matthew-brett
diff --git a/‎numpy/core/tests/test_blasdot.py
Lines changed: 32 additions & 84 deletions b/‎numpy/core/tests/test_blasdot.py
Lines changed: 32 additions & 84 deletions
@@ -170,98 +170,46 @@ def __numpy_ufunc__(self, ufunc, method, pos, inputs, **kwargs):
     assert_equal(c.dot(a), "A")
     assert_raises(TypeError, np.dot, b, c)
     assert_raises(TypeError, c.dot, b)
-    
-    
+
+
 def test_npdot_segfault():
     if sys.platform != 'darwin': return
-    
     # Test for float32 np.dot segfault
     # https://github.com/numpy/numpy/issues/4007
-    
-    def aligned_array(N, align, dtype):
+
+    def aligned_array(shape, align, dtype, order='C'):
+        # Make array shape `shape` with aligned at `align` bytes
         d = dtype()
+        # Make array of correct size with `align` extra bytes
+        N = np.prod(shape)
         tmp = np.zeros(N * d.nbytes + align, dtype=np.uint8)
         address = tmp.__array_interface__["data"][0]
+        # Find offset into array giving desired alignment
         for offset in range(align):
             if (address + offset) % align == 0: break
-        return tmp[offset:offset+N*d.nbytes].view(dtype=dtype)
-
-    m = aligned_array(100,15,np.float32)
-    s = aligned_array(10000,15,np.float32).reshape(100,100)
-    np.dot(s,m) # This always segfaults when the sgemv alignment bug is present
-                                                 
-    # test the sanity of np.dot after applying patch
-    # misaligned arrays trigger SGEMM
-    m0 = np.random.rand(200)
-    s0 = np.random.rand(10000,200)
-    m1 = aligned_array(200,15,np.float32)
-    m1[:] = m0
-    s1 = aligned_array(10000*200,15,np.float32).reshape((10000,200))
-    s2 = aligned_array(10000*200,15,np.float32).reshape((10000,200),order='F')
-    s1[:,:] = s0
-    s2[:,:] = s0
-    desired = np.dot(s0,m0).astype(np.float32)
-    assert_allclose(np.dot(s1,m1),desired,atol=0.01) 
-    assert_allclose(np.dot(s2,m1),desired,atol=0.01)
-    
-    m0 = np.random.rand(200)
-    s0 = np.random.rand(200,10000)
-    m1 = aligned_array(200,15,np.float32)
-    m1[:] = m0
-    s1 = aligned_array(10000*200,15,np.float32).reshape((200,10000))
-    s2 = aligned_array(10000*200,15,np.float32).reshape((200,10000),order='F')
-    s1[:,:] = s0
-    s2[:,:] = s0
-    desired = np.dot(s0.T,m0).astype(np.float32)
-    assert_allclose(np.dot(s1.T,m1),desired,atol=0.01) 
-    assert_allclose(np.dot(s2.T,m1),desired,atol=0.01) 
-    
-    m0 = np.random.rand(89)
-    s0 = np.random.rand(10000,89)
-    m1 = aligned_array(89,15,np.float32)
-    m1[:] = m0
-    s1 = aligned_array(10000*89,15,np.float32).reshape((10000,89))
-    s2 = aligned_array(10000*89,15,np.float32).reshape((10000,89),order='F')
-    s1[:,:] = s0
-    s2[:,:] = s0
-    desired = np.dot(s0,m0).astype(np.float32)
-    assert_allclose(np.dot(s1,m1),desired,atol=0.01) 
-    assert_allclose(np.dot(s2,m1),desired,atol=0.01)
+        tmp = tmp[offset:offset+N*d.nbytes].view(dtype=dtype)
+        return tmp.reshape(shape, order=order)
 
+    m = aligned_array(100, 15, np.float32)
+    s = aligned_array((100, 100), 15, np.float32)
+    # This always segfaults when the sgemv alignment bug is present
+    np.dot(s, m)
     # test the sanity of np.dot after applying patch
-    # 32 byte aligned arrays trigger SGEMV
-    m0 = np.random.rand(200)
-    s0 = np.random.rand(10000,200)
-    m1 = aligned_array(200,32,np.float32)
-    m1[:] = m0
-    s1 = aligned_array(10000*200,32,np.float32).reshape((10000,200))
-    s2 = aligned_array(10000*200,32,np.float32).reshape((10000,200),order='F')
-    s1[:,:] = s0
-    s2[:,:] = s0
-    desired = np.dot(s0,m0).astype(np.float32)
-    assert_allclose(np.dot(s1,m1),desired,atol=0.01) 
-    assert_allclose(np.dot(s2,m1),desired,atol=0.01)
-    
-    m0 = np.random.rand(200)
-    s0 = np.random.rand(200,10000)
-    m1 = aligned_array(200,32,np.float32)
-    m1[:] = m0
-    s1 = aligned_array(10000*200,32,np.float32).reshape((200,10000))
-    s2 = aligned_array(10000*200,32,np.float32).reshape((200,10000),order='F')
-    s1[:,:] = s0
-    s2[:,:] = s0
-    desired = np.dot(s0.T,m0).astype(np.float32)
-    assert_allclose(np.dot(s1.T,m1),desired,atol=0.01) 
-    assert_allclose(np.dot(s2.T,m1),desired,atol=0.01) 
-    
-    m0 = np.random.rand(89)
-    s0 = np.random.rand(10000,89)
-    m1 = aligned_array(89,32,np.float32)
-    m1[:] = m0
-    s1 = aligned_array(10000*89,32,np.float32).reshape((10000,89))
-    s2 = aligned_array(10000*89,32,np.float32).reshape((10000,89),order='F')
-    s1[:,:] = s0
-    s2[:,:] = s0
-    desired = np.dot(s0,m0).astype(np.float32)
-    assert_allclose(np.dot(s1,m1),desired,atol=0.01) 
-    assert_allclose(np.dot(s2,m1),desired,atol=0.01)
+    for align in (15, 32):
+        for m, n, trans in ((10000, 200, False),
+                            (10000, 200, True),
+                            (10000, 89, True)):
+            m0 = np.random.rand(n)
+            s_shape = (n, m) if trans else (m, n)
+            s0 = np.random.rand(*s_shape)
+            m1 = aligned_array(n, align, np.float32)
+            m1[:] = m0
+            s1 = aligned_array(s_shape, align, np.float32)
+            s2 = aligned_array(s_shape, align, np.float32 ,order='F')
+            s1[:, :] = s0
+            s2[:, :] = s0
+            if trans:
+                s0, s1, s2 = s0.T, s1.T, s2.T
+            desired = np.dot(s0, m0).astype(np.float32)
+            assert_allclose(np.dot(s1, m1), desired, rtol=1e-5, atol=1e-5)
+            assert_allclose(np.dot(s2, m1), desired, rtol=1e-5, atol=1e-5)