diff --git a/compute/pzgeqrf.c b/compute/pzgeqrf.c
index 849d7a00568a25988fb8c66819cb2068209a1a22..f5b3b744c50d7b8b64412b0d0b7842557ab1a3bf 100644
--- a/compute/pzgeqrf.c
+++ b/compute/pzgeqrf.c
@@ -33,9 +33,9 @@
#define A(m,n) A, m, n
#define T(m,n) T, m, n
#if defined(CHAMELEON_COPY_DIAG)
-#define D(k) D, k, 0
+#define D(k) D, k, 0
#else
-#define D(k) A, k, k
+#define D(k) D, k, k
#endif
/*******************************************************************************
@@ -62,6 +62,10 @@ void morse_pzgeqrf(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D,
ib = MORSE_IB;
+ if ( D == NULL ) {
+ D = A;
+ }
+
/*
* zgeqrt = A->nb * (ib+1)
* zunmqr = A->nb * ib
@@ -122,28 +126,44 @@ void morse_pzgeqrf(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D,
T(k, k), T->mb,
A(k, n), ldak);
}
+
for (m = k+1; m < A->mt; m++) {
tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
ldam = BLKLDD(A, m);
- MORSE_TASK_ztsqrt(
+
+ RUNTIME_data_migrate( sequence, A(k, k),
+ A->get_rankof( A, m, k ) );
+
+ MORSE_TASK_ztpqrt(
&options,
- tempmm, tempkn, ib, T->nb,
+ tempmm, tempkn, 0, ib, T->nb,
A(k, k), ldak,
A(m, k), ldam,
T(m, k), T->mb);
+
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- MORSE_TASK_ztsmqr(
+
+ RUNTIME_data_migrate( sequence, A(k, n),
+ A->get_rankof( A, m, n ) );
+
+ MORSE_TASK_ztpmqrt(
&options,
MorseLeft, MorseConjTrans,
- A->mb, tempnn, tempmm, tempnn, A->nb, ib, T->nb,
- A(k, n), ldak,
- A(m, n), ldam,
+ tempmm, tempnn, A->nb, 0, ib, T->nb,
A(m, k), ldam,
- T(m, k), T->mb);
+ T(m, k), T->mb,
+ A(k, n), ldak,
+ A(m, n), ldam);
}
}
+ /* Restore the original location of the tiles */
+ for (n = k; n < A->nt; n++) {
+ RUNTIME_data_migrate( sequence, A(k, n),
+ A->get_rankof( A, k, n ) );
+ }
+
RUNTIME_iteration_pop(morse);
}
diff --git a/compute/pzgeqrf_param.c b/compute/pzgeqrf_param.c
index 3845f4627f459ebeda7733d1e976589f30e6c430..99027c88fd645fe1f314364cea7ad19b6d927375 100644
--- a/compute/pzgeqrf_param.c
+++ b/compute/pzgeqrf_param.c
@@ -26,14 +26,9 @@
#include <stdlib.h>
#include "libhqr.h"
-#define A(m,n) A, (m), (n)
-#define TS(m,n) TS, (m), (n)
-#define TT(m,n) TT, (m), (n)
-#if defined(CHAMELEON_COPY_DIAG)
-#define D(m,n) D, (m), (n)
-#else
-#define D(m,n) A, (m), (n)
-#endif
+#define A(m,n) A, (m), (n)
+#define T(m,n) T, (m), (n)
+#define D(m,n) D, (m), (n)
/**
* Parallel tile QR factorization (reduction Householder) - dynamic scheduling
@@ -44,11 +39,12 @@ void morse_pzgeqrf_param( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
{
MORSE_context_t *morse;
MORSE_option_t options;
+ MORSE_desc_t *T;
size_t ws_worker = 0;
size_t ws_host = 0;
int k, m, n, i, p;
- int K;
+ int K, L;
int ldap, ldam;
int tempkmin, tempkn, tempnn, tempmm;
int ib;
@@ -61,6 +57,10 @@ void morse_pzgeqrf_param( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
ib = MORSE_IB;
+ if ( D == NULL ) {
+ D = A;
+ }
+
/*
* zgeqrt = A->nb * (ib+1)
* zunmqr = A->nb * ib
@@ -81,8 +81,7 @@ void morse_pzgeqrf_param( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
ws_worker = chameleon_max( ws_worker, ib * A->nb * 2 );
#endif
- /* Initialisation of tiles */
-
+ /* Initialisation of temporary tiles array */
tiles = (int*)calloc(qrtree->mt, sizeof(int));
ws_worker *= sizeof(MORSE_Complex64_t);
@@ -104,11 +103,13 @@ void morse_pzgeqrf_param( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
tempkmin = chameleon_min(tempmm, tempkn);
ldam = BLKLDD(A, m);
+ T = TS;
+
MORSE_TASK_zgeqrt(
&options,
- tempmm, tempkn, ib, TS->nb,
- A( m, k), ldam,
- TS(m, k), TS->mb);
+ tempmm, tempkn, ib, T->nb,
+ A(m, k), ldam,
+ T(m, k), T->mb);
if ( k < (A->nt-1) ) {
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
@@ -130,10 +131,10 @@ void morse_pzgeqrf_param( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
MORSE_TASK_zunmqr(
&options,
MorseLeft, MorseConjTrans,
- tempmm, tempnn, tempkmin, ib, TS->nb,
- D( m, k), ldam,
- TS(m, k), TS->mb,
- A( m, n), ldam);
+ tempmm, tempnn, tempkmin, ib, T->nb,
+ D(m, k), ldam,
+ T(m, k), T->mb,
+ A(m, n), ldam);
}
}
@@ -149,54 +150,56 @@ void morse_pzgeqrf_param( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
ldam = BLKLDD(A, m);
/* Tiles killed is a TS */
- if(qrtree->gettype(qrtree, k, m) == 0){
- MORSE_TASK_ztsqrt(
- &options,
- tempmm, tempkn, ib, TS->nb,
- A( p, k), ldap,
- A( m, k), ldam,
- TS(m, k), TS->mb);
-
- for (n = k+1; n < A->nt; n++) {
- tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- MORSE_TASK_ztsmqr(
- &options,
- MorseLeft, MorseConjTrans,
- A->nb, tempnn, tempmm, tempnn, A->nb, ib, TS->nb,
- A( p, n), ldap,
- A( m, n), ldam,
- A( m, k), ldam,
- TS(m, k), TS->mb);
- }
+ if (qrtree->gettype(qrtree, k, m) == 0) {
+ T = TS;
+ L = 0;
}
-
- /* Tiles killed is a TT */
else {
- MORSE_TASK_zttqrt(
+ T = TT;
+ L = tempmm;
+ }
+
+ RUNTIME_data_migrate( sequence, A(p, k),
+ A->get_rankof( A, m, k ) );
+ RUNTIME_data_migrate( sequence, A(m, k),
+ A->get_rankof( A, m, k ) );
+
+ MORSE_TASK_ztpqrt(
+ &options,
+ tempmm, tempkn, L, ib, T->nb,
+ A(p, k), ldap,
+ A(m, k), ldam,
+ T(m, k), T->mb);
+
+ for (n = k+1; n < A->nt; n++) {
+ tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
+
+ RUNTIME_data_migrate( sequence, A(p, n),
+ A->get_rankof( A, m, n ) );
+ RUNTIME_data_migrate( sequence, A(m, n),
+ A->get_rankof( A, m, n ) );
+
+ MORSE_TASK_ztpmqrt(
&options,
- tempmm, tempkn, ib, TT->nb,
- A( p, k), ldap,
- A( m, k), ldam,
- TT(m, k), TT->mb);
-
- for (n = k+1; n < A->nt; n++) {
- tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- MORSE_TASK_zttmqr(
- &options,
- MorseLeft, MorseConjTrans,
- A->mb, tempnn, tempmm, tempnn, A->nb, ib, TT->nb,
- A( p, n), ldap,
- A( m, n), ldam,
- A( m, k), ldam,
- TT(m, k), TT->mb);
- }
+ MorseLeft, MorseConjTrans,
+ tempmm, tempnn, A->nb, L, ib, T->nb,
+ A(m, k), ldam,
+ T(m, k), T->mb,
+ A(p, n), ldap,
+ A(m, n), ldam);
}
}
+
+ /* Restore the original location of the tiles */
+ for (n = k; n < A->nt; n++) {
+ RUNTIME_data_migrate( sequence, A(k, n),
+ A->get_rankof( A, k, n ) );
+ }
+
RUNTIME_iteration_pop(morse);
}
free(tiles);
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
- (void)D;
}
diff --git a/compute/pzgeqrfrh.c b/compute/pzgeqrfrh.c
index 30d6a1273cb8092a41348a6d50c9e222f6773891..1024a6a2370063dd88d3c44bfcf1f77100bef41b 100644
--- a/compute/pzgeqrfrh.c
+++ b/compute/pzgeqrfrh.c
@@ -98,6 +98,7 @@ void morse_pzgeqrfrh(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D, int BS,
tempMm = M == A->mt-1 ? A->m-M*A->mb : A->mb;
tempkmin = chameleon_min(tempMm, tempkn);
ldaM = BLKLDD(A, M);
+
MORSE_TASK_zgeqrt(
&options,
tempMm, tempkn, ib, T->nb,
@@ -129,26 +130,35 @@ void morse_pzgeqrfrh(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D, int BS,
T(M, k), T->mb,
A(M, n), ldaM);
}
+
for (m = M+1; m < chameleon_min(M+BS, A->mt); m++) {
tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
ldam = BLKLDD(A, m);
- MORSE_TASK_ztsqrt(
+
+ RUNTIME_data_migrate( sequence, A(M, k),
+ A->get_rankof( A, m, k ) );
+
+ MORSE_TASK_ztpqrt(
&options,
- tempmm, tempkn, ib, T->nb,
+ tempmm, tempkn, 0, ib, T->nb,
A(M, k), ldaM,
A(m, k), ldam,
T(m, k), T->mb);
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- MORSE_TASK_ztsmqr(
+
+ RUNTIME_data_migrate( sequence, A(M, n),
+ A->get_rankof( A, m, n ) );
+
+ MORSE_TASK_ztpmqrt(
&options,
MorseLeft, MorseConjTrans,
- A->nb, tempnn, tempmm, tempnn, A->nb, ib, T->nb,
- A(M, n), ldaM,
- A(m, n), ldam,
+ tempmm, tempnn, A->nb, 0, ib, T->nb,
A(m, k), ldam,
- T(m, k), T->mb);
+ T(m, k), T->mb,
+ A(M, n), ldaM,
+ A(m, n), ldam);
}
}
}
@@ -157,26 +167,45 @@ void morse_pzgeqrfrh(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *D, int BS,
tempMRDm = M+RD == A->mt-1 ? A->m-(M+RD)*A->mb : A->mb;
ldaM = BLKLDD(A, M );
ldaMRD = BLKLDD(A, M+RD);
- MORSE_TASK_zttqrt(
+
+ RUNTIME_data_migrate( sequence, A(M, k),
+ A->get_rankof( A, M+RD, k ) );
+ RUNTIME_data_migrate( sequence, A(M+RD, k),
+ A->get_rankof( A, M+RD, k ) );
+
+ MORSE_TASK_ztpqrt(
&options,
- tempMRDm, tempkn, ib, T->nb,
+ tempMRDm, tempkn, tempMRDm, ib, T->nb,
A (M , k), ldaM,
A (M+RD, k), ldaMRD,
T2(M+RD, k), T->mb);
for (n = k+1; n < A->nt; n++) {
tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- MORSE_TASK_zttmqr(
+
+ RUNTIME_data_migrate( sequence, A(M, n),
+ A->get_rankof( A, M+RD, n ) );
+ RUNTIME_data_migrate( sequence, A(M+RD, n),
+ A->get_rankof( A, M+RD, n ) );
+
+ MORSE_TASK_ztpmqrt(
&options,
MorseLeft, MorseConjTrans,
- A->nb, tempnn, tempMRDm, tempnn, A->nb, ib, T->nb,
- A (M, n), ldaM,
- A (M+RD, n), ldaMRD,
+ tempMRDm, tempnn, A->nb, tempMRDm, ib, T->nb,
A (M+RD, k), ldaMRD,
- T2(M+RD, k), T->mb);
+ T2(M+RD, k), T->mb,
+ A (M, n), ldaM,
+ A (M+RD, n), ldaMRD);
}
}
}
+
+ /* Restore the original location of the tiles */
+ for (n = k; n < A->nt; n++) {
+ RUNTIME_data_migrate( sequence, A(k, n),
+ A->get_rankof( A, k, n ) );
+ }
+
RUNTIME_iteration_pop(morse);
}