Merge pull request #72 from pypr/fix-32bit

Fix tests on 32bit architectures

compyle/array.py

@@ -536,7 +536,7 @@ def update_minmax_gpu(ary_list, only_min=False, only_max=False,
     props = ['ary_%s' % i for i in range(len(ary_list))]
 
     dtype = ary_list[0].dtype
-    ctype = dtype_to_ctype(dtype)
+    ctype = dtype_to_ctype(dtype, backend=backend)
 
     op = 'min' if not only_max else ''
     op += 'max' if not only_min else ''
@@ -687,7 +687,8 @@ def __init__(self, dtype, n=0, allocate=True, backend=None):
             from .cuda import set_context
             set_context()
         self.dtype = dtype
-        self.gptr_type = dtype_to_knowntype(dtype, address='global')
+        self.gptr_type = dtype_to_knowntype(dtype, address='global',
+                                            backend=backend)
         self.minimum = 0
         self.maximum = 0
         self.data = None

compyle/cuda.py

@@ -2,8 +2,8 @@
 """
 from __future__ import print_function
 
-from pytools import Record, RecordWithoutPickling
 import logging
+from pytools import Record, RecordWithoutPickling
 from pytools.persistent_dict import KeyBuilder as KeyBuilderBase
 from pytools.persistent_dict import WriteOncePersistentDict
 from pycuda._cluda import CLUDA_PREAMBLE
@@ -14,7 +14,6 @@
 from compyle.thrust.sort import argsort
 import pycuda.driver as drv
 from pycuda.compiler import SourceModule as _SourceModule
-from pycuda.tools import dtype_to_ctype
 from pytools import memoize
 import numpy as np
 import six
@@ -66,7 +65,8 @@ def add_dtype(self, dtype):
             self.add_dtype(field_dtype)
 
         _, cdecl = match_dtype_to_c_struct(
-            self.device, dtype_to_ctype(dtype), dtype)
+            self.device, dtype_to_ctype(dtype), dtype
+        )
 
         self.declarations.append(cdecl)
         self.declared_dtypes.add(dtype)
@@ -132,7 +132,8 @@ def match_dtype_to_c_struct(device, name, dtype, context=None, use_typedef=False
     c_fields = []
     for field_name, dtype_and_offset in fields:
         field_dtype, offset = dtype_and_offset[:2]
-        c_fields.append("  %s %s;" % (dtype_to_ctype(field_dtype), field_name))
+        c_fields.append("  %s %s;" % (dtype_to_ctype(field_dtype),
+                                      field_name))
 
     if use_typedef:
         c_decl = "typedef struct {\n%s\n} %s;\n\n" % (

compyle/jit.py

@@ -11,7 +11,7 @@
 from .cython_generator import CythonGenerator
 from .transpiler import Transpiler, BUILTINS
 from .types import (dtype_to_ctype, get_declare_info,
-                    dtype_to_knowntype, annotate)
+                    dtype_to_knowntype, annotate, BITS)
 from .extern import Extern
 from .utils import getsourcelines
 from .profile import profile
@@ -44,7 +44,7 @@ def get_ctype_from_arg(arg):
             return 'double'
         else:
             if arg > 2147483648:
-                return 'long'
+                return 'long long' if BITS.startswith('32') else 'long'
             else:
                 return 'int'
 
@@ -362,7 +362,7 @@ def __init__(self, reduce_expr, map_func=None, dtype_out=np.float64,
             self.name = 'reduce'
         self.reduce_expr = reduce_expr
         self.dtype_out = dtype_out
-        self.type = dtype_to_ctype(dtype_out)
+        self.type = dtype_to_ctype(dtype_out, backend)
         if backend == 'cython':
             # On Windows, INFINITY is not defined so we use INFTY which we
             # internally define.
@@ -446,7 +446,7 @@ def __init__(self, input=None, output=None, scan_expr="a+b",
             self.name = 'scan'
         self.scan_expr = scan_expr
         self.dtype = dtype
-        self.type = dtype_to_ctype(dtype)
+        self.type = dtype_to_ctype(dtype, backend)
         if backend == 'cython':
             # On Windows, INFINITY is not defined so we use INFTY which we
             # internally define.
@@ -463,7 +463,9 @@ def __init__(self, input=None, output=None, scan_expr="a+b",
         builtin_symbols = ['item', 'prev_item', 'last_item']
         self.builtin_types = {'i': 'int', 'N': 'int'}
         for sym in builtin_symbols:
-            self.builtin_types[sym] = dtype_to_knowntype(self.dtype)
+            self.builtin_types[sym] = dtype_to_knowntype(
+                self.dtype, backend=backend
+            )
 
     def get_type_info_from_kwargs(self, func, **kwargs):
         type_info = {}
@@ -487,7 +489,9 @@ def _generate_kernel(self, **kwargs):
         if self.input_func is not None:
             arg_types = self.get_type_info_from_kwargs(
                 self.input_func, **kwargs)
-            arg_types['return_'] = dtype_to_knowntype(self.dtype)
+            arg_types['return_'] = dtype_to_knowntype(
+                self.dtype, backend=self.backend
+            )
             helper = AnnotationHelper(self.input_func, arg_types)
             declarations.update(helper.annotate())
             self.input_func = helper.func

compyle/parallel.py

@@ -596,7 +596,7 @@ def __init__(self, reduce_expr, map_func=None, dtype_out=np.float64,
             self.name = 'reduce'
         self.reduce_expr = reduce_expr
         self.dtype_out = dtype_out
-        self.type = dtype_to_ctype(dtype_out)
+        self.type = dtype_to_ctype(dtype_out, backend=backend)
         if backend == 'cython':
             # On Windows, INFINITY is not defined so we use INFTY which we
             # internally define.
@@ -847,7 +847,7 @@ def __init__(self, input=None, output=None, scan_expr="a+b",
             self.name = 'scan'
         self.scan_expr = scan_expr
         self.dtype = dtype
-        self.type = dtype_to_ctype(dtype)
+        self.type = dtype_to_ctype(dtype, backend=backend)
         if backend == 'cython':
             # On Windows, INFINITY is not defined so we use INFTY which we
             # internally define.

compyle/types.py

@@ -1,8 +1,12 @@
 import ast
+import platform
 import sys
 import numpy as np
 
 
+BITS = platform.architecture()[0]
+
+
 def declare(type, num=1):
     """Declare the variable to be of the given type.
 
@@ -179,7 +183,7 @@ def _inject_types_in_module():
     'unsigned short': np.uint16
 }
 
-if sys.platform.startswith('win'):
+if sys.platform.startswith('win') or BITS.startswith('32bit'):
     NP_C_TYPE_MAP[np.dtype(np.int64)] = 'long long'
     NP_C_TYPE_MAP[np.dtype(np.uint64)] = 'unsigned long long'
     C_NP_TYPE_MAP['long long'] = np.int64
@@ -195,19 +199,14 @@ def _inject_types_in_module():
 NP_TYPE_LIST = list(C_NP_TYPE_MAP.values())
 
 
-def dtype_to_ctype(dtype):
-    try:
-        # FIXME: pyopencl depency
-
-        from pyopencl.compyte.dtypes import \
-            dtype_to_ctype as dtype_to_ctype_pyopencl
-        ctype = dtype_to_ctype_pyopencl(dtype)
-    except ValueError:
-        pass
-    except ImportError:
-        pass
-    else:
-        return ctype
+def dtype_to_ctype(dtype, backend=None):
+    if backend in ('opencl', 'cuda'):
+        try:
+            from pyopencl.compyte.dtypes import \
+                dtype_to_ctype as d2c_opencl
+            return d2c_opencl(dtype)
+        except (ValueError, ImportError):
+            pass
     dtype = np.dtype(dtype)
     return NP_C_TYPE_MAP[dtype]
 
@@ -224,8 +223,8 @@ def knowntype_to_ctype(knowntype):
         raise ValueError("Not a vaild known type")
 
 
-def dtype_to_knowntype(dtype, address='scalar'):
-    ctype = dtype_to_ctype(dtype)
+def dtype_to_knowntype(dtype, address='scalar', backend=None):
+    ctype = dtype_to_ctype(dtype, backend=backend)
     if 'unsigned' in ctype:
         ctype = 'u%s' % ctype.replace('unsigned ', '')
     knowntype = ctype.replace(' ', '')