Mercurial > libavcodec.hg

comparison ppc/mpegvideo_altivec.c @ 828:ace3ccd18dd2 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
author michaelni
date Sat, 02 Nov 2002 11:28:08 +0000
parents
children 8bec850dc9c7
comparison
equal deleted inserted replaced
827:770578c6c300 828:ace3ccd18dd2
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
24
25 // Used when initializing constant vectors
26 #define FOUR_INSTANCES(x) x,x,x,x
27
28 // Swaps two variables (used for altivec registers)
29 #define SWAP(a,b) \
30 do { \
31 __typeof__(a) swap_temp=a; \
32 a=b; \
33 b=swap_temp; \
34 } while (0)
35
36 // transposes a matrix consisting of four vectors with four elements each
37 #define TRANSPOSE4(a,b,c,d) \
38 do { \
39 __typeof__(a) _trans_ach = vec_mergeh(a, c); \
40 __typeof__(a) _trans_acl = vec_mergel(a, c); \
41 __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
42 __typeof__(a) _trans_bdl = vec_mergel(b, d); \
43 \
44 a = vec_mergeh(_trans_ach, _trans_bdh); \
45 b = vec_mergel(_trans_ach, _trans_bdh); \
46 c = vec_mergeh(_trans_acl, _trans_bdl); \
47 d = vec_mergel(_trans_acl, _trans_bdl); \
48 } while (0)
49
50 #define TRANSPOSE8(a,b,c,d,e,f,g,h) \
51 do { \
52 __typeof__(a) _A1, _B1, _C1, _D1, _E1, _F1, _G1, _H1; \
53 __typeof__(a) _A2, _B2, _C2, _D2, _E2, _F2, _G2, _H2; \
54 \
55 _A1 = vec_mergeh (a, e); \
56 _B1 = vec_mergel (a, e); \
57 _C1 = vec_mergeh (b, f); \
58 _D1 = vec_mergel (b, f); \
59 _E1 = vec_mergeh (c, g); \
60 _F1 = vec_mergel (c, g); \
61 _G1 = vec_mergeh (d, h); \
62 _H1 = vec_mergel (d, h); \
63 \
64 _A2 = vec_mergeh (_A1, _E1); \
65 _B2 = vec_mergel (_A1, _E1); \
66 _C2 = vec_mergeh (_B1, _F1); \
67 _D2 = vec_mergel (_B1, _F1); \
68 _E2 = vec_mergeh (_C1, _G1); \
69 _F2 = vec_mergel (_C1, _G1); \
70 _G2 = vec_mergeh (_D1, _H1); \
71 _H2 = vec_mergel (_D1, _H1); \
72 \
73 a = vec_mergeh (_A2, _E2); \
74 b = vec_mergel (_A2, _E2); \
75 c = vec_mergeh (_B2, _F2); \
76 d = vec_mergel (_B2, _F2); \
77 e = vec_mergeh (_C2, _G2); \
78 f = vec_mergel (_C2, _G2); \
79 g = vec_mergeh (_D2, _H2); \
80 h = vec_mergel (_D2, _H2); \
81 } while (0)
82
83
84 // Loads a four-byte value (int or float) from the target address
85 // into every element in the target vector. Only works if the
86 // target address is four-byte aligned (which should be always).
87 #define LOAD4(vec, address) \
88 { \
89 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
90 vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
91 vec = vec_ld(0, _load_addr); \
92 vec = vec_perm(vec, vec, _perm_vec); \
93 vec = vec_splat(vec, 0); \
94 }
95
96 int dct_quantize_altivec(MpegEncContext* s,
97 DCTELEM* data, int n,
98 int qscale, int* overflow)
99 {
100 int lastNonZero;
101 vector float row0, row1, row2, row3, row4, row5, row6, row7;
102 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
103 const vector float zero = {FOUR_INSTANCES(0.0f)};
104
105 // Load the data into the row/alt vectors
106 {
107 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
108
109 data0 = vec_ld(0, data);
110 data1 = vec_ld(16, data);
111 data2 = vec_ld(32, data);
112 data3 = vec_ld(48, data);
113 data4 = vec_ld(64, data);
114 data5 = vec_ld(80, data);
115 data6 = vec_ld(96, data);
116 data7 = vec_ld(112, data);
117
118 // Transpose the data before we start
119 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
120
121 // load the data into floating point vectors. We load
122 // the high half of each row into the main row vectors
123 // and the low half into the alt vectors.
124 row0 = vec_ctf(vec_unpackh(data0), 0);
125 alt0 = vec_ctf(vec_unpackl(data0), 0);
126 row1 = vec_ctf(vec_unpackh(data1), 0);
127 alt1 = vec_ctf(vec_unpackl(data1), 0);
128 row2 = vec_ctf(vec_unpackh(data2), 0);
129 alt2 = vec_ctf(vec_unpackl(data2), 0);
130 row3 = vec_ctf(vec_unpackh(data3), 0);
131 alt3 = vec_ctf(vec_unpackl(data3), 0);
132 row4 = vec_ctf(vec_unpackh(data4), 0);
133 alt4 = vec_ctf(vec_unpackl(data4), 0);
134 row5 = vec_ctf(vec_unpackh(data5), 0);
135 alt5 = vec_ctf(vec_unpackl(data5), 0);
136 row6 = vec_ctf(vec_unpackh(data6), 0);
137 alt6 = vec_ctf(vec_unpackl(data6), 0);
138 row7 = vec_ctf(vec_unpackh(data7), 0);
139 alt7 = vec_ctf(vec_unpackl(data7), 0);
140 }
141
142 // The following block could exist as a separate an altivec dct
143 // function. However, if we put it inline, the DCT data can remain
144 // in the vector local variables, as floats, which we'll use during the
145 // quantize step...
146 {
147 const vector float vec_0_298631336 = {FOUR_INSTANCES(0.298631336f)};
148 const vector float vec_0_390180644 = {FOUR_INSTANCES(-0.390180644f)};
149 const vector float vec_0_541196100 = {FOUR_INSTANCES(0.541196100f)};
150 const vector float vec_0_765366865 = {FOUR_INSTANCES(0.765366865f)};
151 const vector float vec_0_899976223 = {FOUR_INSTANCES(-0.899976223f)};
152 const vector float vec_1_175875602 = {FOUR_INSTANCES(1.175875602f)};
153 const vector float vec_1_501321110 = {FOUR_INSTANCES(1.501321110f)};
154 const vector float vec_1_847759065 = {FOUR_INSTANCES(-1.847759065f)};
155 const vector float vec_1_961570560 = {FOUR_INSTANCES(-1.961570560f)};
156 const vector float vec_2_053119869 = {FOUR_INSTANCES(2.053119869f)};
157 const vector float vec_2_562915447 = {FOUR_INSTANCES(-2.562915447f)};
158 const vector float vec_3_072711026 = {FOUR_INSTANCES(3.072711026f)};
159
160
161 int whichPass, whichHalf;
162
163 for(whichPass = 1; whichPass<=2; whichPass++)
164 {
165 for(whichHalf = 1; whichHalf<=2; whichHalf++)
166 {
167 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
168 vector float tmp10, tmp11, tmp12, tmp13;
169 vector float z1, z2, z3, z4, z5;
170
171 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
172 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
173 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
174 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
175 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
176 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
177 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
178 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
179
180 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
181 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
182 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
183 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
184
185
186 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
187 row0 = vec_add(tmp10, tmp11);
188
189 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
190 row4 = vec_sub(tmp10, tmp11);
191
192
193 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
194 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
195
196 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
197 // CONST_BITS-PASS1_BITS);
198 row2 = vec_madd(tmp13, vec_0_765366865, z1);
199
200 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
201 // CONST_BITS-PASS1_BITS);
202 row6 = vec_madd(tmp12, vec_1_847759065, z1);
203
204 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
205 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
206 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
207 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
208
209 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
210 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
211
212 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
213 z3 = vec_madd(z3, vec_1_961570560, z5);
214
215 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
216 z4 = vec_madd(z4, vec_0_390180644, z5);
217
218 // The following adds are rolled into the multiplies above
219 // z3 = vec_add(z3, z5); // z3 += z5;
220 // z4 = vec_add(z4, z5); // z4 += z5;
221
222 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
223 // Wow! It's actually more effecient to roll this multiply
224 // into the adds below, even thought the multiply gets done twice!
225 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
226
227 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
228 // Same with this one...
229 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
230
231 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
232 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
233 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
234
235 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
236 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
237 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
238
239 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
240 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
241 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
242
243 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
244 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
245 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
246
247 // Swap the row values with the alts. If this is the first half,
248 // this sets up the low values to be acted on in the second half.
249 // If this is the second half, it puts the high values back in
250 // the row values where they are expected to be when we're done.
251 SWAP(row0, alt0);
252 SWAP(row1, alt1);
253 SWAP(row2, alt2);
254 SWAP(row3, alt3);
255 SWAP(row4, alt4);
256 SWAP(row5, alt5);
257 SWAP(row6, alt6);
258 SWAP(row7, alt7);
259 }
260
261 if (whichPass == 1)
262 {
263 // transpose the data for the second pass
264
265 // First, block transpose the upper right with lower left.
266 SWAP(row4, alt0);
267 SWAP(row5, alt1);
268 SWAP(row6, alt2);
269 SWAP(row7, alt3);
270
271 // Now, transpose each block of four
272 TRANSPOSE4(row0, row1, row2, row3);
273 TRANSPOSE4(row4, row5, row6, row7);
274 TRANSPOSE4(alt0, alt1, alt2, alt3);
275 TRANSPOSE4(alt4, alt5, alt6, alt7);
276 }
277 }
278 }
279
280 // used after quantise step
281 int oldBaseValue = 0;
282
283 // perform the quantise step, using the floating point data
284 // still in the row/alt registers
285 {
286 const int* biasAddr;
287 const vector signed int* qmat;
288 vector float bias, negBias;
289
290 if (s->mb_intra)
291 {
292 vector signed int baseVector;
293
294 // We must cache element 0 in the intra case
295 // (it needs special handling).
296 baseVector = vec_cts(vec_splat(row0, 0), 0);
297 vec_ste(baseVector, 0, &oldBaseValue);
298
299 qmat = (vector signed int*)s->q_intra_matrix[qscale];
300 biasAddr = &(s->intra_quant_bias);
301 }
302 else
303 {
304 qmat = (vector signed int*)s->q_inter_matrix[qscale];
305 biasAddr = &(s->inter_quant_bias);
306 }
307
308 // Load the bias vector (We add 0.5 to the bias so that we're
309 // rounding when we convert to int, instead of flooring.)
310 {
311 vector signed int biasInt;
312 const vector float negOneFloat = (vector float)(FOUR_INSTANCES(-1.0f));
313 LOAD4(biasInt, biasAddr);
314 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
315 negBias = vec_madd(bias, negOneFloat, zero);
316 }
317
318 {
319 vector float q0, q1, q2, q3, q4, q5, q6, q7;
320
321 q0 = vec_ctf(qmat[0], QMAT_SHIFT);
322 q1 = vec_ctf(qmat[2], QMAT_SHIFT);
323 q2 = vec_ctf(qmat[4], QMAT_SHIFT);
324 q3 = vec_ctf(qmat[6], QMAT_SHIFT);
325 q4 = vec_ctf(qmat[8], QMAT_SHIFT);
326 q5 = vec_ctf(qmat[10], QMAT_SHIFT);
327 q6 = vec_ctf(qmat[12], QMAT_SHIFT);
328 q7 = vec_ctf(qmat[14], QMAT_SHIFT);
329
330 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
331 vec_cmpgt(row0, zero));
332 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
333 vec_cmpgt(row1, zero));
334 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
335 vec_cmpgt(row2, zero));
336 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
337 vec_cmpgt(row3, zero));
338 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
339 vec_cmpgt(row4, zero));
340 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
341 vec_cmpgt(row5, zero));
342 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
343 vec_cmpgt(row6, zero));
344 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
345 vec_cmpgt(row7, zero));
346
347 q0 = vec_ctf(qmat[1], QMAT_SHIFT);
348 q1 = vec_ctf(qmat[3], QMAT_SHIFT);
349 q2 = vec_ctf(qmat[5], QMAT_SHIFT);
350 q3 = vec_ctf(qmat[7], QMAT_SHIFT);
351 q4 = vec_ctf(qmat[9], QMAT_SHIFT);
352 q5 = vec_ctf(qmat[11], QMAT_SHIFT);
353 q6 = vec_ctf(qmat[13], QMAT_SHIFT);
354 q7 = vec_ctf(qmat[15], QMAT_SHIFT);
355
356 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
357 vec_cmpgt(alt0, zero));
358 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
359 vec_cmpgt(alt1, zero));
360 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
361 vec_cmpgt(alt2, zero));
362 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
363 vec_cmpgt(alt3, zero));
364 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
365 vec_cmpgt(alt4, zero));
366 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
367 vec_cmpgt(alt5, zero));
368 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
369 vec_cmpgt(alt6, zero));
370 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
371 vec_cmpgt(alt7, zero));
372 }
373
374
375 }
376
377 // Store the data back into the original block
378 {
379 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
380
381 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
382 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
383 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
384 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
385 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
386 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
387 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
388 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
389
390 {
391 // Clamp for overflow
392 vector signed int max_q_int, min_q_int;
393 vector signed short max_q, min_q;
394
395 LOAD4(max_q_int, &(s->max_qcoeff));
396 LOAD4(min_q_int, &(s->min_qcoeff));
397
398 max_q = vec_pack(max_q_int, max_q_int);
399 min_q = vec_pack(min_q_int, min_q_int);
400
401 data0 = vec_max(vec_min(data0, max_q), min_q);
402 data1 = vec_max(vec_min(data1, max_q), min_q);
403 data2 = vec_max(vec_min(data2, max_q), min_q);
404 data4 = vec_max(vec_min(data4, max_q), min_q);
405 data5 = vec_max(vec_min(data5, max_q), min_q);
406 data6 = vec_max(vec_min(data6, max_q), min_q);
407 data7 = vec_max(vec_min(data7, max_q), min_q);
408 }
409
410 vector bool char zero_01, zero_23, zero_45, zero_67;
411 vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
412 vector signed char negOne = vec_splat_s8(-1);
413 vector signed char* scanPtr =
414 (vector signed char*)(s->intra_scantable.inverse);
415
416 // Determine the largest non-zero index.
417 zero_01 = vec_pack(vec_cmpeq(data0, (vector short)zero),
418 vec_cmpeq(data1, (vector short)zero));
419 zero_23 = vec_pack(vec_cmpeq(data2, (vector short)zero),
420 vec_cmpeq(data3, (vector short)zero));
421 zero_45 = vec_pack(vec_cmpeq(data4, (vector short)zero),
422 vec_cmpeq(data5, (vector short)zero));
423 zero_67 = vec_pack(vec_cmpeq(data6, (vector short)zero),
424 vec_cmpeq(data7, (vector short)zero));
425
426 // 64 biggest values
427 scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
428 scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
429 scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
430 scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
431
432 // 32 largest values
433 scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
434 scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
435
436 // 16 largest values
437 scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
438
439 // 8 largest values
440 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
441 vec_mergel(scanIndices_01, negOne));
442
443 // 4 largest values
444 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
445 vec_mergel(scanIndices_01, negOne));
446
447 // 2 largest values
448 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
449 vec_mergel(scanIndices_01, negOne));
450
451 // largest value
452 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
453 vec_mergel(scanIndices_01, negOne));
454
455 scanIndices_01 = vec_splat(scanIndices_01, 0);
456
457 signed char lastNonZeroChar;
458
459 vec_ste(scanIndices_01, 0, &lastNonZeroChar);
460
461 lastNonZero = lastNonZeroChar;
462
463 // While the data is still in vectors we check for the transpose IDCT permute
464 // and handle it using the vector unit if we can. This is the permute used
465 // by the altivec idct, so it is common when using the altivec dct.
466
467 if ((lastNonZero > 0) && (s->idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
468 {
469 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
470 }
471
472 vec_st(data0, 0, data);
473 vec_st(data1, 16, data);
474 vec_st(data2, 32, data);
475 vec_st(data3, 48, data);
476 vec_st(data4, 64, data);
477 vec_st(data5, 80, data);
478 vec_st(data6, 96, data);
479 vec_st(data7, 112, data);
480 }
481
482 // special handling of block[0]
483 if (s->mb_intra)
484 {
485 if (!s->h263_aic)
486 {
487 if (n < 4)
488 oldBaseValue /= s->y_dc_scale;
489 else
490 oldBaseValue /= s->c_dc_scale;
491 }
492
493 // Divide by 8, rounding the result
494 data[0] = (oldBaseValue + 4) >> 3;
495 }
496
497 // We handled the tranpose permutation above and we don't
498 // need to permute the "no" permutation case.
499 if ((lastNonZero > 0) &&
500 (s->idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
501 (s->idct_permutation_type != FF_NO_IDCT_PERM))
502 {
503 ff_block_permute(data, s->idct_permutation,
504 s->intra_scantable.scantable, lastNonZero);
505 }
506
507 return lastNonZero;
508 }
509