Statistics
| Branch: | Revision:

ffmpeg / libavcodec / ppc / mpegvideo_altivec.c @ 5509bffa

History | View | Annotate | Download (24.7 KB)

1
/*
2
 * Copyright (c) 2002 Dieter Shirley
3
 *
4
 * dct_unquantize_h263_altivec:
5
 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
6
 *
7
 * This library is free software; you can redistribute it and/or
8
 * modify it under the terms of the GNU Lesser General Public
9
 * License as published by the Free Software Foundation; either
10
 * version 2 of the License, or (at your option) any later version.
11
 *
12
 * This library is distributed in the hope that it will be useful,
13
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15
 * Lesser General Public License for more details.
16
 *
17
 * You should have received a copy of the GNU Lesser General Public
18
 * License along with this library; if not, write to the Free Software
19
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20
 */
21

    
22
#include <stdlib.h>
23
#include <stdio.h>
24
#include "../dsputil.h"
25
#include "../mpegvideo.h"
26

    
27
#include "gcc_fixes.h"
28

    
29
#include "dsputil_altivec.h"
30

    
31
// Swaps two variables (used for altivec registers)
32
#define SWAP(a,b) \
33
do { \
34
    __typeof__(a) swap_temp=a; \
35
    a=b; \
36
    b=swap_temp; \
37
} while (0)
38

    
39
// transposes a matrix consisting of four vectors with four elements each
40
#define TRANSPOSE4(a,b,c,d) \
41
do { \
42
  __typeof__(a) _trans_ach = vec_mergeh(a, c); \
43
  __typeof__(a) _trans_acl = vec_mergel(a, c); \
44
  __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
45
  __typeof__(a) _trans_bdl = vec_mergel(b, d); \
46
 \
47
  a = vec_mergeh(_trans_ach, _trans_bdh); \
48
  b = vec_mergel(_trans_ach, _trans_bdh); \
49
  c = vec_mergeh(_trans_acl, _trans_bdl); \
50
  d = vec_mergel(_trans_acl, _trans_bdl); \
51
} while (0)
52

    
53
#define TRANSPOSE8(a,b,c,d,e,f,g,h) \
54
do { \
55
    __typeof__(a)  _A1, _B1, _C1, _D1, _E1, _F1, _G1, _H1; \
56
    __typeof__(a)  _A2, _B2, _C2, _D2, _E2, _F2, _G2, _H2; \
57
 \
58
    _A1 = vec_mergeh (a, e); \
59
    _B1 = vec_mergel (a, e); \
60
    _C1 = vec_mergeh (b, f); \
61
    _D1 = vec_mergel (b, f); \
62
    _E1 = vec_mergeh (c, g); \
63
    _F1 = vec_mergel (c, g); \
64
    _G1 = vec_mergeh (d, h); \
65
    _H1 = vec_mergel (d, h); \
66
 \
67
    _A2 = vec_mergeh (_A1, _E1); \
68
    _B2 = vec_mergel (_A1, _E1); \
69
    _C2 = vec_mergeh (_B1, _F1); \
70
    _D2 = vec_mergel (_B1, _F1); \
71
    _E2 = vec_mergeh (_C1, _G1); \
72
    _F2 = vec_mergel (_C1, _G1); \
73
    _G2 = vec_mergeh (_D1, _H1); \
74
    _H2 = vec_mergel (_D1, _H1); \
75
 \
76
    a = vec_mergeh (_A2, _E2); \
77
    b = vec_mergel (_A2, _E2); \
78
    c = vec_mergeh (_B2, _F2); \
79
    d = vec_mergel (_B2, _F2); \
80
    e = vec_mergeh (_C2, _G2); \
81
    f = vec_mergel (_C2, _G2); \
82
    g = vec_mergeh (_D2, _H2); \
83
    h = vec_mergel (_D2, _H2); \
84
} while (0)
85

    
86

    
87
// Loads a four-byte value (int or float) from the target address
88
// into every element in the target vector.  Only works if the
89
// target address is four-byte aligned (which should be always).
90
#define LOAD4(vec, address) \
91
{ \
92
    __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
93
    vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
94
    vec = vec_ld(0, _load_addr); \
95
    vec = vec_perm(vec, vec, _perm_vec); \
96
    vec = vec_splat(vec, 0); \
97
}
98

    
99

    
100
#ifdef CONFIG_DARWIN
101
#define FOUROF(a) (a)
102
#else
103
// slower, for dumb non-apple GCC
104
#define FOUROF(a) {a,a,a,a}
105
#endif
106
int dct_quantize_altivec(MpegEncContext* s,
107
                        DCTELEM* data, int n,
108
                        int qscale, int* overflow)
109
{
110
    int lastNonZero;
111
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
112
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
113
    const_vector float zero = (const_vector float)FOUROF(0.);
114
    // used after quantise step
115
    int oldBaseValue = 0;
116

    
117
    // Load the data into the row/alt vectors
118
    {
119
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
120

    
121
        data0 = vec_ld(0, data);
122
        data1 = vec_ld(16, data);
123
        data2 = vec_ld(32, data);
124
        data3 = vec_ld(48, data);
125
        data4 = vec_ld(64, data);
126
        data5 = vec_ld(80, data);
127
        data6 = vec_ld(96, data);
128
        data7 = vec_ld(112, data);
129

    
130
        // Transpose the data before we start
131
        TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
132

    
133
        // load the data into floating point vectors.  We load
134
        // the high half of each row into the main row vectors
135
        // and the low half into the alt vectors.
136
        row0 = vec_ctf(vec_unpackh(data0), 0);
137
        alt0 = vec_ctf(vec_unpackl(data0), 0);
138
        row1 = vec_ctf(vec_unpackh(data1), 0);
139
        alt1 = vec_ctf(vec_unpackl(data1), 0);
140
        row2 = vec_ctf(vec_unpackh(data2), 0);
141
        alt2 = vec_ctf(vec_unpackl(data2), 0);
142
        row3 = vec_ctf(vec_unpackh(data3), 0);
143
        alt3 = vec_ctf(vec_unpackl(data3), 0);
144
        row4 = vec_ctf(vec_unpackh(data4), 0);
145
        alt4 = vec_ctf(vec_unpackl(data4), 0);
146
        row5 = vec_ctf(vec_unpackh(data5), 0);
147
        alt5 = vec_ctf(vec_unpackl(data5), 0);
148
        row6 = vec_ctf(vec_unpackh(data6), 0);
149
        alt6 = vec_ctf(vec_unpackl(data6), 0);
150
        row7 = vec_ctf(vec_unpackh(data7), 0);
151
        alt7 = vec_ctf(vec_unpackl(data7), 0);
152
    }
153

    
154
    // The following block could exist as a separate an altivec dct
155
                // function.  However, if we put it inline, the DCT data can remain
156
                // in the vector local variables, as floats, which we'll use during the
157
                // quantize step...
158
    {
159
        const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
160
        const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
161
        const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
162
        const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
163
        const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
164
        const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
165
        const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
166
        const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
167
        const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
168
        const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
169
        const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
170
        const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
171

    
172

    
173
        int whichPass, whichHalf;
174

    
175
        for(whichPass = 1; whichPass<=2; whichPass++)
176
        {
177
            for(whichHalf = 1; whichHalf<=2; whichHalf++)
178
            {
179
                vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
180
                vector float tmp10, tmp11, tmp12, tmp13;
181
                vector float z1, z2, z3, z4, z5;
182

    
183
                tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
184
                tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
185
                tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
186
                tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
187
                tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
188
                tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
189
                tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
190
                tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
191

    
192
                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
193
                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
194
                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
195
                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
196

    
197

    
198
                // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
199
                row0 = vec_add(tmp10, tmp11);
200

    
201
                // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
202
                row4 = vec_sub(tmp10, tmp11);
203

    
204

    
205
                // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
206
                z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
207

    
208
                // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
209
                //                                CONST_BITS-PASS1_BITS);
210
                row2 = vec_madd(tmp13, vec_0_765366865, z1);
211

    
212
                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
213
                //                                CONST_BITS-PASS1_BITS);
214
                row6 = vec_madd(tmp12, vec_1_847759065, z1);
215

    
216
                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
217
                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
218
                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
219
                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
220

    
221
                // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
222
                z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
223

    
224
                // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
225
                z3 = vec_madd(z3, vec_1_961570560, z5);
226

    
227
                // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
228
                z4 = vec_madd(z4, vec_0_390180644, z5);
229

    
230
                // The following adds are rolled into the multiplies above
231
                // z3 = vec_add(z3, z5);  // z3 += z5;
232
                // z4 = vec_add(z4, z5);  // z4 += z5;
233

    
234
                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
235
                // Wow!  It's actually more effecient to roll this multiply
236
                // into the adds below, even thought the multiply gets done twice!
237
                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
238

    
239
                // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
240
                // Same with this one...
241
                // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
242

    
243
                // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
244
                // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
245
                row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
246

    
247
                // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
248
                // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
249
                row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
250

    
251
                // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
252
                // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
253
                row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
254

    
255
                // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
256
                // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
257
                row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
258

    
259
                // Swap the row values with the alts.  If this is the first half,
260
                // this sets up the low values to be acted on in the second half.
261
                // If this is the second half, it puts the high values back in
262
                // the row values where they are expected to be when we're done.
263
                SWAP(row0, alt0);
264
                SWAP(row1, alt1);
265
                SWAP(row2, alt2);
266
                SWAP(row3, alt3);
267
                SWAP(row4, alt4);
268
                SWAP(row5, alt5);
269
                SWAP(row6, alt6);
270
                SWAP(row7, alt7);
271
            }
272

    
273
            if (whichPass == 1)
274
            {
275
                // transpose the data for the second pass
276

    
277
                // First, block transpose the upper right with lower left.
278
                SWAP(row4, alt0);
279
                SWAP(row5, alt1);
280
                SWAP(row6, alt2);
281
                SWAP(row7, alt3);
282

    
283
                // Now, transpose each block of four
284
                TRANSPOSE4(row0, row1, row2, row3);
285
                TRANSPOSE4(row4, row5, row6, row7);
286
                TRANSPOSE4(alt0, alt1, alt2, alt3);
287
                TRANSPOSE4(alt4, alt5, alt6, alt7);
288
            }
289
        }
290
    }
291

    
292
    // perform the quantise step, using the floating point data
293
    // still in the row/alt registers
294
    {
295
        const int* biasAddr;
296
        const vector signed int* qmat;
297
        vector float bias, negBias;
298

    
299
        if (s->mb_intra)
300
        {
301
            vector signed int baseVector;
302

    
303
            // We must cache element 0 in the intra case
304
            // (it needs special handling).
305
            baseVector = vec_cts(vec_splat(row0, 0), 0);
306
            vec_ste(baseVector, 0, &oldBaseValue);
307

    
308
            qmat = (vector signed int*)s->q_intra_matrix[qscale];
309
            biasAddr = &(s->intra_quant_bias);
310
        }
311
        else
312
        {
313
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
314
            biasAddr = &(s->inter_quant_bias);
315
        }
316

    
317
        // Load the bias vector (We add 0.5 to the bias so that we're
318
                                // rounding when we convert to int, instead of flooring.)
319
        {
320
            vector signed int biasInt;
321
            const vector float negOneFloat = (vector float)FOUROF(-1.0f);
322
            LOAD4(biasInt, biasAddr);
323
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
324
            negBias = vec_madd(bias, negOneFloat, zero);
325
        }
326

    
327
        {
328
            vector float q0, q1, q2, q3, q4, q5, q6, q7;
329

    
330
            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
331
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
332
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
333
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
334
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
335
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
336
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
337
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);
338

    
339
            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
340
                    vec_cmpgt(row0, zero));
341
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
342
                    vec_cmpgt(row1, zero));
343
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
344
                    vec_cmpgt(row2, zero));
345
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
346
                    vec_cmpgt(row3, zero));
347
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
348
                    vec_cmpgt(row4, zero));
349
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
350
                    vec_cmpgt(row5, zero));
351
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
352
                    vec_cmpgt(row6, zero));
353
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
354
                    vec_cmpgt(row7, zero));
355

    
356
            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
357
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
358
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
359
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
360
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
361
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
362
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
363
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);
364

    
365
            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
366
                    vec_cmpgt(alt0, zero));
367
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
368
                    vec_cmpgt(alt1, zero));
369
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
370
                    vec_cmpgt(alt2, zero));
371
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
372
                    vec_cmpgt(alt3, zero));
373
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
374
                    vec_cmpgt(alt4, zero));
375
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
376
                    vec_cmpgt(alt5, zero));
377
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
378
                    vec_cmpgt(alt6, zero));
379
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
380
                    vec_cmpgt(alt7, zero));
381
        }
382

    
383

    
384
    }
385

    
386
    // Store the data back into the original block
387
    {
388
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
389

    
390
        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
391
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
392
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
393
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
394
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
395
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
396
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
397
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
398

    
399
        {
400
            // Clamp for overflow
401
            vector signed int max_q_int, min_q_int;
402
            vector signed short max_q, min_q;
403

    
404
            LOAD4(max_q_int, &(s->max_qcoeff));
405
            LOAD4(min_q_int, &(s->min_qcoeff));
406

    
407
            max_q = vec_pack(max_q_int, max_q_int);
408
            min_q = vec_pack(min_q_int, min_q_int);
409

    
410
            data0 = vec_max(vec_min(data0, max_q), min_q);
411
            data1 = vec_max(vec_min(data1, max_q), min_q);
412
            data2 = vec_max(vec_min(data2, max_q), min_q);
413
            data4 = vec_max(vec_min(data4, max_q), min_q);
414
            data5 = vec_max(vec_min(data5, max_q), min_q);
415
            data6 = vec_max(vec_min(data6, max_q), min_q);
416
            data7 = vec_max(vec_min(data7, max_q), min_q);
417
        }
418

    
419
        {
420
        vector bool char zero_01, zero_23, zero_45, zero_67;
421
        vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
422
        vector signed char negOne = vec_splat_s8(-1);
423
        vector signed char* scanPtr =
424
                (vector signed char*)(s->intra_scantable.inverse);
425
        signed char lastNonZeroChar;
426

    
427
        // Determine the largest non-zero index.
428
        zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
429
                vec_cmpeq(data1, (vector signed short)zero));
430
        zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
431
                vec_cmpeq(data3, (vector signed short)zero));
432
        zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
433
                vec_cmpeq(data5, (vector signed short)zero));
434
        zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
435
                vec_cmpeq(data7, (vector signed short)zero));
436

    
437
        // 64 biggest values
438
        scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
439
        scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
440
        scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
441
        scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
442

    
443
        // 32 largest values
444
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
445
        scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
446

    
447
        // 16 largest values
448
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
449

    
450
        // 8 largest values
451
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
452
                vec_mergel(scanIndices_01, negOne));
453

    
454
        // 4 largest values
455
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
456
                vec_mergel(scanIndices_01, negOne));
457

    
458
        // 2 largest values
459
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
460
                vec_mergel(scanIndices_01, negOne));
461

    
462
        // largest value
463
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
464
                vec_mergel(scanIndices_01, negOne));
465

    
466
        scanIndices_01 = vec_splat(scanIndices_01, 0);
467

    
468

    
469
        vec_ste(scanIndices_01, 0, &lastNonZeroChar);
470

    
471
        lastNonZero = lastNonZeroChar;
472

    
473
        // While the data is still in vectors we check for the transpose IDCT permute
474
        // and handle it using the vector unit if we can.  This is the permute used
475
        // by the altivec idct, so it is common when using the altivec dct.
476

    
477
        if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
478
        {
479
            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
480
        }
481

    
482
        vec_st(data0, 0, data);
483
        vec_st(data1, 16, data);
484
        vec_st(data2, 32, data);
485
        vec_st(data3, 48, data);
486
        vec_st(data4, 64, data);
487
        vec_st(data5, 80, data);
488
        vec_st(data6, 96, data);
489
        vec_st(data7, 112, data);
490
        }
491
    }
492

    
493
    // special handling of block[0]
494
    if (s->mb_intra)
495
    {
496
        if (!s->h263_aic)
497
        {
498
            if (n < 4)
499
                oldBaseValue /= s->y_dc_scale;
500
            else
501
                oldBaseValue /= s->c_dc_scale;
502
        }
503

    
504
        // Divide by 8, rounding the result
505
        data[0] = (oldBaseValue + 4) >> 3;
506
    }
507

    
508
    // We handled the tranpose permutation above and we don't
509
    // need to permute the "no" permutation case.
510
    if ((lastNonZero > 0) &&
511
        (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
512
        (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM))
513
    {
514
        ff_block_permute(data, s->dsp.idct_permutation,
515
                s->intra_scantable.scantable, lastNonZero);
516
    }
517

    
518
    return lastNonZero;
519
}
520
#undef FOUROF
521

    
522
/*
523
  AltiVec version of dct_unquantize_h263
524
  this code assumes `block' is 16 bytes-aligned
525
*/
526
void dct_unquantize_h263_altivec(MpegEncContext *s,
527
                                 DCTELEM *block, int n, int qscale)
528
{
529
POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
530
    int i, level, qmul, qadd;
531
    int nCoeffs;
532

    
533
    assert(s->block_last_index[n]>=0);
534

    
535
POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
536

    
537
    qadd = (qscale - 1) | 1;
538
    qmul = qscale << 1;
539

    
540
    if (s->mb_intra) {
541
        if (!s->h263_aic) {
542
            if (n < 4)
543
                block[0] = block[0] * s->y_dc_scale;
544
            else
545
                block[0] = block[0] * s->c_dc_scale;
546
        }else
547
            qadd = 0;
548
        i = 1;
549
        nCoeffs= 63; //does not allways use zigzag table
550
    } else {
551
        i = 0;
552
        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
553
    }
554

    
555
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
556
    for(;i<=nCoeffs;i++) {
557
        level = block[i];
558
        if (level) {
559
            if (level < 0) {
560
                level = level * qmul - qadd;
561
            } else {
562
                level = level * qmul + qadd;
563
            }
564
            block[i] = level;
565
        }
566
    }
567
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
568
    {
569
      register const_vector signed short vczero = (const_vector signed short)vec_splat_s16(0);
570
      short __attribute__ ((aligned(16))) qmul8[] =
571
          {
572
            qmul, qmul, qmul, qmul,
573
            qmul, qmul, qmul, qmul
574
          };
575
      short __attribute__ ((aligned(16))) qadd8[] =
576
          {
577
            qadd, qadd, qadd, qadd,
578
            qadd, qadd, qadd, qadd
579
          };
580
      short __attribute__ ((aligned(16))) nqadd8[] =
581
          {
582
            -qadd, -qadd, -qadd, -qadd,
583
            -qadd, -qadd, -qadd, -qadd
584
          };
585
      register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
586
      register vector bool short blockv_null, blockv_neg;
587
      register short backup_0 = block[0];
588
      register int j = 0;
589

    
590
      qmulv = vec_ld(0, qmul8);
591
      qaddv = vec_ld(0, qadd8);
592
      nqaddv = vec_ld(0, nqadd8);
593

    
594
#if 0 // block *is* 16 bytes-aligned, it seems.
595
      // first make sure block[j] is 16 bytes-aligned
596
      for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
597
        level = block[j];
598
        if (level) {
599
          if (level < 0) {
600
                level = level * qmul - qadd;
601
            } else {
602
                level = level * qmul + qadd;
603
            }
604
            block[j] = level;
605
        }
606
      }
607
#endif
608

    
609
      // vectorize all the 16 bytes-aligned blocks
610
      // of 8 elements
611
      for(; (j + 7) <= nCoeffs ; j+=8)
612
      {
613
        blockv = vec_ld(j << 1, block);
614
        blockv_neg = vec_cmplt(blockv, vczero);
615
        blockv_null = vec_cmpeq(blockv, vczero);
616
        // choose between +qadd or -qadd as the third operand
617
        temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
618
        // multiply & add (block{i,i+7} * qmul [+-] qadd)
619
        temp1 = vec_mladd(blockv, qmulv, temp1);
620
        // put 0 where block[{i,i+7} used to have 0
621
        blockv = vec_sel(temp1, blockv, blockv_null);
622
        vec_st(blockv, j << 1, block);
623
      }
624

    
625
      // if nCoeffs isn't a multiple of 8, finish the job
626
      // using good old scalar units.
627
      // (we could do it using a truncated vector,
628
      // but I'm not sure it's worth the hassle)
629
      for(; j <= nCoeffs ; j++) {
630
        level = block[j];
631
        if (level) {
632
          if (level < 0) {
633
                level = level * qmul - qadd;
634
            } else {
635
                level = level * qmul + qadd;
636
            }
637
            block[j] = level;
638
        }
639
      }
640

    
641
      if (i == 1)
642
      { // cheat. this avoid special-casing the first iteration
643
        block[0] = backup_0;
644
      }
645
    }
646
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
647

    
648
POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
649
}