HDF5: Support for opening scalar datasets

.github/ci/spack-envs/clang15_py311_nompi_h5_ad2/spack.yaml

@@ -8,8 +8,11 @@ spack:
   specs:
   - adios2
   - hdf5
+  - py-h5py
 
   packages:
+    py-h5py:
+      variants: ~mpi
     hdf5:
       variants: ~mpi
     adios2:

.github/ci/spack-envs/clangtidy_nopy_ompi_h5_ad2/spack.yaml

@@ -9,6 +9,7 @@ spack:
   - [email protected]
   - hdf5
   - openmpi
+  - py-h5py
 
   packages:
     adios2:

.github/ci/spack-envs/gcc12_py36_ompi_h5_ad2/spack.yaml

@@ -9,6 +9,7 @@ spack:
   - [email protected]
   - hdf5
   - openmpi
+  - py-h5py
 
   packages:
     adios2:

.github/workflows/linux.yml

@@ -260,7 +260,7 @@ jobs:
     - name: Install
       run: |
         sudo apt-get update
-        sudo apt-get install g++ libopenmpi-dev libhdf5-openmpi-dev python3 python3-numpy python3-mpi4py python3-pandas
+        sudo apt-get install g++ libopenmpi-dev libhdf5-openmpi-dev python3 python3-numpy python3-mpi4py python3-pandas python3-h5py-mpi
 #       TODO ADIOS2
     - name: Build
       env: {CXXFLAGS: -Werror, PKG_CONFIG_PATH: /usr/lib/x86_64-linux-gnu/pkgconfig}
@@ -286,7 +286,7 @@ jobs:
       run: |
         apk update
         apk add hdf5-dev
-        python3.10 -m pip install numpy
+        python3.10 -m pip install numpy h5py
     - name: Build
       env: {CXXFLAGS: -Werror}
       run: |

conda.yml

@@ -24,6 +24,7 @@ dependencies:
   - doxygen
   - git
   - hdf5=*=mpi_openmpi_*
+  - h5py
   - mamba
   - make
   - mpi4py

src/IO/HDF5/HDF5IOHandler.cpp

@@ -1294,15 +1294,24 @@ void HDF5IOHandlerImpl::openDataset(
     *dtype = d;
 
     int ndims = H5Sget_simple_extent_ndims(dataset_space);
-    std::vector<hsize_t> dims(ndims, 0);
-    std::vector<hsize_t> maxdims(ndims, 0);
+    if (ndims == 0)
+    {
+        // Is a scalar. Since the openPMD-api frontend supports no scalar
+        // datasets, return the extent as {1}
+        *parameters.extent = {1};
+    }
+    else
+    {
+        std::vector<hsize_t> dims(ndims, 0);
+        std::vector<hsize_t> maxdims(ndims, 0);
 
-    H5Sget_simple_extent_dims(dataset_space, dims.data(), maxdims.data());
-    Extent e;
-    for (auto const &val : dims)
-        e.push_back(val);
-    auto extent = parameters.extent;
-    *extent = e;
+        H5Sget_simple_extent_dims(dataset_space, dims.data(), maxdims.data());
+        Extent e;
+        for (auto const &val : dims)
+            e.push_back(val);
+        auto &extent = parameters.extent;
+        *extent = e;
+    }
 
     herr_t status;
     status = H5Sclose(dataset_space);
@@ -1555,28 +1564,54 @@ void HDF5IOHandlerImpl::writeDataset(
         "[HDF5] Internal error: Failed to open HDF5 dataset during dataset "
         "write");
 
-    std::vector<hsize_t> start;
-    for (auto const &val : parameters.offset)
-        start.push_back(static_cast<hsize_t>(val));
-    std::vector<hsize_t> stride(start.size(), 1); /* contiguous region */
-    std::vector<hsize_t> count(start.size(), 1); /* single region */
-    std::vector<hsize_t> block;
-    for (auto const &val : parameters.extent)
-        block.push_back(static_cast<hsize_t>(val));
-    memspace =
-        H5Screate_simple(static_cast<int>(block.size()), block.data(), nullptr);
     filespace = H5Dget_space(dataset_id);
-    status = H5Sselect_hyperslab(
-        filespace,
-        H5S_SELECT_SET,
-        start.data(),
-        stride.data(),
-        count.data(),
-        block.data());
-    VERIFY(
-        status == 0,
-        "[HDF5] Internal error: Failed to select hyperslab during dataset "
-        "write");
+    int ndims = H5Sget_simple_extent_ndims(filespace);
+
+    if (ndims == 0)
+    {
+        if (parameters.offset != Offset{0} || parameters.extent != Extent{1})
+        {
+            std::stringstream errorMessage;
+            errorMessage
+                << "HDF5 dataset '" << concrete_h5_file_position(writable)
+                << "' is scalar (dimensionality 0) and must be accessed with "
+                   "offset [0] and extent [1]. Was accessed with offset ";
+            auxiliary::write_vec_to_stream(errorMessage, parameters.offset)
+                << " and extent ";
+            auxiliary::write_vec_to_stream(errorMessage, parameters.extent)
+                << ".";
+            throw error::WrongAPIUsage(errorMessage.str());
+        }
+        memspace = H5Screate_simple(0, nullptr, nullptr);
+        VERIFY(
+            memspace > 0,
+            "[HDF5] Internal error: Failed to create memspace during dataset "
+            "write");
+    }
+    else
+    {
+        std::vector<hsize_t> start;
+        for (auto const &val : parameters.offset)
+            start.push_back(static_cast<hsize_t>(val));
+        std::vector<hsize_t> stride(start.size(), 1); /* contiguous region */
+        std::vector<hsize_t> count(start.size(), 1); /* single region */
+        std::vector<hsize_t> block;
+        for (auto const &val : parameters.extent)
+            block.push_back(static_cast<hsize_t>(val));
+        memspace = H5Screate_simple(
+            static_cast<int>(block.size()), block.data(), nullptr);
+        status = H5Sselect_hyperslab(
+            filespace,
+            H5S_SELECT_SET,
+            start.data(),
+            stride.data(),
+            count.data(),
+            block.data());
+        VERIFY(
+            status == 0,
+            "[HDF5] Internal error: Failed to select hyperslab during dataset "
+            "write");
+    }
 
     void const *data = parameters.data.get();
 
@@ -2013,28 +2048,54 @@ void HDF5IOHandlerImpl::readDataset(
         "[HDF5] Internal error: Failed to open HDF5 dataset during dataset "
         "read");
 
-    std::vector<hsize_t> start;
-    for (auto const &val : parameters.offset)
-        start.push_back(static_cast<hsize_t>(val));
-    std::vector<hsize_t> stride(start.size(), 1); /* contiguous region */
-    std::vector<hsize_t> count(start.size(), 1); /* single region */
-    std::vector<hsize_t> block;
-    for (auto const &val : parameters.extent)
-        block.push_back(static_cast<hsize_t>(val));
-    memspace =
-        H5Screate_simple(static_cast<int>(block.size()), block.data(), nullptr);
     filespace = H5Dget_space(dataset_id);
-    status = H5Sselect_hyperslab(
-        filespace,
-        H5S_SELECT_SET,
-        start.data(),
-        stride.data(),
-        count.data(),
-        block.data());
-    VERIFY(
-        status == 0,
-        "[HDF5] Internal error: Failed to select hyperslab during dataset "
-        "read");
+    int ndims = H5Sget_simple_extent_ndims(filespace);
+
+    if (ndims == 0)
+    {
+        if (parameters.offset != Offset{0} || parameters.extent != Extent{1})
+        {
+            std::stringstream errorMessage;
+            errorMessage
+                << "HDF5 dataset '" << concrete_h5_file_position(writable)
+                << "' is scalar (dimensionality 0) and must be accessed with "
+                   "offset [0] and extent [1]. Was accessed with offset ";
+            auxiliary::write_vec_to_stream(errorMessage, parameters.offset)
+                << " and extent ";
+            auxiliary::write_vec_to_stream(errorMessage, parameters.extent)
+                << ".";
+            throw error::WrongAPIUsage(errorMessage.str());
+        }
+        memspace = H5Screate_simple(0, nullptr, nullptr);
+        VERIFY(
+            memspace > 0,
+            "[HDF5] Internal error: Failed to create memspace during dataset "
+            "read");
+    }
+    else
+    {
+        std::vector<hsize_t> start;
+        for (auto const &val : parameters.offset)
+            start.push_back(static_cast<hsize_t>(val));
+        std::vector<hsize_t> stride(start.size(), 1); /* contiguous region */
+        std::vector<hsize_t> count(start.size(), 1); /* single region */
+        std::vector<hsize_t> block;
+        for (auto const &val : parameters.extent)
+            block.push_back(static_cast<hsize_t>(val));
+        memspace = H5Screate_simple(
+            static_cast<int>(block.size()), block.data(), nullptr);
+        status = H5Sselect_hyperslab(
+            filespace,
+            H5S_SELECT_SET,
+            start.data(),
+            stride.data(),
+            count.data(),
+            block.data());
+        VERIFY(
+            status == 0,
+            "[HDF5] Internal error: Failed to select hyperslab during dataset "
+            "read");
+    }
 
     void *data = parameters.data.get();
 

test/python/unittest/API/APITest.py

@@ -2239,6 +2239,58 @@ def testSeriesConstructors(self):
                 s = io.Series(f, io.Access.create, c)
                 s.close()
 
+    def testScalarHdf5Fields(self):
+        if "hdf5" not in io.variants:
+            return
+        try:
+            import h5py
+        except ImportError:
+            return
+
+        # While the openPMD-api (currently) does not create scalar HDF5
+        # datasets, we should at least try reading and modifying them in files
+        # that were created elsewhere. Scalar here refers to a dataset without
+        # dimension. Interacting with them in the openPMD-api is possible by
+        # specifying a single element, i.e. offset=[0], extent=[1].
+        # For testing this, create a dataset, then use h5py to create a scalar
+        # dataset in the file. Then, open first for reading, then for
+        # modifying.
+
+        file = "../samples/scalar_hdf5.h5"
+        series_write = io.Series(file, io.Access.create)
+        E_x = series_write.write_iterations()[0].meshes["E"]["x"]
+        E_x.reset_dataset(io.Dataset(np.dtype(np.int_), [1]))
+        E_x[:] = np.array([43])
+        series_write.close()
+
+        # Now turn E_x into a scalar
+        with h5py.File(file, "r+") as f:
+            E = f["data"]["0"]["meshes"]["E"]
+            reapply_attributes = \
+                {key: val for key, val in E["x"].attrs.items()}
+            del E["x"]
+            E["x"] = 44
+            for key, val in reapply_attributes.items():
+                E["x"].attrs[key] = val
+
+        series_read = io.Series(file, io.Access.read_only)
+        loaded_from_scalar = series_read.iterations[0].meshes["E"]["x"][:]
+        series_read.flush()
+        self.assertEqual(loaded_from_scalar, np.array([44]))
+        series_read.close()
+
+        series_read_write = io.Series(file, io.Access.read_write)
+        E_x = series_read_write.iterations[0].meshes["E"]["x"]
+        E_x[:] = np.array([45])
+        series_read_write.close()
+
+        series_read_again = io.Series(file, io.Access.read_only)
+        loaded_from_scalar = \
+            series_read_again.iterations[0].meshes["E"]["x"][:]
+        series_read_again.flush()
+        self.assertEqual(loaded_from_scalar, np.array([45]))
+        series_read_again.close()
+
 
 if __name__ == '__main__':
     unittest.main()