/*
* libmad - MPEG audio decoder library
* Copyright (C) 2000-2004 Underbit Technologies, Inc.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* $Id: fixed.h,v 1.38 2004/02/17 02:02:03 rob Exp $
*/
# ifndef LIBMAD_FIXED_H
# define LIBMAD_FIXED_H
typedef int mad_fixed_t; // must be 32 bits
typedef int mad_fixed64hi_t; // must be 32 bits
typedef u32int mad_fixed64lo_t; // must be 32 bits
typedef vlong mad_fixed64_t;
/*
* Fixed-point format: 0xABBBBBBB
* A == whole part (sign + 3 bits)
* B == fractional part (28 bits)
*
* Values are signed two's complement, so the effective range is:
* 0x80000000 to 0x7fffffff
* -8.0 to +7.9999999962747097015380859375
*
* The smallest representable value is:
* 0x00000001 == 0.0000000037252902984619140625 (i.e. about 3.725e-9)
*
* 28 bits of fractional accuracy represent about
* 8.6 digits of decimal accuracy.
*
* Fixed-point numbers can be added or subtracted as normal
* integers, but multiplication requires shifting the 64-bit result
* from 56 fractional bits back to 28 (and rounding.)
*
* Changing the definition of MAD_F_FRACBITS is only partially
* supported, and must be done with care.
*/
/*
* This version should be the most accurate if 64-bit types are supported by
* the compiler, although it may not be the most efficient.
*/
# if defined(OPT_ACCURACY)
# define mad_f_mul(x, y) \
((mad_fixed_t) \
((((mad_fixed64_t) (x) * (y)) + \
(1L << (MAD_F_SCALEBITS - 1))) >> MAD_F_SCALEBITS))
# else
# define mad_f_mul(x, y) \
((mad_fixed_t) (((mad_fixed64_t) (x) * (y)) >> MAD_F_SCALEBITS))
# endif
# if defined(_MSC_VER)
# pragma warning(push)
# pragma warning(disable: 4035) /* no return value */
static __forceinline
mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y)
{
enum {
fracbits = MAD_F_FRACBITS
};
__asm {
mov eax, x
imul y
shrd eax, edx, fracbits
}
/* implicit return of eax */
}
# pragma warning(pop)
# define mad_f_mul mad_f_mul_inline
# define mad_f_scale64
# else
/*
* This Intel version is fast and accurate; the disposition of the least
* significant bit depends on OPT_ACCURACY via mad_f_scale64().
*/
# define MAD_F_MLX(hi, lo, x, y) \
asm ("imull %3" \
: "=a" (lo), "=d" (hi) \
: "%a" (x), "rm" (y) \
: "cc")
# if defined(OPT_ACCURACY)
/*
* This gives best accuracy but is not very fast.
*/
# define MAD_F_MLA(hi, lo, x, y) \
({ mad_fixed64hi_t __hi; \
mad_fixed64lo_t __lo; \
MAD_F_MLX(__hi, __lo, (x), (y)); \
asm ("addl %2,%0\n\t" \
"adcl %3,%1" \
: "=rm" (lo), "=rm" (hi) \
: "r" (__lo), "r" (__hi), "0" (lo), "1" (hi) \
: "cc"); \
})
# endif /* OPT_ACCURACY */
/* --- ARM ----------------------------------------------------------------- */
# elif defined(FPM_ARM)
/*
* This ARM V4 version is as accurate as FPM_64BIT but much faster. The
* least significant bit is properly rounded at no CPU cycle cost!
*/
# if 1
/*
* This is faster than the default implementation via MAD_F_MLX() and
* mad_f_scale64().
*/
# define mad_f_mul(x, y) \
({ mad_fixed64hi_t __hi; \
mad_fixed64lo_t __lo; \
mad_fixed_t __result; \
asm ("smull %0, %1, %3, %4\n\t" \
"movs %0, %0, lsr %5\n\t" \
"adc %2, %0, %1, lsl %6" \
: "=&r" (__lo), "=&r" (__hi), "=r" (__result) \
: "%r" (x), "r" (y), \
"M" (MAD_F_SCALEBITS), "M" (32 - MAD_F_SCALEBITS) \
: "cc"); \
__result; \
})
# endif
/*
* This MIPS version is fast and accurate; the disposition of the least
* significant bit depends on OPT_ACCURACY via mad_f_scale64().
*/
# define MAD_F_MLX(hi, lo, x, y) \
asm ("mult %2,%3" \
: "=l" (lo), "=h" (hi) \
: "%r" (x), "r" (y))
/*
* This SPARC V8 version is fast and accurate; the disposition of the least
* significant bit depends on OPT_ACCURACY via mad_f_scale64().
*/
# define MAD_F_MLX(hi, lo, x, y) \
asm ("smul %2, %3, %0\n\t" \
"rd %%y, %1" \
: "=r" (lo), "=r" (hi) \
: "%r" (x), "rI" (y))
/*
* This PowerPC version is fast and accurate; the disposition of the least
* significant bit depends on OPT_ACCURACY via mad_f_scale64().
*/
# define MAD_F_MLX(hi, lo, x, y) \
do { \
asm ("mullw %0,%1,%2" \
: "=r" (lo) \
: "%r" (x), "r" (y)); \
asm ("mulhw %0,%1,%2" \
: "=r" (hi) \
: "%r" (x), "r" (y)); \
} \
while (0)
# if defined(OPT_ACCURACY)
/*
* This gives best accuracy but is not very fast.
*/
# define MAD_F_MLA(hi, lo, x, y) \
({ mad_fixed64hi_t __hi; \
mad_fixed64lo_t __lo; \
MAD_F_MLX(__hi, __lo, (x), (y)); \
asm ("addc %0,%2,%3\n\t" \
"adde %1,%4,%5" \
: "=r" (lo), "=r" (hi) \
: "%r" (lo), "r" (__lo), \
"%r" (hi), "r" (__hi) \
: "xer"); \
})
# endif
/*
* This version is the most portable but it loses significant accuracy.
* Furthermore, accuracy is biased against the second argument, so care
* should be taken when ordering operands.
*
* The scale factors are constant as this is not used with SSO.
*
* Pre-rounding is required to stay within the limits of compliance.
*/
# if defined(OPT_SPEED)
# define mad_f_mul(x, y) (((x) >> 12) * ((y) >> 16))
# else
# define mad_f_mul(x, y) ((((x) + (1L << 11)) >> 12) * \
(((y) + (1L << 15)) >> 16))
# endif
