xref: /dragonfly/sys/dev/sound/pcm/feeder_rate.c (revision 6e5c5008)
1 /*-
2  * Copyright (c) 2005-2009 Ariff Abdullah <ariff@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 /*
28  * feeder_rate: (Codename: Z Resampler), which means any effort to create
29  *              future replacement for this resampler are simply absurd unless
30  *              the world decide to add new alphabet after Z.
31  *
32  * FreeBSD bandlimited sinc interpolator, technically based on
33  * "Digital Audio Resampling" by Julius O. Smith III
34  *  - http://ccrma.stanford.edu/~jos/resample/
35  *
36  * The Good:
37  * + all out fixed point integer operations, no soft-float or anything like
38  *   that.
39  * + classic polyphase converters with high quality coefficient's polynomial
40  *   interpolators.
41  * + fast, faster, or the fastest of its kind.
42  * + compile time configurable.
43  * + etc etc..
44  *
45  * The Bad:
46  * - The z, z_, and Z_ . Due to mental block (or maybe just 0x7a69), I
47  *   couldn't think of anything simpler than that (feeder_rate_xxx is just
48  *   too long). Expect possible clashes with other zitizens (any?).
49  */
50 
51 #ifdef _KERNEL
52 #ifdef HAVE_KERNEL_OPTION_HEADERS
53 #include "opt_snd.h"
54 #endif
55 #include <dev/sound/pcm/sound.h>
56 #include <dev/sound/pcm/pcm.h>
57 #include "feeder_if.h"
58 
59 #define SND_USE_FXDIV
60 #include "snd_fxdiv_gen.h"
61 
62 SND_DECLARE_FILE("$FreeBSD: head/sys/dev/sound/pcm/feeder_rate.c 267992 2014-06-28 03:56:17Z hselasky $");
63 #endif
64 
65 #include "feeder_rate_gen.h"
66 
67 #if !defined(_KERNEL) && defined(SND_DIAGNOSTIC)
68 #undef Z_DIAGNOSTIC
69 #define Z_DIAGNOSTIC		1
70 #elif defined(_KERNEL)
71 #undef Z_DIAGNOSTIC
72 #endif
73 
74 #ifndef Z_QUALITY_DEFAULT
75 #define Z_QUALITY_DEFAULT	Z_QUALITY_LINEAR
76 #endif
77 
78 #define Z_RESERVOIR		2048
79 #define Z_RESERVOIR_MAX		131072
80 
81 #define Z_SINC_MAX		0x3fffff
82 #define Z_SINC_DOWNMAX		48		/* 384000 / 8000 */
83 
84 #ifdef _KERNEL
85 #define Z_POLYPHASE_MAX		183040		/* 286 taps, 640 phases */
86 #else
87 #define Z_POLYPHASE_MAX		1464320		/* 286 taps, 5120 phases */
88 #endif
89 
90 #define Z_RATE_DEFAULT		48000
91 
92 #define Z_RATE_MIN		FEEDRATE_RATEMIN
93 #define Z_RATE_MAX		FEEDRATE_RATEMAX
94 #define Z_ROUNDHZ		FEEDRATE_ROUNDHZ
95 #define Z_ROUNDHZ_MIN		FEEDRATE_ROUNDHZ_MIN
96 #define Z_ROUNDHZ_MAX		FEEDRATE_ROUNDHZ_MAX
97 
98 #define Z_RATE_SRC		FEEDRATE_SRC
99 #define Z_RATE_DST		FEEDRATE_DST
100 #define Z_RATE_QUALITY		FEEDRATE_QUALITY
101 #define Z_RATE_CHANNELS		FEEDRATE_CHANNELS
102 
103 #define Z_PARANOID		1
104 
105 #define Z_MULTIFORMAT		1
106 
107 #ifdef _KERNEL
108 #undef Z_USE_ALPHADRIFT
109 #define Z_USE_ALPHADRIFT	1
110 #endif
111 
112 #define Z_FACTOR_MIN		1
113 #define Z_FACTOR_MAX		Z_MASK
114 #define Z_FACTOR_SAFE(v)	(!((v) < Z_FACTOR_MIN || (v) > Z_FACTOR_MAX))
115 
116 struct z_info;
117 
118 typedef void (*z_resampler_t)(struct z_info *, uint8_t *);
119 
120 struct z_info {
121 	int32_t rsrc, rdst;	/* original source / destination rates */
122 	int32_t src, dst;	/* rounded source / destination rates */
123 	int32_t channels;	/* total channels */
124 	int32_t bps;		/* bytes-per-sample */
125 	int32_t quality;	/* resampling quality */
126 
127 	int32_t z_gx, z_gy;	/* interpolation / decimation ratio */
128 	int32_t z_alpha;	/* output sample time phase / drift */
129 	uint8_t *z_delay;	/* FIR delay line / linear buffer */
130 	int32_t *z_coeff;	/* FIR coefficients */
131 	int32_t *z_dcoeff;	/* FIR coefficients differences */
132 	int32_t *z_pcoeff;	/* FIR polyphase coefficients */
133 	int32_t z_scale;	/* output scaling */
134 	int32_t z_dx;		/* input sample drift increment */
135 	int32_t z_dy;		/* output sample drift increment */
136 #ifdef Z_USE_ALPHADRIFT
137 	int32_t z_alphadrift;	/* alpha drift rate */
138 	int32_t z_startdrift;	/* buffer start position drift rate */
139 #endif
140 	int32_t z_mask;		/* delay line full length mask */
141 	int32_t z_size;		/* half width of FIR taps */
142 	int32_t z_full;		/* full size of delay line */
143 	int32_t z_alloc;	/* largest allocated full size of delay line */
144 	int32_t z_start;	/* buffer processing start position */
145 	int32_t z_pos;		/* current position for the next feed */
146 #ifdef Z_DIAGNOSTIC
147 	uint32_t z_cycle;	/* output cycle, purely for statistical */
148 #endif
149 	int32_t z_maxfeed;	/* maximum feed to avoid 32bit overflow */
150 
151 	z_resampler_t z_resample;
152 };
153 
154 int feeder_rate_min = Z_RATE_MIN;
155 int feeder_rate_max = Z_RATE_MAX;
156 int feeder_rate_round = Z_ROUNDHZ;
157 int feeder_rate_quality = Z_QUALITY_DEFAULT;
158 
159 static int feeder_rate_polyphase_max = Z_POLYPHASE_MAX;
160 
161 #ifdef _KERNEL
162 static char feeder_rate_presets[] = FEEDER_RATE_PRESETS;
163 SYSCTL_STRING(_hw_snd, OID_AUTO, feeder_rate_presets, CTLFLAG_RD,
164     &feeder_rate_presets, 0, "compile-time rate presets");
165 
166 TUNABLE_INT("hw.snd.feeder_rate_min", &feeder_rate_min);
167 TUNABLE_INT("hw.snd.feeder_rate_max", &feeder_rate_max);
168 TUNABLE_INT("hw.snd.feeder_rate_round", &feeder_rate_round);
169 TUNABLE_INT("hw.snd.feeder_rate_quality", &feeder_rate_quality);
170 
171 TUNABLE_INT("hw.snd.feeder_rate_polyphase_max", &feeder_rate_polyphase_max);
172 SYSCTL_INT(_hw_snd, OID_AUTO, feeder_rate_polyphase_max, CTLFLAG_RW,
173     &feeder_rate_polyphase_max, 0, "maximum allowable polyphase entries");
174 
175 static int
176 sysctl_hw_snd_feeder_rate_min(SYSCTL_HANDLER_ARGS)
177 {
178 	int err, val;
179 
180 	val = feeder_rate_min;
181 	err = sysctl_handle_int(oidp, &val, 0, req);
182 
183 	if (err != 0 || req->newptr == NULL || val == feeder_rate_min)
184 		return (err);
185 
186 	if (!(Z_FACTOR_SAFE(val) && val < feeder_rate_max))
187 		return (EINVAL);
188 
189 	feeder_rate_min = val;
190 
191 	return (0);
192 }
193 SYSCTL_PROC(_hw_snd, OID_AUTO, feeder_rate_min, CTLTYPE_INT | CTLFLAG_RW,
194     0, sizeof(int), sysctl_hw_snd_feeder_rate_min, "I",
195     "minimum allowable rate");
196 
197 static int
198 sysctl_hw_snd_feeder_rate_max(SYSCTL_HANDLER_ARGS)
199 {
200 	int err, val;
201 
202 	val = feeder_rate_max;
203 	err = sysctl_handle_int(oidp, &val, 0, req);
204 
205 	if (err != 0 || req->newptr == NULL || val == feeder_rate_max)
206 		return (err);
207 
208 	if (!(Z_FACTOR_SAFE(val) && val > feeder_rate_min))
209 		return (EINVAL);
210 
211 	feeder_rate_max = val;
212 
213 	return (0);
214 }
215 SYSCTL_PROC(_hw_snd, OID_AUTO, feeder_rate_max, CTLTYPE_INT | CTLFLAG_RW,
216     0, sizeof(int), sysctl_hw_snd_feeder_rate_max, "I",
217     "maximum allowable rate");
218 
219 static int
220 sysctl_hw_snd_feeder_rate_round(SYSCTL_HANDLER_ARGS)
221 {
222 	int err, val;
223 
224 	val = feeder_rate_round;
225 	err = sysctl_handle_int(oidp, &val, 0, req);
226 
227 	if (err != 0 || req->newptr == NULL || val == feeder_rate_round)
228 		return (err);
229 
230 	if (val < Z_ROUNDHZ_MIN || val > Z_ROUNDHZ_MAX)
231 		return (EINVAL);
232 
233 	feeder_rate_round = val - (val % Z_ROUNDHZ);
234 
235 	return (0);
236 }
237 SYSCTL_PROC(_hw_snd, OID_AUTO, feeder_rate_round, CTLTYPE_INT | CTLFLAG_RW,
238     0, sizeof(int), sysctl_hw_snd_feeder_rate_round, "I",
239     "sample rate converter rounding threshold");
240 
241 static int
242 sysctl_hw_snd_feeder_rate_quality(SYSCTL_HANDLER_ARGS)
243 {
244 	struct snddev_info *d;
245 	struct pcm_channel *c;
246 	struct pcm_feeder *f;
247 	int i, err, val;
248 
249 	val = feeder_rate_quality;
250 	err = sysctl_handle_int(oidp, &val, 0, req);
251 
252 	if (err != 0 || req->newptr == NULL || val == feeder_rate_quality)
253 		return (err);
254 
255 	if (val < Z_QUALITY_MIN || val > Z_QUALITY_MAX)
256 		return (EINVAL);
257 
258 	feeder_rate_quality = val;
259 
260 	/*
261 	 * Traverse all available channels on each device and try to
262 	 * set resampler quality if and only if it is exist as
263 	 * part of feeder chains and the channel is idle.
264 	 */
265 	for (i = 0; pcm_devclass != NULL &&
266 	    i < devclass_get_maxunit(pcm_devclass); i++) {
267 		d = devclass_get_softc(pcm_devclass, i);
268 		if (!PCM_REGISTERED(d))
269 			continue;
270 		PCM_LOCK(d);
271 		PCM_WAIT(d);
272 		PCM_ACQUIRE(d);
273 		CHN_FOREACH(c, d, channels.pcm) {
274 			CHN_LOCK(c);
275 			f = chn_findfeeder(c, FEEDER_RATE);
276 			if (f == NULL || f->data == NULL || CHN_STARTED(c)) {
277 				CHN_UNLOCK(c);
278 				continue;
279 			}
280 			(void)FEEDER_SET(f, FEEDRATE_QUALITY, val);
281 			CHN_UNLOCK(c);
282 		}
283 		PCM_RELEASE(d);
284 		PCM_UNLOCK(d);
285 	}
286 
287 	return (0);
288 }
289 SYSCTL_PROC(_hw_snd, OID_AUTO, feeder_rate_quality, CTLTYPE_INT | CTLFLAG_RW,
290     0, sizeof(int), sysctl_hw_snd_feeder_rate_quality, "I",
291     "sample rate converter quality ("__XSTRING(Z_QUALITY_MIN)"=low .. "
292     __XSTRING(Z_QUALITY_MAX)"=high)");
293 #endif	/* _KERNEL */
294 
295 
296 /*
297  * Resampler type.
298  */
299 #define Z_IS_ZOH(i)		((i)->quality == Z_QUALITY_ZOH)
300 #define Z_IS_LINEAR(i)		((i)->quality == Z_QUALITY_LINEAR)
301 #define Z_IS_SINC(i)		((i)->quality > Z_QUALITY_LINEAR)
302 
303 /*
304  * Macroses for accurate sample time drift calculations.
305  *
306  * gy2gx : given the amount of output, return the _exact_ required amount of
307  *         input.
308  * gx2gy : given the amount of input, return the _maximum_ amount of output
309  *         that will be generated.
310  * drift : given the amount of input and output, return the elapsed
311  *         sample-time.
312  */
313 #define _Z_GCAST(x)		((uint64_t)(x))
314 
315 #ifdef __x86_64__
316 #define Z_DIV(x, y)		((x) / (y))
317 #endif
318 
319 #define _Z_GY2GX(i, a, v)						\
320 	Z_DIV(((_Z_GCAST((i)->z_gx) * (v)) + ((i)->z_gy - (a) - 1)),	\
321 	(i)->z_gy)
322 
323 #define _Z_GX2GY(i, a, v)						\
324 	Z_DIV(((_Z_GCAST((i)->z_gy) * (v)) + (a)), (i)->z_gx)
325 
326 #define _Z_DRIFT(i, x, y)						\
327 	((_Z_GCAST((i)->z_gy) * (x)) - (_Z_GCAST((i)->z_gx) * (y)))
328 
329 #define z_gy2gx(i, v)		_Z_GY2GX(i, (i)->z_alpha, v)
330 #define z_gx2gy(i, v)		_Z_GX2GY(i, (i)->z_alpha, v)
331 #define z_drift(i, x, y)	_Z_DRIFT(i, x, y)
332 
333 /*
334  * Macroses for SINC coefficients table manipulations.. whatever.
335  */
336 #define Z_SINC_COEFF_IDX(i)	((i)->quality - Z_QUALITY_LINEAR - 1)
337 
338 #define Z_SINC_LEN(i)							\
339 	((int32_t)(((uint64_t)z_coeff_tab[Z_SINC_COEFF_IDX(i)].len <<	\
340 	    Z_SHIFT) / (i)->z_dy))
341 
342 #define Z_SINC_BASE_LEN(i)						\
343 	((z_coeff_tab[Z_SINC_COEFF_IDX(i)].len - 1) >> (Z_DRIFT_SHIFT - 1))
344 
345 /*
346  * Macroses for linear delay buffer operations. Alignment is not
347  * really necessary since we're not using true circular buffer, but it
348  * will help us guard against possible trespasser. To be honest,
349  * the linear block operations does not need guarding at all due to
350  * accurate drifting!
351  */
352 #define z_align(i, v)		((v) & (i)->z_mask)
353 #define z_next(i, o, v)		z_align(i, (o) + (v))
354 #define z_prev(i, o, v)		z_align(i, (o) - (v))
355 #define z_fetched(i)		(z_align(i, (i)->z_pos - (i)->z_start) - 1)
356 #define z_free(i)		((i)->z_full - (i)->z_pos)
357 
358 /*
359  * Macroses for Bla Bla .. :)
360  */
361 #define z_copy(src, dst, sz)	(void)memcpy(dst, src, sz)
362 #define z_feed(...)		FEEDER_FEED(__VA_ARGS__)
363 
364 static __inline uint32_t
365 z_min(uint32_t x, uint32_t y)
366 {
367 
368 	return ((x < y) ? x : y);
369 }
370 
371 static int32_t
372 z_gcd(int32_t x, int32_t y)
373 {
374 	int32_t w;
375 
376 	while (y != 0) {
377 		w = x % y;
378 		x = y;
379 		y = w;
380 	}
381 
382 	return (x);
383 }
384 
385 static int32_t
386 z_roundpow2(int32_t v)
387 {
388 	int32_t i;
389 
390 	i = 1;
391 
392 	/*
393 	 * Let it overflow at will..
394 	 */
395 	while (i > 0 && i < v)
396 		i <<= 1;
397 
398 	return (i);
399 }
400 
401 /*
402  * Zero Order Hold, the worst of the worst, an insult against quality,
403  * but super fast.
404  */
405 static void
406 z_feed_zoh(struct z_info *info, uint8_t *dst)
407 {
408 #if 0
409 	z_copy(info->z_delay +
410 	    (info->z_start * info->channels * info->bps), dst,
411 	    info->channels * info->bps);
412 #else
413 	uint32_t cnt;
414 	uint8_t *src;
415 
416 	cnt = info->channels * info->bps;
417 	src = info->z_delay + (info->z_start * cnt);
418 
419 	/*
420 	 * This is a bit faster than doing bcopy() since we're dealing
421 	 * with possible unaligned samples.
422 	 */
423 	do {
424 		*dst++ = *src++;
425 	} while (--cnt != 0);
426 #endif
427 }
428 
429 /*
430  * Linear Interpolation. This at least sounds better (perceptually) and fast,
431  * but without any proper filtering which means aliasing still exist and
432  * could become worst with a right sample. Interpolation centered within
433  * Z_LINEAR_ONE between the present and previous sample and everything is
434  * done with simple 32bit scaling arithmetic.
435  */
436 #define Z_DECLARE_LINEAR(SIGN, BIT, ENDIAN)					\
437 static void									\
438 z_feed_linear_##SIGN##BIT##ENDIAN(struct z_info *info, uint8_t *dst)		\
439 {										\
440 	int32_t z;								\
441 	intpcm_t x, y;								\
442 	uint32_t ch;								\
443 	uint8_t *sx, *sy;							\
444 										\
445 	z = ((uint32_t)info->z_alpha * info->z_dx) >> Z_LINEAR_UNSHIFT;		\
446 										\
447 	sx = info->z_delay + (info->z_start * info->channels *			\
448 	    PCM_##BIT##_BPS);							\
449 	sy = sx - (info->channels * PCM_##BIT##_BPS);				\
450 										\
451 	ch = info->channels;							\
452 										\
453 	do {									\
454 		x = _PCM_READ_##SIGN##BIT##_##ENDIAN(sx);			\
455 		y = _PCM_READ_##SIGN##BIT##_##ENDIAN(sy);			\
456 		x = Z_LINEAR_INTERPOLATE_##BIT(z, x, y);			\
457 		_PCM_WRITE_##SIGN##BIT##_##ENDIAN(dst, x);			\
458 		sx += PCM_##BIT##_BPS;						\
459 		sy += PCM_##BIT##_BPS;						\
460 		dst += PCM_##BIT##_BPS;						\
461 	} while (--ch != 0);							\
462 }
463 
464 /*
465  * Userland clipping diagnostic check, not enabled in kernel compilation.
466  * While doing sinc interpolation, unrealistic samples like full scale sine
467  * wav will clip, but for other things this will not make any noise at all.
468  * Everybody should learn how to normalized perceived loudness of their own
469  * music/sounds/samples (hint: ReplayGain).
470  */
471 #ifdef Z_DIAGNOSTIC
472 #define Z_CLIP_CHECK(v, BIT)	do {					\
473 	if ((v) > PCM_S##BIT##_MAX) {					\
474 		fprintf(stderr, "Overflow: v=%jd, max=%jd\n",		\
475 		    (intmax_t)(v), (intmax_t)PCM_S##BIT##_MAX);		\
476 	} else if ((v) < PCM_S##BIT##_MIN) {				\
477 		fprintf(stderr, "Underflow: v=%jd, min=%jd\n",		\
478 		    (intmax_t)(v), (intmax_t)PCM_S##BIT##_MIN);		\
479 	}								\
480 } while (0)
481 #else
482 #define Z_CLIP_CHECK(...)
483 #endif
484 
485 #define Z_CLAMP(v, BIT)							\
486 	(((v) > PCM_S##BIT##_MAX) ? PCM_S##BIT##_MAX :			\
487 	(((v) < PCM_S##BIT##_MIN) ? PCM_S##BIT##_MIN : (v)))
488 
489 /*
490  * Sine Cardinal (SINC) Interpolation. Scaling is done in 64 bit, so
491  * there's no point to hold the plate any longer. All samples will be
492  * shifted to a full 32 bit, scaled and restored during write for
493  * maximum dynamic range (only for downsampling).
494  */
495 #define _Z_SINC_ACCUMULATE(SIGN, BIT, ENDIAN, adv)			\
496 	c += z >> Z_SHIFT;						\
497 	z &= Z_MASK;							\
498 	coeff = Z_COEFF_INTERPOLATE(z, z_coeff[c], z_dcoeff[c]);	\
499 	x = _PCM_READ_##SIGN##BIT##_##ENDIAN(p);			\
500 	v += Z_NORM_##BIT((intpcm64_t)x * coeff);			\
501 	z += info->z_dy;						\
502 	p adv##= info->channels * PCM_##BIT##_BPS
503 
504 /*
505  * XXX GCC4 optimization is such a !@#$%, need manual unrolling.
506  */
507 #if defined(__GNUC__) && __GNUC__ >= 4
508 #define Z_SINC_ACCUMULATE(...)	do {					\
509 	_Z_SINC_ACCUMULATE(__VA_ARGS__);				\
510 	_Z_SINC_ACCUMULATE(__VA_ARGS__);				\
511 } while (0)
512 #define Z_SINC_ACCUMULATE_DECR		2
513 #else
514 #define Z_SINC_ACCUMULATE(...)	do {					\
515 	_Z_SINC_ACCUMULATE(__VA_ARGS__);				\
516 } while (0)
517 #define Z_SINC_ACCUMULATE_DECR		1
518 #endif
519 
520 #define Z_DECLARE_SINC(SIGN, BIT, ENDIAN)					\
521 static void									\
522 z_feed_sinc_##SIGN##BIT##ENDIAN(struct z_info *info, uint8_t *dst)		\
523 {										\
524 	intpcm64_t v;								\
525 	intpcm_t x;								\
526 	uint8_t *p;								\
527 	int32_t coeff, z, *z_coeff, *z_dcoeff;					\
528 	uint32_t c, center, ch, i;						\
529 										\
530 	z_coeff = info->z_coeff;						\
531 	z_dcoeff = info->z_dcoeff;						\
532 	center = z_prev(info, info->z_start, info->z_size);			\
533 	ch = info->channels * PCM_##BIT##_BPS;					\
534 	dst += ch;								\
535 										\
536 	do {									\
537 		dst -= PCM_##BIT##_BPS;						\
538 		ch -= PCM_##BIT##_BPS;						\
539 		v = 0;								\
540 		z = info->z_alpha * info->z_dx;					\
541 		c = 0;								\
542 		p = info->z_delay + (z_next(info, center, 1) *			\
543 		    info->channels * PCM_##BIT##_BPS) + ch;			\
544 		for (i = info->z_size; i != 0; i -= Z_SINC_ACCUMULATE_DECR) 	\
545 			Z_SINC_ACCUMULATE(SIGN, BIT, ENDIAN, +);		\
546 		z = info->z_dy - (info->z_alpha * info->z_dx);			\
547 		c = 0;								\
548 		p = info->z_delay + (center * info->channels *			\
549 		    PCM_##BIT##_BPS) + ch;					\
550 		for (i = info->z_size; i != 0; i -= Z_SINC_ACCUMULATE_DECR) 	\
551 			Z_SINC_ACCUMULATE(SIGN, BIT, ENDIAN, -);		\
552 		if (info->z_scale != Z_ONE)					\
553 			v = Z_SCALE_##BIT(v, info->z_scale);			\
554 		else								\
555 			v >>= Z_COEFF_SHIFT - Z_GUARD_BIT_##BIT;		\
556 		Z_CLIP_CHECK(v, BIT);						\
557 		_PCM_WRITE_##SIGN##BIT##_##ENDIAN(dst, Z_CLAMP(v, BIT));	\
558 	} while (ch != 0);							\
559 }
560 
561 #define Z_DECLARE_SINC_POLYPHASE(SIGN, BIT, ENDIAN)				\
562 static void									\
563 z_feed_sinc_polyphase_##SIGN##BIT##ENDIAN(struct z_info *info, uint8_t *dst)	\
564 {										\
565 	intpcm64_t v;								\
566 	intpcm_t x;								\
567 	uint8_t *p;								\
568 	int32_t ch, i, start, *z_pcoeff;					\
569 										\
570 	ch = info->channels * PCM_##BIT##_BPS;					\
571 	dst += ch;								\
572 	start = z_prev(info, info->z_start, (info->z_size << 1) - 1) * ch;	\
573 										\
574 	do {									\
575 		dst -= PCM_##BIT##_BPS;						\
576 		ch -= PCM_##BIT##_BPS;						\
577 		v = 0;								\
578 		p = info->z_delay + start + ch;					\
579 		z_pcoeff = info->z_pcoeff +					\
580 		    ((info->z_alpha * info->z_size) << 1);			\
581 		for (i = info->z_size; i != 0; i--) {				\
582 			x = _PCM_READ_##SIGN##BIT##_##ENDIAN(p);		\
583 			v += Z_NORM_##BIT((intpcm64_t)x * *z_pcoeff);		\
584 			z_pcoeff++;						\
585 			p += info->channels * PCM_##BIT##_BPS;			\
586 			x = _PCM_READ_##SIGN##BIT##_##ENDIAN(p);		\
587 			v += Z_NORM_##BIT((intpcm64_t)x * *z_pcoeff);		\
588 			z_pcoeff++;						\
589 			p += info->channels * PCM_##BIT##_BPS;			\
590 		}								\
591 		if (info->z_scale != Z_ONE)					\
592 			v = Z_SCALE_##BIT(v, info->z_scale);			\
593 		else								\
594 			v >>= Z_COEFF_SHIFT - Z_GUARD_BIT_##BIT;		\
595 		Z_CLIP_CHECK(v, BIT);						\
596 		_PCM_WRITE_##SIGN##BIT##_##ENDIAN(dst, Z_CLAMP(v, BIT));	\
597 	} while (ch != 0);							\
598 }
599 
600 #define Z_DECLARE(SIGN, BIT, ENDIAN)					\
601 	Z_DECLARE_LINEAR(SIGN, BIT, ENDIAN)				\
602 	Z_DECLARE_SINC(SIGN, BIT, ENDIAN)				\
603 	Z_DECLARE_SINC_POLYPHASE(SIGN, BIT, ENDIAN)
604 
605 #if BYTE_ORDER == LITTLE_ENDIAN || defined(SND_FEEDER_MULTIFORMAT)
606 Z_DECLARE(S, 16, LE)
607 Z_DECLARE(S, 32, LE)
608 #endif
609 #if BYTE_ORDER == BIG_ENDIAN || defined(SND_FEEDER_MULTIFORMAT)
610 Z_DECLARE(S, 16, BE)
611 Z_DECLARE(S, 32, BE)
612 #endif
613 #ifdef SND_FEEDER_MULTIFORMAT
614 Z_DECLARE(S,  8, NE)
615 Z_DECLARE(S, 24, LE)
616 Z_DECLARE(S, 24, BE)
617 Z_DECLARE(U,  8, NE)
618 Z_DECLARE(U, 16, LE)
619 Z_DECLARE(U, 24, LE)
620 Z_DECLARE(U, 32, LE)
621 Z_DECLARE(U, 16, BE)
622 Z_DECLARE(U, 24, BE)
623 Z_DECLARE(U, 32, BE)
624 #endif
625 
626 enum {
627 	Z_RESAMPLER_ZOH,
628 	Z_RESAMPLER_LINEAR,
629 	Z_RESAMPLER_SINC,
630 	Z_RESAMPLER_SINC_POLYPHASE,
631 	Z_RESAMPLER_LAST
632 };
633 
634 #define Z_RESAMPLER_IDX(i)						\
635 	(Z_IS_SINC(i) ? Z_RESAMPLER_SINC : (i)->quality)
636 
637 #define Z_RESAMPLER_ENTRY(SIGN, BIT, ENDIAN)					\
638 	{									\
639 	    AFMT_##SIGN##BIT##_##ENDIAN,					\
640 	    {									\
641 		[Z_RESAMPLER_ZOH]    = z_feed_zoh,				\
642 		[Z_RESAMPLER_LINEAR] = z_feed_linear_##SIGN##BIT##ENDIAN,	\
643 		[Z_RESAMPLER_SINC]   = z_feed_sinc_##SIGN##BIT##ENDIAN,		\
644 		[Z_RESAMPLER_SINC_POLYPHASE]   =				\
645 		    z_feed_sinc_polyphase_##SIGN##BIT##ENDIAN			\
646 	    }									\
647 	}
648 
649 static const struct {
650 	uint32_t format;
651 	z_resampler_t resampler[Z_RESAMPLER_LAST];
652 } z_resampler_tab[] = {
653 #if BYTE_ORDER == LITTLE_ENDIAN || defined(SND_FEEDER_MULTIFORMAT)
654 	Z_RESAMPLER_ENTRY(S, 16, LE),
655 	Z_RESAMPLER_ENTRY(S, 32, LE),
656 #endif
657 #if BYTE_ORDER == BIG_ENDIAN || defined(SND_FEEDER_MULTIFORMAT)
658 	Z_RESAMPLER_ENTRY(S, 16, BE),
659 	Z_RESAMPLER_ENTRY(S, 32, BE),
660 #endif
661 #ifdef SND_FEEDER_MULTIFORMAT
662 	Z_RESAMPLER_ENTRY(S,  8, NE),
663 	Z_RESAMPLER_ENTRY(S, 24, LE),
664 	Z_RESAMPLER_ENTRY(S, 24, BE),
665 	Z_RESAMPLER_ENTRY(U,  8, NE),
666 	Z_RESAMPLER_ENTRY(U, 16, LE),
667 	Z_RESAMPLER_ENTRY(U, 24, LE),
668 	Z_RESAMPLER_ENTRY(U, 32, LE),
669 	Z_RESAMPLER_ENTRY(U, 16, BE),
670 	Z_RESAMPLER_ENTRY(U, 24, BE),
671 	Z_RESAMPLER_ENTRY(U, 32, BE),
672 #endif
673 };
674 
675 #define Z_RESAMPLER_TAB_SIZE						\
676 	((int32_t)(sizeof(z_resampler_tab) / sizeof(z_resampler_tab[0])))
677 
678 static void
679 z_resampler_reset(struct z_info *info)
680 {
681 
682 	info->src = info->rsrc - (info->rsrc % ((feeder_rate_round > 0 &&
683 	    info->rsrc > feeder_rate_round) ? feeder_rate_round : 1));
684 	info->dst = info->rdst - (info->rdst % ((feeder_rate_round > 0 &&
685 	    info->rdst > feeder_rate_round) ? feeder_rate_round : 1));
686 	info->z_gx = 1;
687 	info->z_gy = 1;
688 	info->z_alpha = 0;
689 	info->z_resample = NULL;
690 	info->z_size = 1;
691 	info->z_coeff = NULL;
692 	info->z_dcoeff = NULL;
693 	if (info->z_pcoeff != NULL) {
694 		kfree(info->z_pcoeff, M_DEVBUF);
695 		info->z_pcoeff = NULL;
696 	}
697 	info->z_scale = Z_ONE;
698 	info->z_dx = Z_FULL_ONE;
699 	info->z_dy = Z_FULL_ONE;
700 #ifdef Z_DIAGNOSTIC
701 	info->z_cycle = 0;
702 #endif
703 	if (info->quality < Z_QUALITY_MIN)
704 		info->quality = Z_QUALITY_MIN;
705 	else if (info->quality > Z_QUALITY_MAX)
706 		info->quality = Z_QUALITY_MAX;
707 }
708 
709 #ifdef Z_PARANOID
710 static int32_t
711 z_resampler_sinc_len(struct z_info *info)
712 {
713 	int32_t c, z, len, lmax;
714 
715 	if (!Z_IS_SINC(info))
716 		return (1);
717 
718 	/*
719 	 * A rather careful (or useless) way to calculate filter length.
720 	 * Z_SINC_LEN() itself is accurate enough to do its job. Extra
721 	 * sanity checking is not going to hurt though..
722 	 */
723 	c = 0;
724 	z = info->z_dy;
725 	len = 0;
726 	lmax = z_coeff_tab[Z_SINC_COEFF_IDX(info)].len;
727 
728 	do {
729 		c += z >> Z_SHIFT;
730 		z &= Z_MASK;
731 		z += info->z_dy;
732 	} while (c < lmax && ++len > 0);
733 
734 	if (len != Z_SINC_LEN(info)) {
735 #ifdef _KERNEL
736 		kprintf("%s(): sinc l=%d != Z_SINC_LEN=%d\n",
737 		    __func__, len, Z_SINC_LEN(info));
738 #else
739 		fprintf(stderr, "%s(): sinc l=%d != Z_SINC_LEN=%d\n",
740 		    __func__, len, Z_SINC_LEN(info));
741 		return (-1);
742 #endif
743 	}
744 
745 	return (len);
746 }
747 #else
748 #define z_resampler_sinc_len(i)		(Z_IS_SINC(i) ? Z_SINC_LEN(i) : 1)
749 #endif
750 
751 #define Z_POLYPHASE_COEFF_SHIFT		0
752 
753 /*
754  * Pick suitable polynomial interpolators based on filter oversampled ratio
755  * (2 ^ Z_DRIFT_SHIFT).
756  */
757 #if !(defined(Z_COEFF_INTERP_ZOH) || defined(Z_COEFF_INTERP_LINEAR) ||		\
758     defined(Z_COEFF_INTERP_QUADRATIC) || defined(Z_COEFF_INTERP_HERMITE) ||	\
759     defined(Z_COEFF_INTER_BSPLINE) || defined(Z_COEFF_INTERP_OPT32X) ||		\
760     defined(Z_COEFF_INTERP_OPT16X) || defined(Z_COEFF_INTERP_OPT8X) ||		\
761     defined(Z_COEFF_INTERP_OPT4X) || defined(Z_COEFF_INTERP_OPT2X))
762 #if Z_DRIFT_SHIFT >= 6
763 #define Z_COEFF_INTERP_BSPLINE		1
764 #elif Z_DRIFT_SHIFT >= 5
765 #define Z_COEFF_INTERP_OPT32X		1
766 #elif Z_DRIFT_SHIFT == 4
767 #define Z_COEFF_INTERP_OPT16X		1
768 #elif Z_DRIFT_SHIFT == 3
769 #define Z_COEFF_INTERP_OPT8X		1
770 #elif Z_DRIFT_SHIFT == 2
771 #define Z_COEFF_INTERP_OPT4X		1
772 #elif Z_DRIFT_SHIFT == 1
773 #define Z_COEFF_INTERP_OPT2X		1
774 #else
775 #error "Z_DRIFT_SHIFT screwed!"
776 #endif
777 #endif
778 
779 /*
780  * In classic polyphase mode, the actual coefficients for each phases need to
781  * be calculated based on default prototype filters. For highly oversampled
782  * filter, linear or quadradatic interpolator should be enough. Anything less
783  * than that require 'special' interpolators to reduce interpolation errors.
784  *
785  * "Polynomial Interpolators for High-Quality Resampling of Oversampled Audio"
786  *    by Olli Niemitalo
787  *    - http://www.student.oulu.fi/~oniemita/dsp/deip.pdf
788  *
789  */
790 static int32_t
791 z_coeff_interpolate(int32_t z, int32_t *z_coeff)
792 {
793 	int32_t coeff;
794 #if defined(Z_COEFF_INTERP_ZOH)
795 
796 	/* 1-point, 0th-order (Zero Order Hold) */
797 	z = z;
798 	coeff = z_coeff[0];
799 #elif defined(Z_COEFF_INTERP_LINEAR)
800 	int32_t zl0, zl1;
801 
802 	/* 2-point, 1st-order Linear */
803 	zl0 = z_coeff[0];
804 	zl1 = z_coeff[1] - z_coeff[0];
805 
806 	coeff = Z_RSHIFT((int64_t)zl1 * z, Z_SHIFT) + zl0;
807 #elif defined(Z_COEFF_INTERP_QUADRATIC)
808 	int32_t zq0, zq1, zq2;
809 
810 	/* 3-point, 2nd-order Quadratic */
811 	zq0 = z_coeff[0];
812 	zq1 = z_coeff[1] - z_coeff[-1];
813 	zq2 = z_coeff[1] + z_coeff[-1] - (z_coeff[0] << 1);
814 
815 	coeff = Z_RSHIFT((Z_RSHIFT((int64_t)zq2 * z, Z_SHIFT) +
816 	    zq1) * z, Z_SHIFT + 1) + zq0;
817 #elif defined(Z_COEFF_INTERP_HERMITE)
818 	int32_t zh0, zh1, zh2, zh3;
819 
820 	/* 4-point, 3rd-order Hermite */
821 	zh0 = z_coeff[0];
822 	zh1 = z_coeff[1] - z_coeff[-1];
823 	zh2 = (z_coeff[-1] << 1) - (z_coeff[0] * 5) + (z_coeff[1] << 2) -
824 	    z_coeff[2];
825 	zh3 = z_coeff[2] - z_coeff[-1] + ((z_coeff[0] - z_coeff[1]) * 3);
826 
827 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((int64_t)zh3 * z, Z_SHIFT) +
828 	    zh2) * z, Z_SHIFT) + zh1) * z, Z_SHIFT + 1) + zh0;
829 #elif defined(Z_COEFF_INTERP_BSPLINE)
830 	int32_t zb0, zb1, zb2, zb3;
831 
832 	/* 4-point, 3rd-order B-Spline */
833 	zb0 = Z_RSHIFT(0x15555555LL * (((int64_t)z_coeff[0] << 2) +
834 	    z_coeff[-1] + z_coeff[1]), 30);
835 	zb1 = z_coeff[1] - z_coeff[-1];
836 	zb2 = z_coeff[-1] + z_coeff[1] - (z_coeff[0] << 1);
837 	zb3 = Z_RSHIFT(0x15555555LL * (((z_coeff[0] - z_coeff[1]) * 3) +
838 	    z_coeff[2] - z_coeff[-1]), 30);
839 
840 	coeff = (Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((int64_t)zb3 * z, Z_SHIFT) +
841 	    zb2) * z, Z_SHIFT) + zb1) * z, Z_SHIFT) + zb0 + 1) >> 1;
842 #elif defined(Z_COEFF_INTERP_OPT32X)
843 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
844 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
845 
846 	/* 6-point, 5th-order Optimal 32x */
847 	zoz = z - (Z_ONE >> 1);
848 	zoe1 = z_coeff[1] + z_coeff[0];
849 	zoe2 = z_coeff[2] + z_coeff[-1];
850 	zoe3 = z_coeff[3] + z_coeff[-2];
851 	zoo1 = z_coeff[1] - z_coeff[0];
852 	zoo2 = z_coeff[2] - z_coeff[-1];
853 	zoo3 = z_coeff[3] - z_coeff[-2];
854 
855 	zoc0 = Z_RSHIFT((0x1ac2260dLL * zoe1) + (0x0526cdcaLL * zoe2) +
856 	    (0x00170c29LL * zoe3), 30);
857 	zoc1 = Z_RSHIFT((0x14f8a49aLL * zoo1) + (0x0d6d1109LL * zoo2) +
858 	    (0x008cd4dcLL * zoo3), 30);
859 	zoc2 = Z_RSHIFT((-0x0d3e94a4LL * zoe1) + (0x0bddded4LL * zoe2) +
860 	    (0x0160b5d0LL * zoe3), 30);
861 	zoc3 = Z_RSHIFT((-0x0de10cc4LL * zoo1) + (0x019b2a7dLL * zoo2) +
862 	    (0x01cfe914LL * zoo3), 30);
863 	zoc4 = Z_RSHIFT((0x02aa12d7LL * zoe1) + (-0x03ff1bb3LL * zoe2) +
864 	    (0x015508ddLL * zoe3), 30);
865 	zoc5 = Z_RSHIFT((0x051d29e5LL * zoo1) + (-0x028e7647LL * zoo2) +
866 	    (0x0082d81aLL * zoo3), 30);
867 
868 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
869 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
870 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
871 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
872 #elif defined(Z_COEFF_INTERP_OPT16X)
873 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
874 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
875 
876 	/* 6-point, 5th-order Optimal 16x */
877 	zoz = z - (Z_ONE >> 1);
878 	zoe1 = z_coeff[1] + z_coeff[0];
879 	zoe2 = z_coeff[2] + z_coeff[-1];
880 	zoe3 = z_coeff[3] + z_coeff[-2];
881 	zoo1 = z_coeff[1] - z_coeff[0];
882 	zoo2 = z_coeff[2] - z_coeff[-1];
883 	zoo3 = z_coeff[3] - z_coeff[-2];
884 
885 	zoc0 = Z_RSHIFT((0x1ac2260dLL * zoe1) + (0x0526cdcaLL * zoe2) +
886 	    (0x00170c29LL * zoe3), 30);
887 	zoc1 = Z_RSHIFT((0x14f8a49aLL * zoo1) + (0x0d6d1109LL * zoo2) +
888 	    (0x008cd4dcLL * zoo3), 30);
889 	zoc2 = Z_RSHIFT((-0x0d3e94a4LL * zoe1) + (0x0bddded4LL * zoe2) +
890 	    (0x0160b5d0LL * zoe3), 30);
891 	zoc3 = Z_RSHIFT((-0x0de10cc4LL * zoo1) + (0x019b2a7dLL * zoo2) +
892 	    (0x01cfe914LL * zoo3), 30);
893 	zoc4 = Z_RSHIFT((0x02aa12d7LL * zoe1) + (-0x03ff1bb3LL * zoe2) +
894 	    (0x015508ddLL * zoe3), 30);
895 	zoc5 = Z_RSHIFT((0x051d29e5LL * zoo1) + (-0x028e7647LL * zoo2) +
896 	    (0x0082d81aLL * zoo3), 30);
897 
898 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
899 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
900 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
901 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
902 #elif defined(Z_COEFF_INTERP_OPT8X)
903 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
904 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
905 
906 	/* 6-point, 5th-order Optimal 8x */
907 	zoz = z - (Z_ONE >> 1);
908 	zoe1 = z_coeff[1] + z_coeff[0];
909 	zoe2 = z_coeff[2] + z_coeff[-1];
910 	zoe3 = z_coeff[3] + z_coeff[-2];
911 	zoo1 = z_coeff[1] - z_coeff[0];
912 	zoo2 = z_coeff[2] - z_coeff[-1];
913 	zoo3 = z_coeff[3] - z_coeff[-2];
914 
915 	zoc0 = Z_RSHIFT((0x1aa9b47dLL * zoe1) + (0x053d9944LL * zoe2) +
916 	    (0x0018b23fLL * zoe3), 30);
917 	zoc1 = Z_RSHIFT((0x14a104d1LL * zoo1) + (0x0d7d2504LL * zoo2) +
918 	    (0x0094b599LL * zoo3), 30);
919 	zoc2 = Z_RSHIFT((-0x0d22530bLL * zoe1) + (0x0bb37a2cLL * zoe2) +
920 	    (0x016ed8e0LL * zoe3), 30);
921 	zoc3 = Z_RSHIFT((-0x0d744b1cLL * zoo1) + (0x01649591LL * zoo2) +
922 	    (0x01dae93aLL * zoo3), 30);
923 	zoc4 = Z_RSHIFT((0x02a7ee1bLL * zoe1) + (-0x03fbdb24LL * zoe2) +
924 	    (0x0153ed07LL * zoe3), 30);
925 	zoc5 = Z_RSHIFT((0x04cf9b6cLL * zoo1) + (-0x0266b378LL * zoo2) +
926 	    (0x007a7c26LL * zoo3), 30);
927 
928 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
929 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
930 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
931 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
932 #elif defined(Z_COEFF_INTERP_OPT4X)
933 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
934 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
935 
936 	/* 6-point, 5th-order Optimal 4x */
937 	zoz = z - (Z_ONE >> 1);
938 	zoe1 = z_coeff[1] + z_coeff[0];
939 	zoe2 = z_coeff[2] + z_coeff[-1];
940 	zoe3 = z_coeff[3] + z_coeff[-2];
941 	zoo1 = z_coeff[1] - z_coeff[0];
942 	zoo2 = z_coeff[2] - z_coeff[-1];
943 	zoo3 = z_coeff[3] - z_coeff[-2];
944 
945 	zoc0 = Z_RSHIFT((0x1a8eda43LL * zoe1) + (0x0556ee38LL * zoe2) +
946 	    (0x001a3784LL * zoe3), 30);
947 	zoc1 = Z_RSHIFT((0x143d863eLL * zoo1) + (0x0d910e36LL * zoo2) +
948 	    (0x009ca889LL * zoo3), 30);
949 	zoc2 = Z_RSHIFT((-0x0d026821LL * zoe1) + (0x0b837773LL * zoe2) +
950 	    (0x017ef0c6LL * zoe3), 30);
951 	zoc3 = Z_RSHIFT((-0x0cef1502LL * zoo1) + (0x01207a8eLL * zoo2) +
952 	    (0x01e936dbLL * zoo3), 30);
953 	zoc4 = Z_RSHIFT((0x029fe643LL * zoe1) + (-0x03ef3fc8LL * zoe2) +
954 	    (0x014f5923LL * zoe3), 30);
955 	zoc5 = Z_RSHIFT((0x043a9d08LL * zoo1) + (-0x02154febLL * zoo2) +
956 	    (0x00670dbdLL * zoo3), 30);
957 
958 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
959 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
960 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
961 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
962 #elif defined(Z_COEFF_INTERP_OPT2X)
963 	int32_t zoz, zoe1, zoe2, zoe3, zoo1, zoo2, zoo3;
964 	int32_t zoc0, zoc1, zoc2, zoc3, zoc4, zoc5;
965 
966 	/* 6-point, 5th-order Optimal 2x */
967 	zoz = z - (Z_ONE >> 1);
968 	zoe1 = z_coeff[1] + z_coeff[0];
969 	zoe2 = z_coeff[2] + z_coeff[-1];
970 	zoe3 = z_coeff[3] + z_coeff[-2];
971 	zoo1 = z_coeff[1] - z_coeff[0];
972 	zoo2 = z_coeff[2] - z_coeff[-1];
973 	zoo3 = z_coeff[3] - z_coeff[-2];
974 
975 	zoc0 = Z_RSHIFT((0x19edb6fdLL * zoe1) + (0x05ebd062LL * zoe2) +
976 	    (0x00267881LL * zoe3), 30);
977 	zoc1 = Z_RSHIFT((0x1223af76LL * zoo1) + (0x0de3dd6bLL * zoo2) +
978 	    (0x00d683cdLL * zoo3), 30);
979 	zoc2 = Z_RSHIFT((-0x0c3ee068LL * zoe1) + (0x0a5c3769LL * zoe2) +
980 	    (0x01e2aceaLL * zoe3), 30);
981 	zoc3 = Z_RSHIFT((-0x0a8ab614LL * zoo1) + (-0x0019522eLL * zoo2) +
982 	    (0x022cefc7LL * zoo3), 30);
983 	zoc4 = Z_RSHIFT((0x0276187dLL * zoe1) + (-0x03a801e8LL * zoe2) +
984 	    (0x0131d935LL * zoe3), 30);
985 	zoc5 = Z_RSHIFT((0x02c373f5LL * zoo1) + (-0x01275f83LL * zoo2) +
986 	    (0x0018ee79LL * zoo3), 30);
987 
988 	coeff = Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT((Z_RSHIFT(
989 	    (int64_t)zoc5 * zoz, Z_SHIFT) +
990 	    zoc4) * zoz, Z_SHIFT) + zoc3) * zoz, Z_SHIFT) +
991 	    zoc2) * zoz, Z_SHIFT) + zoc1) * zoz, Z_SHIFT) + zoc0;
992 #else
993 #error "Interpolation type screwed!"
994 #endif
995 
996 #if Z_POLYPHASE_COEFF_SHIFT > 0
997 	coeff = Z_RSHIFT(coeff, Z_POLYPHASE_COEFF_SHIFT);
998 #endif
999 	return (coeff);
1000 }
1001 
1002 static int
1003 z_resampler_build_polyphase(struct z_info *info)
1004 {
1005 	int32_t alpha, c, i, z, idx;
1006 
1007 	/* Let this be here first. */
1008 	if (info->z_pcoeff != NULL) {
1009 		kfree(info->z_pcoeff, M_DEVBUF);
1010 		info->z_pcoeff = NULL;
1011 	}
1012 
1013 	if (feeder_rate_polyphase_max < 1)
1014 		return (ENOTSUP);
1015 
1016 	if (((int64_t)info->z_size * info->z_gy * 2) >
1017 	    feeder_rate_polyphase_max) {
1018 #ifndef _KERNEL
1019 		fprintf(stderr, "Polyphase entries exceed: [%d/%d] %jd > %d\n",
1020 		    info->z_gx, info->z_gy,
1021 		    (intmax_t)info->z_size * info->z_gy * 2,
1022 		    feeder_rate_polyphase_max);
1023 #endif
1024 		return (E2BIG);
1025 	}
1026 
1027 	info->z_pcoeff = kmalloc(sizeof(int32_t) *
1028 	    info->z_size * info->z_gy * 2, M_DEVBUF, M_WAITOK | M_ZERO);
1029 	if (info->z_pcoeff == NULL)
1030 		return (ENOMEM);
1031 
1032 	for (alpha = 0; alpha < info->z_gy; alpha++) {
1033 		z = alpha * info->z_dx;
1034 		c = 0;
1035 		for (i = info->z_size; i != 0; i--) {
1036 			c += z >> Z_SHIFT;
1037 			z &= Z_MASK;
1038 			idx = (alpha * info->z_size * 2) +
1039 			    (info->z_size * 2) - i;
1040 			info->z_pcoeff[idx] =
1041 			    z_coeff_interpolate(z, info->z_coeff + c);
1042 			z += info->z_dy;
1043 		}
1044 		z = info->z_dy - (alpha * info->z_dx);
1045 		c = 0;
1046 		for (i = info->z_size; i != 0; i--) {
1047 			c += z >> Z_SHIFT;
1048 			z &= Z_MASK;
1049 			idx = (alpha * info->z_size * 2) + i - 1;
1050 			info->z_pcoeff[idx] =
1051 			    z_coeff_interpolate(z, info->z_coeff + c);
1052 			z += info->z_dy;
1053 		}
1054 	}
1055 
1056 #ifndef _KERNEL
1057 	fprintf(stderr, "Polyphase: [%d/%d] %d entries\n",
1058 	    info->z_gx, info->z_gy, info->z_size * info->z_gy * 2);
1059 #endif
1060 
1061 	return (0);
1062 }
1063 
1064 static int
1065 z_resampler_setup(struct pcm_feeder *f)
1066 {
1067 	struct z_info *info;
1068 	int64_t gy2gx_max, gx2gy_max;
1069 	uint32_t format;
1070 	int32_t align, i, z_scale;
1071 	int adaptive;
1072 
1073 	info = f->data;
1074 	z_resampler_reset(info);
1075 
1076 	if (info->src == info->dst)
1077 		return (0);
1078 
1079 	/* Shrink by greatest common divisor. */
1080 	i = z_gcd(info->src, info->dst);
1081 	info->z_gx = info->src / i;
1082 	info->z_gy = info->dst / i;
1083 
1084 	/* Too big, or too small. Bail out. */
1085 	if (!(Z_FACTOR_SAFE(info->z_gx) && Z_FACTOR_SAFE(info->z_gy)))
1086 		return (EINVAL);
1087 
1088 	format = f->desc->in;
1089 	adaptive = 0;
1090 	z_scale = 0;
1091 
1092 	/*
1093 	 * Setup everything: filter length, conversion factor, etc.
1094 	 */
1095 	if (Z_IS_SINC(info)) {
1096 		/*
1097 		 * Downsampling, or upsampling scaling factor. As long as the
1098 		 * factor can be represented by a fraction of 1 << Z_SHIFT,
1099 		 * we're pretty much in business. Scaling is not needed for
1100 		 * upsampling, so we just slap Z_ONE there.
1101 		 */
1102 		if (info->z_gx > info->z_gy)
1103 			/*
1104 			 * If the downsampling ratio is beyond sanity,
1105 			 * enable semi-adaptive mode. Although handling
1106 			 * extreme ratio is possible, the result of the
1107 			 * conversion is just pointless, unworthy,
1108 			 * nonsensical noises, etc.
1109 			 */
1110 			if ((info->z_gx / info->z_gy) > Z_SINC_DOWNMAX)
1111 				z_scale = Z_ONE / Z_SINC_DOWNMAX;
1112 			else
1113 				z_scale = ((uint64_t)info->z_gy << Z_SHIFT) /
1114 				    info->z_gx;
1115 		else
1116 			z_scale = Z_ONE;
1117 
1118 		/*
1119 		 * This is actually impossible, unless anything above
1120 		 * overflow.
1121 		 */
1122 		if (z_scale < 1)
1123 			return (E2BIG);
1124 
1125 		/*
1126 		 * Calculate sample time/coefficients index drift. It is
1127 		 * a constant for upsampling, but downsampling require
1128 		 * heavy duty filtering with possible too long filters.
1129 		 * If anything goes wrong, revisit again and enable
1130 		 * adaptive mode.
1131 		 */
1132 z_setup_adaptive_sinc:
1133 		if (info->z_pcoeff != NULL) {
1134 			kfree(info->z_pcoeff, M_DEVBUF);
1135 			info->z_pcoeff = NULL;
1136 		}
1137 
1138 		if (adaptive == 0) {
1139 			info->z_dy = z_scale << Z_DRIFT_SHIFT;
1140 			if (info->z_dy < 1)
1141 				return (E2BIG);
1142 			info->z_scale = z_scale;
1143 		} else {
1144 			info->z_dy = Z_FULL_ONE;
1145 			info->z_scale = Z_ONE;
1146 		}
1147 
1148 #if 0
1149 #define Z_SCALE_DIV	10000
1150 #define Z_SCALE_LIMIT(s, v)						\
1151 	((((uint64_t)(s) * (v)) + (Z_SCALE_DIV >> 1)) / Z_SCALE_DIV)
1152 
1153 		info->z_scale = Z_SCALE_LIMIT(info->z_scale, 9780);
1154 #endif
1155 
1156 		/* Smallest drift increment. */
1157 		info->z_dx = info->z_dy / info->z_gy;
1158 
1159 		/*
1160 		 * Overflow or underflow. Try adaptive, let it continue and
1161 		 * retry.
1162 		 */
1163 		if (info->z_dx < 1) {
1164 			if (adaptive == 0) {
1165 				adaptive = 1;
1166 				goto z_setup_adaptive_sinc;
1167 			}
1168 			return (E2BIG);
1169 		}
1170 
1171 		/*
1172 		 * Round back output drift.
1173 		 */
1174 		info->z_dy = info->z_dx * info->z_gy;
1175 
1176 		for (i = 0; i < Z_COEFF_TAB_SIZE; i++) {
1177 			if (Z_SINC_COEFF_IDX(info) != i)
1178 				continue;
1179 			/*
1180 			 * Calculate required filter length and guard
1181 			 * against possible abusive result. Note that
1182 			 * this represents only 1/2 of the entire filter
1183 			 * length.
1184 			 */
1185 			info->z_size = z_resampler_sinc_len(info);
1186 
1187 			/*
1188 			 * Multiple of 2 rounding, for better accumulator
1189 			 * performance.
1190 			 */
1191 			info->z_size &= ~1;
1192 
1193 			if (info->z_size < 2 || info->z_size > Z_SINC_MAX) {
1194 				if (adaptive == 0) {
1195 					adaptive = 1;
1196 					goto z_setup_adaptive_sinc;
1197 				}
1198 				return (E2BIG);
1199 			}
1200 			info->z_coeff = z_coeff_tab[i].coeff + Z_COEFF_OFFSET;
1201 			info->z_dcoeff = z_coeff_tab[i].dcoeff;
1202 			break;
1203 		}
1204 
1205 		if (info->z_coeff == NULL || info->z_dcoeff == NULL)
1206 			return (EINVAL);
1207 	} else if (Z_IS_LINEAR(info)) {
1208 		/*
1209 		 * Don't put much effort if we're doing linear interpolation.
1210 		 * Just center the interpolation distance within Z_LINEAR_ONE,
1211 		 * and be happy about it.
1212 		 */
1213 		info->z_dx = Z_LINEAR_FULL_ONE / info->z_gy;
1214 	}
1215 
1216 	/*
1217 	 * We're safe for now, lets continue.. Look for our resampler
1218 	 * depending on configured format and quality.
1219 	 */
1220 	for (i = 0; i < Z_RESAMPLER_TAB_SIZE; i++) {
1221 		int ridx;
1222 
1223 		if (AFMT_ENCODING(format) != z_resampler_tab[i].format)
1224 			continue;
1225 		if (Z_IS_SINC(info) && adaptive == 0 &&
1226 		    z_resampler_build_polyphase(info) == 0)
1227 			ridx = Z_RESAMPLER_SINC_POLYPHASE;
1228 		else
1229 			ridx = Z_RESAMPLER_IDX(info);
1230 		info->z_resample = z_resampler_tab[i].resampler[ridx];
1231 		break;
1232 	}
1233 
1234 	if (info->z_resample == NULL)
1235 		return (EINVAL);
1236 
1237 	info->bps = AFMT_BPS(format);
1238 	align = info->channels * info->bps;
1239 
1240 	/*
1241 	 * Calculate largest value that can be fed into z_gy2gx() and
1242 	 * z_gx2gy() without causing (signed) 32bit overflow. z_gy2gx() will
1243 	 * be called early during feeding process to determine how much input
1244 	 * samples that is required to generate requested output, while
1245 	 * z_gx2gy() will be called just before samples filtering /
1246 	 * accumulation process based on available samples that has been
1247 	 * calculated using z_gx2gy().
1248 	 *
1249 	 * Now that is damn confusing, I guess ;-) .
1250 	 */
1251 	gy2gx_max = (((uint64_t)info->z_gy * INT32_MAX) - info->z_gy + 1) /
1252 	    info->z_gx;
1253 
1254 	if ((gy2gx_max * align) > SND_FXDIV_MAX)
1255 		gy2gx_max = SND_FXDIV_MAX / align;
1256 
1257 	if (gy2gx_max < 1)
1258 		return (E2BIG);
1259 
1260 	gx2gy_max = (((uint64_t)info->z_gx * INT32_MAX) - info->z_gy) /
1261 	    info->z_gy;
1262 
1263 	if (gx2gy_max > INT32_MAX)
1264 		gx2gy_max = INT32_MAX;
1265 
1266 	if (gx2gy_max < 1)
1267 		return (E2BIG);
1268 
1269 	/*
1270 	 * Ensure that z_gy2gx() at its largest possible calculated value
1271 	 * (alpha = 0) will not cause overflow further late during z_gx2gy()
1272 	 * stage.
1273 	 */
1274 	if (z_gy2gx(info, gy2gx_max) > _Z_GCAST(gx2gy_max))
1275 		return (E2BIG);
1276 
1277 	info->z_maxfeed = gy2gx_max * align;
1278 
1279 #ifdef Z_USE_ALPHADRIFT
1280 	info->z_startdrift = z_gy2gx(info, 1);
1281 	info->z_alphadrift = z_drift(info, info->z_startdrift, 1);
1282 #endif
1283 
1284 	i = z_gy2gx(info, 1);
1285 	info->z_full = z_roundpow2((info->z_size << 1) + i);
1286 
1287 	/*
1288 	 * Too big to be true, and overflowing left and right like mad ..
1289 	 */
1290 	if ((info->z_full * align) < 1) {
1291 		if (adaptive == 0 && Z_IS_SINC(info)) {
1292 			adaptive = 1;
1293 			goto z_setup_adaptive_sinc;
1294 		}
1295 		return (E2BIG);
1296 	}
1297 
1298 	/*
1299 	 * Increase full buffer size if its too small to reduce cyclic
1300 	 * buffer shifting in main conversion/feeder loop.
1301 	 */
1302 	while (info->z_full < Z_RESERVOIR_MAX &&
1303 	    (info->z_full - (info->z_size << 1)) < Z_RESERVOIR)
1304 		info->z_full <<= 1;
1305 
1306 	/* Initialize buffer position. */
1307 	info->z_mask = info->z_full - 1;
1308 	info->z_start = z_prev(info, info->z_size << 1, 1);
1309 	info->z_pos = z_next(info, info->z_start, 1);
1310 
1311 	/*
1312 	 * Allocate or reuse delay line buffer, whichever makes sense.
1313 	 */
1314 	i = info->z_full * align;
1315 	if (i < 1)
1316 		return (E2BIG);
1317 
1318 	if (info->z_delay == NULL || info->z_alloc < i ||
1319 	    i <= (info->z_alloc >> 1)) {
1320 		if (info->z_delay != NULL)
1321 			kfree(info->z_delay, M_DEVBUF);
1322 		info->z_delay = kmalloc(i, M_DEVBUF, M_WAITOK | M_ZERO);
1323 		if (info->z_delay == NULL)
1324 			return (ENOMEM);
1325 		info->z_alloc = i;
1326 	}
1327 
1328 	/*
1329 	 * Zero out head of buffer to avoid pops and clicks.
1330 	 */
1331 	memset(info->z_delay, sndbuf_zerodata(f->desc->out),
1332 	    info->z_pos * align);
1333 
1334 #ifdef Z_DIAGNOSTIC
1335 	/*
1336 	 * XXX Debuging mess !@#$%^
1337 	 */
1338 #define dumpz(x)	fprintf(stderr, "\t%12s = %10u : %-11d\n",	\
1339 			    "z_"__STRING(x), (uint32_t)info->z_##x,	\
1340 			    (int32_t)info->z_##x)
1341 	fprintf(stderr, "\n%s():\n", __func__);
1342 	fprintf(stderr, "\tchannels=%d, bps=%d, format=0x%08x, quality=%d\n",
1343 	    info->channels, info->bps, format, info->quality);
1344 	fprintf(stderr, "\t%d (%d) -> %d (%d), ",
1345 	    info->src, info->rsrc, info->dst, info->rdst);
1346 	fprintf(stderr, "[%d/%d]\n", info->z_gx, info->z_gy);
1347 	fprintf(stderr, "\tminreq=%d, ", z_gy2gx(info, 1));
1348 	if (adaptive != 0)
1349 		z_scale = Z_ONE;
1350 	fprintf(stderr, "factor=0x%08x/0x%08x (%f)\n",
1351 	    z_scale, Z_ONE, (double)z_scale / Z_ONE);
1352 	fprintf(stderr, "\tbase_length=%d, ", Z_SINC_BASE_LEN(info));
1353 	fprintf(stderr, "adaptive=%s\n", (adaptive != 0) ? "YES" : "NO");
1354 	dumpz(size);
1355 	dumpz(alloc);
1356 	if (info->z_alloc < 1024)
1357 		fprintf(stderr, "\t%15s%10d Bytes\n",
1358 		    "", info->z_alloc);
1359 	else if (info->z_alloc < (1024 << 10))
1360 		fprintf(stderr, "\t%15s%10d KBytes\n",
1361 		    "", info->z_alloc >> 10);
1362 	else if (info->z_alloc < (1024 << 20))
1363 		fprintf(stderr, "\t%15s%10d MBytes\n",
1364 		    "", info->z_alloc >> 20);
1365 	else
1366 		fprintf(stderr, "\t%15s%10d GBytes\n",
1367 		    "", info->z_alloc >> 30);
1368 	fprintf(stderr, "\t%12s   %10d (min output samples)\n",
1369 	    "",
1370 	    (int32_t)z_gx2gy(info, info->z_full - (info->z_size << 1)));
1371 	fprintf(stderr, "\t%12s   %10d (min allocated output samples)\n",
1372 	    "",
1373 	    (int32_t)z_gx2gy(info, (info->z_alloc / align) -
1374 	    (info->z_size << 1)));
1375 	fprintf(stderr, "\t%12s = %10d\n",
1376 	    "z_gy2gx()", (int32_t)z_gy2gx(info, 1));
1377 	fprintf(stderr, "\t%12s = %10d -> z_gy2gx() -> %d\n",
1378 	    "Max", (int32_t)gy2gx_max, (int32_t)z_gy2gx(info, gy2gx_max));
1379 	fprintf(stderr, "\t%12s = %10d\n",
1380 	    "z_gx2gy()", (int32_t)z_gx2gy(info, 1));
1381 	fprintf(stderr, "\t%12s = %10d -> z_gx2gy() -> %d\n",
1382 	    "Max", (int32_t)gx2gy_max, (int32_t)z_gx2gy(info, gx2gy_max));
1383 	dumpz(maxfeed);
1384 	dumpz(full);
1385 	dumpz(start);
1386 	dumpz(pos);
1387 	dumpz(scale);
1388 	fprintf(stderr, "\t%12s   %10f\n", "",
1389 	    (double)info->z_scale / Z_ONE);
1390 	dumpz(dx);
1391 	fprintf(stderr, "\t%12s   %10f\n", "",
1392 	    (double)info->z_dx / info->z_dy);
1393 	dumpz(dy);
1394 	fprintf(stderr, "\t%12s   %10d (drift step)\n", "",
1395 	    info->z_dy >> Z_SHIFT);
1396 	fprintf(stderr, "\t%12s   %10d (scaling differences)\n", "",
1397 	    (z_scale << Z_DRIFT_SHIFT) - info->z_dy);
1398 	fprintf(stderr, "\t%12s = %u bytes\n",
1399 	    "intpcm32_t", sizeof(intpcm32_t));
1400 	fprintf(stderr, "\t%12s = 0x%08x, smallest=%.16lf\n",
1401 	    "Z_ONE", Z_ONE, (double)1.0 / (double)Z_ONE);
1402 #endif
1403 
1404 	return (0);
1405 }
1406 
1407 static int
1408 z_resampler_set(struct pcm_feeder *f, int what, int32_t value)
1409 {
1410 	struct z_info *info;
1411 	int32_t oquality;
1412 
1413 	info = f->data;
1414 
1415 	switch (what) {
1416 	case Z_RATE_SRC:
1417 		if (value < feeder_rate_min || value > feeder_rate_max)
1418 			return (E2BIG);
1419 		if (value == info->rsrc)
1420 			return (0);
1421 		info->rsrc = value;
1422 		break;
1423 	case Z_RATE_DST:
1424 		if (value < feeder_rate_min || value > feeder_rate_max)
1425 			return (E2BIG);
1426 		if (value == info->rdst)
1427 			return (0);
1428 		info->rdst = value;
1429 		break;
1430 	case Z_RATE_QUALITY:
1431 		if (value < Z_QUALITY_MIN || value > Z_QUALITY_MAX)
1432 			return (EINVAL);
1433 		if (value == info->quality)
1434 			return (0);
1435 		/*
1436 		 * If we failed to set the requested quality, restore
1437 		 * the old one. We cannot afford leaving it broken since
1438 		 * passive feeder chains like vchans never reinitialize
1439 		 * itself.
1440 		 */
1441 		oquality = info->quality;
1442 		info->quality = value;
1443 		if (z_resampler_setup(f) == 0)
1444 			return (0);
1445 		info->quality = oquality;
1446 		break;
1447 	case Z_RATE_CHANNELS:
1448 		if (value < SND_CHN_MIN || value > SND_CHN_MAX)
1449 			return (EINVAL);
1450 		if (value == info->channels)
1451 			return (0);
1452 		info->channels = value;
1453 		break;
1454 	default:
1455 		return (EINVAL);
1456 		break;
1457 	}
1458 
1459 	return (z_resampler_setup(f));
1460 }
1461 
1462 static int
1463 z_resampler_get(struct pcm_feeder *f, int what)
1464 {
1465 	struct z_info *info;
1466 
1467 	info = f->data;
1468 
1469 	switch (what) {
1470 	case Z_RATE_SRC:
1471 		return (info->rsrc);
1472 		break;
1473 	case Z_RATE_DST:
1474 		return (info->rdst);
1475 		break;
1476 	case Z_RATE_QUALITY:
1477 		return (info->quality);
1478 		break;
1479 	case Z_RATE_CHANNELS:
1480 		return (info->channels);
1481 		break;
1482 	default:
1483 		break;
1484 	}
1485 
1486 	return (-1);
1487 }
1488 
1489 static int
1490 z_resampler_init(struct pcm_feeder *f)
1491 {
1492 	struct z_info *info;
1493 	int ret;
1494 
1495 	if (f->desc->in != f->desc->out)
1496 		return (EINVAL);
1497 
1498 	info = kmalloc(sizeof(*info), M_DEVBUF, M_WAITOK | M_ZERO);
1499 	if (info == NULL)
1500 		return (ENOMEM);
1501 
1502 	info->rsrc = Z_RATE_DEFAULT;
1503 	info->rdst = Z_RATE_DEFAULT;
1504 	info->quality = feeder_rate_quality;
1505 	info->channels = AFMT_CHANNEL(f->desc->in);
1506 
1507 	f->data = info;
1508 
1509 	ret = z_resampler_setup(f);
1510 	if (ret != 0) {
1511 		if (info->z_pcoeff != NULL)
1512 			kfree(info->z_pcoeff, M_DEVBUF);
1513 		if (info->z_delay != NULL)
1514 			kfree(info->z_delay, M_DEVBUF);
1515 		kfree(info, M_DEVBUF);
1516 		f->data = NULL;
1517 	}
1518 
1519 	return (ret);
1520 }
1521 
1522 static int
1523 z_resampler_free(struct pcm_feeder *f)
1524 {
1525 	struct z_info *info;
1526 
1527 	info = f->data;
1528 	if (info != NULL) {
1529 		if (info->z_pcoeff != NULL)
1530 			kfree(info->z_pcoeff, M_DEVBUF);
1531 		if (info->z_delay != NULL)
1532 			kfree(info->z_delay, M_DEVBUF);
1533 		kfree(info, M_DEVBUF);
1534 	}
1535 
1536 	f->data = NULL;
1537 
1538 	return (0);
1539 }
1540 
1541 static uint32_t
1542 z_resampler_feed_internal(struct pcm_feeder *f, struct pcm_channel *c,
1543     uint8_t *b, uint32_t count, void *source)
1544 {
1545 	struct z_info *info;
1546 	int32_t alphadrift, startdrift, reqout, ocount, reqin, align;
1547 	int32_t fetch, fetched, start, cp;
1548 	uint8_t *dst;
1549 
1550 	info = f->data;
1551 	if (info->z_resample == NULL)
1552 		return (z_feed(f->source, c, b, count, source));
1553 
1554 	/*
1555 	 * Calculate sample size alignment and amount of sample output.
1556 	 * We will do everything in sample domain, but at the end we
1557 	 * will jump back to byte domain.
1558 	 */
1559 	align = info->channels * info->bps;
1560 	ocount = SND_FXDIV(count, align);
1561 	if (ocount == 0)
1562 		return (0);
1563 
1564 	/*
1565 	 * Calculate amount of input samples that is needed to generate
1566 	 * exact amount of output.
1567 	 */
1568 	reqin = z_gy2gx(info, ocount) - z_fetched(info);
1569 
1570 #ifdef Z_USE_ALPHADRIFT
1571 	startdrift = info->z_startdrift;
1572 	alphadrift = info->z_alphadrift;
1573 #else
1574 	startdrift = _Z_GY2GX(info, 0, 1);
1575 	alphadrift = z_drift(info, startdrift, 1);
1576 #endif
1577 
1578 	dst = b;
1579 
1580 	do {
1581 		if (reqin != 0) {
1582 			fetch = z_min(z_free(info), reqin);
1583 			if (fetch == 0) {
1584 				/*
1585 				 * No more free spaces, so wind enough
1586 				 * samples back to the head of delay line
1587 				 * in byte domain.
1588 				 */
1589 				fetched = z_fetched(info);
1590 				start = z_prev(info, info->z_start,
1591 				    (info->z_size << 1) - 1);
1592 				cp = (info->z_size << 1) + fetched;
1593 				z_copy(info->z_delay + (start * align),
1594 				    info->z_delay, cp * align);
1595 				info->z_start =
1596 				    z_prev(info, info->z_size << 1, 1);
1597 				info->z_pos =
1598 				    z_next(info, info->z_start, fetched + 1);
1599 				fetch = z_min(z_free(info), reqin);
1600 #ifdef Z_DIAGNOSTIC
1601 				if (1) {
1602 					static uint32_t kk = 0;
1603 					fprintf(stderr,
1604 					    "Buffer Move: "
1605 					    "start=%d fetched=%d cp=%d "
1606 					    "cycle=%u [%u]\r",
1607 					    start, fetched, cp, info->z_cycle,
1608 					    ++kk);
1609 				}
1610 				info->z_cycle = 0;
1611 #endif
1612 			}
1613 			if (fetch != 0) {
1614 				/*
1615 				 * Fetch in byte domain and jump back
1616 				 * to sample domain.
1617 				 */
1618 				fetched = SND_FXDIV(z_feed(f->source, c,
1619 				    info->z_delay + (info->z_pos * align),
1620 				    fetch * align, source), align);
1621 				/*
1622 				 * Prepare to convert fetched buffer,
1623 				 * or mark us done if we cannot fulfill
1624 				 * the request.
1625 				 */
1626 				reqin -= fetched;
1627 				info->z_pos += fetched;
1628 				if (fetched != fetch)
1629 					reqin = 0;
1630 			}
1631 		}
1632 
1633 		reqout = z_min(z_gx2gy(info, z_fetched(info)), ocount);
1634 		if (reqout != 0) {
1635 			ocount -= reqout;
1636 
1637 			/*
1638 			 * Drift.. drift.. drift..
1639 			 *
1640 			 * Notice that there are 2 methods of doing the drift
1641 			 * operations: The former is much cleaner (in a sense
1642 			 * of mathematical readings of my eyes), but slower
1643 			 * due to integer division in z_gy2gx(). Nevertheless,
1644 			 * both should give the same exact accurate drifting
1645 			 * results, so the later is favourable.
1646 			 */
1647 			do {
1648 				info->z_resample(info, dst);
1649 #if 0
1650 				startdrift = z_gy2gx(info, 1);
1651 				alphadrift = z_drift(info, startdrift, 1);
1652 				info->z_start += startdrift;
1653 				info->z_alpha += alphadrift;
1654 #else
1655 				info->z_alpha += alphadrift;
1656 				if (info->z_alpha < info->z_gy)
1657 					info->z_start += startdrift;
1658 				else {
1659 					info->z_start += startdrift - 1;
1660 					info->z_alpha -= info->z_gy;
1661 				}
1662 #endif
1663 				dst += align;
1664 #ifdef Z_DIAGNOSTIC
1665 				info->z_cycle++;
1666 #endif
1667 			} while (--reqout != 0);
1668 		}
1669 	} while (reqin != 0 && ocount != 0);
1670 
1671 	/*
1672 	 * Back to byte domain..
1673 	 */
1674 	return (dst - b);
1675 }
1676 
1677 static int
1678 z_resampler_feed(struct pcm_feeder *f, struct pcm_channel *c, uint8_t *b,
1679     uint32_t count, void *source)
1680 {
1681 	uint32_t feed, maxfeed, left;
1682 
1683 	/*
1684 	 * Split count to smaller chunks to avoid possible 32bit overflow.
1685 	 */
1686 	maxfeed = ((struct z_info *)(f->data))->z_maxfeed;
1687 	left = count;
1688 
1689 	do {
1690 		feed = z_resampler_feed_internal(f, c, b,
1691 		    z_min(maxfeed, left), source);
1692 		b += feed;
1693 		left -= feed;
1694 	} while (left != 0 && feed != 0);
1695 
1696 	return (count - left);
1697 }
1698 
1699 static struct pcm_feederdesc feeder_rate_desc[] = {
1700 	{ FEEDER_RATE, 0, 0, 0, 0 },
1701 	{ 0, 0, 0, 0, 0 },
1702 };
1703 
1704 static kobj_method_t feeder_rate_methods[] = {
1705 	KOBJMETHOD(feeder_init,		z_resampler_init),
1706 	KOBJMETHOD(feeder_free,		z_resampler_free),
1707 	KOBJMETHOD(feeder_set,		z_resampler_set),
1708 	KOBJMETHOD(feeder_get,		z_resampler_get),
1709 	KOBJMETHOD(feeder_feed,		z_resampler_feed),
1710 	KOBJMETHOD_END
1711 };
1712 
1713 FEEDER_DECLARE(feeder_rate, NULL);
1714