mbior97.cpp

介绍小波的函数

                        }

                        if (x < w / 2)
                                dest[k][x] = mmxround(s);
                        else
                                dest[k][x] = mmxroundTH(s); 	        //is this needed? hi band were TH'ed on transrows

                        dest[k+h2][x] = mmxroundTH(d);          //is this needed? hi band were TH'ed on transrows on x>w/2
                }
        }
}
///////////////////////////////////////////////transforms/////////////////////////////////////////////////////////////////////





//////////////////////////////////////////////synths//////////////////////////////////////////////////////////////////////////
void mBior97::synthrows(char** dest, char** sour, unsigned int w, unsigned int h) const      //w,h of the LO part
{
        char srclo[8] = {0, 0, 0, 0, 0, 0, 0, 0};
        char srchi[8] = {0, 0, 0, 0, 0, 0, 0, 0};

        //int n;
        float s2k, s2k1;

        const vec1D& H2m = getH2m();
        const vec1D& G2m = getG2m();
        const vec1D& H2m1 = getH2m1();
        const vec1D& G2m1 = getG2m1();

        for (unsigned int y = 0; y < 2*h; y++) {

                //k = [0; 2)
                for (unsigned int k = 0; k < 2; k++) {

                        if (k == 0) {
                                srclo[0] = sour[y][2];
                                srclo[1] = sour[y][1];
                                srclo[2] = sour[y][0];
                                srclo[3] = sour[y][1];
                                srclo[4] = sour[y][2];
                                srchi[0] = sour[y][2+w];
                                srchi[1] = sour[y][1+w];
                                srchi[2] = sour[y][0+w];
                                srchi[3] = sour[y][1+w];
                                srchi[4] = sour[y][2+w];
                        } else if (k == 1) {
                                srclo[0] = sour[y][1];
                                srclo[1] = sour[y][0];
                                srclo[2] = sour[y][1];
                                srclo[3] = sour[y][2];
                                srclo[4] = sour[y][3];
                                srchi[0] = sour[y][1+w];
                                srchi[1] = sour[y][0+w];
                                srchi[2] = sour[y][1+w];
                                srchi[3] = sour[y][2+w];
                                srchi[4] = sour[y][3+w];
                        }

                        s2k = convr(H2m.data(), srclo);                //s2k even H
                        s2k += convr(G2m.data(), srchi);

                        s2k1 = convr(H2m1.data(), srclo + 1);          //s2k1 odd H
                        s2k1 += H2m1[H2m1.last()] * float(srclo[0]);
                        s2k1 += convr(G2m1.data(), srchi + 1);         //s2k1 odd G
                        s2k1 += G2m1[G2m1.last()] * float(srchi[0]);

                        dest[y][2*k] = mmxround(2.0f * s2k);
                        dest[y][2*k+1] = mmxround(2.0f * s2k1);
                }


                //k = [2; w-2)
                for (unsigned int k = 2; k < w - 2; k++) {

                        s2k = convr(H2m.data(), &sour[y][k-2]);         //s2k1 odd H
                        s2k += convr(G2m.data(), &sour[y][k-2+w]);      //s2k even G

                        s2k1 = convr(H2m1.data(), &sour[y][k-1]);         //s2k1 odd H
                        s2k1 += H2m1[H2m1.last()] * float(sour[y][k-2]);
                        s2k1 += convr(G2m1.data(), &sour[y][k-1+w]);      //s2k1 odd G
                        s2k1 += G2m1[G2m1.last()] * float(sour[y][k-2+w]);

                        dest[y][2*k] = mmxround(2.0f * s2k);
                        dest[y][2*k+1] = mmxround(2.0f * s2k1);
                }


                //k = [w-2; w)
                for (unsigned int k = w - 2; k < w; k++) {

                        if (k == w - 2) {
                                srclo[0] = sour[y][k-2];
                                srclo[1] = sour[y][k-1];
                                srclo[2] = sour[y][k];
                                srclo[3] = sour[y][k+1];
                                srclo[4] = sour[y][k];
                                srchi[0] = sour[y][k-2+w];
                                srchi[1] = sour[y][k-1+w];
                                srchi[2] = sour[y][k+w];
                                srchi[3] = sour[y][k+1+w];
                                srchi[4] = sour[y][k+w];
                        } else if (k == w - 1) {
                                srclo[0] = sour[y][k-2];
                                srclo[1] = sour[y][k-1];
                                srclo[2] = sour[y][k];
                                srclo[3] = sour[y][k-1];
                                srclo[4] = sour[y][k-2];
                                srchi[0] = sour[y][k-2+w];
                                srchi[1] = sour[y][k-1+w];
                                srchi[2] = sour[y][k+w];
                                srchi[3] = sour[y][k-1+w];
                                srchi[4] = sour[y][k-2+w];
                        }

                        s2k = convr(H2m.data(), srclo);            //s2k even H
                        s2k += convr(G2m.data(), srchi);

                        s2k1 = convr(H2m1.data(), srclo + 1);          //s2k1 odd H
                        s2k1 += H2m1[H2m1.last()] * float(srclo[0]);
                        s2k1 += convr(G2m1.data(), srchi + 1);         //s2k1 odd G
                        s2k1 += G2m1[G2m1.last()] * float(srchi[0]);

                        dest[y][2*k] = mmxround(2.0f * s2k);
                        dest[y][2*k+1] = mmxround(2.0f * s2k1);
                }
        }
}

void mBior97::synthcols(char** dest, char** sour, unsigned int w, unsigned int h) const     //w,h of the LO part
{
        float fz = 0.0f;
        float mul2 = 2.0f;

        int n;
        float s2k, s2k1;
        __m128 ms2k, ms2k1;
        unsigned int w2 = 2 * w;

        const vec1D& H2m = getH2m();
        const vec1D& G2m = getG2m();
        const vec1D& H2m1 = getH2m1();
        const vec1D& G2m1 = getG2m1();

        for (unsigned int x = 0; x < w2 / 4; x++) {      //x<w2/2   x = 4*x
                for (unsigned int k = 0; k < h; k++) {

                        ms2k = _mm_load_ss(&fz);
                        ms2k1 = ms2k;

                        for (int m = -1; m <= 1; m++) {                             //s2k even H
                                n = k - m;
                                if (n < 0) n = 0 - n;
                                if (n >= (int)h) n -= 2 * (1 + n - h);
                                ms2k = _mm_add_ps(ms2k, _mm_mul_ps(_mm_load_ps1(H2m.data(m)), _mm_cvtpi8_ps(*(__m64 *)(&sour[n][4*x]))));
                        }
                        for (int m = -1; m <= 2; m++) {                             //s2k even G
                                n = k - m;
                                if (n < 0) n = 0 - n;
                                if (n >= (int)h) n -= 2 * (1 + n - h);
                                ms2k = _mm_add_ps(ms2k, _mm_mul_ps(_mm_load_ps1(G2m.data(m)), _mm_cvtpi8_ps(*(__m64 *)(&sour[n+h][4*x]))));
                        }

                        for (int m = -2; m <= 1; m++) {                            //s2k1 odd H
                                n = k - m;
                                if (n < 0) n = 0 - n;
                                if (n >= (int)h) n -= 2 * (1 + n - h);
                                ms2k1 = _mm_add_ps(ms2k1, _mm_mul_ps(_mm_load_ps1(H2m1.data(m)), _mm_cvtpi8_ps(*(__m64 *)(&sour[n][4*x]))));
                        }
                        for (int m = -2; m <= 2; m++) {                            //s2k1 odd G
                                n = k - m;
                                if (n < 0) n = 0 - n;
                                if (n >= (int)h) n -= 2 * (1 + n - h);
                                ms2k1 = _mm_add_ps(ms2k1, _mm_mul_ps(_mm_load_ps1(G2m1.data(m)), _mm_cvtpi8_ps(*(__m64 *)(&sour[n+h][4*x]))));
                        }

                        __m128 mmul2 = _mm_load_ps1(&mul2);

                        mmxround4(&dest[2*k][4*x], _mm_mul_ps(ms2k, mmul2));
                        mmxround4(&dest[2*k+1][4*x], _mm_mul_ps(ms2k1, mmul2));
                }
        }
        _mm_empty();


        //odd remainder
        for (unsigned int x = w2 - (w2 % 4); x < w2; x++) {
                for (unsigned int k = 0; k < h; k++) {

                        s2k = 0;
                        s2k1 = 0;

                        for (int m = -1; m <= 1; m++) {        //s2k even H
                                n = k - m;
                                if (n < 0) n = 0 - n;
                                if (n >= (int)h) n -= 2 * (1 + n - h);
                                s2k += H2m[m] * float(sour[n][x]);
                        }
                        for (int m = -1; m <= 2; m++) {        //s2k even G
                                n = k - m;
                                if (n < 0) n = 0 - n;
                                if (n >= (int)h) n -= 2 * (1 + n - h);
                                s2k += G2m[m] * float(sour[n+h][x]);
                        }

                        for (int m = -2; m <= 1; m++) {       //s2k1 odd H
                                n = k - m;
                                if (n < 0) n = 0 - n;
                                if (n >= (int)h) n -= 2 * (1 + n - h);
                                s2k1 += H2m1[m] * float(sour[n][x]);
                        }
                        for (int m = -2; m <= 2; m++) {      //s2k1 odd G
                                n = k - m;
                                if (n < 0) n = 0 - n;
                                if (n >= (int)h) n -= 2 * (1 + n - h);
                                s2k1 += G2m1[m] * float(sour[n+h][x]);
                        }

                        dest[2*k][x] = mmxround(2.0f * s2k);
                        dest[2*k+1][x] = mmxround(2.0f * s2k1);
                }
        }
}
//////////////////////////////////////////////synths//////////////////////////////////////////////////////////////////////////