Statistics
| Branch: | Revision:

ffmpeg / libavcodec / ppc / mpegvideo_altivec.c @ aab34ca0

History | View | Annotate | Download (24.6 KB)

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
    // used after quantise step
112
    int oldBaseValue = 0;
113

    
114
    // Load the data into the row/alt vectors
115
    {
116
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
117

    
118
        data0 = vec_ld(0, data);
119
        data1 = vec_ld(16, data);
120
        data2 = vec_ld(32, data);
121
        data3 = vec_ld(48, data);
122
        data4 = vec_ld(64, data);
123
        data5 = vec_ld(80, data);
124
        data6 = vec_ld(96, data);
125
        data7 = vec_ld(112, data);
126

    
127
        // Transpose the data before we start
128
        TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
129

    
130
        // load the data into floating point vectors.  We load
131
        // the high half of each row into the main row vectors
132
        // and the low half into the alt vectors.
133
        row0 = vec_ctf(vec_unpackh(data0), 0);
134
        alt0 = vec_ctf(vec_unpackl(data0), 0);
135
        row1 = vec_ctf(vec_unpackh(data1), 0);
136
        alt1 = vec_ctf(vec_unpackl(data1), 0);
137
        row2 = vec_ctf(vec_unpackh(data2), 0);
138
        alt2 = vec_ctf(vec_unpackl(data2), 0);
139
        row3 = vec_ctf(vec_unpackh(data3), 0);
140
        alt3 = vec_ctf(vec_unpackl(data3), 0);
141
        row4 = vec_ctf(vec_unpackh(data4), 0);
142
        alt4 = vec_ctf(vec_unpackl(data4), 0);
143
        row5 = vec_ctf(vec_unpackh(data5), 0);
144
        alt5 = vec_ctf(vec_unpackl(data5), 0);
145
        row6 = vec_ctf(vec_unpackh(data6), 0);
146
        alt6 = vec_ctf(vec_unpackl(data6), 0);
147
        row7 = vec_ctf(vec_unpackh(data7), 0);
148
        alt7 = vec_ctf(vec_unpackl(data7), 0);
149
    }
150

    
151
    // The following block could exist as a separate an altivec dct
152
                // function.  However, if we put it inline, the DCT data can remain
153
                // in the vector local variables, as floats, which we'll use during the
154
                // quantize step...
155
    {
156
        const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
157
        const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
158
        const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
159
        const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
160
        const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
161
        const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
162
        const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
163
        const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
164
        const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
165
        const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
166
        const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
167
        const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
168

    
169

    
170
        int whichPass, whichHalf;
171

    
172
        for(whichPass = 1; whichPass<=2; whichPass++)
173
        {
174
            for(whichHalf = 1; whichHalf<=2; whichHalf++)
175
            {
176
                vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
177
                vector float tmp10, tmp11, tmp12, tmp13;
178
                vector float z1, z2, z3, z4, z5;
179

    
180
                tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
181
                tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
182
                tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
183
                tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
184
                tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
185
                tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
186
                tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
187
                tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
188

    
189
                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
190
                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
191
                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
192
                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
193

    
194

    
195
                // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
196
                row0 = vec_add(tmp10, tmp11);
197

    
198
                // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
199
                row4 = vec_sub(tmp10, tmp11);
200

    
201

    
202
                // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
203
                z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
204

    
205
                // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
206
                //                   CONST_BITS-PASS1_BITS);
207
                row2 = vec_madd(tmp13, vec_0_765366865, z1);
208

    
209
                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
210
                //                   CONST_BITS-PASS1_BITS);
211
                row6 = vec_madd(tmp12, vec_1_847759065, z1);
212

    
213
                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
214
                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
215
                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
216
                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
217

    
218
                // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
219
                z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
220

    
221
                // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
222
                z3 = vec_madd(z3, vec_1_961570560, z5);
223

    
224
                // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
225
                z4 = vec_madd(z4, vec_0_390180644, z5);
226

    
227
                // The following adds are rolled into the multiplies above
228
                // z3 = vec_add(z3, z5);  // z3 += z5;
229
                // z4 = vec_add(z4, z5);  // z4 += z5;
230

    
231
                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
232
                // Wow!  It's actually more effecient to roll this multiply
233
                // into the adds below, even thought the multiply gets done twice!
234
                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
235

    
236
                // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
237
                // Same with this one...
238
                // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
239

    
240
                // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
241
                // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
242
                row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
243

    
244
                // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
245
                // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
246
                row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
247

    
248
                // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
249
                // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
250
                row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
251

    
252
                // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
253
                // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
254
                row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
255

    
256
                // Swap the row values with the alts.  If this is the first half,
257
                // this sets up the low values to be acted on in the second half.
258
                // If this is the second half, it puts the high values back in
259
                // the row values where they are expected to be when we're done.
260
                SWAP(row0, alt0);
261
                SWAP(row1, alt1);
262
                SWAP(row2, alt2);
263
                SWAP(row3, alt3);
264
                SWAP(row4, alt4);
265
                SWAP(row5, alt5);
266
                SWAP(row6, alt6);
267
                SWAP(row7, alt7);
268
            }
269

    
270
            if (whichPass == 1)
271
            {
272
                // transpose the data for the second pass
273
                 
274
                // First, block transpose the upper right with lower left.
275
                SWAP(row4, alt0);
276
                SWAP(row5, alt1);
277
                SWAP(row6, alt2);
278
                SWAP(row7, alt3);
279

    
280
                // Now, transpose each block of four
281
                TRANSPOSE4(row0, row1, row2, row3);
282
                TRANSPOSE4(row4, row5, row6, row7);
283
                TRANSPOSE4(alt0, alt1, alt2, alt3);
284
                TRANSPOSE4(alt4, alt5, alt6, alt7);
285
            }
286
        }
287
    }
288

    
289
    // perform the quantise step, using the floating point data
290
    // still in the row/alt registers
291
    {
292
        const int* biasAddr;
293
        const vector signed int* qmat;
294
        vector float bias, negBias;
295

    
296
        if (s->mb_intra)
297
        {
298
            vector signed int baseVector;
299

    
300
            // We must cache element 0 in the intra case
301
            // (it needs special handling).
302
            baseVector = vec_cts(vec_splat(row0, 0), 0);
303
            vec_ste(baseVector, 0, &oldBaseValue);
304

    
305
            qmat = (vector signed int*)s->q_intra_matrix[qscale];
306
            biasAddr = &(s->intra_quant_bias);
307
        }
308
        else
309
        {
310
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
311
            biasAddr = &(s->inter_quant_bias);
312
        }
313

    
314
        // Load the bias vector (We add 0.5 to the bias so that we're
315
                                // rounding when we convert to int, instead of flooring.)
316
        {
317
            vector signed int biasInt;
318
            const vector float negOneFloat = (vector float)FOUROF(-1.0f);
319
            LOAD4(biasInt, biasAddr);
320
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
321
            negBias = vec_madd(bias, negOneFloat, zero);
322
        }
323

    
324
        {
325
            vector float q0, q1, q2, q3, q4, q5, q6, q7;
326

    
327
            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
328
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
329
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
330
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
331
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
332
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
333
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
334
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);
335

    
336
            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
337
                    vec_cmpgt(row0, zero));
338
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
339
                    vec_cmpgt(row1, zero));
340
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
341
                    vec_cmpgt(row2, zero));
342
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
343
                    vec_cmpgt(row3, zero));
344
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
345
                    vec_cmpgt(row4, zero));
346
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
347
                    vec_cmpgt(row5, zero));
348
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
349
                    vec_cmpgt(row6, zero));
350
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
351
                    vec_cmpgt(row7, zero));
352

    
353
            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
354
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
355
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
356
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
357
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
358
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
359
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
360
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);
361

    
362
            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
363
                    vec_cmpgt(alt0, zero));
364
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
365
                    vec_cmpgt(alt1, zero));
366
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
367
                    vec_cmpgt(alt2, zero));
368
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
369
                    vec_cmpgt(alt3, zero));
370
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
371
                    vec_cmpgt(alt4, zero));
372
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
373
                    vec_cmpgt(alt5, zero));
374
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
375
                    vec_cmpgt(alt6, zero));
376
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
377
                    vec_cmpgt(alt7, zero));
378
        }
379

    
380
 
381
    }
382

    
383
    // Store the data back into the original block
384
    {
385
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
386

    
387
        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
388
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
389
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
390
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
391
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
392
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
393
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
394
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
395

    
396
        {
397
            // Clamp for overflow
398
            vector signed int max_q_int, min_q_int;
399
            vector signed short max_q, min_q;
400

    
401
            LOAD4(max_q_int, &(s->max_qcoeff));
402
            LOAD4(min_q_int, &(s->min_qcoeff));
403

    
404
            max_q = vec_pack(max_q_int, max_q_int);
405
            min_q = vec_pack(min_q_int, min_q_int);
406

    
407
            data0 = vec_max(vec_min(data0, max_q), min_q);
408
            data1 = vec_max(vec_min(data1, max_q), min_q);
409
            data2 = vec_max(vec_min(data2, max_q), min_q);
410
            data4 = vec_max(vec_min(data4, max_q), min_q);
411
            data5 = vec_max(vec_min(data5, max_q), min_q);
412
            data6 = vec_max(vec_min(data6, max_q), min_q);
413
            data7 = vec_max(vec_min(data7, max_q), min_q);
414
        }
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
        signed char lastNonZeroChar;
423

    
424
        // Determine the largest non-zero index.
425
        zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
426
                vec_cmpeq(data1, (vector signed short)zero));
427
        zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
428
                vec_cmpeq(data3, (vector signed short)zero));
429
        zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
430
                vec_cmpeq(data5, (vector signed short)zero));
431
        zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
432
                vec_cmpeq(data7, (vector signed short)zero));
433

    
434
        // 64 biggest values
435
        scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
436
        scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
437
        scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
438
        scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
439

    
440
        // 32 largest values
441
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
442
        scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
443

    
444
        // 16 largest values
445
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
446

    
447
        // 8 largest values
448
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
449
                vec_mergel(scanIndices_01, negOne));
450

    
451
        // 4 largest values
452
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
453
                vec_mergel(scanIndices_01, negOne));
454

    
455
        // 2 largest values
456
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
457
                vec_mergel(scanIndices_01, negOne));
458

    
459
        // largest value
460
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
461
                vec_mergel(scanIndices_01, negOne));
462

    
463
        scanIndices_01 = vec_splat(scanIndices_01, 0);
464

    
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

    
490
    // special handling of block[0]
491
    if (s->mb_intra)
492
    {
493
        if (!s->h263_aic)
494
        {
495
            if (n < 4)
496
                oldBaseValue /= s->y_dc_scale;
497
            else
498
                oldBaseValue /= s->c_dc_scale;
499
        }
500

    
501
        // Divide by 8, rounding the result
502
        data[0] = (oldBaseValue + 4) >> 3;
503
    }
504

    
505
    // We handled the tranpose permutation above and we don't
506
    // need to permute the "no" permutation case.
507
    if ((lastNonZero > 0) &&
508
        (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
509
        (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM))
510
    {
511
        ff_block_permute(data, s->dsp.idct_permutation,
512
                s->intra_scantable.scantable, lastNonZero);
513
    }
514

    
515
    return lastNonZero;
516
}
517
#undef FOUROF
518

    
519
/*
520
  AltiVec version of dct_unquantize_h263
521
  this code assumes `block' is 16 bytes-aligned
522
*/
523
void dct_unquantize_h263_altivec(MpegEncContext *s, 
524
                                 DCTELEM *block, int n, int qscale)
525
{
526
POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
527
    int i, level, qmul, qadd;
528
    int nCoeffs;
529
    
530
    assert(s->block_last_index[n]>=0);
531

    
532
POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
533
    
534
    qadd = (qscale - 1) | 1;
535
    qmul = qscale << 1;
536
    
537
    if (s->mb_intra) {
538
        if (!s->h263_aic) {
539
            if (n < 4) 
540
                block[0] = block[0] * s->y_dc_scale;
541
            else
542
                block[0] = block[0] * s->c_dc_scale;
543
        }else
544
            qadd = 0;
545
        i = 1;
546
        nCoeffs= 63; //does not allways use zigzag table 
547
    } else {
548
        i = 0;
549
        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
550
    }
551

    
552
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
553
    for(;i<=nCoeffs;i++) {
554
        level = block[i];
555
        if (level) {
556
            if (level < 0) {
557
                level = level * qmul - qadd;
558
            } else {
559
                level = level * qmul + qadd;
560
            }
561
            block[i] = level;
562
        }
563
    }
564
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
565
    {
566
      register const_vector signed short vczero = (const_vector signed short)vec_splat_s16(0);
567
      short __attribute__ ((aligned(16))) qmul8[] =
568
          {
569
            qmul, qmul, qmul, qmul,
570
            qmul, qmul, qmul, qmul
571
          };
572
      short __attribute__ ((aligned(16))) qadd8[] =
573
          {
574
            qadd, qadd, qadd, qadd,
575
            qadd, qadd, qadd, qadd
576
          };
577
      short __attribute__ ((aligned(16))) nqadd8[] =
578
          {
579
            -qadd, -qadd, -qadd, -qadd,
580
            -qadd, -qadd, -qadd, -qadd
581
          };
582
      register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
583
      register vector bool short blockv_null, blockv_neg;
584
      register short backup_0 = block[0];
585
      register int j = 0;
586
      
587
      qmulv = vec_ld(0, qmul8);
588
      qaddv = vec_ld(0, qadd8);
589
      nqaddv = vec_ld(0, nqadd8);
590

    
591
#if 0 // block *is* 16 bytes-aligned, it seems.
592
      // first make sure block[j] is 16 bytes-aligned
593
      for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
594
        level = block[j];
595
        if (level) {
596
          if (level < 0) {
597
                level = level * qmul - qadd;
598
            } else {
599
                level = level * qmul + qadd;
600
            }
601
            block[j] = level;
602
        }
603
      }
604
#endif
605
      
606
      // vectorize all the 16 bytes-aligned blocks
607
      // of 8 elements
608
      for(; (j + 7) <= nCoeffs ; j+=8)
609
      {
610
        blockv = vec_ld(j << 1, block);
611
        blockv_neg = vec_cmplt(blockv, vczero);
612
        blockv_null = vec_cmpeq(blockv, vczero);
613
        // choose between +qadd or -qadd as the third operand
614
        temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
615
        // multiply & add (block{i,i+7} * qmul [+-] qadd)
616
        temp1 = vec_mladd(blockv, qmulv, temp1);
617
        // put 0 where block[{i,i+7} used to have 0
618
        blockv = vec_sel(temp1, blockv, blockv_null);
619
        vec_st(blockv, j << 1, block);
620
      }
621

    
622
      // if nCoeffs isn't a multiple of 8, finish the job
623
      // using good old scalar units.
624
      // (we could do it using a truncated vector,
625
      // but I'm not sure it's worth the hassle)
626
      for(; j <= nCoeffs ; j++) {
627
        level = block[j];
628
        if (level) {
629
          if (level < 0) {
630
                level = level * qmul - qadd;
631
            } else {
632
                level = level * qmul + qadd;
633
            }
634
            block[j] = level;
635
        }
636
      }
637
      
638
      if (i == 1)
639
      { // cheat. this avoid special-casing the first iteration
640
        block[0] = backup_0;
641
      }
642
    }
643
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
644

    
645
POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
646
}