Now I've got the Overlap & Scrap right! The spectrum is clean.

This commit is contained in:
ha7ilm 2015-11-22 14:34:24 +01:00
parent 646de8b258
commit b6fde4ce86
3 changed files with 39 additions and 17 deletions

6
csdr.c
View file

@ -1725,13 +1725,14 @@ int main(int argc, char *argv[])
//make the filter
float filter_half_bw = 0.5/decimation;
firdes_bandpass_c(taps, ddc.taps_real_length, shift_rate-filter_half_bw, shift_rate+filter_half_bw, window);
fprintf(stderr, "fastddc_inv_cc: preparing a bandpass filter of [%g, %g] cutoff rates. Real transition bandwidth is: %g\n", shift_rate-filter_half_bw, shift_rate+filter_half_bw, 4.0/ddc.taps_length);
firdes_bandpass_c(taps, ddc.taps_length, shift_rate-filter_half_bw, shift_rate+filter_half_bw, window);
fft_execute(plan_taps);
//make FFT plan
complexf* inv_input = (complexf*)fft_malloc(sizeof(complexf)*ddc.fft_inv_size);
complexf* inv_output = (complexf*)fft_malloc(sizeof(complexf)*ddc.fft_inv_size);
fprintf(stderr,"fastddc_apply_cc: benchmarking FFT...");
fprintf(stderr,"fastddc_inv_cc: benchmarking FFT...");
FFT_PLAN_T* plan_inverse = make_fft_c2c(ddc.fft_inv_size, inv_input, inv_output, 0, 1); //inverse, do benchmark
fprintf(stderr," done\n");
@ -1747,6 +1748,7 @@ int main(int argc, char *argv[])
fread(input, sizeof(complexf), ddc.fft_size, stdin);
shift_stat = fastddc_inv_cc(input, output, &ddc, plan_inverse, taps_fft, shift_stat);
fwrite(output, sizeof(complexf), shift_stat.output_size, stdout);
fprintf(stderr, "ss os = %d\n", shift_stat.output_size);
TRY_YIELD;
}
}

View file

@ -44,16 +44,16 @@ int fastddc_init(fastddc_t* ddc, float transition_bw, int decimation, float shif
ddc->post_decimation/=2;
ddc->pre_decimation*=2;
}
ddc->taps_real_length = firdes_filter_len(transition_bw); //the number of non-zero taps
ddc->taps_length = ceil(ddc->taps_real_length/(float)ddc->pre_decimation) * ddc->pre_decimation; //the number of taps must be a multiple of the decimation factor
ddc->taps_min_length = firdes_filter_len(transition_bw); //his is the minimal number of taps to achieve the given transition_bw; we are likely to have more taps than this number.
ddc->taps_length = next_pow2(ceil(ddc->taps_min_length/(float)ddc->pre_decimation) * ddc->pre_decimation) + 1; //the number of taps must be a multiple of the decimation factor
ddc->fft_size = next_pow2(ddc->taps_length * 4); //it is a good rule of thumb for performance (based on the article), but we should do benchmarks
while (ddc->fft_size<ddc->pre_decimation) ddc->fft_size*=2; //fft_size should be a multiple of pre_decimation
while (ddc->fft_size<ddc->pre_decimation) ddc->fft_size*=2; //fft_size should be a multiple of pre_decimation.
ddc->overlap_length = ddc->taps_length - 1;
ddc->input_size = ddc->fft_size - ddc->overlap_length;
ddc->fft_inv_size = ddc->fft_size / ddc->pre_decimation;
//Shift operation in the frequency domain: we can shift by a multiple of v.
ddc->v = ddc->fft_size/ddc->overlap_length; //+-1 ? (or maybe ceil() this?) //TODO: why?
ddc->v = ddc->fft_size/ddc->overlap_length; //overlap factor | +-1 ? (or maybe ceil() this?)
int middlebin=ddc->fft_size / 2;
ddc->startbin = middlebin + middlebin * shift_rate * 2;
//fprintf(stderr, "ddc->startbin=%g\n",(float)ddc->startbin);
@ -64,7 +64,7 @@ int fastddc_init(fastddc_t* ddc, float transition_bw, int decimation, float shif
ddc->pre_shift = ddc->offsetbin/(float)ddc->fft_size;
ddc->dsadata = decimating_shift_addition_init(ddc->post_shift, ddc->post_decimation);
//Overlap is scrapd, not added
//Overlap is scrapped, not added
ddc->scrap=ddc->overlap_length/ddc->pre_decimation; //TODO this is problematic sometimes! overlap_length = 401 :: scrap = 200
ddc->post_input_size=ddc->fft_inv_size-ddc->scrap;
@ -76,13 +76,13 @@ void fastddc_print(fastddc_t* ddc, char* source)
{
fprintf(stderr,
"%s: fastddc_print_sizes(): (fft_size = %d) = (taps_length = %d) + (input_size = %d) - 1\n"
" overlap :: (overlap_length = %d) = taps_length - 1, taps_real_length = %d\n"
" overlap :: (overlap_length = %d) = taps_length - 1, taps_min_length = %d\n"
" decimation :: decimation = (pre_decimation = %d) * (post_decimation = %d), fft_inv_size = %d\n"
" shift :: startbin = %d, offsetbin = %d, v = %d, pre_shift = %g, post_shift = %g\n"
" o&s :: post_input_size = %d, scrap = %d\n"
,
source, ddc->fft_size, ddc->taps_length, ddc->input_size,
ddc->overlap_length, ddc->taps_real_length,
ddc->overlap_length, ddc->taps_min_length,
ddc->pre_decimation, ddc->post_decimation, ddc->fft_inv_size,
ddc->startbin, ddc->offsetbin, ddc->v, ddc->pre_shift, ddc->post_shift,
ddc->post_input_size, ddc->scrap );
@ -120,16 +120,35 @@ decimating_shift_addition_status_t fastddc_inv_cc(complexf* input, complexf* out
}
//Alias & shift & filter at once
fft_swap_sides(input, ddc->fft_size); //this is not very optimal, but now we stick with this slow solution until we got the algorithm working
fft_swap_sides(input, ddc->fft_size); //TODO this is not very optimal, but now we stick with this slow solution until we got the algorithm working
//fprintf(stderr, " === fastddc_inv_cc() ===\n");
for(int i=0;i<ddc->fft_size;i++)
{
int output_index = (ddc->startbin+i)%plan_inverse->size;
int tap_index = (ddc->fft_size+i-ddc->offsetbin)%ddc->fft_size;
cmultadd(inv_input+output_index, input+i, taps_fft+tap_index); //cmultadd(output, input1, input2): complex output += complex input1 * complex input 2
int output_index = (ddc->fft_size+i-ddc->offsetbin)%plan_inverse->size;
int tap_index = i;
//fprintf(stderr, "output_index = %d , tap_index = %d, input index = %d\n", output_index, tap_index, i);
//cmultadd(inv_input+output_index, input+i, taps_fft+tap_index); //cmultadd(output, input1, input2): complex output += complex input1 * complex input 2
// (a+b*i)*(c+d*i) = (ac-bd)+(ad+bc)*i
// a = iof(input,i)
// b = qof(input,i)
// c = iof(taps_fft,i)
// d = qof(taps_fft,i)
//iof(inv_input,output_index) += iof(input,i) * iof(taps_fft,i) - qof(input,i) * qof(taps_fft,i);
//qof(inv_input,output_index) += iof(input,i) * qof(taps_fft,i) + qof(input,i) * iof(taps_fft,i);
iof(inv_input,output_index) += iof(input,i); //no filter
qof(inv_input,output_index) += qof(input,i);
}
//Normalize inv fft bins (now our output level is not higher than the input... but we may optimize this into the later loop when we normalize by size)
for(int i=0;i<plan_inverse->size;i++)
{
iof(inv_input,i)/=ddc->pre_decimation;
qof(inv_input,i)/=ddc->pre_decimation;
}
fft_swap_sides(inv_input,plan_inverse->size);
fft_execute(plan_inverse);
fft_swap_sides(inv_output,plan_inverse->size);
//Normalize data
for(int i=0;i<plan_inverse->size;i++) //@fastddc_inv_cc: normalize by size
@ -140,7 +159,8 @@ decimating_shift_addition_status_t fastddc_inv_cc(complexf* input, complexf* out
//Overlap is scrapped, not added
//Shift correction
shift_stat=decimating_shift_addition_cc(inv_output+ddc->scrap, output, ddc->post_input_size, ddc->dsadata, ddc->post_decimation, shift_stat);
//memcpy(inv_output+ddc->scrap, output
//shift_stat=decimating_shift_addition_cc(inv_output+ddc->scrap, output, ddc->post_input_size, ddc->dsadata, ddc->post_decimation, shift_stat);
shift_stat.output_size = ddc->post_input_size; //bypass shift correction
memcpy(output, inv_output+ddc->scrap, sizeof(complexf)*ddc->post_input_size);
return shift_stat;
}

View file

@ -7,7 +7,7 @@ typedef struct fastddc_s
int pre_decimation;
int post_decimation;
int taps_length;
int taps_real_length;
int taps_min_length;
int overlap_length; //it is taps_length - 1
int fft_size;
int fft_inv_size;