static int find_best_16x16_intra(VP9_COMP *cpi, PREDICTION_MODE *pbest_mode) {
MACROBLOCK *const x = &cpi->td.mb;
MACROBLOCKD *const xd = &x->e_mbd;
PREDICTION_MODE best_mode = -1, mode;
unsigned int best_err = INT_MAX;
// calculate SATD for each intra prediction mode;
// we're intentionally not doing 4x4, we just want a rough estimate
for (mode = DC_PRED; mode <= TM_PRED; mode++) {
unsigned int err;
xd->mi[0]->mbmi.mode = mode;
vp9_predict_intra_block(xd, 0, 2, TX_16X16, mode,
x->plane[0].src.buf, x->plane[0].src.stride,
xd->plane[0].dst.buf, xd->plane[0].dst.stride,
0, 0, 0);
err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
xd->plane[0].dst.buf, xd->plane[0].dst.stride);
// find best
if (err < best_err) {
best_err = err;
best_mode = mode;
}
}
if (pbest_mode)
*pbest_mode = best_mode;
return best_err;
}
static int find_best_16x16_intra
(
VP9_COMP *cpi,
YV12_BUFFER_CONFIG *buf,
int mb_y_offset,
MB_PREDICTION_MODE *pbest_mode
) {
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
MB_PREDICTION_MODE best_mode = -1, mode;
unsigned int best_err = INT_MAX;
// calculate SATD for each intra prediction mode;
// we're intentionally not doing 4x4, we just want a rough estimate
for (mode = DC_PRED; mode <= TM_PRED; mode++) {
unsigned int err;
xd->mode_info_context->mbmi.mode = mode;
vp9_build_intra_predictors_mby(xd);
err = vp9_sad16x16(xd->predictor, 16, buf->y_buffer + mb_y_offset,
buf->y_stride, best_err);
// find best
if (err < best_err) {
best_err = err;
best_mode = mode;
}
}
if (pbest_mode)
*pbest_mode = best_mode;
return best_err;
}
static int do_16x16_motion_search
(
VP9_COMP *cpi,
int_mv *ref_mv,
int_mv *dst_mv,
YV12_BUFFER_CONFIG *buf,
int buf_mb_y_offset,
YV12_BUFFER_CONFIG *ref,
int mb_y_offset
) {
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
unsigned int err, tmp_err;
int_mv tmp_mv;
int n;
for (n = 0; n < 16; n++) {
BLOCKD *d = &xd->block[n];
BLOCK *b = &x->block[n];
b->base_src = &buf->y_buffer;
b->src_stride = buf->y_stride;
b->src = buf->y_stride * (n & 12) + (n & 3) * 4 + buf_mb_y_offset;
d->base_pre = &ref->y_buffer;
d->pre_stride = ref->y_stride;
d->pre = ref->y_stride * (n & 12) + (n & 3) * 4 + mb_y_offset;
}
// Try zero MV first
// FIXME should really use something like near/nearest MV and/or MV prediction
xd->pre.y_buffer = ref->y_buffer + mb_y_offset;
xd->pre.y_stride = ref->y_stride;
err = vp9_sad16x16(ref->y_buffer + mb_y_offset, ref->y_stride,
xd->dst.y_buffer, xd->dst.y_stride, INT_MAX);
dst_mv->as_int = 0;
// Test last reference frame using the previous best mv as the
// starting point (best reference) for the search
tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv);
if (tmp_err < err) {
err = tmp_err;
dst_mv->as_int = tmp_mv.as_int;
}
// If the current best reference mv is not centred on 0,0 then do a 0,0 based search as well
if (ref_mv->as_int) {
unsigned int tmp_err;
int_mv zero_ref_mv, tmp_mv;
zero_ref_mv.as_int = 0;
tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv);
if (tmp_err < err) {
dst_mv->as_int = tmp_mv.as_int;
err = tmp_err;
}
}
return err;
}
static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
const MV *ref_mv,
MV *dst_mv,
int mb_row,
int mb_col) {
MACROBLOCK *const x = &cpi->td.mb;
MACROBLOCKD *const xd = &x->e_mbd;
const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
const vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
const int tmp_col_min = x->mv_col_min;
const int tmp_col_max = x->mv_col_max;
const int tmp_row_min = x->mv_row_min;
const int tmp_row_max = x->mv_row_max;
MV ref_full;
int cost_list[5];
// Further step/diamond searches as necessary
int step_param = mv_sf->reduce_first_step_size;
step_param = MIN(step_param, MAX_MVSEARCH_STEPS - 2);
vp9_set_mv_search_range(x, ref_mv);
ref_full.col = ref_mv->col >> 3;
ref_full.row = ref_mv->row >> 3;
/*cpi->sf.search_method == HEX*/
vp9_hex_search(x, &ref_full, step_param, x->errorperbit, 0,
cond_cost_list(cpi, cost_list),
&v_fn_ptr, 0, ref_mv, dst_mv);
// Try sub-pixel MC
// if (bestsme > error_thresh && bestsme < INT_MAX)
{
int distortion;
unsigned int sse;
cpi->find_fractional_mv_step(
x, dst_mv, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
&v_fn_ptr, 0, mv_sf->subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
NULL, NULL,
&distortion, &sse, NULL, 0, 0);
}
xd->mi[0]->mbmi.mode = NEWMV;
xd->mi[0]->mbmi.mv[0].as_mv = *dst_mv;
vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
/* restore UMV window */
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
x->mv_row_min = tmp_row_min;
x->mv_row_max = tmp_row_max;
return vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
xd->plane[0].dst.buf, xd->plane[0].dst.stride);
}
static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
int_mv *ref_mv,
int_mv *dst_mv,
int mb_row,
int mb_col) {
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
unsigned int best_err;
const int tmp_col_min = x->mv_col_min;
const int tmp_col_max = x->mv_col_max;
const int tmp_row_min = x->mv_row_min;
const int tmp_row_max = x->mv_row_max;
int_mv ref_full;
// Further step/diamond searches as necessary
int step_param = cpi->sf.reduce_first_step_size +
(cpi->speed < 8 ? (cpi->speed > 5 ? 1 : 0) : 2);
step_param = MIN(step_param, (cpi->sf.max_step_search_steps - 2));
vp9_clamp_mv_min_max(x, &ref_mv->as_mv);
ref_full.as_mv.col = ref_mv->as_mv.col >> 3;
ref_full.as_mv.row = ref_mv->as_mv.row >> 3;
/*cpi->sf.search_method == HEX*/
best_err = vp9_hex_search(x, &ref_full, step_param, x->errorperbit,
0, &v_fn_ptr,
0, ref_mv, dst_mv);
// Try sub-pixel MC
// if (bestsme > error_thresh && bestsme < INT_MAX)
{
int distortion;
unsigned int sse;
best_err = cpi->find_fractional_mv_step(
x,
dst_mv, ref_mv,
x->errorperbit, &v_fn_ptr,
0, cpi->sf.subpel_iters_per_step, NULL, NULL,
& distortion, &sse);
}
vp9_set_mbmode_and_mvs(x, NEWMV, dst_mv);
vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
best_err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
xd->plane[0].dst.buf, xd->plane[0].dst.stride,
INT_MAX);
/* restore UMV window */
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
x->mv_row_min = tmp_row_min;
x->mv_row_max = tmp_row_max;
return best_err;
}
static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
MACROBLOCK *const x = &cpi->td.mb;
MACROBLOCKD *const xd = &x->e_mbd;
unsigned int err;
// Try zero MV first
// FIXME should really use something like near/nearest MV and/or MV prediction
err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
dst_mv->as_int = 0;
return err;
}
static int do_16x16_motion_search(VP9_COMP *cpi, const int_mv *ref_mv,
int_mv *dst_mv, int mb_row, int mb_col) {
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
unsigned int err, tmp_err;
int_mv tmp_mv;
// Try zero MV first
// FIXME should really use something like near/nearest MV and/or MV prediction
err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
INT_MAX);
dst_mv->as_int = 0;
// Test last reference frame using the previous best mv as the
// starting point (best reference) for the search
tmp_err = do_16x16_motion_iteration(cpi, &ref_mv->as_mv, &tmp_mv.as_mv,
mb_row, mb_col);
if (tmp_err < err) {
err = tmp_err;
dst_mv->as_int = tmp_mv.as_int;
}
// If the current best reference mv is not centered on 0,0 then do a 0,0
// based search as well.
if (ref_mv->as_int) {
unsigned int tmp_err;
int_mv zero_ref_mv, tmp_mv;
zero_ref_mv.as_int = 0;
tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv.as_mv, &tmp_mv.as_mv,
mb_row, mb_col);
if (tmp_err < err) {
dst_mv->as_int = tmp_mv.as_int;
err = tmp_err;
}
}
return err;
}
static int do_16x16_zerozero_search
(
VP9_COMP *cpi,
int_mv *dst_mv,
YV12_BUFFER_CONFIG *buf,
int buf_mb_y_offset,
YV12_BUFFER_CONFIG *ref,
int mb_y_offset
) {
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
unsigned int err;
int n;
for (n = 0; n < 16; n++) {
BLOCKD *d = &xd->block[n];
BLOCK *b = &x->block[n];
b->base_src = &buf->y_buffer;
b->src_stride = buf->y_stride;
b->src = buf->y_stride * (n & 12) + (n & 3) * 4 + buf_mb_y_offset;
d->base_pre = &ref->y_buffer;
d->pre_stride = ref->y_stride;
d->pre = ref->y_stride * (n & 12) + (n & 3) * 4 + mb_y_offset;
}
// Try zero MV first
// FIXME should really use something like near/nearest MV and/or MV prediction
xd->pre.y_buffer = ref->y_buffer + mb_y_offset;
xd->pre.y_stride = ref->y_stride;
err = vp9_sad16x16(ref->y_buffer + mb_y_offset, ref->y_stride,
xd->dst.y_buffer, xd->dst.y_stride, INT_MAX);
dst_mv->as_int = 0;
return err;
}
static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
int_mv *ref_mv,
int_mv *dst_mv,
int mb_row,
int mb_col) {
MACROBLOCK *const x = &cpi->mb;
MACROBLOCKD *const xd = &x->e_mbd;
BLOCK *b = &x->block[0];
BLOCKD *d = &xd->block[0];
vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
unsigned int best_err;
int tmp_col_min = x->mv_col_min;
int tmp_col_max = x->mv_col_max;
int tmp_row_min = x->mv_row_min;
int tmp_row_max = x->mv_row_max;
int_mv ref_full;
// Further step/diamond searches as necessary
int step_param = cpi->sf.first_step +
(cpi->Speed < 8 ? (cpi->Speed > 5 ? 1 : 0) : 2);
vp9_clamp_mv_min_max(x, ref_mv);
ref_full.as_mv.col = ref_mv->as_mv.col >> 3;
ref_full.as_mv.row = ref_mv->as_mv.row >> 3;
/*cpi->sf.search_method == HEX*/
best_err = vp9_hex_search(
x, b, d,
&ref_full, dst_mv,
step_param,
x->errorperbit,
&v_fn_ptr,
NULL, NULL,
NULL, NULL,
ref_mv);
// Try sub-pixel MC
// if (bestsme > error_thresh && bestsme < INT_MAX)
{
int distortion;
unsigned int sse;
best_err = cpi->find_fractional_mv_step(
x, b, d,
dst_mv, ref_mv,
x->errorperbit, &v_fn_ptr,
NULL, NULL,
& distortion, &sse);
}
vp9_set_mbmode_and_mvs(x, NEWMV, dst_mv);
vp9_build_inter16x16_predictors_mby(xd, xd->predictor, 16, mb_row, mb_col);
best_err = vp9_sad16x16(xd->dst.y_buffer, xd->dst.y_stride,
xd->predictor, 16, INT_MAX);
/* restore UMV window */
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
x->mv_row_min = tmp_row_min;
x->mv_row_max = tmp_row_max;
return best_err;
}
