Statistics
| Branch: | Revision:

ffmpeg / libavcodec / ppc / mpegvideo_altivec.c @ 2912e87a

History | View | Annotate | Download (24.6 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 Libav.
8
 *
9
 * Libav 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
 * Libav 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 Libav; 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 "libavutil/cpu.h"
27
#include "libavcodec/dsputil.h"
28
#include "libavcodec/mpegvideo.h"
29

    
30
#include "util_altivec.h"
31
#include "types_altivec.h"
32
#include "dsputil_altivec.h"
33

    
34
// Swaps two variables (used for altivec registers)
35
#define SWAP(a,b) \
36
do { \
37
    __typeof__(a) swap_temp=a; \
38
    a=b; \
39
    b=swap_temp; \
40
} while (0)
41

    
42
// transposes a matrix consisting of four vectors with four elements each
43
#define TRANSPOSE4(a,b,c,d) \
44
do { \
45
    __typeof__(a) _trans_ach = vec_mergeh(a, c); \
46
    __typeof__(a) _trans_acl = vec_mergel(a, c); \
47
    __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
48
    __typeof__(a) _trans_bdl = vec_mergel(b, d); \
49
                                                 \
50
    a = vec_mergeh(_trans_ach, _trans_bdh);      \
51
    b = vec_mergel(_trans_ach, _trans_bdh);      \
52
    c = vec_mergeh(_trans_acl, _trans_bdl);      \
53
    d = vec_mergel(_trans_acl, _trans_bdl);      \
54
} while (0)
55

    
56

    
57
// Loads a four-byte value (int or float) from the target address
58
// into every element in the target vector.  Only works if the
59
// target address is four-byte aligned (which should be always).
60
#define LOAD4(vec, address) \
61
{ \
62
    __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address);  \
63
    vector unsigned char _perm_vec = vec_lvsl(0,(address));     \
64
    vec = vec_ld(0, _load_addr);                                \
65
    vec = vec_perm(vec, vec, _perm_vec);                        \
66
    vec = vec_splat(vec, 0);                                    \
67
}
68

    
69

    
70
#define FOUROF(a) {a,a,a,a}
71

    
72
static int dct_quantize_altivec(MpegEncContext* s,
73
                         DCTELEM* data, int n,
74
                         int qscale, int* overflow)
75
{
76
    int lastNonZero;
77
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
78
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
79
    const vector float zero = (const vector float)FOUROF(0.);
80
    // used after quantize step
81
    int oldBaseValue = 0;
82

    
83
    // Load the data into the row/alt vectors
84
    {
85
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
86

    
87
        data0 = vec_ld(0, data);
88
        data1 = vec_ld(16, data);
89
        data2 = vec_ld(32, data);
90
        data3 = vec_ld(48, data);
91
        data4 = vec_ld(64, data);
92
        data5 = vec_ld(80, data);
93
        data6 = vec_ld(96, data);
94
        data7 = vec_ld(112, data);
95

    
96
        // Transpose the data before we start
97
        TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
98

    
99
        // load the data into floating point vectors.  We load
100
        // the high half of each row into the main row vectors
101
        // and the low half into the alt vectors.
102
        row0 = vec_ctf(vec_unpackh(data0), 0);
103
        alt0 = vec_ctf(vec_unpackl(data0), 0);
104
        row1 = vec_ctf(vec_unpackh(data1), 0);
105
        alt1 = vec_ctf(vec_unpackl(data1), 0);
106
        row2 = vec_ctf(vec_unpackh(data2), 0);
107
        alt2 = vec_ctf(vec_unpackl(data2), 0);
108
        row3 = vec_ctf(vec_unpackh(data3), 0);
109
        alt3 = vec_ctf(vec_unpackl(data3), 0);
110
        row4 = vec_ctf(vec_unpackh(data4), 0);
111
        alt4 = vec_ctf(vec_unpackl(data4), 0);
112
        row5 = vec_ctf(vec_unpackh(data5), 0);
113
        alt5 = vec_ctf(vec_unpackl(data5), 0);
114
        row6 = vec_ctf(vec_unpackh(data6), 0);
115
        alt6 = vec_ctf(vec_unpackl(data6), 0);
116
        row7 = vec_ctf(vec_unpackh(data7), 0);
117
        alt7 = vec_ctf(vec_unpackl(data7), 0);
118
    }
119

    
120
    // The following block could exist as a separate an altivec dct
121
                // function.  However, if we put it inline, the DCT data can remain
122
                // in the vector local variables, as floats, which we'll use during the
123
                // quantize step...
124
    {
125
        const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
126
        const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
127
        const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
128
        const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
129
        const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
130
        const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
131
        const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
132
        const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
133
        const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
134
        const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
135
        const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
136
        const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
137

    
138

    
139
        int whichPass, whichHalf;
140

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

    
147
                tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
148
                tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
149
                tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
150
                tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
151
                tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
152
                tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
153
                tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
154
                tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
155

    
156
                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
157
                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
158
                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
159
                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
160

    
161

    
162
                // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
163
                row0 = vec_add(tmp10, tmp11);
164

    
165
                // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
166
                row4 = vec_sub(tmp10, tmp11);
167

    
168

    
169
                // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
170
                z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
171

    
172
                // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
173
                //                                CONST_BITS-PASS1_BITS);
174
                row2 = vec_madd(tmp13, vec_0_765366865, z1);
175

    
176
                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
177
                //                                CONST_BITS-PASS1_BITS);
178
                row6 = vec_madd(tmp12, vec_1_847759065, z1);
179

    
180
                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
181
                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
182
                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
183
                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
184

    
185
                // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
186
                z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
187

    
188
                // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
189
                z3 = vec_madd(z3, vec_1_961570560, z5);
190

    
191
                // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
192
                z4 = vec_madd(z4, vec_0_390180644, z5);
193

    
194
                // The following adds are rolled into the multiplies above
195
                // z3 = vec_add(z3, z5);  // z3 += z5;
196
                // z4 = vec_add(z4, z5);  // z4 += z5;
197

    
198
                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
199
                // Wow!  It's actually more efficient to roll this multiply
200
                // into the adds below, even thought the multiply gets done twice!
201
                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
202

    
203
                // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
204
                // Same with this one...
205
                // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
206

    
207
                // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
208
                // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
209
                row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
210

    
211
                // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
212
                // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
213
                row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
214

    
215
                // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
216
                // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
217
                row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
218

    
219
                // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
220
                // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
221
                row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
222

    
223
                // Swap the row values with the alts.  If this is the first half,
224
                // this sets up the low values to be acted on in the second half.
225
                // If this is the second half, it puts the high values back in
226
                // the row values where they are expected to be when we're done.
227
                SWAP(row0, alt0);
228
                SWAP(row1, alt1);
229
                SWAP(row2, alt2);
230
                SWAP(row3, alt3);
231
                SWAP(row4, alt4);
232
                SWAP(row5, alt5);
233
                SWAP(row6, alt6);
234
                SWAP(row7, alt7);
235
            }
236

    
237
            if (whichPass == 1) {
238
                // transpose the data for the second pass
239

    
240
                // First, block transpose the upper right with lower left.
241
                SWAP(row4, alt0);
242
                SWAP(row5, alt1);
243
                SWAP(row6, alt2);
244
                SWAP(row7, alt3);
245

    
246
                // Now, transpose each block of four
247
                TRANSPOSE4(row0, row1, row2, row3);
248
                TRANSPOSE4(row4, row5, row6, row7);
249
                TRANSPOSE4(alt0, alt1, alt2, alt3);
250
                TRANSPOSE4(alt4, alt5, alt6, alt7);
251
            }
252
        }
253
    }
254

    
255
    // perform the quantize step, using the floating point data
256
    // still in the row/alt registers
257
    {
258
        const int* biasAddr;
259
        const vector signed int* qmat;
260
        vector float bias, negBias;
261

    
262
        if (s->mb_intra) {
263
            vector signed int baseVector;
264

    
265
            // We must cache element 0 in the intra case
266
            // (it needs special handling).
267
            baseVector = vec_cts(vec_splat(row0, 0), 0);
268
            vec_ste(baseVector, 0, &oldBaseValue);
269

    
270
            qmat = (vector signed int*)s->q_intra_matrix[qscale];
271
            biasAddr = &(s->intra_quant_bias);
272
        } else {
273
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
274
            biasAddr = &(s->inter_quant_bias);
275
        }
276

    
277
        // Load the bias vector (We add 0.5 to the bias so that we're
278
                                // rounding when we convert to int, instead of flooring.)
279
        {
280
            vector signed int biasInt;
281
            const vector float negOneFloat = (vector float)FOUROF(-1.0f);
282
            LOAD4(biasInt, biasAddr);
283
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
284
            negBias = vec_madd(bias, negOneFloat, zero);
285
        }
286

    
287
        {
288
            vector float q0, q1, q2, q3, q4, q5, q6, q7;
289

    
290
            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
291
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
292
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
293
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
294
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
295
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
296
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
297
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);
298

    
299
            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
300
                    vec_cmpgt(row0, zero));
301
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
302
                    vec_cmpgt(row1, zero));
303
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
304
                    vec_cmpgt(row2, zero));
305
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
306
                    vec_cmpgt(row3, zero));
307
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
308
                    vec_cmpgt(row4, zero));
309
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
310
                    vec_cmpgt(row5, zero));
311
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
312
                    vec_cmpgt(row6, zero));
313
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
314
                    vec_cmpgt(row7, zero));
315

    
316
            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
317
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
318
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
319
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
320
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
321
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
322
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
323
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);
324

    
325
            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
326
                    vec_cmpgt(alt0, zero));
327
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
328
                    vec_cmpgt(alt1, zero));
329
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
330
                    vec_cmpgt(alt2, zero));
331
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
332
                    vec_cmpgt(alt3, zero));
333
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
334
                    vec_cmpgt(alt4, zero));
335
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
336
                    vec_cmpgt(alt5, zero));
337
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
338
                    vec_cmpgt(alt6, zero));
339
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
340
                    vec_cmpgt(alt7, zero));
341
        }
342

    
343

    
344
    }
345

    
346
    // Store the data back into the original block
347
    {
348
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
349

    
350
        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
351
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
352
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
353
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
354
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
355
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
356
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
357
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
358

    
359
        {
360
            // Clamp for overflow
361
            vector signed int max_q_int, min_q_int;
362
            vector signed short max_q, min_q;
363

    
364
            LOAD4(max_q_int, &(s->max_qcoeff));
365
            LOAD4(min_q_int, &(s->min_qcoeff));
366

    
367
            max_q = vec_pack(max_q_int, max_q_int);
368
            min_q = vec_pack(min_q_int, min_q_int);
369

    
370
            data0 = vec_max(vec_min(data0, max_q), min_q);
371
            data1 = vec_max(vec_min(data1, max_q), min_q);
372
            data2 = vec_max(vec_min(data2, max_q), min_q);
373
            data4 = vec_max(vec_min(data4, max_q), min_q);
374
            data5 = vec_max(vec_min(data5, max_q), min_q);
375
            data6 = vec_max(vec_min(data6, max_q), min_q);
376
            data7 = vec_max(vec_min(data7, max_q), min_q);
377
        }
378

    
379
        {
380
        vector bool char zero_01, zero_23, zero_45, zero_67;
381
        vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67;
382
        vector signed char negOne = vec_splat_s8(-1);
383
        vector signed char* scanPtr =
384
                (vector signed char*)(s->intra_scantable.inverse);
385
        signed char lastNonZeroChar;
386

    
387
        // Determine the largest non-zero index.
388
        zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
389
                vec_cmpeq(data1, (vector signed short)zero));
390
        zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
391
                vec_cmpeq(data3, (vector signed short)zero));
392
        zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
393
                vec_cmpeq(data5, (vector signed short)zero));
394
        zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
395
                vec_cmpeq(data7, (vector signed short)zero));
396

    
397
        // 64 biggest values
398
        scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01);
399
        scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23);
400
        scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45);
401
        scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67);
402

    
403
        // 32 largest values
404
        scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23);
405
        scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67);
406

    
407
        // 16 largest values
408
        scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45);
409

    
410
        // 8 largest values
411
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
412
                vec_mergel(scanIndexes_01, negOne));
413

    
414
        // 4 largest values
415
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
416
                vec_mergel(scanIndexes_01, negOne));
417

    
418
        // 2 largest values
419
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
420
                vec_mergel(scanIndexes_01, negOne));
421

    
422
        // largest value
423
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
424
                vec_mergel(scanIndexes_01, negOne));
425

    
426
        scanIndexes_01 = vec_splat(scanIndexes_01, 0);
427

    
428

    
429
        vec_ste(scanIndexes_01, 0, &lastNonZeroChar);
430

    
431
        lastNonZero = lastNonZeroChar;
432

    
433
        // While the data is still in vectors we check for the transpose IDCT permute
434
        // and handle it using the vector unit if we can.  This is the permute used
435
        // by the altivec idct, so it is common when using the altivec dct.
436

    
437
        if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) {
438
            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
439
        }
440

    
441
        vec_st(data0, 0, data);
442
        vec_st(data1, 16, data);
443
        vec_st(data2, 32, data);
444
        vec_st(data3, 48, data);
445
        vec_st(data4, 64, data);
446
        vec_st(data5, 80, data);
447
        vec_st(data6, 96, data);
448
        vec_st(data7, 112, data);
449
        }
450
    }
451

    
452
    // special handling of block[0]
453
    if (s->mb_intra) {
454
        if (!s->h263_aic) {
455
            if (n < 4)
456
                oldBaseValue /= s->y_dc_scale;
457
            else
458
                oldBaseValue /= s->c_dc_scale;
459
        }
460

    
461
        // Divide by 8, rounding the result
462
        data[0] = (oldBaseValue + 4) >> 3;
463
    }
464

    
465
    // We handled the transpose permutation above and we don't
466
    // need to permute the "no" permutation case.
467
    if ((lastNonZero > 0) &&
468
        (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
469
        (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) {
470
        ff_block_permute(data, s->dsp.idct_permutation,
471
                s->intra_scantable.scantable, lastNonZero);
472
    }
473

    
474
    return lastNonZero;
475
}
476

    
477
/* AltiVec version of dct_unquantize_h263
478
   this code assumes `block' is 16 bytes-aligned */
479
static void dct_unquantize_h263_altivec(MpegEncContext *s,
480
                                 DCTELEM *block, int n, int qscale)
481
{
482
    int i, level, qmul, qadd;
483
    int nCoeffs;
484

    
485
    assert(s->block_last_index[n]>=0);
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) = qmul;
508
        DECLARE_ALIGNED(16, short, qadd8) = qadd;
509
        register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
510
        register vector bool short blockv_null, blockv_neg;
511
        register short backup_0 = block[0];
512
        register int j = 0;
513

    
514
        qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
515
        qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
516
        nqaddv = vec_sub(vczero, qaddv);
517

    
518
#if 0   // block *is* 16 bytes-aligned, it seems.
519
        // first make sure block[j] is 16 bytes-aligned
520
        for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
521
            level = block[j];
522
            if (level) {
523
                if (level < 0) {
524
                    level = level * qmul - qadd;
525
                } else {
526
                    level = level * qmul + qadd;
527
                }
528
                block[j] = level;
529
            }
530
        }
531
#endif
532

    
533
        // vectorize all the 16 bytes-aligned blocks
534
        // of 8 elements
535
        for(; (j + 7) <= nCoeffs ; j+=8) {
536
            blockv = vec_ld(j << 1, block);
537
            blockv_neg = vec_cmplt(blockv, vczero);
538
            blockv_null = vec_cmpeq(blockv, vczero);
539
            // choose between +qadd or -qadd as the third operand
540
            temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
541
            // multiply & add (block{i,i+7} * qmul [+-] qadd)
542
            temp1 = vec_mladd(blockv, qmulv, temp1);
543
            // put 0 where block[{i,i+7} used to have 0
544
            blockv = vec_sel(temp1, blockv, blockv_null);
545
            vec_st(blockv, j << 1, block);
546
        }
547

    
548
        // if nCoeffs isn't a multiple of 8, finish the job
549
        // using good old scalar units.
550
        // (we could do it using a truncated vector,
551
        // but I'm not sure it's worth the hassle)
552
        for(; j <= nCoeffs ; 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

    
564
        if (i == 1) {
565
            // cheat. this avoid special-casing the first iteration
566
            block[0] = backup_0;
567
        }
568
    }
569
}
570

    
571

    
572
void MPV_common_init_altivec(MpegEncContext *s)
573
{
574
    if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)) return;
575

    
576
    if (s->avctx->lowres==0) {
577
        if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
578
            (s->avctx->idct_algo == FF_IDCT_ALTIVEC)) {
579
            s->dsp.idct_put = idct_put_altivec;
580
            s->dsp.idct_add = idct_add_altivec;
581
            s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
582
        }
583
    }
584

    
585
    // Test to make sure that the dct required alignments are met.
586
    if ((((long)(s->q_intra_matrix) & 0x0f) != 0) ||
587
        (((long)(s->q_inter_matrix) & 0x0f) != 0)) {
588
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned "
589
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
590
        return;
591
    }
592

    
593
    if (((long)(s->intra_scantable.inverse) & 0x0f) != 0) {
594
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned "
595
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
596
        return;
597
    }
598

    
599

    
600
    if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
601
            (s->avctx->dct_algo == FF_DCT_ALTIVEC)) {
602
#if 0 /* seems to cause trouble under some circumstances */
603
        s->dct_quantize = dct_quantize_altivec;
604
#endif
605
        s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
606
        s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
607
    }
608
}