Updated zScatter_B3d() to support uneven block row partition of {A,B}…

… in 3D interface. Only complex16 is implemented.
xiaoyeli · Sep 21, 2021 · fd71062 · fd71062
1 parent 7375880
commit fd71062
Show file tree

Hide file tree

Showing 36 changed files with 386 additions and 74 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -21,7 +21,7 @@ list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
 option(enable_doc       "Build doxygen documentation" OFF)
 option(enable_double    "Enable double precision library" ON)
 option(enable_single    "Enable single precision library" OFF)
-option(enable_complex16 "Enable complex16 precision library" OFF)
+option(enable_complex16 "Enable complex16 precision library" ON)
 option(enable_tests  "Build tests" ON)
 option(enable_examples  "Build examples" ON)
 option(XSDK_ENABLE_Fortran "Enable Fortran" ON)
@@ -224,9 +224,9 @@ if (enable_openmp)
   if(OPENMP_FOUND)
     set(CMAKE_C_FLAGS "${OpenMP_C_FLAGS} ${CMAKE_C_FLAGS}")
     set(CMAKE_CXX_FLAGS "${OpenMP_CXX_FLAGS} ${CMAKE_CXX_FLAGS}")
-# On edison, OpenMP_EXE_LINKER_FLAGS is empty
-#  set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
-    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_C_FLAGS}")
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
+# The following causes problem with cmake/3.20.+
+# set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_C_FLAGS}")
     message("-- OpenMP_EXE_LINKER_FLAGS='${OpenMP_EXE_LINKER_FLAGS}'")
     message("-- CMAKE_EXE_LINKER_FLAGS='${CMAKE_EXE_LINKER_FLAGS}'")
   endif()

diff --git a/EXAMPLE/Makefile b/EXAMPLE/Makefile
@@ -53,7 +53,7 @@ ZEXM1	= pzdrive1.o zcreate_matrix.o
 ZEXM2	= pzdrive2.o zcreate_matrix.o zcreate_matrix_perturbed.o
 ZEXM3	= pzdrive3.o zcreate_matrix.o
 ZEXM4	= pzdrive4.o zcreate_matrix.o
-ZEXM3D	= pzdrive3d.o zcreate_matrix.o zcreate_matrix3d.o
+ZEXM3D	= pzdrive3d.o zcreate_matrix.o zcreate_matrix3d.o #znrformat_loc3d.o
 ZEXM3D1	= pzdrive3d1.o zcreate_matrix.o zcreate_matrix3d.o
 ZEXM3D2	= pzdrive3d2.o zcreate_matrix.o zcreate_matrix3d.o
 ZEXM3D3	= pzdrive3d3.o zcreate_matrix.o zcreate_matrix3d.o

diff --git a/EXAMPLE/pddrive1.c b/EXAMPLE/pddrive1.c
@@ -73,7 +73,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init_thread( &argc, &argv, MPI_THREAD_MULTIPLE, &omp_mpi_level); 
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pddrive1_ABglobal.c b/EXAMPLE/pddrive1_ABglobal.c
@@ -72,6 +72,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
 
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {

diff --git a/EXAMPLE/pddrive2_ABglobal.c b/EXAMPLE/pddrive2_ABglobal.c
@@ -72,7 +72,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pddrive3_ABglobal.c b/EXAMPLE/pddrive3_ABglobal.c
@@ -78,7 +78,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pddrive4.c b/EXAMPLE/pddrive4.c
@@ -74,7 +74,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init_thread( &argc, &argv, MPI_THREAD_MULTIPLE, &omp_mpi_level); 
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     MPI_Comm_size( MPI_COMM_WORLD, &nprocs );
     if ( nprocs < 10 ) {
 	fprintf(stderr, "Requires at least 10 processes\n");

diff --git a/EXAMPLE/pddrive4_ABglobal.c b/EXAMPLE/pddrive4_ABglobal.c
@@ -71,6 +71,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif	
     MPI_Comm_size( MPI_COMM_WORLD, &nprocs );
     if ( nprocs < 10 ) {
 	fprintf(stderr, "Requires at least 10 processes\n");

diff --git a/EXAMPLE/pddrive_ABglobal.c b/EXAMPLE/pddrive_ABglobal.c
@@ -73,7 +73,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pddrive_spawn.c b/EXAMPLE/pddrive_spawn.c
@@ -82,7 +82,12 @@ int main(int argc, char *argv[])
     //MPI_Init( &argc, &argv );
     MPI_Init_thread( &argc, &argv, MPI_THREAD_MULTIPLE, &omp_mpi_level); 
 	MPI_Comm_get_parent(&parent);   	
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif	
 
 
 #if ( VAMPIR>=1 )

diff --git a/EXAMPLE/pzdrive.c b/EXAMPLE/pzdrive.c
@@ -74,7 +74,12 @@ int main(int argc, char *argv[])
        ------------------------------------------------------------*/
     //MPI_Init( &argc, &argv );
     MPI_Init_thread( &argc, &argv, MPI_THREAD_MULTIPLE, &omp_mpi_level); 
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
 
 #if ( VAMPIR>=1 )
     VT_traceoff(); 

diff --git a/EXAMPLE/pzdrive1.c b/EXAMPLE/pzdrive1.c
@@ -72,7 +72,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init_thread( &argc, &argv, MPI_THREAD_MULTIPLE, &omp_mpi_level); 
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pzdrive1_ABglobal.c b/EXAMPLE/pzdrive1_ABglobal.c
@@ -71,7 +71,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pzdrive2_ABglobal.c b/EXAMPLE/pzdrive2_ABglobal.c
@@ -71,7 +71,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pzdrive3_ABglobal.c b/EXAMPLE/pzdrive3_ABglobal.c
@@ -77,7 +77,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pzdrive3d.c b/EXAMPLE/pzdrive3d.c
@@ -183,6 +183,7 @@ main (int argc, char *argv[])
        INITIALIZE THE SUPERLU PROCESS GRID.
        ------------------------------------------------------------ */
     superlu_gridinit3d (MPI_COMM_WORLD, nprow, npcol, npdep, &grid);
+    //    grid.rankorder = 1;
 
     if(grid.iam==0) {
 	MPI_Query_thread(&omp_mpi_level);

diff --git a/EXAMPLE/pzdrive4.c b/EXAMPLE/pzdrive4.c
@@ -73,7 +73,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init_thread( &argc, &argv, MPI_THREAD_MULTIPLE, &omp_mpi_level); 
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     MPI_Comm_size( MPI_COMM_WORLD, &nprocs );
     if ( nprocs < 10 ) {
 	fprintf(stderr, "Requires at least 10 processes\n");

diff --git a/EXAMPLE/pzdrive4_ABglobal.c b/EXAMPLE/pzdrive4_ABglobal.c
@@ -70,6 +70,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif	
     MPI_Comm_size( MPI_COMM_WORLD, &nprocs );
     if ( nprocs < 10 ) {
 	fprintf(stderr, "Requires at least 10 processes\n");

diff --git a/EXAMPLE/pzdrive_ABglobal.c b/EXAMPLE/pzdrive_ABglobal.c
@@ -72,7 +72,12 @@ int main(int argc, char *argv[])
        INITIALIZE MPI ENVIRONMENT. 
        ------------------------------------------------------------*/
     MPI_Init( &argc, &argv );
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif
     /* Parse command line argv[]. */
     for (cpp = argv+1; *cpp; ++cpp) {
 	if ( **cpp == '-' ) {

diff --git a/EXAMPLE/pzdrive_spawn.c b/EXAMPLE/pzdrive_spawn.c
@@ -82,7 +82,12 @@ int main(int argc, char *argv[])
     //MPI_Init( &argc, &argv );
     MPI_Init_thread( &argc, &argv, MPI_THREAD_MULTIPLE, &omp_mpi_level); 
 	MPI_Comm_get_parent(&parent);   	
-
+#ifdef GPU_ACC
+    int rank, devs;
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+    cudaGetDeviceCount(&devs);
+    cudaSetDevice(rank % devs);
+#endif	
 
 
 #if ( VAMPIR>=1 )

diff --git a/EXAMPLE/zcreate_matrix3d.c b/EXAMPLE/zcreate_matrix3d.c
@@ -341,17 +341,44 @@ int zcreate_matrix_postfix3d(SuperMatrix *A, int nrhs, doublecomplex **rhs,
     nzval[0] = 0.1;
 #endif
 
-    /* Compute the number of rows to be distributed to local process */
-    m_loc = m / (grid3d->nprow * grid3d->npcol* grid3d->npdep);
-    m_loc_fst = m_loc;
-    /* When m / procs is not an integer */
-    if ((m_loc * grid3d->nprow * grid3d->npcol* grid3d->npdep) != m)
-    {
-        /*m_loc = m_loc+1;
-          m_loc_fst = m_loc;*/
-        if (iam == (grid3d->nprow * grid3d->npcol* grid3d->npdep - 1)) /* last proc. gets all*/
-            m_loc = m - m_loc * (grid3d->nprow * grid3d->npcol* grid3d->npdep - 1);
-    }
+//    /* Compute the number of rows to be distributed to local process */
+//    m_loc = m / (grid3d->nprow * grid3d->npcol* grid3d->npdep);
+//    m_loc_fst = m_loc;
+//    /* When m / procs is not an integer */
+//    if ((m_loc * grid3d->nprow * grid3d->npcol* grid3d->npdep) != m)
+//    {
+//        /*m_loc = m_loc+1;
+//          m_loc_fst = m_loc;*/
+//        if (iam == (grid3d->nprow * grid3d->npcol* grid3d->npdep - 1)) /* last proc. gets all*/
+//            m_loc = m - m_loc * (grid3d->nprow * grid3d->npcol* grid3d->npdep - 1);
+//    }
+
+    switch(iam) {
+      case 0:
+       m_loc=111; fst_row=0;
+       break;
+      case 1:
+       m_loc=84; fst_row=111;
+       break;
+      case 2:
+       m_loc=108; fst_row=195;
+       break;
+      case 3:
+       m_loc=84; fst_row=303;
+       break;
+      case 4:
+       m_loc=108; fst_row=387;
+       break;
+      case 5:
+       m_loc=84; fst_row=495;
+       break;
+      case 6:
+       m_loc=108; fst_row=579;
+       break;
+      case 7:
+       m_loc=84; fst_row=687;
+       break;
+     }
 
     /* Create compressed column matrix for GA. */
     zCreate_CompCol_Matrix_dist(&GA, m, n, nnz, nzval, rowind, colptr,
@@ -379,7 +406,7 @@ int zcreate_matrix_postfix3d(SuperMatrix *A, int nrhs, doublecomplex **rhs,
         for (j = colptr[i]; j < colptr[i + 1]; ++j) ++marker[rowind[j]];
     /* Set up row pointers */
     rowptr[0] = 0;
-    fst_row = iam * m_loc_fst;
+//    fst_row = iam * m_loc_fst;
     nnz_loc = 0;
     for (j = 0; j < m_loc; ++j)
     {

diff --git a/SRC/CMakeLists.txt b/SRC/CMakeLists.txt
@@ -294,6 +294,12 @@ if(CUDAToolkit_FOUND)  # this is found in top-level CMakeLists.txt
     target_link_libraries(superlu_dist CUDA::cudart CUDA::cublas)
 endif()
 
+# This is recommended by modern cmake:
+# https://cliutils.gitlab.io/modern-cmake/chapters/packages/OpenMP.html
+if(OpenMP_FOUND) # this is found in top-level CMakeLists.txt
+  target_link_libraries(superlu_dist OpenMP::OpenMP_C)
+endif()
+
 target_compile_definitions(superlu_dist PRIVATE SUPERLU_DIST_EXPORTS)
 if(MSVC AND BUILD_SHARED_LIBS)
   set_target_properties(superlu_dist PROPERTIES

diff --git a/SRC/dlustruct_gpu.h b/SRC/dlustruct_gpu.h
@@ -78,16 +78,17 @@ typedef struct //LUstruct_gpu_
     int_t   *LrowindPtr;      /* A single vector */
 
     double  *LnzvalVec;       /* A single vector */
-    int_t   *LnzvalPtr;       /* A single vector */
-    int_t   *LnzvalPtr_host;  /* A single vector */
+    int_t   *LnzvalPtr;        /* A single vector */
+    int_t   *LnzvalPtr_host;   /* A single vector */
 
     int_t   *UrowindVec;            /* A single vector */
     int_t   *UrowindPtr;            /* A single vector */
     int_t   *UrowindPtr_host;       /* A single vector */
     int_t   *UnzvalPtr_host;
 
     double  *UnzvalVec;       /* A single vector */
-    int_t   *UnzvalPtr;      /* A single vector */
+    int_t   *UnzvalPtr;        /* A single vector */
+
     /*gpu pointers for easy block accesses */
     local_l_blk_info_t *local_l_blk_infoVec;
     int_t *local_l_blk_infoPtr;
@@ -109,7 +110,6 @@ typedef struct //LUstruct_gpu_
     int_t *xsup;
     gridinfo_t *grid;
 
-
     double ScatterMOPCounter;
     double ScatterMOPTimer;
     double GemmFLOPCounter;

diff --git a/SRC/dnrformat_loc3d.c b/SRC/dnrformat_loc3d.c
@@ -80,6 +80,8 @@ void dGatherNRformat_loc3d
 	nnz_counts_int = SUPERLU_MALLOC(grid3d->npdep * sizeof(int));
 	row_counts_int = SUPERLU_MALLOC(grid3d->npdep * sizeof(int));
 	b_counts_int = SUPERLU_MALLOC(grid3d->npdep * sizeof(int));
+
+	/* Gathered to layer 0. Other procs do not have these counts */
 	MPI_Gather(&A->nnz_loc, 1, mpi_int_t, nnz_counts,
 		   1, mpi_int_t, 0, grid3d->zscp.comm);
 	MPI_Gather(&A->m_loc, 1, mpi_int_t, row_counts,