Improve performance of host to device memory copy

cupy/core/core.pyx

@@ -3,13 +3,13 @@
 from __future__ import division
 import sys
 
+import ctypes
 import numpy
 import six
 
 import cupy
 from cupy.core import flags
 from cupy.cuda import device
-from cupy.cuda import stream
 
 try:
     from cupy.cuda import thrust
@@ -2120,6 +2120,7 @@ cpdef ndarray array(obj, dtype=None, bint copy=True, str order='K',
     # TODO(beam2d): Support subok options
     cdef Py_ssize_t nvidem
     cdef ndarray a, src
+    cdef size_t nbytes
     if subok:
         raise NotImplementedError
     if isinstance(obj, ndarray):
@@ -2139,24 +2140,26 @@ cpdef ndarray array(obj, dtype=None, bint copy=True, str order='K',
                 a = a.view()
             a.shape = (1,) * (ndmin - ndim) + a.shape
     else:
-        if order == 'K':
-            order = 'A'
+        if order is not None and len(order) == 1 and order in 'KAka':
+            if isinstance(obj, numpy.ndarray) and obj.flags.f_contiguous:
+                order = 'F'
+            else:
+                order = 'C'
         a_cpu = numpy.array(obj, dtype=dtype, copy=False, order=order,
                             ndmin=ndmin)
-        order = 'C' if a_cpu.flags.c_contiguous else 'F'
         a_dtype = a_cpu.dtype
         if a_dtype.char not in '?bhilqBHILQefdFD':
             raise ValueError('Unsupported dtype %s' % a_dtype)
         a = ndarray(a_cpu.shape, dtype=a_dtype, order=order)
         if a_cpu.ndim == 0:
             a.fill(a_cpu[()])
             return a
-        mem = pinned_memory.alloc_pinned_memory(a.nbytes)
-        src_cpu = numpy.frombuffer(mem, a_cpu.dtype,
-                                   a_cpu.size).reshape(a_cpu.shape)
-        src_cpu[...] = a_cpu
+        nbytes = a.nbytes
+        mem = pinned_memory.alloc_pinned_memory(nbytes)
+        src_cpu = numpy.frombuffer(mem, a_dtype, a_cpu.size)
+        src_cpu[:] = a_cpu.ravel(order)
         stream = stream_module.get_current_stream()
-        a.set(src_cpu, stream)
+        a.data.copy_from_host_async(ctypes.c_void_p(mem.ptr), nbytes)
         pinned_memory._add_to_watch_list(stream.record(), mem)
     return a
 

cupy/logic/type_test.py

@@ -86,7 +86,7 @@ def isfortran(a):
     >>> cupy.isfortran(a)
     False
 
-    >>> b = cupy.array([[1, 2, 3], [4, 5, 6]], order='FORTRAN')
+    >>> b = cupy.array([[1, 2, 3], [4, 5, 6]], order='F')
     >>> b
     array([[1, 2, 3],
            [4, 5, 6]])
@@ -111,7 +111,7 @@ def isfortran(a):
 
     C-ordered arrays evaluate as False even if they are also FORTRAN-ordered.
 
-    >>> cupy.isfortran(np.array([1, 2], order='FORTRAN'))
+    >>> cupy.isfortran(np.array([1, 2], order='F'))
     False
 
     """

tests/cupy_tests/creation_tests/test_from_data.py

@@ -22,6 +22,15 @@ def test_array_from_numpy(self, xp, dtype, order):
         a = testing.shaped_arange((2, 3, 4), numpy, dtype)
         return xp.array(a, order=order)
 
+    @testing.for_orders('CFAK')
+    @testing.for_all_dtypes()
+    @testing.with_requires('numpy>=1.10')
+    @testing.numpy_cupy_array_equal()
+    def test_array_from_numpy_broad_cast(self, xp, dtype, order):
+        a = testing.shaped_arange((2, 1, 4), numpy, dtype)
+        a = numpy.broadcast_to(a, (2, 3, 4))
+        return xp.array(a, order=order)
+
     @testing.for_orders('CFAK')
     @testing.for_all_dtypes()
     @testing.numpy_cupy_array_equal()