numba-cuda 0.5.0__tar.gz → 0.6.0__tar.gz

{numba_cuda-0.5.0 → numba_cuda-0.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: numba-cuda
-Version: 0.5.0
+Version: 0.6.0
 Summary: CUDA target for Numba
 Author: Anaconda Inc., NVIDIA Corporation
 License: BSD 2-clause

numba_cuda-0.6.0/numba_cuda/VERSION

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

{numba_cuda-0.5.0 → numba_cuda-0.6.0}/numba_cuda/numba/cuda/compiler.py

@@ -1,14 +1,17 @@
 from llvmlite import ir
 from numba.core.typing.templates import ConcreteTemplate
+from numba.core import ir as numba_ir
 from numba.core import (cgutils, types, typing, funcdesc, config, compiler,
                         sigutils, utils)
 from numba.core.compiler import (sanitize_compile_result_entries, CompilerBase,
                                  DefaultPassBuilder, Flags, Option,
                                  CompileResult)
 from numba.core.compiler_lock import global_compiler_lock
-from numba.core.compiler_machinery import (LoweringPass,
+from numba.core.compiler_machinery import (FunctionPass, LoweringPass,
                                            PassManager, register_pass)
+from numba.core.interpreter import Interpreter
 from numba.core.errors import NumbaInvalidConfigWarning
+from numba.core.untyped_passes import TranslateByteCode
 from numba.core.typed_passes import (IRLegalization, NativeLowering,
                                      AnnotateTypes)
 from warnings import warn
@@ -143,13 +146,74 @@ class CreateLibrary(LoweringPass):
         return True
 
 
+class CUDABytecodeInterpreter(Interpreter):
+    # Based on the superclass implementation, but names the resulting variable
+    # "$bool<N>" instead of "bool<N>" - see Numba PR #9888:
+    # https://github.com/numba/numba/pull/9888
+    #
+    # This can be removed once that PR is available in an upstream Numba
+    # release.
+    def _op_JUMP_IF(self, inst, pred, iftrue):
+        brs = {
+            True: inst.get_jump_target(),
+            False: inst.next,
+        }
+        truebr = brs[iftrue]
+        falsebr = brs[not iftrue]
+
+        name = "$bool%s" % (inst.offset)
+        gv_fn = numba_ir.Global("bool", bool, loc=self.loc)
+        self.store(value=gv_fn, name=name)
+
+        callres = numba_ir.Expr.call(self.get(name), (self.get(pred),), (),
+                                     loc=self.loc)
+
+        pname = "$%spred" % (inst.offset)
+        predicate = self.store(value=callres, name=pname)
+        bra = numba_ir.Branch(cond=predicate, truebr=truebr, falsebr=falsebr,
+                              loc=self.loc)
+        self.current_block.append(bra)
+
+
+@register_pass(mutates_CFG=True, analysis_only=False)
+class CUDATranslateBytecode(FunctionPass):
+    _name = "cuda_translate_bytecode"
+
+    def __init__(self):
+        FunctionPass.__init__(self)
+
+    def run_pass(self, state):
+        func_id = state['func_id']
+        bc = state['bc']
+        interp = CUDABytecodeInterpreter(func_id)
+        func_ir = interp.interpret(bc)
+        state['func_ir'] = func_ir
+        return True
+
+
 class CUDACompiler(CompilerBase):
     def define_pipelines(self):
         dpb = DefaultPassBuilder
         pm = PassManager('cuda')
 
         untyped_passes = dpb.define_untyped_pipeline(self.state)
-        pm.passes.extend(untyped_passes.passes)
+
+        # Rather than replicating the whole untyped passes definition in
+        # numba-cuda, it seems cleaner to take the pass list and replace the
+        # TranslateBytecode pass with our own.
+
+        def replace_translate_pass(implementation, description):
+            if implementation is TranslateByteCode:
+                return (CUDATranslateBytecode, description)
+            else:
+                return (implementation, description)
+
+        cuda_untyped_passes = [
+            replace_translate_pass(implementation, description)
+            for implementation, description in untyped_passes.passes
+        ]
+
+        pm.passes.extend(cuda_untyped_passes)
 
         typed_passes = dpb.define_typed_pipeline(self.state)
         pm.passes.extend(typed_passes.passes)
@@ -352,6 +416,18 @@ def kernel_fixup(kernel, debug):
     kernel.return_value = ir.ReturnValue(kernel, ir.VoidType())
     kernel.args = kernel.args[1:]
 
+    # If debug metadata is present, remove the return value from it
+
+    if kernel_metadata := getattr(kernel, 'metadata', None):
+        if dbg_metadata := kernel_metadata.get('dbg', None):
+            for name, value in dbg_metadata.operands:
+                if name == "type":
+                    type_metadata = value
+                    for tm_name, tm_value in type_metadata.operands:
+                        if tm_name == 'types':
+                            types = tm_value
+                            types.operands = types.operands[1:]
+
     # Mark as a kernel for NVVM
 
     nvvm.set_cuda_kernel(kernel)

numba_cuda-0.6.0/numba_cuda/numba/cuda/debuginfo.py

@@ -0,0 +1,44 @@
+from llvmlite import ir
+from numba.core import types
+from numba.core.debuginfo import DIBuilder
+from numba.cuda.types import GridGroup
+
+_BYTE_SIZE = 8
+
+
+class CUDADIBuilder(DIBuilder):
+
+    def _var_type(self, lltype, size, datamodel=None):
+        is_bool = False
+        is_grid_group = False
+
+        if isinstance(lltype, ir.IntType):
+            if datamodel is None:
+                if size == 1:
+                    name = str(lltype)
+                    is_bool = True
+            else:
+                name = str(datamodel.fe_type)
+                if isinstance(datamodel.fe_type, types.Boolean):
+                    is_bool = True
+                elif isinstance(datamodel.fe_type, GridGroup):
+                    is_grid_group = True
+
+        if is_bool or is_grid_group:
+            m = self.module
+            bitsize = _BYTE_SIZE * size
+            # Boolean type workaround until upstream Numba is fixed
+            if is_bool:
+                ditok = "DW_ATE_boolean"
+            # GridGroup type should use numba.cuda implementation
+            elif is_grid_group:
+                ditok = "DW_ATE_unsigned"
+
+            return m.add_debug_info('DIBasicType', {
+                'name': name,
+                'size': bitsize,
+                'encoding': ir.DIToken(ditok),
+            })
+
+        # For other cases, use upstream Numba implementation
+        return super()._var_type(lltype, size, datamodel=datamodel)

{numba_cuda-0.5.0 → numba_cuda-0.6.0}/numba_cuda/numba/cuda/dispatcher.py

@@ -4,8 +4,9 @@ import re
 import sys
 import ctypes
 import functools
+from collections import defaultdict
 
-from numba.core import config, serialize, sigutils, types, typing, utils
+from numba.core import config, ir, serialize, sigutils, types, typing, utils
 from numba.core.caching import Cache, CacheImpl
 from numba.core.compiler_lock import global_compiler_lock
 from numba.core.dispatcher import Dispatcher
@@ -42,6 +43,55 @@ cuda_fp16_math_funcs = ['hsin', 'hcos',
 reshape_funcs = ['nocopy_empty_reshape', 'numba_attempt_nocopy_reshape']
 
 
+def get_cres_link_objects(cres):
+    """Given a compile result, return a set of all linkable code objects that
+    are required for it to be fully linked."""
+
+    link_objects = set()
+
+    # List of calls into declared device functions
+    device_func_calls = [
+        (name, v) for name, v in cres.fndesc.typemap.items() if (
+            isinstance(v, cuda_types.CUDADispatcher)
+        )
+    ]
+
+    # List of tuples with SSA name of calls and corresponding signature
+    call_signatures = [
+        (call.func.name, sig)
+        for call, sig in cres.fndesc.calltypes.items() if (
+            isinstance(call, ir.Expr) and call.op == 'call'
+        )
+    ]
+
+    # Map SSA names to all invoked signatures
+    call_signature_d = defaultdict(list)
+    for name, sig in call_signatures:
+        call_signature_d[name].append(sig)
+
+    # Add the link objects from the current function's callees
+    for name, v in device_func_calls:
+        for sig in call_signature_d.get(name, []):
+            called_cres = v.dispatcher.overloads[sig.args]
+            called_link_objects = get_cres_link_objects(called_cres)
+            link_objects.update(called_link_objects)
+
+    # From this point onwards, we are only interested in ExternFunction
+    # declarations - these are the calls made directly in this function to
+    # them.
+    for name, v in cres.fndesc.typemap.items():
+        if not isinstance(v, Function):
+            continue
+
+        if not isinstance(v.typing_key, ExternFunction):
+            continue
+
+        for obj in v.typing_key.link:
+            link_objects.add(obj)
+
+    return link_objects
+
+
 class _Kernel(serialize.ReduceMixin):
     '''
     CUDA Kernel specialized for a given set of argument types. When called, this
@@ -159,15 +209,8 @@ class _Kernel(serialize.ReduceMixin):
 
         self.maybe_link_nrt(link, tgt_ctx, asm)
 
-        for k, v in cres.fndesc.typemap.items():
-            if not isinstance(v, Function):
-                continue
-
-            if not isinstance(v.typing_key, ExternFunction):
-                continue
-
-            for obj in v.typing_key.link:
-                lib.add_linking_file(obj)
+        for obj in get_cres_link_objects(cres):
+            lib.add_linking_file(obj)
 
         for filepath in link:
             lib.add_linking_file(filepath)
@@ -267,7 +310,11 @@ class _Kernel(serialize.ReduceMixin):
         """
         cufunc = self._codelibrary.get_cufunc()
 
-        if hasattr(self, "target_context") and self.target_context.enable_nrt:
+        if (
+            hasattr(self, "target_context")
+            and self.target_context.enable_nrt
+            and config.CUDA_NRT_STATS
+        ):
             rtsys.ensure_initialized()
             rtsys.set_memsys_to_module(cufunc.module)
             # We don't know which stream the kernel will be launched on, so

{numba_cuda-0.5.0 → numba_cuda-0.6.0}/numba_cuda/numba/cuda/target.py

@@ -3,8 +3,7 @@ from functools import cached_property
 import llvmlite.binding as ll
 from llvmlite import ir
 
-from numba.core import (cgutils, config, debuginfo, itanium_mangler, types,
-                        typing, utils)
+from numba.core import cgutils, config, itanium_mangler, types, typing
 from numba.core.dispatcher import Dispatcher
 from numba.core.base import BaseContext
 from numba.core.callconv import BaseCallConv, MinimalCallConv
@@ -12,7 +11,8 @@ from numba.core.typing import cmathdecl
 from numba.core import datamodel
 
 from .cudadrv import nvvm
-from numba.cuda import codegen, nvvmutils, ufuncs
+from numba.cuda import codegen, ufuncs
+from numba.cuda.debuginfo import CUDADIBuilder
 from numba.cuda.models import cuda_data_manager
 
 # -----------------------------------------------------------------------------
@@ -80,7 +80,7 @@ class CUDATargetContext(BaseContext):
 
     @property
     def DIBuilder(self):
-        return debuginfo.DIBuilder
+        return CUDADIBuilder
 
     @property
     def enable_boundscheck(self):
@@ -150,136 +150,6 @@ class CUDATargetContext(BaseContext):
         return itanium_mangler.mangle(name, argtypes, abi_tags=abi_tags,
                                       uid=uid)
 
-    def prepare_cuda_kernel(self, codelib, fndesc, debug, lineinfo,
-                            nvvm_options, filename, linenum,
-                            max_registers=None, lto=False):
-        """
-        Adapt a code library ``codelib`` with the numba compiled CUDA kernel
-        with name ``fname`` and arguments ``argtypes`` for NVVM.
-        A new library is created with a wrapper function that can be used as
-        the kernel entry point for the given kernel.
-
-        Returns the new code library and the wrapper function.
-
-        Parameters:
-
-        codelib:       The CodeLibrary containing the device function to wrap
-                       in a kernel call.
-        fndesc:        The FunctionDescriptor of the source function.
-        debug:         Whether to compile with debug.
-        lineinfo:      Whether to emit line info.
-        nvvm_options:  Dict of NVVM options used when compiling the new library.
-        filename:      The source filename that the function is contained in.
-        linenum:       The source line that the function is on.
-        max_registers: The max_registers argument for the code library.
-        """
-        kernel_name = itanium_mangler.prepend_namespace(
-            fndesc.llvm_func_name, ns='cudapy',
-        )
-        library = self.codegen().create_library(f'{codelib.name}_kernel_',
-                                                entry_name=kernel_name,
-                                                nvvm_options=nvvm_options,
-                                                max_registers=max_registers,
-                                                lto=lto
-                                                )
-        library.add_linking_library(codelib)
-        wrapper = self.generate_kernel_wrapper(library, fndesc, kernel_name,
-                                               debug, lineinfo, filename,
-                                               linenum)
-        return library, wrapper
-
-    def generate_kernel_wrapper(self, library, fndesc, kernel_name, debug,
-                                lineinfo, filename, linenum):
-        """
-        Generate the kernel wrapper in the given ``library``.
-        The function being wrapped is described by ``fndesc``.
-        The wrapper function is returned.
-        """
-
-        argtypes = fndesc.argtypes
-        arginfo = self.get_arg_packer(argtypes)
-        argtys = list(arginfo.argument_types)
-        wrapfnty = ir.FunctionType(ir.VoidType(), argtys)
-        wrapper_module = self.create_module("cuda.kernel.wrapper")
-        fnty = ir.FunctionType(ir.IntType(32),
-                               [self.call_conv.get_return_type(types.pyobject)]
-                               + argtys)
-        func = ir.Function(wrapper_module, fnty, fndesc.llvm_func_name)
-
-        prefixed = itanium_mangler.prepend_namespace(func.name, ns='cudapy')
-        wrapfn = ir.Function(wrapper_module, wrapfnty, prefixed)
-        builder = ir.IRBuilder(wrapfn.append_basic_block(''))
-
-        if debug or lineinfo:
-            directives_only = lineinfo and not debug
-            debuginfo = self.DIBuilder(module=wrapper_module,
-                                       filepath=filename,
-                                       cgctx=self,
-                                       directives_only=directives_only)
-            debuginfo.mark_subprogram(
-                wrapfn, kernel_name, fndesc.args, argtypes, linenum,
-            )
-            debuginfo.mark_location(builder, linenum)
-
-        # Define error handling variable
-        def define_error_gv(postfix):
-            name = wrapfn.name + postfix
-            gv = cgutils.add_global_variable(wrapper_module, ir.IntType(32),
-                                             name)
-            gv.initializer = ir.Constant(gv.type.pointee, None)
-            return gv
-
-        gv_exc = define_error_gv("__errcode__")
-        gv_tid = []
-        gv_ctaid = []
-        for i in 'xyz':
-            gv_tid.append(define_error_gv("__tid%s__" % i))
-            gv_ctaid.append(define_error_gv("__ctaid%s__" % i))
-
-        callargs = arginfo.from_arguments(builder, wrapfn.args)
-        status, _ = self.call_conv.call_function(
-            builder, func, types.void, argtypes, callargs)
-
-        if debug:
-            # Check error status
-            with cgutils.if_likely(builder, status.is_ok):
-                builder.ret_void()
-
-            with builder.if_then(builder.not_(status.is_python_exc)):
-                # User exception raised
-                old = ir.Constant(gv_exc.type.pointee, None)
-
-                # Use atomic cmpxchg to prevent rewriting the error status
-                # Only the first error is recorded
-
-                xchg = builder.cmpxchg(gv_exc, old, status.code,
-                                       'monotonic', 'monotonic')
-                changed = builder.extract_value(xchg, 1)
-
-                # If the xchange is successful, save the thread ID.
-                sreg = nvvmutils.SRegBuilder(builder)
-                with builder.if_then(changed):
-                    for dim, ptr, in zip("xyz", gv_tid):
-                        val = sreg.tid(dim)
-                        builder.store(val, ptr)
-
-                    for dim, ptr, in zip("xyz", gv_ctaid):
-                        val = sreg.ctaid(dim)
-                        builder.store(val, ptr)
-
-        builder.ret_void()
-
-        nvvm.set_cuda_kernel(wrapfn)
-        library.add_ir_module(wrapper_module)
-        if debug or lineinfo:
-            debuginfo.finalize()
-        library.finalize()
-
-        if config.DUMP_LLVM:
-            utils.dump_llvm(fndesc, wrapper_module)
-
-        return library.get_function(wrapfn.name)
-
     def make_constant_array(self, builder, aryty, arr):
         """
         Unlike the parent version.  This returns a a pointer in the constant

{numba_cuda-0.5.0 → numba_cuda-0.6.0}/numba_cuda/numba/cuda/tests/cudapy/test_debuginfo.py

@@ -72,6 +72,57 @@ class TestCudaDebugInfo(CUDATestCase):
         def f(x):
             x[0] = 0
 
+    def test_issue_9888(self):
+        # Compiler created symbol should not be emitted in DILocalVariable
+        # See Numba Issue #9888 https://github.com/numba/numba/pull/9888
+        sig = (types.boolean,)
+
+        @cuda.jit(sig, debug=True, opt=False)
+        def f(cond):
+            if cond:
+                x = 1  # noqa: F841
+            else:
+                x = 0  # noqa: F841
+
+        llvm_ir = f.inspect_llvm(sig)
+        # A varible name starting with "bool" in the debug metadata
+        pat = r'!DILocalVariable\(.*name:\s+\"bool'
+        match = re.compile(pat).search(llvm_ir)
+        self.assertIsNone(match, msg=llvm_ir)
+
+    def test_bool_type(self):
+        sig = (types.int32, types.int32)
+
+        @cuda.jit("void(int32, int32)", debug=True, opt=False)
+        def f(x, y):
+            z = x == y  # noqa: F841
+
+        llvm_ir = f.inspect_llvm(sig)
+
+        # extract the metadata node id from `type` field of DILocalVariable
+        pat = r'!DILocalVariable\(.*name:\s+"z".*type:\s+!(\d+)'
+        match = re.compile(pat).search(llvm_ir)
+        self.assertIsNotNone(match, msg=llvm_ir)
+        mdnode_id = match.group(1)
+
+        # verify the DIBasicType has correct encoding attribute DW_ATE_boolean
+        pat = rf'!{mdnode_id}\s+=\s+!DIBasicType\(.*DW_ATE_boolean'
+        match = re.compile(pat).search(llvm_ir)
+        self.assertIsNotNone(match, msg=llvm_ir)
+
+    def test_grid_group_type(self):
+        sig = (types.int32,)
+
+        @cuda.jit(sig, debug=True, opt=False)
+        def f(x):
+            grid = cuda.cg.this_grid()  # noqa: F841
+
+        llvm_ir = f.inspect_llvm(sig)
+
+        pat = r'!DIBasicType\(.*DW_ATE_unsigned, name: "GridGroup", size: 64'
+        match = re.compile(pat).search(llvm_ir)
+        self.assertIsNotNone(match, msg=llvm_ir)
+
     @unittest.skip("Wrappers no longer exist")
     def test_wrapper_has_debuginfo(self):
         sig = (types.int32[::1],)
@@ -217,6 +268,36 @@ 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):
+        sig = (types.int32, types.int32)
+
+        @cuda.jit("void(int32, int32)", debug=True, opt=False)
+        def f(x, y):
+            z = x + y  # noqa: F841
+
+        llvm_ir = f.inspect_llvm(sig)
+
+        # extract the metadata node id from `types` field of DISubroutineType
+        pat = r'!DISubroutineType\(types:\s+!(\d+)\)'
+        match = re.compile(pat).search(llvm_ir)
+        self.assertIsNotNone(match, msg=llvm_ir)
+        mdnode_id = match.group(1)
+
+        # extract the metadata node ids from the flexible node of types
+        pat = rf'!{mdnode_id}\s+=\s+!{{\s+!(\d+),\s+!(\d+)\s+}}'
+        match = re.compile(pat).search(llvm_ir)
+        self.assertIsNotNone(match, msg=llvm_ir)
+        mdnode_id1 = match.group(1)
+        mdnode_id2 = match.group(2)
+
+        # verify each of the two metadata nodes match expected type
+        pat = rf'!{mdnode_id1}\s+=\s+!DIBasicType\(.*DW_ATE_signed,\s+name:\s+"int32"'  # noqa: E501
+        match = re.compile(pat).search(llvm_ir)
+        self.assertIsNotNone(match, msg=llvm_ir)
+        pat = rf'!{mdnode_id2}\s+=\s+!DIBasicType\(.*DW_ATE_signed,\s+name:\s+"int32"'  # noqa: E501
+        match = re.compile(pat).search(llvm_ir)
+        self.assertIsNotNone(match, msg=llvm_ir)
+
 
 if __name__ == '__main__':
     unittest.main()

{numba_cuda-0.5.0 → numba_cuda-0.6.0}/numba_cuda/numba/cuda/tests/cudapy/test_device_func.py

@@ -205,6 +205,14 @@ int times2(int *out, int a)
 }
 """)
 
+times3_cu = cuda.CUSource("""
+extern "C" __device__
+int times3(int *out, int a)
+{
+  *out = a * 3;
+  return 0;
+}
+""")
 
 times4_cu = cuda.CUSource("""
 extern "C" __device__
@@ -351,6 +359,123 @@ class TestDeclareDevice(CUDATestCase):
         kernel[1, 1](x, 1)
         np.testing.assert_equal(x[0], 323845807)
 
+    def test_declared_in_called_function(self):
+        times2 = cuda.declare_device('times2', 'int32(int32)', link=times2_cu)
+
+        @cuda.jit
+        def device_func(x):
+            return times2(x)
+
+        @cuda.jit
+        def kernel(r, x):
+            i = cuda.grid(1)
+            if i < len(r):
+                r[i] = device_func(x[i])
+
+        x = np.arange(10, dtype=np.int32)
+        r = np.empty_like(x)
+
+        kernel[1, 32](r, x)
+
+        np.testing.assert_equal(r, x * 2)
+
+    def test_declared_in_called_function_twice(self):
+        times2 = cuda.declare_device('times2', 'int32(int32)', link=times2_cu)
+
+        @cuda.jit
+        def device_func_1(x):
+            return times2(x)
+
+        @cuda.jit
+        def device_func_2(x):
+            return device_func_1(x)
+
+        @cuda.jit
+        def kernel(r, x):
+            i = cuda.grid(1)
+            if i < len(r):
+                r[i] = device_func_2(x[i])
+
+        x = np.arange(10, dtype=np.int32)
+        r = np.empty_like(x)
+
+        kernel[1, 32](r, x)
+
+        np.testing.assert_equal(r, x * 2)
+
+    def test_declared_in_called_function_two_calls(self):
+        times2 = cuda.declare_device('times2', 'int32(int32)', link=times2_cu)
+
+        @cuda.jit
+        def device_func(x):
+            return times2(x)
+
+        @cuda.jit
+        def kernel(r, x):
+            i = cuda.grid(1)
+            if i < len(r):
+                r[i] = device_func(x[i]) + device_func(x[i] + i)
+
+        x = np.arange(10, dtype=np.int32)
+        r = np.empty_like(x)
+
+        kernel[1, 32](r, x)
+
+        np.testing.assert_equal(r, x * 6)
+
+    def test_call_declared_function_twice(self):
+        times2 = cuda.declare_device('times2', 'int32(int32)', link=times2_cu)
+
+        @cuda.jit
+        def kernel(r, x):
+            i = cuda.grid(1)
+            if i < len(r):
+                r[i] = times2(x[i]) + times2(x[i] + i)
+
+        x = np.arange(10, dtype=np.int32)
+        r = np.empty_like(x)
+
+        kernel[1, 32](r, x)
+
+        np.testing.assert_equal(r, x * 6)
+
+    def test_declared_in_called_function_and_parent(self):
+        times2 = cuda.declare_device('times2', 'int32(int32)', link=times2_cu)
+
+        @cuda.jit
+        def device_func(x):
+            return times2(x)
+
+        @cuda.jit
+        def kernel(r, x):
+            i = cuda.grid(1)
+            if i < len(r):
+                r[i] = device_func(x[i]) + times2(x[i])
+
+        x = np.arange(10, dtype=np.int32)
+        r = np.empty_like(x)
+
+        kernel[1, 32](r, x)
+
+        np.testing.assert_equal(r, x * 4)
+
+    def test_call_two_different_declared_functions(self):
+        times2 = cuda.declare_device('times2', 'int32(int32)', link=times2_cu)
+        times3 = cuda.declare_device('times3', 'int32(int32)', link=times3_cu)
+
+        @cuda.jit
+        def kernel(r, x):
+            i = cuda.grid(1)
+            if i < len(r):
+                r[i] = times2(x[i]) + times3(x[i])
+
+        x = np.arange(10, dtype=np.int32)
+        r = np.empty_like(x)
+
+        kernel[1, 32](r, x)
+
+        np.testing.assert_equal(r, x * 5)
+
 
 if __name__ == '__main__':
     unittest.main()

{numba_cuda-0.5.0 → numba_cuda-0.6.0}/numba_cuda/numba/cuda/tests/nrt/test_nrt.py

@@ -171,7 +171,10 @@ class TestNrtStatistics(CUDATestCase):
             arr = cuda_arange(5 * tmp[0]) # noqa: F841
             return None
 
-        with override_config('CUDA_ENABLE_NRT', True):
+        with (
+            override_config('CUDA_ENABLE_NRT', True),
+            override_config('CUDA_NRT_STATS', True)
+        ):
             # Switch on stats
             rtsys.memsys_enable_stats()
             # check the stats are on

{numba_cuda-0.5.0 → numba_cuda-0.6.0}/numba_cuda/numba/cuda/tests/nrt/test_nrt_refct.py

@@ -18,7 +18,10 @@ class TestNrtRefCt(EnableNRTStatsMixin, CUDATestCase):
         super(TestNrtRefCt, self).tearDown()
 
     def run(self, result=None):
-        with override_config("CUDA_ENABLE_NRT", True):
+        with (
+            override_config("CUDA_ENABLE_NRT", True),
+            override_config('CUDA_NRT_STATS', True)
+        ):
             super(TestNrtRefCt, self).run(result)
 
     def test_no_return(self):