Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
3960 | Serge | 1 | /* |
2 | dct64_sse: MMX/SSE optimized dct64 |
||
3 | |||
4 | copyright 1995-2009 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 Taihei Monma |
||
7 | */ |
||
8 | |||
9 | #include "mangle.h" |
||
10 | |||
11 | #define ARG(n) (8+n*4)(%ebp) |
||
12 | #define TEMP(n) (4+n*16)(%esp) |
||
13 | #define TEMP_BYTE(n) (4+n)(%esp) |
||
14 | |||
15 | /* |
||
16 | void dct64_sse(short *out0, short *out1, real *samples); |
||
17 | */ |
||
18 | |||
19 | #ifndef __APPLE__ |
||
20 | .section .rodata |
||
21 | #else |
||
22 | .data |
||
23 | #endif |
||
24 | ALIGN16 |
||
25 | pnpn: |
||
26 | .long 0 |
||
27 | .long -2147483648 |
||
28 | .long 0 |
||
29 | .long -2147483648 |
||
30 | ALIGN16 |
||
31 | mask: |
||
32 | .long -1 |
||
33 | .long -1 |
||
34 | .long -1 |
||
35 | .long 0 |
||
36 | |||
37 | .text |
||
38 | ALIGN16 |
||
39 | .globl ASM_NAME(dct64_sse) |
||
40 | ASM_NAME(dct64_sse): |
||
41 | pushl %ebp |
||
42 | movl %esp, %ebp |
||
43 | |||
44 | andl $-16, %esp /* align the stack at 16 bytes */ |
||
45 | subl $128, %esp /* reserve space for temporal store */ |
||
46 | pushl %ebx |
||
47 | |||
48 | movl ARG(0), %ecx |
||
49 | movl ARG(1), %ebx |
||
50 | movl ARG(2), %eax |
||
51 | |||
52 | MOVUAPS (%eax), %xmm7 |
||
53 | MOVUAPS 16(%eax), %xmm6 |
||
54 | MOVUAPS 112(%eax), %xmm0 |
||
55 | MOVUAPS 96(%eax), %xmm1 |
||
56 | shufps $0x1b, %xmm0, %xmm0 |
||
57 | shufps $0x1b, %xmm1, %xmm1 |
||
58 | movaps %xmm7, %xmm4 |
||
59 | movaps %xmm6, %xmm5 |
||
60 | addps %xmm0, %xmm4 |
||
61 | addps %xmm1, %xmm5 |
||
62 | subps %xmm0, %xmm7 |
||
63 | subps %xmm1, %xmm6 |
||
64 | movaps %xmm4, TEMP(0) |
||
65 | movaps %xmm5, TEMP(1) |
||
66 | |||
67 | MOVUAPS 32(%eax), %xmm2 |
||
68 | MOVUAPS 48(%eax), %xmm3 |
||
69 | MOVUAPS 80(%eax), %xmm0 |
||
70 | MOVUAPS 64(%eax), %xmm1 |
||
71 | shufps $0x1b, %xmm0, %xmm0 |
||
72 | shufps $0x1b, %xmm1, %xmm1 |
||
73 | movaps %xmm2, %xmm5 |
||
74 | movaps %xmm3, %xmm4 |
||
75 | addps %xmm0, %xmm2 |
||
76 | addps %xmm1, %xmm3 |
||
77 | subps %xmm0, %xmm5 |
||
78 | subps %xmm1, %xmm4 |
||
79 | |||
80 | mulps ASM_NAME(costab_mmxsse), %xmm7 |
||
81 | mulps ASM_NAME(costab_mmxsse)+16, %xmm6 |
||
82 | mulps ASM_NAME(costab_mmxsse)+32, %xmm5 |
||
83 | mulps ASM_NAME(costab_mmxsse)+48, %xmm4 |
||
84 | |||
85 | shufps $0x1b, %xmm2, %xmm2 |
||
86 | shufps $0x1b, %xmm3, %xmm3 |
||
87 | shufps $0x1b, %xmm4, %xmm4 |
||
88 | shufps $0x1b, %xmm5, %xmm5 |
||
89 | movaps TEMP(0), %xmm0 |
||
90 | movaps TEMP(1), %xmm1 |
||
91 | subps %xmm3, %xmm0 |
||
92 | subps %xmm2, %xmm1 |
||
93 | addps TEMP(0), %xmm3 |
||
94 | addps TEMP(1), %xmm2 |
||
95 | movaps %xmm3, TEMP(0) |
||
96 | movaps %xmm2, TEMP(1) |
||
97 | movaps %xmm6, %xmm2 |
||
98 | movaps %xmm7, %xmm3 |
||
99 | subps %xmm5, %xmm6 |
||
100 | subps %xmm4, %xmm7 |
||
101 | addps %xmm3, %xmm4 |
||
102 | addps %xmm2, %xmm5 |
||
103 | mulps ASM_NAME(costab_mmxsse)+64, %xmm0 |
||
104 | mulps ASM_NAME(costab_mmxsse)+80, %xmm1 |
||
105 | mulps ASM_NAME(costab_mmxsse)+80, %xmm6 |
||
106 | mulps ASM_NAME(costab_mmxsse)+64, %xmm7 |
||
107 | |||
108 | movaps TEMP(0), %xmm2 |
||
109 | movaps TEMP(1), %xmm3 |
||
110 | shufps $0x1b, %xmm3, %xmm3 |
||
111 | shufps $0x1b, %xmm5, %xmm5 |
||
112 | shufps $0x1b, %xmm1, %xmm1 |
||
113 | shufps $0x1b, %xmm6, %xmm6 |
||
114 | movaps %xmm0, TEMP(1) |
||
115 | subps %xmm3, %xmm2 |
||
116 | subps %xmm1, %xmm0 |
||
117 | addps TEMP(0), %xmm3 |
||
118 | addps TEMP(1), %xmm1 |
||
119 | movaps %xmm3, TEMP(0) |
||
120 | movaps %xmm1, TEMP(2) |
||
121 | movaps %xmm5, %xmm1 |
||
122 | movaps %xmm4, %xmm5 |
||
123 | movaps %xmm7, %xmm3 |
||
124 | subps %xmm1, %xmm5 |
||
125 | subps %xmm6, %xmm7 |
||
126 | addps %xmm1, %xmm4 |
||
127 | addps %xmm3, %xmm6 |
||
128 | mulps ASM_NAME(costab_mmxsse)+96, %xmm2 |
||
129 | mulps ASM_NAME(costab_mmxsse)+96, %xmm0 |
||
130 | mulps ASM_NAME(costab_mmxsse)+96, %xmm5 |
||
131 | mulps ASM_NAME(costab_mmxsse)+96, %xmm7 |
||
132 | movaps %xmm2, TEMP(1) |
||
133 | movaps %xmm0, TEMP(3) |
||
134 | |||
135 | movaps %xmm4, %xmm2 |
||
136 | movaps %xmm5, %xmm3 |
||
137 | shufps $0x44, %xmm6, %xmm2 |
||
138 | shufps $0xbb, %xmm7, %xmm5 |
||
139 | shufps $0xbb, %xmm6, %xmm4 |
||
140 | shufps $0x44, %xmm7, %xmm3 |
||
141 | movaps %xmm2, %xmm6 |
||
142 | movaps %xmm3, %xmm7 |
||
143 | subps %xmm4, %xmm2 |
||
144 | subps %xmm5, %xmm3 |
||
145 | addps %xmm6, %xmm4 |
||
146 | addps %xmm7, %xmm5 |
||
147 | movaps ASM_NAME(costab_mmxsse)+112, %xmm0 |
||
148 | movlhps %xmm0, %xmm0 |
||
149 | mulps %xmm0, %xmm2 |
||
150 | mulps %xmm0, %xmm3 |
||
151 | movaps %xmm0, TEMP(4) |
||
152 | movaps %xmm4, %xmm6 |
||
153 | movaps %xmm5, %xmm7 |
||
154 | shufps $0x14, %xmm2, %xmm4 |
||
155 | shufps $0xbe, %xmm2, %xmm6 |
||
156 | shufps $0x14, %xmm3, %xmm5 |
||
157 | shufps $0xbe, %xmm3, %xmm7 |
||
158 | movaps %xmm5, TEMP(5) |
||
159 | movaps %xmm7, TEMP(7) |
||
160 | |||
161 | movaps TEMP(0), %xmm0 |
||
162 | movaps TEMP(1), %xmm1 |
||
163 | movaps %xmm0, %xmm2 |
||
164 | movaps %xmm1, %xmm3 |
||
165 | shufps $0x44, TEMP(2), %xmm2 |
||
166 | shufps $0xbb, TEMP(3), %xmm1 |
||
167 | shufps $0xbb, TEMP(2), %xmm0 |
||
168 | shufps $0x44, TEMP(3), %xmm3 |
||
169 | movaps %xmm2, %xmm5 |
||
170 | movaps %xmm3, %xmm7 |
||
171 | subps %xmm0, %xmm2 |
||
172 | subps %xmm1, %xmm3 |
||
173 | addps %xmm5, %xmm0 |
||
174 | addps %xmm7, %xmm1 |
||
175 | mulps TEMP(4), %xmm2 |
||
176 | mulps TEMP(4), %xmm3 |
||
177 | movaps %xmm0, %xmm5 |
||
178 | movaps %xmm1, %xmm7 |
||
179 | shufps $0x14, %xmm2, %xmm0 |
||
180 | shufps $0xbe, %xmm2, %xmm5 |
||
181 | shufps $0x14, %xmm3, %xmm1 |
||
182 | shufps $0xbe, %xmm3, %xmm7 |
||
183 | |||
184 | movaps %xmm0, TEMP(0) |
||
185 | movaps %xmm1, TEMP(1) |
||
186 | movaps %xmm5, TEMP(2) |
||
187 | movaps %xmm7, TEMP(3) |
||
188 | |||
189 | movss ASM_NAME(costab_mmxsse)+120, %xmm5 |
||
190 | shufps $0x00, %xmm5, %xmm5 |
||
191 | xorps pnpn, %xmm5 |
||
192 | |||
193 | movaps %xmm4, %xmm0 |
||
194 | movaps %xmm6, %xmm1 |
||
195 | unpcklps TEMP(5), %xmm4 |
||
196 | unpckhps TEMP(5), %xmm0 |
||
197 | unpcklps TEMP(7), %xmm6 |
||
198 | unpckhps TEMP(7), %xmm1 |
||
199 | movaps %xmm4, %xmm2 |
||
200 | movaps %xmm6, %xmm3 |
||
201 | unpcklps %xmm0, %xmm4 |
||
202 | unpckhps %xmm0, %xmm2 |
||
203 | unpcklps %xmm1, %xmm6 |
||
204 | unpckhps %xmm1, %xmm3 |
||
205 | movaps %xmm4, %xmm0 |
||
206 | movaps %xmm6, %xmm1 |
||
207 | subps %xmm2, %xmm0 |
||
208 | subps %xmm3, %xmm1 |
||
209 | addps %xmm2, %xmm4 |
||
210 | addps %xmm3, %xmm6 |
||
211 | mulps %xmm5, %xmm0 |
||
212 | mulps %xmm5, %xmm1 |
||
213 | movaps %xmm5, TEMP(5) |
||
214 | movaps %xmm4, %xmm5 |
||
215 | movaps %xmm6, %xmm7 |
||
216 | unpcklps %xmm0, %xmm4 |
||
217 | unpckhps %xmm0, %xmm5 |
||
218 | unpcklps %xmm1, %xmm6 |
||
219 | unpckhps %xmm1, %xmm7 |
||
220 | |||
221 | movaps TEMP(0), %xmm0 |
||
222 | movaps TEMP(2), %xmm2 |
||
223 | movaps %xmm4, TEMP(4) |
||
224 | movaps %xmm6, TEMP(6) |
||
225 | |||
226 | movaps %xmm0, %xmm4 |
||
227 | movaps %xmm2, %xmm6 |
||
228 | unpcklps TEMP(1), %xmm0 |
||
229 | unpckhps TEMP(1), %xmm4 |
||
230 | unpcklps TEMP(3), %xmm2 |
||
231 | unpckhps TEMP(3), %xmm6 |
||
232 | movaps %xmm0, %xmm1 |
||
233 | movaps %xmm2, %xmm3 |
||
234 | unpcklps %xmm4, %xmm0 |
||
235 | unpckhps %xmm4, %xmm1 |
||
236 | unpcklps %xmm6, %xmm2 |
||
237 | unpckhps %xmm6, %xmm3 |
||
238 | movaps %xmm0, %xmm4 |
||
239 | movaps %xmm2, %xmm6 |
||
240 | subps %xmm1, %xmm4 |
||
241 | subps %xmm3, %xmm6 |
||
242 | addps %xmm1, %xmm0 |
||
243 | addps %xmm3, %xmm2 |
||
244 | mulps TEMP(5), %xmm4 |
||
245 | mulps TEMP(5), %xmm6 |
||
246 | movaps %xmm0, %xmm1 |
||
247 | movaps %xmm2, %xmm3 |
||
248 | unpcklps %xmm4, %xmm0 |
||
249 | unpckhps %xmm4, %xmm1 |
||
250 | unpcklps %xmm6, %xmm2 |
||
251 | unpckhps %xmm6, %xmm3 |
||
252 | |||
253 | movaps %xmm0, TEMP(0) |
||
254 | movaps %xmm1, TEMP(1) |
||
255 | movaps %xmm2, TEMP(2) |
||
256 | movaps %xmm3, TEMP(3) |
||
257 | movaps %xmm5, TEMP(5) |
||
258 | movaps %xmm7, TEMP(7) |
||
259 | |||
260 | movss TEMP_BYTE(12), %xmm0 |
||
261 | movss TEMP_BYTE(28), %xmm1 |
||
262 | movss TEMP_BYTE(44), %xmm2 |
||
263 | movss TEMP_BYTE(60), %xmm3 |
||
264 | addss TEMP_BYTE(8), %xmm0 |
||
265 | addss TEMP_BYTE(24), %xmm1 |
||
266 | addss TEMP_BYTE(40), %xmm2 |
||
267 | addss TEMP_BYTE(56), %xmm3 |
||
268 | movss %xmm0, TEMP_BYTE(8) |
||
269 | movss %xmm1, TEMP_BYTE(24) |
||
270 | movss %xmm2, TEMP_BYTE(40) |
||
271 | movss %xmm3, TEMP_BYTE(56) |
||
272 | movss TEMP_BYTE(76), %xmm0 |
||
273 | movss TEMP_BYTE(92), %xmm1 |
||
274 | movss TEMP_BYTE(108), %xmm2 |
||
275 | movss TEMP_BYTE(124), %xmm3 |
||
276 | addss TEMP_BYTE(72), %xmm0 |
||
277 | addss TEMP_BYTE(88), %xmm1 |
||
278 | addss TEMP_BYTE(104), %xmm2 |
||
279 | addss TEMP_BYTE(120), %xmm3 |
||
280 | movss %xmm0, TEMP_BYTE(72) |
||
281 | movss %xmm1, TEMP_BYTE(88) |
||
282 | movss %xmm2, TEMP_BYTE(104) |
||
283 | movss %xmm3, TEMP_BYTE(120) |
||
284 | |||
285 | movaps TEMP_BYTE(16), %xmm1 |
||
286 | movaps TEMP_BYTE(48), %xmm3 |
||
287 | movaps TEMP_BYTE(80), %xmm5 |
||
288 | movaps TEMP_BYTE(112), %xmm7 |
||
289 | movaps %xmm1, %xmm0 |
||
290 | movaps %xmm3, %xmm2 |
||
291 | movaps %xmm5, %xmm4 |
||
292 | movaps %xmm7, %xmm6 |
||
293 | shufps $0x1e, %xmm0, %xmm0 |
||
294 | shufps $0x1e, %xmm2, %xmm2 |
||
295 | shufps $0x1e, %xmm4, %xmm4 |
||
296 | shufps $0x1e, %xmm6, %xmm6 |
||
297 | andps mask, %xmm0 |
||
298 | andps mask, %xmm2 |
||
299 | andps mask, %xmm4 |
||
300 | andps mask, %xmm6 |
||
301 | addps %xmm0, %xmm1 |
||
302 | addps %xmm2, %xmm3 |
||
303 | addps %xmm4, %xmm5 |
||
304 | addps %xmm6, %xmm7 |
||
305 | |||
306 | movaps TEMP_BYTE(32), %xmm2 |
||
307 | movaps TEMP_BYTE(96), %xmm6 |
||
308 | movaps %xmm2, %xmm0 |
||
309 | movaps %xmm6, %xmm4 |
||
310 | shufps $0x1e, %xmm0, %xmm0 |
||
311 | shufps $0x1e, %xmm4, %xmm4 |
||
312 | andps mask, %xmm0 |
||
313 | andps mask, %xmm4 |
||
314 | addps %xmm3, %xmm2 |
||
315 | addps %xmm0, %xmm3 |
||
316 | addps %xmm7, %xmm6 |
||
317 | addps %xmm4, %xmm7 |
||
318 | |||
319 | movaps TEMP_BYTE(0), %xmm0 |
||
320 | movaps TEMP_BYTE(64), %xmm4 |
||
321 | |||
322 | cvtps2pi %xmm0, %mm0 |
||
323 | cvtps2pi %xmm1, %mm1 |
||
324 | movhlps %xmm0, %xmm0 |
||
325 | movhlps %xmm1, %xmm1 |
||
326 | cvtps2pi %xmm0, %mm2 |
||
327 | cvtps2pi %xmm1, %mm3 |
||
328 | packssdw %mm2, %mm0 |
||
329 | packssdw %mm3, %mm1 |
||
330 | |||
331 | cvtps2pi %xmm2, %mm2 |
||
332 | cvtps2pi %xmm3, %mm3 |
||
333 | movhlps %xmm2, %xmm2 |
||
334 | movhlps %xmm3, %xmm3 |
||
335 | cvtps2pi %xmm2, %mm4 |
||
336 | cvtps2pi %xmm3, %mm5 |
||
337 | packssdw %mm4, %mm2 |
||
338 | packssdw %mm5, %mm3 |
||
339 | |||
340 | movd %mm0, %eax |
||
341 | movd %mm1, %edx |
||
342 | movw %ax, 512(%ecx) |
||
343 | movw %dx, 384(%ecx) |
||
344 | shrl $16, %eax |
||
345 | shrl $16, %edx |
||
346 | movw %ax, (%ecx) |
||
347 | movw %ax, (%ebx) |
||
348 | movw %dx, 128(%ebx) |
||
349 | |||
350 | movd %mm2, %eax |
||
351 | movd %mm3, %edx |
||
352 | movw %ax, 448(%ecx) |
||
353 | movw %dx, 320(%ecx) |
||
354 | shrl $16, %eax |
||
355 | shrl $16, %edx |
||
356 | movw %ax, 64(%ebx) |
||
357 | movw %dx, 192(%ebx) |
||
358 | |||
359 | psrlq $32, %mm0 |
||
360 | psrlq $32, %mm1 |
||
361 | movd %mm0, %eax |
||
362 | movd %mm1, %edx |
||
363 | movw %ax, 256(%ecx) |
||
364 | movw %dx, 128(%ecx) |
||
365 | shrl $16, %eax |
||
366 | shrl $16, %edx |
||
367 | movw %ax, 256(%ebx) |
||
368 | movw %dx, 384(%ebx) |
||
369 | |||
370 | psrlq $32, %mm2 |
||
371 | psrlq $32, %mm3 |
||
372 | movd %mm2, %eax |
||
373 | movd %mm3, %edx |
||
374 | movw %ax, 192(%ecx) |
||
375 | movw %dx, 64(%ecx) |
||
376 | shrl $16, %eax |
||
377 | shrl $16, %edx |
||
378 | movw %ax, 320(%ebx) |
||
379 | movw %dx, 448(%ebx) |
||
380 | |||
381 | movaps %xmm4, %xmm0 |
||
382 | shufps $0x1e, %xmm0, %xmm0 |
||
383 | movaps %xmm5, %xmm1 |
||
384 | andps mask, %xmm0 |
||
385 | |||
386 | addps %xmm6, %xmm4 |
||
387 | addps %xmm7, %xmm5 |
||
388 | addps %xmm1, %xmm6 |
||
389 | addps %xmm0, %xmm7 |
||
390 | |||
391 | cvtps2pi %xmm4, %mm0 |
||
392 | cvtps2pi %xmm5, %mm1 |
||
393 | movhlps %xmm4, %xmm4 |
||
394 | movhlps %xmm5, %xmm5 |
||
395 | cvtps2pi %xmm4, %mm2 |
||
396 | cvtps2pi %xmm5, %mm3 |
||
397 | packssdw %mm2, %mm0 |
||
398 | packssdw %mm3, %mm1 |
||
399 | |||
400 | cvtps2pi %xmm6, %mm2 |
||
401 | cvtps2pi %xmm7, %mm3 |
||
402 | movhlps %xmm6, %xmm6 |
||
403 | movhlps %xmm7, %xmm7 |
||
404 | cvtps2pi %xmm6, %mm4 |
||
405 | cvtps2pi %xmm7, %mm5 |
||
406 | packssdw %mm4, %mm2 |
||
407 | packssdw %mm5, %mm3 |
||
408 | |||
409 | movd %mm0, %eax |
||
410 | movd %mm2, %edx |
||
411 | movw %ax, 480(%ecx) |
||
412 | movw %dx, 416(%ecx) |
||
413 | shrl $16, %eax |
||
414 | shrl $16, %edx |
||
415 | movw %ax, 32(%ebx) |
||
416 | movw %dx, 96(%ebx) |
||
417 | |||
418 | psrlq $32, %mm0 |
||
419 | psrlq $32, %mm2 |
||
420 | movd %mm0, %eax |
||
421 | movd %mm2, %edx |
||
422 | movw %ax, 224(%ecx) |
||
423 | movw %dx, 160(%ecx) |
||
424 | shrl $16, %eax |
||
425 | shrl $16, %edx |
||
426 | movw %ax, 288(%ebx) |
||
427 | movw %dx, 352(%ebx) |
||
428 | |||
429 | movd %mm1, %eax |
||
430 | movd %mm3, %edx |
||
431 | movw %ax, 352(%ecx) |
||
432 | movw %dx, 288(%ecx) |
||
433 | shrl $16, %eax |
||
434 | shrl $16, %edx |
||
435 | movw %ax, 160(%ebx) |
||
436 | movw %dx, 224(%ebx) |
||
437 | |||
438 | psrlq $32, %mm1 |
||
439 | psrlq $32, %mm3 |
||
440 | movd %mm1, %eax |
||
441 | movd %mm3, %edx |
||
442 | movw %ax, 96(%ecx) |
||
443 | movw %dx, 32(%ecx) |
||
444 | shrl $16, %eax |
||
445 | shrl $16, %edx |
||
446 | movw %ax, 416(%ebx) |
||
447 | movw %dx, 480(%ebx) |
||
448 | |||
449 | popl %ebx |
||
450 | movl %ebp, %esp |
||
451 | popl %ebp |
||
452 | ret |
||
453 | |||
454 | NONEXEC_STACK |