diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index efdc38131..74cad0049 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -8,36 +8,21 @@ variables:
   BUILD_ROOT: ${CI_PROJECT_DIR}
 
 stages:
-  - l_build
-  - q_allocate_resources
-  - q_build
-  - q_release_resources
+  - allocate
+  - build
+  - release
 
 ####
 # Template
-.srun_build_script:
+.build_script:
   script:
     #Use pre-existing allocation if any
-    - JOBID=$(squeue -h --name=${PROJECT_ALLOC_NAME} --format=%A)
+    - JOBID=$(if [[ "$SYS_TYPE" == "toss_3_x86_64_ib" ]]; then squeue -h --name=${PROJECT_ALLOC_NAME} --format=%A; fi)
     - ASSIGN_ID=$(if [[ -n "${JOBID}" ]]; then echo "--jobid=${JOBID}"; fi)
-    - EXEC_PREFIX="srun -p pdebug ${ASSIGN_ID} -t 10 -N 1 --mpibind=off"
    #BUILD + TEST
     - echo -e "section_start:$(date +%s):build_and_test\r\e[0K
       Build and test ${CI_PROJECT_NAME}"
-    - ${EXEC_PREFIX} .gitlab/build_and_test.sh
-    - echo -e "section_end:$(date +%s):build_and_test\r\e[0K"
-  artifacts:
-    reports:
-      junit: build/junit.xml
-
-
-.build_blueos_3_ppc64le_ib_p9_script:
-  script:
-    - EXEC_PREFIX="lalloc 1 -W 10 -q pdebug"
-    #BUILD + TEST
-    - echo -e "section_start:$(date +%s):build_and_test\r\e[0K
-      Build and test ${CI_PROJECT_NAME}"
-    - ${EXEC_PREFIX} .gitlab/build_and_test.sh
+    - ${ALLOC_COMMAND} .gitlab/build_and_test.sh
     - echo -e "section_end:$(date +%s):build_and_test\r\e[0K"
   artifacts:
     reports:
diff --git a/.gitlab/build_lassen.yml b/.gitlab/build_lassen.yml
index 0cd35459e..3096cc79a 100644
--- a/.gitlab/build_lassen.yml
+++ b/.gitlab/build_lassen.yml
@@ -20,33 +20,35 @@
 ####
 # Template
 .build_on_lassen:
-  stage: l_build
-  extends: [.build_blueos_3_ppc64le_ib_p9_script, .on_lassen]
+  stage: build
+  variables:
+    ALLOC_COMMAND: "lalloc 1 -W 10 -q pdebug"
+  extends: [.build_script, .on_lassen]
   needs: []
 
 ####
 # Build jobs
-clang_upstream_link_with_nvcc (lassen):
+lassen-clang_upstream_link_with_nvcc:
   variables:
     HOST_CONFIG: "clang@upstream_link_with_nvcc.cmake"
   extends: [.build_on_lassen, .with_cuda]
 
-clang_upstream_nvcc_c++17 (lassen):
+lassen-clang_upstream_nvcc_c++17:
   variables:
     HOST_CONFIG: "clang@upstream_nvcc_c++17.cmake"
   extends: [.build_on_lassen, .with_cuda]
 
-clang_upstream_nvcc_c++17_no_separable (lassen):
+lassen-clang_upstream_nvcc_c++17_no_separable:
   variables:
     HOST_CONFIG: "clang@upstream_nvcc_c++17_no_separable.cmake"
   extends: [.build_on_lassen, .with_cuda]
 
-clang_upstream_nvcc_xlf (lassen):
+lassen-clang_upstream_nvcc_xlf:
   variables:
     HOST_CONFIG: "clang@upstream_nvcc_xlf.cmake"
   extends: [.build_on_lassen, .with_cuda]
 
-pgi_20.4_nvcc (lassen):
+lassen-pgi_20.4_nvcc:
   variables:
     HOST_CONFIG: "pgi@20.4_nvcc.cmake"
   extends: [.build_on_lassen, .with_cuda]
diff --git a/.gitlab/build_quartz.yml b/.gitlab/build_quartz.yml
index f9cc757d4..e93d1610e 100644
--- a/.gitlab/build_quartz.yml
+++ b/.gitlab/build_quartz.yml
@@ -13,23 +13,23 @@
 
 ####
 # In pre-build phase, allocate a node for builds
-allocate_resources_build_quartz:
+quartz_allocate:
   variables:
     GIT_STRATEGY: none
   extends: [.on_quartz]
-  stage: q_allocate_resources
+  stage: allocate
   script:
-    - salloc -p pdebug -N 1 -c 36 -t 30 --no-shell --job-name=${PROJECT_ALLOC_NAME} --mpibind=off
+    - salloc -p pdebug -N 1 -t 10 --no-shell --job-name=${PROJECT_ALLOC_NAME} --mpibind=off
   needs: []
 
 ####
 # In post-build phase, deallocate resources
 # Note : make sure this is run even on build phase failure
-release_resources_build_quartz:
+quartz_release:
   variables:
     GIT_STRATEGY: none
   extends: [.on_quartz]
-  stage: q_release_resources
+  stage: release
   script:
     - export JOBID=$(squeue -h --name=${PROJECT_ALLOC_NAME} --format=%A)
     - if [[ -n "${JOBID}" ]]; then scancel ${JOBID}; fi
@@ -37,27 +37,29 @@ release_resources_build_quartz:
 ####
 # Template
 .build_on_quartz:
-  stage: q_build
-  extends: [.srun_build_script, .on_quartz]
+  stage: build
+  variables:
+    ALLOC_COMMAND: "srun -p pdebug -t 30 -N 1 ${ASSIGN_ID}"
+  extends: [.build_script, .on_quartz]
 
 ####
 # Build jobs
-clang_4_0_0_libcxx (quartz):
+quartz-clang_4_0_0_libcxx:
   variables:
     HOST_CONFIG: "clang@4.0.0-libcxx.cmake"
   extends: [.build_on_quartz]
 
-clang_6_0_0_static_analysis (quartz):
+quartz-clang_6_0_0_static_analysis:
   variables:
     HOST_CONFIG: "clang@6.0.0-static-analysis.cmake"
   extends: [.build_on_quartz]
 
-gcc_8_3_1 (quartz):
+quartz-gcc_8_3_1:
   variables:
     HOST_CONFIG: "gcc@8.3.1.cmake"
   extends: [.build_on_quartz]
 
-pgi_20_1 (quartz):
+quartz-pgi_20_1:
   variables:
     HOST_CONFIG: "pgi@20.1.cmake"
   extends: [.build_on_quartz]
diff --git a/RELEASE-NOTES.md b/RELEASE-NOTES.md
index 14045ab33..7bb9847bb 100644
--- a/RELEASE-NOTES.md
+++ b/RELEASE-NOTES.md
@@ -9,9 +9,20 @@ The project release numbers follow [Semantic Versioning](http://semver.org/spec/
 
 ## [Unreleased] - Release date yyyy-mm-dd
 
+## [Version 0.4.1] - Release date 2021-07-20
+
+### Added
+- Added compilation of HIP with clang using ``ENABLE_CLANG_HIP`` and ``BLT_CLANG_HIP_ARCH``
+
+### Changed
+- XL: Use compiler flag `-std=c++14` instead of `-std=c++1y` when `BLT_CXX_STD` is set to `c++14`
+
+### Fixed
+- Simpified the clang-format version regex that was causing hangs on some version strings.
+
 ## [Version 0.4.0] - Release date 2021-04-09
 
-## Added
+### Added
 - Added variable ``BLT_CMAKE_IMPLICIT_LINK_DIRECTORIES_EXCLUDE`` for filtering
   link directories implicitly added by CMake. See the following example host-config:
   ``host-configs/llnl/blueos_3_ppc64le_ib_p9/clang@upstream_nvcc_c++17.cmake``
@@ -202,7 +213,8 @@ The project release numbers follow [Semantic Versioning](http://semver.org/spec/
 
 
 
-[Unreleased]:    https://github.com/LLNL/blt/compare/v0.4.0...develop
+[Unreleased]:    https://github.com/LLNL/blt/compare/v0.4.1...develop
+[Version 0.4.1]: https://github.com/LLNL/blt/compare/v0.4.0...v0.4.1
 [Version 0.4.0]: https://github.com/LLNL/blt/compare/v0.3.6...v0.4.0
 [Version 0.3.6]: https://github.com/LLNL/blt/compare/v0.3.5...v0.3.6
 [Version 0.3.5]: https://github.com/LLNL/blt/compare/v0.3.0...v0.3.5
diff --git a/SetupBLT.cmake b/SetupBLT.cmake
index 322bde168..6402d4dd1 100644
--- a/SetupBLT.cmake
+++ b/SetupBLT.cmake
@@ -4,7 +4,7 @@
 # SPDX-License-Identifier: (BSD-3-Clause)
 
 if (NOT BLT_LOADED)
-    set(BLT_VERSION "0.4.0" CACHE STRING "")
+    set(BLT_VERSION "0.4.1" CACHE STRING "")
     mark_as_advanced(BLT_VERSION)
     message(STATUS "BLT Version: ${BLT_VERSION}")
 
diff --git a/cmake/BLTMacros.cmake b/cmake/BLTMacros.cmake
index 5fd6be7e3..e41d12837 100644
--- a/cmake/BLTMacros.cmake
+++ b/cmake/BLTMacros.cmake
@@ -644,7 +644,7 @@ macro(blt_add_library)
             set(_lib_type "STATIC")
         endif()
 
-        if (ENABLE_HIP)
+        if (ENABLE_HIP AND NOT ENABLE_CLANG_HIP)
             blt_add_hip_library(NAME         ${arg_NAME}
                                 SOURCES      ${arg_SOURCES}
                                 HEADERS      ${arg_HEADERS}
@@ -775,7 +775,7 @@ macro(blt_add_executable)
         message(FATAL_ERROR "blt_add_executable(NAME ${arg_NAME} ...) given with no sources")
     endif()
 
-    if (ENABLE_HIP)
+    if (ENABLE_HIP AND NOT ENABLE_CLANG_HIP)
         blt_add_hip_executable(NAME         ${arg_NAME}
                                SOURCES      ${arg_SOURCES}
                                HEADERS      ${arg_HEADERS}
diff --git a/cmake/BLTOptions.cmake b/cmake/BLTOptions.cmake
index edd1e5168..fd9cef252 100644
--- a/cmake/BLTOptions.cmake
+++ b/cmake/BLTOptions.cmake
@@ -53,11 +53,17 @@ option(ENABLE_FORTRAN      "Enables Fortran compiler support" ${_fortran_already
 option(ENABLE_MPI          "Enables MPI support" OFF)
 option(ENABLE_OPENMP       "Enables OpenMP compiler support" OFF)
 option(ENABLE_CUDA         "Enable CUDA support" OFF)
-option(ENABLE_CLANG_CUDA   "Enable Clang's native CUDA support" OFF)
+cmake_dependent_option(ENABLE_CLANG_CUDA   "Enable Clang's native CUDA support" OFF
+                       "ENABLE_CUDA" OFF)
 mark_as_advanced(ENABLE_CLANG_CUDA)
 set(BLT_CLANG_CUDA_ARCH "sm_30" CACHE STRING "Compute architecture to use when generating CUDA code with Clang")
 mark_as_advanced(BLT_CLANG_CUDA_ARCH)
 option(ENABLE_HIP         "Enable HIP support" OFF)
+cmake_dependent_option(ENABLE_CLANG_HIP   "Enable Clang's native HIP support" OFF
+                      "ENABLE_HIP" OFF)
+mark_as_advanced(ENABLE_CLANG_HIP)
+set(BLT_CLANG_HIP_ARCH "gfx906" CACHE STRING "AMDGPU ISA target to use when generating HIP code with Clang" )
+mark_as_advanced(BLT_CLANG_HIP_ARCH)
 option(ENABLE_HCC         "Enable HCC support" OFF)
 set(BLT_ROCM_ARCH "gfx900" CACHE STRING "gfx architecture to use when generating ROCm code")
 
diff --git a/cmake/SetupCodeChecks.cmake b/cmake/SetupCodeChecks.cmake
index 8e3facda9..9eb09c991 100644
--- a/cmake/SetupCodeChecks.cmake
+++ b/cmake/SetupCodeChecks.cmake
@@ -684,7 +684,8 @@ macro(blt_add_clangformat_target)
                         OUTPUT_VARIABLE _version_str
                         OUTPUT_STRIP_TRAILING_WHITESPACE)
         # The version number is the last token - can contain non-numeric
-        string(REGEX MATCH "([0-9a-zA-Z\\-]+(\\.)?)+$" _clangformat_version ${_version_str})
+        string(REGEX MATCH "([0-9a-zA-Z\\-]+\\.[0-9a-zA-Z\\-]+\\.?[0-9a-zA-Z\\-]?)"
+               _clangformat_version ${_version_str})
         # The user may only specify a part of the version (e.g. just the maj ver)
         # so check for substring
         string(FIND "${_clangformat_version}" ${BLT_REQUIRED_CLANGFORMAT_VERSION} VERSION_POS)
diff --git a/cmake/SetupCompilerOptions.cmake b/cmake/SetupCompilerOptions.cmake
index db38469bd..9f3453106 100644
--- a/cmake/SetupCompilerOptions.cmake
+++ b/cmake/SetupCompilerOptions.cmake
@@ -268,7 +268,7 @@ if (BLT_CXX_STD)
         blt_append_custom_compiler_flag(
             FLAGS_VAR CMAKE_CXX_FLAGS
             DEFAULT " "
-            XL "-std=c++1y"
+            XL "-std=c++14"
             PGI "--c++14")
     elseif( BLT_CXX_STD STREQUAL c++17)
         # Error out on what does not support C++17
diff --git a/cmake/thirdparty/SetupHIP.cmake b/cmake/thirdparty/SetupHIP.cmake
index 5437fb580..cc4c7e8f0 100644
--- a/cmake/thirdparty/SetupHIP.cmake
+++ b/cmake/thirdparty/SetupHIP.cmake
@@ -15,24 +15,108 @@ find_package(HIP REQUIRED)
 message(STATUS "HIP version:      ${HIP_VERSION_STRING}")
 message(STATUS "HIP platform:     ${HIP_PLATFORM}")
 
-if("${HIP_PLATFORM}" STREQUAL "hcc" OR "${HIP_PLATFORM}" STREQUAL "amd")
-	set(HIP_RUNTIME_DEFINE "__HIP_PLATFORM_HCC__")
-elseif("${HIP_PLATFORM}" STREQUAL "nvcc")
-	set(HIP_RUNTIME_DEFINE "__HIP_PLATFORM_NVCC__")
+set(HIP_RUNTIME_INCLUDE_DIRS "${HIP_ROOT_DIR}/include")
+if(${HIP_PLATFORM} STREQUAL "hcc")
+	set(HIP_RUNTIME_DEFINES "-D__HIP_PLATFORM_HCC__")
+    find_library(HIP_RUNTIME_LIBRARIES NAMES hip_hcc libhip_hcc
+                PATHS ${HIP_ROOT_DIR}/lib
+                NO_DEFAULT_PATH
+                NO_CMAKE_ENVIRONMENT_PATH
+                NO_CMAKE_PATH
+                NO_SYSTEM_ENVIRONMENT_PATH
+                NO_CMAKE_SYSTEM_PATH)
+    set(HIP_RUNTIME_LIBRARIES "${HIP_ROOT_DIR}/lib/libhip_hcc.so")
+elseif(${HIP_PLATFORM} STREQUAL "clang" OR ${HIP_PLATFORM} STREQUAL "amd")
+    set(HIP_RUNTIME_DEFINES "-D__HIP_PLATFORM_HCC__;-D__HIP_ROCclr__;-D__HIP_PLATFORM_AMD__")
+    find_library(HIP_RUNTIME_LIBRARIES NAMES amdhip64 libamdhip64
+                PATHS ${HIP_ROOT_DIR}/lib
+                NO_DEFAULT_PATH
+                NO_CMAKE_ENVIRONMENT_PATH
+                NO_CMAKE_PATH
+                NO_SYSTEM_ENVIRONMENT_PATH
+                NO_CMAKE_SYSTEM_PATH)
+elseif(${HIP_PLATFORM} STREQUAL "nvcc" OR ${HIP_PLATFORM} STREQUAL "nvidia")
+    set(HIP_RUNTIME_DEFINES "-D__HIP_PLATFORM_NVCC__;-D__HIP_PLATFORM_NVIDIA__")
+    if (${CMAKE_VERSION} VERSION_LESS "3.17.0")
+        find_package(CUDA)
+        find_library(HIP_RUNTIME_LIBRARIES NAMES cudart libcudart
+            PATHS ${CMAKE_CUDA_IMPLICIT_LINK_DIRECTORIES}
+            NO_DEFAULT_PATH
+            NO_CMAKE_ENVIRONMENT_PATH
+            NO_CMAKE_PATH
+            NO_SYSTEM_ENVIRONMENT_PATH
+            NO_CMAKE_SYSTEM_PATH)
+        set(HIP_RUNTIME_INCLUDE_DIRS "${HIP_RUNTIME_INCLUDE_DIRS};${CUDA_INCLUDE_DIRS}")
+    else()
+        find_package(CUDAToolkit)
+        find_library(HIP_RUNTIME_LIBRARIES NAMES cudart libcudart
+            PATHS ${CUDAToolkit_LIBRARY_DIR}
+            NO_DEFAULT_PATH
+            NO_CMAKE_ENVIRONMENT_PATH
+            NO_CMAKE_PATH
+            NO_SYSTEM_ENVIRONMENT_PATH
+            NO_CMAKE_SYSTEM_PATH)
+        set(HIP_RUNTIME_INCLUDE_DIRS "${HIP_RUNTIME_INCLUDE_DIRS};${CUDAToolkit_INCLUDE_DIR}")
+    endif()
 endif()
+
 if ( IS_DIRECTORY "${HIP_ROOT_DIR}/hcc/include" ) # this path only exists on older rocm installs
-        set(HIP_RUNTIME_INCLUDE_DIRS "${HIP_ROOT_DIR}/include;${HIP_ROOT_DIR}/hcc/include" CACHE STRING "")
+    set(HIP_RUNTIME_INCLUDE_DIRS "${HIP_ROOT_DIR}/include;${HIP_ROOT_DIR}/hcc/include" CACHE STRING "")
 else()
-        set(HIP_RUNTIME_INCLUDE_DIRS "${HIP_ROOT_DIR}/include" CACHE STRING "")
+    set(HIP_RUNTIME_INCLUDE_DIRS "${HIP_ROOT_DIR}/include" CACHE STRING "")
 endif()
-set(HIP_RUNTIME_COMPILE_FLAGS "${HIP_RUNTIME_COMPILE_FLAGS};-D${HIP_RUNTIME_DEFINE};-Wno-unused-parameter")
+set(HIP_RUNTIME_COMPILE_FLAGS "${HIP_RUNTIME_COMPILE_FLAGS};-Wno-unused-parameter")
+
+set(_hip_compile_flags " ")
+if (ENABLE_CLANG_HIP)
+    if (NOT (${HIP_PLATFORM} STREQUAL "clang"))
+        message(FATAL_ERROR "ENABLE_CLANG_HIP requires HIP_PLATFORM=clang")
+    endif()
+    set(_hip_compile_flags -x;hip)
+    # Using clang HIP, we need to construct a few CPP defines and compiler flags
+    foreach(_arch ${BLT_CLANG_HIP_ARCH})
+        string(TOUPPER ${_arch} _UPARCH)
+        string(TOLOWER ${_arch} _lowarch)
+        list(APPEND _hip_compile_flags "--offload-arch=${_lowarch}")
+        set(_hip_compile_defines "${HIP_RUNTIME_DEFINES};-D__HIP_ARCH_${_UPARCH}__=1")
+    endforeach(_arch)
+    
+    # We need to pass rocm path as well, for certain bitcode libraries.
+    # First see if we were given it, then see if it exists in the environment.
+    # If not, don't try to guess but print a warning and hope the compiler knows where it is.
+    if (NOT ROCM_PATH)
+        find_path(ROCM_PATH
+            bin/rocminfo
+            ENV ROCM_DIR
+            ENV ROCM_PATH
+            ${HIP_ROOT_DIR}/../
+            ${ROCM_ROOT_DIR}
+            /opt/rocm)
+    endif()
+
+    if(DEFINED ROCM_PATH)
+        list(APPEND _hip_compile_flags "--rocm-path=${ROCM_PATH}")
+    else()
+        message(WARN "ROCM_PATH not set or found! This is typically required for Clang HIP Compilation")
+    endif()
+
+    message(STATUS "Clang HIP Enabled. Clang flags for HIP compilation: ${_hip_compile_flags}")
+    message(STATUS "Defines for HIP compilation: ${_hip_compile_defines}")
+
+    blt_import_library(NAME             hip
+                       DEFINES          ${_hip_compile_defines}
+                       COMPILE_FLAGS    ${_hip_compile_flags}
+                       DEPENDS_ON       ${HIP_RUNTIME_LIBRARIES})
+else()
 
 # depend on 'hip', if you need to use hip
 # headers, link to hip libs, and need to run your source
 # through a hip compiler (hipcc)
 # This is currently used only as an indicator for blt_add_hip* -- FindHIP/hipcc will handle resolution
 # of all required HIP-related includes/libraries/flags.
-blt_import_library(NAME      hip)
+    blt_import_library(NAME      hip)
+endif()
+
 
 # depend on 'hip_runtime', if you only need to use hip
 # headers or link to hip libs, but don't need to run your source
@@ -41,5 +125,6 @@ blt_import_library(NAME          hip_runtime
                    INCLUDES      ${HIP_RUNTIME_INCLUDE_DIRS}
                    DEFINES       ${HIP_RUNTIME_DEFINES}
                    COMPILE_FLAGS ${HIP_RUNTIME_COMPILE_FLAGS}
+                   DEPENDS_ON    ${HIP_RUNTIME_LIBRARIES}
                    TREAT_INCLUDES_AS_SYSTEM ON
                    EXPORTABLE    ${BLT_EXPORT_THIRDPARTY})
diff --git a/cmake/thirdparty/SetupMPI.cmake b/cmake/thirdparty/SetupMPI.cmake
index 812df35e0..c04e0b6ef 100644
--- a/cmake/thirdparty/SetupMPI.cmake
+++ b/cmake/thirdparty/SetupMPI.cmake
@@ -31,7 +31,13 @@ set(_mpi_includes )
 set(_mpi_libraries )
 set(_mpi_link_flags )
 
-message(STATUS "Enable FindMPI:  ${ENABLE_FIND_MPI}")
+
+if(ENABLE_FIND_MPI)
+    message(STATUS "FindMPI Enabled  (ENABLE_FIND_MPI == ON)")
+else()
+    message(STATUS "FindMPI Disabled (ENABLE_FIND_MPI == OFF) ")
+endif()
+
 
 if (ENABLE_FIND_MPI)
     find_package(MPI REQUIRED)
diff --git a/docs/_static/style.css b/docs/_static/style.css
index 775d1a851..ebe26836b 100644
--- a/docs/_static/style.css
+++ b/docs/_static/style.css
@@ -33,5 +33,5 @@
 }
 
 a img.logo {
-  scale: 75%
+  transform: scale(.75)
 }
diff --git a/docs/tutorial/third_party_libraries.rst b/docs/tutorial/third_party_libraries.rst
index e9f5255d3..761087ffb 100644
--- a/docs/tutorial/third_party_libraries.rst
+++ b/docs/tutorial/third_party_libraries.rst
@@ -81,8 +81,8 @@ searching for the include directories and then for the library itself.  Finally
     blt_import_library(NAME       lua
                        TREAT_INCLUDES_AS_SYSTEM ON
                        DEFINES    HAVE_LUA=1
-                       INCLUDES   ${LUA_LIBRARY}
-                       LIBRARIES  ${LUA_INCLUDE_DIR}
+                       INCLUDES   ${LUA_INCLUDE_DIR}
+                       LIBRARIES  ${LUA_LIBRARY}
                        EXPORTABLE ON)
 
 Then ``lua`` is available to be used in the ``DEPENDS_ON`` list in the following
@@ -114,10 +114,11 @@ new imported library's ``NAME`` in the rest of your project.
 
 .. code-block:: cmake
 
-    # FindLua.cmake takes in LUA_DIR, which the directory where Lua was installed to
-    # and fills variables: LUA_FOUND, LUA_LIBRARIES, and LUA_INCLUDE_DIR
-    find_package(Lua NO_DEFAULT_PATH
-                     PATHS ${LUA_DIR})
+    # FindLua.cmake takes in LUA_DIR as an environment variable, which is the directory 
+    # where Lua was installed to and fills variables: LUA_FOUND, LUA_LIBRARIES, and LUA_INCLUDE_DIR
+    set(ENV{LUA_DIR} ${LUA_DIR})
+
+    find_package(Lua)
 
     if (NOT LUA_FOUND)
         MESSAGE(FATAL_ERROR "Could not find Lua in the provided LUA_DIR: ${LUA_DIR}")
@@ -126,8 +127,8 @@ new imported library's ``NAME`` in the rest of your project.
     blt_import_library(NAME       lua
                        TREAT_INCLUDES_AS_SYSTEM ON
                        DEFINES    HAVE_LUA=1
-                       INCLUDES   ${LUA_LIBRARIES}
-                       LIBRARIES  ${LUA_INCLUDE_DIR}
+                       INCLUDES   ${LUA_INCLUDE_DIR}
+                       LIBRARIES  ${LUA_LIBRARIES}
                        EXPORTABLE ON)
 
 
diff --git a/host-configs/llnl/toss_4_x86_64_ib_cray/cce@12.0.0_clang_hip.cmake b/host-configs/llnl/toss_4_x86_64_ib_cray/cce@12.0.0_clang_hip.cmake
new file mode 100644
index 000000000..3c9f904c5
--- /dev/null
+++ b/host-configs/llnl/toss_4_x86_64_ib_cray/cce@12.0.0_clang_hip.cmake
@@ -0,0 +1,55 @@
+# Copyright (c) 2017-2021, Lawrence Livermore National Security, LLC and
+# other BLT Project Developers. See the top-level LICENSE file for details
+# 
+# SPDX-License-Identifier: (BSD-3-Clause)
+
+#------------------------------------------------------------------------------
+# Example host-config file for the rznevada cluster at LLNL
+#------------------------------------------------------------------------------
+# This file provides CMake with paths / details for:
+#  C,C++, & Fortran compilers + MPI & HIP
+# using tce wrappers, rather than HPE Cray PE compiler drivers
+#------------------------------------------------------------------------------
+
+#------------------------------------------------------------------------------
+# HPE Cray cce@12.0.0 compilers
+#------------------------------------------------------------------------------
+# _blt_tutorial_compiler_config_start
+set(CCE_HOME "/usr/tce/packages/cce-tce/cce-12.0.0")
+set(CMAKE_C_COMPILER   "${CCE_HOME}/bin/craycc" CACHE PATH "")
+set(CMAKE_CXX_COMPILER "${CCE_HOME}/bin/crayCC" CACHE PATH "")
+
+# Fortran support
+set(ENABLE_FORTRAN ON CACHE BOOL "")
+set(CMAKE_Fortran_COMPILER "${CCE_HOME}/bin/crayftn" CACHE PATH "")
+# _blt_tutorial_compiler_config_end
+
+#------------------------------------------------------------------------------
+# MPI Support
+#------------------------------------------------------------------------------
+# _blt_tutorial_mpi_config_start
+set(ENABLE_MPI ON CACHE BOOL "")
+
+set(MPI_HOME "/usr/tce/packages/cray-mpich-tce/cray-mpich-8.1.5.6-cce-12.0.0/")
+set(MPI_C_COMPILER "${MPI_HOME}/bin/mpicc" CACHE PATH "")
+
+set(MPI_CXX_COMPILER "${MPI_HOME}/bin/mpicxx" CACHE PATH "")
+
+set(MPI_Fortran_COMPILER "${MPI_HOME}/bin/mpif90" CACHE PATH "")
+# _blt_tutorial_mpi_config_end
+
+#------------------------------------------------------------------------------
+# HIP support
+#------------------------------------------------------------------------------
+# _blt_tutorial_hip_config_start
+set(ENABLE_HIP ON CACHE BOOL "")
+set(ENABLE_CLANG_HIP ON CACHE BOOL "")
+
+set(ROCM_PATH "/opt/rocm-4.1.0/" CACHE PATH "")
+set(HIP_PLATFORM "clang" CACHE STRIING "")
+
+set(BLT_CLANG_HIP_ARCH "gfx908" CACHE STRING "")
+
+
+# _blt_tutorial_hip_config_end
+