comparison vp8.c @ 12237:f0c4dc49c8f1 libavcodec

VP8: smarter prefetching Don't prefetch reference frames that were used less than 1/32th of the time so far in the frame. This helps speed up to ~2% on videos that, in many frames, make near-zero (but not entirely zero) use of golden and/or alt-refs. This is a very common property of videos encoded by libvpx.

12236:cabcd751b1e5

           s->put_pixels_tab[1 + (block_w == 4)]);
}

/* Fetch pixels for estimated mv 4 macroblocks ahead.
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int ref)
{
    /* Don't prefetch refs that haven't been used yet this frame. */
    if (s->ref_count[ref-1]) {
        int x_off = mb_x << 4, y_off = mb_y << 4;
        int mx = mb->mv.x + x_off + 8;
        int my = mb->mv.y + y_off;
        uint8_t **src= s->framep[ref]->data;
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;

    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
        uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
        uint8_t *dst[3] = {
            curframe->data[0] + 16*mb_y*s->linesize,
            curframe->data[1] +  8*mb_y*s->uvlinesize,
            curframe->data[2] +  8*mb_y*s->uvlinesize
        };

                for (y = 0; y < 16>>!!i; y++)
                    dst[i][y*curframe->linesize[i]-1] = 129;
        if (mb_y)
            memset(s->top_border, 129, sizeof(*s->top_border));

        for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
            uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
            uint8_t *segment_mb = segment_map+mb_x;

            /* Prefetch the current frame, 4 MBs ahead */
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);

            decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);

            prefetch_motion(s, mb, mb_x, mb_y, VP56_FRAME_PREVIOUS);

            if (!mb->skip)
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);

            if (mb->mode <= MODE_I4x4)
                intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
            else
                inter_predict(s, dst, mb, mb_x, mb_y);

            prefetch_motion(s, mb, mb_x, mb_y, VP56_FRAME_GOLDEN);

            if (!mb->skip) {
                idct_mb(s, dst[0], dst[1], dst[2], mb);
            } else {
                AV_ZERO64(s->left_nnz);

            }

            if (s->deblock_filter)
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);

            prefetch_motion(s, mb, mb_x, mb_y, VP56_FRAME_GOLDEN2);

            dst[0] += 16;
            dst[1] += 8;
            dst[2] += 8;
            mb++;
        }
        if (s->deblock_filter) {
            if (s->filter.simple)
                filter_mb_row_simple(s, mb_y);
            else

12237:f0c4dc49c8f1

           s->put_pixels_tab[1 + (block_w == 4)]);
}

/* Fetch pixels for estimated mv 4 macroblocks ahead.
 * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
static inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
{
    /* Don't prefetch refs that haven't been used very often this frame. */
    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
        int x_off = mb_x << 4, y_off = mb_y << 4;
        int mx = mb->mv.x + x_off + 8;
        int my = mb->mv.y + y_off;
        uint8_t **src= s->framep[ref]->data;
        int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;

    for (mb_y = 0; mb_y < s->mb_height; mb_y++) {
        VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];
        VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;
        uint8_t *intra4x4 = s->intra4x4_pred_mode + 4*mb_y*s->b4_stride;
        uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride;
        int mb_xy = mb_y * s->mb_stride;
        uint8_t *dst[3] = {
            curframe->data[0] + 16*mb_y*s->linesize,
            curframe->data[1] +  8*mb_y*s->uvlinesize,
            curframe->data[2] +  8*mb_y*s->uvlinesize
        };

                for (y = 0; y < 16>>!!i; y++)
                    dst[i][y*curframe->linesize[i]-1] = 129;
        if (mb_y)
            memset(s->top_border, 129, sizeof(*s->top_border));

        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
            uint8_t *intra4x4_mb = s->keyframe ? intra4x4 + 4*mb_x : s->intra4x4_pred_mode_mb;
            uint8_t *segment_mb = segment_map+mb_x;

            /* Prefetch the current frame, 4 MBs ahead */
            s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);
            s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);

            decode_mb_mode(s, mb, mb_x, mb_y, intra4x4_mb, segment_mb);

            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);

            if (!mb->skip)
                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);

            if (mb->mode <= MODE_I4x4)
                intra_predict(s, dst, mb, intra4x4_mb, mb_x, mb_y);
            else
                inter_predict(s, dst, mb, mb_x, mb_y);

            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);

            if (!mb->skip) {
                idct_mb(s, dst[0], dst[1], dst[2], mb);
            } else {
                AV_ZERO64(s->left_nnz);

            }

            if (s->deblock_filter)
                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);

            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);

            dst[0] += 16;
            dst[1] += 8;
            dst[2] += 8;
        }
        if (s->deblock_filter) {
            if (s->filter.simple)
                filter_mb_row_simple(s, mb_y);
            else