Skip to content

Chameleon
Chameleon
Commit 2d2e102a authored by PRUVOST Florent's avatar PRUVOST Florent
Browse files
Options

fix the problem of diag tile copy appearing in QUARK case

parent 97fe07d3
Hide whitespace changes
Inline Side-by-side
Showing with 114 additions and 6 deletions
compute/pzgelqf.c
View file @ 2d2e102a
......@@ -32,7 +32,11 @@
#define A(m,n) A, m, n
#define T(m,n) T, m, n
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(k) DIAG, k, 0
#else
#define DIAG(k) A, k, k
#endif
/***************************************************************************//**
* Parallel tile LQ factorization - dynamic scheduling
......@@ -112,11 +116,13 @@ void morse_pzgelqf(MORSE_desc_t *A, MORSE_desc_t *T,
A(k, k), ldak,
T(k, k), T->mb);
if ( k < (A->mt-1) ) {
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, A->mb, A->nb, A->nb,
A(k, k), ldak,
DIAG(k), A->mb );
DIAG(k), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzgelqfrh.c
View file @ 2d2e102a
......@@ -36,7 +36,11 @@
#define A(m,n) A, (m), (n)
#define T(m,n) T, (m), (n)
#define T2(m,n) T, (m), (n)+A->nt
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(m,n) DIAG, ((n)/BS), 0
#else
#define DIAG(m,n) A, (m), (n)
#endif
/***************************************************************************//**
* Parallel tile LQ factorization (reduction Householder) - dynamic scheduling
......@@ -117,11 +121,13 @@ void morse_pzgelqfrh(MORSE_desc_t *A, MORSE_desc_t *T, int BS,
A(k, N), ldak,
T(k, N), T->mb);
if ( k < (A->mt-1) ) {
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkm, tempNn, A->nb,
A(k, N), ldak,
DIAG(k, N), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzgeqrf.c
View file @ 2d2e102a
......@@ -32,7 +32,11 @@
#define A(m,n) A, m, n
#define T(m,n) T, m, n
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(k) DIAG, k, 0
#else
#define DIAG(k) A, k, k
#endif
/***************************************************************************//**
* Parallel tile QR factorization - dynamic scheduling
......@@ -107,11 +111,13 @@ void morse_pzgeqrf(MORSE_desc_t *A, MORSE_desc_t *T,
A(k, k), ldak,
T(k, k), T->mb);
if ( k < (A->nt-1) ) {
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, A->mb, A->nb, A->nb,
A(k, k), ldak,
DIAG(k), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzgeqrfrh.c
View file @ 2d2e102a
......@@ -34,7 +34,11 @@
#define A(m,n) A, (m), (n)
#define T(m,n) T, (m), (n)
#define T2(m,n) T, (m), ((n)+A->nt)
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(m,n) DIAG, ((m)/BS), 0
#else
#define DIAG(m,n) A, (m), (n)
#endif
/***************************************************************************//**
* Parallel tile QR factorization (reduction Householder) - dynamic scheduling
......@@ -116,11 +120,13 @@ void morse_pzgeqrfrh(MORSE_desc_t *A, MORSE_desc_t *T, int BS,
A(M, k), ldaM,
T(M, k), T->mb);
if ( k < (A->nt-1) ) {
MORSE_TASK_zlacpy(
&options,
MorseLower, tempMm, A->nb, A->nb,
A(M, k), ldaM,
DIAG(M, k), ldaM );
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempMm, A->nb, A->nb,
A(M, k), ldaM,
DIAG(M, k), ldaM );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzgetrf_incpiv.c
View file @ 2d2e102a
......@@ -33,7 +33,11 @@
#include "common.h"
#define A(_m_,_n_) A, _m_, _n_
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(_k_) DIAG, _k_, 0
#else
#define DIAG(_k_) A, _k_, _k_
#endif
#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_))])
......@@ -86,11 +90,13 @@ void morse_pzgetrf_incpiv(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
k == A->mt-1, A->nb*k);
if ( k < (minMNT-1) ) {
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpperLower, tempkm, tempkn, A->nb,
A(k, k), ldak,
DIAG(k), ldak);
#endif
}
for (n = k+1; n < A->nt; n++) {
......
compute/pzunglq.c
View file @ 2d2e102a
......@@ -33,7 +33,11 @@
#define A(m,n) A, m, n
#define Q(m,n) Q, m, n
#define T(m,n) T, m, n
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(k) DIAG, k, 0
#else
#define DIAG(k) A, k, k
#endif
/***************************************************************************//**
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
......@@ -112,11 +116,13 @@ void morse_pzunglq(MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T,
T(k, n), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempkn, A->nb,
A(k, k), ldak,
DIAG(k), A->mb );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzunglqrh.c
View file @ 2d2e102a
......@@ -33,7 +33,11 @@
#define Q(m,n) Q, (m), (n)
#define T(m,n) T, (m), (n)
#define T2(m,n) T, (m), (n)+(A->nt)
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(m,n) DIAG, ((n)/BS), 0
#else
#define DIAG(m,n) A, (m), (n)
#endif
/**
* Parallel construction of Q using tile V (application to identity;
......@@ -135,11 +139,13 @@ void morse_pzunglqrh(MORSE_desc_t *A, MORSE_desc_t *Q,
T(k, n), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempNn, A->nb,
A(k, N), ldak,
DIAG(k, N), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzungqr.c
View file @ 2d2e102a
......@@ -33,7 +33,11 @@
#define A(m,n) A, m, n
#define Q(m,n) Q, m, n
#define T(m,n) T, m, n
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(k) DIAG, k, 0
#else
#define DIAG(k) A, k, k
#endif
/***************************************************************************//**
* Parallel construction of Q using tile V (application to identity) - dynamic scheduling
......@@ -108,11 +112,13 @@ void morse_pzungqr(MORSE_desc_t *A, MORSE_desc_t *Q, MORSE_desc_t *T,
T(m, k), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempkm, tempkmin, A->nb,
A(k, k), ldak,
DIAG(k), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzungqrrh.c
View file @ 2d2e102a
......@@ -35,7 +35,11 @@
#define Q(m,n) Q, (m), (n)
#define T(m,n) T, (m), (n)
#define T2(m,n) T, (m), (n)+(A->nt)
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(m,n) DIAG, ((m)/BS), 0
#else
#define DIAG(m,n) A, (m), (n)
#endif
/**
* Parallel construction of Q using tile V (application to identity;
......@@ -141,11 +145,13 @@ void morse_pzungqrrh(MORSE_desc_t *A, MORSE_desc_t *Q,
T(m, k), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempMm, tempkmin, A->nb,
A(M, k), ldaM,
DIAG(M, k), ldaM );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzunmlq.c
View file @ 2d2e102a
......@@ -34,7 +34,11 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
#define T(m,n) T, m, n
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(k) DIAG, k, 0
#else
#define DIAG(k) A, k, k
#endif
/***************************************************************************//**
* Parallel application of Q using tile V - LQ factorization - dynamic scheduling
......@@ -103,11 +107,13 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempkm, A->nb,
A(k, k), ldak,
DIAG(k), A->mb );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -166,11 +172,13 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
T(k, m), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempkm, A->nb,
A(k, k), ldak,
DIAG(k), A->mb );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -215,11 +223,13 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
T(k, n), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempkn, A->nb,
A(k, k), ldak,
DIAG(k), A->mb );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -248,11 +258,13 @@ void morse_pzunmlq(MORSE_enum side, MORSE_enum trans,
tempkn = k == B->nt -1 ? B->n -k*B->nb : B->nb;
tempkmin = k == minMT-1 ? minM-k*A->mb : A->mb;
ldak = BLKLDD(A, k);
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempkn, A->nb,
A(k, k), ldak,
DIAG(k), A->mb );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzunmlqrh.c
View file @ 2d2e102a
......@@ -35,7 +35,11 @@
#define B(m,n) B, (m), (n)
#define T(m,n) T, (m), (n)
#define T2(m,n) T, (m), (n)+A->nt
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(m,n) DIAG, ((n)/BS), 0
#else
#define DIAG(m,n) A, (m), (n)
#endif
/***************************************************************************//**
* Parallel application of Q using tile V - LQ factorization (reduction
......@@ -106,11 +110,13 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
tempkmin = min(tempkm,tempNn);
ldaN = BLKLDD(A, N);
ldbN = BLKLDD(B, N);
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempNn, A->nb,
A(k, N), ldak,
DIAG(k, N), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -216,11 +222,13 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
T(k, m), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempNn, A->nb,
A(k, N), ldak,
DIAG(k, N), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -291,11 +299,13 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
T(k, n), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempNn, A->nb,
A(k, N), ldak,
DIAG(k, N), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -328,11 +338,13 @@ void morse_pzunmlqrh(MORSE_enum side, MORSE_enum trans,
tempNn = N == A->nt-1 ? A->n-N*A->nb : A->nb;
tempkmin = min(tempkm,tempNn);
ldaN = BLKLDD(A, N);
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseUpper, tempkmin, tempNn, A->nb,
A(k, N), ldaN,
DIAG(k, N), ldaN );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzunmqr.c
View file @ 2d2e102a
......@@ -34,7 +34,11 @@
#define A(m,n) A, m, n
#define B(m,n) B, m, n
#define T(m,n) T, m, n
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(k) DIAG, k, 0
#else
#define DIAG(k) A, k, k
#endif
/***************************************************************************//**
* Parallel application of Q using tile V - QR factorization - dynamic scheduling
......@@ -103,11 +107,13 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
ldak = BLKLDD(A, k);
ldbk = BLKLDD(B, k);
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempkm, tempkmin, A->nb,
A(k, k), ldak,
DIAG(k), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -168,11 +174,13 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
T(m, k), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempkm, tempkmin, A->nb,
A(k, k), ldak,
DIAG(k), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -219,11 +227,13 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
T(n, k), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempkn, tempkmin, A->nb,
A(k, k), ldak,
DIAG(k), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -252,11 +262,13 @@ void morse_pzunmqr(MORSE_enum side, MORSE_enum trans,
tempkn = k == B->nt-1 ? B->n-k*B->nb : B->nb;
tempkmin = k == minMT-1 ? minM-k*A->nb : A->nb;
ldak = BLKLDD(A, k);
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempkn, tempkmin, A->nb,
A(k, k), ldak,
DIAG(k), ldak );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
compute/pzunmqrrh.c
View file @ 2d2e102a
......@@ -35,7 +35,11 @@
#define B(m,n) B, (m), (n)
#define T(m,n) T, (m), (n)
#define T2(m,n) T, (m), ((n)+A->nt)
#if defined(CHAMELEON_COPY_DIAG)
#define DIAG(m,n) DIAG, ((m)/BS), 0
#else
#define DIAG(m,n) A, (m), (n)
#endif
/***************************************************************************//**
* Parallel application of Q using tile V - QR factorization (reduction
......@@ -105,11 +109,13 @@ void morse_pzunmqrrh(MORSE_enum side, MORSE_enum trans,
tempkmin = min(tempMm, tempkn);
ldaM = BLKLDD(A, M);
ldbM = BLKLDD(B, M);
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempMm, tempkmin, A->nb,
A(M, k), ldaM,
DIAG(M, k), ldaM );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -218,11 +224,13 @@ void morse_pzunmqrrh(MORSE_enum side, MORSE_enum trans,
T(m, k), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempMm, tempkmin, A->nb,
A(M, k), ldaM,
DIAG(M, k), ldaM );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -295,11 +303,13 @@ void morse_pzunmqrrh(MORSE_enum side, MORSE_enum trans,
T(n, k), T->mb);
}
}
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempMm, tempkmin, A->nb,
A(M, k), ldaM,
DIAG(M, k), ldaM );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......@@ -331,11 +341,13 @@ void morse_pzunmqrrh(MORSE_enum side, MORSE_enum trans,
tempMm = M == A->mt-1 ? A->m-M*A->mb : A->mb;
tempkmin = min(tempMm, tempkn);
ldaM = BLKLDD(A, M);
#if defined(CHAMELEON_COPY_DIAG)
MORSE_TASK_zlacpy(
&options,
MorseLower, tempMm, tempkmin, A->nb,
A(M, k), ldaM,
DIAG(M, k), ldaM );
#endif
#if defined(CHAMELEON_USE_MAGMA)
MORSE_TASK_zlaset(
&options,
......
control/common.h
View file @ 2d2e102a
......@@ -136,6 +136,12 @@
#define MORSE_MPI_SIZE morse->mpi_comm_size
#endif
/*******************************************************************************
* Activate copy of diagonal tile (StarPU only) for some tile algorithms (pz)
**/
#if defined(CHAMELEON_SCHED_STARPU)
#define CHAMELEON_COPY_DIAG
#endif
/*******************************************************************************
* IPT internal define
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment