|
4147
|
1 /*
|
|
|
2 * mpg123_synth_1to1 works the same way as the c version of this
|
|
|
3 * file. only two types of changes have been made:
|
|
|
4 * - reordered floating point instructions to
|
|
|
5 * prevent pipline stalls
|
|
|
6 * - made WRITE_SAMPLE use integer instead of
|
|
|
7 * (slower) floating point
|
|
|
8 * all kinds of x86 processors should benefit from these
|
|
|
9 * modifications.
|
|
|
10 *
|
|
|
11 * useful sources of information on optimizing x86 code include:
|
|
|
12 *
|
|
|
13 * Intel Architecture Optimization Manual
|
|
|
14 * http://www.intel.com/design/pentium/manuals/242816.htm
|
|
|
15 *
|
|
|
16 * Cyrix 6x86 Instruction Set Summary
|
|
|
17 * ftp://ftp.cyrix.com/6x86/6x-dbch6.pdf
|
|
|
18 *
|
|
|
19 * AMD-K5 Processor Software Development
|
|
|
20 * http://www.amd.com/products/cpg/techdocs/appnotes/20007e.pdf
|
|
|
21 *
|
|
|
22 * Stefan Bieschewski <stb@acm.org>
|
|
|
23 *
|
|
|
24 * $Id$
|
|
|
25 */
|
|
|
26 #define real float /* ugly - but only way */
|
|
|
27
|
|
|
28 static long buffs[1088];
|
|
|
29 static long bo=1;
|
|
|
30
|
|
|
31 int synth_1to1_pent(real *bandPtr, int channel, short *samples)
|
|
|
32 {
|
|
|
33 real tmp[3];
|
|
|
34 register int retval;
|
|
|
35 __asm __volatile(
|
|
|
36 " movl %1,%%eax\n\t"/*bandPtr*/
|
|
|
37 " movl %3,%%esi\n\t"
|
|
|
38 " xorl %%edi,%%edi\n\t"
|
|
|
39 " movl bo,%%ebp\n\t"
|
|
|
40 " cmpl %%edi,%2\n\t"
|
|
|
41 " jne .L48\n\t"
|
|
|
42 " decl %%ebp\n\t"
|
|
|
43 " andl $15,%%ebp\n\t"
|
|
|
44 " movl %%ebp,bo\n\t"
|
|
|
45 " movl $buffs,%%ecx\n\t"
|
|
|
46 " jmp .L49\n\t"
|
|
|
47 ".L48:\n\t"
|
|
|
48 " addl $2,%%esi\n\t"
|
|
|
49 " movl $buffs+2176,%%ecx\n\t"
|
|
|
50 ".L49:\n\t"
|
|
|
51 " testl $1,%%ebp\n\t"
|
|
|
52 " je .L50\n\t"
|
|
|
53 " movl %%ecx,%%ebx\n\t"
|
|
|
54 " movl %%ebp,%4\n\t"
|
|
|
55 " pushl %%eax\n\t"
|
|
|
56 " movl 4+%4,%%edx\n\t"
|
|
|
57 " leal (%%ebx,%%edx,4),%%eax\n\t"
|
|
|
58 " pushl %%eax\n\t"
|
|
|
59 " movl 8+%4,%%eax\n\t"
|
|
|
60 " incl %%eax\n\t"
|
|
|
61 " andl $15,%%eax\n\t"
|
|
|
62 " leal 1088(,%%eax,4),%%eax\n\t"
|
|
|
63 " addl %%ebx,%%eax\n\t"
|
|
|
64 " jmp .L74\n\t"
|
|
|
65 ".L50:\n\t"
|
|
|
66 " leal 1088(%%ecx),%%ebx\n\t"
|
|
|
67 " leal 1(%%ebp),%%edx\n\t"
|
|
|
68 " movl %%edx,%4\n\t"
|
|
|
69 " pushl %%eax\n\t"
|
|
|
70 " leal 1092(%%ecx,%%ebp,4),%%eax\n\t"
|
|
|
71 " pushl %%eax\n\t"
|
|
|
72 " leal (%%ecx,%%ebp,4),%%eax\n\t"
|
|
|
73 ".L74:\n\t"
|
|
|
74 " pushl %%eax\n\t"
|
|
|
75 " call dct64\n\t"
|
|
|
76 " addl $12,%%esp\n\t"
|
|
|
77 " movl %4,%%edx\n\t"
|
|
|
78 " leal 0(,%%edx,4),%%edx\n\t"
|
|
|
79 " movl $decwin+64,%%eax\n\t"
|
|
|
80 " movl %%eax,%%ecx\n\t"
|
|
|
81 " subl %%edx,%%ecx\n\t"
|
|
|
82 " movl $16,%%ebp\n\t"
|
|
|
83 ".L55:\n\t"
|
|
|
84 " flds (%%ecx)\n\t"
|
|
|
85 " fmuls (%%ebx)\n\t"
|
|
|
86 " flds 4(%%ecx)\n\t"
|
|
|
87 " fmuls 4(%%ebx)\n\t"
|
|
|
88 " fxch %%st(1)\n\t"
|
|
|
89 " flds 8(%%ecx)\n\t"
|
|
|
90 " fmuls 8(%%ebx)\n\t"
|
|
|
91 " fxch %%st(2)\n\t"
|
|
|
92 " fsubrp %%st,%%st(1)\n\t"
|
|
|
93 " flds 12(%%ecx)\n\t"
|
|
|
94 " fmuls 12(%%ebx)\n\t"
|
|
|
95 " fxch %%st(2)\n\t"
|
|
|
96 " faddp %%st,%%st(1)\n\t"
|
|
|
97 " flds 16(%%ecx)\n\t"
|
|
|
98 " fmuls 16(%%ebx)\n\t"
|
|
|
99 " fxch %%st(2)\n\t"
|
|
|
100 " fsubrp %%st,%%st(1)\n\t"
|
|
|
101 " flds 20(%%ecx)\n\t"
|
|
|
102 " fmuls 20(%%ebx)\n\t"
|
|
|
103 " fxch %%st(2)\n\t"
|
|
|
104 " faddp %%st,%%st(1)\n\t"
|
|
|
105 " flds 24(%%ecx)\n\t"
|
|
|
106 " fmuls 24(%%ebx)\n\t"
|
|
|
107 " fxch %%st(2)\n\t"
|
|
|
108 " fsubrp %%st,%%st(1)\n\t"
|
|
|
109 " flds 28(%%ecx)\n\t"
|
|
|
110 " fmuls 28(%%ebx)\n\t"
|
|
|
111 " fxch %%st(2)\n\t"
|
|
|
112 " faddp %%st,%%st(1)\n\t"
|
|
|
113 " flds 32(%%ecx)\n\t"
|
|
|
114 " fmuls 32(%%ebx)\n\t"
|
|
|
115 " fxch %%st(2)\n\t"
|
|
|
116 " fsubrp %%st,%%st(1)\n\t"
|
|
|
117 " flds 36(%%ecx)\n\t"
|
|
|
118 " fmuls 36(%%ebx)\n\t"
|
|
|
119 " fxch %%st(2)\n\t"
|
|
|
120 " faddp %%st,%%st(1)\n\t"
|
|
|
121 " flds 40(%%ecx)\n\t"
|
|
|
122 " fmuls 40(%%ebx)\n\t"
|
|
|
123 " fxch %%st(2)\n\t"
|
|
|
124 " fsubrp %%st,%%st(1)\n\t"
|
|
|
125 " flds 44(%%ecx)\n\t"
|
|
|
126 " fmuls 44(%%ebx)\n\t"
|
|
|
127 " fxch %%st(2)\n\t"
|
|
|
128 " faddp %%st,%%st(1)\n\t"
|
|
|
129 " flds 48(%%ecx)\n\t"
|
|
|
130 " fmuls 48(%%ebx)\n\t"
|
|
|
131 " fxch %%st(2)\n\t"
|
|
|
132 " fsubrp %%st,%%st(1)\n\t"
|
|
|
133 " flds 52(%%ecx)\n\t"
|
|
|
134 " fmuls 52(%%ebx)\n\t"
|
|
|
135 " fxch %%st(2)\n\t"
|
|
|
136 " faddp %%st,%%st(1)\n\t"
|
|
|
137 " flds 56(%%ecx)\n\t"
|
|
|
138 " fmuls 56(%%ebx)\n\t"
|
|
|
139 " fxch %%st(2)\n\t"
|
|
|
140 " fsubrp %%st,%%st(1)\n\t"
|
|
|
141 " flds 60(%%ecx)\n\t"
|
|
|
142 " fmuls 60(%%ebx)\n\t"
|
|
|
143 " fxch %%st(2)\n\t"
|
|
|
144 " subl $4,%%esp\n\t"
|
|
|
145 " faddp %%st,%%st(1)\n\t"
|
|
|
146 " fxch %%st(1)\n\t"
|
|
|
147 " fsubrp %%st,%%st(1)\n\t"
|
|
|
148 " fistpl (%%esp)\n\t"
|
|
|
149 " popl %%eax\n\t"
|
|
|
150 " cmpl $32767,%%eax\n\t"
|
|
|
151 " jg 1f\n\t"
|
|
|
152 " cmpl $-32768,%%eax\n\t"
|
|
|
153 " jl 2f\n\t"
|
|
|
154 " movw %%ax,(%%esi)\n\t"
|
|
|
155 " jmp 4f\n\t"
|
|
|
156 "1: movw $32767,(%%esi)\n\t"
|
|
|
157 " jmp 3f\n\t"
|
|
|
158 "2: movw $-32768,(%%esi)\n\t"
|
|
|
159 "3: incl %%edi\n\t"
|
|
|
160 "4:\n\t"
|
|
|
161 ".L54:\n\t"
|
|
|
162 " addl $64,%%ebx\n\t"
|
|
|
163 " subl $-128,%%ecx\n\t"
|
|
|
164 " addl $4,%%esi\n\t"
|
|
|
165 " decl %%ebp\n\t"
|
|
|
166 " jnz .L55\n\t"
|
|
|
167 " flds (%%ecx)\n\t"
|
|
|
168 " fmuls (%%ebx)\n\t"
|
|
|
169 " flds 8(%%ecx)\n\t"
|
|
|
170 " fmuls 8(%%ebx)\n\t"
|
|
|
171 " flds 16(%%ecx)\n\t"
|
|
|
172 " fmuls 16(%%ebx)\n\t"
|
|
|
173 " fxch %%st(2)\n\t"
|
|
|
174 " faddp %%st,%%st(1)\n\t"
|
|
|
175 " flds 24(%%ecx)\n\t"
|
|
|
176 " fmuls 24(%%ebx)\n\t"
|
|
|
177 " fxch %%st(2)\n\t"
|
|
|
178 " faddp %%st,%%st(1)\n\t"
|
|
|
179 " flds 32(%%ecx)\n\t"
|
|
|
180 " fmuls 32(%%ebx)\n\t"
|
|
|
181 " fxch %%st(2)\n\t"
|
|
|
182 " faddp %%st,%%st(1)\n\t"
|
|
|
183 " flds 40(%%ecx)\n\t"
|
|
|
184 " fmuls 40(%%ebx)\n\t"
|
|
|
185 " fxch %%st(2)\n\t"
|
|
|
186 " faddp %%st,%%st(1)\n\t"
|
|
|
187 " flds 48(%%ecx)\n\t"
|
|
|
188 " fmuls 48(%%ebx)\n\t"
|
|
|
189 " fxch %%st(2)\n\t"
|
|
|
190 " faddp %%st,%%st(1)\n\t"
|
|
|
191 " flds 56(%%ecx)\n\t"
|
|
|
192 " fmuls 56(%%ebx)\n\t"
|
|
|
193 " fxch %%st(2)\n\t"
|
|
|
194 " subl $4,%%esp\n\t"
|
|
|
195 " faddp %%st,%%st(1)\n\t"
|
|
|
196 " fxch %%st(1)\n\t"
|
|
|
197 " faddp %%st,%%st(1)\n\t"
|
|
|
198 " fistpl (%%esp)\n\t"
|
|
|
199 " popl %%eax\n\t"
|
|
|
200 " cmpl $32767,%%eax\n\t"
|
|
|
201 " jg 1f\n\t"
|
|
|
202 " cmpl $-32768,%%eax\n\t"
|
|
|
203 " jl 2f\n\t"
|
|
|
204 " movw %%ax,(%%esi)\n\t"
|
|
|
205 " jmp 4f\n\t"
|
|
|
206 "1: movw $32767,(%%esi)\n\t"
|
|
|
207 " jmp 3f\n\t"
|
|
|
208 "2: movw $-32768,(%%esi)\n\t"
|
|
|
209 "3: incl %%edi\n\t"
|
|
|
210 "4:\n\t"
|
|
|
211 ".L62:\n\t"
|
|
|
212 " addl $-64,%%ebx\n\t"
|
|
|
213 " addl $4,%%esi\n\t"
|
|
|
214 " movl %4,%%edx\n\t"
|
|
|
215 " leal -128(%%ecx,%%edx,8),%%ecx\n\t"
|
|
|
216 " movl $15,%%ebp\n\t"
|
|
|
217 ".L68:\n\t"
|
|
|
218 " flds -4(%%ecx)\n\t"
|
|
|
219 " fchs\n\t"
|
|
|
220 " fmuls (%%ebx)\n\t"
|
|
|
221 " flds -8(%%ecx)\n\t"
|
|
|
222 " fmuls 4(%%ebx)\n\t"
|
|
|
223 " fxch %%st(1)\n\t"
|
|
|
224 " flds -12(%%ecx)\n\t"
|
|
|
225 " fmuls 8(%%ebx)\n\t"
|
|
|
226 " fxch %%st(2)\n\t"
|
|
|
227 " fsubrp %%st,%%st(1)\n\t"
|
|
|
228 " flds -16(%%ecx)\n\t"
|
|
|
229 " fmuls 12(%%ebx)\n\t"
|
|
|
230 " fxch %%st(2)\n\t"
|
|
|
231 " fsubrp %%st,%%st(1)\n\t"
|
|
|
232 " flds -20(%%ecx)\n\t"
|
|
|
233 " fmuls 16(%%ebx)\n\t"
|
|
|
234 " fxch %%st(2)\n\t"
|
|
|
235 " fsubrp %%st,%%st(1)\n\t"
|
|
|
236 " flds -24(%%ecx)\n\t"
|
|
|
237 " fmuls 20(%%ebx)\n\t"
|
|
|
238 " fxch %%st(2)\n\t"
|
|
|
239 " fsubrp %%st,%%st(1)\n\t"
|
|
|
240 " flds -28(%%ecx)\n\t"
|
|
|
241 " fmuls 24(%%ebx)\n\t"
|
|
|
242 " fxch %%st(2)\n\t"
|
|
|
243 " fsubrp %%st,%%st(1)\n\t"
|
|
|
244 " flds -32(%%ecx)\n\t"
|
|
|
245 " fmuls 28(%%ebx)\n\t"
|
|
|
246 " fxch %%st(2)\n\t"
|
|
|
247 " fsubrp %%st,%%st(1)\n\t"
|
|
|
248 " flds -36(%%ecx)\n\t"
|
|
|
249 " fmuls 32(%%ebx)\n\t"
|
|
|
250 " fxch %%st(2)\n\t"
|
|
|
251 " fsubrp %%st,%%st(1)\n\t"
|
|
|
252 " flds -40(%%ecx)\n\t"
|
|
|
253 " fmuls 36(%%ebx)\n\t"
|
|
|
254 " fxch %%st(2)\n\t"
|
|
|
255 " fsubrp %%st,%%st(1)\n\t"
|
|
|
256 " flds -44(%%ecx)\n\t"
|
|
|
257 " fmuls 40(%%ebx)\n\t"
|
|
|
258 " fxch %%st(2)\n\t"
|
|
|
259 " fsubrp %%st,%%st(1)\n\t"
|
|
|
260 " flds -48(%%ecx)\n\t"
|
|
|
261 " fmuls 44(%%ebx)\n\t"
|
|
|
262 " fxch %%st(2)\n\t"
|
|
|
263 " fsubrp %%st,%%st(1)\n\t"
|
|
|
264 " flds -52(%%ecx)\n\t"
|
|
|
265 " fmuls 48(%%ebx)\n\t"
|
|
|
266 " fxch %%st(2)\n\t"
|
|
|
267 " fsubrp %%st,%%st(1)\n\t"
|
|
|
268 " flds -56(%%ecx)\n\t"
|
|
|
269 " fmuls 52(%%ebx)\n\t"
|
|
|
270 " fxch %%st(2)\n\t"
|
|
|
271 " fsubrp %%st,%%st(1)\n\t"
|
|
|
272 " flds -60(%%ecx)\n\t"
|
|
|
273 " fmuls 56(%%ebx)\n\t"
|
|
|
274 " fxch %%st(2)\n\t"
|
|
|
275 " fsubrp %%st,%%st(1)\n\t"
|
|
|
276 " flds (%%ecx)\n\t"
|
|
|
277 " fmuls 60(%%ebx)\n\t"
|
|
|
278 " fxch %%st(2)\n\t"
|
|
|
279 " subl $4,%%esp\n\t"
|
|
|
280 " fsubrp %%st,%%st(1)\n\t"
|
|
|
281 " fxch %%st(1)\n\t"
|
|
|
282 " fsubrp %%st,%%st(1)\n\t"
|
|
|
283 " fistpl (%%esp)\n\t"
|
|
|
284 " popl %%eax\n\t"
|
|
|
285 " cmpl $32767,%%eax\n\t"
|
|
|
286 " jg 1f\n\t"
|
|
|
287 " cmpl $-32768,%%eax\n\t"
|
|
|
288 " jl 2f\n\t"
|
|
|
289 " movw %%ax,(%%esi)\n\t"
|
|
|
290 " jmp 4f\n\t"
|
|
|
291 "1: movw $32767,(%%esi)\n\t"
|
|
|
292 " jmp 3f\n\t"
|
|
|
293 "2: movw $-32768,(%%esi)\n\t"
|
|
|
294 "3: incl %%edi\n\t"
|
|
|
295 "4:\n\t"
|
|
|
296 ".L67:\n\t"
|
|
|
297 " addl $-64,%%ebx\n\t"
|
|
|
298 " addl $-128,%%ecx\n\t"
|
|
|
299 " addl $4,%%esi\n\t"
|
|
|
300 " decl %%ebp\n\t"
|
|
|
301 " jnz .L68\n\t"
|
|
|
302 " movl %%edi,%%eax\n\t"
|
|
|
303 :"=a"(retval)
|
|
|
304 :"m"(bandPtr),"m"(channel),"m"(samples),"m"(tmp[0])
|
|
|
305 :"memory","%ebp","%edi","%esi","%ebx");
|
|
|
306 return retval;
|
|
|
307 }
|
