Statistics
| Branch: | Revision:

ffmpeg / libavcodec / ppc / mpegvideo_altivec.c @ d6267d02

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
#define FOUROF(a) AVV(a,a,a,a)
70

    
71
int dct_quantize_altivec(MpegEncContext* s,
72
                        DCTELEM* data, int n,
73
                        int qscale, int* overflow)
74
{
75
    int lastNonZero;
76
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
77
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
78
    const vector float zero = (const vector float)FOUROF(0.);
79
    // used after quantize step
80
    int oldBaseValue = 0;
81

    
82
    // Load the data into the row/alt vectors
83
    {
84
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
85

    
86
        data0 = vec_ld(0, data);
87
        data1 = vec_ld(16, data);
88
        data2 = vec_ld(32, data);
89
        data3 = vec_ld(48, data);
90
        data4 = vec_ld(64, data);
91
        data5 = vec_ld(80, data);
92
        data6 = vec_ld(96, data);
93
        data7 = vec_ld(112, data);
94

    
95
        // Transpose the data before we start
96
        TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
97

    
98
        // load the data into floating point vectors.  We load
99
        // the high half of each row into the main row vectors
100
        // and the low half into the alt vectors.
101
        row0 = vec_ctf(vec_unpackh(data0), 0);
102
        alt0 = vec_ctf(vec_unpackl(data0), 0);
103
        row1 = vec_ctf(vec_unpackh(data1), 0);
104
        alt1 = vec_ctf(vec_unpackl(data1), 0);
105
        row2 = vec_ctf(vec_unpackh(data2), 0);
106
        alt2 = vec_ctf(vec_unpackl(data2), 0);
107
        row3 = vec_ctf(vec_unpackh(data3), 0);
108
        alt3 = vec_ctf(vec_unpackl(data3), 0);
109
        row4 = vec_ctf(vec_unpackh(data4), 0);
110
        alt4 = vec_ctf(vec_unpackl(data4), 0);
111
        row5 = vec_ctf(vec_unpackh(data5), 0);
112
        alt5 = vec_ctf(vec_unpackl(data5), 0);
113
        row6 = vec_ctf(vec_unpackh(data6), 0);
114
        alt6 = vec_ctf(vec_unpackl(data6), 0);
115
        row7 = vec_ctf(vec_unpackh(data7), 0);
116
        alt7 = vec_ctf(vec_unpackl(data7), 0);
117
    }
118

    
119
    // The following block could exist as a separate an altivec dct
120
                // function.  However, if we put it inline, the DCT data can remain
121
                // in the vector local variables, as floats, which we'll use during the
122
                // quantize step...
123
    {
124
        const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
125
        const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
126
        const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
127
        const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
128
        const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
129
        const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
130
        const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
131
        const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
132
        const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
133
        const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
134
        const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
135
        const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
136

    
137

    
138
        int whichPass, whichHalf;
139

    
140
        for(whichPass = 1; whichPass<=2; whichPass++)
141
        {
142
            for(whichHalf = 1; whichHalf<=2; whichHalf++)
143
            {
144
                vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
145
                vector float tmp10, tmp11, tmp12, tmp13;
146
                vector float z1, z2, z3, z4, z5;
147

    
148
                tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
149
                tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
150
                tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
151
                tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
152
                tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
153
                tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
154
                tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
155
                tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
156

    
157
                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
158
                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
159
                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
160
                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
161

    
162

    
163
                // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
164
                row0 = vec_add(tmp10, tmp11);
165

    
166
                // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
167
                row4 = vec_sub(tmp10, tmp11);
168

    
169

    
170
                // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
171
                z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
172

    
173
                // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
174
                //                                CONST_BITS-PASS1_BITS);
175
                row2 = vec_madd(tmp13, vec_0_765366865, z1);
176

    
177
                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
178
                //                                CONST_BITS-PASS1_BITS);
179
                row6 = vec_madd(tmp12, vec_1_847759065, z1);
180

    
181
                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
182
                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
183
                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
184
                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
185

    
186
                // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
187
                z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
188

    
189
                // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
190
                z3 = vec_madd(z3, vec_1_961570560, z5);
191

    
192
                // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
193
                z4 = vec_madd(z4, vec_0_390180644, z5);
194

    
195
                // The following adds are rolled into the multiplies above
196
                // z3 = vec_add(z3, z5);  // z3 += z5;
197
                // z4 = vec_add(z4, z5);  // z4 += z5;
198

    
199
                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
200
                // Wow!  It's actually more effecient to roll this multiply
201
                // into the adds below, even thought the multiply gets done twice!
202
                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
203

    
204
                // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
205
                // Same with this one...
206
                // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
207

    
208
                // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
209
                // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
210
                row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
211

    
212
                // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
213
                // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
214
                row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
215

    
216
                // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
217
                // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
218
                row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
219

    
220
                // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
221
                // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
222
                row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
223

    
224
                // Swap the row values with the alts.  If this is the first half,
225
                // this sets up the low values to be acted on in the second half.
226
                // If this is the second half, it puts the high values back in
227
                // the row values where they are expected to be when we're done.
228
                SWAP(row0, alt0);
229
                SWAP(row1, alt1);
230
                SWAP(row2, alt2);
231
                SWAP(row3, alt3);
232
                SWAP(row4, alt4);
233
                SWAP(row5, alt5);
234
                SWAP(row6, alt6);
235
                SWAP(row7, alt7);
236
            }
237

    
238
            if (whichPass == 1)
239
            {
240
                // transpose the data for the second pass
241

    
242
                // First, block transpose the upper right with lower left.
243
                SWAP(row4, alt0);
244
                SWAP(row5, alt1);
245
                SWAP(row6, alt2);
246
                SWAP(row7, alt3);
247

    
248
                // Now, transpose each block of four
249
                TRANSPOSE4(row0, row1, row2, row3);
250
                TRANSPOSE4(row4, row5, row6, row7);
251
                TRANSPOSE4(alt0, alt1, alt2, alt3);
252
                TRANSPOSE4(alt4, alt5, alt6, alt7);
253
            }
254
        }
255
    }
256

    
257
    // perform the quantize step, using the floating point data
258
    // still in the row/alt registers
259
    {
260
        const int* biasAddr;
261
        const vector signed int* qmat;
262
        vector float bias, negBias;
263

    
264
        if (s->mb_intra)
265
        {
266
            vector signed int baseVector;
267

    
268
            // We must cache element 0 in the intra case
269
            // (it needs special handling).
270
            baseVector = vec_cts(vec_splat(row0, 0), 0);
271
            vec_ste(baseVector, 0, &oldBaseValue);
272

    
273
            qmat = (vector signed int*)s->q_intra_matrix[qscale];
274
            biasAddr = &(s->intra_quant_bias);
275
        }
276
        else
277
        {
278
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
279
            biasAddr = &(s->inter_quant_bias);
280
        }
281

    
282
        // Load the bias vector (We add 0.5 to the bias so that we're
283
                                // rounding when we convert to int, instead of flooring.)
284
        {
285
            vector signed int biasInt;
286
            const vector float negOneFloat = (vector float)FOUROF(-1.0f);
287
            LOAD4(biasInt, biasAddr);
288
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
289
            negBias = vec_madd(bias, negOneFloat, zero);
290
        }
291

    
292
        {
293
            vector float q0, q1, q2, q3, q4, q5, q6, q7;
294

    
295
            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
296
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
297
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
298
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
299
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
300
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
301
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
302
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);
303

    
304
            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
305
                    vec_cmpgt(row0, zero));
306
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
307
                    vec_cmpgt(row1, zero));
308
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
309
                    vec_cmpgt(row2, zero));
310
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
311
                    vec_cmpgt(row3, zero));
312
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
313
                    vec_cmpgt(row4, zero));
314
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
315
                    vec_cmpgt(row5, zero));
316
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
317
                    vec_cmpgt(row6, zero));
318
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
319
                    vec_cmpgt(row7, zero));
320

    
321
            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
322
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
323
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
324
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
325
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
326
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
327
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
328
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);
329

    
330
            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
331
                    vec_cmpgt(alt0, zero));
332
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
333
                    vec_cmpgt(alt1, zero));
334
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
335
                    vec_cmpgt(alt2, zero));
336
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
337
                    vec_cmpgt(alt3, zero));
338
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
339
                    vec_cmpgt(alt4, zero));
340
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
341
                    vec_cmpgt(alt5, zero));
342
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
343
                    vec_cmpgt(alt6, zero));
344
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
345
                    vec_cmpgt(alt7, zero));
346
        }
347

    
348

    
349
    }
350

    
351
    // Store the data back into the original block
352
    {
353
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
354

    
355
        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
356
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
357
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
358
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
359
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
360
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
361
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
362
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
363

    
364
        {
365
            // Clamp for overflow
366
            vector signed int max_q_int, min_q_int;
367
            vector signed short max_q, min_q;
368

    
369
            LOAD4(max_q_int, &(s->max_qcoeff));
370
            LOAD4(min_q_int, &(s->min_qcoeff));
371

    
372
            max_q = vec_pack(max_q_int, max_q_int);
373
            min_q = vec_pack(min_q_int, min_q_int);
374

    
375
            data0 = vec_max(vec_min(data0, max_q), min_q);
376
            data1 = vec_max(vec_min(data1, max_q), min_q);
377
            data2 = vec_max(vec_min(data2, max_q), min_q);
378
            data4 = vec_max(vec_min(data4, max_q), min_q);
379
            data5 = vec_max(vec_min(data5, max_q), min_q);
380
            data6 = vec_max(vec_min(data6, max_q), min_q);
381
            data7 = vec_max(vec_min(data7, max_q), min_q);
382
        }
383

    
384
        {
385
        vector bool char zero_01, zero_23, zero_45, zero_67;
386
        vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
387
        vector signed char negOne = vec_splat_s8(-1);
388
        vector signed char* scanPtr =
389
                (vector signed char*)(s->intra_scantable.inverse);
390
        signed char lastNonZeroChar;
391

    
392
        // Determine the largest non-zero index.
393
        zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
394
                vec_cmpeq(data1, (vector signed short)zero));
395
        zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
396
                vec_cmpeq(data3, (vector signed short)zero));
397
        zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
398
                vec_cmpeq(data5, (vector signed short)zero));
399
        zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
400
                vec_cmpeq(data7, (vector signed short)zero));
401

    
402
        // 64 biggest values
403
        scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
404
        scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
405
        scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
406
        scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
407

    
408
        // 32 largest values
409
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
410
        scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
411

    
412
        // 16 largest values
413
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
414

    
415
        // 8 largest values
416
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
417
                vec_mergel(scanIndices_01, negOne));
418

    
419
        // 4 largest values
420
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
421
                vec_mergel(scanIndices_01, negOne));
422

    
423
        // 2 largest values
424
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
425
                vec_mergel(scanIndices_01, negOne));
426

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

    
431
        scanIndices_01 = vec_splat(scanIndices_01, 0);
432

    
433

    
434
        vec_ste(scanIndices_01, 0, &lastNonZeroChar);
435

    
436
        lastNonZero = lastNonZeroChar;
437

    
438
        // While the data is still in vectors we check for the transpose IDCT permute
439
        // and handle it using the vector unit if we can.  This is the permute used
440
        // by the altivec idct, so it is common when using the altivec dct.
441

    
442
        if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
443
        {
444
            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
445
        }
446

    
447
        vec_st(data0, 0, data);
448
        vec_st(data1, 16, data);
449
        vec_st(data2, 32, data);
450
        vec_st(data3, 48, data);
451
        vec_st(data4, 64, data);
452
        vec_st(data5, 80, data);
453
        vec_st(data6, 96, data);
454
        vec_st(data7, 112, data);
455
        }
456
    }
457

    
458
    // special handling of block[0]
459
    if (s->mb_intra)
460
    {
461
        if (!s->h263_aic)
462
        {
463
            if (n < 4)
464
                oldBaseValue /= s->y_dc_scale;
465
            else
466
                oldBaseValue /= s->c_dc_scale;
467
        }
468

    
469
        // Divide by 8, rounding the result
470
        data[0] = (oldBaseValue + 4) >> 3;
471
    }
472

    
473
    // We handled the transpose permutation above and we don't
474
    // need to permute the "no" permutation case.
475
    if ((lastNonZero > 0) &&
476
        (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
477
        (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM))
478
    {
479
        ff_block_permute(data, s->dsp.idct_permutation,
480
                s->intra_scantable.scantable, lastNonZero);
481
    }
482

    
483
    return lastNonZero;
484
}
485

    
486
/*
487
  AltiVec version of dct_unquantize_h263
488
  this code assumes `block' is 16 bytes-aligned
489
*/
490
void dct_unquantize_h263_altivec(MpegEncContext *s,
491
                                 DCTELEM *block, int n, int qscale)
492
{
493
POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
494
    int i, level, qmul, qadd;
495
    int nCoeffs;
496

    
497
    assert(s->block_last_index[n]>=0);
498

    
499
POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
500

    
501
    qadd = (qscale - 1) | 1;
502
    qmul = qscale << 1;
503

    
504
    if (s->mb_intra) {
505
        if (!s->h263_aic) {
506
            if (n < 4)
507
                block[0] = block[0] * s->y_dc_scale;
508
            else
509
                block[0] = block[0] * s->c_dc_scale;
510
        }else
511
            qadd = 0;
512
        i = 1;
513
        nCoeffs= 63; //does not always use zigzag table
514
    } else {
515
        i = 0;
516
        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
517
    }
518

    
519
    {
520
      register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
521
      DECLARE_ALIGNED_16(short, qmul8[]) =
522
          {
523
            qmul, qmul, qmul, qmul,
524
            qmul, qmul, qmul, qmul
525
          };
526
      DECLARE_ALIGNED_16(short, qadd8[]) =
527
          {
528
            qadd, qadd, qadd, qadd,
529
            qadd, qadd, qadd, qadd
530
          };
531
      DECLARE_ALIGNED_16(short, nqadd8[]) =
532
          {
533
            -qadd, -qadd, -qadd, -qadd,
534
            -qadd, -qadd, -qadd, -qadd
535
          };
536
      register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
537
      register vector bool short blockv_null, blockv_neg;
538
      register short backup_0 = block[0];
539
      register int j = 0;
540

    
541
      qmulv = vec_ld(0, qmul8);
542
      qaddv = vec_ld(0, qadd8);
543
      nqaddv = vec_ld(0, nqadd8);
544

    
545
#if 0 // block *is* 16 bytes-aligned, it seems.
546
      // first make sure block[j] is 16 bytes-aligned
547
      for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
548
        level = block[j];
549
        if (level) {
550
          if (level < 0) {
551
                level = level * qmul - qadd;
552
            } else {
553
                level = level * qmul + qadd;
554
            }
555
            block[j] = level;
556
        }
557
      }
558
#endif
559

    
560
      // vectorize all the 16 bytes-aligned blocks
561
      // of 8 elements
562
      for(; (j + 7) <= nCoeffs ; j+=8)
563
      {
564
        blockv = vec_ld(j << 1, block);
565
        blockv_neg = vec_cmplt(blockv, vczero);
566
        blockv_null = vec_cmpeq(blockv, vczero);
567
        // choose between +qadd or -qadd as the third operand
568
        temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
569
        // multiply & add (block{i,i+7} * qmul [+-] qadd)
570
        temp1 = vec_mladd(blockv, qmulv, temp1);
571
        // put 0 where block[{i,i+7} used to have 0
572
        blockv = vec_sel(temp1, blockv, blockv_null);
573
        vec_st(blockv, j << 1, block);
574
      }
575

    
576
      // if nCoeffs isn't a multiple of 8, finish the job
577
      // using good old scalar units.
578
      // (we could do it using a truncated vector,
579
      // but I'm not sure it's worth the hassle)
580
      for(; j <= nCoeffs ; j++) {
581
        level = block[j];
582
        if (level) {
583
          if (level < 0) {
584
                level = level * qmul - qadd;
585
            } else {
586
                level = level * qmul + qadd;
587
            }
588
            block[j] = level;
589
        }
590
      }
591

    
592
      if (i == 1)
593
      { // cheat. this avoid special-casing the first iteration
594
        block[0] = backup_0;
595
      }
596
    }
597
POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
598
}
599

    
600

    
601
extern void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block);
602
extern void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block);
603

    
604
void MPV_common_init_altivec(MpegEncContext *s)
605
{
606
    if (mm_flags & (MM_ALTIVEC == 0)) return;
607

    
608
    if (s->avctx->lowres==0)
609
    {
610
        if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
611
                (s->avctx->idct_algo == FF_IDCT_ALTIVEC))
612
        {
613
            s->dsp.idct_put = idct_put_altivec;
614
            s->dsp.idct_add = idct_add_altivec;
615
            s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
616
        }
617
    }
618

    
619
    // Test to make sure that the dct required alignments are met.
620
    if ((((long)(s->q_intra_matrix) & 0x0f) != 0) ||
621
        (((long)(s->q_inter_matrix) & 0x0f) != 0))
622
    {
623
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned "
624
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
625
        return;
626
    }
627

    
628
    if (((long)(s->intra_scantable.inverse) & 0x0f) != 0)
629
    {
630
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned "
631
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
632
        return;
633
    }
634

    
635

    
636
    if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
637
            (s->avctx->dct_algo == FF_DCT_ALTIVEC))
638
    {
639
#if 0 /* seems to cause trouble under some circumstances */
640
        s->dct_quantize = dct_quantize_altivec;
641
#endif
642
        s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
643
        s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
644
    }
645
}