comparison h264.c @ 4118:a1786732cd62 libavcodec

merge clip() into the alpha/beta/tc0 tables (10% faster filter_mb_fast() on P3)

4117:c647eff78d91

}


static void filter_mb_edgev( H264Context *h, uint8_t *pix, int stride, int16_t bS[4], int qp ) {
    int i, d;
    const int index_a = clip( qp + h->slice_alpha_c0_offset, 0, 51 );
    const int alpha = alpha_table[index_a];
    const int beta  = beta_table[clip( qp + h->slice_beta_offset, 0, 51 )];

    if( bS[0] < 4 ) {
        int8_t tc[4];
        for(i=0; i<4; i++)
            tc[i] = bS[i] ? tc0_table[index_a][bS[i] - 1] : -1;
        h->s.dsp.h264_h_loop_filter_luma(pix, stride, alpha, beta, tc);
    } else {
        /* 16px edge length, because bS=4 is triggered by being at
         * the edge of an intra MB, so all 4 bS are the same */
            for( d = 0; d < 16; d++ ) {

            }
    }
}
static void filter_mb_edgecv( H264Context *h, uint8_t *pix, int stride, int16_t bS[4], int qp ) {
    int i;
    const int index_a = clip( qp + h->slice_alpha_c0_offset, 0, 51 );
    const int alpha = alpha_table[index_a];
    const int beta  = beta_table[clip( qp + h->slice_beta_offset, 0, 51 )];

    if( bS[0] < 4 ) {
        int8_t tc[4];
        for(i=0; i<4; i++)
            tc[i] = bS[i] ? tc0_table[index_a][bS[i] - 1] + 1 : 0;
        h->s.dsp.h264_h_loop_filter_chroma(pix, stride, alpha, beta, tc);
    } else {
        h->s.dsp.h264_h_loop_filter_chroma_intra(pix, stride, alpha, beta);
    }
}

        if( bS[bS_index] == 0 ) {
            continue;
        }

        qp_index = MB_FIELD ? (i >> 3) : (i & 1);
        index_a = clip( qp[qp_index] + h->slice_alpha_c0_offset, 0, 51 );
        alpha = alpha_table[index_a];
        beta  = beta_table[clip( qp[qp_index] + h->slice_beta_offset, 0, 51 )];

        if( bS[bS_index] < 4 ) {
            const int tc0 = tc0_table[index_a][bS[bS_index] - 1];
            const int p0 = pix[-1];
            const int p1 = pix[-2];
            const int p2 = pix[-3];
            const int q0 = pix[0];
            const int q1 = pix[1];

        if( bS[bS_index] == 0 ) {
            continue;
        }

        qp_index = MB_FIELD ? (i >> 2) : (i & 1);
        index_a = clip( qp[qp_index] + h->slice_alpha_c0_offset, 0, 51 );
        alpha = alpha_table[index_a];
        beta  = beta_table[clip( qp[qp_index] + h->slice_beta_offset, 0, 51 )];

        if( bS[bS_index] < 4 ) {
            const int tc = tc0_table[index_a][bS[bS_index] - 1] + 1;
            const int p0 = pix[-1];
            const int p1 = pix[-2];
            const int q0 = pix[0];
            const int q1 = pix[1];

    }
}

static void filter_mb_edgeh( H264Context *h, uint8_t *pix, int stride, int16_t bS[4], int qp ) {
    int i, d;
    const int index_a = clip( qp + h->slice_alpha_c0_offset, 0, 51 );
    const int alpha = alpha_table[index_a];
    const int beta  = beta_table[clip( qp + h->slice_beta_offset, 0, 51 )];
    const int pix_next  = stride;

    if( bS[0] < 4 ) {
        int8_t tc[4];
        for(i=0; i<4; i++)
            tc[i] = bS[i] ? tc0_table[index_a][bS[i] - 1] : -1;
        h->s.dsp.h264_v_loop_filter_luma(pix, stride, alpha, beta, tc);
    } else {
        /* 16px edge length, see filter_mb_edgev */
            for( d = 0; d < 16; d++ ) {
                const int p0 = pix[-1*pix_next];

    }
}

static void filter_mb_edgech( H264Context *h, uint8_t *pix, int stride, int16_t bS[4], int qp ) {
    int i;
    const int index_a = clip( qp + h->slice_alpha_c0_offset, 0, 51 );
    const int alpha = alpha_table[index_a];
    const int beta  = beta_table[clip( qp + h->slice_beta_offset, 0, 51 )];

    if( bS[0] < 4 ) {
        int8_t tc[4];
        for(i=0; i<4; i++)
            tc[i] = bS[i] ? tc0_table[index_a][bS[i] - 1] + 1 : 0;
        h->s.dsp.h264_v_loop_filter_chroma(pix, stride, alpha, beta, tc);
    } else {
        h->s.dsp.h264_v_loop_filter_chroma_intra(pix, stride, alpha, beta);
    }
}

4118:a1786732cd62

}


static void filter_mb_edgev( H264Context *h, uint8_t *pix, int stride, int16_t bS[4], int qp ) {
    int i, d;
    const int index_a = qp + h->slice_alpha_c0_offset;
    const int alpha = (alpha_table+52)[index_a];
    const int beta  = (beta_table+52)[qp + h->slice_beta_offset];

    if( bS[0] < 4 ) {
        int8_t tc[4];
        for(i=0; i<4; i++)
            tc[i] = bS[i] ? (tc0_table+52)[index_a][bS[i] - 1] : -1;
        h->s.dsp.h264_h_loop_filter_luma(pix, stride, alpha, beta, tc);
    } else {
        /* 16px edge length, because bS=4 is triggered by being at
         * the edge of an intra MB, so all 4 bS are the same */
            for( d = 0; d < 16; d++ ) {

            }
    }
}
static void filter_mb_edgecv( H264Context *h, uint8_t *pix, int stride, int16_t bS[4], int qp ) {
    int i;
    const int index_a = qp + h->slice_alpha_c0_offset;
    const int alpha = (alpha_table+52)[index_a];
    const int beta  = (beta_table+52)[qp + h->slice_beta_offset];

    if( bS[0] < 4 ) {
        int8_t tc[4];
        for(i=0; i<4; i++)
            tc[i] = bS[i] ? (tc0_table+52)[index_a][bS[i] - 1] + 1 : 0;
        h->s.dsp.h264_h_loop_filter_chroma(pix, stride, alpha, beta, tc);
    } else {
        h->s.dsp.h264_h_loop_filter_chroma_intra(pix, stride, alpha, beta);
    }
}

        if( bS[bS_index] == 0 ) {
            continue;
        }

        qp_index = MB_FIELD ? (i >> 3) : (i & 1);
        index_a = qp[qp_index] + h->slice_alpha_c0_offset;
        alpha = (alpha_table+52)[index_a];
        beta  = (beta_table+52)[qp[qp_index] + h->slice_beta_offset];

        if( bS[bS_index] < 4 ) {
            const int tc0 = (tc0_table+52)[index_a][bS[bS_index] - 1];
            const int p0 = pix[-1];
            const int p1 = pix[-2];
            const int p2 = pix[-3];
            const int q0 = pix[0];
            const int q1 = pix[1];

        if( bS[bS_index] == 0 ) {
            continue;
        }

        qp_index = MB_FIELD ? (i >> 2) : (i & 1);
        index_a = qp[qp_index] + h->slice_alpha_c0_offset;
        alpha = (alpha_table+52)[index_a];
        beta  = (beta_table+52)[qp[qp_index] + h->slice_beta_offset];

        if( bS[bS_index] < 4 ) {
            const int tc = (tc0_table+52)[index_a][bS[bS_index] - 1] + 1;
            const int p0 = pix[-1];
            const int p1 = pix[-2];
            const int q0 = pix[0];
            const int q1 = pix[1];

    }
}

static void filter_mb_edgeh( H264Context *h, uint8_t *pix, int stride, int16_t bS[4], int qp ) {
    int i, d;
    const int index_a = qp + h->slice_alpha_c0_offset;
    const int alpha = (alpha_table+52)[index_a];
    const int beta  = (beta_table+52)[qp + h->slice_beta_offset];
    const int pix_next  = stride;

    if( bS[0] < 4 ) {
        int8_t tc[4];
        for(i=0; i<4; i++)
            tc[i] = bS[i] ? (tc0_table+52)[index_a][bS[i] - 1] : -1;
        h->s.dsp.h264_v_loop_filter_luma(pix, stride, alpha, beta, tc);
    } else {
        /* 16px edge length, see filter_mb_edgev */
            for( d = 0; d < 16; d++ ) {
                const int p0 = pix[-1*pix_next];

    }
}

static void filter_mb_edgech( H264Context *h, uint8_t *pix, int stride, int16_t bS[4], int qp ) {
    int i;
    const int index_a = qp + h->slice_alpha_c0_offset;
    const int alpha = (alpha_table+52)[index_a];
    const int beta  = (beta_table+52)[qp + h->slice_beta_offset];

    if( bS[0] < 4 ) {
        int8_t tc[4];
        for(i=0; i<4; i++)
            tc[i] = bS[i] ? (tc0_table+52)[index_a][bS[i] - 1] + 1 : 0;
        h->s.dsp.h264_v_loop_filter_chroma(pix, stride, alpha, beta, tc);
    } else {
        h->s.dsp.h264_v_loop_filter_chroma_intra(pix, stride, alpha, beta);
    }
}