Added SYCL2020 compatibility

Author: Cielo

.gitignore

@@ -0,0 +1,4 @@
+*.swp
+*.swo
+/build/
+/.ccls-cache/

CMakeLists.txt

@@ -29,28 +29,59 @@ set(MYBUILDTYPE Release Debug RelWithDebInfo MinSizeRel)
 set(CMAKE_BUILD_TYPE Release CACHE STRING "Choose the type of build")
 set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS ${MYBUILDTYPE})
 
-# Unfortunately, there is no proper support of SyCL in CMake as of now.
+# Unfortunately, there is no proper support of SYCL in CMake as of now.
 # Compilation without passing any environment variables was given as a
 # requirement, hence, we set the compiler here directly (CMake discourages this).
 set(SYCL oneAPI CACHE STRING "Select the SYCL target architecture")
-set(MYSYCL oneAPI oneAPIold LLVM hipSYCL)
+set(MYSYCL oneAPI oneAPIold LLVM OpenSYCL)
 set_property(CACHE SYCL PROPERTY STRINGS ${MYSYCL})
 target_compile_definitions(echo PRIVATE SYCL=${SYCL})
-# TODO: So far the only target for LLVM is NVIDIA, and for hipSYCL omp/cpu. Add the others!
-if (SYCL STREQUAL LLVM)
-  set(CMAKE_CXX_COMPILER "clang++")
-  message(STATUS "Configuring Echo with LLVM (NVIDIA) Toolchain")
-  target_compile_options(echo PUBLIC -fsycl -fsycl-targets=nvptx64-nvidia-cuda)
-  target_link_options   (echo PUBLIC -fsycl -fsycl-targets=nvptx64-nvidia-cuda)
-  message(STATUS "Set device target to NVIDIA PTX64")
-elseif (SYCL STREQUAL hipSYCL )
+target_compile_options(echo PUBLIC -fsycl)
+target_link_options   (echo PUBLIC -fsycl)
+
+# A bit cumbersome to support all!
+if     (SYCL STREQUAL OpenSYCL )
   set(CMAKE_CXX_COMPILER "syclcc-clang")
-  message(STATUS "Configuring Echo with hipSycl Toolchain")
-else()
-  set(CMAKE_CXX_COMPILER "dpcpp")
-  message(STATUS "Configuring Echo with LLVM (Intel) Toolchain")
-  target_compile_options(echo PUBLIC -fsycl -fsycl-unnamed-lambda -xHost)
-  target_link_options   (echo PUBLIC -fsycl)
+  message(STATUS "Configuring Echo with OpenSycl Toolchain")
+elseif (SYCL STREQUAL LLVM)
+  set(CMAKE_CXX_COMPILER "clang++")
+  set(SYCL_LLVM_CHAIN Intel CACHE STRING "Intel/NVidia/AMD")
+  set(MYCHAIN AMD Intel NVidia)
+  set_property(CACHE SYCL_LLVM_CHAIN PROPERTY STRINGS ${MYCHAIN})
+  message(STATUS "Configuring Echo with LLVM (${SYCL_LLVM_CHAIN}) Toolchain")
+  if     (SYCL_LLVM_CHAIN STREQUAL AMD)
+    target_compile_options(echo PUBLIC -fsycl-targets=amdgcn-amd-amdhsa)
+    target_link_options   (echo PUBLIC -fsycl-targets=amdgcn-amd-amdhsa)
+    set(SYCL_LLVM_ARCH gfx90a CACHE STRING "Run sycl-ls when in doubt.")
+    set(MYOFFLOAD_ARCH gfx90a gfx908 gfx906)
+    set_property(CACHE SYCL_LLVM_ARCH PROPERTY STRINGS ${MYOFFLOAD_ARCH})
+    target_compile_options(echo PUBLIC -Xsycl-target-backend --offload-arch=${SYCL_LLVM_ARCH})
+    target_link_options   (echo PUBLIC -Xsycl-target-backend --offload-arch=${SYCL_LLVM_ARCH})
+    message(STATUS "Set device target to AMDGCN ${SYCL_LLVM_ARCH}")
+  elseif (SYCL_LLVM_CHAIN STREQUAL NVidia)
+    # Atm only AMD needs the arch. specified. Once needed for NVIDIA as well, do it here.
+    unset(SYCL_LLVM_ARCH  CACHE)
+    target_compile_options(echo PUBLIC -fsycl-targets=nvptx64-nvidia-cuda)
+    target_link_options   (echo PUBLIC -fsycl-targets=nvptx64-nvidia-cuda)
+    include(CheckLanguage)
+    check_language(CUDA)
+    if   (CMAKE_CUDA_COMPILER)
+      enable_language(CUDA)
+      target_compile_options(echo PUBLIC --cuda-path=${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}/../../..)
+      target_link_options   (echo PUBLIC --cuda-path=${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}/../../..)
+      #target_compile_options(echo PUBLIC --cuda-path=${CMAKE_CUDA_TOOLKIT_LIBRARY_ROOT})
+      #target_link_options   (echo PUBLIC --cuda-path=${CMAKE_CUDA_TOOLKIT_LIBRARY_ROOT})
+      message(STATUS "Added explicit CUDA language.")
+    else  ( )
+      message(WARNING "Explicit CUDA not found. With Codeplay NVidia plugin you should be fine. Otherwise...")
+    endif ( )
+    message(STATUS "Set device target to NVIDIA PTX64.")
+  endif  ( )
+else   ( )
+  unset(SYCL_LLVM_CHAIN CACHE)
+  unset(SYCL_LLVM_ARCH  CACHE)
+  set(CMAKE_CXX_COMPILER "icpx")
+  message(STATUS "Configuring Echo with oneAPI (Intel) Toolchain")
   message(STATUS "Set device target to SPIR-V Just-in-Time Compilation")
   #-- Stuff specific for oneAPI (for the moment)
   option(ENABLE_MPI "Compile with MPI besides DPC++" OFF)
@@ -96,16 +127,26 @@ else()
 endif()
 
 # -- DPC++ and Runtime
-option(GPU "Uses device selector rather than host" OFF)
-if(GPU)
-  set(SUFFIX "${SUFFIX}_gpu")
-  target_compile_definitions(echo PRIVATE GPU)
-endif(GPU)
+set(SYCL_DEVICE CPU CACHE STRING "Select the primary SYCL device (default/CPU/GPU/etc.). Overridden by forceDevice in echo.par")
+set(MYDEVICE DEF CPU GPU ACC FPGA)
+set_property(CACHE SYCL_DEVICE PROPERTY STRINGS ${MYDEVICE})
+target_compile_definitions(echo PRIVATE DEVICE=DEV_${SYCL_DEVICE})
+if(NOT SYCL_DEVICE STREQUAL DEF )
+  string(TOLOWER ${SYCL_DEVICE} DEVSUFFIX)
+  set(SUFFIX "${SUFFIX}_${DEVSUFFIX}")
+endif()
+# TODO: Should be done way better than this! Also a10sx and s10sx were options...
+if(SYCL_DEVICE STREQUAL FPGA )
+  message(WARNING "VERY experimental Intel FPGA workflow (probably not yet supported)!")
+  target_compile_options(echo PUBLIC -fsycl-link -fintelfpga -Xshardware -Xsboard-package=/glob/development-tools/oneapi/oneapi/intel_s10sx_pac -Xsboard=pac_s10_usm)
+  target_link_options   (echo PUBLIC -fsycl-link -fintelfpga -Xshardware -Xsboard-package=/glob/development-tools/oneapi/oneapi/intel_s10sx_pac -Xsboard=pac_s10_usm)
+endif()
 
+# -- Compiling
 try_compile(COMPILE_SUCCEEDED
-            ${CMAKE_BINARY_DIR}/check
-            SOURCES ${CMAKE_SOURCE_DIR}/cmake/checks/fscheck.cpp
-            CXX_STANDARD 17
+  ${CMAKE_BINARY_DIR}/check
+  SOURCES ${CMAKE_SOURCE_DIR}/cmake/checks/fscheck.cpp
+  CXX_STANDARD 17
 )
 if(COMPILE_SUCCEEDED)
   message(STATUS "C++ filesystem API available. Using that.")
@@ -116,15 +157,14 @@ endif()
 CHECK_INCLUDE_FILE_CXX("filesystem.h" FSH_FOUND)
 CHECK_INCLUDE_FILE_CXX(filesystem FS_FOUND)
 
-
 option(ENABLE_MEMORY_SANITATION "Enable memory sanitizer to find and track invalid memory accesses. Currently disabled." OFF)
-#if (ENABLE_MEMORY_SANITATION AND NOT GPU)
-#  message(STATUS "Memory sanitation enabled. Linking appropriate libraries.")
-#  target_compile_options(echo PUBLIC  -fno-omit-frame-pointer -fsanitize=address -fsanitize=undefined -fsanitize-address-use-after-scope -Wuninitialized -g)
-#  target_link_libraries(echo PUBLIC  -fno-omit-frame-pointer -fsanitize=address -fsanitize=undefined -g)
-#elseif(ENABLE_MEMORY_SANITATION)
-#  message(WARNING "Memory sanitation is only available with disabled GPU for now.")
-#endif()
+if (ENABLE_MEMORY_SANITATION) # AND NOT GPU)
+  message(STATUS "Memory sanitation enabled. Linking appropriate libraries.")
+  target_compile_options(echo PUBLIC -fno-omit-frame-pointer -Wuninitialized -g) # -fsanitize=address -fsanitize=undefined   -fsanitize-address-use-after-scope
+  target_link_libraries (echo PUBLIC -fno-omit-frame-pointer -g) # -fsanitize=address -fsanitize=undefined
+elseif(ENABLE_MEMORY_SANITATION)
+  message(WARNING "Memory sanitation is only available with disabled GPU for now.")
+endif()
 
 option(SINGLE_PRECISION "Change field from double to float" OFF)
 if(SINGLE_PRECISION)

DeviceConfig.cpp

@@ -13,17 +13,54 @@
 #include <sstream>
 
 DeviceConfig::DeviceConfig() {
-  for (auto const &p : mysycl::platform::get_platforms()) {
-    for (auto dev : p.get_devices()) {
+  for (auto const &p : mysycl::platform::get_platforms())
+    for ( auto dev : p.get_devices() )
       devices.push_back(dev);
-    }
+}
+
+device DeviceConfig::deviceWith(int id){
+  Logger *log = Logger::getInstance(); log->setPar(true);
+  device temp;
+  if( (id>=0) && (id<devices.size()) ){
+    ((*log)+2)<<TAG <<"Looking at device #"<<id; log->fl();
+    temp = devices[id];
+  }else{
+    ((*log)+2)<<TAG <<"Looking at DPEcho default device."; log->fl();
+    listDevices();
+#if   SYCL==oneAPI || SYCL==LLVM
+#if   DEVICE==DEV_CPU
+    temp = mysycl::device(mysycl::cpu_selector_v); // mysycl::cpu_selector_v         fallBackSel;
+#elif DEVICE==DEV_GPU
+    temp = mysycl::device(mysycl::gpu_selector_v); // mysycl::gpu_selector_v         fallBackSel;
+#elif DEVICE==DEV_ACC
+    temp = mysycl::device(mysycl::accelerator_selector_v); // mysycl::accelerator_selector_v fallBackSel;
+#elif DEVICE==DEV_FPGA
+    temp = mysycl::device(mysycl::accelerator_selector_v); //    mysycl::accelerator_selector_v fallBackSel;
+#else // host is deprecated in SYCL2020
+    temp = mysycl::device(mysycl::default_selector_v); //    mysycl::default_selector_v     fallBackSel;
+#endif
+#else
+#if   DEVICE==DEV_CPU
+    mysycl::cpu_selector         fallBackSel;
+#elif DEVICE==DEV_GPU
+    mysycl::gpu_selector         fallBackSel;
+#elif DEVICE==DEV_ACC
+    mysycl::accelerator_selector fallBackSel;
+#elif DEVICE==DEV_FPGA
+    mysycl::accelerator_selector fallBackSel;
+#else // host is deprecated in SYCL2020
+    mysycl::default_selector     fallBackSel;
+#endif
+    temp = mysycl::device(fallBackSel);
+#endif
   }
+  printTargetInfo(temp);
+  return temp;
 }
 
 void DeviceConfig::listDevices() {
-  Logger *log = Logger::getInstance();
-  log->setPar(true);
-  ((*log) + 2) << TAG <<"\n\t# SYCL devices:\t" << devices.size() ;
+  Logger *log = Logger::getInstance(); log->setPar(false);
+  ((*log) + 2) << TAG <<"\n\t# Available SYCL devices:\t" << devices.size() ;
   for (size_t i = 0; i < devices.size(); i++) {
     bool hasDpSupport = devices[i].has(aspect::fp64);
     (*log) <<"\n\t- Device #" << i << ":\t"
@@ -64,19 +101,17 @@ device DeviceConfig::debugDevice() {
   throw std::runtime_error("No debug device is available on this machine!");
 }
 
-void  DeviceConfig::printTargetInfo ( mysycl::queue q) {
-  Logger *Log = Logger::getInstance(); Log->setPar(false);
-  auto dev = q.get_device();
+void  DeviceConfig::printTargetInfo(device dev) {
+  Logger *Log = Logger::getInstance(); Log->setPar(true);
   *Log+0<<TAG
-        << "\n\tHardware "   << dev.get_info<info::device::name>() << " is " << (dev.is_host()? "HOST ":"")
+        << "\n\tHardware "   << dev.get_info<info::device::name>() // << " is " << (dev.is_host()? "HOST ":"")
         << (dev.is_cpu()? "CPU ":"") << (dev.is_gpu()? "GPU ":"") << (dev.is_accelerator()? " ACCELERATOR ":"")
-        << "\n\tMax Compute Units  : " << dev.get_info<info::device::max_compute_units>  ();
-#if SYCL <= ONEAPI
+        << "\n\tMax Compute Units  : " << dev.get_info<info::device::max_compute_units>  ();   Log->fl();
+#if SYCL<=oneAPI
   *Log+0<< "\n\tMax Work Group Size: " << dev.get_info<info::device::max_work_group_size>()
         << "\n\tGlobal Memory / GB : " << dev.get_info<info::device::global_mem_size>    ()/pow(1024.0, 3)
-        << "\n\tLocal  Memory / kB : " << dev.get_info<info::device::local_mem_size>     ()/1024.0      ;
+        << "\n\tLocal  Memory / kB : " << dev.get_info<info::device::local_mem_size>     ()/1024.0      ;  Log->fl();
 #else
-  Log->fl();
   *Log+18<<"\n\tMax Work Group Size, Global and Local Memory queries are handled differently outside oneAPI.";
 #endif
   Log->fl();

DeviceConfig.hpp

@@ -17,16 +17,16 @@
 
 struct DeviceConfig {
   private:
-  std::vector<device> devices;
+    std::vector<device> devices;
 
   public:
-  DeviceConfig();
-  void printTargetInfo(mysycl::queue);
-  void listDevices();
-  device deviceWith(int id);
-  device debugDevice();
-  std::vector<device> gpus();
-  std::vector<device> cpus();
+    DeviceConfig();
+    void printTargetInfo(mysycl::device);
+    void listDevices();
+    device deviceWith(int imd);
+    device debugDevice();
+    std::vector<device> gpus();
+    std::vector<device> cpus();
 };
 
 #endif

Domain.hpp

@@ -21,8 +21,6 @@ class Domain {
   private:
     Logger *Log;
     mysycl::queue qq;  field *bufL, *bufR; // Tools for the BCex
-    // Most of the following could be public const and save us the functions below
-    // IF we pass values from constructor instead of reading echo.par in there.
     int cartDims_[3], cartPeriodic_[3], cartCoords_[3], bcType_[3];
     field boxMin_[3], boxMax_[3], boxSize_[3]; // Global info, physical
     field locMin_[3], locMax_[3], locSize_[3]; // This rank info, physical

Physics.cpp

@@ -117,11 +117,6 @@ SYCL_EXTERNAL void cons2prim(id<1> myId, unsigned n, field_array u, field_array
 //-- Fluxes and characteristic velocities.
 //    IMPORTANT: all local quantities, they have been sampled --> access simply by eg. f[VX]
 void physicalFlux(int dir, Metric &g, field vD[FLD_TOT], field uD[FLD_TOT], field f[FLD_TOT], field vf[2], field vt[2] ){
-
-  static const CONSTANT char FMTA[] = "%s: %lf %lf %lf %lf %lf %lf %lf %lf %lf \n";
-  static const CONSTANT char FMTV[] = "%s: %lf %lf %lf\n";
-  static const CONSTANT char FMTF[] = "%s: %lf \n";
-
   field alpha = g.alpha(), betai[3], gCov[9], gCon[9];
   g.beta(betai);
   g.g3DCov(gCov);

Problem.cpp

@@ -14,13 +14,13 @@
 
 using namespace std;
 
-Problem::Problem(mysycl::queue qx, Grid *grid, Domain *D, field_array &fld ){
+Problem::Problem(mysycl::queue qx, string &confFile, Grid *grid, Domain *D, field_array &fld ){
   Log = Logger::getInstance(); grid_ = grid; D_ = D; N_ = grid_->nht;
   iOut_ = 0; iStep_ = 0; nStep_ = 0;  dumpHalos = false; locSize = 1;
   tMax_   = 1.0; dt_ = 0.0; t_ = 0.0, tOut_=0.025, cfl_ = 0.8/3.0; // Divide by 3 as it's 3D
   qq = qx;
   stepTime_.init();
-  std::ifstream inFile("echo.par");   std::string key, val;
+  std::ifstream inFile(confFile);   std::string key, val;
   Log->setPar(false); *Log+3<<TAG<<"Reading input: ";
   while (std::getline(inFile, key, ' ') && std::getline(inFile, val)){
     if(!key.compare("tMax" )){ tMax_  = static_cast<field>(stod(val)); *Log<<"\n\ttMax  "<<tMax_  ; continue;}
@@ -63,7 +63,7 @@ void Problem::dump(field_array &v){ // Asynchronous output
   } else {  // Manual indexing necessary
     field *vt[FLD_TOT]; for (int i = 0; i < FLD_TOT; i++) vt[i] = v[i];
     field *outt[FLD_TOT]; for (int i = 0; i < FLD_TOT; i++) outt[i] = out[i];
-    qq.parallel_for(range(gr.n[0], gr.n[1], gr.n[2]), [=](item<3> it) {
+    qq.parallel_for<class parForDump>(range(gr.n[0], gr.n[1], gr.n[2]), [=](item<3> it) {
       auto iOut= it.get_linear_id(); // Output array has NH halo scope here
       auto iV  = globLinId(it.get_id(), gr.nh, gr.h); // v has WH indexing; offset by halos
       for(int iVar=0; iVar<FLD_TOT; ++iVar)
@@ -83,7 +83,7 @@ void Problem::dump(field_array &v){ // Asynchronous output
 
 void Problem::InitConstWH(field *v, field val) { // HOST CODE: kernel for initialization.
   if(!v){ Log->Error("%s Array was not initialized.", TAG); return; }
-  qq.parallel_for(range<3>(grid_->nh[0], grid_->nh[1], grid_->nh[2]), [=, gr = *(this->grid_)](item<3> it) {
+  qq.parallel_for<class parForInitConstWH>(range<3>(grid_->nh[0], grid_->nh[1], grid_->nh[2]), [=, gr = *(this->grid_)](item<3> it) {
     int offset[3] = {0,0,0};
     auto iV  = globLinId(it, gr.nh, offset); // v has WH indexing; offset by halos
     v[iV] = val;
@@ -92,7 +92,7 @@ void Problem::InitConstWH(field *v, field val) { // HOST CODE: kernel for initia
 
 void Problem::InitConstNH(field *v, field val) { // HOST CODE: kernel for initialization.
   if(!v){ Log->Error("%s Array was not initialized.", TAG); return; }
-  qq.parallel_for(range<3>(grid_->n[0], grid_->n[1], grid_->n[2]), [=, gr = *(this->grid_)](item<3> it) {
+  qq.parallel_for<class parForInitConstNH>(range<3>(grid_->n[0], grid_->n[1], grid_->n[2]), [=, gr = *(this->grid_)](item<3> it) {
     auto iV  = globLinId(it, gr.nh, gr.h); // v has WH indexing; offset by halos
     v[iV] = val;
   });
@@ -147,7 +147,7 @@ void Problem::Alfven(field_array &v, field_array &u){ // HOST CODE: Initializing
   //-- Device code
   field bS[]={D_->boxSize(0), D_->boxSize(1), D_->boxSize(2)};
   Grid gr = *grid_; // For ease of lambda capture
-  qq.parallel_for(range(gr.n[0], gr.n[1], gr.n[2]), [=, NN=N_](item<3> it) {
+  qq.parallel_for<class parForProblemAlfven>(range(gr.n[0], gr.n[1], gr.n[2]), [=, NN=N_](item<3> it) {
     field phi = 0.0, bx, by, bz, vx, vy, vz;
     auto i = globLinId(it, gr.nh, gr.h); // Addressing fld: WH indexing
 

Problem.hpp

@@ -31,7 +31,7 @@ class Problem {
     bool dumpHalos;
     field *out[FLD_TOT]; // Just to print
 
-    Problem(mysycl::queue q, Grid *g, Domain *f, field_array &out);
+    Problem(mysycl::queue q, std::string &confFile, Grid *g, Domain *f, field_array &out);
     void InitRampWH (field *);
     void InitRampNH (field *);
     void InitConstWH(field *, field );