HDFView cannot read attribute with variable length sequence data

HDF5 Library Library Development
1

I have another issue with HDFView. With my library I am able to create all kinds of attributes and h5dump reads all of them properly. HDFView can also open all attributes except the most complex one: A compound with some simple fields and three variable-length sequences. HDFView crashes with the following error:
hs_err_pid117345.log (148.3 KB)

The data I am trying to write as an attribute is a C# object called “DataspaceMessage”. So this is some kind of HDF5 inception where the attribute data itself is representing an internal HDF5 object message. Please do not get confused by this, it was just a good candidate to test my generic attribute writing algorithm. So please ignore the name “DataspaceMessage” and just treat it as an arbitrary compound type.

The output of h5dump of the same file is correct:

HDF5 "DataspaceMessage.h5" {
GROUP "/" {
   ATTRIBUTE "DataspaceMessage" {
      DATATYPE  H5T_COMPOUND {
         H5T_STD_U8LE "Rank";
         H5T_ENUM {
            H5T_STD_U8LE;
            "None"             0;
            "DimensionMaxSizes" 1;
            "PermuationIndices" 2;
         } "Flags";
         H5T_ENUM {
            H5T_STD_U8LE;
            "Scalar"           0;
            "Simple"           1;
            "Null"             2;
         } "Type";
         H5T_VLEN { H5T_STD_U64LE} "DimensionSizes";
         H5T_VLEN { H5T_STD_U64LE} "DimensionMaxSizes";
         H5T_VLEN { H5T_STD_U64LE} "PermutationIndices";
         H5T_STD_U8LE "Version";
      }
      DATASPACE  SCALAR
      DATA {
      (0): {
            1,
            DimensionMaxSizes,
            Simple,
            (10, 20, 30),
            (20, 40, 60),
            (40, 80, 120),
            1
         }
      }
   }
}
}

The output of myHDF5 is partially OK but it is ignoring the variable-length data:

{
  "name": "",
  "path": "/",
  "kind": "group",
  "attributes": [
    {
      "name": "DataspaceMessage",
      "shape": [],
      "type": {
        "class": "Compound",
        "fields": {
          "Rank": {
            "class": "Integer (unsigned)",
            "size": 8,
            "endianness": "little-endian"
          },
          "Flags": {
            "class": "Enumeration",
            "mapping": {
              "None": 0,
              "DimensionMaxSizes": 1,
              "PermuationIndices": 2
            },
            "base": {
              "class": "Integer (unsigned)",
              "size": 8,
              "endianness": "little-endian"
            }
          },
          "Type": {
            "class": "Enumeration",
            "mapping": {
              "Scalar": 0,
              "Simple": 1,
              "Null": 2
            },
            "base": {
              "class": "Integer (unsigned)",
              "size": 8,
              "endianness": "little-endian"
            }
          },
          "DimensionSizes": {
            "class": "Unknown"
          },
          "DimensionMaxSizes": {
            "class": "Unknown"
          },
          "PermutationIndices": {
            "class": "Unknown"
          },
          "Version": {
            "class": "Integer (unsigned)",
            "size": 8,
            "endianness": "little-endian"
          }
        }
      }
    }
  ]
}

Data:

[1,1,1,{"0":3,"1":0,"2":0,"3":0,"4":216,"5":171,"6":13,"7":0},{"0":3,"1":0,"2":0,"3":0,"4":248,"5":171,"6":13,"7":0},{"0":3,"1":0,"2":0,"3":0,"4":24,"5":172,"6":13,"7":0},1]

I have attached a test file (DataspaceMessage.h5 (5 KB)) which contains the problematic attribute.

I have no idea what is going wrong in HDFView but I guess it is not able to handle the variable length sequence data. Is there a way to increase the HDFView log level via command line parameters, i.e. without recompiling the application?

Thank you
Vincent

2

Yes, complex compound datatype translation between Java and C are still a work in progress - The code where it fails, in the JNI library of HDF5 is not unique to attributes, it would fail for both attributes and datasets. The translation is because of the need to use Java containers to express C containers and be able to translate both directions.
Improvements have been made since the 3.1 release.
Yes, there is a way to change the command line to use logging levels. There is a file in the hdfview repo: docs/USING_logging.txt that explains how.

3

Thanks for the explanation and for the hint to the USING_logging.text file. I have seen it before but thought the steps need to be applied on the source code. I will try to apply it on the distributed application tomorrow.