Altivec on non darwin systems patch by Romain Dolbeau
[libav.git] / libavcodec / ppc / mpegvideo_altivec.c
1 /*
2 * Copyright (c) 2002 Dieter Shirley
3 *
4 * This library is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either
7 * version 2 of the License, or (at your option) any later version.
8 *
9 * This library is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * Lesser General Public License for more details.
13 *
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with this library; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17 */
18
19 #include <stdlib.h>
20 #include <stdio.h>
21 #include "../dsputil.h"
22 #include "../mpegvideo.h"
23 #include "dsputil_altivec.h"
24
25 // Swaps two variables (used for altivec registers)
26 #define SWAP(a,b) \
27 do { \
28 __typeof__(a) swap_temp=a; \
29 a=b; \
30 b=swap_temp; \
31 } while (0)
32
33 // transposes a matrix consisting of four vectors with four elements each
34 #define TRANSPOSE4(a,b,c,d) \
35 do { \
36 __typeof__(a) _trans_ach = vec_mergeh(a, c); \
37 __typeof__(a) _trans_acl = vec_mergel(a, c); \
38 __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
39 __typeof__(a) _trans_bdl = vec_mergel(b, d); \
40 \
41 a = vec_mergeh(_trans_ach, _trans_bdh); \
42 b = vec_mergel(_trans_ach, _trans_bdh); \
43 c = vec_mergeh(_trans_acl, _trans_bdl); \
44 d = vec_mergel(_trans_acl, _trans_bdl); \
45 } while (0)
46
47 #define TRANSPOSE8(a,b,c,d,e,f,g,h) \
48 do { \
49 __typeof__(a) _A1, _B1, _C1, _D1, _E1, _F1, _G1, _H1; \
50 __typeof__(a) _A2, _B2, _C2, _D2, _E2, _F2, _G2, _H2; \
51 \
52 _A1 = vec_mergeh (a, e); \
53 _B1 = vec_mergel (a, e); \
54 _C1 = vec_mergeh (b, f); \
55 _D1 = vec_mergel (b, f); \
56 _E1 = vec_mergeh (c, g); \
57 _F1 = vec_mergel (c, g); \
58 _G1 = vec_mergeh (d, h); \
59 _H1 = vec_mergel (d, h); \
60 \
61 _A2 = vec_mergeh (_A1, _E1); \
62 _B2 = vec_mergel (_A1, _E1); \
63 _C2 = vec_mergeh (_B1, _F1); \
64 _D2 = vec_mergel (_B1, _F1); \
65 _E2 = vec_mergeh (_C1, _G1); \
66 _F2 = vec_mergel (_C1, _G1); \
67 _G2 = vec_mergeh (_D1, _H1); \
68 _H2 = vec_mergel (_D1, _H1); \
69 \
70 a = vec_mergeh (_A2, _E2); \
71 b = vec_mergel (_A2, _E2); \
72 c = vec_mergeh (_B2, _F2); \
73 d = vec_mergel (_B2, _F2); \
74 e = vec_mergeh (_C2, _G2); \
75 f = vec_mergel (_C2, _G2); \
76 g = vec_mergeh (_D2, _H2); \
77 h = vec_mergel (_D2, _H2); \
78 } while (0)
79
80
81 // Loads a four-byte value (int or float) from the target address
82 // into every element in the target vector. Only works if the
83 // target address is four-byte aligned (which should be always).
84 #define LOAD4(vec, address) \
85 { \
86 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
87 vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
88 vec = vec_ld(0, _load_addr); \
89 vec = vec_perm(vec, vec, _perm_vec); \
90 vec = vec_splat(vec, 0); \
91 }
92
93
94 #ifdef CONFIG_DARWIN
95 #define FOUROF(a) (a)
96 #else
97 // slower, for dumb non-apple GCC
98 #define FOUROF(a) {a,a,a,a}
99 #endif
100 int dct_quantize_altivec(MpegEncContext* s,
101 DCTELEM* data, int n,
102 int qscale, int* overflow)
103 {
104 int lastNonZero;
105 vector float row0, row1, row2, row3, row4, row5, row6, row7;
106 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
107 const vector float zero = (const vector float)FOUROF(0.);
108
109 // Load the data into the row/alt vectors
110 {
111 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
112
113 data0 = vec_ld(0, data);
114 data1 = vec_ld(16, data);
115 data2 = vec_ld(32, data);
116 data3 = vec_ld(48, data);
117 data4 = vec_ld(64, data);
118 data5 = vec_ld(80, data);
119 data6 = vec_ld(96, data);
120 data7 = vec_ld(112, data);
121
122 // Transpose the data before we start
123 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
124
125 // load the data into floating point vectors. We load
126 // the high half of each row into the main row vectors
127 // and the low half into the alt vectors.
128 row0 = vec_ctf(vec_unpackh(data0), 0);
129 alt0 = vec_ctf(vec_unpackl(data0), 0);
130 row1 = vec_ctf(vec_unpackh(data1), 0);
131 alt1 = vec_ctf(vec_unpackl(data1), 0);
132 row2 = vec_ctf(vec_unpackh(data2), 0);
133 alt2 = vec_ctf(vec_unpackl(data2), 0);
134 row3 = vec_ctf(vec_unpackh(data3), 0);
135 alt3 = vec_ctf(vec_unpackl(data3), 0);
136 row4 = vec_ctf(vec_unpackh(data4), 0);
137 alt4 = vec_ctf(vec_unpackl(data4), 0);
138 row5 = vec_ctf(vec_unpackh(data5), 0);
139 alt5 = vec_ctf(vec_unpackl(data5), 0);
140 row6 = vec_ctf(vec_unpackh(data6), 0);
141 alt6 = vec_ctf(vec_unpackl(data6), 0);
142 row7 = vec_ctf(vec_unpackh(data7), 0);
143 alt7 = vec_ctf(vec_unpackl(data7), 0);
144 }
145
146 // The following block could exist as a separate an altivec dct
147 // function. However, if we put it inline, the DCT data can remain
148 // in the vector local variables, as floats, which we'll use during the
149 // quantize step...
150 {
151 const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
152 const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
153 const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
154 const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
155 const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
156 const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
157 const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
158 const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
159 const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
160 const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
161 const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
162 const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
163
164
165 int whichPass, whichHalf;
166
167 for(whichPass = 1; whichPass<=2; whichPass++)
168 {
169 for(whichHalf = 1; whichHalf<=2; whichHalf++)
170 {
171 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
172 vector float tmp10, tmp11, tmp12, tmp13;
173 vector float z1, z2, z3, z4, z5;
174
175 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
176 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
177 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
178 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
179 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
180 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
181 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
182 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
183
184 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
185 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
186 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
187 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
188
189
190 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
191 row0 = vec_add(tmp10, tmp11);
192
193 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
194 row4 = vec_sub(tmp10, tmp11);
195
196
197 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
198 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
199
200 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
201 // CONST_BITS-PASS1_BITS);
202 row2 = vec_madd(tmp13, vec_0_765366865, z1);
203
204 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
205 // CONST_BITS-PASS1_BITS);
206 row6 = vec_madd(tmp12, vec_1_847759065, z1);
207
208 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
209 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
210 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
211 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
212
213 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
214 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
215
216 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
217 z3 = vec_madd(z3, vec_1_961570560, z5);
218
219 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
220 z4 = vec_madd(z4, vec_0_390180644, z5);
221
222 // The following adds are rolled into the multiplies above
223 // z3 = vec_add(z3, z5); // z3 += z5;
224 // z4 = vec_add(z4, z5); // z4 += z5;
225
226 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
227 // Wow! It's actually more effecient to roll this multiply
228 // into the adds below, even thought the multiply gets done twice!
229 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
230
231 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
232 // Same with this one...
233 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
234
235 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
236 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
237 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
238
239 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
240 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
241 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
242
243 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
244 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
245 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
246
247 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
248 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
249 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
250
251 // Swap the row values with the alts. If this is the first half,
252 // this sets up the low values to be acted on in the second half.
253 // If this is the second half, it puts the high values back in
254 // the row values where they are expected to be when we're done.
255 SWAP(row0, alt0);
256 SWAP(row1, alt1);
257 SWAP(row2, alt2);
258 SWAP(row3, alt3);
259 SWAP(row4, alt4);
260 SWAP(row5, alt5);
261 SWAP(row6, alt6);
262 SWAP(row7, alt7);
263 }
264
265 if (whichPass == 1)
266 {
267 // transpose the data for the second pass
268
269 // First, block transpose the upper right with lower left.
270 SWAP(row4, alt0);
271 SWAP(row5, alt1);
272 SWAP(row6, alt2);
273 SWAP(row7, alt3);
274
275 // Now, transpose each block of four
276 TRANSPOSE4(row0, row1, row2, row3);
277 TRANSPOSE4(row4, row5, row6, row7);
278 TRANSPOSE4(alt0, alt1, alt2, alt3);
279 TRANSPOSE4(alt4, alt5, alt6, alt7);
280 }
281 }
282 }
283
284 // used after quantise step
285 int oldBaseValue = 0;
286
287 // perform the quantise step, using the floating point data
288 // still in the row/alt registers
289 {
290 const int* biasAddr;
291 const vector signed int* qmat;
292 vector float bias, negBias;
293
294 if (s->mb_intra)
295 {
296 vector signed int baseVector;
297
298 // We must cache element 0 in the intra case
299 // (it needs special handling).
300 baseVector = vec_cts(vec_splat(row0, 0), 0);
301 vec_ste(baseVector, 0, &oldBaseValue);
302
303 qmat = (vector signed int*)s->q_intra_matrix[qscale];
304 biasAddr = &(s->intra_quant_bias);
305 }
306 else
307 {
308 qmat = (vector signed int*)s->q_inter_matrix[qscale];
309 biasAddr = &(s->inter_quant_bias);
310 }
311
312 // Load the bias vector (We add 0.5 to the bias so that we're
313 // rounding when we convert to int, instead of flooring.)
314 {
315 vector signed int biasInt;
316 const vector float negOneFloat = (vector float)FOUROF(-1.0f);
317 LOAD4(biasInt, biasAddr);
318 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
319 negBias = vec_madd(bias, negOneFloat, zero);
320 }
321
322 {
323 vector float q0, q1, q2, q3, q4, q5, q6, q7;
324
325 q0 = vec_ctf(qmat[0], QMAT_SHIFT);
326 q1 = vec_ctf(qmat[2], QMAT_SHIFT);
327 q2 = vec_ctf(qmat[4], QMAT_SHIFT);
328 q3 = vec_ctf(qmat[6], QMAT_SHIFT);
329 q4 = vec_ctf(qmat[8], QMAT_SHIFT);
330 q5 = vec_ctf(qmat[10], QMAT_SHIFT);
331 q6 = vec_ctf(qmat[12], QMAT_SHIFT);
332 q7 = vec_ctf(qmat[14], QMAT_SHIFT);
333
334 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
335 vec_cmpgt(row0, zero));
336 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
337 vec_cmpgt(row1, zero));
338 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
339 vec_cmpgt(row2, zero));
340 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
341 vec_cmpgt(row3, zero));
342 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
343 vec_cmpgt(row4, zero));
344 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
345 vec_cmpgt(row5, zero));
346 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
347 vec_cmpgt(row6, zero));
348 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
349 vec_cmpgt(row7, zero));
350
351 q0 = vec_ctf(qmat[1], QMAT_SHIFT);
352 q1 = vec_ctf(qmat[3], QMAT_SHIFT);
353 q2 = vec_ctf(qmat[5], QMAT_SHIFT);
354 q3 = vec_ctf(qmat[7], QMAT_SHIFT);
355 q4 = vec_ctf(qmat[9], QMAT_SHIFT);
356 q5 = vec_ctf(qmat[11], QMAT_SHIFT);
357 q6 = vec_ctf(qmat[13], QMAT_SHIFT);
358 q7 = vec_ctf(qmat[15], QMAT_SHIFT);
359
360 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
361 vec_cmpgt(alt0, zero));
362 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
363 vec_cmpgt(alt1, zero));
364 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
365 vec_cmpgt(alt2, zero));
366 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
367 vec_cmpgt(alt3, zero));
368 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
369 vec_cmpgt(alt4, zero));
370 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
371 vec_cmpgt(alt5, zero));
372 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
373 vec_cmpgt(alt6, zero));
374 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
375 vec_cmpgt(alt7, zero));
376 }
377
378
379 }
380
381 // Store the data back into the original block
382 {
383 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
384
385 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
386 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
387 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
388 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
389 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
390 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
391 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
392 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
393
394 {
395 // Clamp for overflow
396 vector signed int max_q_int, min_q_int;
397 vector signed short max_q, min_q;
398
399 LOAD4(max_q_int, &(s->max_qcoeff));
400 LOAD4(min_q_int, &(s->min_qcoeff));
401
402 max_q = vec_pack(max_q_int, max_q_int);
403 min_q = vec_pack(min_q_int, min_q_int);
404
405 data0 = vec_max(vec_min(data0, max_q), min_q);
406 data1 = vec_max(vec_min(data1, max_q), min_q);
407 data2 = vec_max(vec_min(data2, max_q), min_q);
408 data4 = vec_max(vec_min(data4, max_q), min_q);
409 data5 = vec_max(vec_min(data5, max_q), min_q);
410 data6 = vec_max(vec_min(data6, max_q), min_q);
411 data7 = vec_max(vec_min(data7, max_q), min_q);
412 }
413
414 vector bool char zero_01, zero_23, zero_45, zero_67;
415 vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
416 vector signed char negOne = vec_splat_s8(-1);
417 vector signed char* scanPtr =
418 (vector signed char*)(s->intra_scantable.inverse);
419
420 // Determine the largest non-zero index.
421 zero_01 = vec_pack(vec_cmpeq(data0, (vector short)zero),
422 vec_cmpeq(data1, (vector short)zero));
423 zero_23 = vec_pack(vec_cmpeq(data2, (vector short)zero),
424 vec_cmpeq(data3, (vector short)zero));
425 zero_45 = vec_pack(vec_cmpeq(data4, (vector short)zero),
426 vec_cmpeq(data5, (vector short)zero));
427 zero_67 = vec_pack(vec_cmpeq(data6, (vector short)zero),
428 vec_cmpeq(data7, (vector short)zero));
429
430 // 64 biggest values
431 scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
432 scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
433 scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
434 scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
435
436 // 32 largest values
437 scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
438 scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
439
440 // 16 largest values
441 scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
442
443 // 8 largest values
444 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
445 vec_mergel(scanIndices_01, negOne));
446
447 // 4 largest values
448 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
449 vec_mergel(scanIndices_01, negOne));
450
451 // 2 largest values
452 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
453 vec_mergel(scanIndices_01, negOne));
454
455 // largest value
456 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
457 vec_mergel(scanIndices_01, negOne));
458
459 scanIndices_01 = vec_splat(scanIndices_01, 0);
460
461 signed char lastNonZeroChar;
462
463 vec_ste(scanIndices_01, 0, &lastNonZeroChar);
464
465 lastNonZero = lastNonZeroChar;
466
467 // While the data is still in vectors we check for the transpose IDCT permute
468 // and handle it using the vector unit if we can. This is the permute used
469 // by the altivec idct, so it is common when using the altivec dct.
470
471 if ((lastNonZero > 0) && (s->idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
472 {
473 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
474 }
475
476 vec_st(data0, 0, data);
477 vec_st(data1, 16, data);
478 vec_st(data2, 32, data);
479 vec_st(data3, 48, data);
480 vec_st(data4, 64, data);
481 vec_st(data5, 80, data);
482 vec_st(data6, 96, data);
483 vec_st(data7, 112, data);
484 }
485
486 // special handling of block[0]
487 if (s->mb_intra)
488 {
489 if (!s->h263_aic)
490 {
491 if (n < 4)
492 oldBaseValue /= s->y_dc_scale;
493 else
494 oldBaseValue /= s->c_dc_scale;
495 }
496
497 // Divide by 8, rounding the result
498 data[0] = (oldBaseValue + 4) >> 3;
499 }
500
501 // We handled the tranpose permutation above and we don't
502 // need to permute the "no" permutation case.
503 if ((lastNonZero > 0) &&
504 (s->idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
505 (s->idct_permutation_type != FF_NO_IDCT_PERM))
506 {
507 ff_block_permute(data, s->idct_permutation,
508 s->intra_scantable.scantable, lastNonZero);
509 }
510
511 return lastNonZero;
512 }
513 #undef FOUROF
514
515 /*
516 AltiVec version of dct_unquantize_h263
517 this code assumes `block' is 16 bytes-aligned
518 */
519 void dct_unquantize_h263_altivec(MpegEncContext *s,
520 DCTELEM *block, int n, int qscale)
521 {
522 POWERPC_TBL_DECLARE(altivec_dct_unquantize_h263_num, 1);
523 int i, level, qmul, qadd;
524 int nCoeffs;
525
526 assert(s->block_last_index[n]>=0);
527
528 POWERPC_TBL_START_COUNT(altivec_dct_unquantize_h263_num, 1);
529
530 qadd = (qscale - 1) | 1;
531 qmul = qscale << 1;
532
533 if (s->mb_intra) {
534 if (!s->h263_aic) {
535 if (n < 4)
536 block[0] = block[0] * s->y_dc_scale;
537 else
538 block[0] = block[0] * s->c_dc_scale;
539 }else
540 qadd = 0;
541 i = 1;
542 nCoeffs= 63; //does not allways use zigzag table
543 } else {
544 i = 0;
545 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
546 }
547
548 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
549 for(;i<=nCoeffs;i++) {
550 level = block[i];
551 if (level) {
552 if (level < 0) {
553 level = level * qmul - qadd;
554 } else {
555 level = level * qmul + qadd;
556 }
557 block[i] = level;
558 }
559 }
560 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
561 {
562 register const vector short vczero = (const vector short)vec_splat_s16(0);
563 short __attribute__ ((aligned(16))) qmul8[] =
564 {
565 qmul, qmul, qmul, qmul,
566 qmul, qmul, qmul, qmul
567 };
568 short __attribute__ ((aligned(16))) qadd8[] =
569 {
570 qadd, qadd, qadd, qadd,
571 qadd, qadd, qadd, qadd
572 };
573 short __attribute__ ((aligned(16))) nqadd8[] =
574 {
575 -qadd, -qadd, -qadd, -qadd,
576 -qadd, -qadd, -qadd, -qadd
577 };
578 register vector short blockv, qmulv, qaddv, nqaddv, temp1;
579 register vector bool short blockv_null, blockv_neg;
580 register short backup_0 = block[0];
581 register int j = 0;
582
583 qmulv = vec_ld(0, qmul8);
584 qaddv = vec_ld(0, qadd8);
585 nqaddv = vec_ld(0, nqadd8);
586
587 #if 0 // block *is* 16 bytes-aligned, it seems.
588 // first make sure block[j] is 16 bytes-aligned
589 for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
590 level = block[j];
591 if (level) {
592 if (level < 0) {
593 level = level * qmul - qadd;
594 } else {
595 level = level * qmul + qadd;
596 }
597 block[j] = level;
598 }
599 }
600 #endif
601
602 // vectorize all the 16 bytes-aligned blocks
603 // of 8 elements
604 for(; (j + 7) <= nCoeffs ; j+=8)
605 {
606 blockv = vec_ld(j << 1, block);
607 blockv_neg = vec_cmplt(blockv, vczero);
608 blockv_null = vec_cmpeq(blockv, vczero);
609 // choose between +qadd or -qadd as the third operand
610 temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
611 // multiply & add (block{i,i+7} * qmul [+-] qadd)
612 temp1 = vec_mladd(blockv, qmulv, temp1);
613 // put 0 where block[{i,i+7} used to have 0
614 blockv = vec_sel(temp1, blockv, blockv_null);
615 vec_st(blockv, j << 1, block);
616 }
617
618 // if nCoeffs isn't a multiple of 8, finish the job
619 // using good old scalar units.
620 // (we could do it using a truncated vector,
621 // but I'm not sure it's worth the hassle)
622 for(; j <= nCoeffs ; j++) {
623 level = block[j];
624 if (level) {
625 if (level < 0) {
626 level = level * qmul - qadd;
627 } else {
628 level = level * qmul + qadd;
629 }
630 block[j] = level;
631 }
632 }
633
634 if (i == 1)
635 { // cheat. this avoid special-casing the first iteration
636 block[0] = backup_0;
637 }
638 }
639 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
640
641 POWERPC_TBL_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
642 }