Fix CMake for TSP-stdpar example

TSP-stdpar can be built using g++ or clang
for the CPU and using nvc++ for the GPU.
The CMake file now has the necessary code.
There is a README file explaining the compilation
process as well.

day4/examples/TSP-stdpar/CMakeLists.txt

 cmake_minimum_required(VERSION 3.20)
 
-set(CMAKE_CXX_COMPILER nvc++)
 set(CMAKE_CXX_STANDARD 20)
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -stdpar")
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 project(TSP-stdpar CXX)
+message(STATUS "Compiler: ${CMAKE_CXX_COMPILER_ID}")
+if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -stdlib=libc++")
+elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "NVHPC")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -stdpar")
+endif()
 include(cmake/CPM.cmake)
 CPMAddPackage("gh:bfgroup/Lyra#1.6.1")
 add_executable(tsp TSP.cc)
 target_link_libraries(tsp PRIVATE bfg::lyra)
+if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang"
+		OR "${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
+find_package(TBB REQUIRED)
+target_link_libraries(tsp PRIVATE TBB::tbb TBB::tbbmalloc)
+endif()
 

day4/examples/TSP-stdpar/README

+TSP example for the GPU with -stdpar
+====================================
+
+This example is meant to demonstrate solving
+problems using standard library algorithms on
+a GPU. This is NOT a CUDA program, but just
+a regular C++ code, which can be built using
+g++ or clang++ for the CPU. You can see that
+by bulding it like this (for CPU):
+
+mkdir build_cpu
+cd build_cpu
+CXX=g++ cmake -DCMAKE_BUILD_TYPE=Release ..
+make
+./tsp -r -n 13
+
+Without changing any code, you should also be
+able to build the same program for the GPU.
+First make sure that the NVidia NVHPC tool kit
+is in the relevant PATH variables. This is not
+the regular CUDA compiler, nvcc, but instead
+the nvc++ compiler. In many supercomputing
+centres, that compiler is added to the PATH
+when you load a module, usually having NVHPC
+in its name. Here is how you build for the GPU:
+
+ml NVHPC CMake
+mkdir build_gpu
+cd build_gpu
+CXX=nvc++ cmake -DCMAKE_CXX_FLAGS="-gpu=cc70" -DCMAKE_BUILD_TYPE=Release ..
+make
+./tsp -r -n 15
+
+Note 1: We need a GPU with a compute capability of
+at least 70 for std::execution::par to be converted
+into GPU code. That's why the -gpu=cc70 option.
+Without that, the compilation goes through very quickly,
+but it is only being built for serial CPU execution.
+
+Note 2: When building for more recent GPUs, automatic
+GPU code generation does take place. But, as of version
+12 of the nvc++ compiler, linking fails with a message
+
+nvlink: Undefined reference to '__blt_pgi_clzl'
+
+If you successfully build and run this on a GPU, with
+-gpu=cc70 or higher, I would like to hear about what
+you did to fix it!
+