fix chunking bug identified by @jakebolewski

src/JLD.jl

@@ -7,7 +7,7 @@ using HDF5
 # Add methods to...
 import HDF5: close, dump, exists, file, getindex, setindex!, g_create, g_open, o_delete, name, names, read, size, write,
              HDF5ReferenceObj, HDF5BitsKind, ismmappable, readmmap
-import Base: length, endof, show, done, next, start, delete!
+import Base: length, endof, show, done, next, start, delete!, sizeof
 
 # .jld files written before v"0.4.0-dev+1419" might have Uint32 instead of UInt32 as the typename string.
 # See julia issue #8907
@@ -37,6 +37,7 @@ immutable JldDatatype
     dtype::HDF5Datatype
     index::Int
 end
+sizeof(T::JldDatatype) = sizeof(T.dtype)
 
 immutable JldWriteSession
     persist::Vector{Any} # To hold objects that should not be garbage-collected
@@ -471,26 +472,24 @@ write(parent::Union(JldFile, JldGroup), name::ByteString,
 # Pick whether to use compact or default storage based on data size
 const COMPACT_PROPERTIES = p_create(HDF5.H5P_DATASET_CREATE)
 HDF5.h5p_set_layout(COMPACT_PROPERTIES.id, HDF5.H5D_COMPACT)
-function dset_create_properties(parent, sz::Int, obj)
+function dset_create_properties(parent, sz::Int, obj, chunk=Int[])
     sz <= 8192 && return COMPACT_PROPERTIES
-    if iscompressed(parent)
-        chunk = HDF5.heuristic_chunk(obj)
-        if !isempty(chunk)
-            p = p_create(HDF5.H5P_DATASET_CREATE)
-            p["chunk"] = chunk
-            p["blosc"] = 5
-            return p
-        end
+    if iscompressed(parent) && !isempty(chunk)
+        p = p_create(HDF5.H5P_DATASET_CREATE)
+        p["chunk"] = chunk
+        p["blosc"] = 5
+        return p
+    else
+        return HDF5.DEFAULT_PROPERTIES
     end
-    return HDF5.DEFAULT_PROPERTIES
 end
 
 # Write "basic" types
 function _write{T<:Union(HDF5BitsKind, ByteString)}(parent::Union(JldFile, JldGroup),
                                                     name::ByteString,
                                                     data::Union(T, Array{T}),
                                                     wsession::JldWriteSession)
-    dset, dtype = d_create(parent.plain, bytestring(name), data, HDF5._link_properties(name), dset_create_properties(parent, sizeof(data), data))
+    dset, dtype = d_create(parent.plain, bytestring(name), data, HDF5._link_properties(name), dset_create_properties(parent, sizeof(data), data, HDF5.heuristic_chunk(data)))
     try
         # Write the attribute
         isa(data, Array) && isempty(data) && a_write(dset, "dims", [size(data)...])
@@ -511,8 +510,9 @@ function _write{T}(parent::Union(JldFile, JldGroup),
     buf = h5convert_array(f, data, dtype, wsession)
     dims = convert(Array{HDF5.Hsize, 1}, [reverse(size(data))...])
     dspace = dataspace(data)
+    chunk = HDF5.heuristic_chunk(dtype, size(data))
     try
-        dset = d_create(parent.plain, path, dtype.dtype, dspace, HDF5._link_properties(path), dset_create_properties(parent, sizeof(buf), buf))
+        dset = d_create(parent.plain, path, dtype.dtype, dspace, HDF5._link_properties(path), dset_create_properties(parent, sizeof(buf), buf, chunk))
         if dtype == JLD_REF_TYPE
             a_write(dset, "julia eltype", full_typename(f, T))
         end

src/plain.jl

@@ -515,7 +515,7 @@ include("blosc_filter.jl")
 register_blosc()
 
 # heuristic chunk layout (return empty array to disable chunking)
-function heuristic_chunk(T::Type, shape)
+function heuristic_chunk(T, shape)
     Ts = sizeof(T)
     sz = prod(shape)
     sz == 0 && return Int[] # never return a zero-size chunk

test/jld.jl

@@ -184,6 +184,7 @@ nonearr = Array(Union(), 5)
 
 # some data big enough to ensure that compression is used:
 Abig = kron(eye(10), rand(20,20))
+Bbig = Any[i for i=1:3000]
 Sbig = "A test string "^1000
 
 iseq(x,y) = isequal(x,y)
@@ -337,6 +338,7 @@ for compress in (true,false)
     @write fid none
     @write fid nonearr
     @write fid Abig
+    @write fid Bbig
     @write fid Sbig
     # Make sure we can create groups (i.e., use HDF5 features)
     g = g_create(fid, "mygroup")
@@ -449,6 +451,7 @@ for compress in (true,false)
         @check fidr none
         @check fidr nonearr
         @check fidr Abig
+        @check fidr Bbig
         @check fidr Sbig
 
         x1 = read(fidr, "group1/x")