release/cuda core v0.5.1

.github/workflows/ci.yml

@@ -15,6 +15,7 @@ on:
     branches:
       - "pull-request/[0-9]+"
       - "main"
+      - "release/*"
 
 jobs:
   ci-vars:

cuda_core/cuda/core/_layout.pxd

@@ -111,7 +111,8 @@ cdef class _StridedLayout:
     # ==============================
 
     cdef inline int _init(_StridedLayout self, BaseLayout& base, int itemsize, bint divide_strides=False) except -1 nogil:
-        _validate_itemsize(itemsize)
+        if itemsize <= 0:
+            raise ValueError("itemsize must be positive")
 
         if base.strides != NULL and divide_strides:
             _divide_strides(base, itemsize)
@@ -123,7 +124,8 @@ cdef class _StridedLayout:
         return 0
 
     cdef inline stride_t _init_dense(_StridedLayout self, BaseLayout& base, int itemsize, OrderFlag order_flag, axis_vec_t* stride_order=NULL) except -1 nogil:
-        _validate_itemsize(itemsize)
+        if itemsize <= 0:
+            raise ValueError("itemsize must be positive")
 
         cdef stride_t volume
         if order_flag == ORDER_C:
@@ -643,14 +645,6 @@ cdef inline bint _normalize_axis(integer_t& axis, integer_t extent) except -1 no
     return True
 
 
-cdef inline int _validate_itemsize(int itemsize) except -1 nogil:
-    if itemsize <= 0:
-        raise ValueError("itemsize must be positive")
-    if itemsize & (itemsize - 1):
-        raise ValueError("itemsize must be a power of two")
-    return 0
-
-
 cdef inline bint _is_unique(BaseLayout& base, axis_vec_t& stride_order) except -1 nogil:
     if base.strides == NULL:
         return True

cuda_core/cuda/core/_layout.pyx

@@ -29,7 +29,7 @@ cdef class _StridedLayout:
         Otherwise, the strides are assumed to be implicitly C-contiguous and the resulting
         layout's :attr:`strides` will be None.
     itemsize : int
-        The number of bytes per single element (dtype size). Must be a power of two.
+        The number of bytes per single element (dtype size).
     divide_strides : bool, optional
         If True, the provided :attr:`strides` will be divided by the :attr:`itemsize`.
 
@@ -40,7 +40,7 @@ cdef class _StridedLayout:
     Attributes
     ----------
     itemsize : int
-        The number of bytes per single element (dtype size). Must be a power of two.
+        The number of bytes per single element (dtype size).
     slice_offset : int
         The offset (as a number of elements, not bytes) of the element at
         index ``(0,) * ndim``. See also :attr:`slice_offset_in_bytes`.
@@ -636,7 +636,6 @@ cdef class _StridedLayout:
         In either case, the ``volume * itemsize`` of the layout remains the same.
 
         The conversion is subject to the following constraints:
-            * The old and new itemsizes must be powers of two.
             * The extent at ``axis`` must be a positive integer.
             * The stride at ``axis`` must be 1.
 
@@ -1214,10 +1213,10 @@ cdef inline int64_t gcd(int64_t a, int64_t b) except? -1 nogil:
 
 cdef inline int pack_extents(BaseLayout& out_layout, stride_t& out_slice_offset, BaseLayout& in_layout, stride_t slice_offset, int itemsize, int new_itemsize, intptr_t data_ptr, bint keep_dim, int axis) except -1 nogil:
     cdef int ndim = in_layout.ndim
-    if new_itemsize <= 0 or new_itemsize & (new_itemsize - 1):
-        raise ValueError(f"new itemsize must be a power of two, got {new_itemsize}.")
-    if itemsize <= 0 or itemsize & (itemsize - 1):
-        raise ValueError(f"itemsize must be a power of two, got {itemsize}.")
+    if new_itemsize <= 0:
+        raise ValueError(f"new itemsize must be greater than zero, got {new_itemsize}.")
+    if itemsize <= 0:
+        raise ValueError(f"itemsize must be greater than zero, got {itemsize}.")
     if new_itemsize <= itemsize:
         if new_itemsize == itemsize:
             return 1
@@ -1270,10 +1269,10 @@ cdef inline int unpack_extents(BaseLayout &out_layout, BaseLayout &in_layout, in
     cdef int ndim = in_layout.ndim
     if not _normalize_axis(axis, ndim):
         raise ValueError(f"Invalid axis: {axis} out of range for {ndim}D tensor")
-    if new_itemsize <= 0 or new_itemsize & (new_itemsize - 1):
-        raise ValueError(f"new itemsize must be a power of two, got {new_itemsize}.")
-    if itemsize <= 0 or itemsize & (itemsize - 1):
-        raise ValueError(f"itemsize must be a power of two, got {itemsize}.")
+    if new_itemsize <= 0:
+        raise ValueError(f"new itemsize must be greater than zero, got {new_itemsize}.")
+    if itemsize <= 0:
+        raise ValueError(f"itemsize must be greater than zero, got {itemsize}.")
     if new_itemsize >= itemsize:
         if new_itemsize == itemsize:
             return 1
@@ -1301,10 +1300,10 @@ cdef inline int unpack_extents(BaseLayout &out_layout, BaseLayout &in_layout, in
 
 cdef inline int max_compatible_itemsize(BaseLayout& layout, stride_t slice_offset, int itemsize, int max_itemsize, intptr_t data_ptr, int axis) except? -1 nogil:
     cdef int ndim = layout.ndim
-    if max_itemsize <= 0 or max_itemsize & (max_itemsize - 1):
-        raise ValueError(f"max_itemsize must be a power of two, got {max_itemsize}.")
-    if itemsize <= 0 or itemsize & (itemsize - 1):
-        raise ValueError(f"itemsize must be a power of two, got {itemsize}.")
+    if max_itemsize <= 0:
+        raise ValueError(f"max_itemsize must be greater than zero, got {max_itemsize}.")
+    if itemsize <= 0:
+        raise ValueError(f"itemsize must be greater than zero, got {itemsize}.")
     if not _normalize_axis(axis, ndim):
         raise ValueError(f"Invalid axis: {axis} out of range for {ndim}D tensor")
     if max_itemsize < itemsize:

cuda_core/cuda/core/_memoryview.pyx

@@ -137,22 +137,22 @@ cdef class StridedMemoryView:
 
     @classmethod
     def from_dlpack(cls, obj: object, stream_ptr: int | None=None) -> StridedMemoryView:
-        cdef StridedMemoryView buf
-        with warnings.catch_warnings():
-            warnings.simplefilter("ignore")
-            buf = cls()
+        cdef StridedMemoryView buf = StridedMemoryView.__new__(cls)
         view_as_dlpack(obj, stream_ptr, buf)
         return buf
 
     @classmethod
     def from_cuda_array_interface(cls, obj: object, stream_ptr: int | None=None) -> StridedMemoryView:
-        cdef StridedMemoryView buf
-        with warnings.catch_warnings():
-            warnings.simplefilter("ignore")
-            buf = cls()
+        cdef StridedMemoryView buf = StridedMemoryView.__new__(cls)
         view_as_cai(obj, stream_ptr, buf)
         return buf
 
+    @classmethod
+    def from_array_interface(cls, obj: object) -> StridedMemoryView:
+        cdef StridedMemoryView buf = StridedMemoryView.__new__(cls)
+        view_as_array_interface(obj, buf)
+        return buf
+
     @classmethod
     def from_any_interface(cls, obj: object, stream_ptr: int | None = None) -> StridedMemoryView:
         if check_has_dlpack(obj):
@@ -365,8 +365,7 @@ cdef class StridedMemoryView:
             if self.dl_tensor != NULL:
                 self._dtype = dtype_dlpack_to_numpy(&self.dl_tensor.dtype)
             elif self.metadata is not None:
-                # TODO: this only works for built-in numeric types
-                self._dtype = _typestr2dtype[self.metadata["typestr"]]
+                self._dtype = _typestr2dtype(self.metadata["typestr"])
         return self._dtype
 
 
@@ -486,25 +485,14 @@ cdef StridedMemoryView view_as_dlpack(obj, stream_ptr, view=None):
     return buf
 
 
-_builtin_numeric_dtypes = [
-    numpy.dtype("uint8"),
-    numpy.dtype("uint16"),
-    numpy.dtype("uint32"),
-    numpy.dtype("uint64"),
-    numpy.dtype("int8"),
-    numpy.dtype("int16"),
-    numpy.dtype("int32"),
-    numpy.dtype("int64"),
-    numpy.dtype("float16"),
-    numpy.dtype("float32"),
-    numpy.dtype("float64"),
-    numpy.dtype("complex64"),
-    numpy.dtype("complex128"),
-    numpy.dtype("bool"),
-]
-# Doing it once to avoid repeated overhead
-_typestr2dtype = {dtype.str: dtype for dtype in _builtin_numeric_dtypes}
-_typestr2itemsize = {dtype.str: dtype.itemsize for dtype in _builtin_numeric_dtypes}
+@functools.lru_cache
+def _typestr2dtype(str typestr):
+    return numpy.dtype(typestr)
+
+
+@functools.lru_cache
+def _typestr2itemsize(str typestr):
+    return _typestr2dtype(typestr).itemsize
 
 
 cdef object dtype_dlpack_to_numpy(DLDataType* dtype):
@@ -609,6 +597,23 @@ cpdef StridedMemoryView view_as_cai(obj, stream_ptr, view=None):
     return buf
 
 
+cpdef StridedMemoryView view_as_array_interface(obj, view=None):
+    cdef dict data = obj.__array_interface__
+    if data["version"] < 3:
+        raise BufferError("only NumPy Array Interface v3 or above is supported")
+    if data.get("mask") is not None:
+        raise BufferError("mask is not supported")
+
+    cdef StridedMemoryView buf = StridedMemoryView() if view is None else view
+    buf.exporting_obj = obj
+    buf.metadata = data
+    buf.dl_tensor = NULL
+    buf.ptr, buf.readonly = data["data"]
+    buf.is_device_accessible = False
+    buf.device_id = handle_return(driver.cuCtxGetDevice())
+    return buf
+
+
 def args_viewable_as_strided_memory(tuple arg_indices):
     """
     Decorator to create proxy objects to :obj:`StridedMemoryView` for the
@@ -664,7 +669,7 @@ cdef _StridedLayout layout_from_cai(object metadata):
     cdef _StridedLayout layout = _StridedLayout.__new__(_StridedLayout)
     cdef object shape = metadata["shape"]
     cdef object strides = metadata.get("strides")
-    cdef int itemsize = _typestr2itemsize[metadata["typestr"]]
+    cdef int itemsize = _typestr2itemsize(metadata["typestr"])
     layout.init_from_tuple(shape, strides, itemsize, True)
     return layout
 

cuda_core/cuda/core/_version.py

@@ -2,4 +2,4 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
-__version__ = "0.5.0"
+__version__ = "0.5.1"

cuda_core/docs/nv-versions.json

@@ -3,6 +3,10 @@
         "version": "latest",
         "url": "https://nvidia.github.io/cuda-python/cuda-core/latest/"
     },
+    {
+        "version": "0.5.1",
+        "url": "https://nvidia.github.io/cuda-python/cuda-core/0.5.1/"
+    },
     {
         "version": "0.5.0",
         "url": "https://nvidia.github.io/cuda-python/cuda-core/0.5.0/"

cuda_core/pixi.toml

@@ -68,7 +68,7 @@ cu12 = { features = ["cu12", "test", "cython-tests"], solve-group = "cu12" }
 # TODO: check if these can be extracted from pyproject.toml
 [package]
 name = "cuda-core"
-version = "0.5.0"
+version = "0.5.1"
 
 [package.build]
 backend = { name = "pixi-build-python", version = "*" }

cuda_core/tests/conftest.py

@@ -72,7 +72,7 @@ def init_cuda():
             driver.cuDevicePrimaryCtxSetFlags(device.device_id, driver.CUctx_flags.CU_CTX_SCHED_BLOCKING_SYNC)
         )
 
-    yield
+    yield device
     _ = _device_unset_current()