Add embedded-oriented integer version of DFT.

This commit is contained in:
cnlohr 2015-02-10 00:57:10 -05:00
parent 7148254c41
commit 2e26b747fd
10 changed files with 333 additions and 25 deletions

View file

@ -5,7 +5,8 @@ SOUND:=sound.o sound_alsa.o sound_pulse.o sound_null.o
OUTS := OutputVoronoi.o DisplayArray.o OutputLinear.o DisplayPie.o DisplayNetwork.o DisplayUSB2812.o DisplayDMX.o
WINGCC:=i586-mingw32msvc-gcc
WINGCCFLAGS:= -lwinmm -lgdi32 -lws2_32 -O2 -ffast-math -g
WINGCCFLAGS:= -O2 -ffast-math -Wl,--relax -Wl,--gc-sections -ffunction-sections -fdata-sections -s
WINLDFLAGS:=-lwinmm -lgdi32 -lws2_32
RAWDRAWLIBS:=-lX11 -lm -lpthread -lXinerama -lXext
LDLIBS:=-lpthread -lasound -lm -lpulse-simple -lpulse
@ -16,8 +17,8 @@ colorchord : os_generic.o main.o dft.o decompose.o filter.o color.o sort.o note
gcc -o $@ $^ $(CFLAGS) $(LDLIBS) $(EXTRALIBS) $(RAWDRAWLIBS)
colorchord.exe : os_generic.c main.c dft.c decompose.c filter.c color.c sort.c notefinder.c util.c outdrivers.c DrawFunctions.c parameters.c chash.c WinDriver.c sound.c sound_null.c sound_win.c OutputVoronoi.c DisplayArray.c OutputLinear.c DisplayPie.c DisplayNetwork.c
$(WINGCC) -o $@ $^ $(WINGCCFLAGS)
$(WINGCC) $(WINGCCFLAGS) -o $@ $^ $(WINLDFLAGS)
clean :
rm -rf *.o *~ colorchord
rm -rf *.o *~ colorchord colorchord.exe

View file

@ -5,6 +5,7 @@
#include "DrawFunctions.h"
#include <windows.h>
#include <stdlib.h>
#include <stdio.h>
#include <malloc.h> //for alloca
static HINSTANCE lhInstance;
@ -193,6 +194,8 @@ void CNFGSetup( const char * name_of_window, int width, int height )
InternalHandleResize();
}
void WindowsTerm();
void CNFGHandleInput()
{
int ldown = 0;
@ -204,6 +207,13 @@ void CNFGHandleInput()
switch( msg.message )
{
case WM_QUIT:
case WM_DESTROY:
case WM_CLOSE:
printf( "Close\n" );
WindowsTerm();
TerminateProcess( 0, 0 );
break;
case WM_MOUSEMOVE:
HandleMotion( (msg.lParam & 0xFFFF), (msg.lParam>>16) & 0xFFFF, ( (msg.wParam & 0x01)?1:0) | ((msg.wParam & 0x02)?2:0) | ((msg.wParam & 0x10)?4:0) );
break;

Binary file not shown.

255
dft.c
View file

@ -1,7 +1,10 @@
#include "dft.h"
#include <math.h>
#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
void DoDFT( float * outbins, float * frequencies, int bins, float * databuffer, int place_in_data_buffer, int size_of_data_buffer, float q )
{
@ -41,6 +44,8 @@ void DoDFTQuick( float * outbins, float * frequencies, int bins, const float * d
for( i = 0; i < bins; i++ )
{
int flirts = 0;
float freq = frequencies[i];
float phi = 0;
int ftq = freq * q;
@ -66,6 +71,7 @@ void DoDFTQuick( float * outbins, float * frequencies, int bins, const float * d
binqtyc += cv;
phi += advance;
flirts++;
}
float amp = sqrtf( binqtys * binqtys + binqtyc * binqtyc );
@ -73,3 +79,250 @@ void DoDFTQuick( float * outbins, float * frequencies, int bins, const float * d
}
}
////////////////////////////DFT Progressive is for embedded systems, primarily.
static float * gbinqtys;
static float * gbinqtyc;
static float * phis;
static float * gfrequencies;
static float * goutbins;
static float * lastbins;
static float * advances;
static int gbins;
static float gq;
static float gspeedup;
#define PROGIIR .005
void HandleProgressive( float sample )
{
int i;
for( i = 0; i < gbins; i++ )
{
float thiss = sinf( phis[i] ) * sample;
float thisc = cosf( phis[i] ) * sample;
float s = gbinqtys[i] = gbinqtys[i] * (1.-PROGIIR) + thiss * PROGIIR;
float c = gbinqtyc[i] = gbinqtyc[i] * (1.-PROGIIR) + thisc * PROGIIR;
phis[i] += advances[i];
if( phis[i] > 6.283 ) phis[i]-=6.283;
goutbins[i] = sqrtf( s * s + c * c );
}
}
void DoDFTProgressive( float * outbins, float * frequencies, int bins, const float * databuffer, int place_in_data_buffer, int size_of_data_buffer, float q, float speedup )
{
int i;
static int last_place;
if( gbins != bins )
{
if( gbinqtys ) free( gbinqtys );
if( gbinqtyc ) free( gbinqtyc );
if( phis ) free( phis );
if( lastbins ) free( lastbins );
if( advances ) free( advances );
gbinqtys = malloc( sizeof(float)*bins );
gbinqtyc = malloc( sizeof(float)*bins );
phis = malloc( sizeof(float)*bins );
lastbins = malloc( sizeof(float)*bins );
advances = malloc( sizeof(float)*bins );
memset( gbinqtys, 0, sizeof(float)*bins );
memset( gbinqtyc, 0, sizeof(float)*bins );
memset( phis, 0, sizeof(float)*bins );
memset( lastbins, 0, sizeof(float)*bins );
}
memcpy( outbins, lastbins, sizeof(float)*bins );
for( i = 0; i < bins; i++ )
{
float freq = frequencies[i];
advances[i] = 3.14159*2.0/freq;
}
gbins = bins;
gfrequencies = frequencies;
goutbins = outbins;
gspeedup = speedup;
gq = q;
place_in_data_buffer = (place_in_data_buffer+1)%size_of_data_buffer;
int didrun = 0;
for( i = last_place; i != place_in_data_buffer; i = (i+1)%size_of_data_buffer )
{
float fin = ((float)((int)(databuffer[i] * 127))) / 127.0; //simulate 8-bit input (it looks FINE!)
HandleProgressive( fin );
didrun = 1;
}
last_place = place_in_data_buffer;
if( didrun )
{
memcpy( lastbins, outbins, sizeof(float)*bins );
}
/* for( i = 0; i < bins; i++ )
{
printf( "%0.2f ", outbins[i]*100 );
}
printf( "\n" );*/
}
/////////////////////////////INTEGER DFT
#define PROGIIR .005
//NOTES to self:
//
// Let's say we want to try this on an AVR.
// 24 bins, 5 octaves = 120 bins.
// 20 MHz clock / 4.8k sps = 4096 IPS = 34 clocks per bin = :(
// We can do two at the same time, this frees us up some
static uint8_t donefirstrun;
static int8_t sintable[512]; //Actually [sin][cos] pairs.
//LDD instruction on AVR can read with constant offset. We can set Y to be the place in the buffer, and read with offset.
static uint16_t * datspace; //(advances,places,isses,icses)
//
void HandleProgressiveInt( int8_t sample1, int8_t sample2 )
{
int i;
uint16_t startpl = 0;
int16_t ts, tc;
int16_t tmp1;
int8_t s1, c1;
uint16_t ipl, localipl, adv;
//startpl maps to 'Y'
//
//Estimated 68 minimum instructions... So for two pairs each... just under 5ksps, theoretical.
//Running overall at ~2kHz.
for( i = 0; i < gbins; i++ ) //Loop, fixed size = 3 + 2 cycles 5
{
//12 cycles MIN
adv = datspace[startpl++]; //Read, indirect from RAM (and increment) 2+2 cycles 4
ipl = datspace[startpl++]; //Read, indirect from RAM (and increment) 2+2 cycles 4
//13 cycles MIN
ipl += adv; //Advance, 16bit += 16bit, 1 + 1 cycles 2
localipl = (ipl>>8)<<1; //Select upper 8 bits 1 cycles 1
// need to load Z with 'sintable' and add localipl 2
s1 = sintable[localipl++]; //Read s1 component out of table. 2+2 cycles 2
c1 = sintable[localipl++]; //Read c1 component out of table. 2 cycles 2
ts = (s1 * sample1); // 8 x 8 multiply signed + copy R1 out. zero MSB ts 2
tc = (c1 * sample1); // 8 x 8 multiply signed + copy R1 out. zero MSB tc 2
//15 cycles MIN
ipl += adv; //Advance, 16bit += 16bit, 1 + 1 cycles 2
localipl = (ipl>>8)<<1; //Select upper 8 bits 1 cycles 1
// need to load Z with 'sintable' and add localipl 2
s1 = sintable[localipl++]; //Read s1 component out of table. 2 cycles 2
c1 = sintable[localipl++]; //Read c1 component out of table. 2 cycles 2
ts += (s1 * sample2); // 8 x 8 multiply signed + add R1 out. 3
tc += (c1 * sample2); // 8 x 8 multiply signed + add R1 out. 3
//Add TS and TC to the datspace stuff. (24 instructions)
tmp1 = datspace[startpl]; //Read out, sin component. 4
tmp1 -= tmp1>>6; //Subtract from the MSB (with carry) 2
tmp1 += ts>>6; //Add MSBs with carry 2
datspace[startpl++] = tmp1; //Store values back 4
tmp1 = datspace[startpl]; //Read out, sin component. 4
tmp1 -= tmp1>>6; //Subtract from the MSB (with carry) 2
tmp1 += tc>>6; //Add MSBs with carry 2
datspace[startpl++] = tmp1; //Store values back 4
datspace[startpl-3] = ipl; //Store values back 4
}
}
void DoDFTProgressiveInteger( float * outbins, float * frequencies, int bins, const float * databuffer, int place_in_data_buffer, int size_of_data_buffer, float q, float speedup )
{
int i;
static int last_place;
if( !donefirstrun )
{
donefirstrun = 1;
for( i = 0; i < 256; i++ )
{
sintable[i*2+0] = (int8_t)((sinf( i / 256.0 * 6.283 ) * 127.0));
sintable[i*2+1] = (int8_t)((cosf( i / 256.0 * 6.283 ) * 127.0));
}
}
if( gbins != bins )
{
gbins = bins;
if( datspace ) free( datspace );
datspace = malloc( bins * 2 * 4 );
}
for( i = 0; i < bins; i++ )
{
float freq = frequencies[i];
datspace[i*4] = 65536.0/freq;
}
for( i = last_place; i != ( place_in_data_buffer&0xffffe ); i = (i+2)%size_of_data_buffer )
{
int8_t ifr1 = (int8_t)( ((databuffer[i+0]) ) * 127 );
int8_t ifr2 = (int8_t)( ((databuffer[i+1]) ) * 127 );
// printf( "%d %d\n", i, place_in_data_buffer&0xffffe );
HandleProgressiveInt( ifr1, ifr2 );
}
last_place = place_in_data_buffer&0xfffe;
//Extract bins.
for( i = 0; i < bins; i++ )
{
int16_t isps = datspace[i*4+2];
int16_t ispc = datspace[i*4+3];
int16_t mux = ( (isps/256) * (isps/256)) + ((ispc/256) * (ispc/256));
// printf( "%d (%d %d)\n", mux, isps, ispc );
outbins[i] = sqrt( mux )/100.0;
}
// printf( "\n");
}

7
dft.h
View file

@ -12,5 +12,12 @@ void DoDFT( float * outbins, float * frequencies, int bins, float * databuffer,
//Speedup = target number of data points
void DoDFTQuick( float * outbins, float * frequencies, int bins, const float * databuffer, int place_in_data_buffer, int size_of_data_buffer, float q, float speedup );
//An unusual tool to do a "progressive" DFT, using data from previous rounds.
void DoDFTProgressive( float * outbins, float * frequencies, int bins, const float * databuffer, int place_in_data_buffer, int size_of_data_buffer, float q, float speedup );
//A progressive DFT that's done using only low-bit integer math.
void DoDFTProgressiveInteger( float * outbins, float * frequencies, int bins, const float * databuffer, int place_in_data_buffer, int size_of_data_buffer, float q, float speedup );
#endif

5
integerprog.conf Normal file
View file

@ -0,0 +1,5 @@
do_progressive_dft = 1
samplerate = 8000
buffer = 128

9
main.c
View file

@ -14,8 +14,15 @@
#include "outdrivers.h"
#include "parameters.h"
struct SoundDriver * sd;
#ifdef WIN32
#include <windows.h>
void WindowsTerm()
{
CloseSound( sd );
}
#endif
int lastfps;
@ -244,7 +251,7 @@ int main(int argc, char ** argv)
//Initialize Sound
struct SoundDriver * sd = InitSound( sound_source, &SoundCB );
sd = InitSound( sound_source, &SoundCB );
if( !sd )
{

View file

@ -25,6 +25,7 @@ struct NoteFinder * CreateNoteFinder( int spsRec )
ret->decompose_iterations = 1000;
ret->dft_speedup = 300;
ret->dft_q = 16;
ret->do_progressive_dft = 0;
ret->default_sigma = 1.4;
ret->note_jumpability = 2.5;
ret->note_combine_distance = 0.5;
@ -58,6 +59,7 @@ struct NoteFinder * CreateNoteFinder( int spsRec )
RegisterValue( "note_minimum_new_distribution_value", PAFLOAT, &ret->note_minimum_new_distribution_value, sizeof( ret->note_minimum_new_distribution_value ) );
RegisterValue( "note_out_chop", PAFLOAT, &ret->note_out_chop, sizeof( ret->note_out_chop ) );
RegisterValue( "dft_iir", PAFLOAT, &ret->dft_iir, sizeof( ret->dft_iir ) );
RegisterValue( "do_progressive_dft", PAINT, &ret->do_progressive_dft, sizeof( ret->do_progressive_dft ) );
AddCallback( "freqbins", ChangeNFParameters, ret );
AddCallback( "octaves", ChangeNFParameters, ret );
@ -175,7 +177,20 @@ void RunNoteFinder( struct NoteFinder * nf, const float * audio_stream, int head
//This DFT function does not wavelet or anything.
nf->StartTime = OGGetAbsoluteTime();
DoDFTQuick( dftbins, nf->frequencies, freqs, audio_stream, head, buffersize, nf->dft_q, nf->dft_speedup );
switch( nf->do_progressive_dft )
{
case 0:
DoDFTQuick( dftbins, nf->frequencies, freqs, audio_stream, head, buffersize, nf->dft_q, nf->dft_speedup );
break;
case 1:
DoDFTProgressive( dftbins, nf->frequencies, freqs, audio_stream, head, buffersize, nf->dft_q, nf->dft_speedup );
break;
case 2:
DoDFTProgressiveInteger( dftbins, nf->frequencies, freqs, audio_stream, head, buffersize, nf->dft_q, nf->dft_speedup );
break;
default:
fprintf( stderr, "Error: No DFT Seleced\n" );
}
nf->DFTTime = OGGetAbsoluteTime();
for( i = 0; i < freqs; i++ )

View file

@ -16,6 +16,7 @@ struct NoteFinder
int filter_iter;// = 1;
int decompose_iterations;// = 1000;
float amplify; // =1 (amplify input across the board)
int do_progressive_dft; //= 1
//at 300, there is still some minimal aliasing at higher frequencies. Increase this for less low-end distortion
//NOTE: This /should/ get fixed, as we /should/ be decimating the input data intelligently with lower octaves.

View file

@ -1,5 +1,3 @@
//XXX THIS DRIVER IS INCOMPLETE XXX
#include <windows.h>
#include "parameters.h"
#include "sound.h"
@ -12,7 +10,7 @@
#pragma comment(lib,"winmm.lib")
#endif
#define BUFFS 3
#define BUFFS 2
struct SoundDriverWin
{
@ -27,9 +25,7 @@ struct SoundDriverWin
int buffer;
int isEnding;
int Cbuff;
int GOBUFF;
int OLDBUFF;
int recording;
@ -42,6 +38,7 @@ static struct SoundDriverWin * w;
void CloseSoundWin( struct SoundDriverWin * r )
{
int i;
if( r )
{
waveInStop(r->hMyWave);
@ -65,6 +62,7 @@ int SoundStateWin( struct SoundDriverWin * soundobject )
void CALLBACK HANDLEMIC(HWAVEIN hwi,UINT umsg, DWORD dwi, DWORD hdr, DWORD dwparm)
{
int ctr;
int ob;
long cValue;
unsigned int maxWave=0;
@ -80,18 +78,20 @@ void CALLBACK HANDLEMIC(HWAVEIN hwi,UINT umsg, DWORD dwi, DWORD hdr, DWORD dwpar
break;
case MM_WIM_DATA:
w->OLDBUFF=w->GOBUFF;
w->GOBUFF=w->Cbuff;
w->Cbuff = (w->Cbuff+1)%3;
waveInPrepareHeader(w->hMyWave,&(w->WavBuff[w->Cbuff]),sizeof(WAVEHDR));
waveInAddBuffer(w->hMyWave,&(w->WavBuff[w->Cbuff]),sizeof(WAVEHDR));
// printf( "Mic Data.\n");
ob = (w->GOBUFF+(BUFFS))%BUFFS;
// waveInPrepareHeader(w->hMyWave,&(w->WavBuff[w->Cbuff]),sizeof(WAVEHDR));
for (ctr=0;ctr<w->buffer * w->channelsRec;ctr++) {
float cv = (uint16_t)(((uint8_t)w->WavBuff[w->GOBUFF].lpData[ctr*2+1])*256+((uint8_t)w->WavBuff[w->GOBUFF].lpData[ctr*2])+32768)-32768;
float cv = (uint16_t)(((uint8_t)w->WavBuff[ob].lpData[ctr*2+1])*256+((uint8_t)w->WavBuff[ob].lpData[ctr*2])+32768)-32768;
cv /= 32768;
// if( ctr < 3 ) cv = -1;
// buffer[(w->buffer * w->channelsRec)-ctr-1] = cv;
buffer[ctr] = cv;
}
waveInAddBuffer(w->hMyWave,&(w->WavBuff[w->GOBUFF]),sizeof(WAVEHDR));
w->GOBUFF = ( w->GOBUFF + 1 ) % BUFFS;
int playbacksamples; //Unused
w->callback( 0, buffer, w->buffer, &playbacksamples, (struct SoundDriver*)w );
@ -113,23 +113,34 @@ static struct SoundDriverWin * InitWinSound( struct SoundDriverWin * r )
w = r;
printf( "WFMT: %d %d %d\n", r->channelsRec, r->spsRec,
r->spsRec * r->channelsRec );
wfmt.wFormatTag = WAVE_FORMAT_PCM;
wfmt.nChannels = r->channelsRec;
wfmt.nSamplesPerSec = r->spsRec;
wfmt.nBlockAlign = 1;
wfmt.nAvgBytesPerSec = r->spsRec * r->channelsRec;
wfmt.nBlockAlign = r->channelsRec * 2;
wfmt.wBitsPerSample = 16;
wfmt.nAvgBytesPerSec = 0;
wfmt.cbSize = 0;
long dwdevice;
dwdevice = GetParameterI( "wininput", WAVE_MAPPER );
int p = waveInOpen(&r->hMyWave, dwdevice, &wfmt,(DWORD)(void*)(HANDLEMIC) , 0, CALLBACK_FUNCTION);
printf( "Wave Devs: %d; WAVE_MAPPER: %d; Selected Input: %d\n", waveInGetNumDevs(), WAVE_MAPPER, dwdevice );
printf( "WIO: %d\n", p );
printf( "waveInOpen: %p, %p\n", &r->hMyWave, &wfmt );
int p = waveInOpen(&r->hMyWave, dwdevice, &wfmt,(DWORD)(void*)(&HANDLEMIC) , 0, CALLBACK_FUNCTION);
printf( "WIO: %d\n", p ); //On real windows, returns 11
for ( i=0;i<BUFFS;i++)
{
memset( &(r->WavBuff[i]), 0, sizeof(r->WavBuff[i]) );
(r->WavBuff[i]).dwBufferLength = r->buffer*2*r->channelsRec;
(r->WavBuff[i]).dwLoops = 1;
(r->WavBuff[i]).lpData=(char*) malloc(r->buffer*r->channelsRec*2);
waveInPrepareHeader(r->hMyWave,&(r->WavBuff[i]),sizeof(WAVEHDR));
waveInAddBuffer(r->hMyWave,&(r->WavBuff[i]),sizeof(WAVEHDR));
@ -137,7 +148,7 @@ static struct SoundDriverWin * InitWinSound( struct SoundDriverWin * r )
p = waveInStart(r->hMyWave);
printf( "WIS: %d\n", p );
printf( "WIS: %d\n", p ); //On real windows returns 5.
return r;
}
@ -154,14 +165,12 @@ void * InitSoundWin( SoundCBType cb )
r->spsRec = GetParameterI( "samplerate", 44100 );
r->channelsRec = GetParameterI( "channels", 2 );
r->buffer = GetParameterI( "buffer", 1024 );
r->buffer = GetParameterI( "buffer", 384 );
r->recording = 0;
r->isEnding = 0;
printf( "Buffer: %d\n", r->buffer );
r->Cbuff=0;
r->GOBUFF=0;
r->OLDBUFF=0;
return InitWinSound(r);
}