Statistics
| Branch: | Revision:

ffmpeg / libavcodec / ppc / mpegvideo_altivec.c @ 5137235e

History | View | Annotate | Download (25.1 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 file is part of FFmpeg.
8
 *
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.
13
 *
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.
18
 *
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
22
 */
23

    
24
#include <stdlib.h>
25
#include <stdio.h>
26
#include "libavcodec/dsputil.h"
27
#include "libavcodec/mpegvideo.h"
28

    
29
#include "dsputil_ppc.h"
30
#include "util_altivec.h"
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

    
54
// Loads a four-byte value (int or float) from the target address
55
// into every element in the target vector.  Only works if the
56
// target address is four-byte aligned (which should be always).
57
#define LOAD4(vec, address) \
58
{ \
59
    __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address);  \
60
    vector unsigned char _perm_vec = vec_lvsl(0,(address));     \
61
    vec = vec_ld(0, _load_addr);                                \
62
    vec = vec_perm(vec, vec, _perm_vec);                        \
63
    vec = vec_splat(vec, 0);                                    \
64
}
65

    
66

    
67
#define FOUROF(a) {a,a,a,a}
68

    
69
int dct_quantize_altivec(MpegEncContext* s,
70
                         DCTELEM* data, int n,
71
                         int qscale, int* overflow)
72
{
73
    int lastNonZero;
74
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
75
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
76
    const vector float zero = (const vector float)FOUROF(0.);
77
    // used after quantize step
78
    int oldBaseValue = 0;
79

    
80
    // Load the data into the row/alt vectors
81
    {
82
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
83

    
84
        data0 = vec_ld(0, data);
85
        data1 = vec_ld(16, data);
86
        data2 = vec_ld(32, data);
87
        data3 = vec_ld(48, data);
88
        data4 = vec_ld(64, data);
89
        data5 = vec_ld(80, data);
90
        data6 = vec_ld(96, data);
91
        data7 = vec_ld(112, data);
92

    
93
        // Transpose the data before we start
94
        TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
95

    
96
        // load the data into floating point vectors.  We load
97
        // the high half of each row into the main row vectors
98
        // and the low half into the alt vectors.
99
        row0 = vec_ctf(vec_unpackh(data0), 0);
100
        alt0 = vec_ctf(vec_unpackl(data0), 0);
101
        row1 = vec_ctf(vec_unpackh(data1), 0);
102
        alt1 = vec_ctf(vec_unpackl(data1), 0);
103
        row2 = vec_ctf(vec_unpackh(data2), 0);
104
        alt2 = vec_ctf(vec_unpackl(data2), 0);
105
        row3 = vec_ctf(vec_unpackh(data3), 0);
106
        alt3 = vec_ctf(vec_unpackl(data3), 0);
107
        row4 = vec_ctf(vec_unpackh(data4), 0);
108
        alt4 = vec_ctf(vec_unpackl(data4), 0);
109
        row5 = vec_ctf(vec_unpackh(data5), 0);
110
        alt5 = vec_ctf(vec_unpackl(data5), 0);
111
        row6 = vec_ctf(vec_unpackh(data6), 0);
112
        alt6 = vec_ctf(vec_unpackl(data6), 0);
113
        row7 = vec_ctf(vec_unpackh(data7), 0);
114
        alt7 = vec_ctf(vec_unpackl(data7), 0);
115
    }
116

    
117
    // The following block could exist as a separate an altivec dct
118
                // function.  However, if we put it inline, the DCT data can remain
119
                // in the vector local variables, as floats, which we'll use during the
120
                // quantize step...
121
    {
122
        const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
123
        const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
124
        const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
125
        const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
126
        const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
127
        const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
128
        const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
129
        const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
130
        const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
131
        const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
132
        const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
133
        const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
134

    
135

    
136
        int whichPass, whichHalf;
137

    
138
        for(whichPass = 1; whichPass<=2; whichPass++) {
139
            for(whichHalf = 1; whichHalf<=2; whichHalf++) {
140
                vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
141
                vector float tmp10, tmp11, tmp12, tmp13;
142
                vector float z1, z2, z3, z4, z5;
143

    
144
                tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
145
                tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
146
                tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
147
                tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
148
                tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
149
                tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
150
                tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
151
                tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
152

    
153
                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
154
                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
155
                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
156
                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
157

    
158

    
159
                // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
160
                row0 = vec_add(tmp10, tmp11);
161

    
162
                // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
163
                row4 = vec_sub(tmp10, tmp11);
164

    
165

    
166
                // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
167
                z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
168

    
169
                // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
170
                //                                CONST_BITS-PASS1_BITS);
171
                row2 = vec_madd(tmp13, vec_0_765366865, z1);
172

    
173
                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
174
                //                                CONST_BITS-PASS1_BITS);
175
                row6 = vec_madd(tmp12, vec_1_847759065, z1);
176

    
177
                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
178
                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
179
                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
180
                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
181

    
182
                // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
183
                z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
184

    
185
                // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
186
                z3 = vec_madd(z3, vec_1_961570560, z5);
187

    
188
                // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
189
                z4 = vec_madd(z4, vec_0_390180644, z5);
190

    
191
                // The following adds are rolled into the multiplies above
192
                // z3 = vec_add(z3, z5);  // z3 += z5;
193
                // z4 = vec_add(z4, z5);  // z4 += z5;
194

    
195
                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
196
                // Wow!  It's actually more efficient to roll this multiply
197
                // into the adds below, even thought the multiply gets done twice!
198
                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
199

    
200
                // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
201
                // Same with this one...
202
                // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
203

    
204
                // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
205
                // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
206
                row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
207

    
208
                // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
209
                // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
210
                row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
211

    
212
                // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
213
                // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
214
                row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
215

    
216
                // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
217
                // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
218
                row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
219

    
220
                // Swap the row values with the alts.  If this is the first half,
221
                // this sets up the low values to be acted on in the second half.
222
                // If this is the second half, it puts the high values back in
223
                // the row values where they are expected to be when we're done.
224
                SWAP(row0, alt0);
225
                SWAP(row1, alt1);
226
                SWAP(row2, alt2);
227
                SWAP(row3, alt3);
228
                SWAP(row4, alt4);
229
                SWAP(row5, alt5);
230
                SWAP(row6, alt6);
231
                SWAP(row7, alt7);
232
            }
233

    
234
            if (whichPass == 1) {
235
                // transpose the data for the second pass
236

    
237
                // First, block transpose the upper right with lower left.
238
                SWAP(row4, alt0);
239
                SWAP(row5, alt1);
240
                SWAP(row6, alt2);
241
                SWAP(row7, alt3);
242

    
243
                // Now, transpose each block of four
244
                TRANSPOSE4(row0, row1, row2, row3);
245
                TRANSPOSE4(row4, row5, row6, row7);
246
                TRANSPOSE4(alt0, alt1, alt2, alt3);
247
                TRANSPOSE4(alt4, alt5, alt6, alt7);
248
            }
249
        }
250
    }
251

    
252
    // perform the quantize step, using the floating point data
253
    // still in the row/alt registers
254
    {
255
        const int* biasAddr;
256
        const vector signed int* qmat;
257
        vector float bias, negBias;
258

    
259
        if (s->mb_intra) {
260
            vector signed int baseVector;
261

    
262
            // We must cache element 0 in the intra case
263
            // (it needs special handling).
264
            baseVector = vec_cts(vec_splat(row0, 0), 0);
265
            vec_ste(baseVector, 0, &oldBaseValue);
266

    
267
            qmat = (vector signed int*)s->q_intra_matrix[qscale];
268
            biasAddr = &(s->intra_quant_bias);
269
        } else {
270
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
271
            biasAddr = &(s->inter_quant_bias);
272
        }
273

    
274
        // Load the bias vector (We add 0.5 to the bias so that we're
275
                                // rounding when we convert to int, instead of flooring.)
276
        {
277
            vector signed int biasInt;
278
            const vector float negOneFloat = (vector float)FOUROF(-1.0f);
279
            LOAD4(biasInt, biasAddr);
280
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
281
            negBias = vec_madd(bias, negOneFloat, zero);
282
        }
283

    
284
        {
285
            vector float q0, q1, q2, q3, q4, q5, q6, q7;
286

    
287
            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
288
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
289
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
290
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
291
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
292
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
293
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
294
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);
295

    
296
            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
297
                    vec_cmpgt(row0, zero));
298
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
299
                    vec_cmpgt(row1, zero));
300
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
301
                    vec_cmpgt(row2, zero));
302
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
303
                    vec_cmpgt(row3, zero));
304
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
305
                    vec_cmpgt(row4, zero));
306
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
307
                    vec_cmpgt(row5, zero));
308
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
309
                    vec_cmpgt(row6, zero));
310
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
311
                    vec_cmpgt(row7, zero));
312

    
313
            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
314
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
315
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
316
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
317
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
318
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
319
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
320
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);
321

    
322
            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
323
                    vec_cmpgt(alt0, zero));
324
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
325
                    vec_cmpgt(alt1, zero));
326
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
327
                    vec_cmpgt(alt2, zero));
328
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
329
                    vec_cmpgt(alt3, zero));
330
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
331
                    vec_cmpgt(alt4, zero));
332
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
333
                    vec_cmpgt(alt5, zero));
334
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
335
                    vec_cmpgt(alt6, zero));
336
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
337
                    vec_cmpgt(alt7, zero));
338
        }
339

    
340

    
341
    }
342

    
343
    // Store the data back into the original block
344
    {
345
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
346

    
347
        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
348
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
349
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
350
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
351
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
352
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
353
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
354
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
355

    
356
        {
357
            // Clamp for overflow
358
            vector signed int max_q_int, min_q_int;
359
            vector signed short max_q, min_q;
360

    
361
            LOAD4(max_q_int, &(s->max_qcoeff));
362
            LOAD4(min_q_int, &(s->min_qcoeff));
363

    
364
            max_q = vec_pack(max_q_int, max_q_int);
365
            min_q = vec_pack(min_q_int, min_q_int);
366

    
367
            data0 = vec_max(vec_min(data0, max_q), min_q);
368
            data1 = vec_max(vec_min(data1, max_q), min_q);
369
            data2 = vec_max(vec_min(data2, max_q), min_q);
370
            data4 = vec_max(vec_min(data4, max_q), min_q);
371
            data5 = vec_max(vec_min(data5, max_q), min_q);
372
            data6 = vec_max(vec_min(data6, max_q), min_q);
373
            data7 = vec_max(vec_min(data7, max_q), min_q);
374
        }
375

    
376
        {
377
        vector bool char zero_01, zero_23, zero_45, zero_67;
378
        vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67;
379
        vector signed char negOne = vec_splat_s8(-1);
380
        vector signed char* scanPtr =
381
                (vector signed char*)(s->intra_scantable.inverse);
382
        signed char lastNonZeroChar;
383

    
384
        // Determine the largest non-zero index.
385
        zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
386
                vec_cmpeq(data1, (vector signed short)zero));
387
        zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
388
                vec_cmpeq(data3, (vector signed short)zero));
389
        zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
390
                vec_cmpeq(data5, (vector signed short)zero));
391
        zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
392
                vec_cmpeq(data7, (vector signed short)zero));
393

    
394
        // 64 biggest values
395
        scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01);
396
        scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23);
397
        scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45);
398
        scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67);
399

    
400
        // 32 largest values
401
        scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23);
402
        scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67);
403

    
404
        // 16 largest values
405
        scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45);
406

    
407
        // 8 largest values
408
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
409
                vec_mergel(scanIndexes_01, negOne));
410

    
411
        // 4 largest values
412
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
413
                vec_mergel(scanIndexes_01, negOne));
414

    
415
        // 2 largest values
416
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
417
                vec_mergel(scanIndexes_01, negOne));
418

    
419
        // largest value
420
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
421
                vec_mergel(scanIndexes_01, negOne));
422

    
423
        scanIndexes_01 = vec_splat(scanIndexes_01, 0);
424

    
425

    
426
        vec_ste(scanIndexes_01, 0, &lastNonZeroChar);
427

    
428
        lastNonZero = lastNonZeroChar;
429

    
430
        // While the data is still in vectors we check for the transpose IDCT permute
431
        // and handle it using the vector unit if we can.  This is the permute used
432
        // by the altivec idct, so it is common when using the altivec dct.
433

    
434
        if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) {
435
            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
436
        }
437

    
438
        vec_st(data0, 0, data);
439
        vec_st(data1, 16, data);
440
        vec_st(data2, 32, data);
441
        vec_st(data3, 48, data);
442
        vec_st(data4, 64, data);
443
        vec_st(data5, 80, data);
444
        vec_st(data6, 96, data);
445
        vec_st(data7, 112, data);
446
        }
447
    }
448

    
449
    // special handling of block[0]
450
    if (s->mb_intra) {
451
        if (!s->h263_aic) {
452
            if (n < 4)
453
                oldBaseValue /= s->y_dc_scale;
454
            else
455
                oldBaseValue /= s->c_dc_scale;
456
        }
457

    
458
        // Divide by 8, rounding the result
459
        data[0] = (oldBaseValue + 4) >> 3;
460
    }
461

    
462
    // We handled the transpose permutation above and we don't
463
    // need to permute the "no" permutation case.
464
    if ((lastNonZero > 0) &&
465
        (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
466
        (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) {
467
        ff_block_permute(data, s->dsp.idct_permutation,
468
                s->intra_scantable.scantable, lastNonZero);
469
    }
470

    
471
    return lastNonZero;
472
}
473

    
474
/* AltiVec version of dct_unquantize_h263
475
   this code assumes `block' is 16 bytes-aligned */
476
void dct_unquantize_h263_altivec(MpegEncContext *s,
477
                                 DCTELEM *block, int n, int qscale)
478
{
479
POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
480
    int i, level, qmul, qadd;
481
    int nCoeffs;
482

    
483
    assert(s->block_last_index[n]>=0);
484

    
485
POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
486

    
487
    qadd = (qscale - 1) | 1;
488
    qmul = qscale << 1;
489

    
490
    if (s->mb_intra) {
491
        if (!s->h263_aic) {
492
            if (n < 4)
493
                block[0] = block[0] * s->y_dc_scale;
494
            else
495
                block[0] = block[0] * s->c_dc_scale;
496
        }else
497
            qadd = 0;
498
        i = 1;
499
        nCoeffs= 63; //does not always use zigzag table
500
    } else {
501
        i = 0;
502
        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
503
    }
504

    
505
    {
506
        register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
507
        DECLARE_ALIGNED_16(short, qmul8[]) =
508
            {
509
              qmul, qmul, qmul, qmul,
510
              qmul, qmul, qmul, qmul
511
            };
512
        DECLARE_ALIGNED_16(short, qadd8[]) =
513
            {
514
              qadd, qadd, qadd, qadd,
515
              qadd, qadd, qadd, qadd
516
            };
517
        DECLARE_ALIGNED_16(short, nqadd8[]) =
518
            {
519
              -qadd, -qadd, -qadd, -qadd,
520
              -qadd, -qadd, -qadd, -qadd
521
            };
522
        register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
523
        register vector bool short blockv_null, blockv_neg;
524
        register short backup_0 = block[0];
525
        register int j = 0;
526

    
527
        qmulv = vec_ld(0, qmul8);
528
        qaddv = vec_ld(0, qadd8);
529
        nqaddv = vec_ld(0, nqadd8);
530

    
531
#if 0   // block *is* 16 bytes-aligned, it seems.
532
        // first make sure block[j] is 16 bytes-aligned
533
        for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
534
            level = block[j];
535
            if (level) {
536
                if (level < 0) {
537
                    level = level * qmul - qadd;
538
                } else {
539
                    level = level * qmul + qadd;
540
                }
541
                block[j] = level;
542
            }
543
        }
544
#endif
545

    
546
        // vectorize all the 16 bytes-aligned blocks
547
        // of 8 elements
548
        for(; (j + 7) <= nCoeffs ; j+=8) {
549
            blockv = vec_ld(j << 1, block);
550
            blockv_neg = vec_cmplt(blockv, vczero);
551
            blockv_null = vec_cmpeq(blockv, vczero);
552
            // choose between +qadd or -qadd as the third operand
553
            temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
554
            // multiply & add (block{i,i+7} * qmul [+-] qadd)
555
            temp1 = vec_mladd(blockv, qmulv, temp1);
556
            // put 0 where block[{i,i+7} used to have 0
557
            blockv = vec_sel(temp1, blockv, blockv_null);
558
            vec_st(blockv, j << 1, block);
559
        }
560

    
561
        // if nCoeffs isn't a multiple of 8, finish the job
562
        // using good old scalar units.
563
        // (we could do it using a truncated vector,
564
        // but I'm not sure it's worth the hassle)
565
        for(; j <= nCoeffs ; j++) {
566
            level = block[j];
567
            if (level) {
568
                if (level < 0) {
569
                    level = level * qmul - qadd;
570
                } else {
571
                    level = level * qmul + qadd;
572
                }
573
                block[j] = level;
574
            }
575
        }
576

    
577
        if (i == 1) {
578
            // cheat. this avoid special-casing the first iteration
579
            block[0] = backup_0;
580
        }
581
    }
582
POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
583
}
584

    
585

    
586
void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block);
587
void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block);
588

    
589
void MPV_common_init_altivec(MpegEncContext *s)
590
{
591
    if ((mm_flags & FF_MM_ALTIVEC) == 0) return;
592

    
593
    if (s->avctx->lowres==0) {
594
        if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
595
            (s->avctx->idct_algo == FF_IDCT_ALTIVEC)) {
596
            s->dsp.idct_put = idct_put_altivec;
597
            s->dsp.idct_add = idct_add_altivec;
598
            s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
599
        }
600
    }
601

    
602
    // Test to make sure that the dct required alignments are met.
603
    if ((((long)(s->q_intra_matrix) & 0x0f) != 0) ||
604
        (((long)(s->q_inter_matrix) & 0x0f) != 0)) {
605
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned "
606
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
607
        return;
608
    }
609

    
610
    if (((long)(s->intra_scantable.inverse) & 0x0f) != 0) {
611
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned "
612
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
613
        return;
614
    }
615

    
616

    
617
    if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
618
            (s->avctx->dct_algo == FF_DCT_ALTIVEC)) {
619
#if 0 /* seems to cause trouble under some circumstances */
620
        s->dct_quantize = dct_quantize_altivec;
621
#endif
622
        s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
623
        s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
624
    }
625
}