Merge llvm branch to master

.github/workflows/coverage.yml

@@ -18,7 +18,7 @@ on:
 env:
   CMAKE_BUILD_PARALLEL_LEVEL: 3
   CTEST_PARALLEL_LEVEL: 1
-  DESIRED_CMAKE_VERSION: 3.15.0
+  DESIRED_CMAKE_VERSION: 3.17.0
   PYTHON_VERSION: 3.8
 
 jobs:

.github/workflows/nmodl-ci.yml

@@ -17,7 +17,7 @@ on:
 env:
   CTEST_PARALLEL_LEVEL: 1
   PYTHON_VERSION: 3.8
-  DESIRED_CMAKE_VERSION: 3.15.0
+  DESIRED_CMAKE_VERSION: 3.17.0
 
 jobs:
   ci:

.github/workflows/nmodl-doc.yml

@@ -17,7 +17,7 @@ on:
 env:
   BUILD_TYPE: Release
   PYTHON_VERSION: 3.8
-  DESIRED_CMAKE_VERSION: 3.15.0
+  DESIRED_CMAKE_VERSION: 3.17.0
 
 jobs:
   ci:

.gitlab-ci.yml

@@ -41,7 +41,8 @@ trigger cvf:
 .spack_nmodl:
   variables:
     SPACK_PACKAGE: nmodl
-    SPACK_PACKAGE_SPEC: ~legacy-unit+python
+    SPACK_PACKAGE_SPEC: ~legacy-unit+python+llvm
+    SPACK_INSTALL_EXTRA_FLAGS: -v
 
 spack_setup:
   extends: .spack_setup_ccache
@@ -65,14 +66,6 @@ build:intel:
   variables:
     SPACK_PACKAGE_COMPILER: intel
 
-build:nvhpc:
-  extends:
-    - .spack_build
-    - .spack_nmodl
-  variables:
-    SPACK_PACKAGE_COMPILER: nvhpc
-    SPACK_PACKAGE_DEPENDENCIES: ^bison%gcc^flex%gcc^py-jinja2%gcc^py-sympy%gcc^py-pyyaml%gcc
-
 .nmodl_tests:
   variables:
     # https://github.com/BlueBrain/nmodl/issues/737
@@ -84,8 +77,30 @@ test:intel:
     - .nmodl_tests
   needs: ["build:intel"]
 
-test:nvhpc:
+.benchmark_config:
+  variables:
+    bb5_ntasks: 1
+    bb5_cpus_per_task: 1
+    bb5_memory: 16G
+    bb5_exclusive: full
+    bb5_constraint: volta # V100 GPU node
+
+.build_allocation:
+  variables:
+    bb5_ntasks: 2   # so we block 16 cores
+    bb5_cpus_per_task: 8 # ninja -j {this}
+    bb5_memory: 76G # ~16*384/80
+
+build_cuda:gcc:
+  extends: [.spack_build, .build_allocation]
+  variables:
+    SPACK_PACKAGE: nmodl
+    SPACK_PACKAGE_SPEC: ~legacy-unit+python+llvm+llvm_cuda
+    SPACK_INSTALL_EXTRA_FLAGS: -v
+    SPACK_PACKAGE_COMPILER: gcc
+
+test_benchmark:gcc:
   extends:
+    - .benchmark_config
     - .ctest
-    - .nmodl_tests
-  needs: ["build:nvhpc"]
+  needs: ["build_cuda:gcc"]

.sanitizers/undefined.supp

@@ -1,3 +1,4 @@
 implicit-integer-sign-change:double vector[2] Eigen::internal::pabs<double vector[2]>(double vector[2] const&)
 unsigned-integer-overflow:nmodl::fast_math::vexp(double)
 unsigned-integer-overflow:nmodl::fast_math::vexpm1(double)
+unsigned-integer-overflow:std::mersenne_twister_engine<unsigned long, 32ul, 624ul, 397ul, 31ul, 2567483615ul, 11ul, 4294967295ul, 7ul, 2636928640ul, 15ul, 4022730752ul, 18ul, 1812433253ul>::_M_gen_rand()

CMakeLists.txt

@@ -5,7 +5,7 @@
 # See top-level LICENSE file for details.
 # =============================================================================
 
-cmake_minimum_required(VERSION 3.15 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.17 FATAL_ERROR)
 
 project(NMODL LANGUAGES CXX)
 
@@ -22,6 +22,11 @@ set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
 # =============================================================================
 option(NMODL_ENABLE_PYTHON_BINDINGS "Enable pybind11 based python bindings" ON)
 option(NMODL_ENABLE_LEGACY_UNITS "Use original faraday, R, etc. instead of 2019 nist constants" OFF)
+option(NMODL_ENABLE_LLVM "Enable LLVM based code generation" OFF)
+option(NMODL_ENABLE_LLVM_GPU "Enable LLVM based GPU code generation" OFF)
+option(NMODL_ENABLE_LLVM_CUDA "Enable LLVM CUDA backend to run GPU benchmark" OFF)
+option(NMODL_ENABLE_JIT_EVENT_LISTENERS "Enable JITEventListener for Perf and Vtune" OFF)
+
 if(NMODL_ENABLE_LEGACY_UNITS)
   add_definitions(-DUSE_LEGACY_UNITS)
 endif()
@@ -174,6 +179,21 @@ cpp_cc_find_python_module(sympy 1.3 REQUIRED)
 cpp_cc_find_python_module(textwrap 0.9 REQUIRED)
 cpp_cc_find_python_module(yaml 3.12 REQUIRED)
 
+# =============================================================================
+# Find LLVM dependencies
+# =============================================================================
+if(NMODL_ENABLE_LLVM)
+  include(cmake/LLVMHelper.cmake)
+  include_directories(${LLVM_INCLUDE_DIRS})
+  add_definitions(-DNMODL_LLVM_BACKEND)
+  if(NMODL_ENABLE_LLVM_CUDA)
+    enable_language(CUDA)
+    find_package(CUDAToolkit)
+    include_directories(${CUDAToolkit_INCLUDE_DIRS})
+    add_definitions(-DNMODL_LLVM_CUDA_BACKEND)
+  endif()
+endif()
+
 # =============================================================================
 # Compiler specific flags for external submodules
 # =============================================================================
@@ -207,6 +227,9 @@ set(MEMORYCHECK_COMMAND_OPTIONS
 # do not enable tests if nmodl is used as submodule
 if(NOT NMODL_AS_SUBPROJECT)
   include(CTest)
+  if(NMODL_ENABLE_LLVM)
+    add_subdirectory(test/benchmark)
+  endif()
   add_subdirectory(test/unit)
   add_subdirectory(test/integration)
 endif()
@@ -271,6 +294,19 @@ message(STATUS "Python Bindings     | ${NMODL_ENABLE_PYTHON_BINDINGS}")
 message(STATUS "Flex                | ${FLEX_EXECUTABLE}")
 message(STATUS "Bison               | ${BISON_EXECUTABLE}")
 message(STATUS "Python              | ${PYTHON_EXECUTABLE}")
+message(STATUS "LLVM Codegen        | ${NMODL_ENABLE_LLVM}")
+if(NMODL_ENABLE_LLVM)
+  message(STATUS "  VERSION           | ${LLVM_PACKAGE_VERSION}")
+  message(STATUS "  INCLUDE           | ${LLVM_INCLUDE_DIRS}")
+  message(STATUS "  CMAKE             | ${LLVM_CMAKE_DIR}")
+  message(STATUS "  JIT LISTENERS     | ${NMODL_ENABLE_JIT_EVENT_LISTENERS}")
+endif()
+message(STATUS "LLVM CUDA Codegen   | ${NMODL_ENABLE_LLVM_CUDA}")
+if(NMODL_ENABLE_LLVM_CUDA)
+  message(STATUS "  CUDA VERSION      | ${CUDAToolkit_VERSION}")
+  message(STATUS "  INCLUDE           | ${CUDAToolkit_INCLUDE_DIRS}")
+  message(STATUS "  LIBRARY           | ${CUDAToolkit_LIBRARY_DIR}")
+endif()
 message(STATUS "--------------+--------------------------------------------------------------")
 message(STATUS " See documentation : https://github.com/BlueBrain/nmodl/")
 message(STATUS "--------------+--------------------------------------------------------------")

INSTALL.md

@@ -21,7 +21,7 @@ To build the project from source, a modern C++ compiler with C++14 support is ne
 
 - flex (>=2.6)
 - bison (>=3.0)
-- CMake (>=3.15)
+- CMake (>=3.17)
 - Python (>=3.7)
 - Python packages : jinja2 (>=2.10), pyyaml (>=3.13), pytest (>=4.0.0), sympy (>=1.3), textwrap
 
@@ -31,7 +31,7 @@ Typically the versions of bison and flex provided by the system are outdated and
 To get recent version of all dependencies we recommend using [homebrew](https://brew.sh/):
 
 ```sh
-brew install flex bison cmake python3
+brew install flex bison cmake python3 llvm
 ```
 
 The necessary Python packages can then easily be added using the pip3 command.
@@ -57,7 +57,7 @@ export PATH=/opt/homebrew/opt/flex/bin:/opt/homebrew/opt/bison/bin:$PATH
 On Ubuntu (>=18.04) flex/bison versions are recent enough and are installed along with the system toolchain:
 
 ```sh
-apt-get install flex bison gcc python3 python3-pip
+apt-get install flex bison gcc python3 python3-pip llvm-dev llvm-runtime llvm clang-format clang
 ```
 
 The Python dependencies are installed using:
@@ -79,6 +79,15 @@ cmake .. -DCMAKE_INSTALL_PREFIX=$HOME/nmodl
 make -j && make install
 ```
 
+If `llvm-config` is not in PATH then set LLVM_DIR as:
+
+```sh
+cmake .. -DCMAKE_INSTALL_PREFIX=$HOME/nmodl -DLLVM_DIR=/path/to/llvm/install/lib/cmake/llvm
+
+# on OSX
+cmake .. -DCMAKE_INSTALL_PREFIX=$HOME/nmodl -DLLVM_DIR=`brew --prefix llvm`/lib/cmake/llvm
+```
+
 And set PYTHONPATH as:
 
 ```sh
@@ -132,6 +141,29 @@ export NMODL_WRAPLIB=/opt/nmodl/lib/libpywrapper.so
 **Note**: In order for all unit tests to function correctly when building without linking against libpython we must
 set `NMODL_PYLIB` before running cmake!
 
+### Using CUDA backend to run benchmarks
+
+`NMODL` supports generating code and compiling it for execution on an `NVIDIA` GPU via its benchmark infrastructure using the `LLVM` backend. To enable the `CUDA` backend to compile and execute the GPU code we need to set the following `CMake` flag during compilation of `NMODL`:
+```
+-DNMODL_ENABLE_LLVM_CUDA=ON
+```
+
+To find the need `CUDA` libraries (`cudart` and `nvrtc`) it's needed to have CUDA Toolkit installed on your system.
+This can be done by installing the CUDA Toolkit from the [CUDA Toolkit website](https://developer.nvidia.com/cuda-downloads) or by installing the `CUDA` spack package and loading the corresponding module.
+
+Then given a supported MOD file you can execute the benchmark on GPU in you supported NVIDIA GPU by running the following command:
+```
+./bin/nmodl <file>.mod llvm --no-debug --ir --opt-level-ir 3 gpu --target-arch "sm_80" --name "nvptx64" --math-library libdevice benchmark --run --libs "${CUDA_ROOT}/nvvm/libdevice/libdevice.10.bc" --opt-level-codegen 3 --instance-size 10000000 --repeat 2 --grid-dim-x 4096 --block-dim-x 256
+```
+The above command executes the benchmark on a GPU with `Compute Architecture` `sm_80` and links the generated code to the `libdevice` optimized math library provided by `NVIDIA`.
+Using the above command you can also select the optimization level of the generated code, the instance size of the generated data, the number of repetitions and the grid and block dimensions for the GPU execution.
+
+**Note**: In order for the CUDA backend to be able to compile and execute the generated code on GPU the CUDA Toolkit version installed needs to have the same version as the `CUDA` installed by the NVIDIA driver in the system that will be used to run the benchmark.
+You can find the CUDA Toolkit version by running the following command:
+```
+nvidia-smi
+```
+and noting the `CUDA Version` stated there. For example if `CUDA Version` reported by `nvidia-smi` is CUDA 11.4 you need to install the `CUDA Toolkit 11.4.*` to be able to compile and execute the GPU code.
 
 ## Testing the Installed Module
 

azure-pipelines.yml

@@ -99,17 +99,21 @@ stages:
         url="https://github.com/ispc/ispc/releases/download/${ispc_version}/ispc-${ispc_version}${ispc_version_suffix}-${url_os}.tar.gz";
         mkdir $(pwd)/$CMAKE_PKG/ispc
         wget --quiet --output-document=- $url | tar -xvzf - -C $(pwd)/$CMAKE_PKG/ispc --strip 1;
+        # install llvm nightly (future v13) TODO: this will fail now, FIX this!
+        wget https://apt.llvm.org/llvm.sh
+        chmod +x llvm.sh
+        sudo ./llvm.sh 13
       env:
-        CMAKE_VER: 'v3.15.0'
-        CMAKE_PKG: 'cmake-3.15.0-Linux-x86_64'
+        CMAKE_VER: 'v3.17.0'
+        CMAKE_PKG: 'cmake-3.17.0-Linux-x86_64'
       displayName: 'Install Dependencies'
     - script: |
         export PATH=$(pwd)/$CMAKE_PKG/bin:/home/vsts/.local/bin:$PATH
         export CXX='g++-8'
         mkdir -p $(Build.Repository.LocalPath)/build
         cd $(Build.Repository.LocalPath)/build
         cmake --version
-        cmake .. -DPYTHON_EXECUTABLE=$(which python3.7) -DCMAKE_INSTALL_PREFIX=$HOME/nmodl -DCMAKE_BUILD_TYPE=Release
+        cmake .. -DPYTHON_EXECUTABLE=$(which python3.7) -DCMAKE_INSTALL_PREFIX=$HOME/nmodl -DCMAKE_BUILD_TYPE=Release -DNMODL_ENABLE_LLVM=ON -DLLVM_DIR=/usr/lib/llvm-13/share/llvm/cmake/
         make -j 2
         if [ $? -ne 0 ]
         then
@@ -119,7 +123,7 @@ stages:
         make install #this is needed for the integration tests
         env CTEST_OUTPUT_ON_FAILURE=1 make test
       env:
-        CMAKE_PKG: 'cmake-3.15.0-Linux-x86_64'
+        CMAKE_PKG: 'cmake-3.17.0-Linux-x86_64'
       displayName: 'Build and Run Unit Tests'
     - script: |
         export PATH=$(pwd)/$CMAKE_PKG/bin:/home/vsts/.local/bin:$PATH
@@ -150,7 +154,7 @@ stages:
         fi
         ./bin/nrnivmodl-core $(Build.Repository.LocalPath)/test/integration/mod
       env:
-        CMAKE_PKG: 'cmake-3.15.0-Linux-x86_64'
+        CMAKE_PKG: 'cmake-3.17.0-Linux-x86_64'
         SHELL: 'bash'
       displayName: 'Build Neuron and Run Integration Tests'
     - script: |
@@ -174,25 +178,25 @@ stages:
         fi
         ./bin/nrnivmodl-core $(Build.Repository.LocalPath)/test/integration/mod
       env:
-        CMAKE_PKG: 'cmake-3.15.0-Linux-x86_64'
+        CMAKE_PKG: 'cmake-3.17.0-Linux-x86_64'
       displayName: 'Build CoreNEURON and Run Integration Tests with ISPC compiler'
   - job: 'osx11'
     pool:
-      vmImage: 'macOS-11'
-    displayName: 'MacOS (11), AppleClang 12.0'
+      vmImage: 'macOS-10.15'
+    displayName: 'MacOS (10.15), AppleClang 13.0 (trunk, May 2021)'
     steps:
     - checkout: self
       submodules: True
     - script: |
-        brew install flex bison cmake python@3 gcc@8
+        brew install flex bison cmake python@3 gcc@8 llvm@13
         python3 -m pip install --upgrade pip setuptools
         python3 -m pip install --user 'Jinja2>=2.9.3' 'PyYAML>=3.13' pytest pytest-cov numpy 'sympy>=1.3'
       displayName: 'Install Dependencies'
     - script: |
         export PATH=/usr/local/opt/flex/bin:/usr/local/opt/bison/bin:$PATH;
         mkdir -p $(Build.Repository.LocalPath)/build
         cd $(Build.Repository.LocalPath)/build
-        cmake .. -DPYTHON_EXECUTABLE=$(which python3) -DCMAKE_INSTALL_PREFIX=$HOME/nmodl -DCMAKE_BUILD_TYPE=RelWithDebInfo -DNMODL_ENABLE_PYTHON_BINDINGS=OFF
+        cmake .. -DPYTHON_EXECUTABLE=$(which python3) -DCMAKE_INSTALL_PREFIX=$HOME/nmodl -DCMAKE_BUILD_TYPE=RelWithDebInfo -DNMODL_ENABLE_PYTHON_BINDINGS=OFF -DLLVM_DIR=$(brew --prefix llvm@13)/lib/cmake/llvm -DNMODL_ENABLE_LLVM=ON
         make -j 2
         if [ $? -ne 0 ]
         then
@@ -237,9 +241,11 @@ stages:
         ./bin/nrnivmodl-core $(Build.Repository.LocalPath)/test/integration/mod
       env:
         SHELL: 'bash'
+      condition: false
       displayName: 'Build Neuron and Run Integration Tests'
   - job: 'manylinux_wheels'
     timeoutInMinutes: 45
+    condition: eq(1,2)
     pool:
       vmImage: 'ubuntu-20.04'
     strategy:
@@ -289,6 +295,7 @@ stages:
     - template: ci/upload-wheels.yml
   - job: 'macos_wheels'
     timeoutInMinutes: 45
+    condition: eq(1,2)
     pool:
       vmImage: 'macOS-11'
     strategy: