cbackend.py 26.7 KB
Newer Older
Martin Bauer's avatar
Martin Bauer committed
1
from collections import namedtuple
2
from typing import Set
3

4
import numpy as np
5
6
import sympy as sp
from sympy.core import S
7
from sympy.printing.ccode import C89CodePrinter
8
from pystencils.astnodes import KernelFunction, Node
9
from pystencils.cpu.vectorization import vec_all, vec_any
Martin Bauer's avatar
Martin Bauer committed
10
from pystencils.data_types import (
11
12
    PointerType, VectorType, address_of, cast_func, create_type, get_type_of_expression,
    reinterpret_cast_func, vector_memory_access)
Martin Bauer's avatar
Martin Bauer committed
13
14
from pystencils.fast_approximation import fast_division, fast_inv_sqrt, fast_sqrt
from pystencils.integer_functions import (
15
16
    bit_shift_left, bit_shift_right, bitwise_and, bitwise_or, bitwise_xor,
    int_div, int_power_of_2, modulo_ceil)
17

Martin Bauer's avatar
Martin Bauer committed
18
19
try:
    from sympy.printing.ccode import C99CodePrinter as CCodePrinter
Martin Bauer's avatar
Martin Bauer committed
20
21
except ImportError:
    from sympy.printing.ccode import CCodePrinter  # for sympy versions < 1.1
Martin Bauer's avatar
Martin Bauer committed
22

23
__all__ = ['generate_c', 'CustomCodeNode', 'PrintNode', 'get_headers', 'CustomSympyPrinter']
24

25
26
KERNCRAFT_NO_TERNARY_MODE = False

Martin Bauer's avatar
Fixes    
Martin Bauer committed
27

28
def generate_c(ast_node: Node, signature_only: bool = False, dialect='c', custom_backend=None) -> str:
Martin Bauer's avatar
Martin Bauer committed
29
30
31
32
33
34
35
36
37
    """Prints an abstract syntax tree node as C or CUDA code.

    This function does not need to distinguish between C, C++ or CUDA code, it just prints 'C-like' code as encoded
    in the abstract syntax tree (AST). The AST is built differently for C or CUDA by calling different create_kernel
    functions.

    Args:
        ast_node:
        signature_only:
38
        dialect: 'c' or 'cuda'
Martin Bauer's avatar
Martin Bauer committed
39
40
    Returns:
        C-like code for the ast node and its descendants
Martin Bauer's avatar
Martin Bauer committed
41
    """
42
43
44
45
46
47
    global_declarations = get_global_declarations(ast_node)
    for d in global_declarations:
        if hasattr(ast_node, "global_variables"):
            ast_node.global_variables.update(d.symbols_defined)
        else:
            ast_node.global_variables = d.symbols_defined
48
49
50
    if custom_backend:
        printer = custom_backend
    elif dialect == 'c':
51
52
53
54
        try:
            instruction_set = ast_node.instruction_set
        except Exception:
            instruction_set = None
55
        printer = CBackend(signature_only=signature_only,
56
                           vector_instruction_set=instruction_set)
57
58
59
    elif dialect == 'cuda':
        from pystencils.backends.cuda_backend import CudaBackend
        printer = CudaBackend(signature_only=signature_only)
Stephan Seitz's avatar
Stephan Seitz committed
60
    elif dialect == 'opencl':
61
62
        from pystencils.backends.opencl_backend import OpenClBackend
        printer = OpenClBackend(signature_only=signature_only)
63
    else:
Martin Bauer's avatar
Martin Bauer committed
64
        raise ValueError("Unknown dialect: " + str(dialect))
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
    code = printer(ast_node)
    if not signature_only and isinstance(ast_node, KernelFunction):
        code = "\n" + code
        for declaration in global_declarations:
            code = printer(declaration) + "\n" + code

    return code


def get_global_declarations(ast):
    global_declarations = []

    def visit_node(sub_ast):
        if hasattr(sub_ast, "required_global_declarations"):
            nonlocal global_declarations
            global_declarations += sub_ast.required_global_declarations

        if hasattr(sub_ast, "args"):
            for node in sub_ast.args:
                visit_node(node)

    visit_node(ast)

88
    return sorted(set(global_declarations), key=lambda x: str(x))
89
90


Martin Bauer's avatar
Martin Bauer committed
91
92
def get_headers(ast_node: Node) -> Set[str]:
    """Return a set of header files, necessary to compile the printed C-like code."""
93
94
    headers = set()

Martin Bauer's avatar
Martin Bauer committed
95
96
97
    if isinstance(ast_node, KernelFunction) and ast_node.instruction_set:
        headers.update(ast_node.instruction_set['headers'])

Martin Bauer's avatar
Martin Bauer committed
98
99
100
    if hasattr(ast_node, 'headers'):
        headers.update(ast_node.headers)
    for a in ast_node.args:
101
        if isinstance(a, Node):
Martin Bauer's avatar
Martin Bauer committed
102
            headers.update(get_headers(a))
103

104
105
106
107
    for g in get_global_declarations(ast_node):
        if isinstance(g, Node):
            headers.update(get_headers(g))

108
    return sorted(headers)
109
110


111
112
113
# --------------------------------------- Backend Specific Nodes -------------------------------------------------------


114
class CustomCodeNode(Node):
Martin Bauer's avatar
Martin Bauer committed
115
    def __init__(self, code, symbols_read, symbols_defined, parent=None):
116
        super(CustomCodeNode, self).__init__(parent=parent)
117
        self._code = "\n" + code
118
119
        self._symbols_read = set(symbols_read)
        self._symbols_defined = set(symbols_defined)
120
        self.headers = []
121

122
    def get_code(self, dialect, vector_instruction_set):
123
124
125
126
127
128
129
        return self._code

    @property
    def args(self):
        return []

    @property
Martin Bauer's avatar
Martin Bauer committed
130
    def symbols_defined(self):
131
        return self._symbols_defined
132
133

    @property
Martin Bauer's avatar
Martin Bauer committed
134
    def undefined_symbols(self):
135
        return self._symbols_read - self._symbols_defined
136

137
138
139
140
141
142
    def __eq___(self, other):
        return self._code == other._code

    def __hash__(self):
        return hash(self._code)

143

144
class PrintNode(CustomCodeNode):
Martin Bauer's avatar
Martin Bauer committed
145
146
147
148
    # noinspection SpellCheckingInspection
    def __init__(self, symbol_to_print):
        code = '\nstd::cout << "%s  =  " << %s << std::endl; \n' % (symbol_to_print.name, symbol_to_print.name)
        super(PrintNode, self).__init__(code, symbols_read=[symbol_to_print], symbols_defined=set())
149
        self.headers.append("<iostream>")
150
151
152
153


# ------------------------------------------- Printer ------------------------------------------------------------------

154

Martin Bauer's avatar
Martin Bauer committed
155
156
# noinspection PyPep8Naming
class CBackend:
157

Martin Bauer's avatar
Martin Bauer committed
158
    def __init__(self, sympy_printer=None, signature_only=False, vector_instruction_set=None, dialect='c'):
Martin Bauer's avatar
Martin Bauer committed
159
160
        if sympy_printer is None:
            if vector_instruction_set is not None:
161
                self.sympy_printer = VectorizedCustomSympyPrinter(vector_instruction_set)
162
            else:
163
                self.sympy_printer = CustomSympyPrinter()
164
        else:
Martin Bauer's avatar
Martin Bauer committed
165
            self.sympy_printer = sympy_printer
166

167
        self._vector_instruction_set = vector_instruction_set
168
        self._indent = "   "
169
        self._dialect = dialect
Martin Bauer's avatar
Martin Bauer committed
170
        self._signatureOnly = signature_only
171
172

    def __call__(self, node):
Martin Bauer's avatar
Martin Bauer committed
173
        prev_is = VectorType.instruction_set
174
        VectorType.instruction_set = self._vector_instruction_set
175
        result = str(self._print(node))
Martin Bauer's avatar
Martin Bauer committed
176
        VectorType.instruction_set = prev_is
177
        return result
178
179

    def _print(self, node):
Stephan Seitz's avatar
Stephan Seitz committed
180
181
        if isinstance(node, str):
            return node
182
        for cls in type(node).__mro__:
Martin Bauer's avatar
Martin Bauer committed
183
184
185
            method_name = "_print_" + cls.__name__
            if hasattr(self, method_name):
                return getattr(self, method_name)(node)
186
        raise NotImplementedError(self.__class__.__name__ + " does not support node of type " + node.__class__.__name__)
187

188
189
190
    def _print_Type(self, node):
        return str(node)

191
    def _print_KernelFunction(self, node):
192
        function_arguments = ["%s %s" % (self._print(s.symbol.dtype), s.symbol.name) for s in node.get_parameters()]
193
        launch_bounds = ""
194
        if self._dialect == 'cuda':
195
196
197
198
199
            max_threads = node.indexing.max_threads_per_block()
            if max_threads:
                launch_bounds = "__launch_bounds__({}) ".format(max_threads)
        func_declaration = "FUNC_PREFIX %svoid %s(%s)" % (launch_bounds, node.function_name,
                                                          ", ".join(function_arguments))
200
        if self._signatureOnly:
Martin Bauer's avatar
Martin Bauer committed
201
            return func_declaration
202

203
        body = self._print(node.body)
Martin Bauer's avatar
Martin Bauer committed
204
        return func_declaration + "\n" + body
205
206

    def _print_Block(self, node):
Martin Bauer's avatar
Martin Bauer committed
207
208
        block_contents = "\n".join([self._print(child) for child in node.args])
        return "{\n%s\n}" % (self._indent + self._indent.join(block_contents.splitlines(True)))
209
210

    def _print_PragmaBlock(self, node):
Martin Bauer's avatar
Martin Bauer committed
211
        return "%s\n%s" % (node.pragma_line, self._print_Block(node))
212
213

    def _print_LoopOverCoordinate(self, node):
Martin Bauer's avatar
Martin Bauer committed
214
        counter_symbol = node.loop_counter_name
Martin Bauer's avatar
Martin Bauer committed
215
216
217
218
        start = "int %s = %s" % (counter_symbol, self.sympy_printer.doprint(node.start))
        condition = "%s < %s" % (counter_symbol, self.sympy_printer.doprint(node.stop))
        update = "%s += %s" % (counter_symbol, self.sympy_printer.doprint(node.step),)
        loop_str = "for (%s; %s; %s)" % (start, condition, update)
219

Martin Bauer's avatar
Martin Bauer committed
220
        prefix = "\n".join(node.prefix_lines)
221
222
        if prefix:
            prefix += "\n"
Martin Bauer's avatar
Martin Bauer committed
223
        return "%s%s\n%s" % (prefix, loop_str, self._print(node.body))
224
225

    def _print_SympyAssignment(self, node):
Martin Bauer's avatar
Martin Bauer committed
226
        if node.is_declaration:
Stephan Seitz's avatar
Lint    
Stephan Seitz committed
227
228
229
230
            if node.is_const:
                prefix = 'const '
            else:
                prefix = ''
231
            data_type = prefix + self._print(node.lhs.dtype).replace(' const', '') + " "
232
233
            return "%s%s = %s;" % (data_type, self.sympy_printer.doprint(node.lhs),
                                   self.sympy_printer.doprint(node.rhs))
234
        else:
Martin Bauer's avatar
Martin Bauer committed
235
            lhs_type = get_type_of_expression(node.lhs)
Martin Bauer's avatar
Martin Bauer committed
236
            printed_mask = ""
Martin Bauer's avatar
Martin Bauer committed
237
            if type(lhs_type) is VectorType and isinstance(node.lhs, cast_func):
Martin Bauer's avatar
Martin Bauer committed
238
                arg, data_type, aligned, nontemporal, mask = node.lhs.args
Martin Bauer's avatar
Martin Bauer committed
239
240
241
                instr = 'storeU'
                if aligned:
                    instr = 'stream' if nontemporal else 'storeA'
Martin Bauer's avatar
Martin Bauer committed
242
243
244
245
246
                if mask != True:
                    instr = 'maskStore' if aligned else 'maskStoreU'
                    printed_mask = self.sympy_printer.doprint(mask)
                    if self._vector_instruction_set['dataTypePrefix']['double'] == '__mm256d':
                        printed_mask = "_mm256_castpd_si256({})".format(printed_mask)
Martin Bauer's avatar
Martin Bauer committed
247

248
249
250
251
252
253
                rhs_type = get_type_of_expression(node.rhs)
                if type(rhs_type) is not VectorType:
                    rhs = cast_func(node.rhs, VectorType(rhs_type))
                else:
                    rhs = node.rhs

254
                return self._vector_instruction_set[instr].format("&" + self.sympy_printer.doprint(node.lhs.args[0]),
Martin Bauer's avatar
Martin Bauer committed
255
256
                                                                  self.sympy_printer.doprint(rhs),
                                                                  printed_mask) + ';'
257
            else:
Martin Bauer's avatar
Martin Bauer committed
258
                return "%s = %s;" % (self.sympy_printer.doprint(node.lhs), self.sympy_printer.doprint(node.rhs))
259
260

    def _print_TemporaryMemoryAllocation(self, node):
261
        align = 64
Martin Bauer's avatar
Martin Bauer committed
262
263
264
265
266
267
        np_dtype = node.symbol.dtype.base_type.numpy_dtype
        required_size = np_dtype.itemsize * node.size + align
        size = modulo_ceil(required_size, align)
        code = "{dtype} {name}=({dtype})aligned_alloc({align}, {size}) + {offset};"
        return code.format(dtype=node.symbol.dtype,
                           name=self.sympy_printer.doprint(node.symbol.name),
268
                           size=self.sympy_printer.doprint(size),
Martin Bauer's avatar
Martin Bauer committed
269
270
                           offset=int(node.offset(align)),
                           align=align)
271
272

    def _print_TemporaryMemoryFree(self, node):
273
        align = 64
Martin Bauer's avatar
Martin Bauer committed
274
        return "free(%s - %d);" % (self.sympy_printer.doprint(node.symbol.name), node.offset(align))
275

Martin Bauer's avatar
Martin Bauer committed
276
    def _print_SkipIteration(self, _):
277
        return "continue;"
Martin Bauer's avatar
Martin Bauer committed
278

279
280
    def _print_CustomCodeNode(self, node):
        return node.get_code(self._dialect, self._vector_instruction_set)
281

282
283
284
285
286
287
    def _print_SourceCodeComment(self, node):
        return "/* " + node.text + " */"

    def _print_EmptyLine(self, node):
        return ""

288
    def _print_Conditional(self, node):
289
290
291
        cond_type = get_type_of_expression(node.condition_expr)
        if isinstance(cond_type, VectorType):
            raise ValueError("Problem with Conditional inside vectorized loop - use vec_any or vec_all")
Martin Bauer's avatar
Martin Bauer committed
292
293
        condition_expr = self.sympy_printer.doprint(node.condition_expr)
        true_block = self._print_Block(node.true_block)
Martin Bauer's avatar
Martin Bauer committed
294
        result = "if (%s)\n%s " % (condition_expr, true_block)
Martin Bauer's avatar
Martin Bauer committed
295
296
        if node.false_block:
            false_block = self._print_Block(node.false_block)
Martin Bauer's avatar
Martin Bauer committed
297
            result += "else " + false_block
298
299
        return result

300
301
302
303

# ------------------------------------------ Helper function & classes -------------------------------------------------


Martin Bauer's avatar
Martin Bauer committed
304
# noinspection PyPep8Naming
305
class CustomSympyPrinter(CCodePrinter):
Martin Bauer's avatar
Martin Bauer committed
306

307
    def __init__(self):
Martin Bauer's avatar
Martin Bauer committed
308
        super(CustomSympyPrinter, self).__init__()
309
        self._float_type = create_type("float32")
310
311
312
313
        if 'Min' in self.known_functions:
            del self.known_functions['Min']
        if 'Max' in self.known_functions:
            del self.known_functions['Max']
Martin Bauer's avatar
Martin Bauer committed
314

315
316
    def _print_Pow(self, expr):
        """Don't use std::pow function, for small integer exponents, write as multiplication"""
Martin Bauer's avatar
Martin Bauer committed
317
318
319
        if not expr.free_symbols:
            return self._typed_number(expr.evalf(), get_type_of_expression(expr))

320
        if expr.exp.is_integer and expr.exp.is_number and 0 < expr.exp < 8:
321
            return "(" + self._print(sp.Mul(*[expr.base] * expr.exp, evaluate=False)) + ")"
322
323
        elif expr.exp.is_integer and expr.exp.is_number and - 8 < expr.exp < 0:
            return "1 / ({})".format(self._print(sp.Mul(*[expr.base] * (-expr.exp), evaluate=False)))
324
325
326
327
328
        else:
            return super(CustomSympyPrinter, self)._print_Pow(expr)

    def _print_Rational(self, expr):
        """Evaluate all rationals i.e. print 0.25 instead of 1.0/4.0"""
Martin Bauer's avatar
Martin Bauer committed
329
330
        res = str(expr.evalf().num)
        return res
331
332
333
334
335
336
337
338

    def _print_Equality(self, expr):
        """Equality operator is not printable in default printer"""
        return '((' + self._print(expr.lhs) + ") == (" + self._print(expr.rhs) + '))'

    def _print_Piecewise(self, expr):
        """Print piecewise in one line (remove newlines)"""
        result = super(CustomSympyPrinter, self)._print_Piecewise(expr)
Martin Bauer's avatar
Martin Bauer committed
339
340
        return result.replace("\n", "")

341
    def _print_Function(self, expr):
342
        infix_functions = {
Martin Bauer's avatar
Martin Bauer committed
343
344
345
346
347
            bitwise_xor: '^',
            bit_shift_right: '>>',
            bit_shift_left: '<<',
            bitwise_or: '|',
            bitwise_and: '&',
Martin Bauer's avatar
Martin Bauer committed
348
        }
Martin Bauer's avatar
Martin Bauer committed
349
350
        if hasattr(expr, 'to_c'):
            return expr.to_c(self._print)
351
352
353
        if isinstance(expr, reinterpret_cast_func):
            arg, data_type = expr.args
            return "*((%s)(& %s))" % (PointerType(data_type, restrict=False), self._print(arg))
354
355
356
        elif isinstance(expr, address_of):
            assert len(expr.args) == 1, "address_of must only have one argument"
            return "&(%s)" % self._print(expr.args[0])
357
        elif isinstance(expr, cast_func):
Martin Bauer's avatar
Martin Bauer committed
358
            arg, data_type = expr.args
359
            if isinstance(arg, sp.Number) and arg.is_finite:
360
361
                return self._typed_number(arg, data_type)
            else:
362
363
                return "((%s)(%s))" % (data_type, self._print(arg))
        elif isinstance(expr, fast_division):
364
            return "({})".format(self._print(expr.args[0] / expr.args[1]))
365
        elif isinstance(expr, fast_sqrt):
366
            return "({})".format(self._print(sp.sqrt(expr.args[0])))
367
368
        elif isinstance(expr, vec_any) or isinstance(expr, vec_all):
            return self._print(expr.args[0])
369
        elif isinstance(expr, fast_inv_sqrt):
370
            return "({})".format(self._print(1 / sp.sqrt(expr.args[0])))
371
372
        elif expr.func in infix_functions:
            return "(%s %s %s)" % (self._print(expr.args[0]), infix_functions[expr.func], self._print(expr.args[1]))
373
374
375
376
        elif expr.func == int_power_of_2:
            return "(1 << (%s))" % (self._print(expr.args[0]))
        elif expr.func == int_div:
            return "((%s) / (%s))" % (self._print(expr.args[0]), self._print(expr.args[1]))
377
        else:
378
            return super(CustomSympyPrinter, self)._print_Function(expr)
Martin Bauer's avatar
Martin Bauer committed
379

380
381
    def _typed_number(self, number, dtype):
        res = self._print(number)
382
383
384
385
        if dtype.numpy_dtype == np.float32:
            return res + '.0f' if '.' not in res else res + 'f'
        elif dtype.numpy_dtype == np.float64:
            return res + '.0' if '.' not in res else res
386
387
        else:
            return res
388

Stephan Seitz's avatar
Stephan Seitz committed
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
    def _print_Sum(self, expr):
        template = """[&]() {{
    {dtype} sum = ({dtype}) 0;
    for ( {iterator_dtype} {var} = {start}; {condition}; {var} += {increment} ) {{
        sum += {expr};
    }}
    return sum;
}}()"""
        var = expr.limits[0][0]
        start = expr.limits[0][1]
        end = expr.limits[0][2]
        code = template.format(
            dtype=get_type_of_expression(expr.args[0]),
            iterator_dtype='int',
            var=self._print(var),
            start=self._print(start),
            end=self._print(end),
            expr=self._print(expr.function),
            increment=str(1),
            condition=self._print(var) + ' <= ' + self._print(end)  # if start < end else '>='
        )
        return code

    def _print_Product(self, expr):
        template = """[&]() {{
    {dtype} product = ({dtype}) 1;
    for ( {iterator_dtype} {var} = {start}; {condition}; {var} += {increment} ) {{
        product *= {expr};
    }}
    return product;
}}()"""
        var = expr.limits[0][0]
        start = expr.limits[0][1]
        end = expr.limits[0][2]
        code = template.format(
            dtype=get_type_of_expression(expr.args[0]),
            iterator_dtype='int',
            var=self._print(var),
            start=self._print(start),
            end=self._print(end),
            expr=self._print(expr.function),
            increment=str(1),
            condition=self._print(var) + ' <= ' + self._print(end)  # if start < end else '>='
        )
        return code
434

435
436
437
    def _print_ConditionalFieldAccess(self, node):
        return self._print(sp.Piecewise((node.outofbounds_value, node.outofbounds_condition), (node.access, True)))

438
439
440
    _print_Max = C89CodePrinter._print_Max
    _print_Min = C89CodePrinter._print_Min

441

Martin Bauer's avatar
Martin Bauer committed
442
# noinspection PyPep8Naming
443
444
445
class VectorizedCustomSympyPrinter(CustomSympyPrinter):
    SummandInfo = namedtuple("SummandInfo", ['sign', 'term'])

446
447
    def __init__(self, instruction_set):
        super(VectorizedCustomSympyPrinter, self).__init__()
Martin Bauer's avatar
Martin Bauer committed
448
        self.instruction_set = instruction_set
449

Martin Bauer's avatar
Martin Bauer committed
450
451
452
453
    def _scalarFallback(self, func_name, expr, *args, **kwargs):
        expr_type = get_type_of_expression(expr)
        if type(expr_type) is not VectorType:
            return getattr(super(VectorizedCustomSympyPrinter, self), func_name)(expr, *args, **kwargs)
454
        else:
Martin Bauer's avatar
Martin Bauer committed
455
            assert self.instruction_set['width'] == expr_type.width
456
457
            return None

458
    def _print_Function(self, expr):
459
        if isinstance(expr, vector_memory_access):
Martin Bauer's avatar
Martin Bauer committed
460
            arg, data_type, aligned, _, mask = expr.args
Martin Bauer's avatar
Martin Bauer committed
461
462
            instruction = self.instruction_set['loadA'] if aligned else self.instruction_set['loadU']
            return instruction.format("& " + self._print(arg))
463
        elif isinstance(expr, cast_func):
Martin Bauer's avatar
Martin Bauer committed
464
465
            arg, data_type = expr.args
            if type(data_type) is VectorType:
466
467
468
469
470
471
472
473
474
                if isinstance(arg, sp.Tuple):
                    is_boolean = get_type_of_expression(arg[0]) == create_type("bool")
                    printed_args = [self._print(a) for a in arg]
                    instruction = 'makeVecBool' if is_boolean else 'makeVec'
                    return self.instruction_set[instruction].format(*printed_args)
                else:
                    is_boolean = get_type_of_expression(arg) == create_type("bool")
                    instruction = 'makeVecConstBool' if is_boolean else 'makeVecConst'
                    return self.instruction_set[instruction].format(self._print(arg))
475
        elif expr.func == fast_division:
476
477
            result = self._scalarFallback('_print_Function', expr)
            if not result:
478
479
                result = self.instruction_set['/'].format(self._print(expr.args[0]), self._print(expr.args[1]))
            return result
480
481
482
        elif expr.func == fast_sqrt:
            return "({})".format(self._print(sp.sqrt(expr.args[0])))
        elif expr.func == fast_inv_sqrt:
483
484
485
486
487
488
            result = self._scalarFallback('_print_Function', expr)
            if not result:
                if self.instruction_set['rsqrt']:
                    return self.instruction_set['rsqrt'].format(self._print(expr.args[0]))
                else:
                    return "({})".format(self._print(1 / sp.sqrt(expr.args[0])))
489
490
491
492
493
494
495
496
497
498
499
500
501
        elif isinstance(expr, vec_any):
            expr_type = get_type_of_expression(expr.args[0])
            if type(expr_type) is not VectorType:
                return self._print(expr.args[0])
            else:
                return self.instruction_set['any'].format(self._print(expr.args[0]))
        elif isinstance(expr, vec_all):
            expr_type = get_type_of_expression(expr.args[0])
            if type(expr_type) is not VectorType:
                return self._print(expr.args[0])
            else:
                return self.instruction_set['all'].format(self._print(expr.args[0]))

502
503
        return super(VectorizedCustomSympyPrinter, self)._print_Function(expr)

504
505
506
507
508
    def _print_And(self, expr):
        result = self._scalarFallback('_print_And', expr)
        if result:
            return result

Martin Bauer's avatar
Martin Bauer committed
509
510
511
512
        arg_strings = [self._print(a) for a in expr.args]
        assert len(arg_strings) > 0
        result = arg_strings[0]
        for item in arg_strings[1:]:
Martin Bauer's avatar
Martin Bauer committed
513
            result = self.instruction_set['&'].format(result, item)
514
515
516
517
518
519
520
        return result

    def _print_Or(self, expr):
        result = self._scalarFallback('_print_Or', expr)
        if result:
            return result

Martin Bauer's avatar
Martin Bauer committed
521
522
523
524
        arg_strings = [self._print(a) for a in expr.args]
        assert len(arg_strings) > 0
        result = arg_strings[0]
        for item in arg_strings[1:]:
Martin Bauer's avatar
Martin Bauer committed
525
            result = self.instruction_set['|'].format(result, item)
526
527
        return result

528
    def _print_Add(self, expr, order=None):
529
530
531
        result = self._scalarFallback('_print_Add', expr)
        if result:
            return result
532
533
534
535

        summands = []
        for term in expr.args:
            if term.func == sp.Mul:
Martin Bauer's avatar
Martin Bauer committed
536
                sign, t = self._print_Mul(term, inside_add=True)
537
538
539
540
541
542
543
544
545
546
547
548
549
            else:
                t = self._print(term)
                sign = 1
            summands.append(self.SummandInfo(sign, t))
        # Use positive terms first
        summands.sort(key=lambda e: e.sign, reverse=True)
        # if no positive term exists, prepend a zero
        if summands[0].sign == -1:
            summands.insert(0, self.SummandInfo(1, "0"))

        assert len(summands) >= 2
        processed = summands[0].term
        for summand in summands[1:]:
Martin Bauer's avatar
Martin Bauer committed
550
            func = self.instruction_set['-'] if summand.sign == -1 else self.instruction_set['+']
551
552
553
            processed = func.format(processed, summand.term)
        return processed

554
    def _print_Pow(self, expr):
555
556
557
        result = self._scalarFallback('_print_Pow', expr)
        if result:
            return result
558

559
        one = self.instruction_set['makeVecConst'].format(1.0)
560

561
562
        if expr.exp.is_integer and expr.exp.is_number and 0 < expr.exp < 8:
            return "(" + self._print(sp.Mul(*[expr.base] * expr.exp, evaluate=False)) + ")"
563
        elif expr.exp == -1:
564
            one = self.instruction_set['makeVecConst'].format(1.0)
565
566
567
            return self.instruction_set['/'].format(one, self._print(expr.base))
        elif expr.exp == 0.5:
            return self.instruction_set['sqrt'].format(self._print(expr.base))
568
569
570
        elif expr.exp == -0.5:
            root = self.instruction_set['sqrt'].format(self._print(expr.base))
            return self.instruction_set['/'].format(one, root)
571
572
573
        elif expr.exp.is_integer and expr.exp.is_number and - 8 < expr.exp < 0:
            return self.instruction_set['/'].format(one,
                                                    self._print(sp.Mul(*[expr.base] * (-expr.exp), evaluate=False)))
574
        else:
575
            raise ValueError("Generic exponential not supported: " + str(expr))
576

Martin Bauer's avatar
Martin Bauer committed
577
578
579
580
    def _print_Mul(self, expr, inside_add=False):
        # noinspection PyProtectedMember
        from sympy.core.mul import _keep_coeff

581
582
583
        result = self._scalarFallback('_print_Mul', expr)
        if result:
            return result
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611

        c, e = expr.as_coeff_Mul()
        if c < 0:
            expr = _keep_coeff(-c, e)
            sign = -1
        else:
            sign = 1

        a = []  # items in the numerator
        b = []  # items that are in the denominator (if any)

        # Gather args for numerator/denominator
        for item in expr.as_ordered_factors():
            if item.is_commutative and item.is_Pow and item.exp.is_Rational and item.exp.is_negative:
                if item.exp != -1:
                    b.append(sp.Pow(item.base, -item.exp, evaluate=False))
                else:
                    b.append(sp.Pow(item.base, -item.exp))
            else:
                a.append(item)

        a = a or [S.One]

        a_str = [self._print(x) for x in a]
        b_str = [self._print(x) for x in b]

        result = a_str[0]
        for item in a_str[1:]:
Martin Bauer's avatar
Martin Bauer committed
612
            result = self.instruction_set['*'].format(result, item)
613
614
615
616

        if len(b) > 0:
            denominator_str = b_str[0]
            for item in b_str[1:]:
Martin Bauer's avatar
Martin Bauer committed
617
618
                denominator_str = self.instruction_set['*'].format(denominator_str, item)
            result = self.instruction_set['/'].format(result, denominator_str)
619

Martin Bauer's avatar
Martin Bauer committed
620
        if inside_add:
621
622
623
            return sign, result
        else:
            if sign < 0:
Martin Bauer's avatar
Martin Bauer committed
624
                return self.instruction_set['*'].format(self._print(S.NegativeOne), result)
625
626
627
            else:
                return result

628
    def _print_Relational(self, expr):
629
630
631
        result = self._scalarFallback('_print_Relational', expr)
        if result:
            return result
Martin Bauer's avatar
Martin Bauer committed
632
        return self.instruction_set[expr.rel_op].format(self._print(expr.lhs), self._print(expr.rhs))
633
634

    def _print_Equality(self, expr):
635
636
637
        result = self._scalarFallback('_print_Equality', expr)
        if result:
            return result
Martin Bauer's avatar
Martin Bauer committed
638
        return self.instruction_set['=='].format(self._print(expr.lhs), self._print(expr.rhs))
639
640

    def _print_Piecewise(self, expr):
641
642
643
        result = self._scalarFallback('_print_Piecewise', expr)
        if result:
            return result
644

Martin Bauer's avatar
Martin Bauer committed
645
        if expr.args[-1].cond.args[0] is not sp.sympify(True):
646
647
648
649
650
651
652
653
654
            # We need the last conditional to be a True, otherwise the resulting
            # function may not return a result.
            raise ValueError("All Piecewise expressions must contain an "
                             "(expr, True) statement to be used as a default "
                             "condition. Without one, the generated "
                             "expression may not evaluate to anything under "
                             "some condition.")

        result = self._print(expr.args[-1][0])
Martin Bauer's avatar
Martin Bauer committed
655
        for true_expr, condition in reversed(expr.args[:-1]):
656
            if isinstance(condition, cast_func) and get_type_of_expression(condition.args[0]) == create_type("bool"):
657
658
659
660
661
                if not KERNCRAFT_NO_TERNARY_MODE:
                    result = "(({}) ? ({}) : ({}))".format(self._print(condition.args[0]), self._print(true_expr),
                                                           result)
                else:
                    print("Warning - skipping ternary op")
662
663
664
            else:
                # noinspection SpellCheckingInspection
                result = self.instruction_set['blendv'].format(result, self._print(true_expr), self._print(condition))
665
        return result