/* Function: p7_gbands_Reverse()
* Synopsis: Reverse the band structure arrays, after a backwards DP pass.
*
* Purpose: Our checkpointed DP posterior decoding algorithms that make a
* band structure work backwards in rows from L..1, and so they
* construct a <P7_GBANDS> structure that has its data elements
* in reversed order. Before we can use that structure, we have
* to reverse these arrays.
*
* Args: bnd - band list to reverse.
*
* Returns: <eslOK> on success.
*/
int
p7_gbands_Reverse(P7_GBANDS *bnd)
{
esl_vec_IReverse(bnd->imem, bnd->imem, 2*bnd->nseg);
esl_vec_IReverse(bnd->kmem, bnd->kmem, 2*bnd->nrow);
return eslOK;
}
int
p7_sparsemask_Finish_avx512(P7_SPARSEMASK *sm)
{
#ifdef HAVE_AVX512
int i,r;
int s;
int status;
//printf("calling p7_sparsemask_Finish, sm->ncells = %li\n", sm->ncells);
/* Reverse kmem. */
int *p_AVX_512;
esl_vec_IReverse(sm->kmem_AVX_512, sm->kmem_AVX_512, sm->ncells_AVX_512);
/* Set the k[] pointers; count <S> and <nrow> */
p_AVX_512 = sm->kmem_AVX_512;
sm->S_AVX_512 = sm->nrow_AVX_512 = 0;
for (i = 1; i <= sm->L; i++){
if (sm->n_AVX_512[i])
{
sm->nrow_AVX_512++;
sm->k_AVX_512[i] = p_AVX_512;
p_AVX_512 += sm->n_AVX_512[i];
if (sm->n_AVX_512[i-1] == 0) sm->S_AVX_512++;
}
else
sm->k_AVX_512[i] = NULL;
}
/* Reallocate seg[] if needed. */
if ( (sm->S_AVX_512+2) > sm->salloc_AVX_512)
{
ESL_REALLOC(sm->seg_AVX_512, (sm->S_AVX_512+2) * sizeof(p7_sparsemask_seg_s)); /* +2, for sentinels */
sm->salloc_AVX_512 = sm->S_AVX_512 + 2; // inclusive of sentinels
sm->n_srealloc++;
}
/* Set seg[] coord pairs. */
sm->seg_AVX_512[0].ia = sm->seg_AVX_512[0].ib = -1;
for (s = 1, i = 1; i <= sm->L; i++)
{
if (sm->n_AVX_512[i] && sm->n_AVX_512[i-1] == 0) sm->seg_AVX_512[s].ia = i;
if (sm->n_AVX_512[i] && (i == sm->L || sm->n_AVX_512[i+1] == 0)) sm->seg_AVX_512[s++].ib = i;
}
ESL_DASSERT1(( s == sm->S_AVX_512+1 ));
sm->seg_AVX_512[s].ia = sm->seg_AVX_512[s].ib = sm->L+2;
sm->last_i_AVX_512 = -1;
for (r = 0; r < p7_VNF_AVX_512; r++)
sm->last_k_AVX_512[r] = -1;
// if we're running AVX-512 code and not SSE, need to copy some values into the SSE data structure
// so the downstream code will see them
sm->seg = sm->seg_AVX_512;
sm->k = sm->k_AVX_512;
sm->n = sm->n_AVX_512;
sm->kmem = sm->kmem_AVX_512;
sm->S = sm->S_AVX_512;
sm->nrow = sm->nrow_AVX_512;
sm->ncells = sm->ncells_AVX_512;
return eslOK;
ERROR:
return eslEMEM;
#endif
#ifndef HAVE_AVX512
return eslENORESULT;
#endif
}
