1 /*****************************************************************************
2 * idctaltivec.c : Altivec IDCT module
3 *****************************************************************************
4 * Copyright (C) 1999, 2000 VideoLAN
5 * $Id: idctaltivec.c,v 1.18 2001/09/28 15:08:40 massiot 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 #define MODULE_NAME idctaltivec
27 #include "modules_inner.h"
29 /*****************************************************************************
31 *****************************************************************************/
42 #include "tests.h" /* TestCPU() */
47 #include "modules_export.h"
49 /*****************************************************************************
51 *****************************************************************************/
52 static void idct_getfunctions( function_list_t * p_function_list );
54 /*****************************************************************************
55 * Build configuration tree.
56 *****************************************************************************/
58 ADD_WINDOW( "Configuration for Altivec IDCT module" )
59 ADD_COMMENT( "Ha, ha -- nothing to configure yet" )
63 p_module->i_capabilities = MODULE_CAPABILITY_NULL
64 | MODULE_CAPABILITY_IDCT;
65 p_module->psz_longname = "Altivec IDCT module";
69 idct_getfunctions( &p_module->p_functions->idct );
72 MODULE_DEACTIVATE_START
73 MODULE_DEACTIVATE_STOP
75 /* Following functions are local */
77 /*****************************************************************************
78 * idct_Probe: return a preference score
79 *****************************************************************************/
80 static int idct_Probe( probedata_t *p_data )
82 if( !TestCPU( CPU_CAPABILITY_ALTIVEC ) )
87 if( TestMethod( IDCT_METHOD_VAR, "idctaltivec" )
88 || TestMethod( IDCT_METHOD_VAR, "altivec" ) )
96 /*****************************************************************************
97 * NormScan : This IDCT uses reordered coeffs, so we patch the scan table
98 *****************************************************************************/
99 static void NormScan( u8 ppi_scan[2][64] )
103 for( i = 0; i < 64; i++ )
106 ppi_scan[0][i] = (j >> 3) | ((j & 7) << 3);
109 ppi_scan[1][i] = (j >> 3) | ((j & 7) << 3);
113 /*****************************************************************************
114 * Placeholders for unused functions
115 *****************************************************************************/
116 static void InitIDCT( void ** pp_idct_data )
120 /*****************************************************************************
122 *****************************************************************************/
124 #ifndef HAVE_C_ALTIVEC
126 static int16_t constants[5][8] ATTR_ALIGN(16) = {
127 {23170, 13573, 6518, 21895, -23170, -21895, 32, 31},
128 {16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725},
129 {22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521},
130 {21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692},
131 {19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722}
135 * The asm code is generated with:
137 * gcc-2.95 -fvec -D__BUILD_ALTIVEC_ASM__ -O9 -fomit-frame-pointer -mregnames -S
140 * awk '{args=""; len=split ($2, arg, ",");
141 * for (i=1; i<=len; i++) { a=arg[i]; if (i<len) a=a",";
142 * args = args sprintf ("%-6s", a) }
143 * printf ("\t\"\t%-16s%-24s\\n\"\n", $1, args) }' idct_altivec.s |
146 * I then do some simple trimming on the function prolog/trailers
149 void idct_block_copy_altivec (int16_t * block, uint8_t * dest, int stride)
152 "# stwu %r1, -128(%r1) \n"
154 "# stw %r0, 132(%r1) \n"
155 "# addi %r0, %r1, 128 \n"
158 " addi %r9, %r3, 112 \n"
159 " vspltish %v25, 4 \n"
160 " vxor %v13, %v13, %v13 \n"
161 " lis %r10, constants@ha \n"
162 " lvx %v1, 0, %r9 \n"
163 " la %r10, constants@l(%r10) \n"
164 " lvx %v5, 0, %r3 \n"
165 " addi %r9, %r3, 16 \n"
166 " lvx %v8, 0, %r10 \n"
167 " addi %r11, %r10, 32 \n"
168 " lvx %v12, 0, %r9 \n"
169 " lvx %v6, 0, %r11 \n"
170 " addi %r8, %r3, 48 \n"
171 " vslh %v1, %v1, %v25 \n"
172 " addi %r9, %r3, 80 \n"
173 " lvx %v11, 0, %r8 \n"
174 " vslh %v5, %v5, %v25 \n"
175 " lvx %v0, 0, %r9 \n"
176 " addi %r11, %r10, 64 \n"
177 " vsplth %v3, %v8, 2 \n"
178 " lvx %v7, 0, %r11 \n"
179 " addi %r9, %r3, 96 \n"
180 " vslh %v12, %v12, %v25 \n"
181 " vmhraddshs %v27, %v1, %v6, %v13 \n"
182 " addi %r8, %r3, 32 \n"
183 " vsplth %v2, %v8, 5 \n"
184 " lvx %v1, 0, %r9 \n"
185 " vslh %v11, %v11, %v25 \n"
186 " addi %r3, %r3, 64 \n"
187 " lvx %v9, 0, %r8 \n"
188 " addi %r9, %r10, 48 \n"
189 " vslh %v0, %v0, %v25 \n"
190 " lvx %v4, 0, %r9 \n"
191 " vmhraddshs %v31, %v12, %v6, %v13 \n"
192 " addi %r10, %r10, 16 \n"
193 " vmhraddshs %v30, %v0, %v7, %v13 \n"
194 " lvx %v10, 0, %r3 \n"
195 " vsplth %v19, %v8, 3 \n"
196 " vmhraddshs %v15, %v11, %v7, %v13 \n"
197 " lvx %v12, 0, %r10 \n"
198 " vsplth %v6, %v8, 4 \n"
199 " vslh %v1, %v1, %v25 \n"
200 " vsplth %v11, %v8, 1 \n"
202 " vslh %v9, %v9, %v25 \n"
203 " vsplth %v7, %v8, 0 \n"
204 " vmhraddshs %v18, %v1, %v4, %v13 \n"
205 " vspltw %v8, %v8, 3 \n"
206 " vsubshs %v0, %v13, %v27 \n"
207 " vmhraddshs %v1, %v9, %v4, %v13 \n"
208 " vmhraddshs %v17, %v3, %v31, %v0 \n"
209 " vmhraddshs %v4, %v2, %v15, %v30 \n"
210 " vslh %v10, %v10, %v25 \n"
211 " vmhraddshs %v9, %v5, %v12, %v13 \n"
212 " vspltish %v25, 6 \n"
213 " vmhraddshs %v5, %v10, %v12, %v13 \n"
214 " vmhraddshs %v28, %v19, %v30, %v15 \n"
215 " vmhraddshs %v27, %v3, %v27, %v31 \n"
216 " vsubshs %v0, %v13, %v18 \n"
217 " vmhraddshs %v18, %v11, %v18, %v1 \n"
218 " vaddshs %v30, %v17, %v4 \n"
219 " vmhraddshs %v12, %v11, %v1, %v0 \n"
220 " vsubshs %v4, %v17, %v4 \n"
221 " vaddshs %v10, %v9, %v5 \n"
222 " vsubshs %v17, %v27, %v28 \n"
223 " vaddshs %v27, %v27, %v28 \n"
224 " vsubshs %v1, %v9, %v5 \n"
225 " vaddshs %v28, %v10, %v18 \n"
226 " vsubshs %v18, %v10, %v18 \n"
227 " vaddshs %v10, %v1, %v12 \n"
228 " vsubshs %v1, %v1, %v12 \n"
229 " vsubshs %v12, %v17, %v4 \n"
230 " vaddshs %v4, %v17, %v4 \n"
231 " vmhraddshs %v5, %v7, %v12, %v1 \n"
232 " vmhraddshs %v26, %v6, %v4, %v10 \n"
233 " vmhraddshs %v29, %v6, %v12, %v1 \n"
234 " vmhraddshs %v14, %v7, %v4, %v10 \n"
235 " vsubshs %v12, %v18, %v30 \n"
236 " vaddshs %v9, %v28, %v27 \n"
237 " vaddshs %v16, %v18, %v30 \n"
238 " vsubshs %v10, %v28, %v27 \n"
239 " vmrglh %v31, %v9, %v12 \n"
240 " vmrglh %v30, %v5, %v26 \n"
241 " vmrglh %v15, %v14, %v29 \n"
242 " vmrghh %v5, %v5, %v26 \n"
243 " vmrglh %v27, %v16, %v10 \n"
244 " vmrghh %v9, %v9, %v12 \n"
245 " vmrghh %v18, %v16, %v10 \n"
246 " vmrghh %v1, %v14, %v29 \n"
247 " vmrglh %v14, %v9, %v5 \n"
248 " vmrglh %v16, %v31, %v30 \n"
249 " vmrglh %v10, %v15, %v27 \n"
250 " vmrghh %v9, %v9, %v5 \n"
251 " vmrghh %v26, %v15, %v27 \n"
252 " vmrglh %v27, %v16, %v10 \n"
253 " vmrghh %v12, %v1, %v18 \n"
254 " vmrglh %v29, %v1, %v18 \n"
255 " vsubshs %v0, %v13, %v27 \n"
256 " vmrghh %v5, %v31, %v30 \n"
257 " vmrglh %v31, %v9, %v12 \n"
258 " vmrglh %v30, %v5, %v26 \n"
259 " vmrglh %v15, %v14, %v29 \n"
260 " vmhraddshs %v17, %v3, %v31, %v0 \n"
261 " vmrghh %v18, %v16, %v10 \n"
262 " vmhraddshs %v27, %v3, %v27, %v31 \n"
263 " vmhraddshs %v4, %v2, %v15, %v30 \n"
264 " vmrghh %v1, %v14, %v29 \n"
265 " vmhraddshs %v28, %v19, %v30, %v15 \n"
266 " vmrghh %v0, %v9, %v12 \n"
267 " vsubshs %v13, %v13, %v18 \n"
268 " vmrghh %v5, %v5, %v26 \n"
269 " vmhraddshs %v18, %v11, %v18, %v1 \n"
270 " vaddshs %v9, %v0, %v8 \n"
271 " vaddshs %v30, %v17, %v4 \n"
272 " vmhraddshs %v12, %v11, %v1, %v13 \n"
273 " vsubshs %v4, %v17, %v4 \n"
274 " vaddshs %v10, %v9, %v5 \n"
275 " vsubshs %v17, %v27, %v28 \n"
276 " vaddshs %v27, %v27, %v28 \n"
277 " vsubshs %v1, %v9, %v5 \n"
278 " vaddshs %v28, %v10, %v18 \n"
279 " vsubshs %v18, %v10, %v18 \n"
280 " vaddshs %v10, %v1, %v12 \n"
281 " vsubshs %v1, %v1, %v12 \n"
282 " vsubshs %v12, %v17, %v4 \n"
283 " vaddshs %v4, %v17, %v4 \n"
284 " vaddshs %v9, %v28, %v27 \n"
285 " vmhraddshs %v14, %v7, %v4, %v10 \n"
286 " vsrah %v9, %v9, %v25 \n"
287 " vmhraddshs %v5, %v7, %v12, %v1 \n"
288 " vpkshus %v0, %v9, %v9 \n"
289 " vmhraddshs %v29, %v6, %v12, %v1 \n"
290 " stvewx %v0, 0, %r4 \n"
291 " vaddshs %v16, %v18, %v30 \n"
292 " vsrah %v31, %v14, %v25 \n"
293 " stvewx %v0, %r9, %r4 \n"
294 " add %r4, %r4, %r5 \n"
295 " vsrah %v15, %v16, %v25 \n"
296 " vpkshus %v0, %v31, %v31 \n"
297 " vsrah %v1, %v5, %v25 \n"
298 " stvewx %v0, 0, %r4 \n"
299 " vsubshs %v12, %v18, %v30 \n"
300 " stvewx %v0, %r9, %r4 \n"
301 " vmhraddshs %v26, %v6, %v4, %v10 \n"
302 " vpkshus %v0, %v1, %v1 \n"
303 " add %r4, %r4, %r5 \n"
304 " vsrah %v5, %v12, %v25 \n"
305 " stvewx %v0, 0, %r4 \n"
306 " vsrah %v30, %v29, %v25 \n"
307 " stvewx %v0, %r9, %r4 \n"
308 " vsubshs %v10, %v28, %v27 \n"
309 " vpkshus %v0, %v15, %v15 \n"
310 " add %r4, %r4, %r5 \n"
311 " stvewx %v0, 0, %r4 \n"
312 " vsrah %v18, %v26, %v25 \n"
313 " stvewx %v0, %r9, %r4 \n"
314 " vsrah %v27, %v10, %v25 \n"
315 " vpkshus %v0, %v5, %v5 \n"
316 " add %r4, %r4, %r5 \n"
317 " stvewx %v0, 0, %r4 \n"
318 " stvewx %v0, %r9, %r4 \n"
319 " vpkshus %v0, %v30, %v30 \n"
320 " add %r4, %r4, %r5 \n"
321 " stvewx %v0, 0, %r4 \n"
322 " stvewx %v0, %r9, %r4 \n"
323 " vpkshus %v0, %v18, %v18 \n"
324 " add %r4, %r4, %r5 \n"
325 " stvewx %v0, 0, %r4 \n"
326 " stvewx %v0, %r9, %r4 \n"
327 " add %r4, %r4, %r5 \n"
328 " vpkshus %v0, %v27, %v27 \n"
329 " stvewx %v0, 0, %r4 \n"
330 " stvewx %v0, %r9, %r4 \n"
332 "# addi %r0, %r1, 128 \n"
334 "# lwz %r0, 132(%r1) \n"
336 "# la %r1, 128(%r1) \n"
340 void idct_block_add_altivec (int16_t * block, uint8_t * dest, int stride)
343 "# stwu %r1, -192(%r1) \n"
345 "# stw %r0, 196(%r1) \n"
346 "# addi %r0, %r1, 192 \n"
349 " addi %r9, %r3, 112 \n"
350 " vspltish %v21, 4 \n"
351 " vxor %v1, %v1, %v1 \n"
352 " lvx %v13, 0, %r9 \n"
353 " lis %r10, constants@ha \n"
354 " vspltisw %v3, -1 \n"
355 " la %r10, constants@l(%r10) \n"
356 " lvx %v5, 0, %r3 \n"
357 " addi %r9, %r3, 16 \n"
358 " lvx %v8, 0, %r10 \n"
359 " lvx %v12, 0, %r9 \n"
360 " addi %r11, %r10, 32 \n"
361 " lvx %v6, 0, %r11 \n"
362 " addi %r8, %r3, 48 \n"
363 " vslh %v13, %v13, %v21 \n"
364 " addi %r9, %r3, 80 \n"
365 " lvx %v11, 0, %r8 \n"
366 " vslh %v5, %v5, %v21 \n"
367 " lvx %v0, 0, %r9 \n"
368 " addi %r11, %r10, 64 \n"
369 " vsplth %v2, %v8, 2 \n"
370 " lvx %v7, 0, %r11 \n"
371 " vslh %v12, %v12, %v21 \n"
372 " addi %r9, %r3, 96 \n"
373 " vmhraddshs %v24, %v13, %v6, %v1 \n"
374 " addi %r8, %r3, 32 \n"
375 " vsplth %v17, %v8, 5 \n"
376 " lvx %v13, 0, %r9 \n"
377 " vslh %v11, %v11, %v21 \n"
378 " addi %r3, %r3, 64 \n"
379 " lvx %v10, 0, %r8 \n"
380 " vslh %v0, %v0, %v21 \n"
381 " addi %r9, %r10, 48 \n"
382 " vmhraddshs %v31, %v12, %v6, %v1 \n"
383 " lvx %v4, 0, %r9 \n"
384 " addi %r10, %r10, 16 \n"
385 " vmhraddshs %v26, %v0, %v7, %v1 \n"
386 " lvx %v9, 0, %r3 \n"
387 " vsplth %v16, %v8, 3 \n"
388 " vmhraddshs %v22, %v11, %v7, %v1 \n"
389 " lvx %v6, 0, %r10 \n"
390 " lvsl %v19, 0, %r4 \n"
391 " vsubshs %v12, %v1, %v24 \n"
392 " lvsl %v0, %r5, %r4 \n"
393 " vsplth %v11, %v8, 1 \n"
394 " vslh %v10, %v10, %v21 \n"
395 " vmrghb %v19, %v3, %v19 \n"
396 " lvx %v15, 0, %r4 \n"
397 " vslh %v13, %v13, %v21 \n"
398 " vmrghb %v3, %v3, %v0 \n"
400 " vmhraddshs %v14, %v2, %v31, %v12 \n"
401 " vsplth %v7, %v8, 0 \n"
402 " vmhraddshs %v23, %v13, %v4, %v1 \n"
403 " vsplth %v18, %v8, 4 \n"
404 " vmhraddshs %v27, %v10, %v4, %v1 \n"
405 " vspltw %v8, %v8, 3 \n"
406 " vmhraddshs %v12, %v17, %v22, %v26 \n"
407 " vperm %v15, %v15, %v1, %v19 \n"
408 " vslh %v9, %v9, %v21 \n"
409 " vmhraddshs %v10, %v5, %v6, %v1 \n"
410 " vspltish %v21, 6 \n"
411 " vmhraddshs %v30, %v9, %v6, %v1 \n"
412 " vmhraddshs %v26, %v16, %v26, %v22 \n"
413 " vmhraddshs %v24, %v2, %v24, %v31 \n"
414 " vmhraddshs %v31, %v11, %v23, %v27 \n"
415 " vsubshs %v0, %v1, %v23 \n"
416 " vaddshs %v23, %v14, %v12 \n"
417 " vmhraddshs %v9, %v11, %v27, %v0 \n"
418 " vsubshs %v12, %v14, %v12 \n"
419 " vaddshs %v6, %v10, %v30 \n"
420 " vsubshs %v14, %v24, %v26 \n"
421 " vaddshs %v24, %v24, %v26 \n"
422 " vsubshs %v13, %v10, %v30 \n"
423 " vaddshs %v26, %v6, %v31 \n"
424 " vsubshs %v31, %v6, %v31 \n"
425 " vaddshs %v6, %v13, %v9 \n"
426 " vsubshs %v13, %v13, %v9 \n"
427 " vsubshs %v9, %v14, %v12 \n"
428 " vaddshs %v12, %v14, %v12 \n"
429 " vmhraddshs %v30, %v7, %v9, %v13 \n"
430 " vmhraddshs %v25, %v18, %v12, %v6 \n"
431 " vmhraddshs %v28, %v18, %v9, %v13 \n"
432 " vmhraddshs %v29, %v7, %v12, %v6 \n"
433 " vaddshs %v10, %v26, %v24 \n"
434 " vsubshs %v5, %v31, %v23 \n"
435 " vsubshs %v13, %v26, %v24 \n"
436 " vaddshs %v4, %v31, %v23 \n"
437 " vmrglh %v26, %v30, %v25 \n"
438 " vmrglh %v31, %v10, %v5 \n"
439 " vmrglh %v22, %v29, %v28 \n"
440 " vmrghh %v30, %v30, %v25 \n"
441 " vmrglh %v24, %v4, %v13 \n"
442 " vmrghh %v10, %v10, %v5 \n"
443 " vmrghh %v23, %v4, %v13 \n"
444 " vmrghh %v27, %v29, %v28 \n"
445 " vmrglh %v29, %v10, %v30 \n"
446 " vmrglh %v4, %v31, %v26 \n"
447 " vmrglh %v13, %v22, %v24 \n"
448 " vmrghh %v10, %v10, %v30 \n"
449 " vmrghh %v25, %v22, %v24 \n"
450 " vmrglh %v24, %v4, %v13 \n"
451 " vmrghh %v5, %v27, %v23 \n"
452 " vmrglh %v28, %v27, %v23 \n"
453 " vsubshs %v0, %v1, %v24 \n"
454 " vmrghh %v30, %v31, %v26 \n"
455 " vmrglh %v31, %v10, %v5 \n"
456 " vmrglh %v26, %v30, %v25 \n"
457 " vmrglh %v22, %v29, %v28 \n"
458 " vmhraddshs %v14, %v2, %v31, %v0 \n"
459 " vmrghh %v23, %v4, %v13 \n"
460 " vmhraddshs %v24, %v2, %v24, %v31 \n"
461 " vmhraddshs %v12, %v17, %v22, %v26 \n"
462 " vmrghh %v27, %v29, %v28 \n"
463 " vmhraddshs %v26, %v16, %v26, %v22 \n"
464 " vmrghh %v0, %v10, %v5 \n"
465 " vmhraddshs %v31, %v11, %v23, %v27 \n"
466 " vmrghh %v30, %v30, %v25 \n"
467 " vsubshs %v13, %v1, %v23 \n"
468 " vaddshs %v10, %v0, %v8 \n"
469 " vaddshs %v23, %v14, %v12 \n"
470 " vsubshs %v12, %v14, %v12 \n"
471 " vaddshs %v6, %v10, %v30 \n"
472 " vsubshs %v14, %v24, %v26 \n"
473 " vmhraddshs %v9, %v11, %v27, %v13 \n"
474 " vaddshs %v24, %v24, %v26 \n"
475 " vaddshs %v26, %v6, %v31 \n"
476 " vsubshs %v13, %v10, %v30 \n"
477 " vaddshs %v10, %v26, %v24 \n"
478 " vsubshs %v31, %v6, %v31 \n"
479 " vaddshs %v6, %v13, %v9 \n"
480 " vsrah %v10, %v10, %v21 \n"
481 " vsubshs %v13, %v13, %v9 \n"
482 " vaddshs %v0, %v15, %v10 \n"
483 " vsubshs %v9, %v14, %v12 \n"
484 " vaddshs %v12, %v14, %v12 \n"
485 " vpkshus %v15, %v0, %v0 \n"
486 " stvewx %v15, 0, %r4 \n"
487 " vaddshs %v4, %v31, %v23 \n"
488 " vmhraddshs %v29, %v7, %v12, %v6 \n"
489 " stvewx %v15, %r9, %r4 \n"
490 " add %r4, %r4, %r5 \n"
491 " vsubshs %v5, %v31, %v23 \n"
492 " lvx %v15, 0, %r4 \n"
493 " vmhraddshs %v30, %v7, %v9, %v13 \n"
494 " vsrah %v22, %v4, %v21 \n"
495 " vperm %v15, %v15, %v1, %v3 \n"
496 " vmhraddshs %v28, %v18, %v9, %v13 \n"
497 " vsrah %v31, %v29, %v21 \n"
498 " vsubshs %v13, %v26, %v24 \n"
499 " vaddshs %v0, %v15, %v31 \n"
500 " vsrah %v27, %v30, %v21 \n"
501 " vpkshus %v15, %v0, %v0 \n"
502 " vsrah %v30, %v5, %v21 \n"
503 " stvewx %v15, 0, %r4 \n"
504 " vsrah %v26, %v28, %v21 \n"
505 " stvewx %v15, %r9, %r4 \n"
506 " vmhraddshs %v25, %v18, %v12, %v6 \n"
507 " add %r4, %r4, %r5 \n"
508 " vsrah %v24, %v13, %v21 \n"
509 " lvx %v15, 0, %r4 \n"
510 " vperm %v15, %v15, %v1, %v19 \n"
511 " vsrah %v23, %v25, %v21 \n"
512 " vaddshs %v0, %v15, %v27 \n"
513 " vpkshus %v15, %v0, %v0 \n"
514 " stvewx %v15, 0, %r4 \n"
515 " stvewx %v15, %r9, %r4 \n"
516 " add %r4, %r4, %r5 \n"
517 " lvx %v15, 0, %r4 \n"
518 " vperm %v15, %v15, %v1, %v3 \n"
519 " vaddshs %v0, %v15, %v22 \n"
520 " vpkshus %v15, %v0, %v0 \n"
521 " stvewx %v15, 0, %r4 \n"
522 " stvewx %v15, %r9, %r4 \n"
523 " add %r4, %r4, %r5 \n"
524 " lvx %v15, 0, %r4 \n"
525 " vperm %v15, %v15, %v1, %v19 \n"
526 " vaddshs %v0, %v15, %v30 \n"
527 " vpkshus %v15, %v0, %v0 \n"
528 " stvewx %v15, 0, %r4 \n"
529 " stvewx %v15, %r9, %r4 \n"
530 " add %r4, %r4, %r5 \n"
531 " lvx %v15, 0, %r4 \n"
532 " vperm %v15, %v15, %v1, %v3 \n"
533 " vaddshs %v0, %v15, %v26 \n"
534 " vpkshus %v15, %v0, %v0 \n"
535 " stvewx %v15, 0, %r4 \n"
536 " stvewx %v15, %r9, %r4 \n"
537 " add %r4, %r4, %r5 \n"
538 " lvx %v15, 0, %r4 \n"
539 " vperm %v15, %v15, %v1, %v19 \n"
540 " vaddshs %v0, %v15, %v23 \n"
541 " vpkshus %v15, %v0, %v0 \n"
542 " stvewx %v15, 0, %r4 \n"
543 " stvewx %v15, %r9, %r4 \n"
544 " add %r4, %r4, %r5 \n"
545 " lvx %v15, 0, %r4 \n"
546 " vperm %v15, %v15, %v1, %v3 \n"
547 " vaddshs %v0, %v15, %v24 \n"
548 " vpkshus %v15, %v0, %v0 \n"
549 " stvewx %v15, 0, %r4 \n"
550 " stvewx %v15, %r9, %r4 \n"
552 "# addi %r0, %r1, 192 \n"
554 "# lwz %r0, 196(%r1) \n"
556 "# la %r1, 192(%r1) \n"
560 #endif /* !HAVE_C_ALTIVEC */
561 #endif /* __BUILD_ALTIVEC_ASM__ */
564 #if defined(HAVE_C_ALTIVEC) || defined(__BUILD_ALTIVEC_ASM__)
566 #define vector_s16_t vector signed short
567 #define vector_u16_t vector unsigned short
568 #define vector_s8_t vector signed char
569 #define vector_u8_t vector unsigned char
570 #define vector_s32_t vector signed int
571 #define vector_u32_t vector unsigned int
575 t1 = vec_mradds (a1, vx7, vx1 ); \
576 t8 = vec_mradds (a1, vx1, vec_subs (zero, vx7)); \
577 t7 = vec_mradds (a2, vx5, vx3); \
578 t3 = vec_mradds (ma2, vx3, vx5); \
581 t5 = vec_adds (vx0, vx4); \
582 t0 = vec_subs (vx0, vx4); \
583 t2 = vec_mradds (a0, vx6, vx2); \
584 t4 = vec_mradds (a0, vx2, vec_subs (zero,vx6)); \
585 t6 = vec_adds (t8, t3); \
586 t3 = vec_subs (t8, t3); \
587 t8 = vec_subs (t1, t7); \
588 t1 = vec_adds (t1, t7); \
591 t7 = vec_adds (t5, t2); \
592 t2 = vec_subs (t5, t2); \
593 t5 = vec_adds (t0, t4); \
594 t0 = vec_subs (t0, t4); \
595 t4 = vec_subs (t8, t3); \
596 t3 = vec_adds (t8, t3); \
599 vy0 = vec_adds (t7, t1); \
600 vy7 = vec_subs (t7, t1); \
601 vy1 = vec_mradds (c4, t3, t5); \
602 vy6 = vec_mradds (mc4, t3, t5); \
603 vy2 = vec_mradds (c4, t4, t0); \
604 vy5 = vec_mradds (mc4, t4, t0); \
605 vy3 = vec_adds (t2, t6); \
606 vy4 = vec_subs (t2, t6);
609 vector_s16_t vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7; \
610 vector_s16_t vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7; \
611 vector_s16_t a0, a1, a2, ma2, c4, mc4, zero, bias; \
612 vector_s16_t t0, t1, t2, t3, t4, t5, t6, t7, t8; \
613 vector_u16_t shift; \
615 c4 = vec_splat (constants[0], 0); \
616 a0 = vec_splat (constants[0], 1); \
617 a1 = vec_splat (constants[0], 2); \
618 a2 = vec_splat (constants[0], 3); \
619 mc4 = vec_splat (constants[0], 4); \
620 ma2 = vec_splat (constants[0], 5); \
621 bias = (vector_s16_t)vec_splat ((vector_s32_t)constants[0], 3); \
623 zero = vec_splat_s16 (0); \
624 shift = vec_splat_u16 (4); \
626 vx0 = vec_mradds (vec_sl (block[0], shift), constants[1], zero); \
627 vx1 = vec_mradds (vec_sl (block[1], shift), constants[2], zero); \
628 vx2 = vec_mradds (vec_sl (block[2], shift), constants[3], zero); \
629 vx3 = vec_mradds (vec_sl (block[3], shift), constants[4], zero); \
630 vx4 = vec_mradds (vec_sl (block[4], shift), constants[1], zero); \
631 vx5 = vec_mradds (vec_sl (block[5], shift), constants[4], zero); \
632 vx6 = vec_mradds (vec_sl (block[6], shift), constants[3], zero); \
633 vx7 = vec_mradds (vec_sl (block[7], shift), constants[2], zero); \
637 vx0 = vec_mergeh (vy0, vy4); \
638 vx1 = vec_mergel (vy0, vy4); \
639 vx2 = vec_mergeh (vy1, vy5); \
640 vx3 = vec_mergel (vy1, vy5); \
641 vx4 = vec_mergeh (vy2, vy6); \
642 vx5 = vec_mergel (vy2, vy6); \
643 vx6 = vec_mergeh (vy3, vy7); \
644 vx7 = vec_mergel (vy3, vy7); \
646 vy0 = vec_mergeh (vx0, vx4); \
647 vy1 = vec_mergel (vx0, vx4); \
648 vy2 = vec_mergeh (vx1, vx5); \
649 vy3 = vec_mergel (vx1, vx5); \
650 vy4 = vec_mergeh (vx2, vx6); \
651 vy5 = vec_mergel (vx2, vx6); \
652 vy6 = vec_mergeh (vx3, vx7); \
653 vy7 = vec_mergel (vx3, vx7); \
655 vx0 = vec_adds (vec_mergeh (vy0, vy4), bias); \
656 vx1 = vec_mergel (vy0, vy4); \
657 vx2 = vec_mergeh (vy1, vy5); \
658 vx3 = vec_mergel (vy1, vy5); \
659 vx4 = vec_mergeh (vy2, vy6); \
660 vx5 = vec_mergel (vy2, vy6); \
661 vx6 = vec_mergeh (vy3, vy7); \
662 vx7 = vec_mergel (vy3, vy7); \
666 shift = vec_splat_u16 (6); \
667 vx0 = vec_sra (vy0, shift); \
668 vx1 = vec_sra (vy1, shift); \
669 vx2 = vec_sra (vy2, shift); \
670 vx3 = vec_sra (vy3, shift); \
671 vx4 = vec_sra (vy4, shift); \
672 vx5 = vec_sra (vy5, shift); \
673 vx6 = vec_sra (vy6, shift); \
674 vx7 = vec_sra (vy7, shift);
676 static vector_s16_t constants[5] ATTR_ALIGN(16) = {
677 (vector_s16_t)(23170, 13573, 6518, 21895, -23170, -21895, 32, 31),
678 (vector_s16_t)(16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725),
679 (vector_s16_t)(22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521),
680 (vector_s16_t)(21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692),
681 (vector_s16_t)(19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722)
684 void idct_block_copy_altivec (vector_s16_t * block, unsigned char * dest,
691 #define COPY(dest,src) \
692 tmp = vec_packsu (src, src); \
693 vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \
694 vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
696 COPY (dest, vx0) dest += stride;
697 COPY (dest, vx1) dest += stride;
698 COPY (dest, vx2) dest += stride;
699 COPY (dest, vx3) dest += stride;
700 COPY (dest, vx4) dest += stride;
701 COPY (dest, vx5) dest += stride;
702 COPY (dest, vx6) dest += stride;
706 void idct_block_add_altivec (vector_s16_t * block, unsigned char * dest,
710 vector_s16_t tmp2, tmp3;
713 vector_u8_t p0, p1, p;
717 p0 = vec_lvsl (0, dest);
718 p1 = vec_lvsl (stride, dest);
719 p = vec_splat_u8 (-1);
720 perm0 = vec_mergeh (p, p0);
721 perm1 = vec_mergeh (p, p1);
723 #define ADD(dest,src,perm) \
724 /* *(uint64_t *)&tmp = *(uint64_t *)dest; */ \
725 tmp = vec_ld (0, dest); \
726 tmp2 = (vector_s16_t)vec_perm (tmp, (vector_u8_t)zero, perm); \
727 tmp3 = vec_adds (tmp2, src); \
728 tmp = vec_packsu (tmp3, tmp3); \
729 vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \
730 vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
732 ADD (dest, vx0, perm0) dest += stride;
733 ADD (dest, vx1, perm1) dest += stride;
734 ADD (dest, vx2, perm0) dest += stride;
735 ADD (dest, vx3, perm1) dest += stride;
736 ADD (dest, vx4, perm0) dest += stride;
737 ADD (dest, vx5, perm1) dest += stride;
738 ADD (dest, vx6, perm0) dest += stride;
739 ADD (dest, vx7, perm1)
742 #endif /* __BUILD_ALTIVEC_ASM__ || HAVE_C_ALTIVEC */
744 #ifndef __BUILD_ALTIVEC_ASM__
746 /*****************************************************************************
747 * Functions exported as capabilities. They are declared as static so that
748 * we don't pollute the namespace too much.
749 *****************************************************************************/
750 static void idct_getfunctions( function_list_t * p_function_list )
752 p_function_list->pf_probe = idct_Probe;
753 #define F p_function_list->functions.idct
754 F.pf_idct_init = InitIDCT;
755 F.pf_norm_scan = NormScan;
756 /* FIXME : it would be a nice idea to use sparse IDCT functions */
757 F.pf_sparse_idct_add = idct_block_add_altivec;
758 F.pf_sparse_idct_copy = idct_block_copy_altivec;
759 F.pf_idct_add = idct_block_add_altivec;
760 F.pf_idct_copy = idct_block_copy_altivec;
764 #endif /* __BUILD_ALTIVEC_ASM__ */