Statistics
| Branch: | Revision:

ffmpeg / libavcodec / ppc / mpegvideo_altivec.c @ f62a9a46

History | View | Annotate | Download (24.9 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 "dsputil.h"
27
#include "mpegvideo.h"
28

    
29
#include "gcc_fixes.h"
30

    
31
#include "dsputil_ppc.h"
32
#include "util_altivec.h"
33
// Swaps two variables (used for altivec registers)
34
#define SWAP(a,b) \
35
do { \
36
    __typeof__(a) swap_temp=a; \
37
    a=b; \
38
    b=swap_temp; \
39
} while (0)
40

    
41
// transposes a matrix consisting of four vectors with four elements each
42
#define TRANSPOSE4(a,b,c,d) \
43
do { \
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); \
48
 \
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); \
53
} while (0)
54

    
55

    
56
// Loads a four-byte value (int or float) from the target address
57
// into every element in the target vector.  Only works if the
58
// target address is four-byte aligned (which should be always).
59
#define LOAD4(vec, address) \
60
{ \
61
    __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
62
    vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
63
    vec = vec_ld(0, _load_addr); \
64
    vec = vec_perm(vec, vec, _perm_vec); \
65
    vec = vec_splat(vec, 0); \
66
}
67

    
68

    
69
#ifdef __APPLE_CC__
70
#define FOUROF(a) (a)
71
#else
72
// slower, for dumb non-apple GCC
73
#define FOUROF(a) {a,a,a,a}
74
#endif
75
int dct_quantize_altivec(MpegEncContext* s,
76
                        DCTELEM* data, int n,
77
                        int qscale, int* overflow)
78
{
79
    int lastNonZero;
80
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
81
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
82
    const vector float zero = (const vector float)FOUROF(0.);
83
    // used after quantise step
84
    int oldBaseValue = 0;
85

    
86
    // Load the data into the row/alt vectors
87
    {
88
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
89

    
90
        data0 = vec_ld(0, data);
91
        data1 = vec_ld(16, data);
92
        data2 = vec_ld(32, data);
93
        data3 = vec_ld(48, data);
94
        data4 = vec_ld(64, data);
95
        data5 = vec_ld(80, data);
96
        data6 = vec_ld(96, data);
97
        data7 = vec_ld(112, data);
98

    
99
        // Transpose the data before we start
100
        TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
101

    
102
        // load the data into floating point vectors.  We load
103
        // the high half of each row into the main row vectors
104
        // and the low half into the alt vectors.
105
        row0 = vec_ctf(vec_unpackh(data0), 0);
106
        alt0 = vec_ctf(vec_unpackl(data0), 0);
107
        row1 = vec_ctf(vec_unpackh(data1), 0);
108
        alt1 = vec_ctf(vec_unpackl(data1), 0);
109
        row2 = vec_ctf(vec_unpackh(data2), 0);
110
        alt2 = vec_ctf(vec_unpackl(data2), 0);
111
        row3 = vec_ctf(vec_unpackh(data3), 0);
112
        alt3 = vec_ctf(vec_unpackl(data3), 0);
113
        row4 = vec_ctf(vec_unpackh(data4), 0);
114
        alt4 = vec_ctf(vec_unpackl(data4), 0);
115
        row5 = vec_ctf(vec_unpackh(data5), 0);
116
        alt5 = vec_ctf(vec_unpackl(data5), 0);
117
        row6 = vec_ctf(vec_unpackh(data6), 0);
118
        alt6 = vec_ctf(vec_unpackl(data6), 0);
119
        row7 = vec_ctf(vec_unpackh(data7), 0);
120
        alt7 = vec_ctf(vec_unpackl(data7), 0);
121
    }
122

    
123
    // The following block could exist as a separate an altivec dct
124
                // function.  However, if we put it inline, the DCT data can remain
125
                // in the vector local variables, as floats, which we'll use during the
126
                // quantize step...
127
    {
128
        const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
129
        const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
130
        const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
131
        const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
132
        const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
133
        const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
134
        const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
135
        const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
136
        const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
137
        const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
138
        const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
139
        const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
140

    
141

    
142
        int whichPass, whichHalf;
143

    
144
        for(whichPass = 1; whichPass<=2; whichPass++)
145
        {
146
            for(whichHalf = 1; whichHalf<=2; whichHalf++)
147
            {
148
                vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
149
                vector float tmp10, tmp11, tmp12, tmp13;
150
                vector float z1, z2, z3, z4, z5;
151

    
152
                tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
153
                tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
154
                tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
155
                tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
156
                tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
157
                tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
158
                tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
159
                tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
160

    
161
                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
162
                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
163
                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
164
                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
165

    
166

    
167
                // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
168
                row0 = vec_add(tmp10, tmp11);
169

    
170
                // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
171
                row4 = vec_sub(tmp10, tmp11);
172

    
173

    
174
                // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
175
                z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
176

    
177
                // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
178
                //                                CONST_BITS-PASS1_BITS);
179
                row2 = vec_madd(tmp13, vec_0_765366865, z1);
180

    
181
                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
182
                //                                CONST_BITS-PASS1_BITS);
183
                row6 = vec_madd(tmp12, vec_1_847759065, z1);
184

    
185
                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
186
                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
187
                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
188
                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
189

    
190
                // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
191
                z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
192

    
193
                // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
194
                z3 = vec_madd(z3, vec_1_961570560, z5);
195

    
196
                // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
197
                z4 = vec_madd(z4, vec_0_390180644, z5);
198

    
199
                // The following adds are rolled into the multiplies above
200
                // z3 = vec_add(z3, z5);  // z3 += z5;
201
                // z4 = vec_add(z4, z5);  // z4 += z5;
202

    
203
                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
204
                // Wow!  It's actually more effecient to roll this multiply
205
                // into the adds below, even thought the multiply gets done twice!
206
                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
207

    
208
                // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
209
                // Same with this one...
210
                // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
211

    
212
                // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
213
                // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
214
                row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
215

    
216
                // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
217
                // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
218
                row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
219

    
220
                // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
221
                // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
222
                row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
223

    
224
                // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
225
                // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
226
                row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
227

    
228
                // Swap the row values with the alts.  If this is the first half,
229
                // this sets up the low values to be acted on in the second half.
230
                // If this is the second half, it puts the high values back in
231
                // the row values where they are expected to be when we're done.
232
                SWAP(row0, alt0);
233
                SWAP(row1, alt1);
234
                SWAP(row2, alt2);
235
                SWAP(row3, alt3);
236
                SWAP(row4, alt4);
237
                SWAP(row5, alt5);
238
                SWAP(row6, alt6);
239
                SWAP(row7, alt7);
240
            }
241

    
242
            if (whichPass == 1)
243
            {
244
                // transpose the data for the second pass
245

    
246
                // First, block transpose the upper right with lower left.
247
                SWAP(row4, alt0);
248
                SWAP(row5, alt1);
249
                SWAP(row6, alt2);
250
                SWAP(row7, alt3);
251

    
252
                // Now, transpose each block of four
253
                TRANSPOSE4(row0, row1, row2, row3);
254
                TRANSPOSE4(row4, row5, row6, row7);
255
                TRANSPOSE4(alt0, alt1, alt2, alt3);
256
                TRANSPOSE4(alt4, alt5, alt6, alt7);
257
            }
258
        }
259
    }
260

    
261
    // perform the quantise step, using the floating point data
262
    // still in the row/alt registers
263
    {
264
        const int* biasAddr;
265
        const vector signed int* qmat;
266
        vector float bias, negBias;
267

    
268
        if (s->mb_intra)
269
        {
270
            vector signed int baseVector;
271

    
272
            // We must cache element 0 in the intra case
273
            // (it needs special handling).
274
            baseVector = vec_cts(vec_splat(row0, 0), 0);
275
            vec_ste(baseVector, 0, &oldBaseValue);
276

    
277
            qmat = (vector signed int*)s->q_intra_matrix[qscale];
278
            biasAddr = &(s->intra_quant_bias);
279
        }
280
        else
281
        {
282
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
283
            biasAddr = &(s->inter_quant_bias);
284
        }
285

    
286
        // Load the bias vector (We add 0.5 to the bias so that we're
287
                                // rounding when we convert to int, instead of flooring.)
288
        {
289
            vector signed int biasInt;
290
            const vector float negOneFloat = (vector float)FOUROF(-1.0f);
291
            LOAD4(biasInt, biasAddr);
292
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
293
            negBias = vec_madd(bias, negOneFloat, zero);
294
        }
295

    
296
        {
297
            vector float q0, q1, q2, q3, q4, q5, q6, q7;
298

    
299
            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
300
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
301
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
302
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
303
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
304
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
305
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
306
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);
307

    
308
            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
309
                    vec_cmpgt(row0, zero));
310
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
311
                    vec_cmpgt(row1, zero));
312
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
313
                    vec_cmpgt(row2, zero));
314
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
315
                    vec_cmpgt(row3, zero));
316
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
317
                    vec_cmpgt(row4, zero));
318
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
319
                    vec_cmpgt(row5, zero));
320
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
321
                    vec_cmpgt(row6, zero));
322
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
323
                    vec_cmpgt(row7, zero));
324

    
325
            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
326
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
327
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
328
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
329
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
330
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
331
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
332
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);
333

    
334
            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
335
                    vec_cmpgt(alt0, zero));
336
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
337
                    vec_cmpgt(alt1, zero));
338
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
339
                    vec_cmpgt(alt2, zero));
340
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
341
                    vec_cmpgt(alt3, zero));
342
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
343
                    vec_cmpgt(alt4, zero));
344
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
345
                    vec_cmpgt(alt5, zero));
346
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
347
                    vec_cmpgt(alt6, zero));
348
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
349
                    vec_cmpgt(alt7, zero));
350
        }
351

    
352

    
353
    }
354

    
355
    // Store the data back into the original block
356
    {
357
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
358

    
359
        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
360
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
361
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
362
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
363
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
364
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
365
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
366
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
367

    
368
        {
369
            // Clamp for overflow
370
            vector signed int max_q_int, min_q_int;
371
            vector signed short max_q, min_q;
372

    
373
            LOAD4(max_q_int, &(s->max_qcoeff));
374
            LOAD4(min_q_int, &(s->min_qcoeff));
375

    
376
            max_q = vec_pack(max_q_int, max_q_int);
377
            min_q = vec_pack(min_q_int, min_q_int);
378

    
379
            data0 = vec_max(vec_min(data0, max_q), min_q);
380
            data1 = vec_max(vec_min(data1, max_q), min_q);
381
            data2 = vec_max(vec_min(data2, max_q), min_q);
382
            data4 = vec_max(vec_min(data4, max_q), min_q);
383
            data5 = vec_max(vec_min(data5, max_q), min_q);
384
            data6 = vec_max(vec_min(data6, max_q), min_q);
385
            data7 = vec_max(vec_min(data7, max_q), min_q);
386
        }
387

    
388
        {
389
        vector bool char zero_01, zero_23, zero_45, zero_67;
390
        vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
391
        vector signed char negOne = vec_splat_s8(-1);
392
        vector signed char* scanPtr =
393
                (vector signed char*)(s->intra_scantable.inverse);
394
        signed char lastNonZeroChar;
395

    
396
        // Determine the largest non-zero index.
397
        zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
398
                vec_cmpeq(data1, (vector signed short)zero));
399
        zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
400
                vec_cmpeq(data3, (vector signed short)zero));
401
        zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
402
                vec_cmpeq(data5, (vector signed short)zero));
403
        zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
404
                vec_cmpeq(data7, (vector signed short)zero));
405

    
406
        // 64 biggest values
407
        scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
408
        scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
409
        scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
410
        scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
411

    
412
        // 32 largest values
413
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
414
        scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
415

    
416
        // 16 largest values
417
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
418

    
419
        // 8 largest values
420
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
421
                vec_mergel(scanIndices_01, negOne));
422

    
423
        // 4 largest values
424
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
425
                vec_mergel(scanIndices_01, negOne));
426

    
427
        // 2 largest values
428
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
429
                vec_mergel(scanIndices_01, negOne));
430

    
431
        // largest value
432
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
433
                vec_mergel(scanIndices_01, negOne));
434

    
435
        scanIndices_01 = vec_splat(scanIndices_01, 0);
436

    
437

    
438
        vec_ste(scanIndices_01, 0, &lastNonZeroChar);
439

    
440
        lastNonZero = lastNonZeroChar;
441

    
442
        // While the data is still in vectors we check for the transpose IDCT permute
443
        // and handle it using the vector unit if we can.  This is the permute used
444
        // by the altivec idct, so it is common when using the altivec dct.
445

    
446
        if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
447
        {
448
            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
449
        }
450

    
451
        vec_st(data0, 0, data);
452
        vec_st(data1, 16, data);
453
        vec_st(data2, 32, data);
454
        vec_st(data3, 48, data);
455
        vec_st(data4, 64, data);
456
        vec_st(data5, 80, data);
457
        vec_st(data6, 96, data);
458
        vec_st(data7, 112, data);
459
        }
460
    }
461

    
462
    // special handling of block[0]
463
    if (s->mb_intra)
464
    {
465
        if (!s->h263_aic)
466
        {
467
            if (n < 4)
468
                oldBaseValue /= s->y_dc_scale;
469
            else
470
                oldBaseValue /= s->c_dc_scale;
471
        }
472

    
473
        // Divide by 8, rounding the result
474
        data[0] = (oldBaseValue + 4) >> 3;
475
    }
476

    
477
    // We handled the tranpose permutation above and we don't
478
    // need to permute the "no" permutation case.
479
    if ((lastNonZero > 0) &&
480
        (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
481
        (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM))
482
    {
483
        ff_block_permute(data, s->dsp.idct_permutation,
484
                s->intra_scantable.scantable, lastNonZero);
485
    }
486

    
487
    return lastNonZero;
488
}
489
#undef FOUROF
490

    
491
/*
492
  AltiVec version of dct_unquantize_h263
493
  this code assumes `block' is 16 bytes-aligned
494
*/
495
void dct_unquantize_h263_altivec(MpegEncContext *s,
496
                                 DCTELEM *block, int n, int qscale)
497
{
498
POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
499
    int i, level, qmul, qadd;
500
    int nCoeffs;
501

    
502
    assert(s->block_last_index[n]>=0);
503

    
504
POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
505

    
506
    qadd = (qscale - 1) | 1;
507
    qmul = qscale << 1;
508

    
509
    if (s->mb_intra) {
510
        if (!s->h263_aic) {
511
            if (n < 4)
512
                block[0] = block[0] * s->y_dc_scale;
513
            else
514
                block[0] = block[0] * s->c_dc_scale;
515
        }else
516
            qadd = 0;
517
        i = 1;
518
        nCoeffs= 63; //does not always use zigzag table
519
    } else {
520
        i = 0;
521
        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
522
    }
523

    
524
    {
525
      register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
526
      DECLARE_ALIGNED_16(short, qmul8[]) =
527
          {
528
            qmul, qmul, qmul, qmul,
529
            qmul, qmul, qmul, qmul
530
          };
531
      DECLARE_ALIGNED_16(short, qadd8[]) =
532
          {
533
            qadd, qadd, qadd, qadd,
534
            qadd, qadd, qadd, qadd
535
          };
536
      DECLARE_ALIGNED_16(short, nqadd8[]) =
537
          {
538
            -qadd, -qadd, -qadd, -qadd,
539
            -qadd, -qadd, -qadd, -qadd
540
          };
541
      register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
542
      register vector bool short blockv_null, blockv_neg;
543
      register short backup_0 = block[0];
544
      register int j = 0;
545

    
546
      qmulv = vec_ld(0, qmul8);
547
      qaddv = vec_ld(0, qadd8);
548
      nqaddv = vec_ld(0, nqadd8);
549

    
550
#if 0 // block *is* 16 bytes-aligned, it seems.
551
      // first make sure block[j] is 16 bytes-aligned
552
      for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
553
        level = block[j];
554
        if (level) {
555
          if (level < 0) {
556
                level = level * qmul - qadd;
557
            } else {
558
                level = level * qmul + qadd;
559
            }
560
            block[j] = level;
561
        }
562
      }
563
#endif
564

    
565
      // vectorize all the 16 bytes-aligned blocks
566
      // of 8 elements
567
      for(; (j + 7) <= nCoeffs ; j+=8)
568
      {
569
        blockv = vec_ld(j << 1, block);
570
        blockv_neg = vec_cmplt(blockv, vczero);
571
        blockv_null = vec_cmpeq(blockv, vczero);
572
        // choose between +qadd or -qadd as the third operand
573
        temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
574
        // multiply & add (block{i,i+7} * qmul [+-] qadd)
575
        temp1 = vec_mladd(blockv, qmulv, temp1);
576
        // put 0 where block[{i,i+7} used to have 0
577
        blockv = vec_sel(temp1, blockv, blockv_null);
578
        vec_st(blockv, j << 1, block);
579
      }
580

    
581
      // if nCoeffs isn't a multiple of 8, finish the job
582
      // using good old scalar units.
583
      // (we could do it using a truncated vector,
584
      // but I'm not sure it's worth the hassle)
585
      for(; j <= nCoeffs ; j++) {
586
        level = block[j];
587
        if (level) {
588
          if (level < 0) {
589
                level = level * qmul - qadd;
590
            } else {
591
                level = level * qmul + qadd;
592
            }
593
            block[j] = level;
594
        }
595
      }
596

    
597
      if (i == 1)
598
      { // cheat. this avoid special-casing the first iteration
599
        block[0] = backup_0;
600
      }
601
    }
602
POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
603
}
604

    
605

    
606
extern void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block);
607
extern void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block);
608

    
609
void MPV_common_init_altivec(MpegEncContext *s)
610
{
611
    if (s->avctx->lowres==0)
612
    {
613
        if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
614
                (s->avctx->idct_algo == FF_IDCT_ALTIVEC))
615
        {
616
            s->dsp.idct_put = idct_put_altivec;
617
            s->dsp.idct_add = idct_add_altivec;
618
            s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
619
        }
620
    }
621

    
622
    // Test to make sure that the dct required alignments are met.
623
    if ((((long)(s->q_intra_matrix) & 0x0f) != 0) ||
624
        (((long)(s->q_inter_matrix) & 0x0f) != 0))
625
    {
626
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned "
627
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
628
        return;
629
    }
630

    
631
    if (((long)(s->intra_scantable.inverse) & 0x0f) != 0)
632
    {
633
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned "
634
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
635
        return;
636
    }
637

    
638

    
639
    if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
640
            (s->avctx->dct_algo == FF_DCT_ALTIVEC))
641
    {
642
#if 0 /* seems to cause trouble under some circumstances */
643
        s->dct_quantize = dct_quantize_altivec;
644
#endif
645
        s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
646
        s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
647
    }
648
}