xref: /freebsd/sys/dev/sound/pcm/feeder_rate.c (revision a57ca37dd1848cd42844d9082c4a74c2ed57f68a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2005-2009 Ariff Abdullah <ariff@FreeBSD.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * feeder_rate: (Codename: Z Resampler), which means any effort to create
31  *              future replacement for this resampler are simply absurd unless
32  *              the world decide to add new alphabet after Z.
33  *
34  * FreeBSD bandlimited sinc interpolator, technically based on
35  * "Digital Audio Resampling" by Julius O. Smith III
36  *  - http://ccrma.stanford.edu/~jos/resample/
37  *
38  * The Good:
39  * + all out fixed point integer operations, no soft-float or anything like
40  *   that.
41  * + classic polyphase converters with high quality coefficient's polynomial
42  *   interpolators.
43  * + fast, faster, or the fastest of its kind.
44  * + compile time configurable.
45  * + etc etc..
46  *
47  * The Bad:
48  * - The z, z_, and Z_ . Due to mental block (or maybe just 0x7a69), I
49  *   couldn't think of anything simpler than that (feeder_rate_xxx is just
50  *   too long). Expect possible clashes with other zitizens (any?).
51  */
52 
53 #ifdef _KERNEL
54 #ifdef HAVE_KERNEL_OPTION_HEADERS
55 #include "opt_snd.h"
56 #endif
57 #include <dev/sound/pcm/sound.h>
58 #include <dev/sound/pcm/pcm.h>
59 #include "feeder_if.h"
60 
61 #define SND_USE_FXDIV
62 #include "snd_fxdiv_gen.h"
63 
64 SND_DECLARE_FILE("$FreeBSD$");
65 #endif
66 
67 #include "feeder_rate_gen.h"
68 
69 #if !defined(_KERNEL) && defined(SND_DIAGNOSTIC)
70 #undef Z_DIAGNOSTIC
71 #define Z_DIAGNOSTIC		1
72 #elif defined(_KERNEL)
73 #undef Z_DIAGNOSTIC
74 #endif
75 
76 #ifndef Z_QUALITY_DEFAULT
77 #define Z_QUALITY_DEFAULT	Z_QUALITY_LINEAR
78 #endif
79 
80 #define Z_RESERVOIR		2048
81 #define Z_RESERVOIR_MAX		131072
82 
83 #define Z_SINC_MAX		0x3fffff
84 #define Z_SINC_DOWNMAX		48		/* 384000 / 8000 */
85 
86 #ifdef _KERNEL
87 #define Z_POLYPHASE_MAX		183040		/* 286 taps, 640 phases */
88 #else
89 #define Z_POLYPHASE_MAX		1464320		/* 286 taps, 5120 phases */
90 #endif
91 
92 #define Z_RATE_DEFAULT		48000
93 
94 #define Z_RATE_MIN		FEEDRATE_RATEMIN
95 #define Z_RATE_MAX		FEEDRATE_RATEMAX
96 #define Z_ROUNDHZ		FEEDRATE_ROUNDHZ
97 #define Z_ROUNDHZ_MIN		FEEDRATE_ROUNDHZ_MIN
98 #define Z_ROUNDHZ_MAX		FEEDRATE_ROUNDHZ_MAX
99 
100 #define Z_RATE_SRC		FEEDRATE_SRC
101 #define Z_RATE_DST		FEEDRATE_DST
102 #define Z_RATE_QUALITY		FEEDRATE_QUALITY
103 #define Z_RATE_CHANNELS		FEEDRATE_CHANNELS
104 
105 #define Z_PARANOID		1
106 
107 #define Z_MULTIFORMAT		1
108 
109 #ifdef _KERNEL
110 #undef Z_USE_ALPHADRIFT
111 #define Z_USE_ALPHADRIFT	1
112 #endif
113 
114 #define Z_FACTOR_MIN		1
115 #define Z_FACTOR_MAX		Z_MASK
116 #define Z_FACTOR_SAFE(v)	(!((v) < Z_FACTOR_MIN || (v) > Z_FACTOR_MAX))
117 
118 struct z_info;
119 
120 typedef void (*z_resampler_t)(struct z_info *, uint8_t *);
121 
122 struct z_info {
123 	int32_t rsrc, rdst;	/* original source / destination rates */
124 	int32_t src, dst;	/* rounded source / destination rates */
125 	int32_t channels;	/* total channels */
126 	int32_t bps;		/* bytes-per-sample */
127 	int32_t quality;	/* resampling quality */
128 
129 	int32_t z_gx, z_gy;	/* interpolation / decimation ratio */
130 	int32_t z_alpha;	/* output sample time phase / drift */
131 	uint8_t *z_delay;	/* FIR delay line / linear buffer */
132 	int32_t *z_coeff;	/* FIR coefficients */
133 	int32_t *z_dcoeff;	/* FIR coefficients differences */
134 	int32_t *z_pcoeff;	/* FIR polyphase coefficients */
135 	int32_t z_scale;	/* output scaling */
136 	int32_t z_dx;		/* input sample drift increment */
137 	int32_t z_dy;		/* output sample drift increment */
138 #ifdef Z_USE_ALPHADRIFT
139 	int32_t z_alphadrift;	/* alpha drift rate */
140 	int32_t z_startdrift;	/* buffer start position drift rate */
141 #endif
142 	int32_t z_mask;		/* delay line full length mask */
143 	int32_t z_size;		/* half width of FIR taps */
144 	int32_t z_full;		/* full size of delay line */
145 	int32_t z_alloc;	/* largest allocated full size of delay line */
146 	int32_t z_start;	/* buffer processing start position */
147 	int32_t z_pos;		/* current position for the next feed */
148 #ifdef Z_DIAGNOSTIC
149 	uint32_t z_cycle;	/* output cycle, purely for statistical */
150 #endif
151 	int32_t z_maxfeed;	/* maximum feed to avoid 32bit overflow */
152 
153 	z_resampler_t z_resample;
154 };
155 
156 int feeder_rate_min = Z_RATE_MIN;
157 int feeder_rate_max = Z_RATE_MAX;
158 int feeder_rate_round = Z_ROUNDHZ;
159 int feeder_rate_quality = Z_QUALITY_DEFAULT;
160 
161 static int feeder_rate_polyphase_max = Z_POLYPHASE_MAX;
162 
163 #ifdef _KERNEL
164 static char feeder_rate_presets[] = FEEDER_RATE_PRESETS;
165 SYSCTL_STRING(_hw_snd, OID_AUTO, feeder_rate_presets, CTLFLAG_RD,
166     &feeder_rate_presets, 0, "compile-time rate presets");
167 SYSCTL_INT(_hw_snd, OID_AUTO, feeder_rate_polyphase_max, CTLFLAG_RWTUN,
168     &feeder_rate_polyphase_max, 0, "maximum allowable polyphase entries");
169 
170 static int
171 sysctl_hw_snd_feeder_rate_min(SYSCTL_HANDLER_ARGS)
172 {
173 	int err, val;
174 
175 	val = feeder_rate_min;
176 	err = sysctl_handle_int(oidp, &val, 0, req);
177 
178 	if (err != 0 || req->newptr == NULL || val == feeder_rate_min)
179 		return (err);
180 
181 	if (!(Z_FACTOR_SAFE(val) && val < feeder_rate_max))
182 		return (EINVAL);
183 
184 	feeder_rate_min = val;
185 
186 	return (0);
187 }
188 SYSCTL_PROC(_hw_snd, OID_AUTO, feeder_rate_min,
189     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, sizeof(int),
190     sysctl_hw_snd_feeder_rate_min, "I",
191     "minimum allowable rate");
192 
193 static int
194 sysctl_hw_snd_feeder_rate_max(SYSCTL_HANDLER_ARGS)
195 {
196 	int err, val;
197 
198 	val = feeder_rate_max;
199 	err = sysctl_handle_int(oidp, &val, 0, req);
200 
201 	if (err != 0 || req->newptr == NULL || val == feeder_rate_max)
202 		return (err);
203 
204 	if (!(Z_FACTOR_SAFE(val) && val > feeder_rate_min))
205 		return (EINVAL);
206 
207 	feeder_rate_max = val;
208 
209 	return (0);
210 }
211 SYSCTL_PROC(_hw_snd, OID_AUTO, feeder_rate_max,
212     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, sizeof(int),
213     sysctl_hw_snd_feeder_rate_max, "I",
214     "maximum allowable rate");
215 
216 static int
217 sysctl_hw_snd_feeder_rate_round(SYSCTL_HANDLER_ARGS)
218 {
219 	int err, val;
220 
221 	val = feeder_rate_round;
222 	err = sysctl_handle_int(oidp, &val, 0, req);
223 
224 	if (err != 0 || req->newptr == NULL || val == feeder_rate_round)
225 		return (err);
226 
227 	if (val < Z_ROUNDHZ_MIN || val > Z_ROUNDHZ_MAX)
228 		return (EINVAL);
229 
230 	feeder_rate_round = val - (val % Z_ROUNDHZ);
231 
232 	return (0);
233 }
234 SYSCTL_PROC(_hw_snd, OID_AUTO, feeder_rate_round,
235     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, sizeof(int),
236     sysctl_hw_snd_feeder_rate_round, "I",
237     "sample rate converter rounding threshold");
238 
239 static int
240 sysctl_hw_snd_feeder_rate_quality(SYSCTL_HANDLER_ARGS)
241 {
242 	struct snddev_info *d;
243 	struct pcm_channel *c;
244 	struct pcm_feeder *f;
245 	int i, err, val;
246 
247 	val = feeder_rate_quality;
248 	err = sysctl_handle_int(oidp, &val, 0, req);
249 
250 	if (err != 0 || req->newptr == NULL || val == feeder_rate_quality)
251 		return (err);
252 
253 	if (val < Z_QUALITY_MIN || val > Z_QUALITY_MAX)
254 		return (EINVAL);
255 
256 	feeder_rate_quality = val;
257 
258 	/*
259 	 * Traverse all available channels on each device and try to
260 	 * set resampler quality if and only if it is exist as
261 	 * part of feeder chains and the channel is idle.
262 	 */
263 	for (i = 0; pcm_devclass != NULL &&
264 	    i < devclass_get_maxunit(pcm_devclass); i++) {
265 		d = devclass_get_softc(pcm_devclass, i);
266 		if (!PCM_REGISTERED(d))
267 			continue;
268 		PCM_LOCK(d);
269 		PCM_WAIT(d);
270 		PCM_ACQUIRE(d);
271 		CHN_FOREACH(c, d, channels.pcm) {
272 			CHN_LOCK(c);
273 			f = chn_findfeeder(c, FEEDER_RATE);
274 			if (f == NULL || f->data == NULL || CHN_STARTED(c)) {
275 				CHN_UNLOCK(c);
276 				continue;
277 			}
278 			(void)FEEDER_SET(f, FEEDRATE_QUALITY, val);
279 			CHN_UNLOCK(c);
280 		}
281 		PCM_RELEASE(d);
282 		PCM_UNLOCK(d);
283 	}
284 
285 	return (0);
286 }
287 SYSCTL_PROC(_hw_snd, OID_AUTO, feeder_rate_quality,
288     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, 0, sizeof(int),
289     sysctl_hw_snd_feeder_rate_quality, "I",
290     "sample rate converter quality ("__XSTRING(Z_QUALITY_MIN)"=low .. "
291     __XSTRING(Z_QUALITY_MAX)"=high)");
292 #endif	/* _KERNEL */
293 
294 /*
295  * Resampler type.
296  */
297 #define Z_IS_ZOH(i)		((i)->quality == Z_QUALITY_ZOH)
298 #define Z_IS_LINEAR(i)		((i)->quality == Z_QUALITY_LINEAR)
299 #define Z_IS_SINC(i)		((i)->quality > Z_QUALITY_LINEAR)
300 
301 /*
302  * Macroses for accurate sample time drift calculations.
303  *
304  * gy2gx : given the amount of output, return the _exact_ required amount of
305  *         input.
306  * gx2gy : given the amount of input, return the _maximum_ amount of output
307  *         that will be generated.
308  * drift : given the amount of input and output, return the elapsed
309  *         sample-time.
310  */
311 #define _Z_GCAST(x)		((uint64_t)(x))
312 
313 #if defined(__i386__)
314 /*
315  * This is where i386 being beaten to a pulp. Fortunately this function is
316  * rarely being called and if it is, it will decide the best (hopefully)
317  * fastest way to do the division. If we can ensure that everything is dword
318  * aligned, letting the compiler to call udivdi3 to do the division can be
319  * faster compared to this.
320  *
321  * amd64 is the clear winner here, no question about it.
322  */
323 static __inline uint32_t
324 Z_DIV(uint64_t v, uint32_t d)
325 {
326 	uint32_t hi, lo, quo, rem;
327 
328 	hi = v >> 32;
329 	lo = v & 0xffffffff;
330 
331 	/*
332 	 * As much as we can, try to avoid long division like a plague.
333 	 */
334 	if (hi == 0)
335 		quo = lo / d;
336 	else
337 		__asm("divl %2"
338 		    : "=a" (quo), "=d" (rem)
339 		    : "r" (d), "0" (lo), "1" (hi));
340 
341 	return (quo);
342 }
343 #else
344 #define Z_DIV(x, y)		((x) / (y))
345 #endif
346 
347 #define _Z_GY2GX(i, a, v)						\
348 	Z_DIV(((_Z_GCAST((i)->z_gx) * (v)) + ((i)->z_gy - (a) - 1)),	\
349 	(i)->z_gy)
350 
351 #define _Z_GX2GY(i, a, v)						\
352 	Z_DIV(((_Z_GCAST((i)->z_gy) * (v)) + (a)), (i)->z_gx)
353 
354 #define _Z_DRIFT(i, x, y)						\
355 	((_Z_GCAST((i)->z_gy) * (x)) - (_Z_GCAST((i)->z_gx) * (y)))
356 
357 #define z_gy2gx(i, v)		_Z_GY2GX(i, (i)->z_alpha, v)
358 #define z_gx2gy(i, v)		_Z_GX2GY(i, (i)->z_alpha, v)
359 #define z_drift(i, x, y)	_Z_DRIFT(i, x, y)
360 
361 /*
362  * Macroses for SINC coefficients table manipulations.. whatever.
363  */
364 #define Z_SINC_COEFF_IDX(i)	((i)->quality - Z_QUALITY_LINEAR - 1)
365 
366 #define Z_SINC_LEN(i)							\
367 	((int32_t)(((uint64_t)z_coeff_tab[Z_SINC_COEFF_IDX(i)].len <<	\
368 	    Z_SHIFT) / (i)->z_dy))
369 
370 #define Z_SINC_BASE_LEN(i)						\
371 	((z_coeff_tab[Z_SINC_COEFF_IDX(i)].len - 1) >> (Z_DRIFT_SHIFT - 1))
372 
373 /*
374  * Macroses for linear delay buffer operations. Alignment is not
375  * really necessary since we're not using true circular buffer, but it
376  * will help us guard against possible trespasser. To be honest,
377  * the linear block operations does not need guarding at all due to
378  * accurate drifting!
379  */
380 #define z_align(i, v)		((v) & (i)->z_mask)
381 #define z_next(i, o, v)		z_align(i, (o) + (v))
382 #define z_prev(i, o, v)		z_align(i, (o) - (v))
383 #define z_fetched(i)		(z_align(i, (i)->z_pos - (i)->z_start) - 1)
384 #define z_free(i)		((i)->z_full - (i)->z_pos)
385 
386 /*
387  * Macroses for Bla Bla .. :)
388  */
389 #define z_copy(src, dst, sz)	(void)memcpy(dst, src, sz)
390 #define z_feed(...)		FEEDER_FEED(__VA_ARGS__)
391 
392 static __inline uint32_t
393 z_min(uint32_t x, uint32_t y)
394 {
395 
396 	return ((x < y) ? x : y);
397 }
398 
399 static int32_t
400 z_gcd(int32_t x, int32_t y)
401 {
402 	int32_t w;
403 
404 	while (y != 0) {
405 		w = x % y;
406 		x = y;
407 		y = w;
408 	}
409 
410 	return (x);
411 }
412 
413 static int32_t
414 z_roundpow2(int32_t v)
415 {
416 	int32_t i;
417 
418 	i = 1;
419 
420 	/*
421 	 * Let it overflow at will..
422 	 */
423 	while (i > 0 && i < v)
424 		i <<= 1;
425 
426 	return (i);
427 }
428 
429 /*
430  * Zero Order Hold, the worst of the worst, an insult against quality,
431  * but super fast.
432  */
433 static void
434 z_feed_zoh(struct z_info *info, uint8_t *dst)
435 {
436 #if 0
437 	z_copy(info->z_delay +
438 	    (info->z_start * info->channels * info->bps), dst,
439 	    info->channels * info->bps);
440 #else
441 	uint32_t cnt;
442 	uint8_t *src;
443 
444 	cnt = info->channels * info->bps;
445 	src = info->z_delay + (info->z_start * cnt);
446 
447 	/*
448 	 * This is a bit faster than doing bcopy() since we're dealing
449 	 * with possible unaligned samples.
450 	 */
451 	do {
452 		*dst++ = *src++;
453 	} while (--cnt != 0);
454 #endif
455 }
456 
457 /*
458  * Linear Interpolation. This at least sounds better (perceptually) and fast,
459  * but without any proper filtering which means aliasing still exist and
460  * could become worst with a right sample. Interpolation centered within
461  * Z_LINEAR_ONE between the present and previous sample and everything is
462  * done with simple 32bit scaling arithmetic.
463  */
464 #define Z_DECLARE_LINEAR(SIGN, BIT, ENDIAN)					\
465 static void									\
466 z_feed_linear_##SIGN##BIT##ENDIAN(struct z_info *info, uint8_t *dst)		\
467 {										\
468 	int32_t z;								\
469 	intpcm_t x, y;								\
470 	uint32_t ch;								\
471 	uint8_t *sx, *sy;							\
472 										\
473 	z = ((uint32_t)info->z_alpha * info->z_dx) >> Z_LINEAR_UNSHIFT;		\
474 										\
475 	sx = info->z_delay + (info->z_start * info->channels *			\
476 	    PCM_##BIT##_BPS);							\
477 	sy = sx - (info->channels * PCM_##BIT##_BPS);				\
478 										\
479 	ch = info->channels;							\
480 										\
481 	do {									\
482 		x = _PCM_READ_##SIGN##BIT##_##ENDIAN(sx);			\
483 		y = _PCM_READ_##SIGN##BIT##_##ENDIAN(sy);			\
484 		x = Z_LINEAR_INTERPOLATE_##BIT(z, x, y);			\
485 		_PCM_WRITE_##SIGN##BIT##_##ENDIAN(dst, x);			\
486 		sx += PCM_##BIT##_BPS;						\
487 		sy += PCM_##BIT##_BPS;						\
488 		dst += PCM_##BIT##_BPS;						\
489 	} while (--ch != 0);							\
490 }
491 
492 /*
493  * Userland clipping diagnostic check, not enabled in kernel compilation.
494  * While doing sinc interpolation, unrealistic samples like full scale sine
495  * wav will clip, but for other things this will not make any noise at all.
496  * Everybody should learn how to normalized perceived loudness of their own
497  * music/sounds/samples (hint: ReplayGain).
498  */
499 #ifdef Z_DIAGNOSTIC
500 #define Z_CLIP_CHECK(v, BIT)	do {					\
501 	if ((v) > PCM_S##BIT##_MAX) {					\
502 		fprintf(stderr, "Overflow: v=%jd, max=%jd\n",		\
503 		    (intmax_t)(v), (intmax_t)PCM_S##BIT##_MAX);		\
504 	} else if ((v) < PCM_S##BIT##_MIN) {				\
505 		fprintf(stderr, "Underflow: v=%jd, min=%jd\n",		\
506 		    (intmax_t)(v), (intmax_t)PCM_S##BIT##_MIN);		\
507 	}								\
508 } while (0)
509 #else
510 #define Z_CLIP_CHECK(...)
511 #endif
512 
513 #define Z_CLAMP(v, BIT)							\
514 	(((v) > PCM_S##BIT##_MAX) ? PCM_S##BIT##_MAX :			\
515 	(((v) < PCM_S##BIT##_MIN) ? PCM_S##BIT##_MIN : (v)))
516 
517 /*
518  * Sine Cardinal (SINC) Interpolation. Scaling is done in 64 bit, so
519  * there's no point to hold the plate any longer. All samples will be
520  * shifted to a full 32 bit, scaled and restored during write for
521  * maximum dynamic range (only for downsampling).
522  */
523 #define _Z_SINC_ACCUMULATE(SIGN, BIT, ENDIAN, adv)			\
524 	c += z >> Z_SHIFT;						\
525 	z &= Z_MASK;							\
526 	coeff = Z_COEFF_INTERPOLATE(z, z_coeff[c], z_dcoeff[c]);	\
527 	x = _PCM_READ_##SIGN##BIT##_##ENDIAN(p);			\
528 	v += Z_NORM_##BIT((intpcm64_t)x * coeff);			\
529 	z += info->z_dy;						\
530 	p adv##= info->channels * PCM_##BIT##_BPS
531 
532 /*
533  * XXX GCC4 optimization is such a !@#$%, need manual unrolling.
534  */
535 #if defined(__GNUC__) && __GNUC__ >= 4
536 #define Z_SINC_ACCUMULATE(...)	do {					\
537 	_Z_SINC_ACCUMULATE(__VA_ARGS__);				\
538 	_Z_SINC_ACCUMULATE(__VA_ARGS__);				\
539 } while (0)
540 #define Z_SINC_ACCUMULATE_DECR		2
541 #else
542 #define Z_SINC_ACCUMULATE(...)	do {					\
543 	_Z_SINC_ACCUMULATE(__VA_ARGS__);				\
544 } while (0)
545 #define Z_SINC_ACCUMULATE_DECR		1
546 #endif
547 
548 #define Z_DECLARE_SINC(SIGN, BIT, ENDIAN)					\
549 static void									\
550 z_feed_sinc_##SIGN##BIT##ENDIAN(struct z_info *info, uint8_t *dst)		\
551 {										\
552 	intpcm64_t v;								\
553 	intpcm_t x;								\
554 	uint8_t *p;								\
555 	int32_t coeff, z, *z_coeff, *z_dcoeff;					\
556 	uint32_t c, center, ch, i;						\
557 										\
558 	z_coeff = info->z_coeff;						\
559 	z_dcoeff = info->z_dcoeff;						\
560 	center = z_prev(info, info->z_start, info->z_size);			\
561 	ch = info->channels * PCM_##BIT##_BPS;					\
562 	dst += ch;								\
563 										\
564 	do {									\
565 		dst -= PCM_##BIT##_BPS;						\
566 		ch -= PCM_##BIT##_BPS;						\
567 		v = 0;								\
568 		z = info->z_alpha * info->z_dx;					\
569 		c = 0;								\
570 		p = info->z_delay + (z_next(info, center, 1) *			\
571 		    info->channels * PCM_##BIT##_BPS) + ch;			\
572 		for (i = info->z_size; i != 0; i -= Z_SINC_ACCUMULATE_DECR) 	\
573 			Z_SINC_ACCUMULATE(SIGN, BIT, ENDIAN, +);		\
574 		z = info->z_dy - (info->z_alpha * info->z_dx);			\
575 		c = 0;								\
576 		p = info->z_delay + (center * info->channels *			\
577 		    PCM_##BIT##_BPS) + ch;					\
578 		for (i = info->z_size; i != 0; i -= Z_SINC_ACCUMULATE_DECR) 	\
579 			Z_SINC_ACCUMULATE(SIGN, BIT, ENDIAN, -);		\
580 		if (info->z_scale != Z_ONE)					\
581 			v = Z_SCALE_##BIT(v, info->z_scale);			\
582 		else								\
583 			v >>= Z_COEFF_SHIFT - Z_GUARD_BIT_##BIT;		\
584 		Z_CLIP_CHECK(v, BIT);						\
585 		_PCM_WRITE_##SIGN##BIT##_##ENDIAN(dst, Z_CLAMP(v, BIT));	\
586 	} while (ch != 0);							\
587 }
588 
589 #define Z_DECLARE_SINC_POLYPHASE(SIGN, BIT, ENDIAN)				\
590 static void									\
591 z_feed_sinc_polyphase_##SIGN##BIT##ENDIAN(struct z_info *info, uint8_t *dst)	\
592 {										\
593 	intpcm64_t v;								\
594 	intpcm_t x;								\
595 	uint8_t *p;								\
596 	int32_t ch, i, start, *z_pcoeff;					\
597 										\
598 	ch = info->channels * PCM_##BIT##_BPS;					\
599 	dst += ch;								\
600 	start = z_prev(info, info->z_start, (info->z_size << 1) - 1) * ch;	\
601 										\
602 	do {									\
603 		dst -= PCM_##BIT##_BPS;						\
604 		ch -= PCM_##BIT##_BPS;						\
605 		v = 0;								\
606 		p = info->z_delay + start + ch;					\
607 		z_pcoeff = info->z_pcoeff +					\
608 		    ((info->z_alpha * info->z_size) << 1);			\
609 		for (i = info->z_size; i != 0; i--) {				\
610 			x = _PCM_READ_##SIGN##BIT##_##ENDIAN(p);		\
611 			v += Z_NORM_##BIT((intpcm64_t)x * *z_pcoeff);		\
612 			z_pcoeff++;						\
613 			p += info->channels * PCM_##BIT##_BPS;			\
614 			x = _PCM_READ_##SIGN##BIT##_##ENDIAN(p);		\
615 			v += Z_NORM_##BIT((intpcm64_t)x * *z_pcoeff);		\
616 			z_pcoeff++;						\
617 			p += info->channels * PCM_##BIT##_BPS;			\
618 		}								\
619 		if (info->z_scale != Z_ONE)					\
620 			v = Z_SCALE_##BIT(v, info->z_scale);			\
621 		else								\
622 			v >>= Z_COEFF_SHIFT - Z_GUARD_BIT_##BIT;		\
623 		Z_CLIP_CHECK(v, BIT);						\
624 		_PCM_WRITE_##SIGN##BIT##_##ENDIAN(dst, Z_CLAMP(v, BIT));	\
625 	} while (ch != 0);							\
626 }
627 
628 #define Z_DECLARE(SIGN, BIT, ENDIAN)					\
629 	Z_DECLARE_LINEAR(SIGN, BIT, ENDIAN)				\
630 	Z_DECLARE_SINC(SIGN, BIT, ENDIAN)				\
631 	Z_DECLARE_SINC_POLYPHASE(SIGN, BIT, ENDIAN)
632 
633 #if BYTE_ORDER == LITTLE_ENDIAN || defined(SND_FEEDER_MULTIFORMAT)
634 Z_DECLARE(S, 16, LE)
635 Z_DECLARE(S, 32, LE)
636 #endif
637 #if BYTE_ORDER == BIG_ENDIAN || defined(SND_FEEDER_MULTIFORMAT)
638 Z_DECLARE(S, 16, BE)
639 Z_DECLARE(S, 32, BE)
640 #endif
641 #ifdef SND_FEEDER_MULTIFORMAT
642 Z_DECLARE(S,  8, NE)
643 Z_DECLARE(S, 24, LE)
644 Z_DECLARE(S, 24, BE)
645 Z_DECLARE(U,  8, NE)
646 Z_DECLARE(U, 16, LE)
647 Z_DECLARE(U, 24, LE)
648 Z_DECLARE(U, 32, LE)
649 Z_DECLARE(U, 16, BE)
650 Z_DECLARE(U, 24, BE)
651 Z_DECLARE(U, 32, BE)
652 #endif
653 
654 enum {
655 	Z_RESAMPLER_ZOH,
656 	Z_RESAMPLER_LINEAR,
657 	Z_RESAMPLER_SINC,
658 	Z_RESAMPLER_SINC_POLYPHASE,
659 	Z_RESAMPLER_LAST
660 };
661 
662 #define Z_RESAMPLER_IDX(i)						\
663 	(Z_IS_SINC(i) ? Z_RESAMPLER_SINC : (i)->quality)
664 
665 #define Z_RESAMPLER_ENTRY(SIGN, BIT, ENDIAN)					\
666 	{									\
667 	    AFMT_##SIGN##BIT##_##ENDIAN,					\
668 	    {									\
669 		[Z_RESAMPLER_ZOH]    = z_feed_zoh,				\
670 		[Z_RESAMPLER_LINEAR] = z_feed_linear_##SIGN##BIT##ENDIAN,	\
671 		[Z_RESAMPLER_SINC]   = z_feed_sinc_##SIGN##BIT##ENDIAN,		\
672 		[Z_RESAMPLER_SINC_POLYPHASE]   =				\
673 		    z_feed_sinc_polyphase_##SIGN##BIT##ENDIAN			\
674 	    }									\
675 	}
676 
677 static const struct {
678 	uint32_t format;
679 	z_resampler_t resampler[Z_RESAMPLER_LAST];
680 } z_resampler_tab[] = {
681 #if BYTE_ORDER == LITTLE_ENDIAN || defined(SND_FEEDER_MULTIFORMAT)
682 	Z_RESAMPLER_ENTRY(S, 16, LE),
683 	Z_RESAMPLER_ENTRY(S, 32, LE),
684 #endif
685 #if BYTE_ORDER == BIG_ENDIAN || defined(SND_FEEDER_MULTIFORMAT)
686 	Z_RESAMPLER_ENTRY(S, 16, BE),
687 	Z_RESAMPLER_ENTRY(S, 32, BE),
688 #endif
689 #ifdef SND_FEEDER_MULTIFORMAT
690 	Z_RESAMPLER_ENTRY(S,  8, NE),
691 	Z_RESAMPLER_ENTRY(S, 24, LE),
692 	Z_RESAMPLER_ENTRY(S, 24, BE),
693 	Z_RESAMPLER_ENTRY(U,  8, NE),
694 	Z_RESAMPLER_ENTRY(U, 16, LE),
695 	Z_RESAMPLER_ENTRY(U, 24, LE),
696 	Z_RESAMPLER_ENTRY(U, 32, LE),
697 	Z_RESAMPLER_ENTRY(U, 16, BE),
698 	Z_RESAMPLER_ENTRY(U, 24, BE),
699 	Z_RESAMPLER_ENTRY(U, 32, BE),
700 #endif
701 };
702 
703 #define Z_RESAMPLER_TAB_SIZE						\
704 	((int32_t)(sizeof(z_resampler_tab) / sizeof(z_resampler_tab[0])))
705 
706 static void
707 z_resampler_reset(struct z_info *info)
708 {
709 
710 	info->src = info->rsrc - (info->rsrc % ((feeder_rate_round > 0 &&
711 	    info->rsrc > feeder_rate_round) ? feeder_rate_round : 1));
712 	info->dst = info->rdst - (info->rdst % ((feeder_rate_round > 0 &&
713 	    info->rdst > feeder_rate_round) ? feeder_rate_round : 1));
714 	info->z_gx = 1;
715 	info->z_gy = 1;
716 	info->z_alpha = 0;
717 	info->z_resample = NULL;
718 	info->z_size = 1;
719 	info->z_coeff = NULL;
720 	info->z_dcoeff = NULL;
721 	if (info->z_pcoeff != NULL) {
722 		free(info->z_pcoeff, M_DEVBUF);
723 		info->z_pcoeff = NULL;
724 	}
725 	info->z_scale = Z_ONE;
726 	info->z_dx = Z_FULL_ONE;
727 	info->z_dy = Z_FULL_ONE;
728 #ifdef Z_DIAGNOSTIC
729 	info->z_cycle = 0;
730 #endif
731 	if (info->quality < Z_QUALITY_MIN)
732 		info->quality = Z_QUALITY_MIN;
733 	else if (info->quality > Z_QUALITY_MAX)
734 		info->quality = Z_QUALITY_MAX;
735 }
736 
737 #ifdef Z_PARANOID
738 static int32_t
739 z_resampler_sinc_len(struct z_info *info)
740 {
741 	int32_t c, z, len, lmax;
742 
743 	if (!Z_IS_SINC(info))
744 		return (1);
745 
746 	/*
747 	 * A rather careful (or useless) way to calculate filter length.
748 	 * Z_SINC_LEN() itself is accurate enough to do its job. Extra
749 	 * sanity checking is not going to hurt though..
750 	 */
751 	c = 0;
752 	z = info->z_dy;
753 	len = 0;
754 	lmax = z_coeff_tab[Z_SINC_COEFF_IDX(info)].len;
755 
756 	do {
757 		c += z >> Z_SHIFT;
758 		z &= Z_MASK;
759 		z += info->z_dy;
760 	} while (c < lmax && ++len > 0);
761 
762 	if (len != Z_SINC_LEN(info)) {
763 #ifdef _KERNEL
764 		printf("%s(): sinc l=%d != Z_SINC_LEN=%d\n",
765 		    __func__, len, Z_SINC_LEN(info));
766 #else
767 		fprintf(stderr, "%s(): sinc l=%d != Z_SINC_LEN=%d\n",
768 		    __func__, len, Z_SINC_LEN(info));
769 		return (-1);
770 #endif
771 	}
772 
773 	return (len);
774 }
775 #else
776 #define z_resampler_sinc_len(i)		(Z_IS_SINC(i) ? Z_SINC_LEN(i) : 1)
777 #endif
778 
779 #define Z_POLYPHASE_COEFF_SHIFT		0
780 
781 /*
782  * Pick suitable polynomial interpolators based on filter oversampled ratio
783  * (2 ^ Z_DRIFT_SHIFT).
784  */
785 #if !(defined(Z_COEFF_INTERP_ZOH) || defined(Z_COEFF_INTERP_LINEAR) ||		\
786     defined(Z_COEFF_INTERP_QUADRATIC) || defined(Z_COEFF_INTERP_HERMITE) ||	\
787     defined(Z_COEFF_INTER_BSPLINE) || defined(Z_COEFF_INTERP_OPT32X) ||		\
788     defined(Z_COEFF_INTERP_OPT16X) || defined(Z_COEFF_INTERP_OPT8X) ||		\
789     defined(Z_COEFF_INTERP_OPT4X) || defined(Z_COEFF_INTERP_OPT2X))
790 #if Z_DRIFT_SHIFT >= 6
791 #define Z_COEFF_INTERP_BSPLINE		1
792 #elif Z_DRIFT_SHIFT >= 5
793 #define Z_COEFF_INTERP_OPT32X		1
794 #elif Z_DRIFT_SHIFT == 4
795 #define Z_COEFF_INTERP_OPT16X		1
796 #elif Z_DRIFT_SHIFT == 3
797 #define Z_COEFF_INTERP_OPT8X		1
798 #elif Z_DRIFT_SHIFT == 2
799 #define Z_COEFF_INTERP_OPT4X		1
800 #elif Z_DRIFT_SHIFT == 1
801 #define Z_COEFF_INTERP_OPT2X		1
802 #else
803 #error "Z_DRIFT_SHIFT screwed!"
804 #endif
805 #endif
806 
807 /*
808  * In classic polyphase mode, the actual coefficients for each phases need to
809  * be calculated based on default prototype filters. For highly oversampled
810  * filter, linear or quadradatic interpolator should be enough. Anything less
811  * than that require 'special' interpolators to reduce interpolation errors.
812  *
813  * "Polynomial Interpolators for High-Quality Resampling of Oversampled Audio"
814  *    by Olli Niemitalo
815  *    - http://www.student.oulu.fi/~oniemita/dsp/deip.pdf
816  *
817  */
818 static int32_t
819 z_coeff_interpolate(int32_t z, int32_t *z_coeff)
820 {
821 	int32_t coeff;
822 #if defined(Z_COEFF_INTERP_ZOH)
823 
824 	/* 1-point, 0th-order (Zero Order Hold) */
825 	z = z;
826 	coeff = z_coeff[0];
827 #elif defined(Z_COEFF_INTERP_LINEAR)
828 	int32_t zl0, zl1;
829 
830 	/* 2-point, 1st-order Linear */
831 	zl0 = z_coeff[0];
832 	zl1 = z_coeff[1] - z_coeff[0];
833 
834 	coeff = Z_RSHIFT((int64_t)zl1 * z, Z_SHIFT) + zl0;
835 #elif defined(Z_COEFF_INTERP_QUADRATIC)
836 	int32_t zq0, zq1, zq2;
837 
838 	/* 3-point, 2nd-order Quadratic */
839 	zq0 = z_coeff[0];
840 	zq1 = z_coeff[1] - z_coeff[-1];
841 	zq2 = z_coeff[1] + z_coeff[-1] - (z_coeff[0] << 1);
842 
843 	coeff = Z_RSHIFT((Z_RSHIFT((int64_t)zq2 * z, Z_SHIFT) +
844 	    zq1) * z, Z_SHIFT + 1) + zq0;
845 #elif defined(Z_COEFF_INTERP_HERMITE)
846 	int32_t zh0, zh1, zh2, zh3;
847 
848 	/* 4-point, 3rd-order Hermite */
849 	zh0 = z_coeff[0];
850 	zh1 = z_coeff[1] - z_coeff[-1];
851 	zh2 = (z_coeff[-1] << 1) - (z_coeff[0] * 5) + (z_coeff[1] << 2) -
852 	    z_coeff[2];
853 	zh3 = z_coeff[2] - z_coeff[-1] + ((z_coeff[0] - z_coeff[1]) * 3);
854 
855 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((int64_t)zh3 * z, Z_SHIFT) +
856 	    zh2) * z, Z_SHIFT) + zh1) * z, Z_SHIFT + 1) + zh0;
857 #elif defined(Z_COEFF_INTERP_BSPLINE)
858 	int32_t zb0, zb1, zb2, zb3;
859 
860 	/* 4-point, 3rd-order B-Spline */
861 	zb0 = Z_RSHIFT(0x15555555LL * (((int64_t)z_coeff[0] << 2) +
862 	    z_coeff[-1] + z_coeff[1]), 30);
863 	zb1 = z_coeff[1] - z_coeff[-1];
864 	zb2 = z_coeff[-1] + z_coeff[1] - (z_coeff[0] << 1);
865 	zb3 = Z_RSHIFT(0x15555555LL * (((z_coeff[0] - z_coeff[1]) * 3) +
866 	    z_coeff[2] - z_coeff[-1]), 30);
867 
868 	coeff = (Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((int64_t)zb3 * z, Z_SHIFT) +
869 	    zb2) * z, Z_SHIFT) + zb1) * z, Z_SHIFT) + zb0 + 1) >> 1;
870 #elif defined(Z_COEFF_INTERP_OPT32X)
871 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
872 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
873 
874 	/* 6-point, 5th-order Optimal 32x */
875 	zoz = z - (Z_ONE >> 1);
876 	zoe1 = z_coeff[1] + z_coeff[0];
877 	zoe2 = z_coeff[2] + z_coeff[-1];
878 	zoe3 = z_coeff[3] + z_coeff[-2];
879 	zoo1 = z_coeff[1] - z_coeff[0];
880 	zoo2 = z_coeff[2] - z_coeff[-1];
881 	zoo3 = z_coeff[3] - z_coeff[-2];
882 
883 	zoc0 = Z_RSHIFT((0x1ac2260dLL * zoe1) + (0x0526cdcaLL * zoe2) +
884 	    (0x00170c29LL * zoe3), 30);
885 	zoc1 = Z_RSHIFT((0x14f8a49aLL * zoo1) + (0x0d6d1109LL * zoo2) +
886 	    (0x008cd4dcLL * zoo3), 30);
887 	zoc2 = Z_RSHIFT((-0x0d3e94a4LL * zoe1) + (0x0bddded4LL * zoe2) +
888 	    (0x0160b5d0LL * zoe3), 30);
889 	zoc3 = Z_RSHIFT((-0x0de10cc4LL * zoo1) + (0x019b2a7dLL * zoo2) +
890 	    (0x01cfe914LL * zoo3), 30);
891 	zoc4 = Z_RSHIFT((0x02aa12d7LL * zoe1) + (-0x03ff1bb3LL * zoe2) +
892 	    (0x015508ddLL * zoe3), 30);
893 	zoc5 = Z_RSHIFT((0x051d29e5LL * zoo1) + (-0x028e7647LL * zoo2) +
894 	    (0x0082d81aLL * zoo3), 30);
895 
896 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
897 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
898 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
899 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
900 #elif defined(Z_COEFF_INTERP_OPT16X)
901 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
902 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
903 
904 	/* 6-point, 5th-order Optimal 16x */
905 	zoz = z - (Z_ONE >> 1);
906 	zoe1 = z_coeff[1] + z_coeff[0];
907 	zoe2 = z_coeff[2] + z_coeff[-1];
908 	zoe3 = z_coeff[3] + z_coeff[-2];
909 	zoo1 = z_coeff[1] - z_coeff[0];
910 	zoo2 = z_coeff[2] - z_coeff[-1];
911 	zoo3 = z_coeff[3] - z_coeff[-2];
912 
913 	zoc0 = Z_RSHIFT((0x1ac2260dLL * zoe1) + (0x0526cdcaLL * zoe2) +
914 	    (0x00170c29LL * zoe3), 30);
915 	zoc1 = Z_RSHIFT((0x14f8a49aLL * zoo1) + (0x0d6d1109LL * zoo2) +
916 	    (0x008cd4dcLL * zoo3), 30);
917 	zoc2 = Z_RSHIFT((-0x0d3e94a4LL * zoe1) + (0x0bddded4LL * zoe2) +
918 	    (0x0160b5d0LL * zoe3), 30);
919 	zoc3 = Z_RSHIFT((-0x0de10cc4LL * zoo1) + (0x019b2a7dLL * zoo2) +
920 	    (0x01cfe914LL * zoo3), 30);
921 	zoc4 = Z_RSHIFT((0x02aa12d7LL * zoe1) + (-0x03ff1bb3LL * zoe2) +
922 	    (0x015508ddLL * zoe3), 30);
923 	zoc5 = Z_RSHIFT((0x051d29e5LL * zoo1) + (-0x028e7647LL * zoo2) +
924 	    (0x0082d81aLL * zoo3), 30);
925 
926 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
927 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
928 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
929 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
930 #elif defined(Z_COEFF_INTERP_OPT8X)
931 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
932 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
933 
934 	/* 6-point, 5th-order Optimal 8x */
935 	zoz = z - (Z_ONE >> 1);
936 	zoe1 = z_coeff[1] + z_coeff[0];
937 	zoe2 = z_coeff[2] + z_coeff[-1];
938 	zoe3 = z_coeff[3] + z_coeff[-2];
939 	zoo1 = z_coeff[1] - z_coeff[0];
940 	zoo2 = z_coeff[2] - z_coeff[-1];
941 	zoo3 = z_coeff[3] - z_coeff[-2];
942 
943 	zoc0 = Z_RSHIFT((0x1aa9b47dLL * zoe1) + (0x053d9944LL * zoe2) +
944 	    (0x0018b23fLL * zoe3), 30);
945 	zoc1 = Z_RSHIFT((0x14a104d1LL * zoo1) + (0x0d7d2504LL * zoo2) +
946 	    (0x0094b599LL * zoo3), 30);
947 	zoc2 = Z_RSHIFT((-0x0d22530bLL * zoe1) + (0x0bb37a2cLL * zoe2) +
948 	    (0x016ed8e0LL * zoe3), 30);
949 	zoc3 = Z_RSHIFT((-0x0d744b1cLL * zoo1) + (0x01649591LL * zoo2) +
950 	    (0x01dae93aLL * zoo3), 30);
951 	zoc4 = Z_RSHIFT((0x02a7ee1bLL * zoe1) + (-0x03fbdb24LL * zoe2) +
952 	    (0x0153ed07LL * zoe3), 30);
953 	zoc5 = Z_RSHIFT((0x04cf9b6cLL * zoo1) + (-0x0266b378LL * zoo2) +
954 	    (0x007a7c26LL * zoo3), 30);
955 
956 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
957 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
958 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
959 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
960 #elif defined(Z_COEFF_INTERP_OPT4X)
961 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
962 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
963 
964 	/* 6-point, 5th-order Optimal 4x */
965 	zoz = z - (Z_ONE >> 1);
966 	zoe1 = z_coeff[1] + z_coeff[0];
967 	zoe2 = z_coeff[2] + z_coeff[-1];
968 	zoe3 = z_coeff[3] + z_coeff[-2];
969 	zoo1 = z_coeff[1] - z_coeff[0];
970 	zoo2 = z_coeff[2] - z_coeff[-1];
971 	zoo3 = z_coeff[3] - z_coeff[-2];
972 
973 	zoc0 = Z_RSHIFT((0x1a8eda43LL * zoe1) + (0x0556ee38LL * zoe2) +
974 	    (0x001a3784LL * zoe3), 30);
975 	zoc1 = Z_RSHIFT((0x143d863eLL * zoo1) + (0x0d910e36LL * zoo2) +
976 	    (0x009ca889LL * zoo3), 30);
977 	zoc2 = Z_RSHIFT((-0x0d026821LL * zoe1) + (0x0b837773LL * zoe2) +
978 	    (0x017ef0c6LL * zoe3), 30);
979 	zoc3 = Z_RSHIFT((-0x0cef1502LL * zoo1) + (0x01207a8eLL * zoo2) +
980 	    (0x01e936dbLL * zoo3), 30);
981 	zoc4 = Z_RSHIFT((0x029fe643LL * zoe1) + (-0x03ef3fc8LL * zoe2) +
982 	    (0x014f5923LL * zoe3), 30);
983 	zoc5 = Z_RSHIFT((0x043a9d08LL * zoo1) + (-0x02154febLL * zoo2) +
984 	    (0x00670dbdLL * zoo3), 30);
985 
986 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
987 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
988 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
989 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
990 #elif defined(Z_COEFF_INTERP_OPT2X)
991 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
992 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
993 
994 	/* 6-point, 5th-order Optimal 2x */
995 	zoz = z - (Z_ONE >> 1);
996 	zoe1 = z_coeff[1] + z_coeff[0];
997 	zoe2 = z_coeff[2] + z_coeff[-1];
998 	zoe3 = z_coeff[3] + z_coeff[-2];
999 	zoo1 = z_coeff[1] - z_coeff[0];
1000 	zoo2 = z_coeff[2] - z_coeff[-1];
1001 	zoo3 = z_coeff[3] - z_coeff[-2];
1002 
1003 	zoc0 = Z_RSHIFT((0x19edb6fdLL * zoe1) + (0x05ebd062LL * zoe2) +
1004 	    (0x00267881LL * zoe3), 30);
1005 	zoc1 = Z_RSHIFT((0x1223af76LL * zoo1) + (0x0de3dd6bLL * zoo2) +
1006 	    (0x00d683cdLL * zoo3), 30);
1007 	zoc2 = Z_RSHIFT((-0x0c3ee068LL * zoe1) + (0x0a5c3769LL * zoe2) +
1008 	    (0x01e2aceaLL * zoe3), 30);
1009 	zoc3 = Z_RSHIFT((-0x0a8ab614LL * zoo1) + (-0x0019522eLL * zoo2) +
1010 	    (0x022cefc7LL * zoo3), 30);
1011 	zoc4 = Z_RSHIFT((0x0276187dLL * zoe1) + (-0x03a801e8LL * zoe2) +
1012 	    (0x0131d935LL * zoe3), 30);
1013 	zoc5 = Z_RSHIFT((0x02c373f5LL * zoo1) + (-0x01275f83LL * zoo2) +
1014 	    (0x0018ee79LL * zoo3), 30);
1015 
1016 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
1017 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
1018 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
1019 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
1020 #else
1021 #error "Interpolation type screwed!"
1022 #endif
1023 
1024 #if Z_POLYPHASE_COEFF_SHIFT > 0
1025 	coeff = Z_RSHIFT(coeff, Z_POLYPHASE_COEFF_SHIFT);
1026 #endif
1027 	return (coeff);
1028 }
1029 
1030 static int
1031 z_resampler_build_polyphase(struct z_info *info)
1032 {
1033 	int32_t alpha, c, i, z, idx;
1034 
1035 	/* Let this be here first. */
1036 	if (info->z_pcoeff != NULL) {
1037 		free(info->z_pcoeff, M_DEVBUF);
1038 		info->z_pcoeff = NULL;
1039 	}
1040 
1041 	if (feeder_rate_polyphase_max < 1)
1042 		return (ENOTSUP);
1043 
1044 	if (((int64_t)info->z_size * info->z_gy * 2) >
1045 	    feeder_rate_polyphase_max) {
1046 #ifndef _KERNEL
1047 		fprintf(stderr, "Polyphase entries exceed: [%d/%d] %jd > %d\n",
1048 		    info->z_gx, info->z_gy,
1049 		    (intmax_t)info->z_size * info->z_gy * 2,
1050 		    feeder_rate_polyphase_max);
1051 #endif
1052 		return (E2BIG);
1053 	}
1054 
1055 	info->z_pcoeff = malloc(sizeof(int32_t) *
1056 	    info->z_size * info->z_gy * 2, M_DEVBUF, M_NOWAIT | M_ZERO);
1057 	if (info->z_pcoeff == NULL)
1058 		return (ENOMEM);
1059 
1060 	for (alpha = 0; alpha < info->z_gy; alpha++) {
1061 		z = alpha * info->z_dx;
1062 		c = 0;
1063 		for (i = info->z_size; i != 0; i--) {
1064 			c += z >> Z_SHIFT;
1065 			z &= Z_MASK;
1066 			idx = (alpha * info->z_size * 2) +
1067 			    (info->z_size * 2) - i;
1068 			info->z_pcoeff[idx] =
1069 			    z_coeff_interpolate(z, info->z_coeff + c);
1070 			z += info->z_dy;
1071 		}
1072 		z = info->z_dy - (alpha * info->z_dx);
1073 		c = 0;
1074 		for (i = info->z_size; i != 0; i--) {
1075 			c += z >> Z_SHIFT;
1076 			z &= Z_MASK;
1077 			idx = (alpha * info->z_size * 2) + i - 1;
1078 			info->z_pcoeff[idx] =
1079 			    z_coeff_interpolate(z, info->z_coeff + c);
1080 			z += info->z_dy;
1081 		}
1082 	}
1083 
1084 #ifndef _KERNEL
1085 	fprintf(stderr, "Polyphase: [%d/%d] %d entries\n",
1086 	    info->z_gx, info->z_gy, info->z_size * info->z_gy * 2);
1087 #endif
1088 
1089 	return (0);
1090 }
1091 
1092 static int
1093 z_resampler_setup(struct pcm_feeder *f)
1094 {
1095 	struct z_info *info;
1096 	int64_t gy2gx_max, gx2gy_max;
1097 	uint32_t format;
1098 	int32_t align, i, z_scale;
1099 	int adaptive;
1100 
1101 	info = f->data;
1102 	z_resampler_reset(info);
1103 
1104 	if (info->src == info->dst)
1105 		return (0);
1106 
1107 	/* Shrink by greatest common divisor. */
1108 	i = z_gcd(info->src, info->dst);
1109 	info->z_gx = info->src / i;
1110 	info->z_gy = info->dst / i;
1111 
1112 	/* Too big, or too small. Bail out. */
1113 	if (!(Z_FACTOR_SAFE(info->z_gx) && Z_FACTOR_SAFE(info->z_gy)))
1114 		return (EINVAL);
1115 
1116 	format = f->desc->in;
1117 	adaptive = 0;
1118 	z_scale = 0;
1119 
1120 	/*
1121 	 * Setup everything: filter length, conversion factor, etc.
1122 	 */
1123 	if (Z_IS_SINC(info)) {
1124 		/*
1125 		 * Downsampling, or upsampling scaling factor. As long as the
1126 		 * factor can be represented by a fraction of 1 << Z_SHIFT,
1127 		 * we're pretty much in business. Scaling is not needed for
1128 		 * upsampling, so we just slap Z_ONE there.
1129 		 */
1130 		if (info->z_gx > info->z_gy)
1131 			/*
1132 			 * If the downsampling ratio is beyond sanity,
1133 			 * enable semi-adaptive mode. Although handling
1134 			 * extreme ratio is possible, the result of the
1135 			 * conversion is just pointless, unworthy,
1136 			 * nonsensical noises, etc.
1137 			 */
1138 			if ((info->z_gx / info->z_gy) > Z_SINC_DOWNMAX)
1139 				z_scale = Z_ONE / Z_SINC_DOWNMAX;
1140 			else
1141 				z_scale = ((uint64_t)info->z_gy << Z_SHIFT) /
1142 				    info->z_gx;
1143 		else
1144 			z_scale = Z_ONE;
1145 
1146 		/*
1147 		 * This is actually impossible, unless anything above
1148 		 * overflow.
1149 		 */
1150 		if (z_scale < 1)
1151 			return (E2BIG);
1152 
1153 		/*
1154 		 * Calculate sample time/coefficients index drift. It is
1155 		 * a constant for upsampling, but downsampling require
1156 		 * heavy duty filtering with possible too long filters.
1157 		 * If anything goes wrong, revisit again and enable
1158 		 * adaptive mode.
1159 		 */
1160 z_setup_adaptive_sinc:
1161 		if (info->z_pcoeff != NULL) {
1162 			free(info->z_pcoeff, M_DEVBUF);
1163 			info->z_pcoeff = NULL;
1164 		}
1165 
1166 		if (adaptive == 0) {
1167 			info->z_dy = z_scale << Z_DRIFT_SHIFT;
1168 			if (info->z_dy < 1)
1169 				return (E2BIG);
1170 			info->z_scale = z_scale;
1171 		} else {
1172 			info->z_dy = Z_FULL_ONE;
1173 			info->z_scale = Z_ONE;
1174 		}
1175 
1176 #if 0
1177 #define Z_SCALE_DIV	10000
1178 #define Z_SCALE_LIMIT(s, v)						\
1179 	((((uint64_t)(s) * (v)) + (Z_SCALE_DIV >> 1)) / Z_SCALE_DIV)
1180 
1181 		info->z_scale = Z_SCALE_LIMIT(info->z_scale, 9780);
1182 #endif
1183 
1184 		/* Smallest drift increment. */
1185 		info->z_dx = info->z_dy / info->z_gy;
1186 
1187 		/*
1188 		 * Overflow or underflow. Try adaptive, let it continue and
1189 		 * retry.
1190 		 */
1191 		if (info->z_dx < 1) {
1192 			if (adaptive == 0) {
1193 				adaptive = 1;
1194 				goto z_setup_adaptive_sinc;
1195 			}
1196 			return (E2BIG);
1197 		}
1198 
1199 		/*
1200 		 * Round back output drift.
1201 		 */
1202 		info->z_dy = info->z_dx * info->z_gy;
1203 
1204 		for (i = 0; i < Z_COEFF_TAB_SIZE; i++) {
1205 			if (Z_SINC_COEFF_IDX(info) != i)
1206 				continue;
1207 			/*
1208 			 * Calculate required filter length and guard
1209 			 * against possible abusive result. Note that
1210 			 * this represents only 1/2 of the entire filter
1211 			 * length.
1212 			 */
1213 			info->z_size = z_resampler_sinc_len(info);
1214 
1215 			/*
1216 			 * Multiple of 2 rounding, for better accumulator
1217 			 * performance.
1218 			 */
1219 			info->z_size &= ~1;
1220 
1221 			if (info->z_size < 2 || info->z_size > Z_SINC_MAX) {
1222 				if (adaptive == 0) {
1223 					adaptive = 1;
1224 					goto z_setup_adaptive_sinc;
1225 				}
1226 				return (E2BIG);
1227 			}
1228 			info->z_coeff = z_coeff_tab[i].coeff + Z_COEFF_OFFSET;
1229 			info->z_dcoeff = z_coeff_tab[i].dcoeff;
1230 			break;
1231 		}
1232 
1233 		if (info->z_coeff == NULL || info->z_dcoeff == NULL)
1234 			return (EINVAL);
1235 	} else if (Z_IS_LINEAR(info)) {
1236 		/*
1237 		 * Don't put much effort if we're doing linear interpolation.
1238 		 * Just center the interpolation distance within Z_LINEAR_ONE,
1239 		 * and be happy about it.
1240 		 */
1241 		info->z_dx = Z_LINEAR_FULL_ONE / info->z_gy;
1242 	}
1243 
1244 	/*
1245 	 * We're safe for now, lets continue.. Look for our resampler
1246 	 * depending on configured format and quality.
1247 	 */
1248 	for (i = 0; i < Z_RESAMPLER_TAB_SIZE; i++) {
1249 		int ridx;
1250 
1251 		if (AFMT_ENCODING(format) != z_resampler_tab[i].format)
1252 			continue;
1253 		if (Z_IS_SINC(info) && adaptive == 0 &&
1254 		    z_resampler_build_polyphase(info) == 0)
1255 			ridx = Z_RESAMPLER_SINC_POLYPHASE;
1256 		else
1257 			ridx = Z_RESAMPLER_IDX(info);
1258 		info->z_resample = z_resampler_tab[i].resampler[ridx];
1259 		break;
1260 	}
1261 
1262 	if (info->z_resample == NULL)
1263 		return (EINVAL);
1264 
1265 	info->bps = AFMT_BPS(format);
1266 	align = info->channels * info->bps;
1267 
1268 	/*
1269 	 * Calculate largest value that can be fed into z_gy2gx() and
1270 	 * z_gx2gy() without causing (signed) 32bit overflow. z_gy2gx() will
1271 	 * be called early during feeding process to determine how much input
1272 	 * samples that is required to generate requested output, while
1273 	 * z_gx2gy() will be called just before samples filtering /
1274 	 * accumulation process based on available samples that has been
1275 	 * calculated using z_gx2gy().
1276 	 *
1277 	 * Now that is damn confusing, I guess ;-) .
1278 	 */
1279 	gy2gx_max = (((uint64_t)info->z_gy * INT32_MAX) - info->z_gy + 1) /
1280 	    info->z_gx;
1281 
1282 	if ((gy2gx_max * align) > SND_FXDIV_MAX)
1283 		gy2gx_max = SND_FXDIV_MAX / align;
1284 
1285 	if (gy2gx_max < 1)
1286 		return (E2BIG);
1287 
1288 	gx2gy_max = (((uint64_t)info->z_gx * INT32_MAX) - info->z_gy) /
1289 	    info->z_gy;
1290 
1291 	if (gx2gy_max > INT32_MAX)
1292 		gx2gy_max = INT32_MAX;
1293 
1294 	if (gx2gy_max < 1)
1295 		return (E2BIG);
1296 
1297 	/*
1298 	 * Ensure that z_gy2gx() at its largest possible calculated value
1299 	 * (alpha = 0) will not cause overflow further late during z_gx2gy()
1300 	 * stage.
1301 	 */
1302 	if (z_gy2gx(info, gy2gx_max) > _Z_GCAST(gx2gy_max))
1303 		return (E2BIG);
1304 
1305 	info->z_maxfeed = gy2gx_max * align;
1306 
1307 #ifdef Z_USE_ALPHADRIFT
1308 	info->z_startdrift = z_gy2gx(info, 1);
1309 	info->z_alphadrift = z_drift(info, info->z_startdrift, 1);
1310 #endif
1311 
1312 	i = z_gy2gx(info, 1);
1313 	info->z_full = z_roundpow2((info->z_size << 1) + i);
1314 
1315 	/*
1316 	 * Too big to be true, and overflowing left and right like mad ..
1317 	 */
1318 	if ((info->z_full * align) < 1) {
1319 		if (adaptive == 0 && Z_IS_SINC(info)) {
1320 			adaptive = 1;
1321 			goto z_setup_adaptive_sinc;
1322 		}
1323 		return (E2BIG);
1324 	}
1325 
1326 	/*
1327 	 * Increase full buffer size if its too small to reduce cyclic
1328 	 * buffer shifting in main conversion/feeder loop.
1329 	 */
1330 	while (info->z_full < Z_RESERVOIR_MAX &&
1331 	    (info->z_full - (info->z_size << 1)) < Z_RESERVOIR)
1332 		info->z_full <<= 1;
1333 
1334 	/* Initialize buffer position. */
1335 	info->z_mask = info->z_full - 1;
1336 	info->z_start = z_prev(info, info->z_size << 1, 1);
1337 	info->z_pos = z_next(info, info->z_start, 1);
1338 
1339 	/*
1340 	 * Allocate or reuse delay line buffer, whichever makes sense.
1341 	 */
1342 	i = info->z_full * align;
1343 	if (i < 1)
1344 		return (E2BIG);
1345 
1346 	if (info->z_delay == NULL || info->z_alloc < i ||
1347 	    i <= (info->z_alloc >> 1)) {
1348 		if (info->z_delay != NULL)
1349 			free(info->z_delay, M_DEVBUF);
1350 		info->z_delay = malloc(i, M_DEVBUF, M_NOWAIT | M_ZERO);
1351 		if (info->z_delay == NULL)
1352 			return (ENOMEM);
1353 		info->z_alloc = i;
1354 	}
1355 
1356 	/*
1357 	 * Zero out head of buffer to avoid pops and clicks.
1358 	 */
1359 	memset(info->z_delay, sndbuf_zerodata(f->desc->out),
1360 	    info->z_pos * align);
1361 
1362 #ifdef Z_DIAGNOSTIC
1363 	/*
1364 	 * XXX Debuging mess !@#$%^
1365 	 */
1366 #define dumpz(x)	fprintf(stderr, "\t%12s = %10u : %-11d\n",	\
1367 			    "z_"__STRING(x), (uint32_t)info->z_##x,	\
1368 			    (int32_t)info->z_##x)
1369 	fprintf(stderr, "\n%s():\n", __func__);
1370 	fprintf(stderr, "\tchannels=%d, bps=%d, format=0x%08x, quality=%d\n",
1371 	    info->channels, info->bps, format, info->quality);
1372 	fprintf(stderr, "\t%d (%d) -> %d (%d), ",
1373 	    info->src, info->rsrc, info->dst, info->rdst);
1374 	fprintf(stderr, "[%d/%d]\n", info->z_gx, info->z_gy);
1375 	fprintf(stderr, "\tminreq=%d, ", z_gy2gx(info, 1));
1376 	if (adaptive != 0)
1377 		z_scale = Z_ONE;
1378 	fprintf(stderr, "factor=0x%08x/0x%08x (%f)\n",
1379 	    z_scale, Z_ONE, (double)z_scale / Z_ONE);
1380 	fprintf(stderr, "\tbase_length=%d, ", Z_SINC_BASE_LEN(info));
1381 	fprintf(stderr, "adaptive=%s\n", (adaptive != 0) ? "YES" : "NO");
1382 	dumpz(size);
1383 	dumpz(alloc);
1384 	if (info->z_alloc < 1024)
1385 		fprintf(stderr, "\t%15s%10d Bytes\n",
1386 		    "", info->z_alloc);
1387 	else if (info->z_alloc < (1024 << 10))
1388 		fprintf(stderr, "\t%15s%10d KBytes\n",
1389 		    "", info->z_alloc >> 10);
1390 	else if (info->z_alloc < (1024 << 20))
1391 		fprintf(stderr, "\t%15s%10d MBytes\n",
1392 		    "", info->z_alloc >> 20);
1393 	else
1394 		fprintf(stderr, "\t%15s%10d GBytes\n",
1395 		    "", info->z_alloc >> 30);
1396 	fprintf(stderr, "\t%12s   %10d (min output samples)\n",
1397 	    "",
1398 	    (int32_t)z_gx2gy(info, info->z_full - (info->z_size << 1)));
1399 	fprintf(stderr, "\t%12s   %10d (min allocated output samples)\n",
1400 	    "",
1401 	    (int32_t)z_gx2gy(info, (info->z_alloc / align) -
1402 	    (info->z_size << 1)));
1403 	fprintf(stderr, "\t%12s = %10d\n",
1404 	    "z_gy2gx()", (int32_t)z_gy2gx(info, 1));
1405 	fprintf(stderr, "\t%12s = %10d -> z_gy2gx() -> %d\n",
1406 	    "Max", (int32_t)gy2gx_max, (int32_t)z_gy2gx(info, gy2gx_max));
1407 	fprintf(stderr, "\t%12s = %10d\n",
1408 	    "z_gx2gy()", (int32_t)z_gx2gy(info, 1));
1409 	fprintf(stderr, "\t%12s = %10d -> z_gx2gy() -> %d\n",
1410 	    "Max", (int32_t)gx2gy_max, (int32_t)z_gx2gy(info, gx2gy_max));
1411 	dumpz(maxfeed);
1412 	dumpz(full);
1413 	dumpz(start);
1414 	dumpz(pos);
1415 	dumpz(scale);
1416 	fprintf(stderr, "\t%12s   %10f\n", "",
1417 	    (double)info->z_scale / Z_ONE);
1418 	dumpz(dx);
1419 	fprintf(stderr, "\t%12s   %10f\n", "",
1420 	    (double)info->z_dx / info->z_dy);
1421 	dumpz(dy);
1422 	fprintf(stderr, "\t%12s   %10d (drift step)\n", "",
1423 	    info->z_dy >> Z_SHIFT);
1424 	fprintf(stderr, "\t%12s   %10d (scaling differences)\n", "",
1425 	    (z_scale << Z_DRIFT_SHIFT) - info->z_dy);
1426 	fprintf(stderr, "\t%12s = %u bytes\n",
1427 	    "intpcm32_t", sizeof(intpcm32_t));
1428 	fprintf(stderr, "\t%12s = 0x%08x, smallest=%.16lf\n",
1429 	    "Z_ONE", Z_ONE, (double)1.0 / (double)Z_ONE);
1430 #endif
1431 
1432 	return (0);
1433 }
1434 
1435 static int
1436 z_resampler_set(struct pcm_feeder *f, int what, int32_t value)
1437 {
1438 	struct z_info *info;
1439 	int32_t oquality;
1440 
1441 	info = f->data;
1442 
1443 	switch (what) {
1444 	case Z_RATE_SRC:
1445 		if (value < feeder_rate_min || value > feeder_rate_max)
1446 			return (E2BIG);
1447 		if (value == info->rsrc)
1448 			return (0);
1449 		info->rsrc = value;
1450 		break;
1451 	case Z_RATE_DST:
1452 		if (value < feeder_rate_min || value > feeder_rate_max)
1453 			return (E2BIG);
1454 		if (value == info->rdst)
1455 			return (0);
1456 		info->rdst = value;
1457 		break;
1458 	case Z_RATE_QUALITY:
1459 		if (value < Z_QUALITY_MIN || value > Z_QUALITY_MAX)
1460 			return (EINVAL);
1461 		if (value == info->quality)
1462 			return (0);
1463 		/*
1464 		 * If we failed to set the requested quality, restore
1465 		 * the old one. We cannot afford leaving it broken since
1466 		 * passive feeder chains like vchans never reinitialize
1467 		 * itself.
1468 		 */
1469 		oquality = info->quality;
1470 		info->quality = value;
1471 		if (z_resampler_setup(f) == 0)
1472 			return (0);
1473 		info->quality = oquality;
1474 		break;
1475 	case Z_RATE_CHANNELS:
1476 		if (value < SND_CHN_MIN || value > SND_CHN_MAX)
1477 			return (EINVAL);
1478 		if (value == info->channels)
1479 			return (0);
1480 		info->channels = value;
1481 		break;
1482 	default:
1483 		return (EINVAL);
1484 		break;
1485 	}
1486 
1487 	return (z_resampler_setup(f));
1488 }
1489 
1490 static int
1491 z_resampler_get(struct pcm_feeder *f, int what)
1492 {
1493 	struct z_info *info;
1494 
1495 	info = f->data;
1496 
1497 	switch (what) {
1498 	case Z_RATE_SRC:
1499 		return (info->rsrc);
1500 		break;
1501 	case Z_RATE_DST:
1502 		return (info->rdst);
1503 		break;
1504 	case Z_RATE_QUALITY:
1505 		return (info->quality);
1506 		break;
1507 	case Z_RATE_CHANNELS:
1508 		return (info->channels);
1509 		break;
1510 	default:
1511 		break;
1512 	}
1513 
1514 	return (-1);
1515 }
1516 
1517 static int
1518 z_resampler_init(struct pcm_feeder *f)
1519 {
1520 	struct z_info *info;
1521 	int ret;
1522 
1523 	if (f->desc->in != f->desc->out)
1524 		return (EINVAL);
1525 
1526 	info = malloc(sizeof(*info), M_DEVBUF, M_NOWAIT | M_ZERO);
1527 	if (info == NULL)
1528 		return (ENOMEM);
1529 
1530 	info->rsrc = Z_RATE_DEFAULT;
1531 	info->rdst = Z_RATE_DEFAULT;
1532 	info->quality = feeder_rate_quality;
1533 	info->channels = AFMT_CHANNEL(f->desc->in);
1534 
1535 	f->data = info;
1536 
1537 	ret = z_resampler_setup(f);
1538 	if (ret != 0) {
1539 		if (info->z_pcoeff != NULL)
1540 			free(info->z_pcoeff, M_DEVBUF);
1541 		if (info->z_delay != NULL)
1542 			free(info->z_delay, M_DEVBUF);
1543 		free(info, M_DEVBUF);
1544 		f->data = NULL;
1545 	}
1546 
1547 	return (ret);
1548 }
1549 
1550 static int
1551 z_resampler_free(struct pcm_feeder *f)
1552 {
1553 	struct z_info *info;
1554 
1555 	info = f->data;
1556 	if (info != NULL) {
1557 		if (info->z_pcoeff != NULL)
1558 			free(info->z_pcoeff, M_DEVBUF);
1559 		if (info->z_delay != NULL)
1560 			free(info->z_delay, M_DEVBUF);
1561 		free(info, M_DEVBUF);
1562 	}
1563 
1564 	f->data = NULL;
1565 
1566 	return (0);
1567 }
1568 
1569 static uint32_t
1570 z_resampler_feed_internal(struct pcm_feeder *f, struct pcm_channel *c,
1571     uint8_t *b, uint32_t count, void *source)
1572 {
1573 	struct z_info *info;
1574 	int32_t alphadrift, startdrift, reqout, ocount, reqin, align;
1575 	int32_t fetch, fetched, start, cp;
1576 	uint8_t *dst;
1577 
1578 	info = f->data;
1579 	if (info->z_resample == NULL)
1580 		return (z_feed(f->source, c, b, count, source));
1581 
1582 	/*
1583 	 * Calculate sample size alignment and amount of sample output.
1584 	 * We will do everything in sample domain, but at the end we
1585 	 * will jump back to byte domain.
1586 	 */
1587 	align = info->channels * info->bps;
1588 	ocount = SND_FXDIV(count, align);
1589 	if (ocount == 0)
1590 		return (0);
1591 
1592 	/*
1593 	 * Calculate amount of input samples that is needed to generate
1594 	 * exact amount of output.
1595 	 */
1596 	reqin = z_gy2gx(info, ocount) - z_fetched(info);
1597 
1598 #ifdef Z_USE_ALPHADRIFT
1599 	startdrift = info->z_startdrift;
1600 	alphadrift = info->z_alphadrift;
1601 #else
1602 	startdrift = _Z_GY2GX(info, 0, 1);
1603 	alphadrift = z_drift(info, startdrift, 1);
1604 #endif
1605 
1606 	dst = b;
1607 
1608 	do {
1609 		if (reqin != 0) {
1610 			fetch = z_min(z_free(info), reqin);
1611 			if (fetch == 0) {
1612 				/*
1613 				 * No more free spaces, so wind enough
1614 				 * samples back to the head of delay line
1615 				 * in byte domain.
1616 				 */
1617 				fetched = z_fetched(info);
1618 				start = z_prev(info, info->z_start,
1619 				    (info->z_size << 1) - 1);
1620 				cp = (info->z_size << 1) + fetched;
1621 				z_copy(info->z_delay + (start * align),
1622 				    info->z_delay, cp * align);
1623 				info->z_start =
1624 				    z_prev(info, info->z_size << 1, 1);
1625 				info->z_pos =
1626 				    z_next(info, info->z_start, fetched + 1);
1627 				fetch = z_min(z_free(info), reqin);
1628 #ifdef Z_DIAGNOSTIC
1629 				if (1) {
1630 					static uint32_t kk = 0;
1631 					fprintf(stderr,
1632 					    "Buffer Move: "
1633 					    "start=%d fetched=%d cp=%d "
1634 					    "cycle=%u [%u]\r",
1635 					    start, fetched, cp, info->z_cycle,
1636 					    ++kk);
1637 				}
1638 				info->z_cycle = 0;
1639 #endif
1640 			}
1641 			if (fetch != 0) {
1642 				/*
1643 				 * Fetch in byte domain and jump back
1644 				 * to sample domain.
1645 				 */
1646 				fetched = SND_FXDIV(z_feed(f->source, c,
1647 				    info->z_delay + (info->z_pos * align),
1648 				    fetch * align, source), align);
1649 				/*
1650 				 * Prepare to convert fetched buffer,
1651 				 * or mark us done if we cannot fulfill
1652 				 * the request.
1653 				 */
1654 				reqin -= fetched;
1655 				info->z_pos += fetched;
1656 				if (fetched != fetch)
1657 					reqin = 0;
1658 			}
1659 		}
1660 
1661 		reqout = z_min(z_gx2gy(info, z_fetched(info)), ocount);
1662 		if (reqout != 0) {
1663 			ocount -= reqout;
1664 
1665 			/*
1666 			 * Drift.. drift.. drift..
1667 			 *
1668 			 * Notice that there are 2 methods of doing the drift
1669 			 * operations: The former is much cleaner (in a sense
1670 			 * of mathematical readings of my eyes), but slower
1671 			 * due to integer division in z_gy2gx(). Nevertheless,
1672 			 * both should give the same exact accurate drifting
1673 			 * results, so the later is favourable.
1674 			 */
1675 			do {
1676 				info->z_resample(info, dst);
1677 #if 0
1678 				startdrift = z_gy2gx(info, 1);
1679 				alphadrift = z_drift(info, startdrift, 1);
1680 				info->z_start += startdrift;
1681 				info->z_alpha += alphadrift;
1682 #else
1683 				info->z_alpha += alphadrift;
1684 				if (info->z_alpha < info->z_gy)
1685 					info->z_start += startdrift;
1686 				else {
1687 					info->z_start += startdrift - 1;
1688 					info->z_alpha -= info->z_gy;
1689 				}
1690 #endif
1691 				dst += align;
1692 #ifdef Z_DIAGNOSTIC
1693 				info->z_cycle++;
1694 #endif
1695 			} while (--reqout != 0);
1696 		}
1697 	} while (reqin != 0 && ocount != 0);
1698 
1699 	/*
1700 	 * Back to byte domain..
1701 	 */
1702 	return (dst - b);
1703 }
1704 
1705 static int
1706 z_resampler_feed(struct pcm_feeder *f, struct pcm_channel *c, uint8_t *b,
1707     uint32_t count, void *source)
1708 {
1709 	uint32_t feed, maxfeed, left;
1710 
1711 	/*
1712 	 * Split count to smaller chunks to avoid possible 32bit overflow.
1713 	 */
1714 	maxfeed = ((struct z_info *)(f->data))->z_maxfeed;
1715 	left = count;
1716 
1717 	do {
1718 		feed = z_resampler_feed_internal(f, c, b,
1719 		    z_min(maxfeed, left), source);
1720 		b += feed;
1721 		left -= feed;
1722 	} while (left != 0 && feed != 0);
1723 
1724 	return (count - left);
1725 }
1726 
1727 static struct pcm_feederdesc feeder_rate_desc[] = {
1728 	{ FEEDER_RATE, 0, 0, 0, 0 },
1729 	{ 0, 0, 0, 0, 0 },
1730 };
1731 
1732 static kobj_method_t feeder_rate_methods[] = {
1733 	KOBJMETHOD(feeder_init,		z_resampler_init),
1734 	KOBJMETHOD(feeder_free,		z_resampler_free),
1735 	KOBJMETHOD(feeder_set,		z_resampler_set),
1736 	KOBJMETHOD(feeder_get,		z_resampler_get),
1737 	KOBJMETHOD(feeder_feed,		z_resampler_feed),
1738 	KOBJMETHOD_END
1739 };
1740 
1741 FEEDER_DECLARE(feeder_rate, NULL);
1742