warp-lang 1.9.0__py3-none-win_amd64.whl → 1.10.0rc2__py3-none-win_amd64.whl

warp/_src/build_dll.py

@@ -0,0 +1,640 @@
+# SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import concurrent.futures
+import os
+import platform
+import subprocess
+import sys
+import time
+
+from warp._src.utils import ScopedTimer
+
+verbose_cmd = True  # print command lines before executing them
+
+MIN_CTK_VERSION = (12, 0)
+
+
+def machine_architecture() -> str:
+    """Return a canonical machine architecture string.
+    - "x86_64" for x86-64, aka. AMD64, aka. x64
+    - "aarch64" for AArch64, aka. ARM64
+    """
+    machine = platform.machine()
+    if machine == "x86_64" or machine == "AMD64":
+        return "x86_64"
+    if machine == "aarch64" or machine == "arm64":
+        return "aarch64"
+    raise RuntimeError(f"Unrecognized machine architecture {machine}")
+
+
+def run_cmd(cmd):
+    if verbose_cmd:
+        print(cmd)
+
+    try:
+        return subprocess.check_output(cmd, stderr=subprocess.STDOUT, shell=True)
+    except subprocess.CalledProcessError as e:
+        print("Command failed with exit code:", e.returncode)
+        print("Command output was:")
+        print(e.output.decode())
+        raise e
+
+
+# cut-down version of vcvars64.bat that allows using
+# custom toolchain locations, returns the compiler program path
+def set_msvc_env(msvc_path, sdk_path):
+    if "INCLUDE" not in os.environ:
+        os.environ["INCLUDE"] = ""
+
+    if "LIB" not in os.environ:
+        os.environ["LIB"] = ""
+
+    msvc_path = os.path.abspath(msvc_path)
+    sdk_path = os.path.abspath(sdk_path)
+
+    os.environ["INCLUDE"] += os.pathsep + os.path.join(msvc_path, "include")
+    os.environ["INCLUDE"] += os.pathsep + os.path.join(sdk_path, "include/winrt")
+    os.environ["INCLUDE"] += os.pathsep + os.path.join(sdk_path, "include/um")
+    os.environ["INCLUDE"] += os.pathsep + os.path.join(sdk_path, "include/ucrt")
+    os.environ["INCLUDE"] += os.pathsep + os.path.join(sdk_path, "include/shared")
+
+    os.environ["LIB"] += os.pathsep + os.path.join(msvc_path, "lib/x64")
+    os.environ["LIB"] += os.pathsep + os.path.join(sdk_path, "lib/ucrt/x64")
+    os.environ["LIB"] += os.pathsep + os.path.join(sdk_path, "lib/um/x64")
+
+    os.environ["PATH"] += os.pathsep + os.path.join(msvc_path, "bin/HostX64/x64")
+    os.environ["PATH"] += os.pathsep + os.path.join(sdk_path, "bin/x64")
+
+    return os.path.join(msvc_path, "bin", "HostX64", "x64", "cl.exe")
+
+
+def find_host_compiler():
+    if os.name == "nt":
+        # try and find an installed host compiler (msvc)
+        # runs vcvars and copies back the build environment
+
+        vswhere_path = r"%ProgramFiles(x86)%/Microsoft Visual Studio/Installer/vswhere.exe"
+        vswhere_path = os.path.expandvars(vswhere_path)
+        if not os.path.exists(vswhere_path):
+            return ""
+
+        vs_path = run_cmd(f'"{vswhere_path}" -latest -property installationPath').decode().rstrip()
+        vsvars_path = os.path.join(vs_path, "VC\\Auxiliary\\Build\\vcvars64.bat")
+
+        output = run_cmd(f'"{vsvars_path}" && set').decode()
+
+        for line in output.splitlines():
+            pair = line.split("=", 1)
+            if len(pair) >= 2:
+                os.environ[pair[0]] = pair[1]
+
+        cl_path = run_cmd("where cl.exe").decode("utf-8").rstrip()
+        cl_version = os.environ["VCToolsVersion"].split(".")
+
+        # ensure at least VS2019 version, see list of MSVC versions here https://en.wikipedia.org/wiki/Microsoft_Visual_C%2B%2B
+        cl_required_major = 14
+        cl_required_minor = 29
+
+        if int(cl_version[0]) < cl_required_major or (
+            (int(cl_version[0]) == cl_required_major) and (int(cl_version[1]) < cl_required_minor)
+        ):
+            print(
+                f"Warp: MSVC found but compiler version too old, found {cl_version[0]}.{cl_version[1]}, but must be {cl_required_major}.{cl_required_minor} or higher, kernel host compilation will be disabled."
+            )
+            return ""
+
+        return cl_path
+
+    else:
+        # try and find g++
+        return run_cmd("which g++").decode()
+
+
+def get_cuda_toolkit_version(cuda_home) -> tuple[int, int]:
+    try:
+        # the toolkit version can be obtained by running "nvcc --version"
+        nvcc_path = os.path.join(cuda_home, "bin", "nvcc")
+        nvcc_version_output = subprocess.check_output([nvcc_path, "--version"]).decode("utf-8")
+        # search for release substring (e.g., "release 11.5")
+        import re
+
+        m = re.search(r"release (\d+)\.(\d+)", nvcc_version_output)
+        if m is not None:
+            major, minor = map(int, m.groups())
+            return (major, minor)
+        else:
+            raise Exception("Failed to parse NVCC output")
+
+    except Exception as e:
+        print(f"Warning: Failed to determine CUDA Toolkit version: {e}")
+        return MIN_CTK_VERSION
+
+
+def quote(path):
+    return '"' + path + '"'
+
+
+def add_llvm_bin_to_path(args):
+    """Add the LLVM bin directory to the PATH environment variable if it's set.
+
+    Args:
+        args: The argument namespace containing llvm_path.
+
+    Returns:
+        ``True`` if the PATH was updated, ``False`` otherwise.
+    """
+    if not hasattr(args, "llvm_path") or not args.llvm_path:
+        return False
+
+    # Construct the bin directory path
+    llvm_bin_path = os.path.join(args.llvm_path, "bin")
+
+    # Check if the directory exists
+    if not os.path.isdir(llvm_bin_path):
+        print(f"Warning: LLVM bin directory not found at {llvm_bin_path}")
+        return False
+
+    # Add to PATH environment variable
+    os.environ["PATH"] = llvm_bin_path + os.pathsep + os.environ.get("PATH", "")
+
+    print(f"Added {llvm_bin_path} to PATH")
+    return True
+
+
+def _get_architectures_cu12(
+    ctk_version: tuple[int, int], arch: str, target_platform: str, quick_build: bool = False
+) -> tuple[list[str], list[str]]:
+    """Get architecture flags for CUDA 12.x."""
+    gencode_opts = []
+    clang_arch_flags = []
+
+    if quick_build:
+        gencode_opts = ["-gencode=arch=compute_52,code=compute_52", "-gencode=arch=compute_75,code=compute_75"]
+        clang_arch_flags = ["--cuda-gpu-arch=sm_52", "--cuda-gpu-arch=sm_75"]
+    else:
+        if arch == "aarch64" and target_platform == "linux" and ctk_version == (12, 9):
+            # Skip certain architectures for aarch64 with CUDA 12.9 due to CCCL bug
+            print(
+                "[INFO] Skipping sm_52, sm_60, sm_61, and sm_70 targets for ARM due to a CUDA Toolkit bug. "
+                "See https://nvidia.github.io/warp/installation.html#cuda-12-9-limitation-on-linux-arm-platforms "
+                "for details."
+            )
+        else:
+            gencode_opts.extend(
+                [
+                    "-gencode=arch=compute_52,code=sm_52",  # Maxwell
+                    "-gencode=arch=compute_60,code=sm_60",  # Pascal
+                    "-gencode=arch=compute_61,code=sm_61",
+                    "-gencode=arch=compute_70,code=sm_70",  # Volta
+                ]
+            )
+            clang_arch_flags.extend(
+                [
+                    "--cuda-gpu-arch=sm_52",
+                    "--cuda-gpu-arch=sm_60",
+                    "--cuda-gpu-arch=sm_61",
+                    "--cuda-gpu-arch=sm_70",
+                ]
+            )
+
+        # Desktop architectures
+        gencode_opts.extend(
+            [
+                "-gencode=arch=compute_75,code=sm_75",  # Turing
+                "-gencode=arch=compute_75,code=compute_75",  # Turing (PTX)
+                "-gencode=arch=compute_80,code=sm_80",  # Ampere
+                "-gencode=arch=compute_86,code=sm_86",
+                "-gencode=arch=compute_89,code=sm_89",  # Ada
+                "-gencode=arch=compute_90,code=sm_90",  # Hopper
+            ]
+        )
+        clang_arch_flags.extend(
+            [
+                "--cuda-gpu-arch=sm_75",  # Turing
+                "--cuda-gpu-arch=sm_80",  # Ampere
+                "--cuda-gpu-arch=sm_86",
+                "--cuda-gpu-arch=sm_89",  # Ada
+                "--cuda-gpu-arch=sm_90",  # Hopper
+            ]
+        )
+
+        if ctk_version >= (12, 8):
+            gencode_opts.extend(["-gencode=arch=compute_100,code=sm_100", "-gencode=arch=compute_120,code=sm_120"])
+            clang_arch_flags.extend(["--cuda-gpu-arch=sm_100", "--cuda-gpu-arch=sm_120"])
+
+        # Mobile architectures for aarch64 Linux
+        if arch == "aarch64" and target_platform == "linux":
+            gencode_opts.extend(
+                [
+                    "-gencode=arch=compute_87,code=sm_87",  # Orin
+                    "-gencode=arch=compute_53,code=sm_53",  # X1
+                    "-gencode=arch=compute_62,code=sm_62",  # X2
+                    "-gencode=arch=compute_72,code=sm_72",  # Xavier
+                ]
+            )
+            clang_arch_flags.extend(
+                [
+                    "--cuda-gpu-arch=sm_87",
+                    "--cuda-gpu-arch=sm_53",
+                    "--cuda-gpu-arch=sm_62",
+                    "--cuda-gpu-arch=sm_72",
+                ]
+            )
+
+            # Thor support in CUDA 12.8+
+            if ctk_version >= (12, 8):
+                gencode_opts.append("-gencode=arch=compute_101,code=sm_101")  # Thor (CUDA 12 numbering)
+                clang_arch_flags.append("--cuda-gpu-arch=sm_101")
+
+            if ctk_version >= (12, 9):
+                gencode_opts.append("-gencode=arch=compute_121,code=sm_121")
+                clang_arch_flags.append("--cuda-gpu-arch=sm_121")
+
+        # PTX for future hardware (use highest available compute capability)
+        if ctk_version >= (12, 9):
+            gencode_opts.extend(["-gencode=arch=compute_121,code=compute_121"])
+        elif ctk_version >= (12, 8):
+            gencode_opts.extend(["-gencode=arch=compute_120,code=compute_120"])
+        else:
+            gencode_opts.append("-gencode=arch=compute_90,code=compute_90")
+
+    return gencode_opts, clang_arch_flags
+
+
+def _get_architectures_cu13(
+    ctk_version: tuple[int, int], arch: str, target_platform: str, quick_build: bool = False
+) -> tuple[list[str], list[str]]:
+    """Get architecture flags for CUDA 13.x."""
+    gencode_opts = []
+    clang_arch_flags = []
+
+    if quick_build:
+        gencode_opts = ["-gencode=arch=compute_75,code=compute_75"]
+        clang_arch_flags = ["--cuda-gpu-arch=sm_75"]
+    else:
+        # Desktop architectures
+        gencode_opts.extend(
+            [
+                "-gencode=arch=compute_75,code=sm_75",  # Turing
+                "-gencode=arch=compute_75,code=compute_75",  # Turing (PTX)
+                "-gencode=arch=compute_80,code=sm_80",  # Ampere
+                "-gencode=arch=compute_86,code=sm_86",
+                "-gencode=arch=compute_89,code=sm_89",  # Ada
+                "-gencode=arch=compute_90,code=sm_90",  # Hopper
+                "-gencode=arch=compute_100,code=sm_100",  # Blackwell
+                "-gencode=arch=compute_120,code=sm_120",  # Blackwell
+            ]
+        )
+        clang_arch_flags.extend(
+            [
+                "--cuda-gpu-arch=sm_75",  # Turing
+                "--cuda-gpu-arch=sm_80",  # Ampere
+                "--cuda-gpu-arch=sm_86",
+                "--cuda-gpu-arch=sm_89",  # Ada
+                "--cuda-gpu-arch=sm_90",  # Hopper
+                "--cuda-gpu-arch=sm_100",  # Blackwell
+                "--cuda-gpu-arch=sm_120",  # Blackwell
+            ]
+        )
+
+        # Mobile architectures for aarch64 Linux
+        if arch == "aarch64" and target_platform == "linux":
+            gencode_opts.extend(
+                [
+                    "-gencode=arch=compute_87,code=sm_87",  # Orin
+                    "-gencode=arch=compute_110,code=sm_110",  # Thor
+                    "-gencode=arch=compute_121,code=sm_121",  # Spark
+                ]
+            )
+            clang_arch_flags.extend(
+                [
+                    "--cuda-gpu-arch=sm_87",
+                    "--cuda-gpu-arch=sm_110",
+                    "--cuda-gpu-arch=sm_121",
+                ]
+            )
+
+        # PTX for future hardware (use highest available compute capability)
+        gencode_opts.extend(["-gencode=arch=compute_121,code=compute_121"])
+
+    return gencode_opts, clang_arch_flags
+
+
+def build_dll_for_arch(args, dll_path, cpp_paths, cu_paths, arch, libs: list[str] | None = None, mode=None):
+    mode = args.mode if (mode is None) else mode
+    cuda_home = args.cuda_path
+    cuda_cmd = None
+
+    # Add LLVM bin directory to PATH
+    add_llvm_bin_to_path(args)
+
+    if args.quick or cu_paths is None:
+        cuda_compat_enabled = "WP_ENABLE_CUDA_COMPATIBILITY=0"
+    else:
+        cuda_compat_enabled = "WP_ENABLE_CUDA_COMPATIBILITY=1"
+
+    if libs is None:
+        libs = []
+
+    import pathlib
+
+    warp_home_path = pathlib.Path(__file__).parent.parent
+    warp_home = warp_home_path.resolve()
+
+    if args.verbose:
+        print(f"Building {dll_path}")
+
+    native_dir = os.path.join(warp_home, "native")
+
+    if cu_paths:
+        # check CUDA Toolkit version
+        ctk_version = get_cuda_toolkit_version(cuda_home)
+        if ctk_version < MIN_CTK_VERSION:
+            raise Exception(
+                f"CUDA Toolkit version {MIN_CTK_VERSION[0]}.{MIN_CTK_VERSION[1]}+ is required (found {ctk_version[0]}.{ctk_version[1]} in {cuda_home})"
+            )
+
+        # Get architecture flags based on CUDA version
+        if ctk_version >= (13, 0):
+            gencode_opts, clang_arch_flags = _get_architectures_cu13(ctk_version, arch, sys.platform, args.quick)
+        else:
+            gencode_opts, clang_arch_flags = _get_architectures_cu12(ctk_version, arch, sys.platform, args.quick)
+
+        nvcc_opts = [
+            *gencode_opts,
+            "-t0",  # multithreaded compilation
+            "--extended-lambda",
+        ]
+
+        # Clang options
+        clang_opts = [
+            *clang_arch_flags,
+            "-std=c++17",
+            "-xcuda",
+            f'--cuda-path="{cuda_home}"',
+        ]
+
+        if args.compile_time_trace:
+            if ctk_version >= (12, 8):
+                nvcc_opts.append("--fdevice-time-trace=_build/build_lib_@filename@_compile-time-trace")
+            else:
+                print("Warp warning: CUDA version is less than 12.8, compile_time_trace is not supported")
+
+        if args.fast_math:
+            nvcc_opts.append("--use_fast_math")
+
+    # is the library being built with CUDA enabled?
+    cuda_enabled = "WP_ENABLE_CUDA=1" if (cu_paths is not None) else "WP_ENABLE_CUDA=0"
+
+    if args.libmathdx_path:
+        libmathdx_includes = f' -I"{args.libmathdx_path}/include"'
+        mathdx_enabled = "WP_ENABLE_MATHDX=1"
+    else:
+        libmathdx_includes = ""
+        mathdx_enabled = "WP_ENABLE_MATHDX=0"
+
+    if os.name == "nt":
+        if args.host_compiler:
+            host_linker = os.path.join(os.path.dirname(args.host_compiler), "link.exe")
+        else:
+            raise RuntimeError("Warp build error: No host compiler was found")
+
+        cpp_includes = f' /I"{warp_home_path.parent}/external/llvm-project/out/install/{mode}-{arch}/include"'
+        cpp_includes += f' /I"{warp_home_path.parent}/_build/host-deps/llvm-project/release-{arch}/include"'
+        cuda_includes = f' /I"{cuda_home}/include"' if cu_paths else ""
+        includes = cpp_includes + cuda_includes
+
+        # nvrtc_static.lib is built with /MT and _ITERATOR_DEBUG_LEVEL=0 so if we link it in we must match these options
+        if cu_paths or mode != "debug":
+            runtime = "/MT"
+            iter_dbg = "_ITERATOR_DEBUG_LEVEL=0"
+            debug = "NDEBUG"
+        else:
+            runtime = "/MTd"
+            iter_dbg = "_ITERATOR_DEBUG_LEVEL=2"
+            debug = "_DEBUG"
+
+        cpp_flags = f'/nologo /std:c++17 /GR- {runtime} /D "{debug}" /D "{cuda_enabled}" /D "{mathdx_enabled}" /D "{cuda_compat_enabled}" /D "{iter_dbg}" /I"{native_dir}" {includes} '
+
+        if args.mode == "debug":
+            cpp_flags += "/FS /Zi /Od /D WP_ENABLE_DEBUG=1"
+            linkopts = ["/DLL", "/DEBUG"]
+        elif args.mode == "release":
+            cpp_flags += "/Ox /D WP_ENABLE_DEBUG=0"
+            linkopts = ["/DLL"]
+        else:
+            raise RuntimeError(f"Unrecognized build configuration (debug, release), got: {args.mode}")
+
+        if args.verify_fp:
+            cpp_flags += ' /D "WP_VERIFY_FP"'
+
+        if args.fast_math:
+            cpp_flags += " /fp:fast"
+
+        with concurrent.futures.ThreadPoolExecutor(max_workers=args.jobs) as executor:
+            futures, wall_clock = [], time.perf_counter_ns()
+
+            cpp_cmds = []
+            for cpp_path in cpp_paths:
+                cpp_out = cpp_path + ".obj"
+                linkopts.append(quote(cpp_out))
+                cpp_cmd = f'"{args.host_compiler}" {cpp_flags} -c "{cpp_path}" /Fo"{cpp_out}"'
+                cpp_cmds.append(cpp_cmd)
+
+            if args.jobs <= 1:
+                with ScopedTimer("build", active=args.verbose):
+                    for cpp_cmd in cpp_cmds:
+                        run_cmd(cpp_cmd)
+            else:
+                futures = [executor.submit(run_cmd, cmd=cpp_cmd) for cpp_cmd in cpp_cmds]
+
+            cuda_cmds = []
+            if cu_paths:
+                for cu_path in cu_paths:
+                    cu_out = cu_path + ".o"
+
+                    _nvcc_opts = [
+                        opt.replace("@filename@", os.path.basename(cu_path).replace(".", "_")) for opt in nvcc_opts
+                    ]
+
+                    if mode == "debug":
+                        cuda_cmd = f'"{cuda_home}/bin/nvcc" --std=c++17 --compiler-options=/MT,/Zi,/Od -g -G -O0 -DNDEBUG -D_ITERATOR_DEBUG_LEVEL=0 -I"{native_dir}" -line-info {" ".join(_nvcc_opts)} -DWP_ENABLE_CUDA=1 -D{mathdx_enabled} {libmathdx_includes} -o "{cu_out}" -c "{cu_path}"'
+                    elif mode == "release":
+                        cuda_cmd = f'"{cuda_home}/bin/nvcc" --std=c++17 -O3 {" ".join(_nvcc_opts)} -I"{native_dir}" -DNDEBUG -DWP_ENABLE_CUDA=1 -D{mathdx_enabled} {libmathdx_includes} -o "{cu_out}" -c "{cu_path}"'
+
+                    cuda_cmds.append(cuda_cmd)
+
+                    linkopts.append(quote(cu_out))
+
+                linkopts.append(
+                    f'cudart_static.lib nvrtc_static.lib nvrtc-builtins_static.lib nvptxcompiler_static.lib ws2_32.lib user32.lib /LIBPATH:"{cuda_home}/lib/x64"'
+                )
+
+                if args.libmathdx_path:
+                    linkopts.append(f'nvJitLink_static.lib /LIBPATH:"{args.libmathdx_path}/lib/x64" mathdx_static.lib')
+
+            if args.jobs <= 1:
+                with ScopedTimer("build_cuda", active=args.verbose):
+                    for cuda_cmd in cuda_cmds:
+                        run_cmd(cuda_cmd)
+            else:
+                futures.extend([executor.submit(run_cmd, cmd=cuda_cmd) for cuda_cmd in cuda_cmds])
+
+            if futures:
+                done, pending = concurrent.futures.wait(futures, return_when=concurrent.futures.FIRST_EXCEPTION)
+                for d in done:
+                    if e := d.exception():
+                        for f in pending:
+                            f.cancel()
+                        raise e
+                elapsed = (time.perf_counter_ns() - wall_clock) / 1000000.0
+                print(f"build took {elapsed:.2f} ms ({args.jobs:d} workers)")
+
+        with ScopedTimer("link", active=args.verbose):
+            link_cmd = f'"{host_linker}" {" ".join(linkopts + libs)} /out:"{dll_path}"'
+            run_cmd(link_cmd)
+
+    else:
+        # Unix compilation
+        cuda_compiler = "clang++" if getattr(args, "clang_build_toolchain", False) else "nvcc"
+        cpp_compiler = "clang++" if getattr(args, "clang_build_toolchain", False) else "g++"
+
+        cpp_includes = f' -I"{warp_home_path.parent}/external/llvm-project/out/install/{mode}-{arch}/include"'
+        cpp_includes += f' -I"{warp_home_path.parent}/_build/host-deps/llvm-project/release-{arch}/include"'
+        cuda_includes = f' -I"{cuda_home}/include"' if cu_paths else ""
+        includes = cpp_includes + cuda_includes
+
+        if sys.platform == "darwin":
+            version = f"--target={arch}-apple-macos11"
+        else:
+            if cpp_compiler == "g++":
+                version = "-fabi-version=13"  # GCC 8.2+
+            else:
+                version = ""
+
+        cpp_flags = f'-Werror -Wuninitialized {version} --std=c++17 -fno-rtti -D{cuda_enabled} -D{mathdx_enabled} -D{cuda_compat_enabled} -fPIC -fvisibility=hidden -D_GLIBCXX_USE_CXX11_ABI=0 -I"{native_dir}" {includes} '
+
+        if mode == "debug":
+            cpp_flags += "-O0 -g -D_DEBUG -DWP_ENABLE_DEBUG=1 -fkeep-inline-functions"
+
+        if mode == "release":
+            cpp_flags += "-O3 -DNDEBUG -DWP_ENABLE_DEBUG=0"
+
+        if args.verify_fp:
+            cpp_flags += " -DWP_VERIFY_FP"
+
+        if args.fast_math:
+            cpp_flags += " -ffast-math"
+
+        ld_inputs = []
+
+        with concurrent.futures.ThreadPoolExecutor(max_workers=args.jobs) as executor:
+            futures, wall_clock = [], time.perf_counter_ns()
+
+            cpp_cmds = []
+            for cpp_path in cpp_paths:
+                cpp_out = cpp_path + ".o"
+                ld_inputs.append(quote(cpp_out))
+                cpp_cmd = f'{cpp_compiler} {cpp_flags} -c "{cpp_path}" -o "{cpp_out}"'
+                cpp_cmds.append(cpp_cmd)
+
+            if args.jobs <= 1:
+                with ScopedTimer("build", active=args.verbose):
+                    for cpp_cmd in cpp_cmds:
+                        run_cmd(cpp_cmd)
+            else:
+                futures = [executor.submit(run_cmd, cmd=cpp_cmd) for cpp_cmd in cpp_cmds]
+
+            cuda_cmds = []
+            if cu_paths:
+                for cu_path in cu_paths:
+                    cu_out = cu_path + ".o"
+
+                    _nvcc_opts = [
+                        opt.replace("@filename@", os.path.basename(cu_path).replace(".", "_")) for opt in nvcc_opts
+                    ]
+
+                    if cuda_compiler == "nvcc":
+                        if mode == "debug":
+                            cuda_cmd = f'"{cuda_home}/bin/nvcc" --std=c++17 -g -G -O0 --compiler-options -fPIC,-fvisibility=hidden -D_DEBUG -D_ITERATOR_DEBUG_LEVEL=0 -line-info {" ".join(_nvcc_opts)} -DWP_ENABLE_CUDA=1 -I"{native_dir}" -D{mathdx_enabled} {libmathdx_includes} -o "{cu_out}" -c "{cu_path}"'
+                        elif mode == "release":
+                            cuda_cmd = f'"{cuda_home}/bin/nvcc" --std=c++17 -O3 --compiler-options -fPIC,-fvisibility=hidden {" ".join(_nvcc_opts)} -DNDEBUG -DWP_ENABLE_CUDA=1 -I"{native_dir}" -D{mathdx_enabled} {libmathdx_includes} -o "{cu_out}" -c "{cu_path}"'
+                    else:
+                        # Use Clang compiler
+                        if mode == "debug":
+                            cuda_cmd = f'clang++ -Werror -Wuninitialized -Wno-unknown-cuda-version {" ".join(clang_opts)} -g -O0 -fPIC -fvisibility=hidden -D_DEBUG -D_ITERATOR_DEBUG_LEVEL=0 -DWP_ENABLE_CUDA=1 -I"{native_dir}" -D{mathdx_enabled} {libmathdx_includes} -o "{cu_out}" -c "{cu_path}"'
+                        elif mode == "release":
+                            cuda_cmd = f'clang++ -Werror -Wuninitialized -Wno-unknown-cuda-version {" ".join(clang_opts)} -O3 -fPIC -fvisibility=hidden -DNDEBUG -DWP_ENABLE_CUDA=1 -I"{native_dir}" -D{mathdx_enabled} {libmathdx_includes} -o "{cu_out}" -c "{cu_path}"'
+
+                    cuda_cmds.append(cuda_cmd)
+
+                    ld_inputs.append(quote(cu_out))
+
+                ld_inputs.append(
+                    f'-L"{cuda_home}/lib64" -lcudart_static -lnvrtc_static -lnvrtc-builtins_static -lnvptxcompiler_static -lpthread -ldl -lrt'
+                )
+
+                if args.libmathdx_path:
+                    ld_inputs.append(f"-lnvJitLink_static -L{args.libmathdx_path}/lib -lmathdx_static")
+
+            if args.jobs <= 1:
+                with ScopedTimer("build_cuda", active=args.verbose):
+                    for cuda_cmd in cuda_cmds:
+                        run_cmd(cuda_cmd)
+            else:
+                futures.extend([executor.submit(run_cmd, cmd=cuda_cmd) for cuda_cmd in cuda_cmds])
+
+            if futures:
+                done, pending = concurrent.futures.wait(futures, return_when=concurrent.futures.FIRST_EXCEPTION)
+                for d in done:
+                    if e := d.exception():
+                        for f in pending:
+                            f.cancel()
+                        raise e
+                elapsed = (time.perf_counter_ns() - wall_clock) / 1000000.0
+                print(f"build took {elapsed:.2f} ms ({args.jobs:d} workers)")
+
+        if sys.platform == "darwin":
+            opt_no_undefined = "-Wl,-undefined,error"
+            opt_exclude_libs = ""
+        else:
+            opt_no_undefined = "-Wl,--no-undefined"
+            opt_exclude_libs = "-Wl,--exclude-libs,ALL"
+
+        with ScopedTimer("link", active=args.verbose):
+            origin = "@loader_path" if (sys.platform == "darwin") else "$ORIGIN"
+            link_cmd = f"{cpp_compiler} {version} -shared -Wl,-rpath,'{origin}' {opt_no_undefined} {opt_exclude_libs} -o '{dll_path}' {' '.join(ld_inputs + libs)}"
+            run_cmd(link_cmd)
+
+            # Strip symbols to reduce the binary size
+            if mode == "release":
+                if sys.platform == "darwin":
+                    run_cmd(f"strip -x {dll_path}")  # Strip all local symbols
+                else:  # Linux
+                    # Strip all symbols except for those needed to support debugging JIT-compiled code
+                    run_cmd(
+                        f"strip --strip-all --keep-symbol=__jit_debug_register_code --keep-symbol=__jit_debug_descriptor {dll_path}"
+                    )
+
+
+def build_dll(args, dll_path, cpp_paths, cu_paths, libs=None):
+    if sys.platform == "darwin":
+        # build for ARM64 only (may be cross-compiled from Intel Mac)
+        build_dll_for_arch(args, dll_path, cpp_paths, cu_paths, "aarch64", libs)
+    else:
+        build_dll_for_arch(args, dll_path, cpp_paths, cu_paths, machine_architecture(), libs)