91df5b1f2d
calc_varsamp() gained a transpose of the coefficient table, which showed up as ~0.3 ms on create_varsamp()/setInRate_varsamp() (1.0 ms -> 1.4 ms). It is a one-off setup cost, not on the sample path, but it is easy to trim: every element of hp is written, so malloc0()'s memset of ~1 MB is dead, and h is cold straight out of fir_bandpass(), so walk it along its fast axis and let the prefetcher work. Output is bit-identical; xvarsamp() is unchanged at ~34 us/buffer. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
318 lines
8.3 KiB
C
318 lines
8.3 KiB
C
/* varsamp.c
|
|
|
|
This file is part of a program that implements a Software-Defined Radio.
|
|
|
|
Copyright (C) 2017 Warren Pratt, NR0V
|
|
|
|
This program is free software; you can redistribute it and/or
|
|
modify it under the terms of the GNU General Public License
|
|
as published by the Free Software Foundation; either version 2
|
|
of the License, or (at your option) any later version.
|
|
|
|
This program is distributed in the hope that it will be useful,
|
|
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
GNU General Public License for more details.
|
|
|
|
You should have received a copy of the GNU General Public License
|
|
along with this program; if not, write to the Free Software
|
|
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
|
|
|
|
The author can be reached by email at
|
|
|
|
warren@wpratt.com
|
|
|
|
*/
|
|
|
|
#include "comm.h"
|
|
|
|
void calc_varsamp (VARSAMP a)
|
|
{
|
|
double min_rate, norm_rate;
|
|
// double max_rate;
|
|
double fc_norm_high, fc_norm_low;
|
|
a->nom_ratio = (double)a->out_rate / (double)a->in_rate;
|
|
a->cvar = a->var * a->nom_ratio;
|
|
a->inv_cvar = 1.0 / a->cvar;
|
|
a->old_inv_cvar = a->inv_cvar;
|
|
a->dicvar = 0.0;
|
|
a->delta = fabs (1.0 / a->cvar - 1.0);
|
|
a->fc = a->fcin;
|
|
if (a->out_rate >= a->in_rate)
|
|
{
|
|
min_rate = (double)a->in_rate;
|
|
// max_rate = (double)a->out_rate;
|
|
norm_rate = min_rate;
|
|
}
|
|
else
|
|
{
|
|
min_rate = (double)a->out_rate;
|
|
// max_rate = (double)a->in_rate;
|
|
// norm_rate = max_rate;
|
|
norm_rate = (double)a->in_rate;
|
|
}
|
|
if (a->fc == 0.0) a->fc = 0.95 * 0.45 * min_rate;
|
|
fc_norm_high = a->fc / norm_rate;
|
|
if (a->fc_low < 0.0)
|
|
fc_norm_low = - fc_norm_high;
|
|
else
|
|
fc_norm_low = a->fc_low / norm_rate;
|
|
a->rsize = (int)(140.0 * norm_rate / min_rate);
|
|
a->ncoef = a->rsize + 1;
|
|
a->ncoef += (a->R - 1) * (a->ncoef - 1); // = R * rsize + 1
|
|
{
|
|
/* Store the coefficients transposed into phases. hshift() walks
|
|
h[hidx + m*R] for m = 0..rsize-1, which strides by R doubles -- 8 KB
|
|
at R = 1024 -- over a 1.1 MB table, so every tap is its own cache
|
|
line. Transposing makes each phase contiguous; hshift() interpolates
|
|
between phases hidx and hidx+1, hence R+1 of them. */
|
|
int p, m;
|
|
const int R = a->R, rsize = a->rsize;
|
|
double* h = fir_bandpass(a->ncoef, fc_norm_low, fc_norm_high, (double)R, 1, 0, (double)R * a->gain);
|
|
// every element is written below, so skip malloc0()'s memset of ~1 MB
|
|
a->hp = (double *)_aligned_malloc ((size_t)(R + 1) * rsize * sizeof (double), 16);
|
|
// walk h forward (p is its fast axis) so the prefetcher sees a linear
|
|
// stream; h is cold here, straight from fir_bandpass()
|
|
for (m = 0; m < rsize; m++)
|
|
for (p = 0; p <= R; p++)
|
|
a->hp[(size_t)p * rsize + m] = h[p + (size_t)m * R];
|
|
_aligned_free (h);
|
|
}
|
|
a->ringI = (double *)malloc0(a->rsize * sizeof(double));
|
|
a->ringQ = (double *)malloc0(a->rsize * sizeof(double));
|
|
a->idx_in = a->rsize - 1;
|
|
a->h_offset = 0.0;
|
|
a->hs = (double *)malloc0 (a->rsize * sizeof (double));
|
|
a->isamps = 0.0;
|
|
}
|
|
|
|
void decalc_varsamp (VARSAMP a)
|
|
{
|
|
_aligned_free (a->hs);
|
|
_aligned_free (a->ringQ);
|
|
_aligned_free (a->ringI);
|
|
_aligned_free (a->hp);
|
|
}
|
|
|
|
VARSAMP create_varsamp ( int run, int size, double* in, double* out,
|
|
int in_rate, int out_rate, double fc, double fc_low, int R, double gain, double var, int varmode)
|
|
{
|
|
VARSAMP a = (VARSAMP) malloc0 (sizeof (varsamp));
|
|
|
|
a->run = run;
|
|
a->size = size;
|
|
a->in = in;
|
|
a->out = out;
|
|
a->in_rate = in_rate;
|
|
a->out_rate = out_rate;
|
|
a->fcin = fc;
|
|
a->fc_low = fc_low;
|
|
a->R = R;
|
|
a->gain = gain;
|
|
a->var = var;
|
|
a->varmode = varmode;
|
|
calc_varsamp (a);
|
|
return a;
|
|
}
|
|
|
|
void destroy_varsamp (VARSAMP a)
|
|
{
|
|
decalc_varsamp (a);
|
|
_aligned_free (a);
|
|
}
|
|
|
|
void flush_varsamp (VARSAMP a)
|
|
{
|
|
memset (a->ringI, 0, a->rsize * sizeof (double));
|
|
memset (a->ringQ, 0, a->rsize * sizeof (double));
|
|
a->idx_in = a->rsize - 1;
|
|
a->h_offset = 0.0;
|
|
a->isamps = 0.0;
|
|
}
|
|
|
|
/* Accumulate n taps of a unit-stride complex dot product into *pI / *pQ.
|
|
|
|
Four independent accumulator pairs keep the FMAs off a single dependency
|
|
chain and let the vectorizer in: an 'I += h[j]*x[j]' reduction cannot be
|
|
reassociated without -ffast-math, which this library must not enable (it
|
|
relies on IEEE semantics for 0/0 = NaN and x/0 = Inf). */
|
|
static inline void varsamp_dot (const double* WDSP_RESTRICT hp,
|
|
const double* WDSP_RESTRICT xI, const double* WDSP_RESTRICT xQ,
|
|
int n, double* pI, double* pQ)
|
|
{
|
|
double i0 = 0.0, i1 = 0.0, i2 = 0.0, i3 = 0.0;
|
|
double q0 = 0.0, q1 = 0.0, q2 = 0.0, q3 = 0.0;
|
|
int j = 0;
|
|
for (; j <= n - 4; j += 4)
|
|
{
|
|
i0 += hp[j + 0] * xI[j + 0]; q0 += hp[j + 0] * xQ[j + 0];
|
|
i1 += hp[j + 1] * xI[j + 1]; q1 += hp[j + 1] * xQ[j + 1];
|
|
i2 += hp[j + 2] * xI[j + 2]; q2 += hp[j + 2] * xQ[j + 2];
|
|
i3 += hp[j + 3] * xI[j + 3]; q3 += hp[j + 3] * xQ[j + 3];
|
|
}
|
|
for (; j < n; j++)
|
|
{
|
|
i0 += hp[j] * xI[j];
|
|
q0 += hp[j] * xQ[j];
|
|
}
|
|
*pI += (i0 + i1) + (i2 + i3);
|
|
*pQ += (q0 + q1) + (q2 + q3);
|
|
}
|
|
|
|
void hshift (VARSAMP a)
|
|
{
|
|
int m;
|
|
int hidx;
|
|
double frac, pos;
|
|
const int rsize = a->rsize;
|
|
const double* WDSP_RESTRICT h0;
|
|
const double* WDSP_RESTRICT h1;
|
|
double* WDSP_RESTRICT hs = a->hs;
|
|
/* h_offset is normalized to [0,1) by the caller, so hidx is in [0, R-1]
|
|
and phase hidx+1 <= R exists. */
|
|
pos = (double)a->R * a->h_offset;
|
|
hidx = (int)(pos);
|
|
frac = pos - (double)hidx;
|
|
h0 = a->hp + (size_t)hidx * rsize;
|
|
h1 = h0 + rsize;
|
|
for (m = 0; m < rsize; m++)
|
|
hs[rsize - 1 - m] = h0[m] + frac * (h1[m] - h0[m]);
|
|
}
|
|
|
|
int xvarsamp (VARSAMP a, double var)
|
|
{
|
|
int outsamps = 0;
|
|
uint64_t* picvar;
|
|
uint64_t N;
|
|
a->var = var;
|
|
a->old_inv_cvar = a->inv_cvar;
|
|
a->cvar = a->var * a->nom_ratio;
|
|
a->inv_cvar = 1.0 / a->cvar;
|
|
if (a->varmode)
|
|
{
|
|
a->dicvar = (a->inv_cvar - a->old_inv_cvar) / (double)a->size;
|
|
a->inv_cvar = a->old_inv_cvar;
|
|
}
|
|
else a->dicvar = 0.0;
|
|
if (a->run)
|
|
{
|
|
int i, n1;
|
|
double I, Q;
|
|
const int rsize = a->rsize;
|
|
/* a->hs is rewritten by hshift() on every output sample, so it must not
|
|
be hoisted behind a restrict pointer here; varsamp_dot() re-reads it. */
|
|
const double* in = a->in;
|
|
double* out = a->out;
|
|
double* WDSP_RESTRICT ringI = a->ringI;
|
|
double* WDSP_RESTRICT ringQ = a->ringQ;
|
|
int idx_in = a->idx_in;
|
|
|
|
for (i = 0; i < a->size; i++)
|
|
{
|
|
ringI[idx_in] = in[2 * i + 0];
|
|
ringQ[idx_in] = in[2 * i + 1];
|
|
a->inv_cvar += a->dicvar;
|
|
picvar = (uint64_t*)(&a->inv_cvar);
|
|
N = *picvar & 0xffffffffffff0000;
|
|
a->inv_cvar = *((double *)&N);
|
|
a->delta = 1.0 - a->inv_cvar;
|
|
while (a->isamps < 1.0)
|
|
{
|
|
I = 0.0;
|
|
Q = 0.0;
|
|
hshift (a);
|
|
a->h_offset += a->delta;
|
|
while (a->h_offset >= 1.0) a->h_offset -= 1.0;
|
|
while (a->h_offset < 0.0) a->h_offset += 1.0;
|
|
/* the ring wraps at most once over rsize taps; split it so both
|
|
halves are unit-stride */
|
|
n1 = rsize - idx_in;
|
|
varsamp_dot (a->hs, ringI + idx_in, ringQ + idx_in, n1, &I, &Q);
|
|
if (n1 < rsize)
|
|
varsamp_dot (a->hs + n1, ringI, ringQ, rsize - n1, &I, &Q);
|
|
out[2 * outsamps + 0] = I;
|
|
out[2 * outsamps + 1] = Q;
|
|
outsamps++;
|
|
a->isamps += a->inv_cvar;
|
|
}
|
|
a->isamps -= 1.0;
|
|
if (--idx_in < 0) idx_in = rsize - 1;
|
|
}
|
|
a->idx_in = idx_in;
|
|
}
|
|
else if (a->in != a->out)
|
|
memcpy (a->out, a->in, a->size * sizeof (complex));
|
|
return outsamps;
|
|
}
|
|
|
|
void setBuffers_varsamp (VARSAMP a, double* in, double* out)
|
|
{
|
|
a->in = in;
|
|
a->out = out;
|
|
}
|
|
|
|
void setSize_varsamp (VARSAMP a, int size)
|
|
{
|
|
a->size = size;
|
|
flush_varsamp (a);
|
|
}
|
|
|
|
void setInRate_varsamp (VARSAMP a, int rate)
|
|
{
|
|
decalc_varsamp (a);
|
|
a->in_rate = rate;
|
|
calc_varsamp (a);
|
|
}
|
|
|
|
void setOutRate_varsamp (VARSAMP a, int rate)
|
|
{
|
|
decalc_varsamp (a);
|
|
a->out_rate = rate;
|
|
calc_varsamp (a);
|
|
}
|
|
|
|
void setFCLow_varsamp (VARSAMP a, double fc_low)
|
|
{
|
|
if (fc_low != a->fc_low)
|
|
{
|
|
decalc_varsamp (a);
|
|
a->fc_low = fc_low;
|
|
calc_varsamp (a);
|
|
}
|
|
}
|
|
|
|
void setBandwidth_varsamp (VARSAMP a, double fc_low, double fc_high)
|
|
{
|
|
if (fc_low != a->fc_low || fc_high != a->fcin)
|
|
{
|
|
decalc_varsamp (a);
|
|
a->fc_low = fc_low;
|
|
a->fcin = fc_high;
|
|
calc_varsamp (a);
|
|
}
|
|
}
|
|
|
|
// exported calls
|
|
|
|
PORT
|
|
void* create_varsampV (int in_rate, int out_rate, int R)
|
|
{
|
|
return (void *)create_varsamp (1, 0, 0, 0, in_rate, out_rate, 0.0, -1.0, R, 1.0, 1.0, 1);
|
|
}
|
|
|
|
PORT
|
|
void xvarsampV (double* input, double* output, int numsamps, double var, int* outsamps, void* ptr)
|
|
{
|
|
VARSAMP a = (VARSAMP)ptr;
|
|
a->in = input;
|
|
a->out = output;
|
|
a->size = numsamps;
|
|
*outsamps = xvarsamp(a, var);
|
|
}
|
|
|
|
PORT
|
|
void destroy_varsampV (void* ptr)
|
|
{
|
|
destroy_varsamp ( (VARSAMP)ptr );
|
|
}
|