CUDA exception handling

cmake/FindCUDACompilerClang.cmake

@@ -77,7 +77,6 @@ function(find_gpu_library)
   endif()
 endfunction(find_gpu_library)
 
-find_gpu_library(VARNAME CUDA_LIBRARY NAMES cuda REQUIRED)
 find_gpu_library(VARNAME CUDART_LIBRARY NAMES cudart REQUIRED)
 find_gpu_library(VARNAME CUFFT_LIBRARY NAMES cufft REQUIRED)
 

cmake/FindCUDACompilerNVCC.cmake

@@ -71,7 +71,6 @@ function(find_gpu_library)
   endif()
 endfunction(find_gpu_library)
 
-find_gpu_library(VARNAME CUDA_LIBRARY NAMES cuda REQUIRED)
 find_gpu_library(VARNAME CUDART_LIBRARY NAMES cudart REQUIRED)
 find_gpu_library(VARNAME CUDA_CUFFT_LIBRARIES NAMES cufft REQUIRED)
 

doc/sphinx/system_setup.rst

@@ -231,16 +231,33 @@ For more information please check :class:`espressomd.cuda_init.CudaInitHandle`.
 List available CUDA devices
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-If you want to list available CUDA devices
-you should access :attr:`espressomd.cuda_init.CudaInitHandle.device_list`, e.g., ::
+If you want to list available CUDA devices, you should call
+:meth:`espressomd.cuda_init.CudaInitHandle.list_devices`::
 
-    system = espressomd.System(box_l=[1, 1, 1])
-
-    print(system.cuda_init_handle.device_list)
+    >>> import espressomd
+    >>> system = espressomd.System(box_l=[1, 1, 1])
+    >>> print(system.cuda_init_handle.list_devices())
+    {0: 'GeForce RTX 2080', 1: 'GeForce GT 730'}
 
-This attribute is read only and will return a dictionary containing
+This method returns a dictionary containing
 the device id as key and the device name as its value.
 
+To get more details on the CUDA devices for each MPI node, call
+:meth:`espressomd.cuda_init.CudaInitHandle.list_devices_properties`::
+
+    >>> import pprint
+    >>> import espressomd
+    >>> system = espressomd.System(box_l=[1, 1, 1])
+    >>> pprint.pprint(system.cuda_init_handle.list_devices_properties())
+    {'seraue': {0: {'name': 'GeForce RTX 2080',
+                    'compute_capability': (7, 5),
+                    'cores': 46,
+                    'total_memory': 8370061312},
+                1: {'name': 'GeForce GT 730',
+                    'compute_capability': (3, 5),
+                    'cores': 2,
+                    'total_memory': 1014104064}}}
+
 .. _Selection of CUDA device:
 
 Selection of CUDA device
@@ -250,9 +267,9 @@ When you start ``pypresso`` your first GPU should be selected.
 If you wanted to use the second GPU, this can be done
 by setting :attr:`espressomd.cuda_init.CudaInitHandle.device` as follows::
 
-    system = espressomd.System(box_l=[1, 1, 1])
-
-    system.cuda_init_handle.device = 1
+    >>> import espressomd
+    >>> system = espressomd.System(box_l=[1, 1, 1])
+    >>> system.cuda_init_handle.device = 1
 
 Setting a device id outside the valid range or a device
 which does not meet the minimum requirements will raise

maintainer/CI/build_cmake.sh

@@ -228,7 +228,7 @@ if [ "${run_checks}" = true ]; then
 
     # fail if built with CUDA but no compatible GPU was found
     if [ "${with_cuda}" = true ] && [ "${hide_gpu}" != true ]; then
-        ./pypresso -c "import espressomd;assert espressomd.gpu_available(), 'No GPU available'" || exit 1
+        ./pypresso -c "import espressomd.cuda_init as gpu;gpu.CudaInitHandle().device = 0" || exit 1
     fi
 
     # unit tests

src/core/CellStructure.hpp

@@ -459,22 +459,21 @@ struct CellStructure {
 
 public:
   /**
-   * @brief Set the particle decomposition to
-   *        AtomDecomposition.
+   * @brief Set the particle decomposition to AtomDecomposition.
    *
-   *        @param comm Communicator to use.
-   *        @param box Box Geometry
+   * @param comm Communicator to use.
+   * @param box Box Geometry
    */
   void set_atom_decomposition(boost::mpi::communicator const &comm,
                               BoxGeometry const &box);
 
   /**
-   * @brief Set the particle decomposition to
-   *        DomainDecomposition.
+   * @brief Set the particle decomposition to DomainDecomposition.
    *
-   *        @param comm Cartesian communicator to use.
-   *        @param box Box Geometry
-   *        @param local_geo Geometry of the local box.
+   * @param comm Cartesian communicator to use.
+   * @param range Interaction range.
+   * @param box Box Geometry
+   * @param local_geo Geometry of the local box.
    */
   void set_domain_decomposition(boost::mpi::communicator const &comm,
                                 double range, BoxGeometry const &box,

src/core/DomainDecomposition.hpp

@@ -66,12 +66,10 @@
  *
  */
 struct DomainDecomposition : public ParticleDecomposition {
-  /** Grind dimensions per node. */
+  /** Grid dimensions per node. */
   Utils::Vector3i cell_grid = {};
-  /** cell size. */
+  /** Cell size. */
   Utils::Vector3d cell_size = {};
-
-private:
   /** Offset in global grid */
   Utils::Vector3i cell_offset = {};
   /** linked cell grid with ghost frame. */
@@ -120,16 +118,16 @@ struct DomainDecomposition : public ParticleDecomposition {
   }
 
 private:
-  /** Fill local_cells list and ghost_cells list for use with domain
+  /** Fill @c m_local_cells list and @c m_ghost_cells list for use with domain
    *  decomposition.
    */
   void mark_cells();
 
   /** Fill a communication cell pointer list. Fill the cell pointers of
    *  all cells which are inside a rectangular subgrid of the 3D cell
    *  grid starting from the
-   *  lower left corner lc up to the high top corner hc. The cell
-   *  pointer list part_lists must already be large enough.
+   *  lower left corner @p lc up to the high top corner @p hc. The cell
+   *  pointer list @p part_lists must already be large enough.
    *  \param part_lists  List of cell pointers to store the result.
    *  \param lc          lower left corner of the subgrid.
    *  \param hc          high up corner of the subgrid.
@@ -159,10 +157,10 @@ struct DomainDecomposition : public ParticleDecomposition {
   /**
    * @brief Split particle list by direction.
    *
-   * Moves all particles from src into left
-   * and right depending if they belong to
-   * the left or right side from local node
-   * in direction dir.
+   * Moves all particles from @p src into @p left
+   * or @p right depending on whether they belong
+   * to the left or right side of the local node
+   * in direction @p dir.
    *
    * @param src Particles to sort.
    * @param left Particles that should go to the left
@@ -185,36 +183,36 @@ struct DomainDecomposition : public ParticleDecomposition {
    *  @brief Calculate cell grid dimensions, cell sizes and number of cells.
    *
    *  Calculates the cell grid, based on the local box size and the range.
-   *  If the number of cells is larger than max_num_cells,
-   *  it increases max_range until the number of cells is
-   *  smaller or equal max_num_cells. It sets:
-   *  cell_grid,
-   *  ghost_cell_grid,
-   *  cell_size, and
-   *  inv_cell_size.
+   *  If the number of cells is larger than @c max_num_cells,
+   *  it increases @c max_range until the number of cells is
+   *  smaller or equal to @c max_num_cells. It sets:
+   *  @c cell_grid,
+   *  @c ghost_cell_grid,
+   *  @c cell_size, and
+   *  @c inv_cell_size.
    *
-   *  @param range Required interacting range. All pairs closer
-   *         than this distance are found.
+   *  @param range interaction range. All pairs closer
+   *               than this distance are found.
    */
   void create_cell_grid(double range);
 
   /** Init cell interactions for cell system domain decomposition.
    *  Initializes the interacting neighbor cell list of a cell.
    *  This list of interacting neighbor cells is used by the Verlet
-   * algorithm.
+   *  algorithm.
    */
   void init_cell_interactions();
 
-  /** Create communicators for cell structure domain decomposition. (see \ref
-   *  GhostCommunicator)
+  /** Create communicators for cell structure domain decomposition (see \ref
+   *  GhostCommunicator).
    */
   GhostCommunicator prepare_comm();
 
   /** Maximal number of cells per node. In order to avoid memory
-   *  problems due to the cell grid one has to specify the maximal
+   *  problems due to the cell grid, one has to specify the maximal
    *  number of cells. If the number of cells is larger
-   *  than max_num_cells the cell grid is reduced.
-   *  max_num_cells has to be larger than 27, e.g. one inner cell.
+   *  than @c max_num_cells, the cell grid is reduced.
+   *  @c max_num_cells has to be larger than 27, e.g. one inner cell.
    */
   static constexpr int max_num_cells = 32768;
 };

src/core/EspressoSystemInterface_cuda.cu

@@ -19,7 +19,7 @@
 
 #include "EspressoSystemInterface.hpp"
 #include "cuda_interface.hpp"
-#include "cuda_utils.hpp"
+#include "cuda_utils.cuh"
 #include "errorhandling.hpp"
 
 #include <cuda.h>

src/core/actor/DipolarBarnesHut.hpp

@@ -27,6 +27,7 @@
 #include "DipolarBarnesHut_cuda.cuh"
 #include "SystemInterface.hpp"
 #include "cuda_interface.hpp"
+#include "cuda_utils.hpp"
 #include "electrostatics_magnetostatics/dipole.hpp"
 #include "errorhandling.hpp"
 
@@ -38,7 +39,7 @@ typedef float dds_float;
 class DipolarBarnesHut : public Actor {
 public:
   DipolarBarnesHut(SystemInterface &s, float epssq, float itolsq) {
-    k = static_cast<float>(dipole.prefactor);
+    m_k = static_cast<float>(dipole.prefactor);
     m_epssq = epssq;
     m_itolsq = itolsq;
     setBHPrecision(&m_epssq, &m_itolsq);
@@ -53,34 +54,44 @@ class DipolarBarnesHut : public Actor {
   };
 
   void computeForces(SystemInterface &s) override {
-    allocBHmemCopy(static_cast<int>(s.npart_gpu()), &m_bh_data);
+    try {
+      allocBHmemCopy(static_cast<int>(s.npart_gpu()), &m_bh_data);
+    } catch (cuda_runtime_error const &err) {
+      runtimeErrorMsg() << "DipolarBarnesHut: " << err.what();
+      return;
+    }
 
     fillConstantPointers(s.rGpuBegin(), s.dipGpuBegin(), m_bh_data);
     initBHgpu(m_bh_data.blocks);
     buildBoxBH(m_bh_data.blocks);
     buildTreeBH(m_bh_data.blocks);
     summarizeBH(m_bh_data.blocks);
     sortBH(m_bh_data.blocks);
-    if (forceBH(&m_bh_data, k, s.fGpuBegin(), s.torqueGpuBegin())) {
+    if (forceBH(&m_bh_data, m_k, s.fGpuBegin(), s.torqueGpuBegin())) {
       runtimeErrorMsg() << "kernels encountered a functional error";
     }
   };
   void computeEnergy(SystemInterface &s) override {
-    allocBHmemCopy(static_cast<int>(s.npart_gpu()), &m_bh_data);
+    try {
+      allocBHmemCopy(static_cast<int>(s.npart_gpu()), &m_bh_data);
+    } catch (cuda_runtime_error const &err) {
+      runtimeErrorMsg() << "DipolarBarnesHut: " << err.what();
+      return;
+    }
 
     fillConstantPointers(s.rGpuBegin(), s.dipGpuBegin(), m_bh_data);
     initBHgpu(m_bh_data.blocks);
     buildBoxBH(m_bh_data.blocks);
     buildTreeBH(m_bh_data.blocks);
     summarizeBH(m_bh_data.blocks);
     sortBH(m_bh_data.blocks);
-    if (energyBH(&m_bh_data, k, (&(((CUDA_energy *)s.eGpu())->dipolar)))) {
+    if (energyBH(&m_bh_data, m_k, (&(((CUDA_energy *)s.eGpu())->dipolar)))) {
       runtimeErrorMsg() << "kernels encountered a functional error";
     }
   };
 
-protected:
-  float k;
+private:
+  float m_k;
   float m_epssq;
   float m_itolsq;
   BHData m_bh_data = {0,       0,       0,       nullptr, nullptr,

src/core/actor/DipolarBarnesHut_cuda.cu

@@ -28,7 +28,8 @@
 #include "DipolarBarnesHut_cuda.cuh"
 
 #include "cuda_init.hpp"
-#include "cuda_utils.hpp"
+#include "cuda_utils.cuh"
+#include "errorhandling.hpp"
 
 #include <thrust/device_ptr.h>
 #include <thrust/reduce.h>
@@ -1181,11 +1182,8 @@ void allocBHmemCopy(int nbodies, BHData *bh_data) {
 
   bh_data->nbodies = nbodies;
 
-  int devID = -1;
-  EspressoGpuDevice dev;
-
-  devID = cuda_get_device();
-  cuda_get_device_props(devID, dev);
+  auto const devID = cuda_get_device();
+  EspressoGpuDevice const dev = cuda_get_device_props(devID);
 
   bh_data->blocks = dev.n_cores;
   // Each node corresponds to a split of the cubic box in 3D space to equal