Renesas / mbed-dsp-neon

Fork of mbed-dsp. CMSIS-DSP library of supporting NEON

Dependents:   mbed-os-example-cmsis_dsp_neon

Fork of mbed-dsp by mbed official

Information

Japanese version is available in lower part of this page.
このページの後半に日本語版が用意されています．

CMSIS-DSP of supporting NEON

What is this ?

A library for CMSIS-DSP of supporting NEON.
We supported the NEON to CMSIS-DSP Ver1.4.3(CMSIS V4.1) that ARM supplied, has achieved the processing speed improvement.
If you use the mbed-dsp library, you can use to replace this library.
CMSIS-DSP of supporting NEON is provied as a library.

Library Creation environment

CMSIS-DSP library of supporting NEON was created by the following environment.

  • Compiler
    ARMCC Version 5.03
  • Compile option switch[C Compiler]
   -DARM_MATH_MATRIX_CHECK -DARM_MATH_ROUNDING -O3 -Otime --cpu=Cortex-A9 --littleend --arm 
   --apcs=/interwork --no_unaligned_access --fpu=vfpv3_fp16 --fpmode=fast --apcs=/hardfp 
   --vectorize --asm
  • Compile option switch[Assembler]
   --cpreproc --cpu=Cortex-A9 --littleend --arm --apcs=/interwork --no_unaligned_access 
   --fpu=vfpv3_fp16 --fpmode=fast --apcs=/hardfp


Effects of NEON support

In the data which passes to each function, large size will be expected more effective than small size.
Also if the data is a multiple of 16, effect will be expected in every function in the CMSIS-DSP.


NEON対応CMSIS-DSP

概要

NEON対応したCMSIS-DSPのライブラリです。
ARM社提供のCMSIS-DSP Ver1.4.3(CMSIS V4.1)をターゲットにNEON対応を行ない、処理速度向上を実現しております。
mbed-dspライブラリを使用している場合は、本ライブラリに置き換えて使用することができます。
NEON対応したCMSIS-DSPはライブラリで提供します。

ライブラリ作成環境

NEON対応CMSIS-DSPライブラリは、以下の環境で作成しています。

  • コンパイラ
    ARMCC Version 5.03
  • コンパイルオプションスイッチ[C Compiler]
   -DARM_MATH_MATRIX_CHECK -DARM_MATH_ROUNDING -O3 -Otime --cpu=Cortex-A9 --littleend --arm 
   --apcs=/interwork --no_unaligned_access --fpu=vfpv3_fp16 --fpmode=fast --apcs=/hardfp 
   --vectorize --asm
  • コンパイルオプションスイッチ[Assembler]
   --cpreproc --cpu=Cortex-A9 --littleend --arm --apcs=/interwork --no_unaligned_access 
   --fpu=vfpv3_fp16 --fpmode=fast --apcs=/hardfp


NEON対応による効果について

CMSIS-DSP内の各関数へ渡すデータは、小さいサイズよりも大きいサイズの方が効果が見込めます。
また、16の倍数のデータであれば、CMSIS-DSP内のどの関数でも効果が見込めます。


cmsis_dsp/BasicMathFunctions/arm_dot_prod_q31.c@2:da51fb522205, 2013-05-30 (annotated)

Committer:
emilmont
Date:
Thu May 30 17:10:11 2013 +0100
Revision:
2:da51fb522205
Parent:
1:fdd22bb7aa52
Child:
3:7a284390b0ce
Keep "cmsis-dsp" module in synch with its source

Who changed what in which revision?

UserRevisionLine numberNew contents of line
emilmont 1:fdd22bb7aa52 1 /* ----------------------------------------------------------------------
emilmont 1:fdd22bb7aa52 2 * Copyright (C) 2010 ARM Limited. All rights reserved.
emilmont 1:fdd22bb7aa52 3 *
emilmont 1:fdd22bb7aa52 4 * $Date: 15. February 2012
emilmont 2:da51fb522205 5 * $Revision: V1.1.0
emilmont 1:fdd22bb7aa52 6 *
emilmont 2:da51fb522205 7 * Project: CMSIS DSP Library
emilmont 2:da51fb522205 8 * Title: arm_dot_prod_q31.c
emilmont 1:fdd22bb7aa52 9 *
emilmont 2:da51fb522205 10 * Description: Q31 dot product.
emilmont 1:fdd22bb7aa52 11 *
emilmont 1:fdd22bb7aa52 12 * Target Processor: Cortex-M4/Cortex-M3/Cortex-M0
emilmont 1:fdd22bb7aa52 13 *
emilmont 1:fdd22bb7aa52 14 * Version 1.1.0 2012/02/15
emilmont 1:fdd22bb7aa52 15 * Updated with more optimizations, bug fixes and minor API changes.
emilmont 1:fdd22bb7aa52 16 *
emilmont 1:fdd22bb7aa52 17 * Version 1.0.10 2011/7/15
emilmont 1:fdd22bb7aa52 18 * Big Endian support added and Merged M0 and M3/M4 Source code.
emilmont 1:fdd22bb7aa52 19 *
emilmont 1:fdd22bb7aa52 20 * Version 1.0.3 2010/11/29
emilmont 1:fdd22bb7aa52 21 * Re-organized the CMSIS folders and updated documentation.
emilmont 1:fdd22bb7aa52 22 *
emilmont 1:fdd22bb7aa52 23 * Version 1.0.2 2010/11/11
emilmont 1:fdd22bb7aa52 24 * Documentation updated.
emilmont 1:fdd22bb7aa52 25 *
emilmont 1:fdd22bb7aa52 26 * Version 1.0.1 2010/10/05
emilmont 1:fdd22bb7aa52 27 * Production release and review comments incorporated.
emilmont 1:fdd22bb7aa52 28 *
emilmont 1:fdd22bb7aa52 29 * Version 1.0.0 2010/09/20
emilmont 1:fdd22bb7aa52 30 * Production release and review comments incorporated.
emilmont 1:fdd22bb7aa52 31 *
emilmont 1:fdd22bb7aa52 32 * Version 0.0.7 2010/06/10
emilmont 1:fdd22bb7aa52 33 * Misra-C changes done
emilmont 1:fdd22bb7aa52 34 * -------------------------------------------------------------------- */
emilmont 1:fdd22bb7aa52 35
emilmont 1:fdd22bb7aa52 36 #include "arm_math.h"
emilmont 1:fdd22bb7aa52 37
emilmont 1:fdd22bb7aa52 38 /**
emilmont 1:fdd22bb7aa52 39 * @ingroup groupMath
emilmont 1:fdd22bb7aa52 40 */
emilmont 1:fdd22bb7aa52 41
emilmont 1:fdd22bb7aa52 42 /**
emilmont 1:fdd22bb7aa52 43 * @addtogroup dot_prod
emilmont 1:fdd22bb7aa52 44 * @{
emilmont 1:fdd22bb7aa52 45 */
emilmont 1:fdd22bb7aa52 46
emilmont 1:fdd22bb7aa52 47 /**
emilmont 1:fdd22bb7aa52 48 * @brief Dot product of Q31 vectors.
emilmont 1:fdd22bb7aa52 49 * @param[in] *pSrcA points to the first input vector
emilmont 1:fdd22bb7aa52 50 * @param[in] *pSrcB points to the second input vector
emilmont 1:fdd22bb7aa52 51 * @param[in] blockSize number of samples in each vector
emilmont 1:fdd22bb7aa52 52 * @param[out] *result output result returned here
emilmont 1:fdd22bb7aa52 53 * @return none.
emilmont 1:fdd22bb7aa52 54 *
emilmont 1:fdd22bb7aa52 55 * <b>Scaling and Overflow Behavior:</b>
emilmont 1:fdd22bb7aa52 56 * \par
emilmont 1:fdd22bb7aa52 57 * The intermediate multiplications are in 1.31 x 1.31 = 2.62 format and these
emilmont 1:fdd22bb7aa52 58 * are truncated to 2.48 format by discarding the lower 14 bits.
emilmont 1:fdd22bb7aa52 59 * The 2.48 result is then added without saturation to a 64-bit accumulator in 16.48 format.
emilmont 1:fdd22bb7aa52 60 * There are 15 guard bits in the accumulator and there is no risk of overflow as long as
emilmont 1:fdd22bb7aa52 61 * the length of the vectors is less than 2^16 elements.
emilmont 1:fdd22bb7aa52 62 * The return result is in 16.48 format.
emilmont 1:fdd22bb7aa52 63 */
emilmont 1:fdd22bb7aa52 64
emilmont 1:fdd22bb7aa52 65 void arm_dot_prod_q31(
emilmont 1:fdd22bb7aa52 66 q31_t * pSrcA,
emilmont 1:fdd22bb7aa52 67 q31_t * pSrcB,
emilmont 1:fdd22bb7aa52 68 uint32_t blockSize,
emilmont 1:fdd22bb7aa52 69 q63_t * result)
emilmont 1:fdd22bb7aa52 70 {
emilmont 1:fdd22bb7aa52 71 q63_t sum = 0; /* Temporary result storage */
emilmont 1:fdd22bb7aa52 72 uint32_t blkCnt; /* loop counter */
emilmont 1:fdd22bb7aa52 73
emilmont 1:fdd22bb7aa52 74
emilmont 1:fdd22bb7aa52 75 #ifndef ARM_MATH_CM0
emilmont 1:fdd22bb7aa52 76
emilmont 1:fdd22bb7aa52 77 /* Run the below code for Cortex-M4 and Cortex-M3 */
emilmont 1:fdd22bb7aa52 78 q31_t inA1, inA2, inA3, inA4;
emilmont 1:fdd22bb7aa52 79 q31_t inB1, inB2, inB3, inB4;
emilmont 1:fdd22bb7aa52 80
emilmont 1:fdd22bb7aa52 81 /*loop Unrolling */
emilmont 1:fdd22bb7aa52 82 blkCnt = blockSize >> 2u;
emilmont 1:fdd22bb7aa52 83
emilmont 1:fdd22bb7aa52 84 /* First part of the processing with loop unrolling. Compute 4 outputs at a time.
emilmont 1:fdd22bb7aa52 85 ** a second loop below computes the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 86 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 87 {
emilmont 1:fdd22bb7aa52 88 /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */
emilmont 1:fdd22bb7aa52 89 /* Calculate dot product and then store the result in a temporary buffer. */
emilmont 1:fdd22bb7aa52 90 inA1 = *pSrcA++;
emilmont 1:fdd22bb7aa52 91 inA2 = *pSrcA++;
emilmont 1:fdd22bb7aa52 92 inA3 = *pSrcA++;
emilmont 1:fdd22bb7aa52 93 inA4 = *pSrcA++;
emilmont 1:fdd22bb7aa52 94 inB1 = *pSrcB++;
emilmont 1:fdd22bb7aa52 95 inB2 = *pSrcB++;
emilmont 1:fdd22bb7aa52 96 inB3 = *pSrcB++;
emilmont 1:fdd22bb7aa52 97 inB4 = *pSrcB++;
emilmont 1:fdd22bb7aa52 98
emilmont 1:fdd22bb7aa52 99 sum += ((q63_t) inA1 * inB1) >> 14u;
emilmont 1:fdd22bb7aa52 100 sum += ((q63_t) inA2 * inB2) >> 14u;
emilmont 1:fdd22bb7aa52 101 sum += ((q63_t) inA3 * inB3) >> 14u;
emilmont 1:fdd22bb7aa52 102 sum += ((q63_t) inA4 * inB4) >> 14u;
emilmont 1:fdd22bb7aa52 103
emilmont 1:fdd22bb7aa52 104 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 105 blkCnt--;
emilmont 1:fdd22bb7aa52 106 }
emilmont 1:fdd22bb7aa52 107
emilmont 1:fdd22bb7aa52 108 /* If the blockSize is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 109 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 110 blkCnt = blockSize % 0x4u;
emilmont 1:fdd22bb7aa52 111
emilmont 1:fdd22bb7aa52 112 #else
emilmont 1:fdd22bb7aa52 113
emilmont 1:fdd22bb7aa52 114 /* Run the below code for Cortex-M0 */
emilmont 1:fdd22bb7aa52 115
emilmont 1:fdd22bb7aa52 116 /* Initialize blkCnt with number of samples */
emilmont 1:fdd22bb7aa52 117 blkCnt = blockSize;
emilmont 1:fdd22bb7aa52 118
emilmont 1:fdd22bb7aa52 119 #endif /* #ifndef ARM_MATH_CM0 */
emilmont 1:fdd22bb7aa52 120
emilmont 1:fdd22bb7aa52 121
emilmont 1:fdd22bb7aa52 122 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 123 {
emilmont 1:fdd22bb7aa52 124 /* C = A[0]* B[0] + A[1]* B[1] + A[2]* B[2] + .....+ A[blockSize-1]* B[blockSize-1] */
emilmont 1:fdd22bb7aa52 125 /* Calculate dot product and then store the result in a temporary buffer. */
emilmont 1:fdd22bb7aa52 126 sum += ((q63_t) * pSrcA++ * *pSrcB++) >> 14u;
emilmont 1:fdd22bb7aa52 127
emilmont 1:fdd22bb7aa52 128 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 129 blkCnt--;
emilmont 1:fdd22bb7aa52 130 }
emilmont 1:fdd22bb7aa52 131
emilmont 1:fdd22bb7aa52 132 /* Store the result in the destination buffer in 16.48 format */
emilmont 1:fdd22bb7aa52 133 *result = sum;
emilmont 1:fdd22bb7aa52 134 }
emilmont 1:fdd22bb7aa52 135
emilmont 1:fdd22bb7aa52 136 /**
emilmont 1:fdd22bb7aa52 137 * @} end of dot_prod group
emilmont 1:fdd22bb7aa52 138 */