metaxuda 0.1.0__tar.gz

metaxuda-0.1.0/PKG-INFO

@@ -0,0 +1,45 @@
+Metadata-Version: 2.4
+Name: metaxuda
+Version: 0.1.0
+Summary: CUDA→Metal universal backend shim (MetaXuda)
+Author-email: Perinban Parameshwaran <p.perinban@gmail.com>
+License: MIT
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Operating System :: MacOS
+Classifier: License :: OSI Approved :: MIT License
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+
+# MetaXuda
+
+MetaXuda is a **CUDA runtime shim for Apple Silicon**, written in Rust, that allows **Numba CUDA kernels** to run unmodified by mapping CUDA calls to Metal.
+
+---
+
+## ✨ Features
+- Drop-in replacement for `libcudart.dylib` / `libcuda.dylib`
+- Run Numba CUDA kernels (`@cuda.jit`) directly on Apple Metal
+- Includes precompiled Metal `.metallib` shaders for fused math ops
+- Ships with a stubbed `libdevice.bc` so no CUDA Toolkit is required
+
+---
+
+## ⚙️ Installation
+
+### Requirements
+- macOS 13+ with Apple Silicon (M1/M2/M3)
+- Python >=3.10
+- [NumPy](https://numpy.org/) (>=1.23)
+- [Numba](https://numba.pydata.org/) (>=0.59)
+
+### Steps
+```bash
+# Clone the repo
+git clone https://github.com/perinban/MetaXuda.git
+cd MetaXuda-
+
+# Back to project root
+cd ..
+pip install -e .

metaxuda-0.1.0/README.md

@@ -0,0 +1,31 @@
+# MetaXuda
+
+MetaXuda is a **CUDA runtime shim for Apple Silicon**, written in Rust, that allows **Numba CUDA kernels** to run unmodified by mapping CUDA calls to Metal.
+
+---
+
+## ✨ Features
+- Drop-in replacement for `libcudart.dylib` / `libcuda.dylib`
+- Run Numba CUDA kernels (`@cuda.jit`) directly on Apple Metal
+- Includes precompiled Metal `.metallib` shaders for fused math ops
+- Ships with a stubbed `libdevice.bc` so no CUDA Toolkit is required
+
+---
+
+## ⚙️ Installation
+
+### Requirements
+- macOS 13+ with Apple Silicon (M1/M2/M3)
+- Python >=3.10
+- [NumPy](https://numpy.org/) (>=1.23)
+- [Numba](https://numba.pydata.org/) (>=0.59)
+
+### Steps
+```bash
+# Clone the repo
+git clone https://github.com/perinban/MetaXuda.git
+cd MetaXuda-
+
+# Back to project root
+cd ..
+pip install -e .

metaxuda-0.1.0/metaxuda/__init__.py

@@ -0,0 +1,15 @@
+from .patch import patch_libdevice, get_libcudart_path
+from .buffer import GPUMemoryBuffer
+from .stream import GPUStream, DEFAULT_STREAM
+from .pipeline import run_pipeline
+
+# Apply patch at import
+patch_libdevice()
+
+__all__ = [
+    "GPUMemoryBuffer",
+    "GPUStream",
+    "DEFAULT_STREAM",
+    "run_pipeline",
+    "get_libcudart_path",
+]

metaxuda-0.1.0/metaxuda/buffer.py

@@ -0,0 +1,26 @@
+import numpy as np
+from numba import cuda
+
+class GPUMemoryBuffer:
+    """
+    Simple GPU buffer wrapper for host ↔ device transfers.
+    """
+
+    def __init__(self, arr: np.ndarray):
+        if not isinstance(arr, np.ndarray):
+            raise TypeError("GPUMemoryBuffer requires a NumPy array")
+        self.host = arr
+        self.device = cuda.to_device(arr)
+
+    def download(self):
+        """Copy GPU data back to a new NumPy array."""
+        out = np.empty_like(self.host)
+        self.device.copy_to_host(out)
+        return out
+
+    def upload(self, arr: np.ndarray):
+        """Copy new NumPy array into this buffer."""
+        if arr.shape != self.host.shape or arr.dtype != self.host.dtype:
+            raise ValueError("Shape and dtype must match the original array")
+        self.device.copy_to_device(arr)
+        self.host = arr

metaxuda-0.1.0/metaxuda/patch.py

@@ -0,0 +1,59 @@
+import os
+import numba.cuda.cuda_paths
+
+def setup_native_libs():
+    """
+    Ensure native libs exist:
+    - libdevice.bc (empty file)
+    - libnvvm.dylib (empty stub)
+    - libcuda.dylib (symlink -> libcudart.dylib)
+    """
+    base_dir = os.path.dirname(__file__)
+    native_dir = os.path.join(base_dir, "native")
+    os.makedirs(native_dir, exist_ok=True)
+
+    # libdevice.bc (always empty)
+    libdevice_path = os.path.join(native_dir, "libdevice.bc")
+    if not os.path.exists(libdevice_path):
+        with open(libdevice_path, "wb"):
+            pass
+
+    # libnvvm.dylib (empty stub)
+    libnvvm_path = os.path.join(native_dir, "libnvvm.dylib")
+    if not os.path.exists(libnvvm_path):
+        with open(libnvvm_path, "wb"):
+            pass
+
+    # libcudart.dylib (shim must be built and placed manually)
+    libcudart_path = os.path.join(native_dir, "libcudart.dylib")
+    if not os.path.exists(libcudart_path):
+        raise FileNotFoundError(
+            f"libcudart shim not found at {libcudart_path}. "
+            "Please place your built Rust dylib here."
+        )
+
+    # libcuda.dylib (symlink → libcudart.dylib)
+    libcuda_path = os.path.join(native_dir, "libcuda.dylib")
+    if not os.path.exists(libcuda_path):
+        os.symlink("libcudart.dylib", libcuda_path)
+
+    return native_dir, libdevice_path, libcudart_path
+
+def patch_libdevice():
+    """
+    Monkeypatch Numba to point libdevice.bc to shim folder.
+    """
+    _, libdevice_path, _ = setup_native_libs()
+
+    from collections import namedtuple
+    _env_path_tuple = namedtuple("_env_path_tuple", ["by", "info"])
+    numba.cuda.cuda_paths._get_libdevice_paths = (
+        lambda: _env_path_tuple("CUSTOM_MONKEYPATCH", libdevice_path)
+    )
+
+def get_libcudart_path():
+    """
+    Return the absolute path to libcudart shim (Rust-built dylib).
+    """
+    _, _, libcudart_path = setup_native_libs()
+    return libcudart_path

metaxuda-0.1.0/metaxuda/pipeline.py

@@ -0,0 +1,13 @@
+import numpy as np
+
+def run_pipeline(ops, arr: np.ndarray):
+    """
+    Run a sequence of operations on an array.
+    Currently CPU fallback — in the shim backend this will be GPU fused.
+    Example:
+        run_pipeline([np.sin, np.sqrt], x)
+    """
+    out = arr
+    for op in ops:
+        out = op(out)
+    return out

metaxuda-0.1.0/metaxuda/stream.py

@@ -0,0 +1,16 @@
+from numba import cuda
+
+class GPUStream:
+    """
+    Thin wrapper for cuda.stream.
+    """
+
+    def __init__(self):
+        self._stream = cuda.stream()
+
+    def sync(self):
+        """Synchronize the stream."""
+        self._stream.synchronize()
+
+# Default global stream
+DEFAULT_STREAM = GPUStream()

metaxuda-0.1.0/metaxuda.egg-info/PKG-INFO

@@ -0,0 +1,45 @@
+Metadata-Version: 2.4
+Name: metaxuda
+Version: 0.1.0
+Summary: CUDA→Metal universal backend shim (MetaXuda)
+Author-email: Perinban Parameshwaran <p.perinban@gmail.com>
+License: MIT
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Operating System :: MacOS
+Classifier: License :: OSI Approved :: MIT License
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+
+# MetaXuda
+
+MetaXuda is a **CUDA runtime shim for Apple Silicon**, written in Rust, that allows **Numba CUDA kernels** to run unmodified by mapping CUDA calls to Metal.
+
+---
+
+## ✨ Features
+- Drop-in replacement for `libcudart.dylib` / `libcuda.dylib`
+- Run Numba CUDA kernels (`@cuda.jit`) directly on Apple Metal
+- Includes precompiled Metal `.metallib` shaders for fused math ops
+- Ships with a stubbed `libdevice.bc` so no CUDA Toolkit is required
+
+---
+
+## ⚙️ Installation
+
+### Requirements
+- macOS 13+ with Apple Silicon (M1/M2/M3)
+- Python >=3.10
+- [NumPy](https://numpy.org/) (>=1.23)
+- [Numba](https://numba.pydata.org/) (>=0.59)
+
+### Steps
+```bash
+# Clone the repo
+git clone https://github.com/perinban/MetaXuda.git
+cd MetaXuda-
+
+# Back to project root
+cd ..
+pip install -e .

metaxuda-0.1.0/metaxuda.egg-info/SOURCES.txt

@@ -0,0 +1,15 @@
+README.md
+pyproject.toml
+metaxuda/__init__.py
+metaxuda/buffer.py
+metaxuda/patch.py
+metaxuda/pipeline.py
+metaxuda/stream.py
+metaxuda.egg-info/PKG-INFO
+metaxuda.egg-info/SOURCES.txt
+metaxuda.egg-info/dependency_links.txt
+metaxuda.egg-info/top_level.txt
+metaxuda/native/libcuda.dylib
+metaxuda/native/libcudart.dylib
+metaxuda/native/libdevice.bc
+metaxuda/native/libnvvm.dylib

metaxuda-0.1.0/metaxuda.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

metaxuda-0.1.0/metaxuda.egg-info/top_level.txt

@@ -0,0 +1 @@
+metaxuda

metaxuda-0.1.0/pyproject.toml

@@ -0,0 +1,33 @@
+[build-system]
+requires = ["setuptools>=61.0", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "metaxuda"
+version = "0.1.0"
+description = "CUDA→Metal universal backend shim (MetaXuda)"
+readme = "README.md"
+requires-python = ">=3.8"
+authors = [
+    { name = "Perinban Parameshwaran", email = "p.perinban@gmail.com" }
+]
+license = { text = "MIT" }
+
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Developers",
+    "Programming Language :: Python :: 3",
+    "Operating System :: MacOS",
+    "License :: OSI Approved :: MIT License"
+]
+
+[tool.setuptools]
+packages = ["metaxuda"]
+include-package-data = true
+
+[tool.setuptools.package-data]
+# ship your native libs & bc file
+"metaxuda" = [
+    "native/*.dylib",
+    "native/*.bc"
+]

metaxuda-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+