| 3960 |
Serge |
1 |
/* |
|
|
2 |
decode_i586_dither: asm synth with dither noise |
|
|
3 |
|
|
|
4 |
copyright ?-2007 by the mpg123 project - free software under the terms of the LGPL 2.1 |
|
|
5 |
see COPYING and AUTHORS files in distribution or http://mpg123.org |
|
|
6 |
initially written by Stefan Bieschewski as decode_i586.s without dither |
|
|
7 |
|
|
|
8 |
This version uses "circular" 64k dither noise. |
|
|
9 |
(Patch by Adrian ) |
|
|
10 |
|
|
|
11 |
Thomas learned something about assembler and the stack while making this one thread safe (removing static data). |
|
|
12 |
*/ |
|
|
13 |
|
|
|
14 |
#include "mangle.h" |
|
|
15 |
|
|
|
16 |
.data |
|
|
17 |
#ifndef __APPLE__ |
|
|
18 |
.section .rodata |
|
|
19 |
#endif |
|
|
20 |
ALIGN8 |
|
|
21 |
.LC0: |
|
|
22 |
.long 0x0,0x40dfffc0 |
|
|
23 |
ALIGN8 |
|
|
24 |
.LC1: |
|
|
25 |
.long 0x0,0xc0e00000 |
|
|
26 |
ALIGN8 |
|
|
27 |
.text |
|
|
28 |
/* int synth_1to1_i586_asm_dither(real *bandPtr, int channel, unsigned char *out, unsigned char *buffs, int bo_and_ditherindex[2], real *decwin, real* dithernoise); */ |
|
|
29 |
.globl ASM_NAME(synth_1to1_i586_asm_dither) |
|
|
30 |
ASM_NAME(synth_1to1_i586_asm_dither): |
|
|
31 |
subl $16,%esp |
|
|
32 |
pushl %ebp |
|
|
33 |
pushl %edi |
|
|
34 |
pushl %esi |
|
|
35 |
pushl %ebx |
|
|
36 |
/* stack: 0(%esp)=%ebx 4=esi 8=edi 12=ebp 16,20,24,28=local 32=back 36=bandptr 40=channel 44=out 48=buffs 52=bo 56=decwin 60=dithernoise */ |
|
|
37 |
#define BANDPTR 36(%esp) |
|
|
38 |
#define CHANNEL 40(%esp) |
|
|
39 |
#define OUT 44(%esp) |
|
|
40 |
#define BUFFS 48(%esp) |
|
|
41 |
#define BO 52(%esp) |
|
|
42 |
#define DECWIN 56(%esp) |
|
|
43 |
#define DITHERNOISE 60(%esp) |
|
|
44 |
/*#define DITHERNOISE $(ASM_NAME(dithernoise))*/ |
|
|
45 |
#define LOC0 16(%esp) |
|
|
46 |
#define LOC1 20(%esp) |
|
|
47 |
#define LOC2 24(%esp) |
|
|
48 |
#define DITHERINDEX 28(%esp) |
|
|
49 |
/* During application of the dithering, we need the shifted locations because there's an additional value on the stack. */ |
|
|
50 |
#define DITHERNOISE2 64(%esp) |
|
|
51 |
#define DITHERINDEX2 32(%esp) |
|
|
52 |
|
|
|
53 |
movl BANDPTR,%eax |
|
|
54 |
movl OUT,%esi |
|
|
55 |
movl BO, %ebx |
|
|
56 |
movl (%ebx),%ebp /* get bo value */ |
|
|
57 |
movl 4(%ebx),%edi; /* get the ditherindex behind bo */ |
|
|
58 |
movl %edi,DITHERINDEX |
|
|
59 |
xorl %edi,%edi |
|
|
60 |
cmpl %edi,CHANNEL |
|
|
61 |
jne .L48 |
|
|
62 |
decl %ebp |
|
|
63 |
andl $15,%ebp |
|
|
64 |
movl %ebp,(%ebx) /* save bo back */ |
|
|
65 |
movl BUFFS,%ecx |
|
|
66 |
jmp .L49 |
|
|
67 |
.L48: |
|
|
68 |
/* In stereo mode , "rewind" dither pointer 32 samples , so 2nd channel */ |
|
|
69 |
/* has same dither values. Tested OK for mono and stereo MP2 and MP3 */ |
|
|
70 |
subl $128,DITHERINDEX /* better move to %edi for the two calculations? */ |
|
|
71 |
andl $0x0003fffc,DITHERINDEX |
|
|
72 |
addl $2,%esi |
|
|
73 |
movl BUFFS,%ecx |
|
|
74 |
addl $2176,%ecx |
|
|
75 |
.L49: |
|
|
76 |
/* now the call of dct64 is prepared, stuff pushed to the stack, but soon after it's removed again */ |
|
|
77 |
testl $1,%ebp |
|
|
78 |
je .L50 |
|
|
79 |
movl %ecx,%ebx |
|
|
80 |
movl %ebp,LOC0 |
|
|
81 |
pushl %eax |
|
|
82 |
movl LOC1,%edx |
|
|
83 |
leal (%ebx,%edx,4),%eax |
|
|
84 |
pushl %eax |
|
|
85 |
movl LOC2,%eax |
|
|
86 |
incl %eax |
|
|
87 |
andl $15,%eax |
|
|
88 |
leal 1088(,%eax,4),%eax |
|
|
89 |
addl %ebx,%eax |
|
|
90 |
jmp .L74 |
|
|
91 |
.L50: |
|
|
92 |
leal 1088(%ecx),%ebx |
|
|
93 |
leal 1(%ebp),%edx |
|
|
94 |
movl %edx,LOC0 |
|
|
95 |
pushl %eax |
|
|
96 |
leal 1092(%ecx,%ebp,4),%eax |
|
|
97 |
pushl %eax |
|
|
98 |
leal (%ecx,%ebp,4),%eax |
|
|
99 |
.L74: |
|
|
100 |
pushl %eax |
|
|
101 |
call ASM_NAME(dct64_i386) |
|
|
102 |
addl $12,%esp |
|
|
103 |
/* Now removed the parameters. |
|
|
104 |
stack: 0(%esp)=%ebx 4=esi 8=edi 12=ebp 16,20,24,28=local 32=back 36=bandptr 40=channel 44=out 48=buffs 52=bo 56=decwin 60=dithernoise */ |
|
|
105 |
movl LOC0,%edx |
|
|
106 |
leal 0(,%edx,4),%edx |
|
|
107 |
/* movl ASM_VALUE(decwin)+64,%eax */ |
|
|
108 |
movl DECWIN,%eax |
|
|
109 |
addl $64,%eax |
|
|
110 |
movl %eax,%ecx |
|
|
111 |
subl %edx,%ecx |
|
|
112 |
movl $16,%ebp |
|
|
113 |
.L55: |
|
|
114 |
flds (%ecx) |
|
|
115 |
fmuls (%ebx) |
|
|
116 |
flds 4(%ecx) |
|
|
117 |
fmuls 4(%ebx) |
|
|
118 |
fxch %st(1) |
|
|
119 |
flds 8(%ecx) |
|
|
120 |
fmuls 8(%ebx) |
|
|
121 |
fxch %st(2) |
|
|
122 |
fsubrp %st,%st(1) |
|
|
123 |
flds 12(%ecx) |
|
|
124 |
fmuls 12(%ebx) |
|
|
125 |
fxch %st(2) |
|
|
126 |
faddp %st,%st(1) |
|
|
127 |
flds 16(%ecx) |
|
|
128 |
fmuls 16(%ebx) |
|
|
129 |
fxch %st(2) |
|
|
130 |
fsubrp %st,%st(1) |
|
|
131 |
flds 20(%ecx) |
|
|
132 |
fmuls 20(%ebx) |
|
|
133 |
fxch %st(2) |
|
|
134 |
faddp %st,%st(1) |
|
|
135 |
flds 24(%ecx) |
|
|
136 |
fmuls 24(%ebx) |
|
|
137 |
fxch %st(2) |
|
|
138 |
fsubrp %st,%st(1) |
|
|
139 |
flds 28(%ecx) |
|
|
140 |
fmuls 28(%ebx) |
|
|
141 |
fxch %st(2) |
|
|
142 |
faddp %st,%st(1) |
|
|
143 |
flds 32(%ecx) |
|
|
144 |
fmuls 32(%ebx) |
|
|
145 |
fxch %st(2) |
|
|
146 |
fsubrp %st,%st(1) |
|
|
147 |
flds 36(%ecx) |
|
|
148 |
fmuls 36(%ebx) |
|
|
149 |
fxch %st(2) |
|
|
150 |
faddp %st,%st(1) |
|
|
151 |
flds 40(%ecx) |
|
|
152 |
fmuls 40(%ebx) |
|
|
153 |
fxch %st(2) |
|
|
154 |
fsubrp %st,%st(1) |
|
|
155 |
flds 44(%ecx) |
|
|
156 |
fmuls 44(%ebx) |
|
|
157 |
fxch %st(2) |
|
|
158 |
faddp %st,%st(1) |
|
|
159 |
flds 48(%ecx) |
|
|
160 |
fmuls 48(%ebx) |
|
|
161 |
fxch %st(2) |
|
|
162 |
fsubrp %st,%st(1) |
|
|
163 |
flds 52(%ecx) |
|
|
164 |
fmuls 52(%ebx) |
|
|
165 |
fxch %st(2) |
|
|
166 |
faddp %st,%st(1) |
|
|
167 |
flds 56(%ecx) |
|
|
168 |
fmuls 56(%ebx) |
|
|
169 |
fxch %st(2) |
|
|
170 |
fsubrp %st,%st(1) |
|
|
171 |
flds 60(%ecx) |
|
|
172 |
fmuls 60(%ebx) |
|
|
173 |
fxch %st(2) |
|
|
174 |
subl $4,%esp |
|
|
175 |
faddp %st,%st(1) |
|
|
176 |
fxch %st(1) |
|
|
177 |
fsubrp %st,%st(1) |
|
|
178 |
|
|
|
179 |
addl $4,DITHERINDEX2 |
|
|
180 |
andl $0x0003fffc,DITHERINDEX2 |
|
|
181 |
movl DITHERNOISE2,%edi |
|
|
182 |
addl DITHERINDEX2,%edi |
|
|
183 |
|
|
|
184 |
fadds (%edi) |
|
|
185 |
|
|
|
186 |
/* fistpl and popl as a unit keep the stack unchanged */ |
|
|
187 |
fistpl (%esp) |
|
|
188 |
popl %eax |
|
|
189 |
cmpl $32767,%eax |
|
|
190 |
jg 1f |
|
|
191 |
cmpl $-32768,%eax |
|
|
192 |
jl 2f |
|
|
193 |
movw %ax,(%esi) |
|
|
194 |
jmp 4f |
|
|
195 |
1: movw $32767,(%esi) |
|
|
196 |
jmp 3f |
|
|
197 |
2: movw $-32768,(%esi) |
|
|
198 |
3: |
|
|
199 |
/* incl %edi */ |
|
|
200 |
4: |
|
|
201 |
.L54: |
|
|
202 |
addl $64,%ebx |
|
|
203 |
subl $-128,%ecx |
|
|
204 |
addl $4,%esi |
|
|
205 |
decl %ebp |
|
|
206 |
jnz .L55 |
|
|
207 |
flds (%ecx) |
|
|
208 |
fmuls (%ebx) |
|
|
209 |
flds 8(%ecx) |
|
|
210 |
fmuls 8(%ebx) |
|
|
211 |
flds 16(%ecx) |
|
|
212 |
fmuls 16(%ebx) |
|
|
213 |
fxch %st(2) |
|
|
214 |
faddp %st,%st(1) |
|
|
215 |
flds 24(%ecx) |
|
|
216 |
fmuls 24(%ebx) |
|
|
217 |
fxch %st(2) |
|
|
218 |
faddp %st,%st(1) |
|
|
219 |
flds 32(%ecx) |
|
|
220 |
fmuls 32(%ebx) |
|
|
221 |
fxch %st(2) |
|
|
222 |
faddp %st,%st(1) |
|
|
223 |
flds 40(%ecx) |
|
|
224 |
fmuls 40(%ebx) |
|
|
225 |
fxch %st(2) |
|
|
226 |
faddp %st,%st(1) |
|
|
227 |
flds 48(%ecx) |
|
|
228 |
fmuls 48(%ebx) |
|
|
229 |
fxch %st(2) |
|
|
230 |
faddp %st,%st(1) |
|
|
231 |
flds 56(%ecx) |
|
|
232 |
fmuls 56(%ebx) |
|
|
233 |
fxch %st(2) |
|
|
234 |
subl $4,%esp |
|
|
235 |
faddp %st,%st(1) |
|
|
236 |
fxch %st(1) |
|
|
237 |
faddp %st,%st(1) |
|
|
238 |
|
|
|
239 |
addl $4,DITHERINDEX2 |
|
|
240 |
andl $0x0003fffc,DITHERINDEX2 |
|
|
241 |
movl DITHERNOISE2,%edi |
|
|
242 |
addl DITHERINDEX2,%edi |
|
|
243 |
|
|
|
244 |
fadds (%edi) |
|
|
245 |
/* fistpl and popl as a unit keep the stack unchanged */ |
|
|
246 |
fistpl (%esp) |
|
|
247 |
popl %eax |
|
|
248 |
cmpl $32767,%eax |
|
|
249 |
jg 1f |
|
|
250 |
cmpl $-32768,%eax |
|
|
251 |
jl 2f |
|
|
252 |
movw %ax,(%esi) |
|
|
253 |
jmp 4f |
|
|
254 |
1: movw $32767,(%esi) |
|
|
255 |
jmp 3f |
|
|
256 |
2: movw $-32768,(%esi) |
|
|
257 |
3: |
|
|
258 |
/* incl %edi */ |
|
|
259 |
4: |
|
|
260 |
.L62: |
|
|
261 |
addl $-64,%ebx |
|
|
262 |
addl $4,%esi |
|
|
263 |
movl LOC0,%edx |
|
|
264 |
leal -128(%ecx,%edx,8),%ecx |
|
|
265 |
movl $15,%ebp |
|
|
266 |
.L68: |
|
|
267 |
flds -4(%ecx) |
|
|
268 |
fchs |
|
|
269 |
fmuls (%ebx) |
|
|
270 |
flds -8(%ecx) |
|
|
271 |
fmuls 4(%ebx) |
|
|
272 |
fxch %st(1) |
|
|
273 |
flds -12(%ecx) |
|
|
274 |
fmuls 8(%ebx) |
|
|
275 |
fxch %st(2) |
|
|
276 |
fsubrp %st,%st(1) |
|
|
277 |
flds -16(%ecx) |
|
|
278 |
fmuls 12(%ebx) |
|
|
279 |
fxch %st(2) |
|
|
280 |
fsubrp %st,%st(1) |
|
|
281 |
flds -20(%ecx) |
|
|
282 |
fmuls 16(%ebx) |
|
|
283 |
fxch %st(2) |
|
|
284 |
fsubrp %st,%st(1) |
|
|
285 |
flds -24(%ecx) |
|
|
286 |
fmuls 20(%ebx) |
|
|
287 |
fxch %st(2) |
|
|
288 |
fsubrp %st,%st(1) |
|
|
289 |
flds -28(%ecx) |
|
|
290 |
fmuls 24(%ebx) |
|
|
291 |
fxch %st(2) |
|
|
292 |
fsubrp %st,%st(1) |
|
|
293 |
flds -32(%ecx) |
|
|
294 |
fmuls 28(%ebx) |
|
|
295 |
fxch %st(2) |
|
|
296 |
fsubrp %st,%st(1) |
|
|
297 |
flds -36(%ecx) |
|
|
298 |
fmuls 32(%ebx) |
|
|
299 |
fxch %st(2) |
|
|
300 |
fsubrp %st,%st(1) |
|
|
301 |
flds -40(%ecx) |
|
|
302 |
fmuls 36(%ebx) |
|
|
303 |
fxch %st(2) |
|
|
304 |
fsubrp %st,%st(1) |
|
|
305 |
flds -44(%ecx) |
|
|
306 |
fmuls 40(%ebx) |
|
|
307 |
fxch %st(2) |
|
|
308 |
fsubrp %st,%st(1) |
|
|
309 |
flds -48(%ecx) |
|
|
310 |
fmuls 44(%ebx) |
|
|
311 |
fxch %st(2) |
|
|
312 |
fsubrp %st,%st(1) |
|
|
313 |
flds -52(%ecx) |
|
|
314 |
fmuls 48(%ebx) |
|
|
315 |
fxch %st(2) |
|
|
316 |
fsubrp %st,%st(1) |
|
|
317 |
flds -56(%ecx) |
|
|
318 |
fmuls 52(%ebx) |
|
|
319 |
fxch %st(2) |
|
|
320 |
fsubrp %st,%st(1) |
|
|
321 |
flds -60(%ecx) |
|
|
322 |
fmuls 56(%ebx) |
|
|
323 |
fxch %st(2) |
|
|
324 |
fsubrp %st,%st(1) |
|
|
325 |
flds (%ecx) |
|
|
326 |
fmuls 60(%ebx) |
|
|
327 |
fxch %st(2) |
|
|
328 |
subl $4,%esp |
|
|
329 |
fsubrp %st,%st(1) |
|
|
330 |
fxch %st(1) |
|
|
331 |
fsubrp %st,%st(1) |
|
|
332 |
|
|
|
333 |
addl $4,DITHERINDEX2 |
|
|
334 |
andl $0x0003fffc,DITHERINDEX2 |
|
|
335 |
movl DITHERNOISE2,%edi |
|
|
336 |
addl DITHERINDEX2,%edi |
|
|
337 |
|
|
|
338 |
fadds (%edi) |
|
|
339 |
/* fistpl and popl as a unit keep the stack unchanged */ |
|
|
340 |
fistpl (%esp) |
|
|
341 |
popl %eax |
|
|
342 |
cmpl $32767,%eax |
|
|
343 |
jg 1f |
|
|
344 |
cmpl $-32768,%eax |
|
|
345 |
jl 2f |
|
|
346 |
movw %ax,(%esi) |
|
|
347 |
jmp 4f |
|
|
348 |
1: movw $32767,(%esi) |
|
|
349 |
jmp 3f |
|
|
350 |
2: movw $-32768,(%esi) |
|
|
351 |
3: |
|
|
352 |
/* incl %edi */ |
|
|
353 |
4: |
|
|
354 |
.L67: |
|
|
355 |
addl $-64,%ebx |
|
|
356 |
addl $-128,%ecx |
|
|
357 |
addl $4,%esi |
|
|
358 |
decl %ebp |
|
|
359 |
jnz .L68 |
|
|
360 |
/* return ipv edi 0 in eax */ |
|
|
361 |
movl $0,%eax |
|
|
362 |
/* save ditherindex */ |
|
|
363 |
movl BO,%ebx |
|
|
364 |
movl DITHERINDEX,%esi |
|
|
365 |
movl %esi,4(%ebx); |
|
|
366 |
/* stack: 0=ebx 4=esi 8=edi 12=ebp 16,20,24,28=local 32=back 36=bandptr 40=channel 44=out 48=buffs 52=bo */ |
|
|
367 |
popl %ebx |
|
|
368 |
popl %esi |
|
|
369 |
popl %edi |
|
|
370 |
popl %ebp |
|
|
371 |
addl $16,%esp |
|
|
372 |
/* The stack must be now: 0=back 4=bandptr 8=channel 12=out 16=buffs 20=bo */ |
|
|
373 |
ret |
|
|
374 |
|
|
|
375 |
NONEXEC_STACK |