ae3170d91a2b15178e98c0d2679a573a0c3736d1
[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
24 #include "gcc_fixes.h"
25
26 #include "dsputil_altivec.h"
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
97 #ifdef CONFIG_DARWIN
98 #define FOUROF(a) (a)
99 #else
100 // slower, for dumb non-apple GCC
101 #define FOUROF(a) {a,a,a,a}
102 #endif
103 int dct_quantize_altivec(MpegEncContext* s,
104 DCTELEM* data, int n,
105 int qscale, int* overflow)
106 {
107 int lastNonZero;
108 vector float row0, row1, row2, row3, row4, row5, row6, row7;
109 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
110 const vector float zero = (const vector float)FOUROF(0.);
111
112 // Load the data into the row/alt vectors
113 {
114 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
115
116 data0 = vec_ld(0, data);
117 data1 = vec_ld(16, data);
118 data2 = vec_ld(32, data);
119 data3 = vec_ld(48, data);
120 data4 = vec_ld(64, data);
121 data5 = vec_ld(80, data);
122 data6 = vec_ld(96, data);
123 data7 = vec_ld(112, data);
124
125 // Transpose the data before we start
126 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
127
128 // load the data into floating point vectors. We load
129 // the high half of each row into the main row vectors
130 // and the low half into the alt vectors.
131 row0 = vec_ctf(vec_unpackh(data0), 0);
132 alt0 = vec_ctf(vec_unpackl(data0), 0);
133 row1 = vec_ctf(vec_unpackh(data1), 0);
134 alt1 = vec_ctf(vec_unpackl(data1), 0);
135 row2 = vec_ctf(vec_unpackh(data2), 0);
136 alt2 = vec_ctf(vec_unpackl(data2), 0);
137 row3 = vec_ctf(vec_unpackh(data3), 0);
138 alt3 = vec_ctf(vec_unpackl(data3), 0);
139 row4 = vec_ctf(vec_unpackh(data4), 0);
140 alt4 = vec_ctf(vec_unpackl(data4), 0);
141 row5 = vec_ctf(vec_unpackh(data5), 0);
142 alt5 = vec_ctf(vec_unpackl(data5), 0);
143 row6 = vec_ctf(vec_unpackh(data6), 0);
144 alt6 = vec_ctf(vec_unpackl(data6), 0);
145 row7 = vec_ctf(vec_unpackh(data7), 0);
146 alt7 = vec_ctf(vec_unpackl(data7), 0);
147 }
148
149 // The following block could exist as a separate an altivec dct
150 // function. However, if we put it inline, the DCT data can remain
151 // in the vector local variables, as floats, which we'll use during the
152 // quantize step...
153 {
154 const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
155 const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
156 const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
157 const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
158 const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
159 const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
160 const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
161 const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
162 const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
163 const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
164 const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
165 const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
166
167
168 int whichPass, whichHalf;
169
170 for(whichPass = 1; whichPass<=2; whichPass++)
171 {
172 for(whichHalf = 1; whichHalf<=2; whichHalf++)
173 {
174 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
175 vector float tmp10, tmp11, tmp12, tmp13;
176 vector float z1, z2, z3, z4, z5;
177
178 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
179 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
180 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
181 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
182 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
183 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
184 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
185 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
186
187 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
188 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
189 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
190 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
191
192
193 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
194 row0 = vec_add(tmp10, tmp11);
195
196 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
197 row4 = vec_sub(tmp10, tmp11);
198
199
200 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
201 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
202
203 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
204 // CONST_BITS-PASS1_BITS);
205 row2 = vec_madd(tmp13, vec_0_765366865, z1);
206
207 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
208 // CONST_BITS-PASS1_BITS);
209 row6 = vec_madd(tmp12, vec_1_847759065, z1);
210
211 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
212 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
213 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
214 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
215
216 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
217 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
218
219 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
220 z3 = vec_madd(z3, vec_1_961570560, z5);
221
222 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
223 z4 = vec_madd(z4, vec_0_390180644, z5);
224
225 // The following adds are rolled into the multiplies above
226 // z3 = vec_add(z3, z5); // z3 += z5;
227 // z4 = vec_add(z4, z5); // z4 += z5;
228
229 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
230 // Wow! It's actually more effecient to roll this multiply
231 // into the adds below, even thought the multiply gets done twice!
232 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
233
234 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
235 // Same with this one...
236 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
237
238 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
239 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
240 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
241
242 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
243 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
244 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
245
246 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
247 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
248 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
249
250 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
251 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
252 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
253
254 // Swap the row values with the alts. If this is the first half,
255 // this sets up the low values to be acted on in the second half.
256 // If this is the second half, it puts the high values back in
257 // the row values where they are expected to be when we're done.
258 SWAP(row0, alt0);
259 SWAP(row1, alt1);
260 SWAP(row2, alt2);
261 SWAP(row3, alt3);
262 SWAP(row4, alt4);
263 SWAP(row5, alt5);
264 SWAP(row6, alt6);
265 SWAP(row7, alt7);
266 }
267
268 if (whichPass == 1)
269 {
270 // transpose the data for the second pass
271
272 // First, block transpose the upper right with lower left.
273 SWAP(row4, alt0);
274 SWAP(row5, alt1);
275 SWAP(row6, alt2);
276 SWAP(row7, alt3);
277
278 // Now, transpose each block of four
279 TRANSPOSE4(row0, row1, row2, row3);
280 TRANSPOSE4(row4, row5, row6, row7);
281 TRANSPOSE4(alt0, alt1, alt2, alt3);
282 TRANSPOSE4(alt4, alt5, alt6, alt7);
283 }
284 }
285 }
286
287 // used after quantise step
288 int oldBaseValue = 0;
289
290 // perform the quantise step, using the floating point data
291 // still in the row/alt registers
292 {
293 const int* biasAddr;
294 const vector signed int* qmat;
295 vector float bias, negBias;
296
297 if (s->mb_intra)
298 {
299 vector signed int baseVector;
300
301 // We must cache element 0 in the intra case
302 // (it needs special handling).
303 baseVector = vec_cts(vec_splat(row0, 0), 0);
304 vec_ste(baseVector, 0, &oldBaseValue);
305
306 qmat = (vector signed int*)s->q_intra_matrix[qscale];
307 biasAddr = &(s->intra_quant_bias);
308 }
309 else
310 {
311 qmat = (vector signed int*)s->q_inter_matrix[qscale];
312 biasAddr = &(s->inter_quant_bias);
313 }
314
315 // Load the bias vector (We add 0.5 to the bias so that we're
316 // rounding when we convert to int, instead of flooring.)
317 {
318 vector signed int biasInt;
319 const vector float negOneFloat = (vector float)FOUROF(-1.0f);
320 LOAD4(biasInt, biasAddr);
321 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
322 negBias = vec_madd(bias, negOneFloat, zero);
323 }
324
325 {
326 vector float q0, q1, q2, q3, q4, q5, q6, q7;
327
328 q0 = vec_ctf(qmat[0], QMAT_SHIFT);
329 q1 = vec_ctf(qmat[2], QMAT_SHIFT);
330 q2 = vec_ctf(qmat[4], QMAT_SHIFT);
331 q3 = vec_ctf(qmat[6], QMAT_SHIFT);
332 q4 = vec_ctf(qmat[8], QMAT_SHIFT);
333 q5 = vec_ctf(qmat[10], QMAT_SHIFT);
334 q6 = vec_ctf(qmat[12], QMAT_SHIFT);
335 q7 = vec_ctf(qmat[14], QMAT_SHIFT);
336
337 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
338 vec_cmpgt(row0, zero));
339 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
340 vec_cmpgt(row1, zero));
341 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
342 vec_cmpgt(row2, zero));
343 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
344 vec_cmpgt(row3, zero));
345 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
346 vec_cmpgt(row4, zero));
347 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
348 vec_cmpgt(row5, zero));
349 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
350 vec_cmpgt(row6, zero));
351 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
352 vec_cmpgt(row7, zero));
353
354 q0 = vec_ctf(qmat[1], QMAT_SHIFT);
355 q1 = vec_ctf(qmat[3], QMAT_SHIFT);
356 q2 = vec_ctf(qmat[5], QMAT_SHIFT);
357 q3 = vec_ctf(qmat[7], QMAT_SHIFT);
358 q4 = vec_ctf(qmat[9], QMAT_SHIFT);
359 q5 = vec_ctf(qmat[11], QMAT_SHIFT);
360 q6 = vec_ctf(qmat[13], QMAT_SHIFT);
361 q7 = vec_ctf(qmat[15], QMAT_SHIFT);
362
363 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
364 vec_cmpgt(alt0, zero));
365 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
366 vec_cmpgt(alt1, zero));
367 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
368 vec_cmpgt(alt2, zero));
369 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
370 vec_cmpgt(alt3, zero));
371 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
372 vec_cmpgt(alt4, zero));
373 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
374 vec_cmpgt(alt5, zero));
375 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
376 vec_cmpgt(alt6, zero));
377 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
378 vec_cmpgt(alt7, zero));
379 }
380
381
382 }
383
384 // Store the data back into the original block
385 {
386 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
387
388 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
389 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
390 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
391 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
392 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
393 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
394 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
395 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
396
397 {
398 // Clamp for overflow
399 vector signed int max_q_int, min_q_int;
400 vector signed short max_q, min_q;
401
402 LOAD4(max_q_int, &(s->max_qcoeff));
403 LOAD4(min_q_int, &(s->min_qcoeff));
404
405 max_q = vec_pack(max_q_int, max_q_int);
406 min_q = vec_pack(min_q_int, min_q_int);
407
408 data0 = vec_max(vec_min(data0, max_q), min_q);
409 data1 = vec_max(vec_min(data1, max_q), min_q);
410 data2 = vec_max(vec_min(data2, max_q), min_q);
411 data4 = vec_max(vec_min(data4, max_q), min_q);
412 data5 = vec_max(vec_min(data5, max_q), min_q);
413 data6 = vec_max(vec_min(data6, max_q), min_q);
414 data7 = vec_max(vec_min(data7, max_q), min_q);
415 }
416
417 vector bool char zero_01, zero_23, zero_45, zero_67;
418 vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
419 vector signed char negOne = vec_splat_s8(-1);
420 vector signed char* scanPtr =
421 (vector signed char*)(s->intra_scantable.inverse);
422
423 // Determine the largest non-zero index.
424 zero_01 = vec_pack(vec_cmpeq(data0, (vector short)zero),
425 vec_cmpeq(data1, (vector short)zero));
426 zero_23 = vec_pack(vec_cmpeq(data2, (vector short)zero),
427 vec_cmpeq(data3, (vector short)zero));
428 zero_45 = vec_pack(vec_cmpeq(data4, (vector short)zero),
429 vec_cmpeq(data5, (vector short)zero));
430 zero_67 = vec_pack(vec_cmpeq(data6, (vector short)zero),
431 vec_cmpeq(data7, (vector short)zero));
432
433 // 64 biggest values
434 scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
435 scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
436 scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
437 scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
438
439 // 32 largest values
440 scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
441 scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
442
443 // 16 largest values
444 scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
445
446 // 8 largest values
447 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
448 vec_mergel(scanIndices_01, negOne));
449
450 // 4 largest values
451 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
452 vec_mergel(scanIndices_01, negOne));
453
454 // 2 largest values
455 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
456 vec_mergel(scanIndices_01, negOne));
457
458 // largest value
459 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
460 vec_mergel(scanIndices_01, negOne));
461
462 scanIndices_01 = vec_splat(scanIndices_01, 0);
463
464 signed char lastNonZeroChar;
465
466 vec_ste(scanIndices_01, 0, &lastNonZeroChar);
467
468 lastNonZero = lastNonZeroChar;
469
470 // While the data is still in vectors we check for the transpose IDCT permute
471 // and handle it using the vector unit if we can. This is the permute used
472 // by the altivec idct, so it is common when using the altivec dct.
473
474 if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
475 {
476 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
477 }
478
479 vec_st(data0, 0, data);
480 vec_st(data1, 16, data);
481 vec_st(data2, 32, data);
482 vec_st(data3, 48, data);
483 vec_st(data4, 64, data);
484 vec_st(data5, 80, data);
485 vec_st(data6, 96, data);
486 vec_st(data7, 112, data);
487 }
488
489 // special handling of block[0]
490 if (s->mb_intra)
491 {
492 if (!s->h263_aic)
493 {
494 if (n < 4)
495 oldBaseValue /= s->y_dc_scale;
496 else
497 oldBaseValue /= s->c_dc_scale;
498 }
499
500 // Divide by 8, rounding the result
501 data[0] = (oldBaseValue + 4) >> 3;
502 }
503
504 // We handled the tranpose permutation above and we don't
505 // need to permute the "no" permutation case.
506 if ((lastNonZero > 0) &&
507 (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
508 (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM))
509 {
510 ff_block_permute(data, s->dsp.idct_permutation,
511 s->intra_scantable.scantable, lastNonZero);
512 }
513
514 return lastNonZero;
515 }
516 #undef FOUROF
517
518 /*
519 AltiVec version of dct_unquantize_h263
520 this code assumes `block' is 16 bytes-aligned
521 */
522 void dct_unquantize_h263_altivec(MpegEncContext *s,
523 DCTELEM *block, int n, int qscale)
524 {
525 POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
526 int i, level, qmul, qadd;
527 int nCoeffs;
528
529 assert(s->block_last_index[n]>=0);
530
531 POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
532
533 qadd = (qscale - 1) | 1;
534 qmul = qscale << 1;
535
536 if (s->mb_intra) {
537 if (!s->h263_aic) {
538 if (n < 4)
539 block[0] = block[0] * s->y_dc_scale;
540 else
541 block[0] = block[0] * s->c_dc_scale;
542 }else
543 qadd = 0;
544 i = 1;
545 nCoeffs= 63; //does not allways use zigzag table
546 } else {
547 i = 0;
548 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
549 }
550
551 #ifdef ALTIVEC_USE_REFERENCE_C_CODE
552 for(;i<=nCoeffs;i++) {
553 level = block[i];
554 if (level) {
555 if (level < 0) {
556 level = level * qmul - qadd;
557 } else {
558 level = level * qmul + qadd;
559 }
560 block[i] = level;
561 }
562 }
563 #else /* ALTIVEC_USE_REFERENCE_C_CODE */
564 {
565 register const vector short vczero = (const vector short)vec_splat_s16(0);
566 short __attribute__ ((aligned(16))) qmul8[] =
567 {
568 qmul, qmul, qmul, qmul,
569 qmul, qmul, qmul, qmul
570 };
571 short __attribute__ ((aligned(16))) qadd8[] =
572 {
573 qadd, qadd, qadd, qadd,
574 qadd, qadd, qadd, qadd
575 };
576 short __attribute__ ((aligned(16))) nqadd8[] =
577 {
578 -qadd, -qadd, -qadd, -qadd,
579 -qadd, -qadd, -qadd, -qadd
580 };
581 register vector short blockv, qmulv, qaddv, nqaddv, temp1;
582 register vector bool short blockv_null, blockv_neg;
583 register short backup_0 = block[0];
584 register int j = 0;
585
586 qmulv = vec_ld(0, qmul8);
587 qaddv = vec_ld(0, qadd8);
588 nqaddv = vec_ld(0, nqadd8);
589
590 #if 0 // block *is* 16 bytes-aligned, it seems.
591 // first make sure block[j] is 16 bytes-aligned
592 for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
593 level = block[j];
594 if (level) {
595 if (level < 0) {
596 level = level * qmul - qadd;
597 } else {
598 level = level * qmul + qadd;
599 }
600 block[j] = level;
601 }
602 }
603 #endif
604
605 // vectorize all the 16 bytes-aligned blocks
606 // of 8 elements
607 for(; (j + 7) <= nCoeffs ; j+=8)
608 {
609 blockv = vec_ld(j << 1, block);
610 blockv_neg = vec_cmplt(blockv, vczero);
611 blockv_null = vec_cmpeq(blockv, vczero);
612 // choose between +qadd or -qadd as the third operand
613 temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
614 // multiply & add (block{i,i+7} * qmul [+-] qadd)
615 temp1 = vec_mladd(blockv, qmulv, temp1);
616 // put 0 where block[{i,i+7} used to have 0
617 blockv = vec_sel(temp1, blockv, blockv_null);
618 vec_st(blockv, j << 1, block);
619 }
620
621 // if nCoeffs isn't a multiple of 8, finish the job
622 // using good old scalar units.
623 // (we could do it using a truncated vector,
624 // but I'm not sure it's worth the hassle)
625 for(; j <= nCoeffs ; j++) {
626 level = block[j];
627 if (level) {
628 if (level < 0) {
629 level = level * qmul - qadd;
630 } else {
631 level = level * qmul + qadd;
632 }
633 block[j] = level;
634 }
635 }
636
637 if (i == 1)
638 { // cheat. this avoid special-casing the first iteration
639 block[0] = backup_0;
640 }
641 }
642 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */
643
644 POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
645 }