perf(acoustic): Optimize classifier and demo, remove dead code

- Cache Ort::MemoryInfo and I/O names in gunshot_classifier (avoid per-inference alloc)
- Reuse single FeatureExtractor across all files in demo_offline
- Add per-file and total inference timing to demo_offline
- Remove third_party/kiss_fft directory (fully replaced by inline FFT)
- Update fft_utils.h comments to remove stale kiss_fft references
- Verify 20/20 validation accuracy maintained, ~160ms avg per file
zhaochang_branch
赵昌 17 hours ago
parent 8dddc3fd48
commit 40379d174f

@ -15,13 +15,13 @@ void compute_power_spectrum(const float* frame, int n_fft, float* power_out);
/**
* @brief FFT
*
* kiss_fft FFT
* FFT
*/
class FftUtils {
public:
/**
* @brief FFT
* @param nfft FFT 2 kiss_fft
* @param nfft FFT
*/
explicit FftUtils(std::size_t nfft);

@ -100,7 +100,9 @@ struct GunshotClassifier::Impl {
}
}
if (!config.model_path.empty()) {
if (!LoadModel(config.model_path)) return false;
if (!LoadModel(config.model_path)) {
return false;
}
}
loaded = session != nullptr;
return loaded;
@ -120,15 +122,17 @@ struct GunshotClassifier::Impl {
}
}
Ort::MemoryInfo memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
// Cached memory info and I/O names (created once on first use)
static Ort::MemoryInfo memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
static const char* input_names[] = {"input"};
static const char* output_names[] = {"output"};
std::vector<int64_t> input_shape = {1, n_frames, n_mels};
Ort::Value input_tensor = Ort::Value::CreateTensor<float>(
memory_info, input_data.data(), input_data.size(),
input_shape.data(), input_shape.size());
// Run inference
const char* input_names[] = {"input"};
const char* output_names[] = {"output"};
auto output_tensors = session->Run(
Ort::RunOptions{nullptr},
input_names, &input_tensor, 1,

@ -7,6 +7,7 @@
#include <cmath>
#include <numeric>
#include <map>
#include <chrono>
#include "acoustic_analyzer/core/feature_extractor.h"
#include "acoustic_analyzer/core/gunshot_classifier.h"
@ -50,7 +51,8 @@ float compute_spl(const std::vector<float>& audio) {
}
Prediction process_file(const std::string& path,
GunshotClassifier& classifier) {
GunshotClassifier& classifier,
FeatureExtractor& extractor) {
Prediction result;
result.file_path = path;
result.true_label = get_parent_folder_name(path);
@ -71,7 +73,6 @@ Prediction process_file(const std::string& path,
return result;
}
FeatureExtractor extractor(sr, 2048, 512, 64, 0.0f, 8000.0f, 0.97f);
Eigen::MatrixXf mel = extractor.MelSpectrogram(audio[0]);
auto [label, confidence] = classifier.Predict(mel);
@ -217,12 +218,23 @@ int main(int argc, char** argv) {
}
std::cout << "Found " << files.size() << " WAV file(s)." << std::endl;
FeatureExtractor extractor(16000, 2048, 512, 64, 0.0f, 8000.0f, 0.97f);
std::vector<Prediction> results;
results.reserve(files.size());
double total_ms = 0.0;
for (const auto& f : files) {
results.push_back(process_file(f, classifier));
auto t0 = std::chrono::steady_clock::now();
results.push_back(process_file(f, classifier, extractor));
auto t1 = std::chrono::steady_clock::now();
total_ms += std::chrono::duration<double, std::milli>(t1 - t0).count();
}
std::cout << "\nTotal inference time: " << std::fixed << std::setprecision(2)
<< total_ms << " ms"
<< " | Avg per file: " << (files.empty() ? 0.0 : total_ms / files.size())
<< " ms" << std::endl;
print_report(results);
return 0;
}

@ -1,424 +0,0 @@
/*
* Copyright (c) 2003-2010, Mark Borgerding. All rights reserved.
* This file is part of KISS FFT - https://github.com/mborgerding/kissfft
*
* SPDX-License-Identifier: BSD-3-Clause
* See COPYING file for more information.
*/
#include <stdint.h>
#include "_kiss_fft_guts.h"
/* The guts header contains all the multiplication and addition macros that are defined for
fixed or floating point complex numbers. It also delares the kf_ internal functions.
*/
static void kf_bfly2(
kiss_fft_cpx * Fout,
const size_t fstride,
const kiss_fft_cfg st,
int m
)
{
kiss_fft_cpx * Fout2;
kiss_fft_cpx * tw1 = st->twiddles;
kiss_fft_cpx t;
Fout2 = Fout + m;
do{
C_FIXDIV(*Fout,2); C_FIXDIV(*Fout2,2);
C_MUL (t, *Fout2 , *tw1);
tw1 += fstride;
C_SUB( *Fout2 , *Fout , t );
C_ADDTO( *Fout , t );
++Fout2;
++Fout;
}while (--m);
}
static void kf_bfly4(
kiss_fft_cpx * Fout,
const size_t fstride,
const kiss_fft_cfg st,
const size_t m
)
{
kiss_fft_cpx *tw1,*tw2,*tw3;
kiss_fft_cpx scratch[6];
size_t k=m;
const size_t m2=2*m;
const size_t m3=3*m;
tw3 = tw2 = tw1 = st->twiddles;
do {
C_FIXDIV(*Fout,4); C_FIXDIV(Fout[m],4); C_FIXDIV(Fout[m2],4); C_FIXDIV(Fout[m3],4);
C_MUL(scratch[0],Fout[m] , *tw1 );
C_MUL(scratch[1],Fout[m2] , *tw2 );
C_MUL(scratch[2],Fout[m3] , *tw3 );
C_SUB( scratch[5] , *Fout, scratch[1] );
C_ADDTO(*Fout, scratch[1]);
C_ADD( scratch[3] , scratch[0] , scratch[2] );
C_SUB( scratch[4] , scratch[0] , scratch[2] );
C_SUB( Fout[m2], *Fout, scratch[3] );
tw1 += fstride;
tw2 += fstride*2;
tw3 += fstride*3;
C_ADDTO( *Fout , scratch[3] );
if(st->inverse) {
Fout[m].r = scratch[5].r - scratch[4].i;
Fout[m].i = scratch[5].i + scratch[4].r;
Fout[m3].r = scratch[5].r + scratch[4].i;
Fout[m3].i = scratch[5].i - scratch[4].r;
}else{
Fout[m].r = scratch[5].r + scratch[4].i;
Fout[m].i = scratch[5].i - scratch[4].r;
Fout[m3].r = scratch[5].r - scratch[4].i;
Fout[m3].i = scratch[5].i + scratch[4].r;
}
++Fout;
}while(--k);
}
static void kf_bfly3(
kiss_fft_cpx * Fout,
const size_t fstride,
const kiss_fft_cfg st,
size_t m
)
{
size_t k=m;
const size_t m2 = 2*m;
kiss_fft_cpx *tw1,*tw2;
kiss_fft_cpx scratch[5];
kiss_fft_cpx epi3;
epi3 = st->twiddles[fstride*m];
tw1=tw2=st->twiddles;
do{
C_FIXDIV(*Fout,3); C_FIXDIV(Fout[m],3); C_FIXDIV(Fout[m2],3);
C_MUL(scratch[1],Fout[m] , *tw1);
C_MUL(scratch[2],Fout[m2] , *tw2);
C_ADD(scratch[3],scratch[1],scratch[2]);
C_SUB(scratch[0],scratch[1],scratch[2]);
tw1 += fstride;
tw2 += fstride*2;
Fout[m].r = Fout->r - HALF_OF(scratch[3].r);
Fout[m].i = Fout->i - HALF_OF(scratch[3].i);
C_MULBYSCALAR( scratch[0] , epi3.i );
C_ADDTO(*Fout,scratch[3]);
Fout[m2].r = Fout[m].r + scratch[0].i;
Fout[m2].i = Fout[m].i - scratch[0].r;
Fout[m].r -= scratch[0].i;
Fout[m].i += scratch[0].r;
++Fout;
}while(--k);
}
static void kf_bfly5(
kiss_fft_cpx * Fout,
const size_t fstride,
const kiss_fft_cfg st,
int m
)
{
kiss_fft_cpx *Fout0,*Fout1,*Fout2,*Fout3,*Fout4;
int u;
kiss_fft_cpx scratch[13];
kiss_fft_cpx * twiddles = st->twiddles;
kiss_fft_cpx *tw;
kiss_fft_cpx ya,yb;
ya = twiddles[fstride*m];
yb = twiddles[fstride*2*m];
Fout0=Fout;
Fout1=Fout0+m;
Fout2=Fout0+2*m;
Fout3=Fout0+3*m;
Fout4=Fout0+4*m;
tw=st->twiddles;
for ( u=0; u<m; ++u ) {
C_FIXDIV( *Fout0,5); C_FIXDIV( *Fout1,5); C_FIXDIV( *Fout2,5); C_FIXDIV( *Fout3,5); C_FIXDIV( *Fout4,5);
scratch[0] = *Fout0;
C_MUL(scratch[1] ,*Fout1, tw[u*fstride]);
C_MUL(scratch[2] ,*Fout2, tw[2*u*fstride]);
C_MUL(scratch[3] ,*Fout3, tw[3*u*fstride]);
C_MUL(scratch[4] ,*Fout4, tw[4*u*fstride]);
C_ADD( scratch[7],scratch[1],scratch[4]);
C_SUB( scratch[10],scratch[1],scratch[4]);
C_ADD( scratch[8],scratch[2],scratch[3]);
C_SUB( scratch[9],scratch[2],scratch[3]);
Fout0->r += scratch[7].r + scratch[8].r;
Fout0->i += scratch[7].i + scratch[8].i;
scratch[5].r = scratch[0].r + S_MUL(scratch[7].r,ya.r) + S_MUL(scratch[8].r,yb.r);
scratch[5].i = scratch[0].i + S_MUL(scratch[7].i,ya.r) + S_MUL(scratch[8].i,yb.r);
scratch[6].r = S_MUL(scratch[10].i,ya.i) + S_MUL(scratch[9].i,yb.i);
scratch[6].i = -S_MUL(scratch[10].r,ya.i) - S_MUL(scratch[9].r,yb.i);
C_SUB(*Fout1,scratch[5],scratch[6]);
C_ADD(*Fout4,scratch[5],scratch[6]);
scratch[11].r = scratch[0].r + S_MUL(scratch[7].r,yb.r) + S_MUL(scratch[8].r,ya.r);
scratch[11].i = scratch[0].i + S_MUL(scratch[7].i,yb.r) + S_MUL(scratch[8].i,ya.r);
scratch[12].r = - S_MUL(scratch[10].i,yb.i) + S_MUL(scratch[9].i,ya.i);
scratch[12].i = S_MUL(scratch[10].r,yb.i) - S_MUL(scratch[9].r,ya.i);
C_ADD(*Fout2,scratch[11],scratch[12]);
C_SUB(*Fout3,scratch[11],scratch[12]);
++Fout0;++Fout1;++Fout2;++Fout3;++Fout4;
}
}
/* perform the butterfly for one stage of a mixed radix FFT */
static void kf_bfly_generic(
kiss_fft_cpx * Fout,
const size_t fstride,
const kiss_fft_cfg st,
int m,
int p
)
{
int u,k,q1,q;
kiss_fft_cpx * twiddles = st->twiddles;
kiss_fft_cpx t;
int Norig = st->nfft;
kiss_fft_cpx * scratch = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC(sizeof(kiss_fft_cpx)*p);
if (scratch == NULL){
KISS_FFT_ERROR("Memory allocation failed.");
return;
}
for ( u=0; u<m; ++u ) {
k=u;
for ( q1=0 ; q1<p ; ++q1 ) {
scratch[q1] = Fout[ k ];
C_FIXDIV(scratch[q1],p);
k += m;
}
k=u;
for ( q1=0 ; q1<p ; ++q1 ) {
int twidx=0;
Fout[ k ] = scratch[0];
for (q=1;q<p;++q ) {
twidx += fstride * k;
if (twidx>=Norig) twidx-=Norig;
C_MUL(t,scratch[q] , twiddles[twidx] );
C_ADDTO( Fout[ k ] ,t);
}
k += m;
}
}
KISS_FFT_TMP_FREE(scratch);
}
static
void kf_work(
kiss_fft_cpx * Fout,
const kiss_fft_cpx * f,
const size_t fstride,
int in_stride,
int * factors,
const kiss_fft_cfg st
)
{
kiss_fft_cpx * Fout_beg=Fout;
const int p=*factors++; /* the radix */
const int m=*factors++; /* stage's fft length/p */
const kiss_fft_cpx * Fout_end = Fout + p*m;
#ifdef _OPENMP
// use openmp extensions at the
// top-level (not recursive)
if (fstride==1 && p<=5 && m!=1)
{
int k;
// execute the p different work units in different threads
# pragma omp parallel for
for (k=0;k<p;++k)
kf_work( Fout +k*m, f+ fstride*in_stride*k,fstride*p,in_stride,factors,st);
// all threads have joined by this point
switch (p) {
case 2: kf_bfly2(Fout,fstride,st,m); break;
case 3: kf_bfly3(Fout,fstride,st,m); break;
case 4: kf_bfly4(Fout,fstride,st,m); break;
case 5: kf_bfly5(Fout,fstride,st,m); break;
default: kf_bfly_generic(Fout,fstride,st,m,p); break;
}
return;
}
#endif
if (m==1) {
do{
*Fout = *f;
f += fstride*in_stride;
}while(++Fout != Fout_end );
}else{
do{
// recursive call:
// DFT of size m*p performed by doing
// p instances of smaller DFTs of size m,
// each one takes a decimated version of the input
kf_work( Fout , f, fstride*p, in_stride, factors,st);
f += fstride*in_stride;
}while( (Fout += m) != Fout_end );
}
Fout=Fout_beg;
// recombine the p smaller DFTs
switch (p) {
case 2: kf_bfly2(Fout,fstride,st,m); break;
case 3: kf_bfly3(Fout,fstride,st,m); break;
case 4: kf_bfly4(Fout,fstride,st,m); break;
case 5: kf_bfly5(Fout,fstride,st,m); break;
default: kf_bfly_generic(Fout,fstride,st,m,p); break;
}
}
/* facbuf is populated by p1,m1,p2,m2, ...
where
p[i] * m[i] = m[i-1]
m0 = n */
static
void kf_factor(int n,int * facbuf)
{
int p=4;
double floor_sqrt;
floor_sqrt = floor( sqrt((double)n) );
/*factor out powers of 4, powers of 2, then any remaining primes */
do {
while (n % p) {
switch (p) {
case 4: p = 2; break;
case 2: p = 3; break;
default: p += 2; break;
}
if (p > floor_sqrt)
p = n; /* no more factors, skip to end */
}
n /= p;
*facbuf++ = p;
*facbuf++ = n;
} while (n > 1);
}
/*
*
* User-callable function to allocate all necessary storage space for the fft.
*
* The return value is a contiguous block of memory, allocated with malloc. As such,
* It can be freed with free(), rather than a kiss_fft-specific function.
* */
kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem )
{
KISS_FFT_ALIGN_CHECK(mem)
kiss_fft_cfg st=NULL;
// check for overflow condition {memneeded > SIZE_MAX}.
if (nfft >= (SIZE_MAX - 2*sizeof(struct kiss_fft_state))/sizeof(kiss_fft_cpx))
return NULL;
size_t memneeded = KISS_FFT_ALIGN_SIZE_UP(sizeof(struct kiss_fft_state)
+ sizeof(kiss_fft_cpx)*(nfft-1)); /* twiddle factors*/
if ( lenmem==NULL ) {
st = ( kiss_fft_cfg)KISS_FFT_MALLOC( memneeded );
}else{
if (mem != NULL && *lenmem >= memneeded)
st = (kiss_fft_cfg)mem;
*lenmem = memneeded;
}
if (st) {
int i;
st->nfft=nfft;
st->inverse = inverse_fft;
for (i=0;i<nfft;++i) {
const double pi=3.141592653589793238462643383279502884197169399375105820974944;
double phase = -2*pi*i / nfft;
if (st->inverse)
phase *= -1;
kf_cexp(st->twiddles+i, phase );
}
kf_factor(nfft,st->factors);
}
return st;
}
void kiss_fft_stride(kiss_fft_cfg st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int in_stride)
{
if (fin == fout) {
//NOTE: this is not really an in-place FFT algorithm.
//It just performs an out-of-place FFT into a temp buffer
if (fout == NULL){
KISS_FFT_ERROR("fout buffer NULL.");
return;
}
kiss_fft_cpx * tmpbuf = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC( sizeof(kiss_fft_cpx)*st->nfft);
if (tmpbuf == NULL){
KISS_FFT_ERROR("Memory allocation error.");
return;
}
kf_work(tmpbuf,fin,1,in_stride, st->factors,st);
memcpy(fout,tmpbuf,sizeof(kiss_fft_cpx)*st->nfft);
KISS_FFT_TMP_FREE(tmpbuf);
}else{
kf_work( fout, fin, 1,in_stride, st->factors,st );
}
}
void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
{
kiss_fft_stride(cfg,fin,fout,1);
}
void kiss_fft_cleanup(void)
{
// nothing needed any more
}
int kiss_fft_next_fast_size(int n)
{
while(1) {
int m=n;
while ( (m%2) == 0 ) m/=2;
while ( (m%3) == 0 ) m/=3;
while ( (m%5) == 0 ) m/=5;
if (m<=1)
break; /* n is completely factorable by twos, threes, and fives */
n++;
}
return n;
}

@ -1,184 +0,0 @@
/*
* Copyright (c) 2003-2010, Mark Borgerding. All rights reserved.
* This file is part of KISS FFT - https://github.com/mborgerding/kissfft
*
* SPDX-License-Identifier: BSD-3-Clause
* See COPYING file for more information.
*/
#ifndef KISS_FFT_H
#define KISS_FFT_H
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <string.h>
// Define KISS_FFT_SHARED macro to properly export symbols
#ifdef KISS_FFT_SHARED
# ifdef _WIN32
# ifdef KISS_FFT_BUILD
# define KISS_FFT_API __declspec(dllexport)
# else
# define KISS_FFT_API __declspec(dllimport)
# endif
# else
# define KISS_FFT_API __attribute__ ((visibility ("default")))
# endif
#else
# define KISS_FFT_API
#endif
#ifdef __cplusplus
extern "C" {
#endif
/*
ATTENTION!
If you would like a :
-- a utility that will handle the caching of fft objects
-- real-only (no imaginary time component ) FFT
-- a multi-dimensional FFT
-- a command-line utility to perform ffts
-- a command-line utility to perform fast-convolution filtering
Then see kfc.h kiss_fftr.h kiss_fftnd.h fftutil.c kiss_fastfir.c
in the tools/ directory.
*/
/* User may override KISS_FFT_MALLOC and/or KISS_FFT_FREE. */
#ifdef USE_SIMD
#ifdef HAVE_LASX
# include <lasxintrin.h>
# define kiss_fft_scalar __m256
# ifndef KISS_FFT_MALLOC
# define KISS_FFT_MALLOC(nbytes) aligned_alloc(32, KISS_FFT_ALIGN_SIZE_UP(nbytes))
# define KISS_FFT_ALIGN_CHECK(ptr)
# define KISS_FFT_ALIGN_SIZE_UP(size) ((size + 31UL) & ~0x1FUL)
# endif
# ifndef KISS_FFT_FREE
# define KISS_FFT_FREE free
# endif
#elif defined(HAVE_LSX)
# include <lsxintrin.h>
# define kiss_fft_scalar __m128
# ifndef KISS_FFT_MALLOC
# define KISS_FFT_MALLOC(nbytes) aligned_alloc(16, KISS_FFT_ALIGN_SIZE_UP(nbytes))
# define KISS_FFT_ALIGN_CHECK(ptr)
# define KISS_FFT_ALIGN_SIZE_UP(size) ((size + 15UL) & ~0xFUL)
# endif
# ifndef KISS_FFT_FREE
# define KISS_FFT_FREE free
# endif
#else
# include <xmmintrin.h>
# define kiss_fft_scalar __m128
# ifndef KISS_FFT_MALLOC
# define KISS_FFT_MALLOC(nbytes) _mm_malloc(nbytes,16)
# define KISS_FFT_ALIGN_CHECK(ptr)
# define KISS_FFT_ALIGN_SIZE_UP(size) ((size + 15UL) & ~0xFUL)
# endif
# ifndef KISS_FFT_FREE
# define KISS_FFT_FREE _mm_free
# endif
#endif
#else
# define KISS_FFT_ALIGN_CHECK(ptr)
# define KISS_FFT_ALIGN_SIZE_UP(size) (size)
# ifndef KISS_FFT_MALLOC
# define KISS_FFT_MALLOC malloc
# endif
# ifndef KISS_FFT_FREE
# define KISS_FFT_FREE free
# endif
#endif
#ifdef FIXED_POINT
#include <stdint.h>
# if (FIXED_POINT == 32)
# define kiss_fft_scalar int32_t
# else
# define kiss_fft_scalar int16_t
# endif
#else
# ifndef kiss_fft_scalar
/* default is float */
# define kiss_fft_scalar float
# endif
#endif
typedef struct {
kiss_fft_scalar r;
kiss_fft_scalar i;
}kiss_fft_cpx;
typedef struct kiss_fft_state* kiss_fft_cfg;
/*
* kiss_fft_alloc
*
* Initialize a FFT (or IFFT) algorithm's cfg/state buffer.
*
* typical usage: kiss_fft_cfg mycfg=kiss_fft_alloc(1024,0,NULL,NULL);
*
* The return value from fft_alloc is a cfg buffer used internally
* by the fft routine or NULL.
*
* If lenmem is NULL, then kiss_fft_alloc will allocate a cfg buffer using malloc.
* The returned value should be free()d when done to avoid memory leaks.
*
* The state can be placed in a user supplied buffer 'mem':
* If lenmem is not NULL and mem is not NULL and *lenmem is large enough,
* then the function places the cfg in mem and the size used in *lenmem
* and returns mem.
*
* If lenmem is not NULL and ( mem is NULL or *lenmem is not large enough),
* then the function returns NULL and places the minimum cfg
* buffer size in *lenmem.
* */
kiss_fft_cfg KISS_FFT_API kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem);
/*
* kiss_fft(cfg,in_out_buf)
*
* Perform an FFT on a complex input buffer.
* for a forward FFT,
* fin should be f[0] , f[1] , ... ,f[nfft-1]
* fout will be F[0] , F[1] , ... ,F[nfft-1]
* Note that each element is complex and can be accessed like
f[k].r and f[k].i
* */
void KISS_FFT_API kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout);
/*
A more generic version of the above function. It reads its input from every Nth sample.
* */
void KISS_FFT_API kiss_fft_stride(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int fin_stride);
/* If kiss_fft_alloc allocated a buffer, it is one contiguous
buffer and can be simply free()d when no longer needed*/
#define kiss_fft_free KISS_FFT_FREE
/*
Cleans up some memory that gets managed internally. Not necessary to call, but it might clean up
your compiler output to call this before you exit.
*/
void KISS_FFT_API kiss_fft_cleanup(void);
/*
* Returns the smallest integer k, such that k>=n and k has only "fast" factors (2,3,5)
*/
int KISS_FFT_API kiss_fft_next_fast_size(int n);
/* for real ffts, we need an even size */
#define kiss_fftr_next_fast_size_real(n) \
(kiss_fft_next_fast_size( ((n)+1)>>1)<<1)
#ifdef __cplusplus
}
#endif
#endif
Loading…
Cancel
Save