Expose device_func decorator at numba_dpex.experimental module.

    - To get the experimental device_func decorator to be usable
      standalone the decorator is now available via the
      numba_dpex.experimental module.
    - The DpexExpKernelTarget allows dynamic_globals to make it
      possible to call device_func from kernel.

Author: Diptorup Deb

numba_dpex/experimental/__init__.py

@@ -8,7 +8,7 @@
 
 from numba.core.imputils import Registry
 
-from .decorators import kernel
+from .decorators import device_func, kernel
 from .kernel_dispatcher import KernelDispatcher
 from .launcher import call_kernel, call_kernel_async
 from .literal_intenum_type import IntEnumLiteral
@@ -28,6 +28,7 @@ def dpex_dispatcher_const(context):
 
 
 __all__ = [
+    "device_func",
     "kernel",
     "call_kernel",
     "call_kernel_async",

numba_dpex/experimental/target.py

@@ -89,6 +89,8 @@ class DpexExpKernelTargetContext(DpexKernelTargetContext):
     they are stable enough to be migrated to DpexKernelTargetContext.
     """
 
+    allow_dynamic_globals = True
+
     def __init__(self, typingctx, target=DPEX_KERNEL_EXP_TARGET_NAME):
         super().__init__(typingctx, target)
         self.data_model_manager = exp_dmm