Topic/update parsec to 14b823a7ff9b

cmake_modules/FindLAPACKE.cmake

@@ -48,6 +48,7 @@
 # Copyright (c) 2019-2024 The University of Tennessee and The University
 #                         of Tennessee Research Foundation.  All rights
 #                         reserved.
+# Copyright (c) 2026      NVIDIA Corporation.  All rights reserved.
 #
 # $COPYRIGHT$
 #
@@ -120,6 +121,19 @@ set(LAPACKE_FIND_ALL_COMPONENTS 0)
 # ==============================================================================
 
 macro(_find_library_with_header component incname)
+  if(LAPACKE_${component}_LIB)
+    get_filename_component(_lapacke_lib_ext "${LAPACKE_${component}_LIB}" EXT)
+    if(IS_DIRECTORY "${LAPACKE_${component}_LIB}" AND NOT _lapacke_lib_ext STREQUAL ".framework")
+      list(APPEND LAPACKE_SEARCH_PATHS "${LAPACKE_${component}_LIB}")
+      get_filename_component(_lapacke_lib_parent "${LAPACKE_${component}_LIB}" DIRECTORY)
+      list(APPEND LAPACKE_SEARCH_PATHS "${_lapacke_lib_parent}")
+      unset(LAPACKE_${component}_LIB CACHE)
+      unset(LAPACKE_${component}_LIB)
+      unset(_lapacke_lib_parent)
+    endif()
+    unset(_lapacke_lib_ext)
+  endif()
+
   find_library(LAPACKE_${component}_LIB
     NAMES ${ARGN}
     NAMES_PER_DIR

contrib/build_with_dplasma/common.c

@@ -2,6 +2,7 @@
  * Copyright (c) 2009-2020 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
+ * Copyright (c) 2026      NVIDIA Corporation.  All rights reserved.
  *
  */
 #include "parsec/runtime.h"
@@ -548,6 +549,23 @@ static void print_arguments(int* iparam)
     }
 }
 
+static void get_parsec_arguments(int argc, char **argv,
+                                 int *parsec_argc, char ***parsec_argv)
+{
+    int idx;
+
+    *parsec_argc = 0;
+    *parsec_argv = NULL;
+
+    for( idx = 1; idx < argc; idx++ ) {
+        if( 0 == strcmp(argv[idx], "--") ) {
+            *parsec_argc = argc - idx - 1;
+            *parsec_argv = (0 < *parsec_argc) ? &argv[idx + 1] : NULL;
+            return;
+        }
+    }
+}
+
 
 
 
@@ -633,19 +651,13 @@ parsec_context_t* setup_parsec(int argc, char **argv, int *iparam)
 
     TIME_START();
 
-    /* Once we got out arguments, we should pass whatever is left down */
-    int parsec_argc, idx;
-    char** parsec_argv = (char**)calloc(argc, sizeof(char*));
-    parsec_argv[0] = argv[0];  /* the app name */
-    for( idx = parsec_argc = 1;
-         (idx < argc) && (0 != strcmp(argv[idx], "--")); idx++);
-    if( idx != argc ) {
-        for( parsec_argc = 1, idx++; idx < argc;
-             parsec_argv[parsec_argc] = argv[idx], parsec_argc++, idx++);
-    }
+    /* PaRSEC only accepts its own arguments, without argv[0] or "--". */
+    int parsec_argc;
+    char **parsec_argv;
+    get_parsec_arguments(argc, argv, &parsec_argc, &parsec_argv);
     parsec_context_t* ctx = parsec_init(iparam[IPARAM_NCORES],
-                                      &parsec_argc, &parsec_argv);
-    free(parsec_argv);
+                                      (0 < parsec_argc) ? &parsec_argc : NULL,
+                                      (0 < parsec_argc) ? &parsec_argv : NULL);
     if( NULL == ctx ) {
         /* Failed to correctly initialize. In a correct scenario report
          * upstream, but in this particular case bail out.
@@ -678,4 +690,3 @@ void cleanup_parsec(parsec_context_t* parsec, int *iparam)
 #endif
     (void)iparam;
 }
-

examples/dqr_driver.c

@@ -2,6 +2,7 @@
  * Copyright (c) 2016-2022 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
+ * Copyright (c) 2026      NVIDIA Corporation.  All rights reserved.
  *
  *  Compile with $CC $(pkg-config --cflags dplasma) -c dqr_driver.c
  *  Link with $CC $(pkg-config --libs dplasma) -o dqr_driver dqr_driver.o
@@ -11,6 +12,7 @@
 #include "parsec/data_dist/matrix/two_dim_rectangle_cyclic.h"
 #include <getopt.h>
 #include <math.h>
+#include <string.h>
 #include <sys/time.h>
 
 int main(int argc, char ** argv)
@@ -23,6 +25,8 @@ int main(int argc, char ** argv)
     int P = 0;
     int ch;
     int cores = -1;
+    int parsec_argc = 0;
+    char **parsec_argv = NULL;
     parsec_matrix_block_cyclic_t dcA;
     parsec_matrix_block_cyclic_t dcWork;
 
@@ -50,6 +54,14 @@ int main(int argc, char ** argv)
     rank = 0;
 #endif
 
+     for( ch = 1; ch < argc; ch++ ) {
+         if( 0 == strcmp(argv[ch], "--") ) {
+             parsec_argc = argc - ch - 1;
+             parsec_argv = (0 < parsec_argc) ? &argv[ch + 1] : NULL;
+             break;
+         }
+     }
+
      while ((ch = getopt_long(argc, argv, "M:N:m:n:P:c:h", longopts, NULL)) != -1)
              switch (ch) {
              case 'M':
@@ -101,8 +113,10 @@ int main(int argc, char ** argv)
 
 
     /** Initialize PaRSEC with the required number of cores,
-     *  and pass all arguments after '--' to PaRSEC, if there are some */
-    parsec = parsec_init(cores, &argc, &argv);
+     *  and pass only PaRSEC arguments after '--', without argv[0]. */
+    parsec = parsec_init(cores,
+                         (0 < parsec_argc) ? &parsec_argc : NULL,
+                         (0 < parsec_argc) ? &parsec_argv : NULL);
 
     /** Declare a Matrix A as a (4, 1)-2D-Tile cyclic matrix of size M x N real
      *  doubles, tiled in tiles of mb x nb, and distributed over PxQ processes,

src/cuda/lapack_cuda_stage_in.c

@@ -2,6 +2,7 @@
  * Copyright (c) 2020-2024 The University of Tennessee and The University
  *                         of Tennessee Research Foundation. All rights
  *                         reserved.
+ * Copyright (c) 2026      NVIDIA Corporation.  All rights reserved.
  *
  */
 
@@ -33,14 +34,14 @@ dplasma_cuda_lapack_stage_in(parsec_gpu_task_t *gtask,
         if(flow_mask & (1U << i)){
             copy_in = task->data[i].data_in;
             copy_out = task->data[i].data_out;
-            ddc = (dplasma_data_collection_t*)gtask->flow_dc[i];
+            ddc = (dplasma_data_collection_t*)gtask->flow_info[i].flow_dc;
             assert(ddc != NULL);
             elem_sz = parsec_datadist_getsizeoftype(ddc->dc_original->mtype);
             in_elem_dev = (parsec_device_gpu_module_t*)parsec_mca_device_get( copy_in->device_index);
             if( (in_elem_dev->super.type == PARSEC_DEV_CUDA) || (ddc->dc_original->storage != PARSEC_MATRIX_LAPACK)){
                 ret = (cudaError_t)cudaMemcpyAsync( copy_out->device_private,
                                                     copy_in->device_private,
-                                                    gtask->flow_nb_elts[i],
+                                                    gtask->flow_info[i].flow_span,
                                                     (in_elem_dev->super.type != PARSEC_DEV_CUDA)?
                                                             cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice,
                                                     cuda_stream->cuda_stream);
@@ -109,15 +110,15 @@ dplasma_cuda_lapack_stage_out(parsec_gpu_task_t *gtask,
         if(flow_mask & (1U << i)){
             copy_in = task->data[i].data_out;
             copy_out = copy_in->original->device_copies[0];
-            ddc = (dplasma_data_collection_t*)gtask->flow_dc[i];
+            ddc = (dplasma_data_collection_t*)gtask->flow_info[i].flow_dc;
             assert(ddc != NULL);
             elem_sz = parsec_datadist_getsizeoftype(ddc->dc_original->mtype);
             out_elem_dev = (parsec_device_gpu_module_t*)parsec_mca_device_get( copy_out->device_index);
 
             if( (out_elem_dev->super.type == PARSEC_DEV_CUDA) || (ddc->dc_original->storage != PARSEC_MATRIX_LAPACK)){
                 ret = (cudaError_t)cudaMemcpyAsync( copy_out->device_private,
                                                     copy_in->device_private,
-                                                    gtask->flow_nb_elts[i],
+                                                    gtask->flow_info[i].flow_span,
                                                     out_elem_dev->super.type != PARSEC_DEV_CUDA ?
                                                             cudaMemcpyDeviceToHost : cudaMemcpyDeviceToDevice,
                                                     cuda_stream->cuda_stream);

src/dplasmaaux.c

@@ -98,7 +98,7 @@ dplasma_aux_getGEMMLookahead( parsec_tiled_matrix_t *A )
      * look ahead based on the global information to get the same one on all
      * nodes.
      */
-    int nbunits = vpmap_get_nb_total_threads() * A->super.nodes;
+    int nbunits = parsec_vpmap_get_nb_total_threads() * A->super.nodes;
     double alpha =  3. * (double)nbunits / ( A->mt * A->nt );
 
     if ( A->super.nodes == 1 ) {

src/scalapack_wrappers/dplasma_wrapper_parsec_init.c

@@ -2,13 +2,12 @@
  * Copyright (c) 2021-2023 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
+ * Copyright (c) 2026      NVIDIA Corporation.  All rights reserved.
  */
 
 #include "common.h"
 
 
-#define APP_NAME "SCALAPACK_WRAPPED_CALL"
-
 static void parsec_init_wrapper_internal(){
   if( parsec_ctx == NULL ){
 
@@ -27,16 +26,6 @@ static void parsec_init_wrapper_internal(){
     if(var_ncores!=NULL){
         ncores = atoi(var_ncores);
     }
-    int parsec_argc = 1;
-
-    char** parsec_argv = (char**)calloc(parsec_argc+1, sizeof(char*));
-    int i;
-    for(i=0; i<parsec_argc; i++){
-        parsec_argv[i] = (char*)malloc(sizeof(char)*80);
-    }
-    parsec_argv[parsec_argc] = NULL;
-    sprintf(parsec_argv[0], APP_NAME);
-
     int rank,size;
     MPI_Comm_rank(MPI_COMM_WORLD, &rank);
     MPI_Comm_size(MPI_COMM_WORLD, &size);
@@ -45,17 +34,12 @@ static void parsec_init_wrapper_internal(){
     SYNC_TIME_START();
 #endif
 
-    parsec_ctx = parsec_init(ncores, &parsec_argc, &parsec_argv);
+    parsec_ctx = parsec_init(ncores, NULL, NULL);
 
 #ifdef MEASURE_INTERNAL_TIMES
     SYNC_TIME_PRINT(rank, ("PaRSEC initialized\n"));
 #endif
 
-    for(i=0; i<parsec_argc; i++){
-        free(parsec_argv[i]);
-    }
-    free(parsec_argv);
-
   }
 }
 
@@ -128,4 +112,3 @@ void parsec_wrapper_devices_reset_load_(void){
         parsec_devices_reset_load(parsec_ctx);
 	}
 }
-

src/utils/dplasma_lapack_adtt.c

@@ -288,7 +288,7 @@ static parsec_data_t* data_of(parsec_data_collection_t *desc, ...)
                 "data_of CP %p [old type %p] loc %d -> dtt %p target_shape %d layout %d",
                 cp, cp->dtt, loc, adt->opaque_dtt, info.shape, info.layout);
             dt = parsec_data_create_with_type( dt->dc,
-                                               dt->key, cp->device_private, dt->nb_elts,
+                                               dt->key, cp->device_private, dt->span,
                                                adt->opaque_dtt);
         }
     }
@@ -392,4 +392,3 @@ void dplasma_unwrap_data_collection(dplasma_data_collection_t *ddc)
 {
     free(ddc);
 }
-

src/zgebrd_ge2gb_wrapper.c

@@ -310,7 +310,7 @@ dplasma_zgebrd_ge2gb_New( int ib,
     dplasma_qrtree_t *qrtre0, *qrtree, *lqtree;
     int P, Q, cores;
 
-    cores = dplasma_imax( vpmap_get_nb_total_threads(), 1 );
+    cores = dplasma_imax( parsec_vpmap_get_nb_total_threads(), 1 );
     qrtree = malloc( sizeof(dplasma_qrtree_t) );
     lqtree = malloc( sizeof(dplasma_qrtree_t) );
 
@@ -613,4 +613,3 @@ dplasma_zgebrd_ge2gb( parsec_context_t *parsec, int ib,
 
     return 0;
 }
-

src/zgetrf_1d.jdf

@@ -57,7 +57,7 @@ descIPIV    [type = "parsec_tiled_matrix_t*" hidden = on default = "((dplasma_da
 
 INFO      [type = "int*"]
 
-nbmaxthrd [type = "int" hidden=on default="( dplasma_imax( 1, dplasma_imin( vpmap_get_nb_threads_in_vp(0) - 1, 48 ) ) )" ] /* 48 is the actual limit of the kernel */
+nbmaxthrd [type = "int" hidden=on default="( dplasma_imax( 1, dplasma_imin( parsec_vpmap_get_vp_threads(0) - 1, 48 ) ) )" ] /* 48 is the actual limit of the kernel */
 KT        [type = "int" hidden=on default="( dplasma_imin( descA->mt, descA->nt )-1 )" ]
 
 getrfdata [type = "CORE_zgetrf_data_t *" hidden = on default = "NULL"]

src/zgetrf_1d_wrapper.c

@@ -83,7 +83,7 @@ dplasma_zgetrf_1d_New( parsec_tiled_matrix_t *A,
                        int *INFO )
 {
     parsec_zgetrf_1d_taskpool_t *parsec_getrf_1d;
-    int nbthreads = dplasma_imax( 1, vpmap_get_nb_threads_in_vp(0) - 1 );
+    int nbthreads = dplasma_imax( 1, parsec_vpmap_get_vp_threads(0) - 1 );
     dplasma_data_collection_t * ddc_A = dplasma_wrap_data_collection((parsec_tiled_matrix_t*)A);
     dplasma_data_collection_t * ddc_IPIV = dplasma_wrap_data_collection((parsec_tiled_matrix_t*)IPIV);
 

src/zgetrf_qrf.jdf

@@ -52,7 +52,7 @@ INFO     [type = "int*"]
 param_p   [type = int default="((parsec_matrix_block_cyclic_t*)descA)->grid.rows" hidden=on ]
 param_q   [type = int default="((parsec_matrix_block_cyclic_t*)descA)->grid.cols" hidden=on ]
 minMNT    [type = int default="dplasma_imin( descA->mt-1, descA->nt-1 )"    hidden=on ]
-nbmaxthrd [type = "int" default="( dplasma_imax( 1, dplasma_imin( vpmap_get_nb_threads_in_vp(0) - 1, 48 ) ) )" hidden=on] /* 48 is the actual limit of the kernel */
+nbmaxthrd [type = "int" default="( dplasma_imax( 1, dplasma_imin( parsec_vpmap_get_vp_threads(0) - 1, 48 ) ) )" hidden=on] /* 48 is the actual limit of the kernel */
 
 getrfdata [type = "CORE_zgetrf_data_t *" hidden = on default = "NULL"]
 

src/zgetrf_qrf_wrapper.c

@@ -166,7 +166,7 @@ dplasma_zgetrf_qrf_New( dplasma_qrtree_t *qrtree,
     int ib = TS->mb;
     size_t sizeW = 1;
     size_t sizeReduceVec = 1;
-    int nbthreads = dplasma_imax( 1, vpmap_get_nb_threads_in_vp(0) - 1 );
+    int nbthreads = dplasma_imax( 1, parsec_vpmap_get_vp_threads(0) - 1 );
 
     /*
      * Compute W size according to criteria used.

tests/common.c

@@ -2,6 +2,7 @@
  * Copyright (c) 2009-2024 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
+ * Copyright (c) 2026      NVIDIA Corporation.  All rights reserved.
  *
  */
 #include "parsec/runtime.h"
@@ -580,6 +581,23 @@ static void print_arguments(int* iparam)
     }
 }
 
+static void get_parsec_arguments(int argc, char **argv,
+                                 int *parsec_argc, char ***parsec_argv)
+{
+    int idx;
+
+    *parsec_argc = 0;
+    *parsec_argv = NULL;
+
+    for( idx = 1; idx < argc; idx++ ) {
+        if( 0 == strcmp(argv[idx], "--") ) {
+            *parsec_argc = argc - idx - 1;
+            *parsec_argv = (0 < *parsec_argc) ? &argv[idx + 1] : NULL;
+            return;
+        }
+    }
+}
+
 
 
 
@@ -671,19 +689,13 @@ parsec_context_t* setup_parsec(int argc, char **argv, int *iparam)
 
     TIME_START();
 
-    /* Once we got out arguments, we should pass whatever is left down */
-    int parsec_argc, idx;
-    char** parsec_argv = (char**)calloc(argc, sizeof(char*));
-    parsec_argv[0] = argv[0];  /* the app name */
-    for( idx = parsec_argc = 1;
-         (idx < argc) && (0 != strcmp(argv[idx], "--")); idx++);
-    if( idx != argc ) {
-        for( parsec_argc = 1, idx++; idx < argc;
-             parsec_argv[parsec_argc] = argv[idx], parsec_argc++, idx++);
-    }
+    /* PaRSEC only accepts its own arguments, without argv[0] or "--". */
+    int parsec_argc;
+    char **parsec_argv;
+    get_parsec_arguments(argc, argv, &parsec_argc, &parsec_argv);
     parsec_context_t* ctx = parsec_init(iparam[IPARAM_NCORES],
-                                        &parsec_argc, &parsec_argv);
-    free(parsec_argv);
+                                        (0 < parsec_argc) ? &parsec_argc : NULL,
+                                        (0 < parsec_argc) ? &parsec_argv : NULL);
     if( NULL == ctx ) {
         /* Failed to correctly initialize. In a correct scenario report
          * upstream, but in this particular case bail out.
@@ -778,4 +790,3 @@ void cleanup_parsec(parsec_context_t* parsec, int *iparam)
 #endif
     (void)iparam;
 }
-