numba-cuda 0.22.0__cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl

numba_cuda/numba/cuda/cudadrv/nvvm.py

@@ -0,0 +1,756 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: BSD-2-Clause
+
+"""
+This is a direct translation of nvvm.h
+"""
+
+import logging
+import re
+import sys
+import warnings
+from ctypes import c_void_p, c_int, POINTER, c_char_p, c_size_t, byref, c_char
+
+import threading
+
+from llvmlite import ir
+
+from .error import NvvmError, NvvmSupportError, NvvmWarning
+from .libs import get_libdevice, open_libdevice, open_cudalib
+from numba.cuda import cgutils
+
+
+logger = logging.getLogger(__name__)
+
+ADDRSPACE_GENERIC = 0
+ADDRSPACE_GLOBAL = 1
+ADDRSPACE_SHARED = 3
+ADDRSPACE_CONSTANT = 4
+ADDRSPACE_LOCAL = 5
+
+# Opaque handle for compilation unit
+nvvm_program = c_void_p
+
+# Result code
+nvvm_result = c_int
+
+RESULT_CODE_NAMES = """
+NVVM_SUCCESS
+NVVM_ERROR_OUT_OF_MEMORY
+NVVM_ERROR_PROGRAM_CREATION_FAILURE
+NVVM_ERROR_IR_VERSION_MISMATCH
+NVVM_ERROR_INVALID_INPUT
+NVVM_ERROR_INVALID_PROGRAM
+NVVM_ERROR_INVALID_IR
+NVVM_ERROR_INVALID_OPTION
+NVVM_ERROR_NO_MODULE_IN_PROGRAM
+NVVM_ERROR_COMPILATION
+""".split()
+
+for i, k in enumerate(RESULT_CODE_NAMES):
+    setattr(sys.modules[__name__], k, i)
+
+_datalayout = (
+    "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-"
+    "i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-"
+    "v64:64:64-v128:128:128-n16:32:64"
+)
+
+
+def is_available():
+    """
+    Return if libNVVM is available
+    """
+    try:
+        NVVM()
+    except NvvmSupportError:
+        return False
+    else:
+        return True
+
+
+_nvvm_lock = threading.Lock()
+
+
+class NVVM(object):
+    """Process-wide singleton."""
+
+    _PROTOTYPES = {
+        # nvvmResult nvvmVersion(int *major, int *minor)
+        "nvvmVersion": (nvvm_result, POINTER(c_int), POINTER(c_int)),
+        # nvvmResult nvvmCreateProgram(nvvmProgram *cu)
+        "nvvmCreateProgram": (nvvm_result, POINTER(nvvm_program)),
+        # nvvmResult nvvmDestroyProgram(nvvmProgram *cu)
+        "nvvmDestroyProgram": (nvvm_result, POINTER(nvvm_program)),
+        # nvvmResult nvvmAddModuleToProgram(nvvmProgram cu, const char *buffer,
+        #                                   size_t size, const char *name)
+        "nvvmAddModuleToProgram": (
+            nvvm_result,
+            nvvm_program,
+            c_char_p,
+            c_size_t,
+            c_char_p,
+        ),
+        # nvvmResult nvvmLazyAddModuleToProgram(nvvmProgram cu,
+        #                                       const char* buffer,
+        #                                       size_t size,
+        #                                       const char *name)
+        "nvvmLazyAddModuleToProgram": (
+            nvvm_result,
+            nvvm_program,
+            c_char_p,
+            c_size_t,
+            c_char_p,
+        ),
+        # nvvmResult nvvmCompileProgram(nvvmProgram cu, int numOptions,
+        #                          const char **options)
+        "nvvmCompileProgram": (
+            nvvm_result,
+            nvvm_program,
+            c_int,
+            POINTER(c_char_p),
+        ),
+        # nvvmResult nvvmGetCompiledResultSize(nvvmProgram cu,
+        #                                      size_t *bufferSizeRet)
+        "nvvmGetCompiledResultSize": (
+            nvvm_result,
+            nvvm_program,
+            POINTER(c_size_t),
+        ),
+        # nvvmResult nvvmGetCompiledResult(nvvmProgram cu, char *buffer)
+        "nvvmGetCompiledResult": (nvvm_result, nvvm_program, c_char_p),
+        # nvvmResult nvvmGetProgramLogSize(nvvmProgram cu,
+        #                                      size_t *bufferSizeRet)
+        "nvvmGetProgramLogSize": (nvvm_result, nvvm_program, POINTER(c_size_t)),
+        # nvvmResult nvvmGetProgramLog(nvvmProgram cu, char *buffer)
+        "nvvmGetProgramLog": (nvvm_result, nvvm_program, c_char_p),
+        # nvvmResult nvvmIRVersion (int* majorIR, int* minorIR, int* majorDbg,
+        #                           int* minorDbg )
+        "nvvmIRVersion": (
+            nvvm_result,
+            POINTER(c_int),
+            POINTER(c_int),
+            POINTER(c_int),
+            POINTER(c_int),
+        ),
+        # nvvmResult nvvmVerifyProgram (nvvmProgram prog, int numOptions,
+        #                               const char** options)
+        "nvvmVerifyProgram": (
+            nvvm_result,
+            nvvm_program,
+            c_int,
+            POINTER(c_char_p),
+        ),
+    }
+
+    # Singleton reference
+    __INSTANCE = None
+
+    def __new__(cls):
+        with _nvvm_lock:
+            if cls.__INSTANCE is None:
+                cls.__INSTANCE = inst = object.__new__(cls)
+                try:
+                    inst.driver = open_cudalib("nvvm")
+                except OSError as e:
+                    cls.__INSTANCE = None
+                    errmsg = "libNVVM cannot be found. Please install the cuda-toolkit conda package:\n%s"
+                    raise NvvmSupportError(errmsg % e)
+
+                # Find & populate functions
+                for name, proto in inst._PROTOTYPES.items():
+                    func = getattr(inst.driver, name)
+                    func.restype = proto[0]
+                    func.argtypes = proto[1:]
+                    setattr(inst, name, func)
+
+        return cls.__INSTANCE
+
+    def __init__(self):
+        ir_versions = self.get_ir_version()
+        self._majorIR = ir_versions[0]
+        self._minorIR = ir_versions[1]
+        self._majorDbg = ir_versions[2]
+        self._minorDbg = ir_versions[3]
+
+    @property
+    def data_layout(self):
+        return _datalayout
+
+    def get_version(self):
+        major = c_int()
+        minor = c_int()
+        err = self.nvvmVersion(byref(major), byref(minor))
+        self.check_error(err, "Failed to get version.")
+        return major.value, minor.value
+
+    def get_ir_version(self):
+        majorIR = c_int()
+        minorIR = c_int()
+        majorDbg = c_int()
+        minorDbg = c_int()
+        err = self.nvvmIRVersion(
+            byref(majorIR), byref(minorIR), byref(majorDbg), byref(minorDbg)
+        )
+        self.check_error(err, "Failed to get IR version.")
+        return majorIR.value, minorIR.value, majorDbg.value, minorDbg.value
+
+    def check_error(self, error, msg, exit=False):
+        if error:
+            exc = NvvmError(msg, RESULT_CODE_NAMES[error])
+            if exit:
+                print(exc)
+                sys.exit(1)
+            else:
+                raise exc
+
+
+class CompilationUnit(object):
+    """
+    A CompilationUnit is a set of LLVM modules that are compiled to PTX or
+    LTO-IR with NVVM.
+
+    Compilation options are accepted as a dict mapping option names to values,
+    with the following considerations:
+
+    - Underscores (`_`) in option names are converted to dashes (`-`), to match
+      NVVM's option name format.
+    - Options that take a value will be emitted in the form "-<name>=<value>".
+    - Booleans passed as option values will be converted to integers.
+    - Options which take no value (such as `-gen-lto`) should have a value of
+      `None` and will be emitted in the form "-<name>".
+
+    For documentation on NVVM compilation options, see the CUDA Toolkit
+    Documentation:
+
+    https://docs.nvidia.com/cuda/libnvvm-api/index.html#_CPPv418nvvmCompileProgram11nvvmProgramiPPKc
+    """
+
+    def __init__(self, options):
+        self.driver = NVVM()
+        self._handle = nvvm_program()
+        err = self.driver.nvvmCreateProgram(byref(self._handle))
+        self.driver.check_error(err, "Failed to create CU")
+
+        def stringify_option(k, v):
+            k = k.replace("_", "-")
+
+            if v is None:
+                return f"-{k}".encode("utf-8")
+
+            if isinstance(v, bool):
+                v = int(v)
+
+            return f"-{k}={v}".encode("utf-8")
+
+        options = [stringify_option(k, v) for k, v in options.items()]
+        option_ptrs = (c_char_p * len(options))(*[c_char_p(x) for x in options])
+
+        # We keep both the options and the pointers to them so that options are
+        # not destroyed before we've used their values
+        self.options = options
+        self.option_ptrs = option_ptrs
+        self.n_options = len(options)
+
+    def __del__(self):
+        driver = NVVM()
+        err = driver.nvvmDestroyProgram(byref(self._handle))
+        driver.check_error(err, "Failed to destroy CU", exit=True)
+
+    def add_module(self, buffer):
+        """
+        Add a module level NVVM IR to a compilation unit.
+        - The buffer should contain an NVVM module IR either in the bitcode
+          representation (LLVM3.0) or in the text representation.
+        """
+        err = self.driver.nvvmAddModuleToProgram(
+            self._handle, buffer, len(buffer), None
+        )
+        self.driver.check_error(err, "Failed to add module")
+
+    def lazy_add_module(self, buffer):
+        """
+        Lazily add an NVVM IR module to a compilation unit.
+        The buffer should contain NVVM module IR either in the bitcode
+        representation or in the text representation.
+        """
+        err = self.driver.nvvmLazyAddModuleToProgram(
+            self._handle, buffer, len(buffer), None
+        )
+        self.driver.check_error(err, "Failed to add module")
+
+    def verify(self):
+        """
+        Run the NVVM verifier on all code added to the compilation unit.
+        """
+        err = self.driver.nvvmVerifyProgram(
+            self._handle, self.n_options, self.option_ptrs
+        )
+        self._try_error(err, "Failed to verify\n")
+
+    def compile(self):
+        """
+        Compile all modules added to the compilation unit and return the
+        resulting PTX or LTO-IR (depending on the options).
+        """
+        err = self.driver.nvvmCompileProgram(
+            self._handle, self.n_options, self.option_ptrs
+        )
+        self._try_error(err, "Failed to compile\n")
+
+        # Get result
+        result_size = c_size_t()
+        err = self.driver.nvvmGetCompiledResultSize(
+            self._handle, byref(result_size)
+        )
+
+        self._try_error(err, "Failed to get size of compiled result.")
+
+        output_buffer = (c_char * result_size.value)()
+        err = self.driver.nvvmGetCompiledResult(self._handle, output_buffer)
+        self._try_error(err, "Failed to get compiled result.")
+
+        # Get log
+        self.log = self.get_log()
+        if self.log:
+            warnings.warn(self.log, category=NvvmWarning)
+
+        return output_buffer[:]
+
+    def _try_error(self, err, msg):
+        self.driver.check_error(err, "%s\n%s" % (msg, self.get_log()))
+
+    def get_log(self):
+        reslen = c_size_t()
+        err = self.driver.nvvmGetProgramLogSize(self._handle, byref(reslen))
+        self.driver.check_error(err, "Failed to get compilation log size.")
+
+        if reslen.value > 1:
+            logbuf = (c_char * reslen.value)()
+            err = self.driver.nvvmGetProgramLog(self._handle, logbuf)
+            self.driver.check_error(err, "Failed to get compilation log.")
+
+            return logbuf.value.decode("utf8")  # populate log attribute
+
+        return ""
+
+
+MISSING_LIBDEVICE_FILE_MSG = """Missing libdevice file.
+``cuda-nvcc`` and ``cuda-nvrtc`` are required:
+
+    $ conda install -c conda-forge cuda-nvcc cuda-nvrtc "cuda-version>=12.0"
+"""
+
+
+class LibDevice(object):
+    _cache_ = None
+
+    def __init__(self):
+        if self._cache_ is None:
+            if get_libdevice() is None:
+                raise RuntimeError(MISSING_LIBDEVICE_FILE_MSG)
+            self._cache_ = open_libdevice()
+
+        self.bc = self._cache_
+
+    def get(self):
+        return self.bc
+
+
+cas_nvvm = """
+    %cas_success = cmpxchg volatile {Ti}* %iptr, {Ti} %old, {Ti} %new monotonic monotonic
+    %cas = extractvalue {{ {Ti}, i1 }} %cas_success, 0
+"""  # noqa: E501
+
+
+# Translation of code from CUDA Programming Guide v6.5, section B.12
+ir_numba_atomic_binary_template = """
+define internal {T} @___numba_atomic_{T}_{FUNC}({T}* %ptr, {T} %val) alwaysinline {{
+entry:
+    %iptr = bitcast {T}* %ptr to {Ti}*
+    %old2 = load volatile {Ti}, {Ti}* %iptr
+    br label %attempt
+
+attempt:
+    %old = phi {Ti} [ %old2, %entry ], [ %cas, %attempt ]
+    %dold = bitcast {Ti} %old to {T}
+    %dnew = {OP} {T} %dold, %val
+    %new = bitcast {T} %dnew to {Ti}
+    {CAS}
+    %repeat = icmp ne {Ti} %cas, %old
+    br i1 %repeat, label %attempt, label %done
+
+done:
+    %result = bitcast {Ti} %old to {T}
+    ret {T} %result
+}}
+"""  # noqa: E501
+
+ir_numba_atomic_inc_template = """
+define internal {T} @___numba_atomic_{Tu}_inc({T}* %iptr, {T} %val) alwaysinline {{
+entry:
+    %old2 = load volatile {T}, {T}* %iptr
+    br label %attempt
+
+attempt:
+    %old = phi {T} [ %old2, %entry ], [ %cas, %attempt ]
+    %bndchk = icmp ult {T} %old, %val
+    %inc = add {T} %old, 1
+    %new = select i1 %bndchk, {T} %inc, {T} 0
+    {CAS}
+    %repeat = icmp ne {T} %cas, %old
+    br i1 %repeat, label %attempt, label %done
+
+done:
+    ret {T} %old
+}}
+"""  # noqa: E501
+
+ir_numba_atomic_dec_template = """
+define internal {T} @___numba_atomic_{Tu}_dec({T}* %iptr, {T} %val) alwaysinline {{
+entry:
+    %old2 = load volatile {T}, {T}* %iptr
+    br label %attempt
+
+attempt:
+    %old = phi {T} [ %old2, %entry ], [ %cas, %attempt ]
+    %dec = add {T} %old, -1
+    %bndchk = icmp ult {T} %dec, %val
+    %new = select i1 %bndchk, {T} %dec, {T} %val
+    {CAS}
+    %repeat = icmp ne {T} %cas, %old
+    br i1 %repeat, label %attempt, label %done
+
+done:
+    ret {T} %old
+}}
+"""  # noqa: E501
+
+ir_numba_atomic_minmax_template = """
+define internal {T} @___numba_atomic_{T}_{NAN}{FUNC}({T}* %ptr, {T} %val) alwaysinline {{
+entry:
+    %ptrval = load volatile {T}, {T}* %ptr
+    ; Return early when:
+    ; - For nanmin / nanmax when val is a NaN
+    ; - For min / max when val or ptr is a NaN
+    %early_return = fcmp uno {T} %val, %{PTR_OR_VAL}val
+    br i1 %early_return, label %done, label %lt_check
+
+lt_check:
+    %dold = phi {T} [ %ptrval, %entry ], [ %dcas, %attempt ]
+    ; Continue attempts if dold less or greater than val (depending on whether min or max)
+    ; or if dold is NaN (for nanmin / nanmax)
+    %cmp = fcmp {OP} {T} %dold, %val
+    br i1 %cmp, label %attempt, label %done
+
+attempt:
+    ; Attempt to swap in the value
+    %old = bitcast {T} %dold to {Ti}
+    %iptr = bitcast {T}* %ptr to {Ti}*
+    %new = bitcast {T} %val to {Ti}
+    {CAS}
+    %dcas = bitcast {Ti} %cas to {T}
+    br label %lt_check
+
+done:
+    ret {T} %ptrval
+}}
+"""  # noqa: E501
+
+
+def ir_cas(Ti):
+    return cas_nvvm.format(Ti=Ti)
+
+
+def ir_numba_atomic_binary(T, Ti, OP, FUNC):
+    params = dict(T=T, Ti=Ti, OP=OP, FUNC=FUNC, CAS=ir_cas(Ti))
+    return ir_numba_atomic_binary_template.format(**params)
+
+
+def ir_numba_atomic_minmax(T, Ti, NAN, OP, PTR_OR_VAL, FUNC):
+    params = dict(
+        T=T,
+        Ti=Ti,
+        NAN=NAN,
+        OP=OP,
+        PTR_OR_VAL=PTR_OR_VAL,
+        FUNC=FUNC,
+        CAS=ir_cas(Ti),
+    )
+
+    return ir_numba_atomic_minmax_template.format(**params)
+
+
+def ir_numba_atomic_inc(T, Tu):
+    return ir_numba_atomic_inc_template.format(T=T, Tu=Tu, CAS=ir_cas(T))
+
+
+def ir_numba_atomic_dec(T, Tu):
+    return ir_numba_atomic_dec_template.format(T=T, Tu=Tu, CAS=ir_cas(T))
+
+
+def llvm_replace(llvmir):
+    replacements = [
+        (
+            'declare double @"___numba_atomic_double_add"(double* %".1", double %".2")',  # noqa: E501
+            ir_numba_atomic_binary(T="double", Ti="i64", OP="fadd", FUNC="add"),
+        ),
+        (
+            'declare float @"___numba_atomic_float_sub"(float* %".1", float %".2")',  # noqa: E501
+            ir_numba_atomic_binary(T="float", Ti="i32", OP="fsub", FUNC="sub"),
+        ),
+        (
+            'declare double @"___numba_atomic_double_sub"(double* %".1", double %".2")',  # noqa: E501
+            ir_numba_atomic_binary(T="double", Ti="i64", OP="fsub", FUNC="sub"),
+        ),
+        (
+            'declare i64 @"___numba_atomic_u64_inc"(i64* %".1", i64 %".2")',
+            ir_numba_atomic_inc(T="i64", Tu="u64"),
+        ),
+        (
+            'declare i64 @"___numba_atomic_u64_dec"(i64* %".1", i64 %".2")',
+            ir_numba_atomic_dec(T="i64", Tu="u64"),
+        ),
+        (
+            'declare float @"___numba_atomic_float_max"(float* %".1", float %".2")',  # noqa: E501
+            ir_numba_atomic_minmax(
+                T="float",
+                Ti="i32",
+                NAN="",
+                OP="nnan olt",
+                PTR_OR_VAL="ptr",
+                FUNC="max",
+            ),
+        ),
+        (
+            'declare double @"___numba_atomic_double_max"(double* %".1", double %".2")',  # noqa: E501
+            ir_numba_atomic_minmax(
+                T="double",
+                Ti="i64",
+                NAN="",
+                OP="nnan olt",
+                PTR_OR_VAL="ptr",
+                FUNC="max",
+            ),
+        ),
+        (
+            'declare float @"___numba_atomic_float_min"(float* %".1", float %".2")',  # noqa: E501
+            ir_numba_atomic_minmax(
+                T="float",
+                Ti="i32",
+                NAN="",
+                OP="nnan ogt",
+                PTR_OR_VAL="ptr",
+                FUNC="min",
+            ),
+        ),
+        (
+            'declare double @"___numba_atomic_double_min"(double* %".1", double %".2")',  # noqa: E501
+            ir_numba_atomic_minmax(
+                T="double",
+                Ti="i64",
+                NAN="",
+                OP="nnan ogt",
+                PTR_OR_VAL="ptr",
+                FUNC="min",
+            ),
+        ),
+        (
+            'declare float @"___numba_atomic_float_nanmax"(float* %".1", float %".2")',  # noqa: E501
+            ir_numba_atomic_minmax(
+                T="float",
+                Ti="i32",
+                NAN="nan",
+                OP="ult",
+                PTR_OR_VAL="",
+                FUNC="max",
+            ),
+        ),
+        (
+            'declare double @"___numba_atomic_double_nanmax"(double* %".1", double %".2")',  # noqa: E501
+            ir_numba_atomic_minmax(
+                T="double",
+                Ti="i64",
+                NAN="nan",
+                OP="ult",
+                PTR_OR_VAL="",
+                FUNC="max",
+            ),
+        ),
+        (
+            'declare float @"___numba_atomic_float_nanmin"(float* %".1", float %".2")',  # noqa: E501
+            ir_numba_atomic_minmax(
+                T="float",
+                Ti="i32",
+                NAN="nan",
+                OP="ugt",
+                PTR_OR_VAL="",
+                FUNC="min",
+            ),
+        ),
+        (
+            'declare double @"___numba_atomic_double_nanmin"(double* %".1", double %".2")',  # noqa: E501
+            ir_numba_atomic_minmax(
+                T="double",
+                Ti="i64",
+                NAN="nan",
+                OP="ugt",
+                PTR_OR_VAL="",
+                FUNC="min",
+            ),
+        ),
+        ("immarg", ""),
+    ]
+
+    for decl, fn in replacements:
+        llvmir = llvmir.replace(decl, fn)
+
+    llvmir = llvm150_to_70_ir(llvmir)
+
+    return llvmir
+
+
+def compile_ir(llvmir, **options):
+    if isinstance(llvmir, str):
+        llvmir = [llvmir]
+
+    if options.pop("fastmath", False):
+        options.update(
+            {
+                "ftz": True,
+                "fma": True,
+                "prec_div": False,
+                "prec_sqrt": False,
+            }
+        )
+
+    cu = CompilationUnit(options)
+
+    for mod in llvmir:
+        mod = llvm_replace(mod)
+        cu.add_module(mod.encode("utf8"))
+    cu.verify()
+
+    # We add libdevice following verification so that it is not subject to the
+    # verifier's requirements
+    libdevice = LibDevice()
+    cu.lazy_add_module(libdevice.get())
+
+    return cu.compile()
+
+
+re_attributes_def = re.compile(r"^attributes #\d+ = \{ ([\w\s]+)\ }")
+
+
+def llvm150_to_70_ir(ir):
+    """
+    Convert LLVM 15.0 IR for LLVM 7.0.
+    """
+    buf = []
+    for line in ir.splitlines():
+        if line.startswith("attributes #"):
+            # Remove function attributes unsupported by LLVM 7.0
+            m = re_attributes_def.match(line)
+            attrs = m.group(1).split()
+            attrs = " ".join(a for a in attrs if a != "willreturn")
+            line = line.replace(m.group(1), attrs)
+
+        buf.append(line)
+
+    return "\n".join(buf)
+
+
+def set_cuda_kernel(function):
+    """
+    Mark a function as a CUDA kernel. Kernels have the following requirements:
+
+    - Metadata that marks them as a kernel.
+    - Addition to the @llvm.used list, so that they will not be discarded.
+    - The noinline attribute is not permitted, because this causes NVVM to emit
+      a warning, which counts as failing IR verification.
+
+    Presently it is assumed that there is one kernel per module, which holds
+    for Numba-jitted functions. If this changes in future or this function is
+    to be used externally, this function may need modification to add to the
+    @llvm.used list rather than creating it.
+    """
+    module = function.module
+
+    # Add kernel metadata
+    mdstr = ir.MetaDataString(module, "kernel")
+    mdvalue = ir.Constant(ir.IntType(32), 1)
+    md = module.add_metadata((function, mdstr, mdvalue))
+
+    nmd = cgutils.get_or_insert_named_metadata(module, "nvvm.annotations")
+    nmd.add(md)
+
+    # Create the used list
+    ptrty = ir.IntType(8).as_pointer()
+    usedty = ir.ArrayType(ptrty, 1)
+
+    fnptr = function.bitcast(ptrty)
+
+    llvm_used = ir.GlobalVariable(module, usedty, "llvm.used")
+    llvm_used.linkage = "appending"
+    llvm_used.section = "llvm.metadata"
+    llvm_used.initializer = ir.Constant(usedty, [fnptr])
+
+    # Remove 'noinline' if it is present.
+    function.attributes.discard("noinline")
+
+
+def set_launch_bounds(kernel, launch_bounds):
+    # Based on: CUDA C / C++ Programming Guide 12.9, Section 8.38:
+    # https://docs.nvidia.com/cuda/archive/12.9.0/cuda-c-programming-guide/index.html#launch-bounds
+    # PTX ISA Specification Version 8.7, Section 11.4:
+    # https://docs.nvidia.com/cuda/archive/12.8.1/parallel-thread-execution/index.html#performance-tuning-directives
+    # NVVM IR Specification 12.9, Section 13:
+    # https://docs.nvidia.com/cuda/archive/12.9.0/nvvm-ir-spec/index.html#global-property-annotation
+
+    if launch_bounds is None:
+        return
+
+    if isinstance(launch_bounds, int):
+        launch_bounds = (launch_bounds,)
+
+    if (n := len(launch_bounds)) > 3:
+        raise ValueError(
+            f"Got {n} launch bounds: {launch_bounds}. A maximum of three are supported: "
+            "(max_threads_per_block, min_blocks_per_sm, max_blocks_per_cluster)"
+        )
+
+    module = kernel.module
+    nvvm_annotations = cgutils.get_or_insert_named_metadata(
+        module, "nvvm.annotations"
+    )
+
+    # Note that only maxntidx is used even though NVVM IR and PTX allow
+    # maxntidy and maxntidz. This is because the thread block size limit
+    # pertains only to the total number of threads, and therefore bounds on
+    # individual dimensions may be exceeded anyway. To prevent an unsurprising
+    # interface, it is cleaner to only allow setting total size via maxntidx
+    # and assuming y and z to be 1 (as is the case in CUDA C/C++).
+
+    properties = (
+        # Max threads per block
+        "maxntidx",
+        # Min blocks per multiprocessor
+        "minctasm",
+        # Max blocks per cluster
+        "cluster_max_blocks",
+    )
+
+    for prop, bound in zip(properties, launch_bounds):
+        mdstr = ir.MetaDataString(module, prop)
+        mdvalue = ir.Constant(ir.IntType(32), bound)
+        md = module.add_metadata((kernel, mdstr, mdvalue))
+        nvvm_annotations.add(md)
+
+
+def add_ir_version(mod):
+    """Add NVVM IR version to module"""
+    # We specify the IR version to match the current NVVM's IR version
+    i32 = ir.IntType(32)
+    ir_versions = [i32(v) for v in NVVM().get_ir_version()]
+    md_ver = mod.add_metadata(ir_versions)
+    mod.add_named_metadata("nvvmir.version", md_ver)