1 /********************************************************************
3 * THIS FILE IS PART OF THE OggVorbis 'TREMOR' CODEC SOURCE CODE. *
5 * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS *
6 * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
7 * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. *
9 * THE OggVorbis 'TREMOR' SOURCE CODE IS (C) COPYRIGHT 1994-2002 *
10 * BY THE Xiph.Org FOUNDATION http://www.xiph.org/ *
12 ********************************************************************
14 function: miscellaneous math and prototypes
16 ********************************************************************/
18 #include "config-tremor.h"
22 #include "ivorbiscodec.h"
26 #include "asm_mcf5249.h"
29 /* Some prototypes that were not defined elsewhere */
30 void *_vorbis_block_alloc(vorbis_block
*vb
,long bytes
);
31 void _vorbis_block_ripcord(vorbis_block
*vb
);
32 extern int _ilog(unsigned int v
);
37 #ifndef _LOW_ACCURACY_
39 /* #include <sys/types.h> */
41 #if BYTE_ORDER==LITTLE_ENDIAN
49 #elif BYTE_ORDER==BIG_ENDIAN
59 static inline ogg_int32_t
MULT32(ogg_int32_t x
, ogg_int32_t y
) {
61 magic
.whole
= (ogg_int64_t
)x
* y
;
62 return magic
.halves
.hi
;
64 static inline ogg_int32_t
MULT31(ogg_int32_t x
, ogg_int32_t y
) {
65 return MULT32(x
,y
)<<1;
68 static inline ogg_int32_t
MULT31_SHIFT15(ogg_int32_t x
, ogg_int32_t y
) {
70 magic
.whole
= (ogg_int64_t
)x
* y
;
71 return ((ogg_uint32_t
)(magic
.halves
.lo
)>>15) | ((magic
.halves
.hi
)<<17);
75 /* 32 bit multiply, more portable but less accurate */
78 * Note: Precision is biased towards the first argument therefore ordering
79 * is important. Shift values were chosen for the best sound quality after
80 * many listening tests.
84 * For MULT32 and MULT31: The second argument is always a lookup table
85 * value already preshifted from 31 to 8 bits. We therefore take the
86 * opportunity to save on text space and use unsigned char for those
87 * tables in this case.
90 static inline ogg_int32_t
MULT32(ogg_int32_t x
, ogg_int32_t y
) {
91 return (x
>> 9) * y
; /* y preshifted >>23 */
94 static inline ogg_int32_t
MULT31(ogg_int32_t x
, ogg_int32_t y
) {
95 return (x
>> 8) * y
; /* y preshifted >>23 */
98 static inline ogg_int32_t
MULT31_SHIFT15(ogg_int32_t x
, ogg_int32_t y
) {
99 return (x
>> 6) * y
; /* y preshifted >>9 */
104 * This should be used as a memory barrier, forcing all cached values in
105 * registers to wr writen back to memory. Might or might not be beneficial
106 * depending on the architecture and compiler.
111 * The XPROD functions are meant to optimize the cross products found all
112 * over the place in mdct.c by forcing memory operation ordering to avoid
113 * unnecessary register reloads as soon as memory is being written to.
114 * However this is only beneficial on CPUs with a sane number of general
115 * purpose registers which exclude the Intel x86. On Intel, better let the
116 * compiler actually reload registers directly from original memory by using
120 /* replaced XPROD32 with a macro to avoid memory reference
121 _x, _y are the results (must be l-values) */
122 #define XPROD32(_a, _b, _t, _v, _x, _y) \
123 { (_x)=MULT32(_a,_t)+MULT32(_b,_v); \
124 (_y)=MULT32(_b,_t)-MULT32(_a,_v); }
129 #define XPROD31(_a, _b, _t, _v, _x, _y) \
130 { *(_x)=MULT31(_a,_t)+MULT31(_b,_v); \
131 *(_y)=MULT31(_b,_t)-MULT31(_a,_v); }
132 #define XNPROD31(_a, _b, _t, _v, _x, _y) \
133 { *(_x)=MULT31(_a,_t)-MULT31(_b,_v); \
134 *(_y)=MULT31(_b,_t)+MULT31(_a,_v); }
138 static inline void XPROD31(ogg_int32_t a
, ogg_int32_t b
,
139 ogg_int32_t t
, ogg_int32_t v
,
140 ogg_int32_t
*x
, ogg_int32_t
*y
)
142 *x
= MULT31(a
, t
) + MULT31(b
, v
);
143 *y
= MULT31(b
, t
) - MULT31(a
, v
);
146 static inline void XNPROD31(ogg_int32_t a
, ogg_int32_t b
,
147 ogg_int32_t t
, ogg_int32_t v
,
148 ogg_int32_t
*x
, ogg_int32_t
*y
)
150 *x
= MULT31(a
, t
) - MULT31(b
, v
);
151 *y
= MULT31(b
, t
) + MULT31(a
, v
);
158 /* generic misc.h has symmetrical versions of vect_add_right_left
159 and vect_add_left_right (since symmetrical versions of
160 vect_mult_fw and vect_mult_bw i.e. both use MULT31) */
162 void vect_add_right_left(ogg_int32_t
*x
, const ogg_int32_t
*y
, int n
)
171 void vect_add_left_right(ogg_int32_t
*x
, const ogg_int32_t
*y
, int n
)
173 vect_add_right_left(x
,y
,n
);
177 void vect_mult_fw(ogg_int32_t
*data
, LOOKUP_T
*window
, int n
)
180 *data
= MULT31(*data
, *window
);
188 void vect_mult_bw(ogg_int32_t
*data
, LOOKUP_T
*window
, int n
)
191 *data
= MULT31(*data
, *window
);
198 /* generic memcpy is probably optimal */
199 static inline void vect_copy(ogg_int32_t
*x
, const ogg_int32_t
*y
, int n
)
201 memcpy(x
,y
,n
*sizeof(ogg_int32_t
));
210 static inline ogg_int32_t
CLIP_TO_15(ogg_int32_t x
) {
212 ret
-= ((x
<=32767)-1)&(x
-32767);
213 ret
-= ((x
>=-32768)-1)&(x
+32768);
219 static inline ogg_int32_t
VFLOAT_MULT(ogg_int32_t a
,ogg_int32_t ap
,
220 ogg_int32_t b
,ogg_int32_t bp
,
223 #ifndef _LOW_ACCURACY_
228 return (a
>>15)*(b
>>16);
234 static inline ogg_int32_t
VFLOAT_MULTI(ogg_int32_t a
,ogg_int32_t ap
,
238 int ip
=_ilog(abs(i
))-31;
239 return VFLOAT_MULT(a
,ap
,i
<<-ip
,ip
,p
);
242 static inline ogg_int32_t
VFLOAT_ADD(ogg_int32_t a
,ogg_int32_t ap
,
243 ogg_int32_t b
,ogg_int32_t bp
,
254 /* yes, this can leak a bit. */
260 b
=(b
+(1<<(shift
-1)))>>shift
;
269 a
=(a
+(1<<(shift
-1)))>>shift
;
276 if((a
&0xc0000000)==0xc0000000 ||