unpack_from_memoryviews

python/pylibcudf/pylibcudf/contiguous_split.pxd

@@ -3,6 +3,7 @@
 from libcpp.memory cimport unique_ptr
 from pylibcudf.libcudf.contiguous_split cimport packed_columns
 
+from .gpumemoryview cimport gpumemoryview
 from .table cimport Table
 
 
@@ -15,3 +16,5 @@ cdef class PackedColumns:
 cpdef PackedColumns pack(Table input)
 
 cpdef Table unpack(PackedColumns input)
+
+cpdef Table unpack_from_memoryviews(memoryview metadata, gpumemoryview gpu_data)

python/pylibcudf/pylibcudf/contiguous_split.pyx

@@ -14,6 +14,7 @@ from pylibcudf.libcudf.table.table_view cimport table_view
 
 from .gpumemoryview cimport gpumemoryview
 from .table cimport Table
+from .utils cimport int_to_void_ptr
 
 from types import SimpleNamespace
 
@@ -123,5 +124,41 @@ cpdef Table unpack(PackedColumns input):
         Copy of the packed columns.
     """
     cdef table_view v = cpp_unpack(dereference(input.c_obj))
-    cdef unique_ptr[table] t = make_unique[table](v)
+    cdef unique_ptr[table] t = make_unique[table](v)  # Copy
+    return Table.from_libcudf(move(t))
+
+
+cpdef Table unpack_from_memoryviews(memoryview metadata, gpumemoryview gpu_data):
+    """Deserialize the result of `pack`.
+
+    Copies the result of a serialized table into a table.
+    Contrary to the libcudf C++ function, the returned table is a copy
+    of the serialized data.
+
+    For details, see :cpp:func:`cudf::unpack`.
+
+    Parameters
+    ----------
+    metadata : memoryview
+        The packed metadata to unpack.
+    gpu_data : gpumemoryview
+        The packed gpu_data to unpack.
+
+    Returns
+    -------
+    Table
+        Copy of the packed columns.
+    """
+    if metadata.nbytes == 0:
+        if gpu_data.__cuda_array_interface__["data"][0] != 0:
+            raise ValueError("expect an empty gpu_data when unpackking an empty table")
+        return Table.from_libcudf(make_unique[table](table_view()))
+
+    # Extract the raw data pointers
+    cdef const uint8_t[::1] _metadata = metadata
+    cdef const uint8_t* metadata_ptr = &_metadata[0]
+    cdef const uint8_t* gpu_data_ptr = <uint8_t*>int_to_void_ptr(gpu_data.ptr)
+
+    cdef table_view v = cpp_unpack(metadata_ptr, gpu_data_ptr)
+    cdef unique_ptr[table] t = make_unique[table](v)  # Copy
     return Table.from_libcudf(move(t))

python/pylibcudf/pylibcudf/libcudf/contiguous_split.pxd

@@ -26,3 +26,8 @@ cdef extern from "cudf/contiguous_split.hpp" namespace "cudf" nogil:
     cdef packed_columns pack (const table_view& input) except +
 
     cdef table_view unpack (const packed_columns& input) except +
+
+    cdef table_view unpack (
+        const uint8_t* metadata,
+        const uint8_t* gpu_data
+    ) except +

python/pylibcudf/pylibcudf/tests/test_contiguous_split.py

@@ -1,5 +1,6 @@
 # Copyright (c) 2024, NVIDIA CORPORATION.
 
+import cupy
 import pyarrow as pa
 import pylibcudf as plc
 import pytest
@@ -19,5 +20,13 @@
 def test_pack_and_unpack(arrow_tbl):
     plc_tbl = plc.interop.from_arrow(arrow_tbl)
     packed = plc.contiguous_split.pack(plc_tbl)
+
     res = plc.contiguous_split.unpack(packed)
     assert_table_eq(arrow_tbl, res)
+
+    # Copy the buffers to simulate IO
+    metadata = memoryview(bytes(packed.metadata))
+    gpu_data = plc.gpumemoryview(cupy.array(packed.gpu_data, copy=True))
+
+    res = plc.contiguous_split.unpack_from_memoryviews(metadata, gpu_data)
+    assert_table_eq(arrow_tbl, res)