Writing Neuroglancer Precomputed data with png compression not working correctly

[LarsKoeppel]

Hi,

i generated 2 versions of a dataset with tensorstore. And a 3erd version of the same data with cloud-volume.

The first version uses jpg compression and can be opened normaly:
jpg version

The second version uses png compression and the image is not fully loaded(border chunks missing):
png version

The 3rd version was created using cloud-volume(loads normaly):
cloud-volume

Neuroglancer is not able to show the chunks that overlap the border of the image from the png version made with tensorstore.
I verified that all chunks are loaded and that the naming of the cunks is identical. I also check this in the Firefox and Edge browser - same result.

I used this config for tensorstore to create this dataset. Did i miss something in the config?

tensorstore::Open({{"driver", "neuroglancer_precomputed"},
                                       {"kvstore", {
                                            {"driver", "file"},
                                            {"path", <path>}
                                        }},
                                       {"multiscale_metadata", {
                                            {"data_type", "uint8}, 
                                            {"num_channels", 1},
                                            {"type", "image"},
                                        }},
                                       {"scale_metadata", {
                                            {"resolution", {resolution[0], resolution[1], resolution[2]}},
                                            {"encoding", "png"},
                                            {"png_level", 6},
                                            {"chunk_size", {chunkSize[0], chunkSize[1], chunkSize[2]}},
                                            {"size", {sizeX, sizeY, sizeZ}},
                                        }},
                                       {"scale_index", index},
                                      },
                                      context,
                                      tensorstore::OpenMode::create,
                                      tensorstore::RecheckCached{false},
                                      tensorstore::ReadWriteMode::write).value();

I used this code to generate the cloud volume version. So the data is stored in the chunks but not in a format that neuroglancer can read.

>>> from cloudvolume import CloudVolume
>>> import tensorstore as ts
>>> dataset = ts.open(
...             {
...                 "driver": "neuroglancer_precomputed",
...                 "kvstore": {
...                     "driver": "file",
...                     "path": <path>,
...                 },
...             }
...         ).result()
>>> data = dataset[:,:,0].read().result()
>>> volume_path = 'file://<output_path>'
>>> info = CloudVolume.create_new_info(
...     num_channels=1,                    # number of channels
...     layer_type='image',                 # 'image' for image data
...     data_type='uint8',                  # data type of your NumPy array
...     encoding='png',                     # use PNG compression
...     resolution=[1, 1, 1],              # in nanometers
...     voxel_offset=[0, 0, 0],             # where the data starts
...     chunk_size=[1024, 1024, 1],             # size of each chunk in voxels
...     volume_size=data.shape,           # volume size in voxels
... )
>>> vol = CloudVolume(volume_path, info=info)
>>> vol.commit_info()  # write the metadata to disk
>>> vol[:,:,0] = data

I am happy to help with more information.