2 * Copyright (c) 2002 Dieter Shirley
4 * dct_unquantize_h263_altivec:
5 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg 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 GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 #include "../dsputil.h"
27 #include "../mpegvideo.h"
29 #include "gcc_fixes.h"
31 #include "dsputil_altivec.h"
33 // Swaps two variables (used for altivec registers)
36 __typeof__(a) swap_temp=a; \
41 // transposes a matrix consisting of four vectors with four elements each
42 #define TRANSPOSE4(a,b,c,d) \
44 __typeof__(a) _trans_ach = vec_mergeh(a, c); \
45 __typeof__(a) _trans_acl = vec_mergel(a, c); \
46 __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
47 __typeof__(a) _trans_bdl = vec_mergel(b, d); \
49 a = vec_mergeh(_trans_ach, _trans_bdh); \
50 b = vec_mergel(_trans_ach, _trans_bdh); \
51 c = vec_mergeh(_trans_acl, _trans_bdl); \
52 d = vec_mergel(_trans_acl, _trans_bdl); \
55 #define TRANSPOSE8(a,b,c,d,e,f,g,h) \
57 __typeof__(a) _A1, _B1, _C1, _D1, _E1, _F1, _G1, _H1; \
58 __typeof__(a) _A2, _B2, _C2, _D2, _E2, _F2, _G2, _H2; \
60 _A1 = vec_mergeh (a, e); \
61 _B1 = vec_mergel (a, e); \
62 _C1 = vec_mergeh (b, f); \
63 _D1 = vec_mergel (b, f); \
64 _E1 = vec_mergeh (c, g); \
65 _F1 = vec_mergel (c, g); \
66 _G1 = vec_mergeh (d, h); \
67 _H1 = vec_mergel (d, h); \
69 _A2 = vec_mergeh (_A1, _E1); \
70 _B2 = vec_mergel (_A1, _E1); \
71 _C2 = vec_mergeh (_B1, _F1); \
72 _D2 = vec_mergel (_B1, _F1); \
73 _E2 = vec_mergeh (_C1, _G1); \
74 _F2 = vec_mergel (_C1, _G1); \
75 _G2 = vec_mergeh (_D1, _H1); \
76 _H2 = vec_mergel (_D1, _H1); \
78 a = vec_mergeh (_A2, _E2); \
79 b = vec_mergel (_A2, _E2); \
80 c = vec_mergeh (_B2, _F2); \
81 d = vec_mergel (_B2, _F2); \
82 e = vec_mergeh (_C2, _G2); \
83 f = vec_mergel (_C2, _G2); \
84 g = vec_mergeh (_D2, _H2); \
85 h = vec_mergel (_D2, _H2); \
89 // Loads a four-byte value (int or float) from the target address
90 // into every element in the target vector. Only works if the
91 // target address is four-byte aligned (which should be always).
92 #define LOAD4(vec, address) \
94 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
95 vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
96 vec = vec_ld(0, _load_addr); \
97 vec = vec_perm(vec, vec, _perm_vec); \
98 vec = vec_splat(vec, 0); \
103 #define FOUROF(a) (a)
105 // slower, for dumb non-apple GCC
106 #define FOUROF(a) {a,a,a,a}
108 int dct_quantize_altivec(MpegEncContext* s,
109 DCTELEM* data, int n,
110 int qscale, int* overflow)
113 vector float row0, row1, row2, row3, row4, row5, row6, row7;
114 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
115 const_vector float zero = (const_vector float)FOUROF(0.);
116 // used after quantise step
117 int oldBaseValue = 0;
119 // Load the data into the row/alt vectors
121 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
123 data0 = vec_ld(0, data);
124 data1 = vec_ld(16, data);
125 data2 = vec_ld(32, data);
126 data3 = vec_ld(48, data);
127 data4 = vec_ld(64, data);
128 data5 = vec_ld(80, data);
129 data6 = vec_ld(96, data);
130 data7 = vec_ld(112, data);
132 // Transpose the data before we start
133 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
135 // load the data into floating point vectors. We load
136 // the high half of each row into the main row vectors
137 // and the low half into the alt vectors.
138 row0 = vec_ctf(vec_unpackh(data0), 0);
139 alt0 = vec_ctf(vec_unpackl(data0), 0);
140 row1 = vec_ctf(vec_unpackh(data1), 0);
141 alt1 = vec_ctf(vec_unpackl(data1), 0);
142 row2 = vec_ctf(vec_unpackh(data2), 0);
143 alt2 = vec_ctf(vec_unpackl(data2), 0);
144 row3 = vec_ctf(vec_unpackh(data3), 0);
145 alt3 = vec_ctf(vec_unpackl(data3), 0);
146 row4 = vec_ctf(vec_unpackh(data4), 0);
147 alt4 = vec_ctf(vec_unpackl(data4), 0);
148 row5 = vec_ctf(vec_unpackh(data5), 0);
149 alt5 = vec_ctf(vec_unpackl(data5), 0);
150 row6 = vec_ctf(vec_unpackh(data6), 0);
151 alt6 = vec_ctf(vec_unpackl(data6), 0);
152 row7 = vec_ctf(vec_unpackh(data7), 0);
153 alt7 = vec_ctf(vec_unpackl(data7), 0);
156 // The following block could exist as a separate an altivec dct
157 // function. However, if we put it inline, the DCT data can remain
158 // in the vector local variables, as floats, which we'll use during the
161 const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
162 const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
163 const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
164 const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
165 const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
166 const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
167 const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
168 const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
169 const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
170 const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
171 const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
172 const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
175 int whichPass, whichHalf;
177 for(whichPass = 1; whichPass<=2; whichPass++)
179 for(whichHalf = 1; whichHalf<=2; whichHalf++)
181 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
182 vector float tmp10, tmp11, tmp12, tmp13;
183 vector float z1, z2, z3, z4, z5;
185 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
186 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
187 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
188 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
189 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
190 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
191 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
192 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
194 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
195 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
196 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
197 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
200 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
201 row0 = vec_add(tmp10, tmp11);
203 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
204 row4 = vec_sub(tmp10, tmp11);
207 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
208 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
210 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
211 // CONST_BITS-PASS1_BITS);
212 row2 = vec_madd(tmp13, vec_0_765366865, z1);
214 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
215 // CONST_BITS-PASS1_BITS);
216 row6 = vec_madd(tmp12, vec_1_847759065, z1);
218 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
219 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
220 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
221 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
223 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
224 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
226 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
227 z3 = vec_madd(z3, vec_1_961570560, z5);
229 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
230 z4 = vec_madd(z4, vec_0_390180644, z5);
232 // The following adds are rolled into the multiplies above
233 // z3 = vec_add(z3, z5); // z3 += z5;
234 // z4 = vec_add(z4, z5); // z4 += z5;
236 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
237 // Wow! It's actually more effecient to roll this multiply
238 // into the adds below, even thought the multiply gets done twice!
239 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
241 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
242 // Same with this one...
243 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
245 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
246 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
247 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
249 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
250 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
251 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
253 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
254 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
255 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
257 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
258 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
259 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
261 // Swap the row values with the alts. If this is the first half,
262 // this sets up the low values to be acted on in the second half.
263 // If this is the second half, it puts the high values back in
264 // the row values where they are expected to be when we're done.
277 // transpose the data for the second pass
279 // First, block transpose the upper right with lower left.
285 // Now, transpose each block of four
286 TRANSPOSE4(row0, row1, row2, row3);
287 TRANSPOSE4(row4, row5, row6, row7);
288 TRANSPOSE4(alt0, alt1, alt2, alt3);
289 TRANSPOSE4(alt4, alt5, alt6, alt7);
294 // perform the quantise step, using the floating point data
295 // still in the row/alt registers
298 const vector signed int* qmat;
299 vector float bias, negBias;
303 vector signed int baseVector;
305 // We must cache element 0 in the intra case
306 // (it needs special handling).
307 baseVector = vec_cts(vec_splat(row0, 0), 0);
308 vec_ste(baseVector, 0, &oldBaseValue);
310 qmat = (vector signed int*)s->q_intra_matrix[qscale];
311 biasAddr = &(s->intra_quant_bias);
315 qmat = (vector signed int*)s->q_inter_matrix[qscale];
316 biasAddr = &(s->inter_quant_bias);
319 // Load the bias vector (We add 0.5 to the bias so that we're
320 // rounding when we convert to int, instead of flooring.)
322 vector signed int biasInt;
323 const vector float negOneFloat = (vector float)FOUROF(-1.0f);
324 LOAD4(biasInt, biasAddr);
325 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
326 negBias = vec_madd(bias, negOneFloat, zero);
330 vector float q0, q1, q2, q3, q4, q5, q6, q7;
332 q0 = vec_ctf(qmat[0], QMAT_SHIFT);
333 q1 = vec_ctf(qmat[2], QMAT_SHIFT);
334 q2 = vec_ctf(qmat[4], QMAT_SHIFT);
335 q3 = vec_ctf(qmat[6], QMAT_SHIFT);
336 q4 = vec_ctf(qmat[8], QMAT_SHIFT);
337 q5 = vec_ctf(qmat[10], QMAT_SHIFT);
338 q6 = vec_ctf(qmat[12], QMAT_SHIFT);
339 q7 = vec_ctf(qmat[14], QMAT_SHIFT);
341 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
342 vec_cmpgt(row0, zero));
343 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
344 vec_cmpgt(row1, zero));
345 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
346 vec_cmpgt(row2, zero));
347 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
348 vec_cmpgt(row3, zero));
349 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
350 vec_cmpgt(row4, zero));
351 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
352 vec_cmpgt(row5, zero));
353 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
354 vec_cmpgt(row6, zero));
355 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
356 vec_cmpgt(row7, zero));
358 q0 = vec_ctf(qmat[1], QMAT_SHIFT);
359 q1 = vec_ctf(qmat[3], QMAT_SHIFT);
360 q2 = vec_ctf(qmat[5], QMAT_SHIFT);
361 q3 = vec_ctf(qmat[7], QMAT_SHIFT);
362 q4 = vec_ctf(qmat[9], QMAT_SHIFT);
363 q5 = vec_ctf(qmat[11], QMAT_SHIFT);
364 q6 = vec_ctf(qmat[13], QMAT_SHIFT);
365 q7 = vec_ctf(qmat[15], QMAT_SHIFT);
367 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
368 vec_cmpgt(alt0, zero));
369 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
370 vec_cmpgt(alt1, zero));
371 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
372 vec_cmpgt(alt2, zero));
373 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
374 vec_cmpgt(alt3, zero));
375 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
376 vec_cmpgt(alt4, zero));
377 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
378 vec_cmpgt(alt5, zero));
379 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
380 vec_cmpgt(alt6, zero));
381 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
382 vec_cmpgt(alt7, zero));
388 // Store the data back into the original block
390 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
392 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
393 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
394 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
395 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
396 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
397 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
398 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
399 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
402 // Clamp for overflow
403 vector signed int max_q_int, min_q_int;
404 vector signed short max_q, min_q;
406 LOAD4(max_q_int, &(s->max_qcoeff));
407 LOAD4(min_q_int, &(s->min_qcoeff));
409 max_q = vec_pack(max_q_int, max_q_int);
410 min_q = vec_pack(min_q_int, min_q_int);
412 data0 = vec_max(vec_min(data0, max_q), min_q);
413 data1 = vec_max(vec_min(data1, max_q), min_q);
414 data2 = vec_max(vec_min(data2, max_q), min_q);
415 data4 = vec_max(vec_min(data4, max_q), min_q);
416 data5 = vec_max(vec_min(data5, max_q), min_q);
417 data6 = vec_max(vec_min(data6, max_q), min_q);
418 data7 = vec_max(vec_min(data7, max_q), min_q);
422 vector bool char zero_01, zero_23, zero_45, zero_67;
423 vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
424 vector signed char negOne = vec_splat_s8(-1);
425 vector signed char* scanPtr =
426 (vector signed char*)(s->intra_scantable.inverse);
427 signed char lastNonZeroChar;
429 // Determine the largest non-zero index.
430 zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
431 vec_cmpeq(data1, (vector signed short)zero));
432 zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
433 vec_cmpeq(data3, (vector signed short)zero));
434 zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
435 vec_cmpeq(data5, (vector signed short)zero));
436 zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
437 vec_cmpeq(data7, (vector signed short)zero));
440 scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
441 scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
442 scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
443 scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
446 scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
447 scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
450 scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
453 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
454 vec_mergel(scanIndices_01, negOne));
457 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
458 vec_mergel(scanIndices_01, negOne));
461 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
462 vec_mergel(scanIndices_01, negOne));
465 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
466 vec_mergel(scanIndices_01, negOne));
468 scanIndices_01 = vec_splat(scanIndices_01, 0);
471 vec_ste(scanIndices_01, 0, &lastNonZeroChar);
473 lastNonZero = lastNonZeroChar;
475 // While the data is still in vectors we check for the transpose IDCT permute
476 // and handle it using the vector unit if we can. This is the permute used
477 // by the altivec idct, so it is common when using the altivec dct.
479 if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
481 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
484 vec_st(data0, 0, data);
485 vec_st(data1, 16, data);
486 vec_st(data2, 32, data);
487 vec_st(data3, 48, data);
488 vec_st(data4, 64, data);
489 vec_st(data5, 80, data);
490 vec_st(data6, 96, data);
491 vec_st(data7, 112, data);
495 // special handling of block[0]
501 oldBaseValue /= s->y_dc_scale;
503 oldBaseValue /= s->c_dc_scale;
506 // Divide by 8, rounding the result
507 data[0] = (oldBaseValue + 4) >> 3;
510 // We handled the tranpose permutation above and we don't
511 // need to permute the "no" permutation case.
512 if ((lastNonZero > 0) &&
513 (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
514 (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM))
516 ff_block_permute(data, s->dsp.idct_permutation,
517 s->intra_scantable.scantable, lastNonZero);
525 AltiVec version of dct_unquantize_h263
526 this code assumes `block' is 16 bytes-aligned
528 void dct_unquantize_h263_altivec(MpegEncContext *s,
529 DCTELEM *block, int n, int qscale)
531 POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
532 int i, level, qmul, qadd;
535 assert(s->block_last_index[n]>=0);
537 POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
539 qadd = (qscale - 1) | 1;
545 block[0] = block[0] * s->y_dc_scale;
547 block[0] = block[0] * s->c_dc_scale;
551 nCoeffs= 63; //does not allways use zigzag table
554 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
558 register const_vector signed short vczero = (const_vector signed short)vec_splat_s16(0);
559 short __attribute__ ((aligned(16))) qmul8[] =
561 qmul, qmul, qmul, qmul,
562 qmul, qmul, qmul, qmul
564 short __attribute__ ((aligned(16))) qadd8[] =
566 qadd, qadd, qadd, qadd,
567 qadd, qadd, qadd, qadd
569 short __attribute__ ((aligned(16))) nqadd8[] =
571 -qadd, -qadd, -qadd, -qadd,
572 -qadd, -qadd, -qadd, -qadd
574 register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
575 register vector bool short blockv_null, blockv_neg;
576 register short backup_0 = block[0];
579 qmulv = vec_ld(0, qmul8);
580 qaddv = vec_ld(0, qadd8);
581 nqaddv = vec_ld(0, nqadd8);
583 #if 0 // block *is* 16 bytes-aligned, it seems.
584 // first make sure block[j] is 16 bytes-aligned
585 for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
589 level = level * qmul - qadd;
591 level = level * qmul + qadd;
598 // vectorize all the 16 bytes-aligned blocks
600 for(; (j + 7) <= nCoeffs ; j+=8)
602 blockv = vec_ld(j << 1, block);
603 blockv_neg = vec_cmplt(blockv, vczero);
604 blockv_null = vec_cmpeq(blockv, vczero);
605 // choose between +qadd or -qadd as the third operand
606 temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
607 // multiply & add (block{i,i+7} * qmul [+-] qadd)
608 temp1 = vec_mladd(blockv, qmulv, temp1);
609 // put 0 where block[{i,i+7} used to have 0
610 blockv = vec_sel(temp1, blockv, blockv_null);
611 vec_st(blockv, j << 1, block);
614 // if nCoeffs isn't a multiple of 8, finish the job
615 // using good old scalar units.
616 // (we could do it using a truncated vector,
617 // but I'm not sure it's worth the hassle)
618 for(; j <= nCoeffs ; j++) {
622 level = level * qmul - qadd;
624 level = level * qmul + qadd;
631 { // cheat. this avoid special-casing the first iteration
635 POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);