numba-cuda 0.6.0__tar.gz → 0.8.0__tar.gz

{numba_cuda-0.6.0 → numba_cuda-0.8.0}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: numba-cuda
-Version: 0.6.0
+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-0.6.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/compiler.py

@@ -427,6 +427,8 @@ def kernel_fixup(kernel, debug):
                         if tm_name == 'types':
                             types = tm_value
                             types.operands = types.operands[1:]
+                            if config.DUMP_LLVM:
+                                types._clear_string_cache()
 
     # Mark as a kernel for NVVM
 

{numba_cuda-0.6.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/cudadrv/nvvm.py

@@ -199,12 +199,52 @@ class NVVM(object):
 
 
 class CompilationUnit(object):
-    def __init__(self):
+    """
+    A CompilationUnit is a set of LLVM modules that are compiled to PTX or
+    LTO-IR with NVVM.
+
+    Compilation options are accepted as a dict mapping option names to values,
+    with the following considerations:
+
+    - Underscores (`_`) in option names are converted to dashes (`-`), to match
+      NVVM's option name format.
+    - Options that take a value will be emitted in the form "-<name>=<value>".
+    - Booleans passed as option values will be converted to integers.
+    - Options which take no value (such as `-gen-lto`) should have a value of
+      `None` and will be emitted in the form "-<name>".
+
+    For documentation on NVVM compilation options, see the CUDA Toolkit
+    Documentation:
+
+    https://docs.nvidia.com/cuda/libnvvm-api/index.html#_CPPv418nvvmCompileProgram11nvvmProgramiPPKc
+    """
+
+    def __init__(self, options):
         self.driver = NVVM()
         self._handle = nvvm_program()
         err = self.driver.nvvmCreateProgram(byref(self._handle))
         self.driver.check_error(err, 'Failed to create CU')
 
+        def stringify_option(k, v):
+            k = k.replace('_', '-')
+
+            if v is None:
+                return f'-{k}'.encode('utf-8')
+
+            if isinstance(v, bool):
+                v = int(v)
+
+            return f'-{k}={v}'.encode('utf-8')
+
+        options = [stringify_option(k, v) for k, v in options.items()]
+        option_ptrs = (c_char_p * len(options))(*[c_char_p(x) for x in options])
+
+        # We keep both the options and the pointers to them so that options are
+        # not destroyed before we've used their values
+        self.options = options
+        self.option_ptrs = option_ptrs
+        self.n_options = len(options)
+
     def __del__(self):
         driver = NVVM()
         err = driver.nvvmDestroyProgram(byref(self._handle))
@@ -230,60 +270,35 @@ class CompilationUnit(object):
                                                      len(buffer), None)
         self.driver.check_error(err, 'Failed to add module')
 
-    def compile(self, **options):
-        """Perform Compilation.
-
-        Compilation options are accepted as keyword arguments, with the
-        following considerations:
-
-        - Underscores (`_`) in option names are converted to dashes (`-`), to
-          match NVVM's option name format.
-        - Options that take a value will be emitted in the form
-          "-<name>=<value>".
-        - Booleans passed as option values will be converted to integers.
-        - Options which take no value (such as `-gen-lto`) should have a value
-          of `None` passed in and will be emitted in the form "-<name>".
-
-        For documentation on NVVM compilation options, see the CUDA Toolkit
-        Documentation:
-
-        https://docs.nvidia.com/cuda/libnvvm-api/index.html#_CPPv418nvvmCompileProgram11nvvmProgramiPPKc
+    def verify(self):
         """
-
-        def stringify_option(k, v):
-            k = k.replace('_', '-')
-
-            if v is None:
-                return f'-{k}'
-
-            if isinstance(v, bool):
-                v = int(v)
-
-            return f'-{k}={v}'
-
-        options = [stringify_option(k, v) for k, v in options.items()]
-
-        c_opts = (c_char_p * len(options))(*[c_char_p(x.encode('utf8'))
-                                             for x in options])
-        # verify
-        err = self.driver.nvvmVerifyProgram(self._handle, len(options), c_opts)
+        Run the NVVM verifier on all code added to the compilation unit.
+        """
+        err = self.driver.nvvmVerifyProgram(self._handle, self.n_options,
+                                            self.option_ptrs)
         self._try_error(err, 'Failed to verify\n')
 
-        # compile
-        err = self.driver.nvvmCompileProgram(self._handle, len(options), c_opts)
+    def compile(self):
+        """
+        Compile all modules added to the compilation unit and return the
+        resulting PTX or LTO-IR (depending on the options).
+        """
+        err = self.driver.nvvmCompileProgram(self._handle, self.n_options,
+                                             self.option_ptrs)
         self._try_error(err, 'Failed to compile\n')
 
-        # get result
-        reslen = c_size_t()
-        err = self.driver.nvvmGetCompiledResultSize(self._handle, byref(reslen))
+        # Get result
+        result_size = c_size_t()
+        err = self.driver.nvvmGetCompiledResultSize(self._handle,
+                                                    byref(result_size))
 
         self._try_error(err, 'Failed to get size of compiled result.')
 
-        output_buffer = (c_char * reslen.value)()
+        output_buffer = (c_char * result_size.value)()
         err = self.driver.nvvmGetCompiledResult(self._handle, output_buffer)
         self._try_error(err, 'Failed to get compiled result.')
 
-        # get log
+        # Get log
         self.log = self.get_log()
         if self.log:
             warnings.warn(self.log, category=NvvmWarning)
@@ -615,40 +630,44 @@ def llvm_replace(llvmir):
     for decl, fn in replacements:
         llvmir = llvmir.replace(decl, fn)
 
-    llvmir = llvm140_to_70_ir(llvmir)
+    llvmir = llvm150_to_70_ir(llvmir)
 
     return llvmir
 
 
-def compile_ir(llvmir, **opts):
+def compile_ir(llvmir, **options):
     if isinstance(llvmir, str):
         llvmir = [llvmir]
 
-    if opts.pop('fastmath', False):
-        opts.update({
+    if options.pop('fastmath', False):
+        options.update({
             'ftz': True,
             'fma': True,
             'prec_div': False,
             'prec_sqrt': False,
         })
 
-    cu = CompilationUnit()
-    libdevice = LibDevice()
+    cu = CompilationUnit(options)
 
     for mod in llvmir:
         mod = llvm_replace(mod)
         cu.add_module(mod.encode('utf8'))
+    cu.verify()
+
+    # We add libdevice following verification so that it is not subject to the
+    # verifier's requirements
+    libdevice = LibDevice()
     cu.lazy_add_module(libdevice.get())
 
-    return cu.compile(**opts)
+    return cu.compile()
 
 
 re_attributes_def = re.compile(r"^attributes #\d+ = \{ ([\w\s]+)\ }")
 
 
-def llvm140_to_70_ir(ir):
+def llvm150_to_70_ir(ir):
     """
-    Convert LLVM 14.0 IR for LLVM 7.0.
+    Convert LLVM 15.0 IR for LLVM 7.0.
     """
     buf = []
     for line in ir.splitlines():

{numba_cuda-0.6.0 → 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-0.6.0 → 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-0.6.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/tests/cudadrv/test_nvjitlink.py

@@ -261,7 +261,8 @@ class TestLinker(CUDATestCase):
 
 
 @unittest.skipIf(
-    not PYNVJITLINK_INSTALLED, reason="Pynvjitlink is not installed"
+    not PYNVJITLINK_INSTALLED or not TEST_BIN_DIR,
+    reason="pynvjitlink not enabled"
 )
 class TestLinkerUsage(CUDATestCase):
     """Test that whether pynvjitlink can be enabled by both environment variable

{numba_cuda-0.6.0 → numba_cuda-0.8.0}/numba_cuda/numba/cuda/tests/cudapy/test_debuginfo.py

@@ -1,4 +1,4 @@
-from numba.tests.support import override_config
+from numba.tests.support import (override_config, captured_stdout)
 from numba.cuda.testing import skip_on_cudasim
 from numba import cuda
 from numba.core import types
@@ -268,7 +268,7 @@ class TestCudaDebugInfo(CUDATestCase):
         three_device_fns(kernel_debug=False, leaf_debug=True)
         three_device_fns(kernel_debug=False, leaf_debug=False)
 
-    def test_kernel_args_types(self):
+    def _test_kernel_args_types(self):
         sig = (types.int32, types.int32)
 
         @cuda.jit("void(int32, int32)", debug=True, opt=False)
@@ -298,6 +298,15 @@ class TestCudaDebugInfo(CUDATestCase):
         match = re.compile(pat).search(llvm_ir)
         self.assertIsNotNone(match, msg=llvm_ir)
 
+    def test_kernel_args_types(self):
+        self._test_kernel_args_types()
+
+    def test_kernel_args_types_dump(self):
+        # see issue#135
+        with override_config('DUMP_LLVM', 1):
+            with captured_stdout():
+                self._test_kernel_args_types()
+
 
 if __name__ == '__main__':
     unittest.main()

{numba_cuda-0.6.0 → 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-0.6.0 → 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-0.6.0 → 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-0.6.0 → 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: 0.6.0
+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-0.6.0 → 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-0.6.0/numba_cuda/VERSION

@@ -1 +0,0 @@
-0.6.0