numba-cuda 0.3.0__tar.gz → 0.4.0__tar.gz

{numba_cuda-0.3.0 → numba_cuda-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: numba-cuda
-Version: 0.3.0
+Version: 0.4.0
 Summary: CUDA target for Numba
 Author: Anaconda Inc., NVIDIA Corporation
 License: BSD 2-clause

numba_cuda-0.4.0/numba_cuda/VERSION

@@ -0,0 +1 @@
+0.4.0

{numba_cuda-0.3.0 → numba_cuda-0.4.0}/numba_cuda/numba/cuda/cudadrv/devicearray.py

@@ -570,10 +570,13 @@ class DeviceNDArray(DeviceNDArrayBase):
         '''
         return self._dummy.is_c_contig
 
-    def __array__(self, dtype=None):
+    def __array__(self, dtype=None, copy=None):
         """
         :return: an `numpy.ndarray`, so copies to the host.
         """
+        if copy is False:
+            msg = "`copy=False` is not supported. A copy is always created."
+            raise ValueError(msg)
         if dtype:
             return self.copy_to_host().__array__(dtype)
         else:

{numba_cuda-0.3.0 → numba_cuda-0.4.0}/numba_cuda/numba/cuda/cudadrv/driver.py

@@ -18,7 +18,6 @@ import functools
 import warnings
 import logging
 import threading
-import traceback
 import asyncio
 import pathlib
 import subprocess
@@ -40,6 +39,7 @@ from .drvapi import API_PROTOTYPES
 from .drvapi import cu_occupancy_b2d_size, cu_stream_callback_pyobj, cu_uuid
 from .mappings import FILE_EXTENSION_MAP
 from .linkable_code import LinkableCode, LTOIR, Fatbin, Object
+from numba.cuda.utils import _readenv
 from numba.cuda.cudadrv import enums, drvapi, nvrtc
 
 try:
@@ -66,25 +66,6 @@ _py_decref.argtypes = [ctypes.py_object]
 _py_incref.argtypes = [ctypes.py_object]
 
 
-def _readenv(name, ctor, default):
-    value = os.environ.get(name)
-    if value is None:
-        return default() if callable(default) else default
-    try:
-        if ctor is bool:
-            return value.lower() in {'1', "true"}
-        return ctor(value)
-    except Exception:
-        warnings.warn(
-            f"Environment variable '{name}' is defined but its associated "
-            f"value '{value}' could not be parsed.\n"
-            "The parse failed with exception:\n"
-            f"{traceback.format_exc()}",
-            RuntimeWarning
-        )
-        return default
-
-
 _MVC_ERROR_MESSAGE = (
     "Minor version compatibility requires ptxcompiler and cubinlinker packages "
     "to be available"

{numba_cuda-0.3.0 → numba_cuda-0.4.0}/numba_cuda/numba/cuda/cudadrv/nvrtc.py

@@ -266,7 +266,11 @@ def compile(src, name, cc, ltoir=False):
     cudadrv_path = os.path.dirname(os.path.abspath(__file__))
     numba_cuda_path = os.path.dirname(cudadrv_path)
     numba_include = f'-I{numba_cuda_path}'
-    options = [arch, *cuda_include, numba_include, '-rdc', 'true']
+
+    nrt_path = os.path.join(numba_cuda_path, "runtime")
+    nrt_include = f'-I{nrt_path}'
+
+    options = [arch, *cuda_include, numba_include, nrt_include, '-rdc', 'true']
 
     if ltoir:
         options.append("-dlto")

{numba_cuda-0.3.0 → numba_cuda-0.4.0}/numba_cuda/numba/cuda/dispatcher.py

@@ -21,6 +21,7 @@ from numba.cuda.descriptor import cuda_target
 from numba.cuda.errors import (missing_launch_config_msg,
                                normalize_kernel_dimensions)
 from numba.cuda import types as cuda_types
+from numba.cuda.runtime.nrt import rtsys
 
 from numba import cuda
 from numba import _dispatcher
@@ -253,7 +254,14 @@ class _Kernel(serialize.ReduceMixin):
         """
         Force binding to current CUDA context
         """
-        self._codelibrary.get_cufunc()
+        cufunc = self._codelibrary.get_cufunc()
+
+        if hasattr(self, "target_context") and self.target_context.enable_nrt:
+            rtsys.ensure_initialized()
+            rtsys.set_memsys_to_module(cufunc.module)
+            # We don't know which stream the kernel will be launched on, so
+            # we force synchronize here.
+            cuda.synchronize()
 
     @property
     def regs_per_thread(self):

numba_cuda-0.4.0/numba_cuda/numba/cuda/runtime/__init__.py

@@ -0,0 +1 @@
+from numba.cuda.runtime.nrt import rtsys # noqa: F401

numba_cuda-0.4.0/numba_cuda/numba/cuda/runtime/memsys.cu

@@ -0,0 +1,94 @@
+#include "memsys.cuh"
+
+__device__ size_t memsys_size = sizeof(NRT_MemSys);
+
+namespace detail
+{
+  void __device__ check_memsys()
+  {
+    if (TheMSys == nullptr)
+    {
+      assert(false && "TheMSys pointer is null. Please use NRT_MemSys_set to set pointer first.");
+    }
+  }
+}
+
+extern "C" __global__ void NRT_MemSys_set(NRT_MemSys *memsys_ptr)
+{
+  TheMSys = memsys_ptr;
+}
+
+extern "C" __global__ void NRT_MemSys_read(uint64_t *managed_memsys)
+{
+  detail::check_memsys();
+  managed_memsys[0] = TheMSys->stats.alloc;
+  managed_memsys[1] = TheMSys->stats.free;
+  managed_memsys[2] = TheMSys->stats.mi_alloc;
+  managed_memsys[3] = TheMSys->stats.mi_free;
+}
+
+extern "C" __global__ void NRT_MemSys_read_alloc(uint64_t *managed_result)
+{
+  detail::check_memsys();
+  managed_result[0] = TheMSys->stats.alloc;
+}
+
+extern "C" __global__ void NRT_MemSys_read_free(uint64_t *managed_result)
+{
+  detail::check_memsys();
+  managed_result[0] = TheMSys->stats.free;
+}
+
+extern "C" __global__ void NRT_MemSys_read_mi_alloc(uint64_t *managed_result)
+{
+  detail::check_memsys();
+  managed_result[0] = TheMSys->stats.mi_alloc;
+}
+
+extern "C" __global__ void NRT_MemSys_read_mi_free(uint64_t *managed_result)
+{
+  detail::check_memsys();
+  managed_result[0] = TheMSys->stats.mi_free;
+}
+
+extern "C" __global__ void NRT_MemSys_init(void)
+{
+  detail::check_memsys();
+  TheMSys->stats.enabled = false;
+  TheMSys->stats.alloc = 0;
+  TheMSys->stats.free = 0;
+  TheMSys->stats.mi_alloc = 0;
+  TheMSys->stats.mi_free = 0;
+}
+
+extern "C" __global__ void NRT_MemSys_enable_stats(void)
+{
+  detail::check_memsys();
+  TheMSys->stats.enabled = true;
+}
+
+extern "C" __global__ void NRT_MemSys_disable_stats(void)
+{
+  detail::check_memsys();
+  TheMSys->stats.enabled = false;
+}
+
+extern "C" __global__ void NRT_MemSys_stats_enabled(uint8_t *enabled)
+{
+  detail::check_memsys();
+  *enabled = static_cast<uint8_t>(TheMSys->stats.enabled);
+}
+
+extern "C" __global__ void NRT_MemSys_print(void)
+{
+  if (TheMSys != nullptr)
+  {
+    printf("TheMSys->stats.enabled %d\n", TheMSys->stats.enabled);
+    printf("TheMSys->stats.alloc %lu\n", TheMSys->stats.alloc.load());
+    printf("TheMSys->stats.free %lu\n", TheMSys->stats.free.load());
+    printf("TheMSys->stats.mi_alloc %lu\n", TheMSys->stats.mi_alloc.load());
+    printf("TheMSys->stats.mi_free %lu\n", TheMSys->stats.mi_free.load());
+  } else {
+    printf("TheMsys is null.\n");
+  }
+}

numba_cuda-0.4.0/numba_cuda/numba/cuda/runtime/memsys.cuh

@@ -0,0 +1,17 @@
+#include <cuda/atomic>
+
+// Globally needed variables
+struct NRT_MemSys {
+    struct {
+      bool enabled;
+      cuda::atomic<size_t, cuda::thread_scope_device> alloc;
+      cuda::atomic<size_t, cuda::thread_scope_device> free;
+      cuda::atomic<size_t, cuda::thread_scope_device> mi_alloc;
+      cuda::atomic<size_t, cuda::thread_scope_device> mi_free;
+    } stats;
+  };
+
+/* The Memory System object */
+__device__ NRT_MemSys* TheMSys;
+
+extern "C" __global__ void NRT_MemSys_set(NRT_MemSys *memsys_ptr);

{numba_cuda-0.3.0 → numba_cuda-0.4.0}/numba_cuda/numba/cuda/runtime/nrt.cu

@@ -3,6 +3,8 @@
 
 #include <cuda/atomic>
 
+#include "memsys.cuh"
+
 typedef void (*NRT_dtor_function)(void* ptr, size_t size, void* info);
 typedef void (*NRT_dealloc_func)(void* ptr, void* dealloc_info);
 
@@ -18,29 +20,21 @@ struct MemInfo {
 };
 }
 
-// Globally needed variables
-struct NRT_MemSys {
-  struct {
-    bool enabled;
-    cuda::atomic<size_t, cuda::thread_scope_device> alloc;
-    cuda::atomic<size_t, cuda::thread_scope_device> free;
-    cuda::atomic<size_t, cuda::thread_scope_device> mi_alloc;
-    cuda::atomic<size_t, cuda::thread_scope_device> mi_free;
-  } stats;
-};
+extern "C" __global__ void NRT_MemSys_set(NRT_MemSys *memsys_ptr)
+{
+  TheMSys = memsys_ptr;
+}
 
 static __device__ void *nrt_allocate_meminfo_and_data_align(size_t size, unsigned align, NRT_MemInfo **mi);
 static __device__ void *nrt_allocate_meminfo_and_data(size_t size, NRT_MemInfo **mi_out);
 extern "C" __device__ void* NRT_Allocate_External(size_t size);
 
-/* The Memory System object */
-__device__ NRT_MemSys* TheMSys;
-
 extern "C" __device__ void* NRT_Allocate(size_t size)
 {
   void* ptr = NULL;
   ptr       = malloc(size);
-//  if (TheMSys->stats.enabled) { TheMSys->stats.alloc++; }
+  if (TheMSys && TheMSys->stats.enabled) { 
+    TheMSys->stats.alloc.fetch_add(1, cuda::memory_order_relaxed); }
   return ptr;
 }
 
@@ -49,14 +43,14 @@ extern "C" __device__ void NRT_MemInfo_init(NRT_MemInfo* mi,
                                             size_t size,
                                             NRT_dtor_function dtor,
                                             void* dtor_info)
-//                                            NRT_MemSys* TheMSys)
 {
   mi->refct     = 1; /* starts with 1 refct */
   mi->dtor      = dtor;
   mi->dtor_info = dtor_info;
   mi->data      = data;
   mi->size      = size;
-//  if (TheMSys->stats.enabled) { TheMSys->stats.mi_alloc++; }
+ if (TheMSys && TheMSys->stats.enabled) { 
+  TheMSys->stats.mi_alloc.fetch_add(1, cuda::memory_order_relaxed); }
 }
 
 extern "C"
@@ -71,7 +65,8 @@ __device__ NRT_MemInfo* NRT_MemInfo_new(
 extern "C" __device__ void NRT_Free(void* ptr)
 {
   free(ptr);
-  //if (TheMSys->stats.enabled) { TheMSys->stats.free++; }
+  if (TheMSys && TheMSys->stats.enabled) {
+    TheMSys->stats.free.fetch_add(1, cuda::memory_order_relaxed); }
 }
 
 extern "C" __device__ void NRT_dealloc(NRT_MemInfo* mi)
@@ -82,8 +77,10 @@ extern "C" __device__ void NRT_dealloc(NRT_MemInfo* mi)
 extern "C" __device__ void NRT_MemInfo_destroy(NRT_MemInfo* mi)
 {
   NRT_dealloc(mi);
-  //if (TheMSys->stats.enabled) { TheMSys->stats.mi_free++; }
+  if (TheMSys && TheMSys->stats.enabled) { 
+    TheMSys->stats.mi_free.fetch_add(1, cuda::memory_order_relaxed); }
 }
+
 extern "C" __device__ void NRT_MemInfo_call_dtor(NRT_MemInfo* mi)
 {
   if (mi->dtor) /* We have a destructor */
@@ -158,10 +155,10 @@ extern "C" __device__ void* NRT_Allocate_External(size_t size) {
     ptr = malloc(size);
     //NRT_Debug(nrt_debug_print("NRT_Allocate_External bytes=%zu ptr=%p\n", size, ptr));
 
-    //if (TheMSys.stats.enabled)
-    //{
-    //    TheMSys.stats.alloc++;
-    //}
+    if (TheMSys && TheMSys->stats.enabled)
+    {
+       TheMSys->stats.alloc.fetch_add(1, cuda::memory_order_relaxed);
+    }
     return ptr;
 }
 

numba_cuda-0.4.0/numba_cuda/numba/cuda/runtime/nrt.py

@@ -0,0 +1,318 @@
+import ctypes
+import os
+from functools import wraps
+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 import devices
+from numba.cuda.api import get_current_device
+from numba.cuda.utils import _readenv
+
+
+# Check environment variable or config for NRT statistics enablement
+NRT_STATS = (
+    _readenv("NUMBA_CUDA_NRT_STATS", bool, False) or
+    getattr(config, "NUMBA_CUDA_NRT_STATS", False)
+)
+if not hasattr(config, "NUMBA_CUDA_NRT_STATS"):
+    config.CUDA_NRT_STATS = NRT_STATS
+
+
+# Check environment variable or config for NRT enablement
+ENABLE_NRT = (
+    _readenv("NUMBA_CUDA_ENABLE_NRT", bool, False) or
+    getattr(config, "NUMBA_CUDA_ENABLE_NRT", False)
+)
+if not hasattr(config, "NUMBA_CUDA_ENABLE_NRT"):
+    config.CUDA_ENABLE_NRT = ENABLE_NRT
+
+
+# Protect method to ensure NRT memory allocation and initialization
+def _alloc_init_guard(method):
+    """
+    Ensure NRT memory allocation and initialization before running the method
+    """
+    @wraps(method)
+    def wrapper(self, *args, **kwargs):
+        self.ensure_allocated()
+        self.ensure_initialized()
+        return method(self, *args, **kwargs)
+    return wrapper
+
+
+class _Runtime:
+    """Singleton class for Numba CUDA runtime"""
+    _instance = None
+
+    def __new__(cls, *args, **kwargs):
+        if cls._instance is None:
+            cls._instance = super(_Runtime, cls).__new__(cls, *args, **kwargs)
+        return cls._instance
+
+    def __init__(self):
+        """Initialize memsys module and variable"""
+        self._memsys_module = None
+        self._memsys = None
+        self._initialized = False
+
+    def _compile_memsys_module(self):
+        """
+        Compile memsys.cu and create a module from it in the current context
+        """
+        # Define the path for memsys.cu
+        memsys_mod = os.path.join(
+            os.path.dirname(os.path.abspath(__file__)),
+            "memsys.cu"
+        )
+        cc = get_current_device().compute_capability
+
+        # Create a new linker instance and add the cu file
+        linker = Linker.new(cc=cc)
+        linker.add_cu_file(memsys_mod)
+
+        # Complete the linker and create a module from it
+        cubin = linker.complete()
+        ctx = devices.get_context()
+        module = ctx.create_module_image(cubin)
+
+        # Set the memsys module
+        self._memsys_module = module
+
+    def ensure_allocated(self, stream=None):
+        """
+        If memsys is not allocated, allocate it; otherwise, perform a no-op
+        """
+        if self._memsys is not None:
+            return
+
+        # Allocate the memsys
+        self.allocate(stream)
+
+    def allocate(self, stream=None):
+        """
+        Allocate memsys on global memory
+        """
+        from numba.cuda import device_array
+
+        # Check if memsys module is defined
+        if self._memsys_module is None:
+            # Compile the memsys module if not defined
+            self._compile_memsys_module()
+
+        # Allocate space for NRT_MemSys
+        ptr, nbytes = self._memsys_module.get_global_symbol("memsys_size")
+        memsys_size = ctypes.c_uint64()
+        driver.cuMemcpyDtoH(ctypes.addressof(memsys_size),
+                            ptr.device_ctypes_pointer, nbytes)
+        self._memsys = device_array(
+            (memsys_size.value,), dtype="i1", stream=stream)
+        self.set_memsys_to_module(self._memsys_module, stream=stream)
+
+    def _single_thread_launch(self, module, stream, name, params=()):
+        """
+        Launch the specified kernel with only 1 thread
+        """
+        if stream is None:
+            stream = cuda.default_stream()
+
+        func = module.get_function(name)
+        launch_kernel(
+            func.handle,
+            1, 1, 1,
+            1, 1, 1,
+            0,
+            stream.handle,
+            params,
+            cooperative=False
+        )
+
+    def ensure_initialized(self, stream=None):
+        """
+        If memsys is not initialized, initialize memsys
+        """
+        if self._initialized:
+            return
+
+        # Initialize the memsys
+        self.initialize(stream)
+
+    def initialize(self, stream=None):
+        """
+        Launch memsys initialization kernel
+        """
+        self.ensure_allocated()
+
+        self._single_thread_launch(
+            self._memsys_module, stream, "NRT_MemSys_init")
+        self._initialized = True
+
+        if config.CUDA_NRT_STATS:
+            self.memsys_enable_stats()
+
+    @_alloc_init_guard
+    def memsys_enable_stats(self, stream=None):
+        """
+        Enable memsys statistics
+        """
+        self._single_thread_launch(
+            self._memsys_module, stream, "NRT_MemSys_enable_stats")
+
+    @_alloc_init_guard
+    def memsys_disable_stats(self, stream=None):
+        """
+        Disable memsys statistics
+        """
+        self._single_thread_launch(
+            self._memsys_module, stream, "NRT_MemSys_disable_stats")
+
+    @_alloc_init_guard
+    def memsys_stats_enabled(self, stream=None):
+        """
+        Return a boolean indicating whether memsys is enabled. Synchronizes
+        context
+        """
+        enabled_ar = cuda.managed_array(1, np.uint8)
+
+        self._single_thread_launch(
+            self._memsys_module,
+            stream,
+            "NRT_MemSys_stats_enabled",
+            (enabled_ar.device_ctypes_pointer,)
+        )
+
+        cuda.synchronize()
+        return bool(enabled_ar[0])
+
+    @_alloc_init_guard
+    def _copy_memsys_to_host(self, stream):
+        """
+        Copy all statistics of memsys to the host
+        """
+        dt = np.dtype([
+            ('alloc', np.uint64),
+            ('free', np.uint64),
+            ('mi_alloc', np.uint64),
+            ('mi_free', np.uint64)
+        ])
+
+        stats_for_read = cuda.managed_array(1, dt)
+
+        self._single_thread_launch(
+            self._memsys_module,
+            stream,
+            "NRT_MemSys_read",
+            [stats_for_read.device_ctypes_pointer]
+        )
+        cuda.synchronize()
+
+        return stats_for_read[0]
+
+    @_alloc_init_guard
+    def get_allocation_stats(self, stream=None):
+        """
+        Get the allocation statistics
+        """
+        enabled = self.memsys_stats_enabled(stream)
+        if not enabled:
+            raise RuntimeError("NRT stats are disabled.")
+        memsys = self._copy_memsys_to_host(stream)
+        return _nrt_mstats(
+            alloc=memsys["alloc"],
+            free=memsys["free"],
+            mi_alloc=memsys["mi_alloc"],
+            mi_free=memsys["mi_free"]
+        )
+
+    @_alloc_init_guard
+    def _get_single_stat(self, stat, stream=None):
+        """
+        Get a single stat from the memsys
+        """
+        got = cuda.managed_array(1, np.uint64)
+        self._single_thread_launch(
+            self._memsys_module,
+            stream,
+            f"NRT_MemSys_read_{stat}",
+            [got.device_ctypes_pointer]
+        )
+
+        cuda.synchronize()
+        return got[0]
+
+    @_alloc_init_guard
+    def memsys_get_stats_alloc(self, stream=None):
+        """
+        Get the allocation statistic
+        """
+        enabled = self.memsys_stats_enabled(stream)
+        if not enabled:
+            raise RuntimeError("NRT stats are disabled.")
+
+        return self._get_single_stat("alloc")
+
+    @_alloc_init_guard
+    def memsys_get_stats_free(self, stream=None):
+        """
+        Get the free statistic
+        """
+        enabled = self.memsys_stats_enabled(stream)
+        if not enabled:
+            raise RuntimeError("NRT stats are disabled.")
+
+        return self._get_single_stat("free")
+
+    @_alloc_init_guard
+    def memsys_get_stats_mi_alloc(self, stream=None):
+        """
+        Get the mi alloc statistic
+        """
+        enabled = self.memsys_stats_enabled(stream)
+        if not enabled:
+            raise RuntimeError("NRT stats are disabled.")
+
+        return self._get_single_stat("mi_alloc")
+
+    @_alloc_init_guard
+    def memsys_get_stats_mi_free(self, stream=None):
+        """
+        Get the mi free statistic
+        """
+        enabled = self.memsys_stats_enabled(stream)
+        if not enabled:
+            raise RuntimeError("NRT stats are disabled.")
+
+        return self._get_single_stat("mi_free")
+
+    def set_memsys_to_module(self, module, stream=None):
+        """
+        Set the memsys module. The module must contain `NRT_MemSys_set` kernel,
+        and declare a pointer to NRT_MemSys structure.
+        """
+        if self._memsys is None:
+            raise RuntimeError(
+                "Please allocate NRT Memsys first before setting to module.")
+
+        self._single_thread_launch(
+            module,
+            stream,
+            "NRT_MemSys_set",
+            [self._memsys.device_ctypes_pointer,]
+        )
+
+    @_alloc_init_guard
+    def print_memsys(self, stream=None):
+        """
+        Print the current statistics of memsys, for debugging purposes
+        """
+        cuda.synchronize()
+        self._single_thread_launch(
+            self._memsys_module,
+            stream,
+            "NRT_MemSys_print"
+        )
+
+
+# Create an instance of the runtime
+rtsys = _Runtime()

{numba_cuda-0.3.0 → numba_cuda-0.4.0}/numba_cuda/numba/cuda/tests/__init__.py

@@ -49,6 +49,7 @@ def load_tests(loader, tests, pattern):
         if gpus and gpus[0].compute_capability >= (2, 0):
             suite.addTests(load_testsuite(loader, join(this_dir, 'cudadrv')))
             suite.addTests(load_testsuite(loader, join(this_dir, 'cudapy')))
+            suite.addTests(load_testsuite(loader, join(this_dir, 'nrt')))
             suite.addTests(load_testsuite(loader, join(this_dir,
                                                        'doc_examples')))
         else:

{numba_cuda-0.3.0 → numba_cuda-0.4.0}/numba_cuda/numba/cuda/tests/cudadrv/test_cuda_ndarray.py

@@ -4,6 +4,7 @@ from numba.cuda.cudadrv import devicearray
 from numba import cuda
 from numba.cuda.testing import unittest, CUDATestCase
 from numba.cuda.testing import skip_on_cudasim
+from numba.tests.support import IS_NUMPY_2
 
 
 class TestCudaNDArray(CUDATestCase):
@@ -456,6 +457,36 @@ class TestCudaNDArray(CUDATestCase):
         dev_array_from_host.copy_to_device(dev_array)
 
 
+class TestArrayMethod(CUDATestCase):
+    """Tests of the __array__() method via np.array"""
+
+    def test_np_array(self):
+        dev_array = cuda.to_device(np.asarray([1.0, 2.0, 3.0]))
+        host_array = np.array(dev_array)
+        np.testing.assert_equal(dev_array.copy_to_host(), host_array)
+
+    def test_np_array_dtype(self):
+        dtype = np.int32
+        dev_array = cuda.to_device(np.asarray([1.0, 2.0, 3.0]))
+        host_array = np.array(dev_array, dtype=dtype)
+        np.testing.assert_equal(
+            host_array,
+            dev_array.copy_to_host().astype(dtype)
+        )
+
+    @unittest.skipUnless(IS_NUMPY_2, "NumPy 1.x does not pass copy kwarg")
+    def test_np_array_copy_false(self):
+        dev_array = cuda.to_device(np.asarray([1.0, 2.0, 3.0]))
+        with self.assertRaisesRegex(ValueError, "`copy=False` is not"):
+            np.array(dev_array, copy=False)
+
+    @unittest.skipUnless(IS_NUMPY_2, "NumPy 1.x does not pass copy kwarg")
+    def test_np_array_copy_true(self):
+        dev_array = cuda.to_device(np.asarray([1.0, 2.0, 3.0]))
+        host_array = np.array(dev_array)
+        np.testing.assert_equal(dev_array.copy_to_host(), host_array)
+
+
 class TestRecarray(CUDATestCase):
     def test_recarray(self):
         # From issue #4111