mpegvideo_altivec.c

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

/*
 * Copyright (c) 2002 Dieter Shirley
 *
 * dct_unquantize_h263_altivec:
 * Copyright (c) 2003 Romain Dolbeau <[e-mail]>
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * 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
 * License along with this library; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 */

#include <stdlib.h>
#include <stdio.h>
#include "../dsputil.h"
#include "../mpegvideo.h"

#include "gcc_fixes.h"
 
#include "dsputil_altivec.h"

// 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 { \
  __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); \
} while (0)

#define TRANSPOSE8(a,b,c,d,e,f,g,h) \
do { \
    __typeof__(a)  _A1, _B1, _C1, _D1, _E1, _F1, _G1, _H1; \
    __typeof__(a)  _A2, _B2, _C2, _D2, _E2, _F2, _G2, _H2; \
 \
    _A1 = vec_mergeh (a, e); \
    _B1 = vec_mergel (a, e); \
    _C1 = vec_mergeh (b, f); \
    _D1 = vec_mergel (b, f); \
    _E1 = vec_mergeh (c, g); \
    _F1 = vec_mergel (c, g); \
    _G1 = vec_mergeh (d, h); \
    _H1 = vec_mergel (d, h); \
 \
    _A2 = vec_mergeh (_A1, _E1); \
    _B2 = vec_mergel (_A1, _E1); \
    _C2 = vec_mergeh (_B1, _F1); \
    _D2 = vec_mergel (_B1, _F1); \
    _E2 = vec_mergeh (_C1, _G1); \
    _F2 = vec_mergel (_C1, _G1); \
    _G2 = vec_mergeh (_D1, _H1); \
    _H2 = vec_mergel (_D1, _H1); \
 \
    a = vec_mergeh (_A2, _E2); \
    b = vec_mergel (_A2, _E2); \
    c = vec_mergeh (_B2, _F2); \
    d = vec_mergel (_B2, _F2); \
    e = vec_mergeh (_C2, _G2); \
    f = vec_mergel (_C2, _G2); \
    g = vec_mergeh (_D2, _H2); \
    h = vec_mergel (_D2, _H2); \
} 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) \
{ \
    __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); \
}


#ifdef CONFIG_DARWIN
#define FOUROF(a) (a)
#else
// slower, for dumb non-apple GCC
#define FOUROF(a) {a,a,a,a}
#endif
int dct_quantize_altivec(MpegEncContext* s, 
                        DCTELEM* data, int n,
                        int qscale, int* overflow)
{
    int lastNonZero;
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
    const_vector float zero = (const_vector float)FOUROF(0.);
    // used after quantise step
    int oldBaseValue = 0;

    // 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
		// 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...
    {
        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);


        int whichPass, whichHalf;

        for(whichPass = 1; whichPass<=2; whichPass++)
        {
            for(whichHalf = 1; whichHalf<=2; whichHalf++)
            {
                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),
                //		   CONST_BITS-PASS1_BITS);
                row2 = vec_madd(tmp13, vec_0_765366865, z1);

                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
                //		   CONST_BITS-PASS1_BITS);
                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) */
                // Wow!  It's actually more effecient to roll this multiply
                // 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);
            }

            if (whichPass == 1)
            {
                // transpose the data for the second pass
                 
                // 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);
            }
        }
    }

    // perform the quantise step, using the floating point data
    // still in the row/alt registers
    {
        const int* biasAddr;
        const vector signed int* qmat;
        vector float bias, negBias;

        if (s->mb_intra)
        {
            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);
        }
        else
        {
            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
				// rounding when we convert to int, instead of flooring.)
        {
            vector signed int biasInt;
            const vector float negOneFloat = (vector float)FOUROF(-1.0f);
            LOAD4(biasInt, biasAddr);
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
            negBias = vec_madd(bias, negOneFloat, zero);
        }