|
562
|
1 #ifndef MPG123_H_OPTIMIZE
|
|
|
2 #define MPG123_H_OPTIMIZE
|
|
|
3 /*
|
|
|
4 optimize: get a grip on the different optimizations
|
|
|
5
|
|
|
6 copyright 2007 by the mpg123 project - free software under the terms of the LGPL 2.1
|
|
|
7 see COPYING and AUTHORS files in distribution or http://mpg123.org
|
|
|
8 initially written by Thomas Orgis, taking from mpg123.[hc]
|
|
|
9
|
|
|
10 for building mpg123 with one optimization only, you have to choose exclusively between
|
|
|
11 OPT_GENERIC (generic C code for everyone)
|
|
|
12 OPT_I386 (Intel i386)
|
|
|
13 OPT_I486 (...)
|
|
|
14 OPT_I586 (Intel Pentium)
|
|
|
15 OPT_I586_DITHER (Intel Pentium with dithering/noise shaping for enhanced quality)
|
|
|
16 OPT_MMX (Intel Pentium and compatibles with MMX, fast, but not the best accuracy)
|
|
|
17 OPT_3DNOW (AMD 3DNow!, K6-2/3, Athlon, compatibles...)
|
|
|
18 OPT_ALTIVEC (Motorola/IBM PPC with AltiVec under MacOSX)
|
|
|
19
|
|
|
20 or you define OPT_MULTI and give a combination which makes sense (do not include i486, do not mix altivec and x86).
|
|
|
21
|
|
|
22 I still have to examine the dynamics of this here together with REAL_IS_FIXED.
|
|
|
23 */
|
|
|
24
|
|
|
25 /* this is included in mpg123.h, which includes config.h */
|
|
|
26 #ifdef CCALIGN
|
|
|
27 #define ALIGNED(a) __attribute__((aligned(a)))
|
|
|
28 #else
|
|
|
29 #define ALIGNED(a)
|
|
|
30 #endif
|
|
|
31
|
|
|
32 /* the optimizations only cover the synth1to1 mode and the dct36 function */
|
|
|
33 /* the first two types are needed in set_synth_functions regardless of optimizations */
|
|
|
34 typedef int (*func_synth)(real *,int, mpg123_handle *,int );
|
|
|
35 typedef int (*func_synth_mono)(real *, mpg123_handle *);
|
|
|
36 typedef void (*func_dct36)(real *,real *,real *,real *,real *);
|
|
|
37 typedef void (*func_dct64)(real *,real *,real *);
|
|
|
38 typedef void (*func_make_decode_tables)(mpg123_handle*);
|
|
|
39 typedef real (*func_init_layer3_gainpow2)(mpg123_handle*, int);
|
|
|
40 typedef real* (*func_init_layer2_table)(mpg123_handle*, real*, double);
|
|
|
41 typedef int (*func_synth_pent)(real *,int,unsigned char *);
|
|
|
42
|
|
|
43 /* last headaches about getting mmx hardcode out */
|
|
|
44 real init_layer3_gainpow2(mpg123_handle *fr, int i);
|
|
|
45 real* init_layer2_table(mpg123_handle *fr, real *table, double m);
|
|
|
46 void make_decode_tables(mpg123_handle *fr);
|
|
|
47 void prepare_decode_tables(void); /* perhaps not best place here */
|
|
|
48
|
|
|
49 /* only 3dnow replaces that one, it's internal to layer3.c otherwise */
|
|
|
50 void dct36(real *,real *,real *,real *,real *);
|
|
|
51 #define opt_dct36(fr) dct36
|
|
|
52 /* only mmx replaces those */
|
|
|
53 #define opt_make_decode_tables(fr) make_decode_tables(fr)
|
|
|
54 #define opt_decwin(fr) (fr)->decwin
|
|
|
55 #define opt_init_layer3_gainpow2(fr) init_layer3_gainpow2
|
|
|
56 #define opt_init_layer2_table(fr) init_layer2_table
|
|
|
57
|
|
|
58 #ifdef OPT_GENERIC
|
|
|
59 #define PENTIUM_FALLBACK
|
|
|
60 void dct64(real *,real *,real *);
|
|
|
61 int synth_1to1(real *bandPtr,int channel, mpg123_handle *fr, int final);
|
|
|
62 int synth_1to1_8bit(real *bandPtr,int channel, mpg123_handle *fr, int final);
|
|
|
63 int synth_1to1_mono(real *, mpg123_handle *fr);
|
|
|
64 int synth_1to1_mono2stereo (real *, mpg123_handle *fr);
|
|
|
65 int synth_1to1_8bit_mono (real *, mpg123_handle *fr);
|
|
|
66 int synth_1to1_8bit_mono2stereo (real *, mpg123_handle *fr);
|
|
|
67 #ifndef OPT_MULTI
|
|
|
68 #define defopt generic
|
|
|
69 #define opt_dct64(fr) dct64
|
|
|
70 #define opt_synth_1to1(fr) synth_1to1
|
|
|
71 #define opt_synth_1to1_mono(fr) synth_1to1_mono
|
|
|
72 #define opt_synth_1to1_mono2stereo(fr) synth_1to1_mono2stereo
|
|
|
73 #define opt_synth_1to1_8bit(fr) synth_1to1_8bit
|
|
|
74 #define opt_synth_1to1_8bit_mono(fr) synth_1to1_8bit_mono
|
|
|
75 #define opt_synth_1to1_8bit_mono2stereo(fr) synth_1to1_8bit_mono2stereo
|
|
|
76 #endif
|
|
|
77 #endif
|
|
|
78
|
|
|
79 /* i486 is special... always alone! */
|
|
|
80 #ifdef OPT_I486
|
|
|
81 #define OPT_X86
|
|
|
82 #define OPT_I386_SYNTH
|
|
|
83 #define defopt ivier
|
|
|
84 int synth_1to1_486(real *bandPtr, int channel, mpg123_handle *fr, int nb_blocks);
|
|
|
85 #ifdef OPT_MULTI
|
|
|
86 #error "i486 can only work alone!"
|
|
|
87 #endif
|
|
|
88 #define opt_synth_1to1(fr) synth_1to1_i386
|
|
|
89 #define FIR_BUFFER_SIZE 128
|
|
|
90 #define FIR_SIZE 16
|
|
|
91 void dct64_i486(int *a,int *b,real *c); /* not used generally */
|
|
|
92 #endif
|
|
|
93
|
|
|
94 #ifdef OPT_I386
|
|
|
95 #define PENTIUM_FALLBACK
|
|
|
96 #define OPT_X86
|
|
|
97 #define OPT_I386_SYNTH
|
|
|
98 #ifndef OPT_MULTI
|
|
|
99 #ifndef defopt
|
|
|
100 #define defopt idrei
|
|
|
101 #endif
|
|
|
102 #define opt_synth_1to1(fr) synth_1to1_i386
|
|
|
103 #endif
|
|
|
104 #endif
|
|
|
105
|
|
|
106 #ifdef OPT_I386_SYNTH
|
|
|
107 int synth_1to1_i386(real *bandPtr, int channel, mpg123_handle *fr, int final);
|
|
|
108 #endif
|
|
|
109
|
|
|
110 #ifdef OPT_I586
|
|
|
111 #define PENTIUM_FALLBACK
|
|
|
112 #define OPT_PENTIUM
|
|
|
113 #define OPT_X86
|
|
|
114 int synth_1to1_i586(real *bandPtr, int channel, mpg123_handle *fr, int final);
|
|
|
115 int synth_1to1_i586_asm(real *bandPtr, int channel, unsigned char *out, unsigned char *buffs, int *bo, real *decwin);
|
|
|
116 #ifndef OPT_MULTI
|
|
|
117 #define defopt ifuenf
|
|
|
118 #define opt_synth_1to1(fr) synth_1to1_i586
|
|
|
119 #define opt_synth_1to1_i586_asm(fr) synth_1to1_i586_asm
|
|
|
120 #endif
|
|
|
121 #endif
|
|
|
122
|
|
|
123 #ifdef OPT_I586_DITHER
|
|
|
124 #define PENTIUM_FALLBACK
|
|
|
125 #define OPT_PENTIUM
|
|
|
126 #define OPT_X86
|
|
|
127 int synth_1to1_i586(real *bandPtr, int channel, mpg123_handle *fr, int final);
|
|
|
128 int synth_1to1_i586_asm_dither(real *bandPtr, int channel, unsigned char *out, unsigned char *buffs, int *bo, real *decwin);
|
|
|
129 #ifndef OPT_MULTI
|
|
|
130 #define defopt ifuenf_dither
|
|
|
131 #define opt_synth_1to1(fr) synth_1to1_i586
|
|
|
132 #define opt_synth_1to1_i586_asm(fr) synth_1to1_i586_asm_dither
|
|
|
133 #endif
|
|
|
134 #endif
|
|
|
135
|
|
|
136 /* That one has by far the most ugly hacks to make it cooperative. */
|
|
|
137 #ifdef OPT_MMX
|
|
|
138 #define OPT_MMXORSSE
|
|
|
139 #define OPT_X86
|
|
|
140 real init_layer3_gainpow2_mmx(mpg123_handle *fr, int i);
|
|
|
141 real* init_layer2_table_mmx(mpg123_handle *fr, real *table, double m);
|
|
|
142 /* I think one can optimize storage here with the normal decwin */
|
|
|
143 extern real decwin_mmx[512+32];
|
|
|
144 void dct64_mmx(real *,real *,real *);
|
|
|
145 int synth_1to1_mmx(real *bandPtr, int channel, mpg123_handle *fr, int final);
|
|
|
146 void make_decode_tables_mmx(mpg123_handle *fr); /* tabinit_mmx.s */
|
|
|
147 void make_decode_tables_mmx_asm(long scaleval, float* decwin_mmx, float *decwins); /* tabinit_mmx.s */
|
|
|
148 /* these are in asm, dct64 called directly there */
|
|
|
149 void dct64_MMX(short *a,short *b,real *c);
|
|
|
150 int synth_1to1_MMX(real *bandPtr, int channel, short *out, short *buffs, int *bo, float *decwins);
|
|
|
151 #ifndef OPT_MULTI
|
|
|
152 #define defopt mmx
|
|
|
153 /* #undef opt_decwin
|
|
|
154 #define opt_decwin(fr) decwin_mmx */
|
|
|
155 #define opt_dct64(fr) dct64_mmx
|
|
|
156 #define opt_synth_1to1(fr) synth_1to1_mmx
|
|
|
157 #define opt_
|
|
|
158 #undef opt_make_decode_tables
|
|
|
159 #define opt_make_decode_tables(fr) make_decode_tables_mmx(fr)
|
|
|
160 #undef opt_init_layer3_gainpow2
|
|
|
161 #define opt_init_layer3_gainpow2(fr) init_layer3_gainpow2_mmx
|
|
|
162 #undef opt_init_layer2_table
|
|
|
163 #define opt_init_layer2_table(fr) init_layer2_table_mmx
|
|
|
164 #define OPT_MMX_ONLY
|
|
|
165 #endif
|
|
|
166 #endif
|
|
|
167
|
|
|
168 /* first crude hack into our source */
|
|
|
169 #ifdef OPT_SSE
|
|
|
170 #define OPT_MMXORSSE
|
|
|
171 #define OPT_MPLAYER
|
|
|
172 #define OPT_X86
|
|
|
173 real init_layer3_gainpow2_mmx(mpg123_handle *fr, int i);
|
|
|
174 real* init_layer2_table_mmx(mpg123_handle *fr, real *table, double m);
|
|
|
175 /* I think one can optimize storage here with the normal decwin */
|
|
|
176 extern real decwin_mmx[512+32];
|
|
|
177 void dct64_mmx(real *,real *,real *);
|
|
|
178 void dct64_sse(real *,real *,real *);
|
|
|
179 int synth_1to1_sse(real *bandPtr, int channel, mpg123_handle *fr, int final);
|
|
|
180 void synth_1to1_sse_asm(real *bandPtr, int channel, short *samples, short *buffs, int *bo, real *decwin);
|
|
|
181 void make_decode_tables_mmx(mpg123_handle *fr); /* tabinit_mmx.s */
|
|
|
182 void make_decode_tables_mmx_asm(long scaleval, float* decwin_mmx, float *decwins); /* tabinit_mmx.s */
|
|
|
183 /* ugly! */
|
|
|
184 extern func_dct64 mpl_dct64;
|
|
|
185 #ifndef OPT_MULTI
|
|
|
186 #define defopt sse
|
|
|
187 #define opt_mpl_dct64(fr) dct64_sse
|
|
|
188 /* #undef opt_decwin
|
|
|
189 #define opt_decwin(fr) decwin_mmx */
|
|
|
190 #define opt_dct64(fr) dct64_mmx /* dct64_sse is silent in downsampling modes */
|
|
|
191 #define opt_synth_1to1(fr) synth_1to1_sse /* that will use dct64_sse */
|
|
|
192 #undef opt_make_decode_tables
|
|
|
193 #define opt_make_decode_tables(fr) make_decode_tables_mmx(fr)
|
|
|
194 #undef opt_init_layer3_gainpow2
|
|
|
195 #define opt_init_layer3_gainpow2(fr) init_layer3_gainpow2_mmx
|
|
|
196 #undef opt_init_layer2_table
|
|
|
197 #define opt_init_layer2_table(fr) init_layer2_table_mmx
|
|
|
198 #define OPT_MMX_ONLY /* watch out! */
|
|
|
199 #endif
|
|
|
200 #endif
|
|
|
201
|
|
|
202 /* first crude hack into our source */
|
|
|
203 #ifdef OPT_3DNOWEXT
|
|
|
204 #define OPT_MMXORSSE
|
|
|
205 #define OPT_MPLAYER
|
|
|
206 #define OPT_X86
|
|
|
207 real init_layer3_gainpow2_mmx(mpg123_handle *fr, int i);
|
|
|
208 real* init_layer2_table_mmx(mpg123_handle *fr, real *table, double m);
|
|
|
209 /* I think one can optimize storage here with the normal decwin */
|
|
|
210 extern real decwin_mmx[512+32];
|
|
|
211 void dct64_mmx(real *,real *,real *);
|
|
|
212 void dct64_3dnowext(real *,real *,real *);
|
|
|
213 void dct36_3dnowext(real *,real *,real *,real *,real *);
|
|
|
214 int synth_1to1_3dnowext(real *bandPtr, int channel, mpg123_handle *fr, int final);
|
|
|
215 void synth_1to1_3dnowext_asm(real *bandPtr, int channel, short *samples, short *buffs, int *bo, real *decwin);
|
|
|
216 void make_decode_tables_mmx(mpg123_handle *fr); /* tabinit_mmx.s */
|
|
|
217 void make_decode_tables_mmx_asm(long scaleval, float* decwin_mmx, float *decwins); /* tabinit_mmx.s */
|
|
|
218 /* ugly! */
|
|
|
219 extern func_dct64 mpl_dct64;
|
|
|
220 #ifndef OPT_MULTI
|
|
|
221 #define defopt dreidnowext
|
|
|
222 #define opt_mpl_dct64(fr) dct64_3dnowext
|
|
|
223 #undef opt_dct36
|
|
|
224 #define opt_dct36(fr) dct36_3dnowext
|
|
|
225 /* #undef opt_decwin
|
|
|
226 #define opt_decwin(fr) decwin_mmx */
|
|
|
227 #define opt_dct64(fr) dct64_mmx /* dct64_sse is silent in downsampling modes */
|
|
|
228 #define opt_synth_1to1(fr) synth_1to1_3dnowext /* that will use dct64_3dnowext */
|
|
|
229 #undef opt_make_decode_tables
|
|
|
230 #define opt_make_decode_tables(fr) make_decode_tables_mmx(fr)
|
|
|
231 #undef opt_init_layer3_gainpow2
|
|
|
232 #define opt_init_layer3_gainpow2(fr) init_layer3_gainpow2_mmx
|
|
|
233 #undef opt_init_layer2_table
|
|
|
234 #define opt_init_layer2_table(fr) init_layer2_table_mmx
|
|
|
235 #define OPT_MMX_ONLY /* watch out! */
|
|
|
236 #endif
|
|
|
237 #endif
|
|
|
238
|
|
|
239
|
|
|
240 #ifndef OPT_MMX_ONLY
|
|
|
241 extern real *pnts[5];
|
|
|
242 extern real decwin[512+32];
|
|
|
243 #endif
|
|
|
244 #ifdef OPT_MPLAYER
|
|
|
245 extern const int costab_mmxsse[];
|
|
|
246 #endif
|
|
|
247
|
|
|
248 /* 3dnow used to use synth_1to1_i586 for mono / 8bit conversion - was that intentional? */
|
|
|
249 /* I'm trying to skip the pentium code here ... until I see that that is indeed a bad idea */
|
|
|
250 #ifdef OPT_3DNOW
|
|
|
251 #define K6_FALLBACK /* a fallback for 3DNowExt */
|
|
|
252 #define OPT_X86
|
|
|
253 void dct36_3dnow(real *,real *,real *,real *,real *);
|
|
|
254 void do_equalizer_3dnow(real *bandPtr,int channel, real equalizer[2][32]);
|
|
|
255 int synth_1to1_3dnow(real *bandPtr, int channel, mpg123_handle *fr, int final);
|
|
|
256 int synth_1to1_3dnow_asm(real *bandPtr, int channel, unsigned char *out, unsigned char *buffs, int *bo, real *decwin);
|
|
|
257 #ifndef OPT_MULTI
|
|
|
258 #define defopt dreidnow
|
|
|
259 #undef opt_dct36
|
|
|
260 #define opt_dct36(fr) dct36_3dnow
|
|
|
261 #define opt_synth_1to1(fr) synth_1to1_3dnow
|
|
|
262 #endif
|
|
|
263 #endif
|
|
|
264
|
|
|
265 #ifdef OPT_X86
|
|
|
266 /* these have to be merged back into one! */
|
|
|
267 unsigned int getcpuid();
|
|
|
268 unsigned int getextcpuflags();
|
|
|
269 unsigned int getstdcpuflags();
|
|
|
270 unsigned int getstd2cpuflags();
|
|
|
271
|
|
|
272 void dct64_i386(real *,real *,real *);
|
|
|
273 int synth_1to1_mono_i386(real *, mpg123_handle *fr);
|
|
|
274 int synth_1to1_mono2stereo_i386(real *, mpg123_handle *fr);
|
|
|
275 int synth_1to1_8bit_i386(real *,int, mpg123_handle *fr, int final);
|
|
|
276 int synth_1to1_8bit_mono_i386(real *, mpg123_handle *fr);
|
|
|
277 int synth_1to1_8bit_mono2stereo_i386(real *, mpg123_handle *fr);
|
|
|
278 #ifndef OPT_MULTI
|
|
|
279 #ifndef opt_dct64
|
|
|
280 #define opt_dct64(fr) dct64_i386 /* default one even for 3dnow and i486 in decode_2to1, decode_ntom */
|
|
|
281 #endif
|
|
|
282 #define opt_synth_1to1_mono(fr) synth_1to1_mono_i386
|
|
|
283 #define opt_synth_1to1_mono2stereo(fr) synth_1to1_mono2stereo_i386
|
|
|
284 #define opt_synth_1to1_8bit(fr) synth_1to1_8bit_i386
|
|
|
285 #define opt_synth_1to1_8bit_mono(fr) synth_1to1_8bit_mono_i386
|
|
|
286 #define opt_synth_1to1_8bit_mono2stereo(fr) synth_1to1_8bit_mono2stereo_i386
|
|
|
287 #endif
|
|
|
288 #endif
|
|
|
289
|
|
|
290 #ifdef OPT_ALTIVEC
|
|
|
291 void dct64_altivec(real *out0,real *out1,real *samples);
|
|
|
292 int synth_1to1_altivec(real *,int,mpg123_handle *, int);
|
|
|
293 int synth_1to1_mono_altivec(real *,mpg123_handle *);
|
|
|
294 int synth_1to1_mono2stereo_altivec(real *, mpg123_handle *);
|
|
|
295 int synth_1to1_8bit_altivec(real *,int,mpg123_handle *,int);
|
|
|
296 int synth_1to1_8bit_mono_altivec(real *,mpg123_handle *);
|
|
|
297 int synth_1to1_8bit_mono2stereo_altivec(real *,mpg123_handle *);
|
|
|
298 #ifndef OPT_MULTI
|
|
|
299 #define defopt altivec
|
|
|
300 #define opt_dct64(fr) dct64_altivec
|
|
|
301 #define opt_synth_1to1(fr) synth_1to1_altivec
|
|
|
302 #define opt_synth_1to1_mono(fr) synth_1to1_mono_altivec
|
|
|
303 #define opt_synth_1to1_mono2stereo(fr) synth_1to1_mono2stereo_altivec
|
|
|
304 #define opt_synth_1to1_8bit(fr) synth_1to1_8bit_altivec
|
|
|
305 #define opt_synth_1to1_8bit_mono(fr) synth_1to1_8bit_mono_altivec
|
|
|
306 #define opt_synth_1to1_8bit_mono2stereo(fr) synth_1to1_8bit_mono2stereo_altivec
|
|
|
307 #endif
|
|
|
308 #endif
|
|
|
309
|
|
|
310 /* used for multi opt mode and the single 3dnow mode to have the old 3dnow test flag still working */
|
|
|
311 void check_decoders(void);
|
|
|
312
|
|
|
313 #ifdef OPT_MULTI
|
|
|
314 #ifdef OPT_X86
|
|
|
315 extern struct cpuflags cf;
|
|
|
316 #endif
|
|
|
317 #define defopt nodec
|
|
|
318 /* a simple global struct to hold the decoding function pointers, could be localized later if really wanted */
|
|
|
319
|
|
|
320 #define opt_synth_1to1(fr) ((fr)->cpu_opts.synth_1to1)
|
|
|
321 #define opt_synth_1to1_mono(fr) ((fr)->cpu_opts.synth_1to1_mono)
|
|
|
322 #define opt_synth_1to1_mono2stereo(fr) ((fr)->cpu_opts.synth_1to1_mono2stereo)
|
|
|
323 #define opt_synth_1to1_8bit(fr) ((fr)->cpu_opts.synth_1to1_8bit)
|
|
|
324 #define opt_synth_1to1_8bit_mono(fr) ((fr)->cpu_opts.synth_1to1_8bit_mono)
|
|
|
325 #define opt_synth_1to1_8bit_mono2stereo(fr) ((fr)->cpu_opts.synth_1to1_8bit_mono2stereo)
|
|
|
326 #ifdef OPT_PENTIUM
|
|
|
327 #define opt_synth_1to1_i586_asm(fr) ((fr)->cpu_opts.synth_1to1_i586_asm)
|
|
|
328 #endif
|
|
|
329 #ifdef OPT_MMXORSSE
|
|
|
330 #undef opt_make_decode_tables
|
|
|
331 #define opt_make_decode_tables(fr) ((fr)->cpu_opts.make_decode_tables)(fr)
|
|
|
332 /* #undef opt_decwin
|
|
|
333 #define opt_decwin(fr) (fr)->cpu_opts.decwin */
|
|
|
334 #undef opt_init_layer3_gainpow2
|
|
|
335 #define opt_init_layer3_gainpow2(fr) ((fr)->cpu_opts.init_layer3_gainpow2)
|
|
|
336 #undef opt_init_layer2_table
|
|
|
337 #define opt_init_layer2_table(fr) ((fr)->cpu_opts.init_layer2_table)
|
|
|
338 #endif
|
|
|
339 #ifdef OPT_3DNOW
|
|
|
340 #undef opt_dct36
|
|
|
341 #define opt_dct36(fr) ((fr)->cpu_opts.dct36)
|
|
|
342 #endif
|
|
|
343 #define opt_dct64(fr) ((fr)->cpu_opts.dct64)
|
|
|
344 #ifdef OPT_MPLAYER
|
|
|
345 #define opt_mpl_dct64(fr) ((fr)->cpu_opts.mpl_dct64)
|
|
|
346 #endif
|
|
|
347 #endif
|
|
|
348
|
|
|
349 #endif /* MPG123_H_OPTIMIZE */
|
|
|
350
|
