15_t)__QSUB16(0, in2)), ((in1 > 0) ? in1 : (q15_t)__QSUB16(0, in1)), 16); #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ /* Decrement the loop counter */ blkCnt--; } pDst = (q15_t *)simd; /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; while(blkCnt > 0u) { /* C = |A| */ /* Read the input */ in1 = *pSrc++; /* Calculate absolute value of input and then store the result in the destination buffer. */ *pDst++ = (in1 > 0) ? in1 : (q15_t)__QSUB16(0, in1); /* Decrement the loop counter */ blkCnt--; } #else /* Run the below code for Cortex-M0 */ q15_t in; /* Temporary input variable */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; while(blkCnt > 0u) { /* C = |A| */ /* Read the input */ in = *pSrc++; /* Calculate absolute value of input and then store the result in the destination buffer. */ *pDst++ = (in > 0) ? in : ((in == (q15_t) 0x8000) ? 0x7fff : -in); /* Decrement the loop counter */ blkCnt--; } #endif /* #ifndef ARM_MATH_CM0_FAMILY */ } 1. 对于Q15格式的数据,饱和运算会使得数据0x8000变成0x7fff。 2. __SIMD32_TYPE的定义在文件arm_math.h中,具体定义如下: #define __SIMD32_TYPE int32_t __packed SIMD就是咱们上期教程所将的单指令多数据流。简单的理解就是__SIMD32_TYPE就是定义了一个int32_t类型的数据,__packed的含义就是实现字节的对齐功能,方便两个16位数据的都存入到这个数据类型中。 3. 函数__SIMD32_CONST的定义如下: #define __SIMD32_CONST(addr) ((__SIMD32_TYPE *)(addr)) 4. 函数__PKHBT的定义在文件core_cm4_simd.h,定义如下: #define __PKHBT(ARG1,ARG2,ARG3) ( ((((uint32_t)(ARG1)) ) & 0x0000FFFFUL) | ((((uint32_t)(ARG2)) << (ARG3)) & 0xFFFF0000UL) ) 这个宏定义的作用就是将将两个16位的数据合并成32位数据。但是有一点要特别说明__PKHBT也是CM4内核支持的SIMD指令,上面的宏定义的C函数会被MDK自动识别并调用相应的PKHBT指令。__QSUB16用于实现16位数据的饱和减法。 |
|
|
|
|
|
8.1.4 arm_abs_q7
这个函数用于求8位定点数的绝对值,源代码分析如下: 复制代码 /** * @brief Q7 vector absolute value. * @param[in] *pSrc points to the input buffer * @param[out] *pDst points to the output buffer * @param[in] blockSize number of samples in each vector * @return none. * * par Conditions for optimum performance * Input and output buffers should be aligned by 32-bit * * * Scaling and Overflow Behavior: (1) * par * The function uses saturating arithmetic. * The Q7 value -1 (0x80) will be saturated to the maximum allowable positive value 0x7F. */ void arm_abs_q7( q7_t * pSrc, q7_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counter */ q7_t in; /* Input value1 */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ q31_t in1, in2, in3, in4; /* temporary input variables */ q31_t out1, out2, out3, out4; /* temporary output variables */ /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = |A| */ /* Read inputs */ in1 = (q31_t) * pSrc; in2 = (q31_t) * (pSrc + 1); in3 = (q31_t) * (pSrc + 2); /* find absolute value */ out1 = (in1 > 0) ? in1 : (q31_t)__QSUB8(0, in1); (2) /* read input */ in4 = (q31_t) * (pSrc + 3); /* find absolute value */ out2 = (in2 > 0) ? in2 : (q31_t)__QSUB8(0, in2); /* store result to destination */ *pDst = (q7_t) out1; /* find absolute value */ out3 = (in3 > 0) ? in3 : (q31_t)__QSUB8(0, in3); /* find absolute value */ out4 = (in4 > 0) ? in4 : (q31_t)__QSUB8(0, in4); /* store result to destination */ *(pDst + 1) = (q7_t) out2; /* store result to destination */ *(pDst + 2) = (q7_t) out3; /* store result to destination */ *(pDst + 3) = (q7_t) out4; /* update pointers to process next samples */ pSrc += 4u; pDst += 4u; /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; #else /* Run the below code for Cortex-M0 */ blkCnt = blockSize; #endif // #define ARM_MATH_CM0_FAMILY while(blkCnt > 0u) { /* C = |A| */ /* Read the input */ in = *pSrc++; /* Store the Absolute result in the destination buffer */ *pDst++ = (in > 0) ? in : ((in == (q7_t) 0x80) ? 0x7f : -in); /* Decrement the loop counter */ blkCnt--; } } 1. 由于饱和运算,0x80求绝对值将变成数据0x7F。 2. __QSUB8用以实现8位数的饱和减法运算。 |
|
|
|
|
|
8.1.5 实例讲解
实验目的: 1. 四种数据类型数据绝对值求解 实验内容: 1. 按下按键K1, 串口打印输出结果 实验现象: 通过窗口上位机软件SecureCRT(V5光盘里面有此软件)查看打印信息现象如下: 程序设计: [url=]复制代码[/url]
(1)到(4)实现相应格式下绝对值的求解。这里只求了一个数,大家可以尝试求解一个数组的绝对值。 |
|
|
|
|
|
8.2 求和(Vector Addition)
这部分函数主要用于求和,公式描述如下: pDst[n] = pSrcA[n] + pSrcB[n], 0 <= n < blockSize. 8.2.1 arm_add_f32 这个函数用于求32位浮点数的和,源代码分析如下: 复制代码 /** * @brief Floating-point vector addition. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[out] *pDst points to the output vector * @param[in] blockSize number of samples in each vector * @return none. */ void arm_add_f32( float32_t * pSrcA, float32_t * pSrcB, float32_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counter */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ float32_t inA1, inA2, inA3, inA4; /* temporary input variabels */ float32_t inB1, inB2, inB3, inB4; /* temporary input variables */ /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ /* read four inputs from sourceA and four inputs from sourceB */ inA1 = *pSrcA; inB1 = *pSrcB; inA2 = *(pSrcA + 1); inB2 = *(pSrcB + 1); inA3 = *(pSrcA + 2); inB3 = *(pSrcB + 2); inA4 = *(pSrcA + 3); inB4 = *(pSrcB + 3); /* C = A + B */ (1) /* add and store result to destination */ *pDst = inA1 + inB1; *(pDst + 1) = inA2 + inB2; *(pDst + 2) = inA3 + inB3; *(pDst + 3) = inA4 + inB4; /* update pointers to process next samples */ pSrcA += 4u; pSrcB += 4u; pDst += 4u; /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; #endif /* #ifndef ARM_MATH_CM0_FAMILY */ while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ *pDst++ = (*pSrcA++) + (*pSrcB++); /* Decrement the loop counter */ blkCnt--; } } 1. 这部分的代码比较简单,只是求解两个数的和。 |
|
|
|
|
|
8.2.2 arm_add_q31
这个函数用于求32位定点数的和,源代码分析如下: 复制代码 /** * @brief Q31 vector addition. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[out] *pDst points to the output vector * @param[in] blockSize number of samples in each vector * @return none. * * Scaling and Overflow Behavior: (1) * par * The function uses saturating arithmetic. * Results outside of the allowable Q31 range[0x80000000 0x7FFFFFFF] will be saturated. */ void arm_add_q31( q31_t * pSrcA, q31_t * pSrcB, q31_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counter */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ q31_t inA1, inA2, inA3, inA4; q31_t inB1, inB2, inB3, inB4; /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ inA1 = *pSrcA++; inA2 = *pSrcA++; inB1 = *pSrcB++; inB2 = *pSrcB++; inA3 = *pSrcA++; inA4 = *pSrcA++; inB3 = *pSrcB++; inB4 = *pSrcB++; *pDst++ = __QADD(inA1, inB1); (2) *pDst++ = __QADD(inA2, inB2); *pDst++ = __QADD(inA3, inB3); *pDst++ = __QADD(inA4, inB4); /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ *pDst++ = __QADD(*pSrcA++, *pSrcB++); /* Decrement the loop counter */ blkCnt--; } #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ *pDst++ = (q31_t) clip_q63_to_q31((q63_t) * pSrcA++ + *pSrcB++); (3) /* Decrement the loop counter */ blkCnt--; } #endif /* #ifndef ARM_MATH_CM0_FAMILY */ } 1. 这个函数也是饱和运算,输出结果的范围[0x80000000 0x7FFFFFFF],超出这个结果将产生饱和结果。 2. __QADD实现32位数的加法。 3. 函数clip_q63_to_q31的定义在文件arm_math.h里面 static __INLINE q31_t clip_q63_to_q31( q63_t x) { return ((q31_t) (x >> 32) != ((q31_t) x >> 31)) ? ((0x7FFFFFFF ^ ((q31_t) (x >> 63)))) : (q31_t) x; } 这个函数的作用是实现饱和结果。 |
|
|
|
|
|
8.2.3 arm_add_q15
这个函数用于求16位定点数的和,源代码分析如下: 复制代码 /** * @brief Q15 vector addition. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[out] *pDst points to the output vector * @param[in] blockSize number of samples in each vector * @return none. * * Scaling and Overflow Behavior: (1) * par * The function uses saturating arithmetic. * Results outside of the allowable Q15 range [0x8000 0x7FFF] will be saturated. */ void arm_add_q15( q15_t * pSrcA, q15_t * pSrcB, q15_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counter */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ q31_t inA1, inA2, inB1, inB2; /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A + B */ (2) /* Add and then store the results in the destination buffer. */ inA1 = *__SIMD32(pSrcA)++; inA2 = *__SIMD32(pSrcA)++; inB1 = *__SIMD32(pSrcB)++; inB2 = *__SIMD32(pSrcB)++; *__SIMD32(pDst)++ = __QADD16(inA1, inB1); *__SIMD32(pDst)++ = __QADD16(inA2, inB2); /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ *pDst++ = (q15_t) __QADD16(*pSrcA++, *pSrcB++); /* Decrement the loop counter */ blkCnt--; } #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ *pDst++ = (q15_t) __SSAT(((q31_t) * pSrcA++ + *pSrcB++), 16); (3) /* Decrement the loop counter */ blkCnt--; } #endif /* #ifndef ARM_MATH_CM0_FAMILY */ } 1. 这个函数也是饱和运算,输出结果的范围[0x8000 0x7FFF],超出这个结果将产生饱和结果。 2. 函数inA1 = *__SIMD32(pSrcA)++仅需要一条SIMD指令即可完成将两个16位数存到32位的变量inA1中。 3. __SSAT也是SIMD指令,这里是将结果饱和到16位精度。 |
|
|
|
|
|
8.2.4 arm_add_q7
这个函数用于求8位定点数的绝对值,源代码分析如下: 复制代码 /** * @brief Q7 vector addition. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[out] *pDst points to the output vector * @param[in] blockSize number of samples in each vector * @return none. * * Scaling and Overflow Behavior: (1) * par * The function uses saturating arithmetic. * Results outside of the allowable Q7 range [0x80 0x7F] will be saturated. */ void arm_add_q7( q7_t * pSrcA, q7_t * pSrcB, q7_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counter */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ (2) *__SIMD32(pDst)++ = __QADD8(*__SIMD32(pSrcA)++, *__SIMD32(pSrcB)++); /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ *pDst++ = (q7_t) __SSAT(*pSrcA++ + *pSrcB++, 8); /* Decrement the loop counter */ blkCnt--; } #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; while(blkCnt > 0u) { /* C = A + B */ /* Add and then store the results in the destination buffer. */ *pDst++ = (q7_t) __SSAT((q15_t) * pSrcA++ + *pSrcB++, 8); /* Decrement the loop counter */ blkCnt--; } #endif /* #ifndef ARM_MATH_CM0_FAMILY */ } 1. 这个函数也是饱和运算,输出结果的范围[0x80 0x7F],超出这个结果将产生饱和。 2. 这里通过SIMD指令实现4组8位数的加法。 |
|
|
|
|
|
8.2.5 实例讲解
实验目的: 1. 四种类似数据的求和 实验内容: 1. 按下按键K2, 串口打印输出结果 实验现象: 通过窗口上位机软件SecureCRT(V5光盘里面有此软件)查看打印信息现象如下: |
|
|
|
|
|
程序设计:
复制代码 /* ********************************************************************************************************* * 函 数 名: DSP_ABS * 功能说明: 加法 * 形 参:无 * 返 回 值: 无 ********************************************************************************************************* */ static void DSP_Add(void) { static float32_t pSrcA; static float32_t pSrcB; static float32_t pDst; static q31_t pSrcA1; static q31_t pSrcB1; static q31_t pDst1; static q15_t pSrcA2; static q15_t pSrcB2; static q15_t pDst2; static q7_t pSrcA3; static q7_t pSrcB3; static q7_t pDst3; pSrcA--; arm_add_f32(&pSrcA, &pSrcB, &pDst, 1); printf("arm_add_f32 = %frn", pDst); pSrcA1--; arm_add_q31(&pSrcA1, &pSrcB1, &pDst1, 1); printf("arm_add_q31 = %drn", pDst1); pSrcA2--; arm_add_q15(&pSrcA2, &pSrcB2, &pDst2, 1); printf("arm_add_q15 = %drn", pDst2); pSrcA3--; arm_add_q7(&pSrcA3, &pSrcB3, &pDst3, 1); printf("arm_add_q7 = %drn", pDst3); printf("***********************************rn"); } |
|
|
|
|
|
8.3 点乘(Vector Dot Product)
这部分函数主要用于点乘,公式描述如下: sum = pSrcA[0]*pSrcB[0] + pSrcA[1]*pSrcB[1] + ... + pSrcA[blockSize-1]*pSrcB[blockSize-1] 8.3.1 arm_dot_prod_f32 这个函数用于求32位浮点数的点乘,源代码分析如下: 复制代码 /** * @defgroup dot_prod Vector Dot Product * * Computes the dot product of two vectors. * The vectors are multiplied element-by-element and then summed. * *
* * There are separate functions for floating-point, Q7, Q15, and Q31 data types. */ /** * @addtogroup dot_prod * @{ */ /** * @brief Dot product of floating-point vectors. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[in] blockSize number of samples in each vector * @param[out] *result output result returned here * @return none. */ void arm_dot_prod_f32( float32_t * pSrcA, float32_t * pSrcB, uint32_t blockSize, float32_t * result) { float32_t sum = 0.0f; /* Temporary result storage */ (1) uint32_t blkCnt; /* loop counter */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ /* Calculate dot product and then store the result in a temporary buffer */ sum += (*pSrcA++) * (*pSrcB++); (2) sum += (*pSrcA++) * (*pSrcB++); sum += (*pSrcA++) * (*pSrcB++); sum += (*pSrcA++) * (*pSrcB++); /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; #endif /* #ifndef ARM_MATH_CM0_FAMILY */ while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ /* Calculate dot product and then store the result in a temporary buffer. */ sum += (*pSrcA++) * (*pSrcB++); /* Decrement the loop counter */ blkCnt--; } /* Store the result back in the destination buffer */ *result = sum; } 1. 由于CM4上带的FPU是单精度的,所以初始化float32_t类型的浮点数时需要在数据的末尾加上f。 2. 类似函数sum += (*pSrcA++) * (*pSrcB++)最终会通过浮点的MAC(乘累加)实现,从而加快执行时间。 |
|
|
|
|
|
8.3.2 arm_dot_prod_q31
这个函数用于求32位定点数的点乘,源代码分析如下: 复制代码 /** * @brief Dot product of Q31 vectors. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[in] blockSize number of samples in each vector * @param[out] *result output result returned here * @return none. * * Scaling and Overflow Behavior: (1) * par * The intermediate multiplications are in 1.31 x 1.31 = 2.62 format and these * are truncated to 2.48 format by discarding the lower 14 bits. * The 2.48 result is then added without saturation to a 64-bit accumulator in 16.48 format. * There are 15 guard bits in the accumulator and there is no risk of overflow as long as * the length of the vectors is less than 2^16 elements. * The return result is in 16.48 format. */ void arm_dot_prod_q31( q31_t * pSrcA, q31_t * pSrcB, uint32_t blockSize, q63_t * result) { q63_t sum = 0; /* Temporary result storage */ uint32_t blkCnt; /* loop counter */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ q31_t inA1, inA2, inA3, inA4; q31_t inB1, inB2, inB3, inB4; /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ /* Calculate dot product and then store the result in a temporary buffer. */ inA1 = *pSrcA++; inA2 = *pSrcA++; inA3 = *pSrcA++; inA4 = *pSrcA++; inB1 = *pSrcB++; inB2 = *pSrcB++; inB3 = *pSrcB++; inB4 = *pSrcB++; sum += ((q63_t) inA1 * inB1) >> 14u; (2) sum += ((q63_t) inA2 * inB2) >> 14u; sum += ((q63_t) inA3 * inB3) >> 14u; sum += ((q63_t) inA4 * inB4) >> 14u; /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; #endif /* #ifndef ARM_MATH_CM0_FAMILY */ while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ /* Calculate dot product and then store the result in a temporary buffer. */ sum += ((q63_t) * pSrcA++ * *pSrcB++) >> 14u; /* Decrement the loop counter */ blkCnt--; } /* Store the result in the destination buffer in 16.48 format */ *result = sum; } 1. 两个Q31格式的32位数相乘,那么输出结果的格式是1.31*1.31 = 2.62。实际应用中基本不需要这么高的精度,这个函数将低14位的数据截取掉,反应在函数中就是两个数的乘积左移14位,也就是定点数的小数点也左移14位,那么最终的结果的格式是16.48。所以只要乘累加的个数小于2^16就没有输出结果溢出的危险(不知道这里为什么不是2^14,留作以后解决)。 2. 将获取的结果左移14位。 |
|
|
|
|
|
8.3.3 arm_dot_prod_q15
这个函数用于求16位定点数的点乘,源代码分析如下: 复制代码 /** * @brief Dot product of Q15 vectors. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[in] blockSize number of samples in each vector * @param[out] *result output result returned here * @return none. * * Scaling and Overflow Behavior: (1) * par * The intermediate multiplications are in 1.15 x 1.15 = 2.30 format and these * results are added to a 64-bit accumulator in 34.30 format. * Nonsaturating additions are used and given that there are 33 guard bits in the accumulator * there is no risk of overflow. * The return result is in 34.30 format. */ void arm_dot_prod_q15( q15_t * pSrcA, q15_t * pSrcB, uint32_t blockSize, q63_t * result) { q63_t sum = 0; /* Temporary result storage */ uint32_t blkCnt; /* loop counter */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ (2) /* Calculate dot product and then store the result in a temporary buffer. */ sum = __SMLALD(*__SIMD32(pSrcA)++, *__SIMD32(pSrcB)++, sum); sum = __SMLALD(*__SIMD32(pSrcA)++, *__SIMD32(pSrcB)++, sum); /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ /* Calculate dot product and then store the results in a temporary buffer. */ sum = __SMLALD(*pSrcA++, *pSrcB++, sum); /* Decrement the loop counter */ blkCnt--; } #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ /* Calculate dot product and then store the results in a temporary buffer. */ sum += (q63_t) ((q31_t) * pSrcA++ * *pSrcB++); /* Decrement the loop counter */ blkCnt--; } #endif /* #ifndef ARM_MATH_CM0_FAMILY */ /* Store the result in the destination buffer in 34.30 format */ *result = sum; } 1. 两个Q15格式的数据相乘,那么输出结果的格式是1.15*1.15 = 2.30,这个函数将输出结果赋值给了64位变量,那么输出结果就是34.30格式。所以基本没有溢出的危险。 2. __SMLALD也是SIMD指令,实现两个16位数相乘,并把结果累加给64位变量。 |
|
|
|
|
|
8.3.4 arm_dot_prod_q7
这个函数用于求8位定点数的点乘,源代码分析如下: 复制代码 /** * @brief Dot product of Q7 vectors. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[in] blockSize number of samples in each vector * @param[out] *result output result returned here * @return none. * * Scaling and Overflow Behavior: (1) * par * The intermediate multiplications are in 1.7 x 1.7 = 2.14 format and these * results are added to an accumulator in 18.14 format. * Nonsaturating additions are used and there is no danger of wrap around as long as * the vectors are less than 2^18 elements long. * The return result is in 18.14 format. */ void arm_dot_prod_q7( q7_t * pSrcA, q7_t * pSrcB, uint32_t blockSize, q31_t * result) { uint32_t blkCnt; /* loop counter */ q31_t sum = 0; /* Temporary variables to store output */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ q31_t input1, input2; /* Temporary variables to store input */ q31_t inA1, inA2, inB1, inB2; /* Temporary variables to store input */ /*loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* read 4 samples at a time from sourceA */ (2) input1 = *__SIMD32(pSrcA)++; /* read 4 samples at a time from sourceB */ input2 = *__SIMD32(pSrcB)++; /* extract two q7_t samples to q15_t samples */ inA1 = __SXTB16(__ROR(input1, 8)); (3) /* extract reminaing two samples */ inA2 = __SXTB16(input1); /* extract two q7_t samples to q15_t samples */ inB1 = __SXTB16(__ROR(input2, 8)); /* extract reminaing two samples */ inB2 = __SXTB16(input2); /* multiply and accumulate two samples at a time */ sum = __SMLAD(inA1, inB1, sum); (4) sum = __SMLAD(inA2, inB2, sum); /* Decrement the loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ /* Dot product and then store the results in a temporary buffer. */ sum = __SMLAD(*pSrcA++, *pSrcB++, sum); /* Decrement the loop counter */ blkCnt--; } #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; while(blkCnt > 0u) { /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */ /* Dot product and then store the results in a temporary buffer. */ sum += (q31_t) ((q15_t) * pSrcA++ * *pSrcB++); /* Decrement the loop counter */ blkCnt--; } #endif /* #ifndef ARM_MATH_CM0_FAMILY */ /* Store the result in the destination buffer in 18.14 format */ *result = sum; } 1. 两个Q8格式的数据相乘,那么输出结果就是1.7*1.7 = 2.14格式。这里将最终结果赋值给了32位的变量,那么最终的格式就是18.14。如果乘累加的个数小于2^18那么就不会有溢出的危险(感觉这里应该是2^16)。 2. 一次读取4个8位的数据。 3. __SXTB16也是SIMD指令,用于将两个8位的有符号数扩展成16位。__ROR用于实现数据的循环右移。 4. __SMLAD也是SIMD指令,用于实现如下功能: sum = __SMLAD(x, y, z) sum = z + ((short)(x>>16) * (short)(y>>16)) + ((short)x * (short)y) |
|
|
|
|
|
8.3.5 实例讲解
实验目的: 1. 四种类型数据的点乘。 实验内容: 1. 按下按键K3, 串口打印输出结果 实验现象: 通过窗口上位机软件SecureCRT(V5光盘里面有此软件)查看打印信息现象如下: |
|
|
|
|
|
程序设计:
复制代码 /* ********************************************************************************************************* * 函 数 名: DSP_DotProduct * 功能说明: 乘积 * 形 参:无 * 返 回 值: 无 ********************************************************************************************************* */ static void DSP_DotProduct(void) { static float32_t pSrcA[5] = {1.0f,1.0f,1.0f,1.0f,1.0f}; static float32_t pSrcB[5] = {1.0f,1.0f,1.0f,1.0f,1.0f}; static float32_t result; static q31_t pSrcA1[5] = {0x7ffffff0,1,1,1,1}; static q31_t pSrcB1[5] = {1,1,1,1,1}; static q63_t result1; static q15_t pSrcA2[5] = {1,1,1,1,1}; static q15_t pSrcB2[5] = {1,1,1,1,1}; static q63_t result2; static q7_t pSrcA3[5] = {1,1,1,1,1}; static q7_t pSrcB3[5] = {1,1,1,1,1}; static q31_t result3; pSrcA[0] -= 1.1f; arm_dot_prod_f32(pSrcA, pSrcB, 5, &result); printf("arm_dot_prod_f32 = %frn", result); pSrcA1[0] -= 0xffff; arm_dot_prod_q31(pSrcA1, pSrcB1, 5, &result1); printf("arm_dot_prod_q31 = %lldrn", result1); pSrcA2[0] -= 1; arm_dot_prod_q15(pSrcA2, pSrcB2, 5, &result2); printf("arm_dot_prod_q15 = %lldrn", result2); pSrcA3[0] -= 1; arm_dot_prod_q7(pSrcA3, pSrcB3, 5, &result3); printf("arm_dot_prod_q7 = %drn", result3); printf("***********************************rn"); } |
|
|
|
|
|
8.4 乘法(Vector Multiplication)
这部分函数主要用于乘法,公式描述如下: pDst[n] = pSrcA[n] * pSrcB[n], 0 <= n < blockSize. 8.4.1 arm_mult_f32 这个函数用于求32位浮点数的乘法,源代码分析如下: 复制代码 /** * @brief Floating-point vector multiplication. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[out] *pDst points to the output vector * @param[in] blockSize number of samples in each vector * @return none. */ void arm_mult_f32( float32_t * pSrcA, float32_t * pSrcB, float32_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counters */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ float32_t inA1, inA2, inA3, inA4; /* temporary input variables */ float32_t inB1, inB2, inB3, inB4; /* temporary input variables */ float32_t out1, out2, out3, out4; /* temporary output variables */ /* loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A * B */ /* Multiply the inputs and store the results in output buffer */ (1) /* read sample from sourceA */ inA1 = *pSrcA; /* read sample from sourceB */ inB1 = *pSrcB; /* read sample from sourceA */ inA2 = *(pSrcA + 1); /* read sample from sourceB */ inB2 = *(pSrcB + 1); /* out = sourceA * sourceB */ out1 = inA1 * inB1; /* read sample from sourceA */ inA3 = *(pSrcA + 2); /* read sample from sourceB */ inB3 = *(pSrcB + 2); /* out = sourceA * sourceB */ out2 = inA2 * inB2; /* read sample from sourceA */ inA4 = *(pSrcA + 3); /* store result to destination buffer */ *pDst = out1; /* read sample from sourceB */ inB4 = *(pSrcB + 3); /* out = sourceA * sourceB */ out3 = inA3 * inB3; /* store result to destination buffer */ *(pDst + 1) = out2; /* out = sourceA * sourceB */ out4 = inA4 * inB4; /* store result to destination buffer */ *(pDst + 2) = out3; /* store result to destination buffer */ *(pDst + 3) = out4; /* update pointers to process next samples */ pSrcA += 4u; pSrcB += 4u; pDst += 4u; /* Decrement the blockSize loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; #endif /* #ifndef ARM_MATH_CM0_FAMILY */ while(blkCnt > 0u) { /* C = A * B */ /* Multiply the inputs and store the results in output buffer */ *pDst++ = (*pSrcA++) * (*pSrcB++); /* Decrement the blockSize loop counter */ blkCnt--; } } 1. 浮点的32位乘法比较简单,这里依然是以4次的计算为一组。 |
|
|
|
|
|
8.4.2 arm_mult_q31
这个函数用于求32位定点数的乘法,源代码分析如下: 复制代码 /** * @brief Q31 vector multiplication. * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[out] *pDst points to the output vector * @param[in] blockSize number of samples in each vector * @return none. * * Scaling and Overflow Behavior: (1) * par * The function uses saturating arithmetic. * Results outside of the allowable Q31 range[0x80000000 0x7FFFFFFF] will be saturated. */ void arm_mult_q31( q31_t * pSrcA, q31_t * pSrcB, q31_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counters */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ q31_t inA1, inA2, inA3, inA4; /* temporary input variables */ q31_t inB1, inB2, inB3, inB4; /* temporary input variables */ q31_t out1, out2, out3, out4; /* temporary output variables */ /* loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* C = A * B */ /* Multiply the inputs and then store the results in the destination buffer. */ inA1 = *pSrcA++; inA2 = *pSrcA++; inA3 = *pSrcA++; inA4 = *pSrcA++; inB1 = *pSrcB++; inB2 = *pSrcB++; inB3 = *pSrcB++; inB4 = *pSrcB++; out1 = ((q63_t) inA1 * inB1) >> 32; (2) out2 = ((q63_t) inA2 * inB2) >> 32; out3 = ((q63_t) inA3 * inB3) >> 32; out4 = ((q63_t) inA4 * inB4) >> 32; out1 = __SSAT(out1, 31); (3) out2 = __SSAT(out2, 31); out3 = __SSAT(out3, 31); out4 = __SSAT(out4, 31); *pDst++ = out1 << 1u; (4) *pDst++ = out2 << 1u; *pDst++ = out3 << 1u; *pDst++ = out4 << 1u; /* Decrement the blockSize loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; #endif /* #ifndef ARM_MATH_CM0_FAMILY */ while(blkCnt > 0u) { /* C = A * B */ /* Multiply the inputs and then store the results in the destination buffer. */ *pDst++ = (q31_t) clip_q63_to_q31(((q63_t) (*pSrcA++) * (*pSrcB++)) >> 31); /* Decrement the blockSize loop counter */ blkCnt--; } } 1. 这个函数使用了饱和算法。 所得结果是Q31格式,范围Q31 range[0x80000000 0x7FFFFFFF]。 2. 所得乘积左移32位。 3. 实现31位精度的饱和运算。 4. 右移一位,保证所得结果是Q31格式。 |
|
|
|
|
|
8.4.3 arm_mult_q15
这个函数用于求16位定点数的乘法,源代码分析如下: 复制代码 /** * @brief Q15 vector multiplication * @param[in] *pSrcA points to the first input vector * @param[in] *pSrcB points to the second input vector * @param[out] *pDst points to the output vector * @param[in] blockSize number of samples in each vector * @return none. * * Scaling and Overflow Behavior: (1) * par * The function uses saturating arithmetic. * Results outside of the allowable Q15 range [0x8000 0x7FFF] will be saturated. */ void arm_mult_q15( q15_t * pSrcA, q15_t * pSrcB, q15_t * pDst, uint32_t blockSize) { uint32_t blkCnt; /* loop counters */ #ifndef ARM_MATH_CM0_FAMILY /* Run the below code for Cortex-M4 and Cortex-M3 */ q31_t inA1, inA2, inB1, inB2; /* temporary input variables */ q15_t out1, out2, out3, out4; /* temporary output variables */ q31_t mul1, mul2, mul3, mul4; /* temporary variables */ /* loop Unrolling */ blkCnt = blockSize >> 2u; /* First part of the processing with loop unrolling. Compute 4 outputs at a time. ** a second loop below computes the remaining 1 to 3 samples. */ while(blkCnt > 0u) { /* read two samples at a time from sourceA */ inA1 = *__SIMD32(pSrcA)++; (2) /* read two samples at a time from sourceB */ inB1 = *__SIMD32(pSrcB)++; /* read two samples at a time from sourceA */ inA2 = *__SIMD32(pSrcA)++; /* read two samples at a time from sourceB */ inB2 = *__SIMD32(pSrcB)++; /* multiply mul = sourceA * sourceB */ mul1 = (q31_t) ((q15_t) (inA1 >> 16) * (q15_t) (inB1 >> 16)); (3) mul2 = (q31_t) ((q15_t) inA1 * (q15_t) inB1); mul3 = (q31_t) ((q15_t) (inA2 >> 16) * (q15_t) (inB2 >> 16)); mul4 = (q31_t) ((q15_t) inA2 * (q15_t) inB2); /* saturate result to 16 bit */ out1 = (q15_t) __SSAT(mul1 >> 15, 16); (4) out2 = (q15_t) __SSAT(mul2 >> 15, 16); out3 = (q15_t) __SSAT(mul3 >> 15, 16); out4 = (q15_t) __SSAT(mul4 >> 15, 16); /* store the result */ #ifndef ARM_MATH_BIG_ENDIAN *__SIMD32(pDst)++ = __PKHBT(out2, out1, 16); (5) *__SIMD32(pDst)++ = __PKHBT(out4, out3, 16); #else *__SIMD32(pDst)++ = __PKHBT(out2, out1, 16); *__SIMD32(pDst)++ = __PKHBT(out4, out3, 16); #endif // #ifndef ARM_MATH_BIG_ENDIAN /* Decrement the blockSize loop counter */ blkCnt--; } /* If the blockSize is not a multiple of 4, compute any remaining output samples here. ** No loop unrolling is used. */ blkCnt = blockSize % 0x4u; #else /* Run the below code for Cortex-M0 */ /* Initialize blkCnt with number of samples */ blkCnt = blockSize; #endif /* #ifndef ARM_MATH_CM0_FAMILY */ while(blkCnt > 0u) { /* C = A * B */ /* Multiply the inputs and store the result in the destination buffer */ *pDst++ = (q15_t) __SSAT((((q31_t) (*pSrcA++) * (*pSrcB++)) >> 15), 16); /* Decrement the blockSize loop counter */ blkCnt--; } } 1. 这个函数使用了饱和算法。 所得结果是Q15格式,范围 [0x8000 0x7FFF]。 2. 一次读取两个Q15格式的数据。 3. 将四组数的乘积保存到Q31格式的变量mul1,mul2,mul3,mul4。 4. 丢弃32位数据的低15位,并把最终结果饱和到16位精度。 5. 通过SIMD指令__PKHBT将两个Q15格式的数据保存的结果数组中,从而一个指令周期就能完成两个数据的存储。 |
|
|
|
|
|
386 浏览 0 评论
求助一下关于51系列单片机的Timer0的计时问题,TH0、TL0+1的时间是怎么算的?
1670 浏览 1 评论
【RA-Eco-RA4E2-64PIN-V1.0开发板试用】开箱+Keil环境搭建+点灯+点亮OLED
1123 浏览 0 评论
【敏矽微ME32G070开发板免费体验】使用coremark测试敏矽微ME32G070 跑分
1005 浏览 0 评论
【敏矽微ME32G070开发板免费体验】开箱+点灯+点亮OLED
1229 浏览 2 评论
【youyeetoo X1 windows 开发板体验】少儿AI智能STEAM积木平台
12013 浏览 31 评论
小黑屋| 手机版| Archiver| 电子发烧友 ( 湘ICP备2023018690号 )
GMT+8, 2024-12-23 09:14 , Processed in 1.270404 second(s), Total 69, Slave 62 queries .
Powered by 电子发烧友网
© 2015 bbs.elecfans.com
关注我们的微信
下载发烧友APP
电子发烧友观察
版权所有 © 湖南华秋数字科技有限公司
电子发烧友 (电路图) 湘公网安备 43011202000918 号 电信与信息服务业务经营许可证:合字B2-20210191 工商网监 湘ICP备2023018690号