static void copy_block_in(bench_real *in,
int rnk, ptrdiff_t *n, ptrdiff_t *start,
ptrdiff_t is, ptrdiff_t *os, ptrdiff_t vn,
const bench_real *out)
{
ptrdiff_t i;
if (rnk == 0) {
for (i = 0; i < vn; ++i)
in[i] = out[i];
}
else if (rnk == 1) { /* this case is just an optimization */
ptrdiff_t j;
out += start[0] * os[0];
for (j = 0; j < n[0]; ++j) {
for (i = 0; i < vn; ++i)
in[i] = out[i];
in += is;
out += os[0];
}
}
else {
/* we should do n[0] for locality, but this way is simpler to code */
for (i = 0; i < n[rnk - 1]; ++i)
copy_block_in(in + i * is,
rnk - 1, n, start, is * n[rnk - 1], os, vn,
out + (start[rnk - 1] + i) * os[rnk - 1]);
}
}
static void
restore_dc_values (SchroEncoderFrame *frame, int16_t *dc_values,
SchroLowDelay *lowdelay,
int slice_x, int slice_y)
{
SchroFrameData block;
schro_frame_data_get_codeblock (&block, lowdelay->luma_subbands + 0,
slice_x, slice_y,
lowdelay->n_horiz_slices, lowdelay->n_vert_slices);
copy_block_in (&block, dc_values);
dc_values += block.width * block.height;
schro_frame_data_get_codeblock (&block, lowdelay->chroma1_subbands + 0,
slice_x, slice_y,
lowdelay->n_horiz_slices, lowdelay->n_vert_slices);
copy_block_in (&block, dc_values);
dc_values += block.width * block.height;
schro_frame_data_get_codeblock (&block, lowdelay->chroma2_subbands + 0,
slice_x, slice_y,
lowdelay->n_horiz_slices, lowdelay->n_vert_slices);
copy_block_in (&block, dc_values);
}
static void do_scatter_in(bench_real *in)
{
bench_real *ali;
int i;
if (all_local_in_alloc) {
bench_free(all_local_in);
all_local_in = (bench_real*) bench_malloc(iNtot*sizeof(bench_real));
all_local_in_alloc = 0;
}
ali = all_local_in;
for (i = 0; i < n_pes; ++i) {
copy_block_in(ali,
rnk, all_local_ni + i * rnk,
all_local_starti + i * rnk,
vn, istrides, vn,
in);
ali += isend_cnt[i];
}
MPI_Scatterv(all_local_in, isend_cnt, isend_off, BENCH_MPI_TYPE,
local_in, isend_cnt[my_pe], BENCH_MPI_TYPE,
0, MPI_COMM_WORLD);
}
