diff --git a/compute/pzbuild.c b/compute/pzbuild.c
index 1b8d361a45ae0c91a6844576fe2b4ee74782bfd7..f4e3a6c0b8f0bdb26eb7676ca5aa0561e92ee7e2 100644
--- a/compute/pzbuild.c
+++ b/compute/pzbuild.c
@@ -30,7 +30,7 @@
#include "control/common.h"
#define A(m, n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile matrix generation
*
*******************************************************************************
@@ -86,6 +86,5 @@ void morse_pzbuild( MORSE_enum uplo, MORSE_desc_t *A, void *user_data, void* use
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_finalize( &options, morse);
}
diff --git a/compute/pzgelqf.c b/compute/pzgelqf.c
index 3821a120983f2a5cda5082eaeeb0321d5c419e93..2d28d719c13d1855e0ab7fb5c698fc2dc6388743 100644
--- a/compute/pzgelqf.c
+++ b/compute/pzgelqf.c
@@ -38,7 +38,7 @@
#define D(k) A, k, k
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile LQ factorization - dynamic scheduling
**/
void morse_pzgelqf(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D,
@@ -153,10 +153,7 @@ void morse_pzgelqf(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpper, T );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzgelqf_param.c b/compute/pzgelqf_param.c
index 88588051e1ef4e3d3baac9b254aefe02ee841d07..8a889b05c659d96c106362d8b1c82d7fbf0ff569 100644
--- a/compute/pzgelqf_param.c
+++ b/compute/pzgelqf_param.c
@@ -196,12 +196,8 @@ void morse_pzgelqf_param( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpper, TS );
- MORSE_TASK_flush_desc( &options, MorseUpper, TT );
free(tiles);
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzgelqfrh.c b/compute/pzgelqfrh.c
index 141d94195d1800aa0e9158d3e4a97f97f78e26ff..ad11204ad69da663e4c680896ffe286436fb90cf 100644
--- a/compute/pzgelqfrh.c
+++ b/compute/pzgelqfrh.c
@@ -181,10 +181,7 @@ void morse_pzgelqfrh(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D, int BS,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpper, T );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzgemm.c b/compute/pzgemm.c
index e001af357c58b8938c74085a7ce89c4d2204020f..c52fcc281c57374bb2249d21faf9b4c343da1771 100644
--- a/compute/pzgemm.c
+++ b/compute/pzgemm.c
@@ -31,7 +31,7 @@
#define A(m, n) A, m, n
#define B(m, n) B, m, n
#define C(m, n) C, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile matrix-matrix multiplication - dynamic scheduling
**/
void morse_pzgemm(MORSE_enum transA, MORSE_enum transB,
@@ -134,19 +134,17 @@ void morse_pzgemm(MORSE_enum transA, MORSE_enum transB,
}
}
}
- MORSE_TASK_flush_data( &options, C(m, n) );
+ RUNTIME_data_flush( sequence, C(m, n) );
}
if (transA == MorseNoTrans) {
for (k = 0; k < A->nt; k++) {
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
}
} else {
for (k = 0; k < A->mt; k++) {
- MORSE_TASK_flush_data( &options, A(k, m) );
+ RUNTIME_data_flush( sequence, A(k, m) );
}
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzgeqrf.c b/compute/pzgeqrf.c
index 1ccd8bb99dae30fec44fe4a3e150a368b21c6fe2..849d7a00568a25988fb8c66819cb2068209a1a22 100644
--- a/compute/pzgeqrf.c
+++ b/compute/pzgeqrf.c
@@ -38,7 +38,7 @@
#define D(k) A, k, k
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile QR factorization - dynamic scheduling
**/
void morse_pzgeqrf(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D,
@@ -147,10 +147,7 @@ void morse_pzgeqrf(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseLower, T );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzgeqrf_param.c b/compute/pzgeqrf_param.c
index 2382953a802dab0a2f219ad1941dc1c659c29092..3845f4627f459ebeda7733d1e976589f30e6c430 100644
--- a/compute/pzgeqrf_param.c
+++ b/compute/pzgeqrf_param.c
@@ -195,12 +195,8 @@ void morse_pzgeqrf_param( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseLower, TS );
- MORSE_TASK_flush_desc( &options, MorseLower, TT );
free(tiles);
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzgeqrfrh.c b/compute/pzgeqrfrh.c
index aeec53fabbb299ca258319ef5c5f25baa8635dd6..a2ea808b95b26cf679df93e81c1a827e2de1436f 100644
--- a/compute/pzgeqrfrh.c
+++ b/compute/pzgeqrfrh.c
@@ -40,7 +40,7 @@
#define D(m,n) A, (m), (n)
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile QR factorization (reduction Householder) - dynamic scheduling
**/
void morse_pzgeqrfrh(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D, int BS,
@@ -180,10 +180,7 @@ void morse_pzgeqrfrh(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D, int BS,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseLower, T );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzgetrf_incpiv.c b/compute/pzgetrf_incpiv.c
index 9b01a7552c952aed98a79bda85af22545d81efa1..092d258faf2ec47c7c630c9d1767bcbec1bb3774 100644
--- a/compute/pzgetrf_incpiv.c
+++ b/compute/pzgetrf_incpiv.c
@@ -42,7 +42,7 @@
#define L(_m_,_n_) L, _m_, _n_
#define IPIV(_m_,_n_) &(IPIV[(int64_t)A->mb*((int64_t)(_m_)+(int64_t)A->mt*(int64_t)(_n_))])
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile LU factorization - dynamic scheduling
**/
void morse_pzgetrf_incpiv(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
@@ -136,11 +136,8 @@ void morse_pzgetrf_incpiv(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseLower, L );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
morse_desc_mat_free(DIAG);
free(DIAG);
diff --git a/compute/pzgetrf_nopiv.c b/compute/pzgetrf_nopiv.c
index 1385462496fec63ad7c1286fd2d5f2c936a15fc9..e620dcdfa07b64f339d962fe455871fb739364ce 100644
--- a/compute/pzgetrf_nopiv.c
+++ b/compute/pzgetrf_nopiv.c
@@ -29,7 +29,7 @@
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile LU factorization with no pivoting - dynamic scheduling
**/
void morse_pzgetrf_nopiv(MORSE_desc_t *A,
@@ -104,7 +104,5 @@ void morse_pzgetrf_nopiv(MORSE_desc_t *A,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzhemm.c b/compute/pzhemm.c
index a8582ee597f84c26b237e4b83fad0fffb1917e85..7e5ca07095bb0a933e66897bca79179883e6c178 100644
--- a/compute/pzhemm.c
+++ b/compute/pzhemm.c
@@ -31,7 +31,7 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
#define C(m,n) C, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile Hermitian matrix-matrix multiplication - dynamic scheduling
**/
void morse_pzhemm(MORSE_enum side, MORSE_enum uplo,
@@ -226,9 +226,5 @@ void morse_pzhemm(MORSE_enum side, MORSE_enum uplo,
}
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, C );
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzher2k.c b/compute/pzher2k.c
index 9bde1020aaf3c705ac60b6378804ba787ec7fd23..63b02741095ad2a6eba0c54e3ae65e134d61f380 100644
--- a/compute/pzher2k.c
+++ b/compute/pzher2k.c
@@ -31,7 +31,7 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
#define C(m,n) C, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile Hermitian rank-k update - dynamic scheduling
**/
void morse_pzher2k(MORSE_enum uplo, MORSE_enum trans,
@@ -214,9 +214,5 @@ void morse_pzher2k(MORSE_enum uplo, MORSE_enum trans,
}
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, C );
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzherk.c b/compute/pzherk.c
index e3dd914f372242d87ac4037e6b524b9575b65ccc..b499177f11ba862d8339e523abbb6277cc8f68e4 100644
--- a/compute/pzherk.c
+++ b/compute/pzherk.c
@@ -30,7 +30,7 @@
#define A(m,n) A, m, n
#define C(m,n) C, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile Hermitian rank-k update - dynamic scheduling
**/
void morse_pzherk(MORSE_enum uplo, MORSE_enum trans,
@@ -173,7 +173,5 @@ void morse_pzherk(MORSE_enum uplo, MORSE_enum trans,
}
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, uplo, C );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzhetrd_he2hb.c b/compute/pzhetrd_he2hb.c
index 0d8ed1d2f3160e9b5e008f0081a56cdebb8c538c..74c3396149dfc3a4e1d34e129ca2676dcb31f7f8 100644
--- a/compute/pzhetrd_he2hb.c
+++ b/compute/pzhetrd_he2hb.c
@@ -38,7 +38,7 @@
#define E(m, n) A, m, n
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile BAND Tridiagonal Reduction - dynamic scheduler
**/
void morse_pzhetrd_he2hb(MORSE_enum uplo,
@@ -436,10 +436,8 @@ void morse_pzhetrd_he2hb(MORSE_enum uplo,
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
MORSE_Sequence_Wait(sequence);
morse_desc_mat_free(D);
diff --git a/compute/pzlacpy.c b/compute/pzlacpy.c
index 220c8b9dbdd6c12e94afcb23a12ac58413a5ffd8..b13b15222513827741298d84bff4abc56060512e 100644
--- a/compute/pzlacpy.c
+++ b/compute/pzlacpy.c
@@ -30,11 +30,11 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
-/***************************************************************************//**
+/*******************************************************************************
*
**/
-/***************************************************************************//**
+/*******************************************************************************
*
**/
void morse_pzlacpy(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
@@ -129,7 +129,5 @@ void morse_pzlacpy(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
}
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzlag2c.c b/compute/pzlag2c.c
index ec193214e4aa00e0391f2e4d51f4516499f92a78..5ca24e2b313b97b9e13f3aa78ca9cdee3bb0cb75 100644
--- a/compute/pzlag2c.c
+++ b/compute/pzlag2c.c
@@ -32,11 +32,11 @@
#define B(m,n) B, m, n
#define SA(m,n) SA, m, n
#define SB(m,n) SB, m, n
-/***************************************************************************//**
+/*******************************************************************************
*
**/
-/***************************************************************************//**
+/*******************************************************************************
*
**/
void morse_pclag2z(MORSE_desc_t *SA, MORSE_desc_t *B,
@@ -67,7 +67,5 @@ void morse_pclag2z(MORSE_desc_t *SA, MORSE_desc_t *B,
B(m, n), ldbm);
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, SA );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzlange.c b/compute/pzlange.c
index 83eee09a6c8511743c0e15b7a6118dcc8cc80e78..f893a4d9b45d1d0f04bbc06397b7c469b69fe880 100644
--- a/compute/pzlange.c
+++ b/compute/pzlange.c
@@ -35,15 +35,12 @@
#define VECNORMS_STEP1(m, n) VECNORMS_STEP1, m, n
#define VECNORMS_STEP2(m, n) VECNORMS_STEP2, m, n
#define RESULT(m, n) RESULT, m, n
-/***************************************************************************//**
- *
- **/
-
-/***************************************************************************//**
- *
- **/
-void morse_pzlange(MORSE_enum norm, MORSE_desc_t *A, double *result,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+
+/*****************************************************************************
+ *
+ **/
+void morse_pzlange( MORSE_enum norm, MORSE_desc_t *A, double *result,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_desc_t *VECNORMS_STEP1 = NULL;
MORSE_desc_t *VECNORMS_STEP2 = NULL;
@@ -412,5 +409,4 @@ void morse_pzlange(MORSE_enum norm, MORSE_desc_t *A, double *result,
MORSE_Desc_Destroy( &(RESULT) );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzlanhe.c b/compute/pzlanhe.c
index a637a283d648f92a4c760ccb87d636cd9b41acec..af3b5bbab0bdff3c67545b22427c59730915d31e 100644
--- a/compute/pzlanhe.c
+++ b/compute/pzlanhe.c
@@ -35,11 +35,11 @@
#define VECNORMS_STEP1(m, n) VECNORMS_STEP1, m, n
#define VECNORMS_STEP2(m, n) VECNORMS_STEP2, m, n
#define RESULT(m, n) RESULT, m, n
-/***************************************************************************//**
+/*******************************************************************************
*
**/
-/***************************************************************************//**
+/*******************************************************************************
*
**/
void morse_pzlanhe(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *result,
@@ -456,5 +456,4 @@ void morse_pzlanhe(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *re
}
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzlansy.c b/compute/pzlansy.c
index cdb4c1dcab6d3c545f8d43052e294cc391609e5c..f71eb5cb0f410729e187318be9bad6441ca876b6 100644
--- a/compute/pzlansy.c
+++ b/compute/pzlansy.c
@@ -35,11 +35,11 @@
#define VECNORMS_STEP1(m, n) VECNORMS_STEP1, m, n
#define VECNORMS_STEP2(m, n) VECNORMS_STEP2, m, n
#define RESULT(m, n) RESULT, m, n
-/***************************************************************************//**
+/*******************************************************************************
*
**/
-/***************************************************************************//**
+/*******************************************************************************
*
**/
void morse_pzlansy(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *result,
@@ -465,5 +465,4 @@ void morse_pzlansy(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *re
}
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzlantr.c b/compute/pzlantr.c
index 7f6ded725d09a4913259ba0a65da32d6a3359bc2..7c1e28b71c2c739c5cd13229f35151d544a2c10b 100644
--- a/compute/pzlantr.c
+++ b/compute/pzlantr.c
@@ -32,7 +32,7 @@
#define VECNORMS_STEP1(m, n) VECNORMS_STEP1, m, n
#define VECNORMS_STEP2(m, n) VECNORMS_STEP2, m, n
#define RESULT(m, n) RESULT, m, n
-/***************************************************************************//**
+/*******************************************************************************
*
**/
void morse_pzlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
@@ -665,5 +665,4 @@ void morse_pzlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
}
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzlascal.c b/compute/pzlascal.c
index 2659c3d7e9c43cb131b75a1c59bbd26f10996da5..1809a5a12f15d5ea778a542fd0857d71de31c342 100644
--- a/compute/pzlascal.c
+++ b/compute/pzlascal.c
@@ -97,6 +97,5 @@ void morse_pzlascal(MORSE_enum uplo, MORSE_Complex64_t alpha, MORSE_desc_t *A,
}
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzlaset.c b/compute/pzlaset.c
index a14ba24f0906bd491d0e16ccb6f41e1e42e63857..f0c84a852747e92dc52de55d6de1134b05a33ac2 100644
--- a/compute/pzlaset.c
+++ b/compute/pzlaset.c
@@ -30,7 +30,7 @@
#include "control/common.h"
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel initialization a 2-D array A to BETA on the diagonal and
* ALPHA on the offdiagonals.
**/
@@ -118,6 +118,5 @@ void morse_pzlaset(MORSE_enum uplo,
A(j, j), ldaj);
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzlaset2.c b/compute/pzlaset2.c
index c56661594e293db4fb28ececedaedc9c28e19dc3..96b681245be044ad9aad739b17ccb3ecfe12aa25 100644
--- a/compute/pzlaset2.c
+++ b/compute/pzlaset2.c
@@ -30,7 +30,7 @@
#include "control/common.h"
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel initializztion a 2-D array A to
* ALPHA on the offdiagonals.
**/
@@ -108,6 +108,5 @@ void morse_pzlaset2(MORSE_enum uplo, MORSE_Complex64_t alpha,
}
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzlauum.c b/compute/pzlauum.c
index 9ff23bebea324af581de9af2c7bd96ca5841e9b9..d7052fd75fef5dcf2f07664dfabdac0c7c60a5e8 100644
--- a/compute/pzlauum.c
+++ b/compute/pzlauum.c
@@ -31,7 +31,7 @@
#include "control/common.h"
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel UU' or L'L operation - dynamic scheduling
**/
void morse_pzlauum(MORSE_enum uplo, MORSE_desc_t *A,
@@ -76,7 +76,7 @@ void morse_pzlauum(MORSE_enum uplo, MORSE_desc_t *A,
}
}
for (n = 0; n < k; n++) {
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
MORSE_TASK_ztrmm(
&options,
MorseLeft, uplo, MorseConjTrans, MorseNonUnit,
@@ -84,7 +84,7 @@ void morse_pzlauum(MORSE_enum uplo, MORSE_desc_t *A,
1.0, A(k, k), ldak,
A(k, n), ldak);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
MORSE_TASK_zlauum(
&options,
uplo, tempkm, A->mb,
@@ -121,7 +121,7 @@ void morse_pzlauum(MORSE_enum uplo, MORSE_desc_t *A,
}
for (m = 0; m < k; m++) {
ldam = BLKLDD(A, m);
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
MORSE_TASK_ztrmm(
&options,
MorseRight, uplo, MorseConjTrans, MorseNonUnit,
@@ -129,13 +129,12 @@ void morse_pzlauum(MORSE_enum uplo, MORSE_desc_t *A,
1.0, A(k, k), ldak,
A(m, k), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
MORSE_TASK_zlauum(
&options,
uplo, tempkn, A->mb,
A(k, k), ldak);
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzplghe.c b/compute/pzplghe.c
index d0b0d16e0f4a66f7fc98a6b3a432fc2e22792f69..239e5947ea86d24758900d0bf4999674162a2442 100644
--- a/compute/pzplghe.c
+++ b/compute/pzplghe.c
@@ -32,7 +32,7 @@
#include "control/common.h"
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* morse_pzplghe - Generate a random hermitian (positive definite if 'bump' is large enough) half-matrix by tiles.
**/
void morse_pzplghe( double bump, MORSE_enum uplo, MORSE_desc_t *A,
@@ -97,6 +97,5 @@ void morse_pzplghe( double bump, MORSE_enum uplo, MORSE_desc_t *A,
}
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzplgsy.c b/compute/pzplgsy.c
index 3c8c288bf80a57c30b3e992bcc2857634ab5870d..c5111e5674a0743d0df5e0dcc7170038d50b1304 100644
--- a/compute/pzplgsy.c
+++ b/compute/pzplgsy.c
@@ -32,7 +32,7 @@
#include "control/common.h"
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* morse_pzplgsy - Generate a random symmetric (positive definite if 'bump' is large enough) half-matrix by tiles.
**/
void morse_pzplgsy( MORSE_Complex64_t bump, MORSE_enum uplo, MORSE_desc_t *A,
@@ -97,6 +97,5 @@ void morse_pzplgsy( MORSE_Complex64_t bump, MORSE_enum uplo, MORSE_desc_t *A,
}
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzplrnt.c b/compute/pzplrnt.c
index 1e2fc8c61d8d45778549a08e326e184969e9c6a9..d0467737980d0ad57d6db98805943afddca3294b 100644
--- a/compute/pzplrnt.c
+++ b/compute/pzplrnt.c
@@ -29,7 +29,7 @@
#include "control/common.h"
#define A(m, n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* morse_pzplghe - Generate a random matrix by tiles.
**/
void morse_pzplrnt( MORSE_desc_t *A, unsigned long long int seed,
@@ -60,6 +60,5 @@ void morse_pzplrnt( MORSE_desc_t *A, unsigned long long int seed,
A->m, m*A->mb, n*A->nb, seed );
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzpotrf.c b/compute/pzpotrf.c
index 662c96fd1050671274d6db890c692104720a445b..ae7f858a3e7aa78caba17dfe0084a16d45d2dd04 100644
--- a/compute/pzpotrf.c
+++ b/compute/pzpotrf.c
@@ -32,7 +32,7 @@
#include "control/common.h"
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile Cholesky factorization - dynamic scheduling
**/
void morse_pzpotrf(MORSE_enum uplo, MORSE_desc_t *A,
@@ -84,7 +84,7 @@ void morse_pzpotrf(MORSE_enum uplo, MORSE_desc_t *A,
zone, A(k, k), ldak,
A(m, k), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
@@ -111,7 +111,7 @@ void morse_pzpotrf(MORSE_enum uplo, MORSE_desc_t *A,
A(n, k), ldan,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(n, k) );
+ RUNTIME_data_flush( sequence, A(n, k) );
}
RUNTIME_iteration_pop(morse);
}
@@ -144,7 +144,7 @@ void morse_pzpotrf(MORSE_enum uplo, MORSE_desc_t *A,
zone, A(k, k), ldak,
A(k, n), ldak);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (m = k+1; m < A->mt; m++) {
tempmm = m == A->mt-1 ? A->m - m*A->mb : A->mb;
@@ -170,7 +170,7 @@ void morse_pzpotrf(MORSE_enum uplo, MORSE_desc_t *A,
A(k, n), ldak,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, m) );
+ RUNTIME_data_flush( sequence, A(k, m) );
}
RUNTIME_iteration_pop(morse);
diff --git a/compute/pzpotrimm.c b/compute/pzpotrimm.c
index abbbd1d03bd6c2e6d618c39ccdf42d9d12c03adc..fd4956e21f2f732056034e54999807ad5606cdb2 100644
--- a/compute/pzpotrimm.c
+++ b/compute/pzpotrimm.c
@@ -31,7 +31,7 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
#define C(m,n) C, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile Cholesky factorization - dynamic scheduling
**/
void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *C,
@@ -85,7 +85,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
zone, A(k, k), ldak,
A(m, k), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
@@ -108,7 +108,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
A(n, k), ldan,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(n, k) );
+ RUNTIME_data_flush( sequence, A(n, k) );
}
RUNTIME_iteration_pop(morse);
@@ -143,10 +143,10 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
A(k, n), ldak,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
}
for (n = 0; n < k; n++) {
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
MORSE_TASK_ztrsm(
&options,
MorseLeft, uplo, MorseNoTrans, MorseNonUnit,
@@ -154,7 +154,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
zone, A(k, k), ldak,
A(k, n), ldak);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
MORSE_TASK_ztrtri(
&options,
uplo, MorseNonUnit,
@@ -192,7 +192,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
}
}
for (n = 0; n < k; n++) {
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
MORSE_TASK_ztrmm(
&options,
MorseLeft, uplo, MorseConjTrans, MorseNonUnit,
@@ -200,7 +200,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
1.0, A(k, k), ldak,
A(k, n), ldak);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
MORSE_TASK_zlauum(
&options,
uplo, tempkm, A->mb,
@@ -257,10 +257,10 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
}
}
}
- MORSE_TASK_flush_data( &options, B(m, k) );
+ RUNTIME_data_flush( sequence, B(m, k) );
}
for (n = 0; n <= k; n++) {
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
}
RUNTIME_iteration_pop(morse);
@@ -293,7 +293,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
zone, A(k, k), ldak,
A(k, n), ldak);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (m = k+1; m < A->mt; m++) {
tempmm = m == A->mt-1 ? A->m - m*A->mb : A->mb;
@@ -317,7 +317,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
A(k, n), ldak,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, m) );
+ RUNTIME_data_flush( sequence, A(k, m) );
}
RUNTIME_iteration_pop(morse);
@@ -351,11 +351,11 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
A(k, n), ldak,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
}
for (m = 0; m < k; m++) {
ldam = BLKLDD(A, m);
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
MORSE_TASK_ztrsm(
&options,
MorseRight, uplo, MorseNoTrans, MorseNonUnit,
@@ -363,7 +363,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
zone, A(k, k), ldak,
A(m, k), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
MORSE_TASK_ztrtri(
&options,
uplo, MorseNonUnit,
@@ -403,7 +403,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
}
for (m = 0; m < k; m++) {
ldam = BLKLDD(A, m);
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
MORSE_TASK_ztrmm(
&options,
MorseRight, uplo, MorseConjTrans, MorseNonUnit,
@@ -411,7 +411,7 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
1.0, A(k, k), ldak,
A(m, k), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
MORSE_TASK_zlauum(
&options,
uplo, tempkn, A->mb,
@@ -468,10 +468,10 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
}
}
}
- MORSE_TASK_flush_data( &options, B(m, k) );
+ RUNTIME_data_flush( sequence, B(m, k) );
}
for (m = 0; m <= k; m++) {
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
}
RUNTIME_iteration_pop(morse);
@@ -479,5 +479,4 @@ void morse_pzpotrimm(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_de
}
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzsymm.c b/compute/pzsymm.c
index 71b95895790cd969316790ce7f871b51da7cb46d..fd608f65d9fac24e3967c5933be0c3ce7430d723 100644
--- a/compute/pzsymm.c
+++ b/compute/pzsymm.c
@@ -31,7 +31,7 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
#define C(m,n) C, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile symmetric matrix-matrix multiplication - dynamic scheduling
**/
void morse_pzsymm(MORSE_enum side, MORSE_enum uplo,
@@ -141,16 +141,16 @@ void morse_pzsymm(MORSE_enum side, MORSE_enum uplo,
}
}
}
- MORSE_TASK_flush_data( &options, B(k, n) );
+ RUNTIME_data_flush( sequence, B(k, n) );
}
if (uplo == MorseLower) {
for (n = 0; n <= k; n++) {
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
}
}
else {
for (m = 0; m <= k; m++) {
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
}
}
}
@@ -242,20 +242,19 @@ void morse_pzsymm(MORSE_enum side, MORSE_enum uplo,
}
}
}
- MORSE_TASK_flush_data( &options, B(m, k) );
+ RUNTIME_data_flush( sequence, B(m, k) );
}
if (uplo == MorseLower) {
for (n = 0; n <= k; n++) {
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
}
}
else {
for (m = 0; m <= k; m++) {
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
}
}
}
}
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzsyr2k.c b/compute/pzsyr2k.c
index 68e4d6ee10c6f207e25a2087dd4b301eac21be44..9ed43086c37a4de4de61c6c2f23cf7c3dfd4b991 100644
--- a/compute/pzsyr2k.c
+++ b/compute/pzsyr2k.c
@@ -31,7 +31,7 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
#define C(m,n) C, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile Hermitian rank-k update - dynamic scheduling
**/
void morse_pzsyr2k(MORSE_enum uplo, MORSE_enum trans,
@@ -213,9 +213,5 @@ void morse_pzsyr2k(MORSE_enum uplo, MORSE_enum trans,
}
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, uplo, C );
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
diff --git a/compute/pzsyrk.c b/compute/pzsyrk.c
index 49508895c5aabd53caa555efff8b967757febb7f..a942fa8d475415759fbac0f0918710e50a4eb323 100644
--- a/compute/pzsyrk.c
+++ b/compute/pzsyrk.c
@@ -31,7 +31,7 @@
#define A(m,n) A, m, n
#define C(m,n) C, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile symmetric rank-k update - dynamic scheduling
**/
void morse_pzsyrk(MORSE_enum uplo, MORSE_enum trans,
@@ -172,7 +172,5 @@ void morse_pzsyrk(MORSE_enum uplo, MORSE_enum trans,
}
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, uplo, C );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzsytrf.c b/compute/pzsytrf.c
index 2ea509a512d9ed1afd16da1ad445bd8d496e84da..6e257b5b2080ace71fc02161256600a0645ba53b 100644
--- a/compute/pzsytrf.c
+++ b/compute/pzsytrf.c
@@ -31,7 +31,7 @@
#include "control/common.h"
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile Cholesky factorization - dynamic scheduling
**/
void morse_pzsytrf(MORSE_enum uplo, MORSE_desc_t *A,
@@ -80,7 +80,7 @@ void morse_pzsytrf(MORSE_enum uplo, MORSE_desc_t *A,
zone, A(k, k), ldak,
A(m, k), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
@@ -103,7 +103,7 @@ void morse_pzsytrf(MORSE_enum uplo, MORSE_desc_t *A,
A(n, k), ldan,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(n, k) );
+ RUNTIME_data_flush( sequence, A(n, k) );
}
RUNTIME_iteration_pop(morse);
@@ -133,7 +133,7 @@ void morse_pzsytrf(MORSE_enum uplo, MORSE_desc_t *A,
zone, A(k, k), ldak,
A(k, n), ldak);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (m = k+1; m < A->mt; m++) {
tempmm = m == A->mt-1 ? A->m - m*A->mb : A->mb;
@@ -157,14 +157,13 @@ void morse_pzsytrf(MORSE_enum uplo, MORSE_desc_t *A,
A(k, n), ldak,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, m) );
+ RUNTIME_data_flush( sequence, A(k, m) );
}
RUNTIME_iteration_pop(morse);
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pztile.c b/compute/pztile.c
index c5222001063655c1d9e706b715101c53a759fd32..6b9b49a8233d3ebfbc8a0c967aa13b7bd0c60da8 100644
--- a/compute/pztile.c
+++ b/compute/pztile.c
@@ -72,10 +72,9 @@ void morse_pzlapack_to_tile(MORSE_Complex64_t *Af77, int ldaf77, MORSE_desc_t *A
}
}
+ RUNTIME_desc_flush( &B, sequence );
RUNTIME_sequence_wait( morse, sequence );
RUNTIME_options_finalize( &options, morse );
- MORSE_TASK_flush_all();
- RUNTIME_desc_getoncpu( &B );
RUNTIME_desc_destroy( &B );
}
@@ -122,10 +121,9 @@ void morse_pztile_to_lapack(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int ldaf77
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
+ RUNTIME_desc_flush( &B, sequence );
RUNTIME_sequence_wait( morse, sequence );
RUNTIME_options_finalize( &options, morse );
- RUNTIME_desc_getoncpu( &B );
RUNTIME_desc_destroy( &B );
}
diff --git a/compute/pztile2band.c b/compute/pztile2band.c
index 22b8ca68fea6be025cc4df1e59b58df84c68f785..bfd4f5cf1861e02b1d19c823083fde827abd68de 100644
--- a/compute/pztile2band.c
+++ b/compute/pztile2band.c
@@ -30,7 +30,7 @@
#define A(m,n) A, m, n
#define B(m, n) B, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel copy of a band matrix from full NxN tile storage to band storage (LDABxN).
**/
void morse_pztile2band(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
@@ -118,7 +118,6 @@ void morse_pztile2band(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
}
}
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
}
#undef B
#undef A
diff --git a/compute/pztpgqrt.c b/compute/pztpgqrt.c
index fe4b413eae2c407fe39e804c6d7aaa53cdee9e3c..a070dc85157fe4e1c0623c823617e4c826fe9ce3 100644
--- a/compute/pztpgqrt.c
+++ b/compute/pztpgqrt.c
@@ -36,7 +36,7 @@
#define D(k) V1, k, k
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile QR factorization - dynamic scheduling
**/
void morse_pztpgqrt( int L,
@@ -162,13 +162,6 @@ void morse_pztpgqrt( int L,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, V1 );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, V2 );
- MORSE_TASK_flush_desc( &options, MorseLower, T1 );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, T2 );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, Q1 );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, Q2 );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
(void)D;
diff --git a/compute/pztpqrt.c b/compute/pztpqrt.c
index 6101c0934b551cc1a2d5cba6388ad6e68353e7da..47415ff055e846d4e690709174ec926e5b347028 100644
--- a/compute/pztpqrt.c
+++ b/compute/pztpqrt.c
@@ -28,7 +28,7 @@
#define B(m,n) B, m, n
#define T(m,n) T, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile QR factorization - dynamic scheduling
**/
void morse_pztpqrt( int L, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *T,
@@ -111,9 +111,6 @@ void morse_pztpqrt( int L, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *T,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseLower, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pztradd.c b/compute/pztradd.c
index 8d121b7e9d1b916129c3de29d30803993e552633..aa5dca0e588b2ef31384b276f43f37c8b6712109 100644
--- a/compute/pztradd.c
+++ b/compute/pztradd.c
@@ -30,7 +30,7 @@
#define A(m, n) A, m, n
#define B(m, n) B, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile matrix-matrix multiplication - dynamic scheduling
**/
void morse_pztradd(MORSE_enum uplo, MORSE_enum trans,
@@ -194,7 +194,5 @@ void morse_pztradd(MORSE_enum uplo, MORSE_enum trans,
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, uplo, B );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pztrmm.c b/compute/pztrmm.c
index b9f6cb0063da5bfba8a0dbd20257d0b5fcf53363..b616bd55a478c65aae58e4c5f0bef8ee9bc17f67 100644
--- a/compute/pztrmm.c
+++ b/compute/pztrmm.c
@@ -33,7 +33,7 @@
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile triangular matrix-matrix multiplication - dynamic scheduling
**/
void morse_pztrmm(MORSE_enum side, MORSE_enum uplo,
@@ -311,7 +311,5 @@ void morse_pztrmm(MORSE_enum side, MORSE_enum uplo,
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pztrsm.c b/compute/pztrsm.c
index c861890ccdee3d1e3995cabedf8ddc8d27091737..727b7e5dbeb10cf313792071a2d7300f6c879502 100644
--- a/compute/pztrsm.c
+++ b/compute/pztrsm.c
@@ -32,7 +32,7 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile triangular solve - dynamic scheduling
**/
void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum diag,
@@ -75,7 +75,7 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
lalpha, A(B->mt-1-k, B->mt-1-k), ldak, /* lda * tempkm */
B(B->mt-1-k, n), ldbk); /* ldb * tempnn */
}
- MORSE_TASK_flush_data( &options, A(B->mt-1-k, B->mt-1-k) );
+ RUNTIME_data_flush( sequence, A(B->mt-1-k, B->mt-1-k) );
for (m = k+1; m < B->mt; m++) {
ldam = BLKLDD(A, B->mt-1-m);
ldbm = BLKLDD(B, B->mt-1-m);
@@ -89,10 +89,10 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
B(B->mt-1-k, n ), ldbk,
lalpha, B(B->mt-1-m, n ), ldbm);
}
- MORSE_TASK_flush_data( &options, A(B->mt-1-m, B->mt-1-k) );
+ RUNTIME_data_flush( sequence, A(B->mt-1-m, B->mt-1-k) );
}
for (n = 0; n < B->nt; n++) {
- MORSE_TASK_flush_data( &options, B(B->mt-1-k, n) );
+ RUNTIME_data_flush( sequence, B(B->mt-1-k, n) );
}
}
}
@@ -114,7 +114,7 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
lalpha, A(k, k), ldak,
B(k, n), ldbk);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (m = k+1; m < B->mt; m++) {
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
@@ -128,10 +128,10 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
B(k, n), ldbk,
lalpha, B(m, n), ldbm);
}
- MORSE_TASK_flush_data( &options, A(k, m) );
+ RUNTIME_data_flush( sequence, A(k, m) );
}
for (n = 0; n < B->nt; n++) {
- MORSE_TASK_flush_data( &options, B(k, n) );
+ RUNTIME_data_flush( sequence, B(k, n) );
}
}
@@ -156,7 +156,7 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
lalpha, A(k, k), ldak,
B(k, n), ldbk);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(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);
@@ -171,10 +171,10 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
B(k, n), ldbk,
lalpha, B(m, n), ldbm);
}
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
}
for (n = 0; n < B->nt; n++) {
- MORSE_TASK_flush_data( &options, B(k, n) );
+ RUNTIME_data_flush( sequence, B(k, n) );
}
}
}
@@ -196,7 +196,7 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
lalpha, A(B->mt-1-k, B->mt-1-k), ldak,
B(B->mt-1-k, n), ldbk);
}
- MORSE_TASK_flush_data( &options, A(B->mt-1-k, B->mt-1-k) );
+ RUNTIME_data_flush( sequence, A(B->mt-1-k, B->mt-1-k) );
for (m = k+1; m < B->mt; m++) {
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, B->mt-1-m);
@@ -210,10 +210,10 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
B(B->mt-1-k, n ), ldbk,
lalpha, B(B->mt-1-m, n ), ldbm);
}
- MORSE_TASK_flush_data( &options, A(B->mt-1-k, B->mt-1-m) );
+ RUNTIME_data_flush( sequence, A(B->mt-1-k, B->mt-1-m) );
}
for (n = 0; n < B->nt; n++) {
- MORSE_TASK_flush_data( &options, B(B->mt-1-k, n) );
+ RUNTIME_data_flush( sequence, B(B->mt-1-k, n) );
}
}
}
@@ -239,7 +239,7 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
lalpha, A(k, k), ldak, /* lda * tempkn */
B(m, k), ldbm); /* ldb * tempkn */
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (m = 0; m < B->mt; m++) {
tempmm = m == B->mt-1 ? B->m-m*B->mb : B->mb;
ldbm = BLKLDD(B, m);
@@ -253,10 +253,10 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
A(k, n), ldak, /* lda * tempnn */
lalpha, B(m, n), ldbm); /* ldb * tempnn */
}
- MORSE_TASK_flush_data( &options, B(m, k) );
+ RUNTIME_data_flush( sequence, B(m, k) );
}
for (n = k+1; n < B->nt; n++) {
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
}
}
}
@@ -276,7 +276,7 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
tempmm, tempkn, A->mb,
alpha, A(B->nt-1-k, B->nt-1-k), ldak, /* lda * tempkn */
B( m, B->nt-1-k), ldbm); /* ldb * tempkn */
- MORSE_TASK_flush_data( &options, A(B->nt-1-k, B->nt-1-k) );
+ RUNTIME_data_flush( sequence, A(B->nt-1-k, B->nt-1-k) );
for (n = k+1; n < B->nt; n++) {
ldan = BLKLDD(A, B->nt-1-n);
@@ -288,10 +288,10 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
A(B->nt-1-n, B->nt-1-k), ldan, /* A->mb * tempkn (Never last row) */
zone, B(m, B->nt-1-n), ldbm); /* ldb * B->nb */
}
- MORSE_TASK_flush_data( &options, B(m, B->nt-1-k) );
+ RUNTIME_data_flush( sequence, B(m, B->nt-1-k) );
}
for (n = k+1; n < B->nt; n++) {
- MORSE_TASK_flush_data( &options, A(B->nt-1-n, B->nt-1-k) );
+ RUNTIME_data_flush( sequence, A(B->nt-1-n, B->nt-1-k) );
}
}
}
@@ -314,7 +314,7 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
tempmm, tempkn, A->mb,
lalpha, A(B->nt-1-k, B->nt-1-k), ldak, /* lda * tempkn */
B( m, B->nt-1-k), ldbm); /* ldb * tempkn */
- MORSE_TASK_flush_data( &options, A(B->nt-1-k, B->nt-1-k) );
+ RUNTIME_data_flush( sequence, A(B->nt-1-k, B->nt-1-k) );
for (n = k+1; n < B->nt; n++) {
MORSE_TASK_zgemm(
@@ -325,10 +325,10 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
A(B->nt-1-k, B->nt-1-n), ldak, /* lda * B->nb */
lalpha, B(m, B->nt-1-n), ldbm); /* ldb * B->nb */
}
- MORSE_TASK_flush_data( &options, B(m, B->nt-1-k) );
+ RUNTIME_data_flush( sequence, B(m, B->nt-1-k) );
}
for (n = k+1; n < B->nt; n++) {
- MORSE_TASK_flush_data( &options, A(B->nt-1-k, B->nt-1-n) );
+ RUNTIME_data_flush( sequence, A(B->nt-1-k, B->nt-1-n) );
}
}
}
@@ -348,7 +348,7 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
tempmm, tempkn, A->mb,
alpha, A(k, k), ldak, /* lda * tempkn */
B(m, k), ldbm); /* ldb * tempkn */
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
for (n = k+1; n < B->nt; n++) {
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
@@ -361,10 +361,10 @@ void morse_pztrsm(MORSE_enum side, MORSE_enum uplo, MORSE_enum trans, MORSE_enum
A(n, k), ldan, /* ldan * tempkn */
zone, B(m, n), ldbm); /* ldb * tempnn */
}
- MORSE_TASK_flush_data( &options, B(m, k) );
+ RUNTIME_data_flush( sequence, B(m, k) );
}
for (n = k+1; n < B->nt; n++) {
- MORSE_TASK_flush_data( &options, A(n, k) );
+ RUNTIME_data_flush( sequence, A(n, k) );
}
}
diff --git a/compute/pztrsmpl.c b/compute/pztrsmpl.c
index 4c51e7a5eef96cee0880901ce9277b8a9459461b..7314ad60fe731751f16e02b72e1987b5a2f78191 100644
--- a/compute/pztrsmpl.c
+++ b/compute/pztrsmpl.c
@@ -34,7 +34,7 @@
#define B(m,n) B, m, n
#define L(m,n) L, m, n
#define IPIV(m,n) &(IPIV[(int64_t)A->nb*((int64_t)(m)+(int64_t)A->mt*(int64_t)(n))])
-/***************************************************************************//**
+/*******************************************************************************
* Parallel forward substitution for tile LU - dynamic scheduling
**/
void morse_pztrsmpl(MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *L, int *IPIV,
@@ -87,7 +87,5 @@ void morse_pztrsmpl(MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *L, int *IPIV
}
}
}
- MORSE_TASK_flush_desc( &options, MorseLower, L );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pztrtri.c b/compute/pztrtri.c
index fbea4fe03db5e43c968f9bbc90cc767520b185c8..e212d3805516147fb1067991dabfbd22961cff07 100644
--- a/compute/pztrtri.c
+++ b/compute/pztrtri.c
@@ -31,7 +31,7 @@
#include "control/common.h"
#define A(m,n) A, m, n
-/***************************************************************************//**
+/*******************************************************************************
* Parallel tile triangular matrix inverse - dynamic scheduling
**/
void morse_pztrtri(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
@@ -82,10 +82,10 @@ void morse_pztrtri(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
A(k, n), ldak,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
}
for (n = 0; n < k; n++) {
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
MORSE_TASK_ztrsm(
&options,
MorseLeft, uplo, MorseNoTrans, diag,
@@ -93,7 +93,7 @@ void morse_pztrtri(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
zone, A(k, k), ldak,
A(k, n), ldak);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
MORSE_TASK_ztrtri(
&options,
uplo, diag,
@@ -133,11 +133,11 @@ void morse_pztrtri(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
A(k, n), ldak,
zone, A(m, n), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, n) );
+ RUNTIME_data_flush( sequence, A(k, n) );
}
for (m = 0; m < k; m++) {
ldam = BLKLDD(A, m);
- MORSE_TASK_flush_data( &options, A(m, k) );
+ RUNTIME_data_flush( sequence, A(m, k) );
MORSE_TASK_ztrsm(
&options,
MorseRight, uplo, MorseNoTrans, diag,
@@ -145,7 +145,7 @@ void morse_pztrtri(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
zone, A(k, k), ldak,
A(m, k), ldam);
}
- MORSE_TASK_flush_data( &options, A(k, k) );
+ RUNTIME_data_flush( sequence, A(k, k) );
MORSE_TASK_ztrtri(
&options,
uplo, diag,
@@ -155,6 +155,5 @@ void morse_pztrtri(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
RUNTIME_iteration_pop(morse);
}
}
- MORSE_TASK_flush_desc( &options, uplo, A );
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzunglq.c b/compute/pzunglq.c
index d764b3a71b19b785f06eaecad117c8827eafc467..13a034ce3d64ecf8f1c5f1abf47e41a09e494141 100644
--- a/compute/pzunglq.c
+++ b/compute/pzunglq.c
@@ -39,7 +39,7 @@
#define D(k) A, k, k
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
**/
void morse_pzunglq(MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T, MORSE_desc_t *D,
@@ -142,10 +142,6 @@ void morse_pzunglq(MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T, MORSE_desc
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, Q );
- MORSE_TASK_flush_desc( &options, MorseUpper, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
(void)D;
diff --git a/compute/pzunglq_param.c b/compute/pzunglq_param.c
index 6da5c30702d06b7db6a3d7cf928a953ddbeab1df..adc3512ff57fa88d9b79b1272eddee14cdad87c4 100644
--- a/compute/pzunglq_param.c
+++ b/compute/pzunglq_param.c
@@ -169,14 +169,8 @@ void morse_pzunglq_param(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_des
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, Q );
- MORSE_TASK_flush_desc( &options, MorseUpper, TS );
- MORSE_TASK_flush_desc( &options, MorseUpper, TT );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
free(tiles);
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzunglqrh.c b/compute/pzunglqrh.c
index bd76fe5be8ce970c6aa6a202a8b21ce0b368adec..72836940f2ef8f4e87a7b218357e2f2b17e9fff2 100644
--- a/compute/pzunglqrh.c
+++ b/compute/pzunglqrh.c
@@ -164,10 +164,6 @@ void morse_pzunglqrh(MORSE_desc_t *A, MORSE_desc_t *Q,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, Q );
- MORSE_TASK_flush_desc( &options, MorseUpper, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
(void)D;
diff --git a/compute/pzungqr.c b/compute/pzungqr.c
index c2e0829d166aca1856a11a8a7789b1e6f022dbaa..389f79e854029cefd083ca19ea681d9838c8b231 100644
--- a/compute/pzungqr.c
+++ b/compute/pzungqr.c
@@ -39,7 +39,7 @@
#define D(k) A, k, k
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
**/
void morse_pzungqr(MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T, MORSE_desc_t *D,
@@ -142,10 +142,6 @@ void morse_pzungqr(MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T, MORSE_desc
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, Q );
- MORSE_TASK_flush_desc( &options, MorseLower, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
(void)D;
diff --git a/compute/pzungqr_param.c b/compute/pzungqr_param.c
index 13b1be1c6a3f910b1062cbeafcf019c7bb219403..a0849b634ee1679ed5ea6c32fb3a486a9dd9bfff 100644
--- a/compute/pzungqr_param.c
+++ b/compute/pzungqr_param.c
@@ -175,14 +175,8 @@ void morse_pzungqr_param(const libhqr_tree_t *qrtree,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, Q );
- MORSE_TASK_flush_desc( &options, MorseLower, TS );
- MORSE_TASK_flush_desc( &options, MorseLower, TT );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
free(tiles);
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzungqrrh.c b/compute/pzungqrrh.c
index c383858d2313afdb47ad6a4820752542b95d50d1..1d0516298a0ec3dbbcfd1aa783b29e2a0c697e63 100644
--- a/compute/pzungqrrh.c
+++ b/compute/pzungqrrh.c
@@ -169,12 +169,7 @@ void morse_pzungqrrh(MORSE_desc_t *A, MORSE_desc_t *Q,
RUNTIME_iteration_pop(morse);
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, Q );
- MORSE_TASK_flush_desc( &options, MorseLower, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzunmlq.c b/compute/pzunmlq.c
index c6547f3dc0720bc747d7cdccc2b48afc4035cdf3..9c9cfd679c8532eb2f94d5d5fba7b25b04d6ff71 100644
--- a/compute/pzunmlq.c
+++ b/compute/pzunmlq.c
@@ -40,7 +40,7 @@
#define D(k) A, k, k
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel application of Q using tile V - LQ factorization - dynamic scheduling
**/
void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
@@ -313,10 +313,6 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, MorseUpper, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
(void)D;
diff --git a/compute/pzunmlq_param.c b/compute/pzunmlq_param.c
index 0b5ef65f8c77cd1fddb726cf2ec672d060c48d9f..5ad102180aa5557f3bc9083af56074806dec95a5 100644
--- a/compute/pzunmlq_param.c
+++ b/compute/pzunmlq_param.c
@@ -432,15 +432,9 @@ void morse_pzunmlq_param(const libhqr_tree_t *qrtree,
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, MorseUpper, TS );
- MORSE_TASK_flush_desc( &options, MorseUpper, TT );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
free(tiles);
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzunmlqrh.c b/compute/pzunmlqrh.c
index ac288111c5c74ee1c8e4b52a14790e2eb7df2a32..8444f0bb3899178caf82269875e4f725130c0bc9 100644
--- a/compute/pzunmlqrh.c
+++ b/compute/pzunmlqrh.c
@@ -41,7 +41,7 @@
#define D(m,n) A, (m), (n)
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel application of Q using tile V - LQ factorization (reduction
* Householder) - dynamic scheduling
**/
@@ -408,12 +408,7 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, MorseUpper, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzunmqr.c b/compute/pzunmqr.c
index 86dc63bd780ea02071078ebe6aa836c70645a169..4e81716ddf0439cde04ed3465422ebefec0a860b 100644
--- a/compute/pzunmqr.c
+++ b/compute/pzunmqr.c
@@ -40,7 +40,7 @@
#define D(k) A, k, k
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel application of Q using tile V - QR factorization - dynamic scheduling
**/
void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
@@ -318,10 +318,6 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, MorseLower, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
(void)D;
diff --git a/compute/pzunmqr_param.c b/compute/pzunmqr_param.c
index 5f7b20032020d22c7dea3bdfef1b78a74ccca9dd..ae9b46d5e9755f0cd03c46e367dadb52d0b8084b 100644
--- a/compute/pzunmqr_param.c
+++ b/compute/pzunmqr_param.c
@@ -431,15 +431,9 @@ void morse_pzunmqr_param(const libhqr_tree_t *qrtree,
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, MorseLower, TS );
- MORSE_TASK_flush_desc( &options, MorseLower, TT );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
free(tiles);
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- MORSE_TASK_flush_all();
(void)D;
}
diff --git a/compute/pzunmqrrh.c b/compute/pzunmqrrh.c
index 80f9879db09d4fe3162794f99b11b23a445ed599..ca7097ca99f51439bf2ec33627a36812fd01ead5 100644
--- a/compute/pzunmqrrh.c
+++ b/compute/pzunmqrrh.c
@@ -41,13 +41,13 @@
#define D(m,n) A, (m), (n)
#endif
-/***************************************************************************//**
+/*******************************************************************************
* Parallel application of Q using tile V - QR factorization (reduction
* Householder) - dynamic scheduling
**/
-void morse_pzunmqrrh(MORSE_enum side, MORSE_enum trans,
- MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *T, MORSE_desc_t *D, int BS,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+void morse_pzunmqrrh( MORSE_enum side, MORSE_enum trans,
+ MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *T, MORSE_desc_t *D, int BS,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_option_t options;
@@ -413,10 +413,6 @@ void morse_pzunmqrrh(MORSE_enum side, MORSE_enum trans,
}
}
- MORSE_TASK_flush_desc( &options, MorseUpperLower, A );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, B );
- MORSE_TASK_flush_desc( &options, MorseLower, T );
- MORSE_TASK_flush_desc( &options, MorseUpperLower, D );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
(void)D;
diff --git a/compute/zbuild.c b/compute/zbuild.c
index 8262d017bb1ebd8aad84d98a41395c034688f3b2..71fadaac9434b5d8357a9bddac6360fb0cfefe02 100644
--- a/compute/zbuild.c
+++ b/compute/zbuild.c
@@ -5,10 +5,6 @@
* All rights reserved.
* @copyright (c) 2012-2014 Inria. All rights reserved.
* @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
- *
- **/
-
-/**
*
* @file zbuild.c
*
@@ -29,7 +25,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -89,61 +86,62 @@ int MORSE_zbuild( MORSE_enum uplo, int M, int N,
MORSE_Complex64_t *A, int LDA,
void *user_data, void* user_build_callback)
{
- int NB;
- int status;
- MORSE_context_t *morse;
- MORSE_sequence_t *sequence = NULL;
- MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ int NB;
+ int status;
+ MORSE_context_t *morse;
+ MORSE_sequence_t *sequence = NULL;
+ MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ MORSE_desc_t descA;
- morse = morse_context_self();
- if (morse == NULL) {
- morse_fatal_error("MORSE_zbuild", "MORSE not initialized");
- return MORSE_ERR_NOT_INITIALIZED;
- }
- /* Check input arguments */
- if (M < 0) {
- morse_error("MORSE_zbuild", "illegal value of M");
- return -1;
- }
- if (N < 0) {
- morse_error("MORSE_zbuild", "illegal value of N");
- return -2;
- }
- if (LDA < chameleon_max(1, M)) {
- morse_error("MORSE_zbuild", "illegal value of LDA");
- return -4;
- }
- /* Quick return */
- if (chameleon_min(M, N) == 0)
- return MORSE_SUCCESS;
+ morse = morse_context_self();
+ if (morse == NULL) {
+ morse_fatal_error("MORSE_zbuild", "MORSE not initialized");
+ return MORSE_ERR_NOT_INITIALIZED;
+ }
+ /* Check input arguments */
+ if (M < 0) {
+ morse_error("MORSE_zbuild", "illegal value of M");
+ return -1;
+ }
+ if (N < 0) {
+ morse_error("MORSE_zbuild", "illegal value of N");
+ return -2;
+ }
+ if (LDA < chameleon_max(1, M)) {
+ morse_error("MORSE_zbuild", "illegal value of LDA");
+ return -4;
+ }
+ /* Quick return */
+ if (chameleon_min(M, N) == 0)
+ return MORSE_SUCCESS;
- /* Tune NB depending on M, N & NRHS; Set NBNB */
- status = morse_tune(MORSE_FUNC_ZGEMM, M, N, 0);
- if (status != MORSE_SUCCESS) {
- morse_error("MORSE_zbuild", "morse_tune() failed");
- return status;
- }
+ /* Tune NB depending on M, N & NRHS; Set NBNB */
+ status = morse_tune(MORSE_FUNC_ZGEMM, M, N, 0);
+ if (status != MORSE_SUCCESS) {
+ morse_error("MORSE_zbuild", "morse_tune() failed");
+ return status;
+ }
- /* Set NT */
- NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
- morse_zdesc_alloc(descA, NB, NB, LDA, N, 0, 0, M, N, morse_desc_mat_free(&descA));
+ /* Set NT */
+ NB = MORSE_NB;
+ morse_sequence_create(morse, &sequence);
+ morse_zdesc_alloc(descA, NB, NB, LDA, N, 0, 0, M, N, morse_desc_mat_free(&descA));
- /* Call the tile interface */
- MORSE_zbuild_Tile_Async(uplo, &descA, user_data, user_build_callback, sequence, &request );
+ /* Call the tile interface */
+ MORSE_zbuild_Tile_Async(uplo, &descA, user_data, user_build_callback, sequence, &request );
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
+ morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
+ morse_sequence_wait(morse, sequence);
+ morse_desc_mat_free(&descA);
- status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ status = sequence->status;
+ morse_sequence_destroy(morse, sequence);
- return status;
+ return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -187,25 +185,26 @@ int MORSE_zbuild( MORSE_enum uplo, int M, int N,
int MORSE_zbuild_Tile( MORSE_enum uplo, MORSE_desc_t *A,
void *user_data, void* user_build_callback )
{
- MORSE_context_t *morse;
- MORSE_sequence_t *sequence = NULL;
- MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- int status;
+ MORSE_context_t *morse;
+ MORSE_sequence_t *sequence = NULL;
+ MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ int status;
- morse = morse_context_self();
- if (morse == NULL) {
- morse_fatal_error("MORSE_zbuild_Tile", "MORSE not initialized");
- return MORSE_ERR_NOT_INITIALIZED;
- }
- morse_sequence_create(morse, &sequence);
- MORSE_zbuild_Tile_Async( uplo, A, user_data, user_build_callback, sequence, &request );
- morse_sequence_wait(morse, sequence);
- status = sequence->status;
- morse_sequence_destroy(morse, sequence);
- return status;
+ morse = morse_context_self();
+ if (morse == NULL) {
+ morse_fatal_error("MORSE_zbuild_Tile", "MORSE not initialized");
+ return MORSE_ERR_NOT_INITIALIZED;
+ }
+ morse_sequence_create(morse, &sequence);
+ MORSE_zbuild_Tile_Async( uplo, A, user_data, user_build_callback, sequence, &request );
+ morse_sequence_wait(morse, sequence);
+ status = sequence->status;
+ morse_sequence_destroy(morse, sequence);
+ return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
@@ -252,38 +251,38 @@ int MORSE_zbuild_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A,
MORSE_sequence_t *sequence,
MORSE_request_t *request)
{
- MORSE_context_t *morse;
+ MORSE_context_t *morse;
- morse = morse_context_self();
- if (morse == NULL) {
- morse_fatal_error("MORSE_zbuild_Tile", "MORSE not initialized");
- return MORSE_ERR_NOT_INITIALIZED;
- }
- if (sequence == NULL) {
- morse_fatal_error("MORSE_zbuild_Tile", "NULL sequence");
- return MORSE_ERR_UNALLOCATED;
- }
- if (request == NULL) {
- morse_fatal_error("MORSE_zbuild_Tile", "NULL request");
- return MORSE_ERR_UNALLOCATED;
- }
- /* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
- request->status = MORSE_SUCCESS;
- else
- return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ morse = morse_context_self();
+ if (morse == NULL) {
+ morse_fatal_error("MORSE_zbuild_Tile", "MORSE not initialized");
+ return MORSE_ERR_NOT_INITIALIZED;
+ }
+ if (sequence == NULL) {
+ morse_fatal_error("MORSE_zbuild_Tile", "NULL sequence");
+ return MORSE_ERR_UNALLOCATED;
+ }
+ if (request == NULL) {
+ morse_fatal_error("MORSE_zbuild_Tile", "NULL request");
+ return MORSE_ERR_UNALLOCATED;
+ }
+ /* Check sequence status */
+ if (sequence->status == MORSE_SUCCESS)
+ request->status = MORSE_SUCCESS;
+ else
+ return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
- /* Check descriptors for correctness */
- if (morse_desc_check(A) != MORSE_SUCCESS) {
- morse_error("MORSE_zbuild_Tile", "invalid descriptor");
- return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
- }
+ /* Check descriptors for correctness */
+ if (morse_desc_check(A) != MORSE_SUCCESS) {
+ morse_error("MORSE_zbuild_Tile", "invalid descriptor");
+ return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
+ }
- /* Quick return */
- if (chameleon_min( A->m, A->n ) == 0)
- return MORSE_SUCCESS;
+ /* Quick return */
+ if (chameleon_min( A->m, A->n ) == 0)
+ return MORSE_SUCCESS;
- morse_pzbuild(uplo, A, user_data, user_build_callback, sequence, request);
+ morse_pzbuild(uplo, A, user_data, user_build_callback, sequence, request);
- return MORSE_SUCCESS;
+ return MORSE_SUCCESS;
}
diff --git a/compute/zgeadd.c b/compute/zgeadd.c
index 9aea3d4c73bed941982ed456415e7b093a6da422..b6e85b32297b0b56bbd9b3e0d9aaa6a46230e657 100644
--- a/compute/zgeadd.c
+++ b/compute/zgeadd.c
@@ -26,7 +26,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -184,7 +185,8 @@ int MORSE_zgeadd(MORSE_enum trans, int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -247,16 +249,17 @@ int MORSE_zgeadd_Tile(MORSE_enum trans,
}
morse_sequence_create(morse, &sequence);
MORSE_zgeadd_Tile_Async(trans, alpha, A, beta, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgelqf.c b/compute/zgelqf.c
index 6fb6606b14c249906af6c883978856eb4afd65ca..55b9284837f7af7fd60f17702fcba108942c4d1b 100644
--- a/compute/zgelqf.c
+++ b/compute/zgelqf.c
@@ -30,7 +30,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -146,7 +147,8 @@ int MORSE_zgelqf(int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -198,15 +200,16 @@ int MORSE_zgelqf_Tile(MORSE_desc_t *A, MORSE_desc_t *T)
}
morse_sequence_create(morse, &sequence);
MORSE_zgelqf_Tile_Async(A, T, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgelqf_param.c b/compute/zgelqf_param.c
index 2e46686816f4d5423371b8a120df9cce725c110a..d650ba8ad5f3b180d22f0011dd334ce27b3b55b8 100644
--- a/compute/zgelqf_param.c
+++ b/compute/zgelqf_param.c
@@ -197,8 +197,8 @@ int MORSE_zgelqf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zgelqf_param_Tile_Async(qrtree, A, TS, TT, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zgelqs.c b/compute/zgelqs.c
index 23cdd4d022d705afab6a683d282ecce90f8f3cc5..b9410fd3d107681ccdbeac7bd54ed92537ba0580 100644
--- a/compute/zgelqs.c
+++ b/compute/zgelqs.c
@@ -222,9 +222,9 @@ int MORSE_zgelqs_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B)
}
morse_sequence_create(morse, &sequence);
MORSE_zgelqs_Tile_Async(A, T, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zgelqs_param.c b/compute/zgelqs_param.c
index cf2e853634609914c3c960a3d7e8b5508e40a55e..ac5c26520b0d471eabea1a527630959d135dc55a 100644
--- a/compute/zgelqs_param.c
+++ b/compute/zgelqs_param.c
@@ -227,9 +227,9 @@ int MORSE_zgelqs_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zgelqs_param_Tile_Async(qrtree, A, TS, TT, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zgels.c b/compute/zgels.c
index f80cc4ef65b4268e41b92453a924553020d5b610..4ab15cb6d82b1b12fb3cdc18df73a7f04abcb674 100644
--- a/compute/zgels.c
+++ b/compute/zgels.c
@@ -217,7 +217,8 @@ int MORSE_zgels(MORSE_enum trans, int M, int N, int NRHS,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -285,16 +286,17 @@ int MORSE_zgels_Tile(MORSE_enum trans, MORSE_desc_t *A,
}
morse_sequence_create(morse, &sequence);
MORSE_zgels_Tile_Async(trans, A, T, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgels_param.c b/compute/zgels_param.c
index d862c63267b877f276740af544336a90087503b5..9e91871be2973d89a78be91aa3c2e788074d757f 100644
--- a/compute/zgels_param.c
+++ b/compute/zgels_param.c
@@ -291,9 +291,9 @@ int MORSE_zgels_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum trans, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zgels_param_Tile_Async(qrtree, trans, A, TS, TT, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zgemm.c b/compute/zgemm.c
index 0f931d8dc4636d67112d7e452b522738c789731b..332b540061849362a78b9c4dc8d893e39ae56bbb 100644
--- a/compute/zgemm.c
+++ b/compute/zgemm.c
@@ -50,7 +50,8 @@
*
*/
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -247,7 +248,8 @@ int MORSE_zgemm(MORSE_enum transA, MORSE_enum transB, int M, int N, int K,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -319,17 +321,18 @@ int MORSE_zgemm_Tile(MORSE_enum transA, MORSE_enum transB,
}
morse_sequence_create(morse, &sequence);
MORSE_zgemm_Tile_Async(transA, transB, alpha, A, B, beta, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgeqrf.c b/compute/zgeqrf.c
index de992ffa755e19cac65af0dcb5cc62332ea68a4d..5e500472b0dd8c3ed530b94c9e230b7833bd8172 100644
--- a/compute/zgeqrf.c
+++ b/compute/zgeqrf.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -145,7 +146,8 @@ int MORSE_zgeqrf(int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -197,8 +199,8 @@ int MORSE_zgeqrf_Tile(MORSE_desc_t *A, MORSE_desc_t *T)
}
morse_sequence_create(morse, &sequence);
MORSE_zgeqrf_Tile_Async(A, T, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zgeqrf_param.c b/compute/zgeqrf_param.c
index 8e84f2a2bc421df67aa4e4e1f4695b53cf2f29e1..a67f508db78185e1264e6935a0cfa379df024c1c 100644
--- a/compute/zgeqrf_param.c
+++ b/compute/zgeqrf_param.c
@@ -206,8 +206,8 @@ int MORSE_zgeqrf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zgeqrf_param_Tile_Async(qrtree, A, TS, TT, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zgeqrs.c b/compute/zgeqrs.c
index 32e618550fb35324d019990f64afa0a9eeca51fd..29b962f4f6c5f20483682f54eff2822ee9402edc 100644
--- a/compute/zgeqrs.c
+++ b/compute/zgeqrs.c
@@ -30,7 +30,8 @@
#include "control/common.h"
#include <stdlib.h>
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -168,7 +169,8 @@ int MORSE_zgeqrs(int M, int N, int NRHS,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -219,17 +221,18 @@ int MORSE_zgeqrs_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B)
}
morse_sequence_create(morse, &sequence);
MORSE_zgeqrs_Tile_Async(A, T, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( T, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(T);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgeqrs_param.c b/compute/zgeqrs_param.c
index 611b136fa92b6e770d903401fc53ff13104c06de..6e6cb28b394e3b1d815757968cd6dcbf5ae56ae6 100644
--- a/compute/zgeqrs_param.c
+++ b/compute/zgeqrs_param.c
@@ -216,11 +216,11 @@ int MORSE_zgeqrs_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zgeqrs_param_Tile_Async(qrtree, A, TS, TT, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( TS, sequence );
+ RUNTIME_desc_flush( TT, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(TS);
- RUNTIME_desc_getoncpu(TT);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zgesv_incpiv.c b/compute/zgesv_incpiv.c
index 02b1129c26bc3254c380fe70da57de57ae03add3..ca0ca1e6a74dfa96cce66e4616ebd1be54f4af4c 100644
--- a/compute/zgesv_incpiv.c
+++ b/compute/zgesv_incpiv.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -166,7 +167,8 @@ int MORSE_zgesv_incpiv(int N, int NRHS,
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -224,16 +226,17 @@ int MORSE_zgesv_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_d
}
morse_sequence_create(morse, &sequence);
MORSE_zgesv_incpiv_Tile_Async(A, L, IPIV, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgesv_nopiv.c b/compute/zgesv_nopiv.c
index 31735f19c856c8c77c8b03a6df813cd41d5e9310..bc0d9bdf9d3fd297954c2a6cc14d692c59002705 100644
--- a/compute/zgesv_nopiv.c
+++ b/compute/zgesv_nopiv.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -165,7 +166,8 @@ int MORSE_zgesv_nopiv(int N, int NRHS,
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -216,16 +218,17 @@ int MORSE_zgesv_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
}
morse_sequence_create(morse, &sequence);
MORSE_zgesv_nopiv_Tile_Async(A, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgesvd.c b/compute/zgesvd.c
index b2d7e48d430c43bfef24fbde3dbd6a404bc10620..968c0a01b427303b55a50c8574231d407222a849 100644
--- a/compute/zgesvd.c
+++ b/compute/zgesvd.c
@@ -31,7 +31,8 @@
#include <coreblas/lapacke.h>
#endif
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -238,7 +239,8 @@ int MORSE_zgesvd(MORSE_enum jobu, MORSE_enum jobvt,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -362,7 +364,8 @@ int MORSE_zgesvd_Tile(MORSE_enum jobu, MORSE_enum jobvt,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgetrf_incpiv.c b/compute/zgetrf_incpiv.c
index 169fb67f05fd55e87abbdbe6e377d85fb1f48237..aaf96f147bfdf88424c2200fd29eb645e005aa47 100644
--- a/compute/zgetrf_incpiv.c
+++ b/compute/zgetrf_incpiv.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -147,7 +148,8 @@ int MORSE_zgetrf_incpiv(int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -202,15 +204,16 @@ int MORSE_zgetrf_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV)
}
morse_sequence_create(morse, &sequence);
MORSE_zgetrf_incpiv_Tile_Async(A, L, IPIV, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgetrf_nopiv.c b/compute/zgetrf_nopiv.c
index ea987f595f701c29cf171e573311d2782f66b990..5e56c4a8b6bd8dffd399385ebc61bcf2ac52ea36 100644
--- a/compute/zgetrf_nopiv.c
+++ b/compute/zgetrf_nopiv.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -143,7 +144,8 @@ int MORSE_zgetrf_nopiv(int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -190,15 +192,16 @@ int MORSE_zgetrf_nopiv_Tile(MORSE_desc_t *A)
}
morse_sequence_create(morse, &sequence);
MORSE_zgetrf_nopiv_Tile_Async(A, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgetrs_incpiv.c b/compute/zgetrs_incpiv.c
index 63f9564a2fca9fa0a01d874abd33874343dfc420..195b0895d444eb00ca8e7b5ad70bff8bd06a5800 100644
--- a/compute/zgetrs_incpiv.c
+++ b/compute/zgetrs_incpiv.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -172,7 +173,8 @@ int MORSE_zgetrs_incpiv(MORSE_enum trans, int N, int NRHS,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -227,16 +229,17 @@ int MORSE_zgetrs_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zgetrs_incpiv_Tile_Async(A, L, IPIV, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zgetrs_nopiv.c b/compute/zgetrs_nopiv.c
index 02b029146f903eba98da6b6b0a4722102a57046d..f92c66abf05da77ee81aa117a5ac2f1d21411e03 100644
--- a/compute/zgetrs_nopiv.c
+++ b/compute/zgetrs_nopiv.c
@@ -30,7 +30,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -166,7 +167,8 @@ int MORSE_zgetrs_nopiv(MORSE_enum trans, int N, int NRHS,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -215,16 +217,17 @@ int MORSE_zgetrs_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
}
morse_sequence_create(morse, &sequence);
MORSE_zgetrs_nopiv_Tile_Async(A, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zheevd.c b/compute/zheevd.c
index e4a85b98b380b7217225252031dd29b7cb4a21c5..8094b590b35474f80f0740ba8a04389eeb557b6b 100644
--- a/compute/zheevd.c
+++ b/compute/zheevd.c
@@ -30,7 +30,8 @@
#include <coreblas/lapacke.h>
#endif
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -175,7 +176,8 @@ int MORSE_zheevd(MORSE_enum jobz, MORSE_enum uplo, int N,
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -257,17 +259,17 @@ int MORSE_zheevd_Tile(MORSE_enum jobz, MORSE_enum uplo,
}
morse_sequence_create(morse, &sequence);
MORSE_zheevd_Tile_Async(jobz, uplo, A, W, T, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( T, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(T);
-
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zhemm.c b/compute/zhemm.c
index a9581e25a62a77ca307ee2250b95384aebfd9bcb..5b6b31852f6335f10dde0a428acb5988abb6338d 100644
--- a/compute/zhemm.c
+++ b/compute/zhemm.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -210,7 +211,8 @@ int MORSE_zhemm(MORSE_enum side, MORSE_enum uplo, int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -284,17 +286,18 @@ int MORSE_zhemm_Tile(MORSE_enum side, MORSE_enum uplo,
}
morse_sequence_create(morse, &sequence);
MORSE_zhemm_Tile_Async(side, uplo, alpha, A, B, beta, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zher2k.c b/compute/zher2k.c
index 6dd8448041eb2c5bdc1de9aec24e3fadb65acc2c..e8f2e6d892b9bf8be8b7c8f3fff666afce446ece 100644
--- a/compute/zher2k.c
+++ b/compute/zher2k.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -214,7 +215,8 @@ int MORSE_zher2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -283,17 +285,18 @@ int MORSE_zher2k_Tile(MORSE_enum uplo, MORSE_enum trans,
}
morse_sequence_create(morse, &sequence);
MORSE_zher2k_Tile_Async(uplo, trans, alpha, A, B, beta, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zherk.c b/compute/zherk.c
index b52cdff217dd0e6297c40ab209bf58241614b1cb..6fbe72d552b780604b3b3026e7c68e3cb7452227 100644
--- a/compute/zherk.c
+++ b/compute/zherk.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -194,7 +195,8 @@ int MORSE_zherk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -259,16 +261,17 @@ int MORSE_zherk_Tile(MORSE_enum uplo, MORSE_enum trans,
}
morse_sequence_create(morse, &sequence);
MORSE_zherk_Tile_Async(uplo, trans, alpha, A, beta, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zhetrd.c b/compute/zhetrd.c
index e3390e8fb0f56e45a2c21e35288bbefc2ed79d94..e69bce8a9eb5f25e76f796a2cd2857f65df93daf 100644
--- a/compute/zhetrd.c
+++ b/compute/zhetrd.c
@@ -30,7 +30,8 @@
#include <coreblas/lapacke.h>
#endif
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -189,7 +190,8 @@ int MORSE_zhetrd(MORSE_enum jobz, MORSE_enum uplo, int N,
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -291,7 +293,8 @@ int MORSE_zhetrd_Tile(MORSE_enum jobz, MORSE_enum uplo,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zlacpy.c b/compute/zlacpy.c
index c19edef30700232549a7caa946cc7085bb5f8e7b..3088442702a338f26aa23fbba300a5aa8e9bd9a9 100644
--- a/compute/zlacpy.c
+++ b/compute/zlacpy.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -158,7 +159,8 @@ int MORSE_zlacpy(MORSE_enum uplo, int M, int N,
return MORSE_SUCCESS;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -216,7 +218,8 @@ int MORSE_zlacpy_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
return MORSE_SUCCESS;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zlange.c b/compute/zlange.c
index c48e992e2a01a3344e6a40863e4364323164e344..82831613fd334461bc4f2a1e1cf437d6eb328a03 100644
--- a/compute/zlange.c
+++ b/compute/zlange.c
@@ -26,7 +26,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -146,8 +147,8 @@ double MORSE_zlange(MORSE_enum norm, int M, int N,
/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
+ RUNTIME_desc_flush( &descA, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(&descA);
morse_desc_mat_free(&descA);
/* } else {*/
/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
@@ -158,7 +159,8 @@ double MORSE_zlange(MORSE_enum norm, int M, int N,
return value;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -215,7 +217,8 @@ double MORSE_zlange_Tile(MORSE_enum norm, MORSE_desc_t *A)
return value;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zlanhe.c b/compute/zlanhe.c
index e03146bc65b14abd3c2e0c0d85a75539ea2501d8..e666221ace06405bf247ff7faab0d9b2cc7ec18d 100644
--- a/compute/zlanhe.c
+++ b/compute/zlanhe.c
@@ -26,7 +26,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -146,8 +147,8 @@ double MORSE_zlanhe(MORSE_enum norm, MORSE_enum uplo, int N,
/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
+ RUNTIME_desc_flush( &descA, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(&descA);
morse_desc_mat_free(&descA);
/* } else {*/
/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
@@ -158,7 +159,8 @@ double MORSE_zlanhe(MORSE_enum norm, MORSE_enum uplo, int N,
return value;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -219,7 +221,8 @@ double MORSE_zlanhe_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A)
return value;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zlansy.c b/compute/zlansy.c
index 5a141e1937c61f6fcab82419ba8fb455ace72cb0..35df4fc8912e33279b04fa6af22974ec005a1e11 100644
--- a/compute/zlansy.c
+++ b/compute/zlansy.c
@@ -26,7 +26,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -156,7 +157,8 @@ double MORSE_zlansy(MORSE_enum norm, MORSE_enum uplo, int N,
return value;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -217,7 +219,8 @@ double MORSE_zlansy_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A)
return value;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zlantr.c b/compute/zlantr.c
index 69563ac643ac00e1c2d08f7082eab70d1182bd9f..3f5e8f1eda3637f3c91541e8da96c816a732dbf6 100644
--- a/compute/zlantr.c
+++ b/compute/zlantr.c
@@ -26,7 +26,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -176,7 +177,8 @@ double MORSE_zlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
return value;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -239,7 +241,8 @@ double MORSE_zlantr_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag, MORS
return value;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zlascal.c b/compute/zlascal.c
index 7fba2b6ee94e810165747545fe309fef2018f860..c86ca1da02ebad9066a30f628521088f95ad0d94 100644
--- a/compute/zlascal.c
+++ b/compute/zlascal.c
@@ -14,7 +14,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -138,7 +139,8 @@ int MORSE_zlascal(MORSE_enum uplo, int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -198,7 +200,8 @@ int MORSE_zlascal_Tile(MORSE_enum uplo,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zlaset.c b/compute/zlaset.c
index c90ba7526368db6e7438813d7b7ebfd08b39ad56..fab26797f9a563522dcb8d744d81070a9bc5dc0b 100644
--- a/compute/zlaset.c
+++ b/compute/zlaset.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -149,7 +150,8 @@ int MORSE_zlaset(MORSE_enum uplo, int M, int N,
return MORSE_SUCCESS;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -205,7 +207,8 @@ int MORSE_zlaset_Tile(MORSE_enum uplo,
return MORSE_SUCCESS;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zlauum.c b/compute/zlauum.c
index f96cab0c6b36d7894df2181d0fa5cdc6bddcdcaa..8ef3f032d56c4805b9e8fad10267a7cc1eb97a26 100644
--- a/compute/zlauum.c
+++ b/compute/zlauum.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -145,7 +146,8 @@ int MORSE_zlauum(MORSE_enum uplo, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -199,15 +201,16 @@ int MORSE_zlauum_Tile(MORSE_enum uplo, MORSE_desc_t *A)
}
morse_sequence_create(morse, &sequence);
MORSE_zlauum_Tile_Async(uplo, A, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zplghe.c b/compute/zplghe.c
index d7c884989890c5bc30ea21ab762242c004a3a123..0d7af992710735c1e17462eb6e5d6d8f71bb6c36 100644
--- a/compute/zplghe.c
+++ b/compute/zplghe.c
@@ -31,7 +31,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -129,7 +130,8 @@ int MORSE_zplghe( double bump, MORSE_enum uplo, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -190,7 +192,8 @@ int MORSE_zplghe_Tile( double bump, MORSE_enum uplo, MORSE_desc_t *A,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zplgsy.c b/compute/zplgsy.c
index 3684d39dca07057d90d9d58cf09edcbd2b4c837a..7c8e8aeda0f5d1e96d384d40b529751a266875c8 100644
--- a/compute/zplgsy.c
+++ b/compute/zplgsy.c
@@ -31,7 +31,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -129,7 +130,8 @@ int MORSE_zplgsy( MORSE_Complex64_t bump, MORSE_enum uplo, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -191,7 +193,8 @@ int MORSE_zplgsy_Tile( MORSE_Complex64_t bump, MORSE_enum uplo,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zplrnt.c b/compute/zplrnt.c
index 707469c9eb10c99f4ffead2d911280dcfff218e1..45e141e11fb5191fe9b2d40b94c7f7a58904f278 100644
--- a/compute/zplrnt.c
+++ b/compute/zplrnt.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -126,7 +127,8 @@ int MORSE_zplrnt( int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -181,7 +183,8 @@ int MORSE_zplrnt_Tile( MORSE_desc_t *A,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zposv.c b/compute/zposv.c
index b0c9cb25d7a8b551c16419cba70aa33ce9e7ecf6..5fe1ab279a4feb38ac097f80ca68bfed70807a54 100644
--- a/compute/zposv.c
+++ b/compute/zposv.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -179,7 +180,8 @@ int MORSE_zposv(MORSE_enum uplo, int N, int NRHS,
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -242,16 +244,17 @@ int MORSE_zposv_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
}
morse_sequence_create(morse, &sequence);
MORSE_zposv_Tile_Async(uplo, A, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zpotrf.c b/compute/zpotrf.c
index e69c04b8605ca4df2a1f2e77ecc42123e22082e3..08261848912dfe3913b30bc45f3e0ffd3adb1d9d 100644
--- a/compute/zpotrf.c
+++ b/compute/zpotrf.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -152,7 +153,8 @@ int MORSE_zpotrf(MORSE_enum uplo, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -211,15 +213,16 @@ int MORSE_zpotrf_Tile(MORSE_enum uplo, MORSE_desc_t *A)
}
morse_sequence_create(morse, &sequence);
MORSE_zpotrf_Tile_Async(uplo, A, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zpotri.c b/compute/zpotri.c
index 5581e027475583c814e6f1693ad944c2870d27e8..801f077cc45b83a80792ea563a91551e4be1f9e2 100644
--- a/compute/zpotri.c
+++ b/compute/zpotri.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -142,7 +143,8 @@ int MORSE_zpotri(MORSE_enum uplo, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -199,15 +201,16 @@ int MORSE_zpotri_Tile(MORSE_enum uplo, MORSE_desc_t *A)
}
morse_sequence_create(morse, &sequence);
MORSE_zpotri_Tile_Async(uplo, A, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zpotrimm.c b/compute/zpotrimm.c
index c779d47c6c9d285de923c001ea9930e98ccd6e8a..180725746aa845cb28b197c915d64304e6e226e1 100644
--- a/compute/zpotrimm.c
+++ b/compute/zpotrimm.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -160,7 +161,8 @@ int MORSE_zpotrimm(MORSE_enum uplo, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -217,15 +219,16 @@ int MORSE_zpotrimm_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE
}
morse_sequence_create(morse, &sequence);
MORSE_zpotrimm_Tile_Async(uplo, A, B, C, sequence, &request);
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zpotrs.c b/compute/zpotrs.c
index 13309e1c524623bdaa8b87b3d783ab9e616acba3..0494b688426132276bb426e911b8645bbd439e35 100644
--- a/compute/zpotrs.c
+++ b/compute/zpotrs.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -163,7 +164,8 @@ int MORSE_zpotrs(MORSE_enum uplo, int N, int NRHS,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -217,16 +219,17 @@ int MORSE_zpotrs_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
}
morse_sequence_create(morse, &sequence);
MORSE_zpotrs_Tile_Async(uplo, A, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zsymm.c b/compute/zsymm.c
index 3bd567a9e486c3eb9f319d6cec1f35e26e332dcd..344f04da6f07ecd8b3be760a95e8d3621aaee504 100644
--- a/compute/zsymm.c
+++ b/compute/zsymm.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -210,7 +211,8 @@ int MORSE_zsymm(MORSE_enum side, MORSE_enum uplo, int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -284,17 +286,18 @@ int MORSE_zsymm_Tile(MORSE_enum side, MORSE_enum uplo,
}
morse_sequence_create(morse, &sequence);
MORSE_zsymm_Tile_Async(side, uplo, alpha, A, B, beta, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zsyr2k.c b/compute/zsyr2k.c
index bc098a162414536c726763b6f341f0b5b826eb5f..50c9b234e36efeb66d03c3347ebc1c6e88a56f2d 100644
--- a/compute/zsyr2k.c
+++ b/compute/zsyr2k.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -214,7 +215,8 @@ int MORSE_zsyr2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -283,17 +285,18 @@ int MORSE_zsyr2k_Tile(MORSE_enum uplo, MORSE_enum trans,
}
morse_sequence_create(morse, &sequence);
MORSE_zsyr2k_Tile_Async(uplo, trans, alpha, A, B, beta, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zsyrk.c b/compute/zsyrk.c
index 3623506f22dba7f484f5e63f5b0399a62b2f0af4..4b29d6b7f93eb1e267d901ae11a897b6119bc634 100644
--- a/compute/zsyrk.c
+++ b/compute/zsyrk.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -194,7 +195,8 @@ int MORSE_zsyrk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -259,16 +261,17 @@ int MORSE_zsyrk_Tile(MORSE_enum uplo, MORSE_enum trans,
}
morse_sequence_create(morse, &sequence);
MORSE_zsyrk_Tile_Async(uplo, trans, alpha, A, beta, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zsysv.c b/compute/zsysv.c
index db51e5c515299716402007d979312469279f75f1..d39bda8bfff3341c1453dcd894478e17d9de50e2 100644
--- a/compute/zsysv.c
+++ b/compute/zsysv.c
@@ -31,7 +31,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -176,7 +177,8 @@ int MORSE_zsysv(MORSE_enum uplo, int N, int NRHS,
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -235,16 +237,17 @@ int MORSE_zsysv_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
}
morse_sequence_create(morse, &sequence);
MORSE_zsysv_Tile_Async(uplo, A, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zsytrf.c b/compute/zsytrf.c
index c157d1248004068d209507dac66451ce7122fb9a..c2e3eeead99ea9237859064a13a66558fdc25360 100644
--- a/compute/zsytrf.c
+++ b/compute/zsytrf.c
@@ -25,7 +25,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -147,7 +148,8 @@ int MORSE_zsytrf(MORSE_enum uplo, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -200,15 +202,16 @@ int MORSE_zsytrf_Tile(MORSE_enum uplo, MORSE_desc_t *A)
}
morse_sequence_create(morse, &sequence);
MORSE_zsytrf_Tile_Async(uplo, A, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zsytrs.c b/compute/zsytrs.c
index 508d404c634dbb376946ea5591b18021c34563fb..df6f7d9dea81ba12ef7e6f443d14f3f9105b2796 100644
--- a/compute/zsytrs.c
+++ b/compute/zsytrs.c
@@ -31,7 +31,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -162,7 +163,8 @@ int MORSE_zsytrs(MORSE_enum uplo, int N, int NRHS,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -213,16 +215,17 @@ int MORSE_zsytrs_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
}
morse_sequence_create(morse, &sequence);
MORSE_zsytrs_Tile_Async(uplo, A, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/ztile.c b/compute/ztile.c
index 5d95c523d3207505aa71a833088414ab7410d59f..5377a78643d356f397c3e46011c097a8113a157d 100644
--- a/compute/ztile.c
+++ b/compute/ztile.c
@@ -26,7 +26,8 @@
**/
#include "control/common.h"
-/*******************************************************************************
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -82,15 +83,16 @@ int MORSE_zLapack_to_Tile(MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *A)
morse_pzlapack_to_tile( Af77, LDA, A, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
RUNTIME_sequence_wait( morse, sequence );
- RUNTIME_desc_getoncpu( A );
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/*******************************************************************************
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
@@ -152,7 +154,8 @@ int MORSE_zLapack_to_Tile_Async(MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *
return MORSE_SUCCESS;
}
-/*******************************************************************************
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -207,14 +210,16 @@ int MORSE_zTile_to_Lapack(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int LDA)
morse_sequence_create(morse, &sequence);
morse_pztile_to_lapack( A, Af77, LDA, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
RUNTIME_sequence_wait( morse, sequence );
- RUNTIME_desc_getoncpu( A );
+
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/*******************************************************************************
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/ztpgqrt.c b/compute/ztpgqrt.c
index 2df25d010ceccca96fb923b6c765e41204d520d1..bf6fee28d640f5e6345299766d1e983c74a5a7d3 100644
--- a/compute/ztpgqrt.c
+++ b/compute/ztpgqrt.c
@@ -296,9 +296,9 @@ int MORSE_ztpgqrt_Tile( int L,
}
morse_sequence_create(morse, &sequence);
MORSE_ztpgqrt_Tile_Async(L, V1, T1, V2, T2, Q1, Q2, sequence, &request);
+ RUNTIME_desc_flush( Q1, sequence );
+ RUNTIME_desc_flush( Q2, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(Q1);
- RUNTIME_desc_getoncpu(Q2);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/ztpqrt.c b/compute/ztpqrt.c
index adb72fe9aef7004c9eae4548e30859802c55b7fd..3298638e02705a80365fc645a57c1c906e76c021 100644
--- a/compute/ztpqrt.c
+++ b/compute/ztpqrt.c
@@ -268,9 +268,9 @@ int MORSE_ztpqrt_Tile( int L, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *T
}
morse_sequence_create(morse, &sequence);
MORSE_ztpqrt_Tile_Async(L, A, B, T, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/ztradd.c b/compute/ztradd.c
index d35eb823c13d08cfb25848e724d2054036aa8532..7bbf05f99cc7828f7c8f7201cf1d63258c5a605d 100644
--- a/compute/ztradd.c
+++ b/compute/ztradd.c
@@ -26,7 +26,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -194,7 +195,8 @@ int MORSE_ztradd(MORSE_enum uplo, MORSE_enum trans, int M, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -263,16 +265,17 @@ int MORSE_ztradd_Tile(MORSE_enum uplo, MORSE_enum trans,
}
morse_sequence_create(morse, &sequence);
MORSE_ztradd_Tile_Async(uplo, trans, alpha, A, beta, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/ztrmm.c b/compute/ztrmm.c
index 17e5d44039b514938445003df3a9c13ed7b125ba..434621d6da5963f885d2e35142b9d1db28b8f22f 100644
--- a/compute/ztrmm.c
+++ b/compute/ztrmm.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -207,7 +208,8 @@ int MORSE_ztrmm(MORSE_enum side, MORSE_enum uplo,
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -285,16 +287,17 @@ int MORSE_ztrmm_Tile(MORSE_enum side, MORSE_enum uplo,
}
morse_sequence_create(morse, &sequence);
MORSE_ztrmm_Tile_Async(side, uplo, transA, diag, alpha, A, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/ztrsm.c b/compute/ztrsm.c
index 5a6b21980c2166cc4f9cd083074b2c59a4a566d7..db4e89dec01cd316b7933dd5faf1c798923cf1e1 100644
--- a/compute/ztrsm.c
+++ b/compute/ztrsm.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -205,7 +206,8 @@ int MORSE_ztrsm(MORSE_enum side, MORSE_enum uplo,
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -283,16 +285,17 @@ int MORSE_ztrsm_Tile(MORSE_enum side, MORSE_enum uplo,
}
morse_sequence_create(morse, &sequence);
MORSE_ztrsm_Tile_Async(side, uplo, transA, diag, alpha, A, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/ztrsmpl.c b/compute/ztrsmpl.c
index 86ef9bb4d3bffa2f0d228fe147e79927bb0d82de..0439834108f356cdb5ed263047c219f069596656 100644
--- a/compute/ztrsmpl.c
+++ b/compute/ztrsmpl.c
@@ -29,7 +29,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -160,7 +161,8 @@ int MORSE_ztrsmpl(int N, int NRHS,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -212,16 +214,17 @@ int MORSE_ztrsmpl_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t
}
morse_sequence_create(morse, &sequence);
MORSE_ztrsmpl_Tile_Async(A, L, IPIV, B, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( B, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(B);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/ztrtri.c b/compute/ztrtri.c
index 28cab09ef59ac5d24e8642972c909a57cf8c0e9d..2f2d36c6040991076ab17946577becfde41f311a 100644
--- a/compute/ztrtri.c
+++ b/compute/ztrtri.c
@@ -28,7 +28,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -154,7 +155,8 @@ int MORSE_ztrtri(MORSE_enum uplo, MORSE_enum diag, int N,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -218,15 +220,16 @@ int MORSE_ztrtri_Tile(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A)
}
morse_sequence_create(morse, &sequence);
MORSE_ztrtri_Tile_Async(uplo, diag, A, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zunglq.c b/compute/zunglq.c
index 0ee4758b652a79c973ae9c54278ca92be62c155c..7ececcb759dc1614ec1ec3b72f2691c541c6c841 100644
--- a/compute/zunglq.c
+++ b/compute/zunglq.c
@@ -30,7 +30,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -166,7 +167,8 @@ int MORSE_zunglq(int M, int N, int K,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -214,16 +216,17 @@ int MORSE_zunglq_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
}
morse_sequence_create(morse, &sequence);
MORSE_zunglq_Tile_Async(A, T, Q, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( Q, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(Q);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zunglq_param.c b/compute/zunglq_param.c
index d64930f4119ada76f12b9623e8230ed43378a1bd..5b6d30e6065194ea3f825b7304f889c8f9be2bb9 100644
--- a/compute/zunglq_param.c
+++ b/compute/zunglq_param.c
@@ -213,9 +213,9 @@ int MORSE_zunglq_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zunglq_param_Tile_Async(qrtree, A, TS, TT, Q, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( Q, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(Q);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zungqr.c b/compute/zungqr.c
index 122b1b911adce9a201c93dbfcc62a9c41ef6405a..822dc4e568f7c9bca64af7c37746b57105aeba7f 100644
--- a/compute/zungqr.c
+++ b/compute/zungqr.c
@@ -30,7 +30,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -165,7 +166,8 @@ int MORSE_zungqr(int M, int N, int K,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -213,16 +215,17 @@ int MORSE_zungqr_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
}
morse_sequence_create(morse, &sequence);
MORSE_zungqr_Tile_Async(A, T, Q, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( Q, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(Q);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zungqr_param.c b/compute/zungqr_param.c
index 8b4f7e46103dcf0fd3cd92492ced8a764c2de654..af88ea598d0b1f9c13bfd1c328f5f5ec2e1170d3 100644
--- a/compute/zungqr_param.c
+++ b/compute/zungqr_param.c
@@ -214,9 +214,9 @@ int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zungqr_param_Tile_Async(qrtree, A, TS, TT, Q, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( Q, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(Q);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zunmlq.c b/compute/zunmlq.c
index 33cf72a309ad39382c9c15b17b60caf59b46dcb0..81535c19d79ff71a24345609b2e319ad90c4056a 100644
--- a/compute/zunmlq.c
+++ b/compute/zunmlq.c
@@ -269,9 +269,9 @@ int MORSE_zunmlq_Tile(MORSE_enum side, MORSE_enum trans,
}
morse_sequence_create(morse, &sequence);
MORSE_zunmlq_Tile_Async(side, trans, A, T, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zunmlq_param.c b/compute/zunmlq_param.c
index cace1d697c507ea985807303f74c1b50846dc3b0..b6f746d02e23a78cbfdd0853a34953d6d04fdac5 100644
--- a/compute/zunmlq_param.c
+++ b/compute/zunmlq_param.c
@@ -268,9 +268,9 @@ int MORSE_zunmlq_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zunmlq_param_Tile_Async(qrtree, side, trans, A, TS, TT, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/compute/zunmqr.c b/compute/zunmqr.c
index a32d603674ffd69e6baf2dc2175d923a6b0dd808..4464aaecbe90f5d6bbf86bf4c644aa89ccbe87d5 100644
--- a/compute/zunmqr.c
+++ b/compute/zunmqr.c
@@ -30,7 +30,8 @@
**/
#include "control/common.h"
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t
*
@@ -207,7 +208,8 @@ int MORSE_zunmqr(MORSE_enum side, MORSE_enum trans, int M, int N, int K,
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile
*
@@ -271,16 +273,17 @@ int MORSE_zunmqr_Tile(MORSE_enum side, MORSE_enum trans,
}
morse_sequence_create(morse, &sequence);
MORSE_zunmqr_Tile_Async(side, trans, A, T, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
return status;
}
-/***************************************************************************//**
+/**
+ ********************************************************************************
*
* @ingroup MORSE_Complex64_t_Tile_Async
*
diff --git a/compute/zunmqr_param.c b/compute/zunmqr_param.c
index 2a101ace364e8b379d92255b11be078925c17670..d285caabf0ce422062a6b621355c85bf309c4de4 100644
--- a/compute/zunmqr_param.c
+++ b/compute/zunmqr_param.c
@@ -274,9 +274,9 @@ int MORSE_zunmqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_
}
morse_sequence_create(morse, &sequence);
MORSE_zunmqr_param_Tile_Async(qrtree, side, trans, A, TS, TT, C, sequence, &request);
+ RUNTIME_desc_flush( A, sequence );
+ RUNTIME_desc_flush( C, sequence );
morse_sequence_wait(morse, sequence);
- RUNTIME_desc_getoncpu(A);
- RUNTIME_desc_getoncpu(C);
status = sequence->status;
morse_sequence_destroy(morse, sequence);
diff --git a/control/control.c b/control/control.c
index fb742b5b491646be64a30fa46818d365bf056ca7..58bc403c18402dd3d42151042325de561eab1afe 100644
--- a/control/control.c
+++ b/control/control.c
@@ -134,7 +134,7 @@ int MORSE_Finalize(void)
morse_error("MORSE_Finalize()", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- MORSE_TASK_flush_all();
+ RUNTIME_flush();
# if !defined(CHAMELEON_SIMULATION)
RUNTIME_barrier(morse);
# endif
diff --git a/control/descriptor.c b/control/descriptor.c
index bec07637c288d0db4bc7e8acd4f1d348046586f2..de7d216bdda0750193727395791c2fe9d3631479 100644
--- a/control/descriptor.c
+++ b/control/descriptor.c
@@ -151,7 +151,7 @@ MORSE_desc_t morse_desc_init_user(MORSE_enum dtyp, int mb, int nb, int bsiz,
desc.occurences = 0;
desc.use_mat = 1;
desc.alloc_mat = 1;
- desc.register_mat = 1;
+ desc.register_mat = (morse->ncudas > 0) ? 1 : 0;
desc.ooc = 0;
desc.myrank = RUNTIME_comm_rank( morse );
@@ -809,8 +809,7 @@ int MORSE_Desc_Release (MORSE_desc_t *desc) {
*
* @ingroup Descriptor
*
- * MORSE_Desc_Getoncpu - Apply an acquire and a release on the data of the
- * descriptors so that its values are up-to-date in the main memory.
+ * MORSE_Desc_Flush - Flushes the data in the sequence when they won't be reused. This calls cleans up the distributed communication caches, and transfer the data back to the CPU.
*
******************************************************************************
*
@@ -823,8 +822,11 @@ int MORSE_Desc_Release (MORSE_desc_t *desc) {
* \retval MORSE_SUCCESS successful exit
*
*****************************************************************************/
-int MORSE_Desc_Getoncpu(MORSE_desc_t *desc) {
- return RUNTIME_desc_getoncpu( desc );
+int MORSE_Desc_Flush( MORSE_desc_t *desc,
+ MORSE_sequence_t *sequence )
+{
+ RUNTIME_desc_flush( desc, sequence );
+ return MORSE_SUCCESS;
}
/**
diff --git a/doc/orgmode/chapters/using.org b/doc/orgmode/chapters/using.org
index 5bc65eaed097a7b8d10bfbe517598325dfbf4881..7bf39140f57702191dee06f459a506e2754c254f 100644
--- a/doc/orgmode/chapters/using.org
+++ b/doc/orgmode/chapters/using.org
@@ -1004,7 +1004,7 @@
Ensure that all data are up-to-date in main memory (even if some
tasks have been processed on GPUs).
#+begin_src
- int MORSE_Desc_Getoncpu(MORSE_desc_t *desc);
+ int MORSE_Desc_Flush(MORSE_desc_t *desc, MORSE_sequence_t *sequence);
#+end_src
Set the sizes for the MPI tags. Default value: tag_width=31,
diff --git a/doc/texinfo/chapters/using.texi b/doc/texinfo/chapters/using.texi
index bb50d74f50044dd41a8be8524bd0d4a148ebba8b..ae7dbe1009264ce26b1230e041ad92f9bd64bed2 100644
--- a/doc/texinfo/chapters/using.texi
+++ b/doc/texinfo/chapters/using.texi
@@ -830,7 +830,7 @@ int MORSE_Desc_Destroy (MORSE_desc_t **desc);
Ensure that all data are up-to-date in main memory (even if some tasks have
been processed on GPUs)
@verbatim
-int MORSE_Desc_Getoncpu(MORSE_desc_t *desc);
+int MORSE_Desc_Flush(MORSE_desc_t *desc, MORSE_sequence_t *sequence);
@end verbatim
@node Options routines
diff --git a/example/lapack_to_morse/step4.c b/example/lapack_to_morse/step4.c
index b4111cdab378e21ea1abcc60f41892a71e14661c..3c37878e56eeca485e2eb922c5e036236bd59e35 100644
--- a/example/lapack_to_morse/step4.c
+++ b/example/lapack_to_morse/step4.c
@@ -141,15 +141,15 @@ int main(int argc, char *argv[]) {
*/
MORSE_dpotrs_Tile_Async( UPLO, descA, descX, sequence, &request);
+ /* Ensure that all data processed on the gpus we are depending on are back
+ * in main memory */
+ MORSE_Desc_Flush( descA, sequence );
+ MORSE_Desc_Flush( descX, sequence );
+
/* Synchronization barrier (the runtime ensures that all submitted tasks
* have been terminated */
MORSE_Sequence_Wait(sequence);
- /* Ensure that all data processed on the gpus we are depending on are back
- * in main memory */
- RUNTIME_desc_getoncpu(descA);
- RUNTIME_desc_getoncpu(descX);
-
status = sequence->status;
if ( status != 0 ) {
fprintf(stderr, "Error in computation (%d)\n", status);
diff --git a/example/lapack_to_morse/step5.c b/example/lapack_to_morse/step5.c
index 46aa1d29ae495cbfa57a4f6b0890dbdb528db50c..addff8e7cb34534a4fc4589a94b2884d815b9bdc 100644
--- a/example/lapack_to_morse/step5.c
+++ b/example/lapack_to_morse/step5.c
@@ -145,15 +145,15 @@ int main(int argc, char *argv[]) {
*/
MORSE_dpotrs_Tile_Async( UPLO, descA, descX, sequence, &request);
+ /* Ensure that all data processed on the gpus we are depending on are back
+ * in main memory */
+ MORSE_Desc_Flush( descA, sequence );
+ MORSE_Desc_Flush( descX, sequence );
+
/* Synchronization barrier (the runtime ensures that all submitted tasks
* have been terminated */
MORSE_Sequence_Wait(sequence);
- /* Ensure that all data processed on the gpus we are depending on are back
- * in main memory */
- RUNTIME_desc_getoncpu(descA);
- RUNTIME_desc_getoncpu(descX);
-
status = sequence->status;
if ( status != 0 ) {
fprintf(stderr, "Error in computation (%d)\n", status);
diff --git a/example/lapack_to_morse/step6.c b/example/lapack_to_morse/step6.c
index 68850db3f7e3af4374aee9e2f9f63b9abf89d8cd..074d4cfa8e054d9ddea83381e692d03ae7b140e2 100644
--- a/example/lapack_to_morse/step6.c
+++ b/example/lapack_to_morse/step6.c
@@ -167,15 +167,15 @@ int main(int argc, char *argv[]) {
*/
MORSE_dpotrs_Tile_Async( UPLO, descA, descX, sequence, &request);
+ /* Ensure that all data processed on the gpus we are depending on are back
+ * in main memory */
+ MORSE_Desc_Flush( descA, sequence );
+ MORSE_Desc_Flush( descX, sequence );
+
/* Synchronization barrier (the runtime ensures that all submitted tasks
* have been terminated */
MORSE_Sequence_Wait(sequence);
- /* Ensure that all data processed on the gpus we are depending on are back
- * in main memory */
- RUNTIME_desc_getoncpu(descA);
- RUNTIME_desc_getoncpu(descX);
-
status = sequence->status;
if ( status != 0 ) {
fprintf(stderr, "Error in computation (%d)\n", status);
diff --git a/example/lapack_to_morse/step7.c b/example/lapack_to_morse/step7.c
index 15e00f5875bc32da15788188c8fd6e0e3fa7f115..fbb48e4e50deaa1ad1d48b9e3aad6963c6a0b72a 100644
--- a/example/lapack_to_morse/step7.c
+++ b/example/lapack_to_morse/step7.c
@@ -175,15 +175,15 @@ int main(int argc, char *argv[]) {
*/
MORSE_dpotrs_Tile_Async( UPLO, descA, descX, sequence, &request);
+ /* Ensure that all data processed on the gpus we are depending on are back
+ * in main memory */
+ MORSE_Desc_Flush( descA, sequence );
+ MORSE_Desc_Flush( descX, sequence );
+
/* Synchronization barrier (the runtime ensures that all submitted tasks
* have been terminated */
MORSE_Sequence_Wait(sequence);
- /* Ensure that all data processed on the gpus we are depending on are back
- * in main memory */
- RUNTIME_desc_getoncpu(descA);
- RUNTIME_desc_getoncpu(descX);
-
status = sequence->status;
if ( status != 0 ) {
fprintf(stderr, "Error in computation (%d)\n", status);
diff --git a/include/chameleon/morse_runtime.h b/include/chameleon/morse_runtime.h
index c9f098f3f7ae8d78189a8503f08f0e319c5ccd26..76c83ce85505d1afac466de41a3d5c1f84b5ab49 100644
--- a/include/chameleon/morse_runtime.h
+++ b/include/chameleon/morse_runtime.h
@@ -394,52 +394,73 @@ int
RUNTIME_desc_release( const MORSE_desc_t *desc );
/**
- * @brief Make sure the data is brought back to the CPU main memory at the
- * end of the algorithm.
+ * @brief Flush all pieces of data from a descriptor.
+ *
+ * This function marks all pieces of data from the descriptor as unused in the future to:
+ * - cleanup the data from the distributed cache if present.
+ * - mark for eviction from the GPU if space is needed
+ * - move the data back to the main memory
*
* @param[in] desc
- * The descriptor to release.
+ * The descriptor to flush.
*
- * @retval MORSE_SUCCESS on success
+ * @param[in] sequence
+ * The sequence in which the data is used.
*/
-int
-RUNTIME_desc_getoncpu( const MORSE_desc_t *desc );
+void
+RUNTIME_desc_flush( const MORSE_desc_t *desc,
+ const MORSE_sequence_t *sequence );
/**
- * @brief Make sure the data will be brought back to the CPU main memory at the
- * end of the algorithm.
+ * @brief Flush all data submitted to the runtime systems from the distributed
+ * cache, and/or mark for eviction from the GPU memory.
*
- * This function is a asynchronous call that submit the data movement from
- * remote memory to the main memory. This call must be completed by a call to
- * RUNTIME_sequence_wait() to ensure that all data have been moved.
- * Users should avoid to call this function as it sequentially moves back the
- * data from outside the main memory to main memory, and should prefer
- * RUNTIME_desc_getoncpu_async().
+ * This function flushes all data from the distributed cache of the runtime system.
+ */
+void
+RUNTIME_flush( );
+
+/**
+ * @brief Flush a single piece of data.
*
- * @param[in] desc
- * The descriptor to release.
+ * This function marks a piece of data as unused in the future to:
+ * - cleanup the data from the distributed cache if present.
+ * - mark for eviction from the GPU if space is needed
+ * - move the data back to the main memory
*
* @param[in] sequence
- * The sequence to which submit the data movements
+ * The sequence in which the data is used.
*
- * @retval MORSE_SUCCESS on success
+ * @param[in] A
+ * The descriptor to which the piece of data belongs.
+ *
+ * @param[in] Am
+ * The row coordinate of the piece of data in the matrix
+ *
+ * @param[in] An
+ * The column coordinate of the piece of data in the matrix
*/
-int
-RUNTIME_desc_getoncpu_async( const MORSE_desc_t *desc,
- MORSE_sequence_t *sequence );
+void
+RUNTIME_data_flush( const MORSE_sequence_t *sequence,
+ const MORSE_desc_t *A, int Am, int An );
/**
* @brief Get the pointer to the data or the runtime handler associated to the
* piece of data (m, n) in desc.
*
- * @param[in] desc
- * The descriptor to release.
+ * @param[in] A
+ * The descriptor to which belongs the piece of data.
*
- * @retval MORSE_SUCCESS on success
+ * @param[in] Am
+ * The row coordinate of the piece of data in the matrix
+ *
+ * @param[in] An
+ * The column coordinate of the piece of data in the matrix
+ *
+ * @retval The runtime handler address of the piece of data.
*/
void *
-RUNTIME_desc_getaddr( const MORSE_desc_t *desc,
- int m, int n );
+RUNTIME_data_getaddr( const MORSE_desc_t *A, int Am, int An );
/**
* @}
diff --git a/include/chameleon/morse_tasks.h b/include/chameleon/morse_tasks.h
index 82abf2c7853cbf5269d8b1887386dbd48440762a..a048e391eb40996b9429eb55bfc63ccaa16fa5b9 100644
--- a/include/chameleon/morse_tasks.h
+++ b/include/chameleon/morse_tasks.h
@@ -46,15 +46,6 @@ void MORSE_TASK_stile_zero(const MORSE_option_t *options,
int X1, int X2, int Y1, int Y2,
const MORSE_desc_t *A, int Am, int An, int lda);
-/*
- * Mark a data as unused after this call
- */
-void MORSE_TASK_flush_data( const MORSE_option_t *options,
- const MORSE_desc_t *A, int Am, int An);
-void MORSE_TASK_flush_desc( const MORSE_option_t *options,
- MORSE_enum uplo, const MORSE_desc_t *A );
-void MORSE_TASK_flush_all();
-
#include "chameleon/morse_tasks_z.h"
#include "chameleon/morse_tasks_d.h"
#include "chameleon/morse_tasks_c.h"
diff --git a/include/morse.h b/include/morse.h
index 6d745743fdebf87f8c14dd903fb2f6b7e1982aa1..525c27b0671748655578d01f2bcaa5ccebe23c77 100644
--- a/include/morse.h
+++ b/include/morse.h
@@ -114,7 +114,7 @@ int MORSE_Desc_Create_OOC_User(MORSE_desc_t **desc, MORSE_enum dtyp,
int MORSE_Desc_Destroy (MORSE_desc_t **desc);
int MORSE_Desc_Acquire (MORSE_desc_t *desc);
int MORSE_Desc_Release (MORSE_desc_t *desc);
-int MORSE_Desc_Getoncpu(MORSE_desc_t *desc);
+int MORSE_Desc_Flush (MORSE_desc_t *desc, MORSE_sequence_t *sequence);
void MORSE_user_tag_size(int, int) ;
/* Workspaces */
diff --git a/runtime/parsec/include/chameleon_parsec.h b/runtime/parsec/include/chameleon_parsec.h
index 76f515905084369d2d597eda5ee143dfac04172f..4dcce7564eb3bfb5d35f96683dd55a49422a73c9 100644
--- a/runtime/parsec/include/chameleon_parsec.h
+++ b/runtime/parsec/include/chameleon_parsec.h
@@ -22,8 +22,8 @@ typedef struct morse_parsec_desc_s morse_parsec_desc_t;
/*
* Access to block pointer and leading dimension
*/
-#define RTBLKADDR( desc, type, m, n ) ( parsec_dtd_tile_of( (parsec_data_collection_t *) ((desc)->schedopt), \
- ((parsec_data_collection_t *) (desc)->schedopt)->data_key((desc)->schedopt, m, n) ))
+#define RTBLKADDR( desc, type, m, n ) ( parsec_dtd_tile_of( (parsec_data_collection_t *) ((desc)->schedopt), \
+ ((parsec_data_collection_t *) (desc)->schedopt)->data_key((desc)->schedopt, m, n) ))
#define RUNTIME_BEGIN_ACCESS_DECLARATION
diff --git a/runtime/quark/CMakeLists.txt b/runtime/quark/CMakeLists.txt
index c6412acc493b121e49b3a4a329796b5411dad93d..db92c2a9bbc2755ca15b7a896432782c7c7cb83e 100644
--- a/runtime/quark/CMakeLists.txt
+++ b/runtime/quark/CMakeLists.txt
@@ -86,7 +86,6 @@ set(RUNTIME_COMMON
control/runtime_descriptor.c
control/runtime_options.c
control/runtime_profiling.c
- codelets/codelet_dataflush.c
${RUNTIME_COMMON_GENERATED}
)
diff --git a/runtime/quark/codelets/codelet_dataflush.c b/runtime/quark/codelets/codelet_dataflush.c
deleted file mode 100644
index cbec78c35a0739516c61a3c7d36a99c2cfbb3148..0000000000000000000000000000000000000000
--- a/runtime/quark/codelets/codelet_dataflush.c
+++ /dev/null
@@ -1,53 +0,0 @@
-/**
- *
- * @copyright (c) 2009-2014 The University of Tennessee and The University
- * of Tennessee Research Foundation.
- * All rights reserved.
- * @copyright (c) 2012-2014 Inria. All rights reserved.
- * @copyright (c) 2012-2014, 2016 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
- *
- **/
-
-/**
- *
- * @file codelet_dataflush.c
- *
- * MORSE codelets kernel
- * MORSE is a software package provided by Univ. of Tennessee,
- * Univ. of California Berkeley and Univ. of Colorado Denver
- *
- * @version 2.5.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for MORSE 1.0.0
- * @author Mathieu Faverge
- *
- * @date 2014-02-05
- *
- **/
-#include "chameleon_quark.h"
-
-void MORSE_TASK_flush_data( const MORSE_option_t *options,
- const MORSE_desc_t *A, int Am, int An )
-{
- (void)options; (void)A; (void)Am; (void)An;
-
- /*
- * This is useful for StarPU and PaRSEC MPI implementation, if it happens in
- * Quark, it will need to be done carefuly to not break other runtimes.
- */
-}
-
-void MORSE_TASK_flush_desc( const MORSE_option_t *options,
- MORSE_enum uplo, const MORSE_desc_t *A )
-{
- (void)options; (void)uplo; (void)A;
-
- /*
- * This is useful for StarPU and PaRSEC MPI implementation, if it happens in
- * Quark, it will need to be done carefuly to not break other runtimes.
- */
-}
-
-void MORSE_TASK_flush_all()
-{
-}
diff --git a/runtime/quark/control/runtime_descriptor.c b/runtime/quark/control/runtime_descriptor.c
index 2b08dd846a3916e1013243867373b91f0ac9470e..3832d9926ee7b00fa86b9ae04d7f52c8a7d7f4bc 100644
--- a/runtime/quark/control/runtime_descriptor.c
+++ b/runtime/quark/control/runtime_descriptor.c
@@ -64,21 +64,34 @@ int RUNTIME_desc_release( const MORSE_desc_t *desc )
return MORSE_SUCCESS;
}
-int RUNTIME_desc_getoncpu( const MORSE_desc_t *desc )
+void
+RUNTIME_desc_flush( const MORSE_desc_t *desc,
+ const MORSE_sequence_t *sequence )
{
(void)desc;
- return MORSE_SUCCESS;
+ (void)sequence;
+ return;
}
-int RUNTIME_desc_getoncpu_async( const MORSE_desc_t *desc,
- MORSE_sequence_t *sequence )
+
+void
+RUNTIME_flush( )
+{
+ return;
+}
+
+void
+RUNTIME_data_flush( const MORSE_sequence_t *sequence,
+ const MORSE_desc_t *A, int Am, int An )
{
- (void)desc;
(void)sequence;
- return MORSE_SUCCESS;
+ (void)A;
+ (void)Am;
+ (void)An;
+ return;
}
-void *RUNTIME_desc_getaddr( const MORSE_desc_t *desc, int m, int n )
+void *RUNTIME_data_getaddr( const MORSE_desc_t *desc, int m, int n )
{
return desc->get_blkaddr( desc, m, n );
}
diff --git a/runtime/quark/include/chameleon_quark.h b/runtime/quark/include/chameleon_quark.h
index 50d11f52188dcdd8759820714aa9486cfb5be11b..231e8acf39a1d1499b19ebab8a5dd71ef63764d6 100644
--- a/runtime/quark/include/chameleon_quark.h
+++ b/runtime/quark/include/chameleon_quark.h
@@ -40,7 +40,7 @@ typedef struct quark_option_s {
/*
* Access to block pointer and leading dimension
*/
-#define RTBLKADDR( desc, type, m, n ) ( (type*)RUNTIME_desc_getaddr( desc, m, n ) )
+#define RTBLKADDR( desc, type, m, n ) ( (type*)RUNTIME_data_getaddr( desc, m, n ) )
#define RUNTIME_BEGIN_ACCESS_DECLARATION
diff --git a/runtime/starpu/CMakeLists.txt b/runtime/starpu/CMakeLists.txt
index 2bc2c94d14005fc9c0831f98920bac3b470c5e3e..a6d06b1ad6e193823f0ef152e10fd08216349a2d 100644
--- a/runtime/starpu/CMakeLists.txt
+++ b/runtime/starpu/CMakeLists.txt
@@ -95,7 +95,6 @@ set(RUNTIME_COMMON
control/runtime_options.c
control/runtime_profiling.c
control/runtime_workspace.c
- codelets/codelet_dataflush.c
${RUNTIME_COMMON_GENERATED}
)
diff --git a/runtime/starpu/codelets/codelet_dataflush.c b/runtime/starpu/codelets/codelet_dataflush.c
deleted file mode 100644
index 976c5dfca15c2b05f62590d68f2e4cb53a2ec5d9..0000000000000000000000000000000000000000
--- a/runtime/starpu/codelets/codelet_dataflush.c
+++ /dev/null
@@ -1,133 +0,0 @@
-/**
- *
- * @copyright (c) 2009-2014 The University of Tennessee and The University
- * of Tennessee Research Foundation.
- * All rights reserved.
- * @copyright (c) 2012-2014 Inria. All rights reserved.
- * @copyright (c) 2012-2016 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
- *
- **/
-
-/**
- *
- * @file codelet_dataflush.c
- *
- * MORSE codelets kernel
- * MORSE is a software package provided by Univ. of Tennessee,
- * Univ. of California Berkeley and Univ. of Colorado Denver
- *
- * @version 2.5.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for MORSE 1.0.0
- * @author Mathieu Faverge
- * @author Marc Sergent
- * @date 2014-02-05
- *
- **/
-#include "chameleon_starpu.h"
-#include "chameleon/morse_tasks.h"
-
-#ifdef HAVE_STARPU_DATA_WONT_USE
-#elif defined HAVE_STARPU_IDLE_PREFETCH
-static void data_flush(void *handle)
-{
- starpu_data_idle_prefetch_on_node(handle, STARPU_MAIN_RAM, 1);
- starpu_data_release_on_node(handle, -1);
-}
-#else
-static void data_release(void *handle)
-{
- starpu_data_release(handle);
-}
-#endif
-
-#ifdef CHAMELEON_USE_MPI
-#ifdef HAVE_STARPU_MPI_CACHED_RECEIVE
-int RUNTIME_desc_iscached(const MORSE_desc_t *A, int Am, int An)
-{
- starpu_data_handle_t *ptrtile = (starpu_data_handle_t*)(A->schedopt);
- ptrtile += ((int64_t)(A->lmt) * (int64_t)An + (int64_t)Am);
-
- if (!(*ptrtile))
- return 0;
-
- return starpu_mpi_cached_receive(*ptrtile);
-}
-#endif
-#endif
-
-void MORSE_TASK_flush_data( const MORSE_option_t *options,
- const MORSE_desc_t *A, int Am, int An )
-{
- (void)options;
-
- {
- starpu_data_handle_t *ptrtile = (starpu_data_handle_t*)(A->schedopt);
- ptrtile += ((int64_t)(A->lmt) * (int64_t)An + (int64_t)Am);
-
- if (*ptrtile != NULL)
- {
-#if defined(CHAMELEON_USE_MPI)
- starpu_mpi_cache_flush(MPI_COMM_WORLD, *ptrtile);
-#endif
-
- if ( A->myrank == A->get_rankof( A, Am, An ) )
- {
- /* Push data to main memory when we have time to */
-#ifdef HAVE_STARPU_DATA_WONT_USE
- starpu_data_wont_use(*ptrtile);
-#elif defined HAVE_STARPU_IDLE_PREFETCH
- starpu_data_acquire_on_node_cb(*ptrtile, -1, STARPU_R, data_flush, *ptrtile);
-#else
- starpu_data_acquire_cb(*ptrtile, STARPU_R, data_release, *ptrtile);
-#endif
- }
- }
- }
-}
-
-void MORSE_TASK_flush_desc( const MORSE_option_t *options,
- MORSE_enum uplo, const MORSE_desc_t *A )
-{
- int m, n;
-
- switch (uplo) {
- /*
- * MorseUpper
- */
- case MorseUpper:
- for (m = 0; m < A->mt; m++) {
- for (n = m; n < A->nt; n++) {
- MORSE_TASK_flush_data( options, A, m, n );
- }
- }
- break;
- /*
- * MorseLower
- */
- case MorseLower:
- for (m = 0; m < A->mt; m++) {
- for (n = 0; n < chameleon_min(m+1, A->nt); n++) {
- MORSE_TASK_flush_data( options, A, m, n );
- }
- }
- break;
- /*
- * MorseUpperLower
- */
- case MorseUpperLower:
- default:
- for (m = 0; m < A->mt; m++) {
- for (n = 0; n < A->nt; n++) {
- MORSE_TASK_flush_data( options, A, m, n );
- }
- }
- }
-}
-
-void MORSE_TASK_flush_all()
-{
-#if defined(CHAMELEON_USE_MPI)
- starpu_mpi_cache_flush_all_data(MPI_COMM_WORLD);
-#endif
-}
diff --git a/runtime/starpu/control/runtime_descriptor.c b/runtime/starpu/control/runtime_descriptor.c
index 39282f386736479b9b94b472f16e3e673c5d96aa..ee2350065243b40bc15d943d26f052875a7292f2 100644
--- a/runtime/starpu/control/runtime_descriptor.c
+++ b/runtime/starpu/control/runtime_descriptor.c
@@ -127,6 +127,8 @@ void RUNTIME_desc_create( MORSE_desc_t *desc )
rc = cudaHostRegister( desc->mat, size, cudaHostRegisterPortable );
if ( rc != cudaSuccess )
{
+ /* Disable the unregister as register failed */
+ desc->register_mat = 0;
morse_warning("RUNTIME_desc_create(StarPU): cudaHostRegister - ", cudaGetErrorString( rc ));
}
}
@@ -294,76 +296,96 @@ int RUNTIME_desc_release( const MORSE_desc_t *desc )
}
/*******************************************************************************
- * Get data on cpu - Synchronous call
+ * Flush cached data
**/
-int RUNTIME_desc_getoncpu( const MORSE_desc_t *desc )
+void RUNTIME_flush()
+{
+#if defined(CHAMELEON_USE_MPI)
+ starpu_mpi_cache_flush_all_data(MPI_COMM_WORLD);
+#endif
+}
+
+/*****************************************************************************
+ * Different implementations of the flush call based on StarPU version
+ */
+#ifdef HAVE_STARPU_DATA_WONT_USE
+
+static inline void
+chameleon_starpu_data_wont_use( starpu_data_handle_t handle ) {
+ starpu_data_wont_use( handle );
+}
+
+#elif defined HAVE_STARPU_IDLE_PREFETCH
+
+static inline void
+chameleon_starpu_data_flush( starpu_data_handle_t handle)
+{
+ starpu_data_idle_prefetch_on_node(handle, STARPU_MAIN_RAM, 1);
+ starpu_data_release_on_node(handle, -1);
+}
+
+static inline void
+chameleon_starpu_data_wont_use( starpu_data_handle_t handle ) {
+ starpu_data_acquire_on_node_cb( handle, -1, STARPU_R,
+ chameleon_starpu_data_flush, handle );
+}
+
+#else
+
+static inline void
+chameleon_starpu_data_wont_use( starpu_data_handle_t handle ) {
+ starpu_data_acquire_cb( handle, STARPU_R,
+ (void (*)(void*))&starpu_data_release, handle );
+
+#endif
+
+void RUNTIME_desc_flush( const MORSE_desc_t *desc,
+ const MORSE_sequence_t *sequence )
{
starpu_data_handle_t *handle = (starpu_data_handle_t*)(desc->schedopt);
int lmt = desc->lmt;
int lnt = desc->lnt;
int m, n;
- if ( desc->ooc ) {
- /* May not even fit */
- morse_warning( "RUNTIME_desc_getoncpu(StarPU)",
- "Try to get an out-of-core matrix on main memory. Cancelled as it might not fit" );
- return MORSE_SUCCESS;
- }
-
- for (n = 0; n < lnt; n++) {
- for (m = 0; m < lmt; m++)
+ for (n = 0; n < lnt; n++)
+ {
+ for (m = 0; m < lmt; m++, handle++)
{
- if ( (*handle == NULL) ||
- !morse_desc_islocal( desc, m, n ) )
- {
- handle++;
+ if ( *handle == NULL ) {
continue;
}
- starpu_data_acquire(*handle, STARPU_R);
- starpu_data_release(*handle);
- handle++;
+#if defined(CHAMELEON_USE_MPI)
+ starpu_mpi_cache_flush( MPI_COMM_WORLD, *handle );
+#endif
+ if ( morse_desc_islocal( desc, m, n ) ) {
+ chameleon_starpu_data_wont_use( *handle );
+ }
}
}
- return MORSE_SUCCESS;
+
+ (void)sequence;
}
-/*******************************************************************************
- * Get data on cpu - Asynchronous call
- **/
-int RUNTIME_desc_getoncpu_async( const MORSE_desc_t *desc,
- MORSE_sequence_t *sequence )
+void RUNTIME_data_flush( const MORSE_sequence_t *sequence,
+ const MORSE_desc_t *A, int Am, int An )
{
- starpu_data_handle_t *handle = (starpu_data_handle_t*)(desc->schedopt);
- int lmt = desc->lmt;
- int lnt = desc->lnt;
- int m, n;
+ starpu_data_handle_t *handle = (starpu_data_handle_t*)(A->schedopt);
+ handle += ((int64_t)(A->lmt) * (int64_t)An + (int64_t)Am);
- if ( desc->ooc ) {
- /* May not even fit */
- morse_warning( "RUNTIME_desc_getoncpu_async(StarPU)",
- "Try to get an out-of-core matrix on main memory. Cancelled as it might not fit" );
- return MORSE_SUCCESS;
+ if (*handle == NULL) {
+ return;
}
- for (n = 0; n < lnt; n++) {
- for (m = 0; m < lmt; m++)
- {
- if ( (*handle == NULL) ||
- !morse_desc_islocal( desc, m, n ) )
- {
- handle++;
- continue;
- }
+#if defined(CHAMELEON_USE_MPI)
+ starpu_mpi_cache_flush( MPI_COMM_WORLD, *handle );
+#endif
- starpu_data_acquire_cb( *handle, STARPU_R,
- (void (*)(void*))&starpu_data_release, *handle );
- handle++;
- }
+ if ( morse_desc_islocal( A, Am, An ) ) {
+ chameleon_starpu_data_wont_use( *handle );
}
(void)sequence;
- return MORSE_SUCCESS;
}
/*******************************************************************************
@@ -375,7 +397,7 @@ int RUNTIME_desc_getoncpu_async( const MORSE_desc_t *desc,
#define STARPU_MAIN_RAM 0
#endif
-void *RUNTIME_desc_getaddr( const MORSE_desc_t *desc, int m, int n )
+void *RUNTIME_data_getaddr( const MORSE_desc_t *desc, int m, int n )
{
int64_t im = m + (desc->i / desc->mb);
int64_t jn = n + (desc->j / desc->nb);
diff --git a/runtime/starpu/include/chameleon_starpu.h.in b/runtime/starpu/include/chameleon_starpu.h.in
index 82a71f96bde9802b9ec23e303f977c50b31a6d65..7567425e6c5819f0fd4c03f0b132ca6518ced614 100644
--- a/runtime/starpu/include/chameleon_starpu.h.in
+++ b/runtime/starpu/include/chameleon_starpu.h.in
@@ -117,24 +117,31 @@ typedef struct starpu_conf starpu_conf_t;
/**
* Access to block pointer and leading dimension
*/
-#define RTBLKADDR( desc, type, m, n ) ( (starpu_data_handle_t)RUNTIME_desc_getaddr( desc, m, n ) )
+#define RTBLKADDR( desc, type, m, n ) ( (starpu_data_handle_t)RUNTIME_data_getaddr( desc, m, n ) )
void RUNTIME_set_reduction_methods(starpu_data_handle_t handle, MORSE_enum dtyp);
-#ifdef CHAMELEON_USE_MPI
-#ifdef HAVE_STARPU_MPI_CACHED_RECEIVE
-int RUNTIME_desc_iscached(const MORSE_desc_t *A, int Am, int An);
-#endif
-#endif
-#if defined(CHAMELEON_USE_MPI)
-# if defined(HAVE_STARPU_MPI_CACHED_RECEIVE)
-# define RUNTIME_ACCESS_WRITE_CACHED(A, Am, An) do { if (RUNTIME_desc_iscached(A, Am, An)) __morse_need_submit = 1; } while(0)
-# else
-# warning "WAR dependencies need starpu_mpi_cached_receive support from StarPU 1.2.1 or greater"
-# define RUNTIME_ACCESS_WRITE_CACHED(A, Am, An)
-# endif
+#if defined(CHAMELEON_USE_MPI) && defined(HAVE_STARPU_MPI_CACHED_RECEIVE)
+static inline int
+chameleon_starpu_data_iscached(const MORSE_desc_t *A, int Am, int An)
+{
+ starpu_data_handle_t *ptrtile = (starpu_data_handle_t*)(A->schedopt);
+ ptrtile += ((int64_t)(A->lmt) * (int64_t)An + (int64_t)Am);
+
+ if (!(*ptrtile))
+ return 0;
+
+ return starpu_mpi_cached_receive(*ptrtile);
+}
+
+#define RUNTIME_ACCESS_WRITE_CACHED(A, Am, An) do { \
+ if (chameleon_starpu_data_iscached(A, Am, An)) __morse_need_submit = 1; } while(0)
+
#else
-#define RUNTIME_ACCESS_WRITE_CACHED(A, Am, An)
+
+#warning "WAR dependencies need starpu_mpi_cached_receive support from StarPU 1.2.1 or greater"
+#define RUNTIME_ACCESS_WRITE_CACHED(A, Am, An) do {} while (0)
+
#endif
#ifdef CHAMELEON_ENABLE_PRUNING_STATS
diff --git a/timing/time_zgetri_tile.c b/timing/time_zgetri_tile.c
index 471f30f621b2dfa57501d99b7d96d30fd7fe5121..ebd0a74f2e15017556e556491705eef4b58a7f6f 100644
--- a/timing/time_zgetri_tile.c
+++ b/timing/time_zgetri_tile.c
@@ -40,10 +40,10 @@ static int check_getri_factorization(MORSE_desc_t *descA1, MORSE_desc_t *descA2,
MORSE_Complex64_t *b = (MORSE_Complex64_t *)malloc((descA1->m)*sizeof(MORSE_Complex64_t));
MORSE_Complex64_t *x = (MORSE_Complex64_t *)malloc((descA1->m)*sizeof(MORSE_Complex64_t));
- MORSE_Desc_Create(&descB, b, MorseComplexDouble, descA1->mb, descA1->nb, descA1->bsiz,
- descA1->m, 1, 0, 0, descA1->m, 1, 1, 1);
- MORSE_Desc_Create(&descX, x, MorseComplexDouble, descA1->mb, descA1->nb, descA1->bsiz,
- descA1->m, 1, 0, 0, descA1->m, 1, 1, 1);
+ MORSE_Desc_Create(&descB, b, MorseComplexDouble, descA1->mb, descA1->nb, descA1->bsiz,
+ descA1->m, 1, 0, 0, descA1->m, 1, 1, 1);
+ MORSE_Desc_Create(&descX, x, MorseComplexDouble, descA1->mb, descA1->nb, descA1->bsiz,
+ descA1->m, 1, 0, 0, descA1->m, 1, 1, 1);
MORSE_zplrnt_Tile( descX, 537 );
MORSE_zlacpy_Tile( MorseUpperLower, descX, descB);
@@ -55,9 +55,9 @@ static int check_getri_factorization(MORSE_desc_t *descA1, MORSE_desc_t *descA2,
Bnorm = MORSE_zlange_Tile(MorseInfNorm, descB, work);
MORSE_zgemm_Tile( MorseNoTrans, MorseNoTrans,
- (MORSE_Complex64_t)1., descA1, descX,
+ (MORSE_Complex64_t)1., descA1, descX,
(MORSE_Complex64_t)-1., descB);
-
+
Rnorm = MORSE_zlange_Tile(MorseInfNorm, descB, work);
if (getenv("MORSE_TESTING_VERBOSE"))
@@ -96,26 +96,26 @@ static int check_getri_inverse(MORSE_desc_t *descA1, MORSE_desc_t *descA2, int *
double eps = LAPACKE_dlamch_work('e');
MORSE_desc_t *descW;
- MORSE_Desc_Create(&descW, work, MorseComplexDouble, descA1->mb, descA1->nb, descA1->bsiz,
+ MORSE_Desc_Create(&descW, work, MorseComplexDouble, descA1->mb, descA1->nb, descA1->bsiz,
descA1->m, descA1->n, 0, 0, descA1->m, descA1->n);
-
+
MORSE_zlaset_Tile( MorseUpperLower, (MORSE_Complex64_t)0., (MORSE_Complex64_t)1., descW);
- MORSE_zgemm_Tile( MorseNoTrans, MorseNoTrans,
- (MORSE_Complex64_t)-1., descA2, descA1,
+ MORSE_zgemm_Tile( MorseNoTrans, MorseNoTrans,
+ (MORSE_Complex64_t)-1., descA2, descA1,
(MORSE_Complex64_t)1., descW);
Anorm = MORSE_zlange_Tile(MorseInfNorm, descA1, W);
Ainvnorm = MORSE_zlange_Tile(MorseInfNorm, descA2, W);
Rnorm = MORSE_zlange_Tile(MorseInfNorm, descW, W);
-
- dparam[IPARAM_ANORM] = Anorm;
- dparam[IPARAM_BNORM] = Ainvnorm;
+
+ dparam[IPARAM_ANORM] = Anorm;
+ dparam[IPARAM_BNORM] = Ainvnorm;
result = Rnorm / ( (Anorm*Ainvnorm)*descA1->m*eps ) ;
dparam[IPARAM_RES] = Rnorm;
if ( isnan(Ainvnorm) || isinf(Ainvnorm) || isnan(result) || isinf(result) || (result > 60.0) ) {
- dparam[IPARAM_XNORM] = -1.;
+ dparam[IPARAM_XNORM] = -1.;
}
else{
dparam[IPARAM_XNORM] = 0.;
@@ -129,12 +129,12 @@ static int check_getri_inverse(MORSE_desc_t *descA1, MORSE_desc_t *descA2, int *
}
static int
-RunTest(int *iparam, double *dparam, morse_time_t *t_)
+RunTest(int *iparam, double *dparam, morse_time_t *t_)
{
MORSE_desc_t descW;
int ret = 0;
PASTE_CODE_IPARAM_LOCALS( iparam );
-
+
if ( M != N ) {
fprintf(stderr, "This timing works only with M == N\n");
return -1;
@@ -155,69 +155,65 @@ RunTest(int *iparam, double *dparam, morse_time_t *t_)
/* MORSE ZGETRF / ZTRTRI / ZTRSMRV */
{
#if defined(TRACE_BY_SEQUENCE)
- MORSE_sequence_t *sequence[4];
- MORSE_request_t request[4] = { MORSE_REQUEST_INITIALIZER,
- MORSE_REQUEST_INITIALIZER,
- MORSE_REQUEST_INITIALIZER,
- MORSE_REQUEST_INITIALIZER };
-
- MORSE_Sequence_Create(&sequence[0]);
- MORSE_Sequence_Create(&sequence[1]);
- MORSE_Sequence_Create(&sequence[2]);
- MORSE_Sequence_Create(&sequence[3]);
+ MORSE_sequence_t *sequence;
+ MORSE_request_t request[4] = { MORSE_REQUEST_INITIALIZER,
+ MORSE_REQUEST_INITIALIZER,
+ MORSE_REQUEST_INITIALIZER,
+ MORSE_REQUEST_INITIALIZER };
+
+ MORSE_Sequence_Create(&sequence);
if ( ! iparam[IPARAM_ASYNC] ) {
-
+
START_TIMING();
- MORSE_zgetrf_Tile_Async(descA, piv, sequence[0], &request[0]);
- MORSE_Sequence_Wait(sequence[0]);
-
- MORSE_ztrtri_Tile_Async(MorseUpper, MorseNonUnit, descA, sequence[1], &request[1]);
- MORSE_Sequence_Wait(sequence[1]);
-
- MORSE_ztrsmrv_Tile_Async(MorseRight, MorseLower, MorseNoTrans, MorseUnit,
- (MORSE_Complex64_t) 1.0, descA, &descW, sequence[2], &request[2]);
- MORSE_Sequence_Wait(sequence[2]);
-
- MORSE_zlaswpc_Tile_Async(descA, 1, descA->m, piv, -1, sequence[3], &request[3]);
- MORSE_Sequence_Wait(sequence[3]);
- MORSE_Desc_Getoncpu( descA );
+ MORSE_zgetrf_Tile_Async( descA, piv, sequence, &request[0] );
+ MORSE_Sequence_Wait(sequence);
+
+ MORSE_ztrtri_Tile_Async( MorseUpper, MorseNonUnit, descA, sequence, &request[1] );
+ MORSE_Sequence_Wait(sequence);
+
+ MORSE_ztrsmrv_Tile_Async( MorseRight, MorseLower, MorseNoTrans, MorseUnit,
+ (MORSE_Complex64_t) 1.0, descA, &descW,
+ sequence, &request[2] );
+ MORSE_Sequence_Wait(sequence);
+
+ MORSE_zlaswpc_Tile_Async( descA, 1, descA->m, piv, -1,
+ sequence, &request[3] );
+ MORSE_Sequence_Wait(sequence);
+ MORSE_Desc_Flush( descA, sequence );
STOP_TIMING();
} else {
START_TIMING();
- MORSE_zgetrf_Tile_Async( descA, piv, sequence[0], &request[0]);
- MORSE_ztrtri_Tile_Async( MorseUpper, MorseNonUnit,
- descA, sequence[1], &request[1]);
- MORSE_ztrsmrv_Tile_Async(MorseRight, MorseLower, MorseNoTrans, MorseUnit,
- (MORSE_Complex64_t) 1.0,
- descA, &descW, sequence[2], &request[2]);
- MORSE_zlaswpc_Tile_Async(descA, 1, descA->m, piv, -1,
- sequence[3], &request[3]);
-
+ MORSE_zgetrf_Tile_Async( descA, piv, sequence, &request[0]);
+ MORSE_ztrtri_Tile_Async( MorseUpper, MorseNonUnit,
+ descA, sequence, &request[1] );
+ MORSE_ztrsmrv_Tile_Async( MorseRight, MorseLower, MorseNoTrans, MorseUnit,
+ (MORSE_Complex64_t) 1.0,
+ descA, &descW, sequence, &request[2] );
+ MORSE_zlaswpc_Tile_Async( descA, 1, descA->m, piv, -1,
+ sequence, &request[3] );
+
/* Wait for everything */
- MORSE_Sequence_Wait(sequence[0]);
- MORSE_Sequence_Wait(sequence[1]);
- MORSE_Sequence_Wait(sequence[2]);
- MORSE_Sequence_Wait(sequence[3]);
- MORSE_Desc_Getoncpu( descA );
+ MORSE_Sequence_Wait( sequence );
+ MORSE_Desc_Flush( descA, sequence );
STOP_TIMING();
-
+
}
MORSE_Sequence_Destroy(sequence[0]);
MORSE_Sequence_Destroy(sequence[1]);
MORSE_Sequence_Destroy(sequence[2]);
MORSE_Sequence_Destroy(sequence[3]);
-
+
#else
if ( ! iparam[IPARAM_ASYNC] ) {
START_TIMING();
MORSE_zgetrf_Tile(descA, piv);
MORSE_ztrtri_Tile(MorseUpper, MorseNonUnit, descA);
- MORSE_ztrsmrv_Tile(MorseRight, MorseLower, MorseNoTrans, MorseUnit,
+ MORSE_ztrsmrv_Tile(MorseRight, MorseLower, MorseNoTrans, MorseUnit,
(MORSE_Complex64_t) 1.0, descA, &descW);
MORSE_zlaswpc_Tile(descA, 1, descA->m, piv, -1);
STOP_TIMING();
@@ -225,28 +221,28 @@ RunTest(int *iparam, double *dparam, morse_time_t *t_)
} else {
MORSE_sequence_t *sequence;
- MORSE_request_t request[2] = { MORSE_REQUEST_INITIALIZER,
+ MORSE_request_t request[2] = { MORSE_REQUEST_INITIALIZER,
MORSE_REQUEST_INITIALIZER };
MORSE_Sequence_Create(&sequence);
-
+
START_TIMING();
MORSE_zgetrf_Tile_Async(descA, piv, sequence, &request[0]);
MORSE_zgetri_Tile_Async(descA, piv, &descW, sequence, &request[1]);
MORSE_Sequence_Wait(sequence);
- MORSE_Desc_Getoncpu( descA );
+ MORSE_Desc_Flush( descA, sequence );
STOP_TIMING();
-
- MORSE_Sequence_Destroy(sequence);
+
+ MORSE_Sequence_Destroy(sequence);
}
#endif
}
-
+
/* Check the solution */
if ( check )
{
ret = check_getri_inverse(descA2, descA, piv, dparam);
-
+
PASTE_CODE_FREE_MATRIX( descA2 );
}
@@ -256,4 +252,3 @@ RunTest(int *iparam, double *dparam, morse_time_t *t_)
return ret;
}
-
diff --git a/timing/time_zpotri_tile.c b/timing/time_zpotri_tile.c
index 07ef6007f8a71bf8fd275301ce0c3af344ca26b2..3a2d59e5b6aebc91b591a7427cbd7f3394c551ea 100644
--- a/timing/time_zpotri_tile.c
+++ b/timing/time_zpotri_tile.c
@@ -50,40 +50,36 @@ RunTest(int *iparam, double *dparam, morse_time_t *t_)
*/
#if defined(TRACE_BY_SEQUENCE)
{
- MORSE_sequence_t *sequence[3];
+ MORSE_sequence_t *sequence;
MORSE_request_t request[3] = { MORSE_REQUEST_INITIALIZER,
MORSE_REQUEST_INITIALIZER,
MORSE_REQUEST_INITIALIZER };
- MORSE_Sequence_Create(&sequence[0]);
- MORSE_Sequence_Create(&sequence[1]);
- MORSE_Sequence_Create(&sequence[2]);
+ MORSE_Sequence_Create(&sequence);
if ( ! iparam[IPARAM_ASYNC] ) {
START_TIMING();
- MORSE_zpotrf_Tile_Async(uplo, descA, sequence[0], &request[0]);
- MORSE_Sequence_Wait(sequence[0]);
+ MORSE_zpotrf_Tile_Async(uplo, descA, sequence, &request[0]);
+ MORSE_Sequence_Wait(sequence);
- MORSE_ztrtri_Tile_Async(uplo, MorseNonUnit, descA, sequence[1], &request[1]);
- MORSE_Sequence_Wait(sequence[1]);
+ MORSE_ztrtri_Tile_Async(uplo, MorseNonUnit, descA, sequence, &request[1]);
+ MORSE_Sequence_Wait(sequence);
- MORSE_zlauum_Tile_Async(uplo, descA, sequence[2], &request[2]);
- MORSE_Sequence_Wait(sequence[2]);
- MORSE_Desc_Getoncpu( descA );
+ MORSE_zlauum_Tile_Async(uplo, descA, sequence, &request[2]);
+ MORSE_Sequence_Wait(sequence);
+ MORSE_Desc_Flush( descA, sequence );
STOP_TIMING();
} else {
START_TIMING();
- MORSE_zpotrf_Tile_Async(uplo, descA, sequence[0], &request[0]);
- MORSE_ztrtri_Tile_Async(uplo, MorseNonUnit, descA, sequence[1], &request[1]);
- MORSE_zlauum_Tile_Async(uplo, descA, sequence[2], &request[2]);
-
- MORSE_Sequence_Wait(sequence[0]);
- MORSE_Sequence_Wait(sequence[1]);
- MORSE_Sequence_Wait(sequence[2]);
- MORSE_Desc_Getoncpu( descA );
+ MORSE_zpotrf_Tile_Async(uplo, descA, sequence, &request[0]);
+ MORSE_ztrtri_Tile_Async(uplo, MorseNonUnit, descA, sequence, &request[1]);
+ MORSE_zlauum_Tile_Async(uplo, descA, sequence, &request[2]);
+
+ MORSE_Sequence_Wait(sequence);
+ MORSE_Desc_Flush( descA, sequence );
STOP_TIMING();
}
@@ -113,7 +109,7 @@ RunTest(int *iparam, double *dparam, morse_time_t *t_)
MORSE_zpotrf_Tile_Async(uplo, descA, sequence, &request[0]);
MORSE_zpotri_Tile_Async(uplo, descA, sequence, &request[1]);
MORSE_Sequence_Wait(sequence);
- MORSE_Desc_Getoncpu( descA );
+ MORSE_Desc_Flush( descA, sequence );
STOP_TIMING();
MORSE_Sequence_Destroy(sequence);
diff --git a/timing/time_zsytrf_tile.c b/timing/time_zsytrf_tile.c
index 423b77b7068c5ee9e63b085c97294cd7db411677..54cf336554d29761dbf573a0480d5b2523cb184b 100644
--- a/timing/time_zsytrf_tile.c
+++ b/timing/time_zsytrf_tile.c
@@ -58,17 +58,16 @@ RunTest(int *iparam, double *dparam, morse_time_t *t_)
dparam[IPARAM_ANORM] = MORSE_zlange_Tile(MorseInfNorm, descAC);
dparam[IPARAM_BNORM] = MORSE_zlange_Tile(MorseInfNorm, descB);
dparam[IPARAM_XNORM] = MORSE_zlange_Tile(MorseInfNorm, descX);
- MORSE_zgemm_Tile( MorseNoTrans, MorseNoTrans, 1.0, descAC, descX, -1.0, descB );
- dparam[IPARAM_RES] = MORSE_zlange_Tile(MorseInfNorm, descB);
+ MORSE_zgemm_Tile( MorseNoTrans, MorseNoTrans, 1.0, descAC, descX, -1.0, descB );
+ dparam[IPARAM_RES] = MORSE_zlange_Tile(MorseInfNorm, descB);
PASTE_CODE_FREE_MATRIX( descB );
PASTE_CODE_FREE_MATRIX( descAC );
PASTE_CODE_FREE_MATRIX( descX );
- }
-
+ }
+
PASTE_CODE_FREE_MATRIX( descA );
- MORSE_TASK_flush_all();
return 0;
}