sys/dev/sound/pcm/feeder_rate.c

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