1 /*
2 dct64_altivec.c: Discrete Cosine Tansform (DCT) for Altivec
3
4 copyright ?-2006 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 Michael Hipp
7 altivec optimization by tmkk
8 */
9
10 /*
11 * Discrete Cosine Tansform (DCT) for subband synthesis
12 *
13 * -funroll-loops (for gcc) will remove the loops for better performance
14 * using loops in the source-code enhances readabillity
15 *
16 *
17 * TODO: write an optimized version for the down-sampling modes
18 * (in these modes the bands 16-31 (2:1) or 8-31 (4:1) are zero
19 */
20
21 #include "mpg123lib_intern.h"
22
23 #ifndef __APPLE__
24 #include <altivec.h>
25 #endif
26
dct64_altivec(real * out0,real * out1,real * samples)27 void dct64_altivec(real *out0,real *out1,real *samples)
28 {
29 ALIGNED(16) real bufs[32];
30
31 {
32 register real *b1,*costab;
33
34 vector unsigned char vinvert,vperm1,vperm2,vperm3,vperm4;
35 vector float v1,v2,v3,v4,v5,v6,v7,v8;
36 vector float vbs1,vbs2,vbs3,vbs4,vbs5,vbs6,vbs7,vbs8;
37 vector float vbs9,vbs10,vbs11,vbs12,vbs13,vbs14,vbs15,vbs16;
38 vector float vzero;
39 b1 = samples;
40 costab = pnts[0];
41
42 vzero = vec_xor(vzero,vzero);
43 #ifdef __APPLE__
44 vinvert = (vector unsigned char)(12,13,14,15,8,9,10,11,4,5,6,7,0,1,2,3);
45 #else
46 vinvert = (vector unsigned char){12,13,14,15,8,9,10,11,4,5,6,7,0,1,2,3};
47 #endif
48 vperm1 = vec_lvsl(0,b1);
49 vperm2 = vec_perm(vperm1,vperm1,vinvert);
50
51 v1 = vec_ld(0,b1);
52 v2 = vec_ld(16,b1);
53 v3 = vec_ld(112,b1);
54 v4 = vec_ld(127,b1);
55 v5 = vec_perm(v1,v2,vperm1); /* b1[0,1,2,3] */
56 v6 = vec_perm(v3,v4,vperm2); /* b1[31,30,29,28] */
57
58 vbs1 = vec_add(v5,v6);
59 vbs8 = vec_sub(v5,v6);
60
61 v1 = vec_ld(32,b1);
62 v4 = vec_ld(96,b1);
63 v5 = vec_perm(v2,v1,vperm1); /* b1[4,5,6,7] */
64 v6 = vec_perm(v4,v3,vperm2); /* b1[27,26,25,24] */
65
66 vbs2 = vec_add(v5,v6);
67 vbs7 = vec_sub(v5,v6);
68
69 v2 = vec_ld(48,b1);
70 v3 = vec_ld(80,b1);
71 v5 = vec_perm(v1,v2,vperm1); /* b1[8,9,10,11] */
72 v6 = vec_perm(v3,v4,vperm2); /* b1[23,22,21,20] */
73
74 vbs3 = vec_add(v5,v6);
75 vbs6 = vec_sub(v5,v6);
76
77 v1 = vec_ld(64,b1);
78 v5 = vec_perm(v2,v1,vperm1); /* b1[12,13,14,15] */
79 v6 = vec_perm(v1,v3,vperm2); /* b1[19,18,17,16] */
80
81 vbs4 = vec_add(v5,v6);
82 vbs5 = vec_sub(v5,v6);
83
84 v1 = vec_ld(0,costab);
85 vbs8 = vec_madd(vbs8,v1,vzero);
86 v2 = vec_ld(16,costab);
87 vbs7 = vec_madd(vbs7,v2,vzero);
88 v3 = vec_ld(32,costab);
89 vbs6 = vec_madd(vbs6,v3,vzero);
90 v4 = vec_ld(48,costab);
91 vbs5 = vec_madd(vbs5,v4,vzero);
92 vbs6 = vec_perm(vbs6,vbs6,vinvert);
93 vbs5 = vec_perm(vbs5,vbs5,vinvert);
94
95
96 costab = pnts[1];
97
98 v1 = vec_perm(vbs4,vbs4,vinvert);
99 vbs9 = vec_add(vbs1,v1);
100 v3 = vec_sub(vbs1,v1);
101 v5 = vec_ld(0,costab);
102 v2 = vec_perm(vbs3,vbs3,vinvert);
103 vbs10 = vec_add(vbs2,v2);
104 v4 = vec_sub(vbs2,v2);
105 v6 = vec_ld(16,costab);
106 vbs12 = vec_madd(v3,v5,vzero);
107 vbs11 = vec_madd(v4,v6,vzero);
108
109 v7 = vec_sub(vbs7,vbs6);
110 v8 = vec_sub(vbs8,vbs5);
111 vbs13 = vec_add(vbs5,vbs8);
112 vbs14 = vec_add(vbs6,vbs7);
113 vbs15 = vec_madd(v7,v6,vzero);
114 vbs16 = vec_madd(v8,v5,vzero);
115
116
117 costab = pnts[2];
118
119 v1 = vec_perm(vbs10,vbs10,vinvert);
120 v5 = vec_perm(vbs14,vbs14,vinvert);
121 vbs1 = vec_add(v1,vbs9);
122 vbs5 = vec_add(v5,vbs13);
123 v2 = vec_sub(vbs9,v1);
124 v6 = vec_sub(vbs13,v5);
125 v3 = vec_ld(0,costab);
126 vbs11 = vec_perm(vbs11,vbs11,vinvert);
127 vbs15 = vec_perm(vbs15,vbs15,vinvert);
128 vbs3 = vec_add(vbs11,vbs12);
129 vbs7 = vec_add(vbs15,vbs16);
130 v4 = vec_sub(vbs12,vbs11);
131 v7 = vec_sub(vbs16,vbs15);
132 vbs2 = vec_madd(v2,v3,vzero);
133 vbs4 = vec_madd(v4,v3,vzero);
134 vbs6 = vec_madd(v6,v3,vzero);
135 vbs8 = vec_madd(v7,v3,vzero);
136
137 vbs2 = vec_perm(vbs2,vbs2,vinvert);
138 vbs4 = vec_perm(vbs4,vbs4,vinvert);
139 vbs6 = vec_perm(vbs6,vbs6,vinvert);
140 vbs8 = vec_perm(vbs8,vbs8,vinvert);
141
142
143 costab = pnts[3];
144
145 #ifdef __APPLE__
146 vperm1 = (vector unsigned char)(0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23);
147 vperm2 = (vector unsigned char)(12,13,14,15,8,9,10,11,28,29,30,31,24,25,26,27);
148 vperm3 = (vector unsigned char)(0,1,2,3,4,5,6,7,20,21,22,23,16,17,18,19);
149 #else
150 vperm1 = (vector unsigned char){0,1,2,3,4,5,6,7,16,17,18,19,20,21,22,23};
151 vperm2 = (vector unsigned char){12,13,14,15,8,9,10,11,28,29,30,31,24,25,26,27};
152 vperm3 = (vector unsigned char){0,1,2,3,4,5,6,7,20,21,22,23,16,17,18,19};
153 #endif
154 vperm4 = vec_add(vperm3,vec_splat_u8(8));
155
156 v1 = vec_ld(0,costab);
157 v2 = vec_splat(v1,0);
158 v3 = vec_splat(v1,1);
159 v1 = vec_mergeh(v2,v3);
160
161 v2 = vec_perm(vbs1,vbs3,vperm1);
162 v3 = vec_perm(vbs2,vbs4,vperm1);
163 v4 = vec_perm(vbs1,vbs3,vperm2);
164 v5 = vec_perm(vbs2,vbs4,vperm2);
165 v6 = vec_sub(v2,v4);
166 v7 = vec_sub(v3,v5);
167 v2 = vec_add(v2,v4);
168 v3 = vec_add(v3,v5);
169 v4 = vec_madd(v6,v1,vzero);
170 v5 = vec_nmsub(v7,v1,vzero);
171 vbs9 = vec_perm(v2,v4,vperm3);
172 vbs11 = vec_perm(v2,v4,vperm4);
173 vbs10 = vec_perm(v3,v5,vperm3);
174 vbs12 = vec_perm(v3,v5,vperm4);
175
176 v2 = vec_perm(vbs5,vbs7,vperm1);
177 v3 = vec_perm(vbs6,vbs8,vperm1);
178 v4 = vec_perm(vbs5,vbs7,vperm2);
179 v5 = vec_perm(vbs6,vbs8,vperm2);
180 v6 = vec_sub(v2,v4);
181 v7 = vec_sub(v3,v5);
182 v2 = vec_add(v2,v4);
183 v3 = vec_add(v3,v5);
184 v4 = vec_madd(v6,v1,vzero);
185 v5 = vec_nmsub(v7,v1,vzero);
186 vbs13 = vec_perm(v2,v4,vperm3);
187 vbs15 = vec_perm(v2,v4,vperm4);
188 vbs14 = vec_perm(v3,v5,vperm3);
189 vbs16 = vec_perm(v3,v5,vperm4);
190
191
192 costab = pnts[4];
193
194 v1 = vec_lde(0,costab);
195 #ifdef __APPLE__
196 v2 = (vector float)(1.0f,-1.0f,1.0f,-1.0f);
197 #else
198 v2 = (vector float){1.0f,-1.0f,1.0f,-1.0f};
199 #endif
200 v3 = vec_splat(v1,0);
201 v1 = vec_madd(v2,v3,vzero);
202
203 v2 = vec_mergeh(vbs9,vbs10);
204 v3 = vec_mergel(vbs9,vbs10);
205 v4 = vec_mergeh(vbs11,vbs12);
206 v5 = vec_mergel(vbs11,vbs12);
207 v6 = vec_mergeh(v2,v3);
208 v7 = vec_mergel(v2,v3);
209 v2 = vec_mergeh(v4,v5);
210 v3 = vec_mergel(v4,v5);
211 v4 = vec_sub(v6,v7);
212 v5 = vec_sub(v2,v3);
213 v6 = vec_add(v6,v7);
214 v7 = vec_add(v2,v3);
215 v2 = vec_madd(v4,v1,vzero);
216 v3 = vec_madd(v5,v1,vzero);
217 vbs1 = vec_mergeh(v6,v2);
218 vbs2 = vec_mergel(v6,v2);
219 vbs3 = vec_mergeh(v7,v3);
220 vbs4 = vec_mergel(v7,v3);
221
222 v2 = vec_mergeh(vbs13,vbs14);
223 v3 = vec_mergel(vbs13,vbs14);
224 v4 = vec_mergeh(vbs15,vbs16);
225 v5 = vec_mergel(vbs15,vbs16);
226 v6 = vec_mergeh(v2,v3);
227 v7 = vec_mergel(v2,v3);
228 v2 = vec_mergeh(v4,v5);
229 v3 = vec_mergel(v4,v5);
230 v4 = vec_sub(v6,v7);
231 v5 = vec_sub(v2,v3);
232 v6 = vec_add(v6,v7);
233 v7 = vec_add(v2,v3);
234 v2 = vec_madd(v4,v1,vzero);
235 v3 = vec_madd(v5,v1,vzero);
236 vbs5 = vec_mergeh(v6,v2);
237 vbs6 = vec_mergel(v6,v2);
238 vbs7 = vec_mergeh(v7,v3);
239 vbs8 = vec_mergel(v7,v3);
240
241 vec_st(vbs1,0,bufs);
242 vec_st(vbs2,16,bufs);
243 vec_st(vbs3,32,bufs);
244 vec_st(vbs4,48,bufs);
245 vec_st(vbs5,64,bufs);
246 vec_st(vbs6,80,bufs);
247 vec_st(vbs7,96,bufs);
248 vec_st(vbs8,112,bufs);
249 }
250
251 {
252 register real *b1;
253 register int i;
254
255 for(b1=bufs,i=8;i;i--,b1+=4)
256 b1[2] += b1[3];
257
258 for(b1=bufs,i=4;i;i--,b1+=8)
259 {
260 b1[4] += b1[6];
261 b1[6] += b1[5];
262 b1[5] += b1[7];
263 }
264
265 for(b1=bufs,i=2;i;i--,b1+=16)
266 {
267 b1[8] += b1[12];
268 b1[12] += b1[10];
269 b1[10] += b1[14];
270 b1[14] += b1[9];
271 b1[9] += b1[13];
272 b1[13] += b1[11];
273 b1[11] += b1[15];
274 }
275 }
276
277
278 out0[0x10*16] = bufs[0];
279 out0[0x10*15] = bufs[16+0] + bufs[16+8];
280 out0[0x10*14] = bufs[8];
281 out0[0x10*13] = bufs[16+8] + bufs[16+4];
282 out0[0x10*12] = bufs[4];
283 out0[0x10*11] = bufs[16+4] + bufs[16+12];
284 out0[0x10*10] = bufs[12];
285 out0[0x10* 9] = bufs[16+12] + bufs[16+2];
286 out0[0x10* 8] = bufs[2];
287 out0[0x10* 7] = bufs[16+2] + bufs[16+10];
288 out0[0x10* 6] = bufs[10];
289 out0[0x10* 5] = bufs[16+10] + bufs[16+6];
290 out0[0x10* 4] = bufs[6];
291 out0[0x10* 3] = bufs[16+6] + bufs[16+14];
292 out0[0x10* 2] = bufs[14];
293 out0[0x10* 1] = bufs[16+14] + bufs[16+1];
294 out0[0x10* 0] = bufs[1];
295
296 out1[0x10* 0] = bufs[1];
297 out1[0x10* 1] = bufs[16+1] + bufs[16+9];
298 out1[0x10* 2] = bufs[9];
299 out1[0x10* 3] = bufs[16+9] + bufs[16+5];
300 out1[0x10* 4] = bufs[5];
301 out1[0x10* 5] = bufs[16+5] + bufs[16+13];
302 out1[0x10* 6] = bufs[13];
303 out1[0x10* 7] = bufs[16+13] + bufs[16+3];
304 out1[0x10* 8] = bufs[3];
305 out1[0x10* 9] = bufs[16+3] + bufs[16+11];
306 out1[0x10*10] = bufs[11];
307 out1[0x10*11] = bufs[16+11] + bufs[16+7];
308 out1[0x10*12] = bufs[7];
309 out1[0x10*13] = bufs[16+7] + bufs[16+15];
310 out1[0x10*14] = bufs[15];
311 out1[0x10*15] = bufs[16+15];
312
313 }
314
315
316