1 /*****************************************************************************
2 * idctaltivec.c : Altivec IDCT module
3 *****************************************************************************
4 * Copyright (C) 2001 VideoLAN
5 * $Id: idctaltivec.c,v 1.24 2002/04/19 13:56:11 sam Exp $
7 * Authors: Christophe Massiot <massiot@via.ecp.fr>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or
12 * (at your option) any later version.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA.
22 *****************************************************************************/
24 #ifndef __BUILD_ALTIVEC_ASM__
26 /*****************************************************************************
28 *****************************************************************************/
33 #include <videolan/vlc.h>
37 /*****************************************************************************
39 *****************************************************************************/
40 static void idct_getfunctions( function_list_t * p_function_list );
42 /*****************************************************************************
43 * Build configuration tree.
44 *****************************************************************************/
49 SET_DESCRIPTION( _("Altivec IDCT module") )
50 ADD_CAPABILITY( IDCT, 200 )
51 ADD_REQUIREMENT( ALTIVEC )
52 ADD_SHORTCUT( "altivec" )
53 ADD_SHORTCUT( "idctaltivec" )
57 idct_getfunctions( &p_module->p_functions->idct );
60 MODULE_DEACTIVATE_START
61 MODULE_DEACTIVATE_STOP
63 /* Following functions are local */
65 /*****************************************************************************
66 * NormScan : This IDCT uses reordered coeffs, so we patch the scan table
67 *****************************************************************************/
68 static void NormScan( u8 ppi_scan[2][64] )
72 for( i = 0; i < 64; i++ )
75 ppi_scan[0][i] = (j >> 3) | ((j & 7) << 3);
78 ppi_scan[1][i] = (j >> 3) | ((j & 7) << 3);
82 /*****************************************************************************
83 * Placeholders for unused functions
84 *****************************************************************************/
85 static void InitIDCT( void ** pp_idct_data )
89 /*****************************************************************************
91 *****************************************************************************/
93 #ifndef CAN_COMPILE_C_ALTIVEC
95 static int16_t constants[5][8] ATTR_ALIGN(16) = {
96 {23170, 13573, 6518, 21895, -23170, -21895, 32, 31},
97 {16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725},
98 {22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521},
99 {21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692},
100 {19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722}
104 * The asm code is generated with:
106 * gcc-2.95 -fvec -D__BUILD_ALTIVEC_ASM__ -O9 -fomit-frame-pointer -mregnames -S
109 * awk '{args=""; len=split ($2, arg, ",");
110 * for (i=1; i<=len; i++) { a=arg[i]; if (i<len) a=a",";
111 * args = args sprintf ("%-6s", a) }
112 * printf ("\t\"\t%-16s%-24s\\n\"\n", $1, args) }' idct_altivec.s |
115 * I then do some simple trimming on the function prolog/trailers
118 void idct_block_copy_altivec (int16_t * block, uint8_t * dest, int stride)
121 "# stwu %r1, -128(%r1) \n"
123 "# stw %r0, 132(%r1) \n"
124 "# addi %r0, %r1, 128 \n"
127 " addi %r9, %r3, 112 \n"
128 " vspltish %v25, 4 \n"
129 " vxor %v13, %v13, %v13 \n"
130 " lis %r10, constants@ha \n"
131 " lvx %v1, 0, %r9 \n"
132 " la %r10, constants@l(%r10) \n"
133 " lvx %v5, 0, %r3 \n"
134 " addi %r9, %r3, 16 \n"
135 " lvx %v8, 0, %r10 \n"
136 " addi %r11, %r10, 32 \n"
137 " lvx %v12, 0, %r9 \n"
138 " lvx %v6, 0, %r11 \n"
139 " addi %r8, %r3, 48 \n"
140 " vslh %v1, %v1, %v25 \n"
141 " addi %r9, %r3, 80 \n"
142 " lvx %v11, 0, %r8 \n"
143 " vslh %v5, %v5, %v25 \n"
144 " lvx %v0, 0, %r9 \n"
145 " addi %r11, %r10, 64 \n"
146 " vsplth %v3, %v8, 2 \n"
147 " lvx %v7, 0, %r11 \n"
148 " addi %r9, %r3, 96 \n"
149 " vslh %v12, %v12, %v25 \n"
150 " vmhraddshs %v27, %v1, %v6, %v13 \n"
151 " addi %r8, %r3, 32 \n"
152 " vsplth %v2, %v8, 5 \n"
153 " lvx %v1, 0, %r9 \n"
154 " vslh %v11, %v11, %v25 \n"
155 " addi %r3, %r3, 64 \n"
156 " lvx %v9, 0, %r8 \n"
157 " addi %r9, %r10, 48 \n"
158 " vslh %v0, %v0, %v25 \n"
159 " lvx %v4, 0, %r9 \n"
160 " vmhraddshs %v31, %v12, %v6, %v13 \n"
161 " addi %r10, %r10, 16 \n"
162 " vmhraddshs %v30, %v0, %v7, %v13 \n"
163 " lvx %v10, 0, %r3 \n"
164 " vsplth %v19, %v8, 3 \n"
165 " vmhraddshs %v15, %v11, %v7, %v13 \n"
166 " lvx %v12, 0, %r10 \n"
167 " vsplth %v6, %v8, 4 \n"
168 " vslh %v1, %v1, %v25 \n"
169 " vsplth %v11, %v8, 1 \n"
171 " vslh %v9, %v9, %v25 \n"
172 " vsplth %v7, %v8, 0 \n"
173 " vmhraddshs %v18, %v1, %v4, %v13 \n"
174 " vspltw %v8, %v8, 3 \n"
175 " vsubshs %v0, %v13, %v27 \n"
176 " vmhraddshs %v1, %v9, %v4, %v13 \n"
177 " vmhraddshs %v17, %v3, %v31, %v0 \n"
178 " vmhraddshs %v4, %v2, %v15, %v30 \n"
179 " vslh %v10, %v10, %v25 \n"
180 " vmhraddshs %v9, %v5, %v12, %v13 \n"
181 " vspltish %v25, 6 \n"
182 " vmhraddshs %v5, %v10, %v12, %v13 \n"
183 " vmhraddshs %v28, %v19, %v30, %v15 \n"
184 " vmhraddshs %v27, %v3, %v27, %v31 \n"
185 " vsubshs %v0, %v13, %v18 \n"
186 " vmhraddshs %v18, %v11, %v18, %v1 \n"
187 " vaddshs %v30, %v17, %v4 \n"
188 " vmhraddshs %v12, %v11, %v1, %v0 \n"
189 " vsubshs %v4, %v17, %v4 \n"
190 " vaddshs %v10, %v9, %v5 \n"
191 " vsubshs %v17, %v27, %v28 \n"
192 " vaddshs %v27, %v27, %v28 \n"
193 " vsubshs %v1, %v9, %v5 \n"
194 " vaddshs %v28, %v10, %v18 \n"
195 " vsubshs %v18, %v10, %v18 \n"
196 " vaddshs %v10, %v1, %v12 \n"
197 " vsubshs %v1, %v1, %v12 \n"
198 " vsubshs %v12, %v17, %v4 \n"
199 " vaddshs %v4, %v17, %v4 \n"
200 " vmhraddshs %v5, %v7, %v12, %v1 \n"
201 " vmhraddshs %v26, %v6, %v4, %v10 \n"
202 " vmhraddshs %v29, %v6, %v12, %v1 \n"
203 " vmhraddshs %v14, %v7, %v4, %v10 \n"
204 " vsubshs %v12, %v18, %v30 \n"
205 " vaddshs %v9, %v28, %v27 \n"
206 " vaddshs %v16, %v18, %v30 \n"
207 " vsubshs %v10, %v28, %v27 \n"
208 " vmrglh %v31, %v9, %v12 \n"
209 " vmrglh %v30, %v5, %v26 \n"
210 " vmrglh %v15, %v14, %v29 \n"
211 " vmrghh %v5, %v5, %v26 \n"
212 " vmrglh %v27, %v16, %v10 \n"
213 " vmrghh %v9, %v9, %v12 \n"
214 " vmrghh %v18, %v16, %v10 \n"
215 " vmrghh %v1, %v14, %v29 \n"
216 " vmrglh %v14, %v9, %v5 \n"
217 " vmrglh %v16, %v31, %v30 \n"
218 " vmrglh %v10, %v15, %v27 \n"
219 " vmrghh %v9, %v9, %v5 \n"
220 " vmrghh %v26, %v15, %v27 \n"
221 " vmrglh %v27, %v16, %v10 \n"
222 " vmrghh %v12, %v1, %v18 \n"
223 " vmrglh %v29, %v1, %v18 \n"
224 " vsubshs %v0, %v13, %v27 \n"
225 " vmrghh %v5, %v31, %v30 \n"
226 " vmrglh %v31, %v9, %v12 \n"
227 " vmrglh %v30, %v5, %v26 \n"
228 " vmrglh %v15, %v14, %v29 \n"
229 " vmhraddshs %v17, %v3, %v31, %v0 \n"
230 " vmrghh %v18, %v16, %v10 \n"
231 " vmhraddshs %v27, %v3, %v27, %v31 \n"
232 " vmhraddshs %v4, %v2, %v15, %v30 \n"
233 " vmrghh %v1, %v14, %v29 \n"
234 " vmhraddshs %v28, %v19, %v30, %v15 \n"
235 " vmrghh %v0, %v9, %v12 \n"
236 " vsubshs %v13, %v13, %v18 \n"
237 " vmrghh %v5, %v5, %v26 \n"
238 " vmhraddshs %v18, %v11, %v18, %v1 \n"
239 " vaddshs %v9, %v0, %v8 \n"
240 " vaddshs %v30, %v17, %v4 \n"
241 " vmhraddshs %v12, %v11, %v1, %v13 \n"
242 " vsubshs %v4, %v17, %v4 \n"
243 " vaddshs %v10, %v9, %v5 \n"
244 " vsubshs %v17, %v27, %v28 \n"
245 " vaddshs %v27, %v27, %v28 \n"
246 " vsubshs %v1, %v9, %v5 \n"
247 " vaddshs %v28, %v10, %v18 \n"
248 " vsubshs %v18, %v10, %v18 \n"
249 " vaddshs %v10, %v1, %v12 \n"
250 " vsubshs %v1, %v1, %v12 \n"
251 " vsubshs %v12, %v17, %v4 \n"
252 " vaddshs %v4, %v17, %v4 \n"
253 " vaddshs %v9, %v28, %v27 \n"
254 " vmhraddshs %v14, %v7, %v4, %v10 \n"
255 " vsrah %v9, %v9, %v25 \n"
256 " vmhraddshs %v5, %v7, %v12, %v1 \n"
257 " vpkshus %v0, %v9, %v9 \n"
258 " vmhraddshs %v29, %v6, %v12, %v1 \n"
259 " stvewx %v0, 0, %r4 \n"
260 " vaddshs %v16, %v18, %v30 \n"
261 " vsrah %v31, %v14, %v25 \n"
262 " stvewx %v0, %r9, %r4 \n"
263 " add %r4, %r4, %r5 \n"
264 " vsrah %v15, %v16, %v25 \n"
265 " vpkshus %v0, %v31, %v31 \n"
266 " vsrah %v1, %v5, %v25 \n"
267 " stvewx %v0, 0, %r4 \n"
268 " vsubshs %v12, %v18, %v30 \n"
269 " stvewx %v0, %r9, %r4 \n"
270 " vmhraddshs %v26, %v6, %v4, %v10 \n"
271 " vpkshus %v0, %v1, %v1 \n"
272 " add %r4, %r4, %r5 \n"
273 " vsrah %v5, %v12, %v25 \n"
274 " stvewx %v0, 0, %r4 \n"
275 " vsrah %v30, %v29, %v25 \n"
276 " stvewx %v0, %r9, %r4 \n"
277 " vsubshs %v10, %v28, %v27 \n"
278 " vpkshus %v0, %v15, %v15 \n"
279 " add %r4, %r4, %r5 \n"
280 " stvewx %v0, 0, %r4 \n"
281 " vsrah %v18, %v26, %v25 \n"
282 " stvewx %v0, %r9, %r4 \n"
283 " vsrah %v27, %v10, %v25 \n"
284 " vpkshus %v0, %v5, %v5 \n"
285 " add %r4, %r4, %r5 \n"
286 " stvewx %v0, 0, %r4 \n"
287 " stvewx %v0, %r9, %r4 \n"
288 " vpkshus %v0, %v30, %v30 \n"
289 " add %r4, %r4, %r5 \n"
290 " stvewx %v0, 0, %r4 \n"
291 " stvewx %v0, %r9, %r4 \n"
292 " vpkshus %v0, %v18, %v18 \n"
293 " add %r4, %r4, %r5 \n"
294 " stvewx %v0, 0, %r4 \n"
295 " stvewx %v0, %r9, %r4 \n"
296 " add %r4, %r4, %r5 \n"
297 " vpkshus %v0, %v27, %v27 \n"
298 " stvewx %v0, 0, %r4 \n"
299 " stvewx %v0, %r9, %r4 \n"
301 "# addi %r0, %r1, 128 \n"
303 "# lwz %r0, 132(%r1) \n"
305 "# la %r1, 128(%r1) \n"
309 void idct_block_add_altivec (int16_t * block, uint8_t * dest, int stride)
312 "# stwu %r1, -192(%r1) \n"
314 "# stw %r0, 196(%r1) \n"
315 "# addi %r0, %r1, 192 \n"
318 " addi %r9, %r3, 112 \n"
319 " vspltish %v21, 4 \n"
320 " vxor %v1, %v1, %v1 \n"
321 " lvx %v13, 0, %r9 \n"
322 " lis %r10, constants@ha \n"
323 " vspltisw %v3, -1 \n"
324 " la %r10, constants@l(%r10) \n"
325 " lvx %v5, 0, %r3 \n"
326 " addi %r9, %r3, 16 \n"
327 " lvx %v8, 0, %r10 \n"
328 " lvx %v12, 0, %r9 \n"
329 " addi %r11, %r10, 32 \n"
330 " lvx %v6, 0, %r11 \n"
331 " addi %r8, %r3, 48 \n"
332 " vslh %v13, %v13, %v21 \n"
333 " addi %r9, %r3, 80 \n"
334 " lvx %v11, 0, %r8 \n"
335 " vslh %v5, %v5, %v21 \n"
336 " lvx %v0, 0, %r9 \n"
337 " addi %r11, %r10, 64 \n"
338 " vsplth %v2, %v8, 2 \n"
339 " lvx %v7, 0, %r11 \n"
340 " vslh %v12, %v12, %v21 \n"
341 " addi %r9, %r3, 96 \n"
342 " vmhraddshs %v24, %v13, %v6, %v1 \n"
343 " addi %r8, %r3, 32 \n"
344 " vsplth %v17, %v8, 5 \n"
345 " lvx %v13, 0, %r9 \n"
346 " vslh %v11, %v11, %v21 \n"
347 " addi %r3, %r3, 64 \n"
348 " lvx %v10, 0, %r8 \n"
349 " vslh %v0, %v0, %v21 \n"
350 " addi %r9, %r10, 48 \n"
351 " vmhraddshs %v31, %v12, %v6, %v1 \n"
352 " lvx %v4, 0, %r9 \n"
353 " addi %r10, %r10, 16 \n"
354 " vmhraddshs %v26, %v0, %v7, %v1 \n"
355 " lvx %v9, 0, %r3 \n"
356 " vsplth %v16, %v8, 3 \n"
357 " vmhraddshs %v22, %v11, %v7, %v1 \n"
358 " lvx %v6, 0, %r10 \n"
359 " lvsl %v19, 0, %r4 \n"
360 " vsubshs %v12, %v1, %v24 \n"
361 " lvsl %v0, %r5, %r4 \n"
362 " vsplth %v11, %v8, 1 \n"
363 " vslh %v10, %v10, %v21 \n"
364 " vmrghb %v19, %v3, %v19 \n"
365 " lvx %v15, 0, %r4 \n"
366 " vslh %v13, %v13, %v21 \n"
367 " vmrghb %v3, %v3, %v0 \n"
369 " vmhraddshs %v14, %v2, %v31, %v12 \n"
370 " vsplth %v7, %v8, 0 \n"
371 " vmhraddshs %v23, %v13, %v4, %v1 \n"
372 " vsplth %v18, %v8, 4 \n"
373 " vmhraddshs %v27, %v10, %v4, %v1 \n"
374 " vspltw %v8, %v8, 3 \n"
375 " vmhraddshs %v12, %v17, %v22, %v26 \n"
376 " vperm %v15, %v15, %v1, %v19 \n"
377 " vslh %v9, %v9, %v21 \n"
378 " vmhraddshs %v10, %v5, %v6, %v1 \n"
379 " vspltish %v21, 6 \n"
380 " vmhraddshs %v30, %v9, %v6, %v1 \n"
381 " vmhraddshs %v26, %v16, %v26, %v22 \n"
382 " vmhraddshs %v24, %v2, %v24, %v31 \n"
383 " vmhraddshs %v31, %v11, %v23, %v27 \n"
384 " vsubshs %v0, %v1, %v23 \n"
385 " vaddshs %v23, %v14, %v12 \n"
386 " vmhraddshs %v9, %v11, %v27, %v0 \n"
387 " vsubshs %v12, %v14, %v12 \n"
388 " vaddshs %v6, %v10, %v30 \n"
389 " vsubshs %v14, %v24, %v26 \n"
390 " vaddshs %v24, %v24, %v26 \n"
391 " vsubshs %v13, %v10, %v30 \n"
392 " vaddshs %v26, %v6, %v31 \n"
393 " vsubshs %v31, %v6, %v31 \n"
394 " vaddshs %v6, %v13, %v9 \n"
395 " vsubshs %v13, %v13, %v9 \n"
396 " vsubshs %v9, %v14, %v12 \n"
397 " vaddshs %v12, %v14, %v12 \n"
398 " vmhraddshs %v30, %v7, %v9, %v13 \n"
399 " vmhraddshs %v25, %v18, %v12, %v6 \n"
400 " vmhraddshs %v28, %v18, %v9, %v13 \n"
401 " vmhraddshs %v29, %v7, %v12, %v6 \n"
402 " vaddshs %v10, %v26, %v24 \n"
403 " vsubshs %v5, %v31, %v23 \n"
404 " vsubshs %v13, %v26, %v24 \n"
405 " vaddshs %v4, %v31, %v23 \n"
406 " vmrglh %v26, %v30, %v25 \n"
407 " vmrglh %v31, %v10, %v5 \n"
408 " vmrglh %v22, %v29, %v28 \n"
409 " vmrghh %v30, %v30, %v25 \n"
410 " vmrglh %v24, %v4, %v13 \n"
411 " vmrghh %v10, %v10, %v5 \n"
412 " vmrghh %v23, %v4, %v13 \n"
413 " vmrghh %v27, %v29, %v28 \n"
414 " vmrglh %v29, %v10, %v30 \n"
415 " vmrglh %v4, %v31, %v26 \n"
416 " vmrglh %v13, %v22, %v24 \n"
417 " vmrghh %v10, %v10, %v30 \n"
418 " vmrghh %v25, %v22, %v24 \n"
419 " vmrglh %v24, %v4, %v13 \n"
420 " vmrghh %v5, %v27, %v23 \n"
421 " vmrglh %v28, %v27, %v23 \n"
422 " vsubshs %v0, %v1, %v24 \n"
423 " vmrghh %v30, %v31, %v26 \n"
424 " vmrglh %v31, %v10, %v5 \n"
425 " vmrglh %v26, %v30, %v25 \n"
426 " vmrglh %v22, %v29, %v28 \n"
427 " vmhraddshs %v14, %v2, %v31, %v0 \n"
428 " vmrghh %v23, %v4, %v13 \n"
429 " vmhraddshs %v24, %v2, %v24, %v31 \n"
430 " vmhraddshs %v12, %v17, %v22, %v26 \n"
431 " vmrghh %v27, %v29, %v28 \n"
432 " vmhraddshs %v26, %v16, %v26, %v22 \n"
433 " vmrghh %v0, %v10, %v5 \n"
434 " vmhraddshs %v31, %v11, %v23, %v27 \n"
435 " vmrghh %v30, %v30, %v25 \n"
436 " vsubshs %v13, %v1, %v23 \n"
437 " vaddshs %v10, %v0, %v8 \n"
438 " vaddshs %v23, %v14, %v12 \n"
439 " vsubshs %v12, %v14, %v12 \n"
440 " vaddshs %v6, %v10, %v30 \n"
441 " vsubshs %v14, %v24, %v26 \n"
442 " vmhraddshs %v9, %v11, %v27, %v13 \n"
443 " vaddshs %v24, %v24, %v26 \n"
444 " vaddshs %v26, %v6, %v31 \n"
445 " vsubshs %v13, %v10, %v30 \n"
446 " vaddshs %v10, %v26, %v24 \n"
447 " vsubshs %v31, %v6, %v31 \n"
448 " vaddshs %v6, %v13, %v9 \n"
449 " vsrah %v10, %v10, %v21 \n"
450 " vsubshs %v13, %v13, %v9 \n"
451 " vaddshs %v0, %v15, %v10 \n"
452 " vsubshs %v9, %v14, %v12 \n"
453 " vaddshs %v12, %v14, %v12 \n"
454 " vpkshus %v15, %v0, %v0 \n"
455 " stvewx %v15, 0, %r4 \n"
456 " vaddshs %v4, %v31, %v23 \n"
457 " vmhraddshs %v29, %v7, %v12, %v6 \n"
458 " stvewx %v15, %r9, %r4 \n"
459 " add %r4, %r4, %r5 \n"
460 " vsubshs %v5, %v31, %v23 \n"
461 " lvx %v15, 0, %r4 \n"
462 " vmhraddshs %v30, %v7, %v9, %v13 \n"
463 " vsrah %v22, %v4, %v21 \n"
464 " vperm %v15, %v15, %v1, %v3 \n"
465 " vmhraddshs %v28, %v18, %v9, %v13 \n"
466 " vsrah %v31, %v29, %v21 \n"
467 " vsubshs %v13, %v26, %v24 \n"
468 " vaddshs %v0, %v15, %v31 \n"
469 " vsrah %v27, %v30, %v21 \n"
470 " vpkshus %v15, %v0, %v0 \n"
471 " vsrah %v30, %v5, %v21 \n"
472 " stvewx %v15, 0, %r4 \n"
473 " vsrah %v26, %v28, %v21 \n"
474 " stvewx %v15, %r9, %r4 \n"
475 " vmhraddshs %v25, %v18, %v12, %v6 \n"
476 " add %r4, %r4, %r5 \n"
477 " vsrah %v24, %v13, %v21 \n"
478 " lvx %v15, 0, %r4 \n"
479 " vperm %v15, %v15, %v1, %v19 \n"
480 " vsrah %v23, %v25, %v21 \n"
481 " vaddshs %v0, %v15, %v27 \n"
482 " vpkshus %v15, %v0, %v0 \n"
483 " stvewx %v15, 0, %r4 \n"
484 " stvewx %v15, %r9, %r4 \n"
485 " add %r4, %r4, %r5 \n"
486 " lvx %v15, 0, %r4 \n"
487 " vperm %v15, %v15, %v1, %v3 \n"
488 " vaddshs %v0, %v15, %v22 \n"
489 " vpkshus %v15, %v0, %v0 \n"
490 " stvewx %v15, 0, %r4 \n"
491 " stvewx %v15, %r9, %r4 \n"
492 " add %r4, %r4, %r5 \n"
493 " lvx %v15, 0, %r4 \n"
494 " vperm %v15, %v15, %v1, %v19 \n"
495 " vaddshs %v0, %v15, %v30 \n"
496 " vpkshus %v15, %v0, %v0 \n"
497 " stvewx %v15, 0, %r4 \n"
498 " stvewx %v15, %r9, %r4 \n"
499 " add %r4, %r4, %r5 \n"
500 " lvx %v15, 0, %r4 \n"
501 " vperm %v15, %v15, %v1, %v3 \n"
502 " vaddshs %v0, %v15, %v26 \n"
503 " vpkshus %v15, %v0, %v0 \n"
504 " stvewx %v15, 0, %r4 \n"
505 " stvewx %v15, %r9, %r4 \n"
506 " add %r4, %r4, %r5 \n"
507 " lvx %v15, 0, %r4 \n"
508 " vperm %v15, %v15, %v1, %v19 \n"
509 " vaddshs %v0, %v15, %v23 \n"
510 " vpkshus %v15, %v0, %v0 \n"
511 " stvewx %v15, 0, %r4 \n"
512 " stvewx %v15, %r9, %r4 \n"
513 " add %r4, %r4, %r5 \n"
514 " lvx %v15, 0, %r4 \n"
515 " vperm %v15, %v15, %v1, %v3 \n"
516 " vaddshs %v0, %v15, %v24 \n"
517 " vpkshus %v15, %v0, %v0 \n"
518 " stvewx %v15, 0, %r4 \n"
519 " stvewx %v15, %r9, %r4 \n"
521 "# addi %r0, %r1, 192 \n"
523 "# lwz %r0, 196(%r1) \n"
525 "# la %r1, 192(%r1) \n"
529 #endif /* !CAN_COMPILE_C_ALTIVEC */
530 #endif /* __BUILD_ALTIVEC_ASM__ */
533 #if defined(CAN_COMPILE_C_ALTIVEC) || defined(__BUILD_ALTIVEC_ASM__)
535 #define vector_s16_t vector signed short
536 #define vector_u16_t vector unsigned short
537 #define vector_s8_t vector signed char
538 #define vector_u8_t vector unsigned char
539 #define vector_s32_t vector signed int
540 #define vector_u32_t vector unsigned int
544 t1 = vec_mradds (a1, vx7, vx1 ); \
545 t8 = vec_mradds (a1, vx1, vec_subs (zero, vx7)); \
546 t7 = vec_mradds (a2, vx5, vx3); \
547 t3 = vec_mradds (ma2, vx3, vx5); \
550 t5 = vec_adds (vx0, vx4); \
551 t0 = vec_subs (vx0, vx4); \
552 t2 = vec_mradds (a0, vx6, vx2); \
553 t4 = vec_mradds (a0, vx2, vec_subs (zero,vx6)); \
554 t6 = vec_adds (t8, t3); \
555 t3 = vec_subs (t8, t3); \
556 t8 = vec_subs (t1, t7); \
557 t1 = vec_adds (t1, t7); \
560 t7 = vec_adds (t5, t2); \
561 t2 = vec_subs (t5, t2); \
562 t5 = vec_adds (t0, t4); \
563 t0 = vec_subs (t0, t4); \
564 t4 = vec_subs (t8, t3); \
565 t3 = vec_adds (t8, t3); \
568 vy0 = vec_adds (t7, t1); \
569 vy7 = vec_subs (t7, t1); \
570 vy1 = vec_mradds (c4, t3, t5); \
571 vy6 = vec_mradds (mc4, t3, t5); \
572 vy2 = vec_mradds (c4, t4, t0); \
573 vy5 = vec_mradds (mc4, t4, t0); \
574 vy3 = vec_adds (t2, t6); \
575 vy4 = vec_subs (t2, t6);
578 vector_s16_t vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7; \
579 vector_s16_t vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7; \
580 vector_s16_t a0, a1, a2, ma2, c4, mc4, zero, bias; \
581 vector_s16_t t0, t1, t2, t3, t4, t5, t6, t7, t8; \
582 vector_u16_t shift; \
584 c4 = vec_splat (constants[0], 0); \
585 a0 = vec_splat (constants[0], 1); \
586 a1 = vec_splat (constants[0], 2); \
587 a2 = vec_splat (constants[0], 3); \
588 mc4 = vec_splat (constants[0], 4); \
589 ma2 = vec_splat (constants[0], 5); \
590 bias = (vector_s16_t)vec_splat ((vector_s32_t)constants[0], 3); \
592 zero = vec_splat_s16 (0); \
593 shift = vec_splat_u16 (4); \
595 vx0 = vec_mradds (vec_sl (block[0], shift), constants[1], zero); \
596 vx1 = vec_mradds (vec_sl (block[1], shift), constants[2], zero); \
597 vx2 = vec_mradds (vec_sl (block[2], shift), constants[3], zero); \
598 vx3 = vec_mradds (vec_sl (block[3], shift), constants[4], zero); \
599 vx4 = vec_mradds (vec_sl (block[4], shift), constants[1], zero); \
600 vx5 = vec_mradds (vec_sl (block[5], shift), constants[4], zero); \
601 vx6 = vec_mradds (vec_sl (block[6], shift), constants[3], zero); \
602 vx7 = vec_mradds (vec_sl (block[7], shift), constants[2], zero); \
606 vx0 = vec_mergeh (vy0, vy4); \
607 vx1 = vec_mergel (vy0, vy4); \
608 vx2 = vec_mergeh (vy1, vy5); \
609 vx3 = vec_mergel (vy1, vy5); \
610 vx4 = vec_mergeh (vy2, vy6); \
611 vx5 = vec_mergel (vy2, vy6); \
612 vx6 = vec_mergeh (vy3, vy7); \
613 vx7 = vec_mergel (vy3, vy7); \
615 vy0 = vec_mergeh (vx0, vx4); \
616 vy1 = vec_mergel (vx0, vx4); \
617 vy2 = vec_mergeh (vx1, vx5); \
618 vy3 = vec_mergel (vx1, vx5); \
619 vy4 = vec_mergeh (vx2, vx6); \
620 vy5 = vec_mergel (vx2, vx6); \
621 vy6 = vec_mergeh (vx3, vx7); \
622 vy7 = vec_mergel (vx3, vx7); \
624 vx0 = vec_adds (vec_mergeh (vy0, vy4), bias); \
625 vx1 = vec_mergel (vy0, vy4); \
626 vx2 = vec_mergeh (vy1, vy5); \
627 vx3 = vec_mergel (vy1, vy5); \
628 vx4 = vec_mergeh (vy2, vy6); \
629 vx5 = vec_mergel (vy2, vy6); \
630 vx6 = vec_mergeh (vy3, vy7); \
631 vx7 = vec_mergel (vy3, vy7); \
635 shift = vec_splat_u16 (6); \
636 vx0 = vec_sra (vy0, shift); \
637 vx1 = vec_sra (vy1, shift); \
638 vx2 = vec_sra (vy2, shift); \
639 vx3 = vec_sra (vy3, shift); \
640 vx4 = vec_sra (vy4, shift); \
641 vx5 = vec_sra (vy5, shift); \
642 vx6 = vec_sra (vy6, shift); \
643 vx7 = vec_sra (vy7, shift);
645 static vector_s16_t constants[5] ATTR_ALIGN(16) = {
646 (vector_s16_t)(23170, 13573, 6518, 21895, -23170, -21895, 32, 31),
647 (vector_s16_t)(16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725),
648 (vector_s16_t)(22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521),
649 (vector_s16_t)(21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692),
650 (vector_s16_t)(19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722)
653 void idct_block_copy_altivec (vector_s16_t * block, unsigned char * dest,
660 #define COPY(dest,src) \
661 tmp = vec_packsu (src, src); \
662 vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \
663 vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
665 COPY (dest, vx0) dest += stride;
666 COPY (dest, vx1) dest += stride;
667 COPY (dest, vx2) dest += stride;
668 COPY (dest, vx3) dest += stride;
669 COPY (dest, vx4) dest += stride;
670 COPY (dest, vx5) dest += stride;
671 COPY (dest, vx6) dest += stride;
675 void idct_block_add_altivec (vector_s16_t * block, unsigned char * dest,
679 vector_s16_t tmp2, tmp3;
682 vector_u8_t p0, p1, p;
686 p0 = vec_lvsl (0, dest);
687 p1 = vec_lvsl (stride, dest);
688 p = vec_splat_u8 (-1);
689 perm0 = vec_mergeh (p, p0);
690 perm1 = vec_mergeh (p, p1);
692 #define ADD(dest,src,perm) \
693 /* *(uint64_t *)&tmp = *(uint64_t *)dest; */ \
694 tmp = vec_ld (0, dest); \
695 tmp2 = (vector_s16_t)vec_perm (tmp, (vector_u8_t)zero, perm); \
696 tmp3 = vec_adds (tmp2, src); \
697 tmp = vec_packsu (tmp3, tmp3); \
698 vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \
699 vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
701 ADD (dest, vx0, perm0) dest += stride;
702 ADD (dest, vx1, perm1) dest += stride;
703 ADD (dest, vx2, perm0) dest += stride;
704 ADD (dest, vx3, perm1) dest += stride;
705 ADD (dest, vx4, perm0) dest += stride;
706 ADD (dest, vx5, perm1) dest += stride;
707 ADD (dest, vx6, perm0) dest += stride;
708 ADD (dest, vx7, perm1)
711 #endif /* __BUILD_ALTIVEC_ASM__ || CAN_COMPILE_C_ALTIVEC */
713 #ifndef __BUILD_ALTIVEC_ASM__
715 /*****************************************************************************
716 * Functions exported as capabilities. They are declared as static so that
717 * we don't pollute the namespace too much.
718 *****************************************************************************/
719 static void idct_getfunctions( function_list_t * p_function_list )
721 #define F p_function_list->functions.idct
722 F.pf_idct_init = InitIDCT;
723 F.pf_norm_scan = NormScan;
724 /* FIXME : it would be a nice idea to use sparse IDCT functions */
725 F.pf_sparse_idct_add = idct_block_add_altivec;
726 F.pf_sparse_idct_copy = idct_block_copy_altivec;
727 F.pf_idct_add = idct_block_add_altivec;
728 F.pf_idct_copy = idct_block_copy_altivec;
732 #endif /* __BUILD_ALTIVEC_ASM__ */