278 lines
5.9 KiB
C
278 lines
5.9 KiB
C
/* ----------------------------------------------------------------------
|
|
* Project: CMSIS DSP Library
|
|
* Title: arm_logsumexp_f32.c
|
|
* Description: LogSumExp
|
|
*
|
|
* $Date: 23 April 2021
|
|
* $Revision: V1.9.0
|
|
*
|
|
* Target Processor: Cortex-M and Cortex-A cores
|
|
* -------------------------------------------------------------------- */
|
|
/*
|
|
* Copyright (C) 2010-2021 ARM Limited or its affiliates. All rights reserved.
|
|
*
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
*
|
|
* Licensed under the Apache License, Version 2.0 (the License); you may
|
|
* not use this file except in compliance with the License.
|
|
* You may obtain a copy of the License at
|
|
*
|
|
* www.apache.org/licenses/LICENSE-2.0
|
|
*
|
|
* Unless required by applicable law or agreed to in writing, software
|
|
* distributed under the License is distributed on an AS IS BASIS, WITHOUT
|
|
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
* See the License for the specific language governing permissions and
|
|
* limitations under the License.
|
|
*/
|
|
|
|
#include "dsp/statistics_functions.h"
|
|
#include <limits.h>
|
|
#include <math.h>
|
|
|
|
|
|
/**
|
|
* @addtogroup LogSumExp
|
|
* @{
|
|
*/
|
|
|
|
|
|
/**
|
|
* @brief Computation of the LogSumExp
|
|
*
|
|
* In probabilistic computations, the dynamic of the probability values can be very
|
|
* wide because they come from gaussian functions.
|
|
* To avoid underflow and overflow issues, the values are represented by their log.
|
|
* In this representation, multiplying the original exp values is easy : their logs are added.
|
|
* But adding the original exp values is requiring some special handling and it is the
|
|
* goal of the LogSumExp function.
|
|
*
|
|
* If the values are x1...xn, the function is computing:
|
|
*
|
|
* ln(exp(x1) + ... + exp(xn)) and the computation is done in such a way that
|
|
* rounding issues are minimised.
|
|
*
|
|
* The max xm of the values is extracted and the function is computing:
|
|
* xm + ln(exp(x1 - xm) + ... + exp(xn - xm))
|
|
*
|
|
* @param[in] *in Pointer to an array of input values.
|
|
* @param[in] blockSize Number of samples in the input array.
|
|
* @return LogSumExp
|
|
*
|
|
*/
|
|
|
|
#if defined(ARM_MATH_MVEF) && !defined(ARM_MATH_AUTOVECTORIZE)
|
|
|
|
#include "arm_helium_utils.h"
|
|
#include "arm_vec_math.h"
|
|
|
|
float32_t arm_logsumexp_f32(const float32_t *in, uint32_t blockSize)
|
|
{
|
|
float32_t maxVal;
|
|
const float32_t *pIn;
|
|
int32_t blkCnt;
|
|
float32_t accum=0.0f;
|
|
float32_t tmp;
|
|
|
|
|
|
arm_max_no_idx_f32((float32_t *) in, blockSize, &maxVal);
|
|
|
|
|
|
blkCnt = blockSize;
|
|
pIn = in;
|
|
|
|
|
|
f32x4_t vSum = vdupq_n_f32(0.0f);
|
|
blkCnt = blockSize >> 2;
|
|
while(blkCnt > 0)
|
|
{
|
|
f32x4_t vecIn = vld1q(pIn);
|
|
f32x4_t vecExp;
|
|
|
|
vecExp = vexpq_f32(vsubq_n_f32(vecIn, maxVal));
|
|
|
|
vSum = vaddq_f32(vSum, vecExp);
|
|
|
|
/*
|
|
* Decrement the blockSize loop counter
|
|
* Advance vector source and destination pointers
|
|
*/
|
|
pIn += 4;
|
|
blkCnt --;
|
|
}
|
|
|
|
/* sum + log */
|
|
accum = vecAddAcrossF32Mve(vSum);
|
|
|
|
blkCnt = blockSize & 0x3;
|
|
while(blkCnt > 0)
|
|
{
|
|
tmp = *pIn++;
|
|
accum += expf(tmp - maxVal);
|
|
blkCnt--;
|
|
|
|
}
|
|
|
|
accum = maxVal + log(accum);
|
|
|
|
return (accum);
|
|
}
|
|
|
|
#else
|
|
#if defined(ARM_MATH_NEON) && !defined(ARM_MATH_AUTOVECTORIZE)
|
|
|
|
#include "NEMath.h"
|
|
float32_t arm_logsumexp_f32(const float32_t *in, uint32_t blockSize)
|
|
{
|
|
float32_t maxVal;
|
|
float32_t tmp;
|
|
float32x4_t tmpV, tmpVb;
|
|
float32x4_t maxValV;
|
|
uint32x4_t idxV;
|
|
float32x4_t accumV;
|
|
float32x2_t accumV2;
|
|
|
|
const float32_t *pIn;
|
|
uint32_t blkCnt;
|
|
float32_t accum;
|
|
|
|
pIn = in;
|
|
|
|
blkCnt = blockSize;
|
|
|
|
if (blockSize <= 3)
|
|
{
|
|
maxVal = *pIn++;
|
|
blkCnt--;
|
|
|
|
while(blkCnt > 0)
|
|
{
|
|
tmp = *pIn++;
|
|
|
|
if (tmp > maxVal)
|
|
{
|
|
maxVal = tmp;
|
|
}
|
|
blkCnt--;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
maxValV = vld1q_f32(pIn);
|
|
pIn += 4;
|
|
blkCnt = (blockSize - 4) >> 2;
|
|
|
|
while(blkCnt > 0)
|
|
{
|
|
tmpVb = vld1q_f32(pIn);
|
|
pIn += 4;
|
|
|
|
idxV = vcgtq_f32(tmpVb, maxValV);
|
|
maxValV = vbslq_f32(idxV, tmpVb, maxValV );
|
|
|
|
blkCnt--;
|
|
}
|
|
|
|
accumV2 = vpmax_f32(vget_low_f32(maxValV),vget_high_f32(maxValV));
|
|
accumV2 = vpmax_f32(accumV2,accumV2);
|
|
maxVal = vget_lane_f32(accumV2, 0) ;
|
|
|
|
blkCnt = (blockSize - 4) & 3;
|
|
|
|
while(blkCnt > 0)
|
|
{
|
|
tmp = *pIn++;
|
|
|
|
if (tmp > maxVal)
|
|
{
|
|
maxVal = tmp;
|
|
}
|
|
blkCnt--;
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
maxValV = vdupq_n_f32(maxVal);
|
|
pIn = in;
|
|
accum = 0;
|
|
accumV = vdupq_n_f32(0.0f);
|
|
|
|
blkCnt = blockSize >> 2;
|
|
|
|
while(blkCnt > 0)
|
|
{
|
|
tmpV = vld1q_f32(pIn);
|
|
pIn += 4;
|
|
tmpV = vsubq_f32(tmpV, maxValV);
|
|
tmpV = vexpq_f32(tmpV);
|
|
accumV = vaddq_f32(accumV, tmpV);
|
|
|
|
blkCnt--;
|
|
|
|
}
|
|
accumV2 = vpadd_f32(vget_low_f32(accumV),vget_high_f32(accumV));
|
|
accum = vget_lane_f32(accumV2, 0) + vget_lane_f32(accumV2, 1);
|
|
|
|
blkCnt = blockSize & 0x3;
|
|
while(blkCnt > 0)
|
|
{
|
|
tmp = *pIn++;
|
|
accum += expf(tmp - maxVal);
|
|
blkCnt--;
|
|
|
|
}
|
|
|
|
accum = maxVal + logf(accum);
|
|
|
|
return(accum);
|
|
}
|
|
#else
|
|
float32_t arm_logsumexp_f32(const float32_t *in, uint32_t blockSize)
|
|
{
|
|
float32_t maxVal;
|
|
float32_t tmp;
|
|
const float32_t *pIn;
|
|
uint32_t blkCnt;
|
|
float32_t accum;
|
|
|
|
pIn = in;
|
|
blkCnt = blockSize;
|
|
|
|
maxVal = *pIn++;
|
|
blkCnt--;
|
|
|
|
while(blkCnt > 0)
|
|
{
|
|
tmp = *pIn++;
|
|
|
|
if (tmp > maxVal)
|
|
{
|
|
maxVal = tmp;
|
|
}
|
|
blkCnt--;
|
|
|
|
}
|
|
|
|
blkCnt = blockSize;
|
|
pIn = in;
|
|
accum = 0;
|
|
while(blkCnt > 0)
|
|
{
|
|
tmp = *pIn++;
|
|
accum += expf(tmp - maxVal);
|
|
blkCnt--;
|
|
|
|
}
|
|
accum = maxVal + logf(accum);
|
|
|
|
return(accum);
|
|
}
|
|
#endif
|
|
#endif /* defined(ARM_MATH_MVEF) && !defined(ARM_MATH_AUTOVECTORIZE) */
|
|
|
|
/**
|
|
* @} end of LogSumExp group
|
|
*/
|