From ec9ce5c03e32dfed0478f731ac26840fc44b4404 Mon Sep 17 00:00:00 2001
From: Mathieu Faverge <mathieu.faverge@inria.fr>
Date: Tue, 12 Oct 2021 12:30:17 +0200
Subject: [PATCH] compute: Prepare matrix generation algorithms toward
recursive descriptors
---
compute/pzlatms.c | 4 ++--
compute/pzplghe.c | 8 ++++----
compute/pzplgsy.c | 8 ++++----
compute/pzplrnt.c | 7 ++++---
compute/zplghe.c | 2 +-
compute/zplgsy.c | 2 +-
compute/zplrnt.c | 2 +-
control/compute_z.h | 6 +++---
8 files changed, 20 insertions(+), 19 deletions(-)
diff --git a/compute/pzlatms.c b/compute/pzlatms.c
index 7a1334d9d..ab0cbe473 100644
--- a/compute/pzlatms.c
+++ b/compute/pzlatms.c
@@ -181,7 +181,7 @@ void chameleon_pzlatms( cham_dist_t idist, unsigned long long int seed, cham_sym
/* U is of size A->m by min(A->m, A->n) */
chameleon_zdesc_copy_and_restrict( A, &descU, A->m, minmn );
- chameleon_pzplrnt( &descU, seed, sequence, request );
+ chameleon_pzplrnt( &descU, descU.m, 0, 0, seed, sequence, request );
/* Shift the seed to generate the next random unitary matrix */
#if !defined(CHAMELEON_SIMULATION)
@@ -257,7 +257,7 @@ void chameleon_pzlatms( cham_dist_t idist, unsigned long long int seed, cham_sym
/* V is of size min(A->m, A->n) by A->n */
chameleon_zdesc_copy_and_restrict( A, &descV, minmn, A->n );
- chameleon_pzplrnt( &descV, seed, sequence, request );
+ chameleon_pzplrnt( &descV, descV.m, 0, 0, seed, sequence, request );
/* Apply a QR factorization */
mat.mt = descV.mt;
diff --git a/compute/pzplghe.c b/compute/pzplghe.c
index d20b36a61..7a1f03120 100644
--- a/compute/pzplghe.c
+++ b/compute/pzplghe.c
@@ -28,7 +28,7 @@
* chameleon_pzplghe - Generate a random hermitian (positive definite if 'bump' is large enough) half-matrix by tiles.
*/
void chameleon_pzplghe( double bump, cham_uplo_t uplo, CHAM_desc_t *A,
- unsigned long long int seed,
+ int bigM, int m0, int n0, unsigned long long int seed,
RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
@@ -56,7 +56,7 @@ void chameleon_pzplghe( double bump, cham_uplo_t uplo, CHAM_desc_t *A,
INSERT_TASK_zplghe(
&options,
bump, tempmm, tempnn, A(m, n),
- A->m, m*A->mb, n*A->nb, seed );
+ bigM, m*A->mb + m0, n*A->nb + n0, seed );
}
}
break;
@@ -72,7 +72,7 @@ void chameleon_pzplghe( double bump, cham_uplo_t uplo, CHAM_desc_t *A,
INSERT_TASK_zplghe(
&options,
bump, tempmm, tempnn, A(m, n),
- A->m, m*A->mb, n*A->nb, seed );
+ bigM, m*A->mb + m0, n*A->nb + n0, seed );
}
}
break;
@@ -89,7 +89,7 @@ void chameleon_pzplghe( double bump, cham_uplo_t uplo, CHAM_desc_t *A,
INSERT_TASK_zplghe(
&options,
bump, tempmm, tempnn, A(m, n),
- A->m, m*A->mb, n*A->nb, seed );
+ bigM, m*A->mb + m0, n*A->nb + n0, seed );
}
}
}
diff --git a/compute/pzplgsy.c b/compute/pzplgsy.c
index 726543b5a..b9a035620 100644
--- a/compute/pzplgsy.c
+++ b/compute/pzplgsy.c
@@ -28,7 +28,7 @@
* chameleon_pzplgsy - Generate a random symmetric (positive definite if 'bump' is large enough) half-matrix by tiles.
*/
void chameleon_pzplgsy( CHAMELEON_Complex64_t bump, cham_uplo_t uplo, CHAM_desc_t *A,
- unsigned long long int seed,
+ int bigM, int m0, int n0, unsigned long long int seed,
RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
@@ -56,7 +56,7 @@ void chameleon_pzplgsy( CHAMELEON_Complex64_t bump, cham_uplo_t uplo, CHAM_desc_
INSERT_TASK_zplgsy(
&options,
bump, tempmm, tempnn, A(m, n),
- A->m, m*A->mb, n*A->nb, seed );
+ bigM, m*A->mb + m0, n*A->nb + n0, seed );
}
}
break;
@@ -72,7 +72,7 @@ void chameleon_pzplgsy( CHAMELEON_Complex64_t bump, cham_uplo_t uplo, CHAM_desc_
INSERT_TASK_zplgsy(
&options,
bump, tempmm, tempnn, A(m, n),
- A->m, m*A->mb, n*A->nb, seed );
+ bigM, m*A->mb + m0, n*A->nb + n0, seed );
}
}
break;
@@ -89,7 +89,7 @@ void chameleon_pzplgsy( CHAMELEON_Complex64_t bump, cham_uplo_t uplo, CHAM_desc_
INSERT_TASK_zplgsy(
&options,
bump, tempmm, tempnn, A(m, n),
- A->m, m*A->mb, n*A->nb, seed );
+ bigM, m*A->mb + m0, n*A->nb + n0, seed );
}
}
}
diff --git a/compute/pzplrnt.c b/compute/pzplrnt.c
index 77845a2ed..a0d08d9be 100644
--- a/compute/pzplrnt.c
+++ b/compute/pzplrnt.c
@@ -28,8 +28,9 @@
/**
* chameleon_pzplghe - Generate a random matrix by tiles.
*/
-void chameleon_pzplrnt( CHAM_desc_t *A, unsigned long long int seed,
- RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
+void chameleon_pzplrnt( CHAM_desc_t *A,
+ int bigM, int m0, int n0, unsigned long long int seed,
+ RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
@@ -52,7 +53,7 @@ void chameleon_pzplrnt( CHAM_desc_t *A, unsigned long long int seed,
INSERT_TASK_zplrnt(
&options,
tempmm, tempnn, A(m, n),
- A->m, m*A->mb, n*A->nb, seed );
+ bigM, m*A->mb + m0, n*A->nb + n0, seed );
}
}
RUNTIME_options_finalize(&options, chamctxt);
diff --git a/compute/zplghe.c b/compute/zplghe.c
index 28f849ea4..ea6a7364f 100644
--- a/compute/zplghe.c
+++ b/compute/zplghe.c
@@ -270,7 +270,7 @@ int CHAMELEON_zplghe_Tile_Async( double bump,
if (chameleon_min( A->m, A->n ) == 0)
return CHAMELEON_SUCCESS;
- chameleon_pzplghe( bump, uplo, A, seed, sequence, request );
+ chameleon_pzplghe( bump, uplo, A, A->m, A->i, A->j, seed, sequence, request );
return CHAMELEON_SUCCESS;
}
diff --git a/compute/zplgsy.c b/compute/zplgsy.c
index c09673df7..2f5921517 100644
--- a/compute/zplgsy.c
+++ b/compute/zplgsy.c
@@ -272,7 +272,7 @@ int CHAMELEON_zplgsy_Tile_Async( CHAMELEON_Complex64_t bump,
if (chameleon_min( A->m, A->n ) == 0)
return CHAMELEON_SUCCESS;
- chameleon_pzplgsy( bump, uplo, A, seed, sequence, request );
+ chameleon_pzplgsy( bump, uplo, A, A->m, A->i, A->j, seed, sequence, request );
return CHAMELEON_SUCCESS;
}
diff --git a/compute/zplrnt.c b/compute/zplrnt.c
index 272e5b4a8..4f25f18a0 100644
--- a/compute/zplrnt.c
+++ b/compute/zplrnt.c
@@ -261,7 +261,7 @@ int CHAMELEON_zplrnt_Tile_Async( CHAM_desc_t *A,
if (chameleon_min( A->m, A->n ) == 0)
return CHAMELEON_SUCCESS;
- chameleon_pzplrnt( A, seed, sequence, request );
+ chameleon_pzplrnt( A, A->m, A->i, A->j, seed, sequence, request );
return CHAMELEON_SUCCESS;
}
diff --git a/control/compute_z.h b/control/compute_z.h
index acc3c5445..0e3dc6ae2 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -97,9 +97,9 @@ void chameleon_pzlaswp(CHAM_desc_t *B, int *IPIV, int inc, RUNTIME_sequence_t *s
void chameleon_pzlaswpc(CHAM_desc_t *B, int *IPIV, int inc, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzlatms( cham_dist_t idist, unsigned long long int seed, cham_sym_t sym, double *D, int mode, double cond, double dmax, CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
void chameleon_pzlauum(cham_uplo_t uplo, CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
-void chameleon_pzplghe(double bump, cham_uplo_t uplo, CHAM_desc_t *A, unsigned long long int seed, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
-void chameleon_pzplgsy(CHAMELEON_Complex64_t bump, cham_uplo_t uplo, CHAM_desc_t *A, unsigned long long int seed, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
-void chameleon_pzplrnt(CHAM_desc_t *A, unsigned long long int seed, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
+void chameleon_pzplghe(double bump, cham_uplo_t uplo, CHAM_desc_t *A, int bigM, int m0, int n0, unsigned long long int seed, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
+void chameleon_pzplgsy(CHAMELEON_Complex64_t bump, cham_uplo_t uplo, CHAM_desc_t *A, int bigM, int m0, int n0, unsigned long long int seed, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
+void chameleon_pzplrnt(CHAM_desc_t *A, int bigM, int m0, int n0, unsigned long long int seed, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
void chameleon_pzplrnk(int K, CHAM_desc_t *C, unsigned long long int seedA, unsigned long long int seedB, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request );
void chameleon_pzpotrf(cham_uplo_t uplo, CHAM_desc_t *A, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
void chameleon_pzpotrimm(cham_uplo_t uplo, CHAM_desc_t *A, CHAM_desc_t *B, CHAM_desc_t *C, RUNTIME_sequence_t *sequence, RUNTIME_request_t *request);
--
GitLab