Diffstat (limited to 'core/multimedia/opieplayer/libmad/fixed.h') (more/less context) (ignore whitespace changes)
| -rw-r--r-- | core/multimedia/opieplayer/libmad/fixed.h | 111 |
1 files changed, 67 insertions, 44 deletions
diff --git a/core/multimedia/opieplayer/libmad/fixed.h b/core/multimedia/opieplayer/libmad/fixed.h index c9b98ca..baa7dc5 100644 --- a/core/multimedia/opieplayer/libmad/fixed.h +++ b/core/multimedia/opieplayer/libmad/fixed.h @@ -2,3 +2,3 @@ * libmad - MPEG audio decoder library - * Copyright (C) 2000-2001 Robert Leslie + * Copyright (C) 2000-2004 Underbit Technologies, Inc. * @@ -210,2 +210,17 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) }) +# elif defined(OPT_INTEL) +/* + * Alternate Intel scaling that may or may not perform better. + */ +# define mad_f_scale64(hi, lo) \ + ({ mad_fixed_t __result; \ + asm ("shrl %3,%1\n\t" \ + "shll %4,%2\n\t" \ + "orl %2,%1" \ + : "=rm" (__result) \ + : "0" (lo), "r" (hi), \ + "I" (MAD_F_SCALEBITS), "I" (32 - MAD_F_SCALEBITS) \ + : "cc"); \ + __result; \ + }) # else @@ -234,8 +249,4 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) /* - * There's a bug somewhere, possibly in the compiler, that sometimes makes - * this necessary instead of the default implementation via MAD_F_MLX and - * mad_f_scale64. It may be related to the use (or lack) of - * -finline-functions and/or -fstrength-reduce. - * - * This is also apparently faster than MAD_F_MLX/mad_f_scale64. + * This is faster than the default implementation via MAD_F_MLX() and + * mad_f_scale64(). */ @@ -277,3 +288,3 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) "adc %0, %0, %2, lsl %4" \ - : "=r" (__result) \ + : "=&r" (__result) \ : "r" (lo), "r" (hi), \ @@ -345,14 +356,21 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) /* - * This PowerPC version is tuned for the 4xx embedded processors. It is - * effectively a tuned version of FPM_64BIT. It is a little faster and just - * as accurate. The disposition of the least significant bit depends on - * OPT_ACCURACY via mad_f_scale64(). + * 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) \ - asm ("mulhw %1, %2, %3\n\t" \ - "mullw %0, %2, %3" \ - : "=&r" (lo), "=&r" (hi) \ - : "%r" (x), "r" (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) -# define MAD_F_MLA(hi, lo, x, y) \ +# 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; \ @@ -360,7 +378,10 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) MAD_F_MLX(__hi, __lo, (x), (y)); \ - asm ("addc %0, %2, %3\n\t" \ - "adde %1, %4, %5" \ + asm ("addc %0,%2,%3\n\t" \ + "adde %1,%4,%5" \ : "=r" (lo), "=r" (hi) \ - : "%r" (__lo), "0" (lo), "%r" (__hi), "1" (hi)); \ + : "%r" (lo), "r" (__lo), \ + "%r" (hi), "r" (__hi) \ + : "xer"); \ }) +# endif @@ -368,20 +389,19 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) /* - * This is accurate and ~2 - 2.5 times slower than the unrounded version. - * - * The __volatile__ improves the generated code by another 5% (fewer spills - * to memory); eventually they should be removed. + * This is slower than the truncating version below it. */ # define mad_f_scale64(hi, lo) \ - ({ mad_fixed_t __result; \ - mad_fixed64hi_t __hi_; \ - mad_fixed64lo_t __lo_; \ - asm __volatile__ ("addc %0, %2, %4\n\t" \ - "addze %1, %3" \ - : "=r" (__lo_), "=r" (__hi_) \ - : "r" (lo), "r" (hi), "r" (1 << (MAD_F_SCALEBITS - 1))); \ - asm __volatile__ ("rlwinm %0, %2,32-%3,0,%3-1\n\t" \ - "rlwimi %0, %1,32-%3,%3,31" \ - : "=&r" (__result) \ - : "r" (__lo_), "r" (__hi_), "I" (MAD_F_SCALEBITS)); \ - __result; \ + ({ mad_fixed_t __result, __round; \ + asm ("rotrwi %0,%1,%2" \ + : "=r" (__result) \ + : "r" (lo), "i" (MAD_F_SCALEBITS)); \ + asm ("extrwi %0,%1,1,0" \ + : "=r" (__round) \ + : "r" (__result)); \ + asm ("insrwi %0,%1,%2,0" \ + : "+r" (__result) \ + : "r" (hi), "i" (MAD_F_SCALEBITS)); \ + asm ("add %0,%1,%2" \ + : "=r" (__result) \ + : "%r" (__result), "r" (__round)); \ + __result; \ }) @@ -390,9 +410,11 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) ({ mad_fixed_t __result; \ - asm ("rlwinm %0, %2,32-%3,0,%3-1\n\t" \ - "rlwimi %0, %1,32-%3,%3,31" \ + asm ("rotrwi %0,%1,%2" \ : "=r" (__result) \ - : "r" (lo), "r" (hi), "I" (MAD_F_SCALEBITS)); \ - __result; \ + : "r" (lo), "i" (MAD_F_SCALEBITS)); \ + asm ("insrwi %0,%1,%2,0" \ + : "+r" (__result) \ + : "r" (hi), "i" (MAD_F_SCALEBITS)); \ + __result; \ }) -# endif /* OPT_ACCURACY */ +# endif @@ -430,4 +452,4 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) # define mad_f_mul(x, y) \ - ({ mad_fixed64hi_t __hi; \ - mad_fixed64lo_t __lo; \ + ({ register mad_fixed64hi_t __hi; \ + register mad_fixed64lo_t __lo; \ MAD_F_MLX(__hi, __lo, (x), (y)); \ @@ -471,5 +493,6 @@ mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y) -/* miscellaneous C routines */ +/* C routines */ mad_fixed_t mad_f_abs(mad_fixed_t); +mad_fixed_t mad_f_div(mad_fixed_t, mad_fixed_t); |