Merge pull request #67 from ntrost57/develop

documentation update

clients/benchmarks/CMakeLists.txt

@@ -59,7 +59,13 @@ target_include_directories(rocsparse-bench
       $<BUILD_INTERFACE:${Boost_INCLUDE_DIRS}>
 )
 
-if(HIP_PLATFORM STREQUAL "hcc")
+if(HIP_PLATFORM STREQUAL "nvcc")
+  target_link_libraries(rocsparse-bench
+    PRIVATE
+      ${Boost_LIBRARIES}
+      roc::rocsparse
+  )
+else()
   target_link_libraries(rocsparse-bench
     PRIVATE
       ${Boost_LIBRARIES}
@@ -75,11 +81,3 @@ if(HIP_PLATFORM STREQUAL "hcc")
     )
   endforeach()
 endif()
-
-if(HIP_PLATFORM STREQUAL "nvcc")
-  target_link_libraries(rocsparse-bench
-    PRIVATE
-      ${Boost_LIBRARIES}
-      roc::rocsparse
-  )
-endif()

clients/samples/CMakeLists.txt

@@ -35,7 +35,12 @@ function(add_rocsparse_example EXAMPLE_SOURCE)
       $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include>
   )
 
-  if(HIP_PLATFORM STREQUAL "hcc")
+  if(HIP_PLATFORM STREQUAL "nvcc")
+    target_link_libraries(${EXAMPLE_TARGET}
+      PRIVATE
+        rocsparse
+    )
+  else()
     target_link_libraries(${EXAMPLE_TARGET}
       PRIVATE
         rocsparse
@@ -49,12 +54,6 @@ function(add_rocsparse_example EXAMPLE_SOURCE)
     endforeach()
   endif()
 
-  if(HIP_PLATFORM STREQUAL "nvcc")
-    target_link_libraries(${EXAMPLE_TARGET}
-      PRIVATE
-        rocsparse
-    )
-  endif()
   set_target_properties(${EXAMPLE_TARGET}
     PROPERTIES
       RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/example"

clients/tests/CMakeLists.txt

@@ -113,7 +113,13 @@ target_include_directories(rocsparse-test
       $<BUILD_INTERFACE:${HIP_INCLUDE_DIRS}>
 )
 
-if(HIP_PLATFORM STREQUAL "hcc")
+if(HIP_PLATFORM STREQUAL "nvcc")
+  target_link_libraries(rocsparse-test
+    PRIVATE
+      roc::rocsparse
+      ${GTEST_BOTH_LIBRARIES}
+  )
+else()
   target_link_libraries(rocsparse-test
     PRIVATE
       roc::rocsparse
@@ -130,12 +136,4 @@ if(HIP_PLATFORM STREQUAL "hcc")
   endforeach()
 endif()
 
-if(HIP_PLATFORM STREQUAL "nvcc")
-  target_link_libraries(rocsparse-test
-    PRIVATE
-      roc::rocsparse
-      ${GTEST_BOTH_LIBRARIES}
-  )
-endif()
-
 add_test(rocsparse-test rocsparse-test)

cmake/Dependencies.cmake

@@ -30,15 +30,7 @@ find_package(Git REQUIRED)
 include(cmake/DownloadProject/DownloadProject.cmake)
 
 # HIP configuration
-if(HIP_PLATFORM STREQUAL "hcc")
-  # Workaround until hcc & hip cmake modules fixes symlink logic in their config files.
-  # (Thanks to rocBLAS devs for finding workaround for this problem!)
-  list(APPEND CMAKE_PREFIX_PATH /opt/rocm/hcc /opt/rocm/hip)
-  # Ignore hcc warning: argument unused during compilation: '-isystem /opt/rocm/hip/include'
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wno-unused-command-line-argument")
-  find_package(hcc REQUIRED CONFIG PATHS /opt/rocm)
-  find_package(HIP REQUIRED CONFIG PATHS /opt/rocm)
-elseif(HIP_PLATFORM STREQUAL "nvcc")
+if(HIP_PLATFORM STREQUAL "nvcc")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xcompiler -Wall")
   set(CMAKE_C_COMPILE_OPTIONS_PIC "-Xcompiler ${CMAKE_C_COMPILE_OPTIONS_PIC}" )
   set(CMAKE_CXX_COMPILE_OPTIONS_PIC "-Xcompiler ${CMAKE_CXX_COMPILE_OPTIONS_PIC}" )
@@ -55,45 +47,53 @@ elseif(HIP_PLATFORM STREQUAL "nvcc")
   set(CMAKE_C_COMPILE_OPTIONS_VISIBILITY_INLINES_HIDDEN "-Xcompiler ${CMAKE_C_COMPILE_OPTIONS_VISIBILITY_INLINES_HIDDEN}" )
   set(CMAKE_CXX_COMPILE_OPTIONS_VISIBILITY_INLINES_HIDDEN "-Xcompiler ${CMAKE_CXX_COMPILE_OPTIONS_VISIBILITY_INLINES_HIDDEN}" )
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -arch sm_35")
+else()
+  # Workaround until hcc & hip cmake modules fixes symlink logic in their config files.
+  # (Thanks to rocBLAS devs for finding workaround for this problem!)
+  list(APPEND CMAKE_PREFIX_PATH /opt/rocm/hcc /opt/rocm/hip)
+  # Ignore hcc warning: argument unused during compilation: '-isystem /opt/rocm/hip/include'
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wno-unused-command-line-argument")
+  find_package(hcc REQUIRED CONFIG PATHS /opt/rocm)
+  find_package(HIP REQUIRED CONFIG PATHS /opt/rocm)
 endif()
 
 # rocPRIM package
-if(HIP_PLATFORM STREQUAL "hcc")
-  find_package(ROCPRIM QUIET CONFIG PATHS /opt/rocm)
+if(HIP_PLATFORM STREQUAL "nvcc")
   find_package(HIPCUB QUIET CONFIG PATHS /opt/rocm)
-  if(NOT ROCPRIM_FOUND)
+  if(NOT HIPCUB_FOUND)
     set(ROCPRIM_ROOT ${CMAKE_CURRENT_BINARY_DIR}/rocPRIM CACHE PATH "")
     message(STATUS "Downloading rocPRIM.")
     download_project(PROJ    rocPRIM
          GIT_REPOSITORY      https://github.com/ROCmSoftwarePlatform/rocPRIM.git
          GIT_TAG             master
          INSTALL_DIR         ${ROCPRIM_ROOT}
-         CMAKE_ARGS          -DCMAKE_BUILD_TYPE=RELEASE -DBUILD_TEST=OFF -DCMAKE_INSTALL_PREFIX=<INSTALL_DIR> -DCMAKE_CXX_COMPILER=/opt/rocm/bin/hcc
+         CMAKE_ARGS          -DCMAKE_BUILD_TYPE=RELEASE -DBUILD_TEST=OFF -DCMAKE_INSTALL_PREFIX=<INSTALL_DIR> -DCMAKE_CXX_COMPILER=${HIP_HIPCC_EXECUTABLE}
          LOG_DOWNLOAD        TRUE
          LOG_CONFIGURE       TRUE
          LOG_INSTALL         TRUE
          BUILD_PROJECT       TRUE
          UPDATE_DISCONNECT   TRUE
     )
-  find_package(ROCPRIM REQUIRED CONFIG PATHS ${ROCPRIM_ROOT})
   find_package(HIPCUB REQUIRED CONFIG PATHS ${ROCPRIM_ROOT})
   endif()
-elseif(HIP_PLATFORM STREQUAL "nvcc")
+else()
+  find_package(ROCPRIM QUIET CONFIG PATHS /opt/rocm)
   find_package(HIPCUB QUIET CONFIG PATHS /opt/rocm)
-  if(NOT HIPCUB_FOUND)
+  if(NOT ROCPRIM_FOUND)
     set(ROCPRIM_ROOT ${CMAKE_CURRENT_BINARY_DIR}/rocPRIM CACHE PATH "")
     message(STATUS "Downloading rocPRIM.")
     download_project(PROJ    rocPRIM
          GIT_REPOSITORY      https://github.com/ROCmSoftwarePlatform/rocPRIM.git
          GIT_TAG             master
          INSTALL_DIR         ${ROCPRIM_ROOT}
-         CMAKE_ARGS          -DCMAKE_BUILD_TYPE=RELEASE -DBUILD_TEST=OFF -DCMAKE_INSTALL_PREFIX=<INSTALL_DIR> -DCMAKE_CXX_COMPILER=${HIP_HIPCC_EXECUTABLE}
+         CMAKE_ARGS          -DCMAKE_BUILD_TYPE=RELEASE -DBUILD_TEST=OFF -DCMAKE_INSTALL_PREFIX=<INSTALL_DIR> -DCMAKE_CXX_COMPILER=/opt/rocm/bin/hcc
          LOG_DOWNLOAD        TRUE
          LOG_CONFIGURE       TRUE
          LOG_INSTALL         TRUE
          BUILD_PROJECT       TRUE
          UPDATE_DISCONNECT   TRUE
     )
+  find_package(ROCPRIM REQUIRED CONFIG PATHS ${ROCPRIM_ROOT})
   find_package(HIPCUB REQUIRED CONFIG PATHS ${ROCPRIM_ROOT})
   endif()
 endif()

cmake/SetToolchain.cmake

@@ -45,7 +45,7 @@ if(HIP_PLATFORM STREQUAL "nvcc" OR HIP_COMPILER STREQUAL "clang")
   endif()
   mark_as_advanced(HIP_HIPCC_EXECUTABLE)
   set(CMAKE_CXX_COMPILER ${HIP_HIPCC_EXECUTABLE})
-elseif(HIP_PLATFORM STREQUAL "hcc")
+else()
   # Find HCC executable
   find_program(
       HIP_HCC_EXECUTABLE
@@ -65,6 +65,4 @@ elseif(HIP_PLATFORM STREQUAL "hcc")
   endif()
   mark_as_advanced(HIP_HCC_EXECUTABLE)
   set(CMAKE_CXX_COMPILER ${HIP_HCC_EXECUTABLE})
-else()
-  message(FATAL_ERROR "HIP_PLATFORM must be 'hcc/nvcc' (AMD ROCm platform).")
 endif()

docs/source/library.rst

@@ -766,7 +766,7 @@ rocsparse_csrsv_solve()
   :outline:
 .. doxygenfunction:: rocsparse_dcsrsv_solve
 
-rocsparse_csrsv_analysis_clear()
+rocsparse_csrsv_clear()
 ********************************
 
 .. doxygenfunction:: rocsparse_csrsv_clear
@@ -822,7 +822,7 @@ rocsparse_csrilu0()
   :outline:
 .. doxygenfunction:: rocsparse_dcsrilu0
 
-rocsparse_csrilu0_analysis_clear()
+rocsparse_csrilu0_clear()
 **********************************
 
 .. doxygenfunction:: rocsparse_csrilu0_clear

library/CMakeLists.txt

@@ -50,7 +50,7 @@ add_library(rocsparse ${rocsparse_source} ${rocsparse_headers_public})
 add_library(roc::rocsparse ALIAS rocsparse)
 
 # Target link libraries
-if(HIP_PLATFORM STREQUAL "hcc")
+if(NOT HIP_PLATFORM STREQUAL "nvcc")
   target_link_libraries(rocsparse PRIVATE hip::hip_hcc hip::hip_device hcc::hccshared)
   foreach(target ${AMDGPU_TARGETS})
     target_link_libraries(rocsparse PRIVATE --amdgpu-target=${target})
@@ -68,15 +68,15 @@ target_include_directories(rocsparse
 )
 
 # Target properties
-if(HIP_PLATFORM STREQUAL "hcc")
+if(NOT HIP_PLATFORM STREQUAL "nvcc")
   set_target_properties(rocsparse PROPERTIES VERSION ${rocsparse_VERSION} SOVERSION ${rocsparse_SOVERSION})
   set_target_properties(rocsparse PROPERTIES CXX_VISIBILITY_PRESET "hidden" VISIBILITY_INLINES_HIDDEN ON)
 endif()
 set_target_properties(rocsparse PROPERTIES RUNTIME_OUTPUT_DIRECTORY "${PROJECT_BINARY_DIR}/staging")
 set_target_properties(rocsparse PROPERTIES DEBUG_POSTFIX "-d")
 
 # Target definitions
-if(HIP_PLATFORM STREQUAL "hcc")
+if(NOT HIP_PLATFORM STREQUAL "nvcc")
   target_compile_definitions(rocsparse PRIVATE ROCPRIM_HIP_API=1)
 endif()
 
@@ -95,7 +95,7 @@ rocm_install_targets(TARGETS rocsparse
 # Export targets
 rocm_export_targets(TARGETS roc::rocsparse
                     PREFIX rocsparse
-                    DEPENDS PACKAGE HIP
+                    DEPENDS PACKAGE hip
                     NAMESPACE roc::
 )
 
@@ -114,12 +114,10 @@ endif()
 set(CPACK_RPM_EXCLUDE_FROM_AUTO_FILELIST_ADDITION "\${CPACK_PACKAGING_INSTALL_PREFIX}" "\${CPACK_PACKAGING_INSTALL_PREFIX}/include")
 
 # Package name differs for CUDA backend
-if(HIP_PLATFORM STREQUAL "hcc")
-  set(package_name rocsparse)
-endif()
-
 if(HIP_PLATFORM STREQUAL "nvcc")
   set(package_name rocsparse-alt)
+else()
+  set(package_name rocsparse)
 endif()
 
 set(ROCSPARSE_CONFIG_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR}" CACHE PATH "Path placed into ldconfig file")

library/include/rocsparse-functions.h

@@ -987,6 +987,9 @@ rocsparse_status rocsparse_zcsrmv(rocsparse_handle handle,
  *  \p position can be in host or device memory. If no zero pivot has been found,
  *  \p position is set to -1 and \ref rocsparse_status_success is returned instead.
  *
+ *  \note \p rocsparse_csrsv_zero_pivot is a blocking function. It might influence
+ *  performance negatively.
+ *
  *  @param[in]
  *  handle      handle to the rocsparse library context queue.
  *  @param[in]
@@ -1918,6 +1921,9 @@ rocsparse_status rocsparse_zcsrmm(rocsparse_handle handle,
  *  \p position can be in host or device memory. If no zero pivot has been found,
  *  \p position is set to -1 and \ref rocsparse_status_success is returned instead.
  *
+ *  \note \p rocsparse_csrilu0_zero_pivot is a blocking function. It might influence
+ *  performance negatively.
+ *
  *  @param[in]
  *  handle      handle to the rocsparse library context queue.
  *  @param[in]

library/src/handle.cpp

@@ -223,10 +223,16 @@ rocsparse_status rocsparse_destroy_csrtr_info(rocsparse_csrtr_info info)
     }
 
     // Clean up
-    if(info->row_map != nullptr)
+    if(info->d_row_map != nullptr)
     {
-        RETURN_IF_HIP_ERROR(hipFree(info->row_map));
-        info->row_map = nullptr;
+        RETURN_IF_HIP_ERROR(hipFree(info->d_row_map));
+        info->d_row_map = nullptr;
+    }
+
+    if(info->h_row_map != nullptr)
+    {
+        RETURN_IF_HIP_ERROR(hipHostFree(info->h_row_map));
+        info->h_row_map = nullptr;
     }
 
     if(info->csr_diag_ind != nullptr)

library/src/include/handle.h

@@ -194,7 +194,9 @@ struct _rocsparse_csrtr_info
     // host array to hold number of rows per level
     std::vector<rocsparse_int> rows_per_level;
     // device array to hold row permutation
-    rocsparse_int* row_map = nullptr;
+    rocsparse_int* d_row_map = nullptr;
+    // host array to hold row permutation
+    rocsparse_int* h_row_map = nullptr;
     // device array to hold pointer to diagonal entry
     rocsparse_int* csr_diag_ind = nullptr;
     // device pointer to hold zero pivot