Improve Vectorisation

This MR fixes some bugs caused in the vectorisation

pystencils_tests/test_vectorization_specific.py

@@ -172,13 +172,16 @@ def test_square_root(dtype, instruction_set, field_layout):
 
 @pytest.mark.parametrize('dtype', ('float32', 'float64'))
 @pytest.mark.parametrize('instruction_set', supported_instruction_sets)
-def test_square_root_2(dtype, instruction_set):
+@pytest.mark.parametrize('padding', (True, False))
+def test_square_root_2(dtype, instruction_set, padding):
     x, y = sp.symbols("x y")
     src = ps.fields(f"src: {dtype}[2D]", layout='fzyx')
 
     up = ps.Assignment(src[0, 0], 1 / x + sp.sqrt(y * 0.52 + x ** 2))
 
-    cpu_vec = {'instruction_set': instruction_set, 'assume_inner_stride_one': True, 'assume_aligned': True}
+    cpu_vec = {'instruction_set': instruction_set, 'assume_inner_stride_one': True,
+               'assume_sufficient_line_padding': padding,
+               'assume_aligned': True}
 
     config = ps.CreateKernelConfig(data_type=dtype, default_number_float=dtype, cpu_vectorize_info=cpu_vec)
     ast = ps.create_kernel(up, config=config)
@@ -190,11 +193,13 @@ def test_square_root_2(dtype, instruction_set):
 
 @pytest.mark.parametrize('dtype', ('float32', 'float64'))
 @pytest.mark.parametrize('instruction_set', supported_instruction_sets)
-def test_pow(dtype, instruction_set):
+@pytest.mark.parametrize('padding', (True, False))
+def test_pow(dtype, instruction_set, padding):
     config = pystencils.config.CreateKernelConfig(data_type=dtype,
                                                   default_number_float=dtype,
                                                   cpu_vectorize_info={'instruction_set': instruction_set,
                                                                       'assume_inner_stride_one': True,
+                                                                      'assume_sufficient_line_padding': padding,
                                                                       'assume_aligned': False, 'nontemporal': False})
 
     src_field = ps.Field.create_generic('pdfs', 2, dtype, index_dimensions=1, layout='fzyx', index_shape=(9,))
@@ -211,8 +216,11 @@ def test_pow(dtype, instruction_set):
 
 @pytest.mark.parametrize('dtype', ('float32', 'float64'))
 @pytest.mark.parametrize('instruction_set', supported_instruction_sets)
-def test_issue62(dtype, instruction_set):
-    opt = {'instruction_set': instruction_set, 'assume_aligned': True, 'assume_inner_stride_one': True}
+@pytest.mark.parametrize('padding', (True, False))
+def test_issue62(dtype, instruction_set, padding):
+    opt = {'instruction_set': instruction_set, 'assume_aligned': True,
+           'assume_inner_stride_one': True,
+           'assume_sufficient_line_padding': padding}
 
     dx = sp.Symbol("dx")
     dy = sp.Symbol("dy")
@@ -233,3 +241,30 @@ def test_issue62(dtype, instruction_set):
     print(code)
 
     assert 'pow' not in code
+
+
+# @pytest.mark.parametrize('dtype', ('float32', 'float64'))
+@pytest.mark.parametrize('instruction_set', supported_instruction_sets)
+def test_div_and_unev(instruction_set):
+    dtype = 'float64'
+    opt = {'instruction_set': instruction_set, 'assume_aligned': True, 'assume_inner_stride_one': True,
+           'assume_sufficient_line_padding': False}
+
+    x, y, z = sp.symbols("x y z")
+    rhs = (-4 * x ** 4 * y ** 2 * z ** 2 + (x ** 4 * y ** 2 / 3) + (x ** 4 * z ** 2 / 3)) / x ** 3
+
+    src = ps.fields(f"src: {dtype}[2D]", layout='fzyx')
+
+    up = ps.Assignment(src[0, 0], rhs)
+
+    config = ps.CreateKernelConfig(data_type=dtype,
+                                   default_number_float=dtype,
+                                   cpu_vectorize_info=opt)
+
+    ast = ps.create_kernel(up, config=config)
+    ast.compile()
+    code = ps.get_code_str(ast)
+
+    print(code)
+
+    assert 'pow' not in code