/** * * @file pzunmqr.c * * @copyright 2009-2014 The University of Tennessee and The University of * Tennessee Research Foundation. All rights reserved. * @copyright 2012-2018 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, * Univ. Bordeaux. All rights reserved. * *** * * @brief Chameleon zunmqr parallel algorithm * * @version 1.0.0 * @comment This file has been automatically generated * from Plasma 2.5.0 for CHAMELEON 1.0.0 * @author Hatem Ltaief * @author Jakub Kurzak * @author Azzam Haidar * @author Mathieu Faverge * @author Emmanuel Agullo * @author Cedric Castagnede * @date 2018-11-09 * @precisions normal z -> s d c * */ #include "control/common.h" #define A(m,n) A, m, n #define B(m,n) B, m, n #define T(m,n) T, m, n #define D(k) D, k, k /** * Parallel application of Q using tile V - QR factorization - dynamic scheduling */ void chameleon_pzunmqr( int genD, cham_side_t side, cham_trans_t trans, CHAM_desc_t *A, CHAM_desc_t *B, CHAM_desc_t *T, CHAM_desc_t *D, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request ) { CHAM_context_t *chamctxt; RUNTIME_option_t options; size_t ws_worker = 0; size_t ws_host = 0; int k, m, n; int ldak, ldbk, ldam, ldan, ldbm, lddk; int tempkm, tempnn, tempkmin, tempmm, tempkn; int ib, minMT, minM; chamctxt = chameleon_context_self(); if (sequence->status != CHAMELEON_SUCCESS) { return; } RUNTIME_options_init(&options, chamctxt, sequence, request); ib = CHAMELEON_IB; if (A->m > A->n) { minM = A->n; minMT = A->nt; } else { minM = A->m; minMT = A->mt; } if ( D == NULL ) { D = A; genD = 0; } /* * zunmqr = A->nb * ib * ztpmqrt = A->nb * ib */ ws_worker = A->nb * ib; #if defined(CHAMELEON_USE_CUDA) /* Worker space * * zunmqr = A->nb * ib * ztpmqrt = 2 * A->nb * ib */ ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 ); #endif ws_worker *= sizeof(CHAMELEON_Complex64_t); ws_host *= sizeof(CHAMELEON_Complex64_t); RUNTIME_options_ws_alloc( &options, ws_worker, ws_host ); if (side == ChamLeft ) { if (trans == ChamConjTrans) { /* * ChamLeft / ChamConjTrans */ for (k = 0; k < minMT; k++) { RUNTIME_iteration_push(chamctxt, k); tempkm = k == B->mt-1 ? B->m-k*B->mb : B->mb; tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb; ldak = BLKLDD(A, k); lddk = BLKLDD(D, k); ldbk = BLKLDD(B, k); if ( genD ) { INSERT_TASK_zlacpy( &options, ChamLower, tempkm, tempkmin, A->nb, A(k, k), ldak, D(k), lddk ); #if defined(CHAMELEON_USE_CUDA) INSERT_TASK_zlaset( &options, ChamUpper, tempkm, tempkmin, 0., 1., D(k), lddk ); #endif } for (n = 0; n < B->nt; n++) { tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb; INSERT_TASK_zunmqr( &options, side, trans, tempkm, tempnn, tempkmin, ib, T->nb, D(k), lddk, T(k, k), T->mb, B(k, n), ldbk); } RUNTIME_data_flush( sequence, D(k) ); RUNTIME_data_flush( sequence, T(k, k) ); for (m = k+1; m < B->mt; m++) { tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb; ldam = BLKLDD(A, m); ldbm = BLKLDD(B, m); for (n = 0; n < B->nt; n++) { tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb; RUNTIME_data_migrate( sequence, B(k, n), B->get_rankof( B, m, n ) ); /* TS kernel */ INSERT_TASK_ztpmqrt( &options, side, trans, tempmm, tempnn, tempkmin, 0, ib, T->nb, A(m, k), ldam, T(m, k), T->mb, B(k, n), ldbk, B(m, n), ldbm); } RUNTIME_data_flush( sequence, A(m, k) ); RUNTIME_data_flush( sequence, T(m, k) ); } /* Restore the original location of the tiles */ for (n = 0; n < B->nt; n++) { RUNTIME_data_migrate( sequence, B(k, n), B->get_rankof( B, k, n ) ); } RUNTIME_iteration_pop(chamctxt); } } /* * ChamLeft / ChamNoTrans */ else { for (k = minMT-1; k >= 0; k--) { RUNTIME_iteration_push(chamctxt, k); tempkm = k == B->mt-1 ? B->m-k*B->mb : B->mb; tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb; ldak = BLKLDD(A, k); ldbk = BLKLDD(B, k); for (m = B->mt-1; m > k; m--) { tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb; ldam = BLKLDD(A, m); ldbm = BLKLDD(B, m); for (n = 0; n < B->nt; n++) { tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb; RUNTIME_data_migrate( sequence, B(k, n), B->get_rankof( B, m, n ) ); /* TS kernel */ INSERT_TASK_ztpmqrt( &options, side, trans, tempmm, tempnn, tempkmin, 0, ib, T->nb, A(m, k), ldam, T(m, k), T->mb, B(k, n), ldbk, B(m, n), ldbm); } RUNTIME_data_flush( sequence, A(m, k) ); RUNTIME_data_flush( sequence, T(m, k) ); } if ( genD ) { INSERT_TASK_zlacpy( &options, ChamLower, tempkm, tempkmin, A->nb, A(k, k), ldak, D(k), ldak ); #if defined(CHAMELEON_USE_CUDA) INSERT_TASK_zlaset( &options, ChamUpper, tempkm, tempkmin, 0., 1., D(k), ldak ); #endif } for (n = 0; n < B->nt; n++) { tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb; RUNTIME_data_migrate( sequence, B(k, n), B->get_rankof( B, k, n ) ); INSERT_TASK_zunmqr( &options, side, trans, tempkm, tempnn, tempkmin, ib, T->nb, D(k), ldak, T(k, k), T->mb, B(k, n), ldbk); } RUNTIME_data_flush( sequence, D(k) ); RUNTIME_data_flush( sequence, T(k, k) ); RUNTIME_iteration_pop(chamctxt); } } } /* * ChamRight / ChamConjTrans */ else { if (trans == ChamConjTrans) { for (k = minMT-1; k >= 0; k--) { RUNTIME_iteration_push(chamctxt, k); tempkn = k == B->nt - 1 ? B->n - k * B->nb : B->nb; tempkmin = k == minMT - 1 ? minM - k * A->nb : A->nb; ldak = BLKLDD(A, k); for (n = B->nt-1; n > k; n--) { tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb; ldan = BLKLDD(A, n); for (m = 0; m < B->mt; m++) { tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb; ldbm = BLKLDD(B, m); RUNTIME_data_migrate( sequence, B(m, k), B->get_rankof( B, m, n ) ); /* TS kernel */ INSERT_TASK_ztpmqrt( &options, side, trans, tempmm, tempnn, tempkmin, 0, ib, T->nb, A(n, k), ldan, T(n, k), T->mb, B(m, k), ldbm, B(m, n), ldbm); } RUNTIME_data_flush( sequence, A(n, k) ); RUNTIME_data_flush( sequence, T(n, k) ); } if ( genD ) { INSERT_TASK_zlacpy( &options, ChamLower, tempkn, tempkmin, A->nb, A(k, k), ldak, D(k), ldak ); #if defined(CHAMELEON_USE_CUDA) INSERT_TASK_zlaset( &options, ChamUpper, tempkn, tempkmin, 0., 1., D(k), ldak ); #endif } for (m = 0; m < B->mt; m++) { tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb; ldbm = BLKLDD(B, m); RUNTIME_data_migrate( sequence, B(m, k), B->get_rankof( B, m, k ) ); INSERT_TASK_zunmqr( &options, side, trans, tempmm, tempkn, tempkmin, ib, T->nb, D(k), ldak, T(k, k), T->mb, B(m, k), ldbm); } RUNTIME_data_flush( sequence, D(k) ); RUNTIME_data_flush( sequence, T(k, k) ); RUNTIME_iteration_pop(chamctxt); } } /* * ChamRight / ChamNoTrans */ else { for (k = 0; k < minMT; k++) { RUNTIME_iteration_push(chamctxt, k); tempkn = k == B->nt-1 ? B->n-k*B->nb : B->nb; tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb; ldak = BLKLDD(A, k); if ( genD ) { INSERT_TASK_zlacpy( &options, ChamLower, tempkn, tempkmin, A->nb, A(k, k), ldak, D(k), ldak ); #if defined(CHAMELEON_USE_CUDA) INSERT_TASK_zlaset( &options, ChamUpper, tempkn, tempkmin, 0., 1., D(k), ldak ); #endif } for (m = 0; m < B->mt; m++) { tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb; ldbm = BLKLDD(B, m); INSERT_TASK_zunmqr( &options, side, trans, tempmm, tempkn, tempkmin, ib, T->nb, D(k), ldak, T(k, k), T->mb, B(m, k), ldbm); } RUNTIME_data_flush( sequence, D(k) ); RUNTIME_data_flush( sequence, T(k, k) ); for (n = k+1; n < B->nt; n++) { tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb; ldan = BLKLDD(A, n); for (m = 0; m < B->mt; m++) { tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb; ldbm = BLKLDD(B, m); RUNTIME_data_migrate( sequence, B(m, k), B->get_rankof( B, m, n ) ); /* TS kernel */ INSERT_TASK_ztpmqrt( &options, side, trans, tempmm, tempnn, tempkmin, 0, ib, T->nb, A(n, k), ldan, T(n, k), T->mb, B(m, k), ldbm, B(m, n), ldbm); } RUNTIME_data_flush( sequence, A(n, k) ); RUNTIME_data_flush( sequence, T(n, k) ); } /* Restore the original location of the tiles */ for (m = 0; m < B->mt; m++) { RUNTIME_data_migrate( sequence, B(m, k), B->get_rankof( B, m, k ) ); } RUNTIME_iteration_pop(chamctxt); } } } RUNTIME_options_ws_free(&options); RUNTIME_options_finalize(&options, chamctxt); }