mirrors/pulseaudio
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pulseaudio/pulseaudio.git synced 2025-11-04 13:29:59 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions

module-equalizer-sink:

Browse source 
removed liboil
    added sse2 optimized dsp logic implementation
    cleaned up a bit
This commit is contained in:
Jason Newton 2009-07-27 01:22:26 -07:00
parent 702480a883
commit c7fcc9cc01
2 changed files with 243 additions and 156 deletions
Download patch file Download diff file Expand all files Collapse all files

4
src/Makefile.am
View file

@ -1383,9 +1383,9 @@ module_ladspa_sink_la_LDFLAGS = $(MODULE_LDFLAGS)
module_ladspa_sink_la_LIBADD = $(AM_LIBADD) $(LIBLTDL) libpulsecore-@PA_MAJORMINORMICRO@.la libpulsecommon-@PA_MAJORMINORMICRO@.la libpulse.la
module_equalizer_sink_la_SOURCES = modules/module-equalizer-sink.c
module_equalizer_sink_la_CFLAGS = $(AM_CFLAGS) $(LIBOIL_CFLAGS)
module_equalizer_sink_la_CFLAGS = $(AM_CFLAGS)
module_equalizer_sink_la_LDFLAGS = $(MODULE_LDFLAGS)
module_equalizer_sink_la_LIBADD = $(AM_LIBADD) $(LIBLTDL) $(LIBOIL_LIBS) -lfftw3f libpulsecore-@PA_MAJORMINORMICRO@.la libpulsecommon-@PA_MAJORMINORMICRO@.la libpulse.la
module_equalizer_sink_la_LIBADD = $(AM_LIBADD) $(LIBLTDL) -lfftw3f libpulsecore-@PA_MAJORMINORMICRO@.la libpulsecommon-@PA_MAJORMINORMICRO@.la libpulse.la
module_match_la_SOURCES = modules/module-match.c
module_match_la_LDFLAGS = $(MODULE_LDFLAGS)

395
src/modules/module-equalizer-sink.c
View file

@ -29,11 +29,13 @@ USA.
#include <config.h>
#endif
#include <stdlib.h>
#include <stdio.h>
#include <float.h>
#include <math.h>
#include <fftw3.h>
#include <float.h>
#include <string.h>
#include <malloc.h>
#include <pulse/xmalloc.h>
#include <pulse/i18n.h>
@ -55,6 +57,14 @@ USA.
#include <time.h>
//#undef __SSE2__
#ifdef __SSE2__
#include <xmmintrin.h>
#include <emmintrin.h>
#endif
#include "module-equalizer-sink-symdef.h"
PA_MODULE_AUTHOR("Jason Newton");
@ -82,10 +92,12 @@ struct userdata {
* the latency of the filter, calculated from window_size
* based on constraints of COLA and window function
*/
size_t latency;
size_t latency;//Really just R but made into it's own variable
//for twiddling with pulseaudio
size_t overlap_size;//window_size-R
size_t samples_gathered;
size_t max_output;
size_t max_output;//max amount of samples outputable in a single
//message
size_t target_samples;
float *H;//frequency response filter (magnitude based)
float *W;//windowing function (time domain)
@ -109,75 +121,38 @@ static const char* const valid_modargs[] = {
NULL
};
uint64_t time_diff(struct timespec *timeA_p, struct timespec *timeB_p);
void hanning_normalized_window(float *W,size_t window_size);
void hanning_window(float *W,size_t window_size);
void hamming_window(float *W,size_t window_size);
void blackman_window(float *W,size_t window_size);
void sin_window(float *W,size_t window_size);
void array_out(const char *name,float *a,size_t length);
static void dsp_logic(float *dst,struct userdata *u);
static uint64_t time_diff(struct timespec *timeA_p, struct timespec *timeB_p);
static void hanning_window(float *W,size_t window_size);
static void array_out(const char *name,float *a,size_t length);
static void process_samples(struct userdata *u);
void input_buffer(struct userdata *u,pa_memchunk *in);
static void input_buffer(struct userdata *u,pa_memchunk *in);
void dsp_logic(
float * __restrict__ dst,
float * __restrict__ src,
float * __restrict__ overlap,
const float * __restrict__ H,
const float * __restrict__ W,
fftwf_complex * __restrict__ output_window,
struct userdata *u);
#define v_size 4
#define gettime(x) clock_gettime(CLOCK_MONOTONIC,&x)
#define tdiff(x,y) time_diff(&x,&y)
#define mround(x,y) (x%y==0?x:(x/y+1)*y)
uint64_t time_diff(struct timespec *timeA_p, struct timespec *timeB_p)
{
return ((timeA_p->tv_sec * 1000000000) + timeA_p->tv_nsec) -
((timeB_p->tv_sec * 1000000000) + timeB_p->tv_nsec);
return ((timeA_p->tv_sec * 1000000000ULL) + timeA_p->tv_nsec) -
((timeB_p->tv_sec * 1000000000ULL) + timeB_p->tv_nsec);
}
void hanning_normalized_window(float *W,size_t window_size){
//h = sqrt(2)/2 * (1+cos(t*pi)) ./ sqrt( 1+cos(t*pi).^2 )
float c;
for(size_t i=0;i<window_size;++i){
c=cos(M_PI*i/(window_size-1));
W[i]=sqrt(2.0)/2.0*(1.0+c) / sqrt(1.0+c*c);
}
}
void hanning_window(float *W,size_t window_size){
//h=.5*(1-cos(2*pi*j/(window_size+1)), COLA for R=(M+1)/2
for(size_t i=0;i<window_size;++i){
W[i]=.5*(1-cos(2*M_PI*i/(window_size+1)));
W[i]=(float).5*(1-cos(2*M_PI*i/(window_size+1)));
}
}
void hamming_window(float *W,size_t window_size){
//h=.54-.46*cos(2*pi*j/(window_size-1))
//COLA for R=(M-1)/2,(M-1)/4 etc when endpoints are divided by 2
//or one endpoint is zeroed
float m;
for(size_t i=0;i<window_size;++i){
m=i;
m/=(window_size-1);
W[i]=.54-.46*cos(2*M_PI*m);
}
W[window_size-1]=0;
//W[0]/=2;
//W[window_size-1]/=2;
}
void blackman_window(float *W,size_t window_size){
//h=.42-.5*cos(2*pi*m)+.08*cos(4*pi*m), m=(0:W-1)/(W-1)
//COLA for R=(M-1)/3 when M is odd and R is an integer
//R=M/3 when M is even and R is an integer
float m;
for(size_t i=0;i<window_size;++i){
m=i;
m/=(window_size-1);
W[i]=.42-.5*cos(2*M_PI*m)+.08*cos(4*M_PI*m);
}
}
void sin_window(float *W,size_t window_size){
//h = (cos(t*pi)+1)/2 .* float(abs(t)<1);
for(size_t i=0;i<window_size;++i){
W[i]=sin(M_PI*i/(window_size-1));
}
}
void array_out(const char *name,float *a,size_t length){
FILE *p=fopen(name,"w");
@ -211,9 +186,9 @@ static int sink_process_msg(pa_msgobject *o, int code, void *data, int64_t offse
if (PA_MSGOBJECT(u->master)->process_msg(PA_MSGOBJECT(u->master), PA_SINK_MESSAGE_GET_LATENCY, &usec, 0, NULL) < 0)
usec = 0;
usec+=pa_bytes_to_usec(u->latency*fs,ss);
//usec+=pa_bytes_to_usec(u->latency*fs,ss);
//usec+=pa_bytes_to_usec(u->samples_gathered*fs,ss);
//usec += pa_bytes_to_usec(pa_memblockq_get_length(u->rendered_q), ss);
usec += pa_bytes_to_usec(pa_memblockq_get_length(u->rendered_q), ss);
/* Add the latency internal to our sink input on top */
usec += pa_bytes_to_usec(pa_memblockq_get_length(u->sink_input->thread_info.render_memblockq), &u->master->sample_spec);
*((pa_usec_t*) data) = usec;
@ -276,7 +251,18 @@ static void process_samples(struct userdata *u){
tchunk.length=u->R*fs;
tchunk.memblock=pa_memblock_new(u->core->mempool,tchunk.length);
dst=((float*)pa_memblock_acquire(tchunk.memblock));
dsp_logic(dst,u);
for (size_t c=0;c<u->channels;c++) {
dsp_logic(
u->work_buffer,
u->input[c],
u->overlap_accum[c],
u->H,
u->W,
u->output_window,
u
);
pa_sample_clamp(PA_SAMPLE_FLOAT32NE,dst+c,fs,u->work_buffer,sizeof(float),u->R);
}
pa_memblock_release(tchunk.memblock);
pa_memblockq_push(u->rendered_q, &tchunk);
pa_memblock_unref(tchunk.memblock);
@ -284,54 +270,166 @@ static void process_samples(struct userdata *u){
}
}
static void dsp_logic(float *dst,struct userdata *u){
size_t fs=pa_frame_size(&(u->sink->sample_spec));
//use a linear-phase sliding STFT and overlap-add method (for each channel)
for (size_t c=0;c<u->channels;c++) {
//zero padd the data
memset(u->work_buffer+u->window_size,0,(u->fft_size-u->window_size)*sizeof(float));
//window the data
for(size_t j=0;j<u->window_size;++j){
u->work_buffer[j]=u->W[j]*u->input[c][j];
}
//Processing is done here!
//do fft
fftwf_execute_dft_r2c(u->forward_plan,u->work_buffer,u->output_window);
//perform filtering
for(size_t j=0;j<u->fft_size/2+1;++j){
u->output_window[j][0]*=u->H[j];
u->output_window[j][1]*=u->H[j];
}
//inverse fft
fftwf_execute_dft_c2r(u->inverse_plan,u->output_window,u->work_buffer);
////debug: tests overlaping add
////and negates ALL PREVIOUS processing
////yields a perfect reconstruction if COLA is held
//for(size_t j=0;j<u->window_size;++j){
// u->work_buffer[j]=u->W[j]*u->input[c][j];
//}
typedef float v4sf __attribute__ ((__aligned__(v_size*sizeof(float))));
typedef union float_vector {
float f[v_size];
v4sf v;
#ifdef __SSE2__
__m128 m;
#endif
} float_vector_t;
//overlap add and preserve overlap component from this window (linear phase)
for(size_t j=0;j<u->R;++j){
u->work_buffer[j]+=u->overlap_accum[c][j];
u->overlap_accum[c][j]=u->work_buffer[u->overlap_size+j];
}
////reference implementation
//void dsp_logic(
// float * __restrict__ dst,//used as a temp array too, needs to be fft_length!
// float * __restrict__ src,/*input data w/ overlap at start,
// *automatically cycled in routine
// */
// float * __restrict__ overlap,//The size of the overlap
// const float * __restrict__ H,//The freq. magnitude scalers filter
// const float * __restrict__ W,//The windowing function
// fftwf_complex * __restrict__ output_window,//The transformed window'd src
// struct userdata *u){
// //use a linear-phase sliding STFT and overlap-add method (for each channel)
// //zero padd the data
// memset(dst+u->window_size,0,(u->fft_size-u->window_size)*sizeof(float));
// //window the data
// for(size_t j=0;j<u->window_size;++j){
// dst[j]=W[j]*src[j];
// }
// //Processing is done here!
// //do fft
// fftwf_execute_dft_r2c(u->forward_plan,dst,output_window);
// //perform filtering
// for(size_t j=0;j<u->fft_size/2+1;++j){
// u->output_window[j][0]*=u->H[j];
// u->output_window[j][1]*=u->H[j];
// }
// //inverse fft
// fftwf_execute_dft_c2r(u->inverse_plan,output_window,dst);
// ////debug: tests overlaping add
// ////and negates ALL PREVIOUS processing
// ////yields a perfect reconstruction if COLA is held
// //for(size_t j=0;j<u->window_size;++j){
// // u->work_buffer[j]=u->W[j]*u->input[c][j];
// //}
//
// //overlap add and preserve overlap component from this window (linear phase)
// for(size_t j=0;j<u->overlap_size;++j){
// u->work_buffer[j]+=overlap[j];
// overlap[j]=dst[u->R+j];
// }
// ////debug: tests if basic buffering works
// ////shouldn't modify the signal AT ALL (beyond roundoff)
// //for(size_t j=0;j<u->window_size;++j){
// // u->work_buffer[j]=u->input[c][j];
// //}
//
// //preseve the needed input for the next window's overlap
// memmove(src,src+u->R,
// (u->samples_gathered+u->overlap_size-u->R)*sizeof(float)
// );
//}
////debug: tests if basic buffering works
////shouldn't modify the signal AT ALL (beyond roundoff)
//for(size_t j=0;j<u->window_size;++j){
// u->work_buffer[j]=u->input[c][j];
//}
//regardless of sse enabled, the loops in here assume
//16 byte aligned addresses and memory allocations divisible by v_size
void dsp_logic(
float * __restrict__ dst,//used as a temp array too, needs to be fft_length!
float * __restrict__ src,/*input data w/ overlap at start,
*automatically cycled in routine
*/
float * __restrict__ overlap,//The size of the overlap
const float * __restrict__ H,//The freq. magnitude scalers filter
const float * __restrict__ W,//The windowing function
fftwf_complex * __restrict__ output_window,//The transformed window'd src
struct userdata *u){//Collection of constants
//preseve the needed input for the next window's overlap
memmove(u->input[c],u->input[c]+u->R,
(u->samples_gathered+u->overlap_size-u->R)*sizeof(float)
);
//output the samples that are outputable now
pa_sample_clamp(PA_SAMPLE_FLOAT32NE,dst+c,fs,u->work_buffer,sizeof(float),u->R);
const size_t window_size=mround(u->window_size,v_size);
const size_t fft_h=mround(u->fft_size/2+1,v_size/2);
const size_t R=mround(u->R,v_size);
const size_t overlap_size=mround(u->overlap_size,v_size);
//assert(u->samples_gathered>=u->R);
//zero out the bit beyond the real overlap so we don't add garbage
for(size_t j=overlap_size;j>u->overlap_size;--j){
overlap[j-1]=0;
}
//use a linear-phase sliding STFT and overlap-add method
//zero padd the data
memset(dst+u->window_size,0,(u->fft_size-u->window_size)*sizeof(float));
//window the data
for(size_t j=0;j<window_size;j+=v_size){
//dst[j]=W[j]*src[j];
float_vector_t *d=(float_vector_t*)(dst+j);
float_vector_t *w=(float_vector_t*)(W+j);
float_vector_t *s=(float_vector_t*)(src+j);
#if __SSE2__
d->m=_mm_mul_ps(w->m,s->m);
#else
d->v=w->v*s->v;
#endif
}
//Processing is done here!
//do fft
fftwf_execute_dft_r2c(u->forward_plan,dst,output_window);
//perform filtering - purely magnitude based
for(size_t j=0;j<fft_h;j+=v_size/2){
//output_window[j][0]*=H[j];
//output_window[j][1]*=H[j];
float_vector_t *d=(float_vector_t*)(output_window+j);
float_vector_t h;
h.f[0]=h.f[1]=H[j];
h.f[2]=h.f[3]=H[j+1];
#if __SSE2__
d->m=_mm_mul_ps(d->m,h.m);
#else
d->v=d->v*h->v;
#endif
}
//inverse fft
fftwf_execute_dft_c2r(u->inverse_plan,output_window,dst);
////debug: tests overlaping add
////and negates ALL PREVIOUS processing
////yields a perfect reconstruction if COLA is held
//for(size_t j=0;j<u->window_size;++j){
// dst[j]=W[j]*src[j];
//}
//overlap add and preserve overlap component from this window (linear phase)
for(size_t j=0;j<overlap_size;j+=v_size){
//dst[j]+=overlap[j];
//overlap[j]+=dst[j+R];
float_vector_t *d=(float_vector_t*)(dst+j);
float_vector_t *o=(float_vector_t*)(overlap+j);
#if __SSE2__
d->m=_mm_add_ps(d->m,o->m);
o->m=((float_vector_t*)(dst+u->R+j))->m;
#else
d->v=d->v+o->v;
o->v=((float_vector_t*)(dst+u->R+j))->v;
#endif
}
//memcpy(overlap,dst+u->R,u->overlap_size*sizeof(float));
//////debug: tests if basic buffering works
//////shouldn't modify the signal AT ALL (beyond roundoff)
//for(size_t j=0;j<u->window_size;++j){
// dst[j]=src[j];
//}
//preseve the needed input for the next window's overlap
memmove(src,src+u->R,
(u->overlap_size+u->samples_gathered-u->R)*sizeof(float)
);
}
void input_buffer(struct userdata *u,pa_memchunk *in){
size_t fs=pa_frame_size(&(u->sink->sample_spec));
size_t samples=in->length/fs;
@ -422,31 +520,6 @@ static int sink_input_pop_cb(pa_sink_input *i, size_t nbytes, pa_memchunk *chunk
pa_assert_se(pa_memblockq_peek(u->rendered_q,&tchunk)==0);
*chunk=tchunk;
pa_memblockq_drop(u->rendered_q, chunk->length);
//if(tchunk.length>=nbytes){
//chunk->length=PA_MIN(tchunk.length,nbytes);
//}else{
// size_t copied=0;
// chunk->index=0;
// chunk->length=PA_MIN(nbytes,pa_memblockq_get_length(u->rendered_q));
// chunk->memblock=pa_memblock_new(u->core->mempool,chunk->length);
// uint8_t *dst=(uint8_t*)pa_memblock_acquire(chunk->memblock);
// for(;;){
// size_t l=PA_MIN(tchunk.length,nbytes-copied);
// pa_assert_se(l>0);
// uint8_t *src=(((uint8_t*)pa_memblock_acquire(tchunk.memblock))+tchunk.index);
// memmove(dst+copied,src,l);
// copied+=l;
// pa_memblock_release(tchunk.memblock);
// pa_memblock_unref(tchunk.memblock);
// pa_memblockq_drop(u->rendered_q,l);
// if(copied<chunk->length){
// pa_assert_se(pa_memblockq_peek(u->rendered_q,&tchunk)==0);
// }else{
// break;
// }
// }
// pa_memblock_release(chunk->memblock);
//}
pa_assert_se(chunk->memblock);
//pa_log("gave %ld",chunk->length/fs);
//pa_log("end pop");
@ -509,7 +582,8 @@ static void sink_input_update_max_request_cb(pa_sink_input *i, size_t nbytes) {
return;
size_t fs=pa_frame_size(&(u->sink->sample_spec));
pa_sink_set_max_request_within_thread(u->sink, u->R*fs);
pa_sink_set_max_request_within_thread(u->sink, nbytes);
//pa_sink_set_max_request_within_thread(u->sink, u->R*fs);
}
/* Called from I/O thread context */
@ -523,7 +597,8 @@ static void sink_input_update_sink_latency_range_cb(pa_sink_input *i) {
return;
size_t fs=pa_frame_size(&(u->sink->sample_spec));
pa_sink_set_latency_range_within_thread(u->sink,u->latency*fs ,u->latency*fs );
pa_sink_set_latency_range_within_thread(u->sink, u->master->thread_info.min_latency, u->latency*fs);
//pa_sink_set_latency_range_within_thread(u->sink,u->latency*fs ,u->latency*fs );
//pa_sink_set_latency_range_within_thread(u->sink, i->sink->thread_info.min_latency, i->sink->thread_info.max_latency);
}
@ -557,7 +632,12 @@ static void sink_input_attach_cb(pa_sink_input *i) {
pa_sink_attach_within_thread(u->sink);
size_t fs=pa_frame_size(&(u->sink->sample_spec));
pa_sink_set_latency_range_within_thread(u->sink, u->latency*fs, u->latency*fs);
//pa_sink_set_latency_range_within_thread(u->sink, u->latency*fs, u->latency*fs);
//pa_sink_set_latency_range_within_thread(u->sink,u->latency*fs, u->master->thread_info.max_latency);
//TODO: setting this guy minimizes drop outs but doesn't get rid
//of them completely, figure out why
pa_sink_set_latency_range_within_thread(u->sink, u->master->thread_info.min_latency, u->latency*fs);
//TODO: this guy causes dropouts constantly+rewinds, it's unusable
//pa_sink_set_latency_range_within_thread(u->sink, u->master->thread_info.min_latency, u->master->thread_info.max_latency);
}
@ -605,6 +685,16 @@ static pa_bool_t sink_input_may_move_to_cb(pa_sink_input *i, pa_sink *dest) {
return u->sink != dest;
}
//ensure's memory allocated is a multiple of v_size
//and aligned
static void * alloc(size_t x,size_t s){
size_t f=mround(x*s,sizeof(float)*v_size);
//printf("requested %ld floats=%ld bytes, rem=%ld\n",x,x*sizeof(float),x*sizeof(float)%16);
//printf("giving %ld floats=%ld bytes, rem=%ld\n",f,f*sizeof(float),f*sizeof(float)%16);
return fftwf_malloc(f*s);
}
int pa__init(pa_module*m) {
struct userdata *u;
pa_sample_spec ss;
@ -649,7 +739,7 @@ int pa__init(pa_module*m) {
u->channels=ss.channels;
u->fft_size=pow(2,ceil(log(ss.rate)/log(2)));
pa_log("fft size: %ld",u->fft_size);
u->window_size=7999;
u->window_size=15999;
u->R=(u->window_size+1)/2;
u->overlap_size=u->window_size-u->R;
u->target_samples=1*u->R;
@ -659,32 +749,28 @@ int pa__init(pa_module*m) {
u->conv_buffer.memblock=pa_memblock_new(u->core->mempool,u->target_samples*fs);
u->latency=u->R;
u->H=(float*) fftwf_malloc((u->fft_size/2+1)*sizeof(float));
u->W=(float*) fftwf_malloc((u->window_size)*sizeof(float));
u->work_buffer=(float*) fftwf_malloc(u->fft_size*sizeof(float));
u->H=alloc((u->fft_size/2+1),sizeof(fftwf_complex));
u->W=alloc(u->window_size,sizeof(float));
u->work_buffer=alloc(u->fft_size,sizeof(float));
memset(u->work_buffer,0,u->fft_size*sizeof(float));
u->input=(float **)malloc(sizeof(float *)*u->channels);
u->overlap_accum=(float **)malloc(sizeof(float *)*u->channels);
u->output_buffer=(float **)malloc(sizeof(float *)*u->channels);
for(size_t c=0;c<u->channels;++c){
u->input[c]=(float*) fftwf_malloc((u->target_samples+u->overlap_size)*sizeof(float));
u->input[c]=alloc(u->target_samples+u->overlap_size,sizeof(float));
pa_assert_se(u->input[c]);
memset(u->input[c],0,(u->target_samples+u->overlap_size)*sizeof(float));
pa_assert_se(u->input[c]);
u->overlap_accum[c]=(float*) fftwf_malloc(u->R*sizeof(float));
u->overlap_accum[c]=alloc(u->overlap_size,sizeof(float));
pa_assert_se(u->overlap_accum[c]);
memset(u->overlap_accum[c],0,u->R*sizeof(float));
u->output_buffer[c]=(float*) fftwf_malloc(u->window_size*sizeof(float));
memset(u->overlap_accum[c],0,u->overlap_size*sizeof(float));
u->output_buffer[c]=alloc(u->window_size,sizeof(float));
pa_assert_se(u->output_buffer[c]);
}
u->output_window = (fftwf_complex *) fftwf_malloc(sizeof(fftwf_complex) * (u->fft_size/2+1));
u->output_window=alloc((u->fft_size/2+1),sizeof(fftwf_complex));
u->forward_plan=fftwf_plan_dft_r2c_1d(u->fft_size, u->work_buffer, u->output_window, FFTW_MEASURE);
u->inverse_plan=fftwf_plan_dft_c2r_1d(u->fft_size, u->output_window, u->work_buffer, FFTW_MEASURE);
/*
for(size_t j=0;j<u->window_size;++j){
u->W[j]=.5;
}
*/
hanning_window(u->W,u->window_size);
const int freqs[]={0,25,50,100,200,300,400,800,1500,
@ -735,6 +821,7 @@ int pa__init(pa_module*m) {
}
free(freq_translated);
/* Create sink */
pa_sink_new_data_init(&sink_data);
sink_data.driver = __FILE__;
@ -857,18 +944,18 @@ void pa__done(pa_module*m) {
fftwf_destroy_plan(u->inverse_plan);
fftwf_destroy_plan(u->forward_plan);
fftwf_free(u->output_window);
free(u->output_window);
for(size_t c=0;c<u->channels;++c){
fftwf_free(u->output_buffer[c]);
fftwf_free(u->overlap_accum[c]);
fftwf_free(u->input[c]);
free(u->output_buffer[c]);
free(u->overlap_accum[c]);
free(u->input[c]);
}
free(u->output_buffer);
free(u->overlap_accum);
free(u->input);
fftwf_free(u->work_buffer);
fftwf_free(u->W);
fftwf_free(u->H);
free(u->work_buffer);
free(u->W);
free(u->H);
pa_xfree(u);
}