diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index e01e26e32b69683fa226bb64ddc2b5e840a9fa28..270076d745be68d22de64561940f0b22fa4ed86f 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -14,6 +14,7 @@ include:
- .gitlab/docker.yml
- .gitlab/build.yml
- .gitlab/test_starpu.yml
+ - .gitlab/test_starpu_plafrim.yml
- .gitlab/test_starpu_simgrid.yml
- .gitlab/test_parsec.yml
- .gitlab/test_quark.yml
diff --git a/.gitlab/build.sh b/.gitlab/build.sh
index 0d17fd1444511508505ac74072203612484ff18f..e631b37f26033fb772a37cc2bba77796a841d549 100755
--- a/.gitlab/build.sh
+++ b/.gitlab/build.sh
@@ -59,7 +59,7 @@ case $SYSTEM in
esac
# Compile
-eval '${SCAN}cmake --build build-${VERSION} -j 4 > /dev/null'
+eval '${SCAN}cmake --build build-${VERSION} -j ${CMAKE_BUILD_PARALLEL_LEVEL} > /dev/null'
# Install
cmake --install build-${VERSION}
diff --git a/.gitlab/build.yml b/.gitlab/build.yml
index e5d6743322aec15b494fc88129fb6f2672b926b3..2d449df9d37a7f6c3c927944d96c1497f02e010f 100644
--- a/.gitlab/build.yml
+++ b/.gitlab/build.yml
@@ -2,6 +2,8 @@
.build_script_common:
stage: build
extends: .only-master-mr
+ variables:
+ CMAKE_BUILD_PARALLEL_LEVEL: 4
artifacts:
name: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"
expire_in: 1 week
@@ -10,6 +12,8 @@
.build_script_linux:
extends: .build_script_common
+ variables:
+ SYSTEM: linux
script:
- bash .gitlab/build.sh | tee ${LOGNAME}.log
@@ -56,29 +60,31 @@ build_starpu:
.build_script_guix:
tags: ['plafrim']
extends: .build_script_common
+ variables:
+ SYSTEM: guix
+ CMAKE_BUILD_PARALLEL_LEVEL: 20
script:
- guix describe -f channels
- - guix shell --pure --preserve=SYSTEM --preserve=VERSION --preserve=LOGNAME --preserve=BUILD_OPTIONS
- -D chameleon-${GUIX_CHAMELEON_VARIANT} coreutils bash-minimal --
+ - guix shell --pure --preserve=SYSTEM --preserve=VERSION --preserve=LOGNAME --preserve=BUILD_OPTIONS --preserve=CMAKE_BUILD_PARALLEL_LEVEL
+ -D chameleon-${GPU_BACKEND} ${BLAS} coreutils bash-minimal --
bash .gitlab/build.sh | tee ${LOGNAME}.log
build_starpu_cuda:
extends: .build_script_guix
variables:
- SYSTEM: guix
VERSION: starpu_cuda
- GUIX_CHAMELEON_VARIANT: cuda
+ GPU_BACKEND: cuda
+ BLAS: "--with-input=openblas=intel-oneapi-mkl"
LOGNAME: "chameleon-build-${SYSTEM}-${VERSION}"
- BUILD_OPTIONS: "-DCHAMELEON_USE_CUDA=ON -DCHAMELEON_USE_MPI=ON -DBLA_VENDOR=OpenBLAS"
+ BUILD_OPTIONS: "-DCHAMELEON_USE_CUDA=ON -DCHAMELEON_USE_MPI=OFF -DBLA_VENDOR=Intel10_64lp_seq"
build_starpu_hip:
extends: .build_script_guix
variables:
- SYSTEM: guix
VERSION: starpu_hip
- GUIX_CHAMELEON_VARIANT: hip
+ GPU_BACKEND: hip
LOGNAME: "chameleon-build-${SYSTEM}-${VERSION}"
- BUILD_OPTIONS: "-DCHAMELEON_USE_HIP_ROC=ON -DCHAMELEON_USE_MPI=ON -DBLA_VENDOR=OpenBLAS -DCMAKE_C_COMPILER=gcc -DCMAKE_Fortran_COMPILER=gfortran"
+ BUILD_OPTIONS: "-DCHAMELEON_USE_HIP_ROC=ON -DCHAMELEON_USE_MPI=OFF -DBLA_VENDOR=OpenBLAS -DCMAKE_C_COMPILER=gcc -DCMAKE_Fortran_COMPILER=gfortran"
build_starpu_simgrid:
extends: .build_script_linux
@@ -97,11 +103,13 @@ build_starpu_macosx:
SYSTEM: macosx
VERSION: starpu
LOGNAME: "chameleon-build-${SYSTEM}-${VERSION}"
+ CMAKE_BUILD_PARALLEL_LEVEL: 4
script:
- bash .gitlab/build.sh | tee ${LOGNAME}.log
artifacts:
name: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"
expire_in: 1 week
+ when: always
paths:
- ${LOGNAME}.log
cache:
@@ -119,11 +127,13 @@ build_starpu_macosx:
MSYSTEM: UCRT64
VERSION: starpu
LOGNAME: "chameleon-build-${SYSTEM}-${VERSION}"
+ CMAKE_BUILD_PARALLEL_LEVEL: 4
script:
- bash -lc .gitlab/build.sh | tee "$env:LOGNAME.log"
artifacts:
name: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"
expire_in: 1 week
+ when: always
paths:
- ${LOGNAME}.log
cache:
diff --git a/.gitlab/common.yml b/.gitlab/common.yml
index 9e679f1c4e1b4eabe97865e7d1ff3a9fbaa57244..f4e30eb4b89a16504b85edf3900a0740fb108427 100644
--- a/.gitlab/common.yml
+++ b/.gitlab/common.yml
@@ -37,5 +37,6 @@ variables:
name: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"
expire_in: 1 week
untracked: true
+ when: always
reports:
junit: ${LOGNAME}-junit.xml
diff --git a/.gitlab/sbatch.sh b/.gitlab/sbatch.sh
new file mode 100755
index 0000000000000000000000000000000000000000..d202d068699b939a985bbc120534955f41511ed6
--- /dev/null
+++ b/.gitlab/sbatch.sh
@@ -0,0 +1,29 @@
+#!/bin/bash
+# sbatch.sh : submit slurm jobs and wait for completion before exiting
+set -x
+
+if [ $# -gt 0 ]
+then
+ JOB_NAME=$1
+fi
+JOB_NAME=${JOB_NAME:-chameleon}
+
+# to get kernels execution on both cpus and gpus
+export STARPU_SCHED=random
+
+# execution commands
+sbatch --wait \
+ --job-name="$JOB_NAME" \
+ --output="$JOB_NAME.out" \
+ --nodes=1 \
+ --exclusive --ntasks-per-node=1 --threads-per-core=1 \
+ --constraint="$SLURM_CONSTRAINTS" \
+ --time=01:00:00 \
+ $(dirname "$0")/test.sh
+# get the error code from the last command: sbatch --wait ...
+err=$?
+
+cat $JOB_NAME.out
+
+# exit with error code from the guix command
+exit $err
diff --git a/.gitlab/test.sh b/.gitlab/test.sh
index 0a4a8639fa4a54d1121f149e57264ed3ad2338fa..27ba5484ba023eaec8bad5cf2018da8c4008b97f 100755
--- a/.gitlab/test.sh
+++ b/.gitlab/test.sh
@@ -1,5 +1,4 @@
#!/usr/bin/env bash
-
set -e
set -x
diff --git a/.gitlab/test_starpu.yml b/.gitlab/test_starpu.yml
index efc3bd783aa0b64355500d6bb8a5fa4b233f5e59..1adc4997f75890d7fe557417a8b46855cecba0ce 100644
--- a/.gitlab/test_starpu.yml
+++ b/.gitlab/test_starpu.yml
@@ -1,5 +1,5 @@
---
-.test_starpu:
+.test_script_starpu:
extends: .test_script_linux
needs: [build_starpu]
variables:
@@ -9,7 +9,7 @@
TESTS_RESTRICTION: "-R _${CATEGORY}_${PRECISION}|example -E sytrf|sysv"
test_starpu_master:
- extends: .test_starpu
+ extends: .test_script_starpu
rules:
- if: ($CI_COMMIT_BRANCH == $CI_DEFAULT_BRANCH || $CI_COMMIT_BRANCH =~ /^ci-.*$/) && $CI_PIPELINE_SOURCE != "schedule"
parallel:
@@ -18,7 +18,7 @@ test_starpu_master:
CATEGORY: [shm, mpi]
test_starpu_mr:
- extends: .test_starpu
+ extends: .test_script_starpu
rules:
- if: ($CI_PIPELINE_SOURCE == "merge_request_event" && $CI_MERGE_REQUEST_SOURCE_BRANCH_NAME !~ /^notest-.*$/)
parallel:
@@ -43,6 +43,7 @@ test_starpu_macosx:
artifacts:
name: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"
expire_in: 1 week
+ when: always
paths:
- ${LOGNAME}.log
reports:
@@ -69,6 +70,7 @@ test_starpu_macosx:
artifacts:
name: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"
expire_in: 1 week
+ when: always
paths:
- ${LOGNAME}.log
reports:
diff --git a/.gitlab/test_starpu_plafrim.yml b/.gitlab/test_starpu_plafrim.yml
new file mode 100644
index 0000000000000000000000000000000000000000..c5178e6cac3910716b87ed85d3cb795413c0a0b9
--- /dev/null
+++ b/.gitlab/test_starpu_plafrim.yml
@@ -0,0 +1,46 @@
+---
+.test_script_starpu_plafrim:
+ tags: ['plafrim']
+ stage: test
+ variables:
+ SYSTEM: guix
+ script:
+ - guix describe -f channels | tee guix-channels.scm
+ - guix shell --pure --preserve=SYSTEM --preserve=VERSION --preserve=LOGNAME --preserve=BUILD_OPTIONS --preserve=TESTS_RESTRICTION --preserve=SLURM_CONSTRAINTS --preserve=LD_PRELOAD
+ -D chameleon-${GPU_BACKEND} ${BLAS} slurm coreutils inetutils bash-minimal --
+ bash .gitlab/sbatch.sh ${LOGNAME} | tee ${LOGNAME}.log
+ artifacts:
+ name: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"
+ expire_in: 1 week
+ when: always
+ paths:
+ - guix-channels.scm
+ - ${LOGNAME}.log
+ reports:
+ junit: ${LOGNAME}-junit.xml
+
+.test_script_starpu_plafrim_cuda:
+ extends: .test_script_starpu_plafrim
+ needs: [build_starpu_cuda]
+ variables:
+ VERSION: starpu_cuda
+ GPU_BACKEND: cuda
+ BLAS: "--with-input=openblas=intel-oneapi-mkl"
+ LOGNAME: "chameleon-${SYSTEM}-${VERSION}"
+ BUILD_OPTIONS: "-DCHAMELEON_USE_CUDA=ON -DCHAMELEON_USE_MPI=OFF -DBLA_VENDOR=Intel10_64lp_seq"
+ SLURM_CONSTRAINTS: "sirocco"
+ LD_PRELOAD: "/usr/lib64/libcuda.so"
+
+test_starpu_cuda_master:
+ extends: .test_script_starpu_plafrim_cuda
+ rules:
+ - if: ($CI_COMMIT_BRANCH == $CI_DEFAULT_BRANCH || $CI_COMMIT_BRANCH =~ /^ci-.*$/) && $CI_PIPELINE_SOURCE != "schedule"
+ variables:
+ TESTS_RESTRICTION: "-R test_shm_gpu"
+
+test_starpu_cuda_mr:
+ extends: .test_script_starpu_plafrim_cuda
+ rules:
+ - if: ($CI_PIPELINE_SOURCE == "merge_request_event" && $CI_MERGE_REQUEST_SOURCE_BRANCH_NAME !~ /^notest-.*$/)
+ variables:
+ TESTS_RESTRICTION: "-R test_shm_gpu_d|test_shm_gpu_c"
diff --git a/compute/pztradd.c b/compute/pztradd.c
index 5c092c937be0d58683b8e2ce570e0c49e022d022..78d7a3d282e8a4d08bd57b91ffe2935fe337210e 100644
--- a/compute/pztradd.c
+++ b/compute/pztradd.c
@@ -29,16 +29,17 @@
/**
* Parallel tile matrix-matrix multiplication - dynamic scheduling
*/
-void chameleon_pztradd(cham_uplo_t uplo, cham_trans_t trans,
- CHAMELEON_Complex64_t alpha, CHAM_desc_t *A,
- CHAMELEON_Complex64_t beta, CHAM_desc_t *B,
- RUNTIME_sequence_t *sequence, RUNTIME_request_t *request)
+void chameleon_pztradd( cham_uplo_t uplo, cham_trans_t trans,
+ CHAMELEON_Complex64_t alpha, CHAM_desc_t *A,
+ CHAMELEON_Complex64_t beta, CHAM_desc_t *B,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
int tempmm, tempnn, tempmn, tempnm;
- int m, n;
+ int m, n, minmn;
chamctxt = chameleon_context_self();
if (sequence->status != CHAMELEON_SUCCESS) {
@@ -46,10 +47,12 @@ void chameleon_pztradd(cham_uplo_t uplo, cham_trans_t trans,
}
RUNTIME_options_init(&options, chamctxt, sequence, request);
+ minmn = chameleon_min( B->mt, B->nt );
+
switch(uplo){
case ChamLower:
if (trans == ChamNoTrans) {
- for (n = 0; n < chameleon_min(B->mt,B->nt); n++) {
+ for (n = 0; n < minmn; n++) {
tempnm = n == B->mt-1 ? B->m-n*B->mb : B->mb;
tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
@@ -72,8 +75,8 @@ void chameleon_pztradd(cham_uplo_t uplo, cham_trans_t trans,
}
else {
for (n = 0; n < chameleon_min(B->mt,B->nt); n++) {
- tempnm = n == B->mt-1 ? B->m-n*B->mb : B->mb;
- tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
+ tempnm = n == B->mt-1 ? B->m - n * B->mb : B->mb;
+ tempnn = n == B->nt-1 ? B->n - n * B->nb : B->nb;
INSERT_TASK_ztradd(
&options,
@@ -95,9 +98,9 @@ void chameleon_pztradd(cham_uplo_t uplo, cham_trans_t trans,
break;
case ChamUpper:
if (trans == ChamNoTrans) {
- for (m = 0; m < chameleon_min(B->mt,B->nt); m++) {
- tempmm = m == B->mt-1 ? B->m-B->mb*m : B->nb;
- tempmn = m == B->nt-1 ? B->n-m*B->nb : B->nb;
+ for (m = 0; m < minmn; m++) {
+ tempmm = m == B->mt-1 ? B->m - m * B->mb : B->nb;
+ tempmn = m == B->nt-1 ? B->n - m * B->nb : B->nb;
INSERT_TASK_ztradd(
&options,
@@ -106,7 +109,7 @@ void chameleon_pztradd(cham_uplo_t uplo, cham_trans_t trans,
beta, B(m, m));
for (n = m+1; n < B->nt; n++) {
- tempnn = n == B->nt-1 ? B->n-n*B->nb : B->nb;
+ tempnn = n == B->nt-1 ? B->n - n * B->nb : B->nb;
INSERT_TASK_zgeadd(
&options,
diff --git a/compute/zgepdf_qr.c b/compute/zgepdf_qr.c
index 35622025f4e1e051222f4f4515a08d0c952c9f19..d138b051773768788421ceb691c7c59d3e433617 100644
--- a/compute/zgepdf_qr.c
+++ b/compute/zgepdf_qr.c
@@ -82,10 +82,12 @@ int CHAMELEON_zgepdf_qr_Tile( int doqr, int optid,
CHAM_desc_t *A1, CHAM_desc_t *TS1, CHAM_desc_t *TT1, CHAM_desc_t *Q1,
CHAM_desc_t *A2, CHAM_desc_t *TS2, CHAM_desc_t *TT2, CHAM_desc_t *Q2 )
{
- CHAM_context_t *chamctxt;
+ CHAM_context_t *chamctxt;
RUNTIME_sequence_t *sequence = NULL;
- RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
- int status;
+ RUNTIME_request_t request = RUNTIME_REQUEST_INITIALIZER;
+ CHAM_desc_t D1, *D1ptr = NULL;
+ CHAM_desc_t D2, *D2ptr = NULL;
+ int status;
chamctxt = chameleon_context_self();
if (chamctxt == NULL) {
@@ -94,16 +96,36 @@ int CHAMELEON_zgepdf_qr_Tile( int doqr, int optid,
}
chameleon_sequence_create( chamctxt, &sequence );
+#if defined(CHAMELEON_COPY_DIAG)
+ {
+ int n = A1->n;
+ chameleon_zdesc_copy_and_restrict( A1, &D1, A1->m, n );
+ D1ptr = &D1;
+ chameleon_zdesc_copy_and_restrict( A2, &D2, A2->m, n );
+ D2ptr = &D2;
+ }
+#endif
+
chameleon_pzgepdf_qr( 1, doqr, optid, qrtreeT, qrtreeB,
- A1, TS1, TT1, NULL, Q1,
- A2, TS2, TT2, NULL, Q2,
+ A1, TS1, TT1, D1ptr, Q1,
+ A2, TS2, TT2, D2ptr, Q2,
sequence, &request );
CHAMELEON_Desc_Flush( Q1, sequence );
CHAMELEON_Desc_Flush( Q2, sequence );
+ if ( D1ptr != NULL ) {
+ CHAMELEON_Desc_Flush( D1ptr, sequence );
+ CHAMELEON_Desc_Flush( D2ptr, sequence );
+ }
chameleon_sequence_wait( chamctxt, sequence );
status = sequence->status;
chameleon_sequence_destroy( chamctxt, sequence );
+
+ if ( D1ptr != NULL ) {
+ chameleon_desc_destroy( D1ptr );
+ chameleon_desc_destroy( D2ptr );
+ }
+
return status;
}
diff --git a/coreblas/compute/core_zlascal.c b/coreblas/compute/core_zlascal.c
index 4355e893f0fbf106606661f0a87c27f4151c8634..801977297fcc6d9f49d8468a6e5a2e294e2ccd4e 100644
--- a/coreblas/compute/core_zlascal.c
+++ b/coreblas/compute/core_zlascal.c
@@ -27,9 +27,10 @@
*
* @ingroup CORE_CHAMELEON_Complex64_t
*
- * CORE_zlascal scales a two-dimensional matrix A. As opposite to
- * CORE_zlascl(), no checks is performed to prevent under/overflow. This should
- * have been done at higher level.
+ * CORE_zlascal scales a two-dimensional matrix A.
+ * As opposed to CORE_zlascl(), no checks is performed to prevent
+ * under/overflow and scaling with a complex is possible. This should have been
+ * done at higher level.
*
*******************************************************************************
*
diff --git a/coreblas/compute/core_ztradd.c b/coreblas/compute/core_ztradd.c
index 924ec0fca06c4649b92519cc667cc1a92e1cf50c..de868b032af9b440abf2616d3027c4e6825d988e 100644
--- a/coreblas/compute/core_ztradd.c
+++ b/coreblas/compute/core_ztradd.c
@@ -146,8 +146,7 @@ int CORE_ztradd(cham_uplo_t uplo, cham_trans_t trans, int M, int N,
M, N, 0., 0., B, LDB );
}
else if ( beta != (CHAMELEON_Complex64_t)1. ) {
- LAPACKE_zlascl_work( LAPACK_COL_MAJOR, chameleon_lapack_const(uplo),
- 0, 0, 1., beta, M, N, B, LDB );
+ CORE_zlascal( uplo, M, N, beta, B, LDB );
}
/**
diff --git a/runtime/starpu/codelets/codelet_map.c b/runtime/starpu/codelets/codelet_map.c
index 7461c3f5aed2697acffb09623fc35e9f34224350..0e9120cd3bbe8c0511000213cbe57b5899e5d955 100644
--- a/runtime/starpu/codelets/codelet_map.c
+++ b/runtime/starpu/codelets/codelet_map.c
@@ -228,6 +228,7 @@ void INSERT_TASK_map( const RUNTIME_option_t *options,
int exec = 0;
int i, readonly = 1;
size_t clargs_size = 0;
+ uint32_t where = 0;
void (*callback)(void*);
if ( ( ndata < 0 ) || ( ndata > 3 ) ) {
@@ -275,6 +276,17 @@ void INSERT_TASK_map( const RUNTIME_option_t *options,
(data[i].desc)->get_blktile( data[i].desc, m, n ) );
}
+ /* Where to execute */
+ if ( op_fcts->cpufunc ) {
+ where |= STARPU_CPU;
+ }
+ if ( op_fcts->cudafunc ) {
+ where |= STARPU_CUDA;
+ }
+ if ( op_fcts->hipfunc ) {
+ where |= STARPU_HIP;
+ }
+
/* Insert the task */
switch( ndata ) {
case 1:
@@ -291,6 +303,7 @@ void INSERT_TASK_map( const RUNTIME_option_t *options,
STARPU_PRIORITY, options->priority,
STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, options->workerid,
+ STARPU_EXECUTE_WHERE, where,
STARPU_NAME, cl_name,
0 );
break;
@@ -310,6 +323,7 @@ void INSERT_TASK_map( const RUNTIME_option_t *options,
STARPU_PRIORITY, options->priority,
STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, options->workerid,
+ STARPU_EXECUTE_WHERE, where,
STARPU_NAME, cl_name,
0 );
break;
@@ -330,6 +344,7 @@ void INSERT_TASK_map( const RUNTIME_option_t *options,
STARPU_PRIORITY, options->priority,
STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, options->workerid,
+ STARPU_EXECUTE_WHERE, where,
STARPU_NAME, cl_name,
0 );
break;
diff --git a/runtime/starpu/codelets/codelet_zgemm.c b/runtime/starpu/codelets/codelet_zgemm.c
index b03e2443c8a4fb53517b7d617020f373de76ff6d..438968e0af89c1c3378e00cb22e03235200e7fb8 100644
--- a/runtime/starpu/codelets/codelet_zgemm.c
+++ b/runtime/starpu/codelets/codelet_zgemm.c
@@ -144,6 +144,7 @@ void INSERT_TASK_zgemm_Astat( const RUNTIME_option_t *options,
int accessC;
int exec = 0;
const char *cl_name = "zgemm_Astat";
+ uint32_t where = cl_zgemm.where;
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
@@ -188,6 +189,13 @@ void INSERT_TASK_zgemm_Astat( const RUNTIME_option_t *options,
B->get_blktile( B, Bm, Bn ),
C->get_blktile( C, Cm, Cn ) );
+ /* WARNING: CUDA 12.3 has an issue when k=1 in complex, thus we disable gemm on gpu in these cases */
+#if defined(PRECISION_z) || defined(PRECISION_c)
+ if ( k == 1 ) {
+ where = STARPU_CPU;
+ }
+#endif
+
/* Insert the task */
rt_starpu_insert_task(
&cl_zgemm,
@@ -204,6 +212,7 @@ void INSERT_TASK_zgemm_Astat( const RUNTIME_option_t *options,
STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_NODE, A->get_rankof(A, Am, An),
STARPU_NAME, cl_name,
+ STARPU_EXECUTE_WHERE, where,
0 );
}
@@ -214,7 +223,7 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
const CHAM_desc_t *B, int Bm, int Bn,
CHAMELEON_Complex64_t beta, const CHAM_desc_t *C, int Cm, int Cn )
{
- if ( alpha == 0. ) {
+ if ( alpha == (CHAMELEON_Complex64_t)0. ) {
INSERT_TASK_zlascal( options, ChamUpperLower, m, n, nb,
beta, C, Cm, Cn );
return;
@@ -225,6 +234,7 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
int accessC;
int exec = 0;
const char *cl_name = "zgemm";
+ uint32_t where = cl_zgemm.where;
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
@@ -249,7 +259,8 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
callback = options->profiling ? cl_zgemm_callback : NULL;
/* Reduce the C access if needed */
- accessC = ( beta == 0. ) ? STARPU_W : (STARPU_RW | ((beta == 1.) ? STARPU_COMMUTE : 0));
+ accessC = ( beta == (CHAMELEON_Complex64_t)0. ) ? STARPU_W :
+ (STARPU_RW | ((beta == (CHAMELEON_Complex64_t)1.) ? STARPU_COMMUTE : 0));
/* Refine name */
cl_name = chameleon_codelet_name( cl_name, 3,
@@ -257,6 +268,13 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
B->get_blktile( B, Bm, Bn ),
C->get_blktile( C, Cm, Cn ) );
+ /* WARNING: CUDA 12.3 has an issue when k=1 in complex, thus we disable gemm on gpu in these cases */
+#if defined(PRECISION_z) || defined(PRECISION_c)
+ if ( k == 1 ) {
+ where = STARPU_CPU;
+ }
+#endif
+
/* Insert the task */
rt_starpu_insert_task(
&cl_zgemm,
@@ -274,5 +292,6 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
STARPU_EXECUTE_ON_WORKER, options->workerid,
STARPU_POSSIBLY_PARALLEL, options->parallel,
STARPU_NAME, cl_name,
+ STARPU_EXECUTE_WHERE, where,
0 );
}
diff --git a/runtime/starpu/control/runtime_descriptor_ipiv.c b/runtime/starpu/control/runtime_descriptor_ipiv.c
index 977c8676e3acad779f2c1f9e27ec59c92026e3d1..da2c1ff59c2d9a4a6d8cff6cd15b5eef45966e18 100644
--- a/runtime/starpu/control/runtime_descriptor_ipiv.c
+++ b/runtime/starpu/control/runtime_descriptor_ipiv.c
@@ -96,15 +96,15 @@ void *RUNTIME_ipiv_getaddr( const CHAM_ipiv_t *ipiv, int m )
return (void*)(*handle);
}
- const CHAM_desc_t *A = ipiv->desc;
- int owner = A->get_rankof( A, m, m );
int ncols = (mm == (ipiv->mt-1)) ? ipiv->m - mm * ipiv->mb : ipiv->mb;
starpu_vector_data_register( handle, -1, (uintptr_t)NULL, ncols, sizeof(int) );
#if defined(CHAMELEON_USE_MPI)
{
- int64_t tag = ipiv->mpitag_ipiv + mm;
+ const CHAM_desc_t *A = ipiv->desc;
+ int owner = A->get_rankof( A, m, m );
+ int64_t tag = ipiv->mpitag_ipiv + mm;
starpu_mpi_data_register( *handle, tag, owner );
}
#endif /* defined(CHAMELEON_USE_MPI) */
@@ -173,15 +173,15 @@ void *RUNTIME_perm_getaddr( const CHAM_ipiv_t *ipiv, int m )
return (void*)(*handle);
}
- const CHAM_desc_t *A = ipiv->desc;
- int owner = A->get_rankof( A, m, m );
int ncols = ipiv->mb;
starpu_vector_data_register( handle, -1, (uintptr_t)NULL, ncols, sizeof(int) );
#if defined(CHAMELEON_USE_MPI)
{
- int64_t tag = ipiv->mpitag_perm + mm;
+ const CHAM_desc_t *A = ipiv->desc;
+ int owner = A->get_rankof( A, m, m );
+ int64_t tag = ipiv->mpitag_perm + mm;
starpu_mpi_data_register( *handle, tag, owner );
}
#endif /* defined(CHAMELEON_USE_MPI) */
@@ -202,15 +202,15 @@ void *RUNTIME_invp_getaddr( const CHAM_ipiv_t *ipiv, int m )
return (void*)(*handle);
}
- const CHAM_desc_t *A = ipiv->desc;
- int owner = A->get_rankof( A, m, m );
int ncols = ipiv->mb;
starpu_vector_data_register( handle, -1, (uintptr_t)NULL, ncols, sizeof(int) );
#if defined(CHAMELEON_USE_MPI)
{
- int64_t tag = ipiv->mpitag_invp + mm;
+ const CHAM_desc_t *A = ipiv->desc;
+ int owner = A->get_rankof( A, m, m );
+ int64_t tag = ipiv->mpitag_invp + mm;
starpu_mpi_data_register( *handle, tag, owner );
}
#endif /* defined(CHAMELEON_USE_MPI) */
@@ -303,6 +303,7 @@ void RUNTIME_perm_flushk( const RUNTIME_sequence_t *sequence,
(void)sequence;
(void)ipiv;
(void)m;
+ (void)A;
}
void RUNTIME_ipiv_gather( const RUNTIME_sequence_t *sequence,
@@ -312,14 +313,13 @@ void RUNTIME_ipiv_gather( const RUNTIME_sequence_t *sequence,
int64_t mb = desc->mb;
int64_t tag = chameleon_starpu_tag_book( (int64_t)(desc->mt) );
int rank = CHAMELEON_Comm_rank();
- int owner = rank;
int m;
for (m = 0; m < mt; m++, ipiv += mb) {
starpu_data_handle_t ipiv_src = RUNTIME_ipiv_getaddr( desc, m );
#if defined(CHAMELEON_USE_MPI)
- owner = starpu_mpi_data_get_rank( ipiv_src );
+ int owner = starpu_mpi_data_get_rank( ipiv_src );
if ( node != owner ) {
starpu_mpi_tag_t tag = starpu_mpi_data_get_tag( ipiv_src );
diff --git a/testing/testing_zcheck_aux.c b/testing/testing_zcheck_aux.c
index 9ea85852eb5ad53b05fc5772d3db14a67b018e8f..98ec027e8ea26ac3140d4269421e97f8d89c79f4 100644
--- a/testing/testing_zcheck_aux.c
+++ b/testing/testing_zcheck_aux.c
@@ -453,6 +453,15 @@ int check_zsum_std( run_arg_list_t *args, cham_uplo_t uplo, cham_trans_t trans,
run_arg_add_double( args, "||B||", Binitnorm );
run_arg_add_double( args, "||R||", Rnorm );
+ if ( alpha != 0. ) {
+ Anorm = Anorm * cabs(alpha);
+ }
+ if ( beta != 0. ) {
+ Binitnorm = Binitnorm * cabs(beta);
+ }
+
+ result = Rnorm / (max(Anorm, Binitnorm) * eps);
+
/* Verifies if the result is inside a threshold */
if ( isnan(Rnorm) || isinf(Rnorm) || isnan(result) || isinf(result) || (result > 10.0) ) {
info_solution = 1;
diff --git a/testing/testing_zsysv.c b/testing/testing_zsysv.c
index 9de378761a97377e5e2981f028efee6c2960f8c4..4ae24d4fec3a861b10716f5f241c0c7b66fc62bf 100644
--- a/testing/testing_zsysv.c
+++ b/testing/testing_zsysv.c
@@ -80,7 +80,7 @@ testing_zsysv_desc( run_arg_list_t *args, int check )
/* Check the factorization */
descA0 = CHAMELEON_Desc_Copy( descA, CHAMELEON_MAT_ALLOC_TILE );
- CHAMELEON_zplgsy_Tile( (double)N, uplo, descA0, seedA );
+ CHAMELEON_zplgsy_Tile( (double)N, ChamUpperLower, descA0, seedA );
hres += check_zxxtrf( args, ChamSymmetric, uplo, descA0, descA );
diff --git a/testing/testing_zsytrf.c b/testing/testing_zsytrf.c
index d48e91c9d0ca0996ab705424822adca8f37cfbc7..0bc75fa82f1e5e967bc6a94c4aaaa91ad5f76901 100644
--- a/testing/testing_zsytrf.c
+++ b/testing/testing_zsytrf.c
@@ -63,7 +63,7 @@ testing_zsytrf_desc( run_arg_list_t *args, int check )
/* Checks the factorisation and residue */
if ( check ) {
CHAM_desc_t *descA0 = CHAMELEON_Desc_Copy( descA, CHAMELEON_MAT_ALLOC_TILE );
- CHAMELEON_zplgsy_Tile( (double)N, uplo, descA0, seedA );
+ CHAMELEON_zplgsy_Tile( (double)N, ChamUpperLower, descA0, seedA );
hres += check_zxxtrf( args, ChamSymmetric, uplo, descA0, descA );
diff --git a/testing/testing_ztradd.c b/testing/testing_ztradd.c
index e80d8f3ad4685d5eb0a9a64337c8cd9b8842659c..7ea8b8e7b614474c3b31ea6d479612a11d775873 100644
--- a/testing/testing_ztradd.c
+++ b/testing/testing_ztradd.c
@@ -241,7 +241,7 @@ testing_t test_ztradd;
const char *ztradd_params[] = { "mtxfmt", "nb", "trans", "uplo", "m", "n", "lda",
"ldb", "alpha", "beta", "seedA", "seedB", NULL };
const char *ztradd_output[] = { NULL };
-const char *ztradd_outchk[] = { "RETURN", NULL };
+const char *ztradd_outchk[] = { "||A||", "||B||", "||R||", "RETURN", NULL };
/**
* @brief Testing registration function