numba-cuda 0.7.0__tar.gz → 0.8.0__tar.gz

{numba_cuda-0.7.0 → numba_cuda-0.8.0}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: numba-cuda
-Version: 0.7.0
+Version: 0.8.0
 Summary: CUDA target for Numba
 Author: Anaconda Inc., NVIDIA Corporation
 License: BSD 2-clause
@@ -12,6 +12,7 @@ Requires-Python: >=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: numba>=0.59.1
+Dynamic: license-file
 
 <div align="center"><img src="docs/source/_static/numba-green-icon-rgb.svg" width="200"/></div>
 

numba_cuda-0.8.0/numba_cuda/VERSION

@@ -0,0 +1 @@
+0.8.0

{numba_cuda-0.7.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/dispatcher.py

@@ -968,6 +968,10 @@ class CUDADispatcher(Dispatcher, serialize.ReduceMixin):
 
         A (template, pysig, args, kws) tuple is returned.
         """
+        # Fold keyword arguments and resolve default values
+        pysig, args = self._compiler.fold_argument_types(args, kws)
+        kws = {}
+
         # Ensure an exactly-matching overload is available if we can
         # compile. We proceed with the typing even if we can't compile
         # because we may be able to force a cast on the caller side.

{numba_cuda-0.7.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/runtime/nrt.py

@@ -5,7 +5,8 @@ import numpy as np
 
 from numba import cuda, config
 from numba.core.runtime.nrt import _nrt_mstats
-from numba.cuda.cudadrv.driver import Linker, driver, launch_kernel
+from numba.cuda.cudadrv.driver import (Linker, driver, launch_kernel,
+                                       USE_NV_BINDING)
 from numba.cuda.cudadrv import devices
 from numba.cuda.api import get_current_device
 from numba.cuda.utils import _readenv
@@ -128,6 +129,18 @@ class _Runtime:
             cooperative=False
         )
 
+    def _ctypes_pointer(self, array):
+        """
+        Given an array, return a ctypes pointer to the data suitable for
+        passing to ``launch_kernel``.
+        """
+        ptr = array.device_ctypes_pointer
+
+        if USE_NV_BINDING:
+            ptr = ctypes.c_void_p(int(ptr))
+
+        return ptr
+
     def ensure_initialized(self, stream=None):
         """
         If memsys is not initialized, initialize memsys
@@ -174,12 +187,13 @@ class _Runtime:
         context
         """
         enabled_ar = cuda.managed_array(1, np.uint8)
+        enabled_ptr = self._ctypes_pointer(enabled_ar)
 
         self._single_thread_launch(
             self._memsys_module,
             stream,
             "NRT_MemSys_stats_enabled",
-            (enabled_ar.device_ctypes_pointer,)
+            (enabled_ptr,)
         )
 
         cuda.synchronize()
@@ -198,12 +212,13 @@ class _Runtime:
         ])
 
         stats_for_read = cuda.managed_array(1, dt)
+        stats_ptr = self._ctypes_pointer(stats_for_read)
 
         self._single_thread_launch(
             self._memsys_module,
             stream,
             "NRT_MemSys_read",
-            [stats_for_read.device_ctypes_pointer]
+            [stats_ptr]
         )
         cuda.synchronize()
 
@@ -231,11 +246,13 @@ class _Runtime:
         Get a single stat from the memsys
         """
         got = cuda.managed_array(1, np.uint64)
+        got_ptr = self._ctypes_pointer(got)
+
         self._single_thread_launch(
             self._memsys_module,
             stream,
             f"NRT_MemSys_read_{stat}",
-            [got.device_ctypes_pointer]
+            [got_ptr]
         )
 
         cuda.synchronize()
@@ -294,11 +311,13 @@ class _Runtime:
             raise RuntimeError(
                 "Please allocate NRT Memsys first before setting to module.")
 
+        memsys_ptr = self._ctypes_pointer(self._memsys)
+
         self._single_thread_launch(
             module,
             stream,
             "NRT_MemSys_set",
-            [self._memsys.device_ctypes_pointer,]
+            [memsys_ptr]
         )
 
     @_alloc_init_guard

{numba_cuda-0.7.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/tests/cudapy/test_overload.py

@@ -56,6 +56,10 @@ def target_overloaded_calls_target_overloaded():
     pass
 
 
+def default_values_and_kwargs():
+    pass
+
+
 # To recognise which functions are resolved for a call, we identify each with a
 # prime number. Each function called multiplies a value by its prime (starting
 # with the value 1), and we can check that the result is as expected based on
@@ -185,6 +189,13 @@ def ol_generic_calls_target_overloaded_cuda(x):
     return impl
 
 
+@overload(default_values_and_kwargs)
+def ol_default_values_and_kwargs(out, x, y=5, z=6):
+    def impl(out, x, y=5, z=6):
+        out[0], out[1] = x + y, z
+    return impl
+
+
 @skip_on_cudasim('Overloading not supported in cudasim')
 class TestOverload(CUDATestCase):
     def check_overload(self, kernel, expected):
@@ -330,6 +341,18 @@ class TestOverload(CUDATestCase):
         def cuda_target_attr_use(res, dummy):
             res[0] = dummy.cuda_only
 
+    def test_default_values_and_kwargs(self):
+        """
+        Test default values and kwargs.
+        """
+        @cuda.jit()
+        def kernel(a, b, out):
+            default_values_and_kwargs(out, a, z=b)
+
+        out = np.empty(2, dtype=np.int64)
+        kernel[1,1](1, 2, out)
+        self.assertEqual(tuple(out), (6, 2))
+
 
 if __name__ == '__main__':
     unittest.main()

{numba_cuda-0.7.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/tests/nrt/test_nrt.py

@@ -5,7 +5,6 @@ import numpy as np
 import unittest
 from numba.cuda.testing import CUDATestCase
 
-from numba.cuda.tests.nrt.mock_numpy import cuda_empty, cuda_ones, cuda_arange
 from numba.tests.support import run_in_subprocess, override_config
 
 from numba import cuda
@@ -24,7 +23,7 @@ class TestNrtBasic(CUDATestCase):
 
         @cuda.jit
         def g():
-            x = cuda_empty(10, np.int64)
+            x = np.empty(10, np.int64)
             f(x)
 
         g[1,1]()
@@ -37,7 +36,7 @@ class TestNrtBasic(CUDATestCase):
 
         @cuda.jit
         def g():
-            x = cuda_empty(10, np.int64)
+            x = np.empty(10, np.int64)
             f(x)
 
         g[1,1]()
@@ -66,7 +65,7 @@ class TestNrtBasic(CUDATestCase):
 
         @cuda.jit
         def g(out_ary):
-            x = cuda_empty(10, np.int64)
+            x = np.empty(10, np.int64)
             x[5] = 1
             y = f(x)
             out_ary[0] = y[0]
@@ -97,11 +96,11 @@ class TestNrtStatistics(CUDATestCase):
         src = """if 1:
         from numba import cuda
         from numba.cuda.runtime import rtsys
-        from numba.cuda.tests.nrt.mock_numpy import cuda_arange
+        import numpy as np
 
         @cuda.jit
         def foo():
-            x = cuda_arange(10)[0]
+            x = np.arange(10)[0]
 
         # initialize the NRT before use
         rtsys.initialize()
@@ -167,8 +166,8 @@ class TestNrtStatistics(CUDATestCase):
 
         @cuda.jit
         def foo():
-            tmp = cuda_ones(3)
-            arr = cuda_arange(5 * tmp[0]) # noqa: F841
+            tmp = np.ones(3)
+            arr = np.arange(5 * tmp[0]) # noqa: F841
             return None
 
         with (
@@ -230,6 +229,38 @@ class TestNrtStatistics(CUDATestCase):
                     stats_func()
                 self.assertIn("NRT stats are disabled.", str(raises.exception))
 
+    def test_read_one_stat(self):
+        @cuda.jit
+        def foo():
+            tmp = np.ones(3)
+            arr = np.arange(5 * tmp[0]) # noqa: F841
+            return None
+
+        with (
+            override_config('CUDA_ENABLE_NRT', True),
+            override_config('CUDA_NRT_STATS', True)
+        ):
+
+            # Switch on stats
+            rtsys.memsys_enable_stats()
+
+            # Launch the kernel a couple of times to increase stats
+            foo[1, 1]()
+            foo[1, 1]()
+
+            # Get stats struct and individual stats
+            stats = rtsys.get_allocation_stats()
+            stats_alloc = rtsys.memsys_get_stats_alloc()
+            stats_mi_alloc = rtsys.memsys_get_stats_mi_alloc()
+            stats_free = rtsys.memsys_get_stats_free()
+            stats_mi_free = rtsys.memsys_get_stats_mi_free()
+
+            # Check individual stats match stats struct
+            self.assertEqual(stats.alloc, stats_alloc)
+            self.assertEqual(stats.mi_alloc, stats_mi_alloc)
+            self.assertEqual(stats.free, stats_free)
+            self.assertEqual(stats.mi_free, stats_mi_free)
+
 
 if __name__ == '__main__':
     unittest.main()

{numba_cuda-0.7.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/tests/nrt/test_nrt_refct.py

@@ -4,7 +4,6 @@ from numba.tests.support import override_config
 from numba.cuda.runtime import rtsys
 from numba.cuda.tests.support import EnableNRTStatsMixin
 from numba.cuda.testing import CUDATestCase
-from numba.cuda.tests.nrt.mock_numpy import cuda_empty, cuda_empty_like
 
 from numba import cuda
 
@@ -34,7 +33,7 @@ class TestNrtRefCt(EnableNRTStatsMixin, CUDATestCase):
         @cuda.jit
         def kernel():
             for i in range(n):
-                temp = cuda_empty(2, np.float64) # noqa: F841
+                temp = np.empty(2) # noqa: F841
             return None
 
         init_stats = rtsys.get_allocation_stats()
@@ -51,7 +50,7 @@ class TestNrtRefCt(EnableNRTStatsMixin, CUDATestCase):
         @cuda.jit
         def g(n):
 
-            x = cuda_empty((n, 2), np.float64)
+            x = np.empty((n, 2))
 
             for i in range(n):
                 y = x[i]
@@ -73,13 +72,13 @@ class TestNrtRefCt(EnableNRTStatsMixin, CUDATestCase):
         """
         @cuda.jit
         def if_with_allocation_and_initialization(arr1, test1):
-            tmp_arr = cuda_empty_like(arr1)
+            tmp_arr = np.empty_like(arr1)
 
             for i in range(tmp_arr.shape[0]):
                 pass
 
             if test1:
-                cuda_empty_like(arr1)
+                np.empty_like(arr1)
 
         arr = np.random.random((5, 5))  # the values are not consumed
 

{numba_cuda-0.7.0 → numba_cuda-0.8.0}/numba_cuda.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: numba-cuda
-Version: 0.7.0
+Version: 0.8.0
 Summary: CUDA target for Numba
 Author: Anaconda Inc., NVIDIA Corporation
 License: BSD 2-clause
@@ -12,6 +12,7 @@ Requires-Python: >=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: numba>=0.59.1
+Dynamic: license-file
 
 <div align="center"><img src="docs/source/_static/numba-green-icon-rgb.svg" width="200"/></div>
 

{numba_cuda-0.7.0 → numba_cuda-0.8.0}/numba_cuda.egg-info/SOURCES.txt

@@ -245,7 +245,6 @@ numba_cuda/numba/cuda/tests/nocuda/test_import.py
 numba_cuda/numba/cuda/tests/nocuda/test_library_lookup.py
 numba_cuda/numba/cuda/tests/nocuda/test_nvvm.py
 numba_cuda/numba/cuda/tests/nrt/__init__.py
-numba_cuda/numba/cuda/tests/nrt/mock_numpy.py
 numba_cuda/numba/cuda/tests/nrt/test_nrt.py
 numba_cuda/numba/cuda/tests/nrt/test_nrt_refct.py
 numba_cuda/numba/cuda/tests/test_binary_generation/Makefile

numba_cuda-0.7.0/numba_cuda/VERSION

@@ -1 +0,0 @@
-0.7.0

numba_cuda-0.7.0/numba_cuda/numba/cuda/tests/nrt/mock_numpy.py

@@ -1,146 +0,0 @@
-import math
-
-import numpy as np
-
-from numba.core import errors, types
-from numba.core.extending import overload
-from numba.np.arrayobj import (_check_const_str_dtype, is_nonelike,
-                               ty_parse_dtype, ty_parse_shape, numpy_empty_nd,
-                               numpy_empty_like_nd)
-
-
-# Typical tests for allocation use array construction (e.g. np.zeros, np.empty,
-# etc.) to induce allocations. These don't work in the CUDA target because they
-# need keyword arguments, which are presently not supported properly in the
-# CUDA target.
-#
-# To work around this, we can define our own function, that works like
-# the desired one, except that it uses only positional arguments.
-#
-# Once the CUDA target supports keyword arguments, this workaround will no
-# longer be necessary and the tests in this module should be switched to use
-# the relevant NumPy functions instead.
-def cuda_empty(shape, dtype):
-    pass
-
-
-def cuda_empty_like(arr):
-    pass
-
-
-def cuda_arange(start):
-    pass
-
-
-def cuda_ones(shape):
-    pass
-
-
-@overload(cuda_empty)
-def ol_cuda_empty(shape, dtype):
-    _check_const_str_dtype("empty", dtype)
-    if (dtype is float or
-        (isinstance(dtype, types.Function) and dtype.typing_key is float) or
-            is_nonelike(dtype)): #default
-        nb_dtype = types.double
-    else:
-        nb_dtype = ty_parse_dtype(dtype)
-
-    ndim = ty_parse_shape(shape)
-    if nb_dtype is not None and ndim is not None:
-        retty = types.Array(dtype=nb_dtype, ndim=ndim, layout='C')
-
-        def impl(shape, dtype):
-            return numpy_empty_nd(shape, dtype, retty)
-        return impl
-    else:
-        msg = f"Cannot parse input types to function np.empty({shape}, {dtype})"
-        raise errors.TypingError(msg)
-
-
-@overload(cuda_empty_like)
-def ol_cuda_empty_like(arr):
-
-    if isinstance(arr, types.Array):
-        nb_dtype = arr.dtype
-    else:
-        nb_dtype = arr
-
-    if isinstance(arr, types.Array):
-        layout = arr.layout if arr.layout != 'A' else 'C'
-        retty = arr.copy(dtype=nb_dtype, layout=layout, readonly=False)
-    else:
-        retty = types.Array(nb_dtype, 0, 'C')
-
-    def impl(arr):
-        dtype = None
-        return numpy_empty_like_nd(arr, dtype, retty)
-    return impl
-
-
-def _arange_dtype(*args):
-    bounds = [a for a in args if not isinstance(a, types.NoneType)]
-
-    if any(isinstance(a, types.Complex) for a in bounds):
-        dtype = types.complex128
-    elif any(isinstance(a, types.Float) for a in bounds):
-        dtype = types.float64
-    else:
-        # `np.arange(10).dtype` is always `np.dtype(int)`, aka `np.int_`, which
-        # in all released versions of numpy corresponds to the C `long` type.
-        # Windows 64 is broken by default here because Numba (as of 0.47) does
-        # not differentiate between Python and NumPy integers, so a `typeof(1)`
-        # on w64 is `int64`, i.e. `intp`. This means an arange(<some int>) will
-        # be typed as arange(int64) and the following will yield int64 opposed
-        # to int32. Example: without a load of analysis to work out of the args
-        # were wrapped in NumPy int*() calls it's not possible to detect the
-        # difference between `np.arange(10)` and `np.arange(np.int64(10)`.
-        NPY_TY = getattr(types, "int%s" % (8 * np.dtype(int).itemsize))
-
-        # unliteral these types such that `max` works.
-        unliteral_bounds = [types.unliteral(x) for x in bounds]
-        dtype = max(unliteral_bounds + [NPY_TY,])
-
-    return dtype
-
-
-@overload(cuda_arange)
-def ol_cuda_arange(start):
-    """Simplified arange with just 1 argument."""
-    if (not isinstance(start, types.Number)):
-        return
-
-    start_value = getattr(start, "literal_value", None)
-
-    def impl(start):
-        # Allow for improved performance if given literal arguments.
-        lit_start = start_value if start_value is not None else start
-
-        _step = 1
-        _start, _stop = 0, lit_start
-
-        nitems_c = (_stop - _start) / _step
-        nitems_r = int(math.ceil(nitems_c.real))
-
-        # Binary operator needed for compiler branch pruning.
-        nitems = max(nitems_r, 0)
-
-        arr = cuda_empty(nitems, np.int64)
-        val = _start
-        for i in range(nitems):
-            arr[i] = val + (i * _step)
-        return arr
-
-    return impl
-
-
-@overload(cuda_ones)
-def ol_cuda_ones(shape):
-
-    def impl(shape):
-        arr = cuda_empty(shape, np.float64)
-        arr_flat = arr.flat
-        for idx in range(len(arr_flat)):
-            arr_flat[idx] = 1
-        return arr
-    return impl