Merge branch 'master' of https://github.com/xiaoyeli/superlu_dist

xiaoyeli · Jan 15, 2022 · becc2d9 · becc2d9
2 parents f74d065 + 363203d
commit becc2d9
Show file tree

Hide file tree

Showing 107 changed files with 10,887 additions and 2,462 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -42,6 +42,7 @@ option(TPL_COMBBLAS_LIBRARIES "List of absolute paths to CombBLAS link libraries
 option(TPL_COMBBLAS_INCLUDE_DIRS "List of absolute paths to CombBLAS include directories [].")
 #-- CUDA
 option(TPL_ENABLE_CUDALIB   "Enable the CUDA libraries" OFF)
+option(TPL_ENABLE_HIPLIB   "Enable the HIP libraries" OFF)
 
 ######################################################################
 #
@@ -164,6 +165,7 @@ if(MPI_C_FOUND)
     set(CMAKE_C_FLAGS "${MPI_C_COMPILE_FLAGS} ${CMAKE_C_FLAGS}")
     set(CMAKE_CXX_FLAGS "${MPI_CXX_COMPILE_FLAGS} ${CMAKE_CXX_FLAGS}")
     set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${MPI_C_LINK_FLAGS}" )
+#    message("MPI_C_INCLUDE_DIRS: ${MPI_C_INCLUDE_DIRS}")
 endif()
 if (XSDK_ENABLE_Fortran)
   if(MPI_Fortran_FOUND)
@@ -192,21 +194,18 @@ if (TPL_ENABLE_CUDALIB)   ## want to use cuda
 #     find_package(CUB REQUIRED)
 
     find_package(CUDAToolkit REQUIRED)
+    if(CUDAToolkit_FOUND)
+      target_link_libraries(superlu PUBLIC CUDA::cudart CUDA::cusolver CUDA::cublas)
+    endif()
     message("-- CUDAToolkit_LIBRARY_ROOT='${CUDAToolkit_LIBRARY_ROOT}'")
     if (NOT "${CUDAToolkit_LIBRARY_ROOT}" STREQUAL "")
-    set(CUDA_LIBRARIES "${CUDAToolkit_LIBRARY_ROOT}/lib64/libcudart.so")
-#    set(CUDA_CUBLAS_LIBRARIES "${CUDAToolkit_LIBRARY_ROOT}/lib64/libcublas.so")
-
-if(CUDAToolkit_FOUND)
-  target_link_libraries(superlu PUBLIC CUDA::cudart CUDA::cusolver CUDA::cublas)
-  endif()
-
+       set(CUDA_LIBRARIES "${CUDAToolkit_LIBRARY_ROOT}/lib64/libcudart.so")
+       # expose the following to make.inc, but location may not be reliable
+      set(CUDA_CUBLAS_LIBRARIES "${CUDAToolkit_LIBRARY_ROOT}/lib64/libcublas.so")
     else()
-    message("-- CUDAToolkit_LIBRARY_ROOT empty, not setting CUDA_LIBRARIES")
+	message("-- CUDAToolkit_LIBRARY_ROOT empty, not setting CUDA_LIBRARIES")
     endif()
 
-
-
 #     # The following make.inc exporting does not work
 #     set(CUDA_LIB CUDA::cudart CUDA::cublas CUDA::cusolver)
 # # fix up CUDA library names    
@@ -237,6 +236,75 @@ if (enable_openmp)
     message("-- CMAKE_EXE_LINKER_FLAGS='${CMAKE_EXE_LINKER_FLAGS}'")
   endif()
 endif()
+
+
+
+include(CheckLanguage)
+
+#--------------------- CUDA libraries ---------------------
+if (TPL_ENABLE_CUDALIB)   ## want to use cuda
+  check_language(CUDA)
+  if(CMAKE_CUDA_COMPILER)
+    message("-- Enabled support for CUDA.")
+    enable_language(CUDA)
+    find_package(CUDAToolkit REQUIRED)
+
+    # make.inc exporting not yet tested
+    set(CUDA_LIB CUDA::cudart CUDA::cublas CUDA::cusolver)
+    string (REPLACE ";" " " CUDA_LIB_STR "${CUDA_LIB}")
+    set(CUDA_LIB_EXPORT ${CUDA_LIB_STR}) 
+    set(HAVE_CUDA TRUE)
+    # set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DHAVE_CUDA")
+    # set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -DHAVE_CUDA")
+  else()
+    message("-- CUDA libraries not found.")
+  endif()
+endif()
+
+
+
+#--------------------- HIP libraries ---------------------
+if (TPL_ENABLE_HIPLIB)   ## want to use hip
+  find_package(HIP MODULE)
+  if(HIP_FOUND)
+    message("-- Enabled support for HIP.")
+    find_package(hipblas REQUIRED)
+    find_package(rocsolver REQUIRED)
+    find_package(rocblas REQUIRED)
+    set(HIP_LIB roc::hipblas roc::rocblas roc::rocsolver)
+    message("HIP_LIB ${HIP_LIB}")
+
+    # make.inc exporting not yet tested
+    string (REPLACE ";" " " HIP_LIB_STR "${HIP_LIB}")
+    set(HIP_LIB_EXPORT ${HIP_LIB_STR}) 
+    set(HAVE_HIP TRUE)
+    # set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DHAVE_HIP")
+    # set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DHAVE_HIP")    
+    # set(HIP_HIPCC_FLAGS "${HIP_HIPCC_FLAGS} -DHAVE_HIP")    
+    if (BUILD_SHARED_LIBS)
+    set(HIP_HIPCC_FLAGS "${HIP_HIPCC_FLAGS} -fPIC")    
+    endif()
+
+
+
+
+    # if ("$ENV{HIP_PLATFORM}" STREQUAL "hcc")
+    #   set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DHAVE_HIP -D__HIP_PLATFORM_HCC__")
+    #   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DHAVE_HIP -D__HIP_PLATFORM_HCC__")
+    # elseif ("$ENV{HIP_PLATFORM}" STREQUAL "nvcc")
+    #   set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DHAVE_HIP -D__HIP_PLATFORM_NVCC__")
+    #   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DHAVE_HIP -D__HIP_PLATFORM_NVCC__")
+    # endif()
+
+
+  else()
+    message("-- HIP libraries not found.")
+  endif()
+endif()
+
+
+
+
 #--------------------- BLAS ---------------------
 if(NOT TPL_ENABLE_INTERNAL_BLASLIB)
 #  set(TPL_BLAS_LIBRARIES "" CACHE FILEPATH
@@ -313,6 +381,8 @@ if (TPL_ENABLE_PARMETISLIB)   ## want to use parmetis
     endif()
     set(CMAKE_C_FLAGS "-I${dir} ${CMAKE_C_FLAGS}")
   endforeach()
+
+
 
   message("-- Enabled support for PARMETIS.")
   set(PARMETIS_FOUND TRUE)
@@ -343,7 +413,7 @@ FOREACH(_lib ${_c_libs})
   set(EXTRA_LIB "-l${_lib} ${EXTRA_LIB}")
 ENDFOREACH()
 string (REPLACE ";" " " EXTRA_LIB_STR "${EXTRA_LIB}")
-set(EXTRA_LIB_EXPORT ${EXTRA_LIB_STR})
+set(EXTRA_LIB_EXPORT ${EXTRA_LIB_STR}) 
 message("-- EXTRA_LIB_EXPORT='${EXTRA_LIB_EXPORT}'")
 
 #---------------------- Additional Fortran linker library ---------

diff --git a/EXAMPLE/pddrive.c b/EXAMPLE/pddrive.c
@@ -197,6 +197,13 @@ int main(int argc, char *argv[])
 	options.DiagInv           = NO;
      */
     set_default_options_dist(&options);
+	// options.IterRefine = NOREFINE;
+	// options.DiagInv = YES;
+    // options.ReplaceTinyPivot  = YES;
+
+	// options.Equil = NO; 
+	// options.ColPerm = NATURAL;
+	// options.RowPerm = NOROWPERM;  						  
 #if 0
     options.RowPerm           = LargeDiag_HWPM;
     options.RowPerm = NOROWPERM;
@@ -206,6 +213,10 @@ int main(int argc, char *argv[])
     options.ReplaceTinyPivot = YES;
 #endif
 
+    // options.ParSymbFact       = YES;
+    // options.ColPerm           = PARMETIS;
+
+
     if (!iam) {
 	print_sp_ienv_dist(&options);
 	print_options_dist(&options);
@@ -252,7 +263,7 @@ int main(int argc, char *argv[])
     SUPERLU_FREE(b);
     SUPERLU_FREE(xtrue);
     SUPERLU_FREE(berr);
-    fclose(fp);
+    // fclose(fp);
 
     /* ------------------------------------------------------------
        RELEASE THE SUPERLU PROCESS GRID.

diff --git a/EXAMPLE/pddrive3d.c b/EXAMPLE/pddrive3d.c
@@ -310,6 +310,7 @@ main (int argc, char *argv[])
      */
     set_default_options_dist (&options);
 #if 0
+    options.ReplaceTinyPivot = YES;
     options.RowPerm = NOROWPERM;
     options.IterRefine = NOREFINE;
     options.ColPerm = NATURAL;

diff --git a/SRC/CMakeLists.txt b/SRC/CMakeLists.txt
@@ -1,6 +1,5 @@
 set(headers
     superlu_FCnames.h
-    cublas_utils.h
     dcomplex.h
     machines.h
     psymbfact.h
@@ -9,10 +8,8 @@ set(headers
     supermatrix.h
     util_dist.h
     colamd.h
-    TreeBcast_slu.hpp
-    TreeReduce_slu.hpp	
-    TreeBcast_slu_impl.hpp
-    TreeReduce_slu_impl.hpp	
+    gpu_api_utils.h
+    gpu_wrapper.h
     ${CMAKE_CURRENT_BINARY_DIR}/superlu_dist_config.h
     ${PROJECT_SOURCE_DIR}/SRC/superlu_FortranCInterface.h
 )
@@ -31,7 +28,7 @@ set(sources
   comm.c
   memory.c
   util.c
-  cublas_utils.c
+  gpu_api_utils.c
   superlu_grid.c
   pxerr_dist.c
   superlu_timer.c
@@ -45,17 +42,21 @@ set(sources
   dmach_dist.c
   colamd.c
   superlu_dist_version.c
-  TreeInterface.cpp
+  comm_tree.c
   superlu_grid3d.c    ## 3D code
   supernodal_etree.c
   supernodalForest.c
   trfAux.c 
   communication_aux.c
   treeFactorization.c
-  sec_structs.c
+  sec_structs.c  
 )
 if (HAVE_CUDA)
-  list(APPEND sources cublas_utils.c superlu_gpu_utils.cu)
+  list(APPEND sources superlu_gpu_utils.cu)
+endif()
+
+if (HAVE_HIP)
+  list(APPEND sources superlu_gpu_utils.hip.cpp)
 endif()
 
 if (MSVC)
@@ -67,6 +68,13 @@ set_source_files_properties(superlu_timer.c PROPERTIES COMPILE_FLAGS -O0)
 if(enable_double)
   list(APPEND headers superlu_ddefs.h dlustruct_gpu.h)
 
+if(TPL_ENABLE_CUDALIB)
+  list(APPEND sources pdgstrs_lsum_cuda.cu)
+endif()
+# if(TPL_ENABLE_HIPLIB)
+#   list(APPEND sources pdgstrs_lsum_cuda.hip.cpp)
+# endif()    	
+
   list(APPEND sources
     dlangs_dist.c
     dgsequ_dist.c
@@ -120,6 +128,7 @@ if(enable_double)
 if (HAVE_CUDA)
   list(APPEND sources dsuperlu_gpu.cu)
 endif()
+
 if (HAVE_COMBBLAS)
   list(APPEND sources d_c2cpp_GetHWPM.cpp dHWPM_CombBLAS.hpp)
 endif()
@@ -248,22 +257,49 @@ endif()
 if (HAVE_COMBBLAS)
     list(APPEND sources z_c2cpp_GetHWPM.cpp zHWPM_CombBLAS.hpp)
 endif()
-
 endif() ######### enable compex16
 
+if (TPL_ENABLE_HIPLIB)
+  file(GLOB MyFiles *.hip.cpp)
+  set_source_files_properties(
+    ${MyFiles}
+    PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
 
-add_library(superlu_dist ${sources} ${HEADERS})
+  set(hipsources
+    pdgstrs_lsum_cuda.hip.cpp
+    superlu_gpu_utils.hip.cpp
+    dsuperlu_gpu.hip.cpp
+  )
+  if(enable_single)
+  list(APPEND hipsources ssuperlu_gpu.hip.cpp)
+  endif()
+  if(enable_complex16)
+  list(APPEND hipsources zsuperlu_gpu.hip.cpp)
+  endif()
+  hip_add_library(superlu_dist ${hipsources})
+  if (BUILD_SHARED_LIBS AND BUILD_STATIC_LIBS)
+    hip_add_library(superlu_dist-static STATIC  ${hipsources})
+  endif()
+
+else()
+  add_library(superlu_dist "")
+  if (BUILD_SHARED_LIBS AND BUILD_STATIC_LIBS)
+    add_library(superlu_dist-static STATIC "")
+  endif()
+endif()
+
+
+target_sources(superlu_dist PRIVATE ${sources} ${HEADERS})
 set(targets superlu_dist)
 
 if (BUILD_SHARED_LIBS AND BUILD_STATIC_LIBS)
   # build both shared and static libs
-  add_library(superlu_dist-static STATIC ${sources} ${HEADERS})
+  target_sources(superlu_dist-static PRIVATE ${sources} ${HEADERS})
   list(APPEND targets superlu_dist-static)
 endif()
 
 set(superlu_dist_libs ${MPI_C_LIBRARIES} ${MPI_CXX_LIBRARIES} ${BLAS_LIB} ${LAPACK_LIB}
-    ${PARMETIS_LIB} ${COMBBLAS_LIB} ${CUDA_LIBRARIES} ${CUDA_CUBLAS_LIBRARIES})
-
+    ${PARMETIS_LIB} ${COMBBLAS_LIB} ${CUDA_LIB} ${HIP_LIB}) 
 if (NOT MSVC)
   list(APPEND superlu_dist_libs m)
 endif ()
@@ -272,15 +308,15 @@ foreach(target ${targets})
     target_link_libraries(${target} ${superlu_dist_libs})
     if (HAVE_COMBBLAS)
         set_target_properties(${target} PROPERTIES
-			      CUDA_SEPARABLE_COMPILATION ON
-                              OUTPUT_NAME superlu_dist
+                              CUDA_SEPARABLE_COMPILATION ON
+							  OUTPUT_NAME superlu_dist
                               VERSION ${PROJECT_VERSION}
                               SOVERSION ${VERSION_MAJOR}
 			      LINKER_LANGUAGE CXX
 			      )
     else()
         set_target_properties(${target} PROPERTIES
-			      CUDA_SEPARABLE_COMPILATION ON
+		                      CUDA_SEPARABLE_COMPILATION ON
                               OUTPUT_NAME superlu_dist
                               VERSION ${PROJECT_VERSION}
                               SOVERSION ${VERSION_MAJOR}

diff --git a/SRC/Makefile b/SRC/Makefile
@@ -44,7 +44,6 @@ ALLAUX 	= sp_ienv.o etree.o sp_colorder.o get_perm_c.o \
 # Following are from 3D code
 ALLAUX += superlu_grid3d.o supernodal_etree.o supernodalForest.o \
 	trfAux.o communication_aux.o treeFactorization.o sec_structs.o
-
 #
 # Routines literally taken from SuperLU, but renamed with suffix _dist
 #
@@ -85,7 +84,7 @@ ZPLUSRC += pzgssvx3d.o pzgstrf3d.o ztreeFactorization.o zscatter3d.o \
 	znrformat_loc3d.o ztreeFactorizationGPU.o ##$(FACT3D)
 
 ifeq ($(HAVE_CUDA),TRUE)
-ALLAUX += cublas_utils.o superlu_gpu_utils.o
+ALLAUX += cublas_utils.o superlu_gpu_utils.o pdgstrs_lsum_cuda.o
 DPLUSRC += dsuperlu_gpu.o
 ZPLUSRC += zsuperlu_gpu.o
 endif
@@ -136,19 +135,19 @@ complex16: config_h $(ZSLUSRC) $(ZPLUSRC) $(ALLAUX)
 	$(RANLIB) $(DSUPERLULIB)
 
 pdgstrf.o: dscatter.c dlook_ahead_update.c dSchCompUdt-2Ddynamic.c pdgstrf.c
-	$(CC) $(CFLAGS) $(CDEFS) $(BLASDEF) -I$(INCLUDEDIR) -c pdgstrf.c $(VERBOSE)
+	$(CC) $(CFLAGS) $(CDEFS) $(BLASDEF) $(INCLUDEDIR) -c pdgstrf.c $(VERBOSE)
 
 pzgstrf.o: zscatter.c zlook_ahead_update.c zSchCompUdt-2Ddynamic.c pzgstrf.c
-	$(CC) $(CFLAGS) $(CDEFS) $(BLASDEF) -I$(INCLUDEDIR) -c pzgstrf.c $(VERBOSE)
+	$(CC) $(CFLAGS) $(CDEFS) $(BLASDEF) $(INCLUDEDIR) -c pzgstrf.c $(VERBOSE)
 
 .c.o:
-	$(CC) $(CFLAGS) $(CDEFS) $(BLASDEF) -I$(INCLUDEDIR) -c $< $(VERBOSE)
+	$(CC) $(CFLAGS) $(CDEFS) $(BLASDEF) $(INCLUDEDIR) -c $< $(VERBOSE)
 
 .cu.o:
 	$(NVCC) $(NVCCFLAGS) $(CDEFS) $(BLASDEF) -I$(INCLUDEDIR) -c $< $(VERBOSE)
 
 .cpp.o:
-	$(CXX) $(CXXFLAGS) $(CPPFLAGS) $(CDEFS) $(BLASDEF) -I$(INCLUDEDIR) -c $< $(VERBOSE)
+	$(CXX) $(CXXFLAGS) $(CPPFLAGS) $(CDEFS) $(BLASDEF) $(INCLUDEDIR) -c $< $(VERBOSE)
 
 .f.o:
 	$(FORTRAN) $(FFLAGS) -c $< $(VERBOSE)