From 8755a1601014cc055e4407598fcd698f0a890310 Mon Sep 17 00:00:00 2001
From: Philippe Virouleau <philippe.44@gmail.com>
Date: Thu, 11 Oct 2018 13:54:22 +0200
Subject: [PATCH] Cleanup
---
CMakeLists.txt | 108 +----------------------
compute/zpotrf.c | 1 -
include/chameleon.h | 10 +++
runtime/openmp/CMakeLists.txt | 32 +------
runtime/openmp/control/runtime_options.c | 12 ++-
timing/time_zgeqrf_tile.c | 6 --
timing/timing.c | 67 ++++++++------
7 files changed, 59 insertions(+), 177 deletions(-)
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5f73d2271..cf5f355a5 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -906,120 +906,14 @@ if( CHAMELEON_SCHED_OPENMP )
find_package(OpenMP COMPONENTS ${OPENMP_COMPONENT_LIST})
- # Add definition and include_dir if found
- # TODO
if ( OPENMP_FOUND )
message("-- ${Blue}Add definition CHAMELEON_SCHED_OPENMP"
" - Activate OpenMP in Chameleon${ColourReset}")
- #if (STARPU_INCLUDE_DIRS_DEP)
- #include_directories(${STARPU_INCLUDE_DIRS_DEP})
- #set(CMAKE_REQUIRED_INCLUDES "${STARPU_INCLUDE_DIRS_DEP}")
- #endif()
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_CXX_FLAGS}")
- #if(STARPU_LDFLAGS_OTHER_DEP)
- #set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${STARPU_LDFLAGS_OTHER_DEP}")
- #set(CMAKE_REQUIRED_LDFLAGS "${STARPU_LDFLAGS_OTHER_DEP}")
- #endif()
- #if(STARPU_LIBRARY_DIRS_DEP)
- #list(APPEND CMAKE_INSTALL_RPATH "${STARPU_LIBRARY_DIRS_DEP}")
- #endif()
- #if (STARPU_LIBRARIES_DEP)
- #list(INSERT CHAMELEON_DEP 0 ${STARPU_LIBRARIES_DEP})
- #set(CMAKE_REQUIRED_LIBRARIES "${STARPU_LIBRARIES_DEP}")
- #endif()
- #if (CHAMELEON_SIMULATION)
- #list(APPEND CMAKE_REQUIRED_FLAGS "-include" "starpu_simgrid_wrap.h")
- #endif()
- #string(REPLACE ";" " " CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS}")
- #check_function_exists(starpu_data_idle_prefetch_on_node HAVE_STARPU_IDLE_PREFETCH)
- #if ( HAVE_STARPU_IDLE_PREFETCH )
- #message("-- ${Blue}Add definition HAVE_STARPU_IDLE_PREFETCH${ColourReset}")
- #endif()
- #check_function_exists(starpu_iteration_push HAVE_STARPU_ITERATION_PUSH)
- #if ( HAVE_STARPU_ITERATION_PUSH )
- #message("-- ${Blue}Add definition HAVE_STARPU_ITERATION_PUSH${ColourReset}")
- #endif()
- #check_function_exists(starpu_data_wont_use HAVE_STARPU_DATA_WONT_USE)
- #if ( HAVE_STARPU_DATA_WONT_USE )
- #message("-- ${Blue}Add definition HAVE_STARPU_DATA_WONT_USE${ColourReset}")
- #endif()
- #check_function_exists(starpu_data_set_coordinates HAVE_STARPU_DATA_SET_COORDINATES)
- #if ( HAVE_STARPU_DATA_SET_COORDINATES )
- #message("-- ${Blue}Add definition HAVE_STARPU_DATA_SET_COORDINATES${ColourReset}")
- #endif()
- #check_function_exists(starpu_malloc_on_node_set_default_flags HAVE_STARPU_MALLOC_ON_NODE_SET_DEFAULT_FLAGS)
- #if ( HAVE_STARPU_MALLOC_ON_NODE_SET_DEFAULT_FLAGS )
- #message("-- ${Blue}Add definition HAVE_STARPU_MALLOC_ON_NODE_SET_DEFAULT_FLAGS${ColourReset}")
- #endif()
- #check_function_exists(starpu_mpi_data_migrate HAVE_STARPU_MPI_DATA_MIGRATE)
- #if ( HAVE_STARPU_MPI_DATA_MIGRATE )
- #message("-- ${Blue}Add definition HAVE_STARPU_MPI_DATA_MIGRATE${ColourReset}")
- #elseif(CHAMELEON_USE_MIGRATE)
- #set(CHAMELEON_USE_MIGRATE "OFF")
- #message("-- ${Blue}CHAMELEON_USE_MIGRATE is turned OFF because starpu_mpi_data_migrate not found${ColourReset}")
- #endif()
+ # TODO: check for compiler support for "_Pragma"
else ( OPENMP_FOUND )
message(FATAL_ERROR "Something went wrong when finding OpenMP")
- #if(MORSE_VERBOSE_FIND_PACKAGE)
- #if(NOT HWLOC_FOUND OR NOT HWLOC_LIBRARIES)
- #if (NOT HWLOC_hwloc.h_DIRS)
- #Print_Find_Header_Status(hwloc hwloc.h)
- #endif ()
- #if (NOT HWLOC_hwloc_LIBRARY)
- #Print_Find_Library_Status(hwloc libhwloc)
- #endif ()
- #endif()
- #if(CHAMELEON_ENABLE_TRACING AND (NOT FXT_FOUND OR NOT FXT_LIBRARIES))
- #if (NOT FXT_fxt.h_DIRS)
- #Print_Find_Header_Status(fxt fxt.h)
- #endif ()
- #if (NOT FXT_fxt_LIBRARY)
- #Print_Find_Library_Status(fxt libfxt)
- #endif ()
- #endif()
- #if(CHAMELEON_SIMULATION AND (NOT SIMGRID_FOUND OR NOT SIMGRID_LIBRARIES))
- #if (NOT SIMGRID_simgrid.h_DIRS)
- #Print_Find_Header_Status(simgrid simgrid.h)
- #endif ()
- #if (NOT SIMGRID_simgrid_LIBRARY)
- #Print_Find_Library_Status(simgrid libsimgrid)
- #endif ()
- #endif()
- #if( (NOT STARPU_SHM_FOUND) OR (NOT STARPU_SHM_LIBRARIES) OR
- #( STARPU_LOOK_FOR_MPI AND (NOT STARPU_MPI_FOUND OR NOT STARPU_MPI_LIBRARIES) )
- #)
- #foreach(starpu_hdr ${STARPU_hdrs_to_find})
- #if (NOT STARPU_${starpu_hdr}_INCLUDE_DIRS)
- #Print_Find_Header_Status(starpu ${starpu_hdr})
- #endif ()
- #endforeach()
- #if(STARPU_VERSION_STRING)
- #foreach(starpu_lib ${STARPU_libs_to_find})
- #if (NOT STARPU_${starpu_lib}_LIBRARY)
- #Print_Find_Library_Status(starpu ${starpu_lib})
- #endif ()
- #endforeach()
- #endif ()
- #endif ()
- #else(MORSE_VERBOSE_FIND_PACKAGE)
- #message(WARNING "StarPU library has not been found and MORSE_VERBOSE_FIND_PACKAGE is set to OFF."
- #" Try to activate MORSE_VERBOSE_FIND_PACKAGE option (-DMORSE_VERBOSE_FIND_PACKAGE=ON) to get some hints for the detection")
- #endif(MORSE_VERBOSE_FIND_PACKAGE)
- #if(NOT HWLOC_FOUND OR NOT HWLOC_LIBRARIES)
- #message(FATAL_ERROR "hwloc library is required but has not been found")
- #endif()
- #if(CHAMELEON_SIMULATION AND (NOT SIMGRID_FOUND OR NOT SIMGRID_LIBRARIES))
- #message(FATAL_ERROR "SimGrid library is required but has not been found")
- #endif()
- #if(CHAMELEON_ENABLE_TRACING AND (NOT FXT_FOUND OR NOT FXT_LIBRARIES))
- #message(FATAL_ERROR "FxT library is required but has not been found")
- #endif()
- #if( (NOT STARPU_SHM_FOUND) OR (NOT STARPU_SHM_LIBRARIES) OR
- #( STARPU_LOOK_FOR_MPI AND (NOT STARPU_MPI_FOUND OR NOT STARPU_MPI_LIBRARIES) )
- #)
- #message(FATAL_ERROR "StarPU library is required but has not been found")
- #endif()
endif ( OPENMP_FOUND )
endif( CHAMELEON_SCHED_OPENMP )
diff --git a/compute/zpotrf.c b/compute/zpotrf.c
index f9ff8878f..bb8485337 100644
--- a/compute/zpotrf.c
+++ b/compute/zpotrf.c
@@ -23,7 +23,6 @@
*
*/
#include "control/common.h"
-#include <stdio.h>
/**
********************************************************************************
diff --git a/include/chameleon.h b/include/chameleon.h
index da34cab2a..a8afc08c8 100644
--- a/include/chameleon.h
+++ b/include/chameleon.h
@@ -135,6 +135,16 @@ int CHAMELEON_Sequence_Create (RUNTIME_sequence_t **sequence);
int CHAMELEON_Sequence_Destroy (RUNTIME_sequence_t *sequence);
int CHAMELEON_Sequence_Wait (RUNTIME_sequence_t *sequence);
+#if defined(CHAMELEON_SCHED_OPENMP)
+#define CHAMELEON_INIT(nworkers, ncudas)\
+ CHAMELEON_Init(nworkers, ncudas);\
+ _Pragma("omp parallel")\
+ _Pragma("omp master")
+#else
+#define CHAMELEON_INIT(nworkers, ncudas)\
+ CHAMELEON_Init(nworkers, ncudas);
+#endif
+
END_C_DECLS
#endif /* _chameleon_h_ */
diff --git a/runtime/openmp/CMakeLists.txt b/runtime/openmp/CMakeLists.txt
index d072f05be..6c0ea5ab5 100644
--- a/runtime/openmp/CMakeLists.txt
+++ b/runtime/openmp/CMakeLists.txt
@@ -4,7 +4,7 @@
#
# @copyright 2009-2015 The University of Tennessee and The University of
# Tennessee Research Foundation. All rights reserved.
-# @copyright 2012-2017 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+# @copyright 2012-2018 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
# Univ. Bordeaux. All rights reserved.
#
###
@@ -30,17 +30,6 @@ cmake_minimum_required(VERSION 2.8)
include_directories( ${CMAKE_CURRENT_SOURCE_DIR}/include )
include_directories( ${CMAKE_CURRENT_BINARY_DIR}/include )
-# Generate headers for all possible precisions
-# --------------------------------------------
-#set(RUNTIME_HDRS_GENERATED "")
-#set(ZHDR
- #include/runtime_codelet_z.h
- #)
-
-#precisions_rules_py(RUNTIME_HDRS_GENERATED "${ZHDR}"
- #PRECISIONS "s;d;c;z;ds;zc"
- #TARGETDIR "include")
-
# Define the list of headers
# --------------------------
set(RUNTIME_HDRS
@@ -71,17 +60,6 @@ install(
# Generate the Chameleon common for all possible precisions
# ---------------------------------------------------------
-# FIXME: restore this generation once file is done
-#set(RUNTIME_COMMON_GENERATED "")
-#set(ZSRC
- #control/runtime_zprofiling.c
- #control/runtime_zlocality.c
- #)
-
-#precisions_rules_py(
- #RUNTIME_COMMON_GENERATED "${ZSRC}"
- #PRECISIONS "${CHAMELEON_PRECISION}"
- #TARGETDIR "control")
set(RUNTIME_COMMON
control/runtime_async.c
@@ -90,21 +68,13 @@ set(RUNTIME_COMMON
control/runtime_descriptor.c
control/runtime_options.c
control/runtime_profiling.c
- #control/runtime_workspace.c
${RUNTIME_COMMON_GENERATED}
)
-#set(flags_to_add "")
-#foreach(_prec ${CHAMELEON_PRECISION})
- #set(flags_to_add "${flags_to_add} -DPRECISION_${_prec}")
-#endforeach()
-#set_source_files_properties(control/runtime_profiling.c PROPERTIES COMPILE_FLAGS "${flags_to_add}")
-
# Generate the Chameleon sources for all possible precisions
# ----------------------------------------------------------
set(RUNTIME_SRCS_GENERATED "")
set(ZSRC
- #codelets/codelet_zcallback.c
${CODELETS_ZSRC}
)
diff --git a/runtime/openmp/control/runtime_options.c b/runtime/openmp/control/runtime_options.c
index 83b848012..4511f102b 100644
--- a/runtime/openmp/control/runtime_options.c
+++ b/runtime/openmp/control/runtime_options.c
@@ -50,8 +50,11 @@ int RUNTIME_options_ws_alloc( RUNTIME_option_t *options, size_t worker_size, siz
options->ws_worker = malloc(worker_size* sizeof(char));
options->ws_wsize = worker_size;
}
- // TODO maybe we'll need it at some point
- options->ws_hsize = host_size;
+ if (host_size > 0) {
+ // TODO used for scratch, maybe we can do better than malloc
+ options->ws_host = malloc(host_size * sizeof(char));
+ options->ws_hsize = host_size;
+ }
return CHAMELEON_SUCCESS;
}
@@ -61,6 +64,9 @@ int RUNTIME_options_ws_free( RUNTIME_option_t *options )
free(options->ws_worker);
options->ws_wsize = 0;
}
- options->ws_hsize = 0;
+ if (options->ws_hsize) {
+ free(options->ws_host);
+ options->ws_hsize = 0;
+ }
return CHAMELEON_SUCCESS;
}
diff --git a/timing/time_zgeqrf_tile.c b/timing/time_zgeqrf_tile.c
index e04bfb8a6..b35782a69 100644
--- a/timing/time_zgeqrf_tile.c
+++ b/timing/time_zgeqrf_tile.c
@@ -48,14 +48,8 @@ RunTest(int *iparam, double *dparam, chameleon_time_t *t_)
memset(descT->mat, 0, (descT->llm*descT->lln)*sizeof(ChamComplexDouble));
/* CHAMELEON ZGEQRF */
-#if defined (CHAMELEON_SCHED_OPENMP)
-#pragma omp parallel
-#pragma omp master
-#endif
- {
START_TIMING();
CHAMELEON_zgeqrf_Tile( descA, descT );
- }
STOP_TIMING();
/* Check the solution */
diff --git a/timing/timing.c b/timing/timing.c
index ff268c719..54476685c 100644
--- a/timing/timing.c
+++ b/timing/timing.c
@@ -619,46 +619,25 @@ parse_arguments(int *_argc, char ***_argv, int *iparam, int *start, int *stop, i
} while(-1 != c);
}
+// NOTE: this function is here to cope with the fact that OpenMP parallel
+// regions must not have instructions jumping outside the region (eg: returns)
+
int
-main(int argc, char *argv[]) {
- int i, m, n, mx, nx;
+CHAMELEON_Main(int *iparam, char *prog_name, int start, int stop, int step) {
+
int status;
+ int i, m, n, mx, nx;
int nbnode = 1;
- int start = 500;
- int stop = 5000;
- int step = 500;
- int iparam[IPARAM_SIZEOF];
int success = 0;
- set_iparam_default(iparam);
-
- parse_arguments(&argc, &argv, iparam, &start, &stop, &step);
-
-#if !defined(CHAMELEON_USE_CUDA)
- if (iparam[IPARAM_NCUDAS] != 0){
- fprintf(stderr, "ERROR: CHAMELEON_USE_CUDA is not defined. "
- "The number of CUDA devices must be set to 0 (--gpus=0).\n");
- return EXIT_FAILURE;
- }
-#endif
-
n = iparam[IPARAM_N];
m = iparam[IPARAM_M];
mx = iparam[IPARAM_MX];
nx = iparam[IPARAM_NX];
- /* Initialize CHAMELEON */
- CHAMELEON_Init( iparam[IPARAM_THRDNBR],
- iparam[IPARAM_NCUDAS] );
-#if defined (CHAMELEON_SCHED_OPENMP)
-#pragma omp parallel
-#pragma omp master
-#endif
- {
/* Get the number of threads set by the runtime */
iparam[IPARAM_THRDNBR] = CHAMELEON_GetThreadNbr();
- }
/* Stops profiling here to avoid profiling uninteresting routines.
It will be reactivated in the time_*.c routines with the macro START_TIMING() */
@@ -709,7 +688,7 @@ main(int argc, char *argv[]) {
CHAMELEON_Set(CHAMELEON_TRANSLATION_MODE, iparam[IPARAM_INPLACE]);
if ( CHAMELEON_My_Mpi_Rank() == 0 ) {
- print_header( argv[0], iparam);
+ print_header( prog_name, iparam);
}
if (step < 1) step = 1;
@@ -749,7 +728,37 @@ main(int argc, char *argv[]) {
if (status != CHAMELEON_SUCCESS) return status;
success += status;
}
+}
+
+int
+main(int argc, char *argv[]) {
+ int start = 500;
+ int stop = 5000;
+ int step = 500;
+ int iparam[IPARAM_SIZEOF];
+
+ set_iparam_default(iparam);
+
+ parse_arguments(&argc, &argv, iparam, &start, &stop, &step);
+
+#if !defined(CHAMELEON_USE_CUDA)
+ if (iparam[IPARAM_NCUDAS] != 0){
+ fprintf(stderr, "ERROR: CHAMELEON_USE_CUDA is not defined. "
+ "The number of CUDA devices must be set to 0 (--gpus=0).\n");
+ return EXIT_FAILURE;
+ }
+#endif
+ int return_code;
+
+ /* Initialize CHAMELEON */
+ CHAMELEON_INIT( iparam[IPARAM_THRDNBR],
+ iparam[IPARAM_NCUDAS] );
+
+ {
+ return_code = CHAMELEON_Main(iparam, argv[0], start, stop, step);
+ }
+
CHAMELEON_Finalize();
- return success;
+ return return_code;
}
--
GitLab