kernelcreation.py 13.8 KB
Newer Older
Martin Bauer's avatar
Martin Bauer committed
1
from types import MappingProxyType
2
import sympy as sp
3
import itertools
Martin Bauer's avatar
Martin Bauer committed
4
5
from pystencils.assignment import Assignment
from pystencils.astnodes import LoopOverCoordinate, Conditional, Block, SympyAssignment
Martin Bauer's avatar
Martin Bauer committed
6
from pystencils.cpu.vectorization import vectorize
Martin Bauer's avatar
Martin Bauer committed
7
8
from pystencils.simp.assignment_collection import AssignmentCollection
from pystencils.gpucuda.indexing import indexing_creator_from_params
9
10
from pystencils.transformations import remove_conditionals_in_staggered_kernel, loop_blocking, \
    move_constants_before_loop
Martin Bauer's avatar
Martin Bauer committed
11
12


Martin Bauer's avatar
Martin Bauer committed
13
def create_kernel(assignments, target='cpu', data_type="double", iteration_slice=None, ghost_layers=None,
14
                  skip_independence_check=False,
Martin Bauer's avatar
Martin Bauer committed
15
                  cpu_openmp=False, cpu_vectorize_info=None, cpu_blocking=None,
Martin Bauer's avatar
Martin Bauer committed
16
                  gpu_indexing='block', gpu_indexing_params=MappingProxyType({})):
Martin Bauer's avatar
Martin Bauer committed
17
18
    """
    Creates abstract syntax tree (AST) of kernel, using a list of update equations.
19
20

    Args:
Martin Bauer's avatar
Martin Bauer committed
21
        assignments: can be a single assignment, sequence of assignments or an `AssignmentCollection`
22
23
24
25
26
27
28
        target: 'cpu', 'llvm' or 'gpu'
        data_type: data type used for all untyped symbols (i.e. non-fields), can also be a dict from symbol name
                  to type
        iteration_slice: rectangular subset to iterate over, if not specified the complete non-ghost layer \
                         part of the field is iterated over
        ghost_layers: if left to default, the number of necessary ghost layers is determined automatically
                     a single integer specifies the ghost layer count at all borders, can also be a sequence of
Martin Bauer's avatar
Martin Bauer committed
29
                     pairs ``[(x_lower_gl, x_upper_gl), .... ]``
30
31
        skip_independence_check: don't check that loop iterations are independent. This is needed e.g. for
                                 periodicity kernel, that access the field outside the iteration bounds. Use with care!
32
        cpu_openmp: True or number of threads for OpenMP parallelization, False for no OpenMP
Martin Bauer's avatar
Martin Bauer committed
33
34
        cpu_vectorize_info: a dictionary with keys, 'vector_instruction_set', 'assume_aligned' and 'nontemporal'
                            for documentation of these parameters see vectorize function. Example:
35
                            '{'instruction_set': 'avx512', 'assume_aligned': True, 'nontemporal':True}'
Martin Bauer's avatar
Martin Bauer committed
36
        cpu_blocking: a tuple of block sizes or None if no blocking should be applied
Martin Bauer's avatar
Martin Bauer committed
37
        gpu_indexing: either 'block' or 'line' , or custom indexing class, see `AbstractIndexing`
38
        gpu_indexing_params: dict with indexing parameters (constructor parameters of indexing class)
Martin Bauer's avatar
Martin Bauer committed
39
                             e.g. for 'block' one can specify '{'block_size': (20, 20, 10) }'
40
41

    Returns:
Martin Bauer's avatar
Martin Bauer committed
42
        abstract syntax tree (AST) object, that can either be printed as source code with `show_code` or
43
        can be compiled with through its 'compile()' member
Martin Bauer's avatar
Martin Bauer committed
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59

    Example:
        >>> import pystencils as ps
        >>> import numpy as np
        >>> s, d = ps.fields('s, d: [2D]')
        >>> assignment = ps.Assignment(d[0,0], s[0, 1] + s[0, -1] + s[1, 0] + s[-1, 0])
        >>> ast = ps.create_kernel(assignment, target='cpu', cpu_openmp=True)
        >>> kernel = ast.compile()
        >>> d_arr = np.zeros([5, 5])
        >>> kernel(d=d_arr, s=np.ones([5, 5]))
        >>> d_arr
        array([[0., 0., 0., 0., 0.],
               [0., 4., 4., 4., 0.],
               [0., 4., 4., 4., 0.],
               [0., 4., 4., 4., 0.],
               [0., 0., 0., 0., 0.]])
Martin Bauer's avatar
Martin Bauer committed
60
61
    """
    # ----  Normalizing parameters
Martin Bauer's avatar
Martin Bauer committed
62
    split_groups = ()
Martin Bauer's avatar
Martin Bauer committed
63
64
65
66
67
68
    if isinstance(assignments, AssignmentCollection):
        if 'split_groups' in assignments.simplification_hints:
            split_groups = assignments.simplification_hints['split_groups']
        assignments = assignments.all_assignments
    if isinstance(assignments, Assignment):
        assignments = [assignments]
Martin Bauer's avatar
Martin Bauer committed
69
70
71

    # ----  Creating ast
    if target == 'cpu':
Martin Bauer's avatar
Martin Bauer committed
72
73
        from pystencils.cpu import create_kernel
        from pystencils.cpu import add_openmp
Martin Bauer's avatar
Martin Bauer committed
74
        ast = create_kernel(assignments, type_info=data_type, split_groups=split_groups,
75
76
                            iteration_slice=iteration_slice, ghost_layers=ghost_layers,
                            skip_independence_check=skip_independence_check)
Martin Bauer's avatar
Martin Bauer committed
77
78
79
        omp_collapse = None
        if cpu_blocking:
            omp_collapse = loop_blocking(ast, cpu_blocking)
Martin Bauer's avatar
Martin Bauer committed
80
        if cpu_openmp:
Martin Bauer's avatar
Martin Bauer committed
81
            add_openmp(ast, num_threads=cpu_openmp, collapse=omp_collapse)
Martin Bauer's avatar
Martin Bauer committed
82
        if cpu_vectorize_info:
Martin Bauer's avatar
Martin Bauer committed
83
            if cpu_vectorize_info is True:
84
                vectorize(ast)
Martin Bauer's avatar
Martin Bauer committed
85
86
87
88
            elif isinstance(cpu_vectorize_info, dict):
                vectorize(ast, **cpu_vectorize_info)
            else:
                raise ValueError("Invalid value for cpu_vectorize_info")
Martin Bauer's avatar
Martin Bauer committed
89
90
        return ast
    elif target == 'llvm':
Martin Bauer's avatar
Martin Bauer committed
91
        from pystencils.llvm import create_kernel
Martin Bauer's avatar
Martin Bauer committed
92
        ast = create_kernel(assignments, type_info=data_type, split_groups=split_groups,
Martin Bauer's avatar
Martin Bauer committed
93
                            iteration_slice=iteration_slice, ghost_layers=ghost_layers)
Martin Bauer's avatar
Martin Bauer committed
94
95
        return ast
    elif target == 'gpu':
Martin Bauer's avatar
Martin Bauer committed
96
        from pystencils.gpucuda import create_cuda_kernel
Martin Bauer's avatar
Martin Bauer committed
97
        ast = create_cuda_kernel(assignments, type_info=data_type,
Martin Bauer's avatar
Martin Bauer committed
98
                                 indexing_creator=indexing_creator_from_params(gpu_indexing, gpu_indexing_params),
99
100
                                 iteration_slice=iteration_slice, ghost_layers=ghost_layers,
                                 skip_independence_check=skip_independence_check)
Martin Bauer's avatar
Martin Bauer committed
101
102
103
104
105
        return ast
    else:
        raise ValueError("Unknown target %s. Has to be one of 'cpu', 'gpu' or 'llvm' " % (target,))


Martin Bauer's avatar
Martin Bauer committed
106
def create_indexed_kernel(assignments, index_fields, target='cpu', data_type="double", coordinate_names=('x', 'y', 'z'),
Martin Bauer's avatar
Martin Bauer committed
107
                          cpu_openmp=True, gpu_indexing='block', gpu_indexing_params=MappingProxyType({})):
Martin Bauer's avatar
Martin Bauer committed
108
    """
Martin Bauer's avatar
Martin Bauer committed
109
    Similar to :func:`create_kernel`, but here not all cells of a field are updated but only cells with
Martin Bauer's avatar
Martin Bauer committed
110
111
    coordinates which are stored in an index field. This traversal method can e.g. be used for boundary handling.

Martin Bauer's avatar
Martin Bauer committed
112
    The coordinates are stored in a separated index_field, which is a one dimensional array with struct data type.
Martin Bauer's avatar
Martin Bauer committed
113
    This struct has to contain fields named 'x', 'y' and for 3D fields ('z'). These names are configurable with the
Martin Bauer's avatar
Martin Bauer committed
114
    'coordinate_names' parameter. The struct can have also other fields that can be read and written in the kernel, for
Martin Bauer's avatar
Martin Bauer committed
115
116
    example boundary parameters.

Martin Bauer's avatar
Martin Bauer committed
117
118
    index_fields: list of index fields, i.e. 1D fields with struct data type
    coordinate_names: name of the coordinate fields in the struct data type
Martin Bauer's avatar
Martin Bauer committed
119

Martin Bauer's avatar
Martin Bauer committed
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
    Example:
        >>> import pystencils as ps
        >>> import numpy as np
        >>>
        >>> # Index field stores the indices of the cell to visit together with optional values
        >>> index_arr_dtype = np.dtype([('x', np.int32), ('y', np.int32), ('val', np.double)])
        >>> index_arr = np.array([(1, 1, 0.1), (2, 2, 0.2), (3, 3, 0.3)], dtype=index_arr_dtype)
        >>> idx_field = ps.fields(idx=index_arr)
        >>>
        >>> # Additional values  stored in index field can be accessed in the kernel as well
        >>> s, d = ps.fields('s, d: [2D]')
        >>> assignment = ps.Assignment(d[0,0], 2 * s[0, 1] + 2 * s[1, 0] + idx_field('val'))
        >>> ast = create_indexed_kernel(assignment, [idx_field], coordinate_names=('x', 'y'))
        >>> kernel = ast.compile()
        >>> d_arr = np.zeros([5, 5])
        >>> kernel(s=np.ones([5, 5]), d=d_arr, idx=index_arr)
        >>> d_arr
        array([[0. , 0. , 0. , 0. , 0. ],
               [0. , 4.1, 0. , 0. , 0. ],
               [0. , 0. , 4.2, 0. , 0. ],
               [0. , 0. , 0. , 4.3, 0. ],
               [0. , 0. , 0. , 0. , 0. ]])
    """
    if isinstance(assignments, Assignment):
        assignments = [assignments]
    elif isinstance(assignments, AssignmentCollection):
Martin Bauer's avatar
Martin Bauer committed
146
        assignments = assignments.all_assignments
Martin Bauer's avatar
Martin Bauer committed
147
    if target == 'cpu':
Martin Bauer's avatar
Martin Bauer committed
148
149
150
151
152
153
        from pystencils.cpu import create_indexed_kernel
        from pystencils.cpu import add_openmp
        ast = create_indexed_kernel(assignments, index_fields=index_fields, type_info=data_type,
                                    coordinate_names=coordinate_names)
        if cpu_openmp:
            add_openmp(ast, num_threads=cpu_openmp)
Martin Bauer's avatar
Martin Bauer committed
154
155
156
157
        return ast
    elif target == 'llvm':
        raise NotImplementedError("Indexed kernels are not yet supported in LLVM backend")
    elif target == 'gpu':
Martin Bauer's avatar
Martin Bauer committed
158
        from pystencils.gpucuda import created_indexed_cuda_kernel
Martin Bauer's avatar
Martin Bauer committed
159
160
161
        idx_creator = indexing_creator_from_params(gpu_indexing, gpu_indexing_params)
        ast = created_indexed_cuda_kernel(assignments, index_fields, type_info=data_type,
                                          coordinate_names=coordinate_names, indexing_creator=idx_creator)
Martin Bauer's avatar
Martin Bauer committed
162
163
164
        return ast
    else:
        raise ValueError("Unknown target %s. Has to be either 'cpu' or 'gpu'" % (target,))
165

166

167
168
def create_staggered_kernel(staggered_field, expressions, subexpressions=(), target='cpu',
                            gpu_exclusive_conditions=False, **kwargs):
169
170
    """Kernel that updates a staggered field.

Martin Bauer's avatar
Martin Bauer committed
171
172
    .. image:: /img/staggered_grid.svg

173
    Args:
174
        staggered_field: field where the first index coordinate defines the location of the staggered value
175
176
                can have 1 or 2 index coordinates, in case of two index coordinates at every staggered location
                a vector is stored, expressions parameter has to be a sequence of sequences then
Martin Bauer's avatar
Martin Bauer committed
177
178
                where e.g. ``f[0,0](0)`` is interpreted as value at the left cell boundary, ``f[1,0](0)`` the right cell
                boundary and ``f[0,0](1)`` the southern cell boundary etc.
179
        expressions: sequence of expressions of length dim, defining how the west, southern, (bottom) cell boundary
180
                     should be updated.
181
182
        subexpressions: optional sequence of Assignments, that define subexpressions used in the main expressions
        target: 'cpu' or 'gpu'
183
        gpu_exclusive_conditions: if/else construct to have only one code block for each of 2**dim code paths
184
        kwargs: passed directly to create_kernel, iteration slice and ghost_layers parameters are not allowed
185

186
    Returns:
187
        AST, see `create_kernel`
188
189
    """
    assert 'iteration_slice' not in kwargs and 'ghost_layers' not in kwargs
190
    assert staggered_field.index_dimensions in (1, 2), 'Staggered field must have one or two index dimensions'
191
192
193
194
195
    dim = staggered_field.spatial_dimensions

    counters = [LoopOverCoordinate.get_loop_counter_symbol(i) for i in range(dim)]
    conditions = [counters[i] < staggered_field.shape[i] - 1 for i in range(dim)]
    assert len(expressions) == dim
196
197
198
199
200
    if staggered_field.index_dimensions == 2:
        assert all(len(sublist) == len(expressions[0]) for sublist in expressions), \
            "If staggered field has two index dimensions expressions has to be a sequence of sequences of all the " \
            "same length."

201
    final_assignments = []
202
203
204
205
    last_conditional = None

    def add(condition, dimensions, as_else_block=False):
        nonlocal last_conditional
206
        if staggered_field.index_dimensions == 1:
207
208
209
            assignments = [Assignment(staggered_field(d), expressions[d]) for d in dimensions]
            a_coll = AssignmentCollection(assignments, list(subexpressions))
            a_coll = a_coll.new_filtered([staggered_field(d) for d in dimensions])
210
211
        elif staggered_field.index_dimensions == 2:
            assert staggered_field.has_fixed_index_shape
212
213
214
            assignments = [Assignment(staggered_field(d, i), expr)
                           for d in dimensions
                           for i, expr in enumerate(expressions[d])]
215
            a_coll = AssignmentCollection(assignments, list(subexpressions))
216
217
            a_coll = a_coll.new_filtered([staggered_field(d, i) for i in range(staggered_field.index_shape[1])
                                          for d in dimensions])
218
        sp_assignments = [SympyAssignment(a.lhs, a.rhs) for a in a_coll.all_assignments]
219
        if as_else_block and last_conditional:
220
221
222
            new_cond = Conditional(condition, Block(sp_assignments))
            last_conditional.false_block = Block([new_cond])
            last_conditional = new_cond
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
        else:
            last_conditional = Conditional(condition, Block(sp_assignments))
            final_assignments.append(last_conditional)

    if target == 'cpu' or not gpu_exclusive_conditions:
        for d in range(dim):
            cond = sp.And(*[conditions[i] for i in range(dim) if d != i])
            add(cond, [d])
    elif target == 'gpu':
        full_conditions = [sp.And(*[conditions[i] for i in range(dim) if d != i]) for d in range(dim)]
        for include in itertools.product(*[[1, 0]] * dim):
            case_conditions = sp.And(*[c if value else sp.Not(c) for c, value in zip(full_conditions, include)])
            dimensions_to_include = [i for i in range(dim) if include[i]]
            if dimensions_to_include:
                add(case_conditions, dimensions_to_include, True)
238

239
240
    ghost_layers = [(1, 0)] * dim

Martin Bauer's avatar
Martin Bauer committed
241
242
243
244
    blocking = kwargs.get('cpu_blocking', None)
    if blocking:
        del kwargs['cpu_blocking']

245
246
247
    cpu_vectorize_info = kwargs.get('cpu_vectorize_info', None)
    if cpu_vectorize_info:
        del kwargs['cpu_vectorize_info']
248
249
250
251
    openmp = kwargs.get('cpu_openmp', None)
    if openmp:
        del kwargs['cpu_openmp']

252
    ast = create_kernel(final_assignments, ghost_layers=ghost_layers, target=target, **kwargs)
253

254
255
    if target == 'cpu':
        remove_conditionals_in_staggered_kernel(ast)
256
        move_constants_before_loop(ast)
257
        omp_collapse = None
Martin Bauer's avatar
Martin Bauer committed
258
        if blocking:
259
260
261
262
            omp_collapse = loop_blocking(ast, blocking)
        if openmp:
            from pystencils.cpu import add_openmp
            add_openmp(ast, num_threads=openmp, collapse=omp_collapse, assume_single_outer_loop=False)
263
264
265
266
        if cpu_vectorize_info is True:
            vectorize(ast)
        elif isinstance(cpu_vectorize_info, dict):
            vectorize(ast, **cpu_vectorize_info)
267
    return ast