Use pynvjitlink for CUDA 12+ MVC (#13650)

Fixes #12822

This PR provides minor version compatibility in the CUDA 12.x range through `nvjitlink` via the preliminary [nvjiitlink python binding](https://github.com/gmarkall/nvjitlink). Thus far this PR merely leverages a local installation of the library and should not be merged until `nvjitlink` is hosted on `conda-forge` and cuDF's dependencies are adjusted accordingly, likely as part of this PR.

Authors:
  - https://github.com/brandon-b-miller
  - Ashwin Srinath (https://github.com/shwina)

Approvers:
  - Bradley Dice (https://github.com/bdice)
  - Ashwin Srinath (https://github.com/shwina)

URL: #13650

python/cudf/cudf/tests/test_mvc.py

@@ -0,0 +1,99 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.
+import subprocess
+import sys
+
+import pytest
+
+IS_CUDA_11 = False
+IS_CUDA_12 = False
+try:
+    from ptxcompiler.patch import safe_get_versions
+except ModuleNotFoundError:
+    from cudf.utils._ptxcompiler import safe_get_versions
+
+# do not test cuda 12 if pynvjitlink isn't present
+HAVE_PYNVJITLINK = False
+try:
+    import pynvjitlink  # noqa: F401
+
+    HAVE_PYNVJITLINK = True
+except ModuleNotFoundError:
+    pass
+
+
+versions = safe_get_versions()
+driver_version, runtime_version = versions
+
+if (11, 0) <= driver_version < (12, 0):
+    IS_CUDA_11 = True
+if (12, 0) <= driver_version < (13, 0):
+    IS_CUDA_12 = True
+
+
+TEST_BODY = """
+@numba.cuda.jit
+def test_kernel(x):
+    id = numba.cuda.grid(1)
+    if id < len(x):
+        x[id] += 1
+
+s = cudf.Series([1, 2, 3])
+with _CUDFNumbaConfig():
+    test_kernel.forall(len(s))(s)
+"""
+
+CUDA_11_TEST = (
+    """
+import numba.cuda
+import cudf
+from cudf.utils._numba import _CUDFNumbaConfig, patch_numba_linker_cuda_11
+
+
+patch_numba_linker_cuda_11()
+"""
+    + TEST_BODY
+)
+
+
+CUDA_12_TEST = (
+    """
+import numba.cuda
+import cudf
+from cudf.utils._numba import _CUDFNumbaConfig
+from pynvjitlink.patch import (
+    patch_numba_linker as patch_numba_linker_pynvjitlink,
+)
+
+patch_numba_linker_pynvjitlink()
+"""
+    + TEST_BODY
+)
+
+
+@pytest.mark.parametrize(
+    "test",
+    [
+        pytest.param(
+            CUDA_11_TEST,
+            marks=pytest.mark.skipif(
+                not IS_CUDA_11,
+                reason="Minor Version Compatibility test for CUDA 11",
+            ),
+        ),
+        pytest.param(
+            CUDA_12_TEST,
+            marks=pytest.mark.skipif(
+                not IS_CUDA_12 or not HAVE_PYNVJITLINK,
+                reason="Minor Version Compatibility test for CUDA 12",
+            ),
+        ),
+    ],
+)
+def test_numba_mvc(test):
+    cp = subprocess.run(
+        [sys.executable, "-c", test],
+        capture_output=True,
+        cwd="/",
+    )
+
+    assert cp.returncode == 0

python/cudf/cudf/utils/_numba.py

@@ -7,6 +7,19 @@
 
 from numba import config as numba_config
 
+try:
+    from pynvjitlink.patch import (
+        patch_numba_linker as patch_numba_linker_pynvjitlink,
+    )
+except ImportError:
+
+    def patch_numba_linker_pynvjitlink():
+        warnings.warn(
+            "CUDA Toolkit is newer than CUDA driver. "
+            "Numba features will not work in this configuration. "
+        )
+
+
 CC_60_PTX_FILE = os.path.join(
     os.path.dirname(__file__), "../core/udf/shim_60.ptx"
 )
@@ -65,7 +78,7 @@ def _get_ptx_file(path, prefix):
         return regular_result[1]
 
 
-def _patch_numba_mvc():
+def patch_numba_linker_cuda_11():
     # Enable the config option for minor version compatibility
     numba_config.CUDA_ENABLE_MINOR_VERSION_COMPATIBILITY = 1
 
@@ -106,29 +119,19 @@ def _setup_numba():
     versions = safe_get_versions()
     if versions != NO_DRIVER:
         driver_version, runtime_version = versions
-        if driver_version >= (12, 0) and runtime_version > driver_version:
-            warnings.warn(
-                f"Using CUDA toolkit version {runtime_version} with CUDA "
-                f"driver version {driver_version} requires minor version "
-                "compatibility, which is not yet supported for CUDA "
-                "driver versions 12.0 and above. It is likely that many "
-                "cuDF operations will not work in this state. Please "
-                f"install CUDA toolkit version {driver_version} to "
-                "continue using cuDF."
-            )
-        else:
-            # Support MVC for all CUDA versions in the 11.x range
-            ptx_toolkit_version = _get_cuda_version_from_ptx_file(
-                CC_60_PTX_FILE
-            )
-            # Numba thinks cubinlinker is only needed if the driver is older
-            # than the CUDA runtime, but when PTX files are present, it might
-            # also need to patch because those PTX files may be compiled by
-            # a CUDA version that is newer than the driver as well
-            if (driver_version < ptx_toolkit_version) or (
-                driver_version < runtime_version
-            ):
-                _patch_numba_mvc()
+        ptx_toolkit_version = _get_cuda_version_from_ptx_file(CC_60_PTX_FILE)
+
+        # MVC is required whenever any PTX is newer than the driver
+        # This could be the shipped PTX file or the PTX emitted by
+        # the version of NVVM on the user system, the latter aligning
+        # with the runtime version
+        if (driver_version < ptx_toolkit_version) or (
+            driver_version < runtime_version
+        ):
+            if driver_version < (12, 0):
+                patch_numba_linker_cuda_11()
+            else:
+                patch_numba_linker_pynvjitlink()
 
 
 def _get_cuda_version_from_ptx_file(path):
@@ -171,6 +174,8 @@ def _get_cuda_version_from_ptx_file(path):
         "7.8": (11, 8),
         "8.0": (12, 0),
         "8.1": (12, 1),
+        "8.2": (12, 2),
+        "8.3": (12, 3),
     }
 
     cuda_ver = ver_map.get(version)