diff options
Diffstat (limited to 'Drivers/CMSIS/DSP/Source/FilteringFunctions/arm_correlate_fast_opt_q15.c')
-rw-r--r-- | Drivers/CMSIS/DSP/Source/FilteringFunctions/arm_correlate_fast_opt_q15.c | 365 |
1 files changed, 105 insertions, 260 deletions
diff --git a/Drivers/CMSIS/DSP/Source/FilteringFunctions/arm_correlate_fast_opt_q15.c b/Drivers/CMSIS/DSP/Source/FilteringFunctions/arm_correlate_fast_opt_q15.c index a1b0dbd80..13661cbf2 100644 --- a/Drivers/CMSIS/DSP/Source/FilteringFunctions/arm_correlate_fast_opt_q15.c +++ b/Drivers/CMSIS/DSP/Source/FilteringFunctions/arm_correlate_fast_opt_q15.c @@ -3,13 +3,13 @@ * Title: arm_correlate_fast_opt_q15.c * Description: Fast Q15 Correlation * - * $Date: 27. January 2017 - * $Revision: V.1.5.1 + * $Date: 18. March 2019 + * $Revision: V1.6.0 * * Target Processor: Cortex-M cores * -------------------------------------------------------------------- */ /* - * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved. + * Copyright (C) 2010-2019 ARM Limited or its affiliates. All rights reserved. * * SPDX-License-Identifier: Apache-2.0 * @@ -29,70 +29,61 @@ #include "arm_math.h" /** - * @ingroup groupFilters + @ingroup groupFilters */ /** - * @addtogroup Corr - * @{ + @addtogroup Corr + @{ */ /** - * @brief Correlation of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4. - * @param[in] *pSrcA points to the first input sequence. - * @param[in] srcALen length of the first input sequence. - * @param[in] *pSrcB points to the second input sequence. - * @param[in] srcBLen length of the second input sequence. - * @param[out] *pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. - * @param[in] *pScratch points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. - * @return none. - * - * - * \par Restrictions - * If the silicon does not support unaligned memory access enable the macro UNALIGNED_SUPPORT_DISABLE - * In this case input, output, scratch buffers should be aligned by 32-bit - * - * - * <b>Scaling and Overflow Behavior:</b> - * - * \par - * This fast version uses a 32-bit accumulator with 2.30 format. - * The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit. - * There is no saturation on intermediate additions. - * Thus, if the accumulator overflows it wraps around and distorts the result. - * The input signals should be scaled down to avoid intermediate overflows. - * Scale down one of the inputs by 1/min(srcALen, srcBLen) to avoid overflow since a - * maximum of min(srcALen, srcBLen) number of additions is carried internally. - * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result. - * - * \par - * See <code>arm_correlate_q15()</code> for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion. + @brief Correlation of Q15 sequences (fast version). + @param[in] pSrcA points to the first input sequence + @param[in] srcALen length of the first input sequence + @param[in] pSrcB points to the second input sequence + @param[in] srcBLen length of the second input sequence. + @param[out] pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. + @param[in] pScratch points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + @return none + + @par Scaling and Overflow Behavior + This fast version uses a 32-bit accumulator with 2.30 format. + The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit. + There is no saturation on intermediate additions. + Thus, if the accumulator overflows it wraps around and distorts the result. + The input signals should be scaled down to avoid intermediate overflows. + Scale down one of the inputs by 1/min(srcALen, srcBLen) to avoid overflow since a + maximum of min(srcALen, srcBLen) number of additions is carried internally. + The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result. + + @remark + Refer to \ref arm_correlate_q15() for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion. */ void arm_correlate_fast_opt_q15( - q15_t * pSrcA, - uint32_t srcALen, - q15_t * pSrcB, - uint32_t srcBLen, - q15_t * pDst, - q15_t * pScratch) + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + q15_t * pScratch) { - q15_t *pIn1; /* inputA pointer */ - q15_t *pIn2; /* inputB pointer */ - q31_t acc0, acc1, acc2, acc3; /* Accumulators */ - q15_t *py; /* Intermediate inputB pointer */ - q31_t x1, x2, x3; /* temporary variables for holding input and coefficient values */ - uint32_t j, blkCnt, outBlockSize; /* loop counter */ - int32_t inc = 1; /* Destination address modifier */ - uint32_t tapCnt; - q31_t y1, y2; - q15_t *pScr; /* Intermediate pointers */ - q15_t *pOut = pDst; /* output pointer */ -#ifdef UNALIGNED_SUPPORT_DISABLE - - q15_t a, b; - -#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ + const q15_t *pIn1; /* InputA pointer */ + const q15_t *pIn2; /* InputB pointer */ + q31_t acc0; /* Accumulators */ + q15_t *pOut = pDst; /* Output pointer */ + q15_t *pScr1 = pScratch; /* Temporary pointer for scratch */ + const q15_t *py; /* Intermediate inputB pointer */ + uint32_t j, blkCnt, outBlockSize; /* Loop counter */ + int32_t inc = 1; /* Destination address modifier */ + uint32_t tapCnt; /* Loop count */ + +#if defined (ARM_MATH_LOOPUNROLL) + q31_t acc1, acc2, acc3; /* Accumulators */ + q31_t x1, x2, x3; /* Temporary variables for holding input and coefficient values */ + q31_t y1, y2; /* State variables */ +#endif /* The algorithm implementation is based on the lengths of the inputs. */ /* srcB is always made to slide across srcA. */ @@ -110,10 +101,10 @@ void arm_correlate_fast_opt_q15( if (srcALen >= srcBLen) { /* Initialization of inputA pointer */ - pIn1 = (pSrcA); + pIn1 = pSrcA; /* Initialization of inputB pointer */ - pIn2 = (pSrcB); + pIn2 = pSrcB; /* Number of output samples is calculated */ outBlockSize = (2U * srcALen) - 1U; @@ -126,15 +117,14 @@ void arm_correlate_fast_opt_q15( /* Updating the pointer position to non zero value */ pOut += j; - } else { /* Initialization of inputA pointer */ - pIn1 = (pSrcB); + pIn1 = pSrcB; /* Initialization of inputB pointer */ - pIn2 = (pSrcA); + pIn2 = pSrcA; /* srcBLen is always considered as shorter or equal to srcALen */ j = srcBLen; @@ -147,112 +137,45 @@ void arm_correlate_fast_opt_q15( /* Destination address modifier is set to -1 */ inc = -1; - } - pScr = pScratch; + pScr1 = pScratch; /* Fill (srcBLen - 1U) zeros in scratch buffer */ - arm_fill_q15(0, pScr, (srcBLen - 1U)); + arm_fill_q15(0, pScr1, (srcBLen - 1U)); /* Update temporary scratch pointer */ - pScr += (srcBLen - 1U); + pScr1 += (srcBLen - 1U); -#ifndef UNALIGNED_SUPPORT_DISABLE /* Copy (srcALen) samples in scratch buffer */ - arm_copy_q15(pIn1, pScr, srcALen); + arm_copy_q15(pIn1, pScr1, srcALen); /* Update pointers */ - pScr += srcALen; - -#else - - /* Apply loop unrolling and do 4 Copies simultaneously. */ - j = srcALen >> 2U; - - /* First part of the processing with loop unrolling copies 4 data points at a time. - ** a second loop below copies for the remaining 1 to 3 samples. */ - while (j > 0U) - { - /* copy second buffer in reversal manner */ - *pScr++ = *pIn1++; - *pScr++ = *pIn1++; - *pScr++ = *pIn1++; - *pScr++ = *pIn1++; - - /* Decrement the loop counter */ - j--; - } + pScr1 += srcALen; - /* If the count is not a multiple of 4, copy remaining samples here. - ** No loop unrolling is used. */ - j = srcALen % 0x4U; - - while (j > 0U) - { - /* copy second buffer in reversal manner for remaining samples */ - *pScr++ = *pIn1++; - - /* Decrement the loop counter */ - j--; - } - -#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ - -#ifndef UNALIGNED_SUPPORT_DISABLE /* Fill (srcBLen - 1U) zeros at end of scratch buffer */ - arm_fill_q15(0, pScr, (srcBLen - 1U)); + arm_fill_q15(0, pScr1, (srcBLen - 1U)); /* Update pointer */ - pScr += (srcBLen - 1U); - -#else - -/* Apply loop unrolling and do 4 Copies simultaneously. */ - j = (srcBLen - 1U) >> 2U; - - /* First part of the processing with loop unrolling copies 4 data points at a time. - ** a second loop below copies for the remaining 1 to 3 samples. */ - while (j > 0U) - { - /* copy second buffer in reversal manner */ - *pScr++ = 0; - *pScr++ = 0; - *pScr++ = 0; - *pScr++ = 0; - - /* Decrement the loop counter */ - j--; - } - - /* If the count is not a multiple of 4, copy remaining samples here. - ** No loop unrolling is used. */ - j = (srcBLen - 1U) % 0x4U; - - while (j > 0U) - { - /* copy second buffer in reversal manner for remaining samples */ - *pScr++ = 0; - - /* Decrement the loop counter */ - j--; - } - -#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ + pScr1 += (srcBLen - 1U); /* Temporary pointer for scratch2 */ py = pIn2; /* Actual correlation process starts here */ + +#if defined (ARM_MATH_LOOPUNROLL) + + /* Loop unrolling: Compute 4 outputs at a time */ blkCnt = (srcALen + srcBLen - 1U) >> 2; while (blkCnt > 0) { /* Initialze temporary scratch pointer as scratch1 */ - pScr = pScratch; + pScr1 = pScratch; /* Clear Accumlators */ acc0 = 0; @@ -260,41 +183,42 @@ void arm_correlate_fast_opt_q15( acc2 = 0; acc3 = 0; - /* Read four samples from scratch1 buffer */ - x1 = *__SIMD32(pScr)++; + /* Read two samples from scratch buffer */ + x1 = read_q15x2_ia (&pScr1); - /* Read next four samples from scratch1 buffer */ - x2 = *__SIMD32(pScr)++; + /* Read next two samples from scratch buffer */ + x2 = read_q15x2_ia (&pScr1); tapCnt = (srcBLen) >> 2U; while (tapCnt > 0U) { - -#ifndef UNALIGNED_SUPPORT_DISABLE - /* Read four samples from smaller buffer */ - y1 = _SIMD32_OFFSET(pIn2); - y2 = _SIMD32_OFFSET(pIn2 + 2U); + y1 = read_q15x2_ia ((q15_t **) &pIn2); + y2 = read_q15x2_ia ((q15_t **) &pIn2); + /* multiply and accumlate */ acc0 = __SMLAD(x1, y1, acc0); - acc2 = __SMLAD(x2, y1, acc2); + /* pack input data */ #ifndef ARM_MATH_BIG_ENDIAN x3 = __PKHBT(x2, x1, 0); #else x3 = __PKHBT(x1, x2, 0); #endif + /* multiply and accumlate */ acc1 = __SMLADX(x3, y1, acc1); - x1 = _SIMD32_OFFSET(pScr); + /* Read next two samples from scratch buffer */ + x1 = read_q15x2_ia (&pScr1); + /* multiply and accumlate */ acc0 = __SMLAD(x2, y2, acc0); - acc2 = __SMLAD(x1, y2, acc2); + /* pack input data */ #ifndef ARM_MATH_BIG_ENDIAN x3 = __PKHBT(x1, x2, 0); #else @@ -302,10 +226,9 @@ void arm_correlate_fast_opt_q15( #endif acc3 = __SMLADX(x3, y1, acc3); - acc1 = __SMLADX(x3, y2, acc1); - x2 = _SIMD32_OFFSET(pScr + 2U); + x2 = read_q15x2_ia (&pScr1); #ifndef ARM_MATH_BIG_ENDIAN x3 = __PKHBT(x2, x1, 0); @@ -314,116 +237,33 @@ void arm_correlate_fast_opt_q15( #endif acc3 = __SMLADX(x3, y2, acc3); -#else - /* Read four samples from smaller buffer */ - a = *pIn2; - b = *(pIn2 + 1); - -#ifndef ARM_MATH_BIG_ENDIAN - y1 = __PKHBT(a, b, 16); -#else - y1 = __PKHBT(b, a, 16); -#endif - - a = *(pIn2 + 2); - b = *(pIn2 + 3); -#ifndef ARM_MATH_BIG_ENDIAN - y2 = __PKHBT(a, b, 16); -#else - y2 = __PKHBT(b, a, 16); -#endif - - acc0 = __SMLAD(x1, y1, acc0); - - acc2 = __SMLAD(x2, y1, acc2); - -#ifndef ARM_MATH_BIG_ENDIAN - x3 = __PKHBT(x2, x1, 0); -#else - x3 = __PKHBT(x1, x2, 0); -#endif - - acc1 = __SMLADX(x3, y1, acc1); - - a = *pScr; - b = *(pScr + 1); - -#ifndef ARM_MATH_BIG_ENDIAN - x1 = __PKHBT(a, b, 16); -#else - x1 = __PKHBT(b, a, 16); -#endif - - acc0 = __SMLAD(x2, y2, acc0); - - acc2 = __SMLAD(x1, y2, acc2); - -#ifndef ARM_MATH_BIG_ENDIAN - x3 = __PKHBT(x1, x2, 0); -#else - x3 = __PKHBT(x2, x1, 0); -#endif - - acc3 = __SMLADX(x3, y1, acc3); - - acc1 = __SMLADX(x3, y2, acc1); - - a = *(pScr + 2); - b = *(pScr + 3); - -#ifndef ARM_MATH_BIG_ENDIAN - x2 = __PKHBT(a, b, 16); -#else - x2 = __PKHBT(b, a, 16); -#endif - -#ifndef ARM_MATH_BIG_ENDIAN - x3 = __PKHBT(x2, x1, 0); -#else - x3 = __PKHBT(x1, x2, 0); -#endif - - acc3 = __SMLADX(x3, y2, acc3); - -#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ - - pIn2 += 4U; - - pScr += 4U; - - - /* Decrement the loop counter */ + /* Decrement loop counter */ tapCnt--; } - - /* Update scratch pointer for remaining samples of smaller length sequence */ - pScr -= 4U; - + pScr1 -= 4U; /* apply same above for remaining samples of smaller length sequence */ tapCnt = (srcBLen) & 3U; while (tapCnt > 0U) { - /* accumlate the results */ - acc0 += (*pScr++ * *pIn2); - acc1 += (*pScr++ * *pIn2); - acc2 += (*pScr++ * *pIn2); - acc3 += (*pScr++ * *pIn2++); + acc0 += (*pScr1++ * *pIn2); + acc1 += (*pScr1++ * *pIn2); + acc2 += (*pScr1++ * *pIn2); + acc3 += (*pScr1++ * *pIn2++); - pScr -= 3U; + pScr1 -= 3U; - /* Decrement the loop counter */ + /* Decrement loop counter */ tapCnt--; } blkCnt--; - /* Store the results in the accumulators in the destination buffer. */ *pOut = (__SSAT(acc0 >> 15U, 16)); pOut += inc; @@ -434,22 +274,27 @@ void arm_correlate_fast_opt_q15( *pOut = (__SSAT(acc3 >> 15U, 16)); pOut += inc; - /* Initialization of inputB pointer */ pIn2 = py; pScratch += 4U; - } - + /* Loop unrolling: Compute remaining outputs */ blkCnt = (srcALen + srcBLen - 1U) & 0x3; +#else + + /* Initialize blkCnt with number of samples */ + blkCnt = (srcALen + srcBLen - 1U); + +#endif /* #if defined (ARM_MATH_LOOPUNROLL) */ + /* Calculate correlation for remaining samples of Bigger length sequence */ while (blkCnt > 0) { /* Initialze temporary scratch pointer as scratch1 */ - pScr = pScratch; + pScr1 = pScratch; /* Clear Accumlators */ acc0 = 0; @@ -459,10 +304,11 @@ void arm_correlate_fast_opt_q15( while (tapCnt > 0U) { - acc0 += (*pScr++ * *pIn2++); - acc0 += (*pScr++ * *pIn2++); + /* Read next two samples from scratch buffer */ + acc0 += (*pScr1++ * *pIn2++); + acc0 += (*pScr1++ * *pIn2++); - /* Decrement the loop counter */ + /* Decrement loop counter */ tapCnt--; } @@ -473,28 +319,27 @@ void arm_correlate_fast_opt_q15( { /* accumlate the results */ - acc0 += (*pScr++ * *pIn2++); + acc0 += (*pScr1++ * *pIn2++); - /* Decrement the loop counter */ + /* Decrement loop counter */ tapCnt--; } blkCnt--; - /* Store the result in the accumulator in the destination buffer. */ - + /* The result is in 2.30 format. Convert to 1.15 with saturation. + ** Then store the output in the destination buffer. */ *pOut = (q15_t) (__SSAT((acc0 >> 15), 16)); - pOut += inc; /* Initialization of inputB pointer */ pIn2 = py; pScratch += 1U; - } + } /** - * @} end of Corr group + @} end of Corr group */ |