Statistics
| Branch: | Revision:

ffmpeg / libavcodec / ppc / mpegvideo_altivec.c @ 84dc2d8a

History | View | Annotate | Download (24.8 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
#include "types_altivec.h"
32

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

    
146
                tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
147
                tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
148
                tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
149
                tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
150
                tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
151
                tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
152
                tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
153
                tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
154

    
155
                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
156
                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
157
                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
158
                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
159

    
160

    
161
                // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
162
                row0 = vec_add(tmp10, tmp11);
163

    
164
                // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
165
                row4 = vec_sub(tmp10, tmp11);
166

    
167

    
168
                // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
169
                z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
170

    
171
                // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
172
                //                                CONST_BITS-PASS1_BITS);
173
                row2 = vec_madd(tmp13, vec_0_765366865, z1);
174

    
175
                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
176
                //                                CONST_BITS-PASS1_BITS);
177
                row6 = vec_madd(tmp12, vec_1_847759065, z1);
178

    
179
                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
180
                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
181
                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
182
                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
183

    
184
                // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
185
                z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
186

    
187
                // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
188
                z3 = vec_madd(z3, vec_1_961570560, z5);
189

    
190
                // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
191
                z4 = vec_madd(z4, vec_0_390180644, z5);
192

    
193
                // The following adds are rolled into the multiplies above
194
                // z3 = vec_add(z3, z5);  // z3 += z5;
195
                // z4 = vec_add(z4, z5);  // z4 += z5;
196

    
197
                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
198
                // Wow!  It's actually more efficient to roll this multiply
199
                // into the adds below, even thought the multiply gets done twice!
200
                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
201

    
202
                // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
203
                // Same with this one...
204
                // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
205

    
206
                // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
207
                // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
208
                row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
209

    
210
                // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
211
                // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
212
                row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
213

    
214
                // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
215
                // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
216
                row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
217

    
218
                // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
219
                // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
220
                row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
221

    
222
                // Swap the row values with the alts.  If this is the first half,
223
                // this sets up the low values to be acted on in the second half.
224
                // If this is the second half, it puts the high values back in
225
                // the row values where they are expected to be when we're done.
226
                SWAP(row0, alt0);
227
                SWAP(row1, alt1);
228
                SWAP(row2, alt2);
229
                SWAP(row3, alt3);
230
                SWAP(row4, alt4);
231
                SWAP(row5, alt5);
232
                SWAP(row6, alt6);
233
                SWAP(row7, alt7);
234
            }
235

    
236
            if (whichPass == 1) {
237
                // transpose the data for the second pass
238

    
239
                // First, block transpose the upper right with lower left.
240
                SWAP(row4, alt0);
241
                SWAP(row5, alt1);
242
                SWAP(row6, alt2);
243
                SWAP(row7, alt3);
244

    
245
                // Now, transpose each block of four
246
                TRANSPOSE4(row0, row1, row2, row3);
247
                TRANSPOSE4(row4, row5, row6, row7);
248
                TRANSPOSE4(alt0, alt1, alt2, alt3);
249
                TRANSPOSE4(alt4, alt5, alt6, alt7);
250
            }
251
        }
252
    }
253

    
254
    // perform the quantize step, using the floating point data
255
    // still in the row/alt registers
256
    {
257
        const int* biasAddr;
258
        const vector signed int* qmat;
259
        vector float bias, negBias;
260

    
261
        if (s->mb_intra) {
262
            vector signed int baseVector;
263

    
264
            // We must cache element 0 in the intra case
265
            // (it needs special handling).
266
            baseVector = vec_cts(vec_splat(row0, 0), 0);
267
            vec_ste(baseVector, 0, &oldBaseValue);
268

    
269
            qmat = (vector signed int*)s->q_intra_matrix[qscale];
270
            biasAddr = &(s->intra_quant_bias);
271
        } else {
272
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
273
            biasAddr = &(s->inter_quant_bias);
274
        }
275

    
276
        // Load the bias vector (We add 0.5 to the bias so that we're
277
                                // rounding when we convert to int, instead of flooring.)
278
        {
279
            vector signed int biasInt;
280
            const vector float negOneFloat = (vector float)FOUROF(-1.0f);
281
            LOAD4(biasInt, biasAddr);
282
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
283
            negBias = vec_madd(bias, negOneFloat, zero);
284
        }
285

    
286
        {
287
            vector float q0, q1, q2, q3, q4, q5, q6, q7;
288

    
289
            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
290
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
291
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
292
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
293
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
294
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
295
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
296
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);
297

    
298
            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
299
                    vec_cmpgt(row0, zero));
300
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
301
                    vec_cmpgt(row1, zero));
302
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
303
                    vec_cmpgt(row2, zero));
304
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
305
                    vec_cmpgt(row3, zero));
306
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
307
                    vec_cmpgt(row4, zero));
308
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
309
                    vec_cmpgt(row5, zero));
310
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
311
                    vec_cmpgt(row6, zero));
312
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
313
                    vec_cmpgt(row7, zero));
314

    
315
            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
316
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
317
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
318
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
319
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
320
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
321
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
322
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);
323

    
324
            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
325
                    vec_cmpgt(alt0, zero));
326
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
327
                    vec_cmpgt(alt1, zero));
328
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
329
                    vec_cmpgt(alt2, zero));
330
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
331
                    vec_cmpgt(alt3, zero));
332
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
333
                    vec_cmpgt(alt4, zero));
334
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
335
                    vec_cmpgt(alt5, zero));
336
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
337
                    vec_cmpgt(alt6, zero));
338
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
339
                    vec_cmpgt(alt7, zero));
340
        }
341

    
342

    
343
    }
344

    
345
    // Store the data back into the original block
346
    {
347
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
348

    
349
        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
350
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
351
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
352
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
353
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
354
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
355
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
356
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
357

    
358
        {
359
            // Clamp for overflow
360
            vector signed int max_q_int, min_q_int;
361
            vector signed short max_q, min_q;
362

    
363
            LOAD4(max_q_int, &(s->max_qcoeff));
364
            LOAD4(min_q_int, &(s->min_qcoeff));
365

    
366
            max_q = vec_pack(max_q_int, max_q_int);
367
            min_q = vec_pack(min_q_int, min_q_int);
368

    
369
            data0 = vec_max(vec_min(data0, max_q), min_q);
370
            data1 = vec_max(vec_min(data1, max_q), min_q);
371
            data2 = vec_max(vec_min(data2, max_q), min_q);
372
            data4 = vec_max(vec_min(data4, max_q), min_q);
373
            data5 = vec_max(vec_min(data5, max_q), min_q);
374
            data6 = vec_max(vec_min(data6, max_q), min_q);
375
            data7 = vec_max(vec_min(data7, max_q), min_q);
376
        }
377

    
378
        {
379
        vector bool char zero_01, zero_23, zero_45, zero_67;
380
        vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67;
381
        vector signed char negOne = vec_splat_s8(-1);
382
        vector signed char* scanPtr =
383
                (vector signed char*)(s->intra_scantable.inverse);
384
        signed char lastNonZeroChar;
385

    
386
        // Determine the largest non-zero index.
387
        zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
388
                vec_cmpeq(data1, (vector signed short)zero));
389
        zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
390
                vec_cmpeq(data3, (vector signed short)zero));
391
        zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
392
                vec_cmpeq(data5, (vector signed short)zero));
393
        zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
394
                vec_cmpeq(data7, (vector signed short)zero));
395

    
396
        // 64 biggest values
397
        scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01);
398
        scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23);
399
        scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45);
400
        scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67);
401

    
402
        // 32 largest values
403
        scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23);
404
        scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67);
405

    
406
        // 16 largest values
407
        scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45);
408

    
409
        // 8 largest values
410
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
411
                vec_mergel(scanIndexes_01, negOne));
412

    
413
        // 4 largest values
414
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
415
                vec_mergel(scanIndexes_01, negOne));
416

    
417
        // 2 largest values
418
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
419
                vec_mergel(scanIndexes_01, negOne));
420

    
421
        // largest value
422
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
423
                vec_mergel(scanIndexes_01, negOne));
424

    
425
        scanIndexes_01 = vec_splat(scanIndexes_01, 0);
426

    
427

    
428
        vec_ste(scanIndexes_01, 0, &lastNonZeroChar);
429

    
430
        lastNonZero = lastNonZeroChar;
431

    
432
        // While the data is still in vectors we check for the transpose IDCT permute
433
        // and handle it using the vector unit if we can.  This is the permute used
434
        // by the altivec idct, so it is common when using the altivec dct.
435

    
436
        if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) {
437
            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
438
        }
439

    
440
        vec_st(data0, 0, data);
441
        vec_st(data1, 16, data);
442
        vec_st(data2, 32, data);
443
        vec_st(data3, 48, data);
444
        vec_st(data4, 64, data);
445
        vec_st(data5, 80, data);
446
        vec_st(data6, 96, data);
447
        vec_st(data7, 112, data);
448
        }
449
    }
450

    
451
    // special handling of block[0]
452
    if (s->mb_intra) {
453
        if (!s->h263_aic) {
454
            if (n < 4)
455
                oldBaseValue /= s->y_dc_scale;
456
            else
457
                oldBaseValue /= s->c_dc_scale;
458
        }
459

    
460
        // Divide by 8, rounding the result
461
        data[0] = (oldBaseValue + 4) >> 3;
462
    }
463

    
464
    // We handled the transpose permutation above and we don't
465
    // need to permute the "no" permutation case.
466
    if ((lastNonZero > 0) &&
467
        (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
468
        (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) {
469
        ff_block_permute(data, s->dsp.idct_permutation,
470
                s->intra_scantable.scantable, lastNonZero);
471
    }
472

    
473
    return lastNonZero;
474
}
475

    
476
/* AltiVec version of dct_unquantize_h263
477
   this code assumes `block' is 16 bytes-aligned */
478
void dct_unquantize_h263_altivec(MpegEncContext *s,
479
                                 DCTELEM *block, int n, int qscale)
480
{
481
POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
482
    int i, level, qmul, qadd;
483
    int nCoeffs;
484

    
485
    assert(s->block_last_index[n]>=0);
486

    
487
POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
488

    
489
    qadd = (qscale - 1) | 1;
490
    qmul = qscale << 1;
491

    
492
    if (s->mb_intra) {
493
        if (!s->h263_aic) {
494
            if (n < 4)
495
                block[0] = block[0] * s->y_dc_scale;
496
            else
497
                block[0] = block[0] * s->c_dc_scale;
498
        }else
499
            qadd = 0;
500
        i = 1;
501
        nCoeffs= 63; //does not always use zigzag table
502
    } else {
503
        i = 0;
504
        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
505
    }
506

    
507
    {
508
        register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
509
        DECLARE_ALIGNED(16, short, qmul8) = qmul;
510
        DECLARE_ALIGNED(16, short, qadd8) = qadd;
511
        register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
512
        register vector bool short blockv_null, blockv_neg;
513
        register short backup_0 = block[0];
514
        register int j = 0;
515

    
516
        qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
517
        qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
518
        nqaddv = vec_sub(vczero, qaddv);
519

    
520
#if 0   // block *is* 16 bytes-aligned, it seems.
521
        // first make sure block[j] is 16 bytes-aligned
522
        for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
523
            level = block[j];
524
            if (level) {
525
                if (level < 0) {
526
                    level = level * qmul - qadd;
527
                } else {
528
                    level = level * qmul + qadd;
529
                }
530
                block[j] = level;
531
            }
532
        }
533
#endif
534

    
535
        // vectorize all the 16 bytes-aligned blocks
536
        // of 8 elements
537
        for(; (j + 7) <= nCoeffs ; j+=8) {
538
            blockv = vec_ld(j << 1, block);
539
            blockv_neg = vec_cmplt(blockv, vczero);
540
            blockv_null = vec_cmpeq(blockv, vczero);
541
            // choose between +qadd or -qadd as the third operand
542
            temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
543
            // multiply & add (block{i,i+7} * qmul [+-] qadd)
544
            temp1 = vec_mladd(blockv, qmulv, temp1);
545
            // put 0 where block[{i,i+7} used to have 0
546
            blockv = vec_sel(temp1, blockv, blockv_null);
547
            vec_st(blockv, j << 1, block);
548
        }
549

    
550
        // if nCoeffs isn't a multiple of 8, finish the job
551
        // using good old scalar units.
552
        // (we could do it using a truncated vector,
553
        // but I'm not sure it's worth the hassle)
554
        for(; j <= nCoeffs ; j++) {
555
            level = block[j];
556
            if (level) {
557
                if (level < 0) {
558
                    level = level * qmul - qadd;
559
                } else {
560
                    level = level * qmul + qadd;
561
                }
562
                block[j] = level;
563
            }
564
        }
565

    
566
        if (i == 1) {
567
            // cheat. this avoid special-casing the first iteration
568
            block[0] = backup_0;
569
        }
570
    }
571
POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
572
}
573

    
574

    
575
void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block);
576
void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block);
577

    
578
void MPV_common_init_altivec(MpegEncContext *s)
579
{
580
    if ((mm_flags & FF_MM_ALTIVEC) == 0) return;
581

    
582
    if (s->avctx->lowres==0) {
583
        if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
584
            (s->avctx->idct_algo == FF_IDCT_ALTIVEC)) {
585
            s->dsp.idct_put = idct_put_altivec;
586
            s->dsp.idct_add = idct_add_altivec;
587
            s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
588
        }
589
    }
590

    
591
    // Test to make sure that the dct required alignments are met.
592
    if ((((long)(s->q_intra_matrix) & 0x0f) != 0) ||
593
        (((long)(s->q_inter_matrix) & 0x0f) != 0)) {
594
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned "
595
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
596
        return;
597
    }
598

    
599
    if (((long)(s->intra_scantable.inverse) & 0x0f) != 0) {
600
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned "
601
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
602
        return;
603
    }
604

    
605

    
606
    if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
607
            (s->avctx->dct_algo == FF_DCT_ALTIVEC)) {
608
#if 0 /* seems to cause trouble under some circumstances */
609
        s->dct_quantize = dct_quantize_altivec;
610
#endif
611
        s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
612
        s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
613
    }
614
}