Set default dtype of usm_ndarray depending on capabilities of device (#1265)

* Change default dtype value in usm_ndarray

* Use select_default_device() instead of SyclDevice()

Author: vlad-perevezentsev

dpctl/tensor/_usmarray.pyx

@@ -37,6 +37,7 @@ cimport dpctl.memory as c_dpmem
 cimport dpctl.tensor._dlpack as c_dlpack
 
 import dpctl.tensor._flags as _flags
+from dpctl.tensor._tensor_impl import default_device_fp_type
 
 include "_stride_utils.pxi"
 include "_types.pxi"
@@ -104,7 +105,7 @@ cdef class InternalUSMArrayError(Exception):
 
 
 cdef class usm_ndarray:
-    """ usm_ndarray(shape, dtype="|f8", strides=None, buffer="device", \
+    """ usm_ndarray(shape, dtype=None, strides=None, buffer="device", \
            offset=0, order="C", buffer_ctor_kwargs=dict(), \
            array_namespace=None)
 
@@ -116,6 +117,8 @@ cdef class usm_ndarray:
             Shape of the array to be created.
         dtype (str, dtype):
             Array data type, i.e. the type of array elements.
+            If ``dtype`` has the value ``None``, it is determined by default
+            floating point type supported by target device.
             The supported types are
                * ``bool``
                      boolean type
@@ -134,7 +137,7 @@ cdef class usm_ndarray:
                      double-precision real and complex floating
                      types, supported if target device's property
                      ``has_aspect_fp64`` is ``True``.
-            Default: ``"|f8"``.
+            Default: ``None``.
         strides (tuple, optional):
             Strides of the array to be created in elements.
             If ``strides`` has the value ``None``, it is determined by the
@@ -219,7 +222,7 @@ cdef class usm_ndarray:
                 "Data pointers of cloned and original objects are different.")
         return res
 
-    def __cinit__(self, shape, dtype="|f8", strides=None, buffer='device',
+    def __cinit__(self, shape, dtype=None, strides=None, buffer='device',
                   Py_ssize_t offset=0, order='C',
                   buffer_ctor_kwargs=dict(),
                   array_namespace=None):
@@ -252,6 +255,13 @@ cdef class usm_ndarray:
             except Exception:
                 raise TypeError("Argument shape must be a list or a tuple.")
         nd = len(shape)
+        if dtype is None:
+            q = buffer_ctor_kwargs.get("queue")
+            if q is not None:
+                dtype = default_device_fp_type(q)
+            else:
+                dev = dpctl.select_default_device()
+                dtype = "f8" if dev.has_aspect_fp64 else "f4"
         typenum = dtype_to_typenum(dtype)
         if (typenum < 0):
             if typenum == -2:

dpctl/tests/test_usm_ndarray_ctor.py

@@ -140,6 +140,26 @@ def test_dtypes(dtype):
     assert expected_fmt == actual_fmt
 
 
+@pytest.mark.parametrize("usm_type", ["device", "shared", "host"])
+@pytest.mark.parametrize("buffer_ctor_kwargs", [dict(), {"queue": None}])
+def test_default_dtype(usm_type, buffer_ctor_kwargs):
+    q = get_queue_or_skip()
+    dev = q.get_sycl_device()
+    if buffer_ctor_kwargs:
+        buffer_ctor_kwargs["queue"] = q
+    Xusm = dpt.usm_ndarray(
+        (1,), buffer=usm_type, buffer_ctor_kwargs=buffer_ctor_kwargs
+    )
+    if dev.has_aspect_fp64:
+        expected_dtype = "f8"
+    else:
+        expected_dtype = "f4"
+    assert Xusm.itemsize == dpt.dtype(expected_dtype).itemsize
+    expected_fmt = (dpt.dtype(expected_dtype).str)[1:]
+    actual_fmt = Xusm.__sycl_usm_array_interface__["typestr"][1:]
+    assert expected_fmt == actual_fmt
+
+
 @pytest.mark.parametrize(
     "dtype",
     [