1 /********************************************************************
2 * *
3 * THIS FILE IS PART OF THE OggVorbis 'TREMOR' CODEC SOURCE CODE. *
4 * *
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. *
8 * *
9 * THE OggVorbis 'TREMOR' SOURCE CODE IS (C) COPYRIGHT 1994-2002 *
10 * BY THE Xiph.Org FOUNDATION http://www.xiph.org/ *
11 * *
12 ********************************************************************
13
14 function: miscellaneous math and prototypes
15
16 ********************************************************************/
17
18 #ifndef _V_RANDOM_H_
19 #define _V_RANDOM_H_
20 #include "ivorbiscodec.h"
21 #include "os.h"
22
23 #ifdef _LOW_ACCURACY_
24 # define X(n) (((((n)>>22)+1)>>1) - ((((n)>>22)+1)>>9))
25 # define LOOKUP_T const unsigned char
26 #else
27 # define X(n) (n)
28 # define LOOKUP_T const ogg_int32_t
29 #endif
30
31 #include "asm_arm.h"
32 #include <stdlib.h> /* for abs() */
33
34 #ifndef _V_WIDE_MATH
35 #define _V_WIDE_MATH
36
37 #ifndef _LOW_ACCURACY_
38 /* 64 bit multiply */
39
40 #if !(defined WIN32 && defined WINCE)
41 #include <sys/types.h>
42 #endif
43
44 #if BYTE_ORDER==LITTLE_ENDIAN
45 union magic {
46 struct {
47 ogg_int32_t lo;
48 ogg_int32_t hi;
49 } halves;
50 ogg_int64_t whole;
51 };
52 #elif BYTE_ORDER==BIG_ENDIAN
53 union magic {
54 struct {
55 ogg_int32_t hi;
56 ogg_int32_t lo;
57 } halves;
58 ogg_int64_t whole;
59 };
60 #endif
61
MULT32(ogg_int32_t x,ogg_int32_t y)62 STIN ogg_int32_t MULT32(ogg_int32_t x, ogg_int32_t y) {
63 union magic magic;
64 magic.whole = (ogg_int64_t)x * y;
65 return magic.halves.hi;
66 }
67
MULT31(ogg_int32_t x,ogg_int32_t y)68 STIN ogg_int32_t MULT31(ogg_int32_t x, ogg_int32_t y) {
69 return MULT32(x,y)<<1;
70 }
71
MULT31_SHIFT15(ogg_int32_t x,ogg_int32_t y)72 STIN ogg_int32_t MULT31_SHIFT15(ogg_int32_t x, ogg_int32_t y) {
73 union magic magic;
74 magic.whole = (ogg_int64_t)x * y;
75 return ((ogg_uint32_t)(magic.halves.lo)>>15) | ((magic.halves.hi)<<17);
76 }
77
78 #else
79 /* 32 bit multiply, more portable but less accurate */
80
81 /*
82 * Note: Precision is biased towards the first argument therefore ordering
83 * is important. Shift values were chosen for the best sound quality after
84 * many listening tests.
85 */
86
87 /*
88 * For MULT32 and MULT31: The second argument is always a lookup table
89 * value already preshifted from 31 to 8 bits. We therefore take the
90 * opportunity to save on text space and use unsigned char for those
91 * tables in this case.
92 */
93
MULT32(ogg_int32_t x,ogg_int32_t y)94 STIN ogg_int32_t MULT32(ogg_int32_t x, ogg_int32_t y) {
95 return (x >> 9) * y; /* y preshifted >>23 */
96 }
97
MULT31(ogg_int32_t x,ogg_int32_t y)98 STIN ogg_int32_t MULT31(ogg_int32_t x, ogg_int32_t y) {
99 return (x >> 8) * y; /* y preshifted >>23 */
100 }
101
MULT31_SHIFT15(ogg_int32_t x,ogg_int32_t y)102 STIN ogg_int32_t MULT31_SHIFT15(ogg_int32_t x, ogg_int32_t y) {
103 return (x >> 6) * y; /* y preshifted >>9 */
104 }
105
106 #endif
107
108 /*
109 * This should be used as a memory barrier, forcing all cached values in
110 * registers to wr writen back to memory. Might or might not be beneficial
111 * depending on the architecture and compiler.
112 */
113 #define MB()
114
115 /*
116 * The XPROD functions are meant to optimize the cross products found all
117 * over the place in mdct.c by forcing memory operation ordering to avoid
118 * unnecessary register reloads as soon as memory is being written to.
119 * However this is only beneficial on CPUs with a sane number of general
120 * purpose registers which exclude the Intel x86. On Intel, better let the
121 * compiler actually reload registers directly from original memory by using
122 * macros.
123 */
124
125 #ifdef __i386__
126
127 #define XPROD32(_a, _b, _t, _v, _x, _y) \
128 { *(_x)=MULT32(_a,_t)+MULT32(_b,_v); \
129 *(_y)=MULT32(_b,_t)-MULT32(_a,_v); }
130 #define XPROD31(_a, _b, _t, _v, _x, _y) \
131 { *(_x)=MULT31(_a,_t)+MULT31(_b,_v); \
132 *(_y)=MULT31(_b,_t)-MULT31(_a,_v); }
133 #define XNPROD31(_a, _b, _t, _v, _x, _y) \
134 { *(_x)=MULT31(_a,_t)-MULT31(_b,_v); \
135 *(_y)=MULT31(_b,_t)+MULT31(_a,_v); }
136
137 #else
138
XPROD32(ogg_int32_t a,ogg_int32_t b,ogg_int32_t t,ogg_int32_t v,ogg_int32_t * x,ogg_int32_t * y)139 STIN void XPROD32(ogg_int32_t a, ogg_int32_t b,
140 ogg_int32_t t, ogg_int32_t v,
141 ogg_int32_t *x, ogg_int32_t *y)
142 {
143 *x = MULT32(a, t) + MULT32(b, v);
144 *y = MULT32(b, t) - MULT32(a, v);
145 }
146
XPROD31(ogg_int32_t a,ogg_int32_t b,ogg_int32_t t,ogg_int32_t v,ogg_int32_t * x,ogg_int32_t * y)147 STIN void XPROD31(ogg_int32_t a, ogg_int32_t b,
148 ogg_int32_t t, ogg_int32_t v,
149 ogg_int32_t *x, ogg_int32_t *y)
150 {
151 *x = MULT31(a, t) + MULT31(b, v);
152 *y = MULT31(b, t) - MULT31(a, v);
153 }
154
XNPROD31(ogg_int32_t a,ogg_int32_t b,ogg_int32_t t,ogg_int32_t v,ogg_int32_t * x,ogg_int32_t * y)155 STIN void XNPROD31(ogg_int32_t a, ogg_int32_t b,
156 ogg_int32_t t, ogg_int32_t v,
157 ogg_int32_t *x, ogg_int32_t *y)
158 {
159 *x = MULT31(a, t) - MULT31(b, v);
160 *y = MULT31(b, t) + MULT31(a, v);
161 }
162
163 #endif
164
165 #endif
166
167 #ifndef _V_CLIP_MATH
168 #define _V_CLIP_MATH
169
CLIP_TO_15(ogg_int32_t x)170 STIN ogg_int32_t CLIP_TO_15(ogg_int32_t x) {
171 int ret=x;
172 ret-= ((x<=32767)-1)&(x-32767);
173 ret-= ((x>=-32768)-1)&(x+32768);
174 return(ret);
175 }
176
177 #endif
178
VFLOAT_MULT(ogg_int32_t a,ogg_int32_t ap,ogg_int32_t b,ogg_int32_t bp,ogg_int32_t * p)179 STIN ogg_int32_t VFLOAT_MULT(ogg_int32_t a,ogg_int32_t ap,
180 ogg_int32_t b,ogg_int32_t bp,
181 ogg_int32_t *p){
182 if(a && b){
183 #ifndef _LOW_ACCURACY_
184 *p=ap+bp+32;
185 return MULT32(a,b);
186 #else
187 *p=ap+bp+31;
188 return (a>>15)*(b>>16);
189 #endif
190 }else
191 return 0;
192 }
193
194 int _ilog(unsigned int);
195
VFLOAT_MULTI(ogg_int32_t a,ogg_int32_t ap,ogg_int32_t i,ogg_int32_t * p)196 STIN ogg_int32_t VFLOAT_MULTI(ogg_int32_t a,ogg_int32_t ap,
197 ogg_int32_t i,
198 ogg_int32_t *p){
199
200 int ip=_ilog(abs(i))-31;
201 return VFLOAT_MULT(a,ap,i<<-ip,ip,p);
202 }
203
VFLOAT_ADD(ogg_int32_t a,ogg_int32_t ap,ogg_int32_t b,ogg_int32_t bp,ogg_int32_t * p)204 STIN ogg_int32_t VFLOAT_ADD(ogg_int32_t a,ogg_int32_t ap,
205 ogg_int32_t b,ogg_int32_t bp,
206 ogg_int32_t *p){
207
208 if(!a){
209 *p=bp;
210 return b;
211 }else if(!b){
212 *p=ap;
213 return a;
214 }
215
216 /* yes, this can leak a bit. */
217 if(ap>bp){
218 int shift=ap-bp+1;
219 *p=ap+1;
220 a>>=1;
221 if(shift<32){
222 b=(b+(1<<(shift-1)))>>shift;
223 }else{
224 b=0;
225 }
226 }else{
227 int shift=bp-ap+1;
228 *p=bp+1;
229 b>>=1;
230 if(shift<32){
231 a=(a+(1<<(shift-1)))>>shift;
232 }else{
233 a=0;
234 }
235 }
236
237 a+=b;
238 if((a&0xc0000000)==0xc0000000 ||
239 (a&0xc0000000)==0){
240 a<<=1;
241 (*p)--;
242 }
243 return(a);
244 }
245
246 #endif
247
248
249
250
251