comparison dsputil.c @ 10068:770294452eaf libavcodec

Get rid of large stack array in rd8x8_c()

10067:685af2860d80

static int rd8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
    MpegEncContext * const s= (MpegEncContext *)c;
    const uint8_t *scantable= s->intra_scantable.permutated;
    DECLARE_ALIGNED_8 (uint64_t, aligned_temp[sizeof(DCTELEM)*64/8]);
    DECLARE_ALIGNED_8 (uint64_t, aligned_bak[stride]);
    DCTELEM * const temp= (DCTELEM*)aligned_temp;
    uint8_t * const bak= (uint8_t*)aligned_bak;
    int i, last, run, bits, level, distortion, start_i;
    const int esc_length= s->ac_esc_length;
    uint8_t * length;
    uint8_t * last_length;

    assert(h==8);

    for(i=0; i<8; i++){
        ((uint32_t*)(bak + i*stride))[0]= ((uint32_t*)(src2 + i*stride))[0];
        ((uint32_t*)(bak + i*stride))[1]= ((uint32_t*)(src2 + i*stride))[1];
    }

    s->dsp.diff_pixels(temp, src1, src2, stride);

    s->block_last_index[0/*FIXME*/]= last= s->fast_dct_quantize(s, temp, 0/*FIXME*/, s->qscale, &i);

    bits=0;

            s->dct_unquantize_intra(s, temp, 0, s->qscale);
        else
            s->dct_unquantize_inter(s, temp, 0, s->qscale);
    }

    s->dsp.idct_add(bak, stride, temp);

    distortion= s->dsp.sse[1](NULL, bak, src1, stride, 8);

    return distortion + ((bits*s->qscale*s->qscale*109 + 64)>>7);
}

static int bit8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){

10068:770294452eaf

static int rd8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
    MpegEncContext * const s= (MpegEncContext *)c;
    const uint8_t *scantable= s->intra_scantable.permutated;
    DECLARE_ALIGNED_8 (uint64_t, aligned_temp[sizeof(DCTELEM)*64/8]);
    DECLARE_ALIGNED_8 (uint64_t, aligned_src1[8]);
    DECLARE_ALIGNED_8 (uint64_t, aligned_src2[8]);
    DCTELEM * const temp= (DCTELEM*)aligned_temp;
    uint8_t * const lsrc1 = (uint8_t*)aligned_src1;
    uint8_t * const lsrc2 = (uint8_t*)aligned_src2;
    int i, last, run, bits, level, distortion, start_i;
    const int esc_length= s->ac_esc_length;
    uint8_t * length;
    uint8_t * last_length;

    assert(h==8);

    copy_block8(lsrc1, src1, 8, stride, 8);
    copy_block8(lsrc2, src2, 8, stride, 8);

    s->dsp.diff_pixels(temp, lsrc1, lsrc2, 8);

    s->block_last_index[0/*FIXME*/]= last= s->fast_dct_quantize(s, temp, 0/*FIXME*/, s->qscale, &i);

    bits=0;

            s->dct_unquantize_intra(s, temp, 0, s->qscale);
        else
            s->dct_unquantize_inter(s, temp, 0, s->qscale);
    }

    s->dsp.idct_add(lsrc2, 8, temp);

    distortion= s->dsp.sse[1](NULL, lsrc2, lsrc1, 8, 8);

    return distortion + ((bits*s->qscale*s->qscale*109 + 64)>>7);
}

static int bit8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){