comparison vp8.c @ 12248:121272849def libavcodec

VP8: always_inline some things to force gcc to do the right thing Mostly seems to help in the MC code, which gets a hundred cycles faster.

12247:50a96623366b

    }

    return 0;
}

static inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src,
                            int mb_x, int mb_y)
{
#define MARGIN (16 << 2)
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
}

static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
                          VP56mv near[2], VP56mv *best, uint8_t cnt[4])
{
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
                                  mb - 1 /* left */,
                                  mb + 1 /* top-left */ };
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };

        x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);

    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
}

static const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
{
    if (left == top)
        return vp8_submv_prob[4-!!left];
    if (!top)
        return vp8_submv_prob[2];

/**
 * Split motion vector prediction, 16.4.
 * @returns the number of motion vectors parsed (2, 4 or 16)
 */
static int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
{
    int part_idx = mb->partitioning =
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
    int n, num = vp8_mbsplit_count[part_idx];
    VP8Macroblock *top_mb  = &mb[2];

    }

    return num;
}

static inline void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
                                         int stride, int keyframe)
{
    int x, y, t, l, i;

    if (keyframe) {
        const uint8_t *ctx;

        for (i = 0; i < 16; i++)
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
    }
}

static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
                           uint8_t *intra4x4, uint8_t *segment)
{
    VP56RangeCoder *c = &s->c;

    if (s->segmentation.update_map)
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);

        offset = 0;
    }
    return nonzero;
}

static void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
                             uint8_t t_nnz[9], uint8_t l_nnz[9])
{
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
    int i, x, y, luma_start = 0, luma_ctx = 3;
    int nnz_pred, nnz, nnz_total = 0;
    int segment = s->segment;

        XCHG(top_border+16,    src_cb, 1);
        XCHG(top_border+24,    src_cr, 1);
    }
}

static int check_intra_pred_mode(int mode, int mb_x, int mb_y)
{
    if (mode == DC_PRED8x8) {
        if (!mb_x) {
            mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
        } else if (!mb_y) {

        }
    }
    return mode;
}

static void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
                          uint8_t *intra4x4, int mb_x, int mb_y)
{
    int x, y, mode, nnz, tr;

    // for the first row, we need to run xchg_mb_border to init the top edge to 127
    // otherwise, skip it if we aren't going to deblock

 * @param width width of src/dst plane data
 * @param height height of src/dst plane data
 * @param linesize size of a single line of plane data, including padding
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
 */
static inline void vp8_mc(VP8Context *s, int luma,
                          uint8_t *dst, uint8_t *src, const VP56mv *mv,
                          int x_off, int y_off, int block_w, int block_h,
                          int width, int height, int linesize,
                          vp8_mc_func mc_func[3][3])
{
    if (AV_RN32A(mv)) {
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
        int my = (mv->y << luma)&7, my_idx = idx[my];

        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
    } else
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
}

static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
                               AVFrame *ref_frame, int x_off, int y_off,
                               int bx_off, int by_off,
                               int block_w, int block_h,
                               int width, int height, VP56mv *mv)
{
    VP56mv uvmv = *mv;

    /* Y */
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,

           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;

}

/**
 * Apply motion vectors to prediction buffer, chapter 18.
 */
static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
                          int mb_x, int mb_y)
{
    int x_off = mb_x << 4, y_off = mb_y << 4;
    int width = 16*s->mb_width, height = 16*s->mb_height;
    AVFrame *ref = s->framep[mb->ref_frame];
    VP56mv *bmv = mb->bmv;

                    8, 8, 8, 8, width, height, &bmv[3]);
        break;
    }
}

static void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
{
    int x, y, ch;

    if (mb->mode != MODE_I4x4) {
        uint8_t *y_dst = dst[0];

            }
        }
    }
}

static void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
{
    int interior_limit, filter_level;

    if (s->segmentation.enabled) {
        filter_level = s->segmentation.filter_level[s->segment];

    f->filter_level = filter_level;
    f->inner_limit = interior_limit;
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
}

static void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
{
    int mbedge_lim, bedge_lim, hev_thresh;
    int filter_level = f->filter_level;
    int inner_limit = f->inner_limit;
    int inner_filter = f->inner_filter;

                                             uvlinesize,  bedge_lim,
                                             inner_limit, hev_thresh);
    }
}

static void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
{
    int mbedge_lim, bedge_lim;
    int filter_level = f->filter_level;
    int inner_limit = f->inner_limit;
    int inner_filter = f->inner_filter;

12248:121272849def

    }

    return 0;
}

static av_always_inline
void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y)
{
#define MARGIN (16 << 2)
    dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN),
                     ((s->mb_width  - 1 - mb_x) << 6) + MARGIN);
    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
}

static av_always_inline
void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
{
    VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,
                                  mb - 1 /* left */,
                                  mb + 1 /* top-left */ };
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };

        x = vp8_rac_get_tree(c, vp8_small_mvtree, &p[2]);

    return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;
}

static av_always_inline
const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
{
    if (left == top)
        return vp8_submv_prob[4-!!left];
    if (!top)
        return vp8_submv_prob[2];

/**
 * Split motion vector prediction, 16.4.
 * @returns the number of motion vectors parsed (2, 4 or 16)
 */
static av_always_inline
int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
{
    int part_idx = mb->partitioning =
        vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
    int n, num = vp8_mbsplit_count[part_idx];
    VP8Macroblock *top_mb  = &mb[2];

    }

    return num;
}

static av_always_inline
void decode_intra4x4_modes(VP56RangeCoder *c, uint8_t *intra4x4,
                           int stride, int keyframe)
{
    int x, y, t, l, i;

    if (keyframe) {
        const uint8_t *ctx;

        for (i = 0; i < 16; i++)
            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
    }
}

static av_always_inline
void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
                    uint8_t *intra4x4, uint8_t *segment)
{
    VP56RangeCoder *c = &s->c;

    if (s->segmentation.update_map)
        *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);

        offset = 0;
    }
    return nonzero;
}

static av_always_inline
void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
                      uint8_t t_nnz[9], uint8_t l_nnz[9])
{
    LOCAL_ALIGNED_16(DCTELEM, dc,[16]);
    int i, x, y, luma_start = 0, luma_ctx = 3;
    int nnz_pred, nnz, nnz_total = 0;
    int segment = s->segment;

        XCHG(top_border+16,    src_cb, 1);
        XCHG(top_border+24,    src_cr, 1);
    }
}

static av_always_inline
int check_intra_pred_mode(int mode, int mb_x, int mb_y)
{
    if (mode == DC_PRED8x8) {
        if (!mb_x) {
            mode = mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
        } else if (!mb_y) {

        }
    }
    return mode;
}

static av_always_inline
void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
                   uint8_t *intra4x4, int mb_x, int mb_y)
{
    int x, y, mode, nnz, tr;

    // for the first row, we need to run xchg_mb_border to init the top edge to 127
    // otherwise, skip it if we aren't going to deblock

 * @param width width of src/dst plane data
 * @param height height of src/dst plane data
 * @param linesize size of a single line of plane data, including padding
 * @param mc_func motion compensation function pointers (bilinear or sixtap MC)
 */
static av_always_inline
void vp8_mc(VP8Context *s, int luma,
            uint8_t *dst, uint8_t *src, const VP56mv *mv,
            int x_off, int y_off, int block_w, int block_h,
            int width, int height, int linesize,
            vp8_mc_func mc_func[3][3])
{
    if (AV_RN32A(mv)) {
        static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 };
        int mx = (mv->x << luma)&7, mx_idx = idx[mx];
        int my = (mv->y << luma)&7, my_idx = idx[my];

        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
    } else
        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
}

static av_always_inline
void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
                 AVFrame *ref_frame, int x_off, int y_off,
                 int bx_off, int by_off,
                 int block_w, int block_h,
                 int width, int height, VP56mv *mv)
{
    VP56mv uvmv = *mv;

    /* Y */
    vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off,

           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 av_always_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;

}

/**
 * Apply motion vectors to prediction buffer, chapter 18.
 */
static av_always_inline
void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
                   int mb_x, int mb_y)
{
    int x_off = mb_x << 4, y_off = mb_y << 4;
    int width = 16*s->mb_width, height = 16*s->mb_height;
    AVFrame *ref = s->framep[mb->ref_frame];
    VP56mv *bmv = mb->bmv;

                    8, 8, 8, 8, width, height, &bmv[3]);
        break;
    }
}

static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
{
    int x, y, ch;

    if (mb->mode != MODE_I4x4) {
        uint8_t *y_dst = dst[0];

            }
        }
    }
}

static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
{
    int interior_limit, filter_level;

    if (s->segmentation.enabled) {
        filter_level = s->segmentation.filter_level[s->segment];

    f->filter_level = filter_level;
    f->inner_limit = interior_limit;
    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
}

static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
{
    int mbedge_lim, bedge_lim, hev_thresh;
    int filter_level = f->filter_level;
    int inner_limit = f->inner_limit;
    int inner_filter = f->inner_filter;

                                             uvlinesize,  bedge_lim,
                                             inner_limit, hev_thresh);
    }
}

static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
{
    int mbedge_lim, bedge_lim;
    int filter_level = f->filter_level;
    int inner_limit = f->inner_limit;
    int inner_filter = f->inner_filter;