blob: 5556e15e0fa482b59a62505f5f747229e81e7c84
1 | /* |
2 | * libmad - MPEG audio decoder library |
3 | * Copyright (C) 2000-2004 Underbit Technologies, Inc. |
4 | * |
5 | * This program is free software; you can redistribute it and/or modify |
6 | * it under the terms of the GNU General Public License as published by |
7 | * the Free Software Foundation; either version 2 of the License, or |
8 | * (at your option) any later version. |
9 | * |
10 | * This program is distributed in the hope that it will be useful, |
11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
13 | * GNU General Public License for more details. |
14 | * |
15 | * You should have received a copy of the GNU General Public License |
16 | * along with this program; if not, write to the Free Software |
17 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
18 | * |
19 | * $Id: fixed.h,v 1.38 2004/02/17 02:02:03 rob Exp $ |
20 | */ |
21 | |
22 | # ifndef LIBMAD_FIXED_H |
23 | # define LIBMAD_FIXED_H |
24 | #include "config.h" |
25 | # if SIZEOF_INT >= 4 |
26 | typedef signed int mad_fixed_t; |
27 | |
28 | typedef signed int mad_fixed64hi_t; |
29 | typedef unsigned int mad_fixed64lo_t; |
30 | # else |
31 | typedef signed long mad_fixed_t; |
32 | |
33 | typedef signed long mad_fixed64hi_t; |
34 | typedef unsigned long mad_fixed64lo_t; |
35 | # endif |
36 | |
37 | # if defined(_MSC_VER) |
38 | # define mad_fixed64_t signed __int64 |
39 | # elif 1 || defined(__GNUC__) |
40 | # define mad_fixed64_t signed long long |
41 | # endif |
42 | |
43 | # if defined(FPM_FLOAT) |
44 | typedef double mad_sample_t; |
45 | # else |
46 | typedef mad_fixed_t mad_sample_t; |
47 | # endif |
48 | |
49 | /* |
50 | * Fixed-point format: 0xABBBBBBB |
51 | * A == whole part (sign + 3 bits) |
52 | * B == fractional part (28 bits) |
53 | * |
54 | * Values are signed two's complement, so the effective range is: |
55 | * 0x80000000 to 0x7fffffff |
56 | * -8.0 to +7.9999999962747097015380859375 |
57 | * |
58 | * The smallest representable value is: |
59 | * 0x00000001 == 0.0000000037252902984619140625 (i.e. about 3.725e-9) |
60 | * |
61 | * 28 bits of fractional accuracy represent about |
62 | * 8.6 digits of decimal accuracy. |
63 | * |
64 | * Fixed-point numbers can be added or subtracted as normal |
65 | * integers, but multiplication requires shifting the 64-bit result |
66 | * from 56 fractional bits back to 28 (and rounding.) |
67 | * |
68 | * Changing the definition of MAD_F_FRACBITS is only partially |
69 | * supported, and must be done with care. |
70 | */ |
71 | |
72 | # define MAD_F_FRACBITS 28 |
73 | |
74 | # if MAD_F_FRACBITS == 28 |
75 | # define MAD_F(x) ((mad_fixed_t) (x##L)) |
76 | # else |
77 | # if MAD_F_FRACBITS < 28 |
78 | # warning "MAD_F_FRACBITS < 28" |
79 | # define MAD_F(x) ((mad_fixed_t) \ |
80 | (((x##L) + \ |
81 | (1L << (28 - MAD_F_FRACBITS - 1))) >> \ |
82 | (28 - MAD_F_FRACBITS))) |
83 | # elif MAD_F_FRACBITS > 28 |
84 | # error "MAD_F_FRACBITS > 28 not currently supported" |
85 | # define MAD_F(x) ((mad_fixed_t) \ |
86 | ((x##L) << (MAD_F_FRACBITS - 28))) |
87 | # endif |
88 | # endif |
89 | |
90 | # define MAD_F_MIN ((mad_fixed_t) -0x80000000L) |
91 | # define MAD_F_MAX ((mad_fixed_t) +0x7fffffffL) |
92 | |
93 | # define MAD_F_ONE MAD_F(0x10000000) |
94 | |
95 | # define mad_f_tofixed(x) ((mad_fixed_t) \ |
96 | ((x) * (double) (1L << MAD_F_FRACBITS) + 0.5)) |
97 | # define mad_f_todouble(x) ((double) \ |
98 | ((x) / (double) (1L << MAD_F_FRACBITS))) |
99 | |
100 | # define mad_f_intpart(x) ((x) >> MAD_F_FRACBITS) |
101 | # define mad_f_fracpart(x) ((x) & ((1L << MAD_F_FRACBITS) - 1)) |
102 | /* (x should be positive) */ |
103 | |
104 | # define mad_f_fromint(x) ((x) << MAD_F_FRACBITS) |
105 | |
106 | # define mad_f_add(x, y) ((x) + (y)) |
107 | # define mad_f_sub(x, y) ((x) - (y)) |
108 | |
109 | # if defined(FPM_FLOAT) |
110 | # error "FPM_FLOAT not yet supported" |
111 | |
112 | # undef MAD_F |
113 | # define MAD_F(x) mad_f_todouble(x) |
114 | |
115 | # define mad_f_mul(x, y) ((x) * (y)) |
116 | # define mad_f_scale64 |
117 | |
118 | # undef ASO_ZEROCHECK |
119 | |
120 | # elif defined(FPM_64BIT) |
121 | |
122 | /* |
123 | * This version should be the most accurate if 64-bit types are supported by |
124 | * the compiler, although it may not be the most efficient. |
125 | */ |
126 | # if defined(OPT_ACCURACY) |
127 | # define mad_f_mul(x, y) \ |
128 | ((mad_fixed_t) \ |
129 | ((((mad_fixed64_t) (x) * (y)) + \ |
130 | (1L << (MAD_F_SCALEBITS - 1))) >> MAD_F_SCALEBITS)) |
131 | # else |
132 | # define mad_f_mul(x, y) \ |
133 | ((mad_fixed_t) (((mad_fixed64_t) (x) * (y)) >> MAD_F_SCALEBITS)) |
134 | # endif |
135 | |
136 | # define MAD_F_SCALEBITS MAD_F_FRACBITS |
137 | |
138 | /* --- Intel --------------------------------------------------------------- */ |
139 | |
140 | # elif defined(FPM_INTEL) |
141 | |
142 | # if defined(_MSC_VER) |
143 | # pragma warning(push) |
144 | # pragma warning(disable: 4035) /* no return value */ |
145 | static __forceinline |
146 | mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) |
147 | { |
148 | enum { |
149 | fracbits = MAD_F_FRACBITS |
150 | }; |
151 | |
152 | __asm { |
153 | mov eax, x |
154 | imul y |
155 | shrd eax, edx, fracbits |
156 | } |
157 | |
158 | /* implicit return of eax */ |
159 | } |
160 | # pragma warning(pop) |
161 | |
162 | # define mad_f_mul mad_f_mul_inline |
163 | # define mad_f_scale64 |
164 | # else |
165 | /* |
166 | * This Intel version is fast and accurate; the disposition of the least |
167 | * significant bit depends on OPT_ACCURACY via mad_f_scale64(). |
168 | */ |
169 | # define MAD_F_MLX(hi, lo, x, y) \ |
170 | asm ("imull %3" \ |
171 | : "=a" (lo), "=d" (hi) \ |
172 | : "%a" (x), "rm" (y) \ |
173 | : "cc") |
174 | |
175 | # if defined(OPT_ACCURACY) |
176 | /* |
177 | * This gives best accuracy but is not very fast. |
178 | */ |
179 | # define MAD_F_MLA(hi, lo, x, y) \ |
180 | ({ mad_fixed64hi_t __hi; \ |
181 | mad_fixed64lo_t __lo; \ |
182 | MAD_F_MLX(__hi, __lo, (x), (y)); \ |
183 | asm ("addl %2,%0\n\t" \ |
184 | "adcl %3,%1" \ |
185 | : "=rm" (lo), "=rm" (hi) \ |
186 | : "r" (__lo), "r" (__hi), "0" (lo), "1" (hi) \ |
187 | : "cc"); \ |
188 | }) |
189 | # endif /* OPT_ACCURACY */ |
190 | |
191 | # if defined(OPT_ACCURACY) |
192 | /* |
193 | * Surprisingly, this is faster than SHRD followed by ADC. |
194 | */ |
195 | # define mad_f_scale64(hi, lo) \ |
196 | ({ mad_fixed64hi_t __hi_; \ |
197 | mad_fixed64lo_t __lo_; \ |
198 | mad_fixed_t __result; \ |
199 | asm ("addl %4,%2\n\t" \ |
200 | "adcl %5,%3" \ |
201 | : "=rm" (__lo_), "=rm" (__hi_) \ |
202 | : "0" (lo), "1" (hi), \ |
203 | "ir" (1L << (MAD_F_SCALEBITS - 1)), "ir" (0) \ |
204 | : "cc"); \ |
205 | asm ("shrdl %3,%2,%1" \ |
206 | : "=rm" (__result) \ |
207 | : "0" (__lo_), "r" (__hi_), "I" (MAD_F_SCALEBITS) \ |
208 | : "cc"); \ |
209 | __result; \ |
210 | }) |
211 | # elif defined(OPT_INTEL) |
212 | /* |
213 | * Alternate Intel scaling that may or may not perform better. |
214 | */ |
215 | # define mad_f_scale64(hi, lo) \ |
216 | ({ mad_fixed_t __result; \ |
217 | asm ("shrl %3,%1\n\t" \ |
218 | "shll %4,%2\n\t" \ |
219 | "orl %2,%1" \ |
220 | : "=rm" (__result) \ |
221 | : "0" (lo), "r" (hi), \ |
222 | "I" (MAD_F_SCALEBITS), "I" (32 - MAD_F_SCALEBITS) \ |
223 | : "cc"); \ |
224 | __result; \ |
225 | }) |
226 | # else |
227 | # define mad_f_scale64(hi, lo) \ |
228 | ({ mad_fixed_t __result; \ |
229 | asm ("shrdl %3,%2,%1" \ |
230 | : "=rm" (__result) \ |
231 | : "0" (lo), "r" (hi), "I" (MAD_F_SCALEBITS) \ |
232 | : "cc"); \ |
233 | __result; \ |
234 | }) |
235 | # endif /* OPT_ACCURACY */ |
236 | |
237 | # define MAD_F_SCALEBITS MAD_F_FRACBITS |
238 | # endif |
239 | |
240 | /* --- ARM ----------------------------------------------------------------- */ |
241 | |
242 | # elif defined(FPM_ARM) |
243 | |
244 | /* |
245 | * This ARM V4 version is as accurate as FPM_64BIT but much faster. The |
246 | * least significant bit is properly rounded at no CPU cycle cost! |
247 | */ |
248 | # if 1 |
249 | /* |
250 | * This is faster than the default implementation via MAD_F_MLX() and |
251 | * mad_f_scale64(). |
252 | */ |
253 | # define mad_f_mul(x, y) \ |
254 | ({ mad_fixed64hi_t __hi; \ |
255 | mad_fixed64lo_t __lo; \ |
256 | mad_fixed_t __result; \ |
257 | asm ("smull %0, %1, %3, %4\n\t" \ |
258 | "movs %0, %0, lsr %5\n\t" \ |
259 | "adc %2, %0, %1, lsl %6" \ |
260 | : "=&r" (__lo), "=&r" (__hi), "=r" (__result) \ |
261 | : "%r" (x), "r" (y), \ |
262 | "M" (MAD_F_SCALEBITS), "M" (32 - MAD_F_SCALEBITS) \ |
263 | : "cc"); \ |
264 | __result; \ |
265 | }) |
266 | # endif |
267 | |
268 | # define MAD_F_MLX(hi, lo, x, y) \ |
269 | asm ("smull %0, %1, %2, %3" \ |
270 | : "=&r" (lo), "=&r" (hi) \ |
271 | : "%r" (x), "r" (y)) |
272 | |
273 | # define MAD_F_MLA(hi, lo, x, y) \ |
274 | asm ("smlal %0, %1, %2, %3" \ |
275 | : "+r" (lo), "+r" (hi) \ |
276 | : "%r" (x), "r" (y)) |
277 | |
278 | # define MAD_F_MLN(hi, lo) \ |
279 | asm ("rsbs %0, %2, #0\n\t" \ |
280 | "rsc %1, %3, #0" \ |
281 | : "=r" (lo), "=r" (hi) \ |
282 | : "0" (lo), "1" (hi) \ |
283 | : "cc") |
284 | |
285 | # define mad_f_scale64(hi, lo) \ |
286 | ({ mad_fixed_t __result; \ |
287 | asm ("movs %0, %1, lsr %3\n\t" \ |
288 | "adc %0, %0, %2, lsl %4" \ |
289 | : "=&r" (__result) \ |
290 | : "r" (lo), "r" (hi), \ |
291 | "M" (MAD_F_SCALEBITS), "M" (32 - MAD_F_SCALEBITS) \ |
292 | : "cc"); \ |
293 | __result; \ |
294 | }) |
295 | |
296 | # define MAD_F_SCALEBITS MAD_F_FRACBITS |
297 | |
298 | /* --- MIPS ---------------------------------------------------------------- */ |
299 | |
300 | # elif defined(FPM_MIPS) |
301 | |
302 | /* |
303 | * This MIPS version is fast and accurate; the disposition of the least |
304 | * significant bit depends on OPT_ACCURACY via mad_f_scale64(). |
305 | */ |
306 | # define MAD_F_MLX(hi, lo, x, y) \ |
307 | asm ("mult %2,%3" \ |
308 | : "=l" (lo), "=h" (hi) \ |
309 | : "%r" (x), "r" (y)) |
310 | |
311 | # if defined(HAVE_MADD_ASM) |
312 | # define MAD_F_MLA(hi, lo, x, y) \ |
313 | asm ("madd %2,%3" \ |
314 | : "+l" (lo), "+h" (hi) \ |
315 | : "%r" (x), "r" (y)) |
316 | # elif defined(HAVE_MADD16_ASM) |
317 | /* |
318 | * This loses significant accuracy due to the 16-bit integer limit in the |
319 | * multiply/accumulate instruction. |
320 | */ |
321 | # define MAD_F_ML0(hi, lo, x, y) \ |
322 | asm ("mult %2,%3" \ |
323 | : "=l" (lo), "=h" (hi) \ |
324 | : "%r" ((x) >> 12), "r" ((y) >> 16)) |
325 | # define MAD_F_MLA(hi, lo, x, y) \ |
326 | asm ("madd16 %2,%3" \ |
327 | : "+l" (lo), "+h" (hi) \ |
328 | : "%r" ((x) >> 12), "r" ((y) >> 16)) |
329 | # define MAD_F_MLZ(hi, lo) ((mad_fixed_t) (lo)) |
330 | # endif |
331 | |
332 | # if defined(OPT_SPEED) |
333 | # define mad_f_scale64(hi, lo) \ |
334 | ((mad_fixed_t) ((hi) << (32 - MAD_F_SCALEBITS))) |
335 | # define MAD_F_SCALEBITS MAD_F_FRACBITS |
336 | # endif |
337 | |
338 | /* --- SPARC --------------------------------------------------------------- */ |
339 | |
340 | # elif defined(FPM_SPARC) |
341 | |
342 | /* |
343 | * This SPARC V8 version is fast and accurate; the disposition of the least |
344 | * significant bit depends on OPT_ACCURACY via mad_f_scale64(). |
345 | */ |
346 | # define MAD_F_MLX(hi, lo, x, y) \ |
347 | asm ("smul %2, %3, %0\n\t" \ |
348 | "rd %%y, %1" \ |
349 | : "=r" (lo), "=r" (hi) \ |
350 | : "%r" (x), "rI" (y)) |
351 | |
352 | /* --- PowerPC ------------------------------------------------------------- */ |
353 | |
354 | # elif defined(FPM_PPC) |
355 | |
356 | /* |
357 | * This PowerPC version is fast and accurate; the disposition of the least |
358 | * significant bit depends on OPT_ACCURACY via mad_f_scale64(). |
359 | */ |
360 | # define MAD_F_MLX(hi, lo, x, y) \ |
361 | do { \ |
362 | asm ("mullw %0,%1,%2" \ |
363 | : "=r" (lo) \ |
364 | : "%r" (x), "r" (y)); \ |
365 | asm ("mulhw %0,%1,%2" \ |
366 | : "=r" (hi) \ |
367 | : "%r" (x), "r" (y)); \ |
368 | } \ |
369 | while (0) |
370 | |
371 | # if defined(OPT_ACCURACY) |
372 | /* |
373 | * This gives best accuracy but is not very fast. |
374 | */ |
375 | # define MAD_F_MLA(hi, lo, x, y) \ |
376 | ({ mad_fixed64hi_t __hi; \ |
377 | mad_fixed64lo_t __lo; \ |
378 | MAD_F_MLX(__hi, __lo, (x), (y)); \ |
379 | asm ("addc %0,%2,%3\n\t" \ |
380 | "adde %1,%4,%5" \ |
381 | : "=r" (lo), "=r" (hi) \ |
382 | : "%r" (lo), "r" (__lo), \ |
383 | "%r" (hi), "r" (__hi) \ |
384 | : "xer"); \ |
385 | }) |
386 | # endif |
387 | |
388 | # if defined(OPT_ACCURACY) |
389 | /* |
390 | * This is slower than the truncating version below it. |
391 | */ |
392 | # define mad_f_scale64(hi, lo) \ |
393 | ({ mad_fixed_t __result, __round; \ |
394 | asm ("rotrwi %0,%1,%2" \ |
395 | : "=r" (__result) \ |
396 | : "r" (lo), "i" (MAD_F_SCALEBITS)); \ |
397 | asm ("extrwi %0,%1,1,0" \ |
398 | : "=r" (__round) \ |
399 | : "r" (__result)); \ |
400 | asm ("insrwi %0,%1,%2,0" \ |
401 | : "+r" (__result) \ |
402 | : "r" (hi), "i" (MAD_F_SCALEBITS)); \ |
403 | asm ("add %0,%1,%2" \ |
404 | : "=r" (__result) \ |
405 | : "%r" (__result), "r" (__round)); \ |
406 | __result; \ |
407 | }) |
408 | # else |
409 | # define mad_f_scale64(hi, lo) \ |
410 | ({ mad_fixed_t __result; \ |
411 | asm ("rotrwi %0,%1,%2" \ |
412 | : "=r" (__result) \ |
413 | : "r" (lo), "i" (MAD_F_SCALEBITS)); \ |
414 | asm ("insrwi %0,%1,%2,0" \ |
415 | : "+r" (__result) \ |
416 | : "r" (hi), "i" (MAD_F_SCALEBITS)); \ |
417 | __result; \ |
418 | }) |
419 | # endif |
420 | |
421 | # define MAD_F_SCALEBITS MAD_F_FRACBITS |
422 | |
423 | /* --- Default ------------------------------------------------------------- */ |
424 | |
425 | # elif defined(FPM_DEFAULT) |
426 | |
427 | /* |
428 | * This version is the most portable but it loses significant accuracy. |
429 | * Furthermore, accuracy is biased against the second argument, so care |
430 | * should be taken when ordering operands. |
431 | * |
432 | * The scale factors are constant as this is not used with SSO. |
433 | * |
434 | * Pre-rounding is required to stay within the limits of compliance. |
435 | */ |
436 | # if defined(OPT_SPEED) |
437 | # define mad_f_mul(x, y) (((x) >> 12) * ((y) >> 16)) |
438 | # else |
439 | # define mad_f_mul(x, y) ((((x) + (1L << 11)) >> 12) * \ |
440 | (((y) + (1L << 15)) >> 16)) |
441 | # endif |
442 | |
443 | /* ------------------------------------------------------------------------- */ |
444 | |
445 | # else |
446 | # error "no FPM selected" |
447 | # endif |
448 | |
449 | /* default implementations */ |
450 | |
451 | # if !defined(mad_f_mul) |
452 | # define mad_f_mul(x, y) \ |
453 | ({ register mad_fixed64hi_t __hi; \ |
454 | register mad_fixed64lo_t __lo; \ |
455 | MAD_F_MLX(__hi, __lo, (x), (y)); \ |
456 | mad_f_scale64(__hi, __lo); \ |
457 | }) |
458 | # endif |
459 | |
460 | # if !defined(MAD_F_MLA) |
461 | # define MAD_F_ML0(hi, lo, x, y) ((lo) = mad_f_mul((x), (y))) |
462 | # define MAD_F_MLA(hi, lo, x, y) ((lo) += mad_f_mul((x), (y))) |
463 | # define MAD_F_MLN(hi, lo) ((lo) = -(lo)) |
464 | # define MAD_F_MLZ(hi, lo) ((void) (hi), (mad_fixed_t) (lo)) |
465 | # endif |
466 | |
467 | # if !defined(MAD_F_ML0) |
468 | # define MAD_F_ML0(hi, lo, x, y) MAD_F_MLX((hi), (lo), (x), (y)) |
469 | # endif |
470 | |
471 | # if !defined(MAD_F_MLN) |
472 | # define MAD_F_MLN(hi, lo) ((hi) = ((lo) = -(lo)) ? ~(hi) : -(hi)) |
473 | # endif |
474 | |
475 | # if !defined(MAD_F_MLZ) |
476 | # define MAD_F_MLZ(hi, lo) mad_f_scale64((hi), (lo)) |
477 | # endif |
478 | |
479 | # if !defined(mad_f_scale64) |
480 | # if defined(OPT_ACCURACY) |
481 | # define mad_f_scale64(hi, lo) \ |
482 | ((((mad_fixed_t) \ |
483 | (((hi) << (32 - (MAD_F_SCALEBITS - 1))) | \ |
484 | ((lo) >> (MAD_F_SCALEBITS - 1)))) + 1) >> 1) |
485 | # else |
486 | # define mad_f_scale64(hi, lo) \ |
487 | ((mad_fixed_t) \ |
488 | (((hi) << (32 - MAD_F_SCALEBITS)) | \ |
489 | ((lo) >> MAD_F_SCALEBITS))) |
490 | # endif |
491 | # define MAD_F_SCALEBITS MAD_F_FRACBITS |
492 | # endif |
493 | |
494 | /* C routines */ |
495 | |
496 | mad_fixed_t mad_f_abs(mad_fixed_t); |
497 | mad_fixed_t mad_f_div(mad_fixed_t, mad_fixed_t); |
498 | |
499 | # endif |
500 |