mpegvideo_altivec.c 24.7 KB
Newer Older
1 2 3
/*
 * Copyright (c) 2002 Dieter Shirley
 *
4 5 6
 * dct_unquantize_h263_altivec:
 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
 *
7 8 9
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
10 11
 * modify it under the terms of the GNU Lesser General Public
 * 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 16 17 18 19
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * 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 25
#include <stdlib.h>
#include <stdio.h>
26 27
#include "libavcodec/dsputil.h"
#include "libavcodec/mpegvideo.h"
28

29 30
#include "dsputil_ppc.h"
#include "util_altivec.h"
31
#include "types_altivec.h"
32
#include "dsputil_altivec.h"
33

34 35 36 37 38 39 40 41 42 43 44
// Swaps two variables (used for altivec registers)
#define SWAP(a,b) \
do { \
    __typeof__(a) swap_temp=a; \
    a=b; \
    b=swap_temp; \
} while (0)

// transposes a matrix consisting of four vectors with four elements each
#define TRANSPOSE4(a,b,c,d) \
do { \
45 46 47 48 49 50 51 52 53
    __typeof__(a) _trans_ach = vec_mergeh(a, c); \
    __typeof__(a) _trans_acl = vec_mergel(a, c); \
    __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
    __typeof__(a) _trans_bdl = vec_mergel(b, d); \
                                                 \
    a = vec_mergeh(_trans_ach, _trans_bdh);      \
    b = vec_mergel(_trans_ach, _trans_bdh);      \
    c = vec_mergeh(_trans_acl, _trans_bdl);      \
    d = vec_mergel(_trans_acl, _trans_bdl);      \
54 55 56 57 58 59 60 61
} while (0)


// Loads a four-byte value (int or float) from the target address
// into every element in the target vector.  Only works if the
// target address is four-byte aligned (which should be always).
#define LOAD4(vec, address) \
{ \
62 63 64 65 66
    __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address);  \
    vector unsigned char _perm_vec = vec_lvsl(0,(address));     \
    vec = vec_ld(0, _load_addr);                                \
    vec = vec_perm(vec, vec, _perm_vec);                        \
    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 74
                         DCTELEM* data, int n,
                         int qscale, int* overflow)
75 76 77 78
{
    int lastNonZero;
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
79
    const vector float zero = (const vector float)FOUROF(0.);
Vitor Sessak's avatar
Vitor Sessak committed
80
    // used after quantize step
81
    int oldBaseValue = 0;
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120

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

        data0 = vec_ld(0, data);
        data1 = vec_ld(16, data);
        data2 = vec_ld(32, data);
        data3 = vec_ld(48, data);
        data4 = vec_ld(64, data);
        data5 = vec_ld(80, data);
        data6 = vec_ld(96, data);
        data7 = vec_ld(112, data);

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

        // load the data into floating point vectors.  We load
        // the high half of each row into the main row vectors
        // and the low half into the alt vectors.
        row0 = vec_ctf(vec_unpackh(data0), 0);
        alt0 = vec_ctf(vec_unpackl(data0), 0);
        row1 = vec_ctf(vec_unpackh(data1), 0);
        alt1 = vec_ctf(vec_unpackl(data1), 0);
        row2 = vec_ctf(vec_unpackh(data2), 0);
        alt2 = vec_ctf(vec_unpackl(data2), 0);
        row3 = vec_ctf(vec_unpackh(data3), 0);
        alt3 = vec_ctf(vec_unpackl(data3), 0);
        row4 = vec_ctf(vec_unpackh(data4), 0);
        alt4 = vec_ctf(vec_unpackl(data4), 0);
        row5 = vec_ctf(vec_unpackh(data5), 0);
        alt5 = vec_ctf(vec_unpackl(data5), 0);
        row6 = vec_ctf(vec_unpackh(data6), 0);
        alt6 = vec_ctf(vec_unpackl(data6), 0);
        row7 = vec_ctf(vec_unpackh(data7), 0);
        alt7 = vec_ctf(vec_unpackl(data7), 0);
    }

    // The following block could exist as a separate an altivec dct
121 122 123
                // function.  However, if we put it inline, the DCT data can remain
                // in the vector local variables, as floats, which we'll use during the
                // quantize step...
124
    {
125 126 127 128 129 130 131 132 133 134 135 136
        const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
        const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
        const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
        const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
        const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
        const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
        const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
        const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
        const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
        const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
        const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
        const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
137 138 139 140


        int whichPass, whichHalf;

141 142
        for(whichPass = 1; whichPass<=2; whichPass++) {
            for(whichHalf = 1; whichHalf<=2; whichHalf++) {
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172
                vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
                vector float tmp10, tmp11, tmp12, tmp13;
                vector float z1, z2, z3, z4, z5;

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

                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;


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

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


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

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

                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
177
                //                                CONST_BITS-PASS1_BITS);
178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198
                row6 = vec_madd(tmp12, vec_1_847759065, z1);

                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;

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

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

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

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

                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
Diego Biurrun's avatar
Diego Biurrun committed
199
                // Wow!  It's actually more efficient to roll this multiply
200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236
                // into the adds below, even thought the multiply gets done twice!
                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);

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

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

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

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

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

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

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

240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
                // First, block transpose the upper right with lower left.
                SWAP(row4, alt0);
                SWAP(row5, alt1);
                SWAP(row6, alt2);
                SWAP(row7, alt3);

                // Now, transpose each block of four
                TRANSPOSE4(row0, row1, row2, row3);
                TRANSPOSE4(row4, row5, row6, row7);
                TRANSPOSE4(alt0, alt1, alt2, alt3);
                TRANSPOSE4(alt4, alt5, alt6, alt7);
            }
        }
    }

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

262
        if (s->mb_intra) {
263 264 265 266 267 268 269 270 271
            vector signed int baseVector;

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

            qmat = (vector signed int*)s->q_intra_matrix[qscale];
            biasAddr = &(s->intra_quant_bias);
272
        } else {
273 274 275 276 277
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
            biasAddr = &(s->inter_quant_bias);
        }

        // 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);
282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342
            LOAD4(biasInt, biasAddr);
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
            negBias = vec_madd(bias, negOneFloat, zero);
        }

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

            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);

            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
                    vec_cmpgt(row0, zero));
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
                    vec_cmpgt(row1, zero));
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
                    vec_cmpgt(row2, zero));
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
                    vec_cmpgt(row3, zero));
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
                    vec_cmpgt(row4, zero));
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
                    vec_cmpgt(row5, zero));
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
                    vec_cmpgt(row6, zero));
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
                    vec_cmpgt(row7, zero));

            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);

            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
                    vec_cmpgt(alt0, zero));
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
                    vec_cmpgt(alt1, zero));
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
                    vec_cmpgt(alt2, zero));
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
                    vec_cmpgt(alt3, zero));
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
                    vec_cmpgt(alt4, zero));
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
                    vec_cmpgt(alt5, zero));
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
                    vec_cmpgt(alt6, zero));
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
                    vec_cmpgt(alt7, zero));
        }

343

344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378
    }

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

        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));

        {
            // Clamp for overflow
            vector signed int max_q_int, min_q_int;
            vector signed short max_q, min_q;

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

            max_q = vec_pack(max_q_int, max_q_int);
            min_q = vec_pack(min_q_int, min_q_int);

            data0 = vec_max(vec_min(data0, max_q), min_q);
            data1 = vec_max(vec_min(data1, max_q), min_q);
            data2 = vec_max(vec_min(data2, max_q), min_q);
            data4 = vec_max(vec_min(data4, max_q), min_q);
            data5 = vec_max(vec_min(data5, max_q), min_q);
            data6 = vec_max(vec_min(data6, max_q), min_q);
            data7 = vec_max(vec_min(data7, max_q), min_q);
        }

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;
382 383 384
        vector signed char negOne = vec_splat_s8(-1);
        vector signed char* scanPtr =
                (vector signed char*)(s->intra_scantable.inverse);
385
        signed char lastNonZeroChar;
386 387

        // Determine the largest non-zero index.
388 389 390 391 392 393 394 395
        zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
                vec_cmpeq(data1, (vector signed short)zero));
        zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
                vec_cmpeq(data3, (vector signed short)zero));
        zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
                vec_cmpeq(data5, (vector signed short)zero));
        zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
                vec_cmpeq(data7, (vector signed short)zero));
396 397

        // 64 biggest values
398 399 400 401
        scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01);
        scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23);
        scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45);
        scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67);
402 403

        // 32 largest values
404 405
        scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23);
        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
411 412
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
                vec_mergel(scanIndexes_01, negOne));
413 414

        // 4 largest values
415 416
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
                vec_mergel(scanIndexes_01, negOne));
417 418

        // 2 largest values
419 420
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
                vec_mergel(scanIndexes_01, negOne));
421 422

        // largest value
423 424
        scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
                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

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

437
        if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) {
438 439 440 441 442 443 444 445 446 447 448
            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
        }

        vec_st(data0, 0, data);
        vec_st(data1, 16, data);
        vec_st(data2, 32, data);
        vec_st(data3, 48, data);
        vec_st(data4, 64, data);
        vec_st(data5, 80, data);
        vec_st(data6, 96, data);
        vec_st(data7, 112, data);
449
        }
450 451 452
    }

    // special handling of block[0]
453 454
    if (s->mb_intra) {
        if (!s->h263_aic) {
455 456 457 458 459 460 461 462 463 464
            if (n < 4)
                oldBaseValue /= s->y_dc_scale;
            else
                oldBaseValue /= s->c_dc_scale;
        }

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

Vitor Sessak's avatar
Vitor Sessak committed
465
    // We handled the transpose permutation above and we don't
466 467
    // need to permute the "no" permutation case.
    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,
471 472 473 474 475 476
                s->intra_scantable.scantable, lastNonZero);
    }

    return lastNonZero;
}

477 478
/* AltiVec version of dct_unquantize_h263
   this code assumes `block' is 16 bytes-aligned */
479
static void dct_unquantize_h263_altivec(MpegEncContext *s,
480 481
                                 DCTELEM *block, int n, int qscale)
{
482
POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
483 484
    int i, level, qmul, qadd;
    int nCoeffs;
485

486
    assert(s->block_last_index[n]>=0);
487

488
POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);
489

490 491
    qadd = (qscale - 1) | 1;
    qmul = qscale << 1;
492

493 494
    if (s->mb_intra) {
        if (!s->h263_aic) {
495
            if (n < 4)
496 497 498 499 500 501
                block[0] = block[0] * s->y_dc_scale;
            else
                block[0] = block[0] * s->c_dc_scale;
        }else
            qadd = 0;
        i = 1;
Diego Biurrun's avatar
Diego Biurrun committed
502
        nCoeffs= 63; //does not always use zigzag table
503 504 505 506 507 508
    } else {
        i = 0;
        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
    }

    {
509
        register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
510 511
        DECLARE_ALIGNED(16, short, qmul8) = qmul;
        DECLARE_ALIGNED(16, short, qadd8) = qadd;
512 513 514 515 516
        register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
        register vector bool short blockv_null, blockv_neg;
        register short backup_0 = block[0];
        register int j = 0;

517 518 519
        qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
        qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
        nqaddv = vec_sub(vczero, qaddv);
520 521 522 523 524 525 526 527 528 529 530 531

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

536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563
        // vectorize all the 16 bytes-aligned blocks
        // of 8 elements
        for(; (j + 7) <= nCoeffs ; j+=8) {
            blockv = vec_ld(j << 1, block);
            blockv_neg = vec_cmplt(blockv, vczero);
            blockv_null = vec_cmpeq(blockv, vczero);
            // choose between +qadd or -qadd as the third operand
            temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
            // multiply & add (block{i,i+7} * qmul [+-] qadd)
            temp1 = vec_mladd(blockv, qmulv, temp1);
            // put 0 where block[{i,i+7} used to have 0
            blockv = vec_sel(temp1, blockv, blockv_null);
            vec_st(blockv, j << 1, block);
        }

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

567 568 569 570
        if (i == 1) {
            // cheat. this avoid special-casing the first iteration
            block[0] = backup_0;
        }
571
    }
572
POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
573
}
574 575


576
void MPV_common_init_altivec(MpegEncContext *s)
577
{
578
    if ((mm_flags & FF_MM_ALTIVEC) == 0) return;
579

580
    if (s->avctx->lowres==0) {
581
        if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
582
            (s->avctx->idct_algo == FF_IDCT_ALTIVEC)) {
583 584 585 586 587 588 589 590
            s->dsp.idct_put = idct_put_altivec;
            s->dsp.idct_add = idct_add_altivec;
            s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
        }
    }

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

597
    if (((long)(s->intra_scantable.inverse) & 0x0f) != 0) {
598 599 600 601 602 603 604
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned "
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
        return;
    }


    if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
605
            (s->avctx->dct_algo == FF_DCT_ALTIVEC)) {
606 607 608 609 610 611 612
#if 0 /* seems to cause trouble under some circumstances */
        s->dct_quantize = dct_quantize_altivec;
#endif
        s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
        s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
    }
}