1 /*****************************************************************************
2 * ac3_srfft_sse.c: accelerated SSE ac3 fft functions
3 *****************************************************************************
4 * Copyright (C) 1999-2001 VideoLAN
5 * $Id: ac3_srfft_sse.c,v 1.14 2002/07/31 20:56:51 sam Exp $
7 * Authors: Renaud Dartus <reno@videolan.org>
8 * Aaron Holtzman <aholtzma@engr.uvic.ca>
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA.
23 *****************************************************************************/
25 /*****************************************************************************
27 *****************************************************************************/
33 #include "ac3_imdct.h"
34 #include "ac3_srfft.h"
37 static float hsqrt2_sse[] ATTR_ALIGN(16) =
38 { 0.707106781188, 0.707106781188, -0.707106781188, -0.707106781188 };
40 static float C_1_sse[] ATTR_ALIGN(16) =
41 { -1.0, 1.0, -1.0, 1.0 };
49 static void fft_4_sse (complex_t *x);
50 static void fft_8_sse (complex_t *x);
51 static void fft_asmb_sse (ck_sse_t * ck, int k, complex_t *x, complex_t *wTB,
52 const complex_t *d, const complex_t *d_3);
54 void E_( fft_64p ) ( complex_t *a )
59 fft_8_sse(&a[0]); fft_4_sse(&a[8]); fft_4_sse(&a[12]);
60 fft_asmb_sse(&ck, 2, &a[0], &a[8], &delta16[0], &delta16_3[0]);
62 fft_8_sse(&a[16]), fft_8_sse(&a[24]);
63 fft_asmb_sse(&ck, 4, &a[0], &a[16],&delta32[0], &delta32_3[0]);
65 fft_8_sse(&a[32]); fft_4_sse(&a[40]); fft_4_sse(&a[44]);
66 fft_asmb_sse(&ck, 2, &a[32], &a[40], &delta16[0], &delta16_3[0]);
68 fft_8_sse(&a[48]); fft_4_sse(&a[56]); fft_4_sse(&a[60]);
69 fft_asmb_sse(&ck, 2, &a[48], &a[56], &delta16[0], &delta16_3[0]);
71 fft_asmb_sse(&ck, 8, &a[0], &a[32],&delta64[0], &delta64_3[0]);
74 void E_( fft_128p ) ( complex_t *a )
79 fft_8_sse(&a[0]); fft_4_sse(&a[8]); fft_4_sse(&a[12]);
80 fft_asmb_sse(&ck, 2, &a[0], &a[8], &delta16[0], &delta16_3[0]);
82 fft_8_sse(&a[16]), fft_8_sse(&a[24]);
83 fft_asmb_sse(&ck, 4, &a[0], &a[16],&delta32[0], &delta32_3[0]);
85 fft_8_sse(&a[32]); fft_4_sse(&a[40]); fft_4_sse(&a[44]);
86 fft_asmb_sse(&ck, 2, &a[32], &a[40], &delta16[0], &delta16_3[0]);
88 fft_8_sse(&a[48]); fft_4_sse(&a[56]); fft_4_sse(&a[60]);
89 fft_asmb_sse(&ck, 2, &a[48], &a[56], &delta16[0], &delta16_3[0]);
91 fft_asmb_sse(&ck, 8, &a[0], &a[32],&delta64[0], &delta64_3[0]);
93 fft_8_sse(&a[64]); fft_4_sse(&a[72]); fft_4_sse(&a[76]);
95 fft_asmb_sse(&ck, 2, &a[64], &a[72], &delta16[0], &delta16_3[0]);
97 fft_8_sse(&a[80]); fft_8_sse(&a[88]);
100 fft_asmb_sse(&ck, 4, &a[64], &a[80],&delta32[0], &delta32_3[0]);
102 fft_8_sse(&a[96]); fft_4_sse(&a[104]), fft_4_sse(&a[108]);
103 /* fft_16(&a[96]); */
104 fft_asmb_sse(&ck, 2, &a[96], &a[104], &delta16[0], &delta16_3[0]);
106 fft_8_sse(&a[112]), fft_8_sse(&a[120]);
107 /* fft_32(&a[96]); */
108 fft_asmb_sse(&ck, 4, &a[96], &a[112], &delta32[0], &delta32_3[0]);
110 /* fft_128(&a[0]); */
111 fft_asmb_sse(&ck, 16, &a[0], &a[64], &delta128[0], &delta128_3[0]);
114 static void fft_4_sse (complex_t *x)
116 __asm__ __volatile__ (
118 "movaps (%%eax), %%xmm0\n" /* x[1] | x[0] */
119 "movaps 16(%%eax), %%xmm2\n" /* x[3] | x[2] */
120 "movaps %%xmm0, %%xmm1\n" /* x[1] | x[0] */
121 "addps %%xmm2, %%xmm0\n" /* x[1] + x[3] | x[0] + x[2] */
122 "subps %%xmm2, %%xmm1\n" /* x[1] - x[3] | x[0] - x[2] */
123 "xorps %%xmm6, %%xmm6\n"
124 "movhlps %%xmm1, %%xmm4\n" /* ? | x[1] - x[3] */
125 "movhlps %%xmm0, %%xmm3\n" /* ? | x[1] + x[3] */
126 "subss %%xmm4, %%xmm6\n" /* 0 | -(x[1] - x[3]).re */
127 "movlhps %%xmm1, %%xmm0\n" /* x[0] - x[2] | x[0] + x[2] */
128 "movlhps %%xmm6, %%xmm4\n" /* 0 | -(x[1] - x[3]).re | (x[1] - x[3]).im | (x[3]-x[1]).re */
129 "movaps %%xmm0, %%xmm2\n" /* x[0] - x[2] | x[0] + x[2] */
130 "shufps $0x94, %%xmm4, %%xmm3\n" /* i*(x[1] - x[3]) | x[1] + x[3] */
131 "addps %%xmm3, %%xmm0\n"
132 "subps %%xmm3, %%xmm2\n"
133 "movaps %%xmm0, (%%eax)\n"
134 "movaps %%xmm2, 16(%%eax)\n"
139 static void fft_8_sse (complex_t *x)
141 __asm__ __volatile__ (
144 "movlps (%%eax), %%xmm0\n" /* x[0] */
145 "movlps 32(%%eax), %%xmm1\n" /* x[4] */
146 "movhps 16(%%eax), %%xmm0\n" /* x[2] | x[0] */
147 "movhps 48(%%eax), %%xmm1\n" /* x[6] | x[4] */
148 "movaps %%xmm0, %%xmm2\n" /* x[2] | x[0] */
149 "xorps %%xmm3, %%xmm3\n"
150 "addps %%xmm1, %%xmm0\n" /* x[2] + x[6] | x[0] + x[4] */
151 "subps %%xmm1, %%xmm2\n" /* x[2] - x[6] | x[0] - x[4] */
152 "movhlps %%xmm0, %%xmm5\n" /* x[2] + x[6] */
153 "movhlps %%xmm2, %%xmm4\n" /* x[2] - x[6] */
154 "movlhps %%xmm2, %%xmm0\n" /* x[0] - x[4] | x[0] + x[4] */
155 "subss %%xmm4, %%xmm3\n" /* (x[2]-x[6]).im | -(x[2]-x[6]).re */
156 "movaps %%xmm0, %%xmm7\n" /* x[0] - x[4] | x[0] + x[4] */
157 "movaps %%xmm3, %%xmm4\n" /* (x[2]-x[6]).im | -(x[2]-x[6]).re */
158 "movlps 8(%%eax), %%xmm1\n" /* x[1] */
159 "shufps $0x14, %%xmm4, %%xmm5\n" /* i*(x[2] - x[6]) | x[2] + x[6] */
161 "addps %%xmm5, %%xmm0\n" /* yt = i*(x2-x6)+x0-x4 | x2+x6+x0+x4 */
162 "subps %%xmm5, %%xmm7\n" /* yb = i*(x6-x2)+x0-x4 | -x6-x2+x0+x4 */
164 "movhps 24(%%eax), %%xmm1\n" /* x[3] | x[1] */
165 "movlps 40(%%eax), %%xmm2\n" /* x[5] */
166 "movhps 56(%%eax), %%xmm2\n" /* x[7] | x[5] */
167 "movaps %%xmm1, %%xmm3\n" /* x[3] | x[1] */
168 "addps %%xmm2, %%xmm1\n" /* x[3] + x[7] | x[1] + x[5] */
169 "subps %%xmm2, %%xmm3\n" /* x[3] - x[7] | x[1] - x[5] */
170 "movaps (%%ecx), %%xmm4\n" /* -1/sqrt2 | -1/sqrt2 | 1/sqrt2 | 1/sqrt2 */
171 "movaps %%xmm3, %%xmm6\n" /* x[3] - x[7] | x[1] - x[5] */
172 "mulps %%xmm4, %%xmm3\n" /* -1/s2*(x[3] - x[7]) | 1/s2*(x[1] - x[5]) */
173 "shufps $0xc8, %%xmm4, %%xmm4\n" /* -1/sqrt2 | 1/sqrt2 | -1/sqrt2 | 1/sqrt2 */
174 "shufps $0xb1, %%xmm6, %%xmm6\n" /* (x3-x7).re|(x3-x7).im|(x1-x5).re|(x1-x5).im */
175 "mulps %%xmm4, %%xmm6\n" /* (x7-x3).re/s2|(x3-x7).im/s2|(x5-x1).re/s2|(x1-x5).im/s2 */
176 "addps %%xmm3, %%xmm6\n" /* (-1-i)/sqrt2 * (x[3]-x[7]) | (1-i)/sqrt2 * (x[1] - x[5]) */
177 "movhlps %%xmm1, %%xmm5\n" /* x[3] + x[7] */
178 "movlhps %%xmm6, %%xmm1\n" /* (1+i)/sqrt2 * (x[1]-x[5]) | x[1]+x[5] */
179 "shufps $0xe4, %%xmm6, %%xmm5\n" /* (-1-i)/sqrt2 * (x[3]-x[7]) | x[3]+x[7] */
180 "movaps %%xmm1, %%xmm3\n" /* (1-i)/sqrt2 * (x[1]-x[5]) | x[1]+x[5] */
181 "addps %%xmm5, %%xmm1\n" /* u */
182 "subps %%xmm5, %%xmm3\n" /* v */
183 "movaps %%xmm0, %%xmm2\n" /* yb */
184 "movaps %%xmm7, %%xmm4\n" /* yt */
185 "movaps (%%edx), %%xmm5\n"
186 "mulps %%xmm5, %%xmm3\n"
187 "addps %%xmm1, %%xmm0\n" /* yt + u */
188 "subps %%xmm1, %%xmm2\n" /* yt - u */
189 "shufps $0xb1, %%xmm3, %%xmm3\n" /* -i * v */
190 "movaps %%xmm0, (%%eax)\n"
191 "movaps %%xmm2, 32(%%eax)\n"
192 "addps %%xmm3, %%xmm4\n" /* yb - i*v */
193 "subps %%xmm3, %%xmm7\n" /* yb + i*v */
194 "movaps %%xmm4, 16(%%eax)\n"
195 "movaps %%xmm7, 48(%%eax)\n"
198 : "a" (x), "c" (hsqrt2_sse), "d" (C_1_sse));
201 static void fft_asmb_sse (ck_sse_t * ck, int k, complex_t *x, complex_t *wTB,
202 const complex_t *d, const complex_t *d_3)
206 __asm__ __volatile__ (
209 "movl %%esp, %%ebp\n"
220 "movl 4(%%ecx), %%ebx\n"
221 "movl %%ebx, -4(%%ebp)\n"
222 "movl (%%ecx), %%ecx\n"
224 "movl %%ecx, -8(%%ebp)\n" /* k */
227 "shll $4, %%ecx\n" /* 16k */
229 /* TRANSZERO and TRANS */
231 "movaps (%%eax), %%xmm0\n" /* x[1] | x[0] */
232 "movaps (%%edi), %%xmm1\n" /* wT[1] | wT[0] */
233 "movaps (%%edi, %%ecx), %%xmm2\n" /* wB[1] | wB[0] */
234 "movlps (%%edx), %%xmm3\n" /* d */
235 "movlps (%%esi), %%xmm4\n" /* d3 */
236 "movhlps %%xmm1, %%xmm5\n" /* wT[1] */
237 "movhlps %%xmm2, %%xmm6\n" /* wB[1] */
238 "shufps $0x50, %%xmm3, %%xmm3\n" /* d[1].im | d[1].im | d[1].re | d[1].re */
239 "shufps $0x50, %%xmm4, %%xmm4\n" /* d3[1].im | d3[1].im | d3[i].re | d3[i].re */
240 "movlhps %%xmm5, %%xmm5\n" /* wT[1] | wT[1] */
241 "movlhps %%xmm6, %%xmm6\n" /* wB[1] | wB[1] */
242 "mulps %%xmm3, %%xmm5\n"
243 "mulps %%xmm4, %%xmm6\n"
244 "movhlps %%xmm5, %%xmm7\n" /* wT[1].im * d[1].im | wT[1].re * d[1].im */
245 "movlhps %%xmm6, %%xmm5\n" /* wB[1].im * d3[1].re | wB[1].re * d3[1].re | wT[1].im * d[1].re | wT[1].re * d[1].re */
246 "shufps $0xb1, %%xmm6, %%xmm7\n" /* wB[1].re * d3[1].im | wB[i].im * d3[1].im | wT[1].re * d[1].im | wT[1].im * d[1].im */
247 "movl -4(%%ebp), %%ebx\n"
248 "movaps (%%ebx), %%xmm4\n"
249 "mulps %%xmm4, %%xmm7\n"
250 "addps %%xmm7, %%xmm5\n" /* wB[1] * d3[1] | wT[1] * d[1] */
251 "movlhps %%xmm5, %%xmm1\n" /* d[1] * wT[1] | wT[0] */
252 "shufps $0xe4, %%xmm5, %%xmm2\n" /* d3[1] * wB[1] | wB[0] */
253 "movaps %%xmm1, %%xmm3\n" /* d[1] * wT[1] | wT[0] */
254 "leal (%%eax, %%ecx, 2), %%ebx\n"
255 "addps %%xmm2, %%xmm1\n" /* u */
256 "subps %%xmm2, %%xmm3\n" /* v */
257 "mulps %%xmm4, %%xmm3\n"
258 "movaps (%%eax, %%ecx), %%xmm5\n" /* xk[1] | xk[0] */
259 "shufps $0xb1, %%xmm3, %%xmm3\n" /* -i * v */
260 "movaps %%xmm0, %%xmm2\n" /* x[1] | x[0] */
261 "movaps %%xmm5, %%xmm6\n" /* xk[1] | xk[0] */
262 "addps %%xmm1, %%xmm0\n"
263 "subps %%xmm1, %%xmm2\n"
264 "addps %%xmm3, %%xmm5\n"
265 "subps %%xmm3, %%xmm6\n"
266 "movaps %%xmm0, (%%eax)\n"
267 "movaps %%xmm2, (%%ebx)\n"
268 "movaps %%xmm5, (%%eax, %%ecx)\n"
269 "movaps %%xmm6, (%%ebx, %%ecx)\n"
278 "movaps (%%edi), %%xmm0\n" /* wT[1] | wT[0] */
279 "movaps (%%edx), %%xmm1\n" /* d[1] | d[0] */
281 "movaps (%%edi, %%ecx), %%xmm4\n" /* wB[1] | wB[0] */
282 "movaps (%%esi), %%xmm5\n" /* d3[1] | d3[0] */
284 "movhlps %%xmm0, %%xmm2\n" /* wT[1] */
285 "movhlps %%xmm1, %%xmm3\n" /* d[1] */
287 "movhlps %%xmm4, %%xmm6\n" /* wB[1] */
288 "movhlps %%xmm5, %%xmm7\n" /* d3[1] */
290 "shufps $0x50, %%xmm1, %%xmm1\n" /* d[0].im | d[0].im | d[0].re | d[0].re */
291 "shufps $0x50, %%xmm3, %%xmm3\n" /* d[1].im | d[1].im | d[1].re | d[1].re */
293 "movlhps %%xmm0, %%xmm0\n" /* wT[0] | wT[0] */
294 "shufps $0x50, %%xmm5, %%xmm5\n" /* d3[0].im | d3[0].im | d3[0].re | d3[0].re */
295 "movlhps %%xmm2, %%xmm2\n" /* wT[1] | wT[1] */
296 "shufps $0x50, %%xmm7, %%xmm7\n" /* d3[1].im | d3[1].im | d3[1].re | d3[1].re */
298 "mulps %%xmm1, %%xmm0\n" /* d[0].im * wT[0].im | d[0].im * wT[0].re | d[0].re * wT[0].im | d[0].re * wT[0].re */
299 "mulps %%xmm3, %%xmm2\n" /* d[1].im * wT[1].im | d[1].im * wT[1].re | d[1].re * wT[1].im | d[1].re * wT[1].re */
300 "movlhps %%xmm4, %%xmm4\n" /* wB[0] | wB[0] */
301 "movlhps %%xmm6, %%xmm6\n" /* wB[1] | wB[1] */
303 "movhlps %%xmm0, %%xmm1\n" /* d[0].im * wT[0].im | d[0].im * wT[0].re */
304 "movlhps %%xmm2, %%xmm0\n" /* d[1].re * wT[1].im | d[1].re * wT[1].re | d[0].re * wT[0].im | d[0].re * wT[0].re */
305 "mulps %%xmm5, %%xmm4\n" /* wB[0].im * d3[0].im | wB[0].re * d3[0].im | wB[0].im * d3[0].re | wB[0].re * d3[0].re */
306 "mulps %%xmm7, %%xmm6\n" /* wB[1].im * d3[1].im | wB[1].re * d3[1].im | wB[1].im * d3[1].re | wB[1].re * d3[1].re */
307 "shufps $0xb1, %%xmm2, %%xmm1\n" /* d[1].im * wT[1].re | d[1].im * wT[1].im | d[0].im * wT[0].re | d[0].im * wT[0].im */
308 "movl -4(%%ebp), %%ebx\n"
309 "movaps (%%ebx), %%xmm3\n" /* 1.0 | -1.0 | 1.0 | -1.0 */
311 "movhlps %%xmm4, %%xmm5\n" /* wB[0].im * d3[0].im | wB[0].re * d3[0].im */
312 "mulps %%xmm3, %%xmm1\n" /* d[1].im * wT[1].re | -d[1].im * wT[1].im | d[0].im * wT[0].re | -d[0].im * wT[0].im */
313 "movlhps %%xmm6, %%xmm4\n" /* wB[1].im * d3[1].re | wB[1].re * d3[1].re | wB[0].im * d3[0].re | wB[0].im * d3[0].re */
314 "addps %%xmm1, %%xmm0\n" /* wT[1] * d[1] | wT[0] * d[0] */
316 "shufps $0xb1, %%xmm6, %%xmm5\n" /* wB[1].re * d3[1].im | wB[1].im * d3[1].im | wB[0].re * d3[0].im | wB[0].im * d3[0].im */
317 "mulps %%xmm3, %%xmm5\n" /* wB[1].re * d3[1].im | -wB[1].im * d3[1].im | wB[0].re * d3[0].im | -wB[0].im * d3[0].im */
318 "addps %%xmm5, %%xmm4\n" /* wB[1] * d3[1] | wB[0] * d3[0] */
320 "movaps %%xmm0, %%xmm1\n" /* wT[1] * d[1] | wT[0] * d[0] */
321 "addps %%xmm4, %%xmm0\n" /* u */
322 "subps %%xmm4, %%xmm1\n" /* v */
323 "movaps (%%eax), %%xmm6\n" /* x[1] | x[0] */
324 "leal (%%eax, %%ecx, 2), %%ebx\n"
325 "mulps %%xmm3, %%xmm1\n"
328 "shufps $0xb1, %%xmm1, %%xmm1\n" /* -i * v */
329 "movaps (%%eax, %%ecx), %%xmm7\n" /* xk[1] | xk[0] */
330 "movaps %%xmm6, %%xmm2\n"
331 "movaps %%xmm7, %%xmm4\n"
332 "addps %%xmm0, %%xmm6\n"
333 "subps %%xmm0, %%xmm2\n"
334 "movaps %%xmm6, (%%eax)\n"
335 "movaps %%xmm2, (%%ebx)\n"
336 "addps %%xmm1, %%xmm7\n"
337 "subps %%xmm1, %%xmm4\n"
339 "movaps %%xmm7, (%%eax, %%ecx)\n"
340 "movaps %%xmm4, (%%ebx, %%ecx)\n"
358 : "=a" (x), "=D" (wTB)
359 : "c" (ck), "a" (x), "D" (wTB), "d" (d), "S" (d_3) );