numba-cuda 0.21.1__cp313-cp313-win_amd64.whl → 0.23.0__cp313-cp313-win_amd64.whl

numba_cuda/numba/cuda/lowering.py

@@ -11,6 +11,7 @@ from llvmlite import ir as llvm_ir
 from numba.cuda import HAS_NUMBA
 from numba.cuda.core import ir
 from numba.cuda import debuginfo, cgutils, utils, typing, types
+from numba import cuda
 from numba.cuda.core import (
     ir_utils,
     targetconfig,
@@ -441,7 +442,9 @@ class Lower(BaseLower):
                         # Ensure that the variable is not defined multiple times
                         # in the block
                         [defblk] = var_assign_map[var]
-                        assign_stmts = self.blocks[defblk].find_insts(ir.Assign)
+                        assign_stmts = self.blocks[defblk].find_insts(
+                            ir.assign_types
+                        )
                         assigns = [
                             stmt
                             for stmt in assign_stmts
@@ -468,7 +471,7 @@ class Lower(BaseLower):
             self.builder.position_at_end(bb)
             all_names = set()
             for block in self.blocks.values():
-                for x in block.find_insts(ir.Del):
+                for x in block.find_insts(ir.del_types):
                     if x.value not in all_names:
                         all_names.add(x.value)
             for name in all_names:
@@ -483,9 +486,9 @@ class Lower(BaseLower):
                 self.func_ir,
                 call.func,
             )
-            if defn is not None and isinstance(defn, ir.Global):
+            if defn is not None and isinstance(defn, ir.global_types):
                 if defn.value is eh.exception_check:
-                    if isinstance(block.terminator, ir.Branch):
+                    if isinstance(block.terminator, ir.branch_types):
                         targetblk = self.blkmap[block.terminator.truebr]
                         # NOTE: This hacks in an attribute for call_conv to
                         #       pick up. This hack is no longer needed when
@@ -505,19 +508,19 @@ class Lower(BaseLower):
         self.debuginfo.mark_location(self.builder, self.loc.line)
         self.notify_loc(self.loc)
         self.debug_print(str(inst))
-        if isinstance(inst, ir.Assign):
+        if isinstance(inst, ir.assign_types):
             ty = self.typeof(inst.target.name)
             val = self.lower_assign(ty, inst)
             argidx = None
             # If this is a store from an arg, like x = arg.x then tell debuginfo
             # that this is the arg
-            if isinstance(inst.value, ir.Arg):
+            if isinstance(inst.value, ir.arg_types):
                 # NOTE: debug location is the `def <func>` line
                 self.debuginfo.mark_location(self.builder, self.defn_loc.line)
                 argidx = inst.value.index + 1  # args start at 1
             self.storevar(val, inst.target.name, argidx=argidx)
 
-        elif isinstance(inst, ir.Branch):
+        elif isinstance(inst, ir.branch_types):
             cond = self.loadvar(inst.cond.name)
             tr = self.blkmap[inst.truebr]
             fl = self.blkmap[inst.falsebr]
@@ -529,11 +532,11 @@ class Lower(BaseLower):
             )
             self.builder.cbranch(pred, tr, fl)
 
-        elif isinstance(inst, ir.Jump):
+        elif isinstance(inst, ir.jump_types):
             target = self.blkmap[inst.target]
             self.builder.branch(target)
 
-        elif isinstance(inst, ir.Return):
+        elif isinstance(inst, ir.return_types):
             if self.generator_info:
                 # StopIteration
                 self.genlower.return_from_generator(self)
@@ -551,10 +554,10 @@ class Lower(BaseLower):
             retval = self.context.get_return_value(self.builder, ty, val)
             self.call_conv.return_value(self.builder, retval)
 
-        elif isinstance(inst, ir.PopBlock):
+        elif isinstance(inst, ir.popblock_types):
             pass  # this is just a marker
 
-        elif isinstance(inst, ir.StaticSetItem):
+        elif isinstance(inst, ir.staticsetitem_types):
             signature = self.fndesc.calltypes[inst]
             assert signature is not None
             try:
@@ -572,22 +575,22 @@ class Lower(BaseLower):
                 )
                 return impl(self.builder, (target, inst.index, value))
 
-        elif isinstance(inst, ir.Print):
+        elif isinstance(inst, ir.print_types):
             self.lower_print(inst)
 
-        elif isinstance(inst, ir.SetItem):
+        elif isinstance(inst, ir.setitem_types):
             signature = self.fndesc.calltypes[inst]
             assert signature is not None
             return self.lower_setitem(
                 inst.target, inst.index, inst.value, signature
             )
 
-        elif isinstance(inst, ir.StoreMap):
+        elif isinstance(inst, ir.storemap_types):
             signature = self.fndesc.calltypes[inst]
             assert signature is not None
             return self.lower_setitem(inst.dct, inst.key, inst.value, signature)
 
-        elif isinstance(inst, ir.DelItem):
+        elif isinstance(inst, ir.delitem_types):
             target = self.loadvar(inst.target.name)
             index = self.loadvar(inst.index.name)
 
@@ -613,10 +616,10 @@ class Lower(BaseLower):
 
             return impl(self.builder, (target, index))
 
-        elif isinstance(inst, ir.Del):
+        elif isinstance(inst, ir.del_types):
             self.delvar(inst.value)
 
-        elif isinstance(inst, ir.SetAttr):
+        elif isinstance(inst, ir.setattr_types):
             target = self.loadvar(inst.target.name)
             value = self.loadvar(inst.value.name)
             signature = self.fndesc.calltypes[inst]
@@ -634,16 +637,16 @@ class Lower(BaseLower):
 
             return impl(self.builder, (target, value))
 
-        elif isinstance(inst, ir.DynamicRaise):
+        elif isinstance(inst, ir.dynamicraise_types):
             self.lower_dynamic_raise(inst)
 
-        elif isinstance(inst, ir.DynamicTryRaise):
+        elif isinstance(inst, ir.dynamictryraise_types):
             self.lower_try_dynamic_raise(inst)
 
-        elif isinstance(inst, ir.StaticRaise):
+        elif isinstance(inst, ir.staticraise_types):
             self.lower_static_raise(inst)
 
-        elif isinstance(inst, ir.StaticTryRaise):
+        elif isinstance(inst, ir.statictryraise_types):
             self.lower_static_try_raise(inst)
 
         else:
@@ -695,7 +698,7 @@ class Lower(BaseLower):
         args = []
         nb_types = []
         for exc_arg in exc_args:
-            if isinstance(exc_arg, ir.Var):
+            if isinstance(exc_arg, ir.var_types):
                 # dynamic values
                 typ = self.typeof(exc_arg.name)
                 val = self.loadvar(exc_arg.name)
@@ -727,24 +730,28 @@ class Lower(BaseLower):
     def lower_assign(self, ty, inst):
         value = inst.value
         # In nopython mode, closure vars are frozen like globals
-        if isinstance(value, (ir.Const, ir.Global, ir.FreeVar)):
+        if (
+            isinstance(value, ir.const_types)
+            or isinstance(value, ir.global_types)
+            or isinstance(value, ir.freevar_types)
+        ):
             res = self.context.get_constant_generic(
                 self.builder, ty, value.value
             )
             self.incref(ty, res)
             return res
 
-        elif isinstance(value, ir.Expr):
+        elif isinstance(value, ir.expr_types):
             return self.lower_expr(ty, value)
 
-        elif isinstance(value, ir.Var):
+        elif isinstance(value, ir.var_types):
             val = self.loadvar(value.name)
             oty = self.typeof(value.name)
             res = self.context.cast(self.builder, val, oty, ty)
             self.incref(ty, res)
             return res
 
-        elif isinstance(value, ir.Arg):
+        elif isinstance(value, ir.arg_types):
             # Suspend debug info else all the arg repacking ends up being
             # associated with some line or other and it's actually just a detail
             # of Numba's CC.
@@ -770,7 +777,7 @@ class Lower(BaseLower):
                 self.incref(ty, res)
                 return res
 
-        elif isinstance(value, ir.Yield):
+        elif isinstance(value, ir.yield_types):
             res = self.lower_yield(ty, value)
             self.incref(ty, res)
             return res
@@ -1677,10 +1684,31 @@ class Lower(BaseLower):
 
 
 class CUDALower(Lower):
+    def _is_shared_array_call(self, fnty):
+        # Check if function type is a cuda.shared.array call
+        if not hasattr(fnty, "typing_key"):
+            return False
+        return fnty.typing_key is cuda.shared.array
+
+    def _lower_call_normal(self, fnty, expr, signature):
+        # Set flag for subsequent store to track shared address space
+        if self.context.enable_debuginfo and self._is_shared_array_call(fnty):
+            self._pending_shared_store = True
+
+        return super()._lower_call_normal(fnty, expr, signature)
+
     def storevar(self, value, name, argidx=None):
         """
         Store the value into the given variable.
         """
+        # Track address space for debug info
+        if self.context.enable_debuginfo and self._pending_shared_store:
+            from numba.cuda.cudadrv import nvvm
+
+            self._addrspace_map[name] = nvvm.ADDRSPACE_SHARED
+            if not name.startswith("$") and not name.startswith("."):
+                self._pending_shared_store = False
+
         # Handle polymorphic variables with CUDA_DEBUG_POLY enabled
         if config.CUDA_DEBUG_POLY:
             src_name = name.split(".")[0]
@@ -1792,7 +1820,7 @@ class CUDALower(Lower):
         self.dbg_val_names = set()
 
         if self.context.enable_debuginfo and self._disable_sroa_like_opt:
-            for x in block.find_insts(ir.Assign):
+            for x in block.find_insts(ir.assign_types):
                 if x.target.name.startswith("$"):
                     continue
                 ssa_name = x.target.name
@@ -1806,6 +1834,13 @@ class CUDALower(Lower):
         """
         super().pre_lower()
 
+        # Track address space for debug info
+        self._addrspace_map = {}
+        self._pending_shared_store = False
+        if self.context.enable_debuginfo:
+            self.debuginfo._set_addrspace_map(self._addrspace_map)
+
+        # Track polymorphic variables for debug info
         self.poly_var_typ_map = {}
         self.poly_var_loc_map = {}
         self.poly_var_set = set()
@@ -1818,7 +1853,7 @@ class CUDALower(Lower):
             poly_map = {}
             # pre-scan all blocks
             for block in self.blocks.values():
-                for x in block.find_insts(ir.Assign):
+                for x in block.find_insts(ir.assign_types):
                     if x.target.name.startswith("$"):
                         continue
                     ssa_name = x.target.name

numba_cuda/numba/cuda/memory_management/nrt.py

@@ -13,9 +13,10 @@ from numba.cuda import config, types
 from numba.cuda.cudadrv.driver import (
     _Linker,
     driver,
-    launch_kernel,
+    _to_core_stream,
     _have_nvjitlink,
 )
+from cuda.core.experimental import LaunchConfig, launch
 from numba.cuda.cudadrv import devices
 from numba.cuda.api import get_current_device
 from numba.cuda.utils import _readenv, cached_file_read
@@ -126,7 +127,7 @@ class _Runtime:
         cc = get_current_device().compute_capability
 
         # Create a new linker instance and add the cu file
-        linker = _Linker.new(cc=cc, lto=_have_nvjitlink())
+        linker = _Linker(max_registers=0, cc=cc, lto=_have_nvjitlink())
         linker.add_cu_file(memsys_mod)
 
         # Complete the linker and create a module from it
@@ -179,20 +180,15 @@ class _Runtime:
             stream = cuda.default_stream()
 
         func = module.get_function(name)
-        launch_kernel(
-            func.handle,
-            1,
-            1,
-            1,
-            1,
-            1,
-            1,
-            0,
-            stream.handle.value,
-            params,
-            cooperative=False,
+        config = LaunchConfig(
+            grid=(1, 1, 1),
+            block=(1, 1, 1),
+            shmem_size=0,
+            cooperative_launch=False,
         )
 
+        launch(_to_core_stream(stream), config, func.kernel, *params)
+
     def ensure_initialized(self, stream=None):
         """
         If memsys is not initialized, initialize memsys

numba_cuda/numba/cuda/np/arrayobj.py

@@ -31,6 +31,7 @@ from numba.cuda.np.numpy_support import (
     type_is_scalar,
     lt_complex,
     lt_floats,
+    strides_from_shape,
 )
 from numba.cuda.np.numpy_support import (
     type_can_asarray,
@@ -3642,10 +3643,63 @@ def record_static_setitem_int(context, builder, sig, args):
 def constant_array(context, builder, ty, pyval):
     """
     Create a constant array (mechanism is target-dependent).
+
+    For objects implementing __cuda_array_interface__,
+    the device pointer is embedded directly as a constant. For other arrays,
+    the target-specific mechanism is used.
     """
+    # Check if this is a device array (implements __cuda_array_interface__)
+    if getattr(pyval, "__cuda_array_interface__", None) is not None:
+        return _lower_constant_device_array(context, builder, ty, pyval)
+
     return context.make_constant_array(builder, ty, pyval)
 
 
+def _lower_constant_device_array(context, builder, ty, pyval):
+    """
+    Lower objects with __cuda_array_interface__ by embedding the device
+    pointer as a constant.
+
+    This allows device arrays captured from globals to be used in CUDA
+    kernels and device functions.
+    """
+    interface = pyval.__cuda_array_interface__
+
+    # Hold on to the device array to prevent garbage collection.
+    context.active_code_library.referenced_objects[id(pyval)] = pyval
+
+    shape = interface["shape"]
+    strides = interface.get("strides")
+    data_ptr = interface["data"][0]
+    typestr = interface["typestr"]
+    itemsize = np.dtype(typestr).itemsize
+
+    # Calculate strides if not provided (C-contiguous)
+    if strides is None:
+        strides = strides_from_shape(shape, itemsize, order="C")
+
+    # Embed device pointer as constant
+    llvoidptr = context.get_value_type(types.voidptr)
+    data = context.get_constant(types.uintp, data_ptr).inttoptr(llvoidptr)
+
+    # Build array structure
+    ary = context.make_array(ty)(context, builder)
+    kshape = [context.get_constant(types.intp, s) for s in shape]
+    kstrides = [context.get_constant(types.intp, s) for s in strides]
+
+    context.populate_array(
+        ary,
+        data=builder.bitcast(data, ary.data.type),
+        shape=kshape,
+        strides=kstrides,
+        itemsize=context.get_constant(types.intp, itemsize),
+        parent=None,
+        meminfo=None,
+    )
+
+    return ary._getvalue()
+
+
 @lower_constant(types.Record)
 def constant_record(context, builder, ty, pyval):
     """

numba_cuda/numba/cuda/np/numpy_support.py

@@ -3,7 +3,10 @@
 
 import collections
 import ctypes
+import itertools
+import operator
 import re
+
 import numpy as np
 
 from numba.cuda import types
@@ -17,6 +20,29 @@ from numba.cuda.cgutils import is_nonelike  # noqa: F401
 
 numpy_version = tuple(map(int, np.__version__.split(".")[:2]))
 
+
+def strides_from_shape(
+    shape: tuple[int, ...], itemsize: int, *, order: str
+) -> tuple[int, ...]:
+    """Compute strides for a contiguous array with given shape and order."""
+    if len(shape) == 0:
+        # 0-D arrays have empty strides
+        return ()
+    limits = slice(1, None) if order == "C" else slice(None, -1)
+    transform = reversed if order == "C" else lambda x: x
+    strides = tuple(
+        map(
+            itemsize.__mul__,
+            itertools.accumulate(
+                transform(shape[limits]), operator.mul, initial=1
+            ),
+        )
+    )
+    if order == "F":
+        return strides
+    return strides[::-1]
+
+
 FROM_DTYPE = {
     np.dtype("bool"): types.boolean,
     np.dtype("int8"): types.int8,

numba_cuda/numba/cuda/printimpl.py

@@ -32,6 +32,26 @@ def print_item(ty, context, builder, val):
     )
 
 
+@print_item.register(types.Tuple)
+@print_item.register(types.UniTuple)
+def tuple_print_impl(ty, context, builder, val):
+    formats = []
+    values = []
+
+    for i, argtyp in enumerate(ty.types):
+        argval = builder.extract_value(val, i)
+        argfmt, argvals = print_item(argtyp, context, builder, argval)
+        formats.append(argfmt)
+        values.extend(argvals)
+
+    if len(formats) == 1:
+        base = "({},)"
+    else:
+        base = "({})"
+    rawfmt = base.format(", ".join(formats))
+    return rawfmt, values
+
+
 @print_item.register(types.Integer)
 @print_item.register(types.IntegerLiteral)
 def int_print_impl(ty, context, builder, val):

numba_cuda/numba/cuda/serialize.py

@@ -197,6 +197,16 @@ class NumbaPickler(cloudpickle.CloudPickler):
         # Overridden to disable pickling of certain types
         if type(obj) in self.disabled_types:
             _no_pickle(obj)  # noreturn
+
+        # Prevent pickling of objects implementing __cuda_array_interface__
+        # These contain device pointers that would become stale after unpickling
+        if getattr(obj, "__cuda_array_interface__", None) is not None:
+            raise pickle.PicklingError(
+                "Cannot serialize kernels or device functions referencing "
+                "global device arrays. Pass the array(s) as arguments "
+                "to the kernel instead."
+            )
+
         return super().reducer_override(obj)
 
 

numba_cuda/numba/cuda/stubs.py

@@ -200,17 +200,6 @@ class syncwarp(Stub):
     _description_ = "<warp_sync()>"
 
 
-class vote_sync_intrinsic(Stub):
-    """
-    vote_sync_intrinsic(mask, mode, predictate)
-
-    Nvvm intrinsic for performing a reduce and broadcast across a warp
-    docs.nvidia.com/cuda/nvvm-ir-spec/index.html#nvvm-intrin-warp-level-vote
-    """
-
-    _description_ = "<vote_sync()>"
-
-
 class match_any_sync(Stub):
     """
     match_any_sync(mask, value)

numba_cuda/numba/cuda/tests/benchmarks/test_kernel_launch.py

@@ -36,8 +36,13 @@ pytestmark = pytest.mark.skipif(
         ),
     ],
 )
-def test_one_arg(benchmark, array_func):
-    @cuda.jit("void(float32[:])")
+@pytest.mark.parametrize(
+    "jit",
+    [cuda.jit, cuda.jit("void(float32[::1])")],
+    ids=["dispatch", "signature"],
+)
+def test_one_arg(benchmark, array_func, jit):
+    @jit
     def one_arg(arr1):
         return
 
@@ -78,10 +83,22 @@ def test_one_arg(benchmark, array_func):
         ),
     ],
 )
-def test_many_args(benchmark, array_func):
+@pytest.mark.parametrize(
+    "jit",
+    [
+        cuda.jit,
+        cuda.jit(
+            "void({})".format(
+                ", ".join(["float32[::1]"] * len(string.ascii_lowercase))
+            )
+        ),
+    ],
+    ids=["dispatch", "signature"],
+)
+def test_many_args(benchmark, array_func, jit):
     many_arrs = array_func()
 
-    @cuda.jit("void({})".format(", ".join(["float32[:]"] * len(many_arrs))))
+    @jit
     def many_args(
         a,
         b,

numba_cuda/numba/cuda/tests/cudadrv/test_context_stack.py

@@ -2,7 +2,6 @@
 # SPDX-License-Identifier: BSD-2-Clause
 
 import numbers
-import weakref
 
 from numba import cuda
 from numba.cuda.testing import unittest, CUDATestCase, skip_on_cudasim
@@ -90,8 +89,8 @@ class Test3rdPartyContext(CUDATestCase):
         dev = driver.binding.CUdevice(0)
         binding_hctx = the_driver.cuDevicePrimaryCtxRetain(dev)
         hctx = driver.drvapi.cu_context(int(binding_hctx))
+        ctx = driver.Context(dev, hctx)
         try:
-            ctx = driver.Context(weakref.proxy(self), hctx)
             ctx.push()
             # Check that the context from numba matches the created primary
             # context.