Merge branch 'develop' into main

CHANGELOG.md

@@ -2,8 +2,13 @@
 All notable changes to this project will be documented in this file.
 
 ## Unreleased
+### Added
+- Ability to build Kokkos and RAJA versions against existing packages.
+
 ### Changed
 - RAJA CUDA CMake build issues resolved.
+- Fix CUDA memory limit check.
+- Use long double for `check_solution` in case of large problem size.
 
 ## [v4.0] - 2021-12-22
 

README.md

@@ -100,7 +100,7 @@ The source for each model's implementations are located in `./src/<model>`.
 
 Currently available models are:
 ```
-omp;ocl;std;std20;hip;cuda;kokkos;sycl;sycl2020;acc;raja;tbb;thrust
+omp;ocl;std-data;std-indices;std-ranges;hip;cuda;kokkos;sycl;sycl2020;acc;raja;tbb;thrust
 ```
 
 #### Overriding default flags

src/ci-prepare-bionic.sh

@@ -134,20 +134,26 @@ setup_aocc() {
 
 setup_nvhpc() {
  echo "Preparing Nvidia HPC SDK"
-  local tarball="nvhpc.tar.gz"
+  local nvhpc_ver="22.3"
-#  local url="http://localhost:8000/nvhpc_2021_219_Linux_x86_64_cuda_11.4.tar.gz"
+  local nvhpc_release="2022_223"
-  local url="https://developer.download.nvidia.com/hpc-sdk/21.9/nvhpc_2021_219_Linux_x86_64_cuda_11.4.tar.gz"
+  local cuda_ver="11.6"
+
+  local tarball="nvhpc_$nvhpc_ver.tar.gz"
+
+  local url="https://developer.download.nvidia.com/hpc-sdk/$nvhpc_ver/nvhpc_${nvhpc_release}_Linux_x86_64_cuda_$cuda_ver.tar.gz"
   get_and_untar "$tarball" "$url"
 
-  local sdk_dir="$PWD/nvhpc_2021_219_Linux_x86_64_cuda_11.4/install_components/Linux_x86_64/21.9"
+  local sdk_dir="$PWD/nvhpc_${nvhpc_release}_Linux_x86_64_cuda_$cuda_ver/install_components/Linux_x86_64/$nvhpc_ver"
   local bin_dir="$sdk_dir/compilers/bin"
   "$bin_dir/makelocalrc" "$bin_dir" -x
 
   export_var NVHPC_SDK_DIR "$sdk_dir"
-  export_var NVHPC_CUDA_DIR "$sdk_dir/cuda/11.4"
+  export_var NVHPC_CUDA_DIR "$sdk_dir/cuda/$cuda_ver"
 
   export_var NVHPC_NVCXX "$bin_dir/nvc++"
-  export_var NVHPC_NVCC "$sdk_dir/cuda/11.4/bin/nvcc"
+  export_var NVHPC_NVCC "$bin_dir/nvcc"
+  export_var NVHPC_CUDA_VER "$cuda_ver"
+#  export_var NVHPC_NVCC "$sdk_dir/cuda/$cuda_ver/bin/nvcc"
 
   echo "Installed CUDA versions:"
   ls "$sdk_dir/cuda"

src/ci-test-compile.sh

@@ -122,7 +122,7 @@ run_build() {
 
 AMD_ARCH="gfx_903"
 NV_ARCH="sm_70"
-NV_ARCH_CCXY="cuda11.4,cc80"
+NV_ARCH_CCXY="cuda${NVHPC_CUDA_VER:?},cc80"
 
 build_gcc() {
   local name="gcc_build"
@@ -175,9 +175,9 @@ build_gcc() {
   local current=$("$CMAKE_BIN" --version | head -n 1 | cut -d ' ' -f3)
   local required="3.15.0"
   if [ "$(printf '%s\n' "$required" "$current" | sort -V | head -n1)" = "$required" ]; then
-    run_build $name "${GCC_CXX:?}" thrust "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DSDK_DIR=$NVHPC_CUDA_DIR/include -DTHRUST_IMPL=CUDA -DBACKEND=CUDA"
+    run_build $name "${GCC_CXX:?}" thrust "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DSDK_DIR=$NVHPC_CUDA_DIR/lib64/cmake -DTHRUST_IMPL=CUDA -DBACKEND=CUDA"
-    run_build $name "${GCC_CXX:?}" thrust "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DSDK_DIR=$NVHPC_CUDA_DIR/include -DTHRUST_IMPL=CUDA -DBACKEND=OMP"
+    run_build $name "${GCC_CXX:?}" thrust "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DSDK_DIR=$NVHPC_CUDA_DIR/lib64/cmake -DTHRUST_IMPL=CUDA -DBACKEND=OMP"
-    run_build $name "${GCC_CXX:?}" thrust "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DSDK_DIR=$NVHPC_CUDA_DIR/include -DTHRUST_IMPL=CUDA -DBACKEND=CPP"
+    run_build $name "${GCC_CXX:?}" thrust "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DSDK_DIR=$NVHPC_CUDA_DIR/lib64/cmake -DTHRUST_IMPL=CUDA -DBACKEND=CPP"
 
     # FIXME CUDA Thrust + TBB throws the following error:
     #    /usr/lib/gcc/x86_64-linux-gnu/9/include/avx512fintrin.h(9146): error: identifier "__builtin_ia32_rndscaless_round" is undefined
@@ -187,7 +187,7 @@ build_gcc() {
     #    /usr/lib/gcc/x86_64-linux-gnu/9/include/avx512dqintrin.h(1365): error: identifier "__builtin_ia32_fpclassss" is undefined
     #    /usr/lib/gcc/x86_64-linux-gnu/9/include/avx512dqintrin.h(1372): error: identifier "__builtin_ia32_fpclasssd" is undefined
 
-    #    run_build $name "${GCC_CXX:?}" THRUST "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DSDK_DIR=$NVHPC_CUDA_DIR/include -DTHRUST_IMPL=CUDA -DBACKEND=TBB"
+    #    run_build $name "${GCC_CXX:?}" THRUST "$cxx -DCMAKE_CUDA_COMPILER=${NVHPC_NVCC:?} -DCUDA_ARCH=$NV_ARCH -DSDK_DIR=$NVHPC_CUDA_DIR/lib64/cmake -DTHRUST_IMPL=CUDA -DBACKEND=TBB"
   else
     echo "CMake version ${current} < ${required}, skipping Thrust models"
   fi

src/cuda/CUDAStream.cu

@@ -51,7 +51,7 @@ CUDAStream<T>::CUDAStream(const int ARRAY_SIZE, const int device_index)
   // Check buffers fit on the device
   cudaDeviceProp props;
   cudaGetDeviceProperties(&props, 0);
-  if (props.totalGlobalMem < 3*ARRAY_SIZE*sizeof(T))
+  if (props.totalGlobalMem < std::size_t{3}*ARRAY_SIZE*sizeof(T))
     throw std::runtime_error("Device does not have enough memory for all 3 buffers");
 
   // Create device buffers

src/hip/HIPStream.cpp

@@ -54,7 +54,7 @@ HIPStream<T>::HIPStream(const int ARRAY_SIZE, const int device_index)
   // Check buffers fit on the device
   hipDeviceProp_t props;
   hipGetDeviceProperties(&props, 0);
-  if (props.totalGlobalMem < 3*ARRAY_SIZE*sizeof(T))
+  if (props.totalGlobalMem < std::size_t{3}*ARRAY_SIZE*sizeof(T))
     throw std::runtime_error("Device does not have enough memory for all 3 buffers");
 
   // Create device buffers

src/main.cpp

@@ -487,15 +487,15 @@ void check_solution(const unsigned int ntimes, std::vector<T>& a, std::vector<T>
   goldSum = goldA * goldB * ARRAY_SIZE;
 
   // Calculate the average error
-  double errA = std::accumulate(a.begin(), a.end(), 0.0, [&](double sum, const T val){ return sum + fabs(val - goldA); });
+  long double errA = std::accumulate(a.begin(), a.end(), 0.0, [&](double sum, const T val){ return sum + fabs(val - goldA); });
   errA /= a.size();
-  double errB = std::accumulate(b.begin(), b.end(), 0.0, [&](double sum, const T val){ return sum + fabs(val - goldB); });
+  long double errB = std::accumulate(b.begin(), b.end(), 0.0, [&](double sum, const T val){ return sum + fabs(val - goldB); });
   errB /= b.size();
-  double errC = std::accumulate(c.begin(), c.end(), 0.0, [&](double sum, const T val){ return sum + fabs(val - goldC); });
+  long double errC = std::accumulate(c.begin(), c.end(), 0.0, [&](double sum, const T val){ return sum + fabs(val - goldC); });
   errC /= c.size();
-  double errSum = fabs((sum - goldSum)/goldSum);
+  long double errSum = fabs((sum - goldSum)/goldSum);
 
-  double epsi = std::numeric_limits<T>::epsilon() * 100.0;
+  long double epsi = std::numeric_limits<T>::epsilon() * 100.0;
 
   if (errA > epsi)
     std::cerr

src/std-indices/STDIndicesStream.h

@@ -34,6 +34,7 @@ public:
       iterator& operator++() { num++; return *this; }
       iterator operator++(int) { iterator retval = *this; ++(*this); return retval; }
       iterator operator+(const value_type v) const { return iterator(num + v); }
+      iterator operator+=(int x) { iterator retval = *this; this->num+=x; return retval; }
 
       bool operator==(iterator other) const { return num == other.num; }
       bool operator!=(iterator other) const { return *this != other; }

src/thrust/model.cmake

@@ -53,6 +53,9 @@ macro(setup)
         message(STATUS "NVCC flags: ${CMAKE_CUDA_FLAGS} ${CMAKE_CUDA_FLAGS_${BUILD_TYPE}}")
 
 
+        # XXX NVHPC <= 21.9 has cub-config in `Linux_x86_64/21.9/cuda/11.4/include/cub/cmake`
+        # XXX NVHPC >= 22.3 has cub-config in `Linux_x86_64/22.3/cuda/11.6/lib64/cmake/cub/`
+        # same thing for thrust
         if (SDK_DIR)
             find_package(CUB REQUIRED CONFIG PATHS ${SDK_DIR}/cub)
             find_package(Thrust REQUIRED CONFIG PATHS ${SDK_DIR}/thrust)