From 48c9546337af55d9eb36ef53f5a23c6f6c9a1b3b Mon Sep 17 00:00:00 2001
From: Mathieu Faverge <mathieu.faverge@inria.fr>
Date: Wed, 29 Mar 2023 20:49:18 +0200
Subject: [PATCH] cudablas: Use the returned value from cuda calls to return a
better information
---
cudablas/compute/cuda_zlarfb.c | 7 ++++++-
cudablas/compute/cuda_zparfb.c | 16 +++++++++++++++-
2 files changed, 21 insertions(+), 2 deletions(-)
diff --git a/cudablas/compute/cuda_zlarfb.c b/cudablas/compute/cuda_zlarfb.c
index 5039d0265..8175d1944 100644
--- a/cudablas/compute/cuda_zlarfb.c
+++ b/cudablas/compute/cuda_zlarfb.c
@@ -45,6 +45,7 @@ CUDA_zlarfb( cham_side_t side, cham_trans_t trans,
cublasStatus_t rc;
cham_trans_t transT, notransV, transV;
cham_uplo_t uplo;
+ int info = 0;
/* Check input arguments */
if ((side != ChamLeft) && (side != ChamRight)) {
@@ -111,6 +112,7 @@ CUDA_zlarfb( cham_side_t side, cham_trans_t trans,
V, LDV,
CUBLAS_SADDR(zzero), WORK, LDWORK );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
// W = W T^H = C^H V T^H
CUDA_ztrmm( ChamRight, uplo, transT, ChamNonUnit,
@@ -127,6 +129,7 @@ CUDA_zlarfb( cham_side_t side, cham_trans_t trans,
WORK, LDWORK,
CUBLAS_SADDR(zone), C, LDC );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
}
else {
// Form C H or C H^H
@@ -140,6 +143,7 @@ CUDA_zlarfb( cham_side_t side, cham_trans_t trans,
V, LDV,
CUBLAS_SADDR(zzero), WORK, LDWORK );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
// W = W T = C V T
CUDA_ztrmm( ChamRight, uplo, trans, ChamNonUnit,
@@ -156,7 +160,8 @@ CUDA_zlarfb( cham_side_t side, cham_trans_t trans,
V, LDV,
CUBLAS_SADDR(zone), C, LDC );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
}
- return (rc == CUBLAS_STATUS_SUCCESS) ? CHAMELEON_SUCCESS : CHAMELEON_ERR_UNEXPECTED;
+ return (info == 0) ? CHAMELEON_SUCCESS : CHAMELEON_ERR_UNEXPECTED;
}
diff --git a/cudablas/compute/cuda_zparfb.c b/cudablas/compute/cuda_zparfb.c
index 39aaa7586..8977eb6a1 100644
--- a/cudablas/compute/cuda_zparfb.c
+++ b/cudablas/compute/cuda_zparfb.c
@@ -165,6 +165,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
cham_trans_t transW, transA2;
int wssize = 0;
int wrsize = 0;
+ int info = 0;
/* Check input arguments */
if ((side != ChamLeft) && (side != ChamRight)) {
@@ -315,6 +316,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
A2 /* M2*N2 */, LDA2,
CUBLAS_SADDR(zone), workW /* K *N2 */, ldW );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
if ( workC == NULL ) {
/* W = op(T) * W */
@@ -331,6 +333,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workW + ldW * j, 1,
A1 + LDA1 * j, 1 );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
}
/* A2 = A2 - op(V) * W */
@@ -341,6 +344,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workW /* K * N2 */, ldW,
CUBLAS_SADDR(zone), A2 /* M2 * N2 */, LDA2 );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
} else {
/* Wc = V * op(T) */
@@ -350,6 +354,8 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
CUBLAS_SADDR(zone), workV, ldV,
T, LDT,
CUBLAS_SADDR(zzero), workC, ldC );
+ assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
/* A1 = A1 - opt(T) * W */
rc = cublasZgemm( handle,
@@ -359,6 +365,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workW, ldW,
CUBLAS_SADDR(zone), A1, LDA1 );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
/* A2 = A2 - Wc * W */
rc = cublasZgemm( handle,
@@ -368,6 +375,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workW, ldW,
CUBLAS_SADDR(zone), A2, LDA2 );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
}
}
else {
@@ -463,6 +471,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workV /* K *N2 */, ldV,
CUBLAS_SADDR(zone), workW /* M1*K */, ldW);
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
if ( workC == NULL ) {
/* W = W * op(T) */
@@ -479,6 +488,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workW + ldW * j, 1,
A1 + LDA1 * j, 1 );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
}
/* A2 = A2 - W * op(V) */
@@ -489,6 +499,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workV /* K *N2 */, ldV,
CUBLAS_SADDR(zone), A2 /* M2*N2 */, LDA2);
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
} else {
/* A1 = A1 - W * opt(T) */
@@ -499,6 +510,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
T, LDT,
CUBLAS_SADDR(zone), A1, LDA1 );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
/* Wc = op(T) * V */
rc = cublasZgemm( handle,
@@ -508,6 +520,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workV, ldV,
CUBLAS_SADDR(zzero), workC, ldC );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
/* A2 = A2 - W * Wc */
rc = cublasZgemm( handle,
@@ -517,6 +530,7 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
workC, ldC,
CUBLAS_SADDR(zone), A2, LDA2 );
assert( rc == CUBLAS_STATUS_SUCCESS );
+ info += (rc == CUBLAS_STATUS_SUCCESS) ? 0 : 1;
}
}
}
@@ -525,5 +539,5 @@ CUDA_zparfb( cham_side_t side, cham_trans_t trans,
return CHAMELEON_ERR_NOT_SUPPORTED;
}
- return (rc == CUBLAS_STATUS_SUCCESS) ? CHAMELEON_SUCCESS : CHAMELEON_ERR_UNEXPECTED;
+ return (info == 0) ? CHAMELEON_SUCCESS : CHAMELEON_ERR_UNEXPECTED;
}
--
GitLab