colorchord/embeddedcommon/DFT8Turbo.c

355 lines
10 KiB
C

#include <stdint.h>
#include <stdlib.h>
#include "DFT8Turbo.h"
#include <math.h>
#include <stdio.h>
#define MAX_FREQS (12)
#define OCTAVES (4)
#define INITIAL_DECIMATE 1
//Right now, we need 8*freqs*octaves bytes.
//This is bad.
//What can we do to fix it?
//4x the hits (sin/cos and we need to do it once for each edge)
//8x for selecting a higher octave.
#define FREQREBASE 8.0
#define TARGFREQ 10000.0
/* Tradeoff guide:
* We will optimize for RAM size here.
* INITIAL_DECIMATE; A larger decimation: {NOTE 1}
+) Reduces the bit depth needed for the integral map.
If you use "1" and a fully saturted map (highest note is every sample), it will not overflow a signed 12-bit number.
-) Increases noise.
With full-scale: 0->1 minimal 1->2 minimal 2->3 significantly noticable, 3->4 major.
If sound is quieter, it matters more. I recommend no less than 1.
Also, other things, like frequency of hits can manipulate the maximum bit depth needed for integral map.
* If you weight the bins in advance see "mulmux", you can: {NOTE 2}
+) potentially use shallower bit depth but
-) have to compute the multiply every time you update the bin.
* You can use a modified-square-wave which only integrates for 1/2 of the duty cycle. {NOTE 3}
+) uses 1/2 the integral memory.
-) Not as pretty of an output. See "integral_at"
*TODO: Investigate using all unsigned (to make multiply and/or 12-bit storage easier)
So, the idea here is we would keep a running total of the current ADC value, kept away in a int16_t.
It is constantly summing, so we can take an integral of it. Or rather an integral range.
Over time, we perform operations like adding or subtracting from a current place. It basically is
a DFT where the kernel is computed using square waves (or modified square waves)
*/
//These live in RAM.
int16_t running_integral; //Realistically treat as 12-bits on ramjet8
int16_t integral_at[MAX_FREQS*OCTAVES]; //For ramjet8, make 12-bits
int32_t cossindata[MAX_FREQS*OCTAVES*2]; //Contains COS and SIN data. (32-bit for now, will be 16-bit, potentially even 8.)
uint8_t which_octave_for_op[MAX_FREQS]; //counts up, tells you which ocative you are operating on. PUT IN RAM.
#define NR_OF_OPS (4<<OCTAVES)
//Format is:
// 255 = DO NOT OPERATE
// bits 0..3 unfolded octave, i.e. sin/cos are offset by one.
// bit 4 = add or subtract.
uint8_t optable[NR_OF_OPS]; //PUT IN FLASH
#define ACTIONTABLESIZE 256
uint16_t actiontable[ACTIONTABLESIZE]; //PUT IN FLASH // If there are more than 8 freqbins, this must be a uint16_t, otherwise if more than 16, 32.
uint8_t actiontableplace;
//Format is
uint8_t mulmux[MAX_FREQS*OCTAVES]; //PUT IN FLASH
static int Setup( float * frequencies, int bins )
{
int i;
printf( "BINS: %d\n", bins );
float highestf = frequencies[bins-1];
for( i = 0; i < bins; i++ )
{
mulmux[i] = (uint8_t)( highestf / frequencies[i] * 255 + 0.5 );
printf( "MM: %d %f / %f\n", mulmux[i], frequencies[i], highestf );
}
for( i = bins-MAX_FREQS; i < bins; i++ )
{
int topbin = i - (bins-MAX_FREQS);
float f = frequencies[i]/FREQREBASE;
float hits_per_table = (float)ACTIONTABLESIZE/f;
int dhrpertable = (int)(hits_per_table+.5);//TRICKY: You might think you need to have even number of hits (sin/cos), but you don't! It can flip sin/cos each time through the table!
float err = (TARGFREQ/((float)ACTIONTABLESIZE/dhrpertable) - (float)TARGFREQ/f)/((float)TARGFREQ/f);
//Perform an op every X samples. How well does this map into units of 1024?
printf( "%d %f -> hits per %d: %f %d (%.2f%% error)\n", topbin, f, ACTIONTABLESIZE, (float)ACTIONTABLESIZE/f, dhrpertable, err * 100.0 );
if( dhrpertable >= ACTIONTABLESIZE )
{
fprintf( stderr, "Error: Too many hits.\n" );
exit(0);
}
float advance_per_step = dhrpertable/(float)ACTIONTABLESIZE;
float fvadv = 0.5;
int j;
int countset = 0;
//Tricky: We need to start fadv off at such a place that there won't be a hicchup when going back around to 0.
// I believe this is done by setting fvadv to 0.5 initially. Unsure.
for( j = 0; j < ACTIONTABLESIZE; j++ )
{
if( fvadv >= 0.5 )
{
actiontable[j] |= 1<<topbin;
fvadv -= 1.0;
countset++;
}
fvadv += advance_per_step;
}
printf( " countset: %d\n", countset );
}
//exit(1);
int phaseinop[OCTAVES] = { 0 };
int already_hit_octaveplace[OCTAVES*2] = { 0 };
for( i = 0; i < NR_OF_OPS; i++ )
{
int longestzeroes = 0;
int val = i & ((1<<OCTAVES)-1);
for( longestzeroes = 0; longestzeroes < 255 && ( ((val >> longestzeroes) & 1) == 0 ); longestzeroes++ );
//longestzeroes goes: 255, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, ...
//This isn't great, because we need to also know whether we are attacking the SIN side or the COS side, and if it's + or -.
//We can actually decide that out.
if( longestzeroes == 255 )
{
//This is a nop. Emit a nop.
optable[i] = 255;
}
else
{
longestzeroes = OCTAVES-1-longestzeroes; //Actually do octave 0 least often.
int iop = phaseinop[longestzeroes]++;
int toop = (longestzeroes<<1) | (iop & 1);
//if it's the first time an octave happened this set, flag it. This may be used later in the process.
if( !already_hit_octaveplace[toop] )
{
already_hit_octaveplace[toop] = 1;
toop |= 1<<5;
}
//Handle add/subtract bit.
if( iop & 2 ) toop |= 1<<4;
optable[i] = toop;
//printf( " %d %d %d\n", iop, val, longestzeroes );
}
//printf( "HBT: %d = %d\n", i, optable[i] );
}
//exit(1);
return 0;
}
#if 0
int16_t running_integral;
int16_t integral_at[MAX_FREQS*OCTAVES];
int16_t cossindata[MAX_FREQS*OCTAVES*2]; //Contains COS and SIN data.
uint8_t which_octave_for_op[MAX_FREQS]; //counts up, tells you which ocative you are operating on. PUT IN RAM.
#define NR_OF_OPS (4<<OCTAVES)
//Format is:
// 255 = DO NOT OPERATE
// bits 0..3 unfolded octave, i.e. sin/cos are offset by one.
// bit 4 = add or subtract.
uint8_t optable[NR_OF_OPS]; //PUT IN FLASH
#define ACTIONTABLESIZE 256
uint32_t actiontable[ACTIONTABLESIZE]; //PUT IN FLASH
//Format is
#endif
void Turbo8BitRun( int8_t adcval )
{
running_integral += adcval>>INITIAL_DECIMATE;
#define dprintf( ... )
uint32_t action = actiontable[actiontableplace++];
int n;
dprintf( "%4d ", actiontableplace );
for( n = 0; n < MAX_FREQS; n++ )
{
if( action & (1<<n) )
{
int ao = which_octave_for_op[n];
int op = optable[ao];
ao++;
if( ao >= NR_OF_OPS ) ao = 0;
which_octave_for_op[n] = ao;
if( op == 255 )
{
dprintf( "*" ); //NOP
}
else
{
int octaveplace = op & 0xf;
//Tricky: We share the integral with SIN and COS.
//We don't need to. It would produce a slightly cleaner signal. See: NOTE 3
int intindex = (octaveplace>>1) * MAX_FREQS + n;
//int invoct = OCTAVES-1-octaveplace;
int16_t diff;
if( op & 0x10 ) //ADD
{
diff = integral_at[intindex] - running_integral;
dprintf( "%c", 'a' + octaveplace );
}
else //SUBTRACT
{
diff = running_integral - integral_at[intindex];
dprintf( "%c", 'A' + octaveplace );
}
if( diff > 2000 || diff < -2000 ) printf( "!!!!!!!!!!!! %d !!!!!!!!!!!\n", diff );
integral_at[intindex] = running_integral;
int idx = intindex * 2 + (octaveplace&1);
//if( n == 1 ) printf( "%d %d %d %d\n", n, idx, diff, op & 0x10 );
//dprintf( "%d\n", idx );
#if 0
//Apply IIR operation 1; This is rough because the Q changes and goes higher as a function of frequency. This is probably a bad move.
cossindata[idx] += diff>>4;
if( op & 0x20 )
{
cossindata[idx] = cossindata[idx]
- (cossindata[idx]>>2);
}
#else
//Apply IIR.
//printf( "%d: %d + %d * %d >> 8 - %d\n", idx, cossindata[idx], diff, mulmux[idx/2], cossindata[idx]>>4 );
cossindata[idx] = cossindata[idx]
+ (((int32_t)diff * (int32_t)mulmux[idx/2])>>6)
- (cossindata[idx]>>4)
;
// if( cossindata[idx] > 2047 ) cossindata[idx] = 2047;
// if( cossindata[idx] < -2048 ) cossindata[idx] = -2048;
#endif
// if( cossindata[idx] > 1 ) cossindata[idx]--;
// if( cossindata[idx] < -1 ) cossindata[idx]++;
// if( cossindata[idx] > 16 ) cossindata[idx]-=8;
// if( cossindata[idx] < -16 ) cossindata[idx]+=8;
}
}
else
{
dprintf( " " );
}
}
dprintf( "\n" );
#if 0
uint32_t actions = *(placeintable++);
if( placeintable == &actiontable[ACTIONTABLESIZE] ) placeintable = actiontable;
int b;
for( b = 0; b < MAX_FREQS; b++ )
{
if( ! ((1<<b) & actions) ) continue;
//If we get here, we need to do an action.
int op = which_octave_for_op[b]++;
int sinorcos = op & 1;
op >>= 1;
int octavebit = op & ((1<<OCTAVES)-1);
if( !octavebit ) { continue; } //XXX TRICKY: In our octavebit table, we have 1 0 and 1 1 entry. 2, 3, 4, etc. are ok. So, if we hit a 0, we abort.
int whichoctave = highbit_table[octavebit];
//Ok, actually we need to also know whether you're on SIN or COS.
//if( b == 0 ) printf( "%d\n", whichoctave );
//XXX TODO Optimization: Use a table, since octavebit can only be 0...31.
}
#endif
}
void DoDFT8BitTurbo( float * outbins, float * frequencies, int bins, const float * databuffer, int place_in_data_buffer, int size_of_data_buffer, float q, float speedup )
{
static int is_setup;
if( !is_setup ) { is_setup = 1; Setup( frequencies, bins ); }
static int last_place;
int i;
for( i = last_place; i != place_in_data_buffer; i = (i+1)%size_of_data_buffer )
{
int16_t ifr1 = (int16_t)( ((databuffer[i]) ) * 4095 );
Turbo8BitRun( ifr1>>5 ); //6 = Actually only feed algorithm numbers from -64 to 63.
}
last_place = place_in_data_buffer;
static int idiv;
idiv++;
#if 1
for( i = 0; i < bins; i++ )
{
outbins[i] = 0;
}
for( i = 0; i < bins; i++ )
{
int iss = cossindata[i*2+0]>>8;
int isc = cossindata[i*2+1]>>8;
int issdiv = 0;
int iscdiv = 0;
int FWDOFFSET = 19;//MAX_FREQS*3/2;
if( i < bins-FWDOFFSET )
{
issdiv = cossindata[(i+FWDOFFSET)*2+0]/256;
iscdiv = cossindata[(i+FWDOFFSET)*2+1]/256;
}
int mux = iss * iss + isc * isc;
int muxdiv = issdiv * issdiv + iscdiv * iscdiv;
//if( (idiv % 100) > 50 ) { printf( "*" ); mux -= muxdiv; }
//mux -= muxdiv;
if( mux <= 0 )
{
outbins[i] = 0;
}
else
{
//if( i == 0 )
//printf( "MUX: %d %d = %d\n", isc, iss, mux );
outbins[i] = sqrt((float)mux)/50.0;
if( abs( cossindata[i*2+0] ) > 2000 || abs( cossindata[i*2+1] ) > 2000 )
printf( "%d/%d/%d/%f ", i, cossindata[i*2+0], cossindata[i*2+1],outbins[i] );
//outbins[i] = (cossindata[i*2+0]/10000.0);
}
}
printf( "\n" );
#endif
}